diff --git a/.github/README.md b/.github/README.md
index 2af7aa0..3700e64 100644
--- a/.github/README.md
+++ b/.github/README.md
@@ -15,7 +15,7 @@ Duke Alumnus, B.S. in Mathematics and M.Eng. in Risk Engineering · Sequoia Fell
## Selected Research
- **[Critical Dependence Stress for Diversification Gains under Fixed Marginals](https://www.theodoreoy.com/critical-dependence-stress-for-diversification-gains-under-fixed-marginals/)**
- How much dependence stress reverses a portfolio's Expected Shortfall advantage while every loss distribution stays fixed. The asymptotic reversal budget is at least 30% above the initial-sensitivity prediction.
+ How much dependence stress reverses a portfolio's Expected Shortfall advantage while every loss distribution stays fixed. In the uniform benchmark, the critical budget is at least 30% above the tangent prediction.
- **[Certified Valuation of Arithmetic Asian Options](https://www.theodoreoy.com/certified-valuation-of-arithmetic-asian-options/)**
What an arithmetic Asian option loses when exercise is decided on the geometric average of its fixings. The proved bound is below a hundredth of a cent per barrel for a one-month crude oil option, and under Heston it becomes a certified price interval.
- **[Certified Joint Pricing Errors in Rough Heston](https://www.theodoreoy.com/certified-joint-pricing-errors-in-rough-heston/)**
diff --git a/.github/workflows/pages.yml b/.github/workflows/pages.yml
index 5d29c58..20d37a6 100644
--- a/.github/workflows/pages.yml
+++ b/.github/workflows/pages.yml
@@ -11,7 +11,7 @@ permissions:
concurrency:
group: pages-${{ github.event_name == 'push' && 'production' || github.ref }}
- cancel-in-progress: true
+ cancel-in-progress: ${{ github.event_name != 'push' }}
jobs:
build:
diff --git a/architecture/README.md b/architecture/README.md
index 3a1d3bb..0397d45 100644
--- a/architecture/README.md
+++ b/architecture/README.md
@@ -48,7 +48,8 @@ npm --prefix website run check
The integration applies the site's type roles, a responsive reader layout, major-section
heading semantics, control labels, and export-font settings while preserving SVG
-geometry. It accepts both freshly generated and integrated viewers, and running it again
+geometry. It sets the site's palette with no bullet before the title or the card
+headings, and adds the site's Return Home link, with its return arrow, above the title. It accepts both freshly generated and integrated viewers, and running it again
leaves an integrated viewer unchanged.
The viewer loads `../assets/fonts/inter.css`, preloads regular Inter, and shares the
@@ -61,7 +62,7 @@ keep their SVG coordinate sizes.
SVG exports embed the local regular and italic Inter files with the same optical
settings; PNG and WebM exports render from those prepared fonts. After regeneration,
-check theme switching, search, tracing, keyboard controls, responsive layout, and
+check search, tracing, keyboard controls, responsive layout, and
export. The complete build also verifies the viewer's production security policy and
local assets.
diff --git a/architecture/index.html b/architecture/index.html
index 5109978..abb0c6e 100644
--- a/architecture/index.html
+++ b/architecture/index.html
@@ -4978,6 +4978,29 @@
.toolbar > button, .toolbar > .preset-wrap, .toolbar > .export-wrap { min-width: 0; }
.toolbar button { max-width: 100%; white-space: normal; }
.diagram-container { container: diagram / inline-size; }
+ .site-return {
+ display: inline-flex;
+ min-height: 44px;
+ align-items: center;
+ gap: 10px;
+ margin-bottom: 1rem;
+ color: var(--text);
+ font-size: var(--type-interface);
+ line-height: var(--interface-leading);
+ text-decoration: underline;
+ text-decoration-color: var(--text-dim);
+ text-decoration-thickness: 1px;
+ text-underline-offset: 0.18em;
+ }
+ .site-return svg {
+ flex: 0 0 16px;
+ width: 16px;
+ min-width: 0;
+ height: 16px;
+ transform: none;
+ transition: none;
+ }
+ html[data-embed="true"] .site-return, html[data-present="true"] .site-return { display: none; }
@container viewer (width < 60rem) {
html:not([data-embed="true"]):not([data-present="true"]) .toolbar {
@@ -5077,7 +5100,7 @@
}
body { background: var(--bg); }
#btn-theme, #btn-preset, .preset-wrap { display: none !important; }
- .pulse-dot, .card-dot { background: var(--text) !important; box-shadow: none !important; }
+ .pulse-dot, .card-dot { display: none !important; }
*:not(.pulse-dot):not(.motion-control-dot) { border-radius: 0 !important; }
@@ -5183,6 +5206,7 @@
• Clarinet hero, navigation, contents, article figures, table and figure zoom, and note previews
-
• Architecture search, themes, tracing, and export
+
• Architecture search, tracing, and export
diff --git a/architecture/site.json b/architecture/site.json
index 0859251..938f41d 100644
--- a/architecture/site.json
+++ b/architecture/site.json
@@ -251,7 +251,7 @@
"title": "Focused Interactions",
"items": [
"Clarinet hero, navigation, contents, article figures, table and figure zoom, and note previews",
- "Architecture search, themes, tracing, and export"
+ "Architecture search, tracing, and export"
]
}
]
diff --git a/docs/design.md b/docs/design.md
index 58b0e3a..c5c8652 100644
--- a/docs/design.md
+++ b/docs/design.md
@@ -96,9 +96,10 @@ prose to the reading track. Structural column guides are decorative, hidden from
assistive technology, and omitted on phones. A guide meets the content edge it marks at
every width; the closing guide on tablets sits on the last column's end. Every page
title starts 112px below the top of the content field on desktop and 52px below the
-header on phones. A page title keeps to four fifths of its column (`--title-width`);
-at 540px and below it takes the whole column, so a long title holds a few balanced
-lines on a phone.
+header on phones; the rail's wordmark starts at the same 112px (`--content-top`), so it
+sits on the title's line. A page title keeps to four fifths of its column
+(`--title-width`); below 1024px it takes the whole column, so a long title holds a few
+balanced lines on tablets and phones.
The rail's minimum width follows the interface type size, indent, padding, and
scrollbar allowance, so the rail, reading field, and guides reflow together when text
@@ -240,9 +241,10 @@ with its logic diagram under the record strip.
AI pages present the scope disclosure, Problem and Approach, Key Contributions, and the
numbered manuscript chapters. Selected Research headers hold the title, GitHub link, and
Full Manuscript link; reading begins at Abstract, and appendices and supporting material
-link to the pinned manuscript on GitHub. Chapter headings display Roman numerals and
-subsections letters in reading order, while source numbering and anchors stay intact;
-verified internal section references link to their headings. The three manuscripts link
+link to the pinned manuscript on GitHub. AI and Selected Research headings show the
+manuscript's own numbers (1., 2.1, 7.3.), so "Section 5", "(5.1)" and "Appendix A.5" in
+the text match what the reader sees; legal articles keep their manuscripts' Parts (I.,
+A.). Anchors stay intact, and verified internal section references link to their headings. The three manuscripts link
their own references alike: an equation by its parenthesized number, a statement, table,
or figure by its number. The dependence manuscript carries these links from its TeX
source; the Asian and Rough Heston manuscripts cite in plain text, and each reference to
@@ -306,7 +308,8 @@ Each AI and legal article opens with a `LogicDiagram`, its visual abstract, draw
`app/lib/content/architecture-diagrams.ts`. Four stages read left to right, each holding
one node or one node per lane, with an optional feedback loop to an earlier stage.
Stage labels sit on short rules with blue ticks; nodes are flat panels with a 1px
-border, and the stage holding the decisive mechanism is set in ink. Connectors are CSS
+border, and the stage holding the decisive mechanism is set in ink. A lane's name above a
+node title is in the label role of the diagram's one face. Connectors are CSS
on the same grid as the nodes; the loop is the one blue path, labeled beneath. Node
details accept inline TeX. Below 640px of figure width the diagram becomes a vertical
sequence (parallel nodes side by side from 480px) and the loop a labeled return note.
@@ -331,8 +334,10 @@ Text is HTML at its true size, so labels stay legible at every width; only curve
in a stretched SVG layer whose strokes keep their width, and schematics scale uniformly
so arrows and angles keep their shape. Labels sit in open space beside the marks they
name and never cover a curve. Bar rows align on one grid; in a narrow figure each label
-shares a line with its value above a full-width bar. A title stands alone on its line,
-with any view switch beneath it as one segmented control.
+and its value stand above a full-width bar, and the grid and reference rules run through
+the bar's own row only. A comparison wider than its column scrolls sideways and fades at
+its right edge, as a wide table does. A title stands alone on its line, with any view
+switch beneath it as one segmented control.
A segmented switch swaps views, a readout follows the pointer on line charts, marks
show exact values, the proof-obligation graph accepts a chosen failure, and each figure
@@ -483,4 +488,5 @@ keyboard use, enlarged text, reduced motion, and pages without JavaScript. Use t
shared tokens in `app/globals.css`, focused CSS Modules for local surfaces, and semantic
HTML, and match the existing type, measure, and spacing before adding a new pattern. The
[architecture viewer](../architecture/) is a technical reference drawn in the site's
-palette: paper, ink and hairlines, square corners, one theme.
+palette: paper, ink and hairlines, square corners, one theme, no bullets before titles,
+and the site's Return Home link above its title.
diff --git a/docs/maintenance.md b/docs/maintenance.md
index d758612..bdf27a9 100644
--- a/docs/maintenance.md
+++ b/docs/maintenance.md
@@ -41,7 +41,7 @@ Paths below are relative to `website/`.
| Experience | `content/experience-showcase.json`; AI fields are generated from the manuscript below | `app/lib/content/experience-showcase.ts` | `/experience/`, `/experience/[slug]/` |
| AI research | `content/sources/ai-portfolio/ai-research-and-engineering.md` | `scripts/import-ai-portfolio.mjs`, `app/lib/content/ai-projects.ts` | `/experience/[slug]/` |
| Selected Research | `content/sources/selected-research/` (pinned manuscripts), `content/selected-research/` (generated articles) | Research import scripts, `app/lib/content/selected-research.ts` | `/[paper]/` |
-| Legal research | `content/legal-papers/` structured manuscripts | `app/lib/content/legal-papers.ts` | `/experience/[slug]/` |
+| Legal research | `content/legal-papers/` structured manuscripts, generated from the LAW repositories | `app/lib/content/legal-papers.ts` | `/experience/[slug]/` |
| Blog | `content/blog/.md`, images in `public/assets/blog//` | `app/lib/content/blog.ts` | `/blog/`, `/blog/[slug]/` |
| Architecture | `../architecture/site.json` | Archify and `scripts/integrate-architecture.mjs` | `/architecture/` |
@@ -61,8 +61,8 @@ Every URL belongs to About (`/`, `/coursework/`, `/[paper]/`), Experience
or a detail page from its category. The export holds the content
routes (About, Coursework, three Selected Research articles, Experience and its eighteen
entries, Blog, and one page per post), the architecture viewer, and the 404 page. A work that a
-post presents is published on this domain at its own address named after the work, as
-The Étude Sky is at `/etude-sky/`.
+post presents is published on this domain at its repository's address, as
+The Étude Sky is at `/BLOG-Etude-Sky/`.
### Blog
@@ -116,7 +116,7 @@ The three papers use pinned research inputs and deterministic importers:
| Paper | Input and importer | On-site scope |
| --- | --- | --- |
-| Critical Dependence Stress for Diversification Gains under Fixed Marginals | `content/sources/selected-research/dependence/`; `scripts/import-dependence-research.mjs` | Abstract, chapters 1–5, references, and supporting-material links |
+| Critical Dependence Stress for Diversification Gains under Fixed Marginals | `content/sources/selected-research/dependence/`; `scripts/import-dependence-research.mjs` | Abstract, chapters 1–5 with their four figures, references, and supporting-material links |
| Certified Valuation of Arithmetic Asian Options | `content/sources/selected-research/asian-source.md` and the study results in `asian-studies/`; `scripts/import-selected-research.mjs` | Abstract, chapters 1–9 with their four figures, references, and supporting-material links |
| Certified Joint Pricing Errors in Rough Heston | `content/sources/selected-research/rough-source.md` and the saved banks in `rough-studies/`; `scripts/import-selected-research.mjs` | Abstract, chapters 1–8 with their four figures, references, and supporting-material links |
@@ -157,15 +157,20 @@ return to the corresponding About research entry.
### Legal research
`content/legal-papers/` preserves source titles, authors, dates, Abstracts, chapter
-hierarchies, paragraph runs, citations, acknowledgments, and numbered footnotes.
-The approved source PDFs are external inputs, not served downloads:
+hierarchies, paragraph runs, tables with their captions, citations, acknowledgments, and
+numbered footnotes. Each file is generated by `site_legal_sync.py` from the manuscript in
+its LAW repository, which is the authority for the text; the converter sets the
+footnotes' small-caps fields in true small caps:
-| Structured manuscript | Source PDF |
+| Structured manuscript | Source repository and manuscript |
| --- | --- |
-| `who-may-choose-the-lesser-risk.json` | `ESSAY1.pdf` |
-| `small-states-and-the-governance-of-strategic-space-dependence.json` | `ESSAY2.pdf` |
-| `autonomous-authority-in-space.json` | `ESSAY3.pdf` |
-| `mangrove-restoration-and-the-limits-of-compensatory-mitigation.json` | `ESSAY4.pdf` |
+| `autonomous-authority-in-space.json` | `130U/LAW-Autonomous-Authority-in-Space`, `manuscript/autonomous-authority-in-space.md` |
+| `small-states-and-the-governance-of-strategic-space-dependence.json` | `130U/LAW-Small-States-and-the-Governance-of-Strategic-Space-Dependence`, `manuscript/small-states-and-the-governance-of-strategic-space-dependence.md` |
+| `who-may-choose-the-lesser-risk.json` | `130U/LAW-Who-May-Choose-the-Lesser-Risk`, `manuscript/who-may-choose-the-lesser-risk.md` |
+| `mangrove-restoration-and-the-limits-of-compensatory-mitigation.json` | `130U/LAW-Mangrove-Restoration-and-the-Limits-of-Compensatory-Mitigation`, `manuscript/mangrove-restoration-and-the-limits-of-compensatory-mitigation.md` |
+
+Change wording in the repository and regenerate the file; then refresh
+`content/protected-sources.json` and the visible-copy manifest.
`PaperParagraph.text` must equal the concatenated run text. Preserve emphasis,
capitalization, superscripts, unique chapter anchors, and sequential notes.
@@ -194,7 +199,7 @@ Article figures live in `app/components/figures/`: `kit.tsx` holds the frame, sc
plot, schematic and bar primitives, `selected-research.tsx`, `ai-research.tsx` and
`legal-research.tsx` hold the figures, `dependence-manuscript.tsx` draws the dependence
manuscript's four figures from the formulas in `130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals`
-(`code/generate_mechanism_figures.py` and `calculations/plot_profiles_and_flows.py` at
+(`computation/code/generate_mechanism_figures.py` and `computation/calculations/plot_profiles_and_flows.py` at
the pinned commit), `asian-manuscript.tsx` draws the Asian manuscript's four figures
from its imported study results and from Proposition 4.1 as
`app/lib/content/asian-kink-bound.ts` evaluates it, `rough-manuscript.tsx` draws the
@@ -297,8 +302,9 @@ files for credentials, temporary files, and unintended generated output.
The workflow validates pull requests with read-only repository access. A push to
`main` runs validation and the dependency audit, uploads `website/out/`, and deploys
with Pages write and OIDC permissions confined to the deployment job. Every action
-is SHA-pinned. After publishing, verify the workflow for the exact commit and read
-back the live routes separately.
+is SHA-pinned. A newer pull-request run cancels the older one; production runs queue,
+so a deployment in progress always completes. After publishing, verify the workflow
+for the exact commit and read back the live routes separately.
Review the repository's Dependabot alerts, secret scanning, and push-protection
settings when maintaining a release. The workflow's dependency audit checks the
@@ -306,9 +312,12 @@ installed lockfile on every build.
Each exported HTML page receives hashes for its exact inline scripts before any
script executes. Production CSP blocks arbitrary inline scripts, eval, inline event
-handlers, embedded frames/objects, and form submission; fonts and network fetches
-use the same origin. Inline styles support framework and viewer geometry.
+handlers, embedded frames/objects, media, and form submission; fonts, images, and
+network fetches use the same origin, and only the architecture viewer adds the data and
+blob images its exports draw. Inline styles support framework and viewer geometry.
+Formulas are rendered at build time; a formula that sets a link, style, or class
+attribute stops the build.
GitHub Pages controls HTTP response headers: a meta policy cannot enforce
-`frame-ancestors` or supply host-level security headers. Keep the lockfile, pinned
-dependencies, and scoped overrides aligned. An advisory check is current only when
-the npm endpoint returns a successful result.
+`frame-ancestors` or supply host-level security headers. Keep the lockfile and pinned
+dependencies aligned. An advisory check is current only when the npm endpoint returns
+a successful result.
diff --git a/website/app/components/AboutContent.tsx b/website/app/components/AboutContent.tsx
index 1f1647d..c29a7eb 100644
--- a/website/app/components/AboutContent.tsx
+++ b/website/app/components/AboutContent.tsx
@@ -5,9 +5,9 @@ import styles from "../home.module.css";
const researchSummaries: Record = {
"critical-dependence-stress-for-diversification-gains-under-fixed-marginals":
- "How far can the dependence between positions move before a diversification trade stops reducing Expected Shortfall, when every individual loss distribution stays fixed? I prove that the critical stress budget is quadratic in the trade size and identify its leading coefficient through a variational limit. The sensitivity at zero stress understates that budget. I introduce the marginal-restoration premium to measure the gap, characterize exactly when it is positive, and prove that it is at least 30% in a uniform benchmark. At equal effective costs, continuous flow–cut duality gives the premium's response to background displacement costs.",
+ "How far can the dependence between positions move before a diversification trade stops reducing Expected Shortfall, when every individual loss distribution stays fixed? I prove that the critical stress budget is quadratic in the trade size and identify its leading coefficient through a variational limit. The sensitivity at zero stress can understate that budget. I introduce the marginal-restoration premium to measure the gap, characterize exactly when it is positive, and prove that it is at least 30% in a uniform benchmark. At equal effective costs, continuous flow–cut duality gives the premium's response to background displacement costs.",
"certified-valuation-of-arithmetic-asian-options":
- "What does an arithmetic Asian option lose when exercise is decided on the geometric average of its fixings? I prove that the loss is at most the dispersion of the fixings divided by the level risk that precedes them, and that in Black's model it grows with the cube of volatility and the square of the averaging window. For a one-month crude oil average price option the bound is below a hundredth of a cent per barrel, against 2.6 cents for the classical first-order bound. Under Heston, validated transforms turn the bound into a proved price interval for a twelve-fixing call spread and a proved bound on its Euler discretization error, with executable certificates.",
+ "What does an arithmetic Asian option lose when exercise is decided on the geometric average of its fixings? I prove that the loss is at most the dispersion of the fixings divided by the level risk that precedes them, and that in Black's model the bound grows with the cube of volatility and the square of the averaging window. For a one-month crude oil average price option the bound is below a hundredth of a cent per barrel, against 2.6 cents for the classical first-order bound. Under Heston, validated transforms turn the bound into a proved price interval for a twelve-fixing call spread and a proved bound on its Euler discretization error, with executable certificates.",
"certified-joint-pricing-errors-in-rough-heston":
"How far can stored option prices sit from the exact model prices, measured across a whole portfolio? Every strike at one expiry is priced from the same characteristic function, so its numerical error is a common factor. I prove that a portfolio's exposure to that error at each frequency is the Fourier transform of its payoff: spreads and butterflies hedge it, and baskets over many strikes diversify it. In rough Heston, a residual in the fractional Riccati equation moves the log-characteristic function by at most the residual integrated against a forward variance curve, with constant one. On S&P 500 options, netting the legs certifies a half-point tolerance for all 28 six-month spreads, butterflies and baskets tested, against 13 without.",
};
@@ -41,9 +41,9 @@ export function AboutContent({ hasHero = false }: { hasHero?: boolean }) {
developer, a frontier AI lab, and a global alternative asset
manager. I also build systems of my own:{" "}
a language model{" "}
- trained from scratch on financial text, and{" "}
- a gate{" "}
- that decides which evaluation feedback an agent's optimizer may act on.
+ trained from scratch for financial research, and{" "}
+ four gates{" "}
+ that decide which evaluation feedback may change an agent.
@@ -122,9 +122,10 @@ export function AboutContent({ hasHero = false }: { hasHero?: boolean }) {
-
+
05
+
Contact
Email
diff --git a/website/app/components/experience/LogicDiagram.module.css b/website/app/components/experience/LogicDiagram.module.css
index 609b7e8..6d944ed 100644
--- a/website/app/components/experience/LogicDiagram.module.css
+++ b/website/app/components/experience/LogicDiagram.module.css
@@ -79,12 +79,12 @@
flex: 0 0 auto;
}
+/* A lane's name is a label, in the diagram's one face. */
.tag {
color: var(--muted);
- font-family: var(--font-prose);
- font-size: var(--type-index);
- line-height: 1.4;
- letter-spacing: 0.02em;
+ font-family: var(--font-text);
+ font-size: var(--type-label);
+ line-height: var(--label-leading);
}
.nodeTitle {
diff --git a/website/app/components/figures/Figure.module.css b/website/app/components/figures/Figure.module.css
index d252d2d..1d65a49 100644
--- a/website/app/components/figures/Figure.module.css
+++ b/website/app/components/figures/Figure.module.css
@@ -48,16 +48,11 @@
outline-offset: 3px;
}
-.frame [data-figure-scroll][data-scroll-overflow]::after {
- position: sticky;
- left: 0;
- display: block;
- width: 100%;
- padding-top: 10px;
- color: var(--supporting-ink);
- font-size: var(--type-label);
- line-height: var(--label-leading);
- content: "Scroll to compare →";
+/* A comparison wider than its column fades at the right edge by as much as is
+ left to scroll, as a wide table does. */
+.frame [data-figure-scroll][data-scroll-overflow] {
+ -webkit-mask-image: linear-gradient(to right, #000 calc(100% - var(--edge-fade, 0px)), transparent);
+ mask-image: linear-gradient(to right, #000 calc(100% - var(--edge-fade, 0px)), transparent);
}
/* The title sits alone on its line; a view switch follows it as one segmented
@@ -814,8 +809,10 @@
grid-template-rows: auto auto auto;
}
+ /* The grid and the reference rules run through the bar's own row only, so
+ they never cross a label or a value. */
.bars[data-stack] > .row > .lines {
- grid-row: 1 / 4;
+ grid-row: 3;
}
.bars[data-stack] > .row > .barLabel {
diff --git a/website/app/components/figures/FigureController.tsx b/website/app/components/figures/FigureController.tsx
index 4a9c2d3..43fb84c 100644
--- a/website/app/components/figures/FigureController.tsx
+++ b/website/app/components/figures/FigureController.tsx
@@ -24,10 +24,16 @@ export function enhanceFigure(figure: HTMLElement): () => void {
const frame = figure.firstElementChild as HTMLElement;
const scale = () => frame.getBoundingClientRect().width / frame.offsetWidth || 1;
- // Make a wide comparison's sideways continuation visible on small screens.
+ // Make a wide comparison's sideways continuation visible on small screens:
+ // its right edge fades by as much as is left to scroll.
for (const region of figure.querySelectorAll("[data-figure-scroll]")) {
- const update = () => region.toggleAttribute("data-scroll-overflow", region.scrollWidth > region.clientWidth + 2);
+ const update = () => {
+ const remaining = region.scrollWidth - region.clientWidth - region.scrollLeft;
+ region.toggleAttribute("data-scroll-overflow", region.scrollWidth > region.clientWidth + 2);
+ region.style.setProperty("--edge-fade", `${Math.max(0, Math.min(40, remaining))}px`);
+ };
update();
+ listen(region, "scroll", update);
const observer = new ResizeObserver(update);
observer.observe(region);
if (region.firstElementChild) observer.observe(region.firstElementChild);
diff --git a/website/app/components/figures/ai-research.tsx b/website/app/components/figures/ai-research.tsx
index a675a0d..b43734f 100644
--- a/website/app/components/figures/ai-research.tsx
+++ b/website/app/components/figures/ai-research.tsx
@@ -363,7 +363,7 @@ function ProofObligations() {
const PROTOCOLS = [
{ name: "Final-answer matching only", parts: [1200, 0, 1800, 600], share: "40.0%" },
{ name: "Strict reference-path matching", parts: [600, 600, 0, 2400], share: "100.0%" },
- { name: "Semantic certificate verification", parts: [1200, 0, 0, 2400], share: "100.0%" },
+ { name: "Certificate verification", parts: [1200, 0, 0, 2400], share: "100.0%" },
] as const;
const OUTCOMES = [
{ name: "Valid, accepted", tone: "ink" },
@@ -380,10 +380,10 @@ function ProtocolOutcomes() {
({ label: outcome.name, tone: outcome.tone }))}
>
-
+
{PROTOCOLS.map((protocol) => {
let start = 0;
@@ -491,7 +491,7 @@ const SCENARIO_VIEWS = [
function RegretScenarios() {
const x = linear(-0.72, 0.72);
- const scenarios = [-1, 1].flatMap((s) => [-0.4, 0.4].flatMap((z1) => [-0.2, 0.2].map((z2) => ({ s, z1, z2, delta: Math.round((0.15 * s + z1 + 0.3 * s * z2) * 100) / 100 }))));
+ const scenarios = [-1, 1].flatMap((theta) => [-0.4, 0.4].flatMap((z1) => [-0.2, 0.2].map((z2) => ({ theta, z1, z2, delta: Math.round((0.15 * theta + z1 + 0.3 * theta * z2) * 100) / 100 }))));
const lanes: Record> = {
none: [{ top: 56, label: "All scenarios", regret: "Regret 0.61" }],
shap: [{ top: 56, label: "Unchanged", regret: "Regret 0.61" }],
@@ -509,8 +509,8 @@ function RegretScenarios() {
0$: company $i$ scores higher", tone: "accent", shape: "dot" }, { label: "$\\Delta<0$: company $k$ scores higher", tone: "ink", shape: "ring" }]}
@@ -538,10 +538,10 @@ function RegretScenarios() {
} as CSSProperties;
return (
0 ? kit.accent : kit.ink}`}
data-sign={scenario.delta > 0 ? "positive" : "negative"}
- data-tip={`s = ${scenario.s}, z₁ = ${scenario.z1}, z₂ = ${scenario.z2}\nΔ = ${fixed(scenario.delta, 2)}`}
+ data-tip={`θ = ${scenario.theta}, z₁ = ${scenario.z1}, z₂ = ${scenario.z2}\nΔ = ${fixed(scenario.delta, 2)}`}
style={style}
/>
);
diff --git a/website/app/components/figures/dependence-manuscript.tsx b/website/app/components/figures/dependence-manuscript.tsx
index 040ca98..958fbf8 100644
--- a/website/app/components/figures/dependence-manuscript.tsx
+++ b/website/app/components/figures/dependence-manuscript.tsx
@@ -4,8 +4,8 @@ import styles from "./Figures.module.css";
/* The four figures of "Critical Dependence Stress", drawn natively from the
formulas that produced the manuscript images: Figures 1 and 2 follow
- code/generate_mechanism_figures.py, Figures 3 and 4 follow
- calculations/plot_profiles_and_flows.py (130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals
+ computation/code/generate_mechanism_figures.py, Figures 3 and 4 follow
+ computation/calculations/plot_profiles_and_flows.py (130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals
at the pinned source commit). Captions stay the manuscript's own. */
export type ManuscriptFigureProps = {
diff --git a/website/app/components/figures/rough-manuscript.tsx b/website/app/components/figures/rough-manuscript.tsx
index 8878347..b7a8182 100644
--- a/website/app/components/figures/rough-manuscript.tsx
+++ b/website/app/components/figures/rough-manuscript.tsx
@@ -45,7 +45,7 @@ function ResidualKernelFigure({ manuscript, caption }: ManuscriptFigureProps) {
]}
>
* {
@@ -481,14 +484,6 @@ button:focus-visible {
text-wrap: balance;
}
-.section-kicker {
- margin: 0 0 24px;
- color: var(--muted);
- font-size: var(--type-label);
- line-height: var(--label-leading);
- letter-spacing: 0.015em;
-}
-
.lede {
margin: 27px 0 0;
color: var(--ink);
@@ -943,11 +938,13 @@ button:focus-visible {
}
}
-@media (width <= 540px) {
+@media (width < 1024px) {
:root {
--title-width: 100%;
}
+}
+@media (width <= 540px) {
.site-footer {
display: grid;
gap: 4px;
diff --git a/website/app/home.module.css b/website/app/home.module.css
index f3891df..a391902 100644
--- a/website/app/home.module.css
+++ b/website/app/home.module.css
@@ -8,7 +8,7 @@
.homeFrame .profileSection {
grid-column: 1 / span 11;
- padding-top: 176px;
+ padding-top: var(--content-top);
}
.identity,
diff --git a/website/app/lib/content/architecture-diagrams.ts b/website/app/lib/content/architecture-diagrams.ts
index 25b8241..039dbb1 100644
--- a/website/app/lib/content/architecture-diagrams.ts
+++ b/website/app/lib/content/architecture-diagrams.ts
@@ -26,14 +26,14 @@ export const architectureDiagrams: Readonly>
label: "Allocate",
emphasis: true,
nodes: [
- { tag: "Measure", title: "Reference audit", detail: "Random audits sampled by information per cost, $m_i / c_i$" },
- { tag: "Correct", title: "Review allocation", detail: "Ranked by removable loss $\\Delta_i$ under capacity $H_j$" },
+ { tag: "Measure", title: "Reference audit", detail: "$\\pi_i^* = \\mathrm{clip}\\,\\sqrt{\\nu m_i/(\\lambda c_i - r_i)_+}$: risk is a rebate on the audit's price" },
+ { tag: "Correct", title: "Review allocation", detail: "Ranked by removable loss $\\Delta_i$ under capacity $H_j$, at the same budget price $\\lambda$" },
],
},
{
label: "Decide",
nodes: [
- { title: "Prediction-assisted estimate", detail: "Corrects the selection bias $\\operatorname{Cov}(\\pi, q) > 0$" },
+ { title: "Prediction-assisted estimate", detail: "Corrects the selection bias $\\mathrm{Cov}(\\pi, q) > 0$" },
{ title: "Expert adjudication", detail: "Tiered review from Beta-Binomial reliability posteriors" },
],
},
@@ -46,7 +46,7 @@ export const architectureDiagrams: Readonly>
},
],
loop: { from: 2, to: 1, label: "Adjudications update reliability posteriors and review intensity" },
- caption: "Measurement and correction run as separate decisions over the same submissions. Adjudications update reliability posteriors, acceptance criteria and review intensity.",
+ caption: "One rule prices measurement and correction against the same budget: an audited defect ships corrected, so an item's risk lowers the price of auditing it. Adjudications update reliability posteriors, acceptance criteria and review intensity.",
},
2: {
label: SYSTEM,
@@ -56,13 +56,13 @@ export const architectureDiagrams: Readonly>
label: "Generate",
nodes: [
{ title: "Gaussian and vine copulas", detail: "Corporate fundamentals keep their joint dependence" },
- { title: "Constrained projection", detail: "Accounting identities, ratio bounds and date order hold" },
- { title: "CTGAN and TabDDPM", detail: "Mixed records pass business rules and margin $\\gamma/s$ checks" },
+ { title: "Constrained sampling", detail: "Identities by projection; ratio bounds and date order by transformation" },
+ { title: "CTGAN and TabDDPM", detail: "Mixed-type records pass mandatory business-rule checks" },
{ title: "SDEs and Brownian bridges", detail: "Price and rate paths respect the filtration $\\mathcal{F}_s$" },
],
},
{ label: "Calibrate", emphasis: true, nodes: [{ title: "Batch Zero", detail: "Independent execution, reference solutions and five-level pairwise preferences" }] },
- { label: "Scale", nodes: [{ title: "Batch One", detail: "Large-scale data construction for RLHF and DPO supervision" }] },
+ { label: "Scale", nodes: [{ title: "Batch One", detail: "Variants below $z_{1-\\alpha}$ in $\\gamma/s$ get a recomputed reference; large-scale construction for RLHF and DPO" }] },
],
loop: { from: 2, to: 0, label: "Failures and disagreements revise task specifications" },
caption: "Batch Zero gates each task family before large-scale data construction. Failures and disagreements revise task instructions, reference solutions and acceptance criteria.",
@@ -97,16 +97,16 @@ export const architectureDiagrams: Readonly>
{ label: "Evidence", nodes: [{ title: "Versioned evidence", detail: "Reports, audited statements, disclosures and questionnaires with source records" }] },
{ label: "Score", nodes: [{ title: "Rashomon scoring", detail: "Near-optimal models $\\mathcal{F}_\\alpha$ with SHAP attributions across the set" }] },
{ label: "Bound", nodes: [{ title: "Ranking bounds", detail: "Support functions over the joint uncertainty set $\\Omega(E)$" }] },
- { label: "Query", emphasis: true, nodes: [{ title: "Verification query", detail: "Largest guaranteed cut in worst-case regret per cost, $V(q)/c_q$" }] },
+ { label: "Plan", emphasis: true, nodes: [{ title: "Verification plan", detail: "Knapsack on verified swing per analyst hour, $(1-\\lambda_j)s_j/c_j$" }] },
],
loop: { from: 3, to: 0, label: "Each correction triggers rescoring and attribution updates" },
- caption: "Minimax regret picks the next fact to verify, and each correction triggers rescoring and attribution updates through versioned evidence.",
+ caption: "Minimax regret sets the verification plan, a knapsack on verified swing per analyst hour, and each correction triggers rescoring and attribution updates through versioned evidence.",
},
8: {
label: FRAMEWORK,
stages: [
{ label: "Task", nodes: [{ title: "Autonomous space task", detail: "Delayed spacecraft operations, orbital avoidance, debris capture, SETI" }] },
- { label: "Test", emphasis: true, nodes: [{ title: "Two-threshold test", detail: "Authority to act, then evidence to delegate while human review can still change the outcome" }] },
+ { label: "Test", emphasis: true, nodes: [{ title: "Two-threshold test", detail: "Authority to act, then evidence to delegate, then the point at which human review can still change the outcome" }] },
{
label: "Supervise",
nodes: [
@@ -157,7 +157,7 @@ export const architectureDiagrams: Readonly>
],
},
{ label: "Justify", emphasis: true, nodes: [{ title: "Comparative justification", detail: "Defend alternatives, disclose distributional choices, answer objections from exposed constituencies" }] },
- { label: "Authorize", nodes: [{ title: "Authorization", detail: "Assessment, notification and consultation scaled to expected effects" }] },
+ { label: "Authorize", nodes: [{ title: "Authorization", detail: "Review council and independent science panel; separate concurrence of sponsors and exposed constituencies" }] },
{
label: "Continue",
nodes: [
@@ -168,7 +168,7 @@ export const architectureDiagrams: Readonly>
},
],
loop: { from: 3, to: 0, label: "Periodic reassessment of intervention and nondeployment" },
- caption: "The comparison duty ties risk-risk analysis to authorization. The continuity duty keeps program control, operator succession and the cooling trajectory apart, so a change in political authority does not force abrupt cessation.",
+ caption: "The comparative justification duty ties risk-risk analysis to authorization. The continuity duty keeps program control, operator succession and the cooling trajectory apart, so a change in political authority does not force abrupt cessation.",
},
11: {
label: FRAMEWORK,
@@ -181,7 +181,7 @@ export const architectureDiagrams: Readonly>
{ title: "Compensation for new loss" },
],
},
- { label: "Evidence", nodes: [{ title: "Functional replacement", detail: "Additionality and timing, not canopy growth or construction alone; West Lake, Deep Bay, San Tin" }] },
+ { label: "Evidence", nodes: [{ title: "Functional replacement", detail: "Additionality, timing and a discount for delay and risk, not canopy growth or construction alone; West Lake, Deep Bay, San Tin" }] },
{ label: "Accept", emphasis: true, nodes: [{ title: "Staged acceptance", detail: "Site baselines, functional objectives, monitoring periods, seasonal coverage" }] },
{
label: "Decide",
@@ -193,7 +193,7 @@ export const architectureDiagrams: Readonly>
},
],
loop: { from: 2, to: 1, label: "Delayed or failed recovery triggers funded monitoring and correction" },
- caption: "Engineering completion and ecological performance are accepted separately. Only functions that have been demonstrated can release funds, discharge responsibility or permit the next phase of development.",
+ caption: "Engineering completion and ecological performance are accepted separately: verified works earn payment, demonstrated functions discharge responsibility, and a promised function is discounted for delay and risk and credited in full only when it exists before the loss.",
},
};
diff --git a/website/app/lib/content/article-figures.ts b/website/app/lib/content/article-figures.ts
index 2d0f2ce..a125398 100644
--- a/website/app/lib/content/article-figures.ts
+++ b/website/app/lib/content/article-figures.ts
@@ -95,7 +95,7 @@ const articleFigures: Readonly> = {
"| Residual- and cost-aware (Section 3) | 7.89% / 19.37% / 21.84% | **0.515** | **-17.8%** |",
"| Share of the batch | 60% | 30% | 10% |",
"| Cost of an independent reference audit ($`c`$, routine = 1) | 1 | 1.5 | 3 |",
- "The audit budget is $`2,000`$ cost units",
+ "The audit budget is 2,000 cost units",
"a sampling floor of $`\\pi_{\\min} = 3\\%`$",
],
},
@@ -171,7 +171,7 @@ const articleFigures: Readonly> = {
after: { paragraph: "• Set the shared bias" },
evidence: [
"$`b = 0.2`$ and independent noise to $`\\sigma = 1`$",
- "48.40% at $`n = 100`$ records and 0.000637% at $`n = 1,000`$",
+ "48.40% at $`n = 100`$ records and 0.000637% at $`n = 1{,}000`$",
"$`0.2 + 1.96/\\sqrt{n}`$ gives 97.50% coverage",
],
},
@@ -184,9 +184,9 @@ const articleFigures: Readonly> = {
figure: "regret-scenarios",
after: { paragraph: "Enumerating the eight scenarios" },
evidence: [
- "$`s \\in \\{-1, 1\\}`$",
+ "$`\\theta \\in \\{-1, 1\\}`$",
"$`z_1 \\in \\{-0.4, 0.4\\}`$, $`z_2 \\in \\{-0.2, 0.2\\}`$",
- "$`\\Delta_s = 0.15s + z_1 + 0.3s z_2`$",
+ "$`\\Delta_\\theta = 0.15\\theta + z_1 + 0.3\\theta z_2`$",
"| 0.61 | 0 |",
"| 0 | 0.61 |",
],
@@ -265,8 +265,8 @@ const tableEmphasis: Readonly> = {
// Manuscript figures drawn natively, keyed by the image path the manuscript
// cites. The dependence drawings evaluate the manuscript's formulas
// (130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals at the pinned commit:
-// code/generate_mechanism_figures.py and
-// calculations/plot_profiles_and_flows.py); the Asian drawings plot the study
+// computation/code/generate_mechanism_figures.py and
+// computation/calculations/plot_profiles_and_flows.py); the Asian drawings plot the study
// results of 130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options at the pinned
// commit and evaluate Proposition 4.1; the Rough Heston drawings plot the saved
// banks of 130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston at the pinned commit and
@@ -328,7 +328,7 @@ export const manuscriptFigures: Readonly 3999) return undefined;
- let remaining = number;
- let result = "";
- for (const [value, label] of [[1000, "M"], [900, "CM"], [500, "D"], [400, "CD"], [100, "C"], [90, "XC"], [50, "L"], [40, "XL"], [10, "X"], [9, "IX"], [5, "V"], [4, "IV"], [1, "I"]] as const) {
- while (remaining >= value) { result += label; remaining -= value; }
- }
- return result;
-}
-
-function letter(number: number) {
- if (!Number.isSafeInteger(number) || number < 1 || number > 26) return undefined;
- return String.fromCharCode(64 + number);
-}
-
/** A heading's label in the article contents, where every heading reads as plain text. */
function inlineText(inlines: readonly ArticleInline[]): string {
return inlines.map((inline) => {
@@ -41,52 +25,35 @@ function inlineText(inlines: readonly ArticleInline[]): string {
}).join("");
}
-/** Present existing section identities without renumbering the source or changing its anchors. */
+/**
+ * Present a heading with the manuscript's own number, so that "Section 5",
+ * "(5.1)" and "Appendix A.5" in the text match what the reader sees. A chapter
+ * reads "5." as the papers and the repositories' manuscripts write it; a
+ * subsection keeps the source's form ("5.1" or "5.1."). The source and its
+ * anchors are never renumbered.
+ */
export function formatArticleHeading(heading: ResearchHeading): ArticleHeadingPresentation {
const unchanged = { id: heading.id, displayInlines: heading.inlines, text: inlineText(heading.inlines) };
const first = heading.inlines[0];
if (first?.type !== "text") return unchanged;
- const prefix = first.text.match(heading.level === 2 ? /^([1-9]\d*)\.?\s+/u : /^([1-9]\d*)\.([1-9]\d*)\.?\s+/u);
+ const prefix = first.text.match(heading.level === 2 ? /^([1-9]\d*)\.?\s+/u : /^([1-9]\d*)\.([1-9]\d*)(\.?)\s+/u);
if (!prefix) return unchanged;
- const chapter = roman(Number(prefix[1]));
- const subsection = heading.level === 3 ? letter(Number(prefix[2])) : undefined;
- if (!chapter || (heading.level === 3 && !subsection)) return unchanged;
const sourceNumber = heading.level === 3 ? `${prefix[1]}.${prefix[2]}` : prefix[1];
- const displayNumber = `${subsection ?? chapter}.`;
+ const displayNumber = heading.level === 3 ? `${sourceNumber}${prefix[3]}` : `${sourceNumber}.`;
const displayInlines: ArticleInline[] = [{ type: "text", text: `${displayNumber} ${first.text.slice(prefix[0].length)}` }, ...heading.inlines.slice(1)];
- return {
- id: heading.id, displayInlines, text: inlineText(displayInlines), sourceNumber, displayNumber,
- referenceLabel: subsection ? `${chapter}.${subsection}.` : `${chapter}.`,
- };
+ return { id: heading.id, displayInlines, text: inlineText(displayInlines), sourceNumber, displayNumber };
}
-/** Assign letters in page order, including subsections the manuscript leaves unnumbered. */
+/** Present every heading of an article, checking that each numbered subsection belongs to its chapter. */
export function formatArticleHeadings(blocks: readonly ResearchBlock[]) {
const headings = new Map();
let chapter: ArticleHeadingPresentation | undefined;
- let subsection = 0;
for (const block of blocks) {
if (block.type !== "heading") continue;
- let heading = formatArticleHeading(block);
- if (block.level === 2) {
- chapter = heading;
- subsection = 0;
- } else {
- subsection += 1;
- const label = letter(subsection);
- if (!label) throw new Error("Article subsection numbering exceeds the reviewed A–Z range.");
- if (heading.sourceNumber && heading.sourceNumber.split(".")[0] !== chapter?.sourceNumber) {
- throw new Error(`Source subsection ${heading.sourceNumber} does not belong to its preceding chapter.`);
- }
- const displayNumber = `${label}.`;
- const first = block.inlines[0];
- const displayInlines: ArticleInline[] = first?.type === "text"
- ? [{ type: "text", text: `${displayNumber} ${heading.sourceNumber ? first.text.replace(/^[1-9]\d*\.[1-9]\d*\.?\s+/u, "") : first.text}` }, ...block.inlines.slice(1)]
- : [{ type: "text", text: `${displayNumber} ` }, ...block.inlines];
- heading = {
- ...heading, displayInlines, text: inlineText(displayInlines), displayNumber,
- referenceLabel: chapter?.displayNumber ? `${chapter.displayNumber}${displayNumber}` : displayNumber,
- };
+ const heading = formatArticleHeading(block);
+ if (block.level === 2) chapter = heading;
+ else if (heading.sourceNumber && heading.sourceNumber.split(".")[0] !== chapter?.sourceNumber) {
+ throw new Error(`Source subsection ${heading.sourceNumber} does not belong to its preceding chapter.`);
}
if (headings.has(heading.id)) throw new Error(`Duplicate heading anchor: ${heading.id}`);
headings.set(heading.id, heading);
@@ -190,7 +157,7 @@ export function buildReferenceTargets(slug: string, blocks: readonly ResearchBlo
return { targets, anchors };
}
-type Reference = { start: number; end: number; href: string; title?: string };
+type Reference = { start: number; end: number; href: string };
/**
* Link audited internal Section references and, in a cross-referenced paper,
@@ -218,8 +185,7 @@ export function linkArticleSectionReferences(
const numberOffset = reference.index + reference[0].indexOf(reference[1]);
for (const number of numbers) {
const from = start + numberOffset + number.index;
- const heading = headings.get(number[0])!;
- links.push({ start: from, end: from + number[0].length, href: `#${heading.id}`, title: `Section ${heading.sourceNumber}; web heading ${heading.referenceLabel}` });
+ links.push({ start: from, end: from + number[0].length, href: `#${headings.get(number[0])!.id}` });
}
}
offset = start + context.length;
@@ -247,7 +213,7 @@ export function linkArticleSectionReferences(
for (const link of links) {
if (link.start < offset) continue;
if (link.start > offset) result.push({ type: "text", text: text.slice(offset, link.start) });
- result.push({ type: "link", href: link.href, ...(link.title ? { title: link.title } : {}), inlines: [{ type: "text", text: text.slice(link.start, link.end) }] });
+ result.push({ type: "link", href: link.href, inlines: [{ type: "text", text: text.slice(link.start, link.end) }] });
offset = link.end;
}
if (offset < text.length || !result.length) result.push({ type: "text", text: text.slice(offset) });
diff --git a/website/app/lib/content/research-math.ts b/website/app/lib/content/research-math.ts
index bfa57bc..275c98c 100644
--- a/website/app/lib/content/research-math.ts
+++ b/website/app/lib/content/research-math.ts
@@ -214,10 +214,16 @@ function intrinsicSvg(container: LiteElement, display: boolean) {
return { svg, widthEm, heightEm, depthEm };
}
+// Attributes that \mmlToken could carry into the page's markup; mathematics never needs them.
+const markupAttributes = new Set(["href", "style", "class"]);
+
function assistiveMathml(tex: string, display: boolean) {
const root = output.math.root;
root.walkTree((node) => {
if (node.kind === "merror") throw new Error("Research equation contains a MathML error.");
+ if (node.attributes?.getExplicitNames().some((name) => markupAttributes.has(name))) {
+ throw new Error("Research equation sets a markup attribute.");
+ }
node.attributes?.unset("data-latex");
node.attributes?.unset("data-latex-item");
node.attributes?.unset("id");
diff --git a/website/app/lib/content/selected-research.ts b/website/app/lib/content/selected-research.ts
index 17d820e..a5992f8 100644
--- a/website/app/lib/content/selected-research.ts
+++ b/website/app/lib/content/selected-research.ts
@@ -21,7 +21,7 @@ function readResearch(slug: string, repository = slug) {
const selectedResearch = readResearch("certified-valuation-of-arithmetic-asian-options", "RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options");
const roughHestonResearch = {
...readResearch("certified-joint-pricing-errors-in-rough-heston", "RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston"),
- sourceUrl: "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/V_20261101/ARTICLE.md",
+ sourceUrl: "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/V_202611/ARTICLE.md",
};
const dependenceResearch = readResearch("critical-dependence-stress-for-diversification-gains-under-fixed-marginals", "RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals");
export const selectedResearchProjects = [dependenceResearch, selectedResearch, roughHestonResearch];
diff --git a/website/app/not-found.tsx b/website/app/not-found.tsx
index 774331f..f98072f 100644
--- a/website/app/not-found.tsx
+++ b/website/app/not-found.tsx
@@ -14,7 +14,6 @@ export default function NotFound() {
return (
-
404
Page Not Found
The page you requested does not exist.
diff --git a/website/content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json b/website/content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json
index fd2977a..bcc8c4a 100644
--- a/website/content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json
+++ b/website/content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json
@@ -3,11 +3,11 @@
"title": "Evidence Verification and Robust Comparison in Investment Research",
"sourceUrl": "https://github.com/130U/RESEARCH-Evidence-Verification-and-Robust-Comparison-in-Investment-Research/blob/main/manuscript/evidence-verification-and-robust-comparison-in-investment-research.md",
"sourceEditedAt": "2026-11-19",
- "markdown": "## 1 Feature attribution and the value of verification\n\n[SHAP](https://arxiv.org/abs/1705.07874) splits a model prediction into additive feature contributions through its local accuracy property. Conditional Shapley methods also account for dependence among features, and Rashomon-set methods check how stable explanations are across near-optimal models ([SHAP](https://arxiv.org/abs/1705.07874); [Aas et al.](https://arxiv.org/html/1903.10464v3); [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html)). All of them take the corporate inputs $`x`$ as given. In investment diligence the inputs are the open question, because self-reported disclosures can be polished well past what the underlying facts support, and an analyst needs to know which corrections to the evidence could change a decision. Attribution was not built for that question, and there is now a theorem to the effect: for expressive model classes, attribution methods that are complete and linear, SHAP among them, can do no better than random guessing at inferring how the model behaves under changed inputs ([Bilodeau et al.](https://arxiv.org/abs/2212.11870)). The remedy those authors recommend is to evaluate the model directly on the changes of interest, which is what the bounds of Sections 2 to 4 do.\n\nNeighboring work bears on other parts of the problem. [Aggregate Confusion](https://mitsloan.mit.edu/sustainability-initiative/aggregate-confusion-project) documents how widely ESG ratings diverge across agencies and traces most of the divergence to measurement. [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899) selects pairwise comparison queries under set-based uncertainty, with worst-case utility and regret as objectives. My framework works in institutional research where the facts are still being checked, and it connects model disagreement, evidence changes, ranking stability and the choice of what to verify next.\n\nThe first failure case concerns verification priorities based on raw attribution magnitudes. Take a linear scoring model with a fixed marginal background distribution $`Q`$. The [SHAP](https://arxiv.org/abs/1705.07874) attribution for feature $`j`$ is $`\\phi_j = \\beta_j(x_j - E_Q X_j)`$. If the feature has unresolved factual uncertainty bounded by $`\\left\\vert\\delta_j\\right\\vert \\le d_j`$, its sensitivity to verification is $`\\left\\vert\\beta_j\\right\\vert d_j`$, and over a box of admissible evidence errors the exact worst-case change in the company score is\n\n$$\n\\sup_{\\left\\vert\\delta_j\\right\\vert \\le d_j} \\left\\vert f(x + \\delta) - f(x)\\right\\vert = \\sum_j \\left\\vert\\beta_j\\right\\vert d_j \\tag{1.1}\n$$\n\nThe upper bound comes from the triangle inequality, and it is attained by picking the endpoint of each error interval so that every perturbation pushes the score in the same direction. Apart from the coefficient, that sum depends on how wide the errors are, whereas the attribution depends on where the feature sits relative to the background. Attribution magnitude and verification value are separate calculations. Two cases show the difference:\n• A financial metric may have a large [SHAP](https://arxiv.org/abs/1705.07874) attribution $`\\left\\vert\\phi_j\\right\\vert`$ and come from audited filings that have already been verified ($`d_j \\approx 0`$). Further checks of that field do little to reduce score uncertainty.\n• A self-reported ESG or operating metric may lie near the sector mean ($`x_j \\approx E_Q X_j \\implies \\phi_j \\approx 0`$) and so receive a near-zero [SHAP](https://arxiv.org/abs/1705.07874) attribution. Substantial unresolved reporting error ($`d_j \\gg 0`$) may still be enough to reverse a company ranking.\n\nThe second failure case concerns conventional resampling: redrawing existing records cannot identify a shared reporting bias. Suppose $`n`$ reported observations satisfy $`X_i = \\mu + b + \\varepsilon_i`$, where $`\\varepsilon_i \\sim \\mathcal{N}(0, \\sigma^2)`$ are independent errors and $`b \\neq 0`$ is an unmodeled bias common to all records, as with systematic overstatement across a sector. A conventional nominal 95% normal confidence interval centered on the sample mean $`\\bar{X}`$ covers the true parameter $`\\mu`$ with probability\n\n$$\n\\Pr\\left\\{ \\mu \\in \\left[ \\bar{X} - z\\frac{\\sigma}{\\sqrt{n}}, \\ \\bar{X} + z\\frac{\\sigma}{\\sqrt{n}} \\right] \\right\\} = \\Phi\\left( z - \\frac{b\\sqrt{n}}{\\sigma} \\right) - \\Phi\\left( -z - \\frac{b\\sqrt{n}}{\\sigma} \\right) \\longrightarrow 0 \\quad (\\text{as } n \\to \\infty) \\tag{1.2}\n$$\n\nwhere $`z = \\Phi^{-1}(0.975) \\approx 1.96`$. With a fixed nonzero common bias, a larger sample reduces random variance while the interval's coverage of the true parameter tends to zero, and bootstrapping the same biased records adds no information that identifies $`b`$. The effect is large in practice. A survey with about 250,000 responses a week overestimated US vaccine uptake by 17 percentage points, an estimate no more accurate than a simple random sample of size 10 ([Bradley et al.](https://arxiv.org/abs/2106.05818)). In ESG data, the association between corporate emissions and stock returns reported in earlier work was traced to emissions figures estimated by one data vendor from financial fundamentals, with minimal evidence of any association in the emissions firms disclosed themselves ([Aswani et al.](https://doi.org/10.1093/rof/rfad013)). A shared third-party estimate is the error structure the joint model of Section 2 is built to carry.\n\nBounding the common bias takes external evidence, and once verification supplies a bound $`\\left\\vert b\\right\\vert \\le \\bar{b}`$ the interval can be made both valid and short. The shortest interval with 95% coverage for every admissible bias has half-width\n\n$$\n\\mathrm{cv}_{0.95}\\!\\Big(\\frac{\\bar{b}\\sqrt{n}}{\\sigma}\\Big) \\frac{\\sigma}{\\sqrt{n}}, \\qquad \\mathrm{cv}_{0.95}(t) = \\text{the 95\\% quantile of } \\left\\vert \\mathcal{N}(t, 1)\\right\\vert \\approx t + 1.645 \\ \\text{ for large } t \\tag{1.3}\n$$\n\nwhich is the bias-aware critical value of [Armstrong and Kolesár](https://doi.org/10.3982/QE1199). It also says when to stop collecting. Past\n\n$$\nn^* = \\Big( \\frac{1.645\\, \\sigma}{\\bar{b}} \\Big)^2 \\tag{1.4}\n$$\n\nrecords, the bias bound is the larger part of the half-width, and halving the bound shortens the interval more than doubling the sample does. From there on, analyst time buys more precision in verification than in collection.\n\n## 2 Joint uncertainty over models and evidence\n\nI represent model multiplicity and unresolved evidence jointly to assess whether an ESG or investment ranking stays stable under admissible changes to either.\n\nOn the model side, I fix the prediction target, input definitions and output scale, and let $`\\mathcal{F}_0`$ contain $`M`$ candidate scoring models specified before examining the validation data. The validation set consists of $`n`$ i.i.d. observations from the target evaluation distribution, independent of model construction, and the loss is bounded in $`[0, 1]`$. For a failure probability $`\\alpha \\in (0, 1)`$, Hoeffding's inequality and a union bound give\n\n$$\n\\Pr\\left\\{ \\max_{f \\in \\mathcal{F}_0} \\left\\vert\\hat{L}_n(f) - L(f)\\right\\vert \\le t_n \\right\\} \\ge 1 - \\alpha, \\qquad t_n = \\sqrt{\\frac{\\log(2M / \\alpha)}{2n}} \\tag{2.1}\n$$\n\nLet $`f^*`$ minimize population loss within $`\\mathcal{F}_0`$. I retain every model whose empirical validation loss is within $`2t_n`$ of the empirical minimum, which defines an empirical Rashomon set:\n\n$$\n\\mathcal{F}_\\alpha = \\left\\{ f \\in \\mathcal{F}_0 : \\hat{L}_n(f) \\le \\min_{g \\in \\mathcal{F}_0} \\hat{L}_n(g) + 2t_n \\right\\} \\tag{2.2}\n$$\n\nOn the simultaneous concentration event, $`\\hat{L}_n(f^*) \\le L(f^*) + t_n \\le \\min_g \\hat{L}_n(g) + 2t_n`$, so $`f^* \\in \\mathcal{F}_\\alpha`$ and every retained model has population excess loss at most $`4t_n`$. The construction follows [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html); the finite-class radius and the excess-loss bound are the two-line derivation above.\n\nOn the evidence side, $`\\mathcal{X}(E)`$ represents the questions still unresolved under the current diligence state $`E`$: entity boundaries, reporting periods, physical units, measurement definitions, shared third-party sources and suspected greenwashing claims. Each $`X \\in \\mathcal{X}(E)`$ specifies a joint input configuration for all candidate companies at once, and that is how error dependence across companies and fields is preserved. Over the joint space $`\\Omega(E) = \\mathcal{F}_\\alpha \\times \\mathcal{X}(E)`$, the exact worst-case bounds on the score difference between companies $`i`$ and $`k`$ are\n\n$$\n\\underline{\\Delta}_{ik} = \\inf_{(f, X) \\in \\Omega(E)} [f(x_i) - f(x_k)], \\qquad \\overline{\\Delta}_{ik} = \\sup_{(f, X) \\in \\Omega(E)} [f(x_i) - f(x_k)] \\tag{2.3}\n$$\n\nIf $`\\underline{\\Delta}_{ik} > 0`$, the ordering $`i \\succ k`$ holds for every retained model in $`\\mathcal{F}_\\alpha`$ and every admissible resolution of the evidence in $`\\mathcal{X}(E)`$. Suppose calibration establishes that $`\\mathcal{X}(E)`$ contains the true inputs with probability at least $`1 - \\beta`$ at the specified certification stage. A union bound then gives joint inclusion of $`f^*`$ and the true inputs with probability at least $`1 - \\alpha - \\beta`$, with no independence required between model validation and evidence calibration. The probability of issuing a certificate that is false for $`f^*`$ at the true inputs is at most $`\\alpha + \\beta`$. The object certified is the score ordering under the best model of the prespecified class at the true inputs, and for certification repeated over several stages $`\\beta`$ is calibrated jointly across the stages.\n\n**One event certifies every pair.** The guarantee is a statement about a single inclusion event, so on that event every pair with a positive lower bound is ordered correctly at once. Nothing is paid for the number of comparisons. Three consequences follow.\n\n• The output of diligence is a certified partial order over the companies, and verification can only extend it: resolving evidence shrinks $`\\mathcal{X}(E)`$, so every lower bound stays where it is or rises.\n• Progress has a measure, the share of company pairs that are certified, and a diligence plan can be judged by how far it moves that share per analyst hour.\n• Selecting the top $`K`$ of $`n`$ companies needs only the $`K(n - K)`$ certificates that put each selected company above each rejected one. For five of twenty that is 75 certificates in place of all 190 pairs.\n\nThree neighboring constructions come close and stop short of this. Robust ordinal regression reports the \"necessary\" preferences that hold for every value function compatible with a decision maker's stated judgments, and shows that the necessary relation grows with information ([Greco et al.](https://ideas.repec.org/a/eee/ejores/v191y2008i2p416-436.html)); its uncertainty is entirely on the model side and the inputs are taken as known. Certain predictions hold for every classifier trained on every possible completion of an incomplete data set, and the companion cleaning rule chooses which row a person should fix next ([Karlaš et al.](https://arxiv.org/abs/2005.05117)); its possible worlds are independent across rows, so an error shared by two units cannot cancel, and its selection rule counts worlds, which amounts to a uniform prior over them. Confidence sets for ranks give simultaneous statistical statements about league tables ([Mogstad et al.](https://www.nber.org/papers/w26883)); they pay for simultaneity and treat each unit's estimate as a noisy draw, with no account of unverified inputs. The joint set $`\\Omega(E)`$ carries model multiplicity and cross-company evidence error together, without a prior.\n\n## 3 Computing comparison bounds with support functions\n\nConvex support functions make the joint comparison bounds tractable ([Ben-Tal and Nemirovski](https://www2.isye.gatech.edu/~nemirovs/MOR_RCO_1998.pdf)). Let $`u \\in U`$ collect the unresolved factual quantities, including common benchmark errors and company-specific allowances for reporting error, and write company inputs as $`x_i(u) = \\bar{x}_i + B_i u`$. For an affine scoring model, the pairwise difference is $`\\Delta_f(u) = m_f + a_f^\\top u`$, where $`a_f = (B_i - B_k)^\\top \\beta_f`$, and the support function $`h_U(a) = \\sup_{u \\in U} a^\\top u`$ gives exact comparison bounds:\n\n$$\n\\inf_{u \\in U} \\Delta_f(u) = m_f - h_U(-a_f), \\qquad \\sup_{u \\in U} \\Delta_f(u) = m_f + h_U(a_f) \\tag{3.1}\n$$\n\nFor the ellipsoidal uncertainty set $`U = \\{Cz : \\|z\\|_2 \\le r\\}`$, the Cauchy-Schwarz inequality gives $`h_U(a_f) = r\\|C^\\top a_f\\|_2`$. For an $`\\ell_\\infty`$ box centered at zero with half-widths $`d_j`$, $`h_U(a_f) = \\sum_j \\left\\vert a_{f,j}\\right\\vert d_j`$. I call $`s_j = \\left\\vert a_{f,j}\\right\\vert d_j`$ the *swing* of field $`j`$: the most the unresolved state of that field can move this comparison.\n\nKeeping everything in one joint representation lets shared errors cancel exactly in a pairwise comparison. Let $`S`$ index common-source coordinates on which the two companies have identical loadings, so that $`(B_i-B_k)_{:,S}=0`$. Under the same linear scoring model,\n\n$$\n(a_f)_S=0,\\qquad \\Delta_f(u)=m_f+(a_f)_{S^c}^{\\top}u_{S^c} \\tag{3.2}\n$$\n\nThe common-source contribution cancels, and company-specific uncertainty stays in the support-function bound. If every uncertain direction cancels, $`a_f=0`$ and $`h_U(\\pm a_f)=0`$. Subtracting intervals constructed separately for companies $`i`$ and $`k`$ loses this, because it lets a shared source move in opposite directions at the same time. That combination is incompatible with the joint error structure, and it produces an unnecessarily wide difference bound.\n\n**How many misstatements a ranking survives.** The box lets every unresolved field take its worst value at once. A committee is usually asking something narrower: how many things would have to be wrong together before this ranking fails? If at most $`\\Gamma`$ fields are misstated simultaneously, the support function is the sum of the $`\\Gamma`$ largest swings, the budgeted uncertainty set of [Bertsimas and Sim](https://ideas.repec.org/a/inm/oropre/v52y2004i1p35-53.html). Ordering the swings as $`s_{(1)} \\ge s_{(2)} \\ge \\cdots`$, the *misstatement tolerance* of the ordering $`i \\succ k`$ is\n\n$$\n\\Gamma^*_{ik} = \\max\\Big\\{ \\Gamma : \\sum_{j \\le \\Gamma} s_{(j)} < m_f \\Big\\} \\tag{3.3}\n$$\n\nthe largest number of simultaneous worst-case misstatements the ranking survives. It turns a bound into a sentence: this ranking holds against any two unverified fields being wrong at once, and three could overturn it. It also sets a target for verification, since raising $`\\Gamma^*`$ to the level a committee requires is a question about which swings to remove.\n\nFor a smooth nonlinear model, set $`m_f=\\Delta_f(0)`$ and $`a_f=\\nabla\\Delta_f(0)`$. If $`\\Delta_f(u)`$ is twice continuously differentiable on $`\\|u\\|_2 \\le r`$ and its Hessian has operator norm bounded by $`K_f`$, Taylor's theorem gives\n\n$$\n\\left\\vert\\Delta_f(u) - m_f - a_f^\\top u\\right\\vert \\le \\kappa_f, \\qquad \\kappa_f = \\frac{1}{2}K_f r^2 \\implies \\underline{\\Delta}_f \\ge m_f - r\\|a_f\\|_2 - \\kappa_f \\tag{3.4}\n$$\n\nNonsmooth tree ensembles get their bounds from leaf-region partitioning or mixed-integer linear programming ([Kantchelian et al.](https://proceedings.mlr.press/v48/kantchelian16.html)). The implementation spends computation first on company pairs whose difference bounds are close to the zero decision boundary.\n\n## 4 Selecting queries by the regret they can remove\n\nWhen $`\\underline{\\Delta}_{ik} \\le 0 < \\overline{\\Delta}_{ik}`$, the ranking is unresolved, since a factual correction or a resolution of ambiguous evidence could change the decision. The system then prioritizes the next verification query $`q`$, such as reviewing a supplier contract, requesting original utility bills, or checking a regulatory filing.\n\nLet $`A`$ be a finite set of institutional actions, such as advancing company $`i`$ or company $`k`$ to final diligence, or deferring a rating. Let $`U(a, \\omega)`$ be the prespecified utility of an action in scenario $`\\omega \\in \\Omega`$, where $`\\Omega`$ is nonempty and utilities are uniformly bounded. The baseline minimax regret over deterministic actions is\n\n$$\nR^*(\\Omega) = \\min_{a \\in A} \\sup_{\\omega \\in \\Omega} \\left[ \\max_{b \\in A} U(b, \\omega) - U(a, \\omega) \\right] \\tag{4.1}\n$$\n\nA query $`q`$ returns an observation $`o \\in O_q`$ and restricts the compatible scenarios to a nonempty subset $`\\Omega_{q,o} \\subseteq \\Omega`$. I score the query by its guaranteed reduction in worst-case decision regret:\n\n$$\nV(q) = R^*(\\Omega) - \\sup_{o \\in O_q} R^*(\\Omega_{q,o}) \\ge 0 \\tag{4.2}\n$$\n\nNonnegativity follows from $`\\Omega_{q,o} \\subseteq \\Omega`$: restricting the scenario set cannot increase worst-case regret for any fixed action, and minimizing over $`a \\in A`$ preserves that inequality. Calculating $`V(q)`$ uses the set of compatible scenarios and requires no prior probability that a disclosure is exaggerated. Minimax regret is Savage's criterion, and choosing queries by it comes from utility elicitation ([Boutilier et al.](https://www.cs.toronto.edu/~cebly/Papers/BPPS-aij06-abs.html); [Robust Active Preference Elicitation](https://arxiv.org/html/2003.01899)). Those methods elicit a decision maker's preferences. Here the queries are facts about companies, their answers move the inputs of every model in $`\\mathcal{F}_\\alpha`$ at once, and they cost analyst hours.\n\n**The exact plan for affine scores.** For a choice between two companies under an affine score with interval evidence, the whole problem has a closed form. Write $`m`$ for the current margin, $`h = \\sum_j s_j`$ for the total swing, and $`s_J = \\sum_{j \\in J} s_j`$ for the swing of a set $`J`$ of fields. Verify the fields in $`J`$, let the action be chosen again once their values are known, and take the worst values they could turn out to have. The regret that remains, and the guaranteed value of the set, are\n\n$$\nR_J = \\big( h - \\max(s_J, \\left\\vert m\\right\\vert) \\big)_+, \\qquad V(J) = R_\\varnothing - R_J = (s_J - \\left\\vert m\\right\\vert)_+ \\quad \\text{when } h > \\left\\vert m\\right\\vert \\tag{4.3}\n$$\n\nThe reason is that verification acts twice. It removes $`s_J`$ of swing, and it moves the margin by the revealed values, by as much as $`s_J`$ in the unfavorable direction. Until $`s_J`$ exceeds $`\\left\\vert m\\right\\vert`$ the two effects cancel in the worst case. Four results follow.\n\n1. *The pivot rule.* A single field has guaranteed value only if $`s_j > \\left\\vert m\\right\\vert`$, that is, only if that field alone could reverse the decision, and its value is then $`s_j - \\left\\vert m\\right\\vert`$. Checking a field whose swing is below the margin guarantees nothing by itself, however large its attribution.\n2. *Complementarity, exactly.* Two fields that are each worthless alone are worth $`s_J - \\left\\vert m\\right\\vert`$ together as soon as their combined swing passes the margin. A rule that scores one query at a time ranks them last; the closed form shows what such a rule gives up.\n3. *The plan is a knapsack.* With verification costs $`c_j`$ and a budget of analyst hours, the plan that leaves the least regret maximizes verified swing within the budget. That is a knapsack on the swings, and ordering fields by $`s_j / c_j`$ solves its relaxation.\n4. *Shared sources are worth nothing to a pairwise choice.* A common-source field on which the two companies load identically has $`a_j = 0`$ and zero swing. This is the cancellation of Section 3 read as scheduling advice: when choosing between two companies, analyst hours spent on the benchmark they share cannot move the choice.\n\nAcross the model set $`\\mathcal{F}_\\alpha`$ the same expressions apply with the least favorable model for each action.\n\nVerification narrows an allowance and seldom removes it. In a field experiment on third-party environmental audits, auditors chosen and paid by the audited plants systematically reported readings just below the regulatory standard ([Duflo et al.](https://ideas.repec.org/a/oup/qjecon/v128y2013i4p1499-1545.html)). I therefore let a verified field keep a fraction $`\\lambda_j`$ of its half-width, set by the standing of the source: self-reported, limited assurance, reasonable assurance or primary document. The swing a check removes is $`(1 - \\lambda_j) s_j`$, and that is the quantity that enters the knapsack.\n\nWith two-point evidence in place of intervals the worst observation cannot land in the middle of the range, so a check guarantees more; the third example of Section 6 is of that kind. The interval case is the conservative one. And because a verification rule the scored companies can anticipate invites embellishment where checks are unlikely ([Lyon and Maxwell](https://ideas.repec.org/a/bla/jemstr/v20y2011i1p3-41.html)), the plan is held inside the diligence team and combined with a share of checks drawn at random.\n\n## 5 Propagating factual corrections through scores and attributions\n\nWhen diligence resolves a query and the verified evidence vector changes from $`x`$ to $`x'`$, the system follows versioned dependencies to recompute the affected scores, peer rankings and [SHAP](https://arxiv.org/abs/1705.07874) explanations. With the scoring function $`f`$, background reference $`Q`$ and rubric version held fixed, Shapley efficiency gives the exact identity ([SHAP local accuracy](https://arxiv.org/pdf/1705.07874)):\n\n$$\nf(x) = \\phi_0 + \\sum_j \\phi_j(f, Q, x), \\qquad \\sum_j [\\phi_j(f, Q, x') - \\phi_j(f, Q, x)] = f(x') - f(x) \\tag{5.1}\n$$\n\nApproximate Shapley outputs meet this up to a residual. Suppose their efficiency residuals are $`\\left\\vert\\phi_0+\\sum_j\\phi_j-f(x)\\right\\vert\\le\\eta`$ and $`\\left\\vert\\phi_0+\\sum_j\\phi'_j-f(x')\\right\\vert\\le\\eta'`$ under the same baseline $`\\phi_0`$. The triangle inequality then gives the recomputation check $`\\left\\vert\\sum_j \\Delta \\phi_j - \\Delta f\\right\\vert \\le \\eta + \\eta'`$, which every update has to pass before its explanations reach the committee.\n\nEvery database update records `(field version, evidence document ID, model version, background Q, rubric version)`. With these identifiers, a change caused by a factual correction is told apart from one caused by a model update or a change to the background distribution.\n\n## 6 Worked examples\n\nFive worked examples fix the size of each effect under stated inputs.\n\n1. **Confidence interval coverage under common reporting bias.**\n• Set the shared bias to $`b = 0.2`$ and independent noise to $`\\sigma = 1`$. If the common bias $`b`$ is ignored, a nominal 95% interval has actual coverage of 48.40% at $`n = 100`$ records and 0.000637% at $`n = 1,000`$. Once targeted verification shows $`\\left\\vert b\\right\\vert \\le 0.2`$, the half-width $`0.2 + 1.96/\\sqrt{n}`$ gives 97.50% coverage at $`n = 100`$, and the bias-aware half-width of Section 1 is $`0.3645`$ against $`0.396`$: 8.0% shorter at exactly 95%. The threshold is $`n^* \\approx 68`$ records. Past it, verification is the better use of the next analyst hour.\n2. **Cancellation of shared errors in pairwise rankings.**\n• Two companies have a nominal score difference of $`0.6`$ points and share an additive sector-benchmark error in $`[-1, 1]`$. Subtracting separately constructed company intervals gives $`[-1.4, 2.6]`$, which cannot certify the ordering, whereas the joint support-function calculation cancels the shared error and gives $`[0.6, 0.6]`$, which certifies the ranking.\n3. **Minimax-regret verification compared with [SHAP](https://arxiv.org/abs/1705.07874) magnitude.**\n• Consider a single query, with the same cost for each query option. The example has 8 scenarios, which come from 2 candidate models in the Rashomon set ($`s \\in \\{-1, 1\\}`$) and 4 unresolved evidence configurations ($`z_1 \\in \\{-0.4, 0.4\\}`$, $`z_2 \\in \\{-0.2, 0.2\\}`$). The pairwise score difference is $`\\Delta_s = 0.15s + z_1 + 0.3s z_2`$. A further shared feature has a large [SHAP](https://arxiv.org/abs/1705.07874) attribution but is already verified ($`d_0 = 0`$). The action is to select company $`i`$ or company $`k`$, and regret is measured in model-score units.\n\n**Table 1. Guaranteed regret reduction of each verification query.**\n\n| Verification query | Remaining regret under the worst observation | Guaranteed regret reduction $`V(q)`$ |\n| --- | --- | --- |\n| Recheck the verified field with a large [SHAP](https://arxiv.org/abs/1705.07874) attribution | 0.61 | 0 |\n| Verify the smaller unresolved variable $`z_2`$ | 0.61 | 0 |\n| Verify $`z_1`$, which determines the sign of the score difference | 0 | 0.61 |\n\nEnumerating the eight scenarios shows that rechecking the verified field with a large [SHAP](https://arxiv.org/abs/1705.07874) attribution produces no regret reduction ($`V(q) = 0`$), while verifying $`z_1`$ takes worst-case regret from 0.61 to zero.\n\n4. **A verification plan under an analyst budget.**\n• One company leads another by $`m = 0.6`$ score points. Five fields are unresolved, with swings $`0.9, 0.5, 0.4, 0.3, 0.2`$ and verification costs of $`4, 1, 2, 1, 1`$ analyst hours. The total swing is $`2.3`$, so the ranking is open and the worst-case regret is $`1.7`$. Only the first field passes the pivot rule, with guaranteed value $`0.3`$; each of the others alone guarantees nothing, and the second and third together are worth $`0.3`$. With four hours available:\n\n**Table 2. Two verification plans under a budget of four analyst hours.**\n\n| Plan | Fields verified | Hours | Regret remaining |\n| --- | --- | --- | --- |\n| Largest swing first | 1 | 4 | 1.4 |\n| Swing per hour (the knapsack plan) | 2, 3, 4 | 4 | **1.1** |\n\nThe knapsack plan is the optimum over all plans within the budget, and it leaves 21% less regret than checking the largest field.\n\n5. **Misstatement tolerance.**\n• A ranking has margin $`m = 1.0`$ and six unresolved fields with swings $`0.4, 0.3, 0.3, 0.2, 0.2, 0.1`$. The box bound does not certify it, since the swings sum to $`1.5`$. The two largest sum to $`0.7`$ and the three largest to $`1.0`$, so $`\\Gamma^* = 2`$: the ranking holds against any two of the six fields being misstated at once, and that sentence is what the committee receives.\n\nIn the implementation with the asset manager, institutional diligence requirements became measures of score reliability, pairwise ranking stability and verification priority. The workflow ingests evidence, scores companies across the near-optimal model set, computes the certified partial order and its tolerances, and allocates verification by the plan of Section 4. Each correction triggers rescoring and the recomputation check of Section 5, so the investment committee has a traceable basis for every conclusion it is shown.",
+ "markdown": "## 1 Feature attribution and the value of verification\n\n[SHAP](https://arxiv.org/abs/1705.07874) splits a model prediction into additive feature contributions through its local accuracy property. Conditional Shapley methods also account for dependence among features, and Rashomon-set methods check how stable explanations are across near-optimal models ([SHAP](https://arxiv.org/abs/1705.07874); [Aas et al.](https://arxiv.org/html/1903.10464v3); [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html)). All of them take the corporate inputs $`x`$ as given. In investment diligence the inputs are the open question, because self-reported disclosures can be polished well past what the underlying facts support, and an analyst needs to know which corrections to the evidence could change a decision. Attribution was not built for that question, and there is now a theorem to that effect: for expressive model classes, attribution methods that are complete and linear, SHAP among them, can do no better than random guessing at inferring how the model behaves under changed inputs ([Bilodeau et al.](https://arxiv.org/abs/2212.11870)). The remedy those authors recommend is to evaluate the model directly on the changes of interest, which is what the bounds of Sections 2 to 4 do.\n\nNeighboring work bears on other parts of the problem. [Aggregate Confusion](https://academic.oup.com/rof/article/26/6/1315/6590670) documents how widely ESG ratings diverge across agencies and traces most of the divergence to measurement. [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899) selects pairwise comparison queries under set-based uncertainty, with worst-case utility and regret as objectives. My framework works in institutional research where the facts are still being checked, and it connects model disagreement, evidence changes, ranking stability and the choice of what to verify next.\n\nThe first failure case concerns verification priorities based on raw attribution magnitudes. Take a linear scoring model with a fixed marginal background distribution $`Q`$. The [SHAP](https://arxiv.org/abs/1705.07874) attribution for feature $`j`$ is $`\\phi_j = \\beta_j(x_j - E_Q X_j)`$. If the feature has unresolved factual uncertainty bounded by $`\\left\\vert\\delta_j\\right\\vert \\le d_j`$, its sensitivity to verification is $`\\left\\vert\\beta_j\\right\\vert d_j`$, and over a box of admissible evidence errors the exact worst-case change in the company score is\n\n$$\n\\sup_{\\left\\vert\\delta_j\\right\\vert \\le d_j} \\left\\vert f(x + \\delta) - f(x)\\right\\vert = \\sum_j \\left\\vert\\beta_j\\right\\vert d_j \\tag{1.1}\n$$\n\nThe upper bound comes from the triangle inequality, and it is attained by picking the endpoint of each error interval so that every perturbation pushes the score in the same direction. Apart from the coefficient, that sum depends on how wide the errors are, whereas the attribution depends on where the feature sits relative to the background. Attribution magnitude and verification value are separate calculations. Two cases show the difference:\n• A financial metric may have a large [SHAP](https://arxiv.org/abs/1705.07874) attribution $`\\left\\vert\\phi_j\\right\\vert`$ and come from audited filings that have already been verified ($`d_j \\approx 0`$). Further checks of that field do little to reduce score uncertainty.\n• A self-reported ESG or operating metric may lie near the sector mean ($`x_j \\approx E_Q X_j \\implies \\phi_j \\approx 0`$) and so receive a near-zero [SHAP](https://arxiv.org/abs/1705.07874) attribution. Substantial unresolved reporting error ($`d_j \\gg 0`$) may still be enough to reverse a company ranking.\n\nThe second failure case concerns conventional resampling: redrawing existing records cannot identify a shared reporting bias. Suppose $`n`$ reported observations satisfy $`X_i = \\mu + b + \\varepsilon_i`$, where $`\\varepsilon_i \\sim \\mathcal{N}(0, \\sigma^2)`$ are independent errors and $`b \\neq 0`$ is an unmodeled bias common to all records, as with systematic overstatement across a sector. A conventional nominal 95% normal confidence interval centered on the sample mean $`\\bar{X}`$ covers the true parameter $`\\mu`$ with probability\n\n$$\n\\Pr\\left\\{ \\mu \\in \\left[ \\bar{X} - z\\frac{\\sigma}{\\sqrt{n}}, \\ \\bar{X} + z\\frac{\\sigma}{\\sqrt{n}} \\right] \\right\\} = \\Phi\\left( z - \\frac{b\\sqrt{n}}{\\sigma} \\right) - \\Phi\\left( -z - \\frac{b\\sqrt{n}}{\\sigma} \\right) \\longrightarrow 0 \\quad (\\text{as } n \\to \\infty) \\tag{1.2}\n$$\n\nwhere $`z = \\Phi^{-1}(0.975) \\approx 1.96`$. With a fixed nonzero common bias, a larger sample reduces random variance while the interval's coverage of the true parameter tends to zero, and bootstrapping the same biased records adds no information that identifies $`b`$. The effect is large in practice. A survey with about 250,000 responses a week overestimated US vaccine uptake by 17 percentage points, an estimate no more accurate than a simple random sample of size 10 ([Bradley et al.](https://arxiv.org/abs/2106.05818)). In ESG data, the association between corporate emissions and stock returns reported in earlier work was traced to emissions figures estimated by one data vendor from financial fundamentals, with minimal evidence of any association in the emissions firms disclosed themselves ([Aswani et al.](https://doi.org/10.1093/rof/rfad013)). A shared third-party estimate is the error structure the joint model of Section 2 is built to carry.\n\nBounding the common bias takes external evidence, and once verification supplies a bound $`\\left\\vert b\\right\\vert \\le \\bar{b}`$ the interval can be made both valid and short. The shortest interval with 95% coverage for every admissible bias has half-width\n\n$$\n\\mathrm{cv}_{0.95}\\!\\Big(\\frac{\\bar{b}\\sqrt{n}}{\\sigma}\\Big) \\frac{\\sigma}{\\sqrt{n}}, \\qquad \\mathrm{cv}_{0.95}(t) = \\text{the 95\\% quantile of } \\left\\vert \\mathcal{N}(t, 1)\\right\\vert \\approx t + 1.645 \\ \\text{ for large } t \\tag{1.3}\n$$\n\nwhich is the bias-aware critical value of [Armstrong and Kolesár](https://doi.org/10.3982/QE1199). It also says when to stop collecting. Past\n\n$$\nn^* = \\Big( \\frac{1.645\\, \\sigma}{\\bar{b}} \\Big)^2 \\tag{1.4}\n$$\n\nrecords, the bias bound is the larger part of the half-width, and halving the bound shortens the interval more than doubling the sample does. From there on, analyst time buys more precision in verification than in collection.\n\n## 2 Joint uncertainty over models and evidence\n\nI represent model multiplicity and unresolved evidence jointly to assess whether an ESG or investment ranking stays stable under admissible changes to either.\n\nOn the model side, I fix the prediction target, input definitions and output scale, and let $`\\mathcal{F}_0`$ contain $`M`$ candidate scoring models specified before examining the validation data. The validation set consists of $`n`$ i.i.d. observations from the target evaluation distribution, independent of model construction, and the loss is bounded in $`[0, 1]`$. For a failure probability $`\\alpha \\in (0, 1)`$, Hoeffding's inequality and a union bound give\n\n$$\n\\Pr\\left\\{ \\max_{f \\in \\mathcal{F}_0} \\left\\vert\\hat{L}_n(f) - L(f)\\right\\vert \\le t_n \\right\\} \\ge 1 - \\alpha, \\qquad t_n = \\sqrt{\\frac{\\log(2M / \\alpha)}{2n}} \\tag{2.1}\n$$\n\nLet $`f^*`$ minimize population loss within $`\\mathcal{F}_0`$. I retain every model whose empirical validation loss is within $`2t_n`$ of the empirical minimum, which defines an empirical Rashomon set:\n\n$$\n\\mathcal{F}_\\alpha = \\left\\{ f \\in \\mathcal{F}_0 : \\hat{L}_n(f) \\le \\min_{g \\in \\mathcal{F}_0} \\hat{L}_n(g) + 2t_n \\right\\} \\tag{2.2}\n$$\n\nOn the simultaneous concentration event, $`\\hat{L}_n(f^*) \\le L(f^*) + t_n \\le \\min_g \\hat{L}_n(g) + 2t_n`$, so $`f^* \\in \\mathcal{F}_\\alpha`$ and every retained model has population excess loss at most $`4t_n`$. The construction follows [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html); the finite-class radius and the excess-loss bound are the two-line derivation above.\n\nOn the evidence side, $`\\mathcal{X}(E)`$ represents the questions still unresolved under the current diligence state $`E`$: entity boundaries, reporting periods, physical units, measurement definitions, shared third-party sources and suspected greenwashing claims. Each $`X \\in \\mathcal{X}(E)`$ specifies a joint input configuration for all candidate companies at once, and that is how error dependence across companies and fields is preserved. Over the joint space $`\\Omega(E) = \\mathcal{F}_\\alpha \\times \\mathcal{X}(E)`$, the exact worst-case bounds on the score difference between companies $`i`$ and $`k`$ are\n\n$$\n\\underline{\\Delta}_{ik} = \\inf_{(f, X) \\in \\Omega(E)} [f(x_i) - f(x_k)], \\qquad \\overline{\\Delta}_{ik} = \\sup_{(f, X) \\in \\Omega(E)} [f(x_i) - f(x_k)] \\tag{2.3}\n$$\n\nIf $`\\underline{\\Delta}_{ik} > 0`$, the ordering $`i \\succ k`$ holds for every retained model in $`\\mathcal{F}_\\alpha`$ and every admissible resolution of the evidence in $`\\mathcal{X}(E)`$. Suppose calibration establishes that $`\\mathcal{X}(E)`$ contains the true inputs with probability at least $`1 - \\beta`$ at the specified certification stage. A union bound then gives joint inclusion of $`f^*`$ and the true inputs with probability at least $`1 - \\alpha - \\beta`$, with no independence required between model validation and evidence calibration. The probability of issuing a certificate that is false for $`f^*`$ at the true inputs is at most $`\\alpha + \\beta`$. The object certified is the score ordering under the best model of the prespecified class at the true inputs, and for certification repeated over several stages $`\\beta`$ is calibrated jointly across the stages.\n\n**One event certifies every pair.** The guarantee is a statement about a single inclusion event, so on that event every pair with a positive lower bound is ordered correctly at once. Nothing is paid for the number of comparisons. Three consequences follow.\n\n• The output of diligence is a certified partial order over the companies, and verification can only extend it: resolving evidence shrinks $`\\mathcal{X}(E)`$, so every lower bound stays where it is or rises.\n• Progress has a measure, the share of company pairs that are certified, and a diligence plan can be judged by how far it moves that share per analyst hour.\n• Selecting the top $`K`$ of $`N`$ companies needs only the $`K(N - K)`$ certificates that put each selected company above each rejected one. For five of twenty that is 75 certificates in place of all 190 pairs.\n\nThree neighboring constructions come close and stop short of this. Robust ordinal regression reports the \"necessary\" preferences that hold for every value function compatible with a decision maker's stated judgments, and shows that the necessary relation grows with information ([Greco et al.](https://ideas.repec.org/a/eee/ejores/v191y2008i2p416-436.html)); its uncertainty is entirely on the model side and the inputs are taken as known. Certain predictions hold for every classifier trained on every possible completion of an incomplete data set, and the companion cleaning rule chooses which row a person should fix next ([Karlaš et al.](https://arxiv.org/abs/2005.05117)); its possible worlds are independent across rows, so an error shared by two units cannot cancel, and its selection rule counts worlds, which amounts to a uniform prior over them. Confidence sets for ranks give simultaneous statistical statements about league tables ([Mogstad et al.](https://www.nber.org/papers/w26883)); they pay for simultaneity and treat each unit's estimate as a noisy draw, with no account of unverified inputs. The joint set $`\\Omega(E)`$ carries model multiplicity and cross-company evidence error together, without a prior.\n\n## 3 Computing comparison bounds with support functions\n\nConvex support functions make the joint comparison bounds tractable ([Ben-Tal and Nemirovski](https://www2.isye.gatech.edu/~nemirovs/MOR_RCO_1998.pdf)). Let $`u \\in U`$ collect the unresolved factual quantities, including common benchmark errors and company-specific allowances for reporting error, and write company inputs as $`x_i(u) = \\bar{x}_i + B_i u`$. For an affine scoring model, the pairwise difference is $`\\Delta_f(u) = m_f + a_f^\\top u`$, where $`a_f = (B_i - B_k)^\\top \\beta_f`$, and the support function $`h_U(a) = \\sup_{u \\in U} a^\\top u`$ gives exact comparison bounds:\n\n$$\n\\inf_{u \\in U} \\Delta_f(u) = m_f - h_U(-a_f), \\qquad \\sup_{u \\in U} \\Delta_f(u) = m_f + h_U(a_f) \\tag{3.1}\n$$\n\nFor the ellipsoidal uncertainty set $`U = \\{Cz : \\|z\\|_2 \\le r\\}`$, the Cauchy-Schwarz inequality gives $`h_U(a_f) = r\\|C^\\top a_f\\|_2`$. For an $`\\ell_\\infty`$ box centered at zero with half-widths $`d_j`$, $`h_U(a_f) = \\sum_j \\left\\vert a_{f,j}\\right\\vert d_j`$. I call $`s_j = \\left\\vert a_{f,j}\\right\\vert d_j`$ the *swing* of field $`j`$: the most the unresolved state of that field can move this comparison.\n\nKeeping everything in one joint representation lets shared errors cancel exactly in a pairwise comparison. Let $`S`$ index common-source coordinates on which the two companies have identical loadings, so that $`(B_i-B_k)_{:,S}=0`$. Under the same linear scoring model,\n\n$$\n(a_f)_S=0,\\qquad \\Delta_f(u)=m_f+(a_f)_{S^c}^{\\top}u_{S^c} \\tag{3.2}\n$$\n\nThe common-source contribution cancels, and company-specific uncertainty stays in the support-function bound. If every uncertain direction cancels, $`a_f=0`$ and $`h_U(\\pm a_f)=0`$. Subtracting intervals constructed separately for companies $`i`$ and $`k`$ loses this, because it lets a shared source move in opposite directions at the same time. That combination is incompatible with the joint error structure, and it produces an unnecessarily wide difference bound.\n\n**How many misstatements a ranking survives.** The box lets every unresolved field take its worst value at once. A committee is usually asking something narrower: how many things would have to be wrong together before this ranking fails? If at most $`\\Gamma`$ fields are misstated simultaneously, the support function is the sum of the $`\\Gamma`$ largest swings, the budgeted uncertainty set of [Bertsimas and Sim](https://ideas.repec.org/a/inm/oropre/v52y2004i1p35-53.html). Ordering the swings as $`s_{(1)} \\ge s_{(2)} \\ge \\cdots`$, the *misstatement tolerance* of the ordering $`i \\succ k`$ is\n\n$$\n\\Gamma^*_{ik} = \\max\\Big\\{ \\Gamma : \\sum_{j \\le \\Gamma} s_{(j)} < m_f \\Big\\} \\tag{3.3}\n$$\n\nthe largest number of simultaneous worst-case misstatements the ranking survives. It turns a bound into a sentence: this ranking holds against any two unverified fields being wrong at once, and three could overturn it. It also sets a target for verification, since raising $`\\Gamma^*`$ to the level a committee requires is a question about which swings to remove.\n\nFor a smooth nonlinear model, set $`m_f=\\Delta_f(0)`$ and $`a_f=\\nabla\\Delta_f(0)`$. If $`\\Delta_f(u)`$ is twice continuously differentiable on $`\\|u\\|_2 \\le r`$ and its Hessian has operator norm bounded by $`K_f`$, Taylor's theorem gives\n\n$$\n\\left\\vert\\Delta_f(u) - m_f - a_f^\\top u\\right\\vert \\le \\kappa_f, \\qquad \\kappa_f = \\frac{1}{2}K_f r^2 \\implies \\underline{\\Delta}_f \\ge m_f - r\\|a_f\\|_2 - \\kappa_f \\tag{3.4}\n$$\n\nNonsmooth tree ensembles get their bounds from leaf-region partitioning or mixed-integer linear programming ([Kantchelian et al.](https://proceedings.mlr.press/v48/kantchelian16.html)). The implementation spends computation first on company pairs whose difference bounds are close to the zero decision boundary.\n\n## 4 Selecting queries by the regret they can remove\n\nWhen $`\\underline{\\Delta}_{ik} \\le 0 < \\overline{\\Delta}_{ik}`$, the ranking is unresolved, since a factual correction or a resolution of ambiguous evidence could change the decision. The system then prioritizes the next verification query $`q`$, such as reviewing a supplier contract, requesting original utility bills, or checking a regulatory filing.\n\nLet $`A`$ be a finite set of institutional actions, such as advancing company $`i`$ or company $`k`$ to final diligence, or deferring a rating. Let $`W(a, \\omega)`$ be the prespecified utility of an action in scenario $`\\omega \\in \\Omega`$, where $`\\Omega`$ is nonempty and utilities are uniformly bounded. The baseline minimax regret over deterministic actions is\n\n$$\nR^*(\\Omega) = \\min_{a \\in A} \\sup_{\\omega \\in \\Omega} \\left[ \\max_{b \\in A} W(b, \\omega) - W(a, \\omega) \\right] \\tag{4.1}\n$$\n\nA query $`q`$ returns an observation $`o \\in O_q`$ and restricts the compatible scenarios to a nonempty subset $`\\Omega_{q,o} \\subseteq \\Omega`$. I score the query by its guaranteed reduction in worst-case decision regret:\n\n$$\nV(q) = R^*(\\Omega) - \\sup_{o \\in O_q} R^*(\\Omega_{q,o}) \\ge 0 \\tag{4.2}\n$$\n\nNonnegativity follows from $`\\Omega_{q,o} \\subseteq \\Omega`$: restricting the scenario set cannot increase worst-case regret for any fixed action, and minimizing over $`a \\in A`$ preserves that inequality. Calculating $`V(q)`$ uses the set of compatible scenarios and requires no prior probability that a disclosure is exaggerated. Minimax regret is Savage's criterion, and choosing queries by it comes from utility elicitation ([Boutilier et al.](https://www.cs.toronto.edu/~cebly/Papers/BPPS-aij06-abs.html); [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899)). Those methods elicit a decision maker's preferences. Here the queries are facts about companies, their answers move the inputs of every model in $`\\mathcal{F}_\\alpha`$ at once, and they cost analyst hours.\n\n**The exact plan for affine scores.** For a choice between two companies under an affine score with interval evidence, the whole problem has a closed form. Write $`m`$ for the current margin, $`h = \\sum_j s_j`$ for the total swing, and $`s_J = \\sum_{j \\in J} s_j`$ for the swing of a set $`J`$ of fields. Verify the fields in $`J`$, let the action be chosen again once their values are known, and take the worst values they could turn out to have. The regret that remains, and the guaranteed value of the set, are\n\n$$\nR_J = \\big( h - \\max(s_J, \\left\\vert m\\right\\vert) \\big)_+, \\qquad V(J) = R_\\varnothing - R_J = (s_J - \\left\\vert m\\right\\vert)_+ \\quad \\text{when } h > \\left\\vert m\\right\\vert \\tag{4.3}\n$$\n\nThe reason is that verification acts twice. It removes $`s_J`$ of swing, and it moves the margin by the revealed values, by as much as $`s_J`$ in the unfavorable direction. Until $`s_J`$ exceeds $`\\left\\vert m\\right\\vert`$ the two effects cancel in the worst case. Four results follow.\n\n1. *The pivot rule.* A single field has guaranteed value only if $`s_j > \\left\\vert m\\right\\vert`$, that is, only if that field alone could reverse the decision, and its value is then $`s_j - \\left\\vert m\\right\\vert`$. Checking a field whose swing is below the margin guarantees nothing by itself, however large its attribution.\n2. *Complementarity, exactly.* Two fields that are each worthless alone are worth $`s_J - \\left\\vert m\\right\\vert`$ together as soon as their combined swing passes the margin. A rule that scores one query at a time ranks them last; the closed form shows what such a rule gives up.\n3. *The plan is a knapsack.* With verification costs $`c_j`$ and a budget of analyst hours, the plan that leaves the least regret maximizes verified swing within the budget. That is a knapsack on the swings, and ordering fields by $`s_j / c_j`$ solves its relaxation.\n4. *Shared sources are worth nothing to a pairwise choice.* A common-source field on which the two companies load identically has $`a_j = 0`$ and zero swing. This is the cancellation of Section 3 read as scheduling advice: when choosing between two companies, analyst hours spent on the benchmark they share cannot move the choice.\n\nAcross the model set $`\\mathcal{F}_\\alpha`$ the same expressions apply with the least favorable model for each action.\n\nVerification narrows an allowance and seldom removes it. In a field experiment on third-party environmental audits, auditors chosen and paid by the audited plants systematically reported readings just below the regulatory standard ([Duflo et al.](https://ideas.repec.org/a/oup/qjecon/v128y2013i4p1499-1545.html)). I therefore let a verified field keep a fraction $`\\lambda_j`$ of its half-width, set by the standing of the source: self-reported, limited assurance, reasonable assurance or primary document. The swing a check removes is $`(1 - \\lambda_j) s_j`$, and that is the quantity that enters the knapsack.\n\nWith two-point evidence in place of intervals the worst observation cannot land in the middle of the range, so a check guarantees more; the third example of Section 6 is of that kind. The interval case is the conservative one. And because a verification rule the scored companies can anticipate invites embellishment where checks are unlikely ([Lyon and Maxwell](https://ideas.repec.org/a/bla/jemstr/v20y2011i1p3-41.html)), the plan is held inside the diligence team and combined with a share of checks drawn at random.\n\n## 5 Propagating factual corrections through scores and attributions\n\nWhen diligence resolves a query and the verified evidence vector changes from $`x`$ to $`x'`$, the system follows versioned dependencies to recompute the affected scores, peer rankings and [SHAP](https://arxiv.org/abs/1705.07874) explanations. With the scoring function $`f`$, background reference $`Q`$ and rubric version held fixed, Shapley efficiency gives the exact identity ([SHAP local accuracy](https://arxiv.org/abs/1705.07874)):\n\n$$\nf(x) = \\phi_0 + \\sum_j \\phi_j(f, Q, x), \\qquad \\sum_j [\\phi_j(f, Q, x') - \\phi_j(f, Q, x)] = f(x') - f(x) \\tag{5.1}\n$$\n\nApproximate Shapley outputs meet this up to a residual. Suppose their efficiency residuals are $`\\left\\vert\\phi_0+\\sum_j\\phi_j-f(x)\\right\\vert\\le\\eta`$ and $`\\left\\vert\\phi_0+\\sum_j\\phi'_j-f(x')\\right\\vert\\le\\eta'`$ under the same baseline $`\\phi_0`$. The triangle inequality then gives the recomputation check $`\\left\\vert\\sum_j \\Delta \\phi_j - \\Delta f\\right\\vert \\le \\eta + \\eta'`$, which every update has to pass before its explanations reach the committee.\n\nEvery database update records `(field version, evidence document ID, model version, background Q, rubric version)`. With these identifiers, a change caused by a factual correction is told apart from one caused by a model update or a change to the background distribution.\n\n## 6 Worked examples\n\nFive worked examples fix the size of each effect under stated inputs.\n\n1. **Confidence interval coverage under common reporting bias.**\n• Set the shared bias to $`b = 0.2`$ and independent noise to $`\\sigma = 1`$. If the common bias $`b`$ is ignored, a nominal 95% interval has actual coverage of 48.40% at $`n = 100`$ records and 0.000637% at $`n = 1{,}000`$. Once targeted verification shows $`\\left\\vert b\\right\\vert \\le 0.2`$, the half-width $`0.2 + 1.96/\\sqrt{n}`$ gives 97.50% coverage at $`n = 100`$, and the bias-aware half-width of Section 1 is $`0.3645`$ against $`0.396`$: 8.0% shorter at exactly 95%. The threshold is $`n^* \\approx 68`$ records. Past it, verification is the better use of the next analyst hour.\n2. **Cancellation of shared errors in pairwise rankings.**\n• Two companies have a nominal score difference of $`0.6`$ points and share an additive sector-benchmark error in $`[-1, 1]`$. Subtracting separately constructed company intervals gives $`[-1.4, 2.6]`$, which cannot certify the ordering, whereas the joint support-function calculation cancels the shared error and gives $`[0.6, 0.6]`$, which certifies the ranking.\n3. **Minimax-regret verification compared with [SHAP](https://arxiv.org/abs/1705.07874) magnitude.**\n• Consider a single query, with the same cost for each query option. The example has eight scenarios, which come from two candidate models in the Rashomon set ($`\\theta \\in \\{-1, 1\\}`$) and four unresolved evidence configurations ($`z_1 \\in \\{-0.4, 0.4\\}`$, $`z_2 \\in \\{-0.2, 0.2\\}`$). The pairwise score difference is $`\\Delta_\\theta = 0.15\\theta + z_1 + 0.3\\theta z_2`$. A further shared feature has a large [SHAP](https://arxiv.org/abs/1705.07874) attribution but is already verified ($`d_0 = 0`$). The action is to select company $`i`$ or company $`k`$, and regret is measured in model-score units.\n\n**Table 1. Guaranteed regret reduction of each verification query.**\n\n| Verification query | Remaining regret under the worst observation | Guaranteed regret reduction $`V(q)`$ |\n| --- | --- | --- |\n| Recheck the verified field with a large [SHAP](https://arxiv.org/abs/1705.07874) attribution | 0.61 | 0 |\n| Verify the smaller unresolved variable $`z_2`$ | 0.61 | 0 |\n| Verify $`z_1`$, which determines the sign of the score difference | 0 | 0.61 |\n\nEnumerating the eight scenarios shows that rechecking the verified field with a large [SHAP](https://arxiv.org/abs/1705.07874) attribution produces no regret reduction ($`V(q) = 0`$), while verifying $`z_1`$ takes worst-case regret from 0.61 to zero.\n\n4. **A verification plan under an analyst budget.**\n• One company leads another by $`m = 0.6`$ score points. Five fields are unresolved, with swings $`0.9, 0.5, 0.4, 0.3, 0.2`$ and verification costs of $`4, 1, 2, 1, 1`$ analyst hours. The total swing is $`2.3`$, so the ranking is open and the worst-case regret is $`1.7`$. Only the first field passes the pivot rule, with guaranteed value $`0.3`$; each of the others alone guarantees nothing, and the second and third together are worth $`0.3`$. With four hours available:\n\n**Table 2. Two verification plans under a budget of four analyst hours.**\n\n| Plan | Fields verified | Hours | Regret remaining |\n| --- | --- | --- | --- |\n| Largest swing first | 1 | 4 | 1.4 |\n| Swing per hour (the knapsack plan) | 2, 3, 4 | 4 | **1.1** |\n\nThe knapsack plan is the optimum over all plans within the budget, and it leaves 21% less regret than checking the largest field.\n\n5. **Misstatement tolerance.**\n• A ranking has margin $`m = 1.0`$ and six unresolved fields with swings $`0.4, 0.3, 0.3, 0.2, 0.2, 0.1`$. The box bound does not certify it, since the swings sum to $`1.5`$. The two largest sum to $`0.7`$ and the three largest to $`1.0`$, so $`\\Gamma^* = 2`$: the ranking holds against any two of the six fields being misstated at once, and that sentence is what the committee receives.\n\nIn the implementation with the asset manager, institutional diligence requirements became measures of score reliability, pairwise ranking stability and verification priority. The workflow ingests evidence, scores companies across the near-optimal model set, computes the certified partial order and its tolerances, and allocates verification by the plan of Section 4. Each correction triggers rescoring and the recomputation check of Section 5, so the investment committee has a traceable basis for every conclusion it is shown.",
"sourceDocument": {
"fileName": "ai-research-and-engineering.md",
"path": "content/sources/ai-portfolio/ai-research-and-engineering.md",
- "sha256": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1"
+ "sha256": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0"
},
"repositoryUrl": "https://github.com/130U/RESEARCH-Evidence-Verification-and-Robust-Comparison-in-Investment-Research"
}
diff --git a/website/content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json b/website/content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json
index 11422c3..19d7b87 100644
--- a/website/content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json
+++ b/website/content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json
@@ -3,11 +3,11 @@
"title": "Financial Task Design and Synthetic Data Engineering",
"sourceUrl": "https://github.com/130U/RESEARCH-Financial-Task-Design-and-Synthetic-Data-Engineering/blob/main/manuscript/financial-task-design-and-synthetic-data-engineering.md",
"sourceEditedAt": "2026-11-19",
- "markdown": "## 1 Distributional fit and task validity\n\nModern tabular generators address different distributional problems. [CTGAN](https://arxiv.org/abs/1907.00503) handles mixed discrete and continuous data, multimodality, and class imbalance through mode-specific normalization and conditional vectors, and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) runs Gaussian diffusion on numerical fields and multinomial diffusion on categorical ones ([CTGAN](https://arxiv.org/abs/1907.00503); [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html)). Both learn empirical distributions effectively. A financial task can match its distribution and still be invalid, because part of what makes it valid is structural. Accounting identities have to hold exactly, fiscal periods have to run in chronological order, covenant applicability conditions have to be respected, and the numerical inputs have to agree logically with the reference judgment. Global distributional similarity metrics enforce none of this.\n\nA simple construction shows how far apart the two can drift. Suppose valid financial inputs $`x \\in \\mathbb{R}^d`$ have to satisfy $`m`$ linear accounting or ratio identities $`Ax = b`$, where $`A \\in \\mathbb{R}^{m \\times d}`$ has full row rank, so that the valid population distribution $`P`$ is supported entirely on the affine subspace $`M = \\{x \\in \\mathbb{R}^d : Ax = b\\}`$. Assume this distribution has finite second moments, and let the columns of $`N \\in \\mathbb{R}^{d \\times m}`$ form an orthonormal basis for the normal space orthogonal to $`M`$. Now take a generator that captures $`P`$ along the manifold but adds a small amount of independent noise in the normal directions: $`X_\\sigma = X + \\sigma N \\varepsilon`$, where $`\\varepsilon \\sim \\mathcal{N}(0, I_m)`$ is independent of $`X \\sim P`$. Writing $`Q_\\sigma`$ for the generated distribution, I obtain, for every $`\\sigma > 0`$,\n\n$$\nW_2(P, Q_\\sigma) \\le \\sqrt{E\\|X_\\sigma - X\\|^2} = \\sigma \\sqrt{m} \\longrightarrow 0 \\quad (\\text{as } \\sigma \\to 0^+) \\tag{1.1}\n$$\n\n$$\nP(M) = 1, \\qquad Q_\\sigma(M) = 0 \\tag{1.2}\n$$\n\n$$\n\\|P - Q_\\sigma\\|_{\\text{TV}} := \\sup_B \\left\\vert P(B) - Q_\\sigma(B)\\right\\vert = 1 \\tag{1.3}\n$$\n\nThe first line comes from coupling $`(X, X_\\sigma)`$ on the same probability space. The second holds because $`AN`$ is invertible and a continuous $`m`$-dimensional Gaussian vector equals zero with probability zero, and taking the Borel set $`B = M`$ then gives the third. So the 2-Wasserstein distance $`W_2(P, Q_\\sigma)`$ can tend to zero while the acceptance rate under exact business constraints stays at zero ($`Q_\\sigma(M) = 0`$) and the total variation distance stays at its maximum value of $`1`$. With a numerical tolerance the acceptance rate depends on the ratio of tolerance to off-manifold noise, and Section 2 shows that at a fixed ratio it falls geometrically with each added constraint.\n\nThe construction is one instance of a general gap. Synthetic data is evaluated on three axes, fidelity to the source distribution, utility for a downstream model and privacy, and the sample-level refinements of those metrics still score the inputs ([Alaa et al.](https://proceedings.mlr.press/v162/alaa22a.html)). None of the three has a place for the question a training team has to answer about supervision: is the label attached to this generated case still the right label? Data augmentation has the same blind spot from the other side. Counterfactual augmentation and contrast sets showed that a minimal edit can flip the gold answer, and they restore validity by having a person relabel every edited example ([Kaushik et al.](https://arxiv.org/abs/1909.12434); [Gardner et al.](https://aclanthology.org/2020.findings-emnlp.117/)), which does not scale to generated variants. I add the missing axis, supervision validity, and the remaining sections build the machinery that keeps it high at scale.\n\nThe counterexample also sets how I divided responsibilities in the synthetic data architecture ([SDV constraint-augmented generation](https://docs.sdv.dev/sdv/modeling/constraint-augmented-generation-cag)):\n• Copula models and conditional sampling capture cross-sectional rank and tail dependencies among continuous corporate indicators, including company size, operating history, leverage, and profit margins. They preserve the dependencies that independent column generation would discard.\n• [CTGAN](https://arxiv.org/abs/1907.00503) and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) model mixed-type and multimodal empirical distributions, and explicit business-rule checks follow them.\n• Deterministic variable transformations and manifold projections enforce accounting equalities, ratio bounds, and chronological order.\n\n## 2 Preserving the conditional distribution under business constraints\n\nRejection sampling generates unconstrained candidates and discards those that violate the constraints, and as several financial constraints tighten its computational cost becomes prohibitive. Suppose a candidate draw produces $`m`$ standardized constraint residuals that behave as independent standard normal variables, each of which has to lie within $`[-\\epsilon, \\epsilon]`$. The acceptance probability and the expected number of candidate draws per valid input are\n\n$$\n\\alpha_m(\\epsilon) = [2\\Phi(\\epsilon) - 1]^m \\sim \\left( \\sqrt{\\frac{2}{\\pi}} \\epsilon \\right)^m \\quad (\\epsilon \\to 0^+), \\qquad E[N_{\\text{attempts}}] = \\frac{1}{\\alpha_m(\\epsilon)} \\tag{2.1}\n$$\n\nso each additional accounting or ratio constraint multiplies the expected candidate count. With a tolerance of $`\\epsilon = 0.05`$ residual standard deviations, the exact analytical values are:\n\n**Table 1. Acceptance rate of rejection sampling under simultaneous constraints.**\n\n| Simultaneous constraints ($`m`$) | Rejection-sampling acceptance rate $`\\alpha_m(\\epsilon)`$ | Expected candidate draws per valid input |\n| --- | --- | --- |\n| **1** | 3.9878% | 25.08 |\n| **2** | 0.1590% | 628.84 |\n| **4** | **0.0002529%** | **395,442.59** |\n\nThe alternative is to condition instead of reject: draw once, then correct the draw onto the constraint. In the linear Gaussian setting the covariance-weighted correction returns a sample from the exact conditional distribution. For a Gaussian copula the result holds for constraints that are affine in the latent Gaussian coordinates, since nonlinear marginal transformations do not preserve affine constraints in the observed financial variables. Let the initial unconstrained draw be $`Z \\sim \\mathcal{N}(\\mu, \\Sigma)`$, with $`\\Sigma \\succ 0`$. I apply a covariance-weighted Mahalanobis projection:\n\n$$\nX = \\arg\\min_{x : Ax = b} \\frac{1}{2}(x - Z)^\\top \\Sigma^{-1}(x - Z) = Z + K(b - AZ), \\qquad K = \\Sigma A^\\top (A \\Sigma A^\\top)^{-1} \\tag{2.2}\n$$\n\nThe first-order Lagrangian stationarity condition is $`\\Sigma^{-1}(X - Z) + A^\\top \\lambda = 0`$, and substituting $`AX = b`$ and eliminating the multiplier $`\\lambda`$ gives the closed-form update. Since $`AK = I_m`$, every draw satisfies $`AX = b`$ exactly. Its mean and covariance are\n\n$$\nE[X] = \\mu + K(b - A\\mu) \\tag{2.3}\n$$\n\n$$\n\\mathrm{Cov}(X) = (I - KA)\\Sigma(I - KA)^\\top = \\Sigma - \\Sigma A^\\top (A \\Sigma A^\\top)^{-1} A\\Sigma \\tag{2.4}\n$$\n\nAffine transformations preserve Gaussianity, so these moments identify the exact regular conditional Gaussian distribution of $`Z \\mid AZ = b`$ ([Majumdar and Majumdar](https://arxiv.org/abs/1710.09285)). The update has a long lineage. It is the sampler for a normal distribution truncated on hyperplanes ([Cong, Chen and Zhou](https://doi.org/10.1214/17-BA1052)) and Matheron's rule for pathwise conditioning on a noise-free linear observation ([Wilson et al.](https://arxiv.org/abs/2011.04026)). What I did with it is put accounting identities in the place of the hyperplanes and state where exactness holds, in the latent coordinates of the copula. That also marks the division of labor with constraint layers for deep tabular generators, which enforce linear inequalities in any generator and make no statement about the law of the corrected sample ([Stoian et al.](https://arxiv.org/abs/2402.04823)). Inequalities such as ratio ranges go to transformations and layers; identities go to the projection, which returns the right distribution as well as a feasible point.\n\nOn a reference run of the constrained generator, **200,000 twelve-dimensional balance-sheet vectors subject to 4 simultaneous linear identities** (assets sum to total assets, liabilities sum to total liabilities, assets equal liabilities plus equity, and receivables stand at 60 days of revenue), the maximum absolute constraint residual was **$`3.0 \\times 10^{-15}`$**, and the sample covariance matched the theoretical conditional covariance to **0.359%** in relative Frobenius norm. Every draw is used; the 395,442 candidate draws per valid input in the table above are what rejection would have spent. Fields that accounting definitions fully determine are computed analytically. Range bounds and date ordering are enforced in the latent generation space when they admit bijective transformations, such as logarithmic transforms or positive increments $`t_{k} = t_{k-1} + \\exp(\\xi_k)`$.\n\n**The correction is also a monitor.** The squared Mahalanobis distance by which the projection moves a draw has a known law:\n\n$$\nT = (b - AZ)^\\top (A \\Sigma A^\\top)^{-1} (b - AZ) \\ \\sim\\ \\chi^2_m(\\delta), \\qquad \\delta = (b - A\\mu)^\\top (A \\Sigma A^\\top)^{-1} (b - A\\mu) \\tag{2.5}\n$$\n\nWhen the upstream generator is consistent with the identities, $`\\delta = 0`$ and $`T`$ averages $`m`$. When the upstream model drifts, $`\\delta`$ grows and the average of $`T`$ rises with it. Every batch therefore tests the fit of its own generator while it is being corrected. On the reference run the average of $`T`$ was 4.01 against $`m = 4`$. A 10% drift in the upstream mean of receivables raises $`\\delta`$ to 0.197 and the average to 4.20, which is two standard errors in a batch of 1,000 draws and more than six in a batch of 10,000.\n\n## 3 Using decision margins to determine when reference judgments need recomputation\n\nA synthetic variant of a financial reference case often keeps the business narrative and changes decisive financial figures, covenant thresholds, or macroeconomic inputs. If the original reference answer stays in place after those values change, systematic label errors enter post-training, even when all accounting constraints hold. [FinQA](https://aclanthology.org/2021.emnlp-main.300/) pairs financial questions with executable reasoning programs, and synthetic preference data similarly require a maintained link between inputs, reference conclusions, and the expert's explanation of a preference ([FinQA](https://aclanthology.org/2021.emnlp-main.300/)).\n\nNeighboring work links data generation to feedback in other ways. [WizardMath](https://arxiv.org/pdf/2308.09583v1) couples the evolution of mathematical instructions with reinforcement feedback, and [FinGPT](https://arxiv.org/pdf/2307.10485v1) fine-tunes financial language models on feedback from the market. Templated reasoning sets recompute an answer from the template's variables for every instance ([GSM-Symbolic](https://arxiv.org/abs/2410.05229)), which works when the answer is one formula and leaves evidence and preference untouched. I make the dependencies among a financial scenario, its reference conclusion, the evidence behind that conclusion and the expert preference built on it the object of design, so that a change to the scenario reaches each of them, and only the ones it has to.\n\nConsider a decision rule that chooses between two corporate actions, and let $`d(x)`$ denote the net valuation or metric advantage of action A over action B, so that the correct reference decision is $`\\mathrm{sign}\\,d(x)`$. If $`d`$ is $`L`$-Lipschitz on a feasible neighborhood, a sufficient condition for retaining the original reference decision is that the perturbation $`\\Delta`$ stays strictly within the decision margin:\n\n$$\n\\left\\vert d(x + \\Delta) - d(x)\\right\\vert \\le L\\|\\Delta\\|, \\qquad \\left\\vert d(x)\\right\\vert > L\\|\\Delta\\| \\implies \\mathrm{sign}\\,d(x + \\Delta) = \\mathrm{sign}\\,d(x) \\tag{3.1}\n$$\n\nThis condition certifies the decision label. A loan can stay approved while the calculation behind it, the supporting evidence and the relative quality of two responses all change, so I check reference calculations, evidence and pairwise preferences separately.\n\nMy diagnostic tasks focus on contested decision boundaries where models are prone to error, which means $`\\left\\vert d(x)\\right\\vert`$ is deliberately kept small. For a linear comparison rule $`d(x) = w^\\top x - c`$ with initial margin $`\\gamma = d(x) > 0`$, the stale-label error rate has an exact closed form. Let the perturbation within the feasible manifold satisfy $`\\Delta \\sim \\mathcal{N}(0, \\Omega)`$, with projected variance $`s^2 = w^\\top \\Omega w > 0`$. The probability that the correct label flips while the original decision is retained is\n\n$$\n\\Pr\\{d(x + \\Delta) < 0\\} = \\Pr\\{w^\\top \\Delta < -\\gamma\\} = \\Phi\\left(-\\frac{\\gamma}{s}\\right) \\tag{3.2}\n$$\n\nThe closed form gives the following inherited-label error rates:\n• At a margin-to-noise ratio of $`\\gamma / s = 2`$, the error rate is **2.28%**.\n• At $`\\gamma / s = 1`$, it is **15.87%**.\n• Near a diagnostic decision boundary, at $`\\gamma / s = 0.25`$, it is **40.13%**.\n\n**The formula is exact for covenants.** The linear rule is the common case in finance. A covenant test is a ratio of accounting levels against a threshold, interest coverage $`E / I \\ge c`$ or leverage $`D / E \\le c`$, and with a positive denominator it is the half-space $`E - cI \\ge 0`$ in the levels themselves. For the covenant family the flip probability $`\\Phi(-\\gamma / s)`$ is the exact rate under Gaussian perturbations of the levels, with $`w = (1, -c)`$. When several covenants bind at once, the case flips if any of them does, and the rate is bounded by the sum of the individual rates: two covenants at $`\\gamma / s`$ of 1.5 and 2 give at most 8.96%.\n\n**The certificate is a cache-hit test.** Treat the four objects as nodes of a dependency graph. A perturbed scenario marks its reference conclusion, evidence and preference as stale, and the margin decides which of them a builder may reuse. With a tolerated error $`\\alpha`$ on inherited labels:\n\n$$\n\\text{recompute the variant when } \\frac{\\gamma}{s} < z_{1 - \\alpha}, \\qquad \\text{inherit otherwise, with } \\Pr(\\text{stale label}) \\le \\alpha \\tag{3.3}\n$$\n\nThe thresholds are 1.645 at $`\\alpha = 5\\%`$ and 2.326 at $`\\alpha = 1\\%`$, and a variant at $`\\gamma / s = 3`$ inherits its label with an error of 0.13%. Variants far from the boundary cost no expert time, and the diagnostic variants near it, which are the ones the task family exists for, are always recomputed.\n\nI therefore tie case expansion to reference-answer recomputation and renewed five-level pairwise preference grading when perturbations approach or cross a decision boundary. In that event the workflow recomputes the reference solution and samples new candidate model reasoning trajectories. Domain experts then assess the pair on the five-level preference scale and identify which trajectory recognized the decisive variable whose change reversed the decision and supported its judgment with primary evidence. [HelpSteer2-Preference](https://arxiv.org/abs/2410.01257) gives a comparison point from the methods literature.\n\n**What a graded preference is worth.** A five-level judgment takes an expert a few seconds longer than a binary one, on a task that has already cost about three specialist-hours. Model the judgment as an ordered response to the latent quality gap $`g`$ between two responses, measured in units of judgment noise, with cutpoints $`c_1 < \\dots < c_4`$. The Fisher information one response carries about the gap is\n\n$$\nI(g) = \\sum_{k} \\frac{\\big( \\partial_g\\, p_k(g) \\big)^2}{p_k(g)}, \\qquad p_k(g) = \\Phi(c_k - g) - \\Phi(c_{k-1} - g) \\tag{3.4}\n$$\n\n**Table 2. Fisher information of a binary and a five-level judgment by quality gap.**\n\n| Quality gap $`g`$ | Binary judgment | Five-level judgment (cutpoints at $`\\pm 0.5`$, $`\\pm 1.5`$) | Ratio |\n| --- | --- | --- | --- |\n| 0 (a tie) | 0.637 | 0.912 | 1.43 |\n| 0.5 | 0.581 | 0.902 | 1.55 |\n| 1.0 | 0.439 | 0.863 | 1.97 |\n\nOne graded judgment is worth 1.4 to 2 binary judgments, and the advantage grows with the gap, where a binary label saturates. The same expression designs the scale: at a tie the information is largest with cutpoints at $`\\pm 0.38`$ and $`\\pm 1.24`$, where it reaches 0.920 of the information in a continuous response. The protocol delivers three signals that a training pipeline can consume separately: the direction of the preference, its strength, and the cited evidence and decisive variable.\n\n## 4 Specifying the information available to each temporal task\n\nA cross-sectional company comparison does not require continuous-time stochastic processes. Dynamic tasks such as derivatives valuation, interest-rate term-structure interpolation, and liquidity stress tests over multiple periods do require path generators. For those I use SDEs and neural SDEs to model path distributions beyond single-time marginals ([Neural SDEs as Infinite-Dimensional GANs](https://proceedings.mlr.press/v139/kidger21b.html)). Brownian bridges and stochastic interpolation support conditioning on known intermediate observations or endpoints, following the constructions in [Glasserman, *Monte Carlo Methods in Financial Engineering*, Chapter 3](https://doi.org/10.1007/978-0-387-21617-1). A bridge is the conditioning of Section 2 in one dimension: draw the free path, then correct it onto the known endpoint.\n\nThat conditioning can introduce temporal leakage into evaluation, and I derive the size of the effect in closed form. Consider scaled Brownian motion $`X_u = \\sigma W_u`$, where $`s < t < T`$. Under squared-error loss, the optimal predictor at time $`s`$, given only the historical filtration $`\\mathcal{F}_s`$, has conditional mean squared error $`\\sigma^2(t - s)`$. If the generator conditions paths on the future endpoint $`X_T`$ and inadvertently exposes that conditioning to a forecasting task, Gaussian conditioning gives ([Pitman on the Brownian bridge](https://www.stat.berkeley.edu/~pitman/s205s03/b_bridge.pdf)):\n\n$$\n\\mathrm{Var}(X_t \\mid \\mathcal{F}_s) = \\sigma^2(t - s) \\tag{4.1}\n$$\n\n$$\n\\mathrm{Var}(X_t \\mid \\mathcal{F}_s, X_T) = \\sigma^2(t - s) - \\frac{[\\sigma^2(t - s)]^2}{\\sigma^2(T - s)} = \\sigma^2(t - s)\\frac{T - t}{T - s} \\tag{4.2}\n$$\n\nAt the midpoint $`t = (s + T)/2`$, conditioning on $`X_T`$ halves the conditional variance. The resulting **29.29% apparent reduction in RMSE** ($`1 - 1/\\sqrt{2}`$) comes entirely from the additional endpoint information in this construction, and whether that counts as a flaw depends on the task. It is temporal leakage if the forecasting task does not permit access to that information. If instead the task gives the model both endpoints $`X_s`$ and $`X_T`$ and asks it to audit an interpolation or backcast an intervening accounting period, Brownian bridge conditioning is consistent with the specified information set.\n\nInterest rates and spreads revert to a mean, and for them the leak has its own closed form. For an Ornstein-Uhlenbeck path with mean-reversion speed $`\\kappa`$, at the midpoint of an interval of half-length $`h`$:\n\n$$\n\\frac{\\mathrm{Var}(X_t \\mid \\mathcal{F}_s, X_T)}{\\mathrm{Var}(X_t \\mid \\mathcal{F}_s)} = \\frac{1}{1 + e^{-2\\kappa h}}, \\qquad \\text{apparent RMSE reduction} = 1 - \\big( 1 + e^{-2\\kappa h} \\big)^{-1/2} \\tag{4.3}\n$$\n\n**Table 3. Apparent RMSE reduction from endpoint leakage under mean reversion.**\n\n| $`\\kappa h`$ | 0 (Brownian) | 0.25 | 0.5 | 1 | 2 |\n| --- | --- | --- | --- | --- | --- |\n| Apparent RMSE reduction | 29.29% | 21.10% | 14.50% | 6.15% | 0.90% |\n\nMean reversion discounts the leak, because the endpoint tells less about the middle of a path that forgets. The financial reading is direct: endpoint leakage is most damaging for series close to a unit root, prices and long rates, and nearly harmless for fast-reverting spreads. The information-set rule is therefore enforced with a severity that follows the persistence of the series.\n\nThe task specification keeps three information sets apart:\n1. **Construction information ($`\\mathcal{F}_{\\text{gen}}`$):** the complete path, endpoint, or latent parameters used to generate a scenario.\n2. **Permitted predictor filtration ($`\\mathcal{F}_s`$):** the historical information the task allows the model to observe at decision time $`s`$.\n3. **Evaluation and oracle information ($`\\mathcal{F}_{\\text{eval}}`$):** outcomes realized after $`s`$, reserved for scoring or for explicitly disclosed oracle baselines.\n\n## 5 Batch Zero: independent execution and calibration for post-training data strategy\n\nI created **Batch Zero** around the capability objectives and data requirements of post-training research teams. Before anything scales, it brings task design, synthetic data strategy, and preference supervision under a common set of acceptance criteria. It tests whether domain experts can independently execute a research specification, reproduce reference solutions, and calibrate their preference judgments. The resulting specifications support subsequent data construction.\n• **Independent execution:** before a task family enters large-scale data construction (*Batch One*), doctoral-level domain experts independently write seed cases, generate synthetic variants, verify reference solutions, and assign five-level pairwise preference scores. Implicit assumptions, missing information, and ambiguities in the specification come to light this way.\n• **Calibration and feedback to data strategy:** failures, expert disagreements, and boundary-case errors inform revisions to task instructions, reference solutions, grading rules, and Batch One data acceptance criteria. Task design, execution, revision, and recalibration form an iterative process.\n• **Organizational adoption:** the partner foundation model company subsequently adopted **Batch Zero** across other post-training programs.\n\n**Why the check belongs at the seed.** Expert-written references are less reliable than their reputation. When FinanceReasoning re-examined four public financial reasoning sets, it updated 15.6% of their questions, between 9.72% and 30.00% per set ([Tang et al.](https://arxiv.org/abs/2506.05828)). A synthetic pipeline multiplies whatever error its seeds carry. If a seed's reference solution is wrong with probability $`\\varepsilon`$ and the seed yields $`k`$ variants, a family carries $`k\\varepsilon`$ wrong records in expectation, and they are wrong together. If an independent re-solve catches a wrong reference with probability $`d`$, then $`r`$ re-solves leave\n\n$$\nk\\, \\varepsilon\\, (1 - d)^r \\tag{5.1}\n$$\n\nwrong records per family. With $`k = 20`$, $`\\varepsilon = 0.15`$ and $`d = 0.8`$, that is 3.0 with no re-solve, 0.6 after one and 0.12 after two. One re-solve at the seed does the work of reviewing twenty variants, and that multiplier is what Batch Zero is built on.\n\nFor statistical analysis and evaluation, I treat the task family as the basic unit of effective coverage and uncertainty estimation. Suppose $`G`$ independent reference-case families each yield $`k`$ synthetic variants, giving $`Gk`$ nominal records, and any two variants within a family have correlation $`\\rho`$ and common variance $`\\sigma_Y^2`$. The sample mean then carries the Kish design effect $`1 + (k - 1)\\rho`$, which gives the effective sample size $`N_{\\text{eff}}`$:\n\n$$\n\\mathrm{Var}(\\bar{Y}) = \\frac{Gk\\sigma_Y^2 + Gk(k - 1)\\rho\\sigma_Y^2}{(Gk)^2} = \\frac{\\sigma_Y^2}{Gk}[1 + (k - 1)\\rho] \\implies N_{\\text{eff}} = \\frac{Gk}{1 + (k - 1)\\rho} \\tag{5.2}\n$$\n\nWith $`G = 100`$ independent families, $`k = 20`$ variants per family ($`2,000`$ nominal tasks), and within-family correlation $`\\rho = 0.5`$, the 2,000 records provide the same precision for estimating the mean as about 190 independent records. Ignoring this correlation understates standard errors by a factor of **$`3.24\\times`$**. My data workflow therefore separates training and evaluation sets at the task-family level, and all confidence intervals account for hierarchical clustering at both the family and expert levels.\n\n**How many variants a seed should yield.** The design effect is a diagnosis; the same algebra gives the prescription. A new family costs $`c_G`$, mostly expert time on the seed and its reference, and each further variant costs $`c_v`$. At a fixed budget the variance of the family-clustered mean is proportional to $`(c_G + k c_v)[1 + (k - 1)\\rho] / k`$, and it is smallest at the optimum cluster size of two-stage sampling:\n\n$$\nk^* = \\sqrt{\\frac{c_G}{c_v} \\cdot \\frac{1 - \\rho}{\\rho}} \\tag{5.3}\n$$\n\nWith a seed that costs twenty variants and $`\\rho = 0.5`$, $`k^* \\approx 4.5`$, and the budget that buys 20 variants per family has 40% more variance than the same budget at four or five. Read the other way, moving from 20 variants per family to the optimum cuts the variance by 29% at no cost. Variants beyond $`k^*`$ buy repeated information, and the money is better spent on new families.\n\n## 6 Staged evaluation design and coverage efficiency\n\nStaged, blinded comparisons across independent task families test whether constraint-preserving synthesis and preference supervision meet the data requirements of post-training research teams. The task mix, generator training budget, expert review budget, and final acceptance criteria stay fixed as each method is introduced:\n\n**Table 4. The staged evaluation design.**\n\n| Comparison stage | Method introduced | What the stage measures |\n| --- | --- | --- |\n| **Current data construction baseline** | Unconstrained generation plus expert review and rework | Fully loaded cost of each valid case that is ultimately accepted |\n| **Add manifold-constrained generation** | Hold the generator fixed; apply affine projections and variable transformations | Change in invalid candidates and manual repairs, with the conditional distribution preserved |\n| **Add decision-margin checks and reference-answer recomputation** | Recompute reference judgments below the margin threshold; use $`\\gamma/s`$ to quantify Gaussian label-flip risk | Change in stale-label errors, with valid supervision on both sides of the decision boundary |\n| **Apply Batch Zero acceptance criteria and filtration rules** | Jointly check the temporal information boundaries in $`\\mathcal{F}_s`$, independent execution, reproducibility of reference solutions, and calibration of preference standards | Post-training gains that persist once leakage is excluded and clustering by source task family is accounted for |\n\nRaw acceptance rates favor trivial variants of easy templates, and generation speed alone omits the cost of expert repairs, so I allocate data construction and review resources by scenario-weighted task-family coverage efficiency. Let $`j`$ index predefined cells in the financial capability matrix, with strategic weights $`w_j \\ge 0`$ ($`\\sum_j w_j = 1`$) and required counts of independent valid task families $`n_j^* > 0`$:\n\n$$\n\\mathrm{Coverage} = \\sum_j w_j \\min\\left(1, \\frac{n_j^{\\text{valid}}}{n_j^*}\\right), \\qquad \\mathrm{Efficiency} = \\frac{\\mathrm{Coverage}}{C_{\\text{generation}} + C_{\\text{review}} + C_{\\text{rework}} + C_{\\text{adjudication}}} \\tag{6.1}\n$$\n\nOnce a cell reaches its target of $`n_j^*`$ independent task families, further variants of simple templates add no coverage. Coverage is concave and piecewise linear in the family counts, so the allocation that maximizes it per unit cost fills cells in decreasing order of $`w_j / (n_j^* c_j)`$, with $`c_j`$ the cost of one valid family in cell $`j`$. Four controls enter the efficiency: manifold constraint enforcement, decision-margin label recomputation, separation of temporal information sets, and Batch Zero calibration.\n\nThe team averaged roughly 500 completed financial tasks a week under these rules, and the partner company carried Batch Zero into its other post-training programs.",
+ "markdown": "## 1 Distributional fit and task validity\n\nModern tabular generators address different distributional problems. [CTGAN](https://arxiv.org/abs/1907.00503) handles mixed discrete and continuous data, multimodality and class imbalance through mode-specific normalization and conditional vectors, and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) runs Gaussian diffusion on numerical fields and multinomial diffusion on categorical ones. Both learn empirical distributions effectively. A financial task can match its distribution and still be invalid, because part of what makes it valid is structural. Accounting identities have to hold exactly, fiscal periods have to run in chronological order, covenant applicability conditions have to be respected, and the numerical inputs have to agree logically with the reference judgment. Global distributional similarity metrics enforce none of this.\n\nA simple construction shows how far apart the two can drift. Suppose valid financial inputs $`x \\in \\mathbb{R}^d`$ have to satisfy $`m`$ linear accounting or ratio identities $`Ax = b`$, where $`A \\in \\mathbb{R}^{m \\times d}`$ has full row rank, so that the valid population distribution $`P`$ is supported entirely on the affine subspace $`M = \\{x \\in \\mathbb{R}^d : Ax = b\\}`$. Assume this distribution has finite second moments, and let the columns of $`N \\in \\mathbb{R}^{d \\times m}`$ form an orthonormal basis for the normal space orthogonal to $`M`$. Now take a generator that captures $`P`$ along the manifold but adds a small amount of independent noise in the normal directions: $`X_\\sigma = X + \\sigma N \\varepsilon`$, where $`\\varepsilon \\sim \\mathcal{N}(0, I_m)`$ is independent of $`X \\sim P`$. Writing $`Q_\\sigma`$ for the generated distribution, I obtain, for every $`\\sigma > 0`$,\n\n$$\nW_2(P, Q_\\sigma) \\le \\sqrt{E\\|X_\\sigma - X\\|^2} = \\sigma \\sqrt{m} \\longrightarrow 0 \\quad (\\text{as } \\sigma \\to 0^+) \\tag{1.1}\n$$\n\n$$\nP(M) = 1, \\qquad Q_\\sigma(M) = 0 \\tag{1.2}\n$$\n\n$$\n\\|P - Q_\\sigma\\|_{\\text{TV}} := \\sup_B \\left\\vert P(B) - Q_\\sigma(B)\\right\\vert = 1 \\tag{1.3}\n$$\n\nThe first line comes from coupling $`(X, X_\\sigma)`$ on the same probability space. The second holds because $`AN`$ is invertible and a continuous $`m`$-dimensional Gaussian vector equals zero with probability zero, and taking the Borel set $`B = M`$ then gives the third. So the 2-Wasserstein distance $`W_2(P, Q_\\sigma)`$ can tend to zero while the acceptance rate under exact business constraints stays at zero ($`Q_\\sigma(M) = 0`$) and the total variation distance stays at its maximum value of $`1`$. With a numerical tolerance the acceptance rate depends on the ratio of tolerance to off-manifold noise, and Section 2 shows that at a fixed ratio it falls geometrically with each added constraint.\n\nThe construction is one instance of a general gap. Synthetic data is evaluated on three axes, fidelity to the source distribution, utility for a downstream model and privacy, and the sample-level refinements of those metrics still score the inputs ([Alaa et al.](https://proceedings.mlr.press/v162/alaa22a.html)). None of the three has a place for the question a training team has to answer about supervision: is the label attached to this generated case still the right label? Data augmentation has the same blind spot from the other side. Counterfactual augmentation and contrast sets showed that a minimal edit can flip the gold answer, and they restore validity by having a person relabel every edited example ([Kaushik et al.](https://arxiv.org/abs/1909.12434); [Gardner et al.](https://aclanthology.org/2020.findings-emnlp.117/)), which does not scale to generated variants. I add the missing axis, supervision validity, and the remaining sections build the machinery that keeps it high at scale.\n\nThe counterexample also sets how I divided responsibilities in the synthetic data architecture ([SDV constraint-augmented generation](https://docs.sdv.dev/sdv/modeling/constraint-augmented-generation-cag)):\n• Copula models and conditional sampling capture cross-sectional rank and tail dependencies among continuous corporate indicators, including company size, operating history, leverage and profit margins. They preserve the dependencies that independent column generation would discard.\n• [CTGAN](https://arxiv.org/abs/1907.00503) and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) model mixed-type and multimodal empirical distributions, and explicit business-rule checks follow them.\n• Deterministic variable transformations and manifold projections enforce accounting equalities, ratio bounds and chronological order.\n\n## 2 Preserving the conditional distribution under business constraints\n\nRejection sampling generates unconstrained candidates and discards those that violate the constraints, and as several financial constraints tighten its computational cost becomes prohibitive. Suppose a candidate draw produces $`m`$ standardized constraint residuals that behave as independent standard normal variables, each of which has to lie within $`[-\\epsilon, \\epsilon]`$. The acceptance probability and the expected number of candidate draws per valid input are\n\n$$\n\\alpha_m(\\epsilon) = [2\\Phi(\\epsilon) - 1]^m \\sim \\left( \\sqrt{\\frac{2}{\\pi}} \\epsilon \\right)^m \\quad (\\epsilon \\to 0^+), \\qquad E[N_{\\text{attempts}}] = \\frac{1}{\\alpha_m(\\epsilon)} \\tag{2.1}\n$$\n\nso each additional accounting or ratio constraint multiplies the expected candidate count. With a tolerance of $`\\epsilon = 0.05`$ residual standard deviations, the exact analytical values are:\n\n**Table 1. Acceptance rate of rejection sampling under simultaneous constraints.**\n\n| Simultaneous constraints ($`m`$) | Rejection-sampling acceptance rate $`\\alpha_m(\\epsilon)`$ | Expected candidate draws per valid input |\n| --- | --- | --- |\n| **1** | 3.9878% | 25.08 |\n| **2** | 0.1590% | 628.84 |\n| **4** | **0.0002529%** | **395,442.59** |\n\nThe alternative is to condition instead of reject: draw once, then correct the draw onto the constraint. In the linear Gaussian setting the covariance-weighted correction returns a sample from the exact conditional distribution. For a Gaussian copula the result holds for constraints that are affine in the latent Gaussian coordinates, since nonlinear marginal transformations do not preserve affine constraints in the observed financial variables. Let the initial unconstrained draw be $`Z \\sim \\mathcal{N}(\\mu, \\Sigma)`$, with $`\\Sigma \\succ 0`$. I apply a covariance-weighted Mahalanobis projection:\n\n$$\nX = \\arg\\min_{x : Ax = b} \\frac{1}{2}(x - Z)^\\top \\Sigma^{-1}(x - Z) = Z + K(b - AZ), \\qquad K = \\Sigma A^\\top (A \\Sigma A^\\top)^{-1} \\tag{2.2}\n$$\n\nThe first-order Lagrangian stationarity condition is $`\\Sigma^{-1}(X - Z) + A^\\top \\lambda = 0`$, and substituting $`AX = b`$ and eliminating the multiplier $`\\lambda`$ gives the closed-form update. Since $`AK = I_m`$, every draw satisfies $`AX = b`$ exactly. Its mean and covariance are\n\n$$\nE[X] = \\mu + K(b - A\\mu) \\tag{2.3}\n$$\n\n$$\n\\mathrm{Cov}(X) = (I - KA)\\Sigma(I - KA)^\\top = \\Sigma - \\Sigma A^\\top (A \\Sigma A^\\top)^{-1} A\\Sigma \\tag{2.4}\n$$\n\nAffine transformations preserve Gaussianity, so these moments identify the exact regular conditional Gaussian distribution of $`Z \\mid AZ = b`$ ([Majumdar and Majumdar](https://arxiv.org/abs/1710.09285)). The update has a long lineage. It is the sampler for a normal distribution truncated on hyperplanes ([Cong, Chen and Zhou](https://doi.org/10.1214/17-BA1052)) and Matheron's rule for pathwise conditioning on a noise-free linear observation ([Wilson et al.](https://arxiv.org/abs/2011.04026)). What I did with it is put accounting identities in the place of the hyperplanes and state where exactness holds, in the latent coordinates of the copula. That also marks the division of labor with constraint layers for deep tabular generators, which enforce linear inequalities in any generator and make no statement about the law of the corrected sample ([Stoian et al.](https://arxiv.org/abs/2402.04823)). Inequalities such as ratio ranges go to transformations and layers; identities go to the projection, which returns the right distribution as well as a feasible point.\n\nOn a reference run of the constrained generator, **200,000 twelve-dimensional balance-sheet and income vectors subject to 4 simultaneous linear identities** (assets sum to total assets, liabilities sum to total liabilities, assets equal liabilities plus equity, and receivables stand at 60 days of revenue), the maximum absolute constraint residual was **$`3.0 \\times 10^{-15}`$**, and the sample covariance matched the theoretical conditional covariance to **0.359%** in relative Frobenius norm. Every draw is used; the expected 395,442.59 candidate draws per valid input in Table 1 are what rejection would have spent. Fields that accounting definitions fully determine are computed analytically. Range bounds and date ordering are enforced in the latent generation space when they admit bijective transformations, such as logarithmic transforms or positive increments $`t_{k} = t_{k-1} + \\exp(\\xi_k)`$.\n\n**The correction is also a monitor.** The squared Mahalanobis distance by which the projection moves a draw has a known law:\n\n$$\nT = (b - AZ)^\\top (A \\Sigma A^\\top)^{-1} (b - AZ) \\ \\sim\\ \\chi^2_m(\\delta), \\qquad \\delta = (b - A\\mu)^\\top (A \\Sigma A^\\top)^{-1} (b - A\\mu) \\tag{2.5}\n$$\n\nWhen the upstream generator is consistent with the identities, $`\\delta = 0`$ and $`T`$ averages $`m`$. When the upstream model drifts, $`\\delta`$ grows and the average of $`T`$ rises with it. Every batch therefore tests the fit of its own generator while it is being corrected. On the reference run the average of $`T`$ was 4.01 against $`m = 4`$. A 10% drift in the upstream mean of receivables raises $`\\delta`$ to 0.197 and the average to 4.20, which is two standard errors in a batch of 1,000 draws and more than six in a batch of 10,000.\n\n## 3 Using decision margins to determine when reference judgments need recomputation\n\nA synthetic variant of a financial reference case often keeps the business narrative and changes decisive financial figures, covenant thresholds or macroeconomic inputs. If the original reference answer stays in place after those values change, systematic label errors enter post-training, even when all accounting constraints hold. [FinQA](https://aclanthology.org/2021.emnlp-main.300/) pairs financial questions with executable reasoning programs, and synthetic preference data similarly require a maintained link between inputs, reference conclusions and the expert's explanation of a preference.\n\nNeighboring work links data generation to feedback in other ways. [WizardMath](https://arxiv.org/pdf/2308.09583v1) couples the evolution of mathematical instructions with reinforcement feedback, and [FinGPT](https://arxiv.org/pdf/2307.10485v1) fine-tunes financial language models on feedback from the market. Templated reasoning sets recompute an answer from the template's variables for every instance ([GSM-Symbolic](https://arxiv.org/abs/2410.05229)), which works when the answer is one formula and leaves evidence and preference untouched. I make the dependencies among a financial scenario, its reference conclusion, the evidence behind that conclusion and the expert preference built on it the object of design, so that a change to the scenario reaches each of them, and only the ones it has to.\n\nConsider a decision rule that chooses between two corporate actions, and let $`d(x)`$ denote the net valuation or metric advantage of action A over action B, so that the correct reference decision is $`\\mathrm{sign}\\,d(x)`$. If $`d`$ is $`L`$-Lipschitz on a feasible neighborhood, a sufficient condition for retaining the original reference decision is that the perturbation $`\\Delta`$ stays strictly within the decision margin:\n\n$$\n\\left\\vert d(x + \\Delta) - d(x)\\right\\vert \\le L\\|\\Delta\\|, \\qquad \\left\\vert d(x)\\right\\vert > L\\|\\Delta\\| \\implies \\mathrm{sign}\\,d(x + \\Delta) = \\mathrm{sign}\\,d(x) \\tag{3.1}\n$$\n\nThis condition certifies the decision label. A loan can stay approved while the calculation behind it, the supporting evidence and the relative quality of two responses all change, so I check reference calculations, evidence and pairwise preferences separately.\n\nMy diagnostic tasks focus on contested decision boundaries where models are prone to error, which means $`\\left\\vert d(x)\\right\\vert`$ is deliberately kept small. For a linear comparison rule $`d(x) = w^\\top x - c`$ with initial margin $`\\gamma = d(x) > 0`$, the stale-label error rate has an exact closed form. Let the perturbation within the feasible manifold satisfy $`\\Delta \\sim \\mathcal{N}(0, \\Omega)`$, with projected variance $`s^2 = w^\\top \\Omega w > 0`$. The probability that the correct label flips while the original decision is retained is\n\n$$\n\\Pr\\{d(x + \\Delta) < 0\\} = \\Pr\\{w^\\top \\Delta < -\\gamma\\} = \\Phi\\left(-\\frac{\\gamma}{s}\\right) \\tag{3.2}\n$$\n\nThe closed form gives the following inherited-label error rates:\n• At a margin-to-noise ratio of $`\\gamma / s = 2`$, the error rate is **2.28%**.\n• At $`\\gamma / s = 1`$, it is **15.87%**.\n• Near a diagnostic decision boundary, at $`\\gamma / s = 0.25`$, it is **40.13%**.\n\n**The formula is exact for covenants.** The linear rule is the common case in finance. A covenant test is a ratio of accounting levels against a threshold, interest coverage $`E / I \\ge c`$ or leverage $`D / E \\le c`$, and with a positive denominator it is the half-space $`E - cI \\ge 0`$ in the levels themselves. For the covenant family the flip probability $`\\Phi(-\\gamma / s)`$ is the exact rate under Gaussian perturbations of the levels, with $`w = (1, -c)`$. When several covenants bind at once, the case flips if any of them does, and the rate is bounded by the sum of the individual rates: two covenants at $`\\gamma / s`$ of 1.5 and 2 give at most 8.96%.\n\n**The certificate is a cache-hit test.** Treat the four objects as nodes of a dependency graph. A perturbed scenario marks its reference conclusion, evidence and preference as stale, and the margin decides which of them a builder may reuse. With a tolerated error $`\\alpha`$ on inherited labels:\n\n$$\n\\text{recompute the variant when } \\frac{\\gamma}{s} < z_{1 - \\alpha}, \\qquad \\text{inherit otherwise, with } \\Pr(\\text{stale label}) \\le \\alpha \\tag{3.3}\n$$\n\nThe thresholds are 1.645 at $`\\alpha = 5\\%`$ and 2.326 at $`\\alpha = 1\\%`$, and a variant at $`\\gamma / s = 3`$ inherits its label with an error of 0.13%. Variants far from the boundary cost no expert time, and the diagnostic variants near it, which are the ones the task family exists for, are always recomputed.\n\nI therefore tie case expansion to reference-answer recomputation and renewed five-level pairwise preference grading when perturbations approach or cross a decision boundary. In that event the workflow recomputes the reference solution and samples new candidate model reasoning trajectories. Domain experts then assess the pair on the five-level preference scale and identify which trajectory recognized the decisive variable whose change reversed the decision and supported its judgment with primary evidence. [HelpSteer2-Preference](https://arxiv.org/abs/2410.01257) gives a comparison point from the methods literature.\n\n**What a graded preference is worth.** A five-level judgment takes an expert a few seconds longer than a binary one, on a task that has already cost about three specialist-hours. Model the judgment as an ordered response to the latent quality gap $`g`$ between two responses, measured in units of judgment noise, with cutpoints $`c_1 < \\dots < c_4`$, $`c_0 = -\\infty`$ and $`c_5 = +\\infty`$. The Fisher information one response carries about the gap is\n\n$$\nI(g) = \\sum_{k} \\frac{\\big( \\partial_g\\, p_k(g) \\big)^2}{p_k(g)}, \\qquad p_k(g) = \\Phi(c_k - g) - \\Phi(c_{k-1} - g) \\tag{3.4}\n$$\n\n**Table 2. Fisher information of a binary and a five-level judgment by quality gap.**\n\n| Quality gap $`g`$ | Binary judgment | Five-level judgment (cutpoints at $`\\pm 0.5`$, $`\\pm 1.5`$) | Ratio |\n| --- | --- | --- | --- |\n| 0 (a tie) | 0.637 | 0.912 | 1.43 |\n| 0.5 | 0.581 | 0.902 | 1.55 |\n| 1.0 | 0.439 | 0.863 | 1.97 |\n\nOne graded judgment is worth 1.4 to 2 binary judgments, and the advantage grows with the gap, where a binary label saturates. The same expression designs the scale: at a tie the information is largest with cutpoints at $`\\pm 0.38`$ and $`\\pm 1.24`$, where it reaches 0.920 of the information in a continuous response. The protocol delivers three signals that a training pipeline can consume separately: the direction of the preference, its strength, and the cited evidence and decisive variable.\n\n## 4 Specifying the information available to each temporal task\n\nA cross-sectional company comparison does not require continuous-time stochastic processes. Dynamic tasks such as derivatives valuation, interest-rate term-structure interpolation and liquidity stress tests over multiple periods do require path generators. For those I use SDEs and neural SDEs to model path distributions beyond single-time marginals ([Neural SDEs as Infinite-Dimensional GANs](https://proceedings.mlr.press/v139/kidger21b.html)). Brownian bridges and stochastic interpolation support conditioning on known intermediate observations or endpoints, following the constructions in [Glasserman, *Monte Carlo Methods in Financial Engineering*, Chapter 3](https://doi.org/10.1007/978-0-387-21617-1). A bridge is the conditioning of Section 2 in one dimension: draw the free path, then correct it onto the known endpoint.\n\nThat conditioning can introduce temporal leakage into evaluation, and I derive the size of the effect in closed form. Consider scaled Brownian motion $`X_u = \\sigma W_u`$, where $`s < t < T`$. Under squared-error loss, the optimal predictor at time $`s`$, given only the historical filtration $`\\mathcal{F}_s`$, has conditional mean squared error $`\\sigma^2(t - s)`$. If the generator conditions paths on the future endpoint $`X_T`$ and inadvertently exposes that conditioning to a forecasting task, Gaussian conditioning gives ([Pitman on the Brownian bridge](https://www.stat.berkeley.edu/~pitman/s205s03/b_bridge.pdf)):\n\n$$\n\\mathrm{Var}(X_t \\mid \\mathcal{F}_s) = \\sigma^2(t - s) \\tag{4.1}\n$$\n\n$$\n\\mathrm{Var}(X_t \\mid \\mathcal{F}_s, X_T) = \\sigma^2(t - s) - \\frac{[\\sigma^2(t - s)]^2}{\\sigma^2(T - s)} = \\sigma^2(t - s)\\frac{T - t}{T - s} \\tag{4.2}\n$$\n\nAt the midpoint $`t = (s + T)/2`$, conditioning on $`X_T`$ halves the conditional variance. The resulting **29.29% apparent reduction in RMSE** ($`1 - 1/\\sqrt{2}`$) comes entirely from the additional endpoint information in this construction, and whether that counts as a flaw depends on the task. It is temporal leakage if the forecasting task does not permit access to that information. If instead the task gives the model both endpoints $`X_s`$ and $`X_T`$ and asks it to audit an interpolation or backcast an intervening accounting period, Brownian bridge conditioning is consistent with the specified information set.\n\nInterest rates and spreads revert to a mean, and for them the leak has its own closed form. For an Ornstein-Uhlenbeck path with mean-reversion speed $`\\kappa`$, at the midpoint of an interval of half-length $`h`$:\n\n$$\n\\frac{\\mathrm{Var}(X_t \\mid \\mathcal{F}_s, X_T)}{\\mathrm{Var}(X_t \\mid \\mathcal{F}_s)} = \\frac{1}{1 + e^{-2\\kappa h}}, \\qquad \\text{apparent RMSE reduction} = 1 - \\big( 1 + e^{-2\\kappa h} \\big)^{-1/2} \\tag{4.3}\n$$\n\n**Table 3. Apparent RMSE reduction from endpoint leakage under mean reversion.**\n\n| $`\\kappa h`$ | 0 (Brownian) | 0.25 | 0.5 | 1 | 2 |\n| --- | --- | --- | --- | --- | --- |\n| Apparent RMSE reduction | 29.29% | 21.10% | 14.50% | 6.15% | 0.90% |\n\nMean reversion discounts the leak, because the endpoint tells less about the middle of a path that forgets. The financial reading is direct: endpoint leakage is most damaging for series close to a unit root, prices and long rates, and nearly harmless for fast-reverting spreads. The information-set rule is therefore enforced with a severity that follows the persistence of the series.\n\nThe task specification keeps three information sets apart:\n1. **Construction information ($`\\mathcal{F}_{\\text{gen}}`$):** the complete path, endpoint or latent parameters used to generate a scenario.\n2. **Permitted predictor filtration ($`\\mathcal{F}_s`$):** the historical information the task allows the model to observe at decision time $`s`$.\n3. **Evaluation and oracle information ($`\\mathcal{F}_{\\text{eval}}`$):** outcomes realized after $`s`$, reserved for scoring or for explicitly disclosed oracle baselines.\n\n## 5 Batch Zero: independent execution and calibration for post-training data strategy\n\nI created **Batch Zero** around the capability objectives and data requirements of post-training research teams. Before anything scales, it brings task design, synthetic data strategy and preference supervision under a common set of acceptance criteria. It tests whether domain experts can independently execute a research specification, reproduce reference solutions and calibrate their preference judgments. The resulting specifications support subsequent data construction.\n• **Independent execution:** before a task family enters large-scale data construction (*Batch One*), doctoral-level domain experts independently write seed cases, generate synthetic variants, verify reference solutions and assign five-level pairwise preference scores. Implicit assumptions, missing information and ambiguities in the specification come to light this way.\n• **Calibration and feedback to data strategy:** failures, expert disagreements and boundary-case errors inform revisions to task instructions, reference solutions, grading rules and Batch One data acceptance criteria. Task design, execution, revision and recalibration form an iterative process.\n• **Organizational adoption:** the partner foundation model company subsequently adopted **Batch Zero** across other post-training programs.\n\n**Why the check belongs at the seed.** Expert-written references are less reliable than their reputation. When FinanceReasoning re-examined four public financial reasoning sets, it updated 15.6% of their questions, between 9.72% and 30.00% per set ([Tang et al.](https://arxiv.org/abs/2506.05828)). A synthetic pipeline multiplies whatever error its seeds carry. If a seed's reference solution is wrong with probability $`\\varepsilon`$ and the seed yields $`k`$ variants, a family carries $`k\\varepsilon`$ wrong records in expectation, and they are wrong together. If an independent re-solve catches a wrong reference with probability $`d`$, then $`r`$ re-solves leave\n\n$$\nk\\, \\varepsilon\\, (1 - d)^r \\tag{5.1}\n$$\n\nwrong records per family. With $`k = 20`$, $`\\varepsilon = 0.15`$ and $`d = 0.8`$, that is 3.0 with no re-solve, 0.6 after one and 0.12 after two. One re-solve at the seed does the work of reviewing twenty variants, and that multiplier is what Batch Zero is built on.\n\nFor statistical analysis and evaluation, I treat the task family as the basic unit of effective coverage and uncertainty estimation. Suppose $`G`$ independent reference-case families each yield $`k`$ synthetic variants, giving $`Gk`$ nominal records, and any two variants within a family have correlation $`\\rho`$ and common variance $`\\sigma_Y^2`$. The sample mean then carries the Kish design effect $`1 + (k - 1)\\rho`$, which gives the effective sample size $`N_{\\text{eff}}`$:\n\n$$\n\\mathrm{Var}(\\bar{Y}) = \\frac{Gk\\sigma_Y^2 + Gk(k - 1)\\rho\\sigma_Y^2}{(Gk)^2} = \\frac{\\sigma_Y^2}{Gk}[1 + (k - 1)\\rho] \\implies N_{\\text{eff}} = \\frac{Gk}{1 + (k - 1)\\rho} \\tag{5.2}\n$$\n\nWith $`G = 100`$ independent families, $`k = 20`$ variants per family (2,000 nominal tasks) and within-family correlation $`\\rho = 0.5`$, the 2,000 records provide the same precision for estimating the mean as about 190 independent records. Ignoring this correlation understates standard errors by a factor of **$`3.24\\times`$**. My data workflow therefore separates training and evaluation sets at the task-family level, and all confidence intervals account for hierarchical clustering at both the family and expert levels.\n\n**How many variants a seed should yield.** The design effect is a diagnosis; the same algebra gives the prescription. A new family costs $`c_G`$, mostly expert time on the seed and its reference, and each further variant costs $`c_v`$. At a fixed budget the variance of the family-clustered mean is proportional to $`(c_G + k c_v)[1 + (k - 1)\\rho] / k`$, and it is smallest at the optimum cluster size of two-stage sampling:\n\n$$\nk^* = \\sqrt{\\frac{c_G}{c_v} \\cdot \\frac{1 - \\rho}{\\rho}} \\tag{5.3}\n$$\n\nWith a seed that costs twenty variants and $`\\rho = 0.5`$, $`k^* \\approx 4.5`$, and the budget that buys 20 variants per family has 40% more variance than the same budget at four or five. Read the other way, moving from 20 variants per family to the optimum cuts the variance by 29% at no cost. Variants beyond $`k^*`$ buy repeated information, and the money is better spent on new families.\n\n## 6 Staged evaluation design and coverage efficiency\n\nStaged, blinded comparisons across independent task families test whether constraint-preserving synthesis and preference supervision meet the data requirements of post-training research teams. The task mix, generator training budget, expert review budget and final acceptance criteria stay fixed as each method is introduced:\n\n**Table 4. The staged evaluation design.**\n\n| Comparison stage | Method introduced | What the stage measures |\n| --- | --- | --- |\n| **Current data construction baseline** | Unconstrained generation plus expert review and rework | Fully loaded cost of each valid case that is ultimately accepted |\n| **Add manifold-constrained generation** | Hold the generator fixed; apply affine projections and variable transformations | Change in invalid candidates and manual repairs, with the conditional distribution preserved |\n| **Add decision-margin checks and reference-answer recomputation** | Recompute reference judgments below the margin threshold; use $`\\gamma/s`$ to quantify Gaussian label-flip risk | Change in stale-label errors, with valid supervision on both sides of the decision boundary |\n| **Apply Batch Zero acceptance criteria and filtration rules** | Jointly check the temporal information boundaries in $`\\mathcal{F}_s`$, independent execution, reproducibility of reference solutions and calibration of preference standards | Post-training gains that persist once leakage is excluded and clustering by source task family is accounted for |\n\nRaw acceptance rates favor trivial variants of easy templates, and generation speed alone omits the cost of expert repairs, so I allocate data construction and review resources by scenario-weighted task-family coverage efficiency. Let $`j`$ index predefined cells in the financial capability matrix, with strategic weights $`w_j \\ge 0`$ ($`\\sum_j w_j = 1`$) and required counts of independent valid task families $`n_j^* > 0`$:\n\n$$\n\\mathrm{Coverage} = \\sum_j w_j \\min\\left(1, \\frac{n_j^{\\text{valid}}}{n_j^*}\\right), \\qquad \\mathrm{Efficiency} = \\frac{\\mathrm{Coverage}}{C_{\\text{generation}} + C_{\\text{review}} + C_{\\text{rework}} + C_{\\text{adjudication}}} \\tag{6.1}\n$$\n\nOnce a cell reaches its target of $`n_j^*`$ independent task families, further variants of simple templates add no coverage. Coverage is concave and piecewise linear in the family counts, so the allocation that maximizes it per unit cost fills cells in decreasing order of $`w_j / (n_j^* c_j)`$, with $`c_j`$ the cost of one valid family in cell $`j`$. Four controls enter the efficiency: manifold constraint enforcement, decision-margin label recomputation, separation of temporal information sets and Batch Zero calibration.\n\nThe team averaged roughly 500 completed financial tasks a week under these rules, and the partner company carried Batch Zero into its other post-training programs.",
"sourceDocument": {
"fileName": "ai-research-and-engineering.md",
"path": "content/sources/ai-portfolio/ai-research-and-engineering.md",
- "sha256": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1"
+ "sha256": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0"
},
"repositoryUrl": "https://github.com/130U/RESEARCH-Financial-Task-Design-and-Synthetic-Data-Engineering"
}
diff --git a/website/content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json b/website/content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json
index d55003e..e55448f 100644
--- a/website/content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json
+++ b/website/content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json
@@ -3,11 +3,11 @@
"title": "General Reasoning Evaluation and Layered Verification",
"sourceUrl": "https://github.com/130U/RESEARCH-General-Reasoning-Evaluation-and-Layered-Verification/blob/main/manuscript/general-reasoning-evaluation-and-layered-verification.md",
"sourceEditedAt": "2026-11-19",
- "markdown": "## 1 Correct answers can conceal invalid reasoning and reverse model rankings\n\nChecking the final answer is enough only when the answer is the entire deliverable. In the samples analyzed for *ProcessBench*, the share of solutions with correct final answers that still contain a process error rises with difficulty: 3.5% on GSM8K, 18.8% on MATH, 32.2% on OlympiadBench and 51.8% on Omni-MATH (*[ProcessBench, Section 3.3](https://arxiv.org/html/2412.06559v2)*). Answer accuracy and the validity of a reasoning trace are distinct evaluation targets, and they come apart fastest on the problems that matter most. For a task that also requires a derivation or a specified execution process, I examine domain assumptions, method applicability, supporting evidence, tool compliance, intermediate results and the conclusion. I treat the model's displayed CoT and visible reasoning as observable outputs and cross-check them against execution records.\n\nTo see what this does to a score, suppose problems and model trajectories are drawn from a target capability distribution. Let $`Z = 1`$ indicate that the complete solution, including its reasoning steps and tool execution, is valid, and let $`V = 1`$ indicate that the evaluation protocol accepts it. For model $`m`$, define its true validity rate as $`\\theta_m = \\Pr_m(Z = 1)`$, its false acceptance rate as $`\\alpha_m = \\Pr_m(V = 1 \\mid Z = 0)`$, and its false rejection rate as $`\\beta_m = \\Pr_m(V = 0 \\mid Z = 1)`$. False acceptance can let through an unsupported derivation, or a tool call that produces the correct answer while violating a predefined execution requirement; false rejection can exclude a valid alternative derivation. By the law of total probability:\n\n$$\ns_m = E_m[V] = \\alpha_m + (1 - \\alpha_m - \\beta_m)\\theta_m, \\qquad s_m - \\theta_m = \\alpha_m(1 - \\theta_m) - \\beta_m\\theta_m \\tag{1.1}\n$$\n\nThe term $`\\alpha_m(1 - \\theta_m)`$ credits invalid trajectories, while $`-\\beta_m\\theta_m`$ penalizes valid ones. Adding problems reduces the sampling variance of $`\\hat{s}_m`$ and leaves the structural gap $`s_m - \\theta_m`$ in place. Models also differ in how closely they adhere to evidence requirements, verification steps and tool permissions, and the resulting differences in $`\\alpha_m`$ and $`\\beta_m`$ can reverse their ranking. Suppose Models A and B have true validity rates of $`\\theta_A = 72\\%`$ and $`\\theta_B = 64\\%`$. If invalid trajectories pass the final-answer check with probabilities $`\\alpha_A = 4\\%`$ and $`\\alpha_B = 36\\%`$, with $`\\beta_A = \\beta_B = 0\\%`$, their observed scores are $`s_A = 73.12\\%`$ and $`s_B = 76.96\\%`$, and the model with the lower rate of fully valid solutions ranks higher. Any $`\\alpha_B`$ above 25.3% reverses this pair.\n\nPublic leaderboards have shown the effect at scale. When EvalPlus extended HumanEval's tests 80-fold, pass@k fell by up to 19.3% to 28.9%, models changed places, and 18 of the original reference solutions, 11% of the problems, were found defective ([EvalPlus](https://arxiv.org/abs/2305.01210)).\n\n**When a lead is certified.** Equation (1.1) can be read in the other direction. If a grader's false acceptance is known to be at most $`\\bar{\\alpha}`$ and its false rejection at most $`\\bar{\\beta}`$, a score pins true validity to an interval, and with $`\\bar{\\beta} = 0`$ a lead of $`g = s_A - s_B`$ certifies the order of true validity under one condition:\n\n$$\n\\frac{s - \\bar{\\alpha}}{1 - \\bar{\\alpha}} \\le \\theta \\le \\frac{s}{1 - \\bar{\\beta}}, \\qquad \\theta_A > \\theta_B \\ \\text{ for every admissible grader} \\iff \\bar{\\alpha} < \\frac{g}{1 - s_B} \\tag{1.2}\n$$\n\nThe condition is tight: at any larger $`\\bar{\\alpha}`$ there are error rates within the bound that reverse the order. A lead of three points over a rival scoring 80% is certified only when the grader's false acceptance is below 15%; over a rival scoring 50%, below 6%. The harder the evaluation, the better the grader has to be before a small lead means anything. This turns the reversal above from a warning into a test that can be applied to any published ranking whose grader has a measured error rate.\n\nProcess supervision extends evaluation to intermediate steps (*[Let's Verify Step by Step, Sections 2 and 3](https://arxiv.org/html/2305.20050v1)*). When I apply it to general post-training evaluation, I account for omissions in reference solutions, accept valid alternative methods, and enforce evidence and execution requirements defined before responses are generated. Structural simplification can explain a method or reveal redundant work. A slower solution, or one that takes more attempts, is still a correct solution, and a resource-related rejection requires a hard limit specified in advance.\n\n## 2 Proof obligations organize layered review and allow alternative solutions\n\nA task specification defines the domain assumptions and execution constraints $`\\Gamma_q`$, the target conclusion $`C_q`$, and the proof obligations needed to establish it. A model trajectory $`\\tau`$, which here means its visible reasoning and observable execution records, may get there through different intermediate lemmas, representations or orders of deduction. That is allowed, because review concerns the submitted material and whether it fulfills the task's obligations. A directed acyclic dependency graph (DAG) $`G`$ organizes the local soundness check for each node $`v_j`$, together with a global tool-compliance predicate $`\\mathcal{T}_{\\text{comp}}(\\tau) \\in \\{0, 1\\}`$:\n\n$$\n\\Gamma_q \\cup \\{v_k : k \\in \\mathrm{pa}(j)\\} \\vdash v_j, \\qquad \\Gamma_q \\cup \\{v_j : j \\in \\mathrm{sink}(G)\\} \\vdash C_q, \\qquad \\mathcal{T}_{\\text{comp}}(\\tau) = 1 \\tag{2.1}\n$$\n\nThe symbol $`\\vdash`$ denotes valid derivability within a sound domain inference system, and the condition $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$ certifies that each external tool invocation $`t \\in \\tau`$ belongs to the task's permitted execution environment $`\\mathcal{A}_q`$. If root nodes follow from verified assumptions $`\\Gamma_q`$, every dependency is sound, with its local derivation verified in that system, and $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$, structural induction over a topological ordering of $`G`$ establishes the full conclusion.\n\nThis is the discipline of certifying algorithms carried over to model output. The producer may be arbitrarily complex and is not trusted; it has to hand over a witness, and the checker of that witness is simple enough to be trusted ([McConnell et al.](https://doi.org/10.1016/j.cosrev.2010.09.009)). A model is the least trusted producer there is, which makes the checker the place to spend rigor.\n\nIn production, I divided verification into three layers:\n1. Code and rule checks cover the directly verifiable parts of an answer, required deliverables, and predefined restrictions on tools, execution environments and resources. Tool compliance requires $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$, and a call that violates an explicit hard requirement causes rejection ($`V = 0`$) even when the final answer is correct.\n2. LLM-as-a-Judge assists with semantic screening. It flags insufficient evidence, misapplied rules, conclusions that exceed their premises and gaps in reasoning, and it links each concern to the relevant trace, tool response or intermediate output. The flags direct further review.\n3. Scholars with relevant expertise take the questions that require domain judgment, admit several reasonable interpretations, or remain disputed after the first two checks. They examine the evidence, the applicability of the method and the conclusion, and a question that cannot be resolved from the available evidence stays open.\n\nAll three layers work from the same observable material. I cross-check the model's displayed CoT and visible reasoning against actual tool calls, intermediate outputs, elapsed time and attempt counts. For each attempt I look for new evidence, revised premises or a change in the state of the solution. That is how I identify repeated calls, continued use of failed assumptions and loops that make no progress. A conclusion whose execution records are missing is recorded as unverified. I record efficiency separately from correctness and compliance: long runtimes, repeated attempts or exhaustive methods are analyzed as efficiency, and they affect acceptance only when they violate a predefined hard limit.\n\nMethod applicability has its own check, because pointwise convergence at a fixed parameter value does not establish that an approximation remains valid within an optimization, an integral or an exchange of limits. A claim about the order of accuracy requires an approximation parameter, its domain of validity and a uniform error bound. Where such an issue cannot be decided programmatically, an LLM helps locate the concern and a specialist scholar reviews the assumptions and derivation. Specialist review settles the question with the standing of expert judgment; a certificate settles it by computation.\n\nFor feedback, I record the first failed obligation at node $`v_j`$ and mark its descendant subgraph $`\\mathrm{desc}(v_j)`$ as unverified, since locally correct calculations under a false premise do not repair the full derivation. This treatment resolves the ambiguity after an error discussed in *[ProcessBench, Section 3.1](https://arxiv.org/html/2412.06559v2)*. A suspected error awaiting specialist review remains unresolved until that review is complete.\n\n**When a screen may decide.** The division between the second and third layers has a quantitative form. Let $`\\text{se}`$ be the screen's sensitivity on semantic obligations, the share of invalid steps it flags. Flagged work goes to specialists, and unflagged work goes to specialists with probability $`\\rho`$. With the specialists' verdict as the reference, the system's false acceptance on those obligations, and the random escalation needed to hold it at a target $`\\alpha^*`$, are:\n\n$$\n\\alpha_{\\text{sys}} = (1 - \\text{se})(1 - \\rho), \\qquad \\rho \\ge 1 - \\frac{\\alpha^*}{1 - \\text{se}} \\tag{2.2}\n$$\n\nThe best model tested by [Tyen et al.](https://arxiv.org/abs/2311.08516) located the first mistake in 52.87% of reasoning traces. At a sensitivity of that order and a target of $`\\alpha^* = 5\\%`$, 89% of unflagged work still has to reach a specialist, and the screen saves almost no expert time. At $`\\text{se} = 0.80`$ the share is 75%, at $`0.90`$ it is 50%, and only at $`\\text{se} \\ge 0.95`$ can unflagged work pass without review. That is the rule behind the layer boundary: an LLM judge screens until its measured sensitivity clears $`1 - \\alpha^*`$, and decides after. The randomly escalated share is also the calibration sample, because a screen whose unflagged output is never reviewed has no measured sensitivity at all. A checker earns decisive standing by a clean run: 59 known-invalid items rejected without exception bound its false acceptance at 5% with 95% confidence, and 299 bound it at 1%.\n\nRelated work approaches these questions from several directions. [ToolEval](https://arxiv.org/html/2307.16789v1) scores tool-using agents by pass rate and win rate, both assigned by an LLM evaluator, so its verdicts carry the standing of a screen. [Let's Verify Step by Step](https://arxiv.org/html/2305.20050v1) collected step-level human labels and withheld reference solutions so that labelers would not favor one route. [Math-Shepherd](https://arxiv.org/html/2312.08935v1) estimates the quality of a step from whether sampled continuations reach the correct answer, which scores a step by its outcome; the certificates of Section 3 check the mathematical validity of the evidence a submission actually presents. [ProcessBench](https://arxiv.org/html/2412.06559v2) has since measured how often correct answers conceal process errors on hard problems. My contribution is the organization of acceptance itself, through explicit proof obligations, a dependency structure over them and conditions that can be checked.\n\n## 3 Error certificates that need no reference solution\n\nA posteriori verification rests on one observation: to bound how far a candidate is from the truth, it is enough to have a residual the candidate exposes and a stability constant of the problem. The true solution never enters. I develop the template in full on an optimization family whose constraints and certificates can be checked explicitly, and then state it for the other task types it covers.\n\nConsider the equality-constrained quadratic problem:\n\n$$\nv^* = \\min_{Ax = b} f(x), \\qquad f(x) = \\frac{1}{2}x^\\top H x + c^\\top x, \\qquad Z^\\top H Z \\succeq \\mu I, \\quad \\mu > 0 \\tag{3.1}\n$$\n\nHere $`H`$ is symmetric, $`Ax = b`$ is feasible, the columns of $`Z`$ form an orthonormal basis of $`\\ker A`$, and $`\\mu`$ is a verified lower bound on curvature along feasible directions. The objective has to be strongly convex only on the feasible manifold; $`H`$ may be indefinite across the constraints. For any feasible candidate $`x`$ satisfying $`Ax = b`$, choose any dual multiplier vector $`\\nu`$ and define the stationarity residual $`r = Hx + c + A^\\top \\nu`$. Curvature gives the first inequality below for every feasible displacement $`d \\in \\ker A`$, and completing the square in $`d`$ gives an exact optimality error certificate:\n\n$$\nf(x + d) \\ge f(x) + r^\\top d + \\frac{\\mu}{2}\\|d\\|^2 \\tag{3.2}\n$$\n\n$$\n\\inf_{d \\in \\ker A} \\left( r^\\top d + \\frac{\\mu}{2}\\|d\\|^2 \\right) = -\\frac{\\|P_{\\ker A} r\\|_2^2}{2\\mu} \\tag{3.3}\n$$\n\n$$\n0 \\le f(x) - v^* \\le \\frac{\\|P_{\\ker A} r\\|_2^2}{2\\mu} \\tag{3.4}\n$$\n\nThe operator $`P_{\\ker A}`$ is the orthogonal projection onto the space of feasible directions $`\\ker A`$. It removes residual components normal to the constraints, which do not affect optimality along the feasible manifold (*[Boyd and Vandenberghe, Convex Optimization, Sections 5.5 and 9.1](https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf)*). A zero projected residual, $`P_{\\ker A} r = 0`$, certifies exact optimality, and otherwise the certificate bounds the suboptimality gap in objective units. Two kinds of submission fare differently under this check:\n• Solutions derived through Lagrange multipliers or by eliminating the equality constraints satisfy the same semantic obligations and pass verification.\n• Reporting the correct optimal value $`v^*`$ alone leaves the obligations undischarged. Feasibility and the claimed objective value are checked separately. A submission that claims a KKT multiplier witness must also have its full stationarity equation $`Hx+c+A^\\top\\nu=0`$ verified, because the projected-residual bound cannot validate that multiplier: $`P_{\\ker A}A^\\top\\nu=0`$ for every $`\\nu`$. A valid equality-elimination proof need not supply a multiplier witness. Whether a checker validates the whole of a claimed witness is what the formal verification of certifying computations examines ([Alkassar et al.](https://arxiv.org/abs/1301.7462)), and Section 6 tests for it directly.\n\nThe same three parts, a residual, a verified constant and an inequality between them, give a certificate for each of the following task types.\n\n**Table 1. Error certificates by task type.**\n\n| Task type | Residual the submission exposes | Constant the task designer verifies | Certificate |\n| --- | --- | --- | --- |\n| Equality-constrained quadratic program (above) | $`P_{\\ker A} r`$ | $`\\mu = \\lambda_{\\min}(Z^\\top H Z)`$ | $`f(x) - v^* \\le \\lVert P_{\\ker A} r\\rVert _2^2 / (2\\mu)`$ |\n| Convex program | A primal-feasible $`x`$ and a dual-feasible $`(\\lambda, \\nu)`$ | None | $`f(x) - v^* \\le f(x) - g(\\lambda, \\nu)`$ |\n| Linear system $`Ax = b`$ | $`b - Ax`$ | A lower bound on $`\\sigma_{\\min}(A)`$ | $`\\lVert x - x^*\\rVert _2 \\le \\lVert b - Ax\\rVert _2 / \\sigma_{\\min}(A)`$ |\n| Eigenpair of a symmetric matrix | $`A\\hat{v} - \\hat{\\lambda}\\hat{v}`$ | None | $`\\min_i \\left\\lvert \\lambda_i - \\hat{\\lambda}\\right\\rvert \\le \\lVert A\\hat{v} - \\hat{\\lambda}\\hat{v}\\rVert _2 / \\lVert \\hat{v}\\rVert _2`$ |\n| Fixed point of a contraction $`g`$ | $`g(x) - x`$ | A Lipschitz constant $`L < 1`$ | $`\\lVert x - x^*\\rVert \\le \\lVert g(x) - x\\rVert / (1 - L)`$ |\n\nEach row is a standard result of its field. What I take from them is the organization: every obligation type is paired with a certificate and a checker whose cost is a few matrix products, and acceptance becomes a computation on the submission alone. That removes the reference solution from the trusted base. Reference solutions are a source of grading error in their own right, as the 18 defective references found by [EvalPlus](https://arxiv.org/abs/2305.01210) show, and a grader that matches against them inherits every one.\n\nI carry the lessons from this family over to the structure of general review: check the inputs and assumptions, compliance with the execution specification, support for the derivation, and whether the conclusion exceeds the evidence. Obligations that admit programmatic checks go to rules or certificate verifiers, and other semantic questions go through LLM-assisted screening and specialist review. Structural simplification may explain why a method needs less repeated work, and efficiency gains and validity certificates are evaluated separately.\n\n## 4 Task transformations must propagate error bounds\n\nSymbolic templates parameterize variable domains and answer formulas (*[GSM-Symbolic, Section 3.1](https://arxiv.org/html/2410.05229v1)*). Tasks with formalizable transformations also need verifiable rules for how those transformations affect feasible sets, objective values and acceptance tolerances. For bijective task transformations the tolerance-propagation rule below follows.\n\nLet an original problem and its transformed variant have feasible sets $`F`$ and $`F'`$, objectives $`J`$ and $`J'`$, and finite infima $`v`$ and $`v'`$. Suppose the task designer verifies a bijection $`\\Phi : F \\to F'`$ such that every feasible $`u \\in F`$ satisfies the uniform bound:\n\n$$\n\\left\\vert J'(\\Phi(u)) - aJ(u) - b\\right\\vert \\le \\varepsilon, \\quad a > 0 \\tag{4.1}\n$$\n\n$$\nav + b - \\varepsilon \\le v' \\le av + b + \\varepsilon \\tag{4.2}\n$$\n\n$$\nJ(u) - v \\le \\delta \\implies J'(\\Phi(u)) - v' \\le a\\delta + 2\\varepsilon \\tag{4.3}\n$$\n\nTaking infima in the first line gives the second, and subtracting the lower bound on $`v'`$ from the upper bound on $`J'(\\Phi(u))`$ gives the third. So after a unit conversion, parameter rescaling or numerical transformation with error bounded by $`\\varepsilon`$, the result determines the acceptance tolerance for a previously $`\\delta`$-optimal solution. Keeping the original tolerance $`\\delta`$ can misclassify a valid transformed solution as a model failure, while relaxing it arbitrarily can conceal a real capability gap. When $`\\varepsilon = 0`$, the bijection and positive scaling $`a > 0`$ also preserve whether an optimum is attained. Changes to boundary compactness or admissible control classes require a new attainment proof.\n\nVariants are usually built by several transformations in a row. Applying the bound step by step to a chain with parameters $`(a_1, \\varepsilon_1), \\dots, (a_k, \\varepsilon_k)`$ gives the composition law:\n\n$$\na = \\prod_{j=1}^{k} a_j, \\qquad \\varepsilon = \\sum_{j=1}^{k} \\Big( \\prod_{l > j} a_l \\Big) \\varepsilon_j, \\qquad J(u) - v \\le \\delta \\implies J^{(k)} - v^{(k)} \\le a\\delta + 2\\varepsilon \\tag{4.4}\n$$\n\nEach error is magnified by every scaling applied after it, so the order of a chain is part of its specification. Take an exact unit conversion with $`a = 100`$, an exact rescaling with $`a = 0.5`$, and a numerical discretization with $`\\varepsilon = 0.003`$, applied to a solution that is $`\\delta = 0.01`$ optimal. With the discretization last, the chain has $`(a, \\varepsilon) = (50, 0.003)`$ and the tolerance owed is $`0.506`$. With the discretization first, it has $`(50, 0.15)`$ and the tolerance is $`0.8`$. A grader that reuses the first tolerance for the second chain rejects valid work.\n\n## 5 Grading errors change the post-training objective\n\nStep-level labels can improve post-training only if they measure logical validity and compliance, because systematic grading errors pass into the targets used by process reward models (PRMs) and RLHF policies. Let $`h`$ denote the task input and the model's visible reasoning prefix through the current step. Let $`Z = 1`$ indicate that the step is valid under the problem's assumptions, and let $`V \\in \\{0, 1\\}`$ be the grader's noisy acceptance label, with conditional false acceptance $`\\alpha(h)`$ and false rejection $`\\beta(h)`$. Writing $`p(h) = \\Pr(Z = 1 \\mid h)`$, a process evaluator with logit $`s`$ trained by binary cross-entropy is optimized toward the corrupted conditional target:\n\n$$\n\\tilde{p}(h) = \\alpha(h) + [1 - \\alpha(h) - \\beta(h)]p(h) \\tag{5.1}\n$$\n\n$$\n\\frac{\\partial \\ell(s, V)}{\\partial s} = \\sigma(s) - V \\tag{5.2}\n$$\n\n$$\nE[\\nabla_s \\ell(s, V) - \\nabla_s \\ell(s, Z) \\mid h] = p(h) - \\tilde{p}(h) \\tag{5.3}\n$$\n\nThe expected gradient difference shows which way grading bias moves the training target:\n• If false rejection dominates, $`\\beta(h)p(h) > \\alpha(h)[1-p(h)]`$, then $`\\tilde{p}(h) < p(h)`$. The learned validity target shifts downward, which penalizes valid alternative proofs.\n• If false acceptance dominates, $`\\alpha(h)[1-p(h)] > \\beta(h)p(h)`$, then $`\\tilde{p}(h) > p(h)`$. The target shifts upward, rewarding flawed proofs or noncompliant tool calls. When the weighted contributions are equal, the effects cancel in the conditional mean.\n\nWhen $`\\alpha, \\beta`$ are known constants with $`\\alpha + \\beta < 1`$, class-conditional noise correction removes the bias from the mean estimate (*[Learning with Noisy Labels, Sections 2 and 3](https://papers.nips.cc/paper_files/paper/2013/file/3871bd64012152bfb53fdf04b401193f-Paper.pdf)*). The estimator is the prevalence correction of [Rogan and Gladen](https://pubmed.ncbi.nlm.nih.gov/623091/), and it increases variance by the squared factor below:\n\n$$\n\\hat{\\theta}_{\\text{corr}} = \\frac{\\bar{V} - \\alpha}{1 - \\alpha - \\beta}, \\qquad \\mathrm{Var}(\\hat{\\theta}_{\\text{corr}}) = \\frac{\\mathrm{Var}(\\bar{V})}{(1 - \\alpha - \\beta)^2} \\tag{5.4}\n$$\n\nFor task families whose validity is exactly decidable by a sound and complete certificate verifier, $`\\alpha=\\beta=0`$ and there is nothing to correct.\n\n**What a policy learns from a grader.** Equations (5.1) and (5.4) describe a grader that is only observed. Once its verdict is the reward, the policy moves toward whatever the grader accepts. Model that as reweighting the policy's trajectories by $`e^{\\eta V}`$, with $`\\eta`$ the strength of optimization. A policy that starts at true validity $`\\theta`$ then has:\n\n$$\n\\theta_\\eta = \\frac{\\theta(1-\\beta)e^{\\eta} + \\theta\\beta}{\\big[\\theta(1-\\beta) + (1-\\theta)\\alpha\\big]e^{\\eta} + \\theta\\beta + (1-\\theta)(1-\\alpha)} \\ \\xrightarrow{\\ \\eta \\to \\infty\\ } \\ \\frac{\\theta(1-\\beta)}{\\theta(1-\\beta) + (1-\\theta)\\alpha} \\tag{5.5}\n$$\n\nThe limit is the grader's positive predictive value at the starting policy, and the observed score goes to 100% along the way. For Model B of Section 1 under final-answer matching:\n\n**Table 2. Observed score of Model B under final-answer matching as optimization strengthens.**\n\n| Optimization strength $`\\eta`$ | 0 | 1 | 2 | 4 | Limit |\n| --- | --- | --- | --- | --- | --- |\n| Observed score | 77.0% | 90.1% | 96.1% | 99.5% | 100% |\n| True validity | 64.0% | 74.9% | 79.9% | 82.7% | 83.2% |\n\nTwo results follow, one for each kind of error.\n\n• *False acceptance sets the ceiling.* The policy can reach the grader's positive predictive value and no more: 83.2% for Model B, 98.5% for Model A. With $`\\alpha = 0`$ the ceiling is 100% for any $`\\beta < 1`$.\n• *False rejection removes routes.* A valid route that the grader always rejects loses its share of valid work at rate $`e^{-\\eta}`$. If two valid routes start equally common and the grader accepts only one, the other falls from 50% of valid work to 26.9%, 11.9% and 1.8% at $`\\eta = 1, 2, 4`$. Training against a reference path teaches the model to stop producing correct proofs the reference did not anticipate.\n\nThe cost of false acceptance is the ceiling and the cost of false rejection is diversity. Certificate checking has neither cost on the families it decides, which is the argument for drawing process rewards from certified nodes. It also locates the limit of a theoretical result sometimes read as making process supervision unnecessary: [Jia, Rakhlin and Xie](https://arxiv.org/abs/2502.10581) show that outcome supervision is statistically no harder than process supervision, up to polynomial factors in the horizon, when the outcome reward is the objective. When a correct answer can rest on an invalid derivation, the outcome reward is wrong about validity, and a policy trained on it converges to the limit in equation (5.5).\n\n## 6 A test set with exactly known verdicts\n\nMeasuring a grader's two error rates requires submissions whose validity is known without appeal to any grader. I built such a test set from 600 strongly convex quadratic task families. Using exact rational arithmetic, each family yields 6 algebraic submissions, for 3,600 in total:\n• 1,200 valid certificates, 2 per family: one derived using Lagrange multipliers and one using equality elimination.\n• 2,400 submissions with controlled defects, 4 per family: (1) an infeasible construction, $`Ax \\neq b`$, that yields the exact optimal value $`v^*`$; (2) failure of stationarity along a feasible direction in $`\\ker A`$ while reporting $`v^*`$; (3) a claimed multiplier that fails the stationarity equation while reporting $`v^*`$; and (4) an incorrect final objective value. The first three classes, comprising 1,800 submissions or 75% of all invalid submissions, report the correct final answer $`v^*`$.\n\nThe composition is the instrument's design: three ways of being wrong with the right answer, one way of being wrong with the wrong answer, and two ways of being right. I compared three deterministic acceptance protocols on all 3,600 submissions:\n\n**Table 3. Three acceptance protocols on 3,600 submissions with known verdicts.**\n\n| Acceptance protocol | False acceptance rate on invalid submissions ($`2,400`$) | False rejection rate on valid certificates ($`1,200`$) | Valid share of accepted submissions |\n| --- | --- | --- | --- |\n| Final-answer matching only | 75.0% (accepts 1,800 flawed solutions) | 0.0% | 40.0% |\n| Strict reference-path matching | 0.0% | 50.0% (rejects 600 valid elimination proofs) | 100.0% |\n| Certificate verification | 0.0% | 0.0% | 100.0% |\n\nCertificate verification accepted 1,800 fewer invalid proofs than final-answer matching and retained 600 more valid alternative solutions than strict reference-path matching. Strict error detection and acceptance of valid alternative proofs held together on the same submissions, which neither baseline managed. The verifier also passed all 600 feasible-perturbation checks on the projected-residual bound in Equation (3.2), and it rejected every submission of defect class (3), the multiplier that a partial witness check would have let through.\n\nThe last column is the ceiling of Section 5. A policy trained against final-answer matching on this mix moves from one third valid work to 40% and stops. Trained against the reference path it reaches 100% and loses the elimination route. Trained against certificate checking it reaches 100% with both routes intact.\n\n## 7 Dependence within task families and the economics of certificates\n\nSurface variations of one problem template do not provide independent observations of general reasoning ability. Suppose there are $`F`$ independent task families, each with $`K`$ variants, common marginal variance $`\\sigma^2`$, and common pairwise within-family correlation $`\\rho`$. The sample mean then carries the Kish design effect $`1 + (K - 1)\\rho`$, and its variance and effective sample size $`n_{\\text{eff}}`$ are:\n\n$$\n\\mathrm{Var}(\\bar{Z}) = \\frac{\\sigma^2}{FK}[1 + (K - 1)\\rho], \\qquad n_{\\text{eff}} = \\frac{FK}{1 + (K - 1)\\rho} \\tag{7.1}\n$$\n\nFor $`F = 50`$ families with $`K = 8`$ variants each, there are $`400`$ responses, and if $`\\rho = 0.6`$, those $`400`$ responses provide the same precision as $`n_{\\text{eff}} \\approx 77`$ independent observations. Training and test sets are therefore split by task family, and new tasks vary assumptions, evidentiary relationships or the structure of a derivation, beyond substituting numbers.\n\nCertificates cost engineering once and almost nothing afterwards. Let $`C_0`$ be the upfront engineering cost of the proof-obligation certificates and verifier, $`c_e`$ the expert review cost per submission, $`c_v`$ the automated verifier's unit execution cost, and $`u`$ the fraction of submissions escalated to human adjudication at the same expert unit cost. The certificate workflow is the cheaper one when submission volume $`N`$ exceeds:\n\n$$\nC_0 + N(c_v + uc_e) < Nc_e \\iff N > \\frac{C_0}{(1 - u)c_e - c_v}, \\qquad (1 - u)c_e > c_v \\tag{7.2}\n$$\n\nFor a certificate family whose engineering costs as much as 400 expert reviews, with 15% of submissions escalated and a verifier run costing 2% of a review, the break-even is 482 submissions. Past that volume every further submission releases 83% of an expert review, and those hours go to the semantic obligations that only specialists can settle. The cost model covers the certificate layer; model calls and calibration are costed with the layers they belong to.\n\nThe team produced roughly 500 scientific reasoning tasks a week under this design, and later programs reused its layered structure.",
+ "markdown": "## 1 Correct answers can conceal invalid reasoning and reverse model rankings\n\nChecking the final answer is enough only when the answer is the entire deliverable. In the samples analyzed for *ProcessBench*, the share of solutions with correct final answers that still contain a process error rises with difficulty: 3.5% on GSM8K, 18.8% on MATH, 32.2% on OlympiadBench and 51.8% on Omni-MATH (*[ProcessBench, Section 3.3](https://arxiv.org/html/2412.06559v2)*). Answer accuracy and the validity of a reasoning trace are distinct evaluation targets, and they come apart fastest on the problems that matter most. For a task that also requires a derivation or a specified execution process, I examine domain assumptions, method applicability, supporting evidence, tool compliance, intermediate results and the conclusion. I treat the model's displayed CoT and visible reasoning as observable outputs and cross-check them against execution records.\n\nTo see what this does to a score, suppose problems and model trajectories are drawn from a target capability distribution. Let $`Z = 1`$ indicate that the complete solution, including its reasoning steps and tool execution, is valid, and let $`V = 1`$ indicate that the evaluation protocol accepts it. For model $`m`$, define its true validity rate as $`\\theta_m = \\Pr_m(Z = 1)`$, its false acceptance rate as $`\\alpha_m = \\Pr_m(V = 1 \\mid Z = 0)`$, and its false rejection rate as $`\\beta_m = \\Pr_m(V = 0 \\mid Z = 1)`$. False acceptance can let through an unsupported derivation, or a tool call that produces the correct answer while violating a predefined execution requirement; false rejection can exclude a valid alternative derivation. By the law of total probability:\n\n$$\ns_m = E_m[V] = \\alpha_m + (1 - \\alpha_m - \\beta_m)\\theta_m, \\qquad s_m - \\theta_m = \\alpha_m(1 - \\theta_m) - \\beta_m\\theta_m \\tag{1.1}\n$$\n\nThe term $`\\alpha_m(1 - \\theta_m)`$ credits invalid trajectories, while $`-\\beta_m\\theta_m`$ penalizes valid ones. Adding problems reduces the sampling variance of $`\\hat{s}_m`$ and leaves the structural gap $`s_m - \\theta_m`$ in place. Models also differ in how closely they adhere to evidence requirements, verification steps and tool permissions, and the resulting differences in $`\\alpha_m`$ and $`\\beta_m`$ can reverse their ranking. Suppose Models A and B have true validity rates of $`\\theta_A = 72\\%`$ and $`\\theta_B = 64\\%`$. If invalid trajectories pass the final-answer check with probabilities $`\\alpha_A = 4\\%`$ and $`\\alpha_B = 36\\%`$, with $`\\beta_A = \\beta_B = 0\\%`$, their observed scores are $`s_A = 73.12\\%`$ and $`s_B = 76.96\\%`$, and the model with the lower rate of fully valid solutions ranks higher. Any $`\\alpha_B`$ above 25.3% reverses this pair.\n\nPublic leaderboards have shown the effect at scale. When EvalPlus extended HumanEval's tests 80-fold, pass@k fell by up to 19.3% to 28.9%, models changed places, and 18 of the original reference solutions, 11% of the problems, were found defective ([EvalPlus](https://arxiv.org/abs/2305.01210)).\n\n**When a lead is certified.** Equation (1.1) can be read in the other direction. If a grader's false acceptance is known to be at most $`\\bar{\\alpha}`$ and its false rejection at most $`\\bar{\\beta}`$, a score pins true validity to an interval, and with $`\\bar{\\beta} = 0`$ a lead of $`g = s_A - s_B`$ certifies the order of true validity under one condition:\n\n$$\n\\frac{s - \\bar{\\alpha}}{1 - \\bar{\\alpha}} \\le \\theta \\le \\frac{s}{1 - \\bar{\\beta}}, \\qquad \\theta_A > \\theta_B \\ \\text{ for every admissible grader} \\iff \\bar{\\alpha} < \\frac{g}{1 - s_B} \\tag{1.2}\n$$\n\nThe condition is tight: at any larger $`\\bar{\\alpha}`$ there are error rates within the bound that reverse the order. A lead of three points over a rival scoring 80% is certified only when the grader's false acceptance is below 15%; over a rival scoring 50%, below 6%. The harder the evaluation, the better the grader has to be before a small lead means anything. This turns the reversal above from a warning into a test that can be applied to any published ranking whose grader has a measured error rate.\n\nProcess supervision extends evaluation to intermediate steps (*[Let's Verify Step by Step, Sections 2 and 3](https://arxiv.org/html/2305.20050v1)*). When I apply it to general post-training evaluation, I account for omissions in reference solutions, accept valid alternative methods, and enforce evidence and execution requirements defined before responses are generated. Structural simplification can explain a method or reveal redundant work. A slower solution, or one that takes more attempts, is still a correct solution, and a resource-related rejection requires a hard limit specified in advance.\n\n## 2 Proof obligations organize layered review and allow alternative solutions\n\nA task specification defines the domain assumptions and execution constraints $`\\Gamma_q`$, the target conclusion $`C_q`$, and the proof obligations needed to establish it. A model trajectory $`\\tau`$, which here means its visible reasoning and observable execution records, may get there through different intermediate lemmas, representations or orders of deduction. That is allowed, because review concerns the submitted material and whether it fulfills the task's obligations. A directed acyclic dependency graph (DAG) $`G`$ organizes the local soundness check for each node $`v_j`$, together with a global tool-compliance predicate $`\\mathcal{T}_{\\text{comp}}(\\tau) \\in \\{0, 1\\}`$:\n\n$$\n\\Gamma_q \\cup \\{v_k : k \\in \\mathrm{pa}(j)\\} \\vdash v_j, \\qquad \\Gamma_q \\cup \\{v_j : j \\in \\mathrm{sink}(G)\\} \\vdash C_q, \\qquad \\mathcal{T}_{\\text{comp}}(\\tau) = 1 \\tag{2.1}\n$$\n\nThe symbol $`\\vdash`$ denotes valid derivability within a sound domain inference system, and the condition $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$ certifies that each external tool invocation $`t \\in \\tau`$ belongs to the task's permitted execution environment $`\\mathcal{A}_q`$. If root nodes follow from verified assumptions $`\\Gamma_q`$, every dependency is sound, with its local derivation verified in that system, and $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$, structural induction over a topological ordering of $`G`$ establishes the full conclusion.\n\nThis is the discipline of certifying algorithms carried over to model output. The producer may be arbitrarily complex and is not trusted; it has to hand over a witness, and the checker of that witness is simple enough to be trusted ([McConnell et al.](https://doi.org/10.1016/j.cosrev.2010.09.009)). A model is the least trusted producer there is, which makes the checker the place to spend rigor.\n\nIn production, I divided verification into three layers:\n1. Code and rule checks cover the directly verifiable parts of an answer, required deliverables, and predefined restrictions on tools, execution environments and resources. Tool compliance requires $`\\mathcal{T}_{\\text{comp}}(\\tau) = 1`$, and a call that violates an explicit hard requirement causes rejection ($`V = 0`$) even when the final answer is correct.\n2. LLM-as-a-Judge assists with semantic screening. It flags insufficient evidence, misapplied rules, conclusions that exceed their premises and gaps in reasoning, and it links each concern to the relevant trace, tool response or intermediate output. The flags direct further review.\n3. Scholars with relevant expertise take the questions that require domain judgment, admit several reasonable interpretations, or remain disputed after the first two checks. They examine the evidence, the applicability of the method and the conclusion, and a question that cannot be resolved from the available evidence stays open.\n\nAll three layers work from the same observable material. I cross-check the model's displayed CoT and visible reasoning against actual tool calls, intermediate outputs, elapsed time and attempt counts. For each attempt I look for new evidence, revised premises or a change in the state of the solution. That is how I identify repeated calls, continued use of failed assumptions and loops that make no progress. A conclusion whose execution records are missing is recorded as unverified. I record efficiency separately from correctness and compliance: long runtimes, repeated attempts or exhaustive methods are analyzed as efficiency, and they affect acceptance only when they violate a predefined hard limit.\n\nMethod applicability has its own check, because pointwise convergence at a fixed parameter value does not establish that an approximation remains valid within an optimization, an integral or an exchange of limits. A claim about the order of accuracy requires an approximation parameter, its domain of validity and a uniform error bound. Where such an issue cannot be decided programmatically, an LLM helps locate the concern and a specialist scholar reviews the assumptions and derivation. Specialist review settles the question with the standing of expert judgment; a certificate settles it by computation.\n\nFor feedback, I record the first failed obligation at node $`v_j`$ and mark its descendant subgraph $`\\mathrm{desc}(v_j)`$ as unverified, since locally correct calculations under a false premise do not repair the full derivation. This treatment resolves the ambiguity after an error discussed in *[ProcessBench, Section 3.1](https://arxiv.org/html/2412.06559v2)*. A suspected error awaiting specialist review remains unresolved until that review is complete.\n\n**When a screen may decide.** The division between the second and third layers has a quantitative form. Let $`\\text{se}`$ be the screen's sensitivity on semantic obligations, the share of invalid steps it flags. Flagged work goes to specialists, and unflagged work goes to specialists with probability $`\\rho`$. With the specialists' verdict as the reference, the system's false acceptance on those obligations, and the random escalation needed to hold it at a target $`\\alpha^*`$, are:\n\n$$\n\\alpha_{\\text{sys}} = (1 - \\text{se})(1 - \\rho), \\qquad \\rho \\ge 1 - \\frac{\\alpha^*}{1 - \\text{se}} \\tag{2.2}\n$$\n\nThe best model tested by [Tyen et al.](https://arxiv.org/abs/2311.08516) located the first mistake in 52.87% of reasoning traces. At a sensitivity of that order and a target of $`\\alpha^* = 5\\%`$, 89% of unflagged work still has to reach a specialist, and the screen saves almost no expert time. At $`\\text{se} = 0.80`$ the share is 75%, at $`0.90`$ it is 50%, and only at $`\\text{se} \\ge 0.95`$ can unflagged work pass without review. That is the rule behind the layer boundary: an LLM judge screens until its measured sensitivity clears $`1 - \\alpha^*`$, and decides after. The randomly escalated share is also the calibration sample, because a screen whose unflagged output is never reviewed has no measured sensitivity at all. A checker earns decisive standing by a clean run: 59 known-invalid items rejected without exception bound its false acceptance at 5% with 95% confidence, and 299 bound it at 1%.\n\nRelated work approaches these questions from several directions. [ToolEval](https://arxiv.org/html/2307.16789v1) scores tool-using agents by pass rate and win rate, both assigned by an LLM evaluator, so its verdicts carry the standing of a screen. [Let's Verify Step by Step](https://arxiv.org/html/2305.20050v1) collected step-level human labels and withheld reference solutions so that labelers would not favor one route. [Math-Shepherd](https://arxiv.org/html/2312.08935v1) estimates the quality of a step from whether sampled continuations reach the correct answer, which scores a step by its outcome; the certificates of Section 3 check the mathematical validity of the evidence a submission actually presents. [ProcessBench](https://arxiv.org/html/2412.06559v2) has since measured how often correct answers conceal process errors on hard problems. My contribution is the organization of acceptance itself, through explicit proof obligations, a dependency structure over them and conditions that can be checked.\n\n## 3 Error certificates that need no reference solution\n\nA posteriori verification rests on one observation: to bound how far a candidate is from the truth, it is enough to have a residual the candidate exposes and a stability constant of the problem. The true solution never enters. I develop the template in full on an optimization family whose constraints and certificates can be checked explicitly, and then state it for the other task types it covers.\n\nConsider the equality-constrained quadratic problem:\n\n$$\nv^* = \\min_{Ax = b} f(x), \\qquad f(x) = \\frac{1}{2}x^\\top H x + c^\\top x, \\qquad Z^\\top H Z \\succeq \\mu I, \\quad \\mu > 0 \\tag{3.1}\n$$\n\nHere $`H`$ is symmetric, $`Ax = b`$ is feasible, the columns of $`Z`$ form an orthonormal basis of $`\\ker A`$, and $`\\mu`$ is a verified lower bound on curvature along feasible directions. The objective has to be strongly convex only on the feasible manifold; $`H`$ may be indefinite across the constraints. For any feasible candidate $`x`$ satisfying $`Ax = b`$, choose any dual multiplier vector $`\\nu`$ and define the stationarity residual $`r = Hx + c + A^\\top \\nu`$. Curvature gives the first inequality below for every feasible displacement $`d \\in \\ker A`$, and completing the square in $`d`$ gives an exact optimality error certificate:\n\n$$\nf(x + d) \\ge f(x) + r^\\top d + \\frac{\\mu}{2}\\|d\\|^2 \\tag{3.2}\n$$\n\n$$\n\\inf_{d \\in \\ker A} \\left( r^\\top d + \\frac{\\mu}{2}\\|d\\|^2 \\right) = -\\frac{\\|P_{\\ker A} r\\|_2^2}{2\\mu} \\tag{3.3}\n$$\n\n$$\n0 \\le f(x) - v^* \\le \\frac{\\|P_{\\ker A} r\\|_2^2}{2\\mu} \\tag{3.4}\n$$\n\nThe operator $`P_{\\ker A}`$ is the orthogonal projection onto the space of feasible directions $`\\ker A`$. It removes residual components normal to the constraints, which do not affect optimality along the feasible manifold (*[Boyd and Vandenberghe, Convex Optimization, Sections 5.5 and 9.1](https://web.stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf)*). A zero projected residual, $`P_{\\ker A} r = 0`$, certifies exact optimality, and otherwise the certificate bounds the suboptimality gap in objective units. Two kinds of submission fare differently under this check:\n• Solutions derived through Lagrange multipliers or by eliminating the equality constraints satisfy the same semantic obligations and pass verification.\n• Reporting the correct optimal value $`v^*`$ alone leaves the obligations undischarged. Feasibility and the claimed objective value are checked separately. A submission that claims a KKT multiplier witness must also have its full stationarity equation $`Hx+c+A^\\top\\nu=0`$ verified, because the projected-residual bound cannot validate that multiplier: $`P_{\\ker A}A^\\top\\nu=0`$ for every $`\\nu`$. A valid equality-elimination proof need not supply a multiplier witness. Whether a checker validates the whole of a claimed witness is what the formal verification of certifying computations examines ([Alkassar et al.](https://arxiv.org/abs/1301.7462)), and Section 6 tests for it directly.\n\nThe same three parts, a residual, a verified constant and an inequality between them, give a certificate for each of the following task types.\n\n**Table 1. Error certificates by task type.**\n\n| Task type | Residual the submission exposes | Constant the task designer verifies | Certificate |\n| --- | --- | --- | --- |\n| Equality-constrained quadratic program (above) | $`P_{\\ker A} r`$ | $`\\mu = \\lambda_{\\min}(Z^\\top H Z)`$ | $`f(x) - v^* \\le \\lVert P_{\\ker A} r\\rVert _2^2 / (2\\mu)`$ |\n| Convex program | A primal-feasible $`x`$ and a dual-feasible $`(\\lambda, \\nu)`$ | None | $`f(x) - v^* \\le f(x) - g(\\lambda, \\nu)`$ |\n| Linear system $`Ax = b`$ | $`b - Ax`$ | A lower bound on $`\\sigma_{\\min}(A)`$ | $`\\lVert x - x^*\\rVert _2 \\le \\lVert b - Ax\\rVert _2 / \\sigma_{\\min}(A)`$ |\n| Eigenpair of a symmetric matrix | $`A\\hat{v} - \\hat{\\lambda}\\hat{v}`$ | None | $`\\min_i \\left\\lvert \\lambda_i - \\hat{\\lambda}\\right\\rvert \\le \\lVert A\\hat{v} - \\hat{\\lambda}\\hat{v}\\rVert _2 / \\lVert \\hat{v}\\rVert _2`$ |\n| Fixed point of a contraction $`g`$ | $`g(x) - x`$ | A Lipschitz constant $`L < 1`$ | $`\\lVert x - x^*\\rVert \\le \\lVert g(x) - x\\rVert / (1 - L)`$ |\n\nEach row is a standard result of its field. What I take from them is the organization: every obligation type is paired with a certificate and a checker whose cost is a few matrix products, and acceptance becomes a computation on the submission alone. That removes the reference solution from the trusted base. Reference solutions are a source of grading error in their own right, as the 18 defective references found by [EvalPlus](https://arxiv.org/abs/2305.01210) show, and a grader that matches against them inherits every one.\n\nI carry the lessons from this family over to the structure of general review: check the inputs and assumptions, compliance with the execution specification, support for the derivation, and whether the conclusion exceeds the evidence. Obligations that admit programmatic checks go to rules or certificate verifiers, and other semantic questions go through LLM-assisted screening and specialist review. Structural simplification may explain why a method needs less repeated work, and efficiency gains and validity certificates are evaluated separately.\n\n## 4 Task transformations must propagate error bounds\n\nSymbolic templates parameterize variable domains and answer formulas (*[GSM-Symbolic, Section 3.1](https://arxiv.org/html/2410.05229v1)*). Tasks with formalizable transformations also need verifiable rules for how those transformations affect feasible sets, objective values and acceptance tolerances. For bijective task transformations the tolerance-propagation rule below follows.\n\nLet an original problem and its transformed variant have feasible sets $`F`$ and $`F'`$, objectives $`J`$ and $`J'`$, and finite infima $`v`$ and $`v'`$. Suppose the task designer verifies a bijection $`\\Phi : F \\to F'`$ such that every feasible $`u \\in F`$ satisfies the uniform bound:\n\n$$\n\\left\\vert J'(\\Phi(u)) - aJ(u) - b\\right\\vert \\le \\varepsilon, \\quad a > 0 \\tag{4.1}\n$$\n\n$$\nav + b - \\varepsilon \\le v' \\le av + b + \\varepsilon \\tag{4.2}\n$$\n\n$$\nJ(u) - v \\le \\delta \\implies J'(\\Phi(u)) - v' \\le a\\delta + 2\\varepsilon \\tag{4.3}\n$$\n\nTaking infima in (4.1) gives (4.2), and subtracting the lower bound on $`v'`$ from the upper bound on $`J'(\\Phi(u))`$ gives (4.3). So after a unit conversion, parameter rescaling or numerical transformation with error bounded by $`\\varepsilon`$, the result determines the acceptance tolerance for a previously $`\\delta`$-optimal solution. Keeping the original tolerance $`\\delta`$ can misclassify a valid transformed solution as a model failure, while relaxing it arbitrarily can conceal a real capability gap. When $`\\varepsilon = 0`$, the bijection and positive scaling $`a > 0`$ also preserve whether an optimum is attained. Changes to boundary compactness or admissible control classes require a new attainment proof.\n\nVariants are usually built by several transformations in a row. Applying the bound step by step to a chain with parameters $`(a_1, \\varepsilon_1), \\dots, (a_k, \\varepsilon_k)`$ gives the composition law:\n\n$$\na = \\prod_{j=1}^{k} a_j, \\qquad \\varepsilon = \\sum_{j=1}^{k} \\Big( \\prod_{l > j} a_l \\Big) \\varepsilon_j, \\qquad J(u) - v \\le \\delta \\implies J^{(k)} - v^{(k)} \\le a\\delta + 2\\varepsilon \\tag{4.4}\n$$\n\nEach error is magnified by every scaling applied after it, so the order of a chain is part of its specification. Take an exact unit conversion with $`a = 100`$, an exact rescaling with $`a = 0.5`$, and a numerical discretization with $`\\varepsilon = 0.003`$, applied to a solution that is $`\\delta = 0.01`$ optimal. With the discretization last, the chain has $`(a, \\varepsilon) = (50, 0.003)`$ and the tolerance owed is $`0.506`$. With the discretization first, it has $`(50, 0.15)`$ and the tolerance is $`0.8`$. A grader that reuses the first tolerance for the second chain rejects valid work.\n\n## 5 Grading errors change the post-training objective\n\nStep-level labels can improve post-training only if they measure logical validity and compliance, because systematic grading errors pass into the targets used by process reward models (PRMs) and RLHF policies. Let $`h`$ denote the task input and the model's visible reasoning prefix through the current step. Let $`Z = 1`$ indicate that the step is valid under the problem's assumptions, and let $`V \\in \\{0, 1\\}`$ be the grader's noisy acceptance label, with conditional false acceptance $`\\alpha(h)`$ and false rejection $`\\beta(h)`$. Writing $`p(h) = \\Pr(Z = 1 \\mid h)`$, a process evaluator with logit $`s`$ trained by binary cross-entropy is optimized toward the corrupted conditional target:\n\n$$\n\\tilde{p}(h) = \\alpha(h) + [1 - \\alpha(h) - \\beta(h)]p(h) \\tag{5.1}\n$$\n\n$$\n\\frac{\\partial \\ell(s, V)}{\\partial s} = \\sigma(s) - V \\tag{5.2}\n$$\n\n$$\nE[\\nabla_s \\ell(s, V) - \\nabla_s \\ell(s, Z) \\mid h] = p(h) - \\tilde{p}(h) \\tag{5.3}\n$$\n\nThe expected gradient difference shows which way grading bias moves the training target:\n• If false rejection dominates, $`\\beta(h)p(h) > \\alpha(h)[1-p(h)]`$, then $`\\tilde{p}(h) < p(h)`$. The learned validity target shifts downward, which penalizes valid alternative proofs.\n• If false acceptance dominates, $`\\alpha(h)[1-p(h)] > \\beta(h)p(h)`$, then $`\\tilde{p}(h) > p(h)`$. The target shifts upward, rewarding flawed proofs or noncompliant tool calls. When the weighted contributions are equal, the effects cancel in the conditional mean.\n\nWhen $`\\alpha, \\beta`$ are known constants with $`\\alpha + \\beta < 1`$, class-conditional noise correction removes the bias from the mean estimate (*[Learning with Noisy Labels, Sections 2 and 3](https://papers.nips.cc/paper_files/paper/2013/file/3871bd64012152bfb53fdf04b401193f-Paper.pdf)*). The estimator is the prevalence correction of [Rogan and Gladen](https://pubmed.ncbi.nlm.nih.gov/623091/), and it increases variance by the squared factor below:\n\n$$\n\\hat{\\theta}_{\\text{corr}} = \\frac{\\bar{V} - \\alpha}{1 - \\alpha - \\beta}, \\qquad \\mathrm{Var}(\\hat{\\theta}_{\\text{corr}}) = \\frac{\\mathrm{Var}(\\bar{V})}{(1 - \\alpha - \\beta)^2} \\tag{5.4}\n$$\n\nFor task families whose validity is exactly decidable by a sound and complete certificate verifier, $`\\alpha=\\beta=0`$ and there is nothing to correct.\n\n**What a policy learns from a grader.** Equations (5.1) and (5.4) describe a grader that is only observed. Once its verdict is the reward, the policy moves toward whatever the grader accepts. Model that as reweighting the policy's trajectories by $`e^{\\eta V}`$, with $`\\eta`$ the strength of optimization. A policy that starts at true validity $`\\theta`$ then has:\n\n$$\n\\theta_\\eta = \\frac{\\theta(1-\\beta)e^{\\eta} + \\theta\\beta}{\\big[\\theta(1-\\beta) + (1-\\theta)\\alpha\\big]e^{\\eta} + \\theta\\beta + (1-\\theta)(1-\\alpha)} \\ \\xrightarrow{\\ \\eta \\to \\infty\\ } \\ \\frac{\\theta(1-\\beta)}{\\theta(1-\\beta) + (1-\\theta)\\alpha} \\tag{5.5}\n$$\n\nThe limit is the grader's positive predictive value at the starting policy, and the observed score goes to 100% along the way. For Model B of Section 1 under final-answer matching:\n\n**Table 2. Observed score and true validity of Model B under final-answer matching as optimization strengthens.**\n\n| Optimization strength $`\\eta`$ | 0 | 1 | 2 | 4 | Limit |\n| --- | --- | --- | --- | --- | --- |\n| Observed score | 77.0% | 90.1% | 96.1% | 99.5% | 100% |\n| True validity | 64.0% | 74.9% | 79.9% | 82.7% | 83.2% |\n\nTwo results follow, one for each kind of error.\n\n• *False acceptance sets the ceiling.* The policy can reach the grader's positive predictive value and no more: 83.2% for Model B, 98.5% for Model A. With $`\\alpha = 0`$ the ceiling is 100% for any $`\\beta < 1`$.\n• *False rejection removes routes.* A valid route that the grader always rejects loses its share of valid work at rate $`e^{-\\eta}`$. If two valid routes start equally common and the grader accepts only one, the other falls from 50% of valid work to 26.9%, 11.9% and 1.8% at $`\\eta = 1, 2, 4`$. Training against a reference path teaches the model to stop producing correct proofs the reference did not anticipate.\n\nThe cost of false acceptance is the ceiling and the cost of false rejection is diversity. Certificate checking has neither cost on the families it decides, which is the argument for drawing process rewards from certified nodes. It also locates the limit of a theoretical result sometimes read as making process supervision unnecessary: [Jia, Rakhlin and Xie](https://arxiv.org/abs/2502.10581) show that outcome supervision is statistically no harder than process supervision, up to polynomial factors in the horizon, when the outcome reward is the objective. When a correct answer can rest on an invalid derivation, the outcome reward is wrong about validity, and a policy trained on it converges to the limit in equation (5.5).\n\n## 6 A test set with exactly known verdicts\n\nMeasuring a grader's two error rates requires submissions whose validity is known without appeal to any grader. I built such a test set from 600 strongly convex quadratic task families. Using exact rational arithmetic, each family yields 6 algebraic submissions, for 3,600 in total:\n• 1,200 valid certificates, 2 per family: one derived using Lagrange multipliers and one using equality elimination.\n• 2,400 submissions with controlled defects, 4 per family: (1) an infeasible construction, $`Ax \\neq b`$, that reports the exact optimal value $`v^*`$; (2) failure of stationarity along a feasible direction in $`\\ker A`$ while reporting $`v^*`$; (3) a claimed multiplier that fails the stationarity equation while reporting $`v^*`$; and (4) an incorrect final objective value. The first three classes, comprising 1,800 submissions or 75% of all invalid submissions, report the correct final answer $`v^*`$.\n\nThe composition is the instrument's design: three ways of being wrong with the right answer, one way of being wrong with the wrong answer, and two ways of being right. I compared three deterministic acceptance protocols on all 3,600 submissions:\n\n**Table 3. Three acceptance protocols on 3,600 submissions with known verdicts.**\n\n| Acceptance protocol | False acceptance rate on invalid submissions (2,400) | False rejection rate on valid certificates (1,200) | Valid share of accepted submissions |\n| --- | --- | --- | --- |\n| Final-answer matching only | 75.0% (accepts 1,800 flawed solutions) | 0.0% | 40.0% |\n| Strict reference-path matching | 0.0% | 50.0% (rejects 600 valid elimination proofs) | 100.0% |\n| Certificate verification | 0.0% | 0.0% | 100.0% |\n\nCertificate verification accepted 1,800 fewer invalid proofs than final-answer matching and retained 600 more valid alternative solutions than strict reference-path matching. Strict error detection and acceptance of valid alternative proofs held together on the same submissions, which neither baseline managed. The projected-residual bound in equation (3.4) also held at all 600 feasible perturbations, one per family, and the verifier rejected every submission of defect class (3), the multiplier that a partial witness check would have let through.\n\nThe last column is the ceiling of Section 5. A policy trained against final-answer matching on this mix moves from one third valid work to 40% and stops. Trained against the reference path it reaches 100% and loses the elimination route. Trained against certificate checking it reaches 100% with both routes intact.\n\n## 7 Dependence within task families and the economics of certificates\n\nSurface variations of one problem template do not provide independent observations of general reasoning ability. Suppose there are $`F`$ independent task families, each with $`K`$ variants, common marginal variance $`\\sigma^2`$, and common pairwise within-family correlation $`\\rho_w`$. The sample mean then carries the Kish design effect $`1 + (K - 1)\\rho_w`$, and its variance and effective sample size $`n_{\\text{eff}}`$ are:\n\n$$\n\\mathrm{Var}(\\bar{Z}) = \\frac{\\sigma^2}{FK}[1 + (K - 1)\\rho_w], \\qquad n_{\\text{eff}} = \\frac{FK}{1 + (K - 1)\\rho_w} \\tag{7.1}\n$$\n\nFor $`F = 50`$ families with $`K = 8`$ variants each, there are $`400`$ responses, and if $`\\rho_w = 0.6`$, those $`400`$ responses provide the same precision as $`n_{\\text{eff}} \\approx 77`$ independent observations. Training and test sets are therefore split by task family, and new tasks vary assumptions, evidentiary relationships or the structure of a derivation, beyond substituting numbers.\n\nCertificates cost engineering once and almost nothing afterwards. Let $`C_0`$ be the upfront engineering cost of the proof-obligation certificates and verifier, $`c_e`$ the expert review cost per submission, $`c_v`$ the automated verifier's unit execution cost, and $`u`$ the fraction of submissions escalated to human adjudication at the same expert unit cost. The certificate workflow is the cheaper one when submission volume $`N`$ exceeds:\n\n$$\nC_0 + N(c_v + uc_e) < Nc_e \\iff N > \\frac{C_0}{(1 - u)c_e - c_v}, \\qquad (1 - u)c_e > c_v \\tag{7.2}\n$$\n\nFor a certificate family whose engineering costs as much as 400 expert reviews, with 15% of submissions escalated and a verifier run costing 2% of a review, the break-even is 482 submissions. Past that volume every further submission releases 83% of an expert review, and those hours go to the semantic obligations that only specialists can settle. The cost model covers the certificate layer; model calls and calibration are costed with the layers they belong to.\n\nThe team produced roughly 500 scientific reasoning tasks a week under this design, and later programs reused its layered structure.",
"sourceDocument": {
"fileName": "ai-research-and-engineering.md",
"path": "content/sources/ai-portfolio/ai-research-and-engineering.md",
- "sha256": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1"
+ "sha256": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0"
},
"repositoryUrl": "https://github.com/130U/RESEARCH-General-Reasoning-Evaluation-and-Layered-Verification"
}
diff --git a/website/content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json b/website/content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json
index 36d22f6..446a3c0 100644
--- a/website/content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json
+++ b/website/content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json
@@ -3,11 +3,11 @@
"title": "Large-Scale Domain Expert Feedback Governance and Adaptive Review",
"sourceUrl": "https://github.com/130U/RESEARCH-Large-Scale-Domain-Expert-Feedback-Governance-and-Adaptive-Review/blob/main/manuscript/large-scale-domain-expert-feedback-governance-and-adaptive-review.md",
"sourceEditedAt": "2026-11-19",
- "markdown": "## 1 What a review queue measures\n\nIn a production RLHF pipeline, whatever policy routes submissions to review also changes the statistical properties of the observed sample. To see how, let $`X`$ denote observable task features, including the domain, prompt complexity and contributor history. Let $`Y \\in \\{0, 1\\}`$ indicate whether the original submission contains an error under fixed acceptance criteria, so that the true conditional error probability is $`q(x) = P(Y = 1 \\mid X = x)`$. Write $`S = 1`$ when a submission is selected for an independent reference audit. Conditional on $`X`$, selection is randomized with probability $`\\pi(x) > 0`$, and it is independent of the reference label, which has not yet been observed.\n\nSuppose an operational dashboard estimates population quality from the error fraction among verified submissions. What it actually targets is:\n\n$$\nE[Y \\mid S = 1] = \\frac{E[\\pi(X)q(X)]}{E[\\pi(X)]} \\tag{1.1}\n$$\n\n$$\nE[Y \\mid S = 1] - E[Y] = \\frac{\\mathrm{Cov}(\\pi(X), q(X))}{E[\\pi(X)]} \\tag{1.2}\n$$\n\nWhen review succeeds in targeting higher true error risk, the covariance between audit probability $`\\pi(X)`$ and true defect probability $`q(X)`$ is strictly positive, and the dashboard reads high. A rising detection rate in the reviewed sample is what better targeting looks like, and the batch may be unchanged or improving underneath it. The identities describe a selected item; in a finite batch the observed fraction has a random denominator and converges to this value.\n\nThis is Neyman's argument against purposive selection, applied to a review queue: a sample chosen to serve one purpose fails on the quantities it was not chosen for ([Neyman 1934](https://doi.org/10.2307/2342192)). Financial auditors already act on it. Under the [GAO Financial Audit Manual](https://www.gao.gov/assets/gao-18-601g.pdf) the auditor examines large and unusual items in full, samples the remainder at random, projects only the random sample to the population and records which selection is which. The PRM800K report describes the bias and leaves it there: its authors chose convincing wrong-answer solutions for labeling and noted that the choice skewed the data set toward wrong answers ([Let's Verify Step by Step](https://arxiv.org/abs/2305.20050)).\n\nExpert rankings suffer from the same selection mechanism. Let $`a_{jh}`$ be the probability that expert $`j`$ answers a task in stratum $`h`$ correctly, and let $`p_{jht}`$ describe that expert's assigned task mix in period $`t`$. Even if underlying ability stays constant, observed accuracy moves with the task mix:\n\n$$\nA_{jt} = \\sum_h p_{jht} a_{jh}, \\qquad A_{j,t+1} - A_{jt} = \\sum_h (p_{jh,t+1} - p_{jht}) a_{jh} \\tag{1.3}\n$$\n\nComparisons across periods therefore require a fixed target task mix $`w_h`$ and standardized accuracy:\n\n$$\nA_j^{\\text{std}} = \\sum_h w_h a_{jh} \\tag{1.4}\n$$\n\nOverlapping assignments or shared anchor cases support estimation within each stratum. In the project's production routing engine I used a conjugate Beta-Binomial model for each contributor's reliability within a task stratum. Given a prior $`\\mathrm{Beta}(\\alpha_{0h}, \\beta_{0h})`$ and $`n_{jht}`$ audited tasks with $`k_{jht}`$ verified successes, the posterior $`\\mathrm{Beta}(\\alpha_{0h} + k_{jht}, \\beta_{0h} + n_{jht} - k_{jht})`$ gives shrinkage-regularized point estimates and credible intervals that account for unequal sample sizes across comparable audits. Section 7 sets out how these records drive contributor decisions.\n\nEach piece of this problem has a literature, and each literature stops at one of the two jobs. [Dawid-Skene](https://doi.org/10.2307/2346806) models estimate how reviewers' confusion matrices differ and carry no sampling design. [PPI++](https://arxiv.org/html/2311.01453v1) gives valid inference from a small labeled sample and a large set of predictions when the labels are drawn uniformly. [Active Statistical Inference](https://arxiv.org/html/2403.03208v3) chooses which labels to collect and prices every label the same; [cost-optimal active evaluation](https://arxiv.org/abs/2506.07949) prices two kinds of rater, at one price per kind for all items. [ActiveLab](https://arxiv.org/html/2301.11856v1) decides whether the next annotation should go to a new example or to one already labeled, with a better classifier as the goal. [DeCCaF](https://arxiv.org/html/2403.06906v1) assigns cases to experts under error costs and capacity limits, and the quality of the batch is outside its objective. A production program has to price both jobs in one unit, expert hours, and the sections below do that.\n\n## 2 Measuring a batch with audits and predictions\n\nFix a production batch of $`N`$ original submissions, its acceptance criteria and its latent reference labels $`Y_1, \\dots, Y_N`$. The inferential target is the finite-batch error rate $`\\mu = N^{-1} \\sum_{i=1}^N Y_i`$. Let $`f_i \\in [0, 1]`$ be a fixed, inexpensive risk prediction that combines contributor reliability posteriors, task complexity and automated heuristic checks. Assign audit probability $`\\pi_i \\in [\\pi_{\\min}, 1]`$ and draw $`S_i \\sim \\mathrm{Bernoulli}(\\pi_i)`$ independently across submissions, with the predictions and the sampling probabilities locked before the current batch's reference labels are revealed.\n\nI used a prediction-assisted estimator with inverse-probability-weighted (IPW) residual correction:\n\n$$\n\\hat{\\mu} = \\frac{1}{N} \\sum_{i=1}^N \\left[ f_i + \\frac{S_i}{\\pi_i} (Y_i - f_i) \\right] \\tag{2.1}\n$$\n\nWrite the prediction residual as $`e_i = Y_i - f_i`$ and condition on the fixed batch, the predictions and the sampling probabilities:\n\n$$\n\\hat{\\mu} - \\mu = \\frac{1}{N} \\sum_{i=1}^N \\left( \\frac{S_i}{\\pi_i} - 1 \\right) e_i, \\qquad E_S[\\hat{\\mu} - \\mu] = 0 \\tag{2.2}\n$$\n\nValidity comes from the design and efficiency from the model. The inverse-probability weights restore the population representation that selective verification altered, and the estimate is unbiased for any predictions at all, because the only randomness it relies on is the audit lottery the program itself runs. Better predictions shrink the residuals and with them the number of audits a given precision needs. That division of labor suits production: the estimate stays correct while the risk model is still being improved. The form is the difference estimator of survey sampling, [Horvitz-Thompson](https://doi.org/10.1080/01621459.1952.10483446) weighting applied to residuals, and the augmented IPW estimator of [Robins, Rotnitzky and Zhao](https://doi.org/10.1080/01621459.1994.10476818); [PPI++](https://arxiv.org/html/2311.01453v1) studies it with black-box predictions.\n\nHow much the predictions can save has a known ceiling under uniform sampling: a predictor no more accurate than the work it scores can at most double the effective audit sample ([Dorner et al.](https://arxiv.org/abs/2410.13341)). The gains in Sections 3 and 4 come from where the audits are sent, which is a separate source of efficiency.\n\nTo use the estimator in production, I defined procedures for freezing the audited objects, recording rubric version histories and retaining the telemetry that optimization needs. Original and revised submissions are retained as separate database entities. Telling defects introduced during initial production apart from those corrected through review takes two estimates, $`\\hat{\\mu}_{\\text{raw}}`$ on frozen originals and $`\\hat{\\mu}_{\\text{final}}`$ on post-review deliverables. A change in post-training criteria causes a different confusion, since stricter acceptance standards can be misread as declining contributor ability. A shared set of anchor cases is therefore adjudicated under both rubric versions whenever the criteria change, and submissions from different periods are compared under a common reference version.\n\n## 3 Where audits buy the most information\n\nBecause the sampling is independent Bernoulli, the cross-item covariance terms drop out and the design variance is exactly:\n\n$$\n\\mathrm{Var}_S(\\hat{\\mu}) = \\frac{1}{N^2} \\sum_{i=1}^N \\frac{1 - \\pi_i}{\\pi_i} e_i^2 \\tag{3.1}\n$$\n\nThis expression sets the objective for audit allocation: audit effort should go to the uncertainty that automated risk predictions leave unresolved. For a binary error label with conditional error probability $`q_i`$, the expected squared residual has two components:\n\n$$\nm_i := E[(Y_i - f_i)^2 \\mid X_i] = q_i(1 - q_i) + (q_i - f_i)^2 \\tag{3.2}\n$$\n\nThe first is the conditional Bernoulli variance $`q_i(1 - q_i)`$ and the second is the squared calibration error $`(q_i - f_i)^2`$. A high error probability $`q_i`$ therefore need not carry high inferential value. If a task category is almost certain to contain an error ($`q_i \\approx 1`$) and the predictor recognizes this ($`f_i \\approx 1`$), then $`m_i \\approx 0`$, and an independent reference audit of that item says less about population quality than an audit of a moderately risky item whose outcome is still open ($`q_i \\approx 0.5`$). A repair model and a measurement model look at the same task and want different things from it.\n\nLet $`c_i > 0`$ be the empirical cost of an independent reference audit, calibrated by reconciling platform review logs with active-work records from the team's KPI monitoring system, and let $`B`$ be the expected audit-cost budget. With residual second moments $`m_i > 0`$, the optimal design solves:\n\n$$\n\\min_{\\pi} \\sum_{i=1}^N \\frac{m_i}{\\pi_i} \\quad \\text{s.t.} \\quad \\sum_{i=1}^N c_i \\pi_i \\le B, \\quad \\pi_{\\min} \\le \\pi_i \\le 1 \\tag{3.3}\n$$\n\nThe objective omits the constant $`-\\sum_i m_i`$, which does not depend on the allocation. For probabilities strictly between their bounds, the Lagrangian first-order condition gives the clipped square-root allocation:\n\n$$\n-\\frac{m_i}{\\pi_i^2} + \\lambda c_i = 0 \\implies \\pi_i^* = \\mathrm{clip}_{[\\pi_{\\min}, 1]} \\left( \\sqrt{\\frac{m_i}{\\lambda c_i}} \\right) \\tag{3.4}\n$$\n\nThe multiplier $`\\lambda > 0`$ is set by the budget, feasibility requires $`B \\ge \\pi_{\\min} \\sum_i c_i`$, and a budget that covers a full audit audits everything. The rule samples harder where expected residuals are larger and more lightly where audits cost more, while the floor $`\\pi_{\\min}`$ keeps every part of the task space covered and bounds the IPW weights.\n\nThe square-root form has a long ancestry. With equal costs and strata it is Neyman's allocation; with stratum costs it is the optimum allocation of [Cochran's text](https://www.wiley.com/en-us/Sampling+Techniques%2C+3rd+Edition-p-9780471162407); inclusion probabilities proportional to a model's error standard deviation are the anticipated-variance optimum of [Isaki and Fuller](https://doi.org/10.1080/01621459.1982.10477770); and [Chen and Lumley](https://doi.org/10.1002/sim.8760) set such designs wave by wave from earlier waves. Machine-learning work reached the same rule for labels of equal cost. Expert audits do not cost the same: on the reference batch of Section 6, an independent reference solution for the hardest tasks costs three times a routine one. My rule carries a cost for every item.\n\nProduction allocation uses regularized stratum estimates $`\\hat{m}_i > 0`$ from earlier independent audits or a pilot calibration batch, at the level of `(task stratum × criteria version × cost stratum)`, so that no parameter is fitted to an individual submission. A weekly program is a multiwave design in which each batch's audits set the next batch's allocation. The budget $`B`$ constrains expected audit cost per batch; a workflow that needs an exact audit count uses a fixed-size design with the same first-order inclusion probabilities.\n\n## 4 What the allocation is worth, and how accurate its inputs must be\n\nLet $`C = \\sum_i c_i`$, $`M = \\sum_i m_i`$ and $`S = \\sum_i \\sqrt{m_i c_i}`$. Uniform sampling at the same expected cost uses probability $`p = B / C`$. With no probability at a clipping boundary, averaging the design variance over the conditional label distribution gives:\n\n$$\nV_{\\text{uniform}} = \\frac{1}{N^2} \\left[ \\frac{C M}{B} - M \\right], \\qquad V_{\\text{optimal}} = \\frac{1}{N^2} \\left[ \\frac{S^2}{B} - M \\right] \\tag{4.1}\n$$\n\nIf all items within a task stratum share a sampling probability, substituting that stratum's finite-batch mean squared residual for $`m_i`$ makes these exact finite-batch identities. By the Cauchy-Schwarz inequality $`S^2 \\le C M`$, with equality if and only if $`m_i / c_i`$ is constant across items. The gain exists exactly where residual uncertainty per unit of audit cost varies across the batch.\n\n**The same precision for a fixed fraction of the budget.** A program manager fixes the precision the dashboard needs and asks what it costs. Solving each variance for the budget that reaches a target $`\\tau^2`$:\n\n$$\nB_{\\text{optimal}}(\\tau^2) = \\frac{S^2}{N^2 \\tau^2 + M}, \\qquad B_{\\text{uniform}}(\\tau^2) = \\frac{C M}{N^2 \\tau^2 + M}, \\qquad \\frac{B_{\\text{optimal}}}{B_{\\text{uniform}}} = \\kappa := \\frac{S^2}{C M} \\tag{4.2}\n$$\n\nThe ratio $`\\kappa`$ does not depend on the target. Whatever precision the program asks for, the residual- and cost-aware allocation delivers it for the same fraction of the uniform audit budget, and that fraction is a property of the batch: how unevenly residual uncertainty and audit cost are spread across it. The hours it frees, $`(1 - \\kappa) B_{\\text{uniform}}`$, are available for repair, which is the first link between the two jobs.\n\n**How accurate the residual estimates must be.** The allocation is computed from estimates $`\\hat{m}_i`$. Write $`\\hat{m}_i = m_i (1 + \\varepsilon_i)`$ and weight each item by its share $`w_i = \\sqrt{m_i c_i} / S`$ of the optimal budget. With neither allocation clipped, the variance of the allocation built on the estimates exceeds the optimum by:\n\n$$\nV(\\hat{\\pi}) - V(\\pi^*) = \\frac{S^2}{N^2 B} \\left[ \\Big( \\sum_i w_i \\sqrt{1 + \\varepsilon_i} \\Big) \\Big( \\sum_i \\frac{w_i}{\\sqrt{1 + \\varepsilon_i}} \\Big) - 1 \\right] = \\frac{S^2}{N^2 B} \\left[ \\tfrac{1}{4} \\mathrm{Var}_w(\\varepsilon) + O(\\|\\varepsilon\\|^3) \\right] \\tag{4.3}\n$$\n\nThe first-order terms cancel. Three consequences follow for the specification of the estimation pipeline. Overestimating or underestimating every residual by a common factor costs nothing, since only the dispersion of the relative errors enters. A relative error of 20% across strata raises the variance by about one percent of $`S^2/(N^2 B)`$, so stratum-level estimates carried over from earlier batches need to be accurate only to that order. And the exact expression shows where the rule can be hurt: through $`1/\\sqrt{1 + \\varepsilon_i}`$, by a gross underestimate of a residual somewhere in the task space.\n\nTwo guards bound that last case. The floor $`\\pi_{\\min}`$ bounds every weight. Mixing the computed probabilities with the uniform design at the same budget,\n\n$$\n\\pi_i(\\rho) = (1 - \\rho)\\hat{\\pi}_i + \\rho p, \\tag{4.4}\n$$\n\nbounds the variance for any estimates whatever, because every probability is then at least $`\\rho p`$:\n\n$$\n\\frac{V(\\pi(\\rho))}{V_{\\text{uniform}}} \\le \\frac{1 - \\rho p}{\\rho (1 - p)} \\tag{4.5}\n$$\n\nThis is the defensive mixture of importance sampling ([Hesterberg 1995](https://doi.org/10.1080/00401706.1995.10484303)) written for a Bernoulli audit design. The weight $`\\rho`$ is tuned on earlier batches; [robust active inference](https://arxiv.org/html/2511.08991v1) later gave an asymptotic choice of the same weight.\n\n## 5 One budget, two products\n\nStatistical auditing measures population quality. Operational review has a different job, removing defects from the current batch before it is delivered for training. For task $`i`$ under handling action $`a`$, such as direct acceptance, peer review, expert adjudication or contributor revision, let $`r_{ia}`$ be the expected error loss remaining after the action, $`d_{ia}`$ its operational cost, and $`t_{iaj}`$ the time required from specialist $`j`$, measured through platform records and the team's KPI monitoring system. Binary assignment decisions $`z_{ia} \\in \\{0, 1\\}`$ solve:\n\n$$\n\\min_{z} F(z) = \\sum_{i,a} z_{ia}(r_{ia} + \\eta d_{ia}) \\quad \\text{s.t.} \\quad \\sum_a z_{ia} = 1, \\quad \\sum_{i,a} z_{ia} t_{iaj} \\le H_j \\quad \\forall j \\tag{5.1}\n$$\n\nHere $`H_j`$ is the specialist's available time, and $`\\eta \\ge 0`$ converts resource cost to the scale of error loss.\n\n**Repair follows removable loss.** When review costs are equal and each reviewed task uses one slot, the marginal remediation gain $`\\Delta_i = r_{i0} - r_{i1}`$ determines the optimal priority order. A high-risk task may be hard to repair through one additional review because external evidence is missing, which gives a small $`\\Delta_i`$. A self-contained calculation error in a moderate-risk task may be fully correctable, which gives a large $`\\Delta_k`$. Moving one review slot from task $`i`$ to task $`k`$ changes total expected loss by:\n\n$$\n\\Delta F = (r_{i0} + r_{k1}) - (r_{i1} + r_{k0}) = \\Delta_i - \\Delta_k \\tag{5.2}\n$$\n\nWhenever $`\\Delta_k > \\Delta_i`$, the exchange strictly reduces delivered defect loss, regardless of the initial risk ordering. Optimal triage under a budget has the same threshold structure, with the human's case-level loss in place of the model's risk ([Okati et al.](https://proceedings.nips.cc/paper_files/paper/2021/file/4c4c937b67cc8d785cea1e42ccea185c-Paper.pdf)). A worked example fixes the size of the effect. Two task groups of $`100`$ items each have initial error risks of $`0.40`$ and $`0.15`$, and additional review reduces these risks to $`0.30`$ ($`\\Delta_1 = 0.10`$) and $`0.01`$ ($`\\Delta_2 = 0.14`$). Review costs are equal and capacity is $`100`$ items. Ranking by initial risk leaves $`100(0.30) + 100(0.15) = 45`$ expected errors; ranking by $`\\Delta_i`$ leaves $`100(0.40) + 100(0.01) = 41`$, a reduction of **8.89%** at the same review volume.\n\nThe quantity that has to be calibrated is the post-action risk. Suppose $`\\left\\vert\\hat{r}_{ia} - r_{ia}\\right\\vert \\le \\delta`$ for every feasible action, costs and capacities are known, and $`\\hat{z}`$ minimizes the predicted objective $`\\hat{F}(z)`$. Relative to the true optimal feasible assignment $`z^*`$:\n\n$$\nF(\\hat{z}) \\le \\hat{F}(\\hat{z}) + N\\delta \\le \\hat{F}(z^*) + N\\delta \\le F(z^*) + 2N\\delta \\tag{5.3}\n$$\n\nThat bound is the reason the program ran randomized pilot trials across the permitted review actions: they estimate $`r_{ia}`$ directly.\n\n**The batch program.** Audits and repair draw on the same money. Let $`\\tau^2`$ be the variance ceiling for the quality estimate, $`B_{\\text{total}}`$ the total cost budget, and $`V_{\\hat{m}}(\\pi) = N^{-2}\\sum_i (1 - \\pi_i)\\hat{m}_i / \\pi_i`$ the planning variance under the calibrated residual model. A reference audit produces an independent reference solution, so an audited item found defective ships corrected. Writing $`a_i`$ for the action assigned to item $`i`$, the item ships with expected loss $`(1 - \\pi_i)\\, \\hat{r}_{i a_i}`$, and the batch is planned by a mixed-integer convex program:\n\n$$\n\\min_{z, \\pi} \\sum_{i,a} z_{ia} (1 - \\pi_i)\\, \\hat{r}_{ia} \\quad \\text{s.t.} \\quad \\sum_{i,a} z_{ia} d_{ia} + \\sum_i c_i \\pi_i \\le B_{\\text{total}}, \\quad V_{\\hat{m}}(\\pi) \\le \\tau^2 \\tag{5.4}\n$$\n\n$$\n\\sum_a z_{ia} = 1, \\quad z_{ia} \\in \\{0, 1\\}, \\quad \\sum_{i,a} z_{ia} t_{iaj} \\le H_j^{\\text{review}} \\ \\forall j, \\quad \\pi_{\\min} \\le \\pi_i \\le 1 \\tag{5.5}\n$$\n\nThe hard capacity $`H_j^{\\text{review}}`$ goes to remediation actions, and reference audits are scheduled independently on frozen original submissions.\n\n**The joint rule.** Attach a price $`\\lambda`$ to the budget, $`\\nu`$ to the precision ceiling (in units of $`N^2 V`$) and $`\\kappa_j`$ to each specialist's hours. At those prices the program separates into one small problem per task, and each has a closed form:\n\n$$\n\\pi_i^* = \\mathrm{clip}_{[\\pi_{\\min}, 1]} \\left( \\sqrt{\\frac{\\nu\\, m_i}{\\big( \\lambda c_i - r_{i a_i} \\big)_+}} \\right), \\qquad a_i^* = \\arg\\min_a \\; (1 - \\pi_i)\\, r_{ia} + \\lambda d_{ia} + \\sum_j \\kappa_j t_{iaj} \\tag{5.6}\n$$\n\nwith $`\\pi_i^* = 1`$ when $`r_{i a_i} \\ge \\lambda c_i`$. Three readings of the audit rule carry the argument of this article.\n\n1. *Risk is a rebate.* The item's residual risk $`r_{i a_i}`$ is subtracted from the audit's price $`\\lambda c_i`$. Risk does not drive the audit; it makes the audit cheaper, by the repair the audit performs on the way.\n2. *Measurement alone.* If audits repaired nothing, the rebate would vanish and the rule would be the $`\\sqrt{m_i / c_i}`$ allocation of Section 3.\n3. *Repair alone.* If nobody valued the estimate, $`\\nu = 0`$, the rule audits exactly the items with $`r_{i a_i} \\ge \\lambda c_i`$, a threshold on risk per unit cost. That is risk-first routing. The legacy policy is the joint rule of a program that does not care what its batch quality is.\n\nInitial risk, residual uncertainty per unit cost and removable loss now sit in one pair of formulas, and design-unbiasedness is untouched: the probabilities depend only on quantities fixed before the labels are revealed, and $`\\hat{\\mu}_{\\text{raw}}`$ is still computed on frozen originals.\n\nThe prices also make the program cheap to solve. For fixed $`(\\lambda, \\nu, \\kappa)`$ every task's audit probability and action follow from the two formulas above in one pass, so a weekly batch of ten thousand tasks is a search over $`J + 2`$ multipliers with a linear-time step. The linear relaxation of the assignment splits at most one task per binding budget or capacity constraint, so rounding to whole assignments moves at most $`J + 1`$ tasks, and the gap between the priced bound and the rounded plan certifies the solution. When the program is infeasible, the price that diverges names what is short: the precision target, the specialist hours or the budget.\n\nThe ratio $`\\nu / \\lambda`$ is the exchange rate between the two products: the number of delivered defects the program gives up to tighten the estimate by one unit of variance. Section 6 reads it off the reference batch.\n\n## 6 Results on a reference batch\n\n**Reference batch.** Production records and operating figures are withheld under the nondisclosure agreement. The results below are stated on a reference batch at the program's scale: one week of work, $`N = 10,000`$ submissions in three strata by difficulty, most of them routine and a tenth of them hard, with defect rates and audit costs that rise with difficulty. Every result is given twice: at the reference batch, and as its range over an envelope in which each entry of the batch moves independently by up to 20% in either direction. A conclusion stated in this section holds across the whole envelope; none of them rests on a single operating point.\n\n**Table 1. The reference batch in three strata by difficulty.**\n\n| | Routine | Intermediate | Hardest |\n| --- | --- | --- | --- |\n| Share of the batch | 60% | 30% | 10% |\n| Defect rate of original submissions ($`q`$) | 1% | 10% | 35% |\n| Frozen risk prediction ($`f`$) | 2% | 8% | 30% |\n| Cost of an independent reference audit ($`c`$, routine = 1) | 1 | 1.5 | 3 |\n| Share of defects removed by one additional review | 80% | 80% | 30% |\n\nThe batch defect rate is $`\\mu = 7.10\\%`$. One review costs half of a reference audit in the same stratum. The audit budget is $`2,000`$ cost units, 14.8% of the cost of auditing everything, with a sampling floor of $`\\pi_{\\min} = 3\\%`$.\n\n**Measurement at equal audit budget.** Every policy below uses the estimator of Section 2 and spends the same expected audit budget.\n\n**Table 2. Audit policies at an equal expected audit budget.**\n\n| Audit policy | Audit probability by stratum | Standard deviation of the estimate (percentage points) | Variance relative to uniform | Range over the envelope |\n| --- | --- | --- | --- | --- |\n| Uniform sampling | 14.81% / 14.81% / 14.81% | 0.568 | Baseline | Baseline |\n| Risk-first (probability proportional to $`f_i`$) | 3.00% / 11.56% / 43.33% | 0.657 | +33.8% | -16% to +97% |\n| Residual- and cost-aware (Section 3) | 7.89% / 19.37% / 21.84% | **0.515** | **-17.8%** | -9% to -32% |\n\nThree results hold everywhere in the envelope. The uncorrected error rate in the risk-first queue reads **19.60%** against a batch rate of 7.10%, an overstatement of 2.76 times, and between 1.96 and 4.14 times across the envelope; the correction of Section 2 recovers 7.10% under every policy. The residual- and cost-aware allocation always measures better than uniform sampling. And risk-first routing always measures worse than that allocation, with **62.7%** more variance at the reference batch and 14% to 137% more across the envelope. How risk-first routing compares with random sampling is not stable: it is 34% worse at the reference batch and can fall on either side within the envelope. Its quality as a measurement design depends on how predicted risk happens to line up with residual uncertainty and audit cost, which nothing in the policy controls.\n\n**The price of precision.** The budget ratio of Section 4 is $`\\kappa = 0.848`$: the residual- and cost-aware allocation reaches any precision target with **15.2%** less audit budget than uniform sampling, and with 7.5% to 27.3% less across the envelope.\n\n**Table 3. Audit budget needed to reach each precision target.**\n\n| Target standard deviation (percentage points) | Audit budget, residual- and cost-aware | Audit budget, uniform |\n| --- | --- | --- |\n| 0.50 | 2,100 | 2,475 |\n| 0.40 | 2,974 | 3,506 |\n| 0.30 | 4,399 | 5,185 |\n| 0.25 | 5,419 | 6,387 |\n\n**Calibration tolerance.** With relative errors of alternating sign and size $`e`$ in the three stratum residual estimates, the variance rises above the optimum as follows.\n\n**Table 4. Variance cost of errors in the stratum residual estimates.**\n\n| Relative error in the residual estimates | Variance above the optimum | Quarter-variance law |\n| --- | --- | --- |\n| 10% | 0.30% | 0.30% |\n| 20% | 1.23% | 1.19% |\n| 30% | 2.88% | 2.68% |\n| 50% | 9.2% | 7.4% |\n\nThe stress case is a residual ordering reversed across the strata, which sends the allocation to 23.00% / 11.77% / 3.00% and raises the variance to 199.7% above uniform. With a weight of $`\\rho = 0.7`$ on the uniform design the same wrong estimates cost 17.1%, and the bound of Section 4 guarantees no more than 50.3% for any estimates.\n\n**One budget, two products.** The total budget is $`5,000`$ cost units, 37% of the cost of auditing everything, to be divided between audits and reviews. With no handling the batch ships 710 defects.\n\n**Table 5. One budget of 5,000 cost units divided between audits and reviews.**\n\n| Plan | Audit probability by stratum | Share of each stratum reviewed | Standard deviation of the estimate (percentage points) | Defects shipped |\n| --- | --- | --- | --- | --- |\n| Risk-first for audits and for reviews | 3.00% / 11.56% / 43.33% | 0% / 67% / 100% | 0.657 | 320.9 |\n| Measurement rule of Section 3, then reviews by removable loss | 7.89% / 19.37% / 21.84% | 0% / 100% / 50% | 0.515 | 336.2 |\n| Joint rule at the precision of risk-first routing | 3.31% / 9.57% / 100% | 0% / 61% / 0% | 0.657 | **197.1** |\n| Joint rule at the precision of uniform sampling | 4.20% / 12.74% / 100% | 0% / 52% / 0% | 0.568 | **209.9** |\n| Joint rule with no precision ceiling | 3.00% / 3.00% / 100% | 0% / 75% / 0% | 1.035 | 174.9 |\n\nAt the same budget and the same precision, the joint rule ships **38.6%** fewer defects than risk-first routing, and 19% to 67% fewer across the envelope. Held to the tighter precision of uniform sampling it still ships 34.6% fewer.\n\nTable 5 also shows why the two jobs cannot be planned one after the other. Choosing audits for measurement and then spending the rest on repair gives the best estimate in the table and ships more defects than the legacy policy, because it ignores the repair that audits perform. The joint rule sends every item of the hardest stratum to an independent reference solution: there a review removes 30% of the defects and a reference audit removes all of them, at a risk per unit cost of $`0.35 / 3 = 0.117`$ against a budget price $`\\lambda`$ of 0.093 to 0.096. In the two easier strata the audit probabilities follow the closed form of Section 5: the $`\\nu`$ implied by the routine stratum and the $`\\nu`$ implied by the intermediate stratum agree to six digits in both joint plans.\n\nThe last three rows are the exchange rate. Holding the estimate at 0.657 points costs 22 shipped defects relative to no ceiling, and tightening it to 0.568 costs 13 more. A program that knows this number can decide how much it wants to know.\n\nThe envelope for this table covers the eight entries that enter the comparison: the hardest stratum's share, prediction and audit cost, the defect rates and review effects of the two harder strata, and the ratio of review cost to audit cost.\n\n## 7 Operating the design\n\n**Contributor records.** Tiered review sends more work from weaker contributors to audit, so the number of audits a contributor has is correlated with reliability, and shrinkage toward a common prior misreads exactly that pattern ([Brown 2008](https://arxiv.org/abs/0803.3697)). The random reference audits supply the correction: their selection does not depend on the contributor's record, so reliability estimated on that stream is free of the routing effect and calibrates the posteriors built from adjudications. For decisions about coaching and review intensity the posterior is read through one number per contributor and stratum, the probability of falling below the stratum standard $`a_0`$:\n\n$$\nt_j = \\Pr(a_{jh} < a_0 \\mid \\text{audits}) = I_{a_0}\\big( \\alpha_{0h} + k_j,\\; \\beta_{0h} + n_j - k_j \\big) \\tag{7.1}\n$$\n\nwhere $`I`$ is the regularized incomplete beta function. Contributors enter the action list in decreasing order of $`t_j`$ until the average of $`1 - t_j`$ over the list reaches a set false-selection rate, which treats selection as the compound decision it is ([Gu and Koenker](https://doi.org/10.3982/ECTA19304)). A worked example with prior $`\\mathrm{Beta}(18, 2)`$ and standard $`a_0 = 0.85`$ shows what the rule corrects. A contributor with 33 of 40 audited tasks correct has $`t = 0.47`$; one with 6 of 8 correct has $`t = 0.41`$; one with 57 of 60 correct has $`t = 0.005`$. Raw accuracy would send the second contributor to coaching first on eight audits. The tail probability puts the first one ahead, on five times the evidence.\n\n**What an auditor's error costs.** The estimator measures the rate of audit labels, and audit labels are themselves a measurement. Let the auditor's sensitivity and specificity be estimated from blind re-audits, with $`J = \\text{sens} + \\text{spec} - 1`$. The true rate is recovered by the classical correction $`\\mu = (\\tilde{\\mu} - (1 - \\text{spec})) / J`$ ([Rogan and Gladen](https://pubmed.ncbi.nlm.nih.gov/623091/)), at $`1/J^2`$ times the variance, and the budget identity of Section 4 converts that into audit hours. On the reference batch, holding the true rate to a standard deviation of 0.568 points:\n\n**Table 6. Audit budget with an imperfect auditor, at a standard deviation of 0.568 points.**\n\n| Auditor error | Audit budget relative to an error-free auditor |\n| --- | --- |\n| 5% of defects missed | +5.2% |\n| 2% of clean items flagged | +43.8% |\n| Both | +52.8% |\n\nFalse flags cost about eight times what misses cost here. More than nine items in ten are clean, and a 2% false-flag rate on them nearly triples the residual variance of the routine stratum. The first target of auditor calibration is therefore the false flag.\n\n**Comparability over time.** Each estimate belongs to one frozen criteria version. Dual-version anchors link versions when the criteria change, and the same anchor set serves to detect drift in auditors' standards between changes.\n\n**Evaluation in production.** I defined separate production evaluations for the two jobs, because errors found in a targeted queue measure how well review locates defects and nothing else.\n\n1. **Inference on frozen, adjudicated batches.** Uniform sampling, legacy high-risk routing, cost-aware residual allocation and defensive mixtures are each set against complete independent adjudication, and judged on bias, mean squared error (MSE) and interval coverage within task strata and rubric versions.\n2. **Remediation trials at matched capacity.** Legacy tiered rules are set against marginal-gain ($`\\Delta_i`$) action allocation at equal total specialist hours, and judged on the severe defects remaining in the final post-training corpus and the fully loaded cost per accepted submission.\n\nThe program worked through roughly one batch a week, with more than 10,000 tasks per batch and more than 300 specialists. Contributor reliability records, tiered review workflows and random reference audits became its operating standards for quality measurement and review allocation, and later post-training programs took over the design.",
+ "markdown": "## 1 What a review queue measures\n\nIn a production RLHF pipeline, whatever policy routes submissions to review also changes the statistical properties of the observed sample. To see how, let $`X`$ denote observable task features, including the domain, prompt complexity and contributor history. Let $`Y \\in \\{0, 1\\}`$ indicate whether the original submission contains an error under fixed acceptance criteria, so that the true conditional error probability is $`q(x) = P(Y = 1 \\mid X = x)`$. Write $`S = 1`$ when a submission is selected for an independent reference audit. Conditional on $`X`$, selection is randomized with probability $`\\pi(x) > 0`$, and it is independent of the reference label, which has not yet been observed.\n\nSuppose an operational dashboard estimates population quality from the error fraction among verified submissions. What it actually targets is:\n\n$$\nE[Y \\mid S = 1] = \\frac{E[\\pi(X)q(X)]}{E[\\pi(X)]} \\tag{1.1}\n$$\n\n$$\nE[Y \\mid S = 1] - E[Y] = \\frac{\\mathrm{Cov}(\\pi(X), q(X))}{E[\\pi(X)]} \\tag{1.2}\n$$\n\nWhen review succeeds in targeting higher true error risk, the covariance between audit probability $`\\pi(X)`$ and true defect probability $`q(X)`$ is strictly positive, and the dashboard reads high. A rising detection rate in the reviewed sample is what better targeting looks like, and the batch may be unchanged or improving underneath it. The identities describe a selected item; in a finite batch the observed fraction has a random denominator and converges to this value.\n\nThis is Neyman's argument against purposive selection, applied to a review queue: a sample chosen to serve one purpose fails on the quantities it was not chosen for ([Neyman 1934](https://doi.org/10.2307/2342192)). Financial auditors already act on it. Under the [GAO Financial Audit Manual](https://www.gao.gov/assets/gao-18-601g.pdf) the auditor examines large and unusual items in full, samples the remainder at random, projects only the random sample to the population and records which selection is which. The PRM800K report describes the bias and leaves it there: its authors chose convincing wrong-answer solutions for labeling and noted that the choice skewed the data set toward wrong answers ([Let's Verify Step by Step](https://arxiv.org/abs/2305.20050)).\n\nExpert rankings suffer from the same selection mechanism. Let $`a_{jh}`$ be the probability that expert $`j`$ answers a task in stratum $`h`$ correctly, and let $`p_{jht}`$ describe that expert's assigned task mix in period $`t`$. Even if underlying ability stays constant, observed accuracy moves with the task mix:\n\n$$\nA_{jt} = \\sum_h p_{jht} a_{jh}, \\qquad A_{j,t+1} - A_{jt} = \\sum_h (p_{jh,t+1} - p_{jht}) a_{jh} \\tag{1.3}\n$$\n\nComparisons across periods therefore require a fixed target task mix $`w_h`$ and standardized accuracy:\n\n$$\nA_j^{\\text{std}} = \\sum_h w_h a_{jh} \\tag{1.4}\n$$\n\nOverlapping assignments or shared anchor cases support estimation within each stratum. In the project's production routing engine I used a conjugate Beta-Binomial model for each contributor's reliability within a task stratum. Given a prior $`\\mathrm{Beta}(\\alpha_{0h}, \\beta_{0h})`$ and $`n_{jht}`$ audited tasks with $`k_{jht}`$ verified successes, the posterior $`\\mathrm{Beta}(\\alpha_{0h} + k_{jht}, \\beta_{0h} + n_{jht} - k_{jht})`$ gives shrinkage-regularized point estimates and credible intervals that account for unequal sample sizes across comparable audits. Section 7 sets out how these records drive contributor decisions.\n\nEach piece of this problem has a literature, and each literature stops at one of the two jobs. [Dawid-Skene](https://doi.org/10.2307/2346806) models estimate how reviewers' confusion matrices differ and carry no sampling design. [PPI++](https://arxiv.org/html/2311.01453v1) gives valid inference from a small labeled sample and a large set of predictions when the labels are drawn uniformly. [Active Statistical Inference](https://arxiv.org/html/2403.03208v3) chooses which labels to collect and prices every label the same; [cost-optimal active evaluation](https://arxiv.org/abs/2506.07949) prices two kinds of rater, at one price per kind for all items. [ActiveLab](https://arxiv.org/html/2301.11856v1) decides whether the next annotation should go to a new example or to one already labeled, with a better classifier as the goal. [DeCCaF](https://arxiv.org/html/2403.06906v1) assigns cases to experts under error costs and capacity limits, and the quality of the batch is outside its objective. A production program has to price both jobs in one unit, expert hours, and the sections below do that.\n\n## 2 Measuring a batch with audits and predictions\n\nFix a production batch of $`N`$ original submissions, its acceptance criteria and its latent reference labels $`Y_1, \\dots, Y_N`$. The inferential target is the finite-batch error rate $`\\mu = N^{-1} \\sum_{i=1}^N Y_i`$. Let $`f_i \\in [0, 1]`$ be a fixed, inexpensive risk prediction that combines contributor reliability posteriors, task complexity and automated heuristic checks. Assign audit probability $`\\pi_i \\in [\\pi_{\\min}, 1]`$ and draw $`S_i \\sim \\mathrm{Bernoulli}(\\pi_i)`$ independently across submissions, with the predictions and the sampling probabilities locked before the current batch's reference labels are revealed.\n\nI used a prediction-assisted estimator with inverse-probability-weighted (IPW) residual correction:\n\n$$\n\\hat{\\mu} = \\frac{1}{N} \\sum_{i=1}^N \\left[ f_i + \\frac{S_i}{\\pi_i} (Y_i - f_i) \\right] \\tag{2.1}\n$$\n\nWrite the prediction residual as $`e_i = Y_i - f_i`$ and condition on the fixed batch, the predictions and the sampling probabilities:\n\n$$\n\\hat{\\mu} - \\mu = \\frac{1}{N} \\sum_{i=1}^N \\left( \\frac{S_i}{\\pi_i} - 1 \\right) e_i, \\qquad E_S[\\hat{\\mu} - \\mu] = 0 \\tag{2.2}\n$$\n\nValidity comes from the design and efficiency from the model. The inverse-probability weights restore the population representation that selective verification altered, and the estimate is unbiased for any predictions at all, because the only randomness it relies on is the audit lottery the program itself runs. Better predictions shrink the residuals and with them the number of audits a given precision needs. That division of labor suits production: the estimate stays correct while the risk model is still being improved. The form is the difference estimator of survey sampling, [Horvitz-Thompson](https://doi.org/10.1080/01621459.1952.10483446) weighting applied to residuals, and the augmented IPW estimator of [Robins, Rotnitzky and Zhao](https://doi.org/10.1080/01621459.1994.10476818); [PPI++](https://arxiv.org/html/2311.01453v1) studies it with black-box predictions.\n\nHow much the predictions can save has a known ceiling under uniform sampling: a predictor no more accurate than the work it scores can at most double the effective audit sample ([Dorner et al.](https://arxiv.org/abs/2410.13341)). The gains in Sections 3 and 4 come from where the audits are sent, which is a separate source of efficiency.\n\nTo use the estimator in production, I defined procedures for freezing the audited objects, recording rubric version histories and retaining the telemetry that optimization needs. Original and revised submissions are retained as separate database entities. Telling defects introduced during initial production apart from those corrected through review takes two estimates, $`\\hat{\\mu}_{\\text{raw}}`$ on frozen originals and $`\\hat{\\mu}_{\\text{final}}`$ on post-review deliverables. A change in post-training criteria causes a different confusion, since stricter acceptance standards can be misread as declining contributor ability. A shared set of anchor cases is therefore adjudicated under both rubric versions whenever the criteria change, and submissions from different periods are compared under a common reference version.\n\n## 3 Where audits buy the most information\n\nBecause the sampling is independent Bernoulli, the cross-item covariance terms drop out and the design variance is exactly:\n\n$$\n\\mathrm{Var}_S(\\hat{\\mu}) = \\frac{1}{N^2} \\sum_{i=1}^N \\frac{1 - \\pi_i}{\\pi_i} e_i^2 \\tag{3.1}\n$$\n\nThis expression sets the objective for audit allocation: audit effort should go to the uncertainty that automated risk predictions leave unresolved. For a binary error label with conditional error probability $`q_i`$, the expected squared residual has two components:\n\n$$\nm_i := E[(Y_i - f_i)^2 \\mid X_i] = q_i(1 - q_i) + (q_i - f_i)^2 \\tag{3.2}\n$$\n\nThe first is the conditional Bernoulli variance $`q_i(1 - q_i)`$ and the second is the squared calibration error $`(q_i - f_i)^2`$. A high error probability $`q_i`$ therefore need not carry high inferential value. If a task category is almost certain to contain an error ($`q_i \\approx 1`$) and the predictor recognizes this ($`f_i \\approx 1`$), then $`m_i \\approx 0`$, and an independent reference audit of that item says less about population quality than an audit of a moderately risky item whose outcome is still open ($`q_i \\approx 0.5`$). A repair model and a measurement model look at the same task and want different things from it.\n\nLet $`c_i > 0`$ be the empirical cost of an independent reference audit, calibrated by reconciling platform review logs with active-work records from the team's KPI monitoring system, and let $`B`$ be the expected audit-cost budget. With residual second moments $`m_i > 0`$, the optimal design solves:\n\n$$\n\\min_{\\pi} \\sum_{i=1}^N \\frac{m_i}{\\pi_i} \\quad \\text{s.t.} \\quad \\sum_{i=1}^N c_i \\pi_i \\le B, \\quad \\pi_{\\min} \\le \\pi_i \\le 1 \\tag{3.3}\n$$\n\nThe objective omits the constant $`-\\sum_i m_i`$, which does not depend on the allocation. For probabilities strictly between their bounds, the Lagrangian first-order condition gives the clipped square-root allocation:\n\n$$\n-\\frac{m_i}{\\pi_i^2} + \\lambda c_i = 0 \\implies \\pi_i^* = \\mathrm{clip}_{[\\pi_{\\min}, 1]} \\left( \\sqrt{\\frac{m_i}{\\lambda c_i}} \\right) \\tag{3.4}\n$$\n\nThe multiplier $`\\lambda > 0`$ is set by the budget, feasibility requires $`B \\ge \\pi_{\\min} \\sum_i c_i`$, and a budget that covers a full audit audits everything. The rule samples harder where expected residuals are larger and more lightly where audits cost more, while the floor $`\\pi_{\\min}`$ keeps every part of the task space covered and bounds the IPW weights.\n\nThe square-root form has a long ancestry. With equal costs and strata it is Neyman's allocation; with stratum costs it is the optimum allocation of [Cochran's text](https://www.wiley.com/en-us/Sampling+Techniques%2C+3rd+Edition-p-9780471162407); inclusion probabilities proportional to a model's error standard deviation are the anticipated-variance optimum of [Isaki and Fuller](https://doi.org/10.1080/01621459.1982.10477770); and [Chen and Lumley](https://doi.org/10.1002/sim.8760) set such designs wave by wave from earlier waves. Machine-learning work reached the same rule for labels of equal cost. Expert audits do not cost the same: on the reference batch of Section 6, an independent reference solution for the hardest tasks costs three times a routine one. My rule carries a cost for every item.\n\nProduction allocation uses regularized stratum estimates $`\\hat{m}_i > 0`$ from earlier independent audits or a pilot calibration batch, at the level of `(task stratum × criteria version × cost stratum)`, so that no parameter is fitted to an individual submission. A weekly program is a multiwave design in which each batch's audits set the next batch's allocation. The budget $`B`$ constrains expected audit cost per batch; a workflow that needs an exact audit count uses a fixed-size design with the same first-order inclusion probabilities.\n\n## 4 What the allocation is worth, and how accurate its inputs must be\n\nLet $`C = \\sum_i c_i`$, $`M = \\sum_i m_i`$ and $`S = \\sum_i \\sqrt{m_i c_i}`$. Uniform sampling at the same expected cost uses probability $`p = B / C`$. With no probability at a clipping boundary, averaging the design variance over the conditional label distribution gives:\n\n$$\nV_{\\text{uniform}} = \\frac{1}{N^2} \\left[ \\frac{C M}{B} - M \\right], \\qquad V_{\\text{optimal}} = \\frac{1}{N^2} \\left[ \\frac{S^2}{B} - M \\right] \\tag{4.1}\n$$\n\nIf all items within a task stratum share a sampling probability, substituting that stratum's finite-batch mean squared residual for $`m_i`$ makes these exact finite-batch identities. By the Cauchy-Schwarz inequality $`S^2 \\le C M`$, with equality if and only if $`m_i / c_i`$ is constant across items. The gain exists exactly where residual uncertainty per unit of audit cost varies across the batch.\n\n**The same precision for a fixed fraction of the budget.** A program manager fixes the precision the dashboard needs and asks what it costs. Solving each variance for the budget that reaches a target $`\\tau^2`$:\n\n$$\nB_{\\text{optimal}}(\\tau^2) = \\frac{S^2}{N^2 \\tau^2 + M}, \\qquad B_{\\text{uniform}}(\\tau^2) = \\frac{C M}{N^2 \\tau^2 + M}, \\qquad \\frac{B_{\\text{optimal}}}{B_{\\text{uniform}}} = \\kappa := \\frac{S^2}{C M} \\tag{4.2}\n$$\n\nThe ratio $`\\kappa`$ does not depend on the target. Whatever precision the program asks for, the residual- and cost-aware allocation delivers it for the same fraction of the uniform audit budget, and that fraction is a property of the batch: how unevenly residual uncertainty and audit cost are spread across it. The hours it frees, $`(1 - \\kappa) B_{\\text{uniform}}`$, are available for repair, which is the first link between the two jobs.\n\n**How accurate the residual estimates must be.** The allocation is computed from estimates $`\\hat{m}_i`$. Write $`\\hat{m}_i = m_i (1 + \\varepsilon_i)`$ and weight each item by its share $`w_i = \\sqrt{m_i c_i} / S`$ of the optimal budget. With neither allocation clipped, the variance of the allocation built on the estimates exceeds the optimum by:\n\n$$\nV(\\hat{\\pi}) - V(\\pi^*) = \\frac{S^2}{N^2 B} \\left[ \\Big( \\sum_i w_i \\sqrt{1 + \\varepsilon_i} \\Big) \\Big( \\sum_i \\frac{w_i}{\\sqrt{1 + \\varepsilon_i}} \\Big) - 1 \\right] = \\frac{S^2}{N^2 B} \\left[ \\tfrac{1}{4} \\mathrm{Var}_w(\\varepsilon) + O(\\|\\varepsilon\\|^3) \\right] \\tag{4.3}\n$$\n\nThe first-order terms cancel. Three consequences follow for the specification of the estimation pipeline. Overestimating or underestimating every residual by a common factor costs nothing, since only the dispersion of the relative errors enters. A relative error of 20% across strata raises the variance by about one percent of $`S^2/(N^2 B)`$, so stratum-level estimates carried over from earlier batches need to be accurate only to that order. And the exact expression shows where the rule can be hurt: through $`1/\\sqrt{1 + \\varepsilon_i}`$, by a gross underestimate of a residual somewhere in the task space.\n\nTwo guards bound that last case. The floor $`\\pi_{\\min}`$ bounds every weight. Mixing the computed probabilities with the uniform design at the same budget,\n\n$$\n\\pi_i(\\rho) = (1 - \\rho)\\hat{\\pi}_i + \\rho p, \\tag{4.4}\n$$\n\nbounds the variance for any estimates whatever, because every probability is then at least $`\\rho p`$:\n\n$$\n\\frac{V(\\pi(\\rho))}{V_{\\text{uniform}}} \\le \\frac{1 - \\rho p}{\\rho (1 - p)} \\tag{4.5}\n$$\n\nThis is the defensive mixture of importance sampling ([Hesterberg 1995](https://doi.org/10.1080/00401706.1995.10484303)) written for a Bernoulli audit design. The weight $`\\rho`$ is tuned on earlier batches; [robust active inference](https://arxiv.org/html/2511.08991v1) later gave an asymptotic choice of the same weight.\n\n## 5 One budget, two products\n\nStatistical auditing measures population quality. Operational review has a different job, removing defects from the current batch before it is delivered for training. For task $`i`$ under handling action $`a`$, such as direct acceptance, peer review, expert adjudication or contributor revision, let $`r_{ia}`$ be the expected error loss remaining after the action, $`d_{ia}`$ its operational cost, and $`t_{iaj}`$ the time required from specialist $`j`$, measured through platform records and the team's KPI monitoring system. Binary assignment decisions $`z_{ia} \\in \\{0, 1\\}`$ solve:\n\n$$\n\\min_{z} F(z) = \\sum_{i,a} z_{ia}(r_{ia} + \\eta d_{ia}) \\quad \\text{s.t.} \\quad \\sum_a z_{ia} = 1, \\quad \\sum_{i,a} z_{ia} t_{iaj} \\le H_j \\quad \\forall j \\tag{5.1}\n$$\n\nHere $`H_j`$ is the specialist's available time, and $`\\eta \\ge 0`$ converts resource cost to the scale of error loss.\n\n**Repair follows removable loss.** When review costs are equal and each reviewed task uses one slot, the marginal remediation gain $`\\Delta_i = r_{i0} - r_{i1}`$ determines the optimal priority order. A high-risk task may be hard to repair through one additional review because external evidence is missing, which gives a small $`\\Delta_i`$. A self-contained calculation error in a moderate-risk task may be fully correctable, which gives a large $`\\Delta_k`$. Moving one review slot from task $`i`$ to task $`k`$ changes total expected loss by:\n\n$$\n\\Delta F = (r_{i0} + r_{k1}) - (r_{i1} + r_{k0}) = \\Delta_i - \\Delta_k \\tag{5.2}\n$$\n\nWhenever $`\\Delta_k > \\Delta_i`$, the exchange strictly reduces delivered defect loss, regardless of the initial risk ordering. Optimal triage under a budget has the same threshold structure, with the human's case-level loss in place of the model's risk ([Okati et al.](https://proceedings.nips.cc/paper_files/paper/2021/file/4c4c937b67cc8d785cea1e42ccea185c-Paper.pdf)). A worked example fixes the size of the effect. Two task groups of $`100`$ items each have initial error risks of $`0.40`$ and $`0.15`$, and additional review reduces these risks to $`0.30`$ ($`\\Delta_1 = 0.10`$) and $`0.01`$ ($`\\Delta_2 = 0.14`$). Review costs are equal and capacity is $`100`$ items. Ranking by initial risk leaves $`100(0.30) + 100(0.15) = 45`$ expected errors; ranking by $`\\Delta_i`$ leaves $`100(0.40) + 100(0.01) = 41`$, a reduction of **8.89%** at the same review volume.\n\nThe quantity that has to be calibrated is the post-action risk. Suppose $`\\left\\vert\\hat{r}_{ia} - r_{ia}\\right\\vert \\le \\delta`$ for every feasible action, costs and capacities are known, and $`\\hat{z}`$ minimizes the predicted objective $`\\hat{F}(z)`$. Relative to the true optimal feasible assignment $`z^*`$:\n\n$$\nF(\\hat{z}) \\le \\hat{F}(\\hat{z}) + N\\delta \\le \\hat{F}(z^*) + N\\delta \\le F(z^*) + 2N\\delta \\tag{5.3}\n$$\n\nThat bound is the reason the program ran randomized pilot trials across the permitted review actions: they estimate $`r_{ia}`$ directly.\n\n**The batch program.** Audits and repair draw on the same money. Let $`\\tau^2`$ be the variance ceiling for the quality estimate, $`B_{\\text{total}}`$ the total cost budget, and $`V_{\\hat{m}}(\\pi) = N^{-2}\\sum_i (1 - \\pi_i)\\hat{m}_i / \\pi_i`$ the planning variance under the calibrated residual model. A reference audit produces an independent reference solution, so an audited item found defective ships corrected. Writing $`a_i`$ for the action assigned to item $`i`$, the item ships with expected loss $`(1 - \\pi_i)\\, \\hat{r}_{i a_i}`$, and the batch is planned by a mixed-integer convex program:\n\n$$\n\\min_{z, \\pi} \\sum_{i,a} z_{ia} (1 - \\pi_i)\\, \\hat{r}_{ia} \\quad \\text{s.t.} \\quad \\sum_{i,a} z_{ia} d_{ia} + \\sum_i c_i \\pi_i \\le B_{\\text{total}}, \\quad V_{\\hat{m}}(\\pi) \\le \\tau^2 \\tag{5.4}\n$$\n\n$$\n\\sum_a z_{ia} = 1, \\quad z_{ia} \\in \\{0, 1\\}, \\quad \\sum_{i,a} z_{ia} t_{iaj} \\le H_j^{\\text{review}} \\ \\forall j, \\quad \\pi_{\\min} \\le \\pi_i \\le 1 \\tag{5.5}\n$$\n\nThe hard capacity $`H_j^{\\text{review}}`$ goes to remediation actions, and reference audits are scheduled independently on frozen original submissions.\n\n**The joint rule.** Attach a price $`\\lambda`$ to the budget, $`\\nu`$ to the precision ceiling (in units of $`N^2 V`$) and $`\\gamma_j`$ to each specialist's hours. At those prices the program separates into one small problem per task, and each has a closed form:\n\n$$\n\\pi_i^* = \\mathrm{clip}_{[\\pi_{\\min}, 1]} \\left( \\sqrt{\\frac{\\nu\\, m_i}{\\big( \\lambda c_i - r_{i a_i} \\big)_+}} \\right), \\qquad a_i^* = \\arg\\min_a \\; (1 - \\pi_i)\\, r_{ia} + \\lambda d_{ia} + \\sum_j \\gamma_j t_{iaj} \\tag{5.6}\n$$\n\nwith $`\\pi_i^* = 1`$ when $`r_{i a_i} \\ge \\lambda c_i`$. Three readings of the audit rule carry the argument of this article.\n\n1. *Risk is a rebate.* The item's residual risk $`r_{i a_i}`$ is subtracted from the audit's price $`\\lambda c_i`$. Risk does not drive the audit; it makes the audit cheaper, by the repair the audit performs on the way.\n2. *Measurement alone.* If audits repaired nothing, the rebate would vanish and the rule would be the $`\\sqrt{m_i / c_i}`$ allocation of Section 3.\n3. *Repair alone.* If nobody valued the estimate, $`\\nu = 0`$, the rule audits exactly the items with $`r_{i a_i} \\ge \\lambda c_i`$, a threshold on risk per unit cost. That is risk-first routing. The legacy policy is the joint rule of a program that does not care what its batch quality is.\n\nInitial risk, residual uncertainty per unit cost and removable loss now sit in one pair of formulas, and design-unbiasedness is untouched: the probabilities depend only on quantities fixed before the labels are revealed, and $`\\hat{\\mu}_{\\text{raw}}`$ is still computed on frozen originals.\n\nThe prices also make the program cheap to solve. For fixed $`(\\lambda, \\nu, \\gamma)`$ every task's audit probability and action follow from the two formulas above in one pass, so a weekly batch of ten thousand tasks is a search over $`\\lambda`$, $`\\nu`$ and one price per specialist, with a linear-time step. The linear relaxation of the assignment splits at most one task per binding budget or capacity constraint, so rounding to whole assignments moves at most that many tasks, and the gap between the priced bound and the rounded plan certifies the solution. When the program is infeasible, the price that diverges names what is short: the precision target, the specialist hours or the budget.\n\nThe ratio $`\\nu / \\lambda`$ is the exchange rate between the two products: the number of delivered defects the program gives up to tighten the estimate by one unit of variance. Section 6 reads it off the reference batch.\n\n## 6 Results on a reference batch\n\n**Reference batch.** Production records and operating figures are withheld under the nondisclosure agreement. The results below are stated on a reference batch at the program's scale: one week of work, $`N = 10{,}000`$ submissions in three strata by difficulty, most of them routine and a tenth of them hard, with defect rates and audit costs that rise with difficulty. Each headline result is given twice: at the reference batch, and as its range over an envelope in which each entry of the batch moves independently by up to 20% in either direction.\n\n**Table 1. The reference batch in three strata by difficulty.**\n\n| | Routine | Intermediate | Hardest |\n| --- | --- | --- | --- |\n| Share of the batch | 60% | 30% | 10% |\n| Defect rate of original submissions ($`q`$) | 1% | 10% | 35% |\n| Frozen risk prediction ($`f`$) | 2% | 8% | 30% |\n| Cost of an independent reference audit ($`c`$, routine = 1) | 1 | 1.5 | 3 |\n| Share of defects removed by one additional review | 80% | 80% | 30% |\n\nThe batch defect rate is $`\\mu = 7.10\\%`$. One review costs half of a reference audit in the same stratum. The audit budget is 2,000 cost units, 14.8% of the cost of auditing everything, with a sampling floor of $`\\pi_{\\min} = 3\\%`$.\n\n**Measurement at equal audit budget.** Every policy below uses the estimator of Section 2 and spends the same expected audit budget.\n\n**Table 2. Audit policies at an equal expected audit budget.**\n\n| Audit policy | Audit probability by stratum | Standard deviation of the estimate (percentage points) | Variance relative to uniform | Range over the envelope |\n| --- | --- | --- | --- | --- |\n| Uniform sampling | 14.81% / 14.81% / 14.81% | 0.568 | Baseline | Baseline |\n| Risk-first (probability proportional to $`f_i`$) | 3.00% / 11.56% / 43.33% | 0.657 | +33.8% | -16% to +97% |\n| Residual- and cost-aware (Section 3) | 7.89% / 19.37% / 21.84% | **0.515** | **-17.8%** | -9% to -32% |\n\nThree results hold everywhere in the envelope. The uncorrected error rate in the risk-first queue reads **19.60%** against a batch rate of 7.10%, an overstatement of 2.76 times, and between 1.96 and 4.14 times across the envelope; the correction of Section 2 recovers 7.10% under every policy. The residual- and cost-aware allocation always measures better than uniform sampling. And risk-first routing always measures worse than that allocation, with **62.7%** more variance at the reference batch and 14% to 137% more across the envelope. How risk-first routing compares with random sampling is not stable: it is 34% worse at the reference batch and can fall on either side within the envelope. Its quality as a measurement design depends on how predicted risk happens to line up with residual uncertainty and audit cost, which nothing in the policy controls.\n\n**The price of precision.** The budget ratio of Section 4 is $`\\kappa = 0.848`$: the residual- and cost-aware allocation reaches any precision target with **15.2%** less audit budget than uniform sampling, and with 7.5% to 27.3% less across the envelope.\n\n**Table 3. Audit budget needed to reach each precision target.**\n\n| Target standard deviation (percentage points) | Audit budget, residual- and cost-aware | Audit budget, uniform |\n| --- | --- | --- |\n| 0.50 | 2,100 | 2,475 |\n| 0.40 | 2,974 | 3,506 |\n| 0.30 | 4,399 | 5,185 |\n| 0.25 | 5,419 | 6,387 |\n\n**Calibration tolerance.** With relative errors of alternating sign and size $`e`$ in the three stratum residual estimates, the variance rises above the optimum as follows.\n\n**Table 4. Variance cost of errors in the stratum residual estimates.**\n\n| Relative error in the residual estimates | Variance above the optimum | Quarter-variance law |\n| --- | --- | --- |\n| 10% | 0.30% | 0.30% |\n| 20% | 1.23% | 1.19% |\n| 30% | 2.88% | 2.68% |\n| 50% | 9.2% | 7.4% |\n\nThe stress case is a residual ordering reversed across the strata, which sends the allocation to 23.00% / 11.77% / 3.00% and raises the variance to 199.7% above uniform. With a weight of $`\\rho = 0.7`$ on the uniform design the same wrong estimates cost 17.1%, and the bound of Section 4 guarantees no more than 50.3% for any estimates.\n\n**One budget, two products.** The total budget is 5,000 cost units, 37% of the cost of auditing everything, to be divided between audits and reviews. With no handling the batch ships 710 defects.\n\n**Table 5. One budget of 5,000 cost units divided between audits and reviews.**\n\n| Plan | Audit probability by stratum | Share of each stratum reviewed | Standard deviation of the estimate (percentage points) | Defects shipped |\n| --- | --- | --- | --- | --- |\n| Risk-first for audits and for reviews | 3.00% / 11.56% / 43.33% | 0% / 67% / 100% | 0.657 | 320.9 |\n| Measurement rule of Section 3, then reviews by removable loss | 7.89% / 19.37% / 21.84% | 0% / 100% / 50% | 0.515 | 336.2 |\n| Joint rule at the precision of risk-first routing | 3.31% / 9.57% / 100% | 0% / 61% / 0% | 0.657 | **197.1** |\n| Joint rule at the precision of uniform sampling | 4.20% / 12.74% / 100% | 0% / 52% / 0% | 0.568 | **209.9** |\n| Joint rule with no precision ceiling | 3.00% / 3.00% / 100% | 0% / 75% / 0% | 1.035 | 174.9 |\n\nAt the same budget and the same precision, the joint rule ships **38.6%** fewer defects than risk-first routing, and 19% to 67% fewer across the envelope. Held to the tighter precision of uniform sampling it still ships 34.6% fewer.\n\nTable 5 also shows why the two jobs cannot be planned one after the other. Choosing audits for measurement and then spending the rest on repair gives the best estimate in the table and ships more defects than the legacy policy, because it ignores the repair that audits perform. Held to the sequential plan's own precision of 0.515 points, the joint rule ships 34.7% fewer defects than that plan, and 12.4% to 69.3% fewer across the whole envelope. The joint rule sends every item of the hardest stratum to an independent reference solution: there a review removes 30% of the defects and a reference audit removes all of them, at a risk per unit cost of $`0.35 / 3 = 0.117`$ against a budget price $`\\lambda`$ of 0.093 to 0.096. In the two easier strata the audit probabilities follow the closed form of Section 5: the $`\\nu`$ implied by the routine stratum and the $`\\nu`$ implied by the intermediate stratum agree to six digits in both joint plans.\n\nThe last three rows are the exchange rate. Holding the estimate at 0.657 points costs 22 shipped defects relative to no ceiling, and tightening it to 0.568 costs 13 more. A program that knows this number can decide how much it wants to know.\n\nThe envelope for this table covers the eight entries that enter the comparison: the hardest stratum's share, prediction and audit cost, the defect rates and review effects of the two harder strata, and the ratio of review cost to audit cost.\n\n## 7 Operating the design\n\n**Contributor records.** Tiered review sends more work from weaker contributors to audit, so the number of audits a contributor has is correlated with reliability, and shrinkage toward a common prior misreads exactly that pattern ([Brown 2008](https://arxiv.org/abs/0803.3697)). The random reference audits supply the correction: their selection does not depend on the contributor's record, so reliability estimated on that stream is free of the routing effect and calibrates the posteriors built from adjudications. For decisions about coaching and review intensity the posterior is read through one number per contributor and stratum, the probability of falling below the stratum standard $`a_0`$:\n\n$$\nt_j = \\Pr(a_{jh} < a_0 \\mid \\text{audits}) = I_{a_0}\\big( \\alpha_{0h} + k_j,\\; \\beta_{0h} + n_j - k_j \\big) \\tag{7.1}\n$$\n\nwhere $`I`$ is the regularized incomplete beta function. Contributors enter the action list in decreasing order of $`t_j`$ until the average of $`1 - t_j`$ over the list reaches a set false-selection rate, which treats selection as the compound decision it is ([Gu and Koenker](https://doi.org/10.3982/ECTA19304)). A worked example with prior $`\\mathrm{Beta}(18, 2)`$ and standard $`a_0 = 0.85`$ shows what the rule corrects. A contributor with 33 of 40 audited tasks correct has $`t = 0.47`$; one with 6 of 8 correct has $`t = 0.41`$; one with 57 of 60 correct has $`t = 0.005`$. Raw accuracy would send the second contributor to coaching first on eight audits. The tail probability puts the first one ahead, on five times the evidence.\n\n**What an auditor's error costs.** The estimator measures the rate of audit labels, and audit labels are themselves a measurement. Let the auditor's sensitivity and specificity be estimated from blind re-audits, with $`J = \\text{sens} + \\text{spec} - 1`$. The true rate is recovered by the classical correction $`\\mu = (\\tilde{\\mu} - (1 - \\text{spec})) / J`$ ([Rogan and Gladen](https://pubmed.ncbi.nlm.nih.gov/623091/)), at $`1/J^2`$ times the variance, and the budget identity of Section 4 converts that into audit hours. On the reference batch, holding the true rate to a standard deviation of 0.568 points:\n\n**Table 6. Audit budget with an imperfect auditor, at a standard deviation of 0.568 points.**\n\n| Auditor error | Audit budget relative to an error-free auditor |\n| --- | --- |\n| 5% of defects missed | +5.2% |\n| 2% of clean items flagged | +43.8% |\n| Both | +52.8% |\n\nFalse flags cost about eight times what misses cost here. More than nine items in ten are clean, and a 2% false-flag rate on them nearly triples the residual variance of the routine stratum. The first target of auditor calibration is therefore the false flag.\n\n**Comparability over time.** Each estimate belongs to one frozen criteria version. Dual-version anchors link versions when the criteria change, and the same anchor set serves to detect drift in auditors' standards between changes.\n\n**Evaluation in production.** I defined separate production evaluations for the two jobs, because errors found in a targeted queue measure how well review locates defects and nothing else.\n\n1. **Inference on frozen, adjudicated batches.** Uniform sampling, legacy high-risk routing, cost-aware residual allocation and defensive mixtures are each set against complete independent adjudication, and judged on bias, mean squared error (MSE) and interval coverage within task strata and rubric versions.\n2. **Remediation trials at matched capacity.** Legacy tiered rules are set against marginal-gain ($`\\Delta_i`$) action allocation at equal total specialist hours, and judged on the severe defects remaining in the final post-training corpus and the fully loaded cost per accepted submission.\n\nThe program worked through roughly one batch a week, with more than 10,000 tasks per batch and more than 300 specialists. Contributor reliability records, tiered review workflows and random reference audits became its operating standards for quality measurement and review allocation, and later post-training programs took over the design.",
"sourceDocument": {
"fileName": "ai-research-and-engineering.md",
"path": "content/sources/ai-portfolio/ai-research-and-engineering.md",
- "sha256": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1"
+ "sha256": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0"
},
"repositoryUrl": "https://github.com/130U/RESEARCH-Large-Scale-Domain-Expert-Feedback-Governance-and-Adaptive-Review"
}
diff --git a/website/content/blog/ten-states-of-water.md b/website/content/blog/ten-states-of-water.md
index cc89dff..c7a1e3a 100644
--- a/website/content/blog/ten-states-of-water.md
+++ b/website/content/blog/ten-states-of-water.md
@@ -15,7 +15,7 @@ workDescription: Ten landscapes of Alaska and Yellowstone
Ten States of Water is ten images of Alaska and Yellowstone, with water in a different state in each. The water arrives as tide, stands as a glacier, lies frozen on a lake, falls through a canyon, and leaves a hot spring as steam.
-The site is built to be looked at slowly. A cover opens it, two albums follow, and every image keeps a short reading beside it. I add to the albums as new work is finished.
+The site is built to be looked at slowly. A cover opens it, two albums follow, and every image keeps a short reading beside it.
## The Water
diff --git a/website/content/experience-showcase.json b/website/content/experience-showcase.json
index b9c522a..6835faf 100644
--- a/website/content/experience-showcase.json
+++ b/website/content/experience-showcase.json
@@ -112,7 +112,7 @@
"**I developed checkable error certificates and bounds on solution quality for programmatically verifiable task families.** The projected-residual optimality bound ($`0 \\le f(x) - v^* \\le \\|P_{\\ker A} r\\|_2^2 / (2\\mu)`$) splits feasibility, stationarity and the accuracy of a conclusion into separate verification obligations, and different proofs can pass by satisfying the same ones. The same template, a residual the submission exposes, a constant the designer verifies and an inequality between them, covers convex programs, linear systems, eigenpairs and fixed points. None of these checks needs a reference solution.",
"**I established complementary forms of verification across mathematical certificates, code and rule checks, LLM-assisted screening and specialist adjudication, and later programs reused the layered design.** Code checks directly verifiable answers, required deliverables, tool permissions and other requirements specified in advance. LLM-as-a-Judge helps flag missing evidence, misapplied rules and gaps in reasoning, and disputed domain judgments go to scholars with relevant expertise. I derived when a screen may decide: at screening sensitivity $`\\text{se}`$, a share $`1 - \\alpha^*/(1 - \\text{se})`$ of unflagged work still has to reach specialists to hold the system's false acceptance at $`\\alpha^*`$.",
"**I evaluated answers, visible reasoning and execution evidence together.** I cross-checked the model's displayed CoT against its actual tool calls, return values and intermediate outputs to see whether what it claimed had observable support. From elapsed time, attempt counts and changes in intermediate results I identified repeated operations, continued use of failed assumptions and attempts that added no evidence. Correctness, compliance and efficiency were recorded as separate results, and a resource limit counted toward acceptance only when the task specification set it beforehand.",
- "**I derived how false acceptance ($`\\alpha`$) and false rejection ($`\\beta`$) reverse model rankings, and the condition under which a ranking is certified.** In the worked example, models with true validity rates of 72% and 64% score 73.12% and 76.96%, so the weaker one ranks first. A lead of $`g`$ points over a rival scoring $`s_B`$ certifies the order of true validity exactly when the grader's false-acceptance rate is below $`g / (1 - s_B)`$.",
+ "**I derived how false acceptance ($`\\alpha`$) and false rejection ($`\\beta`$) reverse model rankings, and the condition under which a ranking is certified.** In the worked example, models with true validity rates of 72% and 64% score 73.12% and 76.96%, so the weaker one ranks first. A lead of $`g`$ over a rival scoring $`s_B`$ certifies the order of true validity exactly when the grader's false-acceptance rate is below $`g / (1 - s_B)`$.",
"**I showed that a policy optimized against a grader converges to the grader's positive predictive value.** The observed score goes to 100% while true validity stops at $`\\theta(1-\\beta) / [\\theta(1-\\beta) + (1-\\theta)\\alpha]`$. False acceptance sets the ceiling, and false rejection removes valid routes from what the policy produces. The same analysis gives the bias in the binary cross-entropy gradient used to train process reward models ($`E[\\nabla_s \\ell(s,V) - \\nabla_s \\ell(s,Z) \\mid h] = p(h) - \\tilde{p}(h)`$).",
"**I gave the tolerance-propagation rule ($`a\\delta + 2\\varepsilon`$) for bijective task transformations and its composition law for chains of transformations.** Together they say when a transformed task still supports a valid comparison and what tolerance its solutions are owed.",
"**On a test set of 3,600 exact-arithmetic submissions from 600 strongly convex quadratic task families, every verdict known, final-answer matching accepted 75.0% of the invalid submissions and strict reference-path matching rejected 50.0% of the valid proofs.** Certificate checking accepted all 1,200 valid proofs and rejected all 2,400 invalid submissions, so strict error detection and acceptance of valid alternative routes held together on the same set."
@@ -120,7 +120,7 @@
"notes": [],
"problemAndApproach": {
"title": "Problem and Approach",
- "text": "The question I took up in late 2023 was what should entitle a model's reasoning to be accepted. A correct answer can coexist with unsupported assumptions, misapplied rules or tool use that violates the task specification, and a fixed reference path will reject sound work that takes another route. I designed an acceptance rule around proof obligations: a conclusion stands when its assumptions, derivations and execution evidence support it, whichever valid route the model took, and a failed obligation removes support from everything that depends on it. Programmatic certificates discharge checkable obligations, LLM-assisted screening identifies possible gaps, and specialist scholars adjudicate disputed semantic questions.\n\nTwo things follow from defining acceptance this way. A certificate bounds how far a submission is from the truth using only a residual the submission exposes and a constant the task designer has verified, so the check never consults a reference solution and inherits none of its errors. And once a grader's two error rates are named, they can be priced: I derived the false-acceptance rate below which a leaderboard lead is certified, the level of true validity a policy can reach when the grader is its reward, and the screening accuracy at which an LLM judge may decide instead of screen. I kept refining the design through September 2026. Sections 1 through 7 set out the acceptance criteria, the certificates, the thresholds, the measurements and the related literature.\n\nThe team produced roughly 500 scientific reasoning tasks a week. Each task included a detailed, step-by-step reference solution and repeated verification of the argument."
+ "text": "The question I took up in late 2023 was what should entitle a model's reasoning to be accepted. A correct answer can coexist with unsupported assumptions, misapplied rules or tool use that violates the task specification, and a fixed reference path will reject sound work that takes another route. I designed an acceptance rule around proof obligations: a conclusion stands when its assumptions, derivations and execution evidence support it, whichever valid route the model took, and a failed obligation removes support from everything that depends on it. Programmatic certificates discharge checkable obligations, LLM-assisted screening identifies possible gaps, and specialist scholars adjudicate disputed semantic questions.\n\nTwo things follow from defining acceptance this way. A certificate bounds how far a submission is from the truth using only a residual the submission exposes and a constant the task designer has verified, so the check never consults a reference solution and inherits none of its errors. And once a grader's two error rates are named, they can be priced: I derived the false-acceptance rate below which a leaderboard lead is certified, the level of true validity a policy can reach when the grader is its reward, and the screening sensitivity at which an LLM judge may decide instead of screen. I kept refining the design through September 2026. Sections 1 through 7 set out the acceptance criteria, the certificates, the thresholds, the measurements and the related literature.\n\nThe team produced roughly 500 scientific reasoning tasks a week. Each task included a detailed, step-by-step reference solution and repeated verification of the argument."
}
},
{
@@ -227,19 +227,20 @@
"id": 8,
"categoryId": "03",
"organization": "Autonomous Authority in Space: Risk Tradeoffs and the Law of Delegation",
- "summary": "A spacecraft sometimes has to act before anyone on the ground can approve the command, and waiting carries risks of its own. I formulated a two-threshold test that separates the authority to act from the evidence needed to delegate execution, and applied it to delayed spacecraft operations, orbital avoidance, debris capture, and SETI. I also proposed task-specific authorization records, so that a delegated decision can be examined in advance, revised when its assumptions change, and reconstructed after the event.",
+ "summary": "A spacecraft sometimes has to act before anyone on the ground can approve the command, and waiting carries risks of its own. I formulated a two-threshold test, drawn from the law of delegation, that separates the authority to act from the evidence needed to delegate execution, and applied it to delayed spacecraft operations, orbital avoidance, debris capture, and SETI. I also proposed task-specific authorization records, so that a delegated decision can be examined in advance, revised when its assumptions change, and reconstructed after the event.",
"readMoreLabel": "Read More",
"detailTitle": "Autonomous Authority in Space: Risk Tradeoffs and the Law of Delegation",
"metadata": {
- "position": "Author (Letao Ouyang)",
+ "position": "Author, Letao (Theodore) Ouyang",
"location": "Durham, North Carolina",
"dates": "December 2024; revised November 2026"
},
"contributionsLabel": "Key Contributions",
"keyContributions": [
"I specified a two-threshold decision procedure linking legal authority, evidence for delegation, and the time at which human review can still change the outcome, distinguishing individual approval, advance-constrained action, and fallback behavior as separate supervisory choices.",
- "I designed a task-specific authorization record specifying responsible actors, permitted actions, operating conditions, exposed interests, and reassessment triggers, and connected these terms to independent scrutiny and event records that allow consequential decisions to be reconstructed.",
- "I compared delayed spacecraft operations and orbital avoidance against existing automation and ground review, analyzing how latency, uncertain trajectories, simultaneous maneuvers, and failed coordination affect delegation and redistribute risk among operators.",
+ "I traced each element of the procedure to the law of delegation: the authority threshold to the rule against giving a private actor power over its rivals, the evidence threshold to the requirement of reasoned decision on a record, and the timing test to the rule that a retained final decision must be exercised in time.",
+ "I designed a task-specific authorization record specifying responsible actors, permitted actions, operating conditions, exposed interests, and reassessment triggers, and connected these terms to independent scrutiny and event records that allow consequential decisions to be reconstructed. I showed that the record connects supervision before the event to liability after it: departure from its operating range is evidence of fault under the Liability Convention, and operation within it is evidence of care without being a defense.",
+ "I compared delayed spacecraft operations and orbital avoidance against existing automation and ground review, analyzing how latency, uncertain trajectories, simultaneous maneuvers, and failed coordination affect delegation and redistribute risk among operators, and concluded that a licensing condition should ask for interoperable right-of-way rules and maneuver-intent messages before it asks for any common prediction model.",
"I used debris capture and SETI to distinguish technical capability from permission to affect another party's interests or confirm a scientific interpretation, and proposed scientific verification through alternative explanations, independent observations, and documented evidence.",
"I integrated review delays, compliance costs, disclosure risks, and correlated failures into the proposed controls. I developed proportionate reassessment around material changes, reusable evidence, and legally competent institutions, making the burdens of supervision part of the authorization decision."
],
@@ -253,17 +254,18 @@
"readMoreLabel": "Read More",
"detailTitle": "Small States and the Governance of Strategic Space Dependence",
"metadata": {
- "position": "Author (Letao Ouyang)",
+ "position": "Author, Letao (Theodore) Ouyang",
"location": "Durham, North Carolina",
"dates": "December 2024; revised November 2026"
},
"contributionsLabel": "Key Contributions",
"keyContributions": [
+ "I separated control over a space system into five entitlements (permission, title, operation, interruption and priority, and voice), drawing on property theory and the economics of incomplete contracts, and showed from the three states' administrative records that the residual power to decide in a crisis follows the chokepoint rather than the title.",
"I built a structured comparison around specific public functions and the decisions needed to perform them, defining practical autonomy as the capacity to make and implement a specified lawful public decision. I examined external dependencies, legal and operational rights, domestic capability, and documented outcomes to assess whether retained authority could support an actual public decision.",
"I identified distinct mechanisms in the three cases: retained launch-permission authority in New Zealand, operational capability through divided responsibilities in Norway, and the planning and legislative information required alongside asset ownership in Luxembourg.",
- "I proposed an accountable-acceptance standard requiring decision makers to identify who can interrupt essential services, which information and rights permit a response, and what funded capacity enables action. Material gaps would require reasoned acceptance or revision before commitment.",
+ "I proposed an accountable-acceptance standard requiring decision makers to identify who can interrupt essential services, which information and rights permit a response, and what funded capacity enables action. Material gaps would require reasoned acceptance or revision before commitment. Like environmental review under the National Environmental Policy Act, the standard forbids uninformed dependence and leaves the choice to depend with the responsible authority.",
"I translated the standard into proposed licensing and procurement terms covering operator information, significant changes, suspension, recovery, and service transition, and connected these terms to lifetime costs, technical staffing, and timely legislative scrutiny of financial commitments.",
- "I analyzed collective agency through anti-satellite-testing commitments, UN consensus procedures, and ESA programme and voting rules. I distinguished participation rights, influence, and implementation capacity, connecting national safeguards to collective restraint on debris and cross-border service disruption."
+ "I analyzed collective agency through anti-satellite-testing commitments, UN consensus procedures, and ESA program and voting rules. I distinguished participation rights, influence, and implementation capacity, connecting national safeguards to collective restraint on debris and cross-border service disruption."
],
"notes": []
},
@@ -271,20 +273,20 @@
"id": 10,
"categoryId": "03",
"organization": "Who May Choose the Lesser Risk: Solar Geoengineering and the Legal Duties of Comparison and Continuity",
- "summary": "Who may choose among climate risks for people who cannot leave the resulting climate? I proposed two legal duties for solar geoengineering. Comparative justification connects risk-risk analysis to authorization by requiring proponents to defend their policy alternatives and answer challenges from affected constituencies. Continuity addresses termination risk, which I traced to a separation between political control, operator succession, and the physical cooling trajectory, so that authority can change hands without an abrupt physical stop.",
+ "summary": "Who may choose among climate risks for people who cannot leave the resulting climate? I proposed two legal duties for solar geoengineering. Comparative justification connects risk-risk analysis to authorization by requiring proponents to defend their policy alternatives and answer challenges from affected constituencies. Continuity answers termination risk by separating political control, operator succession, and the physical cooling trajectory, so that authority can change hands without an abrupt physical stop.",
"readMoreLabel": "Read More",
"detailTitle": "Who May Choose the Lesser Risk: Solar Geoengineering and the Legal Duties of Comparison and Continuity",
"metadata": {
- "position": "Author (Letao Ouyang)",
+ "position": "Author, Letao (Theodore) Ouyang",
"location": "Durham, North Carolina",
"dates": "December 2024; revised November 2026"
},
"contributionsLabel": "Key Contributions",
"keyContributions": [
"I separated historical climate conditions, a future without stratospheric aerosol injection, and feasible policy portfolios as distinct comparators, and analyzed how baseline selection, welfare measures, and distributional assumptions shape the case for intervention and the burdens assigned to exposed populations.",
- "I proposed a comparative justification duty requiring proponents to defend alternatives, disclose distributional choices, and answer substantiated objections, and designed powers for independently selected representatives of exposed constituencies to demand redesign or withhold collective approval.",
- "I developed a continuity duty separating program control, operator choice, and the cooling trajectory. It specifies bounded interim succession, independent monitoring, transition finance, and gradual withdrawal so that political authority can change without automatically triggering abrupt physical cessation.",
- "I mapped environmental assessment, notification, and consultation requirements onto the authorization problem. I distinguished limited research from climate-affecting deployment and proposed scrutiny based on expected effects and cumulative scale, including review of related projects together.",
+ "I proposed a comparative justification duty requiring proponents to defend alternatives, disclose distributional choices, and answer substantiated objections, and designed powers for independently selected representatives of exposed constituencies to demand redesign or withhold collective approval through separate concurrence, a device the Montreal Protocol's financial mechanism and reorganization law already use.",
+ "I developed a continuity duty separating program control, operator choice, and the cooling trajectory. It specifies bounded interim succession, independent monitoring, transition finance, and gradual withdrawal so that political authority can change without automatically triggering abrupt physical cessation, and grounds each element in law that already governs systems that cannot safely stop: resolution plans, decommissioning assurance, regulated abandonment, and bridge institutions.",
+ "I showed that existing law, from the Vienna Convention and the Framework Convention to the International Law Commission's texts and the International Court of Justice's decisions, supplies assessment, notification, and consultation but no authority able to refuse, and drew the comparative justification duty at that point. I distinguished limited research from climate-affecting deployment and proposed scrutiny based on expected effects and cumulative scale, including review of related projects together.",
"I formulated a priority rule for serious threats to subsistence and survival, requiring comparison of feasible alternatives, shared authorization, and credible relief, and built periodic reassessment of forgone benefits and harms of nondeployment into the same decision process."
],
"notes": []
@@ -293,21 +295,21 @@
"id": 11,
"categoryId": "03",
"organization": "Mangrove Restoration and the Limits of Compensatory Mitigation: Lessons from Florida for the Greater Bay Area",
- "summary": "Restoring a wetland is one decision, and relying on that restoration to permit the loss of another is a second. I compared Florida with the Greater Bay Area on protection, repair, and compensatory mitigation, and used West Lake, Deep Bay, and Hong Kong's 2024 San Tin approval to test what counts as evidence of functional replacement. I proposed site-specific baselines, staged authorization and ecological acceptance, funded correction, and continuing responsibility until the obligation is discharged.",
+ "summary": "Restoring a wetland is one decision, and relying on that restoration to permit the loss of another is a second. I treated a compensation credit as a credit, to be discounted for delay and default, secured, and released against performance, and applied Florida's time-lag and risk discount to the published timetable of Hong Kong's 2024 San Tin approval: the Florida method would credit between roughly a quarter and seven tenths of the enhancement the assessment counted in full. I proposed project conditions that tie a defensible baseline to staged authorization, funded correction, and continuing responsibility.",
"readMoreLabel": "Read More",
"detailTitle": "Mangrove Restoration and the Limits of Compensatory Mitigation: Lessons from Florida for the Greater Bay Area",
"metadata": {
- "position": "Author (Letao Ouyang)",
+ "position": "Author, Letao (Theodore) Ouyang",
"location": "Durham, North Carolina",
"dates": "December 2024; revised November 2026"
},
"contributionsLabel": "Key Contributions",
"keyContributions": [
- "I organized the legal comparison around habitat protection, repair, and compensation for new loss, tracing duties, approval instruments, and administrative powers across Florida, mainland China, Shenzhen, Hong Kong, and Macao to identify how restoration conditions can take legal effect.",
- "I used West Lake to examine hydrological rehabilitation and Deep Bay and San Tin to assess functional replacement, distinguishing canopy growth, construction, and predicted enhancement from evidence supporting compensation, including additionality and the timing of replacement functions.",
- "I proposed staged approval and acceptance separating engineering completion from ecological performance. I specified site baselines, functional objectives, monitoring periods, seasonal coverage, and review authority, linking damaging development phases to the availability of the functions used to justify them.",
- "I designed arrangements for delayed or failed recovery through funded monitoring, correction, and continuing management. I distinguished contractor performance from ecological uncertainty and examined retention, financial security, and public budget commitments as instruments for sustaining corrective work.",
- "I linked ecological evidence to fund release, discharge of responsibility, and development permission, proposing smaller footprints, revised sequencing, alternative sites, or refusal to credit compensation where promised functions lack credible replacement, while allowing proportionate trials in degraded habitats."
+ "I distinguished restoration as repair from restoration as currency and treated a compensation credit as a credit, matching its four properties to their regulatory counterparts: discounts or ratios for temporal loss, ratios for the likelihood of success, financial assurances, and release of credits against demonstrated performance.",
+ "I applied Florida's uniform mitigation assessment method to San Tin's published timetable, on which pond filling may begin in 2026/27 and the conservation park is to be completed by 2039, and showed that no combination of time lag and risk score credits the promised enhancement in full; the credited share runs from 23 to 70 percent.",
+ "I used West Lake to examine hydrological rehabilitation and Deep Bay, where mapped mangrove cover rose from 8 to 15 percent while mudflat fell from 56 to 43 percent, to separate canopy growth, construction, and predicted enhancement from evidence of functional replacement, including additionality.",
+ "I showed that the timing norm is already written in each legal system compared, in federal mitigation rules, Florida's method, Hong Kong's Technical Memorandum and Deep Bay planning guidelines, and mainland policy and the Wetland Protection Law, so that reform gives these principles a decision to govern: the start of each damaging phase.",
+ "I designed a restoration schedule that separates engineering completion from ecological acceptance, funds monitoring and correction in advance, distinguishes contractor performance from ecological uncertainty, and allows smaller footprints, revised sequencing, alternative sites, or refusal to credit compensation where promised functions lack credible replacement."
],
"notes": []
},
@@ -462,6 +464,6 @@
"aiSourceDocument": {
"fileName": "ai-research-and-engineering.md",
"path": "content/sources/ai-portfolio/ai-research-and-engineering.md",
- "sha256": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1"
+ "sha256": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0"
}
}
diff --git a/website/content/legal-papers/autonomous-authority-in-space.json b/website/content/legal-papers/autonomous-authority-in-space.json
index 5b797ab..e73bb2e 100644
--- a/website/content/legal-papers/autonomous-authority-in-space.json
+++ b/website/content/legal-papers/autonomous-authority-in-space.json
@@ -1,7 +1,7 @@
{
"slug": "autonomous-authority-in-space",
"title": "Autonomous Authority in Space: Risk Tradeoffs and the Law of Delegation",
- "author": "Letao Ouyang",
+ "author": "Letao (Theodore) Ouyang",
"date": "December 2024; revised November 2026",
"repositoryUrl": "https://github.com/130U/LAW-Autonomous-Authority-in-Space",
"sourceUrl": "https://github.com/130U/LAW-Autonomous-Authority-in-Space/blob/main/manuscript/autonomous-authority-in-space.md",
@@ -852,10 +852,10 @@
},
{
"type": "paragraph",
- "text": "State practice has since supplied an instance. In December 2021 China informed the Secretary-General that its space station had twice carried out \"preventive collision avoidance control\" after close encounters with Starlink satellites, and that on the second occasion, because the satellite \"was continuously manoeuvring, the manoeuvre strategy was unknown and orbital errors were hard to be assessed.\" The note closed by quoting Article VI. The United States replied that its Space Command \"did not estimate a significant probability of collision\" and that, because the encounters \"did not meet the threshold of established emergency collision criteria, emergency notifications were not warranted.\"28 The two states did not disagree about physics. They disagreed about whose threshold governed and about what one operator is owed concerning another's maneuver policy, which are the questions an authorization record is meant to answer in advance.",
+ "text": "State practice has supplied an instance. In December 2021 China informed the Secretary-General that its space station had twice carried out \"preventive collision avoidance control\" after close encounters with Starlink satellites, and that on the second occasion, because the satellite \"was continuously manoeuvring, the manoeuvre strategy was unknown and orbital errors were hard to be assessed.\" The note closed by quoting Article VI. The United States replied that its Space Command \"did not estimate a significant probability of collision\" and that, because the encounters \"did not meet the threshold of established emergency collision criteria, emergency notifications were not warranted.\"28 The two states did not disagree about physics. They disagreed about whose threshold governed and about what one operator is owed concerning another's maneuver policy, which are the questions an authorization record is meant to answer in advance.",
"runs": [
{
- "text": "State practice has since supplied an instance. In December 2021 China informed the Secretary-General that its space station had twice carried out \"preventive collision avoidance control\" after close encounters with Starlink satellites, and that on the second occasion, because the satellite \"was continuously manoeuvring, the manoeuvre strategy was unknown and orbital errors were hard to be assessed.\" The note closed by quoting Article VI. The United States replied that its Space Command \"did not estimate a significant probability of collision\" and that, because the encounters \"did not meet the threshold of established emergency collision criteria, emergency notifications were not warranted.\""
+ "text": "State practice has supplied an instance. In December 2021 China informed the Secretary-General that its space station had twice carried out \"preventive collision avoidance control\" after close encounters with Starlink satellites, and that on the second occasion, because the satellite \"was continuously manoeuvring, the manoeuvre strategy was unknown and orbital errors were hard to be assessed.\" The note closed by quoting Article VI. The United States replied that its Space Command \"did not estimate a significant probability of collision\" and that, because the encounters \"did not meet the threshold of established emergency collision criteria, emergency notifications were not warranted.\""
},
{
"text": "28",
@@ -1622,15 +1622,6 @@
"text": "Risk-risk analysis requires evaluation of the controls themselves. Human review, certification, information sharing, and restrictions all impose costs and can create hazards. A defensible arrangement explains why its combination of action and restraint is preferable to feasible alternatives, for the people who actually bear the consequences. A permission resting on assumptions that can no longer be defended should be narrowed, transferred to an adequately supported alternative, or withdrawn."
}
]
- },
- {
- "type": "paragraph",
- "text": "---",
- "runs": [
- {
- "text": "---"
- }
- ]
}
],
"footnotes": [
@@ -1643,7 +1634,14 @@
"smallCaps": true
},
{
- "text": " 13-16 & tbls. 1-2 (2022), https://www.c2g2.net/wp-content/uploads/202203-C2G-RR-Full.pdf. See generally "
+ "text": " 13-16 & tbls. 1-2 (2022), https://www.c2g2.net/wp-content/uploads/202203-C2G-RR-Full.pdf. "
+ },
+ {
+ "text": "See generally",
+ "italic": true
+ },
+ {
+ "text": " "
},
{
"text": "Risk vs. Risk: Tradeoffs in Protecting Health and the Environment",
@@ -1656,14 +1654,14 @@
},
{
"number": 2,
- "text": "Felgenhauer ET AL., supra note 1, at 13-16 & tbls. 1-2.",
+ "text": "Felgenhauer et al., supra note 1, at 13-16 & tbls. 1-2.",
"runs": [
{
- "text": "Felgenhauer",
+ "text": "Felgenhauer et al.",
"smallCaps": true
},
{
- "text": " ET AL., "
+ "text": ", "
},
{
"text": "supra",
@@ -1735,10 +1733,10 @@
},
{
"number": 8,
- "text": "Responsibilities and Obligations of States Sponsoring Persons and Entities with Respect to Activities in the Area, Case No. 17, Advisory Opinion of Feb. 1, 2011, ITLOS Rep. 10, ¶¶ 110, 117.",
+ "text": "Responsibilities and Obligations of States Sponsoring Persons and Entities with Respect to Activities in the Area, Case No. 17, Advisory Opinion of Feb. 1, 2011, 2011 ITLOS Rep. 10, ¶¶ 110, 117.",
"runs": [
{
- "text": "Responsibilities and Obligations of States Sponsoring Persons and Entities with Respect to Activities in the Area, Case No. 17, Advisory Opinion of Feb. 1, 2011, ITLOS Rep. 10, ¶¶ 110, 117."
+ "text": "Responsibilities and Obligations of States Sponsoring Persons and Entities with Respect to Activities in the Area, Case No. 17, Advisory Opinion of Feb. 1, 2011, 2011 ITLOS Rep. 10, ¶¶ 110, 117."
}
]
},
@@ -1747,29 +1745,43 @@
"text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities art. 6, in Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001).",
"runs": [
{
- "text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities art. 6, in Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001)."
+ "text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities art. 6, "
+ },
+ {
+ "text": "in",
+ "italic": true
+ },
+ {
+ "text": " Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001)."
}
]
},
{
"number": 10,
- "text": "Int'l Law Comm'n, Draft Articles on Responsibility of States for Internationally Wrongful Acts, with Commentaries, arts. 2, 4, 8, 55, in Report of the International Law Commission on the Work of Its Fifty-Third Session, at 34, 40, 47-48, 140-41, U.N. Doc. A/56/10 (2001). The articles address general responsibility rules and their relationship to special regimes; the application to Article VI is the analysis advanced here.",
+ "text": "Int'l Law Comm'n, Draft Articles on Responsibility of States for Internationally Wrongful Acts, with Commentaries, arts. 2, 4, 8, 55, in Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001), reprinted in [2001] 2 Y.B. Int'l L. Comm'n 31, 34, 40, 47-48, 140-41. The articles address general responsibility rules and their relationship to special regimes; the application to Article VI is the analysis advanced here.",
"runs": [
{
- "text": "Int'l Law Comm'n, Draft Articles on Responsibility of States for Internationally Wrongful Acts, with Commentaries, arts. 2, 4, 8, 55, in "
+ "text": "Int'l Law Comm'n, Draft Articles on Responsibility of States for Internationally Wrongful Acts, with Commentaries, arts. 2, 4, 8, 55, "
},
{
- "text": "Report of the International Law Commission on the Work of Its Fifty-Third Session",
- "smallCaps": true
+ "text": "in",
+ "italic": true
},
{
- "text": ", at 34, 40, 47-48, 140-41, U.N. Doc. A/56/10 (2001). The articles address general responsibility rules and their relationship to special regimes; the application to Article VI is the analysis advanced here."
+ "text": " Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001), "
+ },
+ {
+ "text": "reprinted in",
+ "italic": true
+ },
+ {
+ "text": " [2001] 2 Y.B. Int'l L. Comm'n 31, 34, 40, 47-48, 140-41. The articles address general responsibility rules and their relationship to special regimes; the application to Article VI is the analysis advanced here."
}
]
},
{
"number": 11,
- "text": "Outer Space Treaty, supra note 7, arts. VI-IX.",
+ "text": "Outer Space Treaty, supra note 7, art. IX.",
"runs": [
{
"text": "Outer Space Treaty, "
@@ -1779,7 +1791,7 @@
"italic": true
},
{
- "text": " note 7, arts. VI-IX."
+ "text": " note 7, art. IX."
}
]
},
@@ -1794,33 +1806,27 @@
},
{
"number": 13,
- "text": "Liability Convention, supra note 12, arts. I, III, VII-XI, XIX.",
+ "text": "Id. arts. I, III, VII-XI, XIX.",
"runs": [
{
- "text": "Liability Convention, "
- },
- {
- "text": "supra",
+ "text": "Id.",
"italic": true
},
{
- "text": " note 12, arts. I, III, VII-XI, XIX."
+ "text": " arts. I, III, VII-XI, XIX."
}
]
},
{
"number": 14,
- "text": "Liability Convention, supra note 12, arts. I, III, VII-XI, XIX.",
+ "text": "Id. art. III.",
"runs": [
{
- "text": "Liability Convention, "
- },
- {
- "text": "supra",
+ "text": "Id.",
"italic": true
},
{
- "text": " note 12, arts. I, III, VII-XI, XIX."
+ "text": " art. III."
}
]
},
@@ -1835,10 +1841,10 @@
},
{
"number": 16,
- "text": "Comm. on the Peaceful Uses of Outer Space, Report of the Committee on the Peaceful Uses of Outer Space, annex II, pmbl. ¶¶ 14-16, guidelines A.2-A.3, B.1-B.4, at 52, 55-56, 59-61, U.N. Doc. A/74/20 (2019) [hereinafter LTS Guidelines].",
+ "text": "Comm. on the Peaceful Uses of Outer Space, Rep. of the Comm. on the Peaceful Uses of Outer Space, annex II, pmbl. ¶¶ 14-16, guidelines A.2-A.3, B.1-B.4, at 52, 55-56, 59-61, U.N. Doc. A/74/20 (2019) [hereinafter LTS Guidelines].",
"runs": [
{
- "text": "Comm. on the Peaceful Uses of Outer Space, Report of the Committee on the Peaceful Uses of Outer Space, annex II, pmbl. ¶¶ 14-16, guidelines A.2-A.3, B.1-B.4, at 52, 55-56, 59-61, U.N. Doc. A/74/20 (2019) [hereinafter LTS Guidelines]."
+ "text": "Comm. on the Peaceful Uses of Outer Space, Rep. of the Comm. on the Peaceful Uses of Outer Space, annex II, pmbl. ¶¶ 14-16, guidelines A.2-A.3, B.1-B.4, at 52, 55-56, 59-61, U.N. Doc. A/74/20 (2019) [hereinafter LTS Guidelines]."
}
]
},
@@ -1853,10 +1859,10 @@
},
{
"number": 18,
- "text": "Mitigation of Orbital Debris in the New Space Age, 85 Fed. Reg. 52,422, 52,424-25, 52,447 (Aug. 25, 2020) (identifying the agency's statutory basis and listing 47 U.S.C. §§ 151, 154(i), 301, 303, 307-310).",
+ "text": "Mitigation of Orbital Debris in the New Space Age, 85 Fed. Reg. at 52,424-25, 52,447 (identifying the agency's statutory basis and listing 47 U.S.C. §§ 151, 154(i), 301, 303, 307-310).",
"runs": [
{
- "text": "Mitigation of Orbital Debris in the New Space Age, 85 Fed. Reg. 52,422, 52,424-25, 52,447 (Aug. 25, 2020) (identifying the agency's statutory basis and listing 47 U.S.C. §§ 151, 154(i), 301, 303, 307-310)."
+ "text": "Mitigation of Orbital Debris in the New Space Age, 85 Fed. Reg. at 52,424-25, 52,447 (identifying the agency's statutory basis and listing 47 U.S.C. §§ 151, 154(i), 301, 303, 307-310)."
}
]
},
@@ -1890,11 +1896,7 @@
"text": "Nat'l Aeronautics & Space Admin., Jet Propulsion Lab., Remote Agent Experiment Meets All Objectives (June 7, 1999), https://www.jpl.nasa.gov/news/remote-agent-experiment-meets-all-objectives/.",
"runs": [
{
- "text": "Nat'l Aeronautics & Space Admin., Jet Propulsion Lab.",
- "smallCaps": true
- },
- {
- "text": ", "
+ "text": "Nat'l Aeronautics & Space Admin., Jet Propulsion Lab., "
},
{
"text": "Remote Agent Experiment Meets All Objectives",
@@ -1930,7 +1932,7 @@
},
{
"number": 23,
- "text": "Raymond Francis et al., AEGIS Autonomous Targeting for the SuperCam Instrument on the Mars 2020 Perseverance Rover, at 3-5 (i-SAIRAS 2024), https://ai.jpl.nasa.gov/public/documents/papers/francis-isairas-2024.pdf.",
+ "text": "Raymond Francis et al., AEGIS Autonomous Targeting for the SuperCam Instrument on the Mars 2020 Perseverance Rover, at 3-5 (i-SAIRAS 2024), https://superintelligence.jpl.nasa.gov/public/documents/papers/francis-isairas-2024.pdf.",
"runs": [
{
"text": "Raymond Francis et al., "
@@ -1940,7 +1942,7 @@
"italic": true
},
{
- "text": ", at 3-5 (i-SAIRAS 2024), https://ai.jpl.nasa.gov/public/documents/papers/francis-isairas-2024.pdf."
+ "text": ", at 3-5 (i-SAIRAS 2024), https://superintelligence.jpl.nasa.gov/public/documents/papers/francis-isairas-2024.pdf."
}
]
},
@@ -1984,17 +1986,14 @@
},
{
"number": 27,
- "text": "Uriot et al., supra note 26, §§ 2, 4.",
+ "text": "Id. §§ 2, 4.",
"runs": [
{
- "text": "Uriot et al., "
- },
- {
- "text": "supra",
+ "text": "Id.",
"italic": true
},
{
- "text": " note 26, §§ 2, 4."
+ "text": " §§ 2, 4."
}
]
},
@@ -2203,14 +2202,14 @@
},
{
"number": 40,
- "text": "Restatement (Third) of Torts: Liability for Physical and Emotional Harm §§ 14, 16 (Am. L. Inst. 2010).",
+ "text": "Restatement (Third) of Torts: Liability for Physical and Emotional Harm §§ 13(b), 16(a) (Am. L. Inst. 2010).",
"runs": [
{
"text": "Restatement (Third) of Torts: Liability for Physical and Emotional Harm",
"smallCaps": true
},
{
- "text": " §§ 14, 16 ("
+ "text": " §§ 13(b), 16(a) ("
},
{
"text": "Am. L. Inst.",
@@ -2271,10 +2270,10 @@
},
{
"number": 44,
- "text": "National Aeronautics & Space Administration, Overview of the DART Mishap Investigation Results 6 (2006), https://www.nasa.gov/pdf/148072main_DART_mishap_overview.pdf.",
+ "text": "Nat'l Aeronautics & Space Admin., Overview of the DART Mishap Investigation Results 6 (2006), https://www.nasa.gov/pdf/148072main_DART_mishap_overview.pdf.",
"runs": [
{
- "text": "National Aeronautics & Space Administration, Overview of the DART Mishap Investigation Results",
+ "text": "Nat'l Aeronautics & Space Admin., Overview of the DART Mishap Investigation Results",
"smallCaps": true
},
{
diff --git a/website/content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json b/website/content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json
index 629855d..60c95eb 100644
--- a/website/content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json
+++ b/website/content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json
@@ -1,7 +1,7 @@
{
"slug": "mangrove-restoration-and-the-limits-of-compensatory-mitigation",
"title": "Mangrove Restoration and the Limits of Compensatory Mitigation: Lessons from Florida for the Greater Bay Area",
- "author": "Letao Ouyang",
+ "author": "Letao (Theodore) Ouyang",
"date": "December 2024; revised November 2026",
"repositoryUrl": "https://github.com/130U/LAW-Mangrove-Restoration-and-the-Limits-of-Compensatory-Mitigation",
"sourceUrl": "https://github.com/130U/LAW-Mangrove-Restoration-and-the-Limits-of-Compensatory-Mitigation/blob/main/manuscript/mangrove-restoration-and-the-limits-of-compensatory-mitigation.md",
@@ -305,10 +305,10 @@
},
{
"type": "paragraph",
- "text": "China's Wetland Protection Law, adopted December 24, 2021, took effect June 1, 2022. Article 34 requires local governments to prepare special mangrove protection plans and provides for mangrove wetlands to be included among important wetlands. It prohibits occupation subject to exceptions for national major projects, disaster prevention and mitigation, and related needs following provincial-or-higher assessment and lawful approval procedures. It also addresses construction that changes the hydrological regime of mangrove estuaries and substantially affects growth: effective measures must reduce the adverse effects.8 These provisions make both the forest and its water regime objects of statutory protection.",
+ "text": "China's Wetland Protection Law, adopted December 24, 2021, took effect June 1, 2022. Article 34 provides for mangrove wetlands to be included among important wetlands; as enacted, it also required local governments to prepare special mangrove protection plans, a duty the Standing Committee of the National People's Congress deleted on November 6, 2026. It prohibits occupation subject to exceptions for national major projects, disaster prevention and mitigation, and related needs following provincial-or-higher assessment and lawful approval procedures. It also addresses construction that changes the hydrological regime of mangrove estuaries and substantially affects growth: effective measures must reduce the adverse effects.8 These provisions make both the forest and its water regime objects of statutory protection.",
"runs": [
{
- "text": "China's Wetland Protection Law, adopted December 24, 2021, took effect June 1, 2022. Article 34 requires local governments to prepare special mangrove protection plans and provides for mangrove wetlands to be included among important wetlands. It prohibits occupation subject to exceptions for national major projects, disaster prevention and mitigation, and related needs following provincial-or-higher assessment and lawful approval procedures. It also addresses construction that changes the hydrological regime of mangrove estuaries and substantially affects growth: effective measures must reduce the adverse effects."
+ "text": "China's Wetland Protection Law, adopted December 24, 2021, took effect June 1, 2022. Article 34 provides for mangrove wetlands to be included among important wetlands; as enacted, it also required local governments to prepare special mangrove protection plans, a duty the Standing Committee of the National People's Congress deleted on November 6, 2026. It prohibits occupation subject to exceptions for national major projects, disaster prevention and mitigation, and related needs following provincial-or-higher assessment and lawful approval procedures. It also addresses construction that changes the hydrological regime of mangrove estuaries and substantially affects growth: effective measures must reduce the adverse effects."
},
{
"text": "8",
@@ -537,10 +537,10 @@
},
{
"type": "paragraph",
- "text": "Macao's Environmental Framework Law, Law 2/91/M, supplies a different starting point. A 2013 Legislative Assembly committee opinion describes article 28(1) as requiring environmental-impact studies for planning capable of affecting the environment, public health, and quality of life.20 The provision does not establish the designated-project permit machinery used in Hong Kong. Any proposed restoration condition must accordingly be traced to the applicable Macao authorization, land arrangement, or public-works contract.",
+ "text": "Macao's Environmental Framework Law, Law 2/91/M, supplies a different starting point. A 2013 Legislative Assembly committee opinion describes Article 28(1) as requiring environmental-impact studies for planning capable of affecting the environment, public health, and quality of life.20 The provision does not establish the designated-project permit machinery used in Hong Kong. Any proposed restoration condition must accordingly be traced to the applicable Macao authorization, land arrangement, or public-works contract.",
"runs": [
{
- "text": "Macao's Environmental Framework Law, Law 2/91/M, supplies a different starting point. A 2013 Legislative Assembly committee opinion describes article 28(1) as requiring environmental-impact studies for planning capable of affecting the environment, public health, and quality of life."
+ "text": "Macao's Environmental Framework Law, Law 2/91/M, supplies a different starting point. A 2013 Legislative Assembly committee opinion describes Article 28(1) as requiring environmental-impact studies for planning capable of affecting the environment, public health, and quality of life."
},
{
"text": "20",
@@ -629,10 +629,10 @@
},
{
"type": "paragraph",
- "text": "Where section 404 applies, the guidelines impose an alternatives requirement. The guidelines restrict discharges when a practicable, less damaging alternative exists, taking cost, technology, and logistics into account, and bar significant degradation. Compensation comes last in that sequence, after avoidance and minimization.24",
+ "text": "Where section 404 applies, the section 404(b)(1) Guidelines impose an alternatives requirement. The guidelines restrict discharges when a practicable, less damaging alternative exists, taking cost, technology, and logistics into account, and bar significant degradation. Compensation comes last in that sequence, after avoidance and minimization.24",
"runs": [
{
- "text": "Where section 404 applies, the guidelines impose an alternatives requirement. The guidelines restrict discharges when a practicable, less damaging alternative exists, taking cost, technology, and logistics into account, and bar significant degradation. Compensation comes last in that sequence, after avoidance and minimization."
+ "text": "Where section 404 applies, the section 404(b)(1) Guidelines impose an alternatives requirement. The guidelines restrict discharges when a practicable, less damaging alternative exists, taking cost, technology, and logistics into account, and bar significant degradation. Compensation comes last in that sequence, after avoidance and minimization."
},
{
"text": "24",
@@ -643,10 +643,10 @@
},
{
"type": "paragraph",
- "text": "The federal rule also states a timing norm. Compensation is to be implemented, \"to the maximum extent practicable, in advance of or concurrent with the activity causing the authorized impacts,\" and the district engineer must require additional compensation \"to offset temporal losses of aquatic functions.\" Ratios above one-to-one are required where necessary to account for, among other things, the likelihood of success and temporal losses, and mitigation-bank credits are released against performance-based milestones.25",
+ "text": "The federal rule also states a timing norm. Compensation is to be implemented, \"to the maximum extent practicable, in advance of or concurrent with the activity causing the authorized impacts,\" and the district engineer is to require, \"to the extent appropriate and practicable,\" additional compensation \"to offset temporal losses of aquatic functions.\" Ratios above one-to-one are required where necessary to account for, among other things, the likelihood of success and temporal losses, and mitigation-bank credits are released against performance-based milestones.25",
"runs": [
{
- "text": "The federal rule also states a timing norm. Compensation is to be implemented, \"to the maximum extent practicable, in advance of or concurrent with the activity causing the authorized impacts,\" and the district engineer must require additional compensation \"to offset temporal losses of aquatic functions.\" Ratios above one-to-one are required where necessary to account for, among other things, the likelihood of success and temporal losses, and mitigation-bank credits are released against performance-based milestones."
+ "text": "The federal rule also states a timing norm. Compensation is to be implemented, \"to the maximum extent practicable, in advance of or concurrent with the activity causing the authorized impacts,\" and the district engineer is to require, \"to the extent appropriate and practicable,\" additional compensation \"to offset temporal losses of aquatic functions.\" Ratios above one-to-one are required where necessary to account for, among other things, the likelihood of success and temporal losses, and mitigation-bank credits are released against performance-based milestones."
},
{
"text": "25",
@@ -740,10 +740,10 @@
},
{
"type": "paragraph",
- "text": "Coordination introduces further legal costs. In 2024, a federal district court vacated EPA's approval of Florida's assumption of section 404 authority after finding defects in the associated Endangered Species Act arrangements, which had left the analysis of effects on particular species to a later, permit-level process. An appeal remained pending in December 2024.31 The decision concerned the adequacy of those arrangements; state assumption remained authorized by statute, and some waters remained federally administered. Consolidating administration could not excuse the species-specific protection required by the governing law.",
+ "text": "Coordination introduces further legal costs. In 2024, a federal district court vacated EPA's approval of Florida's assumption of section 404 authority after finding defects in the associated Endangered Species Act arrangements, which had left the analysis of effects on particular species to a later, permit-level process. In March 2026 a divided panel of the D.C. Circuit affirmed, and the federal defendants and Florida sought rehearing en banc that August.31 The decisions concerned the adequacy of those arrangements; state assumption remained authorized by statute, and some waters remained federally administered. Consolidating administration could not excuse the species-specific protection required by the governing law.",
"runs": [
{
- "text": "Coordination introduces further legal costs. In 2024, a federal district court vacated EPA's approval of Florida's assumption of section 404 authority after finding defects in the associated Endangered Species Act arrangements, which had left the analysis of effects on particular species to a later, permit-level process. An appeal remained pending in December 2024."
+ "text": "Coordination introduces further legal costs. In 2024, a federal district court vacated EPA's approval of Florida's assumption of section 404 authority after finding defects in the associated Endangered Species Act arrangements, which had left the analysis of effects on particular species to a later, permit-level process. In March 2026 a divided panel of the D.C. Circuit affirmed, and the federal defendants and Florida sought rehearing en banc that August."
},
{
"text": "31",
@@ -751,7 +751,7 @@
"footnoteRef": "31"
},
{
- "text": " The decision concerned the adequacy of those arrangements; state assumption remained authorized by statute, and some waters remained federally administered. Consolidating administration could not excuse the species-specific protection required by the governing law."
+ "text": " The decisions concerned the adequacy of those arrangements; state assumption remained authorized by statute, and some waters remained federally administered. Consolidating administration could not excuse the species-specific protection required by the governing law."
}
]
},
@@ -1013,10 +1013,10 @@
},
{
"type": "paragraph",
- "text": "The Director's May 17, 2024 approval imposed specific management duties. Annex 1, condition 1 required submission of a detailed Habitat Creation and Management Plan for approval at least nine months before pond filling. It specified target habitats and species, implementation methods, monitoring, reporting, action and limit levels, and an event/action plan. Other conditions required an interim enhancement plan, an environmental committee, and periodic reporting.44 These conditions address several concerns raised during consultation. Their sufficiency turns on what they require at the point when habitat loss proceeds.",
+ "text": "The Director's May 17, 2024, approval imposed specific management duties. Annex 1, condition 1 required submission of a detailed Habitat Creation and Management Plan for approval at least nine months before pond filling. It specified target habitats and species, implementation methods, monitoring, reporting, action and limit levels, and an event/action plan. Other conditions required an interim enhancement plan, an environmental committee, and periodic reporting.44 These conditions address several concerns raised during consultation. Their sufficiency turns on what they require at the point when habitat loss proceeds.",
"runs": [
{
- "text": "The Director's May 17, 2024 approval imposed specific management duties. Annex 1, condition 1 required submission of a detailed Habitat Creation and Management Plan for approval at least nine months before pond filling. It specified target habitats and species, implementation methods, monitoring, reporting, action and limit levels, and an event/action plan. Other conditions required an interim enhancement plan, an environmental committee, and periodic reporting."
+ "text": "The Director's May 17, 2024, approval imposed specific management duties. Annex 1, condition 1 required submission of a detailed Habitat Creation and Management Plan for approval at least nine months before pond filling. It specified target habitats and species, implementation methods, monitoring, reporting, action and limit levels, and an event/action plan. Other conditions required an interim enhancement plan, an environmental committee, and periodic reporting."
},
{
"text": "44",
@@ -1030,10 +1030,10 @@
},
{
"type": "paragraph",
- "text": "Condition 9 linked pond filling to the commencement of construction of ecologically enhanced fishponds. That milestone demonstrates that construction has begun; it supplies no independent evidence that replacement functions are operating.45 The management and interim-enhancement conditions may reduce the interval of loss. An adequate evaluation must consider their combined operation, including the expected recovery trajectory and the measures available if enhancement falls behind.",
+ "text": "Condition 9 linked pond filling to the commencement of construction of ecologically enhanced fishponds. That milestone demonstrates that construction has begun; it supplies no independent evidence that replacement functions are operating. The detailed plan prepared under condition 1, dated August 2026, keeps that milestone and targets the park's first phase for completion by 2031 and the whole park on or before 2039.45 The management and interim-enhancement conditions may reduce the interval of loss. An adequate evaluation must consider their combined operation, including the expected recovery trajectory and the measures available if enhancement falls behind.",
"runs": [
{
- "text": "Condition 9 linked pond filling to the commencement of construction of ecologically enhanced fishponds. That milestone demonstrates that construction has begun; it supplies no independent evidence that replacement functions are operating."
+ "text": "Condition 9 linked pond filling to the commencement of construction of ecologically enhanced fishponds. That milestone demonstrates that construction has begun; it supplies no independent evidence that replacement functions are operating. The detailed plan prepared under condition 1, dated August 2026, keeps that milestone and targets the park's first phase for completion by 2031 and the whole park on or before 2039."
},
{
"text": "45",
@@ -1064,10 +1064,10 @@
},
{
"type": "paragraph",
- "text": "Florida's method shows what such a timetable is worth. Suppose, for illustration only, that the enhancement were assessed under the Florida rule. A function lost when filling begins and replaced when the park is complete would carry a time lag of about twelve years, and a shorter lag for ponds filled later or enhancement delivered earlier. The share of a promised gain that the rule credits is the reciprocal of the time-lag factor multiplied by the risk score.47",
+ "text": "Florida's method shows what such a timetable is worth. Suppose, for illustration only, that the enhancement were assessed under the Florida rule. A function lost when filling begins and replaced when the park is complete would carry a time lag of about twelve years, and a shorter lag for ponds filled later or enhancement delivered earlier. The share of a promised gain that the rule credits is the reciprocal of the product of the time-lag factor and the risk score.47",
"runs": [
{
- "text": "Florida's method shows what such a timetable is worth. Suppose, for illustration only, that the enhancement were assessed under the Florida rule. A function lost when filling begins and replaced when the park is complete would carry a time lag of about twelve years, and a shorter lag for ponds filled later or enhancement delivered earlier. The share of a promised gain that the rule credits is the reciprocal of the time-lag factor multiplied by the risk score."
+ "text": "Florida's method shows what such a timetable is worth. Suppose, for illustration only, that the enhancement were assessed under the Florida rule. A function lost when filling begins and replaced when the park is complete would carry a time lag of about twelve years, and a shorter lag for ponds filled later or enhancement delivered earlier. The share of a promised gain that the rule credits is the reciprocal of the product of the time-lag factor and the risk score."
},
{
"text": "47",
@@ -1262,10 +1262,10 @@
},
{
"type": "paragraph",
- "text": "On no row is the promise credited in full. Full credit under the Florida rule requires replacement that offsets the impact at or before the time of loss and a record of success, which is the position the Society urged and the sequence the Technical Memorandum prefers. The San Tin assessment counted its predicted enhancement at face value, on an assumption drawn from a scheme that was never built. The table does not show that the assessment was unlawful, and Florida's rule has no force in Hong Kong. It shows the size of the judgment that condition 9 leaves unexamined: whether the start of construction can stand in for function at a discount of zero.",
+ "text": "On no row is the promise credited in full. Full credit under the Florida rule requires replacement that offsets the impact at or before the time of loss and a record of success, which is the position the Society urged and the sequence the Technical Memorandum prefers. The San Tin assessment counted its predicted enhancement at face value, on an assumption drawn from a scheme that was never built. Table 1 does not show that the assessment was unlawful, and Florida's rule has no force in Hong Kong. It shows the size of the judgment that condition 9 leaves unexamined: whether the start of construction can stand in for function at a discount of zero.",
"runs": [
{
- "text": "On no row is the promise credited in full. Full credit under the Florida rule requires replacement that offsets the impact at or before the time of loss and a record of success, which is the position the Society urged and the sequence the Technical Memorandum prefers. The San Tin assessment counted its predicted enhancement at face value, on an assumption drawn from a scheme that was never built. The table does not show that the assessment was unlawful, and Florida's rule has no force in Hong Kong. It shows the size of the judgment that condition 9 leaves unexamined: whether the start of construction can stand in for function at a discount of zero."
+ "text": "On no row is the promise credited in full. Full credit under the Florida rule requires replacement that offsets the impact at or before the time of loss and a record of success, which is the position the Society urged and the sequence the Technical Memorandum prefers. The San Tin assessment counted its predicted enhancement at face value, on an assumption drawn from a scheme that was never built. Table 1 does not show that the assessment was unlawful, and Florida's rule has no force in Hong Kong. It shows the size of the judgment that condition 9 leaves unexamined: whether the start of construction can stand in for function at a discount of zero."
}
]
},
@@ -1685,15 +1685,6 @@
"text": "Environmental law can assign responsibility for uncertainty even when ecological recovery cannot be guaranteed. A defensible baseline, staged decisions, funded correction, and evidence capable of changing legal consequences make that responsibility concrete. The authority permitting habitat loss should then be able to explain what function it expects to replace, when the replacement must become available, and who remains answerable if the expectation proves wrong."
}
]
- },
- {
- "type": "paragraph",
- "text": "---",
- "runs": [
- {
- "text": "---"
- }
- ]
}
],
"footnotes": [
@@ -1712,10 +1703,10 @@
},
{
"number": 2,
- "text": "Wetland Protection Law of the People's Republic of China art. 34 (promulgated by the Standing Comm. Nat'l People's Cong., Dec. 24, 2021, effective June 1, 2022); Roy R. Lewis III, Ecological Engineering for Successful Management and Restoration of Mangrove Forests, 24 Ecological Engineering 403, 413 & fig. 6 (2005).",
+ "text": "Wetland Protection Law of the People's Republic of China art. 34 (promulgated by the Standing Comm. Nat'l People's Cong., Dec. 24, 2021, effective June 1, 2022, amended Nov. 6, 2026); Roy R. Lewis III, Ecological Engineering for Successful Management and Restoration of Mangrove Forests, 24 Ecological Engineering 403, 413 & fig. 6 (2005).",
"runs": [
{
- "text": "Wetland Protection Law of the People's Republic of China art. 34 (promulgated by the Standing Comm. Nat'l People's Cong., Dec. 24, 2021, effective June 1, 2022); Roy R. Lewis III, "
+ "text": "Wetland Protection Law of the People's Republic of China art. 34 (promulgated by the Standing Comm. Nat'l People's Cong., Dec. 24, 2021, effective June 1, 2022, amended Nov. 6, 2026); Roy R. Lewis III, "
},
{
"text": "Ecological Engineering for Successful Management and Restoration of Mangrove Forests",
@@ -1799,7 +1790,7 @@
},
{
"number": 5,
- "text": "Mark Van Hoecke, Methodology of Comparative Legal Research, Law and Method, 2015, at 1, 6-10.",
+ "text": "Mark Van Hoecke, Methodology of Comparative Legal Research, Law and Method, Dec. 2015, at 1, 6-10.",
"runs": [
{
"text": "Mark Van Hoecke, "
@@ -1816,7 +1807,7 @@
"smallCaps": true
},
{
- "text": ", 2015, at 1, 6-10."
+ "text": ", Dec. 2015, at 1, 6-10."
}
]
},
@@ -1844,10 +1835,17 @@
},
{
"number": 8,
- "text": "Wetland Protection Law of the People's Republic of China arts. 34, 65 (2021). Article 34's exception requires the relevant provincial-or-higher assessment and procedures prescribed by law; substantial effects on mangrove growth trigger its hydrological mitigation requirement.",
+ "text": "Wetland Protection Law of the People's Republic of China arts. 34, 65 (2021), amended by Decision of the Standing Committee of the National People's Congress on Amending Fifteen Laws Including the Emergency Response Law item 9(3) (promulgated by the Standing Comm. Nat'l People's Cong., Nov. 6, 2026, effective Nov. 6, 2026) (deleting article 34, paragraph 1, the special-plan duty). Article 34's exception requires the relevant provincial-or-higher assessment and procedures prescribed by law; substantial effects on mangrove growth trigger its hydrological mitigation requirement.",
"runs": [
{
- "text": "Wetland Protection Law of the People's Republic of China arts. 34, 65 (2021). Article 34's exception requires the relevant provincial-or-higher assessment and procedures prescribed by law; substantial effects on mangrove growth trigger its hydrological mitigation requirement."
+ "text": "Wetland Protection Law of the People's Republic of China arts. 34, 65 (2021), "
+ },
+ {
+ "text": "amended by",
+ "italic": true
+ },
+ {
+ "text": " Decision of the Standing Committee of the National People's Congress on Amending Fifteen Laws Including the Emergency Response Law item 9(3) (promulgated by the Standing Comm. Nat'l People's Cong., Nov. 6, 2026, effective Nov. 6, 2026) (deleting article 34, paragraph 1, the special-plan duty). Article 34's exception requires the relevant provincial-or-higher assessment and procedures prescribed by law; substantial effects on mangrove growth trigger its hydrological mitigation requirement."
}
]
},
@@ -1897,14 +1895,10 @@
},
{
"number": 13,
- "text": "Ministry of Natural Resources & National Forestry & Grassland Administration, Special Action Plan for Mangrove Protection and Restoration (2020-2025) pt. I(3), actions 4, 6 (Aug. 14, 2020).",
+ "text": "Ministry of Natural Resources & National Forestry and Grassland Administration, Special Action Plan for Mangrove Protection and Restoration (2020-2025), Zi Ran Zi Fa [2020] No. 135, pt. I(3), actions 4, 6 (Aug. 14, 2020), https://www.gov.cn/zhengce/zhengceku/2020-08/29/content_5538354.htm.",
"runs": [
{
- "text": "Ministry of Natural Resources & National Forestry & Grassland Administration, Special Action Plan for Mangrove Protection and Restoration",
- "smallCaps": true
- },
- {
- "text": " (2020-2025) pt. I(3), actions 4, 6 (Aug. 14, 2020)."
+ "text": "Ministry of Natural Resources & National Forestry and Grassland Administration, Special Action Plan for Mangrove Protection and Restoration (2020-2025), Zi Ran Zi Fa [2020] No. 135, pt. I(3), actions 4, 6 (Aug. 14, 2020), https://www.gov.cn/zhengce/zhengceku/2020-08/29/content_5538354.htm."
}
]
},
@@ -2098,17 +2092,24 @@
},
{
"number": 23,
- "text": "Sackett v. EPA, No. 21-454, slip op. at 21-22, 27-28 (U.S. May 25, 2023); id., slip op. at 8-14 (Kavanaugh, J., concurring in the judgment); Fla. Stat. §§ 373.414(1), (8), 403.9328(2), (5) (2024).",
+ "text": "Sackett v. EPA, 598 U.S. 651, 678-79, 683-84 (2023); id. at 721-28 (Kavanaugh, J., concurring in the judgment); Fla. Stat. §§ 373.414(1), (8), 403.9328(2), (5) (2024).",
"runs": [
{
- "text": "Sackett v. EPA, No. 21-454, slip op. at 21-22, 27-28 (U.S. May 25, 2023); "
+ "text": "Sackett v. EPA, 598 U.S. 651, 678-79, 683-84 (2023); "
},
{
"text": "id.",
"italic": true
},
{
- "text": ", slip op. at 8-14 (Kavanaugh, J., concurring in the judgment); Fla. Stat. §§ 373.414(1), (8), 403.9328(2), (5) (2024)."
+ "text": " at 721-28 (Kavanaugh, J., concurring in the judgment); "
+ },
+ {
+ "text": "Fla. Stat.",
+ "smallCaps": true
+ },
+ {
+ "text": " §§ 373.414(1), (8), 403.9328(2), (5) (2024)."
}
]
},
@@ -2144,16 +2145,24 @@
"text": "Fla. Stat. §§ 403.9324(2), (5)-(6), 403.9325(7), 403.9326(1)(a)-(b), 403.9327(5), (7) (2024). The fringe definition limits depth to fifty feet and excludes specified islands and conservation or mitigation lands; separate professional-trimming provisions allow greater heights subject to additional conditions.",
"runs": [
{
- "text": "Fla. Stat. §§ 403.9324(2), (5)-(6), 403.9325(7), 403.9326(1)(a)-(b), 403.9327(5), (7) (2024). The fringe definition limits depth to fifty feet and excludes specified islands and conservation or mitigation lands; separate professional-trimming provisions allow greater heights subject to additional conditions."
+ "text": "Fla. Stat.",
+ "smallCaps": true
+ },
+ {
+ "text": " §§ 403.9324(2), (5)-(6), 403.9325(7), 403.9326(1)(a)-(b), 403.9327(5), (7) (2024). The fringe definition limits depth to fifty feet and excludes specified islands and conservation or mitigation lands; separate professional-trimming provisions allow greater heights subject to additional conditions."
}
]
},
{
"number": 28,
- "text": "Fla. Stat. § 403.9332(1)(a)-(d) (2024). The provision specifies a five-percent-or-more damage threshold, equivalent canopy within five years, joint and several restoration responsibility, and eighty-percent survival one year after planting, with further planting where necessary.",
+ "text": "Fla. Stat. § 403.9332(1)(a)-(d) (2024). The provision specifies a damage threshold of 5 percent or more, equivalent canopy within five years, joint and several restoration responsibility, and 80-percent survival one year after planting, with further planting where necessary.",
"runs": [
{
- "text": "Fla. Stat. § 403.9332(1)(a)-(d) (2024). The provision specifies a five-percent-or-more damage threshold, equivalent canopy within five years, joint and several restoration responsibility, and eighty-percent survival one year after planting, with further planting where necessary."
+ "text": "Fla. Stat.",
+ "smallCaps": true
+ },
+ {
+ "text": " § 403.9332(1)(a)-(d) (2024). The provision specifies a damage threshold of 5 percent or more, equivalent canopy within five years, joint and several restoration responsibility, and 80-percent survival one year after planting, with further planting where necessary."
}
]
},
@@ -2162,7 +2171,11 @@
"text": "Fla. Stat. § 373.414(18) (2024).",
"runs": [
{
- "text": "Fla. Stat. § 373.414(18) (2024)."
+ "text": "Fla. Stat.",
+ "smallCaps": true
+ },
+ {
+ "text": " § 373.414(18) (2024)."
}
]
},
@@ -2171,16 +2184,48 @@
"text": "Fla. Admin. Code Ann. r. 62-345.600(1)-(3) (2007). The rule states relative functional gain as the mitigation delta divided by the product of the risk score and the time-lag factor, scores risk in quarter-point increments, and tabulates the time-lag factor by years.",
"runs": [
{
- "text": "Fla. Admin. Code Ann. r. 62-345.600(1)-(3) (2007). The rule states relative functional gain as the mitigation delta divided by the product of the risk score and the time-lag factor, scores risk in quarter-point increments, and tabulates the time-lag factor by years."
+ "text": "Fla. Admin. Code Ann.",
+ "smallCaps": true
+ },
+ {
+ "text": " r. 62-345.600(1)-(3) (2007). The rule states relative functional gain as the mitigation delta divided by the product of the risk score and the time-lag factor, scores risk in quarter-point increments, and tabulates the time-lag factor by years."
}
]
},
{
"number": 31,
- "text": "Ctr. for Biological Diversity v. Regan, No. 21-119, slip op. at 6-7, 90-97 (D.D.C. Feb. 15, 2024); Ctr. for Biological Diversity v. Regan, No. 21-119, slip op. at 2-4 (D.D.C. Apr. 12, 2024); Clean Water Act Section 404 Tribal and State Assumption Program, 89 Fed. Reg. 103,454, 103,454 n.1 (Dec. 18, 2024) (noting the pending appeal); 33 U.S.C. § 1344(g)(1) (2023) (state assumption and retained waters).",
+ "text": "Ctr. for Biological Diversity v. Regan, 734 F. Supp. 3d 1, 49, 58, 63 (D.D.C. 2024), aff'd sub nom. Ctr. for Biological Diversity v. Zeldin, 171 F.4th 356, 378-81 (D.C. Cir. 2026); Cassandra J. Barnum, Cong. Rsch. Serv., LSB11470, ESA Section 7 Consultation and Cooperative Federalism 4 (2026) (reporting the petitions for rehearing en banc); Clean Water Act Section 404 Tribal and State Assumption Program, 89 Fed. Reg. 103,454, 103,454 n.1 (Dec. 18, 2024); 33 U.S.C. § 1344(g)(1) (2023) (state assumption and retained waters).",
"runs": [
{
- "text": "Ctr. for Biological Diversity v. Regan, No. 21-119, slip op. at 6-7, 90-97 (D.D.C. Feb. 15, 2024); Ctr. for Biological Diversity v. Regan, No. 21-119, slip op. at 2-4 (D.D.C. Apr. 12, 2024); Clean Water Act Section 404 Tribal and State Assumption Program, 89 Fed. Reg. 103,454, 103,454 n.1 (Dec. 18, 2024) (noting the pending appeal); 33 U.S.C. § 1344(g)(1) (2023) (state assumption and retained waters)."
+ "text": "Ctr. for Biological Diversity v. Regan, 734 F. Supp. 3d 1, 49, 58, 63 (D.D.C. 2024), "
+ },
+ {
+ "text": "aff'd sub nom.",
+ "italic": true
+ },
+ {
+ "text": " Ctr. for Biological Diversity v. Zeldin, 171 F.4th 356, 378-81 (D.C. Cir. 2026); "
+ },
+ {
+ "text": "Cassandra J. Barnum, Cong. Rsch. Serv.",
+ "smallCaps": true
+ },
+ {
+ "text": ", LSB11470, "
+ },
+ {
+ "text": "ESA Section",
+ "smallCaps": true
+ },
+ {
+ "text": " 7 "
+ },
+ {
+ "text": "Consultation and Cooperative Federalism",
+ "smallCaps": true
+ },
+ {
+ "text": " 4 (2026) (reporting the petitions for rehearing en banc); Clean Water Act Section 404 Tribal and State Assumption Program, 89 Fed. Reg. 103,454, 103,454 n.1 (Dec. 18, 2024); 33 U.S.C. § 1344(g)(1) (2023) (state assumption and retained waters)."
}
]
},
@@ -2432,14 +2477,28 @@
},
{
"number": 45,
- "text": "Id. annex 1, condition 9, at 4. The construction milestone concerns ecologically enhanced fishponds at the proposed Sam Po Shue Wetland Conservation Park.",
+ "text": "Id. annex 1, condition 9, at 4; Civil Engineering and Development Department, Detailed Habitat Creation and Management Plan (Ref. Z22-03): San Tin/Lok Ma Chau Development Node, Package 1 §§ 1.1-1.2, tbl. 8.1, at 1-2, 68 (Aug. 2026), https://www.epd.gov.hk/eia/files/applications/en/pp_340/eia_1994/progress/action_182408/Detailed%20Habitat%20Creation%20and%20Management%20Plan%20(Ref.%20Z22-03).pdf. The construction milestone concerns ecologically enhanced fishponds at the proposed Sam Po Shue Wetland Conservation Park.",
"runs": [
{
"text": "Id.",
"italic": true
},
{
- "text": " annex 1, condition 9, at 4. The construction milestone concerns ecologically enhanced fishponds at the proposed Sam Po Shue Wetland Conservation Park."
+ "text": " annex 1, condition 9, at 4; "
+ },
+ {
+ "text": "Civil Engineering and Development Department, Detailed Habitat Creation and Management Plan (Ref.",
+ "smallCaps": true
+ },
+ {
+ "text": " Z22-03): "
+ },
+ {
+ "text": "San Tin/Lok Ma Chau Development Node, Package",
+ "smallCaps": true
+ },
+ {
+ "text": " 1 §§ 1.1-1.2, tbl. 8.1, at 1-2, 68 (Aug. 2026), https://www.epd.gov.hk/eia/files/applications/en/pp_340/eia_1994/progress/action_182408/Detailed%20Habitat%20Creation%20and%20Management%20Plan%20(Ref.%20Z22-03).pdf. The construction milestone concerns ecologically enhanced fishponds at the proposed Sam Po Shue Wetland Conservation Park."
}
]
},
@@ -2468,7 +2527,11 @@
"text": "Fla. Admin. Code Ann. r. 62-345.600(1)-(3) (2007). The percentages are the reciprocal of the product of the two factors, rounded to the nearest whole number.",
"runs": [
{
- "text": "Fla. Admin. Code Ann. r. 62-345.600(1)-(3) (2007). The percentages are the reciprocal of the product of the two factors, rounded to the nearest whole number."
+ "text": "Fla. Admin. Code Ann.",
+ "smallCaps": true
+ },
+ {
+ "text": " r. 62-345.600(1)-(3) (2007). The percentages are the reciprocal of the product of the two factors, rounded to the nearest whole number."
}
]
},
@@ -2528,17 +2591,10 @@
},
{
"number": 52,
- "text": "Wetland Protection Law of the People's Republic of China art. 36 (2021); Ministry of Natural Resources & National Forestry & Grassland Administration, supra note 13, action 2; Guangdong Wetland Protection Regulation arts. 24, 30, 35 (as amended Nov. 30, 2022), Guangdong Provincial People's Congress Standing Committee Gazette, No. 8 of 2022, at 68-70 (Jan. 6, 2023). Article 36 addresses damage to lawful rights; it does not grant every informal user a compensation entitlement.",
+ "text": "Wetland Protection Law of the People's Republic of China art. 36 (2021); Ministry of Natural Resources & National Forestry and Grassland Administration, supra note 13, action 2; Guangdong Wetland Protection Regulation arts. 24, 30, 35 (as amended Nov. 30, 2022), Guangdong Provincial People's Congress Standing Committee Gazette, No. 8 of 2022, at 68-70 (Jan. 6, 2023). Article 36 addresses damage to lawful rights; it does not grant every informal user a compensation entitlement.",
"runs": [
{
- "text": "Wetland Protection Law of the People's Republic of China art. 36 (2021); "
- },
- {
- "text": "Ministry of Natural Resources & National Forestry & Grassland Administration",
- "smallCaps": true
- },
- {
- "text": ", "
+ "text": "Wetland Protection Law of the People's Republic of China art. 36 (2021); Ministry of Natural Resources & National Forestry and Grassland Administration, "
},
{
"text": "supra",
@@ -2581,7 +2637,7 @@
},
{
"number": 54,
- "text": "U.S. Government Accountability Office, GAO-05-898, Wetlands Protection: Corps of Engineers Does Not Have an Effective Oversight Approach to Ensure that Compensatory Mitigation Is Occurring, Highlights (2005). The seven-district review found monitoring requirements in eighty-nine of 152 permittee-mitigation files, receipt evidence in twenty-one, and inspection evidence in fifteen percent of the 152 files.",
+ "text": "U.S. Government Accountability Office, GAO-05-898, Wetlands Protection: Corps of Engineers Does Not Have an Effective Oversight Approach to Ensure that Compensatory Mitigation Is Occurring, Highlights (2005). The seven-district review found monitoring requirements in 89 of 152 permittee-mitigation files, receipt evidence in 21, and inspection evidence in 15 percent of the 152 files.",
"runs": [
{
"text": "U.S. Government Accountability Office, GAO",
@@ -2595,7 +2651,7 @@
"smallCaps": true
},
{
- "text": ", Highlights (2005). The seven-district review found monitoring requirements in eighty-nine of 152 permittee-mitigation files, receipt evidence in twenty-one, and inspection evidence in fifteen percent of the 152 files."
+ "text": ", Highlights (2005). The seven-district review found monitoring requirements in 89 of 152 permittee-mitigation files, receipt evidence in 21, and inspection evidence in 15 percent of the 152 files."
}
]
}
diff --git a/website/content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json b/website/content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json
index 697ddb6..9eb8cad 100644
--- a/website/content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json
+++ b/website/content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json
@@ -1,7 +1,7 @@
{
"slug": "small-states-and-the-governance-of-strategic-space-dependence",
"title": "Small States and the Governance of Strategic Space Dependence",
- "author": "Letao Ouyang",
+ "author": "Letao (Theodore) Ouyang",
"date": "December 2024; revised November 2026",
"repositoryUrl": "https://github.com/130U/LAW-Small-States-and-the-Governance-of-Strategic-Space-Dependence",
"sourceUrl": "https://github.com/130U/LAW-Small-States-and-the-Governance-of-Strategic-Space-Dependence/blob/main/manuscript/small-states-and-the-governance-of-strategic-space-dependence.md",
@@ -39,10 +39,10 @@
},
{
"type": "paragraph",
- "text": "On 18 December 2024, Norway announced the activation of its Arctic military broadband network. The satellites had been manufactured in the United States, launched by SpaceX, and designed to accommodate Norwegian, American, and commercial communications payloads. Space Norway nevertheless owned and operated the satellite platforms. Foreign participation had helped deliver a Norwegian capability; it had not transferred every relevant decision abroad.1 The arrangement raises two questions: which decisions dependence places beyond national control, and which decisions cooperation makes possible.",
+ "text": "On December 18, 2024, Norway announced the activation of its Arctic military broadband network. The satellites had been manufactured in the United States, launched by SpaceX, and designed to accommodate Norwegian, American, and commercial communications payloads. Space Norway nevertheless owned and operated the satellite platforms. Foreign participation had helped deliver a Norwegian capability; it had not transferred every relevant decision abroad.1 The arrangement raises two questions: which decisions dependence places beyond national control, and which decisions cooperation makes possible.",
"runs": [
{
- "text": "On 18 December 2024, Norway announced the activation of its Arctic military broadband network. The satellites had been manufactured in the United States, launched by SpaceX, and designed to accommodate Norwegian, American, and commercial communications payloads. Space Norway nevertheless owned and operated the satellite platforms. Foreign participation had helped deliver a Norwegian capability; it had not transferred every relevant decision abroad."
+ "text": "On December 18, 2024, Norway announced the activation of its Arctic military broadband network. The satellites had been manufactured in the United States, launched by SpaceX, and designed to accommodate Norwegian, American, and commercial communications payloads. Space Norway nevertheless owned and operated the satellite platforms. Foreign participation had helped deliver a Norwegian capability; it had not transferred every relevant decision abroad."
},
{
"text": "1",
@@ -178,10 +178,10 @@
},
{
"type": "paragraph",
- "text": "A third body of thought supplies the vocabulary for what the cases show. Hohfeld analyzed legal relations into distinct entitlements, and Honoré catalogued the incidents of ownership, among them the rights to possess, to use, and to manage, the right to the income, and the power to exclude. Nothing requires one person to hold them all.5 Grossman, Hart, and Moore identified what makes an owner an owner when contracts are incomplete: the residual right to decide in circumstances the contract does not cover.6 This Article reads control over a space system the same way and separates five entitlements. Permission is the power to allow or refuse an activity in advance. Title is ownership of the asset. Operation is command of the system in use. Interruption and priority concern who can stop the service or claim it first. Voice is participation in the rules that govern everyone. Part IV shows these entitlements divided among a regulator, an owner, an operator, and a supplier's government, and it shows that the residual right in a crisis follows the party that holds a chokepoint, such as a control segment, an export licence, or a ground station, whatever the title says.",
+ "text": "A third body of thought supplies the vocabulary for what the cases show. Hohfeld analyzed legal relations into distinct entitlements, and Honoré catalogued the incidents of ownership, among them the rights to possess, to use, and to manage, the right to the income, and the right to the capital. Nothing requires one person to hold them all.5 Grossman, Hart, and Moore identified what makes an owner an owner when contracts are incomplete: the residual right to decide in circumstances the contract does not cover.6 This Article reads control over a space system the same way and separates five entitlements. Permission is the power to allow or refuse an activity in advance. Title is ownership of the asset. Operation is command of the system in use. Interruption and priority concern who can stop the service or claim it first. Voice is participation in the rules that govern everyone. Part IV shows these entitlements divided among a regulator, an owner, an operator, and a supplier's government, and it shows that the residual right in a crisis follows the party that holds a chokepoint, such as a control segment, an export license, or a ground station, whatever the title says.",
"runs": [
{
- "text": "A third body of thought supplies the vocabulary for what the cases show. Hohfeld analyzed legal relations into distinct entitlements, and Honoré catalogued the incidents of ownership, among them the rights to possess, to use, and to manage, the right to the income, and the power to exclude. Nothing requires one person to hold them all."
+ "text": "A third body of thought supplies the vocabulary for what the cases show. Hohfeld analyzed legal relations into distinct entitlements, and Honoré catalogued the incidents of ownership, among them the rights to possess, to use, and to manage, the right to the income, and the right to the capital. Nothing requires one person to hold them all."
},
{
"text": "5",
@@ -197,7 +197,7 @@
"footnoteRef": "6"
},
{
- "text": " This Article reads control over a space system the same way and separates five entitlements. Permission is the power to allow or refuse an activity in advance. Title is ownership of the asset. Operation is command of the system in use. Interruption and priority concern who can stop the service or claim it first. Voice is participation in the rules that govern everyone. Part IV shows these entitlements divided among a regulator, an owner, an operator, and a supplier's government, and it shows that the residual right in a crisis follows the party that holds a chokepoint, such as a control segment, an export licence, or a ground station, whatever the title says."
+ "text": " This Article reads control over a space system the same way and separates five entitlements. Permission is the power to allow or refuse an activity in advance. Title is ownership of the asset. Operation is command of the system in use. Interruption and priority concern who can stop the service or claim it first. Voice is participation in the rules that govern everyone. Part IV shows these entitlements divided among a regulator, an owner, an operator, and a supplier's government, and it shows that the residual right in a crisis follows the party that holds a chokepoint, such as a control segment, an export license, or a ground station, whatever the title says."
}
]
},
@@ -510,10 +510,10 @@
},
{
"type": "paragraph",
- "text": "The official regulatory impact statement explains how this authority can coexist with international cooperation. Relevant foreign licences may reduce duplicative assessment, but New Zealand must still consider its own national-interest, security, and fitness requirements. Tailored conditions permit differentiated decisions. The national-interest refusal power is discretionary; a certificate of significant national-security risk prevents authorization.22 This is a concrete allocation of responsibility: foreign expertise may inform the decision without determining it. Whether officials can apply the distinction consistently remains an empirical question.",
+ "text": "The official regulatory impact statement explains how this authority can coexist with international cooperation. Relevant foreign licenses may reduce duplicative assessment, but New Zealand must still consider its own national-interest, security, and fitness requirements. Tailored conditions permit differentiated decisions. The national-interest refusal power is discretionary; a certificate of significant national-security risk prevents authorization.22 This is a concrete allocation of responsibility: foreign expertise may inform the decision without determining it. Whether officials can apply the distinction consistently remains an empirical question.",
"runs": [
{
- "text": "The official regulatory impact statement explains how this authority can coexist with international cooperation. Relevant foreign licences may reduce duplicative assessment, but New Zealand must still consider its own national-interest, security, and fitness requirements. Tailored conditions permit differentiated decisions. The national-interest refusal power is discretionary; a certificate of significant national-security risk prevents authorization."
+ "text": "The official regulatory impact statement explains how this authority can coexist with international cooperation. Relevant foreign licenses may reduce duplicative assessment, but New Zealand must still consider its own national-interest, security, and fitness requirements. Tailored conditions permit differentiated decisions. The national-interest refusal power is discretionary; a certificate of significant national-security risk prevents authorization."
},
{
"text": "22",
@@ -624,10 +624,10 @@
},
{
"type": "paragraph",
- "text": "Norway sought broadband capacity suited to Arctic conditions. ASBM brought together Space Norway, the Norwegian defence sector, the United States, and commercial communications interests. The satellites' foreign manufacture and launch are well documented. So is the allocation of specific roles: the platforms were owned by Space Norway and carried distinct Norwegian, American, and Viasat payloads.27 This arrangement creates several relationships of dependence, rather than one undifferentiated transfer of national control.",
+ "text": "Norway sought broadband capacity suited to Arctic conditions. ASBM brought together Space Norway, the Norwegian defense sector, the United States, and commercial communications interests. The satellites' foreign manufacture and launch are well documented. So is the allocation of specific roles: the platforms were owned by Space Norway and carried distinct Norwegian, American, and Viasat payloads.27 This arrangement creates several relationships of dependence, rather than one undifferentiated transfer of national control.",
"runs": [
{
- "text": "Norway sought broadband capacity suited to Arctic conditions. ASBM brought together Space Norway, the Norwegian defence sector, the United States, and commercial communications interests. The satellites' foreign manufacture and launch are well documented. So is the allocation of specific roles: the platforms were owned by Space Norway and carried distinct Norwegian, American, and Viasat payloads."
+ "text": "Norway sought broadband capacity suited to Arctic conditions. ASBM brought together Space Norway, the Norwegian defense sector, the United States, and commercial communications interests. The satellites' foreign manufacture and launch are well documented. So is the allocation of specific roles: the platforms were owned by Space Norway and carried distinct Norwegian, American, and Viasat payloads."
},
{
"text": "27",
@@ -684,10 +684,10 @@
},
{
"type": "paragraph",
- "text": "The financing record also shows conditions of the kind this Article proposes. When the Storting considered the state's equity commitment in 2018, the ministry's assessment was that the project must meet the armed forces' need for military broadband under Norwegian control, and that no funds would be released until binding customer agreements were concluded and all capacity had been sold in advance for the life of the project. The project description placed the ground station and operations centre in Norway in order to secure national control of the system.30 The legislature financed a partnership on stated terms about who would control what.",
+ "text": "The financing record also shows conditions of the kind this Article proposes. When the Storting considered the state's equity commitment in 2018, the ministry's assessment was that the project must meet the armed forces' need for military broadband under Norwegian control, and that no funds would be released until binding customer agreements were concluded and all capacity had been sold in advance for the life of the project. The project description placed the ground station and operations center in Norway in order to secure national control of the system.30 The legislature financed a partnership on stated terms about who would control what.",
"runs": [
{
- "text": "The financing record also shows conditions of the kind this Article proposes. When the Storting considered the state's equity commitment in 2018, the ministry's assessment was that the project must meet the armed forces' need for military broadband under Norwegian control, and that no funds would be released until binding customer agreements were concluded and all capacity had been sold in advance for the life of the project. The project description placed the ground station and operations centre in Norway in order to secure national control of the system."
+ "text": "The financing record also shows conditions of the kind this Article proposes. When the Storting considered the state's equity commitment in 2018, the ministry's assessment was that the project must meet the armed forces' need for military broadband under Norwegian control, and that no funds would be released until binding customer agreements were concluded and all capacity had been sold in advance for the life of the project. The project description placed the ground station and operations center in Norway in order to secure national control of the system."
},
{
"text": "30",
@@ -738,10 +738,10 @@
},
{
"type": "paragraph",
- "text": "Luxembourg's LUXEOSys programme supplies an adverse check on the sufficiency of acquiring a national asset. Here the relevant public decision concerned the resources and organization necessary for an observation system, including its ground segment and operation. The Cour des comptes' 2022 special audit identified planning omissions, underestimated staffing, and coordination problems in explaining the increase in the legislatively authorized envelope from EUR170 million to EUR309 million, excluding VAT, in 2020.32 Both figures are expenditure authorizations.",
+ "text": "Luxembourg's LUXEOSys program supplies an adverse check on the sufficiency of acquiring a national asset. Here the relevant public decision concerned the resources and organization necessary for an observation system, including its ground segment and operation. The Cour des comptes' 2022 special audit identified planning omissions, underestimated staffing, and coordination problems in explaining the increase in the legislatively authorized envelope from EUR170 million to EUR309 million, excluding VAT, in 2020.32 Both figures are expenditure authorizations.",
"runs": [
{
- "text": "Luxembourg's LUXEOSys programme supplies an adverse check on the sufficiency of acquiring a national asset. Here the relevant public decision concerned the resources and organization necessary for an observation system, including its ground segment and operation. The Cour des comptes' 2022 special audit identified planning omissions, underestimated staffing, and coordination problems in explaining the increase in the legislatively authorized envelope from EUR170 million to EUR309 million, excluding VAT, in 2020."
+ "text": "Luxembourg's LUXEOSys program supplies an adverse check on the sufficiency of acquiring a national asset. Here the relevant public decision concerned the resources and organization necessary for an observation system, including its ground segment and operation. The Cour des comptes' 2022 special audit identified planning omissions, underestimated staffing, and coordination problems in explaining the increase in the legislatively authorized envelope from EUR170 million to EUR309 million, excluding VAT, in 2020."
},
{
"text": "32",
@@ -755,10 +755,10 @@
},
{
"type": "paragraph",
- "text": "The audit shows what the title did not include. Operation of the system, meaning control and piloting of the satellite and the management centre that handles image requests, was to be performed by a third-party provider at a cost of EUR75 million, and the antennas were to be installed on rented infrastructure at Redu, in Belgium.33 The record before commitment had already flagged the gap in information. In its opinion on the 2018 bill, the Conseil d'État regretted that, on the ground of confidential negotiations, the documentation contained no details, even summary ones, and no indication of the split between investment and operating expenditure.34 Ownership was acquired; the capacity to operate was bought separately, and the legislature authorized the first without an account of the second.",
+ "text": "The audit shows what the title did not include. Operation of the system, meaning control and piloting of the satellite and the management center that handles image requests, was to be performed by a third-party provider at a cost of EUR75 million, and the antennas were to be installed on rented infrastructure at Redu, in Belgium.33 The record before commitment had already flagged the gap in information. In its opinion on the 2018 bill, the Conseil d'État regretted that, on the ground of confidential negotiations, the documentation contained no details, even summary ones, and no indication of the split between investment and operating expenditure.34 Ownership was acquired; the capacity to operate was bought separately, and the legislature authorized the first without an account of the second.",
"runs": [
{
- "text": "The audit shows what the title did not include. Operation of the system, meaning control and piloting of the satellite and the management centre that handles image requests, was to be performed by a third-party provider at a cost of EUR75 million, and the antennas were to be installed on rented infrastructure at Redu, in Belgium."
+ "text": "The audit shows what the title did not include. Operation of the system, meaning control and piloting of the satellite and the management center that handles image requests, was to be performed by a third-party provider at a cost of EUR75 million, and the antennas were to be installed on rented infrastructure at Redu, in Belgium."
},
{
"text": "33",
@@ -780,10 +780,10 @@
},
{
"type": "paragraph",
- "text": "The mechanism concerns information and implementation. Parliament can authorize a satellite purchase while lacking a reliable account of the associated organization and lifetime obligations. Contractors may supply hardware or expertise, but a public buyer still needs the ability to define requirements and integrate the system. Legal title cannot perform those functions. Conversely, obtaining outside expertise can improve national decision-making rather than replace it. The audit leaves open whether purchased imagery or a pooled programme would have achieved the same objectives more cheaply.",
+ "text": "The mechanism concerns information and implementation. Parliament can authorize a satellite purchase while lacking a reliable account of the associated organization and lifetime obligations. Contractors may supply hardware or expertise, but a public buyer still needs the ability to define requirements and integrate the system. Legal title cannot perform those functions. Conversely, obtaining outside expertise can improve national decision-making rather than replace it. The audit leaves open whether purchased imagery or a pooled program would have achieved the same objectives more cheaply.",
"runs": [
{
- "text": "The mechanism concerns information and implementation. Parliament can authorize a satellite purchase while lacking a reliable account of the associated organization and lifetime obligations. Contractors may supply hardware or expertise, but a public buyer still needs the ability to define requirements and integrate the system. Legal title cannot perform those functions. Conversely, obtaining outside expertise can improve national decision-making rather than replace it. The audit leaves open whether purchased imagery or a pooled programme would have achieved the same objectives more cheaply."
+ "text": "The mechanism concerns information and implementation. Parliament can authorize a satellite purchase while lacking a reliable account of the associated organization and lifetime obligations. Contractors may supply hardware or expertise, but a public buyer still needs the ability to define requirements and integrate the system. Legal title cannot perform those functions. Conversely, obtaining outside expertise can improve national decision-making rather than replace it. The audit leaves open whether purchased imagery or a pooled program would have achieved the same objectives more cheaply."
}
]
},
@@ -916,10 +916,10 @@
]
},
{
- "text": "Not at issue: a state programme",
+ "text": "Not at issue: a state program",
"runs": [
{
- "text": "Not at issue: a state programme"
+ "text": "Not at issue: a state program"
}
]
}
@@ -1039,10 +1039,10 @@
},
{
"type": "paragraph",
- "text": "The table states the thesis of this Article in one view. No case combines all the entitlements, and the one that decides a crisis, the power to interrupt, is the one the public record leaves least defined. Where it can be located, it sits with whoever controls a necessary input: the launch state before launch, a payload's own control arrangements, a contracted operator. Residual control follows the chokepoint, and a title or a licence shows where it lies only by coincidence.",
+ "text": "The table states the thesis of this Article in one view. No case combines all the entitlements, and the one that decides a crisis, the power to interrupt, is the one the public record leaves least defined. Where it can be located, it sits with whoever controls a necessary input: the launch state before launch, a payload's own control arrangements, a contracted operator. Residual control follows the chokepoint, and a title or a license shows where it lies only by coincidence.",
"runs": [
{
- "text": "The table states the thesis of this Article in one view. No case combines all the entitlements, and the one that decides a crisis, the power to interrupt, is the one the public record leaves least defined. Where it can be located, it sits with whoever controls a necessary input: the launch state before launch, a payload's own control arrangements, a contracted operator. Residual control follows the chokepoint, and a title or a licence shows where it lies only by coincidence."
+ "text": "The table states the thesis of this Article in one view. No case combines all the entitlements, and the one that decides a crisis, the power to interrupt, is the one the public record leaves least defined. Where it can be located, it sits with whoever controls a necessary input: the launch state before launch, a payload's own control arrangements, a contracted operator. Residual control follows the chokepoint, and a title or a license shows where it lies only by coincidence."
}
]
},
@@ -1102,10 +1102,10 @@
},
{
"type": "paragraph",
- "text": "New Zealand's October 2023 statement is unusually direct evidence of this problem. It supported continuation of the responsible-behaviours process and warned that competing groups would strain the capacity of smaller and emerging space states.38 That is an expressed institutional concern, not a quantified measure of diplomatic exclusion. In December 2024, the Assembly adopted decision 79/512, replacing the two previously mandated groups with a single process while retaining consensus. The recorded vote was 174-2-9.39 Consolidation addressed a procedural problem within the observation period; the record does not establish that New Zealand caused the change or that the new process would produce substantive agreement.",
+ "text": "New Zealand's October 2023 statement is unusually direct evidence of this problem. It supported continuation of the responsible-behaviors process and warned that competing groups would strain the capacity of smaller and emerging space states.38 That is an expressed institutional concern, not a quantified measure of diplomatic exclusion. In December 2024, the Assembly adopted decision 79/512, replacing the two previously mandated groups with a single process while retaining consensus. The recorded vote was 174-2-9.39 Consolidation addressed a procedural problem within the observation period; the record does not establish that New Zealand caused the change or that the new process would produce substantive agreement.",
"runs": [
{
- "text": "New Zealand's October 2023 statement is unusually direct evidence of this problem. It supported continuation of the responsible-behaviours process and warned that competing groups would strain the capacity of smaller and emerging space states."
+ "text": "New Zealand's October 2023 statement is unusually direct evidence of this problem. It supported continuation of the responsible-behaviors process and warned that competing groups would strain the capacity of smaller and emerging space states."
},
{
"text": "38",
@@ -1127,10 +1127,10 @@
},
{
"type": "paragraph",
- "text": "A separate Group of Governmental Experts adopted a consensus report in August 2024. The report expressly distinguished agreement on the document from common views on all proposed substantive elements and did not seek a common understanding of responsible behaviours.40 Consensus remained possible at one level while important disagreements persisted at another. This reinforces the need to identify what was agreed, rather than treating either a failed report or an adopted document as a verdict on the entire system.",
+ "text": "A separate Group of Governmental Experts adopted a consensus report in August 2024. The report expressly distinguished agreement on the document from common views on all proposed substantive elements and did not seek a common understanding of responsible behaviors.40 Consensus remained possible at one level while important disagreements persisted at another. This reinforces the need to identify what was agreed, rather than treating either a failed report or an adopted document as a verdict on the entire system.",
"runs": [
{
- "text": "A separate Group of Governmental Experts adopted a consensus report in August 2024. The report expressly distinguished agreement on the document from common views on all proposed substantive elements and did not seek a common understanding of responsible behaviours."
+ "text": "A separate Group of Governmental Experts adopted a consensus report in August 2024. The report expressly distinguished agreement on the document from common views on all proposed substantive elements and did not seek a common understanding of responsible behaviors."
},
{
"text": "40",
@@ -1144,10 +1144,10 @@
},
{
"type": "paragraph",
- "text": "Regional cooperation offers a different allocation of authority. The ESA Convention combines mandatory activities with optional programmes, contribution rules, industrial participation objectives, and Council voting rights. Each member has one Council vote, but cannot vote on matters concerning exclusively an accepted programme in which it does not participate. Industrial participation is linked to financial contributions, rather than distributed equally.41 Pooling organizes shared technical capacity while participation carries financial commitments and limits on unilateral choice. Whether a separate national programme would perform better requires a programme-specific comparison. ESA supplies an institutional route for collective agency, without guaranteeing equal national influence or returns.",
+ "text": "Regional cooperation offers a different allocation of authority. The ESA Convention combines mandatory activities with optional programs, contribution rules, industrial participation objectives, and Council voting rights. Each member has one Council vote, but cannot vote on matters concerning exclusively an accepted program in which it does not participate. Industrial participation is linked to financial contributions, rather than distributed equally.41 Pooling organizes shared technical capacity while participation carries financial commitments and limits on unilateral choice. Whether a separate national program would perform better requires a program-specific comparison. ESA supplies an institutional route for collective agency, without guaranteeing equal national influence or returns.",
"runs": [
{
- "text": "Regional cooperation offers a different allocation of authority. The ESA Convention combines mandatory activities with optional programmes, contribution rules, industrial participation objectives, and Council voting rights. Each member has one Council vote, but cannot vote on matters concerning exclusively an accepted programme in which it does not participate. Industrial participation is linked to financial contributions, rather than distributed equally."
+ "text": "Regional cooperation offers a different allocation of authority. The ESA Convention combines mandatory activities with optional programs, contribution rules, industrial participation objectives, and Council voting rights. Each member has one Council vote, but cannot vote on matters concerning exclusively an accepted program in which it does not participate. Industrial participation is linked to financial contributions, rather than distributed equally."
},
{
"text": "41",
@@ -1155,7 +1155,7 @@
"footnoteRef": "41"
},
{
- "text": " Pooling organizes shared technical capacity while participation carries financial commitments and limits on unilateral choice. Whether a separate national programme would perform better requires a programme-specific comparison. ESA supplies an institutional route for collective agency, without guaranteeing equal national influence or returns."
+ "text": " Pooling organizes shared technical capacity while participation carries financial commitments and limits on unilateral choice. Whether a separate national program would perform better requires a program-specific comparison. ESA supplies an institutional route for collective agency, without guaranteeing equal national influence or returns."
}
]
},
@@ -1198,7 +1198,7 @@
},
{
"type": "paragraph",
- "text": "The terms such a process would examine are well developed elsewhere. European financial regulation requires firms that rely on outside technology providers for critical functions to \"put in place exit strategies\" and to contract for termination rights, notice periods, and \"a mandatory adequate transition period.\"43 European defence procurement lets a buyer require a tenderer to indicate any restriction on \"disclosure, transfer or use\" arising from export-control or security arrangements, to commit to \"the capacity required to meet additional needs\" in a crisis, and to give notice of changes in its organisation or supply chain.44 This Article does not claim those clauses as its own. Its claim is that their absence from an arrangement for an essential space function should be something a named public decision maker has to explain, in a form that can be reviewed.",
+ "text": "The terms such a process would examine are well developed elsewhere. European financial regulation requires firms that rely on outside technology providers for critical functions to \"put in place exit strategies\" and to contract for termination rights, notice periods, and \"a mandatory adequate transition period.\"43 European defense procurement lets a buyer require a tenderer to indicate any restriction on \"disclosure, transfer or use\" arising from export-control or security arrangements, to commit to \"the capacity required to meet additional needs\" in a crisis, and to give notice of changes in its organization or supply chain.44 This Article does not claim those clauses as its own. Its claim is that their absence from an arrangement for an essential space function should be something a named public decision maker has to explain, in a form that can be reviewed.",
"runs": [
{
"text": "The terms such a process would examine are well developed elsewhere. European financial regulation requires firms that rely on outside technology providers for critical functions to \"put in place exit strategies\" and to contract for termination rights, notice periods, and \"a mandatory adequate transition period.\""
@@ -1209,7 +1209,7 @@
"footnoteRef": "43"
},
{
- "text": " European defence procurement lets a buyer require a tenderer to indicate any restriction on \"disclosure, transfer or use\" arising from export-control or security arrangements, to commit to \"the capacity required to meet additional needs\" in a crisis, and to give notice of changes in its organisation or supply chain."
+ "text": " European defense procurement lets a buyer require a tenderer to indicate any restriction on \"disclosure, transfer or use\" arising from export-control or security arrangements, to commit to \"the capacity required to meet additional needs\" in a crisis, and to give notice of changes in its organization or supply chain."
},
{
"text": "44",
@@ -1338,15 +1338,6 @@
"text": "If this argument is right, the policy objective should be accountable acceptance of important dependencies. Authorities should identify the decisions others can withhold, establish usable response arrangements, and fund the competence to exercise retained rights. Material gaps should lead to reasoned acceptance or revision before commitment. This standard permits beneficial cooperation while requiring an explanation of its limits. It must accompany restraint against harms that national bargains cannot contain. Small states strengthen their agency when cooperation makes lawful choices usable and accountable and when they recognize that other states and civilian users have corresponding claims."
}
]
- },
- {
- "type": "paragraph",
- "text": "---",
- "runs": [
- {
- "text": "---"
- }
- ]
}
],
"footnotes": [
@@ -1402,7 +1393,7 @@
},
{
"number": 3,
- "text": "Tom Long, Small States, Great Power? Gaining Influence Through Intrinsic, Derivative, and Collective Power, 19 Int'l Stud. Rev. 185, 185-205 (2017).",
+ "text": "Tom Long, Small States, Great Power? Gaining Influence Through Intrinsic, Derivative, and Collective Power, 19 Int'l Stud. Rev. 185 (2017).",
"runs": [
{
"text": "Tom Long, "
@@ -1419,7 +1410,7 @@
"smallCaps": true
},
{
- "text": " 185, 185-205 (2017)."
+ "text": " 185 (2017)."
}
]
},
@@ -1472,7 +1463,14 @@
"italic": true
},
{
- "text": ", in "
+ "text": ", "
+ },
+ {
+ "text": "in",
+ "italic": true
+ },
+ {
+ "text": " "
},
{
"text": "Oxford Essays in Jurisprudence",
@@ -1794,10 +1792,10 @@
},
{
"number": 22,
- "text": "New Zealand Ministry of Business, Innovation & Employment, Regulatory Impact Statement: The Outer Space and High-Altitude Activities Act 2017 Regulations 7-8, ¶¶ 7, 17, 19-21 (Aug. 21, 2017).",
+ "text": "New Zealand Ministry of Business, Innovation & Employment, Regulatory Impact Statement: The Outer Space and High-altitude Activities Act 2017 Regulations 7-8, ¶¶ 7, 17, 19-21 (Aug. 21, 2017).",
"runs": [
{
- "text": "New Zealand Ministry of Business, Innovation & Employment, Regulatory Impact Statement: The Outer Space and High-Altitude Activities Act",
+ "text": "New Zealand Ministry of Business, Innovation & Employment, Regulatory Impact Statement: The Outer Space and High-altitude Activities Act",
"smallCaps": true
},
{
@@ -1814,14 +1812,14 @@
},
{
"number": 23,
- "text": "Brendan Horsley, Inspector-General of Intelligence & Security, Report into a Review of GCSB and NZSIS Activity and Assessments Under the Outer Space and High-Altitude Activities Act 2017 3-5, 7-13 (Apr. 20, 2021).",
+ "text": "Brendan Horsley, Inspector-General of Intelligence & Security, Report into a Review of GCSB and NZSIS Activity and Assessments Under the Outer Space and High-altitude Activities Act 2017, at 3-5, 7-13 (Apr. 20, 2021).",
"runs": [
{
- "text": "Brendan Horsley, Inspector-General of Intelligence & Security, Report into a Review of GCSB and NZSIS Activity and Assessments Under the Outer Space and High-Altitude Activities Act",
+ "text": "Brendan Horsley, Inspector-General of Intelligence & Security, Report into a Review of GCSB and NZSIS Activity and Assessments Under the Outer Space and High-altitude Activities Act",
"smallCaps": true
},
{
- "text": " 2017 3-5, 7-13 (Apr. 20, 2021)."
+ "text": " 2017, at 3-5, 7-13 (Apr. 20, 2021)."
}
]
},
@@ -1850,10 +1848,10 @@
},
{
"number": 25,
- "text": "New Zealand Ministry of Business, Innovation & Employment, Statutory Review of the Outer Space and High-Altitude Activities Act 2017, at 9 tbl.1 & ¶ 31 (2022), https://www.mbie.govt.nz/assets/report-on-the-statutory-review-of-the-oshaa-2017.pdf.",
+ "text": "New Zealand Ministry of Business, Innovation & Employment, Statutory Review of the Outer Space and High-altitude Activities Act 2017, at 9 tbl.1 & ¶ 31 (2022), https://www.mbie.govt.nz/assets/report-on-the-statutory-review-of-the-oshaa-2017.pdf.",
"runs": [
{
- "text": "New Zealand Ministry of Business, Innovation & Employment, Statutory Review of the Outer Space and High-Altitude Activities Act",
+ "text": "New Zealand Ministry of Business, Innovation & Employment, Statutory Review of the Outer Space and High-altitude Activities Act",
"smallCaps": true
},
{
@@ -1876,7 +1874,7 @@
},
{
"number": 27,
- "text": "U.S. Space Systems Command, Enhanced Polar System-Recapitalization (EPS-R) Payloads Launched for U.S. Space Force 1-2 (Aug. 15, 2024), https://www.ssc.spaceforce.mil/Portals/3/EPS-R%20Launch%20Norway.pdf.",
+ "text": "U.S. Space Systems Command, supra note 1, at 1-2.",
"runs": [
{
"text": "U.S. Space Systems Command",
@@ -1886,17 +1884,17 @@
"text": ", "
},
{
- "text": "Enhanced Polar System-Recapitalization (EPS-R) Payloads Launched for U.S. Space Force",
+ "text": "supra",
"italic": true
},
{
- "text": " 1-2 (Aug. 15, 2024), https://www.ssc.spaceforce.mil/Portals/3/EPS-R%20Launch%20Norway.pdf."
+ "text": " note 1, at 1-2."
}
]
},
{
"number": 28,
- "text": "Space Norway, ASBM Handover Marks an Important Milestone (Oct. 17, 2024), https://spacenorway.com/press-release/asbm-handover-marks-an-important-milestone/; Norwegian Ministry of Defence, Norway Switches on Arctic Broadband Network (Dec. 18, 2024), https://www.regjeringen.no/en/whats-new/norway-activates-arctic-broadband-network/id3079294/.",
+ "text": "Space Norway, ASBM Handover Marks an Important Milestone (Oct. 17, 2024), https://spacenorway.com/press-release/asbm-handover-marks-an-important-milestone/; Norwegian Ministry of Defence, supra note 1.",
"runs": [
{
"text": "Space Norway",
@@ -1920,11 +1918,11 @@
"text": ", "
},
{
- "text": "Norway Switches on Arctic Broadband Network",
+ "text": "supra",
"italic": true
},
{
- "text": " (Dec. 18, 2024), https://www.regjeringen.no/en/whats-new/norway-activates-arctic-broadband-network/id3079294/."
+ "text": " note 1."
}
]
},
@@ -1972,10 +1970,10 @@
},
{
"number": 32,
- "text": "Cour des comptes du Grand-Duché de Luxembourg, Rapport spécial sur la gestion financière du projet LuxEOSys 8-10, 33-35 (June 16, 2022).",
+ "text": "Cour des comptes du Grand-Duché de Luxembourg, Rapport spécial sur la gestion financière du projet LUXEOSys 8-10, 33-35 (June 16, 2022).",
"runs": [
{
- "text": "Cour des comptes du Grand-Duché de Luxembourg, Rapport spécial sur la gestion financière du projet LuxEOSys",
+ "text": "Cour des comptes du Grand-Duché de Luxembourg, Rapport spécial sur la gestion financière du projet LUXEOSys",
"smallCaps": true
},
{
diff --git a/website/content/legal-papers/who-may-choose-the-lesser-risk.json b/website/content/legal-papers/who-may-choose-the-lesser-risk.json
index 33e29ca..bca6a14 100644
--- a/website/content/legal-papers/who-may-choose-the-lesser-risk.json
+++ b/website/content/legal-papers/who-may-choose-the-lesser-risk.json
@@ -1,7 +1,7 @@
{
"slug": "who-may-choose-the-lesser-risk",
"title": "Who May Choose the Lesser Risk: Solar Geoengineering and the Legal Duties of Comparison and Continuity",
- "author": "Letao Ouyang",
+ "author": "Letao (Theodore) Ouyang",
"date": "December 2024; revised November 2026",
"repositoryUrl": "https://github.com/130U/LAW-Who-May-Choose-the-Lesser-Risk",
"sourceUrl": "https://github.com/130U/LAW-Who-May-Choose-the-Lesser-Risk/blob/main/manuscript/who-may-choose-the-lesser-risk.md",
@@ -39,10 +39,10 @@
},
{
"type": "paragraph",
- "text": "A government confronting intensifying climate damage might reasonably ask whether reflecting additional sunlight could reduce suffering. A person exposed to the resulting intervention could reasonably ask a different question: who authorized that government to choose the climate under which she must live? Neither question disappears when a model predicts lower global temperatures. Solar radiation modification forces risk regulation to confront both the consequences of action and the authority to determine which consequences others must bear.",
+ "text": "A government confronting intensifying climate damage might reasonably ask whether reflecting additional sunlight could reduce suffering. A person exposed to the resulting intervention could reasonably ask a different question: who authorized that government to choose the climate under which she must live? Neither question disappears when a model predicts lower global temperatures. Solar radiation modification (SRM) forces risk regulation to confront both the consequences of action and the authority to determine which consequences others must bear.",
"runs": [
{
- "text": "A government confronting intensifying climate damage might reasonably ask whether reflecting additional sunlight could reduce suffering. A person exposed to the resulting intervention could reasonably ask a different question: who authorized that government to choose the climate under which she must live? Neither question disappears when a model predicts lower global temperatures. Solar radiation modification forces risk regulation to confront both the consequences of action and the authority to determine which consequences others must bear."
+ "text": "A government confronting intensifying climate damage might reasonably ask whether reflecting additional sunlight could reduce suffering. A person exposed to the resulting intervention could reasonably ask a different question: who authorized that government to choose the climate under which she must live? Neither question disappears when a model predicts lower global temperatures. Solar radiation modification (SRM) forces risk regulation to confront both the consequences of action and the authority to determine which consequences others must bear."
}
]
},
@@ -198,10 +198,10 @@
},
{
"type": "paragraph",
- "text": "A second limit concerns causal structure. Physical hazards, political responses, and governance failures interact. If confidence in SAI weakened emissions reductions, that response could increase the cooling subsequently required and the consequences of interruption. This is a plausible mechanism requiring investigation, not an established universal effect of research. The IPCC distinguishes evidence about individual responses from the harder question of effects on societal mitigation, and reports low agreement about a research-to-deployment slippery slope.7 Counting mitigation deterrence, longer deployment, and termination consequences as independent costs could also double-count one causal pathway. An adequate analysis must disclose dependencies as well as enumerate harms.",
+ "text": "A second limit concerns causal structure. Physical hazards, political responses, and governance failures interact. If confidence in SAI weakened emissions reductions, that response could increase the cooling subsequently required and the consequences of interruption. This is a plausible mechanism requiring investigation, not an established universal effect of research. The Intergovernmental Panel on Climate Change (IPCC) distinguishes evidence about individual responses from the harder question of effects on societal mitigation, and reports low agreement about a research-to-deployment slippery slope.7 Counting mitigation deterrence, longer deployment, and termination consequences as independent costs could also double-count one causal pathway. An adequate analysis must disclose dependencies as well as enumerate harms.",
"runs": [
{
- "text": "A second limit concerns causal structure. Physical hazards, political responses, and governance failures interact. If confidence in SAI weakened emissions reductions, that response could increase the cooling subsequently required and the consequences of interruption. This is a plausible mechanism requiring investigation, not an established universal effect of research. The IPCC distinguishes evidence about individual responses from the harder question of effects on societal mitigation, and reports low agreement about a research-to-deployment slippery slope."
+ "text": "A second limit concerns causal structure. Physical hazards, political responses, and governance failures interact. If confidence in SAI weakened emissions reductions, that response could increase the cooling subsequently required and the consequences of interruption. This is a plausible mechanism requiring investigation, not an established universal effect of research. The Intergovernmental Panel on Climate Change (IPCC) distinguishes evidence about individual responses from the harder question of effects on societal mitigation, and reports low agreement about a research-to-deployment slippery slope."
},
{
"text": "7",
@@ -308,7 +308,14 @@
"text": "Consider a hypothetical proposal justified by reducing heat-related harm in a densely populated region. Suppose its analysis also identifies an uncertain threat to rain-dependent livelihoods elsewhere. Comparing deployment only with unchanged emissions and unfunded adaptation may favor SAI. Adding a feasible package of heat protection and livelihood support may change the comparison, even if it supplies less global cooling. Conversely, if those measures cannot reach the most exposed people soon enough, their nominal availability does not defeat the proposal. The duty requires an account of delivery capacity as well as modeled effects. It also asks whether the sponsoring actors could finance protection for the second group without sacrificing the first group's gains. None of these outcomes follows from the designation risk transfer.",
"runs": [
{
- "text": "Consider a hypothetical proposal justified by reducing heat-related harm in a densely populated region. Suppose its analysis also identifies an uncertain threat to rain-dependent livelihoods elsewhere. Comparing deployment only with unchanged emissions and unfunded adaptation may favor SAI. Adding a feasible package of heat protection and livelihood support may change the comparison, even if it supplies less global cooling. Conversely, if those measures cannot reach the most exposed people soon enough, their nominal availability does not defeat the proposal. The duty requires an account of delivery capacity as well as modeled effects. It also asks whether the sponsoring actors could finance protection for the second group without sacrificing the first group's gains. None of these outcomes follows from the designation risk transfer."
+ "text": "Consider a hypothetical proposal justified by reducing heat-related harm in a densely populated region. Suppose its analysis also identifies an uncertain threat to rain-dependent livelihoods elsewhere. Comparing deployment only with unchanged emissions and unfunded adaptation may favor SAI. Adding a feasible package of heat protection and livelihood support may change the comparison, even if it supplies less global cooling. Conversely, if those measures cannot reach the most exposed people soon enough, their nominal availability does not defeat the proposal. The duty requires an account of delivery capacity as well as modeled effects. It also asks whether the sponsoring actors could finance protection for the second group without sacrificing the first group's gains. None of these outcomes follows from the designation "
+ },
+ {
+ "text": "risk transfer",
+ "italic": true
+ },
+ {
+ "text": "."
}
]
},
@@ -419,7 +426,14 @@
"text": "The research-deployment distinction is substantive. Modeling, observation, and limited process experiments can answer questions without producing a climatically significant intervention. A sequence of experiments, however, may build operational capacity, institutions, or political expectations. Regulation should examine physical effects and cumulative program development, rather than accept either the word research as an exemption or the possibility of future use as proof that all inquiry is equivalent to deployment.",
"runs": [
{
- "text": "The research-deployment distinction is substantive. Modeling, observation, and limited process experiments can answer questions without producing a climatically significant intervention. A sequence of experiments, however, may build operational capacity, institutions, or political expectations. Regulation should examine physical effects and cumulative program development, rather than accept either the word research as an exemption or the possibility of future use as proof that all inquiry is equivalent to deployment."
+ "text": "The research-deployment distinction is substantive. Modeling, observation, and limited process experiments can answer questions without producing a climatically significant intervention. A sequence of experiments, however, may build operational capacity, institutions, or political expectations. Regulation should examine physical effects and cumulative program development, rather than accept either the word "
+ },
+ {
+ "text": "research",
+ "italic": true
+ },
+ {
+ "text": " as an exemption or the possibility of future use as proof that all inquiry is equivalent to deployment."
}
]
},
@@ -493,10 +507,10 @@
},
{
"type": "paragraph",
- "text": "The Convention on Biological Diversity's Decision X/33, paragraph 8(w), calls for restraint on climate-related geoengineering that may affect biodiversity, subject to conditions and a limited scientific-research exception. Its chapeau and status matter: this guidance is not a universally binding treaty prohibition. Its conditional scientific-research exception is not permission for unrestricted outdoor testing.20",
+ "text": "Under the Convention on Biological Diversity (CBD), Decision X/33, paragraph 8(w), calls for restraint on climate-related geoengineering that may affect biodiversity, subject to conditions and a limited scientific-research exception. Its chapeau and status matter: this guidance is not a universally binding treaty prohibition. Its conditional scientific-research exception is not permission for unrestricted outdoor testing.20",
"runs": [
{
- "text": "The Convention on Biological Diversity's Decision X/33, paragraph 8(w), calls for restraint on climate-related geoengineering that may affect biodiversity, subject to conditions and a limited scientific-research exception. Its chapeau and status matter: this guidance is not a universally binding treaty prohibition. Its conditional scientific-research exception is not permission for unrestricted outdoor testing."
+ "text": "Under the Convention on Biological Diversity (CBD), Decision X/33, paragraph 8(w), calls for restraint on climate-related geoengineering that may affect biodiversity, subject to conditions and a limited scientific-research exception. Its chapeau and status matter: this guidance is not a universally binding treaty prohibition. Its conditional scientific-research exception is not permission for unrestricted outdoor testing."
},
{
"text": "20",
@@ -543,10 +557,10 @@
},
{
"type": "paragraph",
- "text": "These materials support assessment and preventive responsibility. They do not already create individual standing in a global SAI tribunal, prescribe a voting rule, or impose the proposed financing requirements. Nor should the hostile-use restrictions in ENMOD be mistaken for a general prohibition on peaceful climate intervention; peaceful purpose, in turn, does not exempt an activity from other international law.24",
+ "text": "These materials support assessment and preventive responsibility. They do not already create individual standing in a global SAI tribunal, prescribe a voting rule, or impose the proposed financing requirements. Nor should the hostile-use restrictions in the Environmental Modification Convention (ENMOD) be mistaken for a general prohibition on peaceful climate intervention; peaceful purpose, in turn, does not exempt an activity from other international law.24",
"runs": [
{
- "text": "These materials support assessment and preventive responsibility. They do not already create individual standing in a global SAI tribunal, prescribe a voting rule, or impose the proposed financing requirements. Nor should the hostile-use restrictions in ENMOD be mistaken for a general prohibition on peaceful climate intervention; peaceful purpose, in turn, does not exempt an activity from other international law."
+ "text": "These materials support assessment and preventive responsibility. They do not already create individual standing in a global SAI tribunal, prescribe a voting rule, or impose the proposed financing requirements. Nor should the hostile-use restrictions in the Environmental Modification Convention (ENMOD) be mistaken for a general prohibition on peaceful climate intervention; peaceful purpose, in turn, does not exempt an activity from other international law."
},
{
"text": "24",
@@ -557,7 +571,7 @@
},
{
"type": "paragraph",
- "text": "Three instruments bear on stratospheric aerosols more directly than those usually invoked. The Vienna Convention for the Protection of the Ozone Layer obliges parties to protect against \"adverse effects resulting or likely to result from human activities which modify or are likely to modify the ozone layer\" and to control such activities under their jurisdiction; sulfate injection is an activity the ozone assessment treats as capable of affecting stratospheric ozone.25 The Framework Convention on Climate Change commits parties to \"employ appropriate methods, for example impact assessments,\" to minimize the adverse effects \"of projects or measures undertaken by them to mitigate or adapt to climate change.\"26 And the International Law Commission's guidelines on the protection of the atmosphere provide that intentional large-scale modification \"should only be conducted with prudence and caution, and subject to any applicable rules of international law, including those relating to environmental impact assessment,\" in a commentary that names stratospheric aerosols.27",
+ "text": "Three instruments bear on stratospheric aerosols more directly than those usually invoked. The Vienna Convention for the Protection of the Ozone Layer obliges parties to protect against \"adverse effects resulting or likely to result from human activities which modify or are likely to modify the ozone layer\" and to control such activities under their jurisdiction; sulfate injection is an activity the ozone assessment treats as capable of affecting stratospheric ozone.25 The United Nations Framework Convention on Climate Change (UNFCCC) commits parties to \"employ appropriate methods, for example impact assessments,\" to minimize the adverse effects \"of projects or measures undertaken by them to mitigate or adapt to climate change.\"26 And the International Law Commission's guidelines on the protection of the atmosphere provide that intentional large-scale modification \"should only be conducted with prudence and caution, and subject to any applicable rules of international law, including those relating to environmental impact assessment,\" in a commentary that names stratospheric aerosols.27",
"runs": [
{
"text": "Three instruments bear on stratospheric aerosols more directly than those usually invoked. The Vienna Convention for the Protection of the Ozone Layer obliges parties to protect against \"adverse effects resulting or likely to result from human activities which modify or are likely to modify the ozone layer\" and to control such activities under their jurisdiction; sulfate injection is an activity the ozone assessment treats as capable of affecting stratospheric ozone."
@@ -568,7 +582,7 @@
"footnoteRef": "25"
},
{
- "text": " The Framework Convention on Climate Change commits parties to \"employ appropriate methods, for example impact assessments,\" to minimize the adverse effects \"of projects or measures undertaken by them to mitigate or adapt to climate change.\""
+ "text": " The United Nations Framework Convention on Climate Change (UNFCCC) commits parties to \"employ appropriate methods, for example impact assessments,\" to minimize the adverse effects \"of projects or measures undertaken by them to mitigate or adapt to climate change.\""
},
{
"text": "26",
@@ -1141,15 +1155,6 @@
"text": "A comparative justification duty would make alternatives and distributional choices open to consequential challenge. A continuity duty would require sponsors to confront the obligations created by sustained intervention, while separating changes in political control from dangerous physical interruption. The law already knows how to keep an indispensable system running while control over it changes hands, and the continuity duty asks that the same be arranged for the stratosphere before anyone depends on it. Together, these duties support governed research and withholding deployment authorization where credible comparison, representation, and continuity remain absent. They also require scrutiny of the harms of restraint. Those exposed to the intervention would have no practical means of withdrawing from its effects."
}
]
- },
- {
- "type": "paragraph",
- "text": "---",
- "runs": [
- {
- "text": "---"
- }
- ]
}
],
"footnotes": [
@@ -1178,7 +1183,14 @@
"italic": true
},
{
- "text": ", in "
+ "text": ", "
+ },
+ {
+ "text": "in",
+ "italic": true
+ },
+ {
+ "text": " "
},
{
"text": "Climate Change",
@@ -1198,7 +1210,7 @@
},
{
"number": 3,
- "text": "Bronislaw Szerszynski et al., Why Solar Radiation Management Geoengineering and Democracy Won't Mix, 45 Env't & Plan. A 2809 (2013); Joshua B. Horton et al., Solar Geoengineering and Democracy, 18 Glob. Envtl. Pol. 5 (2018).",
+ "text": "Bronislaw Szerszynski et al., Why Solar Radiation Management Geoengineering and Democracy Won't Mix, 45 Env't & Plan. A 2809 (2013); Joshua B. Horton et al., Solar Geoengineering and Democracy, 18 Glob. Env't Pol. 5 (2018).",
"runs": [
{
"text": "Bronislaw Szerszynski et al., "
@@ -1225,7 +1237,7 @@
"text": ", 18 "
},
{
- "text": "Glob. Envtl. Pol.",
+ "text": "Glob. Env't Pol.",
"smallCaps": true
},
{
@@ -1235,7 +1247,7 @@
},
{
"number": 4,
- "text": "John D. Graham & Jonathan B. Wiener, Confronting Risk Tradeoffs, in Risk vs. Risk: Tradeoffs in Protecting Health and the Environment 1 (John D. Graham & Jonathan B. Wiener eds., 1995); see also Tyler Felgenhauer et al., Solar Radiation Modification: A Risk-Risk Analysis 13-16 & tbls. 1-2 (2022), https://www.c2g2.net/wp-content/uploads/202203-C2G-RR-Full.pdf.",
+ "text": "John D. Graham & Jonathan B. Wiener, Confronting Risk Tradeoffs, in Risk vs. Risk: Tradeoffs in Protecting Health and the Environment 1 (John D. Graham & Jonathan B. Wiener eds., 1995).",
"runs": [
{
"text": "John D. Graham & Jonathan B. Wiener, "
@@ -1245,21 +1257,21 @@
"italic": true
},
{
- "text": ", in "
+ "text": ", "
},
{
- "text": "Risk vs. Risk: Tradeoffs in Protecting Health and the Environment",
- "smallCaps": true
+ "text": "in",
+ "italic": true
},
{
- "text": " 1 (John D. Graham & Jonathan B. Wiener eds., 1995); see also "
+ "text": " "
},
{
- "text": "Tyler Felgenhauer et al., Solar Radiation Modification: A Risk-Risk Analysis",
+ "text": "Risk vs. Risk: Tradeoffs in Protecting Health and the Environment",
"smallCaps": true
},
{
- "text": " 13-16 & tbls. 1-2 (2022), https://www.c2g2.net/wp-content/uploads/202203-C2G-RR-Full.pdf."
+ "text": " 1 (John D. Graham & Jonathan B. Wiener eds., 1995)."
}
]
},
@@ -1303,7 +1315,7 @@
},
{
"number": 8,
- "text": "Jonathan B. Wiener, Disregard and Due Regard, 29 N.Y.U. Envtl. L.J. 437, 454, 460-64 (2021).",
+ "text": "Jonathan B. Wiener, Disregard and Due Regard, 29 N.Y.U. Env't L.J. 437, 454, 460-64 (2021).",
"runs": [
{
"text": "Jonathan B. Wiener, "
@@ -1316,7 +1328,7 @@
"text": ", 29 "
},
{
- "text": "N.Y.U. Envtl. L.J.",
+ "text": "N.Y.U. Env't L.J.",
"smallCaps": true
},
{
@@ -1326,7 +1338,7 @@
},
{
"number": 9,
- "text": "Jane A. Flegal & Aarti Gupta, Evoking Equity as a Rationale for Solar Geoengineering Research? Scrutinizing Emerging Expert Visions of Equity, 18 Int'l Envtl. Agreements 45 (2018).",
+ "text": "Jane A. Flegal & Aarti Gupta, Evoking Equity as a Rationale for Solar Geoengineering Research? Scrutinizing Emerging Expert Visions of Equity, 18 Int'l Env't Agreements 45 (2018).",
"runs": [
{
"text": "Jane A. Flegal & Aarti Gupta, "
@@ -1339,7 +1351,7 @@
"text": ", 18 "
},
{
- "text": "Int'l Envtl. Agreements",
+ "text": "Int'l Env't Agreements",
"smallCaps": true
},
{
@@ -1349,7 +1361,7 @@
},
{
"number": 10,
- "text": "Peter J. Irvine & David W. Keith, Halving Warming with Stratospheric Aerosol Geoengineering Moderates Policy-Relevant Climate Hazards, 15 Envtl. Rsch. Letters 044011, at 2-3, 8, 10 (2020).",
+ "text": "Peter J. Irvine & David W. Keith, Halving Warming with Stratospheric Aerosol Geoengineering Moderates Policy-Relevant Climate Hazards, 15 Env't Rsch. Letters 044011, at 2-3, 8, 10 (2020).",
"runs": [
{
"text": "Peter J. Irvine & David W. Keith, "
@@ -1362,7 +1374,7 @@
"text": ", 15 "
},
{
- "text": "Envtl. Rsch. Letters",
+ "text": "Env't Rsch. Letters",
"smallCaps": true
},
{
@@ -1418,14 +1430,21 @@
},
{
"number": 13,
- "text": "World Meteorological Organization, Executive Summary: Scientific Assessment of Ozone Depletion: 2022, ch. 6 summary (Global Atmosphere Watch Rep. No. 278, 2022).",
+ "text": "World Meteorological Organization, supra note 1, ch. 6 summary.",
"runs": [
{
- "text": "World Meteorological Organization, Executive Summary: Scientific Assessment of Ozone Depletion",
+ "text": "World Meteorological Organization",
"smallCaps": true
},
{
- "text": ": 2022, ch. 6 summary (Global Atmosphere Watch Rep. No. 278, 2022)."
+ "text": ", "
+ },
+ {
+ "text": "supra",
+ "italic": true
+ },
+ {
+ "text": " note 1, ch. 6 summary."
}
]
},
@@ -1499,7 +1518,11 @@
"text": "Fla. Stat. § 403.411(1) (2025); Act of June 20, 2025, ch. 2025-157, 2025 Fla. Laws (effective July 1, 2025).",
"runs": [
{
- "text": "Fla. Stat. § 403.411(1) (2025); Act of June 20, 2025, ch. 2025-157, 2025 Fla. Laws (effective July 1, 2025)."
+ "text": "Fla. Stat.",
+ "smallCaps": true
+ },
+ {
+ "text": " § 403.411(1) (2025); Act of June 20, 2025, ch. 2025-157, 2025 Fla. Laws (effective July 1, 2025)."
}
]
},
@@ -1548,10 +1571,10 @@
},
{
"number": 23,
- "text": "Obligations of States in Respect of Climate Change, Advisory Opinion (I.C.J. July 23, 2025).",
+ "text": "Obligations of States in Respect of Climate Change, Advisory Opinion, ¶ 138 (I.C.J. July 23, 2025).",
"runs": [
{
- "text": "Obligations of States in Respect of Climate Change, Advisory Opinion (I.C.J. July 23, 2025)."
+ "text": "Obligations of States in Respect of Climate Change, Advisory Opinion, ¶ 138 (I.C.J. July 23, 2025)."
}
]
},
@@ -1560,7 +1583,14 @@
"text": "Convention on the Prohibition of Military or Any Other Hostile Use of Environmental Modification Techniques arts. I-III, opened for signature May 18, 1977, 31 U.S.T. 333, 1108 U.N.T.S. 151.",
"runs": [
{
- "text": "Convention on the Prohibition of Military or Any Other Hostile Use of Environmental Modification Techniques arts. I-III, opened for signature May 18, 1977, 31 U.S.T. 333, 1108 U.N.T.S. 151."
+ "text": "Convention on the Prohibition of Military or Any Other Hostile Use of Environmental Modification Techniques arts. I-III, "
+ },
+ {
+ "text": "opened for signature",
+ "italic": true
+ },
+ {
+ "text": " May 18, 1977, 31 U.S.T. 333, 1108 U.N.T.S. 151."
}
]
},
@@ -1608,7 +1638,14 @@
"text": "Int'l Law Comm'n, Draft Guidelines on the Protection of the Atmosphere, guideline 7 & cmt. ¶¶ (3)-(4), in Rep. of the Int'l Law Comm'n on the Work of Its Seventy-Second Session, U.N. Doc. A/76/10 (2021).",
"runs": [
{
- "text": "Int'l Law Comm'n, Draft Guidelines on the Protection of the Atmosphere, guideline 7 & cmt. ¶¶ (3)-(4), in Rep. of the Int'l Law Comm'n on the Work of Its Seventy-Second Session, U.N. Doc. A/76/10 (2021)."
+ "text": "Int'l Law Comm'n, Draft Guidelines on the Protection of the Atmosphere, guideline 7 & cmt. ¶¶ (3)-(4), "
+ },
+ {
+ "text": "in",
+ "italic": true
+ },
+ {
+ "text": " Rep. of the Int'l Law Comm'n on the Work of Its Seventy-Second Session, U.N. Doc. A/76/10 (2021)."
}
]
},
@@ -1617,7 +1654,14 @@
"text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities arts. 9(3), 10(d)-(e), in Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001).",
"runs": [
{
- "text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities arts. 9(3), 10(d)-(e), in Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001)."
+ "text": "Int'l Law Comm'n, Draft Articles on Prevention of Transboundary Harm from Hazardous Activities arts. 9(3), 10(d)-(e), "
+ },
+ {
+ "text": "in",
+ "italic": true
+ },
+ {
+ "text": " Rep. of the Int'l Law Comm'n on the Work of Its Fifty-Third Session, U.N. Doc. A/56/10 (2001)."
}
]
},
@@ -1656,14 +1700,10 @@
},
{
"number": 31,
- "text": "Pulp Mills, 2010 I.C.J. at 71, ¶ 164.",
+ "text": "Pulp Mills on the River Uruguay (Arg. v. Uru.), Judgment, 2010 I.C.J. 14, ¶ 164 (Apr. 20).",
"runs": [
{
- "text": "Pulp Mills",
- "italic": true
- },
- {
- "text": ", 2010 I.C.J. at 71, ¶ 164."
+ "text": "Pulp Mills on the River Uruguay (Arg. v. Uru.), Judgment, 2010 I.C.J. 14, ¶ 164 (Apr. 20)."
}
]
},
@@ -1842,7 +1882,7 @@
},
{
"number": 43,
- "text": "Parker & Irvine, supra note 37, at 456-67; Katharine L. Ricke, Juan B. Moreno-Cruz & Ken Caldeira, Strategic Incentives for Climate Geoengineering Coalitions to Exclude Broad Participation, 8 Envtl. Rsch. Letters 014021 (2013).",
+ "text": "Parker & Irvine, supra note 37, at 456-67; Katharine L. Ricke, Juan B. Moreno-Cruz & Ken Caldeira, Strategic Incentives for Climate Geoengineering Coalitions to Exclude Broad Participation, 8 Env't Rsch. Letters 014021 (2013).",
"runs": [
{
"text": "Parker & Irvine, "
@@ -1862,7 +1902,7 @@
"text": ", 8 "
},
{
- "text": "Envtl. Rsch. Letters",
+ "text": "Env't Rsch. Letters",
"smallCaps": true
},
{
@@ -1911,10 +1951,10 @@
},
{
"number": 46,
- "text": "Certain Activities Carried Out by Nicaragua in the Border Area (Costa Rica v. Nicar.) & Construction of a Road in Costa Rica Along the San Juan River (Nicar. v. Costa Rica), Judgment, 2015 I.C.J. 665, ¶¶ 157-58 (Dec. 16).",
+ "text": "Certain Activities Carried Out by Nicaragua in the Border Area (Costa Rica v. Nicar.) & Construction of a Road in Costa Rica Along the San Juan River (Nicar. v. Costa Rica), Judgment, 2015 I.C.J. 665, ¶¶ 157-59 (Dec. 16).",
"runs": [
{
- "text": "Certain Activities Carried Out by Nicaragua in the Border Area (Costa Rica v. Nicar.) & Construction of a Road in Costa Rica Along the San Juan River (Nicar. v. Costa Rica), Judgment, 2015 I.C.J. 665, ¶¶ 157-58 (Dec. 16)."
+ "text": "Certain Activities Carried Out by Nicaragua in the Border Area (Costa Rica v. Nicar.) & Construction of a Road in Costa Rica Along the San Juan River (Nicar. v. Costa Rica), Judgment, 2015 I.C.J. 665, ¶¶ 157-59 (Dec. 16)."
}
]
},
diff --git a/website/content/protected-sources.json b/website/content/protected-sources.json
index 9de19a4..84fbe2a 100644
--- a/website/content/protected-sources.json
+++ b/website/content/protected-sources.json
@@ -8,35 +8,35 @@
"app/components/experience/ExperienceDetails.tsx": "ba9aeb75ea9459a1c23d3b3315f6c855a154444946498de8c942160a4cbfa778",
"app/lib/content/experience-showcase.ts": "cf8e2a4aa2b7b4838065f99abfa403e19e0b9eb3a8e887cf0ccacda5057e1b8e",
"app/lib/content/experience-domains.ts": "7362085d4857daa83ae476f176163378dde550572d5a1932d2f142d4d1a00c7d",
- "content/experience-showcase.json": "a65f24f7705840d4d028e92b443371b778637b0cad067bc0a9fca51593ee4ffc",
- "app/components/AboutContent.tsx": "bfe39cff698c149b79ae0a1a439fe4970969e7672afafbbe67e80175dbd1f8a4",
+ "content/experience-showcase.json": "2daed72509277574db1786b465bce7dc1d5195d1b121c54b089d7c33e2a47d81",
+ "app/components/AboutContent.tsx": "e0e699c13b2b841795bae735dfbd2d4695370e041c64f695c72ad5408d5d6e39",
"app/lib/content/about.ts": "8d79e72133dcbcae9523771a099f53f0aecb879cea384769ff6bb0144e05626f",
"app/coursework/page.tsx": "a1caf4bd363d652ce401c237e06336be26e15296a3ffa19890b3d1c53b6ab63c",
"app/components/research/ResearchArticle.tsx": "edfb9fdf292670a641d3621fdf44718d34748d3891b8114521637f41ff7672cf",
"app/components/research/ResearchInline.tsx": "2fac0f2ffce4680d89f435be0a0a9923b6983b6dedfa94ec47a4fe687e76c2eb",
"app/components/research/ResearchMath.tsx": "c4ef79f9b23501070f1d1d360982276bc1f4c0ed4e094da979bb1ccc4e9c0d80",
"app/lib/content/research-markdown.ts": "abf3b2b65b4c8be391d3642acdec618794d47d717805771af720abe2dcc1f7f7",
- "app/lib/content/research-math.ts": "d5afb61eeceb4da8445abf2a3a0c33d23e954b81495ac4570c957b3ace94b02c",
+ "app/lib/content/research-math.ts": "866b47ab6199e2f8a7f8516abae15ccc9701313203a666e6b96d28d5e0bafd11",
"app/lib/content/research-number.ts": "0422523c87a5bd67914d6ead0a2411c93e3ced16e2d464f83052645fea5135e7",
"app/[paper]/page.tsx": "118091a0ca8862bba6d3e0baaaaee0c1af0516c9b457a1e893ac7cbc10980a85",
- "app/lib/content/selected-research.ts": "eaf4475432f5f127bec6ec663361a001db650dbb1aa47f1abcf603b9c3f5f5de",
- "content/selected-research/certified-valuation-of-arithmetic-asian-options.json": "e03c4321683b4f06f884af8ca9f281878d890c4290ed996f8b79794de9c7e98c",
- "content/selected-research/certified-joint-pricing-errors-in-rough-heston.json": "5a9cbf7b166b3afb6b438a7795a648ff03b13fc3c5dd92351e68153964e4697c",
- "content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json": "a5e2f46f62766f06dd1c9f436a7a3f654d9e76689103cc19cfb67cb749417c2a",
+ "app/lib/content/selected-research.ts": "fd9df5d39d2b324f0d8f0e7c17527156e05cb5ee6df0b40564cdff0ee7e6f188",
+ "content/selected-research/certified-valuation-of-arithmetic-asian-options.json": "4b0dcfaef7305564f93a7b7f73c746423e89f5c7468bb62575e9f71b3840c669",
+ "content/selected-research/certified-joint-pricing-errors-in-rough-heston.json": "63c829b11f513195e04e7c429311aaafd4c943e599f049127cf3f4f0400ce1f5",
+ "content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json": "45a7f37dc3cd2363279a542ac2a33ab672006cbfc3b6ab0467aa57a286d8f711",
"app/lib/content/legal-papers.ts": "38fc786859b5248e02de09d3fb4fe3a4635ec0b99f3d3ac2ecefb10402b5af4a",
"app/lib/content/metadata-excerpt.ts": "567139663925462cc48584a53dabb91e50e3727acd68afe7ddf61fa2818f0e71",
"app/components/legal/LegalPaperPage.tsx": "e4ff1a4c3b8ee3ffa84dcb4f86180859c510b8567664d027c2e1e7de845fa1bf",
- "content/legal-papers/autonomous-authority-in-space.json": "252d3c26d0f242a9b6094b251da6556b88344497e16b85e87cf85d181524bbcc",
- "content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json": "249397c03626b5637cc7eb06b5320a049af679a10e2b1832110c4e4df0accb18",
- "content/legal-papers/who-may-choose-the-lesser-risk.json": "8af5525fc9ef5ae1a76178cbee3b60af343d25c23fa37a9e5f1024d23a434139",
- "content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json": "98821daabe8929abfab6f9386832558c7c9f69ce9e44efac3614f442bdd83e6e",
+ "content/legal-papers/autonomous-authority-in-space.json": "e7f2afefb89db66af6d946e535981bf1f9a5348149177cc4f9e332d929a5f6fd",
+ "content/legal-papers/small-states-and-the-governance-of-strategic-space-dependence.json": "0371585e25ced8dcc4a4311a76d72bfd3e0b6142b9b15f4943c8e8d2c7d631c0",
+ "content/legal-papers/who-may-choose-the-lesser-risk.json": "92e847cf32ce4592bc86e1aa2d3710f48e8b1df404ee485be3080c306c3bea24",
+ "content/legal-papers/mangrove-restoration-and-the-limits-of-compensatory-mitigation.json": "24a7e7383e914992e1945396cca76d9982431b446426a5c9ece734148c89b07d",
"app/lib/content/ai-projects.ts": "f7562f3b2245db8d13c7d1df88596acb4c6d9c5f4af644c2fa10d10d8a35d383",
"app/lib/content/routes.ts": "566910e8fe1146050029cd0c3da486dda04df8b499f5f934e3c4ea7d042e8a86",
- "content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json": "f6283889a9459804cdec523ebad9f02187945e3d9505913166918e27f99bb135",
- "content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json": "8c8da6df7abd928f75af958d4972d83b1b496974e9229da50e79e70ae0db9578",
- "content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json": "ec78333f774103e8d46c3b24d7fff9a8e6fcf9a01e17a916240a38eb5b31ed1c",
- "content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json": "cd84433c6512dd63009c3acd73d33c7be6fea794b351172b302965b3c3517703",
- "content/sources/ai-portfolio/ai-research-and-engineering.md": "337a665b87865c723552fcb25da4a5492bc811d5cd4b01b36cc34fa10d718bf1",
+ "content/artificial-intelligence/large-scale-domain-expert-feedback-governance-and-adaptive-review.json": "130a0d8bee07315b7c27180285092367a1823e5feb09148d8b41069472e9fe38",
+ "content/artificial-intelligence/financial-task-design-and-synthetic-data-engineering.json": "a0a7779c57c6f3368d8f33c4b77b63c0259041b5c792fa441e69a25ee7fecb1f",
+ "content/artificial-intelligence/general-reasoning-evaluation-and-layered-verification.json": "e6cb94ceaf7166dcd288b1a9aa919ba67ab83b50baee94d4e13ec049d14d4297",
+ "content/artificial-intelligence/evidence-verification-and-robust-comparison-in-investment-research.json": "5111408d7a424527e376e616482273b7eee7d1f91a7a5f49a9d7e203ba81abca",
+ "content/sources/ai-portfolio/ai-research-and-engineering.md": "be8f32d6ed3d2975e4ed53cde38db23fdc45b231618ec337cbf555c2acddf9c0",
"scripts/import-ai-portfolio.mjs": "4a65b406d663f35ede71a4625822f44027c84161e62e34fd9e269e9c11dcff21"
}
}
diff --git a/website/content/selected-research/certified-joint-pricing-errors-in-rough-heston.json b/website/content/selected-research/certified-joint-pricing-errors-in-rough-heston.json
index 9ed977a..abec202 100644
--- a/website/content/selected-research/certified-joint-pricing-errors-in-rough-heston.json
+++ b/website/content/selected-research/certified-joint-pricing-errors-in-rough-heston.json
@@ -2,10 +2,10 @@
"slug": "certified-joint-pricing-errors-in-rough-heston",
"title": "Certified Joint Pricing Errors in Rough Heston",
"description": "Numerical pricing error as a common factor across strikes: a Riccati residual priced by forward variance in rough Heston, the netting of portfolio errors in strike space, and certified tolerances for S&P 500 option positions.",
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/manuscript/merged-heston-en.md",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/manuscript/merged-heston-en.md",
"repositoryUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston",
- "sourceEditedAt": "2026-11-19T03:44:46-05:00",
- "sourceCommit": "46c85b4b256dfbc597ed56b3b2b526681334e60f",
- "sourceBlob": "27bd396d79cfe3d99072f23d9799641c898442e7",
- "markdown": "## Abstract\n\nEvery option on one underlying and one expiry is priced from a single characteristic function, so a numerical error in that function is a common factor across strikes: a position hedged in strike space is hedged against it, and a position spread over many strikes diversifies it. We make this exact and build on it a deterministic certificate, a bound on the distance between a stored vector of prices and the exact model prices in any portfolio direction. Two results carry the construction. The first holds for every model priced by Fourier inversion: the exposure of a portfolio to the transform error at a given frequency is the Fourier transform of its payoff, so that low-frequency error cancels in spreads and butterflies while high-frequency error averages out across strikes at a rate set by the Herfindahl index of the position. The second bounds the transform error in the rough Heston model: a residual in the fractional Riccati equation moves the log-characteristic function by at most the residual integrated against a forward variance curve, the curve of the same model under the measure midway between the risk-neutral and share measures. The constant is one, the leverage effect supplies the damping, the hypothesis on the curve is the condition under which the model is known to exist, and the argument carries over to Volterra Heston models with completely monotone kernels. On S&P 500 quotes of 18 June 2021, the six-month 4400–4500 call spread, quoted about 2.5 index points wide, is certified to 0.115 points. At three months the same spread is certified to 0.233 points when its legs are netted and to 0.252 when they are not, with every input identical, and the two bounds fall on opposite sides of a quarter-point tolerance.\n\n**Keywords:** rough volatility; Volterra Heston models; fractional Riccati equation; Fourier pricing; a posteriori error bounds; model risk; netting.\n\n**JEL classification:** C63, G13, G12.\n\n**MSC 2020:** 91G20, 91G60, 65G20, 45D05, 26A33.\n\n## 1. Introduction\n\nThe characteristic function of the Heston model is known in closed form, and for more than a decade one of its two standard expressions carried a flaw that went largely unnoticed. Written as in Heston (1993), the formula crosses a branch cut of the complex logarithm at long maturities, and an implementation on the principal branch returns a discontinuous transform (Kahl and Jäckel 2005; Albrecher, Mayer, Schoutens and Tistaert 2007). The error entered every strike of a maturity together, because every strike is an integral of the same function. A second episode shows the other side of the same fact. The expansion of Hagan, Kumar, Lesniewski and Woodward (2002) for the SABR model tracks the exact implied volatility closely, strike by strike, yet at low strikes and long expiries the density it implies turns negative: a butterfly, which is a portfolio of three options, receives a negative price (Hagan, Kumar, Lesniewski and Woodward 2014). Accuracy in each price had not delivered accuracy in a direction.\n\nBoth episodes say that the accuracy of an option pricer is a property of directions in strike space. This paper makes that statement quantitative. Its portfolio results hold for every pricer built on a common transform. Its bound on the transform itself is proved for the rough Heston model of El Euch and Rosenbaum (2019), whose transform solves a fractional Riccati equation with no closed form, so that the prices used in practice come from approximations: the rational approximation of Gatheral and Radoičić (2019), fractional Adams schemes (Diethelm, Ford and Freed 2004), Markovian lifts (Abi Jaber and El Euch 2019b; Abi Jaber 2019) and neural-network surrogates (Horvath, Muguruza and Tomas 2021). Their accuracy is usually reported by comparison with a slower solver on a test set. Agreement between two solvers bounds neither.\n\nWe ask what can be proved after the fact. A vector \\(c^{\\rm fast}\\) of rough Heston prices has been computed and stored. Let \\(c^*\\) be the exact model prices of the same contracts, normalised by the forward \\(F_T\\) and the discount factor \\(D\\). For a portfolio \\(w\\) and a monetary tolerance \\(\\tau\\), the task is to prove\n\\[\nDF_T\\left|w^\\top(c^*-c^{\\rm fast})\\right|\\le\\tau.\n\\tag{1.1}\n\\]\nThe stored output is taken as given. An independent reference \\(\\bar c\\) is built with its own error bound, and the stored output is located relative to it:\n\\[\nc^*-c^{\\rm fast}=d+\\operatorname{Re}(Az)+r,\\qquad d=\\bar c-c^{\\rm fast}.\n\\tag{1.2}\n\\]\nHere \\(z_n\\) is the error of the reference transform at the frequency \\(u_n\\), one complex number shared by all strikes; \\(A\\) is the matrix of known pricing coefficients; and \\(r\\) collects quadrature, the infinite tail and arithmetic. Two questions follow. How large can \\(z_n\\) be, and how should the bounds on the \\(z_n\\) be combined for a portfolio?\n\n### 1.1. What a Riccati residual costs\n\nA reference trajectory \\(\\widehat h\\) fails to solve the Riccati equation by a residual \\(r(t)\\), which can be bounded on every time interval by exact arithmetic. The first result converts that residual into an error of the log-characteristic function \\(L_T\\). For a reference in the left half-plane and an admissible forward variance curve,\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\int_0^T\\xi^{\\lambda}(T-t)\\,|r(t)|\\,dt\\le\\int_0^T\\xi(T-t)\\,|r(t)|\\,dt,\n\\qquad \\lambda=\\tfrac12|\\rho|\\nu .\n\\tag{1.3}\n\\]\nThe weight \\(\\xi^\\lambda\\) is a forward variance curve: it is the expected variance of the same model under the measure whose Girsanov kernel is half that of the share measure, which is the measure the Lewis contour sees (Theorem 3.1 and Proposition 3.2). Three features of (1.3) deserve emphasis. The constant is one; there is no Grönwall exponential, although the equation is nonlinear and carries its whole history. The damping rate \\(\\lambda\\) is the amount by which the leverage effect speeds up mean reversion under that measure, so the feature of the model that produces the equity skew is also the one that makes numerical errors contract. And the single hypothesis on the curve, nonnegativity of its Riemann–Liouville derivative, is the condition under which El Euch and Rosenbaum (2018) and Abi Jaber and El Euch (2019a) construct the model with nonnegative variance. The certificate asks nothing of the curve that the model does not already ask.\n\nThe bound is the rough counterpart of variation of constants for the Heston Riccati equation. A fractional equation has no semigroup to restart, and two classical tools take its place: convexity of the Caputo derivative on a modulus and positivity of the scalar fractional resolvent. On the six-month example of Section 7, replacing a uniform bound on the state by (1.3) lowers the complete error bound of a call spread from 1.398 to 0.379 index points, a reduction of 72.91 per cent, with the stored prices and the reference unchanged. The same proof covers every damped contour and Volterra Heston models with completely monotone kernels, the classical and the lifted Heston models among them (Remarks 3.5 and 3.6).\n\n### 1.2. Transform error as a common factor\n\nOnce each \\(z_n\\) lies in a disk of radius \\(\\epsilon_n\\), the set of price errors compatible with (1.2) is a sum of ellipses, one for each frequency, and its support in the direction \\(w\\) can be written down exactly (Theorem 5.1):\n\\[\n\\sup\\,w^\\top\\operatorname{Re}(Az)=\\frac{h_*}\\pi\\sum_n\\omega_n\\,\\epsilon_n\\,\\frac{|\\widehat w(u_n)|}{u_n^2+\\frac14},\n\\qquad\n\\widehat w(u)=\\sum_iw_i\\sqrt{m_i}\\,e^{-iuk_i},\n\\tag{1.4}\n\\]\nwhere \\(k_i=\\log m_i\\) is log-moneyness, \\(h_*\\) is the grid step and \\(\\omega_n\\) are the trapezoidal weights. The function \\(\\widehat w\\) is the Fourier transform of the payoff of the portfolio. A portfolio of options with a common expiry is one payoff, and (1.4) prices its error as one payoff. Bounding each leg separately replaces \\(|\\widehat w(u)|\\) by \\(\\sum_i|w_i|\\sqrt{m_i}\\), and the ratio of the two, which we call the netting ratio, measures what is lost.\n\nThe netting ratio has a simple anatomy. For a vertical spread with log-strike gap \\(\\Delta\\) it is \\(|\\sin(u\\Delta/2)|\\) up to terms of order \\(\\Delta\\): the spread is hedged against transform error of wavelength longer than twice its strike spacing and fully exposed at the Nyquist frequency \\(\\pi/\\Delta\\). For a butterfly the ratio is squared. For any portfolio its mean square over a long band of frequencies is the Herfindahl index of the position (Propositions 5.2 and 5.3). Transform error therefore behaves like a factor model in frequency. Low-frequency error is systematic across strikes and is removed by hedging. High-frequency error is idiosyncratic across strikes and is removed by diversification. On the S&P 500 example the eleven adjacent call spreads keep on average 0.708 of their leg-by-leg node error, against a Herfindahl benchmark of \\(1/\\sqrt2\\approx0.707\\); an equal-weight basket of all twelve calls keeps 0.076.\n\nThis part of the paper is model-free. It applies to any pricer that obtains a strip of options from one transform, whatever the model and whatever the method that produced the transform.\n\n### 1.3. Evidence and economic scale\n\nSection 7 applies the certificate to S&P 500 index options quoted on 18 June 2021 (Jeng and Kiliçman 2021), with the index at 4221.86. An SPX option has a multiplier of 100 dollars per index point and a minimum tick of 0.05 points below a premium of 3.00 and 0.10 above (Cboe Global Markets 2024); in 2025 an average of 3.9 million contracts changed hands each day (Cboe Global Markets 2026). On the sample date the six-month calls between 3700 and 4800 were quoted between 0.8 and 2.5 points wide. A tolerance of a quarter point is 25 dollars a contract and about one fifth of the quoted width of the 4400 call. Our tolerances of 1, 1/2 and 1/4 point thus ask whether numerical error is safely inside the bid–ask spread.\n\nThe answers are as follows. The stored rational-approximation output for the six-month 4400–4500 call spread is certified to 0.115 points, against a quoted width of about 2.5. At three months the netted bound is 0.233 points and the leg-by-leg bound 0.252, from the same stored output, reference centre, 1,025 node radii and remainders; only netting certifies a quarter point. Across 28 spreads, butterflies and baskets at six months, a half-point tolerance is certified for all 28 with netting and for 13 without. Five neighbouring roughness candidates, from \\(H=0.02\\) to \\(H=0.05\\), are ranked by their least-squares distance to the quotes with certified separation in all ten pairs, against seven of ten without netting, so the ranking belongs to the model and the quotes. And at the best candidate the model price of the 4400 call lies at least 0.49 points above the quoted ask: a discrepancy between model and market that the certificate attributes to the model.\n\nThe setting is one that supervisors already describe in portfolio terms. European prudent-valuation rules require a model-risk valuation adjustment for each valuation model and aggregate the individual adjustments with a flat factor of 50 per cent (Commission Delegated Regulation (EU) 2016/101, Article 11 and Annex). The Basel market-risk framework tests, desk by desk, whether the prices of the risk engine track those of the front office, with the stated aim to \"measure the materiality of simplifications\" in the former (Basel Committee on Banking Supervision 2019, MAR32.20). United States guidance asks that model risk be assessed \"both individually and in aggregate\", where the aggregate reflects reliance on common methodologies (Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency 2026). A shared transform is a common methodology in the most literal sense. For the numerical component of valuation uncertainty, (1.4) replaces a flat diversification factor by the exact one.\n\n### 1.4. Relation to the literature\n\nThe rough Heston model arises as the scaling limit of Hawkes-based microstructure models (El Euch, Fukasawa and Rosenbaum 2018) and belongs to the affine Volterra class (Abi Jaber, Larsson and Pulido 2019); its forward variance form is due to El Euch and Rosenbaum (2018) and Gatheral and Keller-Ressel (2019). The empirical case for rough volatility, and the debate around it, rests on precise pricing at the short end, where numerical methods are under most strain (Gatheral, Jaisson and Rosenbaum 2018; Bayer, Friz and Gatheral 2016; Rømer 2022; Cont and Das 2024).\n\nGatheral and Radoičić (2019, 2024) derive rational approximations by matching the short- and long-time behaviour of the fractional Riccati solution, and Jeng and Kiliçman (2020, 2021) study global Padé approximations and their calibration to SPX. The stored output we audit is the third-order approximation of Gatheral and Radoičić; Section 6 records a structural property of it on the parameter set we use. Bayer and Breneis (2023a, 2023b) study Markovian kernel approximations and weak simulation schemes. Boyarchenko, de Innocentis and Levendorskii (2025) study reliable rough Heston pricing and calibration, and their modified Adams method is one of our controls. Ben Hammouda, Ben Romdhane, Samet and Tempone (2026) allocate time-discretisation and quadrature errors in single- and multilevel methods, and Hager and Kreher (2026) expand the model in the Hurst parameter. These works bound or estimate the error of a method. Our object is the error of a given output, with every remainder charged, in a given portfolio direction.\n\nThe analytical tools are established. Caputo convexity is due to Li and Liu (2018), pointwise a posteriori control of fractional equations by positive resolvents to Kopteva (2022), complete monotonicity of Mittag-Leffler functions to Simon (2015), and the strip analysis of the trapezoidal rule to Trefethen and Weideman (2014). What we add is their composition along the valuation map, from the Riccati residual through the forward variance curve to the payoff transform, the identification of the resulting kernel as a forward variance curve under a tilted measure, and the frequency anatomy of portfolio error.\n\nSection 2 states the pricing problem. Section 3 proves the residual bound and Section 4 turns it into transform disks and price bounds. Section 5 treats portfolios. Section 6 describes the stored output and Section 7 the evidence. Appendices A–E contain the model verification, the complete proofs and the experimental record.\n\n## 2. The pricing problem\n\n### 2.1. Model, transform and price\n\nFix a maturity \\(T\\) and let \\(S\\) be the forward price for delivery at \\(T\\), so that \\(S_0=F_T\\). Under the pricing measure \\(\\mathbb Q\\),\n\\[\ndS_t=S_t\\sqrt{V_t}\\,dB_t,\\qquad\nV_t=\\xi(t)+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW_s,\\qquad\nd\\langle B,W\\rangle_t=\\rho\\,dt,\n\\tag{2.1}\n\\]\nwith \\(g_\\beta(t)=t^{\\beta-1}/\\Gamma(\\beta)\\), roughness index \\(\\alpha=H+\\tfrac12\\in(\\tfrac12,1)\\), volatility of variance \\(\\nu>0\\) and correlation \\(-1\\le\\rho\\le0\\). The function \\(\\xi(t)=\\mathbb E V_t\\) is the forward variance curve. In this form the model has no separate mean-reversion parameters: whatever drift the variance has is recorded in \\(\\xi\\), which is in principle observable from variance swaps (El Euch and Rosenbaum 2018; Gatheral and Keller-Ressel 2019).\n\nThroughout, convolutions start at zero, \\((f*g)(t)=\\int_0^tf(t-s)g(s)\\,ds\\), and\n\\[\nI^\\alpha f=g_\\alpha*f,\\qquad D^\\alpha v=g_{1-\\alpha}*v'\\quad(v\\in AC),\n\\tag{2.2}\n\\]\nare the fractional integral and the Caputo derivative. We work on the contour of Lewis (2001), on which the transform is \\(\\phi_T(u-\\tfrac i2)=\\mathbb E[(S_T/F_T)^{1/2+iu}]\\) for real \\(u\\). It is given by a fractional Riccati equation (El Euch and Rosenbaum 2019):\n\\[\nD^\\alpha h=F_u(h),\\quad h(0)=0,\\qquad\nF_u(h)=-\\tfrac12\\left(u^2+\\tfrac14\\right)+\\left(\\tfrac12\\rho\\nu+i\\rho\\nu u\\right)h+\\tfrac12\\nu^2h^2,\n\\tag{2.3}\n\\]\n\\[\n\\phi_T(u-\\tfrac i2)=e^{L_T(u)},\\qquad\nL_T(u)=\\int_0^T\\xi(T-t)\\,D^\\alpha h(t,u)\\,dt .\n\\tag{2.4}\n\\]\nWith \\(m=K/F_T\\), \\(k=\\log m\\) and \\(c=C/(DF_T)\\), the normalised price of a call struck at \\(K\\) is\n\\[\nc=1-\\frac{\\sqrt m}{\\pi}\\int_0^\\infty\\operatorname{Re}\\frac{e^{-iuk}\\,\\phi_T(u-\\tfrac i2)}{u^2+\\frac14}\\,du .\n\\tag{2.5}\n\\]\n\n**Proposition 2.1 (the exact Riccati trajectory).** Let \\(\\alpha\\in(\\tfrac12,1)\\), \\(-1\\le\\rho\\le0\\) and \\(u\\in\\mathbb R\\). Equation (2.3) has a unique global solution. It is absolutely continuous on every finite interval, continuously differentiable at positive times, and satisfies \\(\\operatorname{Re}h(t)<0\\) for \\(t>0\\). With \\(b_u=\\tfrac12(u^2+\\tfrac14)\\) and \\(\\lambda_*=\\tfrac12|\\rho|\\nu\\),\n\\[\n|h(t)|\\le\\frac{b_u}{\\lambda_*}\\left\\{1-E_\\alpha(-\\lambda_*t^\\alpha)\\right\\}\\le\\frac{b_u\\,t^\\alpha}{\\Gamma(1+\\alpha)},\n\\tag{2.6}\n\\]\nwhere \\(E_\\alpha\\) is the Mittag-Leffler function and the last bound remains valid for \\(\\rho=0\\).\n\n**Proof.** Local Volterra contraction gives existence and uniqueness. On the imaginary axis the real part of \\(F_u\\) is at most \\(-\\tfrac18\\) after completing a square, and the fractional maximum principle then prevents the trajectory from leaving the left half-plane. Convexity of a regularised modulus and comparison with the scalar linear equation give (2.6), which excludes blow-up, and the continuation keeps the preceding history. Appendix C.1 gives the details. \\(\\square\\)\n\nThe proposition identifies the object to be approximated without reference to any numerical scheme. Its two conclusions, the invariant half-plane and the scale \\(\\lambda_*\\), are the source of the stability used below.\n\n### 2.2. Two faces of the exponent\n\nThe exponent \\(L_T\\) has a second expression, and the pair is what makes residual control possible.\n\n**Proposition 2.2 (two faces).** Let \\(\\xi\\in AC[0,T]\\) and define\n\\[\nq=\\frac{d}{dt}\\left(g_{1-\\alpha}*\\xi\\right)=\\xi(0)\\,g_{1-\\alpha}+g_{1-\\alpha}*\\xi' .\n\\tag{2.7}\n\\]\nThen \\(q\\in L^1(0,T)\\), \\(g_\\alpha*q=\\xi\\), and for every \\(v\\in AC[0,T]\\) with \\(v(0)=0\\),\n\\[\n\\int_0^T\\xi(T-t)\\,D^\\alpha v(t)\\,dt=\\int_0^Tq(T-t)\\,v(t)\\,dt .\n\\tag{2.8}\n\\]\n\n**Proof.** By Fubini's theorem the left side is \\(((g_{1-\\alpha}*\\xi)*v')(T)\\). The function \\(g_{1-\\alpha}*\\xi\\) is absolutely continuous and vanishes at zero, as does \\(v\\), so integration by parts moves the derivative and gives \\((q*v)(T)\\). The identity \\(g_\\alpha*g_{1-\\alpha}=1\\) gives \\(g_\\alpha*q=\\xi\\). Appendix C.2 records the integrability bounds. \\(\\square\\)\n\nThe function \\(q\\) is the Riemann–Liouville derivative of the curve, and it has a direct meaning. If the variance is written with an explicit drift, \\(V_t=V_0+\\int_0^tg_\\alpha(t-s)\\theta(s)\\,ds+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW_s\\), then \\(\\xi=V_0+g_\\alpha*\\theta\\) and\n\\[\nq(t)=\\frac{V_0\\,t^{-\\alpha}}{\\Gamma(1-\\alpha)}+\\theta(t).\n\\tag{2.9}\n\\]\nThus \\(q\\) is the rate at which variance is supplied: the drift \\(\\theta\\), together with the term a power-law memory requires in order to hold the initial level \\(V_0\\) in place. Applied to the exact solution, (2.8) reads \\(L_T=\\int_0^T\\xi(T-t)F_u(h(t))\\,dt=\\int_0^Tq(T-t)h(t)\\,dt\\), and both forms appear in El Euch and Rosenbaum (2018, Corollaries 3.2 and 3.3). The first face says what the exponent is, the Riccati polynomial accumulated against forward variance. The second says how the exponent responds to an error in the state: linearly, with weight \\(q\\).\n\n**Assumption 2.3 (admissible curve).** \\(\\xi\\in AC[0,T]\\), \\(\\xi(0)\\ge0\\), and \\(q\\ge0\\) almost everywhere on \\((0,T)\\).\n\nUnder this assumption \\(\\xi=g_\\alpha*q\\ge0\\). The assumption is the model's own. In the notation of (2.9) it says that \\(\\theta(t)\\,dt+V_0\\,t^{-\\alpha}dt/\\Gamma(1-\\alpha)\\) is a nonnegative measure, which is the condition imposed by El Euch and Rosenbaum (2018, condition (4)) and by Abi Jaber and El Euch (2019a, Example 2.2) to construct the model with nonnegative variance. Every nondecreasing curve with \\(\\xi(0)\\ge0\\) is admissible. A decreasing curve is admissible as long as it does not fall faster than the memory of the kernel allows: \\(\\xi(t)=V_0(1-ct)\\) is admissible on \\([0,T]\\) exactly when \\(cT\\le1-\\alpha\\) (Corollary C.4). Appendix A verifies the hypotheses of the existence and transform theorems for the curve used in Section 7.\n\n## 3. What a Riccati residual costs\n\n### 3.1. A martingale that fails to be one\n\nLet \\(\\widehat h\\in AC[0,T]\\) with \\(\\widehat h(0)=0\\) be an approximate solution of (2.3), and let\n\\[\nr=D^\\alpha\\widehat h-F_u(\\widehat h)\n\\tag{3.1}\n\\]\nbe its residual. A formal computation shows what \\(r\\) should cost. Write \\(\\xi_t(s)=\\mathbb E[V_s\\mid\\mathcal F_t]\\) for the forward variance curve seen at time \\(t\\), so that \\(d\\xi_t(s)=\\nu g_\\alpha(s-t)\\sqrt{V_t}\\,dW_t\\), put \\(X_t=\\log(S_t/S_0)\\), and consider\n\\[\n\\widehat M_t=\\exp\\left\\{\\left(\\tfrac12+iu\\right)X_t+\\int_t^TD^\\alpha\\widehat h(T-s)\\,\\xi_t(s)\\,ds\\right\\}.\n\\]\nAt \\(t=T\\) this is \\((S_T/S_0)^{1/2+iu}\\) and at \\(t=0\\) it is \\(e^{\\widehat L_T}\\), where \\(\\widehat L_T=\\int_0^T\\xi(T-t)D^\\alpha\\widehat h(t)\\,dt\\). Itô's formula and the identity \\(\\int_t^TD^\\alpha\\widehat h(T-s)\\,g_\\alpha(s-t)\\,ds=\\widehat h(T-t)\\) give\n\\[\nd\\widehat M_t=-\\,r(T-t)\\,V_t\\,\\widehat M_t\\,dt+dN_t\n\\]\nwith \\(N\\) a local martingale, and taking expectations,\n\\[\n\\phi_T(u-\\tfrac i2)-e^{\\widehat L_T}=-\\int_0^Tr(T-t)\\,\\mathbb E\\left[V_t\\widehat M_t\\right]dt .\n\\tag{3.2}\n\\]\nFor the exact solution \\(r=0\\), and this is the usual verification of (2.4). For an approximation the would-be martingale drifts, and it drifts in proportion to variance, because per unit of variance the Riccati polynomial is the generator of the pair \\((X,\\xi)\\) acting on exponentials. A residual at time to maturity \\(T-t\\) therefore costs the variance expected at calendar time \\(t\\), weighted by \\(\\widehat M_t\\). Since \\(|\\widehat M_t|\\) carries the factor \\((S_t/S_0)^{1/2}\\), the weighting is that of a measure tilted halfway towards the share measure.\n\nIdentity (3.2) is the heuristic behind the theorem that follows, which proves the same conclusion for the exponent by deterministic means and with a weight that depends only on the model.\n\n### 3.2. The residual bound\n\nFor \\(\\lambda\\ge0\\) let \\(k_\\lambda(t)=t^{\\alpha-1}E_{\\alpha,\\alpha}(-\\lambda t^\\alpha)\\) be the resolvent kernel of the scalar fractional equation, the solution of \\(k_\\lambda+\\lambda\\,g_\\alpha*k_\\lambda=g_\\alpha\\), with \\(k_0=g_\\alpha\\). For an admissible curve define\n\\[\n\\xi^\\lambda=k_\\lambda*q,\\qquad\\text{equivalently}\\qquad\\xi^\\lambda+\\lambda\\,g_\\alpha*\\xi^\\lambda=\\xi .\n\\tag{3.3}\n\\]\n\n**Theorem 3.1 (a residual is priced by forward variance).** Let \\(\\alpha\\in(\\tfrac12,1)\\), \\(-1\\le\\rho\\le0\\), \\(u\\in\\mathbb R\\), and let \\(\\xi\\) satisfy Assumption 2.3. Let \\(h\\) be the solution of (2.3) and let \\(\\widehat h\\in AC[0,T]\\) satisfy \\(\\widehat h(0)=0\\) and \\(\\operatorname{Re}\\widehat h\\le\\varepsilon\\) on \\([0,T]\\), with \\(\\lambda=\\tfrac12|\\rho|\\nu-\\tfrac12\\nu^2\\varepsilon\\ge0\\). Suppose the residual (3.1) satisfies \\(|r|\\le R\\) almost everywhere for some \\(R\\in L^\\infty(0,T)\\). Then, for \\(0\\le t\\le T\\),\n\\[\n|h(t)-\\widehat h(t)|\\le(k_\\lambda*R)(t),\n\\tag{3.4}\n\\]\nand the exponents \\(L_T=\\int_0^T\\xi(T-t)D^\\alpha h(t)\\,dt\\) and \\(\\widehat L_T=\\int_0^T\\xi(T-t)D^\\alpha\\widehat h(t)\\,dt\\) satisfy\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\int_0^T\\xi^\\lambda(T-t)\\,R(t)\\,dt\\le\\int_0^T\\xi(T-t)\\,R(t)\\,dt .\n\\tag{3.5}\n\\]\nIn particular, if \\(R\\le\\delta\\) then \\(|L_T-\\widehat L_T|\\le\\delta\\int_0^T\\xi(s)\\,ds\\).\n\n**Proof.** Put \\(e=h-\\widehat h\\). Because \\(F_u\\) is quadratic, subtraction gives an exact linear equation,\n\\[\nD^\\alpha e=\\left\\{\\tfrac12\\rho\\nu+i\\rho\\nu u+\\tfrac12\\nu^2(h+\\widehat h)\\right\\}e-r,\\qquad e(0)=0,\n\\tag{3.6}\n\\]\nand by Proposition 2.1 the coefficient has real part at most \\(-\\lambda\\). The Caputo derivative of a convex function of a path is at most the gradient applied to the Caputo derivative of the path. Applied to a regularised modulus this gives \\(D^\\alpha|e|+\\lambda|e|\\le|r|\\le R\\), and positivity of the resolvent \\(k_\\lambda\\) gives (3.4). By Proposition 2.2, \\(L_T-\\widehat L_T=\\int_0^Tq(T-t)e(t)\\,dt\\). Since \\(q\\ge0\\), (3.4) gives \\(|L_T-\\widehat L_T|\\le(q*k_\\lambda*R)(T)=(\\xi^\\lambda*R)(T)\\). Finally \\(g_\\alpha-k_\\lambda=\\lambda\\,g_\\alpha*k_\\lambda\\ge0\\), so \\(\\xi^\\lambda\\le g_\\alpha*q=\\xi\\). Appendix C.2 supplies the regularity, including the passage to the terminal time. \\(\\square\\)\n\nThe bound is global: it holds for residuals of any size and at every frequency. The exact solution enters through the sign of its real part alone, and the weight in (3.5) is fixed by the equation and the curve.\n\n### 3.3. The weight is a forward variance curve\n\nThe kernel \\(\\xi^\\lambda\\) was defined by a Volterra equation. It is also an expectation.\n\n**Proposition 3.2 (the kernel under a tilted measure).** Let \\(\\xi\\) be admissible and, for \\(\\eta\\in[0,1]\\), let \\(\\mathbb Q^\\eta\\) be the measure with density \\(\\mathcal E(\\eta\\int_0^{\\cdot}\\sqrt{V_s}\\,dB_s)_T\\) with respect to \\(\\mathbb Q\\). Under \\(\\mathbb Q^\\eta\\) the variance follows a rough Heston equation with the same supply \\(q\\), the same \\(\\nu\\) and mean-reversion speed \\(\\eta|\\rho|\\nu\\), and\n\\[\n\\mathbb E^{\\mathbb Q^\\eta}[V_t]=\\xi^{\\eta|\\rho|\\nu}(t),\\qquad0\\le t\\le T .\n\\tag{3.7}\n\\]\nIn particular the weight of Theorem 3.1 for a reference in the left half-plane is \\(\\xi^{\\lambda_*}(t)=\\mathbb E^{\\mathbb Q^{1/2}}[V_t]\\).\n\n**Proof.** The stochastic exponential is a true martingale (Abi Jaber, Larsson and Pulido 2019, Lemma 7.3, whose argument applies to an admissible input curve). By Girsanov's theorem \\(W^\\eta=W-\\eta\\rho\\int_0^{\\cdot}\\sqrt{V_s}\\,ds\\) is a Brownian motion under \\(\\mathbb Q^\\eta\\), and (2.1) becomes\n\\[\nV_t=\\xi(t)-\\eta|\\rho|\\nu\\int_0^tg_\\alpha(t-s)V_s\\,ds+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW^\\eta_s .\n\\]\nThe moment bounds for affine Volterra processes make the stochastic integral centred, so \\(m(t)=\\mathbb E^{\\mathbb Q^\\eta}[V_t]\\) solves \\(m+\\eta|\\rho|\\nu\\,g_\\alpha*m=\\xi\\), whose unique solution is \\(\\xi^{\\eta|\\rho|\\nu}\\) by (3.3). \\(\\square\\)\n\nThe measures \\(\\mathbb Q^\\eta\\) interpolate between the pricing measure at \\(\\eta=0\\) and the share measure, whose numéraire is the underlying, at \\(\\eta=1\\). Under the share measure the variance of an equity index reverts faster, by \\(|\\rho|\\nu\\): paths on which the index is high receive more weight, and on those paths the leverage effect makes variance low. This is the adjustment that separates the two probabilities in the formula of Heston (1993). The Lewis contour prices the claim \\(S_T^{1/2+iu}\\), whose modulus sits midway between the two numéraires, and it inherits half of the adjustment. Theorem 3.1 then reads as follows.\n\n*On the Lewis contour, a Riccati residual at time to maturity \\(s\\) changes the log-characteristic function by at most the variance expected at time \\(T-s\\) under \\(\\mathbb Q^{1/2}\\), per unit of residual.*\n\nFour remarks place the statement.\n\n**Remark 3.3 (the classical model).** At \\(\\alpha=1\\) the resolvent is \\(k_\\lambda(t)=e^{-\\lambda t}\\) and \\(\\xi^\\lambda(t)=V_0e^{-\\lambda t}+\\int_0^te^{-\\lambda(t-s)}\\theta(s)\\,ds\\). Inequality (3.5) is then variation of constants for the Heston Riccati equation, in which the error is transported by a flow that contracts at rate \\(\\lambda\\). For \\(\\alpha<1\\) there is no flow to transport it. Convexity of the Caputo derivative and positivity of \\(k_\\lambda\\) recover the same conclusion with the same constant.\n\n**Remark 3.4 (the constant).** The constant one in (3.5) is sharp. At \\(a=0\\), the point of zero frequency on the undamped contour, the exact solution vanishes, \\(\\lambda=0\\), and a residual of constant sign produces \\(\\widehat L_T-L_T=\\int_0^T\\xi(T-t)r(t)\\,dt\\) up to terms of second order in \\(r\\). On the Lewis contour the exact solution adds damping of its own through \\(\\operatorname{Re}h<0\\), growing with the frequency, and Theorem 3.1 uses the part that is uniform in the frequency.\n\n**Remark 3.5 (other contours and positive correlation).** On a contour \\(\\operatorname{Im}a=-\\eta\\) with \\(0<\\eta<1\\) the linear coefficient of the Riccati polynomial has real part \\(\\eta\\rho\\nu\\), the left half-plane is again invariant, and the proof gives (3.5) with \\(\\lambda=\\eta|\\rho|\\nu\\) and the measure \\(\\mathbb Q^\\eta\\). The damping available to an error bound thus rises linearly from nothing at the pricing measure to \\(|\\rho|\\nu\\) at the share measure. For \\(\\rho>0\\) the same argument gives (3.4) and the first inequality of (3.5) with a negative \\(\\lambda\\), the resolvent being positive term by term; the kernel \\(\\xi^\\lambda\\) then lies above \\(\\xi\\), and residuals compound at the rate at which variance runs away under the tilted measure.\n\n**Remark 3.6 (other kernels).** The fractional kernel enters the proof through three facts: \\(g_\\alpha*g_{1-\\alpha}=1\\) with \\(g_{1-\\alpha}\\) nonnegative and nonincreasing, which gives the convexity inequality; positivity of the resolvent \\(k_\\lambda\\); and \\(\\xi=g_\\alpha*q\\) with \\(q\\ge0\\), which is the second face. Every completely monotone kernel \\(K\\) has a resolvent of the first kind with the same monotonicity and a completely monotone resolvent (Gripenberg, Londen and Staffans 1990, Chapter 5), its admissible curves are again of the form \\(\\xi=K*q\\) with \\(q\\) a nonnegative measure (Abi Jaber and El Euch 2019a, Example 2.2), and the left half-plane is invariant for its Riccati–Volterra equation (Abi Jaber, Larsson and Pulido 2019, Theorem 7.1). The argument therefore gives (3.5) for Volterra Heston models with completely monotone kernels, with \\(\\xi^\\lambda\\) defined by \\(\\xi^\\lambda+\\lambda K*\\xi^\\lambda=\\xi\\): the classical Heston model, where \\(K\\equiv1\\) and the statement is Remark 3.3; the lifted Heston model of Abi Jaber (2019), where \\(K\\) is a finite sum of exponentials; and gamma kernels. Appendix C writes out the regularity for the fractional kernel.\n\nFigure 1 shows the two weights for the curve calibrated in Section 7. At six months the half-share curve lies below the risk-neutral one by at most a few per cent, since \\(\\lambda_*T^\\alpha\\approx0.075\\). The large gain in the experiments comes from the first inequality in (3.5), which weights each residual by the variance still to come, in place of a uniform bound on the state.\n\n\n\n### 3.4. Cell weights\n\nIn computation the residual is bounded on closed time cells. If \\(R\\le R_j\\) on \\([a_j,b_j]\\) and the cells cover \\([0,T]\\), then (3.5) gives\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\sum_jR_j\\int_{a_j}^{b_j}\\xi^\\lambda(T-s)\\,ds\\le\\sum_jR_j\\int_{a_j}^{b_j}\\xi(T-s)\\,ds .\n\\tag{3.8}\n\\]\nEach weight is an amount of forward variance, and all of them refer to the history from time zero. For the curve of Section 7, \\(\\xi(t)=\\theta+(V_0-\\theta)E_{\\alpha_0}(-\\lambda_\\xi t^{\\alpha_0})\\), the cumulative weight \\(W_\\lambda(t)=\\int_0^t\\xi^\\lambda(s)\\,ds\\) has the expansion\n\\[\nW_\\lambda(t)=\\sum_{n=0}^\\infty(-\\lambda)^n\\,t^{1+n\\alpha}\n\\left[\\frac{\\theta}{\\Gamma(2+n\\alpha)}+(V_0-\\theta)\\,E_{\\alpha_0,\\,2+n\\alpha}\\left(-\\lambda_\\xi t^{\\alpha_0}\\right)\\right],\n\\tag{3.9}\n\\]\nvalid for \\(\\lambda T^\\alpha<1\\), and the cell \\([a_j,b_j]\\) has weight \\(W_\\lambda(T-a_j)-W_\\lambda(T-b_j)\\). Appendix C.5 bounds the tails of both series, and the implementation encloses every term by outward-rounded dyadic arithmetic.\n\n## 4. From residuals to prices\n\n### 4.1. A reference shaped by the startup expansion\n\nTheorem 3.1 separates the choice of a reference from the validity of its certificate. A useful reference makes its residual accessible on whole time intervals and reproduces the singular start of the exact solution. Write \\(g=D^\\alpha h=F_u(h)\\), so that \\(h=I^\\alpha g\\). Substituting repeatedly into the Volterra form of (2.3) gives, as \\(t\\downarrow0\\),\n\\[\ng(t)=g_0+a_1t^\\alpha+a_2t^{2\\alpha}+O(t^{3\\alpha}),\\qquad g_0=-\\tfrac12\\left(u^2+\\tfrac14\\right),\n\\tag{4.1}\n\\]\nwith coefficients determined by the equation: writing \\(\\gamma_u=\\tfrac12\\rho\\nu+i\\rho\\nu u\\),\n\\[\nb_1=\\frac{g_0}{\\Gamma(1+\\alpha)},\\quad a_1=\\gamma_ub_1,\\qquad\nb_2=\\frac{a_1\\Gamma(1+\\alpha)}{\\Gamma(1+2\\alpha)},\\quad a_2=\\gamma_ub_2+\\tfrac12\\nu^2b_1^2 .\n\\tag{4.2}\n\\]\nThe identity \\(I^\\alpha t^\\beta=\\Gamma(1+\\beta)\\,t^{\\alpha+\\beta}/\\Gamma(1+\\alpha+\\beta)\\) controls each step. We therefore take the reference to be \\(\\widehat h=I^\\alpha\\widehat g\\), where \\(\\widehat g\\) consists of the three startup powers in (4.1) and a continuous piecewise affine remainder. The powers come from the equation, and the affine part can be integrated against the fractional kernel in closed form. Its nodal values may come from any numerical solver, of any accuracy: Theorem 3.1 charges whatever residual results.\n\nAll stored nodes and coefficients are read as the exact dyadic numbers they are. The residual \\(r=\\widehat g-F_u(I^\\alpha\\widehat g)\\) then includes every discrepancy between the stored coefficients and their ideal values. Appendix D.1 bounds it on the startup interval and on every later closed cell, and verifies \\(\\operatorname{Re}\\widehat h\\le0\\) on the whole of \\([0,T]\\). The certificate belongs to the continuous trajectory: the residual is bounded on every closed cell, between the nodes as well as at them.\n\n### 4.2. Transform disks and the price bound\n\nLet \\(\\eta_n\\) bound the exponent error at the frequency \\(u_n\\), obtained from (3.8). Since \\(|\\phi_T(u-\\tfrac i2)|\\le1\\) on the Lewis contour, the reference transform \\(\\widehat\\phi_n=e^{\\widehat L_T(u_n)}\\) satisfies\n\\[\n\\left|\\phi_T(u_n-\\tfrac i2)-\\widehat\\phi_n\\right|\\le\\min\\{1,|\\widehat\\phi_n|\\}\\left(e^{\\eta_n}-1\\right)=:\\epsilon_n .\n\\tag{4.3}\n\\]\nThe passage from the integral (2.5) to a finite sum has three remainders. The first is the trapezoidal rule on the whole axis, which converges geometrically because the integrand is analytic in a strip.\n\n**Theorem 4.1 (strip quadrature).** Let \\(S\\) be a positive martingale, \\(00\\). Replacing the integral in (2.5) by the infinite trapezoidal sum with step \\(h_*\\) changes the normalised price by at most\n\\[\n\\epsilon_{\\rm grid}=\\frac{\\sqrt m\\,e^{a_*|k|}}{(\\tfrac12-a_*)\\left(e^{2\\pi a_*/h_*}-1\\right)} .\n\\tag{4.4}\n\\]\n\nThe second is the tail of the sum beyond the last node. Comparing \\(-\\operatorname{Re}h\\) from below with the solution of a scalar equation gives an envelope \\(|\\phi_T(u-\\tfrac i2)|\\le e^{-c_Vu}\\) for \\(u\\ge V\\), with an explicit \\(c_V>0\\) built from the same forward variance weights, and hence a tail bound \\(\\epsilon_{\\rm tail}\\) for the exact transform (Appendix D.2). The third is arithmetic. With the finite sum\n\\[\n\\widehat c_N=1-\\frac{h_*\\sqrt m}{\\pi}\\left(2\\operatorname{Re}\\widehat\\phi_0+\\sum_{n=1}^N\\frac{\\operatorname{Re}(e^{-iu_nk}\\widehat\\phi_n)}{u_n^2+\\frac14}\\right),\\qquad u_n=nh_*,\n\\tag{4.5}\n\\]\nthe complete price bound for one strike is\n\\[\n|c-\\widehat c_N|\\le\\frac{h_*\\sqrt m}{\\pi}\\left(2\\epsilon_0+\\sum_{n=1}^N\\frac{\\epsilon_n}{u_n^2+\\frac14}\\right)+\\epsilon_{\\rm grid}+\\epsilon_{\\rm tail}+\\epsilon_{\\rm arithmetic} .\n\\tag{4.6}\n\\]\nThe coefficient at the origin is the trapezoidal half-weight times \\(1/(\\tfrac14)\\). A node at which no reference value is computed is given the reference value zero and the radius supplied by the envelope of the exact transform, so every part of the frequency axis is charged exactly once.\n\n## 5. Portfolios: transform error as a common factor\n\nThis section uses the model through formula (2.5) alone. Its statements hold for every positive martingale \\(S\\), every reference transform and every set of radii: for the Heston, Bates and Lévy models as for rough volatility, and for a transform that comes from a closed form, a differential-equation solver, a neural network or the residual bound of Section 3.\n\n### 5.1. The joint error set\n\nFix the model, the maturity and the frequency grid, and consider \\(p\\) strikes with moneyness \\(m_i\\) and log-moneyness \\(k_i\\). From (4.5), the reference prices are \\(\\bar c_i=1+\\operatorname{Re}\\sum_na_{in}\\bar\\phi_n\\) with\n\\[\na_{in}=-\\frac{h_*\\omega_n\\sqrt{m_i}}{\\pi}\\,\\frac{e^{-iu_nk_i}}{u_n^2+\\frac14},\\qquad\\omega_0=\\tfrac12,\\quad\\omega_n=1\\ (n\\ge1),\n\\tag{5.1}\n\\]\nand the exact prices satisfy\n\\[\nc^*-\\bar c=\\operatorname{Re}\\sum_{n=0}^Na_{\\cdot n}z_n+r,\\qquad z_n=\\phi_T(u_n-\\tfrac i2)-\\bar\\phi_n,\\quad r\\in\\mathcal R,\n\\tag{5.2}\n\\]\nwhere \\(\\mathcal R\\) contains the strip, tail and arithmetic remainders. The same complex number \\(z_n\\) enters every strike. The \\(z_n\\) are deterministic unknowns, each confined to a disk.\n\nFor a portfolio \\(w\\in\\mathbb R^p\\) define its payoff transform\n\\[\n\\widehat w(u)=\\sum_{i=1}^pw_i\\sqrt{m_i}\\,e^{-iuk_i},\\qquad\\text{so that}\\qquad\n\\sum_iw_ia_{in}=-\\frac{h_*\\omega_n}{\\pi}\\,\\frac{\\widehat w(u_n)}{u_n^2+\\frac14} .\n\\tag{5.3}\n\\]\nUp to normalisation, \\(\\widehat w(u)/(u^2+\\tfrac14)\\) is the Fourier transform on the Lewis contour of the payoff \\(\\sum_iw_i(S_T-K_i)^+\\). A portfolio of options with one expiry is a single European payoff, and its transform is the sum of the transforms of its legs.\n\n**Theorem 5.1 (joint error set).** Suppose (5.2) holds with \\(|z_n|\\le\\epsilon_n\\) and \\(\\mathcal R\\) compact and convex. Let\n\\[\n\\mathcal E=\\left\\{\\operatorname{Re}\\sum_na_{\\cdot n}z_n:\\ |z_n|\\le\\epsilon_n\\right\\}+\\mathcal R,\\qquad d=\\bar c-c^{\\rm fast}.\n\\tag{5.4}\n\\]\nThen \\(c^*-c^{\\rm fast}\\in d+\\mathcal E\\), and for every \\(w\\in\\mathbb R^p\\) the support function of this set is\n\\[\nh_{d+\\mathcal E}(w)=w^\\top d+\\frac{h_*}\\pi\\sum_{n=0}^N\\omega_n\\,\\epsilon_n\\,\\frac{|\\widehat w(u_n)|}{u_n^2+\\frac14}+h_{\\mathcal R}(w).\n\\tag{5.5}\n\\]\nConsequently \\(DF_T\\,|w^\\top(c^*-c^{\\rm fast})|\\le DF_T\\max\\{h_{d+\\mathcal E}(w),h_{d+\\mathcal E}(-w)\\}\\).\n\n**Proof.** Substituting the actual errors in (5.2) gives the inclusion. The image of the disk \\(|z|\\le\\epsilon\\) under the real-linear map \\(z\\mapsto\\operatorname{Re}(bz)\\) has support \\(\\epsilon|b|\\), attained at \\(z=\\epsilon\\bar b/|b|\\). The disks are independent, supports add under Minkowski sums, and (5.3) identifies \\(b=\\sum_iw_ia_{in}\\). Translation by \\(d\\) adds \\(w^\\top d\\). \\(\\square\\)\n\nGeometrically, each frequency contributes an ellipse to \\(\\mathcal E\\), the image of a disk under \\(z\\mapsto\\operatorname{Re}(a_{\\cdot n}z)\\), and \\(\\mathcal E\\) is the sum of these ellipses and the remainder. Bounding each price separately replaces this body by its bounding box, whose support is\n\\[\nh_{\\rm box}(w)=\\frac{h_*}\\pi\\sum_{n=0}^N\\omega_n\\,\\epsilon_n\\,\\frac{\\sum_i|w_i|\\sqrt{m_i}}{u_n^2+\\frac14}+\\sum_i|w_i|\\rho_i\n\\tag{5.6}\n\\]\nwhen \\(\\mathcal R=\\prod_i[-\\rho_i,\\rho_i]\\). We call (5.5) the joint bound and (5.6), with the same signed centre \\(w^\\top d\\), the leg-by-leg bound. All numerical evaluations of either are outward enclosures.\n\nAny separately proved interval for a price can be intersected with \\(\\bar c+\\mathcal E\\), since the exact vector lies in both sets. Every comparison of the two bounds in Section 7 uses one centre, one set of radii and one remainder, so that the aggregation is the only difference.\n\n### 5.2. The netting ratio\n\nThe two bounds differ only through the factor multiplying each \\(\\epsilon_n\\). Define the netting ratio of \\(w\\) at the frequency \\(u\\) by\n\\[\n\\chi_w(u)=\\frac{|\\widehat w(u)|}{\\sum_i|w_i|\\sqrt{m_i}}\\in[0,1].\n\\tag{5.7}\n\\]\nThe node term of the joint bound is the node term of the leg-by-leg bound with each frequency multiplied by \\(\\chi_w(u_n)\\). By the equality case of the triangle inequality, \\(\\chi_w(u)=1\\) exactly when the numbers \\(w_ie^{-iuk_i}\\) with \\(w_i\\ne0\\) share one argument, so the joint bound is strictly smaller as soon as one node of positive radius sees two legs out of phase.\n\n**Proposition 5.2 (spreads and butterflies).** Let the strikes form a geometric ladder with log-moneyness \\(k,k+\\Delta,\\dots,k+j\\Delta\\), and let \\(w\\) be the \\(j\\)-th difference, with weights \\((-1)^i\\binom ji\\). Then\n\\[\n\\chi_w(u)=\\left(\\tanh^2\\tfrac\\Delta4+\\operatorname{sech}^2\\tfrac\\Delta4\\,\\sin^2\\tfrac{u\\Delta}2\\right)^{j/2}\\le\\left(\\tfrac\\Delta2\\sqrt{u^2+\\tfrac14}\\right)^j .\n\\tag{5.8}\n\\]\nFor the vertical spread (\\(j=1\\)) the ratio equals \\(\\tanh(\\Delta/4)\\) at \\(u=0\\) and reaches one exactly at the odd multiples of \\(\\pi/\\Delta\\). The butterfly (\\(j=2\\)) has the square of the ratio of the spread.\n\n**Proof.** With \\(\\zeta=e^{(1/2-iu)\\Delta}\\) the payoff transform is \\(\\sqrt m\\,e^{-iuk}(1-\\zeta)^j\\) and the gross weight is \\(\\sqrt m\\,(1+e^{\\Delta/2})^j\\). Then \\(|1-\\zeta|^2=(e^{\\Delta/2}-1)^2+4e^{\\Delta/2}\\sin^2(u\\Delta/2)\\), and division by \\((1+e^{\\Delta/2})^2\\) gives the equality. The inequality uses \\(\\tanh x\\le x\\), \\(\\operatorname{sech}x\\le1\\) and \\(|\\sin x|\\le|x|\\). \\(\\square\\)\n\nThe proposition gives the hierarchy of exposures. A single call loads on the transform error at the frequency \\(u\\) with weight \\((u^2+\\tfrac14)^{-1}\\). A vertical spread of width \\(\\Delta\\) loads with weight at most \\(\\Delta\\,(u^2+\\tfrac14)^{-1/2}\\), and a butterfly with weight at most \\(\\Delta^2\\) at every frequency, in units of \\(h_*\\sqrt m\\,(1+e^{\\Delta/2})^j/(2^j\\pi)\\). Each difference in strike removes one power of the frequency from the decay of the payoff transform and returns one power of the strike spacing. This is the error form of the Breeden and Litzenberger (1978) relation: a butterfly reads off the state-price density, and its error is the unweighted sum of the transform errors.\n\nThe proposition also locates the frequencies that matter. A spread is hedged against transform error whose wavelength in log-strike, \\(2\\pi/u\\), is long compared with its width, and it is fully exposed at the Nyquist frequency \\(u=\\pi/\\Delta\\), where the two legs respond in opposite phase. For strikes 100 points apart near 4400, \\(\\Delta\\approx0.0225\\) and \\(\\pi/\\Delta\\approx140\\). For a five-point vertical, the finest spacing commonly listed near the money, \\(\\pi/\\Delta\\approx2800\\). A narrow spread is protected over a band of frequencies that is inversely proportional to its width.\n\nFor portfolios that are not differences the ratio is governed by concentration.\n\n**Proposition 5.3 (the Herfindahl law).** Let the log-strikes \\(k_i\\) be distinct and put \\(\\ell_i=w_i\\sqrt{m_i}\\). Then\n\\[\n\\lim_{U\\to\\infty}\\frac1U\\int_0^U\\chi_w(u)^2\\,du=\\frac{\\sum_i\\ell_i^2}{\\left(\\sum_i|\\ell_i|\\right)^2} .\n\\tag{5.9}\n\\]\n\n**Proof.** Expand \\(|\\widehat w(u)|^2=\\sum_{i,j}\\ell_i\\ell_je^{-iu(k_i-k_j)}\\). The diagonal terms sum to \\(\\sum_i\\ell_i^2\\) and each off-diagonal term has mean zero over long intervals. \\(\\square\\)\n\nThe right side of (5.9) is the Herfindahl index of the position, the reciprocal of its effective number of strikes. Its square root is \\(1/\\sqrt2\\approx0.707\\) for a two-leg spread, \\(\\sqrt6/4\\approx0.612\\) for a butterfly with weights \\((1,-2,1)\\), and \\(1/\\sqrt p\\) for an equal-weight basket of \\(p\\) strikes. At the origin, in contrast, \\(\\chi_w(0)=|\\sum_i\\ell_i|/\\sum_i|\\ell_i|\\) measures the net position. The two regimes have familiar names. Error at low frequency moves all strikes together; it is systematic, and it is removed by hedging. Error at high frequency moves neighbouring strikes independently; it is idiosyncratic, and it is removed by diversification. A long-only basket is fully exposed to the first and almost immune to the second, and a tight spread is the reverse. Figure 2 shows the three profiles.\n\n\n\nThe centre \\(w^\\top d\\) in (5.5) obeys the same logic from the other side. It is the known displacement of the stored output from the reference, and for a smooth approximation it is smooth in strike: it cancels in spreads and accumulates in baskets. Section 7.4 finds both effects in the data.\n\n### 5.3. The calibration objective is a portfolio\n\nCalibration compares model prices with quotes \\(M_i\\) through a least-squares objective\n\\[\nJ(c)=\\frac1{2p}\\sum_{i=1}^p(c_i-M_i)^2 .\n\\]\nIts numerical error is a portfolio error. For the exact prices \\(c^*=\\bar c+e\\) with \\(e\\in\\mathcal E\\),\n\\[\nJ(c^*)-J(\\bar c)=g^\\top e+\\frac{|e|^2}{2p},\\qquad g=\\frac{\\bar c-M}{p},\n\\]\nso to first order the objective is exposed to the transform error as the portfolio whose weights are the pricing residuals.\n\n**Proposition 5.4 (objective enclosure).** If \\(e\\in\\mathcal E\\) and \\(|e_i|\\le\\beta_i\\) for each \\(i\\), then\n\\[\nJ(\\bar c)-h_{\\mathcal E}(-g)\\ \\le\\ J(c^*)\\ \\le\\ J(\\bar c)+h_{\\mathcal E}(g)+\\frac1{2p}\\sum_i\\beta_i^2 .\n\\tag{5.10}\n\\]\n\n**Proof.** The lower bound drops the nonnegative quadratic term and uses \\(g^\\top e\\ge-h_{\\mathcal E}(-g)\\); the upper bound uses \\(g^\\top e\\le h_{\\mathcal E}(g)\\) and \\(|e|^2\\le\\sum_i\\beta_i^2\\). \\(\\square\\)\n\nThe residuals of a reasonable fit change sign across strikes, so the calibration portfolio is long some strikes and short others, and its support under (5.5) is smaller than under (5.6). For a finite set of candidate parameters, a candidate is the certified minimiser when the upper end of its enclosure lies below the lower ends of all the others; in general the candidates whose lower end lies below the smallest upper end form a set that contains every exact minimiser. Each candidate carries its own transform and its own error set.\n\n## 6. The stored output\n\nThe certificate applies to any stored vector of prices once its displacement from the reference is enclosed. In the experiments the stored output is the third-order rational approximation of Gatheral and Radoičić (2019), which writes the Riccati solution as \\(\\widehat H(y)=P(y)/Q(y)\\) in the variable \\(y=\\nu t^\\alpha\\), with cubic \\(P\\) and \\(Q=1+q_1y+q_2y^2+q_3y^3\\), and fixes the six coefficients by matching three terms of the expansion at \\(t=0\\) and three terms of the expansion at infinity. The formula costs a handful of operations per frequency. Its coefficients are complex, and on the parameter set of Section 7 the absence of poles on the positive axis is a theorem.\n\n**Proposition 6.1 (structure of the rational output).** For \\(\\alpha\\in[\\tfrac{13}{25},\\tfrac35]\\), \\(\\rho=-0.7445\\), \\(\\nu>0\\) and every real \\(u\\), the matching system of Gatheral and Radoičić is nonsingular, and\n\\[\n\\operatorname{Re}q_j>0\\ (j=1,2,3),\\qquad\\operatorname{Re}Q(y)\\ge1\\quad(y\\ge0),\\qquad\\operatorname{Re}\\widehat H(y)<0\\quad(y>0).\n\\tag{6.1}\n\\]\n\nThe proof in Appendix B reduces (6.1) to the signs of nine real functions on a compact rectangle in \\((\\alpha,u/(1+u))\\) and verifies them by interval arithmetic on a finite cover. The conclusion is stable under small changes of \\(\\rho\\), since all nine inequalities are strict. The proposition shows that the stored output is well defined and shares the half-plane of the exact solution. Its accuracy is the business of the reference: one construction supplies the price, and an independent one bounds its error.\n\n## 7. Evidence from S&P 500 options\n\n### 7.1. Data and market scale\n\nThe quotes are the bid and ask implied volatilities of twelve SPX options with six months to expiry, strikes 3700 to 4800 in steps of 100, recorded on 18 June 2021 and published by Jeng and Kiliçman (2021). We use the index level 4221.86 of that data set as the forward, with zero carry, and convert each implied volatility to a call price by the Black–Scholes formula in exact arithmetic. Table 1 gives the result in index points. One point is 100 dollars per contract.\n\n**Table 1. Six-month SPX call quotes of 18 June 2021 in index points: bid, ask and quoted width.**\n\n| Strike | Bid | Ask | Width | Strike | Bid | Ask | Width |\n|---:|---:|---:|---:|---:|---:|---:|---:|\n| 3700 | 586.40 | 588.70 | 2.30 | 4300 | 140.00 | 141.50 | 1.50 |\n| 3800 | 502.10 | 504.50 | 2.40 | 4400 | 91.10 | 92.40 | 1.30 |\n| 3900 | 420.20 | 422.60 | 2.40 | 4500 | 55.00 | 56.20 | 1.21 |\n| 4000 | 341.70 | 344.20 | 2.50 | 4600 | 31.60 | 32.60 | 0.99 |\n| 4100 | 267.79 | 270.19 | 2.40 | 4700 | 17.90 | 18.80 | 0.90 |\n| 4200 | 199.69 | 202.20 | 2.51 | 4800 | 10.30 | 11.10 | 0.80 |\n\nThe 4400–4500 call spread is worth about 36.1 points at mid-market, and a trader who crosses both legs pays up to 2.5 points of spread. These are the magnitudes against which a numerical error should be judged. An error of a quarter point is a fifth of the quoted width of the 4400 call, and a tenth of the width of the spread.\n\nThe model parameters are those fitted to this surface by Jeng and Kiliçman (2021): \\(\\nu=0.2897\\), \\(\\rho=-0.7445\\), and the forward variance curve\n\\[\n\\xi(t)=\\theta+(V_0-\\theta)\\,E_{\\alpha_0}\\left(-\\lambda_\\xi t^{\\alpha_0}\\right),\\qquad(\\alpha_0,V_0,\\theta,\\lambda_\\xi)=(0.5286,\\,0.0262,\\,0.0721,\\,0.5037).\n\\tag{7.1}\n\\]\nThe curve rises from an instantaneous volatility of about 16 per cent towards a long-run level of about 27 per cent. It is nondecreasing, hence admissible, and it is held fixed when the roughness index \\(\\alpha\\) of the kernel in (2.1) is varied. The damping rate of Theorem 3.1 is \\(\\lambda_*=\\tfrac12|\\rho|\\nu\\approx0.108\\).\n\nThe stored output is the rational approximation of Section 6 evaluated in binary64, with its own quadrature. The reference of Section 4 uses the Fourier step \\(h_*=\\tfrac18\\), a grid through \\(u=128\\) (1,025 nodes) and 100-bit outward dyadic arithmetic. Three ingredients of the certificate vary across the experiments: the propagation of the residual, either a uniform bound on the state or Theorem 3.1 with one global residual maximum or with 128 time bins; the band on which the reference transform is computed, \\(u\\le64\\) or \\(u\\le128\\), nodes above the band receiving the envelope radius; and the maturity. Appendix E lists the configurations, which carry the labels H0–H4 at six months, Q0–Q2 at three months, and N1, N2 and N2L for the freshly generated nearby fields.\n\n### 7.2. What the forward variance weight buys\n\nTable 2 follows the complete bound for one position, long the 4400 call and short the 4500 call, as the certificate is refined. Every row keeps the stored output, the signed centre, all finite frequencies, the strip, the exact tail and the arithmetic.\n\n**Table 2. Complete certificates for the unit 4400–4500 call spread, in index points.**\n\n| Maturity and certificate | Joint bound | Leg-by-leg bound | Tolerance met |\n|---|---:|---:|---|\n| H0: six months, uniform state bound | 1.397612096 | 2.679930311 | none |\n| H1: six months, Theorem 3.1, global residual | 0.378598956 | 0.476401716 | 1/2 point, both |\n| H2: six months, Theorem 3.1, 128 time bins | 0.367258782 | — | 1/2 point |\n| H3: six months, reference through 128, upper band global | 0.132245064 | 0.158585551 | 1/4 point, both |\n| H4: six months, reference through 128, 128 bins | 0.115215934 | 0.137896018 | 1/4 point, both |\n| Q0: three months, reference through 64, global | 1.207557898 | 1.545191542 | none |\n| Q1: three months, reference through 128, global | 0.351318692 | 0.408293698 | 1/2 point, both |\n| Q2: three months, reference through 128, 128 bins | 0.233318843 | 0.252393939 | 1/4 point, joint only |\n\nThe step from H0 to H1 changes one thing. Both rows use the same residuals of the same reference trajectories. H0 bounds the state error uniformly in time and multiplies by the total mass of the kernel \\(q\\); H1 applies (3.5), which weights each residual by the forward variance that remains. The complete bound falls by 72.91 per cent. Extending the reference from \\(u\\le64\\) to \\(u\\le128\\) (H2 to H3, Q0 to Q1) replaces envelope radii by certified ones and moves the centre, from about \\(-0.167\\) to \\(-0.082\\) points at six months, so that step combines a smaller uncertainty with a different displacement.\n\nTable 3 isolates the propagation on a second bank, generated independently at \\(N_t=2048\\) time cells, with everything else held fixed. The last two rows use identical binwise residual envelopes and differ only in the kernel, \\(\\xi\\) against \\(\\xi^{\\lambda_*}\\).\n\n**Table 3. Matched propagation on bank N2L for the six-month 4400–4500 spread, in index points.**\n\n| Propagation | Stored output, leg-by-leg | Stored output, joint | Reference returned, joint | Node radius of used nodes |\n|---|---:|---:|---:|---:|\n| Uniform state bound | 3.690877564 | 1.829428107 | 1.662665889 | 1.476801373 |\n| (3.5) with \\(\\xi\\), global residual | 0.589267065 | 0.428570987 | 0.261808769 | 0.075944253 |\n| (3.8) with \\(\\xi\\), 128 bins | 0.514096070 | 0.394999331 | 0.228237113 | 0.042372598 |\n| (3.8) with \\(\\xi^{\\lambda_*}\\), 128 bins | 0.509733613 | 0.393082828 | 0.226320610 | 0.040456094 |\n\nThe forward variance weight accounts for almost all of the improvement: the node radius falls from 1.477 to 0.076 points. Time-local envelopes take it to 0.042. The leverage damping of Proposition 3.2 removes a further 4.5 per cent of the node radius, in line with \\(\\lambda_*T^\\alpha\\approx0.075\\) at this horizon; its effect grows with maturity and with \\(|\\rho|\\nu\\). The complete bound for the stored output moves less, because the fixed terms dominate it: the centre, the omitted finite nodes and the remainders together exceed 0.352 points on this bank. The next gain therefore lies in the reference band, and extending it is what takes H2 to H4 in Table 2.\n\n### 7.3. A decision that turns on netting\n\nAt three months a single global residual maximum certifies half a point (Q1). With 128 time bins (Q2) the bound reaches the quarter-point tolerance, and there the aggregation decides. Table 4 is the ledger of Q2. The two columns share the stored output, the reference centre, all 1,025 node radii and every remainder.\n\n**Table 4. Ledger of the three-month 4400–4500 spread (Q2), in index points. Only the aggregation of the node errors differs.**\n\n| Component | Joint | Leg-by-leg |\n|---|---:|---:|\n| Absolute centre \\(|w^\\top d|\\) | 0.149775147 | 0.149775147 |\n| Node support, 1,025 nodes | 0.045461793 | 0.064536890 |\n| Strip remainder | 0.000026654 | 0.000026654 |\n| Exact infinite tail | 0.038055248 | 0.038055248 |\n| Reference arithmetic | \\(4.1\\times10^{-16}\\) | \\(4.1\\times10^{-16}\\) |\n| Complete bound | 0.233318843 | 0.252393939 |\n\nNetting lowers the node term by 30 per cent and the complete bound by 7.56 per cent, from above a quarter point to below it. Nothing else changed.\n\nProposition 5.2 predicts where the saving comes from. Here \\(\\Delta=\\log(4500/4400)=0.02247\\), and the stored ratio of the joint to the leg-by-leg coefficient at each of the 1,025 nodes agrees with (5.8) to fifteen digits. Table 5 splits the node term by frequency.\n\n**Table 5. Node term of the three-month spread by frequency band, in index points.**\n\n| Band of \\(u\\) | Nodes | Joint | Leg-by-leg | Ratio | Share of leg-by-leg charge |\n|---|---:|---:|---:|---:|---:|\n| 0 to 16 | 128 | 0.000074 | 0.000529 | 0.139 | 0.8% |\n| 16 to 32 | 128 | 0.001158 | 0.004100 | 0.282 | 6.4% |\n| 32 to 64 | 256 | 0.010559 | 0.020137 | 0.524 | 31.2% |\n| 64 to 96 | 256 | 0.017651 | 0.022785 | 0.775 | 35.3% |\n| 96 to 128 | 257 | 0.016020 | 0.016986 | 0.943 | 26.3% |\n\nBelow \\(u=16\\) the spread keeps 14 per cent of its leg-by-leg error, and near the top of the grid, which approaches the Nyquist frequency \\(\\pi/\\Delta\\approx140\\), it keeps 94 per cent. The certified radii grow with the frequency, so nine tenths of the charge sits above \\(u=32\\), where a spread of this width is poorly hedged. The saving of 30 per cent is what a 100-point spread can extract from an error concentrated at short wavelengths.\n\nA narrower spread extracts more. Table 6 evaluates the node term of (5.5) and (5.6) with the same 1,025 radii for verticals of other widths struck at 4400. The radii do not depend on the strike, so the table is a statement about the same certified transform.\n\n**Table 6. Node term of a three-month vertical spread struck at 4400, by width, with the certified radii of Q2; index points.**\n\n| Width in points | Nyquist frequency | Joint | Leg-by-leg | Ratio |\n|---:|---:|---:|---:|---:|\n| 5 | 2766 | 0.002711 | 0.064193 | 0.042 |\n| 10 | 1384 | 0.005414 | 0.064211 | 0.084 |\n| 25 | 555 | 0.013403 | 0.064265 | 0.209 |\n| 50 | 278 | 0.025925 | 0.064356 | 0.403 |\n| 100 | 140 | 0.045462 | 0.064537 | 0.704 |\n| 200 | 71 | 0.052476 | 0.064895 | 0.809 |\n| 400 | 36 | 0.044482 | 0.065601 | 0.678 |\n\nFor the five-point vertical the joint node term is about a twenty-fourth of the leg-by-leg term. The joint term grows in proportion to the width while the Nyquist frequency lies beyond the grid, and levels off once it enters. Figure 3 shows the geometry behind Table 4: the joint error set of the two prices and its bounding box.\n\n\n\n### 7.4. Twenty-eight positions\n\nThe six-month experiment fixes 28 portfolio directions on the twelve strikes before any bound is computed (Table 7), and keeps their weights as stated. Table 7 also reports, for bank H1 at \\(\\alpha=0.52\\), how much of the leg-by-leg node term each family keeps and the mean absolute centre.\n\n**Table 7. The 28 fixed positions, with node netting and centre on bank H1 at \\(\\alpha=0.52\\). The netting ratio is the joint node term divided by the leg-by-leg node term.**\n\n| Family | Weights | Count | Netting ratio, min–mean–max | Herfindahl level | Mean centre, points |\n|---|---|---:|---:|---:|---:|\n| Adjacent spreads | \\(e_i-e_{i+1}\\) | 11 | 0.649–0.708–0.771 | 0.707 | 0.108 |\n| Adjacent butterflies | \\(e_i-2e_{i+1}+e_{i+2}\\) | 10 | 0.449–0.524–0.608 | 0.612 | 0.071 |\n| Wide spreads | \\(e_0-e_3,\\ e_3-e_6,\\ e_6-e_9,\\ e_0-e_{11}\\) | 4 | 0.474–0.570–0.651 | 0.707 | 0.235 |\n| Equal-weight baskets | twelve, lower six, upper six strikes | 3 | 0.076–0.156–0.221 | 0.289, 0.408 | 0.344 |\n\nThe adjacent spreads keep 0.708 of their leg-by-leg node error on average, against the Herfindahl level 0.707 of a two-leg position. The butterflies keep 0.524. The baskets keep least of all: an equal-weight position in all twelve calls keeps 0.076 of its leg-by-leg node error, because three quarters of the node charge of this bank sits between \\(u=64\\) and \\(u=96\\), where twelve strikes spread over a quarter of a unit of log-moneyness respond out of phase. The centre runs the other way. The stored rational output lies above the reference at every strike, by 0.04 points at the ends of the ladder and by up to 0.65 points near the money. A displacement of one sign cancels in part in spreads and butterflies and accumulates in baskets, where it makes up 0.344 of a complete bound of 0.356 points. The approximation error of the stored formula is systematic across strikes and the uncertainty of the certificate is idiosyncratic, as Section 5.2 anticipated, and the joint bound accounts for each correctly.\n\nTable 8 counts the positions certified at three tolerances. Within a row the two methods use identical radii.\n\n**Table 8. Positions certified out of 28 on bank H1, by tolerance in index points.**\n\n| Roughness \\(\\alpha\\) | Tolerance | Leg-by-leg | Joint |\n|---|---:|---:|---:|\n| 0.52 | 1/4 | 0 | 2 |\n| 0.52 | 1/2 | 13 | 28 |\n| 0.52 | 1 | 28 | 28 |\n| 0.60 | 1/4 | 1 | 10 |\n| 0.60 | 1/2 | 22 | 28 |\n| 0.60 | 1 | 28 | 28 |\n| 0.90 | 1/4 | 13 | 20 |\n| 0.90 | 1/2 | 25 | 25 |\n| 0.90 | 1 | 27 | 28 |\n\nAt half a point and \\(\\alpha=0.52\\) the joint bound certifies every position and the leg-by-leg bound certifies 13. Netting decides whenever the tolerance is of the order of the bound, which is the case for a tolerance set inside the bid–ask spread. All 28 positions are certified from one bank: once the transform disks exist, an additional position costs 1,025 support terms. Figure 4 shows the count as a function of the tolerance for the three candidates, and Appendix E.3 gives the exact thresholds.\n\n\n\n### 7.5. Is the roughness ranking numerical?\n\nThe roughness fitted to this surface is \\(H=0.0286\\), and candidates a few thousandths apart in \\(H\\) produce prices that differ by fractions of an index point. Whether a calibration can tell them apart depends on the accuracy of the pricer. We test five candidates, \\(\\alpha\\in\\{0.520,0.525,0.530,0.540,0.550\\}\\), with the curve (7.1), the quotes and all other parameters fixed. The design was set in advance: every candidate receives a freshly generated reference field and a complete residual bank at \\(N_t=1024\\), and all five are regenerated at \\(N_t=2048\\) if any pair remains unresolved. The objective is the normalised least-squares distance to the mid prices, enclosed by Proposition 5.4.\n\n**Table 9. Objective enclosures for five roughness candidates, as \\(J\\times10^8\\), and the number of the ten pairs strictly separated.**\n\n| \\(\\alpha\\) | \\(H\\) | \\(N_t=1024\\), joint | \\(N_t=2048\\), joint |\n|---:|---:|---:|---:|\n| 0.520 | 0.020 | [4.825094, 6.778350] | [5.293449, 6.024095] |\n| 0.525 | 0.025 | [5.645378, 7.501414] | [6.097920, 6.798485] |\n| 0.530 | 0.030 | [6.748974, 8.517584] | [7.187004, 7.859529] |\n| 0.540 | 0.040 | [9.807435, 11.422182] | [10.218707, 10.839486] |\n| 0.550 | 0.050 | [14.000658, 15.482012] | [14.387007, 14.960680] |\n| Pairs separated, joint | | 7 of 10 | 10 of 10 |\n| Pairs separated, leg-by-leg | | 3 of 10 | 7 of 10 |\n\nAt the finer level the five enclosures are disjoint and ordered, and \\(\\alpha=0.520\\) is the certified minimiser on this grid. The closest pair, 0.520 against 0.525, is separated by a gap of \\(7.4\\times10^{-10}\\) in the normalised objective. The leg-by-leg enclosures leave three pairs unresolved at the same level, those among 0.520, 0.525 and 0.530. The ordering of roughness candidates half a hundredth apart in \\(H\\) is therefore a certified property of the model and the quotes.\n\n### 7.6. Model misfit without solver doubt\n\nA certificate also settles the opposite question. Each of the five candidates is inconsistent with at least one quoted bid–ask band, and for a certified price this is a statement about the model. At \\(\\alpha=0.52\\), on the coarsest bank H0, the exact model price of the 4400 call lies in an interval whose lower end is 92.887 points, while the quoted ask is 92.396: the model is above the ask by at least 0.491 points. For the 4500 call the excess is at least 0.068 points. At the other ten strikes the certified interval meets the quoted band. A desk that sees its rough Heston calibration miss the 4400 and 4500 strikes on this date can rule out the pricer as the cause.\n\n### 7.7. Auditing an output or replacing it\n\nA stored-output certificate serves the audit of a pricing library or of a valuation that has already been reported. When the output may be replaced and the reference is available, returning the reference is the natural alternative. Table 10 compares the two on bank N2L, together with a reimplementation of the modified Adams scheme of Boyarchenko, de Innocentis and Levendorskii (2025). All rows share the model, the spread, the reference centre, the node radii and the remainders.\n\n**Table 10. Stored and replaceable outputs for the six-month 4400–4500 spread on bank N2L, in index points.**\n\n| Output | Joint bound | Leg-by-leg bound | Quarter point, joint |\n|---|---:|---:|---|\n| Stored rational output | 0.394999331 | 0.514096070 | not certified |\n| Reference returned in binary64 | 0.228237113 | 0.347333852 | certified |\n| Stored output plus stored correction | 0.228237113 | 0.347333852 | certified |\n| Modified Adams, 512 steps | 0.229082759 | 0.348179497 | certified |\n| Modified Adams, 1024 steps | 0.228534686 | 0.347631424 | certified |\n\nThe reference and the corrected output are certified to a quarter point by the joint bound, and the leg-by-leg bound certifies none of the five outputs. The Adams outputs lie within 0.001 points of the reference, so that more than 99.6 per cent of their bound is the uncertainty of the reference itself: at this resolution the scheme is as accurate as the reference can tell.\n\n## 8. Conclusion\n\nThe construction follows the valuation map. The Riccati equation determines the state and, through its invariant half-plane, the stability of errors. The forward variance curve determines what a residual costs. The payoff determines which transform errors a position is exposed to. At each stage an absolute value taken too early would discard structure that the next stage can use, and the financial content of the paper lies in what that structure is: the weight on a residual is expected variance under a share-tilted measure, and the weight on a transform error is the Fourier transform of the payoff.\n\nThe results reach beyond the example. Section 5 is model-free: the payoff transform, the netting ratio and the Herfindahl law govern the numerical error of every pricer built on a common transform, whatever the model and whatever the method behind the transform. Theorem 3.1 holds on every damped contour and, by Remark 3.6, for Volterra Heston models with completely monotone kernels, the classical and the lifted Heston models included. The damping it uses is the frequency-independent part \\(\\tfrac12|\\rho|\\nu\\); the exact solution supplies more, increasing with the frequency, which is where the radii of Table 5 are largest. For a trading desk, a validation team or a supervisor the consequence is practical: the numerical component of valuation uncertainty can be bounded, attributed to frequencies, and netted across a book exactly as the payoffs net.\n\nThe [online article](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/ARTICLE.md), the scientific source and the verification commands accompany the paper. All 27 numerical banks are identified by size and SHA-256 in `SCIENTIFIC-MANIFEST.json`. `verify_source.py` checks the publication identities. `reproduce.py` reads the saved banks and reconstructs the prices, objectives and decisions downstream of them; `--full` also recomputes the structural-sign cover of Proposition 6.1. Appendix E and `COMMANDS.md` distinguish the generators, the readers and the arithmetic they share.\n\n## Appendices and Supporting Material\n\nThis page includes the complete main text and references. Appendices, supplementary proofs, code, and verification material are available in the [GitHub repository](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston) and the [full manuscript](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/manuscript/merged-heston-en.md#appendix-a-rescaled-variables-model-and-transform).\n\n## References\n\nAbi Jaber, E. (2019). Lifting the Heston model. *Quantitative Finance* 19(12), 1995–2013.\n\nAbi Jaber, E. and El Euch, O. (2019a). Markovian structure of the Volterra Heston model. *Statistics & Probability Letters* 149, 63–72. [doi:10.1016/j.spl.2019.01.024](https://doi.org/10.1016/j.spl.2019.01.024). Theorem and example numbers refer to arXiv:1803.00477.\n\nAbi Jaber, E. and El Euch, O. (2019b). Multifactor approximation of rough volatility models. *SIAM Journal on Financial Mathematics* 10(2), 309–349.\n\nAbi Jaber, E., Larsson, M. and Pulido, S. (2019). Affine Volterra processes. *Annals of Applied Probability* 29(5), 3155–3200.\n\nAlbrecher, H., Mayer, P., Schoutens, W. and Tistaert, J. (2007). The little Heston trap. *Wilmott Magazine*, January, 83–92.\n\nBasel Committee on Banking Supervision (2019). *Minimum capital requirements for market risk*. Basel Framework, chapter MAR32, version effective 1 January 2023. Bank for International Settlements.\n\nBayer, C. and Breneis, S. (2023a). Weak Markovian approximations of rough Heston. [arXiv:2309.07023](https://arxiv.org/abs/2309.07023v1).\n\nBayer, C. and Breneis, S. (2023b). Efficient option pricing in the rough Heston model using weak simulation schemes. [arXiv:2310.04146](https://arxiv.org/abs/2310.04146v1).\n\nBayer, C., Friz, P. and Gatheral, J. (2016). Pricing under rough volatility. *Quantitative Finance* 16(6), 887–904.\n\nBen Hammouda, C., Ben Romdhane, A., Samet, M. and Tempone, R. F. (2026). Single- and multilevel quadrature with error control for Fourier pricing under the rough Heston model. [arXiv:2609.00438](https://arxiv.org/abs/2609.00438v1).\n\nBoard of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency (2026). *Supervisory Guidance on Model Risk Management*. SR Letter 26-2, 17 April 2026.\n\nBoyarchenko, S., de Innocentis, M. and Levendorskii, S. (2025). Fast reliable pricing and calibration of the rough Heston model. [arXiv:2508.15080](https://arxiv.org/abs/2508.15080v1).\n\nBreeden, D. T. and Litzenberger, R. H. (1978). Prices of state-contingent claims implicit in option prices. *Journal of Business* 51(4), 621–651.\n\nCboe Global Markets (2024). *S&P 500 Index Options: SPX product specifications*, as of 19 March 2024.\n\nCboe Global Markets (2026). Cboe Global Markets reports trading volume for December and full year 2025. Press release, 6 January 2026.\n\nCommission Delegated Regulation (EU) 2016/101 of 26 October 2015 supplementing Regulation (EU) No 575/2013 with regard to regulatory technical standards for prudent valuation. *Official Journal of the European Union* L 21, 28 January 2016, 54–65.\n\nCont, R. and Das, P. (2024). Rough volatility: fact or artefact? *Sankhya B* 86, 191–223.\n\nDiethelm, K., Ford, N. J. and Freed, A. D. (2004). Detailed error analysis for a fractional Adams method. *Numerical Algorithms* 36, 31–52.\n\nEl Euch, O., Fukasawa, M. and Rosenbaum, M. (2018). The microstructural foundations of leverage effect and rough volatility. *Finance and Stochastics* 22, 241–280.\n\nEl Euch, O. and Rosenbaum, M. (2018). Perfect hedging in rough Heston models. *Annals of Applied Probability* 28(6), 3813–3856. Statement numbers refer to arXiv:1703.05049.\n\nEl Euch, O. and Rosenbaum, M. (2019). The characteristic function of rough Heston models. *Mathematical Finance* 29(1), 3–38.\n\nGatheral, J., Jaisson, T. and Rosenbaum, M. (2018). Volatility is rough. *Quantitative Finance* 18(6), 933–949.\n\nGatheral, J. and Keller-Ressel, M. (2019). Affine forward variance models. *Finance and Stochastics* 23, 501–533.\n\nGatheral, J. and Radoičić, R. (2019). Rational approximation of the rough Heston solution. *International Journal of Theoretical and Applied Finance* 22(3), 1950010. [doi:10.1142/S0219024919500109](https://doi.org/10.1142/S0219024919500109).\n\nGatheral, J. and Radoičić, R. (2024). A generalization of the rational rough Heston approximation. *Quantitative Finance* 24(2), 329–335. [doi:10.1080/14697688.2024.2302055](https://doi.org/10.1080/14697688.2024.2302055).\n\nGripenberg, G., Londen, S.-O. and Staffans, O. (1990). *Volterra Integral and Functional Equations*. Cambridge University Press.\n\nHagan, P. S., Kumar, D., Lesniewski, A. S. and Woodward, D. E. (2002). Managing smile risk. *Wilmott Magazine*, September, 84–108.\n\nHagan, P. S., Kumar, D., Lesniewski, A. S. and Woodward, D. E. (2014). Arbitrage-free SABR. *Wilmott* 2014(69), 60–75.\n\nHager, P. P. and Kreher, D. (2026). Expanding the rough Heston model in H. [arXiv:2606.16619](https://arxiv.org/abs/2606.16619v1).\n\nHeston, S. L. (1993). A closed-form solution for options with stochastic volatility with applications to bond and currency options. *Review of Financial Studies* 6(2), 327–343.\n\nHorvath, B., Muguruza, A. and Tomas, M. (2021). Deep learning volatility: a deep neural network perspective on pricing and calibration in (rough) volatility models. *Quantitative Finance* 21(1), 11–27.\n\nJeng, S. W. and Kiliçman, A. (2020). Series expansion and fourth-order global Padé approximation for a rough Heston solution. *Mathematics* 8(11), 1968. [doi:10.3390/math8111968](https://doi.org/10.3390/math8111968).\n\nJeng, S. W. and Kiliçman, A. (2021). SPX calibration of option approximations under rough Heston model. *Mathematics* 9(21), 2675. [doi:10.3390/math9212675](https://doi.org/10.3390/math9212675).\n\nKahl, C. and Jäckel, P. (2005). Not-so-complex logarithms in the Heston model. *Wilmott Magazine*, September, 94–103.\n\nKopteva, N. (2022). Pointwise-in-time a posteriori error control for time-fractional parabolic equations. *Applied Mathematics Letters* 123, 107515. [doi:10.1016/j.aml.2021.107515](https://doi.org/10.1016/j.aml.2021.107515).\n\nLewis, A. L. (2001). A simple option formula for general jump-diffusion and other exponential Lévy processes. Working paper, Envision Financial Systems and OptionCity.net.\n\nLi, L. and Liu, J.-G. (2018). A generalized definition of Caputo derivatives and its application to fractional ODEs. *SIAM Journal on Mathematical Analysis* 50(3), 2867–2900. [doi:10.1137/17M1160318](https://doi.org/10.1137/17M1160318).\n\nOlver, F. W. J., Lozier, D. W., Boisvert, R. F. and Clark, C. W. (2010). *NIST Handbook of Mathematical Functions*. Cambridge University Press.\n\nRømer, S. E. (2022). Empirical analysis of rough and classical stochastic volatility models to the SPX and VIX markets. *Quantitative Finance* 22(10), 1805–1838.\n\nSimon, T. (2014). Comparing Fréchet and positive stable laws. *Electronic Journal of Probability* 19, paper 16. [arXiv:1310.1888](https://arxiv.org/abs/1310.1888v2).\n\nSimon, T. (2015). Mittag-Leffler functions and complete monotonicity. *Integral Transforms and Special Functions* 26(1), 36–50. [doi:10.1080/10652469.2014.965704](https://doi.org/10.1080/10652469.2014.965704).\n\nTrefethen, L. N. and Weideman, J. A. C. (2014). The exponentially convergent trapezoidal rule. *SIAM Review* 56(3), 385–458. [doi:10.1137/130932132](https://doi.org/10.1137/130932132).\n"
+ "sourceEditedAt": "2026-11-19T11:45:08-05:00",
+ "sourceCommit": "038db24a3aadf9600edef385dfd7508e612d53c8",
+ "sourceBlob": "12b074bc25984ea11f9fc7d75768800c267a43eb",
+ "markdown": "## Abstract\n\nEvery option on one underlying and one expiry is priced from a single characteristic function, so a numerical error in that function is a common factor across strikes: a position hedged in strike space is hedged against it, and a position spread over many strikes diversifies it. We make this exact and build on it a deterministic certificate, a bound on the distance between a stored vector of prices and the exact model prices in any portfolio direction. Two results carry the construction. The first holds for every model priced by Fourier inversion: the exposure of a portfolio to the transform error at a given frequency is the Fourier transform of its payoff, so that low-frequency error cancels in spreads and butterflies while high-frequency error averages out across strikes at a rate set by the Herfindahl index of the position. The second bounds the transform error in the rough Heston model: a residual in the fractional Riccati equation moves the log-characteristic function by at most the residual integrated against a forward variance curve, the curve of the same model under the measure midway between the risk-neutral and share measures. The constant is one, the leverage effect supplies the damping, the hypothesis on the curve is the condition under which the model is known to exist, and the argument carries over to Volterra Heston models with completely monotone kernels. On S&P 500 quotes of 18 June 2021, the six-month 4400–4500 call spread, quoted about 2.5 index points wide, is certified to 0.115 points. At three months the same spread is certified to 0.233 points when its legs are netted and to 0.252 when they are not, with every input identical, and the two bounds fall on opposite sides of a quarter-point tolerance.\n\n**Keywords:** rough volatility; Volterra Heston models; fractional Riccati equation; Fourier pricing; a posteriori error bounds; model risk; netting.\n\n**JEL classification:** C63, G13, G12.\n\n**MSC 2020:** 91G20, 91G60, 65G20, 45D05, 26A33.\n\n## 1. Introduction\n\nThe characteristic function of the Heston model is known in closed form, and for more than a decade one of its two standard expressions carried a flaw that went largely unnoticed. Written as in Heston (1993), the formula crosses a branch cut of the complex logarithm at long maturities, and an implementation on the principal branch returns a discontinuous transform (Kahl and Jäckel 2005; Albrecher, Mayer, Schoutens and Tistaert 2007). The error entered every strike of a maturity together, because every strike is an integral of the same function. A second episode shows the other side of the same fact. The expansion of Hagan, Kumar, Lesniewski and Woodward (2002) for the SABR model tracks the exact implied volatility closely, strike by strike, yet at low strikes and long expiries the density it implies turns negative: a butterfly, which is a portfolio of three options, receives a negative price (Hagan, Kumar, Lesniewski and Woodward 2014). Accuracy in each price had not delivered accuracy in a direction.\n\nBoth episodes say that the accuracy of an option pricer is a property of directions in strike space. This paper makes that statement quantitative. Its portfolio results hold for every pricer built on a common transform. Its bound on the transform itself is proved for the rough Heston model of El Euch and Rosenbaum (2019), whose transform solves a fractional Riccati equation with no closed form, so that the prices used in practice come from approximations: the rational approximation of Gatheral and Radoičić (2019), fractional Adams schemes (Diethelm, Ford and Freed 2004), Markovian lifts (Abi Jaber and El Euch 2019b; Abi Jaber 2019) and neural-network surrogates (Horvath, Muguruza and Tomas 2021). Their accuracy is usually reported by comparison with a slower solver on a test set. Agreement between two solvers bounds neither.\n\nWe ask what can be proved after the fact. A vector \\(c^{\\rm fast}\\) of rough Heston prices has been computed and stored. Let \\(c^*\\) be the exact model prices of the same contracts, normalised by the forward \\(F_T\\) and the discount factor \\(D\\). For a portfolio \\(w\\) and a monetary tolerance \\(\\tau\\), the task is to prove\n\\[\nDF_T\\left|w^\\top(c^*-c^{\\rm fast})\\right|\\le\\tau.\n\\tag{1.1}\n\\]\nThe stored output is taken as given. An independent reference \\(\\bar c\\) is built with its own error bound, and the stored output is located relative to it:\n\\[\nc^*-c^{\\rm fast}=d+\\operatorname{Re}(Az)+r,\\qquad d=\\bar c-c^{\\rm fast}.\n\\tag{1.2}\n\\]\nHere \\(z_n\\) is the error of the reference transform at the frequency \\(u_n\\), one complex number shared by all strikes; \\(A\\) is the matrix of known pricing coefficients; and \\(r\\) collects quadrature, the infinite tail and arithmetic. Two questions follow. How large can \\(z_n\\) be, and how should the bounds on the \\(z_n\\) be combined for a portfolio?\n\n### 1.1. What a Riccati residual costs\n\nA reference trajectory \\(\\widehat h\\) fails to solve the Riccati equation by a residual \\(r(t)\\), which can be bounded on every time interval by exact arithmetic. The first result converts that residual into an error of the log-characteristic function \\(L_T\\). For a reference in the left half-plane and an admissible forward variance curve,\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\int_0^T\\xi^{\\lambda}(T-t)\\,|r(t)|\\,dt\\le\\int_0^T\\xi(T-t)\\,|r(t)|\\,dt,\n\\qquad \\lambda=\\tfrac12|\\rho|\\nu .\n\\tag{1.3}\n\\]\nThe weight \\(\\xi^\\lambda\\) is a forward variance curve: it is the expected variance of the same model under the measure whose Girsanov kernel is half that of the share measure, which is the measure the Lewis contour sees (Theorem 3.1 and Proposition 3.2). Three features of (1.3) deserve emphasis. The constant is one; there is no Grönwall exponential, although the equation is nonlinear and carries its whole history. The damping rate \\(\\lambda\\) is the amount by which the leverage effect speeds up mean reversion under that measure, so the feature of the model that produces the equity skew is also the one that makes numerical errors contract. And the single hypothesis on the curve, nonnegativity of its Riemann–Liouville derivative, is the condition under which El Euch and Rosenbaum (2018) and Abi Jaber and El Euch (2019a) construct the model with nonnegative variance. The certificate asks nothing of the curve that the model does not already ask.\n\nThe bound is the rough counterpart of variation of constants for the Heston Riccati equation. A fractional equation has no semigroup to restart, and two classical tools take its place: convexity of the Caputo derivative on a modulus and positivity of the scalar fractional resolvent. On the six-month example of Section 7, replacing a uniform bound on the state by (1.3) lowers the complete error bound of a call spread from 1.398 to 0.379 index points, a reduction of 72.91 per cent, with the stored prices and the reference unchanged. The same proof covers every damped contour and Volterra Heston models with completely monotone kernels, the classical and the lifted Heston models among them (Remarks 3.5 and 3.6).\n\n### 1.2. Transform error as a common factor\n\nOnce each \\(z_n\\) lies in a disk of radius \\(\\epsilon_n\\), the set of price errors compatible with (1.2) is a sum of ellipses, one for each frequency, and its support in the direction \\(w\\) can be written down exactly (Theorem 5.1):\n\\[\n\\sup\\,w^\\top\\operatorname{Re}(Az)=\\frac{h_*}\\pi\\sum_n\\omega_n\\,\\epsilon_n\\,\\frac{|\\widehat w(u_n)|}{u_n^2+\\frac14},\n\\qquad\n\\widehat w(u)=\\sum_iw_i\\sqrt{m_i}\\,e^{-iuk_i},\n\\tag{1.4}\n\\]\nwhere \\(k_i=\\log m_i\\) is log-moneyness, \\(h_*\\) is the grid step and \\(\\omega_n\\) are the trapezoidal weights. The function \\(\\widehat w\\) is the Fourier transform of the payoff of the portfolio. A portfolio of options with a common expiry is one payoff, and (1.4) prices its error as one payoff. Bounding each leg separately replaces \\(|\\widehat w(u)|\\) by \\(\\sum_i|w_i|\\sqrt{m_i}\\), and the ratio of the two, which we call the netting ratio, measures what is lost.\n\nThe netting ratio has a simple anatomy. For a vertical spread with log-strike gap \\(\\Delta\\) it is \\(|\\sin(u\\Delta/2)|\\) up to terms of order \\(\\Delta\\): the spread is hedged against transform error whose wavelength is long compared with its strike spacing and fully exposed at the Nyquist frequency \\(\\pi/\\Delta\\). For a butterfly the ratio is squared. For any portfolio its mean square over a long band of frequencies is the Herfindahl index of the position (Propositions 5.2 and 5.3). Transform error therefore behaves like a factor model in frequency. Low-frequency error is systematic across strikes and is removed by hedging. High-frequency error is idiosyncratic across strikes and is removed by diversification. On the S&P 500 example the eleven adjacent call spreads keep on average 0.708 of their leg-by-leg node error, against a Herfindahl benchmark of \\(1/\\sqrt2\\approx0.707\\); an equal-weight basket of all twelve calls keeps 0.076.\n\nThis part of the paper is model-free. It applies to any pricer that obtains a strip of options from one transform, whatever the model and whatever the method that produced the transform.\n\n### 1.3. Evidence and economic scale\n\nSection 7 applies the certificate to S&P 500 index options quoted on 18 June 2021 (Jeng and Kiliçman 2021), with the index at 4221.86. An SPX option has a multiplier of 100 dollars per index point and a minimum tick of 0.05 points below a premium of 3.00 and 0.10 above (Cboe Global Markets 2024); in 2025 an average of 3.9 million contracts changed hands each day (Cboe Global Markets 2026). On the sample date the six-month calls between 3700 and 4800 were quoted between 0.8 and 2.5 points wide. A tolerance of a quarter point is 25 dollars a contract and about one fifth of the quoted width of the 4400 call. Our tolerances of 1, 1/2 and 1/4 point thus ask whether numerical error is safely inside the bid–ask spread.\n\nThe answers are as follows. The stored rational-approximation output for the six-month 4400–4500 call spread is certified to 0.115 points, against a quoted width of about 2.5. At three months the netted bound is 0.233 points and the leg-by-leg bound 0.252, from the same stored output, reference centre, 1,025 node radii and remainders; only netting certifies a quarter point. Across 28 spreads, butterflies and baskets at six months, a half-point tolerance is certified for all 28 with netting and for 13 without. Five neighbouring roughness candidates, from \\(H=0.02\\) to \\(H=0.05\\), are ranked by their least-squares distance to the quotes with certified separation in all ten pairs, against seven of ten without netting, so the ranking belongs to the model and the quotes. And at the best candidate the model price of the 4400 call lies at least 0.49 points above the quoted ask: a discrepancy between model and market that the certificate attributes to the model.\n\nThe setting is one that supervisors already describe in portfolio terms. European prudent-valuation rules require a model-risk valuation adjustment for each valuation model and aggregate the individual adjustments with a flat factor of 50 per cent (Commission Delegated Regulation (EU) 2016/101, Article 11 and Annex). The Basel market-risk framework tests, desk by desk, whether the prices of the risk engine track those of the front office, with the stated aim to \"measure the materiality of simplifications\" in the former (Basel Committee on Banking Supervision 2019, MAR32.20). United States guidance asks that model risk be assessed \"both individually and in aggregate\", where the aggregate reflects reliance on common methodologies (Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency 2026). A shared transform is a common methodology in the most literal sense. For the numerical component of valuation uncertainty, (1.4) replaces a flat diversification factor by the exact one.\n\n### 1.4. Relation to the literature\n\nThe rough Heston model arises as the scaling limit of Hawkes-based microstructure models (El Euch, Fukasawa and Rosenbaum 2018) and belongs to the affine Volterra class (Abi Jaber, Larsson and Pulido 2019); its forward variance form is due to El Euch and Rosenbaum (2018) and Gatheral and Keller-Ressel (2019). The empirical case for rough volatility, and the debate around it, rests on precise pricing at the short end, where numerical methods are under most strain (Gatheral, Jaisson and Rosenbaum 2018; Bayer, Friz and Gatheral 2016; Rømer 2022; Cont and Das 2024).\n\nGatheral and Radoičić (2019, 2024) derive rational approximations by matching the short- and long-time behaviour of the fractional Riccati solution, and Jeng and Kiliçman (2020, 2021) study global Padé approximations and their calibration to SPX. The stored output we audit is the third-order approximation of Gatheral and Radoičić; Section 6 records a structural property of it on the parameter set we use. Bayer and Breneis (2023a, 2023b) study Markovian kernel approximations and weak simulation schemes. Boyarchenko, de Innocentis and Levendorskii (2025) study reliable rough Heston pricing and calibration, and their modified Adams method is one of our controls. Ben Hammouda, Ben Romdhane, Samet and Tempone (2026) allocate time-discretisation and quadrature errors in single- and multilevel methods, and Hager and Kreher (2026) expand the model in the Hurst parameter. These works bound or estimate the error of a method. Our object is the error of a given output, with every remainder charged, in a given portfolio direction.\n\nThe analytical tools are established. Caputo convexity is due to Li and Liu (2018), pointwise a posteriori control of fractional equations by positive resolvents to Kopteva (2022), complete monotonicity of Mittag-Leffler functions to Simon (2015), and the strip analysis of the trapezoidal rule to Trefethen and Weideman (2014). What we add is their composition along the valuation map, from the Riccati residual through the forward variance curve to the payoff transform, the identification of the resulting kernel as a forward variance curve under a tilted measure, and the frequency anatomy of portfolio error.\n\nSection 2 states the pricing problem. Section 3 proves the residual bound and Section 4 turns it into transform disks and price bounds. Section 5 treats portfolios. Section 6 describes the stored output and Section 7 the evidence. Appendices A–E contain the model verification, the complete proofs and the experimental record.\n\n## 2. The pricing problem\n\n### 2.1. Model, transform and price\n\nFix a maturity \\(T\\) and let \\(S\\) be the forward price for delivery at \\(T\\), so that \\(S_0=F_T\\). Under the pricing measure \\(\\mathbb Q\\),\n\\[\ndS_t=S_t\\sqrt{V_t}\\,dB_t,\\qquad\nV_t=\\xi(t)+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW_s,\\qquad\nd\\langle B,W\\rangle_t=\\rho\\,dt,\n\\tag{2.1}\n\\]\nwith \\(g_\\beta(t)=t^{\\beta-1}/\\Gamma(\\beta)\\), roughness index \\(\\alpha=H+\\tfrac12\\in(\\tfrac12,1)\\), volatility of variance \\(\\nu>0\\) and correlation \\(-1\\le\\rho\\le0\\). The function \\(\\xi(t)=\\mathbb E V_t\\) is the forward variance curve. In this form the model has no separate mean-reversion parameters: whatever drift the variance has is recorded in \\(\\xi\\), which is in principle observable from variance swaps (El Euch and Rosenbaum 2018; Gatheral and Keller-Ressel 2019).\n\nThroughout, convolutions start at zero, \\((f*g)(t)=\\int_0^tf(t-s)g(s)\\,ds\\), and\n\\[\nI^\\alpha f=g_\\alpha*f,\\qquad D^\\alpha v=g_{1-\\alpha}*v'\\quad(v\\in AC),\n\\tag{2.2}\n\\]\nare the fractional integral and the Caputo derivative. We work on the contour of Lewis (2001), on which the transform is \\(\\phi_T(u-\\tfrac i2)=\\mathbb E[(S_T/F_T)^{1/2+iu}]\\) for real \\(u\\). It is given by a fractional Riccati equation (El Euch and Rosenbaum 2019):\n\\[\nD^\\alpha h=F_u(h),\\quad h(0)=0,\\qquad\nF_u(h)=-\\tfrac12\\left(u^2+\\tfrac14\\right)+\\left(\\tfrac12\\rho\\nu+i\\rho\\nu u\\right)h+\\tfrac12\\nu^2h^2,\n\\tag{2.3}\n\\]\n\\[\n\\phi_T(u-\\tfrac i2)=e^{L_T(u)},\\qquad\nL_T(u)=\\int_0^T\\xi(T-t)\\,D^\\alpha h(t,u)\\,dt .\n\\tag{2.4}\n\\]\nWith \\(m=K/F_T\\), \\(k=\\log m\\) and \\(c=C/(DF_T)\\), the normalised price of a call struck at \\(K\\) is\n\\[\nc=1-\\frac{\\sqrt m}{\\pi}\\int_0^\\infty\\operatorname{Re}\\frac{e^{-iuk}\\,\\phi_T(u-\\tfrac i2)}{u^2+\\frac14}\\,du .\n\\tag{2.5}\n\\]\n\n**Proposition 2.1 (the exact Riccati trajectory).** Let \\(\\alpha\\in(\\tfrac12,1)\\), \\(-1\\le\\rho\\le0\\) and \\(u\\in\\mathbb R\\). Equation (2.3) has a unique global solution. It is absolutely continuous on every finite interval, continuously differentiable at positive times, and satisfies \\(\\operatorname{Re}h(t)<0\\) for \\(t>0\\). With \\(b_u=\\tfrac12(u^2+\\tfrac14)\\) and \\(\\lambda_*=\\tfrac12|\\rho|\\nu\\),\n\\[\n|h(t)|\\le\\frac{b_u}{\\lambda_*}\\left\\{1-E_\\alpha(-\\lambda_*t^\\alpha)\\right\\}\\le\\frac{b_u\\,t^\\alpha}{\\Gamma(1+\\alpha)},\n\\tag{2.6}\n\\]\nwhere \\(E_\\alpha\\) is the Mittag-Leffler function and the last bound remains valid for \\(\\rho=0\\).\n\n**Proof.** Local Volterra contraction gives existence and uniqueness. On the imaginary axis the real part of \\(F_u\\) is at most \\(-\\tfrac18\\) after completing a square, and the fractional maximum principle then prevents the trajectory from leaving the left half-plane. Convexity of a regularised modulus and comparison with the scalar linear equation give (2.6), which excludes blow-up, and the continuation keeps the preceding history. Appendix C.1 gives the details. \\(\\square\\)\n\nThe proposition identifies the object to be approximated without reference to any numerical scheme. Its two conclusions, the invariant half-plane and the scale \\(\\lambda_*\\), are the source of the stability used below.\n\n### 2.2. Two faces of the exponent\n\nThe exponent \\(L_T\\) has a second expression, and the pair is what makes residual control possible.\n\n**Proposition 2.2 (two faces).** Let \\(\\xi\\in AC[0,T]\\) and define\n\\[\nq=\\frac{d}{dt}\\left(g_{1-\\alpha}*\\xi\\right)=\\xi(0)\\,g_{1-\\alpha}+g_{1-\\alpha}*\\xi' .\n\\tag{2.7}\n\\]\nThen \\(q\\in L^1(0,T)\\), \\(g_\\alpha*q=\\xi\\), and for every \\(v\\in AC[0,T]\\) with \\(v(0)=0\\),\n\\[\n\\int_0^T\\xi(T-t)\\,D^\\alpha v(t)\\,dt=\\int_0^Tq(T-t)\\,v(t)\\,dt .\n\\tag{2.8}\n\\]\n\n**Proof.** By Fubini's theorem the left side is \\(((g_{1-\\alpha}*\\xi)*v')(T)\\). The function \\(g_{1-\\alpha}*\\xi\\) is absolutely continuous and vanishes at zero, as does \\(v\\), so integration by parts moves the derivative and gives \\((q*v)(T)\\). The identity \\(g_\\alpha*g_{1-\\alpha}=1\\) gives \\(g_\\alpha*q=\\xi\\). Appendix C.2 records the integrability bounds. \\(\\square\\)\n\nThe function \\(q\\) is the Riemann–Liouville derivative of the curve, and it has a direct meaning. If the variance is written with an explicit drift, \\(V_t=V_0+\\int_0^tg_\\alpha(t-s)\\theta(s)\\,ds+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW_s\\), then \\(\\xi=V_0+g_\\alpha*\\theta\\) and\n\\[\nq(t)=\\frac{V_0\\,t^{-\\alpha}}{\\Gamma(1-\\alpha)}+\\theta(t).\n\\tag{2.9}\n\\]\nThus \\(q\\) is the rate at which variance is supplied: the drift \\(\\theta\\), together with the term a power-law memory requires in order to hold the initial level \\(V_0\\) in place. Applied to the exact solution, (2.8) reads \\(L_T=\\int_0^T\\xi(T-t)F_u(h(t))\\,dt=\\int_0^Tq(T-t)h(t)\\,dt\\), and both forms appear in El Euch and Rosenbaum (2018, Corollaries 3.2 and 3.3). The first face says what the exponent is, the Riccati polynomial accumulated against forward variance. The second says how the exponent responds to an error in the state: linearly, with weight \\(q\\).\n\n**Assumption 2.3 (admissible curve).** \\(\\xi\\in AC[0,T]\\), \\(\\xi(0)\\ge0\\), and \\(q\\ge0\\) almost everywhere on \\((0,T)\\).\n\nUnder this assumption \\(\\xi=g_\\alpha*q\\ge0\\). The assumption is the model's own. In the notation of (2.9) it says that \\(\\theta(t)\\,dt+V_0\\,t^{-\\alpha}dt/\\Gamma(1-\\alpha)\\) is a nonnegative measure, which is the condition imposed by El Euch and Rosenbaum (2018, condition (4)) and by Abi Jaber and El Euch (2019a, Example 2.2) to construct the model with nonnegative variance. Every nondecreasing curve with \\(\\xi(0)\\ge0\\) is admissible. A decreasing curve is admissible as long as it does not fall faster than the memory of the kernel allows: \\(\\xi(t)=V_0(1-ct)\\) is admissible on \\([0,T]\\) exactly when \\(cT\\le1-\\alpha\\) (Corollary C.4). Appendix A verifies the hypotheses of the existence and transform theorems for the curve used in Section 7.\n\n## 3. What a Riccati residual costs\n\n### 3.1. A martingale that fails to be one\n\nLet \\(\\widehat h\\in AC[0,T]\\) with \\(\\widehat h(0)=0\\) be an approximate solution of (2.3), and let\n\\[\nr=D^\\alpha\\widehat h-F_u(\\widehat h)\n\\tag{3.1}\n\\]\nbe its residual. A formal computation shows what \\(r\\) should cost. Write \\(\\xi_t(s)=\\mathbb E[V_s\\mid\\mathcal F_t]\\) for the forward variance curve seen at time \\(t\\), so that \\(d\\xi_t(s)=\\nu g_\\alpha(s-t)\\sqrt{V_t}\\,dW_t\\), put \\(X_t=\\log(S_t/S_0)\\), and consider\n\\[\n\\widehat M_t=\\exp\\left\\{\\left(\\tfrac12+iu\\right)X_t+\\int_t^TD^\\alpha\\widehat h(T-s)\\,\\xi_t(s)\\,ds\\right\\}.\n\\]\nAt \\(t=T\\) this is \\((S_T/S_0)^{1/2+iu}\\) and at \\(t=0\\) it is \\(e^{\\widehat L_T}\\), where \\(\\widehat L_T=\\int_0^T\\xi(T-t)D^\\alpha\\widehat h(t)\\,dt\\). Itô's formula and the identity \\(\\int_t^TD^\\alpha\\widehat h(T-s)\\,g_\\alpha(s-t)\\,ds=\\widehat h(T-t)\\) give\n\\[\nd\\widehat M_t=-\\,r(T-t)\\,V_t\\,\\widehat M_t\\,dt+dN_t\n\\]\nwith \\(N\\) a local martingale, and taking expectations,\n\\[\n\\phi_T(u-\\tfrac i2)-e^{\\widehat L_T}=-\\int_0^Tr(T-t)\\,\\mathbb E\\left[V_t\\widehat M_t\\right]dt .\n\\tag{3.2}\n\\]\nFor the exact solution \\(r=0\\), and this is the usual verification of (2.4). For an approximation the would-be martingale drifts, and it drifts in proportion to variance, because per unit of variance the Riccati polynomial is the generator of the pair \\((X,\\xi)\\) acting on exponentials. A residual at time to maturity \\(T-t\\) therefore costs the variance expected at calendar time \\(t\\), weighted by \\(\\widehat M_t\\). Since \\(|\\widehat M_t|\\) carries the factor \\((S_t/S_0)^{1/2}\\), the weighting is that of a measure tilted halfway towards the share measure.\n\nIdentity (3.2) is the heuristic behind the theorem that follows, which proves the same conclusion for the exponent by deterministic means and with a weight that depends only on the model.\n\n### 3.2. The residual bound\n\nFor \\(\\lambda\\ge0\\) let \\(k_\\lambda(t)=t^{\\alpha-1}E_{\\alpha,\\alpha}(-\\lambda t^\\alpha)\\) be the resolvent kernel of the scalar fractional equation, the solution of \\(k_\\lambda+\\lambda\\,g_\\alpha*k_\\lambda=g_\\alpha\\), with \\(k_0=g_\\alpha\\). For an admissible curve define\n\\[\n\\xi^\\lambda=k_\\lambda*q,\\qquad\\text{equivalently}\\qquad\\xi^\\lambda+\\lambda\\,g_\\alpha*\\xi^\\lambda=\\xi .\n\\tag{3.3}\n\\]\n\n**Theorem 3.1 (a residual is priced by forward variance).** Let \\(\\alpha\\in(\\tfrac12,1)\\), \\(-1\\le\\rho\\le0\\), \\(u\\in\\mathbb R\\), and let \\(\\xi\\) satisfy Assumption 2.3. Let \\(h\\) be the solution of (2.3) and let \\(\\widehat h\\in AC[0,T]\\) satisfy \\(\\widehat h(0)=0\\) and \\(\\operatorname{Re}\\widehat h\\le\\varepsilon\\) on \\([0,T]\\), with \\(\\lambda=\\tfrac12|\\rho|\\nu-\\tfrac12\\nu^2\\varepsilon\\ge0\\). Suppose the residual (3.1) satisfies \\(|r|\\le R\\) almost everywhere for some \\(R\\in L^\\infty(0,T)\\). Then, for \\(0\\le t\\le T\\),\n\\[\n|h(t)-\\widehat h(t)|\\le(k_\\lambda*R)(t),\n\\tag{3.4}\n\\]\nand the exponents \\(L_T=\\int_0^T\\xi(T-t)D^\\alpha h(t)\\,dt\\) and \\(\\widehat L_T=\\int_0^T\\xi(T-t)D^\\alpha\\widehat h(t)\\,dt\\) satisfy\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\int_0^T\\xi^\\lambda(T-t)\\,R(t)\\,dt\\le\\int_0^T\\xi(T-t)\\,R(t)\\,dt .\n\\tag{3.5}\n\\]\nIn particular, if \\(R\\le\\delta\\) then \\(|L_T-\\widehat L_T|\\le\\delta\\int_0^T\\xi(s)\\,ds\\).\n\n**Proof.** Put \\(e=h-\\widehat h\\). Because \\(F_u\\) is quadratic, subtraction gives an exact linear equation,\n\\[\nD^\\alpha e=\\left\\{\\tfrac12\\rho\\nu+i\\rho\\nu u+\\tfrac12\\nu^2(h+\\widehat h)\\right\\}e-r,\\qquad e(0)=0,\n\\tag{3.6}\n\\]\nand by Proposition 2.1 the coefficient has real part at most \\(-\\lambda\\). The Caputo derivative of a convex function of a path is at most the gradient applied to the Caputo derivative of the path. Applied to a regularised modulus this gives \\(D^\\alpha|e|+\\lambda|e|\\le|r|\\le R\\), and positivity of the resolvent \\(k_\\lambda\\) gives (3.4). By Proposition 2.2, \\(L_T-\\widehat L_T=\\int_0^Tq(T-t)e(t)\\,dt\\). Since \\(q\\ge0\\), (3.4) gives \\(|L_T-\\widehat L_T|\\le(q*k_\\lambda*R)(T)=(\\xi^\\lambda*R)(T)\\). Finally \\(g_\\alpha-k_\\lambda=\\lambda\\,g_\\alpha*k_\\lambda\\ge0\\), so \\(\\xi^\\lambda\\le g_\\alpha*q=\\xi\\). Appendix C.2 supplies the regularity, including the passage to the terminal time. \\(\\square\\)\n\nThe bound is global: it holds for residuals of any size and at every frequency. The exact solution enters through the sign of its real part alone, and the weight in (3.5) is fixed by the equation and the curve.\n\n### 3.3. The weight is a forward variance curve\n\nThe kernel \\(\\xi^\\lambda\\) was defined by a Volterra equation. It is also an expectation.\n\n**Proposition 3.2 (the kernel under a tilted measure).** Let \\(\\xi\\) be admissible and, for \\(\\eta\\in[0,1]\\), let \\(\\mathbb Q^\\eta\\) be the measure with density \\(\\mathcal E(\\eta\\int_0^{\\cdot}\\sqrt{V_s}\\,dB_s)_T\\) with respect to \\(\\mathbb Q\\). Under \\(\\mathbb Q^\\eta\\) the variance follows a rough Heston equation with the same supply \\(q\\), the same \\(\\nu\\) and mean-reversion speed \\(\\eta|\\rho|\\nu\\), and\n\\[\n\\mathbb E^{\\mathbb Q^\\eta}[V_t]=\\xi^{\\eta|\\rho|\\nu}(t),\\qquad0\\le t\\le T .\n\\tag{3.7}\n\\]\nIn particular the weight of Theorem 3.1 for a reference in the left half-plane is \\(\\xi^{\\lambda_*}(t)=\\mathbb E^{\\mathbb Q^{1/2}}[V_t]\\).\n\n**Proof.** The stochastic exponential is a true martingale (Abi Jaber, Larsson and Pulido 2019, Lemma 7.3, whose argument applies to an admissible input curve). By Girsanov's theorem \\(W^\\eta=W-\\eta\\rho\\int_0^{\\cdot}\\sqrt{V_s}\\,ds\\) is a Brownian motion under \\(\\mathbb Q^\\eta\\), and (2.1) becomes\n\\[\nV_t=\\xi(t)-\\eta|\\rho|\\nu\\int_0^tg_\\alpha(t-s)V_s\\,ds+\\nu\\int_0^tg_\\alpha(t-s)\\sqrt{V_s}\\,dW^\\eta_s .\n\\]\nThe moment bounds for affine Volterra processes make the stochastic integral centred, so \\(m(t)=\\mathbb E^{\\mathbb Q^\\eta}[V_t]\\) solves \\(m+\\eta|\\rho|\\nu\\,g_\\alpha*m=\\xi\\), whose unique solution is \\(\\xi^{\\eta|\\rho|\\nu}\\) by (3.3). \\(\\square\\)\n\nThe measures \\(\\mathbb Q^\\eta\\) interpolate between the pricing measure at \\(\\eta=0\\) and the share measure, whose numéraire is the underlying, at \\(\\eta=1\\). Under the share measure the variance of an equity index reverts faster, by \\(|\\rho|\\nu\\): paths on which the index is high receive more weight, and on those paths the leverage effect makes variance low. This is the adjustment that separates the two probabilities in the formula of Heston (1993). The Lewis contour prices the claim \\(S_T^{1/2+iu}\\), whose modulus sits midway between the two numéraires, and it inherits half of the adjustment. Theorem 3.1 then reads as follows.\n\n*On the Lewis contour, a Riccati residual at time to maturity \\(s\\) changes the log-characteristic function by at most the variance expected at time \\(T-s\\) under \\(\\mathbb Q^{1/2}\\), per unit of residual.*\n\nFour remarks place the statement.\n\n**Remark 3.3 (the classical model).** At \\(\\alpha=1\\) the resolvent is \\(k_\\lambda(t)=e^{-\\lambda t}\\) and \\(\\xi^\\lambda(t)=V_0e^{-\\lambda t}+\\int_0^te^{-\\lambda(t-s)}\\theta(s)\\,ds\\). Inequality (3.5) is then variation of constants for the Heston Riccati equation, in which the error is transported by a flow that contracts at rate \\(\\lambda\\). For \\(\\alpha<1\\) there is no flow to transport it. Convexity of the Caputo derivative and positivity of \\(k_\\lambda\\) recover the same conclusion with the same constant.\n\n**Remark 3.4 (the constant).** The constant one in (3.5) is sharp. Write \\(a\\) for the argument of \\(\\phi_T\\), so that the Lewis contour is \\(a=u-\\tfrac i2\\). At \\(a=0\\), the point of zero frequency on the undamped contour, the exact solution vanishes, \\(\\lambda=0\\), and a residual of constant sign produces \\(\\widehat L_T-L_T=\\int_0^T\\xi(T-t)r(t)\\,dt\\) up to terms of second order in \\(r\\). On the Lewis contour the exact solution adds damping of its own through \\(\\operatorname{Re}h<0\\), growing with the frequency, and Theorem 3.1 uses the part that is uniform in the frequency.\n\n**Remark 3.5 (other contours and positive correlation).** On a contour \\(\\operatorname{Im}a=-\\eta\\) with \\(0<\\eta<1\\) the linear coefficient of the Riccati polynomial has real part \\(\\eta\\rho\\nu\\), the left half-plane is again invariant, and the proof gives (3.5) with \\(\\lambda=\\eta|\\rho|\\nu\\) and the measure \\(\\mathbb Q^\\eta\\). The damping available to an error bound thus rises linearly from nothing at the pricing measure to \\(|\\rho|\\nu\\) at the share measure. For \\(\\rho>0\\) the same argument gives (3.4) and the first inequality of (3.5) with a negative \\(\\lambda\\), the resolvent being positive term by term; the kernel \\(\\xi^\\lambda\\) then lies above \\(\\xi\\), and residuals compound at the rate at which variance runs away under the tilted measure.\n\n**Remark 3.6 (other kernels).** The fractional kernel enters the proof through three facts: \\(g_\\alpha*g_{1-\\alpha}=1\\) with \\(g_{1-\\alpha}\\) nonnegative and nonincreasing, which gives the convexity inequality; positivity of the resolvent \\(k_\\lambda\\); and \\(\\xi=g_\\alpha*q\\) with \\(q\\ge0\\), which is the second face. Every completely monotone kernel \\(K\\) has a resolvent of the first kind with the same monotonicity and a completely monotone resolvent (Gripenberg, Londen and Staffans 1990, Chapter 5), its admissible curves are again of the form \\(\\xi=K*q\\) with \\(q\\) a nonnegative measure (Abi Jaber and El Euch 2019a, Example 2.2), and the left half-plane is invariant for its Riccati–Volterra equation (Abi Jaber, Larsson and Pulido 2019, Theorem 7.1). The argument therefore gives (3.5) for Volterra Heston models with completely monotone kernels, with \\(\\xi^\\lambda\\) defined by \\(\\xi^\\lambda+\\lambda K*\\xi^\\lambda=\\xi\\): the classical Heston model, where \\(K\\equiv1\\) and the statement is Remark 3.3; the lifted Heston model of Abi Jaber (2019), where \\(K\\) is a finite sum of exponentials; and gamma kernels. Appendix C writes out the regularity for the fractional kernel.\n\nFigure 1 shows the two weights for the curve calibrated in Section 7. At six months the half-share curve lies below the risk-neutral one by at most 7.5 per cent, the order of \\(\\lambda_*T^\\alpha\\approx0.075\\). The large gain in the experiments comes from the first inequality in (3.5), which weights each residual by the variance still to come, in place of a uniform bound on the state.\n\n\n\n### 3.4. Cell weights\n\nIn computation the residual is bounded on closed time cells. If \\(R\\le R_j\\) on \\([a_j,b_j]\\) and the cells cover \\([0,T]\\), then (3.5) gives\n\\[\n\\left|L_T-\\widehat L_T\\right|\\le\\sum_jR_j\\int_{a_j}^{b_j}\\xi^\\lambda(T-s)\\,ds\\le\\sum_jR_j\\int_{a_j}^{b_j}\\xi(T-s)\\,ds .\n\\tag{3.8}\n\\]\nEach weight is an amount of forward variance, and all of them refer to the history from time zero. For the curve of Section 7, \\(\\xi(t)=\\theta+(V_0-\\theta)E_{\\alpha_0}(-\\lambda_\\xi t^{\\alpha_0})\\), the cumulative weight \\(W_\\lambda(t)=\\int_0^t\\xi^\\lambda(s)\\,ds\\) has the expansion\n\\[\nW_\\lambda(t)=\\sum_{n=0}^\\infty(-\\lambda)^n\\,t^{1+n\\alpha}\n\\left[\\frac{\\theta}{\\Gamma(2+n\\alpha)}+(V_0-\\theta)\\,E_{\\alpha_0,\\,2+n\\alpha}\\left(-\\lambda_\\xi t^{\\alpha_0}\\right)\\right],\n\\tag{3.9}\n\\]\nvalid for \\(\\lambda T^\\alpha<1\\), and the cell \\([a_j,b_j]\\) has weight \\(W_\\lambda(T-a_j)-W_\\lambda(T-b_j)\\). Appendix C.5 bounds the tails of both series, and the implementation encloses every term by outward-rounded dyadic arithmetic.\n\n## 4. From residuals to prices\n\n### 4.1. A reference shaped by the startup expansion\n\nTheorem 3.1 separates the choice of a reference from the validity of its certificate. A useful reference makes its residual accessible on whole time intervals and reproduces the singular start of the exact solution. Write \\(g=D^\\alpha h=F_u(h)\\), so that \\(h=I^\\alpha g\\). Substituting repeatedly into the Volterra form of (2.3) gives, as \\(t\\downarrow0\\),\n\\[\ng(t)=g_0+a_1t^\\alpha+a_2t^{2\\alpha}+O(t^{3\\alpha}),\\qquad g_0=-\\tfrac12\\left(u^2+\\tfrac14\\right),\n\\tag{4.1}\n\\]\nwith coefficients determined by the equation: writing \\(\\gamma_u=\\tfrac12\\rho\\nu+i\\rho\\nu u\\),\n\\[\nb_1=\\frac{g_0}{\\Gamma(1+\\alpha)},\\quad a_1=\\gamma_ub_1,\\qquad\nb_2=\\frac{a_1\\Gamma(1+\\alpha)}{\\Gamma(1+2\\alpha)},\\quad a_2=\\gamma_ub_2+\\tfrac12\\nu^2b_1^2 .\n\\tag{4.2}\n\\]\nThe identity \\(I^\\alpha t^\\beta=\\Gamma(1+\\beta)\\,t^{\\alpha+\\beta}/\\Gamma(1+\\alpha+\\beta)\\) controls each step. We therefore take the reference to be \\(\\widehat h=I^\\alpha\\widehat g\\), where \\(\\widehat g\\) consists of the three startup powers in (4.1) and a continuous piecewise affine remainder. The powers come from the equation, and the affine part can be integrated against the fractional kernel in closed form. Its nodal values may come from any numerical solver, of any accuracy: Theorem 3.1 charges whatever residual results.\n\nAll stored nodes and coefficients are read as the exact dyadic numbers they are. The residual \\(r=\\widehat g-F_u(I^\\alpha\\widehat g)\\) then includes every discrepancy between the stored coefficients and their ideal values. Appendix D.1 bounds it on the startup interval and on every later closed cell, and verifies \\(\\operatorname{Re}\\widehat h\\le0\\) on the whole of \\([0,T]\\). The certificate belongs to the continuous trajectory: the residual is bounded on every closed cell, between the nodes as well as at them.\n\n### 4.2. Transform disks and the price bound\n\nLet \\(\\eta_n\\) bound the exponent error at the frequency \\(u_n\\), obtained from (3.8). Since \\(|\\phi_T(u-\\tfrac i2)|\\le1\\) on the Lewis contour, the reference transform \\(\\widehat\\phi_n=e^{\\widehat L_T(u_n)}\\) satisfies\n\\[\n\\left|\\phi_T(u_n-\\tfrac i2)-\\widehat\\phi_n\\right|\\le\\min\\{1,|\\widehat\\phi_n|\\}\\left(e^{\\eta_n}-1\\right)=:\\epsilon_n .\n\\tag{4.3}\n\\]\nThe passage from the integral (2.5) to a finite sum has three remainders. The first is the trapezoidal rule on the whole axis, which converges geometrically because the integrand is analytic in a strip.\n\n**Theorem 4.1 (strip quadrature).** Let \\(S\\) be a positive martingale, \\(00\\). Replacing the integral in (2.5) by the infinite trapezoidal sum with step \\(h_*\\) changes the normalised price by at most\n\\[\n\\epsilon_{\\rm grid}=\\frac{\\sqrt m\\,e^{a_*|k|}}{(\\tfrac12-a_*)\\left(e^{2\\pi a_*/h_*}-1\\right)} .\n\\tag{4.4}\n\\]\n\nThe second is the tail of the sum beyond the last node. Comparing \\(-\\operatorname{Re}h\\) from below with the solution of a scalar equation gives an envelope \\(|\\phi_T(u-\\tfrac i2)|\\le e^{-c_Vu}\\) for \\(u\\ge V\\), with an explicit \\(c_V>0\\) built from the same forward variance weights, and hence a tail bound \\(\\epsilon_{\\rm tail}\\) for the exact transform (Appendix D.2). The third is arithmetic. With the finite sum\n\\[\n\\widehat c_N=1-\\frac{h_*\\sqrt m}{\\pi}\\left(2\\operatorname{Re}\\widehat\\phi_0+\\sum_{n=1}^N\\frac{\\operatorname{Re}(e^{-iu_nk}\\widehat\\phi_n)}{u_n^2+\\frac14}\\right),\\qquad u_n=nh_*,\n\\tag{4.5}\n\\]\nthe complete price bound for one strike is\n\\[\n|c-\\widehat c_N|\\le\\frac{h_*\\sqrt m}{\\pi}\\left(2\\epsilon_0+\\sum_{n=1}^N\\frac{\\epsilon_n}{u_n^2+\\frac14}\\right)+\\epsilon_{\\rm grid}+\\epsilon_{\\rm tail}+\\epsilon_{\\rm arithmetic} .\n\\tag{4.6}\n\\]\nThe coefficient at the origin is the trapezoidal half-weight times \\(1/(\\tfrac14)\\). A node at which no reference value is computed is given the reference value zero and the radius supplied by the envelope of the exact transform, so every part of the frequency axis is charged exactly once.\n\n## 5. Portfolios: transform error as a common factor\n\nThis section uses the model through formula (2.5) alone. Its statements hold for every positive martingale \\(S\\), every reference transform and every set of radii: for the Heston, Bates and Lévy models as for rough volatility, and for a transform that comes from a closed form, a differential-equation solver, a neural network or the residual bound of Section 3.\n\n### 5.1. The joint error set\n\nFix the model, the maturity and the frequency grid, and consider \\(p\\) strikes with moneyness \\(m_i\\) and log-moneyness \\(k_i\\). From (4.5), the reference prices are \\(\\bar c_i=1+\\operatorname{Re}\\sum_na_{in}\\bar\\phi_n\\) with\n\\[\na_{in}=-\\frac{h_*\\omega_n\\sqrt{m_i}}{\\pi}\\,\\frac{e^{-iu_nk_i}}{u_n^2+\\frac14},\\qquad\\omega_0=\\tfrac12,\\quad\\omega_n=1\\ (n\\ge1),\n\\tag{5.1}\n\\]\nand the exact prices satisfy\n\\[\nc^*-\\bar c=\\operatorname{Re}\\sum_{n=0}^Na_{\\cdot n}z_n+r,\\qquad z_n=\\phi_T(u_n-\\tfrac i2)-\\bar\\phi_n,\\quad r\\in\\mathcal R,\n\\tag{5.2}\n\\]\nwhere \\(\\mathcal R\\) contains the strip, tail and arithmetic remainders. The same complex number \\(z_n\\) enters every strike. The \\(z_n\\) are deterministic unknowns, each confined to a disk.\n\nFor a portfolio \\(w\\in\\mathbb R^p\\) define its payoff transform\n\\[\n\\widehat w(u)=\\sum_{i=1}^pw_i\\sqrt{m_i}\\,e^{-iuk_i},\\qquad\\text{so that}\\qquad\n\\sum_iw_ia_{in}=-\\frac{h_*\\omega_n}{\\pi}\\,\\frac{\\widehat w(u_n)}{u_n^2+\\frac14} .\n\\tag{5.3}\n\\]\nUp to normalisation, \\(\\widehat w(u)/(u^2+\\tfrac14)\\) is the Fourier transform on the Lewis contour of the payoff \\(\\sum_iw_i(S_T-K_i)^+\\). A portfolio of options with one expiry is a single European payoff, and its transform is the sum of the transforms of its legs.\n\n**Theorem 5.1 (joint error set).** Suppose (5.2) holds with \\(|z_n|\\le\\epsilon_n\\) and \\(\\mathcal R\\) compact and convex. Let\n\\[\n\\mathcal E=\\left\\{\\operatorname{Re}\\sum_na_{\\cdot n}z_n:\\ |z_n|\\le\\epsilon_n\\right\\}+\\mathcal R,\\qquad d=\\bar c-c^{\\rm fast}.\n\\tag{5.4}\n\\]\nThen \\(c^*-c^{\\rm fast}\\in d+\\mathcal E\\), and for every \\(w\\in\\mathbb R^p\\) the support function of this set is\n\\[\nh_{d+\\mathcal E}(w)=w^\\top d+\\frac{h_*}\\pi\\sum_{n=0}^N\\omega_n\\,\\epsilon_n\\,\\frac{|\\widehat w(u_n)|}{u_n^2+\\frac14}+h_{\\mathcal R}(w).\n\\tag{5.5}\n\\]\nConsequently \\(DF_T\\,|w^\\top(c^*-c^{\\rm fast})|\\le DF_T\\max\\{h_{d+\\mathcal E}(w),h_{d+\\mathcal E}(-w)\\}\\).\n\n**Proof.** Substituting the actual errors in (5.2) gives the inclusion. The image of the disk \\(|z|\\le\\epsilon\\) under the real-linear map \\(z\\mapsto\\operatorname{Re}(bz)\\) has support \\(\\epsilon|b|\\), attained at \\(z=\\epsilon\\bar b/|b|\\). The disks are independent, supports add under Minkowski sums, and (5.3) identifies \\(b=\\sum_iw_ia_{in}\\). Translation by \\(d\\) adds \\(w^\\top d\\). \\(\\square\\)\n\nGeometrically, each frequency contributes an ellipse to \\(\\mathcal E\\), the image of a disk under \\(z\\mapsto\\operatorname{Re}(a_{\\cdot n}z)\\), and \\(\\mathcal E\\) is the sum of these ellipses and the remainder. Bounding each price separately replaces this body by its bounding box, whose support is\n\\[\nh_{\\rm box}(w)=\\frac{h_*}\\pi\\sum_{n=0}^N\\omega_n\\,\\epsilon_n\\,\\frac{\\sum_i|w_i|\\sqrt{m_i}}{u_n^2+\\frac14}+\\sum_i|w_i|\\rho_i\n\\tag{5.6}\n\\]\nwhen \\(\\mathcal R=\\prod_i[-\\rho_i,\\rho_i]\\). We call (5.5) the joint bound and (5.6), with the same signed centre \\(w^\\top d\\), the leg-by-leg bound. All numerical evaluations of either are outward enclosures.\n\nAny separately proved interval for a price can be intersected with \\(\\bar c+\\mathcal E\\), since the exact vector lies in both sets. Every comparison of the two bounds in Section 7 uses one centre, one set of radii and one remainder, so that the aggregation is the only difference.\n\n### 5.2. The netting ratio\n\nThe two bounds differ only through the factor multiplying each \\(\\epsilon_n\\). Define the netting ratio of \\(w\\) at the frequency \\(u\\) by\n\\[\n\\chi_w(u)=\\frac{|\\widehat w(u)|}{\\sum_i|w_i|\\sqrt{m_i}}\\in[0,1].\n\\tag{5.7}\n\\]\nThe node term of the joint bound is the node term of the leg-by-leg bound with each frequency multiplied by \\(\\chi_w(u_n)\\). By the equality case of the triangle inequality, \\(\\chi_w(u)=1\\) exactly when the numbers \\(w_ie^{-iuk_i}\\) with \\(w_i\\ne0\\) share one argument, so the joint bound is strictly smaller as soon as one node of positive radius sees two legs out of phase.\n\n**Proposition 5.2 (spreads and butterflies).** Let the strikes form a geometric ladder with log-moneyness \\(k,k+\\Delta,\\dots,k+j\\Delta\\), and let \\(w\\) be the \\(j\\)-th difference, with weights \\((-1)^i\\binom ji\\). Then\n\\[\n\\chi_w(u)=\\left(\\tanh^2\\tfrac\\Delta4+\\operatorname{sech}^2\\tfrac\\Delta4\\,\\sin^2\\tfrac{u\\Delta}2\\right)^{j/2}\\le\\left(\\tfrac\\Delta2\\sqrt{u^2+\\tfrac14}\\right)^j .\n\\tag{5.8}\n\\]\nFor the vertical spread (\\(j=1\\)) the ratio equals \\(\\tanh(\\Delta/4)\\) at \\(u=0\\) and reaches one exactly at the odd multiples of \\(\\pi/\\Delta\\). The butterfly (\\(j=2\\)) has the square of the ratio of the spread.\n\n**Proof.** With \\(\\zeta=e^{(1/2-iu)\\Delta}\\) the payoff transform is \\(\\sqrt m\\,e^{-iuk}(1-\\zeta)^j\\) and the gross weight is \\(\\sqrt m\\,(1+e^{\\Delta/2})^j\\). Then \\(|1-\\zeta|^2=(e^{\\Delta/2}-1)^2+4e^{\\Delta/2}\\sin^2(u\\Delta/2)\\), and division by \\((1+e^{\\Delta/2})^2\\) gives the equality. The inequality uses \\(\\tanh x\\le x\\), \\(\\operatorname{sech}x\\le1\\) and \\(|\\sin x|\\le|x|\\). \\(\\square\\)\n\nThe proposition gives the hierarchy of exposures. A single call loads on the transform error at the frequency \\(u\\) with weight \\((u^2+\\tfrac14)^{-1}\\). A vertical spread of width \\(\\Delta\\) loads with weight at most \\(\\Delta\\,(u^2+\\tfrac14)^{-1/2}\\), and a butterfly with weight at most \\(\\Delta^2\\) at every frequency, in units of \\(h_*\\sqrt m\\,(1+e^{\\Delta/2})^j/(2^j\\pi)\\). Each difference in strike removes one power of the frequency from the decay of the payoff transform and returns one power of the strike spacing. This is the error form of the Breeden and Litzenberger (1978) relation: a butterfly reads off the state-price density, and its error is the unweighted sum of the transform errors.\n\nThe proposition also locates the frequencies that matter. A spread is hedged against transform error whose wavelength in log-strike, \\(2\\pi/u\\), is long compared with its width, and it is fully exposed at the Nyquist frequency \\(u=\\pi/\\Delta\\), where the two legs respond in opposite phase. For strikes 100 points apart near 4400, \\(\\Delta\\approx0.0225\\) and \\(\\pi/\\Delta\\approx140\\). For a five-point vertical, the finest spacing commonly listed near the money, \\(\\pi/\\Delta\\approx2800\\). A narrow spread is protected over a band of frequencies that is inversely proportional to its width.\n\nFor portfolios that are not differences the ratio is governed by concentration.\n\n**Proposition 5.3 (the Herfindahl law).** Let the log-strikes \\(k_i\\) be distinct and put \\(\\ell_i=w_i\\sqrt{m_i}\\). Then\n\\[\n\\lim_{U\\to\\infty}\\frac1U\\int_0^U\\chi_w(u)^2\\,du=\\frac{\\sum_i\\ell_i^2}{\\left(\\sum_i|\\ell_i|\\right)^2} .\n\\tag{5.9}\n\\]\n\n**Proof.** Expand \\(|\\widehat w(u)|^2=\\sum_{i,j}\\ell_i\\ell_je^{-iu(k_i-k_j)}\\). The diagonal terms sum to \\(\\sum_i\\ell_i^2\\) and each off-diagonal term has mean zero over long intervals. \\(\\square\\)\n\nThe right side of (5.9) is the Herfindahl index of the position, the reciprocal of its effective number of strikes. Its square root is \\(1/\\sqrt2\\approx0.707\\) for a two-leg spread, \\(\\sqrt6/4\\approx0.612\\) for a butterfly with weights \\((1,-2,1)\\), and \\(1/\\sqrt p\\) for an equal-weight basket of \\(p\\) strikes. At the origin, in contrast, \\(\\chi_w(0)=|\\sum_i\\ell_i|/\\sum_i|\\ell_i|\\) measures the net position. The two regimes have familiar names. Error at low frequency moves all strikes together; it is systematic, and it is removed by hedging. Error at high frequency moves neighbouring strikes independently; it is idiosyncratic, and it is removed by diversification. A long-only basket is fully exposed to the first and almost immune to the second, and a tight spread is the reverse. Figure 2 shows the three profiles.\n\n\n\nThe centre \\(w^\\top d\\) in (5.5) obeys the same logic from the other side. It is the known displacement of the stored output from the reference, and for a smooth approximation it is smooth in strike: it cancels in spreads and accumulates in baskets. Section 7.4 finds both effects in the data.\n\n### 5.3. The calibration objective is a portfolio\n\nCalibration compares model prices with quotes \\(M_i\\) through a least-squares objective\n\\[\nJ(c)=\\frac1{2p}\\sum_{i=1}^p(c_i-M_i)^2 .\n\\]\nIts numerical error is a portfolio error. For the exact prices \\(c^*=\\bar c+e\\) with \\(e\\in\\mathcal E\\),\n\\[\nJ(c^*)-J(\\bar c)=g^\\top e+\\frac{|e|^2}{2p},\\qquad g=\\frac{\\bar c-M}{p},\n\\]\nso to first order the objective is exposed to the transform error as the portfolio whose weights are the pricing residuals.\n\n**Proposition 5.4 (objective enclosure).** If \\(e\\in\\mathcal E\\) and \\(|e_i|\\le\\beta_i\\) for each \\(i\\), then\n\\[\nJ(\\bar c)-h_{\\mathcal E}(-g)\\ \\le\\ J(c^*)\\ \\le\\ J(\\bar c)+h_{\\mathcal E}(g)+\\frac1{2p}\\sum_i\\beta_i^2 .\n\\tag{5.10}\n\\]\n\n**Proof.** The lower bound drops the nonnegative quadratic term and uses \\(g^\\top e\\ge-h_{\\mathcal E}(-g)\\); the upper bound uses \\(g^\\top e\\le h_{\\mathcal E}(g)\\) and \\(|e|^2\\le\\sum_i\\beta_i^2\\). \\(\\square\\)\n\nThe residuals of a reasonable fit change sign across strikes, so the calibration portfolio is long some strikes and short others, and its support under (5.5) is smaller than under (5.6). For a finite set of candidate parameters, a candidate is the certified minimiser when the upper end of its enclosure lies below the lower ends of all the others; in general the candidates whose lower end lies below the smallest upper end form a set that contains every exact minimiser. Each candidate carries its own transform and its own error set.\n\n## 6. The stored output\n\nThe certificate applies to any stored vector of prices once its displacement from the reference is enclosed. In the experiments the stored output is the third-order rational approximation of Gatheral and Radoičić (2019), which writes the Riccati solution as \\(\\widehat H(y)=P(y)/Q(y)\\) in the variable \\(y=\\nu t^\\alpha\\), with cubic \\(P\\) and \\(Q=1+q_1y+q_2y^2+q_3y^3\\), and fixes the six coefficients by matching three terms of the expansion at \\(t=0\\) and three terms of the expansion at infinity. The formula costs a handful of operations per frequency. Its coefficients are complex, and on the parameter set of Section 7 the absence of poles on the positive axis is a theorem.\n\n**Proposition 6.1 (structure of the rational output).** For \\(\\alpha\\in[\\tfrac{13}{25},\\tfrac35]\\), \\(\\rho=-0.7445\\), \\(\\nu>0\\) and every real \\(u\\), the matching system of Gatheral and Radoičić is nonsingular, and\n\\[\n\\operatorname{Re}q_j>0\\ (j=1,2,3),\\qquad\\operatorname{Re}Q(y)\\ge1\\quad(y\\ge0),\\qquad\\operatorname{Re}\\widehat H(y)<0\\quad(y>0).\n\\tag{6.1}\n\\]\n\nThe proof in Appendix B reduces (6.1) to the signs of nine real functions on a compact rectangle in \\((\\alpha,u/(1+u))\\) and verifies them by interval arithmetic on a finite cover. The conclusion is stable under small changes of \\(\\rho\\), since all nine inequalities are strict. The proposition shows that the stored output is well defined and shares the half-plane of the exact solution. Its accuracy is the business of the reference: one construction supplies the price, and an independent one bounds its error.\n\n## 7. Evidence from S&P 500 options\n\n### 7.1. Data and market scale\n\nThe quotes are the bid and ask implied volatilities of twelve SPX options with six months to expiry, strikes 3700 to 4800 in steps of 100, recorded on 18 June 2021 and published by Jeng and Kiliçman (2021). We use the index level 4221.86 of that data set as the forward, with zero carry, and convert each implied volatility to a call price by the Black–Scholes formula in rigorous interval arithmetic. Table 1 gives the result in index points. One point is 100 dollars per contract.\n\n**Table 1. Six-month SPX call quotes of 18 June 2021 in index points: bid, ask and quoted width, each width computed from the unrounded quotes.**\n\n| Strike | Bid | Ask | Width | Strike | Bid | Ask | Width |\n|---:|---:|---:|---:|---:|---:|---:|---:|\n| 3700 | 586.40 | 588.70 | 2.30 | 4300 | 140.00 | 141.50 | 1.50 |\n| 3800 | 502.10 | 504.50 | 2.40 | 4400 | 91.10 | 92.40 | 1.30 |\n| 3900 | 420.20 | 422.60 | 2.40 | 4500 | 55.00 | 56.20 | 1.21 |\n| 4000 | 341.70 | 344.20 | 2.50 | 4600 | 31.60 | 32.60 | 0.99 |\n| 4100 | 267.79 | 270.19 | 2.40 | 4700 | 17.90 | 18.80 | 0.90 |\n| 4200 | 199.69 | 202.20 | 2.51 | 4800 | 10.30 | 11.10 | 0.80 |\n\nThe 4400–4500 call spread is worth about 36.1 points at mid-market, and a trader who crosses both legs pays up to 2.5 points of spread. These are the magnitudes against which a numerical error should be judged. An error of a quarter point is a fifth of the quoted width of the 4400 call, and a tenth of the width of the spread.\n\nThe model parameters are those fitted to this surface by Jeng and Kiliçman (2021): \\(\\nu=0.2897\\), \\(\\rho=-0.7445\\), and the forward variance curve\n\\[\n\\xi(t)=\\theta+(V_0-\\theta)\\,E_{\\alpha_0}\\left(-\\lambda_\\xi t^{\\alpha_0}\\right),\\qquad(\\alpha_0,V_0,\\theta,\\lambda_\\xi)=(0.5286,\\,0.0262,\\,0.0721,\\,0.5037).\n\\tag{7.1}\n\\]\nThe curve rises from an instantaneous volatility of about 16 per cent towards a long-run level of about 27 per cent. It is nondecreasing, hence admissible, and it is held fixed when the roughness index \\(\\alpha\\) of the kernel in (2.1) is varied. The damping rate of Theorem 3.1 is \\(\\lambda_*=\\tfrac12|\\rho|\\nu\\approx0.108\\).\n\nThe stored output is the rational approximation of Section 6 evaluated in binary64, with its own quadrature. The reference of Section 4 uses the Fourier step \\(h_*=\\tfrac18\\), a grid through \\(u=128\\) (1,025 nodes) and 100-bit outward dyadic arithmetic. Three ingredients of the certificate vary across the experiments: the propagation of the residual, either a uniform bound on the state or Theorem 3.1 with one global residual maximum or with 128 time bins; the band on which the reference transform is computed, \\(u\\le64\\) or \\(u\\le128\\), nodes above the band receiving the envelope radius; and the maturity. Appendix E lists the configurations, which carry the labels H0–H4 at six months, Q0–Q2 at three months, and N1, N2 and N2L for the freshly generated nearby fields.\n\n### 7.2. What the forward variance weight buys\n\nTable 2 follows the complete bound for one position, long the 4400 call and short the 4500 call, as the certificate is refined. Every row keeps the stored output, the signed centre, all finite frequencies, the strip, the exact tail and the arithmetic.\n\n**Table 2. Complete certificates for the unit 4400–4500 call spread, in index points.**\n\n| Maturity and certificate | Joint bound | Leg-by-leg bound | Tolerance met |\n|---|---:|---:|---|\n| H0: six months, uniform state bound | 1.397612096 | 2.679930311 | none |\n| H1: six months, Theorem 3.1, global residual | 0.378598956 | 0.476401716 | 1/2 point, both |\n| H2: six months, Theorem 3.1, 128 time bins | 0.367258782 | — | 1/2 point |\n| H3: six months, reference through 128, upper band global | 0.132245064 | 0.158585551 | 1/4 point, both |\n| H4: six months, reference through 128, 128 bins | 0.115215934 | 0.137896018 | 1/4 point, both |\n| Q0: three months, reference through 64, global | 1.207557898 | 1.545191542 | none |\n| Q1: three months, reference through 128, global | 0.351318692 | 0.408293698 | 1/2 point, both |\n| Q2: three months, reference through 128, 128 bins | 0.233318843 | 0.252393939 | 1/4 point, joint only |\n\nThe step from H0 to H1 changes one thing. Both rows use the same residuals of the same reference trajectories. H0 bounds the state error uniformly in time and multiplies by the total mass of the kernel \\(q\\); H1 applies (3.5), which weights each residual by the forward variance that remains. The complete bound falls by 72.91 per cent. Extending the reference from \\(u\\le64\\) to \\(u\\le128\\) (H2 to H3, Q0 to Q1) replaces envelope radii by certified ones and moves the centre, from about \\(-0.167\\) to \\(-0.082\\) points at six months, so that step combines a smaller uncertainty with a different displacement.\n\nTable 3 isolates the propagation on a second bank, generated independently at \\(N_t=2048\\) time cells, with everything else held fixed. The last two rows use identical binwise residual envelopes and differ only in the kernel, \\(\\xi\\) against \\(\\xi^{\\lambda_*}\\).\n\n**Table 3. Matched propagation on bank N2L for the six-month 4400–4500 spread, in index points.**\n\n| Propagation | Stored output, leg-by-leg | Stored output, joint | Reference returned, joint | Node radius of used nodes |\n|---|---:|---:|---:|---:|\n| Uniform state bound | 3.690877564 | 1.829428107 | 1.662665889 | 1.476801373 |\n| (3.5) with \\(\\xi\\), global residual | 0.589267065 | 0.428570987 | 0.261808769 | 0.075944253 |\n| (3.8) with \\(\\xi\\), 128 bins | 0.514096070 | 0.394999331 | 0.228237113 | 0.042372598 |\n| (3.8) with \\(\\xi^{\\lambda_*}\\), 128 bins | 0.509733613 | 0.393082828 | 0.226320610 | 0.040456094 |\n\nThe forward variance weight accounts for almost all of the improvement: the node radius falls from 1.477 to 0.076 points. Time-local envelopes take it to 0.042. The leverage damping of Proposition 3.2 removes a further 4.5 per cent of the node radius, in line with \\(\\lambda_*T^\\alpha\\approx0.075\\) at this horizon; its effect grows with maturity and with \\(|\\rho|\\nu\\). The complete bound for the stored output moves less, because the fixed terms dominate it: the centre, the omitted finite nodes and the remainders together exceed 0.352 points on this bank. The next gain therefore lies in the reference band, and extending it is what takes H2 to H4 in Table 2.\n\n### 7.3. A decision that turns on netting\n\nAt three months a single global residual maximum certifies half a point (Q1). With 128 time bins (Q2) the bound reaches the quarter-point tolerance, and there the aggregation decides. Table 4 is the ledger of Q2. The two columns share the stored output, the reference centre, all 1,025 node radii and every remainder.\n\n**Table 4. Ledger of the three-month 4400–4500 spread (Q2), in index points. Only the aggregation of the node errors differs.**\n\n| Component | Joint | Leg-by-leg |\n|---|---:|---:|\n| Absolute centre \\(|w^\\top d|\\) | 0.149775147 | 0.149775147 |\n| Node support, 1,025 nodes | 0.045461793 | 0.064536890 |\n| Strip remainder | 0.000026654 | 0.000026654 |\n| Exact infinite tail | 0.038055248 | 0.038055248 |\n| Reference arithmetic | \\(4.1\\times10^{-16}\\) | \\(4.1\\times10^{-16}\\) |\n| Complete bound | 0.233318843 | 0.252393939 |\n\nNetting lowers the node term by 30 per cent and the complete bound by 7.56 per cent, from above a quarter point to below it. Nothing else changed.\n\nProposition 5.2 predicts where the saving comes from. Here \\(\\Delta=\\log(4500/4400)=0.02247\\), and the stored ratio of the joint to the leg-by-leg coefficient at each of the 1,025 nodes agrees with (5.8) to fifteen digits. Table 5 splits the node term by frequency.\n\n**Table 5. Node term of the three-month spread by frequency band, in index points.**\n\n| Band of \\(u\\) | Nodes | Joint | Leg-by-leg | Ratio | Share of leg-by-leg charge |\n|---|---:|---:|---:|---:|---:|\n| 0 to 16 | 128 | 0.000074 | 0.000529 | 0.139 | 0.8% |\n| 16 to 32 | 128 | 0.001158 | 0.004100 | 0.282 | 6.4% |\n| 32 to 64 | 256 | 0.010559 | 0.020137 | 0.524 | 31.2% |\n| 64 to 96 | 256 | 0.017651 | 0.022785 | 0.775 | 35.3% |\n| 96 to 128 | 257 | 0.016020 | 0.016986 | 0.943 | 26.3% |\n\nBelow \\(u=16\\) the spread keeps 14 per cent of its leg-by-leg error, and near the top of the grid, which approaches the Nyquist frequency \\(\\pi/\\Delta\\approx140\\), it keeps 94 per cent. The certified radii grow with the frequency, so nine tenths of the charge sits above \\(u=32\\), where a spread of this width is poorly hedged. The saving of 30 per cent is what a 100-point spread can extract from an error concentrated at short wavelengths.\n\nA narrower spread extracts more. Table 6 evaluates the node term of (5.5) and (5.6) with the same 1,025 radii for verticals of other widths struck at 4400. The radii do not depend on the strike, so the table is a statement about the same certified transform.\n\n**Table 6. Node term of a three-month vertical spread struck at 4400, by width, with the certified radii of Q2; index points.**\n\n| Width in points | Nyquist frequency | Joint | Leg-by-leg | Ratio |\n|---:|---:|---:|---:|---:|\n| 5 | 2766 | 0.002711 | 0.064193 | 0.042 |\n| 10 | 1384 | 0.005414 | 0.064211 | 0.084 |\n| 25 | 555 | 0.013403 | 0.064265 | 0.209 |\n| 50 | 278 | 0.025925 | 0.064356 | 0.403 |\n| 100 | 140 | 0.045462 | 0.064537 | 0.704 |\n| 200 | 71 | 0.052476 | 0.064895 | 0.809 |\n| 400 | 36 | 0.044482 | 0.065601 | 0.678 |\n\nFor the five-point vertical the joint node term is about a twenty-fourth of the leg-by-leg term. The joint term grows in proportion to the width while the Nyquist frequency lies beyond the grid, and levels off once it enters. Figure 3 shows the geometry behind Table 4: the joint error set of the two prices and its bounding box.\n\n\n\n### 7.4. Twenty-eight positions\n\nThe six-month experiment fixes 28 portfolio directions on the twelve strikes before any bound is computed (Table 7), and keeps their weights as stated. Table 7 also reports, for bank H1 at \\(\\alpha=0.52\\), how much of the leg-by-leg node term each family keeps and the mean absolute centre.\n\n**Table 7. The 28 fixed positions, with node netting and centre on bank H1 at \\(\\alpha=0.52\\). The netting ratio is the joint node term divided by the leg-by-leg node term.**\n\n| Family | Weights | Count | Netting ratio, min–mean–max | Herfindahl level | Mean centre, points |\n|---|---|---:|---:|---:|---:|\n| Adjacent spreads | \\(e_i-e_{i+1}\\) | 11 | 0.649–0.708–0.771 | 0.707 | 0.108 |\n| Adjacent butterflies | \\(e_i-2e_{i+1}+e_{i+2}\\) | 10 | 0.449–0.524–0.608 | 0.612 | 0.071 |\n| Wide spreads | \\(e_0-e_3,\\ e_3-e_6,\\ e_6-e_9,\\ e_0-e_{11}\\) | 4 | 0.474–0.570–0.651 | 0.707 | 0.235 |\n| Equal-weight baskets | twelve, lower six, upper six strikes | 3 | 0.076–0.156–0.221 | 0.289, 0.408 | 0.344 |\n\nThe adjacent spreads keep 0.708 of their leg-by-leg node error on average, against the Herfindahl level 0.707 of a two-leg position. The butterflies keep 0.524. The baskets keep least of all: an equal-weight position in all twelve calls keeps 0.076 of its leg-by-leg node error, because three quarters of the node charge of this bank sits between \\(u=64\\) and \\(u=96\\), where twelve strikes spread over a quarter of a unit of log-moneyness respond out of phase. The centre runs the other way. The stored rational output lies above the reference at every strike, by 0.04 and 0.07 points at the two ends of the ladder and by up to 0.65 points near the money. A displacement of one sign cancels in part in spreads and butterflies and accumulates in baskets, where it makes up 0.344 of a complete bound of 0.356 points. The approximation error of the stored formula is systematic across strikes and the uncertainty of the certificate is idiosyncratic, as Section 5.2 anticipated, and the joint bound accounts for each correctly.\n\nTable 8 counts the positions certified at three tolerances. Within a row the two methods use identical radii.\n\n**Table 8. Positions certified out of 28 on bank H1, by tolerance in index points.**\n\n| Roughness \\(\\alpha\\) | Tolerance | Leg-by-leg | Joint |\n|---|---:|---:|---:|\n| 0.52 | 1/4 | 0 | 2 |\n| 0.52 | 1/2 | 13 | 28 |\n| 0.52 | 1 | 28 | 28 |\n| 0.60 | 1/4 | 1 | 10 |\n| 0.60 | 1/2 | 22 | 28 |\n| 0.60 | 1 | 28 | 28 |\n| 0.90 | 1/4 | 13 | 20 |\n| 0.90 | 1/2 | 25 | 25 |\n| 0.90 | 1 | 27 | 28 |\n\nAt half a point and \\(\\alpha=0.52\\) the joint bound certifies every position and the leg-by-leg bound certifies 13. Netting decides whenever the tolerance is of the order of the bound, which is the case for a tolerance set inside the bid–ask spread. All 28 positions are certified from one bank: once the transform disks exist, an additional position costs 1,025 support terms. Figure 4 shows the count as a function of the tolerance for the three candidates, and Appendix E.3 gives the exact thresholds.\n\n\n\n### 7.5. Is the roughness ranking numerical?\n\nThe roughness fitted to this surface is \\(H=0.0286\\), and candidates a few thousandths apart in \\(H\\) produce prices that differ by fractions of an index point. Whether a calibration can tell them apart depends on the accuracy of the pricer. We test five candidates, \\(\\alpha\\in\\{0.520,0.525,0.530,0.540,0.550\\}\\), with the curve (7.1), the quotes and all other parameters fixed. The design was set in advance: every candidate receives a freshly generated reference field and a complete residual bank at \\(N_t=1024\\), and all five are regenerated at \\(N_t=2048\\) if any pair remains unresolved. The objective is the normalised least-squares distance to the mid prices, enclosed by Proposition 5.4; Table 9 gives the enclosures at both levels.\n\n**Table 9. Objective enclosures for five roughness candidates, as \\(J\\times10^8\\), and the number of the ten pairs strictly separated.**\n\n| \\(\\alpha\\) | \\(H\\) | \\(N_t=1024\\), joint | \\(N_t=2048\\), joint |\n|---:|---:|---:|---:|\n| 0.520 | 0.020 | [4.825094, 6.778350] | [5.293449, 6.024095] |\n| 0.525 | 0.025 | [5.645378, 7.501414] | [6.097920, 6.798485] |\n| 0.530 | 0.030 | [6.748974, 8.517584] | [7.187004, 7.859529] |\n| 0.540 | 0.040 | [9.807435, 11.422182] | [10.218707, 10.839486] |\n| 0.550 | 0.050 | [14.000658, 15.482012] | [14.387007, 14.960680] |\n| Pairs separated, joint | | 7 of 10 | 10 of 10 |\n| Pairs separated, leg-by-leg | | 3 of 10 | 7 of 10 |\n\nAt the finer level the five enclosures are disjoint and ordered, and \\(\\alpha=0.520\\) is the certified minimiser on this grid. The closest pair, 0.520 against 0.525, is separated by a gap of \\(7.4\\times10^{-10}\\) in the normalised objective. The leg-by-leg enclosures leave three pairs unresolved at the same level, those among 0.520, 0.525 and 0.530. The ordering of roughness candidates half a hundredth apart in \\(H\\) is therefore a certified property of the model and the quotes.\n\n### 7.6. Model misfit without solver doubt\n\nA certificate also settles the opposite question. Each of the five candidates is inconsistent with at least one quoted bid–ask band, and for a certified price this is a statement about the model. At \\(\\alpha=0.52\\), on the coarsest bank H0, the exact model price of the 4400 call lies in an interval whose lower end is 92.886 points, while the quoted ask is 92.396: the model is above the ask by at least 0.490 points. For the 4500 call the excess is at least 0.067 points. At the other ten strikes the certified interval meets the quoted band. A desk that sees its rough Heston calibration miss the 4400 and 4500 strikes on this date can rule out the pricer as the cause.\n\n### 7.7. Auditing an output or replacing it\n\nA stored-output certificate serves the audit of a pricing library or of a valuation that has already been reported. When the output may be replaced and the reference is available, returning the reference is the natural alternative. Table 10 compares the two on bank N2L, together with a reimplementation of the modified Adams scheme of Boyarchenko, de Innocentis and Levendorskii (2025). All rows share the model, the spread, the reference centre, the node radii and the remainders.\n\n**Table 10. Stored and replaceable outputs for the six-month 4400–4500 spread on bank N2L, in index points.**\n\n| Output | Joint bound | Leg-by-leg bound | Quarter point, joint |\n|---|---:|---:|---|\n| Stored rational output | 0.394999331 | 0.514096070 | not certified |\n| Reference returned in binary64 | 0.228237113 | 0.347333852 | certified |\n| Stored output plus stored correction | 0.228237113 | 0.347333852 | certified |\n| Modified Adams, 512 steps | 0.229082759 | 0.348179497 | certified |\n| Modified Adams, 1024 steps | 0.228534686 | 0.347631424 | certified |\n\nThe reference and the corrected output are certified to a quarter point by the joint bound, and the leg-by-leg bound certifies none of the five outputs. The Adams outputs lie within 0.001 points of the reference, so that more than 99.6 per cent of their bound is the uncertainty of the reference itself: at this resolution the scheme is as accurate as the reference can tell.\n\n## 8. Conclusion\n\nThe construction follows the valuation map. The Riccati equation determines the state and, through its invariant half-plane, the stability of errors. The forward variance curve determines what a residual costs. The payoff determines which transform errors a position is exposed to. At each stage an absolute value taken too early would discard structure that the next stage can use, and the financial content of the paper lies in what that structure is: the weight on a residual is expected variance under a share-tilted measure, and the weight on a transform error is the Fourier transform of the payoff.\n\nThe results reach beyond the example. Section 5 is model-free: the payoff transform, the netting ratio and the Herfindahl law govern the numerical error of every pricer built on a common transform, whatever the model and whatever the method behind the transform. Theorem 3.1 holds on every damped contour and, by Remark 3.6, for Volterra Heston models with completely monotone kernels, the classical and the lifted Heston models included. The damping it uses is the frequency-independent part \\(\\tfrac12|\\rho|\\nu\\); the exact solution supplies more, increasing with the frequency, which is where the radii of Table 5 are largest. For a trading desk, a validation team or a supervisor the consequence is practical: the numerical component of valuation uncertainty can be bounded, attributed to frequencies, and netted across a book exactly as the payoffs net.\n\nThe [online article](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/ARTICLE.md), the scientific source and the verification commands accompany the paper. All 27 numerical banks are identified by size and SHA-256 in `SCIENTIFIC-MANIFEST.json`. `verify_source.py` checks the publication identities. `reproduce.py` reads the saved banks and reconstructs the prices, objectives and decisions downstream of them; `--full` also recomputes the structural-sign cover of Proposition 6.1. Appendix E and `COMMANDS.md` distinguish the generators, the readers and the arithmetic they share.\n\n## Appendices and supporting material\n\nThis page includes the complete main text and references. Appendices, supplementary proofs, code and verification material are available in the [GitHub repository](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston) and the [full manuscript](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/ARTICLE.md#appendix-a-rescaled-variables-model-and-transform).\n\n## References\n\nAbi Jaber, E. (2019). Lifting the Heston model. *Quantitative Finance* 19(12), 1995–2013.\n\nAbi Jaber, E. and El Euch, O. (2019a). Markovian structure of the Volterra Heston model. *Statistics & Probability Letters* 149, 63–72. [doi:10.1016/j.spl.2019.01.024](https://doi.org/10.1016/j.spl.2019.01.024). Theorem and example numbers refer to arXiv:1803.00477.\n\nAbi Jaber, E. and El Euch, O. (2019b). Multifactor approximation of rough volatility models. *SIAM Journal on Financial Mathematics* 10(2), 309–349.\n\nAbi Jaber, E., Larsson, M. and Pulido, S. (2019). Affine Volterra processes. *Annals of Applied Probability* 29(5), 3155–3200.\n\nAlbrecher, H., Mayer, P., Schoutens, W. and Tistaert, J. (2007). The little Heston trap. *Wilmott Magazine*, January, 83–92.\n\nBasel Committee on Banking Supervision (2019). *Minimum capital requirements for market risk*. Basel Framework, chapter MAR32, version effective 1 January 2023. Bank for International Settlements.\n\nBayer, C. and Breneis, S. (2023a). Weak Markovian approximations of rough Heston. [arXiv:2309.07023](https://arxiv.org/abs/2309.07023v1).\n\nBayer, C. and Breneis, S. (2023b). Efficient option pricing in the rough Heston model using weak simulation schemes. [arXiv:2310.04146](https://arxiv.org/abs/2310.04146v1).\n\nBayer, C., Friz, P. and Gatheral, J. (2016). Pricing under rough volatility. *Quantitative Finance* 16(6), 887–904.\n\nBen Hammouda, C., Ben Romdhane, A., Samet, M. and Tempone, R. F. (2026). Single- and multilevel quadrature with error control for Fourier pricing under the rough Heston model. [arXiv:2609.00438](https://arxiv.org/abs/2609.00438v1).\n\nBoard of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation and Office of the Comptroller of the Currency (2026). *Supervisory Guidance on Model Risk Management*. SR Letter 26-2, 17 April 2026.\n\nBoyarchenko, S., de Innocentis, M. and Levendorskii, S. (2025). Fast reliable pricing and calibration of the rough Heston model. [arXiv:2508.15080](https://arxiv.org/abs/2508.15080v1).\n\nBreeden, D. T. and Litzenberger, R. H. (1978). Prices of state-contingent claims implicit in option prices. *Journal of Business* 51(4), 621–651.\n\nCboe Global Markets (2024). *S&P 500 Index Options: SPX product specifications*, as of 19 March 2024.\n\nCboe Global Markets (2026). Cboe Global Markets reports trading volume for December and full year 2025. Press release, 6 January 2026.\n\nCommission Delegated Regulation (EU) 2016/101 of 26 October 2015 supplementing Regulation (EU) No 575/2013 with regard to regulatory technical standards for prudent valuation. *Official Journal of the European Union* L 21, 28 January 2016, 54–65.\n\nCont, R. and Das, P. (2024). Rough volatility: fact or artefact? *Sankhya B* 86, 191–223.\n\nDiethelm, K., Ford, N. J. and Freed, A. D. (2004). Detailed error analysis for a fractional Adams method. *Numerical Algorithms* 36, 31–52.\n\nEl Euch, O., Fukasawa, M. and Rosenbaum, M. (2018). The microstructural foundations of leverage effect and rough volatility. *Finance and Stochastics* 22, 241–280.\n\nEl Euch, O. and Rosenbaum, M. (2018). Perfect hedging in rough Heston models. *Annals of Applied Probability* 28(6), 3813–3856. Statement numbers refer to arXiv:1703.05049.\n\nEl Euch, O. and Rosenbaum, M. (2019). The characteristic function of rough Heston models. *Mathematical Finance* 29(1), 3–38.\n\nGatheral, J., Jaisson, T. and Rosenbaum, M. (2018). Volatility is rough. *Quantitative Finance* 18(6), 933–949.\n\nGatheral, J. and Keller-Ressel, M. (2019). Affine forward variance models. *Finance and Stochastics* 23, 501–533.\n\nGatheral, J. and Radoičić, R. (2019). Rational approximation of the rough Heston solution. *International Journal of Theoretical and Applied Finance* 22(3), 1950010. [doi:10.1142/S0219024919500109](https://doi.org/10.1142/S0219024919500109).\n\nGatheral, J. and Radoičić, R. (2024). A generalization of the rational rough Heston approximation. *Quantitative Finance* 24(2), 329–335. [doi:10.1080/14697688.2024.2302055](https://doi.org/10.1080/14697688.2024.2302055).\n\nGripenberg, G., Londen, S.-O. and Staffans, O. (1990). *Volterra Integral and Functional Equations*. Cambridge University Press.\n\nHagan, P. S., Kumar, D., Lesniewski, A. S. and Woodward, D. E. (2002). Managing smile risk. *Wilmott Magazine*, September, 84–108.\n\nHagan, P. S., Kumar, D., Lesniewski, A. S. and Woodward, D. E. (2014). Arbitrage-free SABR. *Wilmott* 2014(69), 60–75.\n\nHager, P. P. and Kreher, D. (2026). Expanding the rough Heston model in H. [arXiv:2606.16619](https://arxiv.org/abs/2606.16619v1).\n\nHeston, S. L. (1993). A closed-form solution for options with stochastic volatility with applications to bond and currency options. *Review of Financial Studies* 6(2), 327–343.\n\nHorvath, B., Muguruza, A. and Tomas, M. (2021). Deep learning volatility: a deep neural network perspective on pricing and calibration in (rough) volatility models. *Quantitative Finance* 21(1), 11–27.\n\nJeng, S. W. and Kiliçman, A. (2020). Series expansion and fourth-order global Padé approximation for a rough Heston solution. *Mathematics* 8(11), 1968. [doi:10.3390/math8111968](https://doi.org/10.3390/math8111968).\n\nJeng, S. W. and Kiliçman, A. (2021). SPX calibration of option approximations under rough Heston model. *Mathematics* 9(21), 2675. [doi:10.3390/math9212675](https://doi.org/10.3390/math9212675).\n\nKahl, C. and Jäckel, P. (2005). Not-so-complex logarithms in the Heston model. *Wilmott Magazine*, September, 94–103.\n\nKopteva, N. (2022). Pointwise-in-time a posteriori error control for time-fractional parabolic equations. *Applied Mathematics Letters* 123, 107515. [doi:10.1016/j.aml.2021.107515](https://doi.org/10.1016/j.aml.2021.107515).\n\nLewis, A. L. (2001). A simple option formula for general jump-diffusion and other exponential Lévy processes. Working paper, Envision Financial Systems and OptionCity.net.\n\nLi, L. and Liu, J.-G. (2018). A generalized definition of Caputo derivatives and its application to fractional ODEs. *SIAM Journal on Mathematical Analysis* 50(3), 2867–2900. [doi:10.1137/17M1160318](https://doi.org/10.1137/17M1160318).\n\nOlver, F. W. J., Lozier, D. W., Boisvert, R. F. and Clark, C. W. (2010). *NIST Handbook of Mathematical Functions*. Cambridge University Press.\n\nRømer, S. E. (2022). Empirical analysis of rough and classical stochastic volatility models to the SPX and VIX markets. *Quantitative Finance* 22(10), 1805–1838.\n\nSimon, T. (2014). Comparing Fréchet and positive stable laws. *Electronic Journal of Probability* 19, paper 16. [arXiv:1310.1888](https://arxiv.org/abs/1310.1888v2).\n\nSimon, T. (2015). Mittag-Leffler functions and complete monotonicity. *Integral Transforms and Special Functions* 26(1), 36–50. [doi:10.1080/10652469.2014.965704](https://doi.org/10.1080/10652469.2014.965704).\n\nTrefethen, L. N. and Weideman, J. A. C. (2014). The exponentially convergent trapezoidal rule. *SIAM Review* 56(3), 385–458. [doi:10.1137/130932132](https://doi.org/10.1137/130932132).\n"
}
diff --git a/website/content/selected-research/certified-valuation-of-arithmetic-asian-options.json b/website/content/selected-research/certified-valuation-of-arithmetic-asian-options.json
index f817b02..8598fa1 100644
--- a/website/content/selected-research/certified-valuation-of-arithmetic-asian-options.json
+++ b/website/content/selected-research/certified-valuation-of-arithmetic-asian-options.json
@@ -2,10 +2,10 @@
"slug": "certified-valuation-of-arithmetic-asian-options",
"title": "Certified Valuation of Arithmetic Asian Options: The Cost of Exercising on the Geometric Average",
"description": "The cost of exercising an arithmetic Asian option on the geometric average: a proved bound, its cubic law in Black's model, and a certified Heston price interval with its Euler discretization error.",
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/manuscript/report.md",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/manuscript/report.md",
"repositoryUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options",
- "sourceEditedAt": "2026-11-19T03:29:58-05:00",
- "sourceCommit": "41a814f6e5aa1eddddb4412e57ce68ec506c323e",
- "sourceBlob": "6f573b14c4c172e394688333b22726438bbed461",
- "markdown": "[English PDF](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/paper/paper.pdf) · [Repository](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/README.md) · [Code and evidence](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/docs/EVIDENCE.md)\n\nThe Cost of Exercising on the Geometric Average\n\n**Theodore Ouyang**\n\n[theodore.oy2025@gmail.com](mailto:theodore.oy2025@gmail.com) · [10@alumni.duke.edu](mailto:10@alumni.duke.edu)\n\n2026\n\n**Keywords:** Asian options; average price options; geometric conditioning; Bregman divergence; stochastic volatility; Heston model; Euler discretization; validated numerics; model risk.\n\n**JEL classification:** G13; C63; C65.\n\n**MSC 2020:** 91G20; 91G60; 60H35; 65G30.\n\n## Abstract\n\nAn arithmetic Asian option is worth more than the same payoff when exercise is decided by a scaled geometric average of the fixings. We call the difference the cost of the kink and prove three things about it. It is the expected overshoot on the event that the two averages fall on opposite sides of the strike. Integrated across strikes with the weights that replicate a log contract, it equals a Bregman divergence between the two averages. Whenever the return before the first fixing has a conditionally Gaussian component with volatility $`\\sigma`$, it is at most\n\n$$\n\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/cG)}{\\sigma}\\right],\\qquad d(x)=x-1-\\log x.\n$$\n\nIn Black's model this bound is $`v^3\\tau^2/(360\\sqrt{2\\pi t_1})`$ per unit of strike to leading order, for volatility $`v`$, an averaging window $`\\tau`$ and a lead time $`t_1`$. For a one-month crude oil average price option it is below one hundredth of a cent per barrel and within a factor of 1.1 to 1.5 of the true cost; for an annual-average put on the scale of a sovereign oil hedge it is under one cent per barrel. For a twelve-fixing 95–110 call spread under Heston, validated transforms turn the bound into a proved price interval, $`[6.508371733,\\;6.518868974]`$, and a proved bound of 0.011025 on the distance to the price under a 768-step Euler scheme. The same computation shows that the Euler bias of the proxy value is first order in the step $`h`$, with leading coefficient $`-0.156`$.\n\n## 1. Introduction\n\nMost options written on oil, metals, freight and corporate currency flows pay on an average. A refiner buys crude every day of the month and an airline burns fuel every day of the quarter, so their physical contracts are priced on monthly averages and their hedges follow. The NYMEX average price option on crude oil settles on the mean of the daily futures settlements in a calendar month [[10](#ref-10)]. The London Metal Exchange lists options on the monthly average of its official settlement prices, the price at which physical metal changes hands [[25](#ref-25)]. Mexico's treasury insures its oil revenue with puts on the average price over a year [[35](#ref-35)]; the program for 2009 covered 330 million barrels at 70 dollars and paid out 5.085 billion dollars [[4](#ref-4)].\n\nThe payoff of such a contract depends on the arithmetic average $`A`$ of the fixings, whose law has no closed form in any standard model. The geometric average $`G`$ does, and practice rests on that fact. Kemna and Vorst use the geometric option as a control variate [[21](#ref-21)]. Curran, and Rogers and Shi, condition on $`G`$ and obtain a lower bound that is accurate to several digits in Black's model [[12](#ref-12), [31](#ref-31)]. Fusai and Kyriakou carry the same bound to stochastic volatility and jumps by transform methods [[16](#ref-16)]. Under stochastic volatility the working alternative is simulation on a time grid, which adds a discretization bias to the sampling error.\n\nTwo numbers in this workflow usually come without a proof. One is the distance between the conditioning lower bound and the price. The other is the distance between the price under the discretized dynamics and the price under the model that was calibrated. The classical estimate of the first is of first order in the conditional standard deviation of $`A`$ given $`G`$, which makes it loose exactly where the lower bound is good. The second is known to be of first order in the step for smooth payoffs [[28](#ref-28)], with an unknown constant. A valuation is certified when both distances are replaced by inequalities whose every term has been computed with a rigorous error. This paper does so, and it starts from a description of the first distance that makes its size predictable.\n\n### 1.1 The idea\n\nSuppose the holder of an arithmetic call must decide exercise by looking at a proxy $`B`$ instead of $`A`$: she receives $`A-K`$ when $`B\\gt K`$ and nothing otherwise. She can only lose by this, and she loses on two events. If the proxy is below the strike while $`A`$ is above, she forgoes $`A-K`$. If the proxy is above while $`A`$ is below, she pays $`K-A`$. In both cases the strike lies between the proxy and the truth, so the loss is at most $`|A-B|`$ and it occurs with a probability of order $`|A-B|`$ times the density of the proxy at the strike. The cost is of second order in the basis $`A-B`$. Figure 1 shows the two events for the contract of Section 5.\n\nThe second-order statement becomes an inequality when the density at the strike can be bounded, and here a structural fact helps. The return realized before the first fixing moves every fixing by the same factor. If part of that return is Gaussian given everything else, with volatility $`\\sigma`$, then conditionally the option is a Black–Scholes call on that factor, and the density at the strike is at most $`1/(\\sigma\\sqrt{2\\pi})`$ in log-strike. With $`B=cG`$ the basis depends only on the shape of the path, since $`A/G`$ is unchanged when all fixings are scaled together. The cost of the kink is therefore controlled by a ratio: the dispersion of the fixings about their own average, divided by the level risk that precedes them.\n\n\n\n### 1.2 Results\n\nSection 3 proves the general statements. Proposition 3.1 identifies the cost of the kink with the Bregman divergence generated by the call payoff and shows that its integral across strikes is exact: with weights $`dK`$ it equals $`E(A-B)^2/2`$, and with the weights $`dK/K^2`$ that replicate a log contract it equals $`E\\,d(A/B)`$, where $`d(x)=x-1-\\log x`$. Theorem 3.3 adds the common factor. For every strike,\n\n$$\n\\begin{gathered} 0\\le E(A-K)^+-E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\bigr] \\le\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\right], \\end{gathered} \\tag{1.1}\n$$\n\nand the same bound holds with $`E[(A-B)^2/\\sigma]/(2K\\sqrt{2\\pi})`$ on the right. Each bound is a total mass from Proposition 3.1 multiplied by the largest value a Gaussian density of width $`\\sigma`$ can take. A matching lower bound shows that (1.1) is attained for at-the-money options whose dispersion is small relative to $`\\sigma`$. For $`B=cG`$ the best constant is the mean of $`A/G`$ weighted by $`1/\\sigma`$, and the bound at that scale is $`K/\\sqrt{2\\pi}`$ times $`E[1/\\sigma]`$ times the mean log deviation of $`A/G`$ (Proposition 3.4). Because the Curran, Rogers–Shi and optimized Fusai–Kyriakou lower bounds dominate the proxy value, (1.1) also bounds their errors.\n\nSection 4 evaluates the bound where contracts trade. In Black's model with volatility $`v`$, a lead time $`t_1`$ before the first fixing and a window $`\\tau`$, the bound per unit of strike is\n\n$$\n\\begin{gathered} \\frac{v^3\\tau^2}{360\\sqrt{2\\pi t_1}} \\end{gathered} \\tag{1.2}\n$$\n\nto leading order under continuous monitoring, with an explicit correction for $`n`$ fixings (Proposition 4.1). For a calendar-month crude oil option one to twelve months ahead, ball arithmetic gives price intervals 0.0023 to 0.0081 cents per barrel wide, and simulation puts the true cost at 65 to 94 percent of the bound. The classical first-order bound for the same contracts is 2.6 cents. Under Heston dynamics with a volatility of volatility of 0.7 the cost is about 0.005 to 0.008 cents and the bound 0.006 to 0.013. For an annual-average put on 330 million barrels the bound is 0.91 cents per barrel in Black's model, three million dollars on a premium of 569 million.\n\nSection 5 turns the bound into a certificate for a contract on which it is not small: a 95–110 call spread on twelve monthly fixings under Heston with $`S_0=100`$. The proxy value is computed by validated transform inversion and the weighted second moment by validated quadrature, with projection, tails, periodization and rounding inside the budget. The continuous price lies in\n\n$$\n\\begin{gathered} [6.508371733,\\;6.518868974], \\end{gathered} \\tag{1.3}\n$$\n\nthe price under absorbed Euler with step $`h=1/768`$ lies in $`[6.507844281,\\;6.519014106]`$, and their difference in $`[-0.011024692273,\\;0.010642371599]`$.\n\nSection 6 explains the Euler error that these intervals enclose. One Euler step of the stochastic differential equation acts on exponential-affine functions as one explicit Euler step of the Riccati equation (Proposition 6.1). The finite sums already enclosed by the certificate then display a first-order bias in the proxy value: $`-0.000406686`$, $`-0.000202992`$ and $`-0.000101409`$ at 384, 768 and 1536 steps, with leading coefficient $`-0.156`$. Simulation puts the change in the cost of the kink between the two laws below $`10^{-6}`$. The Euler price at 768 steps is therefore about 0.0002 below the continuous price, well inside the proved bound of 0.011025. On the deterministic-variance face the leading error is a vega multiplied by the quadrature error of the Euler variance (Theorem 6.2), and a coupling of the two laws narrows the allowance for the cost of the kink by a factor of 175.\n\nSection 7 carries the fixed-parameter bounds through a calibration and bounds the displacement of every posterior quantile of the Asian price. Section 8 describes two uses: an acceptance test for pricing engines and a numerical component of a valuation reserve.\n\n### 1.3 Related work\n\nGeometric conditioning starts with Curran [[12](#ref-12)] and Rogers and Shi [[31](#ref-31)]. Rogers and Shi bound the error of their lower bound by half the expected conditional standard deviation of the average, an estimate of first order. Thompson [[33](#ref-33)], Nielsen and Sandmann [[29](#ref-29)] and Lord [[26](#ref-26)] sharpen it in the lognormal model; in Thompson's example the Rogers–Shi upper bound is 9.039 against a lower bound of 8.8275 and his own upper bound of 8.8333. Fusai and Kyriakou [[16](#ref-16)] optimize the exercise threshold, treat stochastic volatility and jump models through the joint transform of the price and its geometric average, and state a conditional-variance error bound of the Rogers–Shi type (their Theorem 4), which they evaluate numerically. Moment matching [[34](#ref-34), [24](#ref-24)] and the Laplace transform of Geman and Yor [[17](#ref-17)] are the other classical routes in the lognormal model.\n\nRelative to this literature, Theorem 3.3 is an a priori error bound of second order that holds beyond the lognormal model and needs only moments of the fixings weighted by $`1/\\sigma`$. It covers every lower bound of the conditioning family at once (Corollary 3.5), and the cubic law of Section 4 gives its size before any computation.\n\nThe decomposition of a convex payoff into calls is due to Breeden and Litzenberger [[6](#ref-6)] and Carr and Madan [[9](#ref-9)], and the weights $`dK/K^2`$ are those of the variance swap [[13](#ref-13)]. In forecast evaluation the same pathwise quantity is the elementary score of Ehm, Gneiting, Jordan and Krüger [[14](#ref-14)], who read it as the regret of an investor acting on a point forecast and prove the mixture representation of Bregman scores; Proposition 3.1 is that representation with exercise in the role of investment.\n\nHeston's affine structure supplies the continuous transforms [[19](#ref-19)], and recursive transforms for discrete geometric averages are available [[22](#ref-22), [30](#ref-30)]. The scheme studied here is the absorption fix in the classification of Lord, Koekkoek and van Dijk [[27](#ref-27)]; Broadie and Kaya [[8](#ref-8)] and Andersen [[1](#ref-1)] give exact and moment-matched alternatives. Mickel and Neuenkirch prove weak order one for the absorbed scheme and smooth payoffs when the Feller index exceeds one [[28](#ref-28)]; Talay and Tubaro [[32](#ref-32)] and Bally and Talay [[2](#ref-2)] are the classical references for weak expansions. Error control for transform inversion follows Lee [[23](#ref-23)], and all enclosures use the ball arithmetic of Arb [[20](#ref-20)].\n\n## 2. Average-price contracts and their two averages\n\n### 2.1 The contract\n\nFix dates $`0\\lt t_1\\lt \\cdots\\lt t_n=T`$ and let $`S_{t_1},\\ldots,S_{t_n}`$ be the fixings of a positive price. Put\n\n$$\n\\begin{gathered} A=\\frac1n\\sum_{i=1}^nS_{t_i},\\qquad G=\\Bigl(\\prod_{i=1}^nS_{t_i}\\Bigr)^{1/n},\\qquad X=\\frac AG . \\end{gathered} \\tag{2.1}\n$$\n\nAn average price call pays $`(A-K)^+`$ at $`T`$ and a put pays $`(K-A)^+`$. We call $`t_1`$ the lead time and $`\\tau=t_n-t_1`$ the window. For the exchange-traded contracts of the introduction the window is a calendar month of daily fixings and the lead time is whatever remains until that month begins. For Mexico's hedge the window is a year.\n\nTwo elementary reductions are used throughout. Put–call parity holds for the average, $`(A-K)^+-(K-A)^+=A-K`$, so a statement about calls is a statement about puts. And a fixing that has been observed is a constant: if $`m`$ fixings with mean $`\\bar a`$ are known, then\n\n$$\n\\begin{gathered} (A-K)^+=\\frac{n-m}{n}\\,\\bigl(A'-K'\\bigr)^+,\\qquad K'=\\frac{nK-m\\bar a}{n-m}, \\end{gathered} \\tag{2.2}\n$$\n\nwhere $`A'`$ is the mean of the remaining fixings. An option inside its window is an option on the fixings still to come with a shifted strike, or a forward if $`K'\\le0`$. The lead time is therefore always the time to the next unknown fixing.\n\n### 2.2 The gap between the averages is dispersion\n\nWrite $`z_i=\\log S_{t_i}`$, $`\\bar z`$ for their mean, $`u_i=z_i-\\bar z`$, and\n\n$$\ns^2=\\frac1n\\sum_{i=1}^nu_i^2,\\qquad m=\\max_i|u_i| .\n$$\n\nThen $`G=e^{\\bar z}`$ and $`X=n^{-1}\\sum_ie^{u_i}`$. Since the $`u_i`$ sum to zero and $`e^u-1-u`$ lies between $`u^2e^{-|u|}/2`$ and $`u^2e^{|u|}/2`$,\n\n$$\n\\begin{gathered} \\tfrac12s^2e^{-m}\\le X-1\\le\\tfrac12s^2e^{m}. \\end{gathered} \\tag{2.3}\n$$\n\nThe arithmetic average is the geometric average marked up by half the dispersion of the log fixings about their own mean. The markup does not change when every fixing is multiplied by the same number: $`X`$ is a function of the shape of the path only.\n\nIn financial terms the position long $`A`$ and short $`G`$ is a claim on the realized variance of the fixing series, with no exposure to the level at first order. In Black's model with volatility $`v`$ and continuous monitoring over a window $`\\tau`$, $`s^2`$ has mean $`v^2\\tau/6`$ and variance $`v^4\\tau^2/45`$ (Section 4), so $`X`$ is about $`1+v^2\\tau/12`$ and its randomness is of order $`v^2\\tau`$ times $`0.07`$. Under stochastic volatility the dispersion inherits the volatility risk, and with a negative correlation between price and variance it is larger on paths that end low. That asymmetry is visible in the numerical profile of Section 5.\n\n## 3. Exercising on a proxy\n\n### 3.1 The cost of the kink\n\nLet $`A`$ and $`B`$ be positive random variables with finite means; $`B`$ is the proxy. For $`K\\gt 0`$ put\n\n$$\n\\begin{gathered} C(K)=E(A-K)^+,\\qquad L_B(K)=E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\bigr],\\qquad R_B(K)=C(K)-L_B(K). \\end{gathered} \\tag{3.1}\n$$\n\n$`L_B(K)`$ is the value of the call when exercise is decided by $`B`$; we call it the proxy value and $`R_B(K)`$ the cost of the kink. For $`B=cG`$ we write $`L_c`$ and $`R_c`$. Prices are discounted at the end.\n\n**Proposition 3.1 (The cost of the kink).** For all positive numbers $`A,B,K`$,\n\n$$\n\\begin{gathered} (A-K)^+-(A-K)\\mathbf 1_{\\{B\\gt K\\}} =(A-K)\\mathbf 1_{\\{B\\le K\\lt A\\}}+(K-A)\\mathbf 1_{\\{A\\le K\\lt B\\}}\\ \\ge0 . \\end{gathered} \\tag{3.2}\n$$\n\nThe put payoff $`(K-A)^+`$ and its proxy version $`(K-A)\\mathbf{1}_{\\{B\\le K\\}}`$ differ by the same quantity. Consequently $`R_B(K)\\ge0`$ is the expected distance from $`A`$ to the strike on the event that the strike separates $`A`$ from $`B`$, for calls and puts alike. If $`\\varphi`$ is convex on $`(0,\\infty)`$ with a locally absolutely continuous derivative, then\n\n$$\n\\begin{gathered} \\int_0^\\infty R_B(K)\\,\\varphi''(K)\\,dK =E\\bigl[\\varphi(A)-\\varphi(B)-\\varphi'(B)(A-B)\\bigr], \\end{gathered} \\tag{3.3}\n$$\n\nboth sides being finite or infinite together. In particular\n\n$$\n\\begin{gathered} \\int_0^\\infty R_B(K)\\,dK=\\tfrac12E(A-B)^2,\\qquad \\int_0^\\infty R_B(K)\\,\\frac{dK}{K^2}=E\\,d(A/B),\\qquad d(x)=x-1-\\log x . \\end{gathered} \\tag{3.4}\n$$\n\n*Proof.* Identity (3.2) is a check of the four cases given by the positions of $`A`$ and $`B`$ relative to $`K`$; the put follows from parity applied to both payoffs. For (3.3), fix $`A`$ and $`B`$ and integrate the right side of (3.2) against $`\\varphi''`$. If $`B\\lt A`$, integration by parts gives\n\n$$\n\\int_B^A(A-K)\\varphi''(K)\\,dK=-(A-B)\\varphi'(B)+\\varphi(A)-\\varphi(B),\n$$\n\nand the case $`A\\lt B`$ gives the same expression. Taking expectations with Tonelli's theorem proves (3.3). The choices $`\\varphi(K)=K^2/2`$ and $`\\varphi(K)=-\\log K`$ give (3.4). $`\\square`$\n\nThe right side of (3.3) is the Bregman divergence of $`\\varphi`$ between $`A`$ and $`B`$ [[7](#ref-7)], and the integrand in (3.2) is that divergence for the call payoff itself. Formula (3.3) is the spanning of a convex payoff by calls [[6](#ref-6), [9](#ref-9)], applied to the pair of payoffs. The two cases in (3.4) are conservation laws: however the cost of the kink is spread across strikes, its total is a moment of the basis. The weights $`dK/K^2`$ are those that replicate a log contract, so a strip of proxy errors weighted like a variance swap sums to the expected log-gap $`E\\,d(A/B)`$. For $`B=cG`$ that gap involves only the shape variable $`X`$.\n\nProposition 3.1 needs no model. It also shows what a model must supply for a bound at a single strike: a limit on how much of the total mass can concentrate there.\n\n### 3.2 A common factor\n\n**Assumption (CF).** There are a $`\\sigma`$-field $`\\mathcal{G}`$, positive $`\\mathcal{G}`$-measurable random variables $`a_1,\\ldots,a_n`$ and $`\\sigma`$, and a random variable $`U`$ such that\n\n$$\n\\begin{gathered} U\\mid\\mathcal G\\sim N(0,\\sigma^2),\\qquad S_{t_i}=a_ie^{U}\\quad(i=1,\\ldots,n). \\end{gathered} \\tag{3.5}\n$$\n\nPut $`a=n^{-1}\\sum_ia_i`$ and $`g=(\\prod_ia_i)^{1/n}`$, so that $`A=ae^U`$, $`G=ge^U`$ and $`X=a/g`$. Given $`\\mathcal{G}`$ the shape of the path is known and its level is lognormal. The assumption holds in the following cases.\n\n1. Black's model with volatility $`v`$: $`U=vW_{t_1}`$, $`\\sigma=v\\sqrt{t_1}`$, and $`\\mathcal{G}`$ is generated by the Brownian increments after $`t_1`$.\n2. Stochastic volatility, $`dS_t/S_t=r\\,dt+\\sqrt{V_t}\\,(\\rho\\,dW_t+\\sqrt{1-\\rho^2}\\,dB_t)`$ with $`V`$ adapted to $`W`$ and $`B`$ independent of $`W`$. Heston, 3/2, Stein–Stein and rough volatility models are of this form.\n3. The same with jumps in price or variance that are independent of $`B`$: the jump path is placed in $`\\mathcal{G}`$.\n4. Time-stepping schemes with Gaussian price innovations, such as the Euler kernel of Section 5.\n\nIn the second case $`\\mathcal{G}`$ is generated by $`W`$ and by the increments of $`B`$ after $`t_1`$, and\n\n$$\nU=\\sqrt{1-\\rho^2}\\int_0^{t_1}\\sqrt{V_t}\\,dB_t,\\qquad \\sigma^2=(1-\\rho^2)\\int_0^{t_1}V_t\\,dt ;\n$$\n\nin the fourth the integrals are replaced by their sums over the grid. The assumption is a scaling property: the Gaussian part of the return before $`t_1`$ multiplies every fixing by the same factor and leaves the later volatility path as it was. Dynamics whose drift and volatility are free of the price level have it, and so does an average over several futures contracts that move in parallel.\n\nLet $`\\Phi`$ and $`\\phi`$ be the standard normal distribution function and density, $`\\phi_\\sigma(u)=\\phi(u/\\sigma)/\\sigma`$, and for $`x,s,K\\gt 0`$\n\n$$\n\\begin{gathered} F_K(x;s)=xe^{s^2/2}\\Phi\\!\\left(\\frac{\\log(x/K)}{s}+s\\right)-K\\Phi\\!\\left(\\frac{\\log(x/K)}{s}\\right), \\qquad \\Gamma_K(x;s)=\\frac{K}{sx^2}\\,\\phi\\!\\left(\\frac{\\log(x/K)}{s}\\right). \\end{gathered} \\tag{3.6}\n$$\n\n$`F_K(x;s)=E(xe^{sZ}-K)^+`$ for a standard normal $`Z`$ is the undiscounted Black–Scholes value of a call on a lognormal factor with total volatility $`s`$, and $`\\Gamma_K=\\partial_x^2F_K`$ is its gamma.\n\n**Proposition 3.2 (Conditional price and conditional cost).** Assume (CF) and let $`B=be^U`$ with $`b\\gt 0`$ $`\\mathcal{G}`$-measurable. Then\n\n$$\n\\begin{gathered} E\\bigl[(A-K)^+\\mid\\mathcal G\\bigr]=F_K(a;\\sigma),\\qquad E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=F_K(b;\\sigma)+(a-b)\\,\\partial_xF_K(b;\\sigma), \\end{gathered} \\tag{3.7}\n$$\n\nand the conditional cost of the kink is\n\n$$\n\\begin{gathered} E\\bigl[(A-K)^+-(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=\\int_b^a(a-x)\\,\\Gamma_K(x;\\sigma)\\,dx . \\end{gathered} \\tag{3.8}\n$$\n\nIn the log-strike $`k=\\log K`$,\n\n$$\n\\begin{gathered} \\int_b^a(a-x)\\,\\Gamma_K(x;\\sigma)\\,dx=K\\,d(a/b)\\,\\bigl(\\pi_{a,b}*\\phi_\\sigma\\bigr)(k), \\end{gathered} \\tag{3.9}\n$$\n\nwhere $`\\pi_{a,b}`$ is the probability density proportional to $`|ae^{-t}-1|`$ on the interval between $`\\log b`$ and $`\\log a`$.\n\n*Proof.* The first identity in (3.7) is the definition of $`F_K`$. Completing the square in $`e^{sz}\\phi(z)`$ gives $`E[e^{sZ}\\mathbf{1}_{\\{Z\\gt q\\}}]=e^{s^2/2}\\Phi(s-q)`$, hence $`\\partial_xF_K(b;\\sigma)=e^{\\sigma^2/2}\\Phi(\\log(b/K)/\\sigma+\\sigma)`$ and\n\n$$\nE\\bigl[(A-B)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=(a-b)\\,\\partial_xF_K(b;\\sigma),\\qquad E\\bigl[(B-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=F_K(b;\\sigma).\n$$\n\nAdding the two gives the second identity. Since $`\\sigma\\gt 0`$, $`F_K(\\cdot;\\sigma)`$ is twice continuously differentiable, and (3.8) is Taylor's formula with integral remainder between $`b`$ and $`a`$. For (3.9) substitute $`x=e^t`$: then $`\\Gamma_K(e^t;\\sigma)e^t=Ke^{-t}\\phi_\\sigma(k-t)`$ and $`(a-e^t)e^{-t}=ae^{-t}-1`$, whose integral between $`\\log b`$ and $`\\log a`$ is $`a/b-1-\\log(a/b)`$. $`\\square`$\n\nFormula (3.8) says that the cost of the kink is a gamma integrated over the basis. Formula (3.9) says where it sits: as a function of the log-strike, the conditional cost per unit of strike is the mass $`d(a/b)`$ spread by a Gaussian of standard deviation $`\\sigma`$ about a point between $`\\log b`$ and $`\\log a`$.\n\n### 3.3 The bound\n\n**Theorem 3.3 (The kink bound).** Assume (CF) and let $`B=be^U`$ with $`b\\gt 0`$ $`\\mathcal{G}`$-measurable. For every $`K\\gt 0`$,\n\n$$\n\\begin{gathered} 0\\le R_B(K)\\le\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\right], \\end{gathered} \\tag{3.10}\n$$\n\n$$\n\\begin{gathered} 0\\le R_B(K)\\le\\frac{1}{2K\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{(A-B)^2}{\\sigma}\\right]. \\end{gathered} \\tag{3.11}\n$$\n\nIn the other direction, with $`m_K=\\max\\{|\\log(K/a)|,|\\log(K/b)|\\}`$,\n\n$$\n\\begin{gathered} R_B(K)\\ge\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\,e^{-m_K^2/(2\\sigma^2)}\\right]. \\end{gathered} \\tag{3.12}\n$$\n\n*Proof.* In (3.9) the convolution is an average of values of $`\\phi_\\sigma`$, each at most $`1/(\\sigma\\sqrt{2\\pi})`$; since $`A/B=a/b`$, taking expectations gives (3.10). The points averaged lie within $`m_K`$ of $`k`$, where $`\\phi_\\sigma\\ge e^{-m_K^2/(2\\sigma^2)}/(\\sigma\\sqrt{2\\pi})`$; this gives (3.12). For (3.11) write $`u=\\log(K/x)`$. Then\n\n$$\n\\Gamma_K(x;\\sigma)=\\frac{e^{2u}}{K}\\,\\phi_\\sigma(u)=\\frac{e^{2\\sigma^2}}{K}\\,\\phi_\\sigma(u-2\\sigma^2)\\le\\frac{e^{2\\sigma^2}}{K\\sigma\\sqrt{2\\pi}} ,\n$$\n\nso the integral in (3.8) is at most $`e^{2\\sigma^2}(a-b)^2/(2K\\sigma\\sqrt{2\\pi})`$, and $`E[(A-B)^2\\mid \\mathcal{G}]=e^{2\\sigma^2}(a-b)^2`$. $`\\square`$\n\nBoth upper bounds have the same form. By (3.4) the cost integrates across strikes to $`E\\,d(A/B)`$ with weights $`dK/K^2`$ and to $`E(A-B)^2/2`$ with weights $`dK`$. Assumption (CF) limits how concentrated in strike the cost can be: no more than a Gaussian of width $`\\sigma`$ in log-strike. Each bound is one of the two masses multiplied by the peak of that Gaussian. The first is dimensionless and uniform in moneyness after division by $`K`$; the second weights the basis by the price level and tends to be the smaller of the two for strikes above the forward. The lower bound (3.12) shows that nothing is lost in (3.10) when the strike is within a fraction of $`\\sigma`$ of both averages on the paths that carry the dispersion.\n\n### 3.4 The best scale\n\n**Proposition 3.4 (Scale).** Assume (CF), $`E[1/\\sigma]\\lt \\infty`$ and $`E[X/\\sigma]\\lt \\infty`$. Let $`E_w`$ denote expectation with weights proportional to $`1/\\sigma`$. Over constants $`c\\gt 0`$, the right side of (3.10) with $`B=cG`$ is smallest at\n\n$$\n\\begin{gathered} c_*=E_w[X]=\\frac{E[X/\\sigma]}{E[1/\\sigma]},\\qquad\\text{where it equals}\\qquad \\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac1\\sigma\\right]\\bigl(\\log E_wX-E_w\\log X\\bigr). \\end{gathered} \\tag{3.13}\n$$\n\nThe right side of (3.11) is smallest at\n\n$$\n\\begin{gathered} c_\\dagger=\\frac{E[AG/\\sigma]}{E[G^2/\\sigma]} . \\end{gathered} \\tag{3.14}\n$$\n\n*Proof.* $`E[d(X/c)/\\sigma]=E[1/\\sigma]\\,(E_wX/c-1-E_w\\log X+\\log c)`$, which is convex in $`1/c`$ with its minimum at $`c=E_wX`$. The second claim is a least-squares projection. $`\\square`$\n\nThe bracket in (3.13) is the mean log deviation of $`X`$, Theil's second measure of dispersion: the gap in Jensen's inequality for the logarithm. It vanishes only if $`X`$ is constant, and for small dispersion it is half the squared coefficient of variation of $`X`$. The optimal scale has a plain meaning: $`c_*-1`$ is the expected markup of the arithmetic over the geometric average, about $`v^2\\tau/12`$ in Black's model. The Heston certificate of Section 5 uses a rational scale near $`c_\\dagger`$, $`1254433/1250000=1.0035464`$; the Black value $`1+v^2\\tau(n+1)/(12n)`$ for its contract is $`1.0037`$.\n\n**Corollary 3.5 (Classical lower bounds).** Under (CF), for every $`c\\gt 0`$,\n\n$$\n\\begin{gathered} 0\\le C(K)-E\\bigl(E[A\\mid G]-K\\bigr)^+\\le C(K)-\\sup_{\\lambda\\gt 0}E\\bigl[(A-K)\\mathbf 1_{\\{G\\gt \\lambda\\}}\\bigr]\\le R_c(K), \\end{gathered} \\tag{3.15}\n$$\n\nso the bounds of Theorem 3.3 with $`B=cG`$ apply to the lower bound of Curran and of Rogers and Shi and to the optimized lower bound of Fusai and Kyriakou.\n\n*Proof.* $`E(E[A\\mid G]-K)^+`$ is the largest value of $`E[(A-K)\\mathbf{1}_H]`$ over events $`H`$ determined by $`G`$, and $`\\{G\\gt \\lambda\\}`$ is such an event; the choice $`\\lambda=K/c`$ gives $`L_c(K)`$. $`\\square`$\n\n### 3.5 Reading the bound\n\nThree remarks connect the inequalities to trading intuition.\n\nThe cost is a gamma. By (3.8), $`R_B(K)`$ is close to $`E[\\Gamma_K(b;\\sigma)(a-b)^2]/2`$ when the basis is small: half the gamma of the conditional option times the squared basis. Figure 3 shows this second-order term next to the exact cost for the contract of Section 5; at the two strikes of that contract they agree to within three percent.\n\nThe two risks enter as a ratio. $`\\sigma`$ is level risk realized before averaging begins and $`d(A/cG)`$ is dispersion risk realized inside the window. A long lead time and a short window make the proxy nearly exact; an average that starts tomorrow leaves one day of smoothing. Only the part of the early return that is independent of volatility counts, which is why $`\\sigma^2`$ carries the factor $`1-\\rho^2`$: with $`\\rho=-0.7`$, a typical equity value, the bound is 1.4 times what it would be at $`\\rho=0`$.\n\nThe bound is attained when dispersion meets the strike. It charges every path as if it sat at the strike, and (3.12) measures what that charge gives up: the factor $`e^{-m_K^2/(2\\sigma^2)}`$ on paths that have moved away. For short windows with long leads the factor is close to one, and bound and cost agree to within about ten percent (Table 1). For long windows a path with large dispersion has usually moved far, and the factor accounts for the difference between the two.\n\n## 4. How large is the cost? A cubic law\n\n### 4.1 Black's model\n\nIn Black's model [[3](#ref-3)] the fixings are $`S_{t_i}=F_i\\exp(-v^2t_i/2+vW_{t_i})`$ with forwards $`F_i`$ and volatility $`v`$, the common factor has the deterministic volatility $`\\sigma=v\\sqrt{t_1}`$, and everything in Theorem 3.3 is a Gaussian moment.\n\n**Proposition 4.1 (Black's model).** Let $`y_i=\\log S_{t_i}-\\log G`$. The vector $`y`$ is Gaussian with means $`m_i`$ and covariances $`c_{ij}`$ given by\n\n$$\n\\begin{aligned} m_i&=\\log F_i-\\tfrac12v^2t_i-\\frac1n\\sum_j\\bigl(\\log F_j-\\tfrac12v^2t_j\\bigr),\\\\ c_{ij}&=v^2\\Bigl[t_i\\wedge t_j-\\frac1n\\sum_k\\bigl(t_i\\wedge t_k+t_j\\wedge t_k\\bigr)+\\frac1{n^2}\\sum_{k,l}t_k\\wedge t_l\\Bigr], \\end{aligned}\n$$\n\nand $`c_{ij}`$ depends on the dates only through the differences $`t_i-t_1`$. Then\n\n$$\n\\begin{gathered} E\\,X=\\frac1n\\sum_ie^{m_i+c_{ii}/2},\\qquad E\\,X^2=\\frac1{n^2}\\sum_{i,j}e^{m_i+m_j+(c_{ii}+c_{jj})/2+c_{ij}}, \\end{gathered} \\tag{4.1}\n$$\n\n$`c_*=E\\,X`$, and for every strike\n\n$$\n\\begin{gathered} R_{c_*}(K)\\le\\frac{K}{v\\sqrt{2\\pi t_1}}\\Bigl[\\log E\\,X-(E\\,X-1)+\\tfrac12E(X-1)^2\\Bigr]. \\end{gathered} \\tag{4.2}\n$$\n\nFor $`n`$ equally spaced fixings over a window $`\\tau`$ with a flat forward curve, as $`v^2\\tau\\to0`$,\n\n$$\n\\begin{gathered} E\\,X=1+\\frac{v^2\\tau}{2}\\,\\mu_n+O(v^4\\tau^2),\\qquad \\mathop{\\mathrm{Var}}X=\\frac{v^4\\tau^2}{4}\\,q_n+O(v^6\\tau^3), \\end{gathered} \\tag{4.3}\n$$\n\n$$\n\\mu_n=\\frac{n+1}{6n},\\qquad q_n=\\frac{(n+1)(2n^2+7)}{90\\,n^2(n-1)},\n$$\n\nand the right side of (4.2) equals\n\n$$\n\\begin{gathered} K\\,\\frac{v^3\\tau^2}{8\\sqrt{2\\pi t_1}}\\,q_n\\,\\bigl(1+O(v^2\\tau)\\bigr) \\;\\longrightarrow\\; K\\,\\frac{v^3\\tau^2}{360\\sqrt{2\\pi t_1}}\\quad(n\\to\\infty). \\end{gathered} \\tag{4.4}\n$$\n\n*Proof.* The moments of $`y`$ are read from the Brownian covariance $`v^2(t_i\\wedge t_j)`$, and (4.1) is the Gaussian exponential moment. Since $`\\sigma`$ is constant, (3.13) gives $`c_*=E\\,X`$. On $`x\\ge1`$, $`\\log x\\ge(x-1)-(x-1)^2/2`$; applying this to $`X\\ge1`$ in $`E\\,d(X/c_*)=\\log E\\,X-E\\log X`$ and inserting the result in (3.10) gives (4.2). The expansions (4.3) and the constants $`\\mu_n`$ and $`q_n`$ are computed in Appendix I. $`\\square`$\n\nFormula (4.4) is a cubic law. The bound grows with the cube of volatility and the square of the window, and it falls with the square root of the lead time. The window enters squared because the dispersion surprise is of order $`v^2\\tau`$ and the cost is quadratic in it; one power of $`v`$ is returned by the smoothing $`1/(v\\sqrt{t_1})`$. Since the law of $`X`$ does not involve $`t_1`$, the dependence on the lead time is exactly $`t_1^{-1/2}`$ at every volatility. The numbers $`1/6`$ and $`1/45`$ behind $`\\mu_n`$ and $`q_n`$ are the mean and variance of $`\\int_0^1(W_t-\\bar W)^2dt`$ for a Brownian motion $`W`$ with time average $`\\bar W`$. Figure 2 plots the exact right side of (4.2) against the leading term.\n\n\n\n### 4.2 Contracts that trade\n\nTable 1 applies Proposition 4.1 to contracts with the terms of traded ones. The proxy value has the closed form of the conditioning literature: $`\\log G`$ is normal and each fixing is lognormal jointly with it, so $`L_c(K)`$ is a sum of $`n+1`$ normal distribution values. Every enclosure in the table is computed in ball arithmetic at 256 bits from exact rational inputs, so that under Black's model with the stated inputs each price lies between the proxy value and the proxy value plus the bound. The inputs are round numbers of the right order for each market. A year has 252 trading days.\n\n**Table 1. Proved price intervals in Black's model.**\n\n| Contract | Proxy value | Bound | Quote steps | Simulated | First order |\n| --- | --- | --- | --- | --- | --- |\n| Crude oil call, 1 month ahead | 3.206408 | 0.0000805 | 0.0081 | 0.0000523 | 0.0259 |\n| Crude oil call, 3 months ahead | 5.058865 | 0.0000462 | 0.0047 | 0.0000388 | 0.0262 |\n| Crude oil call, 6 months ahead | 6.906596 | 0.0000324 | 0.0033 | 0.0000293 | 0.0261 |\n| Crude oil call, 12 months ahead | 9.431726 | 0.0000224 | 0.0023 | 0.0000210 | 0.0256 |\n| Crude oil put at 60, 3 months ahead | 1.315689 | 0.0000396 | 0.0040 | 0.0000278 | 0.0262 |\n| Crude oil call at 80, 3 months ahead | 1.850312 | 0.0000417 | 0.0042 | 0.0000333 | 0.0262 |\n| Copper call, 3 months ahead | 432.887072 | 0.00170 | 0.17 | 0.00142 | 1.473 |\n| EUR/USD call, quarterly average (in pips) | 198.100004 | 0.000767 | 0.0008 | 0.000499 | 0.628 |\n| Annual-average oil put at 70 | 1.723884 | 0.009084 | 0.91 | 0.004235 | 0.5206 |\n| Twelve monthly fixings, call at 95 | 8.216008 | 0.003573 | 0.36 | 0.000750 | 0.1334 |\n| Twelve monthly fixings, call at 110 | 1.994944 | 0.003566 | 0.36 | 0.000762 | 0.1334 |\n| Daily fixings from tomorrow | 4.434265 | 0.013895 | 1.39 | 0.000679 | 0.1053 |\n\nThe crude oil rows are options on a calendar month of 21 daily settlements of a futures price of 70 dollars with 35 percent volatility, struck at 70 unless stated; the copper row has 22 settlements of a 9,500-dollar price with 22 percent volatility; the currency row has 63 fixings of a rate of 1.10 with 8 percent volatility, struck at the money, three months ahead. The annual put has 252 fixings of a 100-dollar price with 40 percent volatility starting four months ahead. The twelve-fixing rows are the contract of Section 5 with its variance frozen at 0.045. The last row has 252 daily fixings of a price of 100 with 20 percent volatility. Interest is 4 percent, except 1 percent in the twelve-fixing rows. The column headed Bound is the proved bound on the cost of the kink, the smaller of (3.10) and (3.11), each at its best scale, rounded up; the proxy value is rounded to the digits shown, and each price lies between the proxy value and the proxy value plus the bound. A quote step is one cent per barrel, one cent per tonne, one pip, or 0.01. The last two columns are computed in binary64 arithmetic: a Monte Carlo estimate of the true cost on $`2^{20}`$ paths, and $`E[\\mathrm{sd}(A\\mid G)]/2`$, the Rogers–Shi form of the classical first-order bound.\n\nFour observations follow.\n\nFor the monthly contracts the proxy value is the price for every commercial purpose. The exchange quotes these options in dollars and cents per barrel; the proved bound is between a 120th and a 440th of a cent. A market maker who exercises on the geometric average gives up less than ten dollars on a hundred contracts of a thousand barrels, whose premium exceeds three hundred thousand.\n\nThe bound is nearly attained. Its ratio to the simulated cost is 1.54, 1.19, 1.10 and 1.07 at one, three, six and twelve months, as the factor in (3.12) predicts: with a long lead the strike is within a small multiple of $`\\sigma`$ of both averages on almost every path. The classical first-order estimate is 2.6 cents at every lead time, between 320 and 1,140 times larger.\n\nScale changes the verdict. The annual put has the terms made public for Mexico's 2009 program [[4](#ref-4)], a one-year average struck at 70 dollars on 330 million barrels, valued here at 40 percent volatility. The bound is 0.91 cents per barrel. On the full volume that is 3.0 million dollars against a proxy value of 569 million, and the simulated cost is 1.4 million. A proved interval of half a percent is a statement worth having on a trade of this size, and the cubic law says what moves it: at 50 percent volatility the bound would be about twice as large.\n\nThe lead time is the resource. The last row is the same kind of contract as the first with the window lengthened to a year and the lead cut to a day. Theorem 3.3 draws on the one day of level risk before the first fixing, and the bound is 1.4 quote steps; the later fixings smooth the kink further, and the cost is a fifteenth of a quote step. For such contracts the theorem stands beside the lognormal bounds of Thompson [[33](#ref-33)] and Lord [[26](#ref-26)], which use the full covariance of the fixings.\n\n### 4.3 Stochastic volatility\n\nTheorem 3.3 holds for every model that satisfies (CF), and under stochastic volatility the dispersion acquires a second source of risk. Table 2 repeats five rows of Table 1 under Heston dynamics, $`dV_t=\\kappa(\\bar v-V_t)\\,dt+\\xi\\sqrt{V_t}\\,dW_t`$, with $`\\kappa=2`$, $`\\rho=-0.3`$ and $`V_0=\\bar v`$ equal to the squared volatility of Table 1; $`\\xi=0.7`$ for the monthly options and $`0.8`$ for the annual put, which puts the Feller index $`2\\kappa\\bar v/\\xi^2`$ at one. The entries are Monte Carlo estimates in which the first-period factor is integrated in closed form on each path (Proposition 3.2), so that the cost and the bound are sample means of smooth functions.\n\n**Table 2. Simulated cost of the kink under Heston dynamics.**\n\n| Contract | Value | Cost of the kink | Bound | Bound over cost | $`E[1/\\sigma]`$ |\n| --- | --- | --- | --- | --- | --- |\n| Crude oil call, 1 month ahead | 3.147 | 0.0000776 | 0.000132 | 1.70 | 10.78 |\n| Crude oil call, 3 months ahead | 4.875 | 0.0000730 | 0.0000942 | 1.29 | 6.56 |\n| Crude oil call, 6 months ahead | 6.561 | 0.0000624 | 0.0000738 | 1.18 | 4.80 |\n| Crude oil call, 12 months ahead | 8.891 | 0.0000496 | 0.0000576 | 1.16 | 3.41 |\n| Annual-average oil put at 70 | 2.053 | 0.00965 | 0.0296 | 3.06 | 5.05 |\n\nEach row uses $`2^{19}`$ paths ($`2^{18}`$ for the annual put) of absorbed Euler with four steps per trading day (two for the annual put). Standard errors are below 0.8 percent of each cost and bound, except 1.1 and 1.6 percent for the cost and bound of the annual put, and about 0.003 for the option values (0.009 for the annual put). The bound is the smaller of (3.10) at $`c_*`$ and (3.11) at $`c_\\dagger`$.\n\nVolatility of volatility roughly doubles the cost of the kink at three months and beyond, and the bound stays within a factor of 1.3 of it. For the monthly options the cost stays below a hundredth of a cent. For the annual put it is 0.96 cents per barrel and the bound 2.96 cents: 3.2 and 9.8 million dollars on 330 million barrels, against a value of 677 million. A long window is where a model with random volatility makes exercising on the geometric average expensive enough to measure, and where the interval of the next section earns its place.\n\n## 5. A proved price interval under Heston\n\nThis section turns Theorem 3.3 into a certificate for a contract whose cost of the kink is not negligible, under the continuous Heston model and under the time-stepping scheme used to simulate it.\n\n### 5.1 Model, scheme and contract\n\nLet $`W,B`$ be independent standard Brownian motions and $`\\theta=(\\kappa,\\bar v,\\xi,\\rho,v_0)`$ with $`S_0,\\kappa,\\bar v,v_0\\gt 0`$, $`\\xi\\ge0`$, $`|\\rho|\\lt 1`$ and $`r\\in \\mathbb{R}`$. The continuous model $`P`$ is\n\n$$\n\\begin{aligned} dV_t&=\\kappa(\\bar v-V_t)\\,dt+\\xi\\sqrt{V_t}\\,dW_t,\\\\ dZ_t&=(r-V_t/2)\\,dt+ \\sqrt{V_t}\\bigl(\\rho\\,dW_t+\\sqrt{1-\\rho^2}\\,dB_t\\bigr),\\\\ S_t&=S_0e^{Z_t},\\qquad Z_0=0,\\quad V_0=v_0. \\end{aligned} \\tag{5.1}\n$$\n\nOn a grid of step $`h`$, with every $`t_i/h`$ an integer, the discrete law $`Q_h`$ is\n\n$$\n\\begin{aligned} V_{j+1}^h&=\\left[(1-\\kappa h)V_j^h+\\kappa\\bar v h+ \\xi\\sqrt{hV_j^h}\\,G_j\\right]^+,\\\\ Z_{j+1}^h&=Z_j^h+(r-V_j^h/2)h+ \\sqrt{hV_j^h}\\bigl(\\rho G_j+\\sqrt{1-\\rho^2}H_j\\bigr), \\end{aligned} \\tag{5.2}\n$$\n\nwith independent standard normal $`G_j,H_j`$ and the same initial values. This is Euler with absorption in the classification of [[27](#ref-27)]: a negative variance proposal is set to zero, and the price update uses the variance at the start of the step.\n\nThe contract is a call spread on the average of (2.1),\n\n$$\n\\begin{gathered} \\psi(A)=(A-K_1)^+-(A-K_2)^+,\\qquad 0\\lt K_1\\lt K_2 . \\end{gathered} \\tag{5.3}\n$$\n\nA spread keeps the payoff bounded and is the form in which capped average-price protection is sold. Write\n\n$$\n\\begin{gathered} p_c=e^{-rT}E_P\\psi(A),\\qquad p_h=e^{-rT}E_{Q_h}\\psi(A),\\qquad e_h=p_h-p_c , \\end{gathered} \\tag{5.4}\n$$\n\nand $`Q=Q_h`$ when the step is understood. The numerical contract has twelve monthly fixings,\n\n$$\n\\begin{gathered} n=12,\\quad t_i=i/12,\\quad T=1,\\quad S_0=100,\\quad r=.01,\\quad K_1=95,\\quad K_2=110, \\end{gathered} \\tag{5.5}\n$$\n\nand the parameter point and scale are\n\n$$\n\\begin{gathered} \\theta_*=(3,\\;9/200,\\;23/100,\\;-11/20,\\;9/200),\\qquad c=\\frac{1254433}{1250000},\\qquad h=\\frac1{768}. \\end{gathered} \\tag{5.6}\n$$\n\nThe long-run volatility is 21.2 percent, the Feller index $`2\\kappa\\bar v/\\xi^2`$ is $`2700/529`$, about 5.1, and averaging begins after one month and lasts eleven. By the cubic law this is a hard case for the bound: the Black row of Table 1 gives 0.36 cents on a price of 100.\n\n### 5.2 The common factor and the boundary\n\nCondition on the whole variance driver $`W`$ and on the increments $`B_t-B_{t_1}`$, $`t\\ge t_1`$. With\n\n$$\nI_P=\\int_0^{t_1}V_t\\,dt,\\qquad I_Q=h\\sum_{j=0}^{t_1/h-1}V_j^h ,\n$$\n\nassumption (CF) holds under both laws:\n\n$$\n\\begin{gathered} U\\mid\\mathcal G\\sim N\\bigl(0,(1-\\rho^2)I_\\nu\\bigr),\\qquad S_{t_i}=e^Ua_i,\\qquad \\nu\\in\\{P,Q\\}. \\end{gathered} \\tag{5.7}\n$$\n\nUnder $`P`$, $`v_0\\gt 0`$ and path continuity give $`I_P\\gt 0`$ almost surely. Under $`Q`$,\n\n$$\n\\begin{gathered} I_Q\\ge hv_0\\gt 0 . \\end{gathered} \\tag{5.8}\n$$\n\nThe scheme has an atom at zero variance; already at the first step\n\n$$\n\\begin{gathered} Q_h(V_1^h=0)= \\Phi\\!\\left( -\\frac{(1-\\kappa h)v_0+\\kappa\\bar v h}{\\xi\\sqrt{hv_0}} \\right)\\in(0,1). \\end{gathered} \\tag{5.9}\n$$\n\nThe smoothing survives the atom: the common factor needs a positive integrated variance over the first month, and (5.8) supplies it on every path. Appendix B accounts for the effect of the atom on the transforms of the scheme.\n\n### 5.3 The proxy value and the second moment\n\nExercising both legs of the spread on $`cG`$ gives the payoff\n\n$$\n\\begin{gathered} \\mathcal L=(A-K_1)\\mathbf 1_{\\{cG\\gt K_1\\}}-(A-K_2)\\mathbf 1_{\\{cG\\gt K_2\\}} =\\psi(cG)+(A-cG)\\mathbf 1_{\\{K_1\\lt cG\\le K_2\\}} . \\end{gathered} \\tag{5.10}\n$$\n\nWith $`R_{K,\\nu}`$ the cost of the kink $`R_c(K)`$ under $`\\nu`$, Proposition 3.1 gives\n\n$$\n\\begin{gathered} E_\\nu\\psi(A)=E_\\nu\\mathcal L+R_{K_1,\\nu}-R_{K_2,\\nu},\\qquad R_{K,\\nu}\\ge0 . \\end{gathered} \\tag{5.11}\n$$\n\nThe proxy value is linear in the two measures $`\\nu(Y\\in dy)`$ and $`E_\\nu[A;Y\\in dy]`$, where $`Y=\\log(G/S_0)`$. With $`y_K=\\log(K/(cS_0))`$, $`F_{0,\\nu}(y)=\\nu(Y\\le y)`$ and $`F_{A,\\nu}(y)=E_\\nu[A\\mathbf{1}_{\\{Y\\le y\\}}]`$,\n\n$$\n\\begin{gathered} E_\\nu\\mathcal L=(K_2-K_1) +K_1F_{0,\\nu}(y_{K_1})-K_2F_{0,\\nu}(y_{K_2}) -F_{A,\\nu}(y_{K_1})+F_{A,\\nu}(y_{K_2}). \\end{gathered} \\tag{5.12}\n$$\n\nBoth distribution functions are inverted from the transforms\n\n$$\n\\begin{gathered} \\widehat\\mu_0(z)=E_\\nu e^{zY},\\qquad \\widehat\\mu_A(z)=\\frac{S_0}{n}\\sum_{i=1}^n E_\\nu e^{Z_{t_i}+zY}, \\end{gathered} \\tag{5.13}\n$$\n\nwhich are exponential-affine in the fixing-date log prices: one loading for the law of the geometric average and one for each term of the arithmetic average, thirteen when $`n=12`$. Under $`P`$ they solve Riccati equations; under $`Q`$ they are computed by an algebraic recursion plus a correction for the positive part (Section 6 and Appendix B).\n\nThe certificate uses the second bound of Theorem 3.3. By (5.7) its moment is\n\n$$\n\\begin{gathered} D_\\nu(c)=E_\\nu\\!\\left[\\frac{(A-cG)^2}{\\sigma}\\right]=\\frac{W_\\nu}{\\sqrt{1-\\rho^2}},\\qquad W_\\nu=E_\\nu\\!\\left[(A-cG)^2I_\\nu^{-1/2}\\right], \\end{gathered} \\tag{5.14}\n$$\n\nand the square expands into products of fixings,\n\n$$\n\\begin{gathered} (A-cG)^2=\\frac1{n^2}\\sum_{i,j}S_{t_i}S_{t_j} -\\frac{2c}{n}\\sum_iS_{t_i}G+c^2G^2 , \\end{gathered} \\tag{5.15}\n$$\n\nwhich for $`n=12`$ are 78 distinct pairs, twelve cross terms and one square: 91 real loadings. The weight $`I^{-1/2}`$ is a Laplace integral of $`e^{-x^2I}`$, which adds a killing rate in the first month. Appendix A encloses $`W_\\nu`$ by validated quadrature of these 91 transforms. The quadratic form is preferred to (3.10) here because its integrand is a finite sum of exponential-affine terms, and because with $`\\rho\\lt 0`$ it is the smaller bound at both strikes.\n\n### 5.4 The certificate\n\n**Theorem 5.1 (Price and gap intervals).** Let $`\\nu\\in\\{P,Q\\}`$ satisfy (5.7) and suppose that\n\n$$\ne^{-rT}E_\\nu\\mathcal L\\in[\\ell_\\nu,u_\\nu],\\qquad D_\\nu(c)\\le d_\\nu .\n$$\n\nWith $`\\gamma=e^{-rT}/(2\\sqrt{2\\pi})`$, the price under $`\\nu`$ satisfies\n\n$$\n\\begin{gathered} e^{-rT}E_\\nu\\psi(A)\\in\\left[\\ell_\\nu-\\frac{\\gamma d_\\nu}{K_2},\\;u_\\nu+\\frac{\\gamma d_\\nu}{K_1}\\right]. \\end{gathered} \\tag{5.16}\n$$\n\nIf moreover $`e^{-rT}(E_Q\\mathcal{L}-E_P\\mathcal{L})\\in[\\ell,u]`$, then\n\n$$\n\\begin{gathered} e_h\\in \\left[ \\ell-\\gamma\\!\\left(\\frac{d_Q}{K_2}+\\frac{d_P}{K_1}\\right), \\quad u+\\gamma\\!\\left(\\frac{d_Q}{K_1}+\\frac{d_P}{K_2}\\right) \\right]. \\end{gathered} \\tag{5.17}\n$$\n\n*Proof.* By (3.11), $`0\\le e^{-rT}R_{K,\\nu}\\le\\gamma d_\\nu/K`$. Insert this in (5.11) for (5.16), and subtract (5.11) for the two laws for (5.17): the lower end collects $`-R_{K_2,Q}-R_{K_1,P}`$ and the upper end $`R_{K_1,Q}+R_{K_2,P}`$. $`\\square`$\n\n**Corollary 5.2 (The numerical certificate).** For the model (5.1), the scheme (5.2) and the data (5.5)–(5.6),\n\n$$\n\\begin{gathered} W_P\\le2.172471592802,\\qquad W_Q\\le2.242189951600, \\end{gathered} \\tag{5.18}\n$$\n\n$$\n\\begin{gathered} e_h\\in[-.011024692273,\\;.010642371599],\\qquad |e_h|\\lt .011025, \\end{gathered} \\tag{5.19}\n$$\n\n$$\n\\begin{gathered} p_c\\in[6.508371733,\\;6.518868974],\\qquad p_h\\in[6.507844281,\\;6.519014106]. \\end{gathered} \\tag{5.20}\n$$\n\n*Proof.* Appendices A–C specify the validated computation. The finite inversion sums for the proxy value are $`6.513251661289`$ under $`P`$ and $`6.513048668980`$ under $`Q`$, each enclosed to 24 digits, and their difference is $`-.000202992309`$. The complete error radius of the difference, covering projection, both frequency tails, periodization and rounding, is at most $`.000594238060`$. Combining these with (5.18) in (5.16) and (5.17) with exact rational endpoints gives the statement. Appendix G describes the programs, the exact endpoints, and the second implementation that recomputes the whole catalog. $`\\square`$\n\nTable 3 lists the budget behind (5.19).\n\n**Table 3. Error budget of the Heston certificate at 768 steps.**\n\n| Quantity | Enclosure or upper bound |\n| --- | --- |\n| Cost of the kink, two laws and two strikes | $`[-.010227461904,.010251125849]`$ |\n| Projection error per transform under $`Q`$ | $`.000000012711979137`$ |\n| Projection contribution to the proxy value | $`.000004310413`$ |\n| Frequency tail, continuous model | $`.000014095712`$ |\n| Frequency tail, discrete model | $`.000184373301`$ |\n| Periodization, one model | $`.000195729318`$ |\n| Total error radius of the proxy-value difference | $`.000594238060`$ |\n| **Absolute bound on the Euler gap** | $`.\\mathbf{0}\\mathbf{1}\\mathbf{1}\\mathbf{0}\\mathbf{2}\\mathbf{5}`$ |\n\nReported widths are computed from exact rational endpoints and rounded outward, so differences of displayed endpoints can disagree with them in the last place.\n\n### 5.5 What the interval contains\n\nThe interval for $`p_c`$ is $`.0105`$ wide, 16 basis points of the price. Its center is the proxy value, known to $`\\pm.00021`$, and almost all of its width is the allowance for the two costs of the kink, $`.005407`$ above and $`.004670`$ below.\n\nA simulation locates the price inside the interval. On $`2^{20}`$ paths with the first-period factor integrated in closed form, the discounted cost of the kink is $`.000813`$ at $`K_1`$ and $`.000711`$ at $`K_2`$, with standard errors below $`4\\cdot10^{-6}`$. The two enter the spread with opposite signs, so the spread exceeds its proxy value by $`.000102`$ with standard error $`.000004`$, and the continuous price is close to $`6.51335`$. The certificate bounds each cost on its own and from one side, so its allowances add. They are about 6.6 times the simulated costs, in line with the Black rows of Table 1, where the ratio is 4.7 before volatility of volatility is added. The lower bound (3.12), evaluated on the same paths, is $`.000777`$ and $`.000682`$, within five percent of the two costs. The distance between the upper bound and the cost is therefore the Gaussian factor in (3.12): on most paths that carry dispersion the strike is several $`\\sigma`$ away from both averages.\n\nFigure 3 shows the cost across strikes. Three features agree with Section 3. The area under the curve, with the weights of (3.4), is $`8.6063\\cdot10^{-6}`$ by numerical integration and $`E\\,d(A/cG)=8.6063\\cdot10^{-6}`$ by direct simulation; with weights $`dK`$ the two sides are $`.0651615`$ and $`.0651612`$. The second-order term $`E[\\Gamma_K(cg;\\sigma)(a-cg)^2]/2`$ is within three percent of the cost at both strikes of the spread. And the profile leans toward low strikes: per unit of strike the cost peaks near $`K=77`$, at 2.2 times its value at $`K=100`$. With $`\\rho\\lt 0`$, paths that end low have had high variance, and high variance is what separates the two averages.\n\n\n\n### 5.6 Scale, steps and a second point\n\nBy Proposition 3.4, applied to the two laws together as they enter (5.17), the exact quadratic allowance is smallest at\n\n$$\n\\begin{gathered} c_\\dagger=1+\\frac{E_P^++E_Q^+}{H_P+H_Q},\\qquad H_\\nu=E_\\nu[G^2I_\\nu^{-1/2}],\\qquad E_\\nu^+=E_\\nu[(A-G)GI_\\nu^{-1/2}]\\ge0, \\end{gathered} \\tag{5.21}\n$$\n\nwhere $`E_\\nu^+\\ge0`$ by the inequality of arithmetic and geometric means. Both moments belong to the catalog of 91 loadings, and Appendix A encloses them with the same quadrature, which gives\n\n$$\n\\begin{gathered} c_\\dagger\\in[1.003427394603,\\;1.003669831097]. \\end{gathered} \\tag{5.22}\n$$\n\nTable 4 compares three scales. The scale of (5.6) is inside the enclosure and the optimum is flat; the choice $`c=1`$, the unscaled geometric average, would double the allowance.\n\n**Table 4. Effect of the scale on the allowance for the cost of the kink.**\n\n| Scale $`c`$ | Width of the allowance at $`h=1/768`$ |\n| --- | --- |\n| 1 | .040551110717 |\n| 1.0035464 (used) | .020478587752 |\n| 1.0035486 (rounded midpoint of (5.22)) | .020478592980 |\n\nTable 5 repeats the certificate at other steps, and Table 6 separates the widths.\n\n**Table 5. Proved intervals for the Euler gap at three steps.**\n\n| Step | Interval for $`e_h`$ | Width |\n| --- | --- | --- |\n| $`1/384`$ | $`[-0.014169989568,\\;0.013839530031]`$ | 0.028009519599 |\n| $`1/768`$ | $`[-0.011024692273,\\;0.010642371599]`$ | 0.021667063872 |\n| $`1/1536`$ | $`[-0.010967952032,\\;0.010766337995]`$ | 0.021734290027 |\n\n**Table 6. Width contributions at the three certified steps.**\n\n| Contribution | $`1/384`$ | $`1/768`$ | $`1/1536`$ |\n| --- | --- | --- | --- |\n| Cost of the kink | 0.026754977094 | 0.020478587752 | 0.020171631177 |\n| Projection | 0.000265787793 | 0.000008620825 | 0.000000265628 |\n| Frequency tail under $`P`$ | 0.000028191423 | 0.000028191423 | 0.000028191423 |\n| Frequency tail under $`Q`$ | 0.000177646019 | 0.000368746602 | 0.000751284529 |\n| Periodization | 0.000782917272 | 0.000782917272 | 0.000782917272 |\n| Rounding and endpoints | $`10^{-12}`$ | $`10^{-12}`$ | $`10^{-12}`$ |\n\nAt the three steps the cost of the kink is 95.5, 94.5 and 92.8 percent of the width. The width settles near $`.0217`$. Between 768 and 1536 steps the allowance for the cost of the kink changes by 1.5 percent, and the inversion terms make up the rest; among them the tail bound for the discrete transforms rests on the floor $`I_Q\\ge hv_0`$ of (5.8) and roughly doubles when the step is halved. The projection argument of Appendix B, whose exponent must exceed 8, holds at each of the three steps. What the Euler error itself does under refinement is the subject of Section 6.\n\nThe certificate also holds at a second parameter point, $`v_0=.04`$ with everything else unchanged and $`h=1/768`$:\n\n**Table 7. The certificate at initial variance .04.**\n\n| Quantity | Enclosure or upper bound |\n| --- | --- |\n| Continuous price | $`[6.499318797594,\\;6.509948941256]`$ |\n| Price under the scheme | $`[6.498150434416,\\;6.510702536853]`$ |\n| Euler gap | $`[-0.011798506840,\\;0.011383739259]`$ |\n| Width of the gap interval | 0.023182246099 |\n\nThe gap interval is 7.0 percent wider than at $`v_0=.045`$, and 83 percent of the increase comes from the discrete frequency tail, which rises from $`.000368746601`$ to $`.001629880668`$: a smaller $`v_0`$ lowers the floor in (5.8) and raises the first-month Laplace bound in (C.7). Theorem 5.1 itself holds for any spread, fixing schedule and parameter point; each certificate is one run of the validated computation that supplies its inputs.\n\n## 6. What discretization does to the price\n\nCorollary 5.2 bounds the Euler gap by $`.011025`$. This section shows that the gap itself is about fifty times smaller, derives its order and sign from the structure of the scheme, and shows how a coupling of the two laws brings the proved allowance down by two orders of magnitude.\n\n### 6.1 Euler on the price is Euler on the Riccati equation\n\nWrite $`\\alpha=\\xi^2/2`$ and $`d=\\kappa\\bar v`$. For a complex stock loading $`q`$, the transforms $`E_P\\exp(qZ_t+bV_t)=\\exp(\\mathcal{A}(t)+\\mathcal{B}(t)v_0)`$ of the continuous model are governed by\n\n$$\n\\begin{gathered} \\mathcal B'=\\alpha\\mathcal B^2-(\\kappa-\\rho\\xi q)\\mathcal B+\\tfrac12(q^2-q),\\qquad \\mathcal A'=rq+d\\mathcal B . \\end{gathered} \\tag{6.1}\n$$\n\n**Proposition 6.1 (The scheme is affine up to its projection).** Let $`v\\ge0`$, let $`\\widetilde V=(1-\\kappa h)v+\\kappa\\bar vh+\\xi\\sqrt{hv}\\,G`$ be the variance proposal of (5.2) from the state $`V_j^h=v`$, and let $`\\Delta Z`$ be the log-price increment of the same step. For complex $`q`$ and $`b`$,\n\n$$\n\\begin{gathered} E\\exp\\bigl(q\\,\\Delta Z+b\\,\\widetilde V\\bigr) =\\exp\\Bigl\\{h(rq+db)+v\\bigl[b+h\\bigl(\\alpha b^2-(\\kappa-\\rho\\xi q)b+\\tfrac12(q^2-q)\\bigr)\\bigr]\\Bigr\\}. \\end{gathered} \\tag{6.2}\n$$\n\nOne step of the scheme therefore acts on exponential-affine functions as one explicit Euler step of size $`h`$ of the system (6.1), followed by the replacement of $`\\widetilde V`$ by $`\\widetilde V^+`$.\n\n*Proof.* The exponent $`q\\,\\Delta Z+b\\widetilde V`$ is Gaussian given $`v`$, with mean $`q(r-v/2)h+b(1-\\kappa h)v+bdh`$ and variance $`hv[(q\\rho+b\\xi)^2+q^2(1-\\rho^2)]`$. Adding half the variance to the mean gives (6.2). $`\\square`$\n\nTwo consequences follow. First, without the positive part the transforms of the scheme would be the explicit Euler iterates of a polynomial ordinary differential equation, and the global error of that method has an asymptotic expansion in powers of $`h`$ [[18](#ref-18)]: for each loading the Euler transform differs from the continuous one by $`b_1h+b_2h^2+\\cdots`$ with computable coefficients. Second, the positive part acts only when the proposal is negative, an event of probability (5.9) at the first step. Appendix B bounds its accumulated effect on each transform by $`1.28\\cdot10^{-8}`$ at $`h=1/768`$ for the parameters (5.6), where the Feller index is 5.1.\n\n### 6.2 The bias of the proxy value is first order\n\nThe certificate already contains the evidence. Its finite inversion sums for the proxy value under $`P`$ and under the algebraic recursion of $`Q`$ are enclosed to 24 digits at each of the three certified steps, and their differences are\n\n$$\n\\begin{gathered} -.000406686,\\qquad-.000202992,\\qquad-.000101409\\qquad\\text{at }h=\\tfrac1{384},\\ \\tfrac1{768},\\ \\tfrac1{1536}. \\end{gathered} \\tag{6.3}\n$$\n\nEach halving of the step halves the difference, with ratios 2.0035 and 2.0017. Table 8 extends the sequence with a binary64 implementation of the same inversion, which reproduces the certified sums to within $`2\\cdot10^{-13}`$.\n\n**Table 8. Euler bias of the proxy value by step.**\n\n| Steps per year | Euler minus continuous proxy value | Bias times steps |\n| --- | --- | --- |\n| 48 | $`-0.003335423`$ | $`-0.160100`$ |\n| 96 | $`-0.001643900`$ | $`-0.157814`$ |\n| 192 | $`-0.000816199`$ | $`-0.156710`$ |\n| 384 (certified sums) | $`-0.000406686`$ | $`-0.156167`$ |\n| 768 (certified sums) | $`-0.000202992`$ | $`-0.155898`$ |\n| 1536 (certified sums) | $`-0.000101409`$ | $`-0.155764`$ |\n| 3072 | $`-0.000050683`$ | $`-0.155697`$ |\n| 6144 | $`-0.000025336`$ | $`-0.155664`$ |\n\nThe last column converges, and the two finest rungs of the full ladder give\n\n$$\n\\begin{gathered} e^{-rT}\\bigl(E_Q\\mathcal L-E_P\\mathcal L\\bigr)\\approx-0.15563\\,h-0.205\\,h^2 . \\end{gathered} \\tag{6.4}\n$$\n\nThe expansion is read from the finite sums and the algebraic recursion. For the proxy values themselves the certificate proves $`e^{-rT}(E_Q\\mathcal{L}-E_P\\mathcal{L})\\in[-.000797230370,\\;.000391245751]`$ at $`h=1/768`$: the finite-sum difference with the complete error radius of Table 3.\n\nThe Euler gap of the option adds the change in the cost of the kink, $`(R_{K_1,Q}-R_{K_1,P})-(R_{K_2,Q}-R_{K_2,P})`$. A coupled simulation estimates it: $`2^{20}`$ paths of the scheme at several steps driven by the same Brownian increments, with the first-period factor integrated in closed form and the finest step, 3072 per year, standing in for $`P`$. At 768 steps the change is $`-0.3\\cdot10^{-7}`$ at $`K_1`$ and $`1.0\\cdot10^{-7}`$ at $`K_2`$, with standard errors $`0.9\\cdot10^{-7}`$ and $`0.6\\cdot10^{-7}`$, and $`-1.2\\cdot10^{-7}`$ for the spread with standard error $`1.2\\cdot10^{-7}`$. At 384 and 192 steps the spread values are $`-5.0\\cdot10^{-7}`$ and $`-9.3\\cdot10^{-7}`$. The kink contributes less than $`10^{-6}`$ at every step.\n\nPutting the two parts together, the price under the scheme at 768 steps is below the continuous price by about $`.000203`$: two hundredths of a cent on an underlying of 100 and 0.3 basis points of the option price. The proved bound of Corollary 5.2 is 54 times larger. Figure 4 puts the three quantities on one scale. The first-order behavior agrees with the theorem of Mickel and Neuenkirch [[28](#ref-28)], who prove weak order one for this scheme and smooth payoffs when the Feller index exceeds one. The conditional integration is what resolves quantities of this size: a direct coupled-path estimate of the gap has a standard error of $`.00047`$ on 32,768 paths at this step.\n\n\n\n### 6.3 Deterministic variance: a vega times a quadrature error\n\nWhen $`\\xi=0`$ the system (6.1) is linear, the variance is deterministic, and the leading Euler error can be written down for every bounded payoff. Assume\n\n$$\n\\begin{gathered} \\xi=0,\\quad\\kappa\\gt 0,\\quad \\bar v,v_0\\gt 0,\\quad t_j=j/12,\\quad h=1/(12m)\\le h_0\\lt 1/\\kappa,\\quad m\\in\\mathbb N, \\end{gathered} \\tag{6.5}\n$$\n\nand let $`D=v_0-\\bar v`$. The accumulated variances of the model and of the scheme are\n\n$$\n\\begin{gathered} I(t)=\\bar vt+\\frac D\\kappa(1-e^{-\\kappa t}),\\qquad I_h(t)=\\bar vt+\\frac D\\kappa[1-(1-\\kappa h)^{t/h}]. \\end{gathered} \\tag{6.6}\n$$\n\nThe twelve components of the log-increment vector $`Y`$ are independent with laws\n\n$$\n\\begin{gathered} N(r/12-s_j/2,s_j),\\qquad s_j=I(t_j)-I(t_{j-1}), \\end{gathered} \\tag{6.7}\n$$\n\nunder $`P`$, and with $`s_j`$ replaced by $`s_{h,j}`$ under $`Q`$; both variances are at least $`u_*=\\min(v_0,\\bar v)/12`$. Define\n\n$$\n\\begin{gathered} a(t)=\\frac{D\\kappa t e^{-\\kappa t}}2, \\qquad B_I(t)=|D|e^{-\\kappa t} \\left[ \\frac{\\kappa^2t}{3(1-\\kappa h_0)} +\\frac{\\kappa^3t^2}{8(1-\\kappa h_0)^2} \\right], \\end{gathered} \\tag{6.8}\n$$\n\n$`d_j=a(t_j)-a(t_{j-1})`$ and $`b_j=B_I(t_j)+B_I(t_{j-1})`$. Appendix D proves\n\n$$\n\\begin{gathered} |s_{h,j}-s_j-hd_j|\\le h^2b_j . \\end{gathered} \\tag{6.9}\n$$\n\nThe function $`ha(t)`$ is the leading error of the Euler variance accumulated to time $`t`$. It is the sum of two effects of opposite sign: the left-endpoint rule overstates the integral of a falling variance by $`hD(1-e^{-\\kappa t})/2`$, and the discrete decay $`(1-\\kappa h)^{t/h}\\lt e^{-\\kappa t}`$ understates it by $`hD(1-e^{-\\kappa t}-\\kappa te^{-\\kappa t})/2`$. The net error is a hump that peaks at $`t=1/\\kappa`$ and has the sign of $`v_0-\\bar v`$.\n\n**Theorem 6.2 (First-order expansion for bounded payoffs).** Under (6.5), let $`g(Y)`$ be a bounded Borel payoff with range in $`[l,l+L]`$, paid at time $`\\tau`$. Then\n\n$$\n\\begin{gathered} p_h(g)-p_c(g)=h\\beta_g+R_h(g), \\end{gathered} \\tag{6.10}\n$$\n\n$$\n\\begin{gathered} \\beta_g=e^{-r\\tau}E_s\\!\\left[g(Y)\\sum_jd_j\\mathcal S_j(Y)\\right],\\qquad |R_h(g)|\\le e^{-r\\tau}\\frac L2\\,C_2h^2, \\end{gathered} \\tag{6.11}\n$$\n\nwhere, with $`G_j`$ the standardized increments,\n\n$$\n\\mathcal S_j=\\frac{G_j^2-1}{2s_j}-\\frac{G_j}{2\\sqrt{s_j}}, \\qquad J(s)=\\frac1{2s^2}+\\frac1{4s},\\qquad H_2(s)=\\frac2{s^2}+\\frac1{2s},\n$$\n\n$$\n\\begin{gathered} C_2= \\sqrt{\\sum_jb_j^2J(u_*)} +\\frac12\\sum_j(|d_j|+h_0b_j)^2H_2(u_*). \\end{gathered} \\tag{6.12}\n$$\n\n*Proof.* The Gaussian density $`p_s`$ is twice differentiable in $`L^1`$ in its variance parameters, with\n\n$$\n\\|Dp_s[z]\\|_1\\le\\sqrt{\\sum_jz_j^2J(s_j)},\\qquad \\|D^2p_s[z,z]\\|_1\\le\\sum_jz_j^2H_2(s_j),\n$$\n\nthe first bound being the square root of the Fisher information form. Inserting (6.9) in the $`L^1`$ Taylor formula gives $`\\|p_{s_h}-p_s-hDp_s[d]\\|_1\\le C_2h^2`$. The density remainder integrates to zero, so the payoff may be centered as $`g-l-L/2`$, which gives (6.11). Appendix D contains the derivative calculations and the dominated convergence. $`\\square`$\n\nThe score $`\\mathcal{S}_j`$ is the derivative of the log density with respect to the variance of month $`j`$, so $`\\beta_g`$ is the sum over months of the payoff's sensitivity to that month's variance times the Euler error $`d_j`$ of that variance. For a put with strike $`K`$ and maturity $`\\tau`$ the sum collapses to\n\n$$\n\\begin{gathered} \\beta_{K,\\tau} =a(\\tau)\\,\\frac{S_0\\phi(d_1)}{2\\sqrt{I(\\tau)}}, \\qquad d_1=\\frac{\\log(S_0/K)+r\\tau+I(\\tau)/2}{\\sqrt{I(\\tau)}} , \\end{gathered} \\tag{6.13}\n$$\n\nthe Black–Scholes sensitivity to total variance multiplied by the variance error: a vega times a quadrature error. All nine calibration puts therefore have a first-order bias with the sign of $`v_0-\\bar v`$. For the Asian spread the coefficient combines the sensitivities to the twelve monthly variances with the twelve errors $`d_j`$, and its sign turns with the initial variance.\n\nOn the family\n\n$$\n\\begin{gathered} \\kappa=3,\\quad\\bar v=.045,\\quad v_0\\in[.03,.06],\\quad h_0=1/768, \\end{gathered} \\tag{6.14}\n$$\n\nthe remainder for the Asian spread is at most $`.000165217`$ uniformly, and a separate finite-step comparison of the two Gaussian laws gives $`|p_h(\\psi)-p_c(\\psi)|\\le.004990924`$. The score bound\n\n$$\n\\begin{gathered} |\\beta_{\\rm Asian}| \\le {15e^{-rT}\\over2}\\sqrt{\\sum_{j=1}^{12}d_j^2J(s_j)} \\end{gathered} \\tag{6.15}\n$$\n\nyields the enclosures of Table 9.\n\n**Table 9. Enclosures of the Asian leading coefficient and of the remainder.**\n\n| Initial variance | Enclosure of $`\\beta_{\\rm Asian}`$ | Upper bound on $`|e_h-h\\beta|`$, $`h=1/768`$ |\n| --- | --- | --- |\n| 0.03 | $`[-3.441491358,3.441491358]`$ | 0.000121667773 |\n| 0.04 | $`[-0.928189505,0.928189505]`$ | 0.000036646884 |\n| 0.05 | $`[-0.785853082,0.785853082]`$ | 0.000034532369 |\n| 0.06 | $`[-2.057270822,2.057270822]`$ | 0.000099684585 |\n\nA sampling estimate of the coefficient itself, with 2,097,152 Gaussian vectors per point, gives $`-0.041015`$, $`-0.011364`$, $`0.008329`$ and $`0.019917`$ in increasing order of initial variance, with standard errors $`0.002079`$, $`0.000566`$, $`0.000483`$ and $`0.001274`$. The coefficient changes sign between $`v_0=.04`$ and $`.05`$, as the vanishing of every $`d_j`$ at $`v_0=\\bar v`$ requires, and the Euler bias on this face is of order $`10^{-5}`$ at 768 steps. Table 10 reports the scaled residual $`(e_h-h\\beta_{\\rm Asian})/h^2`$, estimated with common samples across steps through a Gaussian density ratio.\n\n**Table 10. Sampling estimates of the scaled second-order residual.**\n\n| Initial variance | $`h=1/192`$ | $`1/384`$ | $`1/768`$ | $`1/1536`$ |\n| --- | --- | --- | --- | --- |\n| 0.03 | $`-0.063243`$ | $`-0.062944`$ | $`-0.062796`$ | $`-0.062722`$ |\n| 0.04 | $`-0.017226`$ | $`-0.017145`$ | $`-0.017105`$ | $`-0.017085`$ |\n| 0.05 | 0.011698 | 0.011646 | 0.011620 | 0.011607 |\n| 0.06 | 0.027912 | 0.027785 | 0.027722 | 0.027690 |\n\nAt $`h=1/768`$ the standard errors are $`.003733`$, $`.001000`$, $`.000863`$ and $`.002319`$, and estimates across steps at one point are correlated. The residual is stable in $`h`$ across the four steps, as a second-order expansion predicts; the constants of (6.14) cover $`h\\le1/768`$.\n\nThe two faces fit together. At the stochastic point (5.6), $`v_0=\\bar v`$ and the deterministic mechanism is switched off, so the whole bias $`-0.156h`$ comes from volatility of volatility and correlation. At the second point $`v_0=.04`$ the finite-sum difference is $`-.000221954`$, lower by $`.000019`$; the deterministic coefficient at $`v_0=.04`$ predicts $`-0.011364/768=-.000015`$.\n\n### 6.4 Coupling the two laws\n\nThe interval (5.17) allows for the four costs of the kink one at a time. With identical laws and exact inputs its width is\n\n$$\n\\begin{gathered} 2\\gamma\\left(K_1^{-1}+K_2^{-1}\\right)D_P(c), \\end{gathered} \\tag{6.16}\n$$\n\nabout $`.02`$ here whatever the step, and the widths of Table 5 at the two finer steps are that allowance plus the inversion tails. The Euler gap involves only the change in each cost between the two laws, and comparing the costs path by path keeps their common part out of the allowance. Appendix H proves the inequality under a coupling of the conditional triples $`(a,cg,\\sigma)`$ of the two laws, and evaluates it in closed form on the deterministic face at $`v_0=.04`$: the width of the allowance falls from $`.01538363715`$ to $`.00008815020`$, a factor of 175.\n\n## 7. After calibration\n\nA pricing engine is used twice: to fit parameters to liquid quotes and to value the exotic at the fitted parameters. An error in the engine moves the fitted parameters as well as the price. If the engine misprices the quotes and the exotic in the same direction, refitting absorbs part of the error; if the exotic loads on a different stretch of the variance curve than the quotes do, refitting can add to it. This section bounds the net effect on every quantile of the posterior distribution of the Asian price. Theorem 7.1 holds for any two pricing laws; the numbers are computed for a prior concentrated near deterministic variance and nine synthetic quotes.\n\n### 7.1 Price transfer for small volatility of volatility\n\nSuppose that the deterministic reference variance and its Euler values lie in $`[m,M]`$ with $`0\\lt m\\le M\\lt \\infty`$, that $`\\kappa\\ge\\kappa_-\\gt 0`$, and that\n\n$$\n\\begin{gathered} \\kappa h\\le1,\\qquad 0\\le\\xi\\le\\sqrt{\\kappa_-M/2}. \\end{gathered} \\tag{7.1}\n$$\n\nWith $`C_P=\\sqrt{M/(2\\kappa_-)}`$ and $`C_Q=\\sqrt{2M/\\kappa_-}`$, the Itô isometry for the model and the 1-Lipschitz property of the positive part for the scheme give\n\n$$\n\\begin{gathered} E(V_t^\\xi-v(t))^2\\le\\xi^2C_P^2,\\qquad E(V_j^{\\xi,h}-v_j)^2\\le\\xi^2C_Q^2, \\end{gathered} \\tag{7.2}\n$$\n\nand, with the lower bound $`m`$ and Doob's inequality,\n\n$$\n\\begin{gathered} E\\max|Z_\\ell^\\xi-Z_\\ell^0| \\le\\xi C_\\ell\\left(\\frac T2+2\\sqrt{T/m}\\right),\\qquad \\ell=P,Q. \\end{gathered} \\tag{7.3}\n$$\n\nA put is $`K`$-Lipschitz in the log price, and the spread is $`K_2`$-Lipschitz in the maximum norm of the fixing log prices. So (7.3) gives uniform radii from every calibration price and from the target price to their deterministic-variance counterparts. Appendix E has the proof. On the parameter box\n\n$$\n\\begin{gathered} \\kappa\\in[2.999,3.001],\\quad\\bar v\\in[.04499,.04501],\\\\ v_0\\in[.03001,.05999],\\quad \\xi\\in[10^{-5},4\\cdot10^{-5}], \\quad\\rho\\in[-.8,-.3], \\end{gathered} \\tag{7.4}\n$$\n\nthe resulting bound on the Euler gap is $`.020948`$ for the Asian spread and $`.016702`$ for each of the nine puts with maturities $`1/4,1/2,1`$ and strikes $`90,100,110`$.\n\n### 7.2 The posterior\n\nNumerical error acts on the likelihood and on its normalizing constant [[11](#ref-11)]. Fix $`\\kappa=3`$, $`\\bar v=.045`$, and let\n\n$$\n\\begin{gathered} u=v_0\\sim U[.03,.06],\\quad \\xi\\sim U[10^{-7},10^{-6}],\\quad \\rho\\sim U[-.8,-.3] \\end{gathered} \\tag{7.5}\n$$\n\nbe independent. Write $`\\zeta=(\\xi,\\rho)`$ with prior law $`\\nu`$; in this section and in Appendix F the index $`m\\in\\{P,Q\\}`$ denotes the pricing law. Let $`c_m(u,\\zeta)`$ be the vector of calibration prices and $`J_m(u,\\zeta)`$ the Asian price under law $`m`$. For synthetic quotes $`y`$ and a positive-definite $`\\Sigma_d`$ define\n\n$$\nw_m(u,\\zeta)= \\exp\\!\\left[-\\frac12(c_m-y)^T\\Sigma_d^{-1}(c_m-y)\\right], \\qquad \\bar w_m(u)=\\int w_m(u,\\zeta)\\,\\nu(d\\zeta),\n$$\n\n$$\n\\begin{gathered} Z_m=\\frac1{.03}\\int_{.03}^{.06}\\bar w_m(u)\\,du,\\qquad \\mu_m(du,d\\zeta)=\\frac{w_m(u,\\zeta)}{.03Z_m}\\,du\\,\\nu(d\\zeta). \\end{gathered} \\tag{7.6}\n$$\n\nEach law has its own calibration prices, its own normalizing constant and its own target price.\n\n**Theorem 7.1 (Transfer to every quantile).** Suppose the two marginal posteriors of $`u`$ admit a coupling $`(U_P,U_Q)`$ with $`|U_P-U_Q|\\le\\delta_u`$ almost surely, that the reference targets satisfy\n\n$$\n|J_P^0(u)-J_P^0(v)|\\le L_0|u-v|, \\qquad |J_Q^0(u)-J_P^0(u)|\\le\\varepsilon_0,\n$$\n\nand that $`|J_m(u,\\zeta)-J_m^0(u)|\\le e_{m,J}`$. Then the left generalized quantiles of the Asian price under the two posteriors satisfy\n\n$$\n\\begin{gathered} |q_Q(p)-q_P(p)| \\le \\varepsilon_0+e_{P,J}+e_{Q,J}+L_0\\delta_u \\qquad(0\\lt p\\lt 1). \\end{gathered} \\tag{7.7}\n$$\n\n*Proof.* Let $`K_m(u,d\\zeta)=w_m(u,\\zeta)\\nu(d\\zeta)/\\bar w_m(u)`$. Extend the coupling by drawing $`\\zeta_P`$ from $`K_P(U_P,\\cdot)`$ and $`\\zeta_Q`$ from $`K_Q(U_Q,\\cdot)`$; by Fubini's theorem the two marginals are $`\\mu_P`$ and $`\\mu_Q`$. The triangle inequality gives\n\n$$\n|J_Q(U_Q,\\zeta_Q)-J_P(U_P,\\zeta_P)| \\le\\varepsilon_0+e_{P,J}+e_{Q,J}+L_0\\delta_u=:E,\n$$\n\nhence $`F_P(z-E)\\le F_Q(z)\\le F_P(z+E)`$ for the two distribution functions, and (7.7) follows by taking left inverses. Atoms are allowed. $`\\square`$\n\nThe four terms have distinct meanings. $`\\varepsilon_0`$ is the Euler gap of the target at fixed parameters, $`e_{P,J}+e_{Q,J}`$ pays for replacing a small $`\\xi`$ by zero, and $`L_0\\delta_u`$ is the transport term: the sensitivity of the target to the initial variance times the distance the fitted variance moves when the engine changes.\n\n### 7.3 Numbers\n\nPartition $`[.03,.06]`$ into 4096 cells. On each cell and over the whole $`\\zeta`$-domain the likelihood is enclosed with (7.3) and Black prices at the reference variance; the enclosures survive integration against $`\\nu`$, which gives envelopes of the two marginal distribution functions with their normalizing mass, and an inverse-distribution coupling driven by one uniform variable gives $`\\delta_u`$ (Appendix F). The single-quote case uses the one-year put struck at 100, the nine-quote case the whole vector, both with $`\\Sigma_d=.75^2(.75I_d+.25\\mathbf{1}\\mathbf{1}^T)`$. Then\n\n$$\nL_0\\le213.792201829985,\\quad \\varepsilon_0\\le.004493504474,\\quad e_{P,J}+e_{Q,J}\\le.000393595478,\n$$\n\n$$\n\\begin{array}{c|c|c} \\text{Case}&\\delta_u&\\sup_{0\\lt p\\lt 1}|q_Q(p)-q_P(p)|\\\\ \\hline \\text{Single quote}&3/204800&\\le.008018821658\\\\ \\text{Nine quotes}&3/81920&\\le.012716404217 \\end{array} \\tag{7.8}\n$$\n\nIn volatility terms the fitted initial variance moves by at most $`1.5\\cdot10^{-5}`$ and $`3.7\\cdot10^{-5}`$, less than a hundredth of a volatility point. The transport term is 39 percent of the first bound and 62 percent of the second: with nine quotes the envelopes of the posterior are wider and the coupling takes more steps. Table 11 refines the cells.\n\n**Table 11. Bound on the displacement of every quantile of the Asian price.**\n\n| Cells | Single quote | Nine quotes | Coupling steps, single/nine |\n| --- | --- | --- | --- |\n| 4096 | 0.008018821658 | 0.012716404217 | 2 / 5 |\n| 8192 | 0.006452960804 | 0.010367612938 | 2 / 7 |\n| 16384 | 0.006061495591 | 0.009584682511 | 3 / 12 |\n\nAt 16384 cells the bounds are 24.4 and 24.6 percent smaller than at 4096, with more coupling steps at each refinement. For the continuous-model posterior at 4096 cells, the 2.5, 50 and 97.5 percent quantiles of the initial variance lie in\n\n$$\n\\begin{array}{lccc} \\text{one quote:}&[.0307690,.0307764]&[.0444067,.0444141]&[.0590917,.0590992]\\\\ \\text{nine quotes:}&[.0308203,.0308350]&[.0435131,.0435425]&[.0587182,.0587256] \\end{array}\n$$\n\n## 8. Two uses\n\n### 8.1 An acceptance test for a pricing engine\n\nThe benchmarks for Heston Asian options in the literature are Monte Carlo estimates with standard errors. Corollary 5.2 is a benchmark of a different kind. Any correct implementation of the continuous model for the contract (5.5)–(5.6) must return a value in $`[6.508371733,\\;6.518868974]`$, and any correct implementation of the scheme (5.2) with 768 steps a value whose expectation lies in $`[6.507844281,\\;6.519014106]`$. A library that prints 6.52 for either has an error that no sample size will explain. The interval is 16 basis points of the price wide, enough to catch a wrong day count, a misplaced correlation or a scheme that is not the one documented.\n\nThe gap interval also converts a statement about the scheme into one about the model. If a system encloses the price under the scheme in $`[q_L,q_U]`$, by simulation with a confidence interval or otherwise, then by Theorem 5.1\n\n$$\n\\begin{gathered} p_c\\in[q_L-u,\\;q_U-\\ell] \\end{gathered} \\tag{8.1}\n$$\n\nfor the endpoints $`[\\ell,u]`$ of (5.19).\n\n### 8.2 A numerical line in the valuation reserve\n\nSupervisory guidance on model risk asks that validation establish whether a model performs as expected and identify its limitations and errors [[5](#ref-5)]. European prudent valuation rules require a model risk adjustment built from a range of plausible valuations at 90 percent confidence, or from expert judgment where no range can be produced [[15](#ref-15)]. The numerical part of that uncertainty, the distance between the number a system prints and the price under the model it claims to implement, is usually handled by judgment. The results here replace judgment by an inequality in three situations.\n\nFor the monthly average options of Table 1 the closed-form proxy value is the Black price to within a hundredth of a cent per barrel, proved, and Table 2 puts the cost of the kink at the same order under stochastic volatility. No numerical reserve is needed for the kink.\n\nFor a long window the cubic law gives the size of the question before any computation: evaluate (1.2), multiply by the strike, and compare the result with the quote step. If it is smaller, the proxy value is the price. If it is not, as for the annual put of Table 1 or the spread of Section 5, the interval of Theorem 5.1 is the reserve, and its width, a few tenths of a cent to a cent per hundred of underlying, can be read against the bid–ask spread of the contract.\n\nFor a simulation engine the discretization bias is a systematic error: it has the same sign on every trade valued with the same scheme and does not diversify across a book. Section 6 gives its order and, for the proxy value, its coefficient; Corollary 5.2 gives a bound that holds at the step in use.\n\n## 9. Conclusion\n\nExercising an arithmetic Asian option on the geometric average costs the expected overshoot when the two averages disagree. Across strikes that cost adds up to a Bregman divergence between the averages, and at each strike it is at most the divergence divided by the volatility of the return that precedes the first fixing, times $`K/\\sqrt{2\\pi}`$. The first statement needs no model. The second holds for calls and puts under Black, Heston, rough volatility and jump dynamics alike, and it bounds the error of every classical conditioning lower bound. The ratio of dispersion to level risk decides everything that follows. For a monthly window it is so small that the closed-form proxy value is the price to a hundredth of a cent per barrel, with a bound that is nearly attained. For an annual window it is a cent, and a proof is worth the computation: under Heston the price of a twelve-fixing spread is enclosed in an interval 16 basis points wide, and the price under a 768-step Euler scheme within $`.011025`$ of it.\n\nThe Euler error is small and regular. The scheme is an Euler method for the Riccati equation, its bias in the proxy value is $`-0.156h`$, and the kink adds less than $`10^{-6}`$, because its costs under the two laws nearly cancel. A path-by-path comparison carries the cancellation into a proof: on the deterministic-variance face the coupled allowance is 175 times narrower than the separate one.\n\n## Appendices and Supporting Material\n\nThis page includes the complete main text and references. Appendices, supplementary proofs, code, and verification material are available in the [GitHub repository](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options) and the [full manuscript](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/manuscript/report.md#appendix-a-finite-certification-of-the-weighted-second-moment).\n\n## References\n\n\n\n1. Andersen, Leif. Simple and efficient simulation of the Heston stochastic volatility model. *The Journal of Computational Finance*, 11(3):1–42, 2008. [https://doi.org/10.21314/JCF.2008.189](https://doi.org/10.21314/JCF.2008.189).\n\n\n\n2. Bally, Vlad, and Talay, Denis. The law of the Euler scheme for stochastic differential equations: I. Convergence rate of the distribution function. *Probability Theory and Related Fields*, 104(1):43–60, 1996. [https://doi.org/10.1007/BF01303802](https://doi.org/10.1007/BF01303802).\n\n\n\n3. Black, Fischer. The pricing of commodity contracts. *Journal of Financial Economics*, 3(1–2):167–179, 1976. [https://doi.org/10.1016/0304-405X(76)90024-6](https://doi.org/10.1016/0304-405X(76)90024-6).\n\n\n\n4. Blas, Javier. Uncovering the Secret History of Wall Street's Largest Oil Trade. *Bloomberg Markets*, 4 April 2017.\n\n\n\n5. Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation, and Office of the Comptroller of the Currency. Revised Guidance on Model Risk Management. SR Letter 26-2, 17 April 2026. [https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm](https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm).\n\n\n\n6. Breeden, Douglas T., and Litzenberger, Robert H. Prices of State-Contingent Claims Implicit in Option Prices. *The Journal of Business*, 51(4):621–651, 1978. [https://doi.org/10.1086/296025](https://doi.org/10.1086/296025).\n\n\n\n7. Bregman, L. M. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. *USSR Computational Mathematics and Mathematical Physics*, 7(3):200–217, 1967. [https://doi.org/10.1016/0041-5553(67)90040-7](https://doi.org/10.1016/0041-5553(67)90040-7).\n\n\n\n8. Broadie, Mark, and Kaya, Özgür. Exact Simulation of Stochastic Volatility and Other Affine Jump Diffusion Processes. *Operations Research*, 54(2):217–231, 2006. [https://doi.org/10.1287/opre.1050.0247](https://doi.org/10.1287/opre.1050.0247).\n\n\n\n9. Carr, Peter, and Madan, Dilip. Towards a Theory of Volatility Trading. In R. Jarrow (ed.), *Volatility: New Estimation Techniques for Pricing Derivatives*, 417–427. Risk Books, London, 1998.\n\n\n\n10. CME Group. NYMEX Rulebook, Chapter 341: WTI Average Price Option. [https://www.cmegroup.com/rulebook/NYMEX/3/341.pdf](https://www.cmegroup.com/rulebook/NYMEX/3/341.pdf).\n\n\n\n11. Cotter, S. L., Dashti, M., and Stuart, A. M. Approximation of Bayesian Inverse Problems for PDEs. *SIAM Journal on Numerical Analysis*, 48(1):322–345, 2010. [https://doi.org/10.1137/090770734](https://doi.org/10.1137/090770734).\n\n\n\n12. Curran, Michael. Valuing Asian and Portfolio Options by Conditioning on the Geometric Mean Price. *Management Science*, 40(12):1705–1711, 1994. [https://doi.org/10.1287/mnsc.40.12.1705](https://doi.org/10.1287/mnsc.40.12.1705).\n\n\n\n13. Demeterfi, Kresimir, Derman, Emanuel, Kamal, Michael, and Zou, Joseph. A Guide to Volatility and Variance Swaps. *The Journal of Derivatives*, 6(4):9–32, 1999. [https://doi.org/10.3905/jod.1999.319129](https://doi.org/10.3905/jod.1999.319129).\n\n\n\n14. Ehm, Werner, Gneiting, Tilmann, Jordan, Alexander, and Krüger, Fabian. Of quantiles and expectiles: consistent scoring functions, Choquet representations and forecast rankings. *Journal of the Royal Statistical Society: Series B*, 78(3):505–562, 2016. [https://doi.org/10.1111/rssb.12154](https://doi.org/10.1111/rssb.12154).\n\n\n\n15. European Commission. Commission Delegated Regulation (EU) 2016/101 supplementing Regulation (EU) No 575/2013 with regard to regulatory technical standards for prudent valuation under Article 105(14). *Official Journal of the European Union*, L 21, 28 January 2016. [https://eur-lex.europa.eu/eli/reg_del/2016/101/oj](https://eur-lex.europa.eu/eli/reg_del/2016/101/oj).\n\n\n\n16. Fusai, Gianluca, and Kyriakou, Ioannis. General Optimized Lower and Upper Bounds for Discrete and Continuous Arithmetic Asian Options. *Mathematics of Operations Research*, 41(2):531–559, 2016. [https://doi.org/10.1287/moor.2015.0739](https://doi.org/10.1287/moor.2015.0739).\n\n\n\n17. Geman, Hélyette, and Yor, Marc. Bessel Processes, Asian Options, and Perpetuities. *Mathematical Finance*, 3(4):349–375, 1993. [https://doi.org/10.1111/j.1467-9965.1993.tb00092.x](https://doi.org/10.1111/j.1467-9965.1993.tb00092.x).\n\n\n\n18. Hairer, Ernst, Nørsett, Syvert P., and Wanner, Gerhard. *Solving Ordinary Differential Equations I: Nonstiff Problems*, second edition, Section II.8. Springer, Berlin, 1993. [https://doi.org/10.1007/978-3-540-78862-1](https://doi.org/10.1007/978-3-540-78862-1).\n\n\n\n19. Heston, Steven L. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options. *The Review of Financial Studies*, 6(2):327–343, 1993. [https://doi.org/10.1093/rfs/6.2.327](https://doi.org/10.1093/rfs/6.2.327).\n\n\n\n20. Johansson, Fredrik. Arb: Efficient Arbitrary-Precision Midpoint-Radius Interval Arithmetic. *IEEE Transactions on Computers*, 66(8):1281–1292, 2017. [https://doi.org/10.1109/TC.2017.2690633](https://doi.org/10.1109/TC.2017.2690633).\n\n\n\n21. Kemna, A. G. Z., and Vorst, A. C. F. A pricing method for options based on average asset values. *Journal of Banking and Finance*, 14(1):113–129, 1990. [https://doi.org/10.1016/0378-4266(90)90039-5](https://doi.org/10.1016/0378-4266(90)90039-5).\n\n\n\n22. Kim, Bara, Kim, Jeongsim, Kim, Jerim, and Wee, In Suk. A recursive method for discretely monitored geometric Asian option prices. *Bulletin of the Korean Mathematical Society*, 53(3):733–749, 2016. [https://doi.org/10.4134/BKMS.b150283](https://doi.org/10.4134/BKMS.b150283).\n\n\n\n23. Lee, Roger W. Option Pricing by Transform Methods: Extensions, Unification, and Error Control. *The Journal of Computational Finance*, 7(3):51–86, 2004. [https://doi.org/10.21314/JCF.2004.121](https://doi.org/10.21314/JCF.2004.121).\n\n\n\n24. Levy, Edmond. Pricing European average rate currency options. *Journal of International Money and Finance*, 11(5):474–491, 1992. [https://doi.org/10.1016/0261-5606(92)90013-N](https://doi.org/10.1016/0261-5606(92)90013-N).\n\n\n\n25. London Metal Exchange. Averaging solutions: a guide to trading and hedging average prices on the London Metal Exchange. [https://www.lme.com/-/media/Files/Trading/Contract-types/LME-Averaging-Solutions-Brochure.pdf](https://www.lme.com/-/media/Files/Trading/Contract-types/LME-Averaging-Solutions-Brochure.pdf).\n\n\n\n26. Lord, Roger. Partially exact and bounded approximations for arithmetic Asian options. *The Journal of Computational Finance*, 10(2):1–52, 2006. [https://www.risk.net/node/2160406](https://www.risk.net/node/2160406).\n\n\n\n27. Lord, Roger, Koekkoek, Remmert, and van Dijk, Dick. A comparison of biased simulation schemes for stochastic volatility models. *Quantitative Finance*, 10(2):177–194, 2010. [https://doi.org/10.1080/14697680802392496](https://doi.org/10.1080/14697680802392496).\n\n\n\n28. Mickel, Annalena, and Neuenkirch, Andreas. The weak convergence order of two Euler-type discretization schemes for the log-Heston model. *IMA Journal of Numerical Analysis*, 43(6):3326–3356, 2023. [https://doi.org/10.1093/imanum/drac069](https://doi.org/10.1093/imanum/drac069).\n\n\n\n29. Nielsen, J. Aase, and Sandmann, Klaus. Pricing Bounds on Asian Options. *Journal of Financial and Quantitative Analysis*, 38(2):449–473, 2003. [https://doi.org/10.2307/4126759](https://doi.org/10.2307/4126759).\n\n\n\n30. QuantLib contributors. AnalyticDiscreteGeometricAveragePriceAsianHestonEngine. Source code at a pinned commit. [https://github.com/lballabio/QuantLib/blob/966a4cc101049ca36a888b2ce223aa96d3f3b22d/ql/experimental/asian/analytic_discr_geom_av_price_heston.hpp](https://github.com/lballabio/QuantLib/blob/966a4cc101049ca36a888b2ce223aa96d3f3b22d/ql/experimental/asian/analytic_discr_geom_av_price_heston.hpp).\n\n\n\n31. Rogers, L. C. G., and Shi, Z. The value of an Asian option. *Journal of Applied Probability*, 32(4):1077–1088, 1995. [https://doi.org/10.2307/3215221](https://doi.org/10.2307/3215221).\n\n\n\n32. Talay, Denis, and Tubaro, Luciano. Expansion of the global error for numerical schemes solving stochastic differential equations. *Stochastic Analysis and Applications*, 8(4):483–509, 1990. [https://doi.org/10.1080/07362999008809220](https://doi.org/10.1080/07362999008809220).\n\n\n\n33. Thompson, G. W. P. Fast narrow bounds on the value of Asian options. Judge Institute of Management, University of Cambridge, WP09/2002, 2002. [https://www.jbs.cam.ac.uk/wp-content/uploads/2020/08/wp0209.pdf](https://www.jbs.cam.ac.uk/wp-content/uploads/2020/08/wp0209.pdf).\n\n\n\n34. Turnbull, Stuart M., and Wakeman, Lee Macdonald. A Quick Algorithm for Pricing European Average Options. *Journal of Financial and Quantitative Analysis*, 26(3):377–389, 1991. [https://doi.org/10.2307/2331213](https://doi.org/10.2307/2331213).\n\n\n\n35. Valencia, Fabián. To Hedge or to Self-Insure? The Benefits of Mexico's Oil Hedging Program. *IMF Blog*, 26 March 2018. [https://www.imf.org/external/np/blog/dialogo/032618.pdf](https://www.imf.org/external/np/blog/dialogo/032618.pdf).\n"
+ "sourceEditedAt": "2026-11-19T11:43:56-05:00",
+ "sourceCommit": "dd4b147d822d45a4ca12a66294fae2a32cd252bb",
+ "sourceBlob": "73fd58c109bca144a8809e3fc9832e8ad258cae6",
+ "markdown": "[English PDF](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/paper/paper.pdf) · [Repository](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/README.md) · [Code and evidence](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/docs/EVIDENCE.md)\n\n**Letao (Theodore) Ouyang**\n\n[theodore.oy2025@gmail.com](mailto:theodore.oy2025@gmail.com) · [10@alumni.duke.edu](mailto:10@alumni.duke.edu)\n\n2026\n\n**Keywords:** Asian options; average price options; geometric conditioning; Bregman divergence; stochastic volatility; Heston model; Euler discretization; validated numerics; model risk.\n\n**JEL classification:** G13; C63; C65.\n\n**MSC 2020:** 91G20; 91G60; 60H35; 65G30.\n\n## Abstract\n\nAn arithmetic Asian option is worth more than the same payoff when exercise is decided by a scaled geometric average of the fixings. We call the difference the cost of the kink and prove three things about it. It is the expected overshoot on the event that the two averages fall on opposite sides of the strike. Integrated across strikes with the weights that replicate a log contract, it equals a Bregman divergence between the two averages. Whenever the return before the first fixing has a conditionally Gaussian component with volatility $`\\sigma`$, it is at most\n\n$$\n\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/cG)}{\\sigma}\\right],\\qquad d(x)=x-1-\\log x.\n$$\n\nIn Black's model this bound is $`v^3\\tau^2/(360\\sqrt{2\\pi t_1})`$ per unit of strike to leading order, for volatility $`v`$, an averaging window $`\\tau`$ and a lead time $`t_1`$. For a one-month crude oil average price option it is below one hundredth of a cent per barrel and within a factor of 1.1 to 1.5 of the true cost; for an annual-average put on the scale of a sovereign oil hedge it is under one cent per barrel. For a twelve-fixing 95–110 call spread under Heston, validated transforms turn the bound into a proved price interval, $`[6.508371733,\\;6.518868974]`$, and a proved bound of 0.011025 on the distance to the price under a 768-step Euler scheme. The same computation shows that the Euler bias of the proxy value is first order in the step $`h`$, with leading coefficient $`-0.156`$.\n\n## 1. Introduction\n\nMost options written on oil, metals, freight and corporate currency flows pay on an average. A refiner buys crude every day of the month and an airline burns fuel every day of the quarter, so their physical contracts are priced on monthly averages and their hedges follow. The NYMEX average price option on crude oil settles on the mean of the daily futures settlements in a calendar month [[10](#ref-10)]. The London Metal Exchange lists options on the monthly average of its official settlement prices, the price at which physical metal changes hands [[25](#ref-25)]. Mexico's treasury insures its oil revenue with puts on the average price over a year [[35](#ref-35)]; the program for 2009 covered 330 million barrels at 70 dollars and paid out 5.085 billion dollars [[4](#ref-4)].\n\nThe payoff of such a contract depends on the arithmetic average $`A`$ of the fixings, whose law has no closed form in any standard model. The geometric average $`G`$ does, and practice rests on that fact. Kemna and Vorst use the geometric option as a control variate [[21](#ref-21)]. Curran, and Rogers and Shi, condition on $`G`$ and obtain a lower bound that is accurate to several digits in Black's model [[12](#ref-12), [31](#ref-31)]. Fusai and Kyriakou carry the same bound to stochastic volatility and jumps by transform methods [[16](#ref-16)]. Under stochastic volatility the working alternative is simulation on a time grid, which adds a discretization bias to the sampling error.\n\nTwo numbers in this workflow usually come without a proof. One is the distance between the conditioning lower bound and the price. The other is the distance between the price under the discretized dynamics and the price under the model that was calibrated. The classical estimate of the first is of first order in the conditional standard deviation of $`A`$ given $`G`$, which makes it loose exactly where the lower bound is good. The second is known to be of first order in the step for smooth payoffs [[28](#ref-28)], with an unknown constant. A valuation is certified when both distances are replaced by inequalities whose every term has been computed with a rigorous error. This paper does so, and it starts from a description of the first distance that makes its size predictable.\n\n### 1.1 The idea\n\nSuppose the holder of an arithmetic call must decide exercise by looking at a proxy $`B`$ instead of $`A`$: she receives $`A-K`$ when $`B\\gt K`$ and nothing otherwise. She can only lose by this, and she loses on two events. If the proxy is below the strike while $`A`$ is above, she forgoes $`A-K`$. If the proxy is above while $`A`$ is below, she pays $`K-A`$. In both cases the strike lies between the proxy and the truth, so the loss is at most $`|A-B|`$ and it occurs with a probability of order $`|A-B|`$ times the density of the proxy at the strike. The cost is of second order in the basis $`A-B`$. Figure 1 shows the two events for the contract of Section 5.\n\nThe second-order statement becomes an inequality when the density at the strike can be bounded, and here a structural fact helps. The return realized before the first fixing moves every fixing by the same factor. If part of that return is Gaussian given everything else, with volatility $`\\sigma`$, then conditionally the option is a Black–Scholes call on that factor, and the density at the strike is at most $`1/(\\sigma\\sqrt{2\\pi})`$ in log-strike. With $`B=cG`$ the basis depends only on the shape of the path, since $`A/G`$ is unchanged when all fixings are scaled together. The cost of the kink is therefore controlled by a ratio: the dispersion of the fixings about their own average, divided by the level risk that precedes them.\n\n\n\n### 1.2 Results\n\nSection 3 proves the general statements. Proposition 3.1 identifies the cost of the kink with the Bregman divergence generated by the call payoff and shows that its integral across strikes is exact: with weights $`dK`$ it equals $`E(A-B)^2/2`$, and with the weights $`dK/K^2`$ that replicate a log contract it equals $`E\\,d(A/B)`$, where $`d(x)=x-1-\\log x`$. Theorem 3.3 adds the common factor. For every strike,\n\n$$\n\\begin{gathered} 0\\le E(A-K)^+-E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\bigr] \\le\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\right], \\end{gathered} \\tag{1.1}\n$$\n\nand the same bound holds with $`E[(A-B)^2/\\sigma]/(2K\\sqrt{2\\pi})`$ on the right. Each bound is a total mass from Proposition 3.1 multiplied by the largest value a Gaussian density of width $`\\sigma`$ can take. A matching lower bound shows that (1.1) is attained for at-the-money options whose dispersion is small relative to $`\\sigma`$. For $`B=cG`$ the best constant is the mean of $`A/G`$ weighted by $`1/\\sigma`$, and the bound at that scale is $`K/\\sqrt{2\\pi}`$ times $`E[1/\\sigma]`$ times the mean log deviation of $`A/G`$ (Proposition 3.4). Because the Curran, Rogers–Shi and optimized Fusai–Kyriakou lower bounds dominate the proxy value, (1.1) also bounds their errors.\n\nSection 4 evaluates the bound where contracts trade. In Black's model with volatility $`v`$, a lead time $`t_1`$ before the first fixing and a window $`\\tau`$, the bound per unit of strike is\n\n$$\n\\begin{gathered} \\frac{v^3\\tau^2}{360\\sqrt{2\\pi t_1}} \\end{gathered} \\tag{1.2}\n$$\n\nto leading order under continuous monitoring, with an explicit correction for $`n`$ fixings (Proposition 4.1). For a calendar-month crude oil option one to twelve months ahead, ball arithmetic gives price intervals 0.0023 to 0.0081 cents per barrel wide, and simulation puts the true cost at 65 to 94 percent of the bound. The classical first-order bound for the same contracts is 2.6 cents. Under Heston dynamics with a volatility of volatility of 0.7 the cost is about 0.005 to 0.008 cents and the bound 0.006 to 0.013. For an annual-average put on 330 million barrels the bound is 0.91 cents per barrel in Black's model, three million dollars on a premium of 569 million.\n\nSection 5 turns the bound into a certificate for a contract on which it is not small: a 95–110 call spread on twelve monthly fixings under Heston with $`S_0=100`$. The proxy value is computed by validated transform inversion and the weighted second moment by validated quadrature, with projection, tails, periodization and rounding inside the budget. The continuous price lies in\n\n$$\n\\begin{gathered} [6.508371733,\\;6.518868974], \\end{gathered} \\tag{1.3}\n$$\n\nthe price under absorbed Euler with step $`h=1/768`$ lies in $`[6.507844281,\\;6.519014106]`$, and their difference in $`[-0.011024692273,\\;0.010642371599]`$.\n\nSection 6 explains the Euler error that these intervals enclose. One Euler step of the stochastic differential equation acts on exponential-affine functions as one explicit Euler step of the Riccati equation (Proposition 6.1). The finite sums already enclosed by the certificate then display a first-order bias in the proxy value: $`-0.000406686`$, $`-0.000202992`$ and $`-0.000101409`$ at 384, 768 and 1536 steps, with leading coefficient $`-0.156`$. Simulation puts the change in the cost of the kink between the two laws below $`10^{-6}`$. The Euler price at 768 steps is therefore about 0.0002 below the continuous price, well inside the proved bound of 0.011025. On the deterministic-variance face the leading error is a vega multiplied by the quadrature error of the Euler variance (Theorem 6.2), and a coupling of the two laws narrows the allowance for the cost of the kink by a factor of 175.\n\nSection 7 carries the fixed-parameter bounds through a calibration and bounds the displacement of every posterior quantile of the Asian price. Section 8 describes two uses: an acceptance test for pricing engines and a numerical component of a valuation reserve.\n\n### 1.3 Related work\n\nGeometric conditioning starts with Curran [[12](#ref-12)] and Rogers and Shi [[31](#ref-31)]. Rogers and Shi bound the error of their lower bound by half the expected conditional standard deviation of the average, an estimate of first order. Thompson [[33](#ref-33)], Nielsen and Sandmann [[29](#ref-29)] and Lord [[26](#ref-26)] sharpen it in the lognormal model; in Thompson's example the Rogers–Shi upper bound is 9.039 against a lower bound of 8.8275 and his own upper bound of 8.8333. Fusai and Kyriakou [[16](#ref-16)] optimize the exercise threshold, treat stochastic volatility and jump models through the joint transform of the price and its geometric average, and state a conditional-variance error bound of the Rogers–Shi type (their Theorem 4), which they evaluate numerically. Moment matching [[34](#ref-34), [24](#ref-24)] and the Laplace transform of Geman and Yor [[17](#ref-17)] are the other classical routes in the lognormal model.\n\nRelative to this literature, Theorem 3.3 is an a priori error bound of second order that holds beyond the lognormal model and needs only moments of the fixings weighted by $`1/\\sigma`$. It covers every lower bound of the conditioning family at once (Corollary 3.5), and the cubic law of Section 4 gives its size before any computation.\n\nThe decomposition of a convex payoff into calls is due to Breeden and Litzenberger [[6](#ref-6)] and Carr and Madan [[9](#ref-9)], and the weights $`dK/K^2`$ are those of the variance swap [[13](#ref-13)]. In forecast evaluation the same pathwise quantity is the elementary score of Ehm, Gneiting, Jordan and Krüger [[14](#ref-14)], who read it as the regret of an investor acting on a point forecast and prove the mixture representation of Bregman scores; Proposition 3.1 is that representation with exercise in the role of investment.\n\nHeston's affine structure supplies the continuous transforms [[19](#ref-19)], and recursive transforms for discrete geometric averages are available [[22](#ref-22), [30](#ref-30)]. The scheme studied here is the absorption fix in the classification of Lord, Koekkoek and van Dijk [[27](#ref-27)]; Broadie and Kaya [[8](#ref-8)] and Andersen [[1](#ref-1)] give exact and moment-matched alternatives. Mickel and Neuenkirch prove weak order one for the absorbed scheme and smooth payoffs when the Feller index exceeds one [[28](#ref-28)]; Talay and Tubaro [[32](#ref-32)] and Bally and Talay [[2](#ref-2)] are the classical references for weak expansions. Error control for transform inversion follows Lee [[23](#ref-23)], and all enclosures use the ball arithmetic of Arb [[20](#ref-20)].\n\n## 2. Average-price contracts and their two averages\n\n### 2.1 The contract\n\nFix dates $`0\\lt t_1\\lt \\cdots\\lt t_n=T`$ and let $`S_{t_1},\\ldots,S_{t_n}`$ be the fixings of a positive price. Put\n\n$$\n\\begin{gathered} A=\\frac1n\\sum_{i=1}^nS_{t_i},\\qquad G=\\Bigl(\\prod_{i=1}^nS_{t_i}\\Bigr)^{1/n},\\qquad X=\\frac AG . \\end{gathered} \\tag{2.1}\n$$\n\nAn average price call pays $`(A-K)^+`$ at $`T`$ and a put pays $`(K-A)^+`$. We call $`t_1`$ the lead time and $`\\tau=t_n-t_1`$ the window. For the exchange-traded contracts of the introduction the window is a calendar month of daily fixings and the lead time is whatever remains until that month begins. For Mexico's hedge the window is a year.\n\nTwo elementary reductions are used throughout. Put–call parity holds for the average, $`(A-K)^+-(K-A)^+=A-K`$, so a statement about calls is a statement about puts. And a fixing that has been observed is a constant: if $`m`$ fixings with mean $`\\bar a`$ are known, then\n\n$$\n\\begin{gathered} (A-K)^+=\\frac{n-m}{n}\\,\\bigl(A'-K'\\bigr)^+,\\qquad K'=\\frac{nK-m\\bar a}{n-m}, \\end{gathered} \\tag{2.2}\n$$\n\nwhere $`A'`$ is the mean of the remaining fixings. An option inside its window is an option on the fixings still to come with a shifted strike, or a forward if $`K'\\le0`$. The lead time is therefore always the time to the next unknown fixing.\n\n### 2.2 The gap between the averages is dispersion\n\nWrite $`z_i=\\log S_{t_i}`$, $`\\bar z`$ for their mean, $`u_i=z_i-\\bar z`$, and\n\n$$\ns^2=\\frac1n\\sum_{i=1}^nu_i^2,\\qquad m=\\max_i|u_i| .\n$$\n\nThen $`G=e^{\\bar z}`$ and $`X=n^{-1}\\sum_ie^{u_i}`$. Since the $`u_i`$ sum to zero and $`e^u-1-u`$ lies between $`u^2e^{-|u|}/2`$ and $`u^2e^{|u|}/2`$,\n\n$$\n\\begin{gathered} \\tfrac12s^2e^{-m}\\le X-1\\le\\tfrac12s^2e^{m}. \\end{gathered} \\tag{2.3}\n$$\n\nThe arithmetic average is the geometric average marked up by half the dispersion of the log fixings about their own mean. The markup does not change when every fixing is multiplied by the same number: $`X`$ is a function of the shape of the path only.\n\nIn financial terms the position long $`A`$ and short $`G`$ is a claim on the realized variance of the fixing series, with no exposure to the level at first order. In Black's model with volatility $`v`$ and continuous monitoring over a window $`\\tau`$, $`s^2`$ has mean $`v^2\\tau/6`$ and variance $`v^4\\tau^2/45`$ (Section 4), so $`X`$ is about $`1+v^2\\tau/12`$ and its randomness is of order $`v^2\\tau`$ times $`0.07`$. Under stochastic volatility the dispersion inherits the volatility risk, and with a negative correlation between price and variance it is larger on paths that end low. That asymmetry is visible in the numerical profile of Section 5.\n\n## 3. Exercising on a proxy\n\n### 3.1 The cost of the kink\n\nLet $`A`$ and $`B`$ be positive random variables with finite means; $`B`$ is the proxy. For $`K\\gt 0`$ put\n\n$$\n\\begin{gathered} C(K)=E(A-K)^+,\\qquad L_B(K)=E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\bigr],\\qquad R_B(K)=C(K)-L_B(K). \\end{gathered} \\tag{3.1}\n$$\n\n$`L_B(K)`$ is the value of the call when exercise is decided by $`B`$; we call it the proxy value and $`R_B(K)`$ the cost of the kink. For $`B=cG`$ we write $`L_c`$ and $`R_c`$. Prices are discounted at the end.\n\n**Proposition 3.1 (The cost of the kink).** For all positive numbers $`A,B,K`$,\n\n$$\n\\begin{gathered} (A-K)^+-(A-K)\\mathbf 1_{\\{B\\gt K\\}} =(A-K)\\mathbf 1_{\\{B\\le K\\lt A\\}}+(K-A)\\mathbf 1_{\\{A\\le K\\lt B\\}}\\ \\ge0 . \\end{gathered} \\tag{3.2}\n$$\n\nThe put payoff $`(K-A)^+`$ and its proxy version $`(K-A)\\mathbf{1}_{\\{B\\le K\\}}`$ differ by the same quantity. Consequently $`R_B(K)\\ge0`$ is the expected distance from $`A`$ to the strike on the event that the strike separates $`A`$ from $`B`$, for calls and puts alike. If $`\\varphi`$ is convex on $`(0,\\infty)`$ with a locally absolutely continuous derivative, then\n\n$$\n\\begin{gathered} \\int_0^\\infty R_B(K)\\,\\varphi''(K)\\,dK =E\\bigl[\\varphi(A)-\\varphi(B)-\\varphi'(B)(A-B)\\bigr], \\end{gathered} \\tag{3.3}\n$$\n\nboth sides being finite or infinite together. In particular\n\n$$\n\\begin{gathered} \\int_0^\\infty R_B(K)\\,dK=\\tfrac12E(A-B)^2,\\qquad \\int_0^\\infty R_B(K)\\,\\frac{dK}{K^2}=E\\,d(A/B),\\qquad d(x)=x-1-\\log x . \\end{gathered} \\tag{3.4}\n$$\n\n*Proof.* Identity (3.2) is a check of the four cases given by the positions of $`A`$ and $`B`$ relative to $`K`$; the put follows from parity applied to both payoffs. For (3.3), fix $`A`$ and $`B`$ and integrate the right side of (3.2) against $`\\varphi''`$. If $`B\\lt A`$, integration by parts gives\n\n$$\n\\int_B^A(A-K)\\varphi''(K)\\,dK=-(A-B)\\varphi'(B)+\\varphi(A)-\\varphi(B),\n$$\n\nand the case $`A\\lt B`$ gives the same expression. Taking expectations with Tonelli's theorem proves (3.3). The choices $`\\varphi(K)=K^2/2`$ and $`\\varphi(K)=-\\log K`$ give (3.4). $`\\square`$\n\nThe right side of (3.3) is the Bregman divergence of $`\\varphi`$ between $`A`$ and $`B`$ [[7](#ref-7)], and the integrand in (3.2) is that divergence for the call payoff itself. Formula (3.3) is the spanning of a convex payoff by calls [[6](#ref-6), [9](#ref-9)], applied to the pair of payoffs. The two cases in (3.4) are conservation laws: however the cost of the kink is spread across strikes, its total is a moment of the basis. The weights $`dK/K^2`$ are those that replicate a log contract, so a strip of proxy errors weighted like a variance swap sums to the expected log-gap $`E\\,d(A/B)`$. For $`B=cG`$ that gap involves only the shape variable $`X`$.\n\nProposition 3.1 needs no model. It also shows what a model must supply for a bound at a single strike: a limit on how much of the total mass can concentrate there.\n\n### 3.2 A common factor\n\n**Assumption (CF).** There are a $`\\sigma`$-field $`\\mathcal{G}`$, positive $`\\mathcal{G}`$-measurable random variables $`a_1,\\ldots,a_n`$ and $`\\sigma`$, and a random variable $`U`$ such that\n\n$$\n\\begin{gathered} U\\mid\\mathcal G\\sim N(0,\\sigma^2),\\qquad S_{t_i}=a_ie^{U}\\quad(i=1,\\ldots,n). \\end{gathered} \\tag{3.5}\n$$\n\nPut $`a=n^{-1}\\sum_ia_i`$ and $`g=(\\prod_ia_i)^{1/n}`$, so that $`A=ae^U`$, $`G=ge^U`$ and $`X=a/g`$. Given $`\\mathcal{G}`$ the shape of the path is known and its level is lognormal. The assumption holds in the following cases.\n\n1. Black's model with volatility $`v`$: $`U=vW_{t_1}`$, $`\\sigma=v\\sqrt{t_1}`$, and $`\\mathcal{G}`$ is generated by the Brownian increments after $`t_1`$.\n2. Stochastic volatility, $`dS_t/S_t=r\\,dt+\\sqrt{V_t}\\,(\\rho\\,dW_t+\\sqrt{1-\\rho^2}\\,dB_t)`$ with $`V`$ adapted to $`W`$ and $`B`$ independent of $`W`$. Heston, 3/2, Stein–Stein and rough volatility models are of this form.\n3. The same with jumps in price or variance that are independent of $`B`$: the jump path is placed in $`\\mathcal{G}`$.\n4. Time-stepping schemes with Gaussian price innovations, such as the Euler kernel of Section 5.\n\nIn the second case $`\\mathcal{G}`$ is generated by $`W`$ and by the increments of $`B`$ after $`t_1`$, and\n\n$$\nU=\\sqrt{1-\\rho^2}\\int_0^{t_1}\\sqrt{V_t}\\,dB_t,\\qquad \\sigma^2=(1-\\rho^2)\\int_0^{t_1}V_t\\,dt ;\n$$\n\nin the fourth the integrals are replaced by their sums over the grid. The assumption is a scaling property: the Gaussian part of the return before $`t_1`$ multiplies every fixing by the same factor and leaves the later volatility path as it was. Dynamics whose drift and volatility are free of the price level have it, and so does an average over several futures contracts that move in parallel.\n\nLet $`\\Phi`$ and $`\\phi`$ be the standard normal distribution function and density, $`\\phi_\\sigma(u)=\\phi(u/\\sigma)/\\sigma`$, and for $`x,s,K\\gt 0`$\n\n$$\n\\begin{gathered} F_K(x;s)=xe^{s^2/2}\\Phi\\!\\left(\\frac{\\log(x/K)}{s}+s\\right)-K\\Phi\\!\\left(\\frac{\\log(x/K)}{s}\\right), \\qquad \\Gamma_K(x;s)=\\frac{K}{sx^2}\\,\\phi\\!\\left(\\frac{\\log(x/K)}{s}\\right). \\end{gathered} \\tag{3.6}\n$$\n\n$`F_K(x;s)=E(xe^{sZ}-K)^+`$ for a standard normal $`Z`$ is the undiscounted Black–Scholes value of a call on a lognormal factor with total volatility $`s`$, and $`\\Gamma_K=\\partial_x^2F_K`$ is its gamma.\n\n**Proposition 3.2 (Conditional price and conditional cost).** Assume (CF) and let $`B=be^U`$ with $`b\\gt 0`$ $`\\mathcal{G}`$-measurable. Then\n\n$$\n\\begin{gathered} E\\bigl[(A-K)^+\\mid\\mathcal G\\bigr]=F_K(a;\\sigma),\\qquad E\\bigl[(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=F_K(b;\\sigma)+(a-b)\\,\\partial_xF_K(b;\\sigma), \\end{gathered} \\tag{3.7}\n$$\n\nand the conditional cost of the kink is\n\n$$\n\\begin{gathered} E\\bigl[(A-K)^+-(A-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=\\int_b^a(a-x)\\,\\Gamma_K(x;\\sigma)\\,dx . \\end{gathered} \\tag{3.8}\n$$\n\nIn the log-strike $`k=\\log K`$,\n\n$$\n\\begin{gathered} \\int_b^a(a-x)\\,\\Gamma_K(x;\\sigma)\\,dx=K\\,d(a/b)\\,\\bigl(\\pi_{a,b}*\\phi_\\sigma\\bigr)(k), \\end{gathered} \\tag{3.9}\n$$\n\nwhere $`\\pi_{a,b}`$ is the probability density proportional to $`|ae^{-t}-1|`$ on the interval between $`\\log b`$ and $`\\log a`$.\n\n*Proof.* The first identity in (3.7) is the definition of $`F_K`$. Completing the square in $`e^{sz}\\phi(z)`$ gives $`E[e^{sZ}\\mathbf{1}_{\\{Z\\gt q\\}}]=e^{s^2/2}\\Phi(s-q)`$, hence $`\\partial_xF_K(b;\\sigma)=e^{\\sigma^2/2}\\Phi(\\log(b/K)/\\sigma+\\sigma)`$ and\n\n$$\nE\\bigl[(A-B)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=(a-b)\\,\\partial_xF_K(b;\\sigma),\\qquad E\\bigl[(B-K)\\mathbf 1_{\\{B\\gt K\\}}\\mid\\mathcal G\\bigr]=F_K(b;\\sigma).\n$$\n\nAdding the two gives the second identity. Since $`\\sigma\\gt 0`$, $`F_K(\\cdot;\\sigma)`$ is twice continuously differentiable, and (3.8) is Taylor's formula with integral remainder between $`b`$ and $`a`$. For (3.9) substitute $`x=e^t`$: then $`\\Gamma_K(e^t;\\sigma)e^t=Ke^{-t}\\phi_\\sigma(k-t)`$ and $`(a-e^t)e^{-t}=ae^{-t}-1`$, whose integral between $`\\log b`$ and $`\\log a`$ is $`a/b-1-\\log(a/b)`$. $`\\square`$\n\nFormula (3.8) says that the cost of the kink is a gamma integrated over the basis. Formula (3.9) says where it sits: as a function of the log-strike, the conditional cost per unit of strike is the mass $`d(a/b)`$ spread by a Gaussian of standard deviation $`\\sigma`$ about a point between $`\\log b`$ and $`\\log a`$.\n\n### 3.3 The bound\n\n**Theorem 3.3 (The kink bound).** Assume (CF) and let $`B=be^U`$ with $`b\\gt 0`$ $`\\mathcal{G}`$-measurable. For every $`K\\gt 0`$,\n\n$$\n\\begin{gathered} 0\\le R_B(K)\\le\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\right], \\end{gathered} \\tag{3.10}\n$$\n\n$$\n\\begin{gathered} 0\\le R_B(K)\\le\\frac{1}{2K\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{(A-B)^2}{\\sigma}\\right]. \\end{gathered} \\tag{3.11}\n$$\n\nIn the other direction, with $`m_K=\\max\\{|\\log(K/a)|,|\\log(K/b)|\\}`$,\n\n$$\n\\begin{gathered} R_B(K)\\ge\\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac{d(A/B)}{\\sigma}\\,e^{-m_K^2/(2\\sigma^2)}\\right]. \\end{gathered} \\tag{3.12}\n$$\n\n*Proof.* In (3.9) the convolution is an average of values of $`\\phi_\\sigma`$, each at most $`1/(\\sigma\\sqrt{2\\pi})`$; since $`A/B=a/b`$, taking expectations gives (3.10). The points averaged lie within $`m_K`$ of $`k`$, where $`\\phi_\\sigma\\ge e^{-m_K^2/(2\\sigma^2)}/(\\sigma\\sqrt{2\\pi})`$; this gives (3.12). For (3.11) write $`u=\\log(K/x)`$. Then\n\n$$\n\\Gamma_K(x;\\sigma)=\\frac{e^{2u}}{K}\\,\\phi_\\sigma(u)=\\frac{e^{2\\sigma^2}}{K}\\,\\phi_\\sigma(u-2\\sigma^2)\\le\\frac{e^{2\\sigma^2}}{K\\sigma\\sqrt{2\\pi}} ,\n$$\n\nso the integral in (3.8) is at most $`e^{2\\sigma^2}(a-b)^2/(2K\\sigma\\sqrt{2\\pi})`$, and $`E[(A-B)^2\\mid \\mathcal{G}]=e^{2\\sigma^2}(a-b)^2`$. $`\\square`$\n\nBoth upper bounds have the same form. By (3.4) the cost integrates across strikes to $`E\\,d(A/B)`$ with weights $`dK/K^2`$ and to $`E(A-B)^2/2`$ with weights $`dK`$. Assumption (CF) limits how concentrated in strike the cost can be: no more than a Gaussian of width $`\\sigma`$ in log-strike. Each bound is one of the two masses multiplied by the peak of that Gaussian. The first is dimensionless and uniform in moneyness after division by $`K`$; the second weights the basis by the price level and tends to be the smaller of the two for strikes above the forward. The lower bound (3.12) shows that nothing is lost in (3.10) when the strike is within a fraction of $`\\sigma`$ of both averages on the paths that carry the dispersion.\n\n### 3.4 The best scale\n\n**Proposition 3.4 (Scale).** Assume (CF), $`E[1/\\sigma]\\lt \\infty`$ and $`E[X/\\sigma]\\lt \\infty`$. Let $`E_w`$ denote expectation with weights proportional to $`1/\\sigma`$. Over constants $`c\\gt 0`$, the right side of (3.10) with $`B=cG`$ is smallest at\n\n$$\n\\begin{gathered} c_*=E_w[X]=\\frac{E[X/\\sigma]}{E[1/\\sigma]},\\qquad\\text{where it equals}\\qquad \\frac{K}{\\sqrt{2\\pi}}\\,E\\!\\left[\\frac1\\sigma\\right]\\bigl(\\log E_wX-E_w\\log X\\bigr). \\end{gathered} \\tag{3.13}\n$$\n\nThe right side of (3.11) is smallest at\n\n$$\n\\begin{gathered} c_\\dagger=\\frac{E[AG/\\sigma]}{E[G^2/\\sigma]} . \\end{gathered} \\tag{3.14}\n$$\n\n*Proof.* $`E[d(X/c)/\\sigma]=E[1/\\sigma]\\,(E_wX/c-1-E_w\\log X+\\log c)`$, which is convex in $`1/c`$ with its minimum at $`c=E_wX`$. The second claim is a least-squares projection. $`\\square`$\n\nThe bracket in (3.13) is the mean log deviation of $`X`$, Theil's second measure of dispersion: the gap in Jensen's inequality for the logarithm. It vanishes only if $`X`$ is constant, and for small dispersion it is half the squared coefficient of variation of $`X`$. The optimal scale has a plain meaning: $`c_*-1`$ is the expected markup of the arithmetic over the geometric average, about $`v^2\\tau/12`$ in Black's model. The Heston certificate of Section 5 uses a rational scale near $`c_\\dagger`$, $`1254433/1250000=1.0035464`$; the Black value $`1+v^2\\tau(n+1)/(12n)`$ for its contract is $`1.0037`$.\n\n**Corollary 3.5 (Classical lower bounds).** Under (CF), for every $`c\\gt 0`$,\n\n$$\n\\begin{gathered} 0\\le C(K)-E\\bigl(E[A\\mid G]-K\\bigr)^+\\le C(K)-\\sup_{\\lambda\\gt 0}E\\bigl[(A-K)\\mathbf 1_{\\{G\\gt \\lambda\\}}\\bigr]\\le R_c(K), \\end{gathered} \\tag{3.15}\n$$\n\nso the bounds of Theorem 3.3 with $`B=cG`$ apply to the lower bound of Curran and of Rogers and Shi and to the optimized lower bound of Fusai and Kyriakou.\n\n*Proof.* $`E(E[A\\mid G]-K)^+`$ is the largest value of $`E[(A-K)\\mathbf{1}_H]`$ over events $`H`$ determined by $`G`$, and $`\\{G\\gt \\lambda\\}`$ is such an event; the choice $`\\lambda=K/c`$ gives $`L_c(K)`$. $`\\square`$\n\n### 3.5 Reading the bound\n\nThree remarks connect the inequalities to trading intuition.\n\nThe cost is a gamma. By (3.8), $`R_B(K)`$ is close to $`E[\\Gamma_K(b;\\sigma)(a-b)^2]/2`$ when the basis is small: half the gamma of the conditional option times the squared basis. Figure 3 shows this second-order term next to the exact cost for the contract of Section 5; at the two strikes of that contract they agree to within three percent.\n\nThe two risks enter as a ratio. $`\\sigma`$ is level risk realized before averaging begins and $`d(A/cG)`$ is dispersion risk realized inside the window. A long lead time and a short window make the proxy nearly exact; an average that starts tomorrow leaves one day of smoothing. Only the part of the early return that is independent of volatility counts, which is why $`\\sigma^2`$ carries the factor $`1-\\rho^2`$: with $`\\rho=-0.7`$, a typical equity value, the bound is 1.4 times what it would be at $`\\rho=0`$.\n\nThe bound is attained when dispersion meets the strike. It charges every path as if it sat at the strike, and (3.12) measures what that charge gives up: the factor $`e^{-m_K^2/(2\\sigma^2)}`$ on paths that have moved away. For short windows with long leads the factor is close to one, and bound and cost agree to within about ten percent (Table 1). For long windows a path with large dispersion has usually moved far, and the factor accounts for the difference between the two.\n\n## 4. How large is the cost? A cubic law\n\n### 4.1 Black's model\n\nIn Black's model [[3](#ref-3)] the fixings are $`S_{t_i}=F_i\\exp(-v^2t_i/2+vW_{t_i})`$ with forwards $`F_i`$ and volatility $`v`$, the common factor has the deterministic volatility $`\\sigma=v\\sqrt{t_1}`$, and everything in Theorem 3.3 is a Gaussian moment.\n\n**Proposition 4.1 (Black's model).** Let $`y_i=\\log S_{t_i}-\\log G`$. The vector $`y`$ is Gaussian with means $`m_i`$ and covariances $`c_{ij}`$ given by\n\n$$\n\\begin{aligned} m_i&=\\log F_i-\\tfrac12v^2t_i-\\frac1n\\sum_j\\bigl(\\log F_j-\\tfrac12v^2t_j\\bigr),\\\\ c_{ij}&=v^2\\Bigl[t_i\\wedge t_j-\\frac1n\\sum_k\\bigl(t_i\\wedge t_k+t_j\\wedge t_k\\bigr)+\\frac1{n^2}\\sum_{k,l}t_k\\wedge t_l\\Bigr], \\end{aligned}\n$$\n\nand $`c_{ij}`$ depends on the dates only through the differences $`t_i-t_1`$. Then\n\n$$\n\\begin{gathered} E\\,X=\\frac1n\\sum_ie^{m_i+c_{ii}/2},\\qquad E\\,X^2=\\frac1{n^2}\\sum_{i,j}e^{m_i+m_j+(c_{ii}+c_{jj})/2+c_{ij}}, \\end{gathered} \\tag{4.1}\n$$\n\n$`c_*=E\\,X`$, and for every strike\n\n$$\n\\begin{gathered} R_{c_*}(K)\\le\\frac{K}{v\\sqrt{2\\pi t_1}}\\Bigl[\\log E\\,X-(E\\,X-1)+\\tfrac12E(X-1)^2\\Bigr]. \\end{gathered} \\tag{4.2}\n$$\n\nFor $`n`$ equally spaced fixings over a window $`\\tau`$ with a flat forward curve, as $`v^2\\tau\\to0`$,\n\n$$\n\\begin{gathered} E\\,X=1+\\frac{v^2\\tau}{2}\\,\\mu_n+O(v^4\\tau^2),\\qquad \\mathop{\\mathrm{Var}}X=\\frac{v^4\\tau^2}{4}\\,q_n+O(v^6\\tau^3), \\end{gathered} \\tag{4.3}\n$$\n\n$$\n\\mu_n=\\frac{n+1}{6n},\\qquad q_n=\\frac{(n+1)(2n^2+7)}{90\\,n^2(n-1)},\n$$\n\nand the right side of (4.2) equals\n\n$$\n\\begin{gathered} K\\,\\frac{v^3\\tau^2}{8\\sqrt{2\\pi t_1}}\\,q_n\\,\\bigl(1+O(v^2\\tau)\\bigr) \\;\\longrightarrow\\; K\\,\\frac{v^3\\tau^2}{360\\sqrt{2\\pi t_1}}\\quad(n\\to\\infty). \\end{gathered} \\tag{4.4}\n$$\n\n*Proof.* The moments of $`y`$ are read from the Brownian covariance $`v^2(t_i\\wedge t_j)`$, and (4.1) is the Gaussian exponential moment. Since $`\\sigma`$ is constant, (3.13) gives $`c_*=E\\,X`$. On $`x\\ge1`$, $`\\log x\\ge(x-1)-(x-1)^2/2`$; applying this to $`X\\ge1`$ in $`E\\,d(X/c_*)=\\log E\\,X-E\\log X`$ and inserting the result in (3.10) gives (4.2). The expansions (4.3) and the constants $`\\mu_n`$ and $`q_n`$ are computed in Appendix I. $`\\square`$\n\nFormula (4.4) is a cubic law. The bound grows with the cube of volatility and the square of the window, and it falls with the square root of the lead time. The window enters squared because the dispersion surprise is of order $`v^2\\tau`$ and the cost is quadratic in it; one power of $`v`$ is returned by the smoothing $`1/(v\\sqrt{t_1})`$. Since the law of $`X`$ does not involve $`t_1`$, the dependence on the lead time is exactly $`t_1^{-1/2}`$ at every volatility. The numbers $`1/6`$ and $`1/45`$ behind $`\\mu_n`$ and $`q_n`$ are the mean and variance of $`\\int_0^1(W_t-\\bar W)^2dt`$ for a Brownian motion $`W`$ with time average $`\\bar W`$. Figure 2 plots the exact right side of (4.2) against the leading term.\n\n\n\n### 4.2 Contracts that trade\n\nTable 1 applies Proposition 4.1 to contracts with the terms of traded ones. The proxy value has the closed form of the conditioning literature: $`\\log G`$ is normal and each fixing is lognormal jointly with it, so $`L_c(K)`$ is a sum of $`n+1`$ normal distribution values. Every enclosure in the table is computed in ball arithmetic at 256 bits from exact rational inputs, so that under Black's model with the stated inputs each price lies between the proxy value and the proxy value plus the bound. The inputs are round numbers of the right order for each market. A year has 252 trading days.\n\n**Table 1. Proved price intervals in Black's model.**\n\n| Contract | Proxy value | Bound | Quote steps | Simulated | First order |\n| --- | --- | --- | --- | --- | --- |\n| Crude oil call, 1 month ahead | 3.206408 | 0.0000805 | 0.0081 | 0.0000523 | 0.0259 |\n| Crude oil call, 3 months ahead | 5.058865 | 0.0000462 | 0.0047 | 0.0000388 | 0.0262 |\n| Crude oil call, 6 months ahead | 6.906596 | 0.0000324 | 0.0033 | 0.0000293 | 0.0261 |\n| Crude oil call, 12 months ahead | 9.431726 | 0.0000224 | 0.0023 | 0.0000210 | 0.0256 |\n| Crude oil put at 60, 3 months ahead | 1.315689 | 0.0000396 | 0.0040 | 0.0000278 | 0.0262 |\n| Crude oil call at 80, 3 months ahead | 1.850312 | 0.0000417 | 0.0042 | 0.0000333 | 0.0262 |\n| Copper call, 3 months ahead | 432.887072 | 0.00170 | 0.17 | 0.00142 | 1.473 |\n| EUR/USD call, quarterly average (in pips) | 198.100004 | 0.000767 | 0.0008 | 0.000499 | 0.628 |\n| Annual-average oil put at 70 | 1.723884 | 0.009084 | 0.91 | 0.004235 | 0.5206 |\n| Twelve monthly fixings, call at 95 | 8.216008 | 0.003573 | 0.36 | 0.000750 | 0.1334 |\n| Twelve monthly fixings, call at 110 | 1.994944 | 0.003566 | 0.36 | 0.000762 | 0.1334 |\n| Daily fixings from tomorrow | 4.434265 | 0.013895 | 1.39 | 0.000679 | 0.1053 |\n\nThe crude oil rows are options on a calendar month of 21 daily settlements of a futures price of 70 dollars with 35 percent volatility, struck at 70 unless stated; the copper row has 22 settlements of a 9,500-dollar price with 22 percent volatility; the currency row has 63 fixings of a rate of 1.10 with 8 percent volatility, struck at the money, three months ahead. The annual put has 252 fixings of a 100-dollar price with 40 percent volatility starting four months ahead. The twelve-fixing rows are the contract of Section 5 with its variance frozen at 0.045. The last row has 252 daily fixings of a price of 100 with 20 percent volatility. Interest is 4 percent, except 1 percent in the twelve-fixing rows. The column headed Bound is the proved bound on the cost of the kink, the smaller of (3.10) and (3.11), each at its best scale, rounded up; the proxy value is rounded to the digits shown, and each price lies between the proxy value and the proxy value plus the bound. A quote step is one cent per barrel, one cent per tonne, one pip, or 0.01. The last two columns are computed in binary64 arithmetic: a Monte Carlo estimate of the true cost on $`2^{20}`$ paths, and $`E[\\mathrm{sd}(A\\mid G)]/2`$, the Rogers–Shi form of the classical first-order bound.\n\nFour observations follow.\n\nFor the monthly contracts the proxy value is the price for every commercial purpose. The exchange quotes these options in dollars and cents per barrel; the proved bound is between a 124th and a 446th of a cent. A market maker who exercises on the geometric average gives up less than ten dollars on a hundred contracts of a thousand barrels, whose premium exceeds three hundred thousand.\n\nThe bound is nearly attained. Its ratio to the simulated cost is 1.54, 1.19, 1.10 and 1.07 at one, three, six and twelve months, as the factor in (3.12) predicts: with a long lead the strike is within a small multiple of $`\\sigma`$ of both averages on almost every path. The classical first-order estimate is 2.6 cents at every lead time, between 320 and 1,140 times larger.\n\nScale changes the verdict. The annual put has the terms made public for Mexico's 2009 program [[4](#ref-4)], a one-year average struck at 70 dollars on 330 million barrels, valued here at 40 percent volatility. The bound is 0.91 cents per barrel. On the full volume that is 3.0 million dollars against a proxy value of 569 million, and the simulated cost is 1.4 million. A proved interval of half a percent is a statement worth having on a trade of this size, and the cubic law says what moves it: at 50 percent volatility the bound would be about twice as large.\n\nThe lead time is the resource. The last row is the same kind of contract as the first with the window lengthened to a year and the lead cut to a day. Theorem 3.3 draws on the one day of level risk before the first fixing, and the bound is 1.4 quote steps; the later fixings smooth the kink further, and the cost is a fifteenth of a quote step. For such contracts the theorem stands beside the lognormal bounds of Thompson [[33](#ref-33)] and Lord [[26](#ref-26)], which use the full covariance of the fixings.\n\n### 4.3 Stochastic volatility\n\nTheorem 3.3 holds for every model that satisfies (CF), and under stochastic volatility the dispersion acquires a second source of risk. Table 2 repeats five rows of Table 1 under Heston dynamics, $`dV_t=\\kappa(\\bar v-V_t)\\,dt+\\xi\\sqrt{V_t}\\,dW_t`$, with $`\\kappa=2`$, $`\\rho=-0.3`$ and $`V_0=\\bar v`$ equal to the squared volatility of Table 1; $`\\xi=0.7`$ for the monthly options and $`0.8`$ for the annual put, which puts the Feller index $`2\\kappa\\bar v/\\xi^2`$ at one. The entries are Monte Carlo estimates in which the first-period factor is integrated in closed form on each path (Proposition 3.2), so that the cost and the bound are sample means of smooth functions.\n\n**Table 2. Simulated cost of the kink under Heston dynamics.**\n\n| Contract | Value | Cost of the kink | Bound | Bound over cost | $`E[1/\\sigma]`$ |\n| --- | --- | --- | --- | --- | --- |\n| Crude oil call, 1 month ahead | 3.147 | 0.0000776 | 0.000132 | 1.70 | 10.78 |\n| Crude oil call, 3 months ahead | 4.875 | 0.0000730 | 0.0000942 | 1.29 | 6.56 |\n| Crude oil call, 6 months ahead | 6.561 | 0.0000624 | 0.0000738 | 1.18 | 4.80 |\n| Crude oil call, 12 months ahead | 8.891 | 0.0000496 | 0.0000576 | 1.16 | 3.41 |\n| Annual-average oil put at 70 | 2.053 | 0.00965 | 0.0296 | 3.06 | 5.05 |\n\nEach row uses $`2^{19}`$ paths ($`2^{18}`$ for the annual put) of absorbed Euler with four steps per trading day (two for the annual put). Standard errors are below 0.8 percent of each cost and bound, except 1.2 and 1.6 percent for the cost and bound of the annual put, and about 0.003 for the option values (0.009 for the annual put). The bound is the smaller of (3.10) at $`c_*`$ and (3.11) at $`c_\\dagger`$.\n\nVolatility of volatility roughly doubles the cost of the kink at three months and beyond, and the bound stays within a factor of 1.3 of it. For the monthly options the cost stays below a hundredth of a cent. For the annual put it is 0.96 cents per barrel and the bound 2.96 cents: 3.2 and 9.8 million dollars on 330 million barrels, against a value of 677 million. A long window is where a model with random volatility makes exercising on the geometric average expensive enough to measure, and where the interval of the next section earns its place.\n\n## 5. A proved price interval under Heston\n\nThis section turns Theorem 3.3 into a certificate for a contract whose cost of the kink is not negligible, under the continuous Heston model and under the time-stepping scheme used to simulate it.\n\n### 5.1 Model, scheme and contract\n\nLet $`W,B`$ be independent standard Brownian motions and $`\\theta=(\\kappa,\\bar v,\\xi,\\rho,v_0)`$ with $`S_0,\\kappa,\\bar v,v_0\\gt 0`$, $`\\xi\\ge0`$, $`|\\rho|\\lt 1`$ and $`r\\in \\mathbb{R}`$. The continuous model $`P`$ is\n\n$$\n\\begin{aligned} dV_t&=\\kappa(\\bar v-V_t)\\,dt+\\xi\\sqrt{V_t}\\,dW_t,\\\\ dZ_t&=(r-V_t/2)\\,dt+ \\sqrt{V_t}\\bigl(\\rho\\,dW_t+\\sqrt{1-\\rho^2}\\,dB_t\\bigr),\\\\ S_t&=S_0e^{Z_t},\\qquad Z_0=0,\\quad V_0=v_0. \\end{aligned} \\tag{5.1}\n$$\n\nOn a grid of step $`h`$, with every $`t_i/h`$ an integer, the discrete law $`Q_h`$ is\n\n$$\n\\begin{aligned} V_{j+1}^h&=\\left[(1-\\kappa h)V_j^h+\\kappa\\bar v h+ \\xi\\sqrt{hV_j^h}\\,G_j\\right]^+,\\\\ Z_{j+1}^h&=Z_j^h+(r-V_j^h/2)h+ \\sqrt{hV_j^h}\\bigl(\\rho G_j+\\sqrt{1-\\rho^2}H_j\\bigr), \\end{aligned} \\tag{5.2}\n$$\n\nwith independent standard normal $`G_j,H_j`$ and the same initial values. This is Euler with absorption in the classification of [[27](#ref-27)]: a negative variance proposal is set to zero, and the price update uses the variance at the start of the step.\n\nThe contract is a call spread on the average of (2.1),\n\n$$\n\\begin{gathered} \\psi(A)=(A-K_1)^+-(A-K_2)^+,\\qquad 0\\lt K_1\\lt K_2 . \\end{gathered} \\tag{5.3}\n$$\n\nA spread keeps the payoff bounded and is the form in which capped average-price protection is sold. Write\n\n$$\n\\begin{gathered} p_c=e^{-rT}E_P\\psi(A),\\qquad p_h=e^{-rT}E_{Q_h}\\psi(A),\\qquad e_h=p_h-p_c , \\end{gathered} \\tag{5.4}\n$$\n\nand $`Q=Q_h`$ when the step is understood. The numerical contract has twelve monthly fixings,\n\n$$\n\\begin{gathered} n=12,\\quad t_i=i/12,\\quad T=1,\\quad S_0=100,\\quad r=.01,\\quad K_1=95,\\quad K_2=110, \\end{gathered} \\tag{5.5}\n$$\n\nand the parameter point and scale are\n\n$$\n\\begin{gathered} \\theta_*=(3,\\;9/200,\\;23/100,\\;-11/20,\\;9/200),\\qquad c=\\frac{1254433}{1250000},\\qquad h=\\frac1{768}. \\end{gathered} \\tag{5.6}\n$$\n\nThe long-run volatility is 21.2 percent, the Feller index $`2\\kappa\\bar v/\\xi^2`$ is $`2700/529`$, about 5.1, and averaging begins after one month and lasts eleven. By the cubic law this is a hard case for the bound: the Black row of Table 1 gives 0.36 cents on a price of 100.\n\n### 5.2 The common factor and the boundary\n\nCondition on the whole variance driver $`W`$ and on the increments $`B_t-B_{t_1}`$, $`t\\ge t_1`$. With\n\n$$\nI_P=\\int_0^{t_1}V_t\\,dt,\\qquad I_Q=h\\sum_{j=0}^{t_1/h-1}V_j^h ,\n$$\n\nassumption (CF) holds under both laws:\n\n$$\n\\begin{gathered} U\\mid\\mathcal G\\sim N\\bigl(0,(1-\\rho^2)I_\\nu\\bigr),\\qquad S_{t_i}=e^Ua_i,\\qquad \\nu\\in\\{P,Q\\}. \\end{gathered} \\tag{5.7}\n$$\n\nUnder $`P`$, $`v_0\\gt 0`$ and path continuity give $`I_P\\gt 0`$ almost surely. Under $`Q`$,\n\n$$\n\\begin{gathered} I_Q\\ge hv_0\\gt 0 . \\end{gathered} \\tag{5.8}\n$$\n\nThe scheme has an atom at zero variance; already at the first step\n\n$$\n\\begin{gathered} Q_h(V_1^h=0)= \\Phi\\!\\left( -\\frac{(1-\\kappa h)v_0+\\kappa\\bar v h}{\\xi\\sqrt{hv_0}} \\right)\\in(0,1). \\end{gathered} \\tag{5.9}\n$$\n\nThe smoothing survives the atom: the common factor needs a positive integrated variance over the first month, and (5.8) supplies it on every path. Appendix B accounts for the effect of the atom on the transforms of the scheme.\n\n### 5.3 The proxy value and the second moment\n\nExercising both legs of the spread on $`cG`$ gives the payoff\n\n$$\n\\begin{gathered} \\mathcal L=(A-K_1)\\mathbf 1_{\\{cG\\gt K_1\\}}-(A-K_2)\\mathbf 1_{\\{cG\\gt K_2\\}} =\\psi(cG)+(A-cG)\\mathbf 1_{\\{K_1\\lt cG\\le K_2\\}} . \\end{gathered} \\tag{5.10}\n$$\n\nWith $`R_{K,\\nu}`$ the cost of the kink $`R_c(K)`$ under $`\\nu`$, Proposition 3.1 gives\n\n$$\n\\begin{gathered} E_\\nu\\psi(A)=E_\\nu\\mathcal L+R_{K_1,\\nu}-R_{K_2,\\nu},\\qquad R_{K,\\nu}\\ge0 . \\end{gathered} \\tag{5.11}\n$$\n\nThe proxy value is linear in the two measures $`\\nu(Y\\in dy)`$ and $`E_\\nu[A;Y\\in dy]`$, where $`Y=\\log(G/S_0)`$. With $`y_K=\\log(K/(cS_0))`$, $`F_{0,\\nu}(y)=\\nu(Y\\le y)`$ and $`F_{A,\\nu}(y)=E_\\nu[A\\mathbf{1}_{\\{Y\\le y\\}}]`$,\n\n$$\n\\begin{gathered} E_\\nu\\mathcal L=(K_2-K_1) +K_1F_{0,\\nu}(y_{K_1})-K_2F_{0,\\nu}(y_{K_2}) -F_{A,\\nu}(y_{K_1})+F_{A,\\nu}(y_{K_2}). \\end{gathered} \\tag{5.12}\n$$\n\nBoth distribution functions are inverted from the transforms\n\n$$\n\\begin{gathered} \\widehat\\mu_0(z)=E_\\nu e^{zY},\\qquad \\widehat\\mu_A(z)=\\frac{S_0}{n}\\sum_{i=1}^n E_\\nu e^{Z_{t_i}+zY}, \\end{gathered} \\tag{5.13}\n$$\n\nwhich are exponential-affine in the fixing-date log prices: one loading for the law of the geometric average and one for each term of the arithmetic average, thirteen when $`n=12`$. Under $`P`$ they solve Riccati equations; under $`Q`$ they are computed by an algebraic recursion plus a correction for the positive part (Section 6 and Appendix B).\n\nThe certificate uses the second bound of Theorem 3.3. By (5.7) its moment is\n\n$$\n\\begin{gathered} D_\\nu(c)=E_\\nu\\!\\left[\\frac{(A-cG)^2}{\\sigma}\\right]=\\frac{W_\\nu}{\\sqrt{1-\\rho^2}},\\qquad W_\\nu=E_\\nu\\!\\left[(A-cG)^2I_\\nu^{-1/2}\\right], \\end{gathered} \\tag{5.14}\n$$\n\nand the square expands into products of fixings,\n\n$$\n\\begin{gathered} (A-cG)^2=\\frac1{n^2}\\sum_{i,j}S_{t_i}S_{t_j} -\\frac{2c}{n}\\sum_iS_{t_i}G+c^2G^2 , \\end{gathered} \\tag{5.15}\n$$\n\nwhich for $`n=12`$ are 78 distinct pairs, twelve cross terms and one square: 91 real loadings. The weight $`I^{-1/2}`$ is a Laplace integral of $`e^{-x^2I}`$, which adds a killing rate in the first month. Appendix A encloses $`W_\\nu`$ by validated quadrature of these 91 transforms. The quadratic form is preferred to (3.10) here because its integrand is a finite sum of exponential-affine terms, and because with $`\\rho\\lt 0`$ it is the smaller bound at both strikes.\n\n### 5.4 The certificate\n\n**Theorem 5.1 (Price and gap intervals).** Let $`\\nu\\in\\{P,Q\\}`$ satisfy (5.7) and suppose that\n\n$$\ne^{-rT}E_\\nu\\mathcal L\\in[\\ell_\\nu,u_\\nu],\\qquad D_\\nu(c)\\le d_\\nu .\n$$\n\nWith $`\\gamma=e^{-rT}/(2\\sqrt{2\\pi})`$, the price under $`\\nu`$ satisfies\n\n$$\n\\begin{gathered} e^{-rT}E_\\nu\\psi(A)\\in\\left[\\ell_\\nu-\\frac{\\gamma d_\\nu}{K_2},\\;u_\\nu+\\frac{\\gamma d_\\nu}{K_1}\\right]. \\end{gathered} \\tag{5.16}\n$$\n\nIf moreover $`e^{-rT}(E_Q\\mathcal{L}-E_P\\mathcal{L})\\in[\\ell,u]`$, then\n\n$$\n\\begin{gathered} e_h\\in \\left[ \\ell-\\gamma\\!\\left(\\frac{d_Q}{K_2}+\\frac{d_P}{K_1}\\right), \\quad u+\\gamma\\!\\left(\\frac{d_Q}{K_1}+\\frac{d_P}{K_2}\\right) \\right]. \\end{gathered} \\tag{5.17}\n$$\n\n*Proof.* By (3.11), $`0\\le e^{-rT}R_{K,\\nu}\\le\\gamma d_\\nu/K`$. Insert this in (5.11) for (5.16), and subtract (5.11) for the two laws for (5.17): the lower end collects $`-R_{K_2,Q}-R_{K_1,P}`$ and the upper end $`R_{K_1,Q}+R_{K_2,P}`$. $`\\square`$\n\n**Corollary 5.2 (The numerical certificate).** For the model (5.1), the scheme (5.2) and the data (5.5)–(5.6),\n\n$$\n\\begin{gathered} W_P\\le2.172471592802,\\qquad W_Q\\le2.242189951600, \\end{gathered} \\tag{5.18}\n$$\n\n$$\n\\begin{gathered} e_h\\in[-.011024692273,\\;.010642371599],\\qquad |e_h|\\lt .011025, \\end{gathered} \\tag{5.19}\n$$\n\n$$\n\\begin{gathered} p_c\\in[6.508371733,\\;6.518868974],\\qquad p_h\\in[6.507844281,\\;6.519014106]. \\end{gathered} \\tag{5.20}\n$$\n\n*Proof.* Appendices A–C specify the validated computation. The finite inversion sums for the proxy value are $`6.513251661289`$ under $`P`$ and $`6.513048668980`$ under $`Q`$, each enclosed to 24 digits, and their difference is $`-.000202992309`$. The complete error radius of the difference, covering projection, both frequency tails, periodization and rounding, is at most $`.000594238060`$. Combining these with (5.18) in (5.16) and (5.17) with exact rational endpoints gives the statement. Appendix G describes the programs, the exact endpoints, and the second implementation that recomputes the whole catalog. $`\\square`$\n\nTable 3 lists the budget behind (5.19).\n\n**Table 3. Error budget of the Heston certificate at 768 steps.**\n\n| Quantity | Enclosure or upper bound |\n| --- | --- |\n| Cost of the kink, two laws and two strikes | $`[-.010227461904,.010251125849]`$ |\n| Projection error per transform under $`Q`$ | $`.000000012711979137`$ |\n| Projection contribution to the proxy value | $`.000004310413`$ |\n| Frequency tail, continuous model | $`.000014095712`$ |\n| Frequency tail, discrete model | $`.000184373301`$ |\n| Periodization, one model | $`.000195729318`$ |\n| Total error radius of the proxy-value difference | $`.000594238060`$ |\n| **Absolute bound on the Euler gap** | $`.\\mathbf{0}\\mathbf{1}\\mathbf{1}\\mathbf{0}\\mathbf{2}\\mathbf{5}`$ |\n\nReported widths are computed from exact rational endpoints and rounded outward, so differences of displayed endpoints can disagree with them in the last place.\n\n### 5.5 What the interval contains\n\nThe interval for $`p_c`$ is $`.0105`$ wide, 16 basis points of the price. It is built around the proxy value, known to $`\\pm.00021`$, and almost all of its width is the allowance for the two costs of the kink, $`.005407`$ above and $`.004670`$ below.\n\nA simulation locates the price inside the interval. On $`2^{20}`$ paths with the first-period factor integrated in closed form, the discounted cost of the kink is $`.000813`$ at $`K_1`$ and $`.000711`$ at $`K_2`$, with standard errors below $`4\\cdot10^{-6}`$. The two enter the spread with opposite signs, so the spread exceeds its proxy value by $`.000102`$ with standard error $`.000004`$, and the continuous price is close to $`6.51335`$. The certificate bounds each cost on its own and from one side, so its allowances add. They are about 6.6 times the simulated costs, in line with the Black rows of Table 1, where the ratio is 4.7 before volatility of volatility is added. The lower bound (3.12), evaluated on the same paths, is $`.000777`$ and $`.000682`$, within five percent of the two costs. The distance between the upper bound and the cost is therefore the Gaussian factor in (3.12): on most paths that carry dispersion the strike is several $`\\sigma`$ away from both averages.\n\nFigure 3 shows the cost across strikes. Three features agree with Section 3. The area under the curve, with the weights of (3.4), is $`8.6063\\cdot10^{-6}`$ by numerical integration and $`E\\,d(A/cG)=8.6063\\cdot10^{-6}`$ by direct simulation; with weights $`dK`$ the two sides are $`.0651615`$ and $`.0651612`$. The second-order term $`E[\\Gamma_K(cg;\\sigma)(a-cg)^2]/2`$ is within three percent of the cost at both strikes of the spread. And the profile leans toward low strikes: per unit of strike the cost peaks near $`K=77`$, at 2.2 times its value at $`K=100`$. With $`\\rho\\lt 0`$, paths that end low have had high variance, and high variance is what separates the two averages.\n\n\n\n### 5.6 Scale, steps and a second point\n\nBy Proposition 3.4, applied to the two laws together as they enter (5.17), the exact quadratic allowance is smallest at\n\n$$\n\\begin{gathered} c_\\dagger=1+\\frac{E_P^++E_Q^+}{H_P+H_Q},\\qquad H_\\nu=E_\\nu[G^2I_\\nu^{-1/2}],\\qquad E_\\nu^+=E_\\nu[(A-G)GI_\\nu^{-1/2}]\\ge0, \\end{gathered} \\tag{5.21}\n$$\n\nwhere $`E_\\nu^+\\ge0`$ by the inequality of arithmetic and geometric means. Both moments belong to the catalog of 91 loadings, and Appendix A encloses them with the same quadrature, which gives\n\n$$\n\\begin{gathered} c_\\dagger\\in[1.003427394603,\\;1.003669831097]. \\end{gathered} \\tag{5.22}\n$$\n\nTable 4 compares three scales. The scale of (5.6) is inside the enclosure and the optimum is flat; the choice $`c=1`$, the unscaled geometric average, would double the allowance.\n\n**Table 4. Effect of the scale on the allowance for the cost of the kink.**\n\n| Scale $`c`$ | Width of the allowance at $`h=1/768`$ |\n| --- | --- |\n| 1 | .040551110717 |\n| 1.0035464 (used) | .020478587752 |\n| 1.0035486 (rounded midpoint of (5.22)) | .020478592980 |\n\nTable 5 repeats the certificate at other steps, and Table 6 separates the widths.\n\n**Table 5. Proved intervals for the Euler gap at three steps.**\n\n| Step | Interval for $`e_h`$ | Width |\n| --- | --- | --- |\n| $`1/384`$ | $`[-0.014169989568,\\;0.013839530031]`$ | 0.028009519599 |\n| $`1/768`$ | $`[-0.011024692273,\\;0.010642371599]`$ | 0.021667063872 |\n| $`1/1536`$ | $`[-0.010967952032,\\;0.010766337995]`$ | 0.021734290027 |\n\n**Table 6. Width contributions at the three certified steps.**\n\n| Contribution | $`1/384`$ | $`1/768`$ | $`1/1536`$ |\n| --- | --- | --- | --- |\n| Cost of the kink | 0.026754977094 | 0.020478587752 | 0.020171631177 |\n| Projection | 0.000265787793 | 0.000008620825 | 0.000000265628 |\n| Frequency tail under $`P`$ | 0.000028191423 | 0.000028191423 | 0.000028191423 |\n| Frequency tail under $`Q`$ | 0.000177646019 | 0.000368746602 | 0.000751284529 |\n| Periodization | 0.000782917272 | 0.000782917272 | 0.000782917272 |\n| Rounding and endpoints | $`10^{-12}`$ | $`10^{-12}`$ | $`10^{-12}`$ |\n\nAt the three steps the cost of the kink is 95.5, 94.5 and 92.8 percent of the width. The width settles near $`.0217`$. Between 768 and 1536 steps the allowance for the cost of the kink changes by 1.5 percent, and the inversion terms make up the rest; among them the tail bound for the discrete transforms rests on the floor $`I_Q\\ge hv_0`$ of (5.8) and roughly doubles when the step is halved. The projection argument of Appendix B, whose exponent must exceed 8, holds at each of the three steps. What the Euler error itself does under refinement is the subject of Section 6.\n\nThe certificate also holds at a second parameter point, $`v_0=.04`$ with everything else unchanged and $`h=1/768`$:\n\n**Table 7. The certificate at initial variance .04.**\n\n| Quantity | Enclosure or upper bound |\n| --- | --- |\n| Continuous price | $`[6.499318797594,\\;6.509948941256]`$ |\n| Price under the scheme | $`[6.498150434416,\\;6.510702536853]`$ |\n| Euler gap | $`[-0.011798506840,\\;0.011383739259]`$ |\n| Width of the gap interval | 0.023182246099 |\n\nThe gap interval is 7.0 percent wider than at $`v_0=.045`$, and 83 percent of the increase comes from the discrete frequency tail, which rises from $`.000368746602`$ to $`.001629880669`$: a smaller $`v_0`$ lowers the floor in (5.8) and raises the first-month Laplace bound in (C.7). Theorem 5.1 itself holds for any spread, fixing schedule and parameter point; each certificate is one run of the validated computation that supplies its inputs.\n\n## 6. What discretization does to the price\n\nCorollary 5.2 bounds the Euler gap by $`.011025`$. This section shows that the gap itself is about fifty times smaller, derives its order and sign from the structure of the scheme, and shows how a coupling of the two laws brings the proved allowance down by two orders of magnitude.\n\n### 6.1 Euler on the price is Euler on the Riccati equation\n\nWrite $`\\alpha=\\xi^2/2`$ and $`d=\\kappa\\bar v`$. For a complex stock loading $`q`$, the transforms $`E_P\\exp(qZ_t+bV_t)=\\exp(\\mathcal{A}(t)+\\mathcal{B}(t)v_0)`$ of the continuous model are governed by\n\n$$\n\\begin{gathered} \\mathcal B'=\\alpha\\mathcal B^2-(\\kappa-\\rho\\xi q)\\mathcal B+\\tfrac12(q^2-q),\\qquad \\mathcal A'=rq+d\\mathcal B . \\end{gathered} \\tag{6.1}\n$$\n\n**Proposition 6.1 (The scheme is affine up to its projection).** Let $`v\\ge0`$, let $`\\widetilde V=(1-\\kappa h)v+\\kappa\\bar vh+\\xi\\sqrt{hv}\\,G`$ be the variance proposal of (5.2) from the state $`V_j^h=v`$, and let $`\\Delta Z`$ be the log-price increment of the same step. For complex $`q`$ and $`b`$,\n\n$$\n\\begin{gathered} E\\exp\\bigl(q\\,\\Delta Z+b\\,\\widetilde V\\bigr) =\\exp\\Bigl\\{h(rq+db)+v\\bigl[b+h\\bigl(\\alpha b^2-(\\kappa-\\rho\\xi q)b+\\tfrac12(q^2-q)\\bigr)\\bigr]\\Bigr\\}. \\end{gathered} \\tag{6.2}\n$$\n\nOne step of the scheme therefore acts on exponential-affine functions as one explicit Euler step of size $`h`$ of the system (6.1), followed by the replacement of $`\\widetilde V`$ by $`\\widetilde V^+`$.\n\n*Proof.* The exponent $`q\\,\\Delta Z+b\\widetilde V`$ is Gaussian given $`v`$, with mean $`q(r-v/2)h+b(1-\\kappa h)v+bdh`$ and variance $`hv[(q\\rho+b\\xi)^2+q^2(1-\\rho^2)]`$. Adding half the variance to the mean gives (6.2). $`\\square`$\n\nTwo consequences follow. First, without the positive part the transforms of the scheme would be the explicit Euler iterates of a polynomial ordinary differential equation, and the global error of that method has an asymptotic expansion in powers of $`h`$ [[18](#ref-18)]: for each loading the Euler transform differs from the continuous one by $`b_1h+b_2h^2+\\cdots`$ with computable coefficients. Second, the positive part acts only when the proposal is negative, an event of probability (5.9) at the first step. Appendix B bounds its accumulated effect on each transform by $`1.28\\cdot10^{-8}`$ at $`h=1/768`$ for the parameters (5.6), where the Feller index is 5.1.\n\n### 6.2 The bias of the proxy value is first order\n\nThe certificate already contains the evidence. Its finite inversion sums for the proxy value under $`P`$ and under the algebraic recursion of $`Q`$ are enclosed to 24 digits at each of the three certified steps, and their differences are\n\n$$\n\\begin{gathered} -.000406686,\\qquad-.000202992,\\qquad-.000101409\\qquad\\text{at }h=\\tfrac1{384},\\ \\tfrac1{768},\\ \\tfrac1{1536}. \\end{gathered} \\tag{6.3}\n$$\n\nEach halving of the step halves the difference, with ratios 2.0035 and 2.0017. Table 8 extends the sequence with a binary64 implementation of the same inversion, which reproduces the certified sums to within $`2\\cdot10^{-13}`$.\n\n**Table 8. Euler bias of the proxy value by step.**\n\n| Steps per year | Euler minus continuous proxy value | Bias times steps |\n| --- | --- | --- |\n| 48 | $`-0.003335423`$ | $`-0.160100`$ |\n| 96 | $`-0.001643900`$ | $`-0.157814`$ |\n| 192 | $`-0.000816199`$ | $`-0.156710`$ |\n| 384 (certified sums) | $`-0.000406686`$ | $`-0.156167`$ |\n| 768 (certified sums) | $`-0.000202992`$ | $`-0.155898`$ |\n| 1536 (certified sums) | $`-0.000101409`$ | $`-0.155764`$ |\n| 3072 | $`-0.000050683`$ | $`-0.155697`$ |\n| 6144 | $`-0.000025336`$ | $`-0.155664`$ |\n\nThe last column converges, and the two finest rungs of the full ladder give\n\n$$\n\\begin{gathered} e^{-rT}\\bigl(E_Q\\mathcal L-E_P\\mathcal L\\bigr)\\approx-0.15563\\,h-0.205\\,h^2 . \\end{gathered} \\tag{6.4}\n$$\n\nThe expansion is read from the finite sums and the algebraic recursion. For the proxy values themselves the certificate proves $`e^{-rT}(E_Q\\mathcal{L}-E_P\\mathcal{L})\\in[-.000797230370,\\;.000391245751]`$ at $`h=1/768`$: the finite-sum difference with the complete error radius of Table 3.\n\nThe Euler gap of the option adds the change in the cost of the kink, $`(R_{K_1,Q}-R_{K_1,P})-(R_{K_2,Q}-R_{K_2,P})`$. A coupled simulation estimates it: $`2^{20}`$ paths of the scheme at several steps driven by the same Brownian increments, with the first-period factor integrated in closed form and the finest step, 3072 per year, standing in for $`P`$. At 768 steps the change is $`-0.3\\cdot10^{-7}`$ at $`K_1`$ and $`1.0\\cdot10^{-7}`$ at $`K_2`$, with standard errors $`0.9\\cdot10^{-7}`$ and $`0.6\\cdot10^{-7}`$, and $`-1.2\\cdot10^{-7}`$ for the spread with standard error $`1.2\\cdot10^{-7}`$. At 384 and 192 steps the spread values are $`-5.0\\cdot10^{-7}`$ and $`-9.3\\cdot10^{-7}`$. The kink contributes less than $`10^{-6}`$ at every step.\n\nPutting the two parts together, the price under the scheme at 768 steps is below the continuous price by about $`.000203`$: two hundredths of a cent on an underlying of 100 and 0.3 basis points of the option price. The proved bound of Corollary 5.2 is 54 times larger. Figure 4 puts the three quantities on one scale. The first-order behavior agrees with the theorem of Mickel and Neuenkirch [[28](#ref-28)], who prove weak order one for this scheme and smooth payoffs when the Feller index exceeds one. The conditional integration is what resolves quantities of this size: a direct coupled-path estimate of the gap has a standard error of $`.00047`$ on 32,768 paths at this step.\n\n\n\n### 6.3 Deterministic variance: a vega times a quadrature error\n\nWhen $`\\xi=0`$ the system (6.1) is linear, the variance is deterministic, and the leading Euler error can be written down for every bounded payoff. Assume\n\n$$\n\\begin{gathered} \\xi=0,\\quad\\kappa\\gt 0,\\quad \\bar v,v_0\\gt 0,\\quad t_j=j/12,\\quad h=1/(12m)\\le h_0\\lt 1/\\kappa,\\quad m\\in\\mathbb N, \\end{gathered} \\tag{6.5}\n$$\n\nand let $`D=v_0-\\bar v`$. The accumulated variances of the model and of the scheme are\n\n$$\n\\begin{gathered} I(t)=\\bar vt+\\frac D\\kappa(1-e^{-\\kappa t}),\\qquad I_h(t)=\\bar vt+\\frac D\\kappa[1-(1-\\kappa h)^{t/h}]. \\end{gathered} \\tag{6.6}\n$$\n\nThe twelve components of the log-increment vector $`Y`$ are independent with laws\n\n$$\n\\begin{gathered} N(r/12-s_j/2,s_j),\\qquad s_j=I(t_j)-I(t_{j-1}), \\end{gathered} \\tag{6.7}\n$$\n\nunder $`P`$, and with $`s_j`$ replaced by $`s_{h,j}`$ under $`Q`$; both variances are at least $`u_*=\\min(v_0,\\bar v)/12`$. Define\n\n$$\n\\begin{gathered} a(t)=\\frac{D\\kappa t e^{-\\kappa t}}2, \\qquad B_I(t)=|D|e^{-\\kappa t} \\left[ \\frac{\\kappa^2t}{3(1-\\kappa h_0)} +\\frac{\\kappa^3t^2}{8(1-\\kappa h_0)^2} \\right], \\end{gathered} \\tag{6.8}\n$$\n\n$`d_j=a(t_j)-a(t_{j-1})`$ and $`b_j=B_I(t_j)+B_I(t_{j-1})`$. Appendix D proves\n\n$$\n\\begin{gathered} |s_{h,j}-s_j-hd_j|\\le h^2b_j . \\end{gathered} \\tag{6.9}\n$$\n\nThe function $`ha(t)`$ is the leading error of the Euler variance accumulated to time $`t`$. It is the sum of two effects of opposite sign: the left-endpoint rule overstates the integral of a falling variance by $`hD(1-e^{-\\kappa t})/2`$, and the discrete decay $`(1-\\kappa h)^{t/h}\\lt e^{-\\kappa t}`$ understates it by $`hD(1-e^{-\\kappa t}-\\kappa te^{-\\kappa t})/2`$. The net error is a hump that peaks at $`t=1/\\kappa`$ and has the sign of $`v_0-\\bar v`$.\n\n**Theorem 6.2 (First-order expansion for bounded payoffs).** Under (6.5), let $`g(Y)`$ be a bounded Borel payoff with range in $`[l,l+L]`$, paid at time $`\\tau`$. Then\n\n$$\n\\begin{gathered} p_h(g)-p_c(g)=h\\beta_g+R_h(g), \\end{gathered} \\tag{6.10}\n$$\n\n$$\n\\begin{gathered} \\beta_g=e^{-r\\tau}E_s\\!\\left[g(Y)\\sum_jd_j\\mathcal S_j(Y)\\right],\\qquad |R_h(g)|\\le e^{-r\\tau}\\frac L2\\,C_2h^2, \\end{gathered} \\tag{6.11}\n$$\n\nwhere, with $`G_j`$ the standardized increments,\n\n$$\n\\mathcal S_j=\\frac{G_j^2-1}{2s_j}-\\frac{G_j}{2\\sqrt{s_j}}, \\qquad J(s)=\\frac1{2s^2}+\\frac1{4s},\\qquad H_2(s)=\\frac2{s^2}+\\frac1{2s},\n$$\n\n$$\n\\begin{gathered} C_2= \\sqrt{\\sum_jb_j^2J(u_*)} +\\frac12\\sum_j(|d_j|+h_0b_j)^2H_2(u_*). \\end{gathered} \\tag{6.12}\n$$\n\n*Proof.* The Gaussian density $`p_s`$ is twice differentiable in $`L^1`$ in its variance parameters, with\n\n$$\n\\|Dp_s[z]\\|_1\\le\\sqrt{\\sum_jz_j^2J(s_j)},\\qquad \\|D^2p_s[z,z]\\|_1\\le\\sum_jz_j^2H_2(s_j),\n$$\n\nthe first bound being the square root of the Fisher information form. Inserting (6.9) in the $`L^1`$ Taylor formula gives $`\\|p_{s_h}-p_s-hDp_s[d]\\|_1\\le C_2h^2`$. The density remainder integrates to zero, so the payoff may be centered as $`g-l-L/2`$, which gives (6.11). Appendix D contains the derivative calculations and the dominated convergence. $`\\square`$\n\nThe score $`\\mathcal{S}_j`$ is the derivative of the log density with respect to the variance of month $`j`$, so $`\\beta_g`$ is the sum over months of the payoff's sensitivity to that month's variance times the Euler error $`d_j`$ of that variance. For a put with strike $`K`$ and maturity $`\\tau`$ the sum collapses to\n\n$$\n\\begin{gathered} \\beta_{K,\\tau} =a(\\tau)\\,\\frac{S_0\\phi(d_1)}{2\\sqrt{I(\\tau)}}, \\qquad d_1=\\frac{\\log(S_0/K)+r\\tau+I(\\tau)/2}{\\sqrt{I(\\tau)}} , \\end{gathered} \\tag{6.13}\n$$\n\nthe Black–Scholes sensitivity to total variance multiplied by the variance error: a vega times a quadrature error. All nine calibration puts therefore have a first-order bias with the sign of $`v_0-\\bar v`$. For the Asian spread the coefficient combines the sensitivities to the twelve monthly variances with the twelve errors $`d_j`$, and its sign turns with the initial variance.\n\nOn the family\n\n$$\n\\begin{gathered} \\kappa=3,\\quad\\bar v=.045,\\quad v_0\\in[.03,.06],\\quad h_0=1/768, \\end{gathered} \\tag{6.14}\n$$\n\nthe remainder for the Asian spread is at most $`.000165217`$ uniformly, and a separate finite-step comparison of the two Gaussian laws gives $`|p_h(\\psi)-p_c(\\psi)|\\le.004990924`$. The score bound\n\n$$\n\\begin{gathered} |\\beta_{\\rm Asian}| \\le {15e^{-rT}\\over2}\\sqrt{\\sum_{j=1}^{12}d_j^2J(s_j)} \\end{gathered} \\tag{6.15}\n$$\n\nyields the enclosures of Table 9.\n\n**Table 9. Enclosures of the Asian leading coefficient and of the remainder.**\n\n| Initial variance | Enclosure of $`\\beta_{\\rm Asian}`$ | Upper bound on $`|e_h-h\\beta|`$, $`h=1/768`$ |\n| --- | --- | --- |\n| 0.03 | $`[-3.441491358,3.441491358]`$ | 0.000121667773 |\n| 0.04 | $`[-0.928189505,0.928189505]`$ | 0.000036646884 |\n| 0.05 | $`[-0.785853082,0.785853082]`$ | 0.000034532369 |\n| 0.06 | $`[-2.057270822,2.057270822]`$ | 0.000099684585 |\n\nA sampling estimate of the coefficient itself, with 2,097,152 Gaussian vectors per point, gives $`-0.041015`$, $`-0.011364`$, $`0.008329`$ and $`0.019917`$ in increasing order of initial variance, with standard errors $`0.002079`$, $`0.000566`$, $`0.000483`$ and $`0.001274`$. The coefficient changes sign between $`v_0=.04`$ and $`.05`$, as the vanishing of every $`d_j`$ at $`v_0=\\bar v`$ requires, and the Euler bias on this face is of order $`10^{-5}`$ at 768 steps. Table 10 reports the scaled residual $`(e_h-h\\beta_{\\rm Asian})/h^2`$, estimated with common samples across steps through a Gaussian density ratio.\n\n**Table 10. Sampling estimates of the scaled second-order residual.**\n\n| Initial variance | $`h=1/192`$ | $`1/384`$ | $`1/768`$ | $`1/1536`$ |\n| --- | --- | --- | --- | --- |\n| 0.03 | $`-0.063243`$ | $`-0.062944`$ | $`-0.062796`$ | $`-0.062722`$ |\n| 0.04 | $`-0.017226`$ | $`-0.017145`$ | $`-0.017105`$ | $`-0.017085`$ |\n| 0.05 | 0.011698 | 0.011646 | 0.011620 | 0.011607 |\n| 0.06 | 0.027912 | 0.027785 | 0.027722 | 0.027690 |\n\nAt $`h=1/768`$ the standard errors are $`.003733`$, $`.001000`$, $`.000863`$ and $`.002319`$, and estimates across steps at one point are correlated. The residual is stable in $`h`$ across the four steps, as a second-order expansion predicts; the constants of (6.14) cover $`h\\le1/768`$.\n\nThe two faces fit together. At the stochastic point (5.6), $`v_0=\\bar v`$ and the deterministic mechanism is switched off, so the whole bias $`-0.156h`$ comes from volatility of volatility and correlation. At the second point $`v_0=.04`$ the finite-sum difference is $`-.000221954`$, lower by $`.000019`$; the deterministic coefficient at $`v_0=.04`$ predicts $`-0.011364/768=-.000015`$.\n\n### 6.4 Coupling the two laws\n\nThe interval (5.17) allows for the four costs of the kink one at a time. With identical laws and exact inputs its width is\n\n$$\n\\begin{gathered} 2\\gamma\\left(K_1^{-1}+K_2^{-1}\\right)D_P(c), \\end{gathered} \\tag{6.16}\n$$\n\nabout $`.02`$ here whatever the step, and the widths of Table 5 at the two finer steps are that allowance plus the inversion tails. The Euler gap involves only the change in each cost between the two laws, and comparing the costs path by path keeps their common part out of the allowance. Appendix H proves the inequality under a coupling of the conditional triples $`(a,cg,\\sigma)`$ of the two laws, and evaluates it in closed form on the deterministic face at $`v_0=.04`$: the width of the allowance falls from $`.01538363715`$ to $`.00008815020`$, a factor of 175.\n\n## 7. After calibration\n\nA pricing engine is used twice: to fit parameters to liquid quotes and to value the exotic at the fitted parameters. An error in the engine moves the fitted parameters as well as the price. If the engine misprices the quotes and the exotic in the same direction, refitting absorbs part of the error; if the exotic loads on a different stretch of the variance curve than the quotes do, refitting can add to it. This section bounds the net effect on every quantile of the posterior distribution of the Asian price. Theorem 7.1 holds for any two pricing laws; the numbers are computed for a prior concentrated near deterministic variance and nine synthetic quotes.\n\n### 7.1 Price transfer for small volatility of volatility\n\nSuppose that the deterministic reference variance and its Euler values lie in $`[m,M]`$ with $`0\\lt m\\le M\\lt \\infty`$, that $`\\kappa\\ge\\kappa_-\\gt 0`$, and that\n\n$$\n\\begin{gathered} \\kappa h\\le1,\\qquad 0\\le\\xi\\le\\sqrt{\\kappa_-M/2}. \\end{gathered} \\tag{7.1}\n$$\n\nWith $`C_P=\\sqrt{M/(2\\kappa_-)}`$ and $`C_Q=\\sqrt{2M/\\kappa_-}`$, the Itô isometry for the model and the 1-Lipschitz property of the positive part for the scheme give\n\n$$\n\\begin{gathered} E(V_t^\\xi-v(t))^2\\le\\xi^2C_P^2,\\qquad E(V_j^{\\xi,h}-v_j)^2\\le\\xi^2C_Q^2, \\end{gathered} \\tag{7.2}\n$$\n\nand, with the lower bound $`m`$ and Doob's inequality,\n\n$$\n\\begin{gathered} E\\max|Z_\\ell^\\xi-Z_\\ell^0| \\le\\xi C_\\ell\\left(\\frac T2+2\\sqrt{T/m}\\right),\\qquad \\ell=P,Q. \\end{gathered} \\tag{7.3}\n$$\n\nA put is $`K`$-Lipschitz in the log price, and the spread is $`K_2`$-Lipschitz in the maximum norm of the fixing log prices. So (7.3) gives uniform radii from every calibration price and from the target price to their deterministic-variance counterparts. Appendix E has the proof. On the parameter box\n\n$$\n\\begin{gathered} \\kappa\\in[2.999,3.001],\\quad\\bar v\\in[.04499,.04501],\\\\ v_0\\in[.03001,.05999],\\quad \\xi\\in[10^{-5},4\\cdot10^{-5}], \\quad\\rho\\in[-.8,-.3], \\end{gathered} \\tag{7.4}\n$$\n\nthe resulting bound on the Euler gap is $`.020920`$ for the Asian spread and $`.016701`$ for each of the nine puts with maturities $`1/4,1/2,1`$ and strikes $`90,100,110`$.\n\n### 7.2 The posterior\n\nNumerical error acts on the likelihood and on its normalizing constant [[11](#ref-11)]. Fix $`\\kappa=3`$, $`\\bar v=.045`$, and let\n\n$$\n\\begin{gathered} u=v_0\\sim U[.03,.06],\\quad \\xi\\sim U[10^{-7},10^{-6}],\\quad \\rho\\sim U[-.8,-.3] \\end{gathered} \\tag{7.5}\n$$\n\nbe independent. Write $`\\zeta=(\\xi,\\rho)`$ with prior law $`\\nu`$; in this section and in Appendix F the index $`m\\in\\{P,Q\\}`$ denotes the pricing law. Let $`c_m(u,\\zeta)`$ be the vector of calibration prices and $`J_m(u,\\zeta)`$ the Asian price under law $`m`$. For synthetic quotes $`y`$ and a positive-definite $`\\Sigma_d`$ define\n\n$$\nw_m(u,\\zeta)= \\exp\\!\\left[-\\frac12(c_m-y)^T\\Sigma_d^{-1}(c_m-y)\\right], \\qquad \\bar w_m(u)=\\int w_m(u,\\zeta)\\,\\nu(d\\zeta),\n$$\n\n$$\n\\begin{gathered} Z_m=\\frac1{.03}\\int_{.03}^{.06}\\bar w_m(u)\\,du,\\qquad \\mu_m(du,d\\zeta)=\\frac{w_m(u,\\zeta)}{.03Z_m}\\,du\\,\\nu(d\\zeta). \\end{gathered} \\tag{7.6}\n$$\n\nEach law has its own calibration prices, its own normalizing constant and its own target price.\n\n**Theorem 7.1 (Transfer to every quantile).** Suppose the two marginal posteriors of $`u`$ admit a coupling $`(U_P,U_Q)`$ with $`|U_P-U_Q|\\le\\delta_u`$ almost surely, that the reference targets satisfy\n\n$$\n|J_P^0(u)-J_P^0(v)|\\le L_0|u-v|, \\qquad |J_Q^0(u)-J_P^0(u)|\\le\\varepsilon_0,\n$$\n\nand that $`|J_m(u,\\zeta)-J_m^0(u)|\\le e_{m,J}`$. Then the left generalized quantiles of the Asian price under the two posteriors satisfy\n\n$$\n\\begin{gathered} |q_Q(p)-q_P(p)| \\le \\varepsilon_0+e_{P,J}+e_{Q,J}+L_0\\delta_u \\qquad(0\\lt p\\lt 1). \\end{gathered} \\tag{7.7}\n$$\n\n*Proof.* Let $`K_m(u,d\\zeta)=w_m(u,\\zeta)\\nu(d\\zeta)/\\bar w_m(u)`$. Extend the coupling by drawing $`\\zeta_P`$ from $`K_P(U_P,\\cdot)`$ and $`\\zeta_Q`$ from $`K_Q(U_Q,\\cdot)`$; by Fubini's theorem the two marginals are $`\\mu_P`$ and $`\\mu_Q`$. The triangle inequality gives\n\n$$\n|J_Q(U_Q,\\zeta_Q)-J_P(U_P,\\zeta_P)| \\le\\varepsilon_0+e_{P,J}+e_{Q,J}+L_0\\delta_u=:E,\n$$\n\nhence $`F_P(z-E)\\le F_Q(z)\\le F_P(z+E)`$ for the two distribution functions, and (7.7) follows by taking left inverses. Atoms are allowed. $`\\square`$\n\nThe four terms have distinct meanings. $`\\varepsilon_0`$ is the Euler gap of the target at fixed parameters, $`e_{P,J}+e_{Q,J}`$ pays for replacing a small $`\\xi`$ by zero, and $`L_0\\delta_u`$ is the transport term: the sensitivity of the target to the initial variance times the distance the fitted variance moves when the engine changes.\n\n### 7.3 Numbers\n\nPartition $`[.03,.06]`$ into 4096 cells. On each cell and over the whole $`\\zeta`$-domain the likelihood is enclosed with (7.3) and Black prices at the reference variance; the enclosures survive integration against $`\\nu`$, which gives envelopes of the two marginal distribution functions with their normalizing mass, and an inverse-distribution coupling driven by one uniform variable gives $`\\delta_u`$ (Appendix F). The single-quote case uses the one-year put struck at 100, the nine-quote case the whole vector, both with $`\\Sigma_d=.75^2(.75I_d+.25\\mathbf{1}\\mathbf{1}^T)`$. Then\n\n$$\nL_0\\le213.792201829985,\\quad \\varepsilon_0\\le.004493504474,\\quad e_{P,J}+e_{Q,J}\\le.000393595478,\n$$\n\n$$\n\\begin{array}{c|c|c} \\text{Case}&\\delta_u&\\sup_{0\\lt p\\lt 1}|q_Q(p)-q_P(p)|\\\\ \\hline \\text{Single quote}&3/204800&\\le.008018821658\\\\ \\text{Nine quotes}&3/81920&\\le.012716404217 \\end{array} \\tag{7.8}\n$$\n\nThe fitted initial variance moves by at most $`1.5\\cdot10^{-5}`$ and $`3.7\\cdot10^{-5}`$; in volatility terms that is at most 0.43 and 1.06 hundredths of a volatility point. The transport term is 39 percent of the first bound and 62 percent of the second: with nine quotes the envelopes of the posterior are wider and the coupling takes more steps. Table 11 refines the cells.\n\n**Table 11. Bound on the displacement of every quantile of the Asian price.**\n\n| Cells | Single quote | Nine quotes | Coupling steps, single/nine |\n| --- | --- | --- | --- |\n| 4096 | 0.008018821658 | 0.012716404217 | 2 / 5 |\n| 8192 | 0.006452960804 | 0.010367612938 | 2 / 7 |\n| 16384 | 0.006061495591 | 0.009584682511 | 3 / 12 |\n\nAt 16384 cells the bounds are 24.4 and 24.6 percent smaller than at 4096, with couplings of 3 and 12 steps instead of 2 and 5. For the continuous-model posterior at 4096 cells, the 2.5, 50 and 97.5 percent quantiles of the initial variance lie in\n\n$$\n\\begin{array}{lccc} \\text{one quote:}&[.0307690,.0307764]&[.0444067,.0444141]&[.0590917,.0590992]\\\\ \\text{nine quotes:}&[.0308203,.0308350]&[.0435131,.0435425]&[.0587182,.0587256] \\end{array}\n$$\n\n## 8. Two uses\n\n### 8.1 An acceptance test for a pricing engine\n\nThe benchmarks for Heston Asian options in the literature are Monte Carlo estimates with standard errors. Corollary 5.2 is a benchmark of a different kind. Any correct implementation of the continuous model for the contract (5.5)–(5.6) must return a value in $`[6.508371733,\\;6.518868974]`$, and any correct implementation of the scheme (5.2) with 768 steps a value whose expectation lies in $`[6.507844281,\\;6.519014106]`$. A library that prints 6.52 for either has an error that no sample size will explain. The interval is 16 basis points of the price wide, enough to catch a wrong day count, a misplaced correlation or a scheme that is not the one documented.\n\nThe gap interval also converts a statement about the scheme into one about the model. If a system encloses the price under the scheme in $`[q_L,q_U]`$, by simulation with a confidence interval or otherwise, then by Theorem 5.1\n\n$$\n\\begin{gathered} p_c\\in[q_L-u,\\;q_U-\\ell] \\end{gathered} \\tag{8.1}\n$$\n\nfor the endpoints $`[\\ell,u]`$ of (5.19).\n\n### 8.2 A numerical line in the valuation reserve\n\nSupervisory guidance on model risk asks that validation establish whether a model performs as expected and identify its limitations and errors [[5](#ref-5)]. European prudent valuation rules require a model risk adjustment built from a range of plausible valuations at 90 percent confidence, or from expert judgment where no range can be produced [[15](#ref-15)]. The numerical part of that uncertainty, the distance between the number a system prints and the price under the model it claims to implement, is usually handled by judgment. The results here replace judgment by an inequality in three situations.\n\nFor the monthly average options of Table 1 the closed-form proxy value is the Black price to within a hundredth of a cent per barrel, proved, and Table 2 puts the cost of the kink at the same order under stochastic volatility. No numerical reserve is needed for the kink.\n\nFor a long window the cubic law gives the size of the question before any computation: evaluate (1.2), multiply by the strike, and compare the result with the quote step. If it is smaller, the proxy value is the price. If it is not, as for the annual put of Table 1 or the spread of Section 5, the interval of Theorem 5.1 is the reserve, and its width, a few tenths of a cent to a cent per hundred of underlying, can be read against the bid–ask spread of the contract.\n\nFor a simulation engine the discretization bias is a systematic error: it has the same sign on every trade valued with the same scheme and does not diversify across a book. Section 6 gives its order and, for the proxy value, its coefficient; Corollary 5.2 gives a bound that holds at the step in use.\n\n## 9. Conclusion\n\nExercising an arithmetic Asian option on the geometric average costs the expected overshoot when the two averages disagree. Across strikes that cost adds up to a Bregman divergence between the averages, and at each strike it is at most the divergence divided by the volatility of the return that precedes the first fixing, times $`K/\\sqrt{2\\pi}`$. The first statement needs no model. The second holds for calls and puts under Black, Heston, rough volatility and jump dynamics alike, and it bounds the error of every classical conditioning lower bound. The ratio of dispersion to level risk decides everything that follows. For a monthly window it is so small that the closed-form proxy value is the price to a hundredth of a cent per barrel, with a bound that is nearly attained. For an annual window it is a cent, and a proof is worth the computation: under Heston the price of a twelve-fixing spread is enclosed in an interval 16 basis points wide, and the price under a 768-step Euler scheme within $`.011025`$ of it.\n\nThe Euler error is small and regular. The scheme is an Euler method for the Riccati equation, its bias in the proxy value is $`-0.156h`$, and the kink adds less than $`10^{-6}`$, because its costs under the two laws nearly cancel. A path-by-path comparison carries the cancellation into a proof: on the deterministic-variance face the coupled allowance is 175 times narrower than the separate one.\n\n## Appendices and supporting material\n\nThis page includes the complete main text and references. Appendices, supplementary proofs, code and verification material are available in the [GitHub repository](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options) and the [full manuscript](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/manuscript/report.md#appendix-a-finite-certification-of-the-weighted-second-moment).\n\n## References\n\n\n\n1. Andersen, Leif. Simple and efficient simulation of the Heston stochastic volatility model. *The Journal of Computational Finance*, 11(3):1–42, 2008. [https://doi.org/10.21314/JCF.2008.189](https://doi.org/10.21314/JCF.2008.189).\n\n\n\n2. Bally, Vlad, and Talay, Denis. The law of the Euler scheme for stochastic differential equations: I. Convergence rate of the distribution function. *Probability Theory and Related Fields*, 104(1):43–60, 1996. [https://doi.org/10.1007/BF01303802](https://doi.org/10.1007/BF01303802).\n\n\n\n3. Black, Fischer. The pricing of commodity contracts. *Journal of Financial Economics*, 3(1–2):167–179, 1976. [https://doi.org/10.1016/0304-405X(76)90024-6](https://doi.org/10.1016/0304-405X(76)90024-6).\n\n\n\n4. Blas, Javier. Uncovering the Secret History of Wall Street's Largest Oil Trade. *Bloomberg Markets*, 4 April 2017.\n\n\n\n5. Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation, and Office of the Comptroller of the Currency. Revised Guidance on Model Risk Management. SR Letter 26-2, 17 April 2026. [https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm](https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm).\n\n\n\n6. Breeden, Douglas T., and Litzenberger, Robert H. Prices of State-Contingent Claims Implicit in Option Prices. *The Journal of Business*, 51(4):621–651, 1978. [https://doi.org/10.1086/296025](https://doi.org/10.1086/296025).\n\n\n\n7. Bregman, L. M. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. *USSR Computational Mathematics and Mathematical Physics*, 7(3):200–217, 1967. [https://doi.org/10.1016/0041-5553(67)90040-7](https://doi.org/10.1016/0041-5553(67)90040-7).\n\n\n\n8. Broadie, Mark, and Kaya, Özgür. Exact Simulation of Stochastic Volatility and Other Affine Jump Diffusion Processes. *Operations Research*, 54(2):217–231, 2006. [https://doi.org/10.1287/opre.1050.0247](https://doi.org/10.1287/opre.1050.0247).\n\n\n\n9. Carr, Peter, and Madan, Dilip. Towards a Theory of Volatility Trading. In R. Jarrow (ed.), *Volatility: New Estimation Techniques for Pricing Derivatives*, 417–427. Risk Books, London, 1998.\n\n\n\n10. CME Group. NYMEX Rulebook, Chapter 341: WTI Average Price Option. [https://www.cmegroup.com/rulebook/NYMEX/3/341.pdf](https://www.cmegroup.com/rulebook/NYMEX/3/341.pdf).\n\n\n\n11. Cotter, S. L., Dashti, M., and Stuart, A. M. Approximation of Bayesian Inverse Problems for PDEs. *SIAM Journal on Numerical Analysis*, 48(1):322–345, 2010. [https://doi.org/10.1137/090770734](https://doi.org/10.1137/090770734).\n\n\n\n12. Curran, Michael. Valuing Asian and Portfolio Options by Conditioning on the Geometric Mean Price. *Management Science*, 40(12):1705–1711, 1994. [https://doi.org/10.1287/mnsc.40.12.1705](https://doi.org/10.1287/mnsc.40.12.1705).\n\n\n\n13. Demeterfi, Kresimir, Derman, Emanuel, Kamal, Michael, and Zou, Joseph. A Guide to Volatility and Variance Swaps. *The Journal of Derivatives*, 6(4):9–32, 1999. [https://doi.org/10.3905/jod.1999.319129](https://doi.org/10.3905/jod.1999.319129).\n\n\n\n14. Ehm, Werner, Gneiting, Tilmann, Jordan, Alexander, and Krüger, Fabian. Of quantiles and expectiles: consistent scoring functions, Choquet representations and forecast rankings. *Journal of the Royal Statistical Society: Series B*, 78(3):505–562, 2016. [https://doi.org/10.1111/rssb.12154](https://doi.org/10.1111/rssb.12154).\n\n\n\n15. European Commission. Commission Delegated Regulation (EU) 2016/101 supplementing Regulation (EU) No 575/2013 with regard to regulatory technical standards for prudent valuation under Article 105(14). *Official Journal of the European Union*, L 21, 28 January 2016. [https://eur-lex.europa.eu/eli/reg_del/2016/101/oj](https://eur-lex.europa.eu/eli/reg_del/2016/101/oj).\n\n\n\n16. Fusai, Gianluca, and Kyriakou, Ioannis. General Optimized Lower and Upper Bounds for Discrete and Continuous Arithmetic Asian Options. *Mathematics of Operations Research*, 41(2):531–559, 2016. [https://doi.org/10.1287/moor.2015.0739](https://doi.org/10.1287/moor.2015.0739).\n\n\n\n17. Geman, Hélyette, and Yor, Marc. Bessel Processes, Asian Options, and Perpetuities. *Mathematical Finance*, 3(4):349–375, 1993. [https://doi.org/10.1111/j.1467-9965.1993.tb00092.x](https://doi.org/10.1111/j.1467-9965.1993.tb00092.x).\n\n\n\n18. Hairer, Ernst, Nørsett, Syvert P., and Wanner, Gerhard. *Solving Ordinary Differential Equations I: Nonstiff Problems*, second edition, Section II.8. Springer, Berlin, 1993. [https://doi.org/10.1007/978-3-540-78862-1](https://doi.org/10.1007/978-3-540-78862-1).\n\n\n\n19. Heston, Steven L. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options. *The Review of Financial Studies*, 6(2):327–343, 1993. [https://doi.org/10.1093/rfs/6.2.327](https://doi.org/10.1093/rfs/6.2.327).\n\n\n\n20. Johansson, Fredrik. Arb: Efficient Arbitrary-Precision Midpoint-Radius Interval Arithmetic. *IEEE Transactions on Computers*, 66(8):1281–1292, 2017. [https://doi.org/10.1109/TC.2017.2690633](https://doi.org/10.1109/TC.2017.2690633).\n\n\n\n21. Kemna, A. G. Z., and Vorst, A. C. F. A pricing method for options based on average asset values. *Journal of Banking and Finance*, 14(1):113–129, 1990. [https://doi.org/10.1016/0378-4266(90)90039-5](https://doi.org/10.1016/0378-4266(90)90039-5).\n\n\n\n22. Kim, Bara, Kim, Jeongsim, Kim, Jerim, and Wee, In Suk. A recursive method for discretely monitored geometric Asian option prices. *Bulletin of the Korean Mathematical Society*, 53(3):733–749, 2016. [https://doi.org/10.4134/BKMS.b150283](https://doi.org/10.4134/BKMS.b150283).\n\n\n\n23. Lee, Roger W. Option Pricing by Transform Methods: Extensions, Unification, and Error Control. *The Journal of Computational Finance*, 7(3):51–86, 2004. [https://doi.org/10.21314/JCF.2004.121](https://doi.org/10.21314/JCF.2004.121).\n\n\n\n24. Levy, Edmond. Pricing European average rate currency options. *Journal of International Money and Finance*, 11(5):474–491, 1992. [https://doi.org/10.1016/0261-5606(92)90013-N](https://doi.org/10.1016/0261-5606(92)90013-N).\n\n\n\n25. London Metal Exchange. Averaging solutions: a guide to trading and hedging average prices on the London Metal Exchange. [https://www.lme.com/-/media/Files/Trading/Contract-types/LME-Averaging-Solutions-Brochure.pdf](https://www.lme.com/-/media/Files/Trading/Contract-types/LME-Averaging-Solutions-Brochure.pdf).\n\n\n\n26. Lord, Roger. Partially exact and bounded approximations for arithmetic Asian options. *The Journal of Computational Finance*, 10(2):1–52, 2006. [https://www.risk.net/node/2160406](https://www.risk.net/node/2160406).\n\n\n\n27. Lord, Roger, Koekkoek, Remmert, and van Dijk, Dick. A comparison of biased simulation schemes for stochastic volatility models. *Quantitative Finance*, 10(2):177–194, 2010. [https://doi.org/10.1080/14697680802392496](https://doi.org/10.1080/14697680802392496).\n\n\n\n28. Mickel, Annalena, and Neuenkirch, Andreas. The weak convergence order of two Euler-type discretization schemes for the log-Heston model. *IMA Journal of Numerical Analysis*, 43(6):3326–3356, 2023. [https://doi.org/10.1093/imanum/drac069](https://doi.org/10.1093/imanum/drac069).\n\n\n\n29. Nielsen, J. Aase, and Sandmann, Klaus. Pricing Bounds on Asian Options. *Journal of Financial and Quantitative Analysis*, 38(2):449–473, 2003. [https://doi.org/10.2307/4126759](https://doi.org/10.2307/4126759).\n\n\n\n30. QuantLib contributors. AnalyticDiscreteGeometricAveragePriceAsianHestonEngine. Source code at a pinned commit. [https://github.com/lballabio/QuantLib/blob/966a4cc101049ca36a888b2ce223aa96d3f3b22d/ql/experimental/asian/analytic_discr_geom_av_price_heston.hpp](https://github.com/lballabio/QuantLib/blob/966a4cc101049ca36a888b2ce223aa96d3f3b22d/ql/experimental/asian/analytic_discr_geom_av_price_heston.hpp).\n\n\n\n31. Rogers, L. C. G., and Shi, Z. The value of an Asian option. *Journal of Applied Probability*, 32(4):1077–1088, 1995. [https://doi.org/10.2307/3215221](https://doi.org/10.2307/3215221).\n\n\n\n32. Talay, Denis, and Tubaro, Luciano. Expansion of the global error for numerical schemes solving stochastic differential equations. *Stochastic Analysis and Applications*, 8(4):483–509, 1990. [https://doi.org/10.1080/07362999008809220](https://doi.org/10.1080/07362999008809220).\n\n\n\n33. Thompson, G. W. P. Fast narrow bounds on the value of Asian options. Judge Institute of Management, University of Cambridge, WP09/2002, 2002. [https://www.jbs.cam.ac.uk/wp-content/uploads/2020/08/wp0209.pdf](https://www.jbs.cam.ac.uk/wp-content/uploads/2020/08/wp0209.pdf).\n\n\n\n34. Turnbull, Stuart M., and Wakeman, Lee Macdonald. A Quick Algorithm for Pricing European Average Options. *Journal of Financial and Quantitative Analysis*, 26(3):377–389, 1991. [https://doi.org/10.2307/2331213](https://doi.org/10.2307/2331213).\n\n\n\n35. Valencia, Fabián. To Hedge or to Self-Insure? The Benefits of Mexico's Oil Hedging Program. *IMF Blog*, 26 March 2018. [https://www.imf.org/external/np/blog/dialogo/032618.pdf](https://www.imf.org/external/np/blog/dialogo/032618.pdf).\n"
}
diff --git a/website/content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json b/website/content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json
index 66849f4..eb3d4c9 100644
--- a/website/content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json
+++ b/website/content/selected-research/critical-dependence-stress-for-diversification-gains-under-fixed-marginals.json
@@ -2,10 +2,10 @@
"slug": "critical-dependence-stress-for-diversification-gains-under-fixed-marginals",
"title": "Critical Dependence Stress for Diversification Gains under Fixed Marginals",
"description": "How much dependence stress erases a diversification gain under fixed marginals: a variational limit, a quadratic critical budget, the marginal-restoration premium and projected flow duality.",
- "sourceUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/paper.pdf",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/paper.pdf",
"repositoryUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals",
- "sourceEditedAt": "2026-11-19T03:29:40-05:00",
- "sourceCommit": "a62cc41106e58e5d2df1387a1cf39ebfa65e36e8",
- "sourceBlob": "e9efa283fc4b8efe2cac0ca483fa3e5da81e01e5",
- "markdown": "## Abstract\n\nHow much dependence uncertainty can a diversification benefit withstand when every marginal loss law is fixed? We study the smallest weighted $`1`$-Wasserstein perturbation that reverses the Expected Shortfall ranking of two nearby allocations evaluated under a common stressed law. For a positive continuous density on a compact box, exchangeability of the two traded positions eliminates the first-order gain, so the relevant stress radius is $`\\rho=s\\varepsilon^2`$ for trade size $`\\varepsilon`$. At this scale the limiting risk profile is continuous and strictly increasing, and its unique zero identifies the asymptotic critical budget. We call the excess over the local-efficiency lower bound the *marginal-restoration premium*. This premium vanishes exactly when background displacement is cheapest per unit change in portfolio loss. Otherwise it is strictly positive, even when the profile is exactly linear on an initial interval: in a uniform benchmark that attains the initial slope, the proved lower bound on the critical budget is already $`30\\%`$ above the local-efficiency prediction. At equal effective costs, continuous maximum-flow/minimum-cut duality characterizes the right derivative of the critical budget with respect to the background penalty. The proof rests on a $`\\Gamma`$-limit for the rescaled transport energies and on convergence of the paired risk differences, with the two ES thresholds optimized separately. Coordinatewise quantile restoration incurs a marginal-flux cost, although its leading effects on the two ES values cancel in their difference.\n\n\n\n## 1. Introduction\n\nHow much of a portfolio's diversification benefit survives uncertainty about dependence? A risk desk may have agreed marginal models for its individual positions and still lack an equally firm model for their joint losses with the background book [[31](#ref-pflugpohl2018)]. We ask how a proposed reallocation performs when these marginal models are kept but the coupling changes. More precisely, we look for the smallest dependence perturbation that reverses its Expected Shortfall (ES) advantage when both allocations face the same stressed law. We show that this critical perturbation is of second order in the size of the trade, and that the response at zero budget can understate it: in a uniform benchmark the proved lower bound is already $`30\\%`$ above the prediction from the initial slope.\n\nLet $`X_1,X_2`$ be the unit losses of two positions of equal initial cost. A transfer from holdings $`(1+\\varepsilon,1-\\varepsilon)`$ to $`(1,1)`$, with an unchanged background loss $`\\beta X_3`$, changes the portfolio loss from $`A_\\varepsilon=(1+\\varepsilon)X_1+(1-\\varepsilon)X_2+\\beta X_3=B+\\varepsilon H`$ to $`B`$, where\n\n$$\nB=X_1+X_2+\\beta X_3,\\qquad H=X_1-X_2,\\qquad \\beta>0.\n$$\n\nAt upper-tail probability $`p\\in(0,1)`$, we compare them through $`\\Phi_\\varepsilon(P)=ES_p^P(B)-ES_p^P(A_\\varepsilon)`$, optimizing the two ES thresholds separately under $`P`$ [[32](#ref-rockafellar2002)], so that $`\\Phi_\\varepsilon(P)<0`$ favors the transfer under that law. The background term $`\\beta X_3`$ cancels from the payoff difference $`A_\\varepsilon-B`$, yet we cannot drop it, because its dependence with the traded positions determines which states enter each portfolio's upper tail. A trade assessed in isolation can therefore rank differently once embedded in the whole book [[8](#ref-brandtner2018)].\n\nWriting $`\\Pi`$ for the Fréchet class of joint laws with these marginals, we measure distance from the reference law $`P_0`$ by weighted $`W_1`$ with ground cost $`c_\\gamma(x,y)=|y-x|_\\gamma=\\sum_i\\gamma_i|y_i-x_i|`$, where each $`\\gamma_i>0`$ sets the budget penalty per unit coordinate loss displacement as losses are re-paired across joint states. At budget $`\\rho`$, the stress problem is\n\n$$\nD_\\varepsilon(\\rho)\n=\\sup_{\\substack{P\\in\\Pi\\\\ W_\\gamma(P,P_0)\\le\\rho}}\n\\bigl\\{ES_p^P(B)-ES_p^P(A_\\varepsilon)\\bigr\\},\n\\qquad m_\\varepsilon=-D_\\varepsilon(0).\n$$\n\nUnder an exchangeable reference law, the traded positions have equal marginal ES contributions, so the nominal gain is of second order: $`m_\\varepsilon=(J/p)\\varepsilon^2+o(\\varepsilon^2)`$, with $`J>0`$. For example, when $`\\gamma_1\\le\\min\\{\\gamma_2,\\gamma_3/\\beta\\}`$, the local exchange in Proposition [2.2](#dep-prop-nominal-benchmark) attains an initial stress derivative $`K_\\varepsilon`$ of order one, which suggests the reversal budget $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon=O(\\varepsilon^2)`$. But the derivative at zero budget does not tell us whether this efficiency persists until the nominal gain is exhausted. Since the predicted reversal already occurs at scale $`\\rho=s\\varepsilon^2`$, the tangent at zero is not enough, and we need the risk profile over a nonvanishing interval of the rescaled budget $`s`$.\n\nThree effects contribute at the critical scale. Moving the $`O(\\varepsilon)`$ mass in an $`O(\\varepsilon)`$-layer around the reference tail boundary by distances of order $`\\varepsilon`$ changes both transport cost and tail expectation by $`O(\\varepsilon^2)`$, matching the nominal gain. This local movement need not preserve the marginals. For a fixed regular local plan, the marginal CDF discrepancies are $`O(\\varepsilon^2)`$; by Vallender's formula [[39](#ref-vallender1974)], their $`L^1`$ norms give the one-dimensional restoration costs, even when compensation occurs far from the tail boundary. An $`O(\\varepsilon)`$ shift of either ES threshold also contributes at order $`\\varepsilon^2`$ through the curvature of its objective. We must therefore optimize the local displacement, marginal restoration, and both thresholds together.\n\nWhen the reference law has a positive continuous density on a compact box $`I^2\\times I_3`$ and is exchangeable in its first two coordinates, Theorem [3.1](#dep-thm-critical-limit) gives a finite, continuous profile\n\n$$\nF(s)=\\lim_{\\varepsilon\\downarrow0}\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2),\n\\qquad F(0)=-J/p,\n$$\n\nwith convergence locally uniform in $`s\\ge0`$. The proof rests on a $`\\Gamma`$-limit [[11](#ref-dalmaso1993)] of the transport energies that charges both for the local displacement and for restoring its marginal projections. In the recovery argument, coordinatewise quantile maps attain the restoration cost for each fixed regular local plan. This correction moves losses throughout the box, but its leading effects on the two ES values cancel outside the threshold layer, and approximation then extends the paired-risk convergence to kernels with finite displacement moment.\n\nLet $`\\rho^-_\\varepsilon`$ and $`\\rho^+_\\varepsilon`$ be the infima of budgets permitting a tie and a strict reversal. The strict-growth theorem identifies their common leading coefficient as the unique zero $`s^*`$ of $`F`$:\n\n$$\n\\rho^-_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2),\\qquad\n\\rho^+_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2).\n$$\n\n*Initial efficiency and its limit.* The effective unit costs are $`\\gamma_1,\\gamma_2`$, and $`c_3=\\gamma_3/\\beta`$, with $`c_{\\min}=\\min\\{\\gamma_1,\\gamma_2,c_3\\}`$. If the first traded coordinate is cheapest, Proposition [3.2](#dep-prop-balanced-linear-profile) constructs a positive interval $`[0,s_{\\rm lin}]`$ on which $`F(s)=-J/p+s/(p\\gamma_1)`$. To reach zero at this rate, however, efficient displacement must remain possible throughout the required interval. Theorem [4.2](#dep-thm-zero-premium) characterizes attainment of the universal efficiency bound:\n\n$$\ns^*=Jc_{\\min}\\quad\\Longleftrightarrow\\quad\nc_3\\le\\min\\{\\gamma_1,\\gamma_2\\}.\n$$\n\nWe call the nonnegative gap $`s^*-Jc_{\\min}`$ the *marginal-restoration premium* (normalized by $`Jc_{\\min}`$ when reported as a percentage). When background displacement is cheapest, exchangeability permits opposite shifts to balance on every fixed-background slice. When only traded directions are cheapest, a corner region requires a displacement moment of order $`\\delta^2`$, but its coordinate projections admit only $`O(\\delta^4)`$ of compensating moment near the diagonal (Figure [2](#dep-fig-corner)). Under the attained-slope condition of Proposition [2.2](#dep-prop-nominal-benchmark), with $`c_{\\min}=\\gamma_11`$.\n\n*An analytical benchmark.* Consider independent uniform losses on $`[0,2]^3`$, with $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(1,1,2)`$. Here $`g=3/8`$ and $`J=5/32`$, so Proposition [2.2](#dep-prop-nominal-benchmark) gives $`m_\\varepsilon=(5/16)\\varepsilon^2+o(\\varepsilon^2)`$. By Corollary [5.2](#dep-prop-uniform-variable-cost), the profile is exactly $`F(s)=2s-5/16`$ for $`0\\le s\\le1/96`$, yet its zero satisfies $`13/64\\le s^*\\le29/96`$. So even in this uniform model, where the initial slope is attained, the proved lower bound on reversal is already $`30\\%`$ above the local-efficiency prediction $`5/32`$.\n\n*Sensitivity at equal effective costs.* Along the cost family $`\\gamma_1=\\gamma_2=\\gamma`$, $`c_3=\\gamma(1+\\eta)`$ for $`\\eta\\ge0`$, the baseline $`Jc_{\\min}=\\gamma J`$ is constant, so $`s^*`$ and the marginal-restoration premium have the same right derivative $`\\gamma Ja_+`$ at $`\\eta=0`$. Here $`a_+\\in(0,1)`$ is the least background share among efficient stresses, characterized in Theorem [4.3](#dep-prop-boundary-sensitivity) by continuous maximum-flow/minimum-cut duality; Section 5 shows how this share and the budget sensitivity $`\\gamma Ja_+`$ can move in opposite directions as the reference dependence changes.\n\nOur question lies between extremal risk aggregation and local distributional sensitivity. On the aggregation side, subadditivity bounds aggregate risk [[1](#ref-artzner1999), [25](#ref-mcneil2015)], and rearrangements within a Fréchet class describe how diversification depends on the coupling [[33](#ref-ruschendorf2013), [15](#ref-embrechts2013), [5](#ref-bernard2014), [16](#ref-embrechts2015), [41](#ref-wang2015superadditive)]. Partial information leads to convex-order and mixability constraints [[22](#ref-li2018partial)] but can still permit comonotonic upper tails [[12](#ref-devecchinendelstreicher2026)], and fixed-marginal MES bounds quantify uncertainty in individual tail contributions [[9](#ref-chenfurmanlin2026)]. In allocation comparisons, the symmetry imposed on the candidate laws is decisive. Chen et al. [[10](#ref-chen2022)] and Liu and Liu [[23](#ref-liu2026)] use exchangeability and majorization to identify diversification protection, and Liu and Liu also analyze when complete dependence ambiguity favors concentration and how Wasserstein uncertainty modifies that conclusion. We impose exchangeability at the reference law alone: stressed couplings need not remain exchangeable, and the stationarity that symmetry gives at the reference does not by itself protect the reallocation once stress breaks the symmetry. We use this setting to measure the distance to a specified decision reversal, in the spirit of reverse sensitivity analysis [[28](#ref-pesenti2019), [29](#ref-pesenti2022)].\n\nTransport duality is a natural way to preserve the marginal information [[40](#ref-villani2003), [34](#ref-santambrogio2015)]. In Wasserstein robust optimization [[7](#ref-blanchet2019), [19](#ref-gao2023)], adding marginal constraints confines the perturbation to dependence changes [[18](#ref-gao2017), [13](#ref-eckstein2020)]; such constraints encode calibrated risk-neutral laws in multiasset robust pricing [[14](#ref-eckstein2021multiasset)], and Wasserstein and copula restrictions also arise in benchmark portfolio optimization [[30](#ref-pesentijaimungal2023)]. Here the constraint acts on a common stressed loss law for two allocations. After rescaling, its full coordinate projections remain as a transport penalty, and the efficient allocation of displacement reduces to a capacity-constrained flow problem [[20](#ref-korman2015)].\n\nLocal Wasserstein expansions relate the sensitivity of optimized decisions and AVaR criteria to changes in the reference law [[2](#ref-bartl2021)], and adapted transport brings in the information available in multiperiod decisions [[4](#ref-bartlwiesel2023)]. When a first-order term vanishes, one turns to higher-order expansions: Fiechtner and Blanchet [[17](#ref-fiechtner2026)] derive a second-order regret expansion under a unique empirical-risk minimizer in this situation, and Sauldubois [[35](#ref-sauldubois2026)] develops higher-order Wasserstein and adapted-Wasserstein expansions, including martingale constraints. Dependence restrictions also affect local sensitivity, as in Lam [[21](#ref-lam2018)], while Namkoong [[26](#ref-namkoong2026)] evaluates the first-order response over restricted transport directions. In our problem the trade size and stress radius vanish together, so locating a decision reversal at scale $`\\rho=s\\varepsilon^2`$ requires the full rescaled profile, with global marginal compensation and both optimized ES thresholds retained over the budget interval needed to reach zero.\n\nSection 2 derives the nominal gain and the initial stress slope, and Section 3 takes the variational limit at scale $`\\rho=s\\varepsilon^2`$ and identifies the unique zero of the limiting profile. Section 4 characterizes zero premium and represents the boundary derivative by continuous projected flow–cut duality. Section 5 works through analytical examples and what they imply for reverse stress testing. The analytic proofs are in Appendix A, and the potential estimates and section calculations for the examples are in Appendix B.\n\n**Notation and conventions** We write $`T_\\#\\mu`$ for a pushforward and measure displacements by $`|d|_\\gamma=\\sum_i\\gamma_i|d_i|`$. The constant in $`O_\\alpha(\\varepsilon^k)`$ may depend on the fixed kernel and model, but not on $`\\varepsilon`$. Scalar section integrands are extended by zero outside their domains, and projection identities are equalities of finite signed Borel measures. We use $`\\inf\\varnothing=+\\infty`$; allocation ratios are assigned the value zero where both numerator and denominator vanish.\n\n\n\n## 2. Model formulation and the local benchmark\n\nWe first formulate the comparison for arbitrary integrable marginal laws, which are enough for a worst stressed model to exist. The critical-scale analysis will use a positive continuous density and exchangeability of the two traded positions, both introduced in Section 2.2.\n\n### 2.1 Fixed marginals and a common stress law\n\nLet $`n\\ge2`$, let $`\\mu_1,\\ldots,\\mu_n`$ have finite first moments, and fix $`P_0`$ in the Fréchet class $`\\Pi(\\mu_1,\\ldots,\\mu_n)=\\{P\\in\\mathcal P(\\mathbb R^n):P_i=\\mu_i,\\ 1\\le i\\le n\\}`$.\n\nFor exposure vectors $`b,h\\in\\mathbb R^n`$, the losses before and after the proposed reallocation are $`A_\\varepsilon=B+\\varepsilon H`$ and $`B`$, with $`B=b\\cdot X`$ and $`H=h\\cdot X`$. We use the upper-tail mass $`p\\in(0,1)`$ and the Expected Shortfall representation of Rockafellar and Uryasev [[32](#ref-rockafellar2002)],\n\n\n\n$$\nES_p^P(L)=\\min_{t\\in\\mathbb R}\n\\left\\{t+\\frac1pE_P(L-t)_+\\right\\}.\n\\tag{2.1}\n$$\n\nWe write $`\\mathscr E_L(P;t)=t+p^{-1}E_P(L-t)_+`$ for the threshold objective. For an integrable loss, its minimum is attained on the nonempty compact interval characterized by\n\n\n\n$$\nt\\in\\arg\\min_{u}\\mathscr E_L(P;u)\n\\quad\\Longleftrightarrow\\quad P(L>t)\\le p\\le P(L\\ge t).\n\\tag{2.2}\n$$\n\nWithout a continuity assumption on the stressed loss law, support gaps can leave a nondegenerate interval of minimizing thresholds. We therefore keep the set-valued characterization [(2.2)](#dep-eq-2-2), so that subsequent estimates apply to every minimizer, including for laws with atoms.\n\nUnder a common candidate law $`P\\in\\Pi`$, we compare $`\\Phi_\\varepsilon(P)=ES_p^P(B)-ES_p^P(A_\\varepsilon)`$, minimizing each loss over its own threshold. Maximizing this paired difference tests whether a single dependence scenario can erase the nominal gain, whereas separate maximizations of the two ES values may select different joint laws. We measure changes in dependence with positive weights $`\\gamma_i`$ and ground cost\n\n$$\nc_\\gamma(x,y)=\\sum_{i=1}^n\\gamma_i|y_i-x_i|,\n\\qquad W_\\gamma(P,P_0)=\\min_{\\pi\\in\\Gamma(P_0,P)}\\int c_\\gamma\\,d\\pi.\n$$\n\nWithin $`\\Pi=\\Pi(\\mu_1,\\ldots,\\mu_n)`$, transport changes how component losses are paired across joint states. With the weights and all marginal laws held fixed, we define the worst comparison at budget $`\\rho`$ and the nominal gain by\n\n\n\n$$\n\\mathcal A_\\rho=\\{P\\in\\Pi:W_\\gamma(P,P_0)\\le\\rho\\},\\quad\nD_\\varepsilon(\\rho)=\\max_{P\\in\\mathcal A_\\rho}\\Phi_\\varepsilon(P),\\quad\nm_\\varepsilon=-\\Phi_\\varepsilon(P_0).\n\\tag{2.3}\n$$\n\nProposition [2.1](#dep-prop-general-wellposedness) guarantees attainment in [(2.3)](#dep-eq-2-3) even with atoms, singular dependence, and unbounded marginal supports. A nondecreasing profile can vanish on a whole interval, so we distinguish the infimal budgets for a tie and a strict reversal,\n\n\n\n$$\n\\rho^-_\\varepsilon=\\inf\\{\\rho\\ge0:D_\\varepsilon(\\rho)\\ge0\\},\n\\qquad\n\\rho^+_\\varepsilon=\\inf\\{\\rho\\ge0:D_\\varepsilon(\\rho)>0\\},\n\\tag{2.4}\n$$\n\nwhich, by Theorem [3.3](#dep-thm-strict-crossing), have the same leading asymptotics in the exchangeable box model of Section 2.2.\n\n\n\n**Proposition 2.1 (Existence under integrable marginal laws).** Let $`n\\ge2`$, let $`\\mu_1,\\ldots,\\mu_n`$ have finite first moments, and fix $`P_0\\in\\Pi(\\mu_1,\\ldots,\\mu_n)`$, positive weights $`\\gamma_i`$, exposure vectors $`b,h\\in\\mathbb R^n`$, and $`p\\in(0,1)`$. Then $`\\Pi`$ and every $`\\mathcal A_\\rho`$ are compact for $`W_\\gamma`$, and the maximum in [(2.3)](#dep-eq-2-3) is attained. For each fixed $`\\varepsilon`$, the profile $`D_\\varepsilon`$ is nondecreasing and Lipschitz on $`[0,\\infty)`$.\n\n*Proof.* The fixed-marginal class is tight and weakly closed. Moreover, along any weakly convergent sequence $`P_k\\to P`$ in $`\\Pi`$, the first moments satisfy $`\\int|x|_\\gamma\\,dP_k=\\sum_i\\gamma_i\\int|x_i|\\,d\\mu_i=\\int|x|_\\gamma\\,dP`$, so the Wasserstein convergence criterion [[40](#ref-villani2003)] implies $`W_\\gamma`$-compactness of $`\\Pi`$ and its closed subsets $`\\mathcal A_\\rho`$. For a linear loss $`L=a\\cdot X`$, coupling the laws and minimizing the uniformly Lipschitz hinges over their thresholds gives\n\n$$\n|ES_p^P(L)-ES_p^Q(L)|\\le p^{-1}\\|a\\|_{\\gamma,*}W_\\gamma(P,Q),\n\\qquad \\|a\\|_{\\gamma,*}=\\max_i|a_i|/\\gamma_i.\n$$\n\nApplying this bound to $`B=b\\cdot X`$ and $`A_\\varepsilon=(b+\\varepsilon h)\\cdot X`$, we find that their ES difference $`\\Phi_\\varepsilon`$ is Lipschitz with constant $`C_\\varepsilon=p^{-1}(\\|b\\|_{\\gamma,*}+\\|b+\\varepsilon h\\|_{\\gamma,*})`$. Compactness of $`\\mathcal A_\\rho`$ then ensures attainment. To compare two budgets $`\\rho_2>\\rho_1`$, take a maximizer $`P`$ at $`\\rho_2`$. The mixture $`P_a=(1-a)P_0+aP`$, $`a=\\rho_1/\\rho_2`$, satisfies $`W_\\gamma(P_a,P_0)\\le a\\rho_2`$ and $`W_\\gamma(P_a,P)\\le(1-a)\\rho_2`$, by mixing an optimal plan with the corresponding identity plans. Comparing $`\\Phi_\\varepsilon(P_a)`$ and $`\\Phi_\\varepsilon(P)`$ yields $`0\\le D_\\varepsilon(\\rho_2)-D_\\varepsilon(\\rho_1)\\le C_\\varepsilon(\\rho_2-\\rho_1)`$. ∎\n\nFor a transfer $`H=X_i-X_j`$ between positions with the same marginal law, replacing $`X_j`$ by $`X_i`$ preserves every marginal and gives $`A_\\varepsilon=B`$ at cost $`\\gamma_jE_{P_0}|X_i-X_j|`$; reversing their roles yields\n\n$$\n0\\le\\rho^-_\\varepsilon\n\\le\\min\\{\\gamma_i,\\gamma_j\\}\\,E_{P_0}|X_i-X_j|<\\infty.\n$$\n\nSo the tie budget is finite, but this bound leaves its small-trade scale open. Under the exchangeable box assumptions introduced next, Section 3 shows that both critical budgets are of order $`\\varepsilon^2`$.\n\n\n\n### 2.2 The exchangeable box model\n\nTo isolate how dependence with the rest of a portfolio affects a small reallocation, we keep the two traded losses separate and collect all unchanged positions into their aggregate loss $`\\beta X_3`$. The marginal law of $`X_3`$ counts as retained information, and we specify the transport metric directly on $`(X_1,X_2,X_3)`$. This aggregate model fixes the information used in the stress test: the retained marginal and displacement cost are those of the aggregate loss, a different specification from fixing separately the marginals and transport costs of every underlying background position.\n\n\n\n**Standing assumptions (exchangeable box model)** Let\n\n$$\nI=[\\ell,r],\\qquad I_3=[\\ell_3,r_3],\\qquad\nS=I^2\\times I_3,\\qquad \\Delta_I=r-\\ell>0,\\quad \\Delta_3=r_3-\\ell_3>0.\n$$\n\nWe also write $`I_1=I_2=I`$ when treating coordinate projections together. The reference law has a continuous density $`f>0`$ on the closed box $`S`$, normalized to integrate to one and satisfying $`f(x,y,z)=f(y,x,z)`$. We take\n\n$$\nB=X_1+X_2+\\beta X_3,\\qquad H=X_1-X_2,\\qquad\nA_\\varepsilon=B+\\varepsilon H,\\qquad\n\\beta>0,\\quad 0<\\varepsilon<1,\n$$\n\nwith arbitrary $`p\\in(0,1)`$ and $`\\gamma_i>0`$, so that $`b=(1,1,\\beta)`$ and $`h=(1,-1,0)`$ in the preceding notation. All model parameters remain fixed as $`\\varepsilon\\downarrow0`$. The first two unit positions are expressed at equal initial cost, so the trade changes their holdings from $`(1+\\varepsilon,1-\\varepsilon)`$ to $`(1,1)`$ without changing the initial outlay. The position $`\\beta X_3`$ represents the background book, whose holding remains unchanged during the trade.\n\nThe stressed law may be any coupling in $`\\Pi((P_0)_1,(P_0)_2,(P_0)_3)`$; exchangeability and absolute continuity are assumptions on the reference law only. Reference symmetry makes the equal allocation stationary under $`P_0`$, and equality of the first two marginals gives $`E_P[A_\\varepsilon]-E_P[B]=\\varepsilon(E_P[X_1]-E_P[X_2])=0`$ under every candidate, so before implementation costs the ES comparison distinguishes tail risks between portfolios with the same expected loss.\n\nWe work on a compact box with a positive continuous density because that lets us study this dependence mechanism with uniform control of marginal restoration. Positivity on the closed box gives both the interior mass that local exchanges need and lower bounds on the marginal densities, hence Lipschitz bounds for their inverse CDFs, and these bounds enter the quantile recovery in Section 3.2 and Appendix A.5. Section 5.3 takes up the additional difficulties posed by unbounded marginals.\n\n\n\n**Loss units and effective stress costs** A shift $`\\Delta X_3`$ changes the portfolio loss by $`\\beta\\Delta X_3`$ at cost $`\\gamma_3|\\Delta X_3|=(\\gamma_3/\\beta)|\\beta\\Delta X_3|`$. So $`c_1=\\gamma_1`$, $`c_2=\\gamma_2`$, and $`c_3=\\gamma_3/\\beta`$ are the costs per unit change of $`B`$ along the three coordinate axes, and the ratio $`c_3/\\gamma_1`$ compares a background displacement with a first-position displacement having the same effect on portfolio loss.\n\n**Canonical box normalization** Writing the original coordinates as $`X=T_{\\rm box}(u)=(\\ell+\\kappa u_1,\\ell+\\kappa u_2,\\ell_3+\\kappa_3u_3)`$, with $`\\kappa=\\Delta_I/2`$ and $`\\kappa_3=\\Delta_3/2`$, gives a normalized model on $`[0,2]^3`$ with $`\\widehat\\beta=\\beta\\kappa_3/\\kappa`$ and $`\\widehat\\gamma=(\\kappa\\gamma_1,\\kappa\\gamma_2,\\kappa_3\\gamma_3)`$. By cash invariance and positive homogeneity of ES, the original paired risk is $`\\kappa`$ times its normalized counterpart, while corresponding transport plans have equal cost; hence the critical budgets agree and $`p`$ is unchanged. A common rescaling of all transport weights rescales those budgets.\n\n\n\n### 2.3 Nominal curvature and the exact initial slope\n\nUnder the reference model of Section 2.2, let $`v`$ be the upper-tail $`p`$ threshold of $`B`$. The marginal expected shortfall (MES) of a unit position $`X_i`$, namely $`E_{P_0}[X_i\\mid B>v]`$, is the ES derivative with respect to its holding, and multiplying by that holding gives its Euler contribution [[36](#ref-scaillet2004), [38](#ref-tasche2001)]. The first two holdings in $`B`$ both equal one, so the directional derivative along $`H=X_1-X_2`$ is their MES difference: with $`T_H=E_{P_0}[H1_{\\{B>v\\}}]`$, it equals $`p^{-1}T_H=E_{P_0}[X_1-X_2\\mid B>v]=0`$ by exchangeability. To find the gain from a nonzero trade we then have to differentiate the moving tail boundary, which leads to a conditional variance on the tail section.\n\nWe record the classical section differentiation formula under our continuity assumptions in Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4).\n\nThe tail boundary can be parameterized by\n\n$$\n\\Omega=\\{(x,y)\\in(\\ell,r)^2:\\ell_3\n\nSince $`p\\in(0,1)`$ and $`f>0`$, the threshold satisfies $`2\\ell+\\beta\\ell_3\n\n$$\ng=\\sigma(\\Omega)>0,\\qquad J=\\frac12\\int_\\Omega H^2\\,d\\sigma>0.\n\\tag{2.5}\n$$\n\nHere $`g=f_B(v)`$ is the loss density at the threshold, and $`d\\sigma/g`$ is its normalized conditional section law. Exchangeability makes the conditional mean of $`H`$ zero, giving $`2J/g=\\operatorname{Var}_{P_0}(H\\mid B=v)`$, so the second-order diversification gain is governed by the conditional variance of the trade payoff on $`\\{B=v\\}`$, weighted by the threshold density $`g`$.\n\n\n\n**Proposition 2.2 (Nominal benchmark and exact initial efficiency).** In the exchangeable box model of Section 2.2:\n\n(i) *Nominal gain and threshold curvature.* The reference thresholds and nominal gain satisfy\n\n\n\n$$\nt_\\varepsilon-v=o(\\varepsilon),\\qquad\nm_\\varepsilon=\\frac Jp\\varepsilon^2+o(\\varepsilon^2).\n\\tag{2.6}\n$$\n\nThe threshold of $`B+\\eta H`$ is unique for every $`|\\eta|<1`$, and $`m_\\varepsilon>0`$ for $`0<\\varepsilon<1`$. With $`L_{c,\\varepsilon}=B+c\\varepsilon H`$, the unoptimized objective satisfies, uniformly for $`c=0,1`$ and $`|\\tau|\\le T<\\infty`$,\n\n\n\n$$\n\\mathscr E_{L_{c,\\varepsilon}}(P_0;v+\\varepsilon\\tau)\n=ES_p^{P_0}(B)+\\frac{\\varepsilon^2}{p}\n\\left(c^2J+\\frac g2\\tau^2\\right)+o(\\varepsilon^2).\n\\tag{2.7}\n$$\n\n(ii) *Fixed-marginal Lipschitz bound and initial slope.* For every $`P,Q\\in\\Pi`$,\n\n\n\n$$\n|\\Phi_\\varepsilon(P)-\\Phi_\\varepsilon(Q)|\\le \\kappa_\\varepsilon W_\\gamma(P,Q),\\quad\n\\kappa_\\varepsilon=\\frac1p\\max\\left\\{\\frac{1+\\varepsilon/2}{\\gamma_1},\n\\frac{1-\\varepsilon/2}{\\gamma_2},\\frac\\beta{\\gamma_3}\\right\\}.\n\\tag{2.8}\n$$\n\nFor fixed $`0<\\varepsilon<1`$, if\n\n\n\n$$\n\\frac{1+\\varepsilon/2}{\\gamma_1}\n\\ge\\max\\left\\{\\frac{1-\\varepsilon/2}{\\gamma_2},\\frac\\beta{\\gamma_3}\\right\\},\n\\tag{2.9}\n$$\n\nthen there exists $`r_\\varepsilon>0`$ such that\n\n\n\n$$\nK_\\varepsilon=\\frac{1+\\varepsilon/2}{p\\gamma_1},\\qquad\nD_\\varepsilon(\\rho)=-m_\\varepsilon+K_\\varepsilon\\rho\n\\quad(0\\le\\rho\\le r_\\varepsilon).\n\\tag{2.10}\n$$\n\n*Proof.* For part (i), the coefficients of $`B+\\eta H`$ are positive whenever $`|\\eta|<1`$, and its support is $`[2\\ell+\\beta\\ell_3,2r+\\beta r_3]`$. Since every interior level meets $`\\operatorname{int}S`$, Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4) gives a positive continuous loss density and a unique threshold. The gradients of $`H`$ and $`B+\\eta H`$ are linearly independent, so $`H`$ is nonconstant on each such section. Its conditional variance is therefore strictly positive, and the same lemma implies strict convexity of $`\\eta\\mapsto ES_p^{P_0}(B+\\eta H)`$ on $`(-1,1)`$.\n\nAt $`\\eta=0`$, exchanging the first two coordinates preserves both the reference tail and its boundary measure and reverses $`H`$, giving $`T_H=m_1=0`$. In the notation of Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4), the remaining section moments at $`(0,v)`$ are $`m_0=g`$ and $`m_2=2J`$. Thus the threshold derivative vanishes and the optimized second derivative is $`2J/p`$, which yield [(2.6)](#dep-eq-2-6) by Taylor expansion at zero. Strict convexity makes this stationary point the strict minimum, so $`m_\\varepsilon>0`$ for all $`0<\\varepsilon<1`$. For the unoptimized objective $`\\Psi`$, the Hessian at $`(0,v)`$ is $`\\operatorname{diag}(2J/p,g/p)`$. Expanding at the trade size and threshold displacement $`(c\\varepsilon,\\varepsilon\\tau)`$ gives [(2.7)](#dep-eq-2-7), uniformly on bounded $`\\tau`$-intervals.\n\n*The fixed-marginal Lipschitz bound.* We next establish the uniform Lipschitz bound over $`\\Pi`$. Given any minimizing $`u`$ for $`B`$ under $`Q`$ and any minimizing $`t`$ for $`A_\\varepsilon`$ under $`P`$, let\n\n$$\n\\begin{aligned}\nh_{u,t}&=(B-u)_+-(A_\\varepsilon-t)_+,\\\\\n\\Delta_B&=\\mathscr E_B(P;u)-ES_p^P(B),&\n\\Delta_A&=\\mathscr E_{A_\\varepsilon}(Q;t)-ES_p^Q(A_\\varepsilon).\n\\end{aligned}\n$$\n\nBoth gaps are nonnegative, and direct subtraction gives\n\n\n\n$$\n\\Phi_\\varepsilon(P)-\\Phi_\\varepsilon(Q)\n=p^{-1}\\int h_{u,t}\\,d(P-Q)-\\Delta_B-\\Delta_A.\n\\tag{2.11}\n$$\n\nA direct bound on $`h_{u,t}`$ would use its largest absolute $`x_1`$-slope, $`1+\\varepsilon`$. Because $`P`$ and $`Q`$ share their marginals, adding an integrable separable potential does not change the integral against $`P-Q`$. We can thus choose a representative of $`h_{u,t}`$ modulo marginal potentials with a smaller Lipschitz bound. Here the required adjustment is linear: since $`\\int H\\,d(P-Q)=0`$, we may add any scalar multiple of $`H=X_1-X_2`$. The extreme $`x_1`$-slopes, $`-(1+\\varepsilon)`$ and $`1`$, have midpoint $`-\\varepsilon/2`$, so adding $`\\varepsilon H/2`$ minimizes their largest absolute value among constant shifts and simultaneously centers the $`x_2`$-slopes. Away from the two tail planes, the four regime slopes are\n\n$$\n\\begin{array}{c|rrrr}\n(1_{\\{B>u\\}},1_{\\{A_\\varepsilon>t\\}})&(0,0)&(1,0)&(0,1)&(1,1)\\\\\n\\partial_{x_1}h_{u,t}&0&1&-(1+\\varepsilon)&-\\varepsilon\\\\\n\\partial_{x_1}(h_{u,t}+\\varepsilon H/2)&\\varepsilon/2&1+\\varepsilon/2&-1-\\varepsilon/2&-\\varepsilon/2.\n\\end{array}\n$$\n\nWith the extreme $`x_1`$-slopes centered at zero, their largest absolute value drops to $`1+\\varepsilon/2`$, while the corresponding bounds for $`x_2`$ and $`x_3`$ are $`1-\\varepsilon/2`$ and $`\\beta`$. Setting $`G_{u,t}=h_{u,t}+\\varepsilon H/2`$, which is absolutely continuous along coordinate lines, we obtain\n\n$$\n|G_{u,t}(Y)-G_{u,t}(X)|\n\\le(1+\\varepsilon/2)|Y_1-X_1|\n+(1-\\varepsilon/2)|Y_2-X_2|+\\beta|Y_3-X_3|.\n$$\n\nIntegrating against an optimal transport plan and then interchanging $`P,Q`$ proves [(2.8)](#dep-eq-2-8). The mixing argument in Proposition [2.1](#dep-prop-general-wellposedness), now with this sharp constant, also gives\n\n\n\n$$\n0\\le D_\\varepsilon(\\rho_2)-D_\\varepsilon(\\rho_1)\n\\le \\kappa_\\varepsilon(\\rho_2-\\rho_1)\\qquad(0\\le\\rho_1<\\rho_2).\n\\tag{2.12}\n$$\n\n\n\n*Attaining the initial slope.* Under [(2.9)](#dep-eq-2-9), we construct an attaining exchange near an interior intersection of the reference tail planes $`\\{B=v\\}`$ and $`\\{A_\\varepsilon=t_\\varepsilon\\}`$. To locate such an intersection, fix $`\\varepsilon\\in(0,1)`$ and let $`[m_F,M_F]`$ be the nondegenerate range of $`A_\\varepsilon`$ on $`F_v=\\{B=v\\}\\cap S`$. Both losses increase along each nonconstant segment from the lower vertex $`x_{\\min}=(\\ell,\\ell,\\ell_3)`$. Any point with $`B>v`$ lies beyond an intersection of that segment with $`F_v`$, so $`t_\\varepsilon\\le m_F`$ would imply $`\\{B>v\\}\\subset\\{A_\\varepsilon>t_\\varepsilon\\}`$. Since $`A_\\varepsilon`$ is nonconstant on $`F_v`$, the second tail would also contain an open set below $`B=v`$, contradicting the equal tail probabilities and $`f>0`$. Rays toward the upper corner similarly exclude $`t_\\varepsilon\\ge M_F`$, which leaves $`t_\\varepsilon\\in(m_F,M_F)`$. The relative-interior property of a linear map on a convex section gives an intersection $`c=(x_*,y_*,z_*)\\in\\operatorname{int}S`$.\n\nTo exchange the first-coordinate marginal intervals of two small source cubes in opposite disagreement regions along $`e_1`$, we first choose reference points with the same coordinate $`x_*`$ and seek equal and opposite changes in the two losses. From the interior intersection $`c`$, a displacement $`(0,-\\zeta_0,a\\zeta_0)`$ changes $`B`$ by $`(-1+\\beta a)\\zeta_0`$ and $`A_\\varepsilon`$ by $`(-(1-\\varepsilon)+\\beta a)\\zeta_0`$. Equating their sum to zero determines $`a=(1-\\varepsilon/2)/\\beta`$. Thus the points\n\n$$\nc_\\pm=(x_*,y_*\\mp \\zeta_0,z_*\\pm(1-\\varepsilon/2)\\zeta_0/\\beta)\n$$\n\nhave opposite strict tail memberships for small $`\\zeta_0>0`$: at $`c_+`$, the losses differ from their thresholds by $`-\\varepsilon \\zeta_0/2`$ for $`B`$ and $`\\varepsilon \\zeta_0/2`$ for $`A_\\varepsilon`$, with both signs reversed at $`c_-`$. Choosing $`0<\\delta_x<\\varepsilon\\zeta_0/(1+\\varepsilon)`$ bounds the loss change along either half-translation by the strict tail margin. With this margin fixed, we can take $`\\eta_0>0`$ small enough that the cubes $`E_\\pm`$ of half-width $`\\eta_0`$, centered at $`c_\\pm\\pm(\\delta_x/2)e_1`$, have disjoint second-coordinate intervals and remain inside $`S`$ and their respective tail regions throughout translation by $`\\mp\\delta_xe_1`$.\n\nThe uniform subdensities $`\\nu_\\pm=\\lambda1_{E_\\pm}\\,dx\\,dy\\,dz`$, where $`\\lambda=\\min_Sf`$, have equal mass $`\\mu=\\lambda(2\\eta_0)^3`$ and sum at most $`P_0`$. The translations $`T_\\pm(x)=x\\mp\\delta_xe_1`$ therefore define the admissible perturbation\n\n\n\n$$\nP_\\theta=P_0-\\theta(\\nu_++\\nu_-)\n+\\theta(T_+)_\\#\\nu_++\\theta(T_-)_\\#\\nu_-,\\qquad 0\\le\\theta\\le1.\n\\tag{2.13}\n$$\n\nThe translations exchange the first-coordinate marginal intervals, fix the other coordinates, and cross neither tail plane. Thus $`P_\\theta`$ preserves all marginals and both reference tail probabilities, leaving $`v`$ and $`t_\\varepsilon`$ optimal. Integrating the affine hinges on the two cubes gives\n\n\n\n$$\n\\Phi_\\varepsilon(P_\\theta)-\\Phi_\\varepsilon(P_0)\n=\\frac{(2+\\varepsilon)\\theta\\mu\\delta_x}{p}\n=K_\\varepsilon(2\\gamma_1\\theta\\mu\\delta_x).\n\\tag{2.14}\n$$\n\nThe plan has cost $`2\\gamma_1\\theta\\mu\\delta_x`$, and the sharp bound [(2.8)](#dep-eq-2-8) forces this to equal $`W_\\gamma(P_\\theta,P_0)`$. Varying $`\\theta\\in[0,1]`$ proves [(2.10)](#dep-eq-2-10) with $`r_\\varepsilon=2\\gamma_1\\mu\\delta_x>0`$. ∎\n\nIf $`\\gamma_2\\ge\\gamma_1`$ and $`\\gamma_3\\ge\\beta\\gamma_1`$, condition [(2.9)](#dep-eq-2-9) holds for every $`\\varepsilon\\in(0,1)`$, which suggests the extrapolated reversal budget $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon`$. The construction gives $`r_\\varepsilon>0`$ for each fixed trade size, but the disagreement regions narrow with $`\\varepsilon`$, so it leaves $`r_\\varepsilon/\\varepsilon^2`$ uncontrolled. Indeed, the unscaled bound $`|D_\\varepsilon(\\rho)|\\le\\varepsilon\\|H\\|_\\infty`$ makes both iterated limits in $`(\\rho,\\varepsilon)`$ vanish even though $`K_\\varepsilon\\to1/(p\\gamma_1)>0`$, which is why Section 3 tracks the profile under the joint scaling $`\\rho=s\\varepsilon^2`$.\n\n\n\n**Reference symmetry and stressed dependence** Our stress class allows asymmetric dependence between the two traded positions. If every candidate law were exchangeable, the losses $`B-\\varepsilon H`$ and $`B+\\varepsilon H`$ would have the same distribution, and convexity of ES would give\n\n\n\n$$\nES_p^P(B)\\le\\tfrac12 ES_p^P(B-\\varepsilon H)+\\tfrac12 ES_p^P(B+\\varepsilon H)\n=ES_p^P(A_\\varepsilon).\n\\tag{2.15}\n$$\n\nRequiring exchangeability of every stressed law would thus protect this reallocation [[10](#ref-chen2022), [23](#ref-liu2026)]. We impose it only at the reference, which gives stationarity there while permitting dependence stresses that test the persistence of the gain. Identical marginal laws alone do not imply stationarity. For example, on $`S=[0,2]^3`$ with $`\\beta=1`$ and $`p=1/2`$, the perturbation $`f_\\delta=[1+\\delta(x-y)(z-1)]/8`$, with $`0<|\\delta|<1/2`$, preserves uniform marginals but gives a nonzero tail expectation $`T_H`$. Antisymmetry in $`x,y`$ preserves the law of $`B=x+y+z`$, hence $`v=3`$. Pairing $`z=1\\pm t`$ leaves the strip $`2-t\n\n## 3. Variational limit and the critical profile\n\nUnder the traded-coordinate condition of Proposition [2.2](#dep-prop-nominal-benchmark), a local exchange attains the initial slope. To find out how far that efficiency persists toward reversal, we now hold $`s=\\rho/\\varepsilon^2`$ fixed. The limiting objects introduced below connect the stress profile to its reversal coefficient:\n\n$$\n\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2)\n\\xrightarrow{\\text{locally uniformly}}\nF(s)\n\\xrightarrow{\\text{unique zero}}\ns^*.\n$$\n\nThe first passage comes from Theorem [3.1](#dep-thm-critical-limit), which combines convergence of the transport energies with exact marginal recovery and convergence of the paired risk. Strict growth in Theorem [3.3](#dep-thm-strict-crossing) then identifies the common critical coefficient.\n\n\n\n### 3.1 Local transport and marginal compensation\n\nAn $`O(\\varepsilon)`$-layer around $`\\{B=v\\}`$ contains $`O(\\varepsilon)`$ reference mass, so displacements of order $`\\varepsilon`$ change both the transport cost and the paired hinge expectation at order $`\\varepsilon^2`$. We resolve this layer through the normal coordinate $`u=(B(X)-v)/\\varepsilon`$ and displacement $`d=(Y-X)/\\varepsilon`$, while keeping the cost of any transport outside bounded windows that is needed to restore the marginals.\n\n**Definition (Marginal flux and effective energy)** On bounded $`u`$-windows the rescaled source converges to $`d\\nu=d\\sigma\\,du`$. We disintegrate a limiting plan with this source into a probability kernel $`\\alpha_{\\xi,u}(dd)`$, specifying the rescaled displacement at surface point $`\\xi`$ and normal coordinate $`u`$; stationary sources have kernel $`\\delta_0`$. Writing $`d\\Pi_\\alpha(\\xi,u,d)=\\alpha_{\\xi,u}(dd)\\,d\\nu(\\xi,u)`$, we require a finite weighted displacement moment even though the source measure $`\\nu=\\sigma\\otimes du`$ has infinite total mass:\n\n$$\nI(\\alpha)=\\int|d|_\\gamma\\,d\\Pi_\\alpha<\\infty,\n\\qquad |d|_\\gamma=\\sum_i\\gamma_i|d_i|.\n$$\n\nAcross a coordinate cut $`\\{X_i\\le q\\}`$, upward movements remove mass from below the cut and downward movements add mass to it. The resulting signed flux has a Lebesgue density $`j_i\\in L^1(I_i)`$, as shown by the section formulas in Appendix A.3, and we define\n\n\n\n$$\nj_i(q)dq=(\\xi_i)_\\#(d_i\\,d\\Pi_\\alpha),\\qquad\n\\mathcal C(\\alpha)=I(\\alpha)+\\sum_{i=1}^3\\gamma_i\\|j_i\\|_{L^1(I_i)}.\n\\tag{3.1}\n$$\n\nIn words, $`j_i`$ is the signed displacement moment projected onto the $`i`$-th coordinate, and its negative is the leading marginal CDF discrepancy produced by a regular local plan. The full plan preserves each marginal, so transport outside a bounded normal and displacement window has to compensate the discrepancy within it, and the one-dimensional CDF formula for $`W_1`$ [[39](#ref-vallender1974)] turns that compensation into a cost lower bound. The energy $`\\mathcal C`$ adds this limiting restoration cost to $`I`$ (Figure [1](#dep-fig-tail-layer)). Projection contracts total variation, giving $`I\\le\\mathcal C\\le2I`$, with $`\\mathcal C=I`$ when all projected fluxes vanish and $`\\mathcal C=2I`$ when each $`d_i`$ has a fixed sign $`\\Pi_\\alpha`$-almost everywhere, so that $`\\|j_i\\|_1=\\int|d_i|\\,d\\Pi_\\alpha`$.\n\n\n\n\n\n\n\n### 3.2 The limiting risk and convergence theorem\n\nUnder the same local plan, each loss has its own minimizing threshold. At threshold $`v+\\varepsilon\\tau`$, a displacement $`d`$ changes the leading centered and rescaled loss of $`L_{c,\\varepsilon}=B+c\\varepsilon H`$ from $`u+cH`$ to $`u+b\\cdot d+cH`$. Denote the corresponding hinge increment by $`\\mathfrak q_c(\\tau)=(u+b\\cdot d+cH-\\tau)_+-(u+cH-\\tau)_+`$. We write $`T_c^\\alpha(\\tau)=\\int\\mathfrak q_c(\\tau)\\,d\\Pi_\\alpha`$ for its integral.\n\n**Definition (Limiting paired risk)** Moving the threshold away from $`v`$ contributes the curvature term $`g\\tau^2/2`$ from [(2.7)](#dep-eq-2-7). Adding the local transport increment gives the two limiting threshold objectives, which must still be minimized separately:\n\n\n\n$$\nG_c^\\alpha(\\tau)=\\frac g2\\tau^2+T_c^\\alpha(\\tau),\\qquad\n\\mathcal V(\\alpha)=-\\frac Jp+\n\\frac{\\min_{\\tau_B}G_0^\\alpha(\\tau_B)-\\min_{\\tau_A}G_1^\\alpha(\\tau_A)}p.\n\\tag{3.2}\n$$\n\nThe energy $`\\mathcal C`$ depends only on the absolute displacement moment and the signed coordinate projections, and it is convex under mixtures of kernels. The paired risk also depends on how the displacements are distributed relative to the normal coordinate $`u`$: each of the two hinge objectives optimizes its own threshold, so their difference need not be concave in the common kernel.\n\nWe omit the superscript $`\\alpha`$ when the kernel is clear, and write $`\\kappa_B=\\max_i b_i/\\gamma_i`$, $`\\kappa_H=\\max\\{\\gamma_1^{-1},\\gamma_2^{-1}\\}`$. Since $`|T_c^\\alpha|\\le\\kappa_BI`$, the quadratic term gives coercivity and confines every minimizer to $`g\\tau^2/2\\le2\\kappa_BI`$. Together with convexity from Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103), these properties justify the threshold minimizations in $`\\mathcal V`$. We then optimize the paired value over the energy budget, with attainment proved in Theorem [3.1](#dep-thm-critical-limit):\n\n\n\n$$\nF(s)=\\max_{\\mathcal C(\\alpha)\\le s}\\mathcal V(\\alpha),\\qquad s\\ge0.\n\\tag{3.3}\n$$\n\nWe use vague convergence on $`\\mathscr X=\\mathcal M^+_{\\rm loc}(Z)`$, testing against compactly supported continuous functions on $`Z=I^2\\times\\mathbb R\\times\\mathbb R^3`$, so that compensation may occur outside bounded windows. We lift a plan $`\\pi`$ from $`P_0`$ to a law in $`\\Pi`$ by\n\n$$\n\\mathscr T_\\varepsilon\\pi=\\varepsilon^{-1}\n\\left(X_1,X_2,\\frac{B(X)-v}{\\varepsilon},\\frac{Y-X}{\\varepsilon}\\right)_\\#\\pi.\n$$\n\nHere the factor $`\\varepsilon^{-1}`$ compensates for the thickness of the source layer. The lifted coordinates determine $`(X,Y)`$ for fixed $`\\varepsilon`$, so the rescaled energy can retain the cost of the entire original plan:\n\n$$\n\\begin{aligned}\n\\mathcal E_\\varepsilon(\\Lambda)\n&=\\begin{cases}\n\\displaystyle\\int |d|_\\gamma\\,d\\Lambda\n=\\varepsilon^{-2}\\int c_\\gamma\\,d\\pi,\n&\\begin{gathered}\\Lambda=\\mathscr T_\\varepsilon\\pi,\\\\ \\pi\\in\\Gamma(P_0,P),\\ P\\in\\Pi,\\end{gathered}\\\\\n+\\infty,&\\text{otherwise},\n\\end{cases}\\\\\n\\mathcal E(\\Lambda)\n&=\\begin{cases}\n\\mathcal C(\\alpha),&\\Lambda=\\Pi_\\alpha,\\ I(\\alpha)<\\infty,\\\\\n+\\infty,&\\text{otherwise}.\n\\end{cases}\n\\end{aligned}\n$$\n\nHere we identify $`\\Pi_\\alpha`$ with a measure on $`Z`$ through $`(\\xi,u,d)\\mapsto(\\xi_1,\\xi_2,u,d)`$, using $`\\xi=(x,y,z_v(x,y))`$ on $`\\Omega`$ and assigning zero measure to $`(I^2\\setminus\\Omega)\\times\\mathbb R\\times\\mathbb R^3`$. Because $`\\mathcal E_\\varepsilon`$ includes the whole plan, including transport outside bounded normal windows, minimizing it over plans with target law $`P`$ gives $`\\varepsilon^{-2}W_\\gamma(P,P_0)`$. We formulate the following $`\\Gamma`$-convergence statement in this vague topology [[11](#ref-dalmaso1993)].\n\n\n\n**Theorem 3.1 ($`\\Gamma`$-convergence and the critical risk profile).** Consider the exchangeable box model of [Section 2.2](#dep-model-critical-box).\n\n(i) *Transport plans and compactness.* $`\\mathcal E_\\varepsilon`$ $`\\Gamma`$-converges to $`\\mathcal E`$ as $`\\varepsilon\\downarrow0`$ in the vague topology on $`\\mathscr X`$. Every sequence $`\\varepsilon_n\\downarrow0`$, $`\\Pi_n\\in\\mathscr X`$, with bounded $`\\mathcal E_{\\varepsilon_n}(\\Pi_n)`$ has a vaguely convergent subsequence. If $`\\Pi_n=\\mathscr T_{\\varepsilon_n}\\pi_n\\to\\Pi_\\alpha`$ has bounded energy and $`P_n`$ is the target law of $`\\pi_n`$, then $`\\Phi_{\\varepsilon_n}(P_n)/\\varepsilon_n^2\\to\\mathcal V(\\alpha)`$.\n\n(ii) *Target laws and exact recovery.* For every sequence $`\\varepsilon_n\\downarrow0`$ and every feasible sequence satisfying $`W_\\gamma(P_n,P_0)=O(\\varepsilon_n^2)`$, there are a subsequence and a kernel with finite $`I`$ such that\n\n\n\n$$\n\\mathcal C(\\alpha)\\le\\liminf_n\\frac{W_\\gamma(P_n,P_0)}{\\varepsilon_n^2},\\qquad\n\\frac{\\Phi_{\\varepsilon_n}(P_n)}{\\varepsilon_n^2}\\longrightarrow\\mathcal V(\\alpha).\n\\tag{3.4}\n$$\n\nConversely, every kernel with finite $`I`$ admits a recovery sequence of distributions preserving the original three full marginals, with\n\n\n\n$$\n\\limsup_{\\varepsilon\\downarrow0}\\frac{W_\\gamma(P_\\varepsilon,P_0)}{\\varepsilon^2}\\le\\mathcal C(\\alpha),\\qquad\n\\frac{\\Phi_\\varepsilon(P_\\varepsilon)}{\\varepsilon^2}\\longrightarrow\\mathcal V(\\alpha).\n\\tag{3.5}\n$$\n\n(iii) *The stress profile.* The maximum in [(3.3)](#dep-eq-3-3) is attained. The function $`F`$ is finite, continuous, and nondecreasing, with $`F(0)=-J/p`$ and $`F>0`$ at some finite budget. For every finite $`S_0`$,\n\n\n\n$$\n\\sup_{0\\le s\\le S_0}\\left|\\frac{D_\\varepsilon(s\\varepsilon^2)}{\\varepsilon^2}-F(s)\\right|\\longrightarrow0.\n\\tag{3.6}\n$$\n\n**Proof strategy** Theorem [3.1](#dep-thm-critical-limit)(i) gives $`\\Gamma`$-convergence of the transport energy and convergence of the paired risk along bounded-energy plans, and the exact marginal recovery in part (ii) makes the limiting kernel values asymptotically attainable in $`\\Pi`$. Part (iii) gives uniform convergence of the optimized profiles on compact budget intervals, which, together with the strict crossing in Theorem [3.3](#dep-thm-strict-crossing), locates the reversal budget $`s^*`$.\n\n*Step 1: Localization and the energy lower bound.* Because the two ES thresholds are optimized independently, comparison of the objectives requires a common threshold window for stresses with bounded rescaled energy. In Appendix [A.2](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L2), Lemma [A.2](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-threshold-statement.tex#L1) places every minimizing threshold in an $`O_M(\\varepsilon)`$ window. On bounded windows the lifted plans are locally compact; outside them, marginal conservation requires compensation for the projected local flux. The compensation bound above then gives the energy lower bound for arbitrary admissible plans. Choosing optimal plans yields the corresponding Wasserstein bound for their target laws.\n\n*Step 2: Convergence of the paired risk.* Bounded first moments make each hinge increment integrable, but need not give uniform integrability along convergent plans. The two objectives share the same large-displacement contribution, so we subtract $`\\mathfrak q_0(0)`$ from both. The relative increments are bounded on the common threshold window. Outside a normal window enlarged to account for the bounded reallocation $`H`$, a nonzero relative increment requires a sufficiently large displacement, so the first moment controls the far-field contributions. The baseline integral is finite but need not converge. Since it translates both threshold objectives by the same amount independently of the threshold, it cancels exactly from their difference of minima.\n\n*Step 3: Exact marginal recovery.* Let $`\\mathcal K_{\\rm fin}^c`$ denote probability kernels with compactly supported continuous weights on finitely many fixed displacement vectors and the remaining mass at zero, as in [(A.13)](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L136). For a fixed kernel in this class, local translations yield a preliminary law $`Q_\\varepsilon`$ with uniform marginal CDF expansion $`F_{i,\\varepsilon}-F_i=-\\varepsilon^2j_i+o(\\varepsilon^2)`$, where $`F_i`$ is the reference CDF. The quantile maps $`F_i^{-1}\\circ F_{i,\\varepsilon}`$ restore all three marginals at cost $`\\varepsilon^2\\sum_i\\gamma_i\\|j_i\\|_1+o(\\varepsilon^2)`$, with $`O_\\alpha(\\varepsilon^2)`$ displacements.\n\nTo see why this correction does not alter the leading paired risk, compare its displacement scale with the $`O_\\alpha(\\varepsilon)`$ threshold band. Outside a sufficiently wide band, the paired hinge has gradient $`-\\varepsilon h`$ above both thresholds and zero below both, so multiplying by the restoration displacement gives an $`O_\\alpha(\\varepsilon^3)`$ contribution. Inside the band the gradient is bounded, but the $`Q_\\varepsilon`$-mass is only $`O_\\alpha(\\varepsilon)`$, giving the same order after integration. The restoration error is therefore $`O_\\alpha(\\varepsilon^3)`$, uniformly over all minimizing threshold pairs.\n\nApproximation by $`\\mathcal K_{\\rm fin}^c`$ in cost, paired risk, and the vague plan topology (Appendix A.4), followed by diagonal recovery, gives the general $`o(\\varepsilon^2)`$ error. Compactness, contraction toward the identity, and the common budget-Lipschitz bound then give locally uniform convergence of the optimized profiles.\n\n\n\n### 3.3 Initial linearity and the critical budget\n\n\n\nAt zero threshold shifts, the portfolios disagree on tail membership precisely on $`W=\\{u(u+H)<0\\}`$. The displacement $`\\zeta=-(u+H)1_W`$ moves a source in this wedge to the limiting $`A_\\varepsilon`$-boundary and attains the maximal paired hinge increment per unit of normal movement. When $`c_{\\min}=\\gamma_1`$, we can assign this movement along $`e_1`$ provided its downward and upward moments balance on each $`x`$-fiber. On active sources in $`W`$, $`\\zeta<0`$ for $`y0`$ for $`y>x`$, with $`\\int\\zeta\\,du=-H|H|/2`$. Writing $`k(x,y)=f(x,y,(v-x-y)/\\beta)1_{\\{\\beta\\ell_3\n\n$$\n\\ell_-(x)=\\frac1{2\\beta}\\int_I(x-y)_+^2k(x,y)\\,dy,\\qquad\n\\ell_+(x)=\\frac1{2\\beta}\\int_I(y-x)_+^2k(x,y)\\,dy.\n\\tag{3.7}\n$$\n\nExchangeability gives $`\\int_I\\ell_-=\\int_I\\ell_+=J/2`$, but the two moments need not agree on each fiber. Matching their pointwise minimum cancels the projected flux and gives the budget\n\n$$\n\\begin{aligned}\nq_{\\rm bal}(x)&=\\min\\{\\ell_-(x),\\ell_+(x)\\},&\nQ_{\\rm bal}&=\\int_I q_{\\rm bal}(x)\\,dx,\\\\\ns_{\\rm lin}&=2\\gamma_1Q_{\\rm bal}=\\gamma_1J(1-\\chi)>0,&\n\\chi&=\\|\\ell_--\\ell_+\\|_{L^1(I)}/J.\n\\end{aligned}\n$$\n\n\n\n**Proposition 3.2 (Capacity of the initial linear segment).** In the exchangeable box model of Section 2.2, suppose $`\\gamma_1=\\min\\{\\gamma_1,\\gamma_2,\\gamma_3/\\beta\\}`$. With $`\\ell_\\pm`$ as in [(3.7)](#dep-eq-3-7) and $`s_{\\rm lin}=\\gamma_1J(1-\\chi)>0`$ as above,\n\n$$\nF(s)=-J/p+s/(p\\gamma_1),\\qquad 0\\le s\\le s_{\\rm lin}.\n$$\n\n*Proof.* For $`W=\\{u(u+H)<0\\}`$ and $`\\zeta=-(u+H)1_W`$, select the proportion\n\n$$\na(\\xi)=\n\\begin{cases}\nq_{\\rm bal}(x)/\\ell_-(x),&H>0,\\ \\ell_-(x)>0,\\\\\nq_{\\rm bal}(x)/\\ell_+(x),&H<0,\\ \\ell_+(x)>0,\\\\\n0,&\\text{otherwise},\n\\end{cases}\n$$\n\nand define $`\\alpha^\\theta_{\\xi,u}=\\theta a(\\xi)\\delta_{\\zeta e_1}+(1-\\theta a(\\xi))\\delta_0`$ for $`0\\le\\theta\\le1`$. Because $`0\\le a\\le1`$ and the active sources satisfy $`|u|,|\\zeta|\\le\\Delta_I`$, this is a finite-first-moment probability kernel. Its opposite first-coordinate fluxes equal $`-\\theta q_{\\rm bal}(x)`$ and $`\\theta q_{\\rm bal}(x)`$, while the other coordinates remain fixed, so $`j_i=0`$ for all $`i`$ and $`I(\\alpha^\\theta)=\\mathcal C(\\alpha^\\theta)=2\\theta\\gamma_1Q_{\\rm bal}=\\theta s_{\\rm lin}`$.\n\nUnder this displacement, an active source has target normal losses $`-H`$ for $`B`$ and zero for $`A_\\varepsilon`$. Because $`\\operatorname{sgn}u=\\operatorname{sgn}(-H)\\ne0`$ on $`W`$, its strict $`B`$-tail membership is unchanged and no mass is placed at the $`B`$-threshold. For $`A_\\varepsilon`$, however, the active sources from the two signs of $`H`$ arrive at the kink itself. The resulting atom accounts for the jump between the one-sided derivatives: with $`\\mathfrak m_\\pm=\\int_{\\{\\pm H>0\\}}a(\\xi)|H|\\,d\\sigma`$,\n\n$$\nG'_{0,-}(0)=G'_{0,+}(0)=0,\\qquad\nG'_{1,-}(0)=-\\theta \\mathfrak m_-\\le0,\\qquad\nG'_{1,+}(0)=\\theta \\mathfrak m_+\\ge0.\n$$\n\nBy Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103), zero is globally optimal for both objectives. Evaluating the nonzero wedge integrals, $`H^2/2`$ for $`B`$ on $`H<0`$ and $`-H^2/2`$ for $`A_\\varepsilon`$ on $`H>0`$, we obtain $`T_0(0)=\\theta Q_{\\rm bal}`$ and $`T_1(0)=-\\theta Q_{\\rm bal}`$, which yield $`\\mathcal V(\\alpha^\\theta)=-J/p+\\theta s_{\\rm lin}/(p\\gamma_1)`$. Since $`0<\\chi<1`$ by Lemma [A.3](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L54), we may choose $`\\theta=s/s_{\\rm lin}`$ to attain the upper bound inherited from Proposition [2.2](#dep-prop-nominal-benchmark). ∎\n\n**Strict growth** To identify a common leading budget for a tie and a strict reversal, we improve every bounded-first-moment kernel at arbitrarily small additional cost. If the paired risk were concave in the kernel, mixing with a kernel of larger paired risk would do this directly, but the two threshold objectives are minimized separately and their difference need not be concave along mixtures. We also cannot assume that a kernel leaves stationary mass available for an added displacement, since an everywhere nonzero shift proportional to $`e^{-u^2}e_1`$ already has finite first moment.\n\nWe instead replace a small fraction of the existing kernel on two exchanged normal bands beyond the common threshold window. Reflecting $`u`$ to $`-u`$ crosses that window. Pairing the bands cancels their $`B`$-increments, while opposite signs of $`H`$ produce a positive paired gain per unit replacement fraction, independent of the distance $`U`$. After subtraction of the common baseline the old kernel contributes only an $`O_M(U^{-1})`$ remainder, because it takes a large displacement to affect the threshold window from those bands. So we first choose $`U`$ to make this remainder smaller than the reflection gain, and then reduce the replacement fraction to meet the additional budget. Appendix A.6 verifies the mixing obstruction and carries out this comparison uniformly over the original and modified minimizing thresholds.\n\n\n\n**Theorem 3.3 (Strict improvement by local replacement and a common critical coefficient).** Consider the exchangeable box model of [Section 2.2](#dep-model-critical-box). The limiting profile $`F`$ is strictly increasing and has a unique zero $`s^*>0`$, which gives the common leading coefficient of both pre-limit critical budgets at scale $`\\varepsilon^2`$. More precisely:\n\n(i) *Quantitative local improvement.* For every $`0\\le M<\\infty`$ there exists $`c_M>0`$ such that every kernel satisfying $`I(\\alpha)\\le M`$ and every $`0<\\delta\\le1`$ admit a feasible kernel $`\\widehat\\alpha`$ with\n\n\n\n$$\n\\mathcal C(\\widehat\\alpha)\\le\\mathcal C(\\alpha)+\\delta,\\qquad\n\\mathcal V(\\widehat\\alpha)\\ge\\mathcal V(\\alpha)+c_M\\delta.\n\\tag{3.8}\n$$\n\nThese kernels can be chosen as a family $`(\\widehat\\alpha_\\delta)_{0<\\delta\\le1}`$ satisfying\n\n$$\n\\int |d|_\\gamma\\,|\\widehat\\alpha_\\delta-\\alpha|(dd)\\,d\\nu\n\\xrightarrow[\\delta\\downarrow0]{}0,\n\\qquad\n\\lim_{\\delta\\downarrow0}\n\\frac{\\mathcal V(\\widehat\\alpha_\\delta)-\\mathcal V(\\alpha)}{\\delta}\n\\ge c_M>0.\n$$\n\nFor every finite $`S_0>0`$,\n\n\n\n$$\nF(t)-F(s)\\ge c_{S_0}(t-s)>0\\qquad(0\\le s0`$. Over the finite-first-moment kernel class,\n\n\n\n$$\ns^*=\\min_{\\alpha:\\,\\mathcal V(\\alpha)\\ge0}\\mathcal C(\\alpha)\n=\\min_{\\alpha:\\,\\mathcal V(\\alpha)=0}\\mathcal C(\\alpha)\n=\\inf_{\\alpha:\\,\\mathcal V(\\alpha)>0}\\mathcal C(\\alpha).\n\\tag{3.10}\n$$\n\nThe corresponding pre-limit critical budgets satisfy\n\n\n\n$$\n\\rho^-_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2),\\qquad\n\\rho^+_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2).\n\\tag{3.11}\n$$\n\nThe first two minima are attained, whereas the infimum over strictly positive risk values is not.\n\n\n\n**Execution costs** A positive execution cost proportional to $`\\varepsilon`$ exceeds the second-order nominal gain for small trades. To retain a critical-budget comparison, suppose instead that the deterministic cost paid to implement the new allocation $`B`$ satisfies $`c_{\\mathrm{exec},\\varepsilon}\\ge0`$ and $`c_{\\mathrm{exec},\\varepsilon}/\\varepsilon^2\\to\\theta\\in[0,J/p)`$. With the stress class and metric fixed, cash invariance gives $`D^{\\rm net}_\\varepsilon=D_\\varepsilon+c_{\\mathrm{exec},\\varepsilon}`$, whose rescaled profile converges locally uniformly to $`F+\\theta`$.\n\nApplying the tie and strict-reversal definitions in [(2.4)](#dep-eq-2-4) to the net profile, we obtain by strict growth and local uniform convergence $`\\rho^{\\pm,\\rm net}_\\varepsilon=s_\\theta\\varepsilon^2+o(\\varepsilon^2)`$, where $`s_\\theta`$ is the unique solution of $`F(s_\\theta)=-\\theta`$. As the rescaled execution cost $`\\theta`$ increases to $`J/p`$, this solution decreases continuously from $`s^*`$ to zero. For example, in the cheap-background case of Section 4.1, the explicit linear profile gives $`s_\\theta=c_3(J-p\\theta)`$.\n\n\n\n## 4. Efficiency premiums and projected flow duality\n\nThe critical coefficient reaches the local-efficiency lower bound exactly when background displacement is cheapest per unit change in portfolio loss, and in every other case a strict marginal-restoration premium remains. This section proves that classification and then characterizes the right derivative of the critical budget in the background penalty at equal effective costs. A locally efficient displacement need not preserve the marginal projections, so any unbalanced moment has to be compensated elsewhere. We measure this obstruction by testing each projected flux against a bounded measurable coordinate potential, using the $`L^1`$ dual representation of the restoration cost:\n\n$$\n\\mathcal C(\\alpha)-I(\\alpha)\n=\\sup_{|\\phi_i|\\le\\gamma_i}\\sum_i\\int_{I_i}\\phi_i(q)j_i(q)\\,dq\n=\\sup_{|\\phi_i|\\le\\gamma_i}\\int\\sum_i\\phi_i(\\xi_i)d_i\\,d\\Pi_\\alpha.\n$$\n\nHere $`\\phi_i(q)`$ acts as a dual shadow price for the projected displacement moment at coordinate level $`q`$. For the projections to vanish, opposite signed displacements have to balance at each coordinate level, and then the leading restoration penalty disappears. We first identify the pointwise displacement required for equality in the efficiency bound, and then test its projections near an off-diagonal corner to rule out equality when only traded coordinates are cheapest. At equal effective costs, continuous potentials lead to the flow–cut duality of Section 4.2, and allowing nonzero projected fluxes in the restoration-cost bound leads to the piecewise-constant potential estimates of Section 5.1.\n\n\n\n### 4.1 Local efficiency and the zero-premium classification\n\nWith $`c_1=\\gamma_1`$, $`c_2=\\gamma_2`$, $`c_3=\\gamma_3/\\beta`$, and $`c_{\\min}=\\min\\{c_1,c_2,c_3\\}`$, Proposition [2.2](#dep-prop-nominal-benchmark) and Theorem [3.1](#dep-thm-critical-limit) already give $`F(s)\\le-J/p+s/(pc_{\\min})`$. The issue at zero risk difference is whether a kernel can satisfy the pointwise efficiency conditions while conserving every marginal projection, and we characterize those simultaneous equality conditions below.\n\nProposition [2.2](#dep-prop-nominal-benchmark) gives initial efficiency at each fixed trade size under [(2.9)](#dep-eq-2-9), and when $`\\gamma_1=c_{\\min}`$, Proposition [3.2](#dep-prop-balanced-linear-profile) separately realizes the limiting slope by matching opposite first-coordinate moments on each fiber. When only traded coordinates are cheapest, that matching is enough for the initial linear segment but cannot cover the whole tail section. If the background coordinate is cheapest, exchanging $`x,y`$ and reversing $`u`$ instead balances the required displacement on every background slice, which sustains the efficiency bound up to zero risk.\n\n\n\n**Lemma 4.1 (Local-efficiency bound and critical-level optimality).** In the exchangeable box model of Section 2.2, every kernel with $`I(\\alpha)<\\infty`$ satisfies $`p\\mathcal V(\\alpha)+J\\le\\mathcal C(\\alpha)/c_{\\min}`$, so $`s^*\\ge c_{\\min}J`$. Set $`b=(1,1,\\beta)`$ and $`\\zeta=-(u+H)1_{\\{u(u+H)<0\\}}`$. Then $`s^*=c_{\\min}J`$ if and only if there is a measurable probability kernel satisfying\n\n\n\n$$\n\\begin{aligned}\nb\\cdot d=\\zeta,\\qquad |d|_\\gamma=c_{\\min}|\\zeta|\n&\\quad\\Pi_\\alpha\\text{-almost everywhere},\\\\\n(\\xi_i)_\\#(d_i\\,d\\Pi_\\alpha)=0\n&\\quad i=1,2,3.\n\\end{aligned}\n\\tag{4.1}\n$$\n\nSuch a kernel has $`\\mathcal C=I=c_{\\min}J`$ and $`\\mathcal V=0`$; every kernel with $`\\mathcal V\\ge0`$ and $`\\mathcal C\\le c_{\\min}J`$ satisfies these conditions.\n\n*Proof.* Set $`w=b\\cdot d`$, $`\\Delta_+(t,w)=(t+w)_+-t_+`$, and choose any $`\\tau_A\\in\\arg\\min G_1`$. To compare the risk gain with its transport cost, we evaluate the $`B`$-objective at zero, bound the difference of hinge increments by $`|w|`$, and use $`c_{\\min}|w|\\le |d|_\\gamma`$. Recording the loss in each comparison gives\n\n\n\n$$\n\\begin{aligned}\n\\frac{\\mathcal C}{c_{\\min}}-(p\\mathcal V+J)\n={}&\\underbrace{\\frac{\\mathcal C-I}{c_{\\min}}}_{\\mathrm{I}}\n+\\underbrace{\\left(\\frac I {c_{\\min}}-\\int |w|\\,d\\Pi_\\alpha\\right)}_{\\mathrm{II}}\n+\\underbrace{\\frac g2\\tau_A^2}_{\\mathrm{III}}\\\\\n&+\\underbrace{\\int[|w|-\\Delta_+(u,w)+\\Delta_+(u+H-\\tau_A,w)]\\,d\\Pi_\\alpha}_{\\mathrm{IV}}\\\\\n&+\\underbrace{G_0(0)-\\min_{\\tau_B}G_0(\\tau_B)}_{\\mathrm{V}}.\n\\end{aligned}\n\\tag{4.2}\n$$\n\nThe five terms (I)–(V) in [(4.2)](#dep-eq-4-2) have distinct roles: (I) the marginal-restoration penalty; (II) the excess transport cost over the cheapest normal movement, including cancellation between coordinate displacements; (III) the curvature cost of shifting the $`A_\\varepsilon`$-threshold; (IV) the unsaturated hinge gain; and (V) the gap from evaluating the $`B`$-objective at zero instead of its minimum. For the fourth term, $`\\Delta_+(t,w)\\in[\\min\\{0,w\\},\\max\\{0,w\\}]`$, an interval of length $`|w|`$, so the difference of the two hinge increments cannot exceed $`|w|`$. All five terms are consequently nonnegative, proving the efficiency bound. Since $`\\tau_A`$ was an arbitrary minimizer, equality forces each defect to vanish for every optimal $`A_\\varepsilon`$-threshold.\n\nTo reach zero risk at the local-efficiency budget, none of these defects can remain. Indeed, if $`\\mathcal V\\ge0`$ and $`\\mathcal C\\le c_{\\min}J`$, the left side of [(4.2)](#dep-eq-4-2) is nonpositive, so every nonnegative term must vanish and\n\n\n\n$$\n\\mathcal V=0,\\qquad \\mathcal C=I=c_{\\min}J,\\qquad\n\\int|w|\\,d\\Pi_\\alpha=J,\\qquad j_i=0\\quad(i=1,2,3).\n\\tag{4.3}\n$$\n\nIn addition, $`\\arg\\min G_1=\\{0\\}`$ and $`0\\in\\arg\\min G_0`$. The vanishing metric defect\n\n\n\n$$\n|d|_\\gamma-c_{\\min}|b\\cdot d|\n=\\sum_i(\\gamma_i-c_{\\min}b_i)|d_i|\n+c_{\\min}\\left(\\sum_i b_i|d_i|-\\left|\\sum_i b_id_i\\right|\\right)\n\\tag{4.4}\n$$\n\nallows only cheapest coordinates, with all nonzero components having the sign of $`w`$. At $`\\tau_A=0`$, equality in the hinge bound requires $`u\\ge0`$, $`u+H\\le-w`$ if $`w>0`$, and $`u\\le0`$, $`u+H\\ge-w`$ if $`w<0`$. Thus $`|w|\\le K:=|u+H|1_{\\{u(u+H)<0\\}}`$, with the sign of $`\\zeta`$ wherever $`w\\ne0`$. Since\n\n\n\n$$\n\\int_{\\mathbb R}K\\,du=H^2/2,\\qquad\n\\int_{\\mathbb R}\\zeta\\,du=-H|H|/2,\n\\tag{4.5}\n$$\n\nand every conditional kernel has mass one, $`\\int K\\,d\\Pi_\\alpha=J=\\int|w|\\,d\\Pi_\\alpha`$. Saturation of this nonnegative bound forces\n\n\n\n$$\nb\\cdot d=\\zeta\\quad\\Pi_\\alpha\\text{-almost everywhere},\n\\tag{4.6}\n$$\n\nwhich, with the metric and flux equalities, proves necessity.\n\nFor the converse, the threshold calculation depends only on the normal displacement $`w=\\zeta`$. We integrate over the disagreement wedge and use exchangeability to obtain\n\n\n\n$$\nT_0(0)=J/2,\\qquad T_1(0)=-J/2,\\qquad\n\\int_{H>0}H^2\\,d\\sigma=\\int_{H<0}H^2\\,d\\sigma=J.\n\\tag{4.7}\n$$\n\nBecause the transformed $`B`$-loss retains its strict tail sign and places no atom at zero, we have $`G_0'(0)=0`$. With $`m_\\pm=\\int(\\pm H)_+\\,d\\sigma>0`$, the mass placed at the $`A_\\varepsilon`$-threshold kink gives\n\n\n\n$$\nG_{1,-}'(0)=-m_-<0,\\qquad G_{1,+}'(0)=m_+>0.\n\\tag{4.8}\n$$\n\nBy convexity, these strict one-sided inequalities make zero the unique minimizer of $`G_1`$. For $`G_0`$, the stationary regions $`\\{H>0,u>0\\}`$ and $`\\{H<0,u<0\\}`$ each have section mass $`g/2`$, and their unchanged normal coordinates cover the positive and negative half-lines, respectively. Hence $`\\lambda_0\\ge(g/2)\\,dt`$ on $`\\mathbb R`$, and Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103) yields $`G_0(\\tau)-G_0(0)\\ge g\\tau^2/4`$, so zero is also the unique minimizer of $`G_0`$. Combining these minima with [(4.7)](#dep-eq-4-7) gives $`\\mathcal V=0`$. Under [(4.1)](#dep-eq-4-1), we have $`I=c_{\\min}J`$ and vanishing fluxes, so $`\\mathcal C=c_{\\min}J`$. Together with attainment of the critical cost in Theorem [3.3](#dep-thm-strict-crossing), this proves the equivalence. ∎\n\n\n\n**Theorem 4.2 (Zero-premium classification).** For the exchangeable box model of [Section 2.2](#dep-model-critical-box) and every positive weight vector,\n\n\n\n$$\ns^*=Jc_{\\min}\\quad\\Longleftrightarrow\\quad\nc_3\\le\\min\\{\\gamma_1,\\gamma_2\\}.\n\\tag{4.9}\n$$\n\nIf $`c_3=c_{\\min}`$, the profile is explicitly $`F(s)=-J/p+s/(pc_3)`$ for $`0\\le s\\le c_3J`$.\n\nIf $`\\gamma_2\\ge\\gamma_1`$ and $`c_3>\\gamma_1`$, then\n\n\n\n$$\n1<\\frac{s^*}{\\gamma_1J}\\le\\min\\{1+\\chi,c_3/\\gamma_1\\}<2,\n\\tag{4.10}\n$$\n\nwhere $`\\chi=\\|\\ell_--\\ell_+\\|_{L^1(I)}/J\\in(0,1)`$ is the projected imbalance of Section 3.3.\n\n*Proof of Theorem [4.2](#dep-thm-zero-premium).*\n\nBy Lemma [4.1](#dep-lem-efficiency-bound), it suffices to allocate the prescribed displacement $`\\zeta`$ among cheapest coordinates while preserving every marginal projection. The background coordinate admits such a balance because exchanging the traded losses and reversing the normal coordinate leaves each background slice fixed and reverses the required displacement. For the traded coordinates, the corresponding balance must instead be found along each $`x`$- or $`y`$-fiber. Near an off-diagonal corner, the required moment is bounded below by a positive multiple of $`\\delta^2`$, whereas the opposite-side sources available at those coordinate levels have capacity only $`O(\\delta^4)`$. We prove the background construction first and then use this discrepancy to exclude even randomized splitting among the traded directions.\n\n\n\n**Comparison kernels and the sufficient direction** Assigning the prescribed displacement to one coordinate gives\n\n\n\n$$\n\\alpha^{(1)}=\\delta_{\\zeta e_1},\\qquad\n\\alpha^{(2)}=\\delta_{\\zeta e_2},\\qquad\n\\alpha^{(3)}=\\delta_{(\\zeta/\\beta)e_3}\n\\tag{4.11}\n$$\n\nwith finite first moments and zero risk differences by the normal-displacement calculation [(4.7)](#dep-eq-4-7)–[(4.8)](#dep-eq-4-8) in Lemma [4.1](#dep-lem-efficiency-bound). The first two kernels have respective fluxes $`j_1=\\ell_+-\\ell_-`$ and $`j_2=\\ell_--\\ell_+`$, all other fluxes being zero. To verify the background projection, we use the exchange $`(x,y,u)\\mapsto(y,x,-u)`$, which preserves $`\\nu,\\xi_3`$ and reverses $`\\zeta`$. Absolute integrability then allows us to test against every bounded measurable $`\\varphi`$, obtaining\n\n\n\n$$\n\\int\\varphi(\\xi_3)d_3^{(3)}d\\nu\n=-\\int\\varphi(\\xi_3)d_3^{(3)}d\\nu=0.\n\\tag{4.12}\n$$\n\nAll three projections of $`\\alpha^{(3)}`$ vanish, yielding\n\n\n\n$$\n\\mathcal C(\\alpha^{(1)})=\\gamma_1J(1+\\chi),\\quad\n\\mathcal C(\\alpha^{(2)})=\\gamma_2J(1+\\chi),\\quad\n\\mathcal C(\\alpha^{(3)})=c_3J.\n\\tag{4.13}\n$$\n\nThe background kernel proves sufficiency, including ties, and the least of these costs gives $`s^*\\le J\\min\\{\\gamma_1(1+\\chi),\\gamma_2(1+\\chi),c_3\\}`$, which yields the upper bound in [(4.10)](#dep-eq-4-10) when $`\\gamma_2\\ge\\gamma_1`$ and $`c_3>\\gamma_1`$.\n\nWhen $`c_3=c_{\\min}`$, mix the background kernel $`\\alpha^{(3)}`$ with the identity using weight $`\\theta=s/(c_3J)\\in[0,1]`$. Both kernels have zero projected flux and both threshold objectives attain their minima at zero, by [(4.7)](#dep-eq-4-7)–[(4.8)](#dep-eq-4-8) and $`G_c^{\\delta_0}(\\tau)=g\\tau^2/2`$, so cost and paired risk are affine along the mixture. Its cost is $`\\theta c_3J=s`$, and [(4.7)](#dep-eq-4-7) gives\n\n\n\n$$\n\\mathcal V=-(1-\\theta)J/p.\n\\tag{4.14}\n$$\n\nThus the mixture attains the efficiency bound on the stated interval.\n\n\n\n**The obstruction for arbitrary randomized kernels** For necessity, suppose $`c_3>c_{\\min}`$ and a kernel with $`\\mathcal V\\ge0`$ attains cost $`c_{\\min}J`$. Lemma [4.1](#dep-lem-efficiency-bound) then requires $`d_3=0`$, $`d_1+d_2=\\zeta`$, and both traded components to have the sign of $`\\zeta`$. Since the kernel has finite first moment, its averaged displacement fields $`m_i(x,y)=\\int_{\\mathbb R}\\int d_i\\,\\alpha_{\\xi,u}(dd)\\,du`$, $`i=1,2`$, are integrable on the section and satisfy, by [(4.1)](#dep-eq-4-1),\n\n\n\n$$\n\\begin{aligned}\n&m_1+m_2=-H|H|/2,\\qquad |m_i|\\le H^2/2,\\\\\n&m_i\\ge0\\quad\\text{for}\\ xy.\n\\end{aligned}\n\\tag{4.15}\n$$\n\nProjected conservation gives, for every pair of Borel sets $`A,B\\subset I`$,\n\n\n\n$$\n\\int_{\\Omega\\cap\\{x\\in A\\}}m_1d\\sigma=0,\\qquad\n\\int_{\\Omega\\cap\\{y\\in B\\}}m_2d\\sigma=0.\n\\tag{4.16}\n$$\n\nWe test [(4.15)](#dep-eq-4-15)–[(4.16)](#dep-eq-4-16) near the upper-left off-diagonal corner of $`\\Omega`$, where $`x`$ is small, $`y`$ is large, and the positive moment $`(y-x)^2/2`$ is bounded away from zero. Compensating this moment at the same coordinate levels requires negative moment from $`\\{x>y\\}`$, where the small value of $`x`$ or large value of $`y`$ confines available sources to a small triangle near the diagonal (Figure [2](#dep-fig-corner)).\n\n\n\nThe section base is the intersection of an open square with a diagonal strip, $`\\Omega=(\\ell,r)^2\\cap\\{q_-0.\n\\end{gathered}\n$$\n\nBoth coordinate projections of $`\\Omega`$ equal $`(\\xi_-,\\xi_+)`$, and the interior diagonal point from [Section 2.3](#dep-geometry-interior-diagonal) satisfies $`\\xi_-0`$. Moreover, $`q_*:=\\xi_-+\\xi_+\\in[q_-,q_+]`$, so the corner $`(\\xi_-,\\xi_+)`$ lies in the closed strip, possibly on its boundary. The square diagonal through that corner runs parallel to the strip boundaries. On the corner square $`(x,y)=(\\xi_-+t,\\xi_+-z)`$, $`00`$, independent of $`\\delta`$, such that\n\n\n\n$$\nM_\\delta:=\\int_{E_\\delta}(m_1+m_2)d\\sigma\n\\ge c_\\Omega\\delta^2.\n\\tag{4.17}\n$$\n\nTesting [(4.16)](#dep-eq-4-16) on $`A_\\delta=(\\xi_-,\\xi_-+\\delta)`$ and $`B_\\delta=(\\xi_+-\\delta,\\xi_+)`$ requires $`M_\\delta\\le N_1+N_2`$, where\n\n$$\nN_1=\\int_{\\Omega\\cap\\{x\\in A_\\delta,\\ x>y\\}}(-m_1)d\\sigma,\n\\qquad\nN_2=\\int_{\\Omega\\cap\\{y\\in B_\\delta,\\ x>y\\}}(-m_2)d\\sigma.\n$$\n\nIn $`N_1`$, the conditions $`x\\in A_\\delta`$ and $`y\n\n$$\nN_1+N_2\\le C_\\Omega\\int_0^\\delta\\int_0^x(x-y)^2\\,dy\\,dx\n=\\frac{C_\\Omega}{12}\\delta^4.\n\\tag{4.18}\n$$\n\nHere $`C_\\Omega<\\infty`$ is independent of $`\\delta`$. For small $`\\delta`$, the bound contradicts $`c_\\Omega\\delta^2\\le M_\\delta\\le N_1+N_2`$. Attainment of the critical cost then implies $`s^*>Jc_{\\min}`$ when $`c_3>c_{\\min}`$, and [(4.13)](#dep-eq-4-13) gives [(4.10)](#dep-eq-4-10). ∎\n\n\n\n![Figure 2. Corner demand and return capacity, illustrated on $`I=I_3=[0,2]`$ for $`\\beta=7/4`$, $`v=15/4`$, $`\\delta=0.35`$. The blue region $`E_\\delta`$ stays away from the diagonal and has demand at least a positive multiple of $`\\delta^2`$. Return sources lie in the hatched triangles, which may extend beyond $`\\Omega`$; their area and squared displacement are both $`O(\\delta^2)`$, giving capacity $`O(\\delta^4)`$. The bounds shown, with $`\\lambda=\\min f`$, $`\\Lambda=\\max f`$, and $`\\Delta=\\xi_+-\\xi_-`$, violate $`M_\\delta\\le N_1+N_2`$ for sufficiently small positive $`\\delta`$.](/assets/research/dependence/corner-capacity.png)\n\n\n\n### 4.2 Boundary sensitivity and the flow–cut representation\n\nAt equal effective costs, every efficient kernel has total required displacement $`J`$. At this critical level, the equality conditions and total-moment saturation in Lemma [4.1](#dep-lem-efficiency-bound) fix $`b\\cdot d=\\zeta`$ pointwise almost everywhere, so normal integration and exchange symmetrization reduce the remaining minimum-background allocation problem to shares on the two-dimensional half-section. Raising only the background penalty selects, to first order, the kernels that use the least background displacement, and we can calculate that amount by assigning as much displacement as possible to the traded coordinates under their projected conservation law. On $`D=\\Omega\\cap\\{x\n\n**Theorem 4.3 (Boundary derivative and projected flow duality).** Fix the reference model and $`p,\\beta,\\gamma`$, and set\n\n$$\n\\boldsymbol\\gamma(\\eta)=(\\gamma,\\gamma,\\beta\\gamma(1+\\eta)),\\quad\n\\pi(\\eta)=s^*(\\boldsymbol\\gamma(\\eta))/(\\gamma J)-1,\\quad \\eta\\ge0.\n$$\n\nLet $`\\mathcal C_0`$ be the cost at $`\\eta=0`$, and let $`\\mathcal Z_0=\\{\\alpha:I(\\alpha)<\\infty,\\ \\mathcal V(\\alpha)=0,\\ \\mathcal C_0(\\alpha)=\\gamma J\\}`$. When kernels are identified with their measures $`\\Pi_\\alpha`$, the set $`\\mathcal Z_0`$ is nonempty and compact in the vague topology on $`\\mathscr X`$ from Section 3.2. The following minimum is attained, and\n\n\n\n$$\n\\pi(\\eta)=a_+\\eta+o(\\eta),\\qquad\na_+=\\frac\\beta J\\min_{\\alpha\\in\\mathcal Z_0}\n\\int|d_3|\\,d\\Pi_\\alpha\\in[0,1].\n\\tag{4.19}\n$$\n\nOn $`D=\\Omega\\cap\\{x\n\n$$\na_+=1-\\frac{2\\mathfrak F_\\mu}{J}\\in(0,1),\\qquad\n\\mathfrak F_\\mu=\\inf_{\\substack{\\varphi\\text{ measurable}\\\\|\\varphi|\\le1}}\n\\int_D\\max\\{0,1-\\varphi(x),1+\\varphi(y)\\}\\,d\\mu.\\tag{4.20}\n$$\n\nEquivalently, $`a_+`$ is the supremum of the projected cut deficits\n\n\n\n$$\na_+=\\frac2J\\sup_{U\\cap V=\\varnothing}\n\\left[\\mu(D\\cap(U\\times V))-\n\\mu\\bigl(D\\cap((V\\times I)\\cup(I\\times U))\\bigr)\\right].\\tag{4.21}\n$$\n\nHere $`U,V\\subset I`$ are disjoint measurable sets.\n\nIn words, $`2\\mathfrak F_\\mu`$ is the largest normal-displacement moment that the traded coordinates can accommodate while balancing their marginal projections. The cut formula [(4.21)](#dep-eq-4-21) writes $`a_+`$ as $`2/J`$ times the maximal deficit between the moment required on $`D\\cap(U\\times V)`$ and the compensating capacity on $`D\\cap((V\\times I)\\cup(I\\times U))`$.\n\n*Proof sketch.* For $`\\alpha\\in\\mathcal Z_0`$, normal integration and exchange symmetrization yield shares satisfying $`(x)_\\#(a\\mu)=(y)_\\#(b\\mu)`$ and background amount $`\\beta\\int|d_3|\\,d\\Pi_\\alpha=J-2\\int_D(a+b)\\,d\\mu`$. Conversely, every feasible pair of shares lifts to a kernel in $`\\mathcal Z_0`$, so the minimum background amount is $`J-2\\mathfrak F_\\mu`$ (Appendix [A.8](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-flow-proof.tex#L1)). Dividing by $`J`$ gives $`a_+`$, and Appendix [A.7](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-cost-proof.tex#L1) establishes [(4.19)](#dep-eq-4-19) by comparing perturbed minimizers with a least-background kernel. Sion's theorem with continuous coordinate potentials gives [(4.20)](#dep-eq-4-20); signed layer-cake reduction then gives [(4.21)](#dep-eq-4-21).\n\n\n\n**Dependence on the transport weights** Varying the positive transport weights leaves the finite-first-moment kernel class and $`\\mathcal V`$ unchanged. With $`A_i(\\alpha)=\\int|d_i|\\,d\\Pi_\\alpha+\\|j_i\\|_1`$, the coordinate cost including marginal restoration, we may therefore write\n\n\n\n$$\ns^*(\\gamma)=\\inf_{\\mathcal V(\\alpha)\\ge0}\\sum_i\\gamma_i A_i(\\alpha).\n\\tag{4.22}\n$$\n\nBecause the feasible class is independent of $`\\gamma`$, this infimum of nonnegative linear functions is positively homogeneous, coordinatewise nondecreasing, and concave. In particular, $`r_-\\gamma\\le\\widetilde\\gamma\\le r_+\\gamma`$, with $`0\\max\\{\\widetilde\\gamma_1,\\widetilde\\gamma_2\\}`$ and therefore retains a strict premium.\n\n\n\n## 5. Analytical illustrations and financial implications\n\n\n\n### 5.1 Projection potentials and a uniform reference\n\nAssigning a bounded shadow price $`\\phi_i`$ to each marginal flux lets us bound the strict premium without solving the full risk optimization. For a prescribed normal displacement we first minimize the adjusted coordinate costs $`\\gamma_i\\pm\\phi_i(\\xi_i)`$, and then we compare that minimum with the paired hinge gain.\n\n\n\n**Proposition 5.1 (A projection-potential bound).** In the exchangeable box model, let $`b=(1,1,\\beta)`$ and let $`|\\phi_i|\\le\\gamma_i`$ be bounded measurable potentials on the coordinate intervals such that $`k_\\pm(\\xi)=\\min_{1\\le i\\le3}(\\gamma_i\\pm\\phi_i(\\xi_i))/b_i`$ are bounded below by a positive constant. For every finite-first-moment kernel, every $`\\tau\\in\\arg\\min G_1`$, and every $`\\lambda>0`$,\n\n\n\n$$\n\\mathcal C(\\alpha)\\ge\n\\lambda\\left(J+p\\mathcal V+\\frac g2\\tau^2\\right)\n-\\frac{\\lambda}{2}\\int_\\Omega(H-\\tau)^2\n\\left(\\frac{\\lambda}{k(\\xi,H-\\tau)}-1\\right)_+\\,d\\sigma,\n\\tag{5.1}\n$$\n\nwhere $`k(\\xi,h)=k_+(\\xi)`$ for $`h<0`$ and $`k_-(\\xi)`$ for $`h>0`$. The integrand is zero when $`h=0`$.\n\nThe proof in [Appendix B.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L3) optimizes the hinge gain against the adjusted cost $`k_+w_++k_-(-w)_+`$, where $`w=b\\cdot d`$. To quantify how far the reversal budget can exceed the initial-slope prediction, we use an analytically tractable reference with independent uniform losses on $`[0,2]`$, $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(\\gamma_0,\\gamma_0,2\\gamma_0)`$. Equal exposures and the median tail make the section invariant under central reflection, so $`v=3`$, $`d\\sigma=dx\\,dy/8`$, $`g=3/8`$, and $`J=5/32`$.\n\nIn the cut representation of Theorem [4.3](#dep-prop-boundary-sensitivity), the midpoint partition corresponds to $`\\varphi(q)=1_{[0,1)}(q)-1_{(1,2]}(q)`$. Scaling this potential by $`\\gamma_0/2`$ and using opposite signs on the two traded coordinates carries the same partition over to the marginal-flux penalty in Proposition [5.1](#dep-prop-projection-potential-bound), giving the admissible two-level potentials\n\n$$\n\\phi_1(x)=\\frac{\\gamma_0}{2}\\operatorname{sgn}(1-x),\\qquad\n\\phi_2(y)=\\frac{\\gamma_0}{2}\\operatorname{sgn}(y-1),\\qquad \\phi_3=0.\n$$\n\nTheir signs distinguish low-loss from high-loss coordinate halves. At zero threshold shift, a positive paired hinge gain requires normal movement $`w=b\\cdot d>0`$ when $`xy`$. With amplitude $`\\gamma_0/2`$, the effective cost in this direction across the midpoint is $`k(\\xi,H)=3\\gamma_0/2`$, and both directional costs on either same-side region are $`\\gamma_0/2`$. Choosing $`\\lambda=3\\gamma_0/2`$ in [(5.1)](#dep-eq-5-1) therefore removes the cross-region penalty at zero threshold. The proof below keeps the correction for shifted minimizing thresholds, which makes the bound valid without prescribing either ES threshold.\n\n\n\n**Corollary 5.2 (Uniform reference with costly background displacement).** For this uniform reference,\n\n\n\n$$\n\\frac{13}{64}\\le\\frac{s^*}{\\gamma_0}\\le\\frac{29}{96},\n\\qquad \\frac{s^*}{\\gamma_0J}\\ge\\frac{13}{10}.\n\\tag{5.2}\n$$\n\nThe initial linear profile holds on $`0\\le s\\le\\gamma_0/96`$.\n\n*Proof.* For $`1/2\\le x\\le1`$, the negative-moment section has $`1-x\n\nSetting $`\\gamma_0=1`$, we obtain analytical envelopes for $`F(s)`$ on $`[0,U]`$, where $`U=29/96`$. The balanced kernel has cost $`s_0=1/96`$, zero projected flux, and risk $`-7/24`$, and the first-coordinate comparison kernel has cost $`U`$ and risk zero. Their convex mixtures have affine paired risk because zero minimizes both threshold objectives for each kernel, and affine cost because the balanced kernel has zero flux. Joining these mixtures to the exact initial linear segment realizes the lower envelope\n\n$$\n\\underline F(s)=\n\\begin{cases}\n2s-5/16,&0\\le s\\le1/96,\\\\\ns-29/96,&1/96\\le s\\le29/96.\n\\end{cases}\n$$\n\nTo bound $`F`$ from above, we invert the two threshold estimates and combine them with the efficiency bound:\n\n$$\n\\begin{aligned}\nF(s)&\\le\\min\\left\\{2s-\\frac5{16},\n\\max\\left\\{\\frac43s-\\frac{13}{48},2s-\\frac{11}{16}\\right\\}\\right\\}\n\\\\\n&=\\overline F(s):=\\min\\left\\{2s-\\frac5{16},\\frac43s-\\frac{13}{48}\\right\\}\n\\qquad(0\\le s\\le U).\n\\end{aligned}\n$$\n\nThe last equality holds because the branches inside the maximum meet only at $`s=5/8>U`$. Figure [3](#dep-fig-budget) plots the analytical envelopes $`\\underline F\\le F\\le\\overline F`$ on $`[0,U]`$. The guaranteed linear segment $`[0,s_{\\rm lin}]`$ covers one fifteenth of the tangent prediction $`J=5/32`$, whereas the critical coefficient $`s^*`$ lies between $`13/10`$ and $`29/15`$ times $`J`$. So the initial slope is attained on a positive interval, and the required reversal budget is still at least $`30\\%`$ higher than its tangent prediction.\n\n\n\n![Figure 3. Proved risk envelopes for the independent uniform model with $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(1,1,2)`$. Explicit mixtures attain the lower envelope; the upper envelope follows from midpoint potentials and the efficiency bound. The inset shows the exact initial interval $`[0,1/96]`$ in green. The dashed line extrapolates its slope to $`J=5/32`$, whereas the critical budget lies in $`[13/64,29/96]`$. Both plotted envelopes are piecewise linear analytical formulas.](/assets/research/dependence/budget-comparison.png)\n\n**Numerical approximation.** The computational companion also discretizes a symmetric family of local kernels, using pure-coordinate moves to the two hinge breakpoints and retaining the absolute projected-flux penalty in [(3.1)](#dep-eq-3-1). With both thresholds centered, surface grids of 8, 12, 16, 24 and 32 points per coordinate place the zero of this family near $`0.23`$; the individual estimates increase with the grid from $`0.2292`$ to $`0.2324`$. The calculation reproduces the proved initial linear segment, and the paired risk of each computed kernel, evaluated again at its own minimizing thresholds, vanishes to rounding error. Because the family lies inside the admissible class, its zero is an upper estimate of $`s^*`$, subject to discretization error. On this evidence $`s^*`$ falls in the lower third of the interval in [(5.2)](#dep-eq-5-2), and the reversal premium over the tangent prediction lies between the proved $`30\\%`$ and about $`49\\%`$. The upper figure is numerical and is not a proof; normal-window checks and solver residuals are recorded in the companion's numerical study.\n\nFor finite trade sizes, the companion solves the discretized problem globally on cubic midpoint lattices with up to 24 points per marginal, for trade sizes from $`\\varepsilon=0.2`$ down to $`\\varepsilon=0.025`$. These lattices do not resolve the quadratic scale as the trade size decreases, so the recorded values measure discretization sensitivity, and a convergence study of [(3.6)](#dep-eq-3-6) requires finer grids.\n\n\n\n### 5.2 A dependence family with fixed marginal laws\n\nSection 5.1 used an independent uniform reference with background displacement twice as costly as either traded direction. We now move to the equal-effective-cost boundary to see how changing the reference dependence affects the required background share $`a_+`$ and the budget derivative $`\\gamma_0Ja_+`$. Keeping uniform marginals, $`\\beta=1`$, and $`p=1/2`$, we perturb the independent density by the product of the centered losses $`x+y-2`$ and $`z-1`$:\n\n$$\nf_\\vartheta(x,y,z)=\\frac{1+\\vartheta(x+y-2)(z-1)}8,\\qquad\n|\\vartheta|<\\frac12,\\qquad\nE_\\vartheta[(X_1+X_2-2)(X_3-1)]=\\frac{2\\vartheta}{9}.\n$$\n\nAfter rescaling the losses to $`[0,1]`$, this is a one-parameter subfamily of the trivariate Farlie–Gumbel–Morgenstern family [[27](#ref-otakimura2021), Section 2]: the interactions between each traded coordinate and the background have parameter $`\\vartheta`$, and the other interaction parameters are zero. Centering preserves every marginal, and the parameter restriction keeps the density strictly positive on the cube. Central reflection fixes $`v=3`$, where the section density is $`[1-\\vartheta(x+y-2)^2]/8`$, so increasing $`\\vartheta`$ reduces the section mass away from $`x+y=2`$ even as covariance increases. On the unchanged half-section $`D`$, let $`d\\mu_\\vartheta=w_{\\sigma,\\vartheta}(x,y)\\,dx\\,dy`$, where $`w_{\\sigma,\\vartheta}(x,y)=(y-x)^2\\{1-\\vartheta(x+y-2)^2\\}/16`$. Set $`D_-=D\\cap\\{y<1\\}`$, $`D_+=D\\cap\\{x>1\\}`$, and denote the same-side masses by $`M_{\\vartheta,\\pm}=\\mu_\\vartheta(D_\\pm)`$.\n\n\n\n**Proposition 5.3 (Background share along the dependence family).** For $`|\\vartheta|<1/2`$, the boundary flow is attained and\n\n\n\n$$\nJ_\\vartheta=\\frac{75-14\\vartheta}{480},\\qquad\nM_{\\vartheta,-}=M_{\\vartheta,+}=\\frac{3-2\\vartheta}{1152},\n\\qquad\na_{+,\\vartheta}\n=1-\\frac{4(M_{\\vartheta,-}+M_{\\vartheta,+})}{J_\\vartheta}.\n\\tag{5.3}\n$$\n\nIn particular,\n\n$$\nJ_\\vartheta a_{+,\\vartheta}=\\frac{195-22\\vartheta}{1440},\\qquad\n\\frac{d a_{+,\\vartheta}}{d\\vartheta}\n=\\frac{360}{(75-14\\vartheta)^2}>0,\\qquad\n\\frac{d(J_\\vartheta a_{+,\\vartheta})}{d\\vartheta}\n=-\\frac{11}{720}.\n$$\n\nTo obtain the explicit value, we use all the mass in $`D_-`$ and $`D_+`$ and match their projections in two disjoint parts of the cross-region. The section inequalities in [Appendix B.3](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L49) give the required capacity throughout $`|\\vartheta|<1/2`$, and the midpoint potential gives the same value in the dual problem, so this feasible flow is optimal.\n\nAs reference dependence varies through $`0<\\vartheta<1/2`$, the minimum background share and the corresponding unnormalized displacement moment move in opposite directions:\n\n\n\n$$\na_{+,\\vartheta}-a_{+,0}\n=\\frac{24\\vartheta}{5(75-14\\vartheta)}>0,\\qquad\nJ_\\vartheta a_{+,\\vartheta}-J_0a_{+,0}\n=-\\frac{11\\vartheta}{720}<0.\n\\tag{5.4}\n$$\n\nThe contrast comes from the median plane $`x+y+z=3`$, on which the centered factors satisfy $`z-1=-(x+y-2)`$. Increasing $`\\vartheta`$ thus favors equal signs in the full box but reduces the section density away from $`x+y=2`$ by the factor $`1-\\vartheta(x+y-2)^2`$. Across both exchanged half-sections, the traded-coordinate matching capacity is $`4(M_{\\vartheta,-}+M_{\\vartheta,+})`$. By [(5.3)](#dep-eq-5-3), $`J_\\vartheta`$ decreases faster in absolute units than this capacity, so their difference $`J_\\vartheta a_{+,\\vartheta}`$ declines. In proportion, though, the matching capacity falls faster than $`J_\\vartheta`$ throughout $`0<\\vartheta<1/2`$, which leaves a larger relative share for the background coordinate.\n\nFigure [4](#dep-fig-flow-shares) compares two examples: independence ($`\\vartheta=0`$, model U) and $`\\vartheta=1/4`$ (model D). The triples $`(J_\\vartheta,a_{+,\\vartheta},J_\\vartheta a_{+,\\vartheta})`$ are $`(5/32,13/15,13/96)`$ and $`(143/960,379/429,379/2880)`$, so the minimum background share rises from about $`0.8667`$ to $`0.8834`$ while its unnormalized displacement moment falls from about $`0.1354`$ to $`0.1316`$.\n\n\n\n; all use the same scale from zero to one. On each same-side section we allocate displacement to a rebalanced coordinate and match its projection using the cross-region. We allocate the remainder to the background coordinate. Blank regions lie outside $`D`$.](/assets/research/dependence/optimal-flow-shares.png)\n\n\n\n### 5.3 Implications for reverse stress testing and further questions\n\nAn exact initial sensitivity need not predict how much dependence stress is required to reverse a diversification trade. This bears on reverse sensitivity analysis [[28](#ref-pesenti2019), [29](#ref-pesenti2022)], where the quantity of interest is the distance to a decision reversal. Under the traded-coordinate dominance condition [(2.9)](#dep-eq-2-9), $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon`$ extrapolates an attained initial slope, yet if $`\\gamma_1=c_{\\min}1.\n$$\n\nProposition [3.2](#dep-prop-balanced-linear-profile) attains the initial slope by balancing opposite movements on selected fibers. Equality in the efficiency bound at reversal would require balancing cheapest-coordinate projections across the entire tail section, and under the stated cost condition the corner obstruction of Theorem [4.2](#dep-thm-zero-premium) excludes that. Unlike optimal allocation across full Fréchet or Wasserstein ambiguity sets [[23](#ref-liu2026)], our reverse stress test fixes a proposed trade and measures the least perturbation of a common reference law that reverses its paired ES ranking.\n\nA whole-book comparison across reference dependence structures also has to account for the size of the nominal advantage. In the family of Section 5.2, consider the boundary perturbation $`\\boldsymbol\\gamma(\\eta)=(\\gamma_0,\\gamma_0,\\gamma_0(1+\\eta))`$. The gain-normalized critical budget and its right derivative at equal effective costs satisfy\n\n$$\n\\mathcal R_\\vartheta(\\eta):=\\lim_{\\varepsilon\\downarrow0}\n\\frac{\\rho^\\pm_{\\varepsilon,\\vartheta}(\\eta)}{m_{\\varepsilon,\\vartheta}}\n=\\frac{p\\,s^*_\\vartheta(\\eta)}{J_\\vartheta},\\qquad\n\\mathcal R_\\vartheta'(0+)=p\\gamma_0a_{+,\\vartheta}.\n$$\n\nTheorem [4.3](#dep-prop-boundary-sensitivity) gives the absolute right derivative $`(s^*_\\vartheta)'(0+)=\\gamma_0J_\\vartheta a_{+,\\vartheta}`$. By [(5.4)](#dep-eq-5-4), increasing $`\\vartheta`$ raises the gain-normalized response $`\\mathcal R_\\vartheta'(0+)`$ but lowers the absolute response: $`\\frac{d}{d\\vartheta}(\\gamma_0J_\\vartheta a_{+,\\vartheta})=\\gamma_0(J_\\vartheta'a_{+,\\vartheta}+J_\\vartheta a_{+,\\vartheta}')<0`$, so the decline in the nominal-gain scale outweighs the increase in background share. Reporting both boundary sensitivities to a risk committee separates a greater relative need for background displacement from a smaller absolute response in budget units.\n\nCharacterizing the boundaries between the active and stationary regions of an optimal kernel could help determine the full profile $`F`$ between the analytical envelopes, beyond the guaranteed initial linear interval. A uniform rate for $`\\sup_{s\\le S_0}|\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2)-F(s)|`$, combined with the lower growth estimate in Theorem [3.3](#dep-thm-strict-crossing), would yield finite-trade error bounds for the critical budgets.\n\nIn the compact-box benchmark, a restoration premium can arise from dependence geometry and full marginal conservation even with bounded marginal tails. For unbounded marginals such as integrable Pareto or lognormal laws, the inverse-CDF derivative is unbounded near probability one, so the uniform displacement estimate in the recovery argument would need a tail-weighted substitute. Since one-dimensional $`W_1`$ remains the $`L^1`$ integral of the CDF discrepancy, this loss of a uniform quantile bound does not by itself imply a different restoration-cost exponent. The open problem is to control that integral jointly with the paired-risk error and determine whether the $`\\varepsilon^2`$ critical scale persists under suitable tail assumptions. Higher-dimensional projected-flow representations and estimation of critical budgets from loss data raise related questions, alongside recent work on ES estimation under weak tail assumptions [[3](#ref-bartleckstein2026)].\n\n## Appendices and Supporting Material\n\nThis page includes the complete abstract, five main sections, all four main-text figures, and references. Technical appendices and declarations are available in the [full manuscript](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/paper.pdf) and the [GitHub repository](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals).\n\n## References\n\n\n\n1. P. Artzner, F. Delbaen, J.-M. Eber, and D. Heath, *Coherent measures of risk*, Math. Finance, 9(3) (1999), pp. 203–228, [doi:10.1111/1467-9965.00068](https://doi.org/10.1111/1467-9965.00068).\n\n\n\n2. D. Bartl, S. Drapeau, J. Obłój, and J. Wiesel, *Sensitivity analysis of Wasserstein distributionally robust optimization problems*, Proc. R. Soc. A, 477(2256) (2021), article 20210176, [doi:10.1098/rspa.2021.0176](https://doi.org/10.1098/rspa.2021.0176).\n\n\n\n3. D. Bartl and S. Eckstein, *Optimal nonparametric estimation of the expected shortfall risk*, preprint (2026), to appear in SIAM J. Financial Math., [arXiv:2405.00357v2](https://arxiv.org/abs/2405.00357v2).\n\n\n\n4. D. Bartl and J. Wiesel, *Sensitivity of multiperiod optimization problems with respect to the adapted Wasserstein distance*, SIAM J. Financial Math., 14(2) (2023), pp. 704–720, [doi:10.1137/22M1537746](https://doi.org/10.1137/22M1537746).\n\n\n\n5. C. Bernard, X. Jiang, and R. Wang, *Risk aggregation with dependence uncertainty*, Insur. Math. Econ., 54 (2014), pp. 93–108, [doi:10.1016/j.insmatheco.2013.11.005](https://doi.org/10.1016/j.insmatheco.2013.11.005).\n\n\n\n6. D. P. Bertsekas, *Convex Optimization Theory*, Athena Scientific, Belmont, MA, 2009, [author's online edition](https://web.mit.edu/dimitrib/www/Convex_Theory_Entire_Book.pdf).\n\n\n\n7. J. Blanchet and K. Murthy, *Quantifying distributional model risk via optimal transport*, Math. Oper. Res., 44(2) (2019), pp. 565–600, [doi:10.1287/moor.2018.0936](https://doi.org/10.1287/moor.2018.0936).\n\n\n\n8. M. Brandtner, *Expected shortfall, spectral risk measures, and the aggravating effect of background risk, or: risk vulnerability and the problem of subadditivity*, J. Bank. Financ., 89 (2018), pp. 138–149, [doi:10.1016/j.jbankfin.2018.02.002](https://doi.org/10.1016/j.jbankfin.2018.02.002).\n\n\n\n9. J. Chen, E. Furman, and X. S. Lin, *Static marginal expected shortfall: Systemic risk measurement under dependence uncertainty*, Insur. Math. Econ., 130 (2026), article 103285, [doi:10.1016/j.insmatheco.2026.103285](https://doi.org/10.1016/j.insmatheco.2026.103285).\n\n\n\n10. Y. Chen, P. Liu, Y. Liu, and R. Wang, *Ordering and inequalities for mixtures on risk aggregation*, Math. Finance, 32(1) (2022), pp. 421–451, [doi:10.1111/mafi.12323](https://doi.org/10.1111/mafi.12323).\n\n\n\n11. G. Dal Maso, *An Introduction to $`\\Gamma`$-Convergence*, Progress in Nonlinear Differential Equations and Their Applications 8, Birkhäuser, Boston, 1993, [doi:10.1007/978-1-4612-0327-8](https://doi.org/10.1007/978-1-4612-0327-8).\n\n\n\n12. C. De Vecchi, M. Nendel, and J. Streicher, *Upper comonotonicity and risk aggregation under dependence uncertainty*, Math. Finance, 36(1) (2026), pp. 118–139, [doi:10.1111/mafi.70008](https://doi.org/10.1111/mafi.70008).\n\n\n\n13. S. Eckstein, M. Kupper, and M. Pohl, *Robust risk aggregation with neural networks*, Math. Finance, 30(4) (2020), pp. 1229–1272, [doi:10.1111/mafi.12280](https://doi.org/10.1111/mafi.12280).\n\n\n\n14. S. Eckstein, G. Guo, T. Lim, and J. Obłój, *Robust pricing and hedging of options on multiple assets and its numerics*, SIAM J. Financial Math., 12(1) (2021), pp. 158–188, [doi:10.1137/19M1286256](https://doi.org/10.1137/19M1286256).\n\n\n\n15. P. Embrechts, G. Puccetti, and L. Rüschendorf, *Model uncertainty and VaR aggregation*, J. Bank. Financ., 37(8) (2013), pp. 2750–2764, [doi:10.1016/j.jbankfin.2013.03.014](https://doi.org/10.1016/j.jbankfin.2013.03.014).\n\n\n\n16. P. Embrechts, B. Wang, and R. Wang, *Aggregation-robustness and model uncertainty of regulatory risk measures*, Finance Stoch., 19(4) (2015), pp. 763–790, [doi:10.1007/s00780-015-0273-z](https://doi.org/10.1007/s00780-015-0273-z).\n\n\n\n17. L.-B. Fiechtner and J. Blanchet, *Wasserstein distributionally robust regret optimization*, preprint, 2025, revised 2026, [arXiv:2504.10796v5](https://arxiv.org/abs/2504.10796v5).\n\n\n\n18. R. Gao and A. J. Kleywegt, *Data-driven robust optimization with known marginal distributions*, working paper, 2017, [https://faculty.mccombs.utexas.edu/rui.gao/copula.pdf](https://faculty.mccombs.utexas.edu/rui.gao/copula.pdf).\n\n\n\n19. R. Gao and A. J. Kleywegt, *Distributionally robust stochastic optimization with Wasserstein distance*, Math. Oper. Res., 48(2) (2023), pp. 603–655, [doi:10.1287/moor.2022.1275](https://doi.org/10.1287/moor.2022.1275).\n\n\n\n20. J. Korman, R. J. McCann, and C. Seis, *Dual potentials for capacity constrained optimal transport*, Calc. Var. Partial Differential Equations, 54 (2015), pp. 573–584, [doi:10.1007/s00526-014-0795-9](https://doi.org/10.1007/s00526-014-0795-9).\n\n\n\n21. H. Lam, *Sensitivity to serial dependency of input processes: A robust approach*, Manage. Sci., 64(3) (2018), pp. 1311–1327, [doi:10.1287/mnsc.2016.2667](https://doi.org/10.1287/mnsc.2016.2667).\n\n\n\n22. L. Li, H. Shao, R. Wang, and J. Yang, *Worst-case Range Value-at-Risk with partial information*, SIAM J. Financial Math., 9(1) (2018), pp. 190–218, [doi:10.1137/17M1126138](https://doi.org/10.1137/17M1126138).\n\n\n\n23. P. Liu and Y. Liu, *Portfolio diversification and concentration under dependence uncertainty: A majorization approach*, preprint, 2026, [arXiv:2609.04496v1](https://arxiv.org/abs/2609.04496v1).\n\n\n\n24. R. J. Martin, *Saddlepoint methods in portfolio theory*, in The Oxford Handbook of Credit Derivatives, A. Lipton and A. Rennie, eds., Oxford University Press, Oxford (2011), pp. 533–569, [doi:10.1093/oxfordhb/9780199546787.013.0015](https://doi.org/10.1093/oxfordhb/9780199546787.013.0015).\n\n\n\n25. A. J. McNeil, R. Frey, and P. Embrechts, *Quantitative Risk Management: Concepts, Techniques and Tools*, revised ed., Princeton University Press, Princeton, NJ, 2015.\n\n\n\n26. H. Namkoong, *Local sensitivity under transport restrictions*, preprint, 2026, [arXiv:2606.04276v1](https://arxiv.org/abs/2606.04276v1).\n\n\n\n27. S. Ota and M. Kimura, *Effective estimation algorithm for parameters of multivariate Farlie–Gumbel–Morgenstern copula*, Jpn. J. Stat. Data Sci., 4 (2021), pp. 1049–1078, [doi:10.1007/s42081-021-00118-y](https://doi.org/10.1007/s42081-021-00118-y).\n\n\n\n28. S. M. Pesenti, P. Millossovich, and A. Tsanakas, *Reverse sensitivity testing: What does it take to break the model?*, European J. Oper. Res., 274(2) (2019), pp. 654–670, [doi:10.1016/j.ejor.2018.10.003](https://doi.org/10.1016/j.ejor.2018.10.003).\n\n\n\n29. S. M. Pesenti, *Reverse sensitivity analysis for risk modelling*, Risks, 10(7) (2022), article 141, [doi:10.3390/risks10070141](https://doi.org/10.3390/risks10070141).\n\n\n\n30. S. M. Pesenti and S. Jaimungal, *Portfolio optimization within a Wasserstein ball*, SIAM J. Financial Math., 14(4) (2023), pp. 1175–1214, [doi:10.1137/22M1496803](https://doi.org/10.1137/22M1496803).\n\n\n\n31. G. Ch. Pflug and M. Pohl, *A review on ambiguity in stochastic portfolio optimization*, Set-Valued Var. Anal., 26(4) (2018), pp. 733–757, [doi:10.1007/s11228-017-0458-z](https://doi.org/10.1007/s11228-017-0458-z).\n\n\n\n32. R. T. Rockafellar and S. Uryasev, *Conditional value-at-risk for general loss distributions*, J. Bank. Financ., 26(7) (2002), pp. 1443–1471, [doi:10.1016/S0378-4266(02)00271-6](https://doi.org/10.1016/S0378-4266%2802%2900271-6).\n\n\n\n33. L. Rüschendorf, *Mathematical Risk Analysis: Dependence, Risk Bounds, Optimal Allocations and Portfolios*, Springer Series in Operations Research and Financial Engineering, Springer, Berlin, Heidelberg, 2013, [doi:10.1007/978-3-642-33590-7](https://doi.org/10.1007/978-3-642-33590-7).\n\n\n\n34. F. Santambrogio, *Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling*, Progress in Nonlinear Differential Equations and Their Applications 87, Birkhäuser, Cham, 2015, [doi:10.1007/978-3-319-20828-2](https://doi.org/10.1007/978-3-319-20828-2).\n\n\n\n35. N. Sauldubois, *Higher-order sensitivity analysis of Wasserstein and adapted Wasserstein distributionally robust optimization problems*, preprint, 2026, [arXiv:2609.25345v1](https://arxiv.org/abs/2609.25345v1).\n\n\n\n36. O. Scaillet, *Nonparametric estimation and sensitivity analysis of expected shortfall*, Math. Finance, 14(1) (2004), pp. 115–129, [doi:10.1111/j.0960-1627.2004.00184.x](https://doi.org/10.1111/j.0960-1627.2004.00184.x).\n\n\n\n37. M. Sion, *On general minimax theorems*, Pacific J. Math., 8(1) (1958), pp. 171–176, [doi:10.2140/pjm.1958.8.171](https://doi.org/10.2140/pjm.1958.8.171).\n\n\n\n38. D. Tasche, *Conditional expectation as quantile derivative*, preprint, 2001, [arXiv:math/0104190v1](https://arxiv.org/abs/math/0104190v1).\n\n\n\n39. S. S. Vallender, *Calculation of the Wasserstein distance between probability distributions on the line*, Theory Probab. Appl., 18(4) (1974), pp. 784–786, [doi:10.1137/1118101](https://doi.org/10.1137/1118101).\n\n\n\n40. C. Villani, *Topics in Optimal Transportation*, Graduate Studies in Mathematics 58, American Mathematical Society, Providence, RI, 2003.\n\n\n\n41. R. Wang, V. Bignozzi, and A. Tsanakas, *How superadditive can a risk measure be?*, SIAM J. Financial Math., 6(1) (2015), pp. 776–803, [doi:10.1137/140981046](https://doi.org/10.1137/140981046).\n"
+ "sourceEditedAt": "2026-11-19T11:43:47-05:00",
+ "sourceCommit": "c0782609a95f9887c248631c56174615233c7ae4",
+ "sourceBlob": "72d5cf95d6812da8198cca12f313c59d254ef3d0",
+ "markdown": "## Abstract\n\nHow much dependence uncertainty can a diversification benefit withstand when every marginal loss law is fixed? We study the smallest weighted $`1`$-Wasserstein perturbation that reverses the Expected Shortfall ranking of two nearby allocations evaluated under a common stressed law. For a positive continuous density on a compact box, exchangeability of the two traded positions eliminates the first-order gain, so the relevant stress radius is $`\\rho=s\\varepsilon^2`$ for trade size $`\\varepsilon`$. At this scale the limiting risk profile is continuous and strictly increasing, and its unique zero identifies the asymptotic critical budget. We call the excess over the local-efficiency lower bound the *marginal-restoration premium*. This premium vanishes exactly when background displacement is cheapest per unit change in portfolio loss. Otherwise it is strictly positive, even when the profile is exactly linear on an initial interval: in a uniform benchmark that attains the initial slope, the proved lower bound on the critical budget is already $`30\\%`$ above the local-efficiency prediction. At equal effective costs, continuous maximum-flow/minimum-cut duality characterizes the right derivative of the critical budget with respect to the background penalty. The proof rests on a $`\\Gamma`$-limit for the rescaled transport energies and on convergence of the paired risk differences, with the two ES thresholds optimized separately. Coordinatewise quantile restoration incurs a marginal-flux cost, although its leading effects on the two ES values cancel in their difference.\n\n\n\n## 1. Introduction\n\nHow much of a portfolio's diversification benefit survives uncertainty about dependence? A risk desk may have agreed marginal models for its individual positions and still lack an equally firm model for their joint losses with the background book [[31](#ref-pflugpohl2018)]. We ask how a proposed reallocation performs when these marginal models are kept but the coupling changes. More precisely, we look for the smallest dependence perturbation that reverses its Expected Shortfall (ES) advantage when both allocations face the same stressed law. We show that this critical perturbation is of second order in the size of the trade, and that the response at zero budget can understate it: in a uniform benchmark the proved lower bound is already $`30\\%`$ above the prediction from the initial slope.\n\nLet $`X_1,X_2`$ be the unit losses of two positions of equal initial cost. A transfer from holdings $`(1+\\varepsilon,1-\\varepsilon)`$ to $`(1,1)`$, with an unchanged background loss $`\\beta X_3`$, changes the portfolio loss from $`A_\\varepsilon=(1+\\varepsilon)X_1+(1-\\varepsilon)X_2+\\beta X_3=B+\\varepsilon H`$ to $`B`$, where\n\n$$\nB=X_1+X_2+\\beta X_3,\\qquad H=X_1-X_2,\\qquad \\beta>0.\n$$\n\nAt upper-tail probability $`p\\in(0,1)`$, we compare them through $`\\Phi_\\varepsilon(P)=ES_p^P(B)-ES_p^P(A_\\varepsilon)`$, optimizing the two ES thresholds separately under $`P`$ [[32](#ref-rockafellar2002)], so that $`\\Phi_\\varepsilon(P)<0`$ favors the transfer under that law. The background term $`\\beta X_3`$ cancels from the payoff difference $`A_\\varepsilon-B`$, yet we cannot drop it, because its dependence with the traded positions determines which states enter each portfolio's upper tail. A trade assessed in isolation can therefore rank differently once embedded in the whole book [[8](#ref-brandtner2018)].\n\nWriting $`\\Pi`$ for the Fréchet class of joint laws with these marginals, we measure distance from the reference law $`P_0`$ by weighted $`W_1`$ with ground cost $`c_\\gamma(x,y)=|y-x|_\\gamma=\\sum_i\\gamma_i|y_i-x_i|`$, where each $`\\gamma_i>0`$ sets the budget penalty per unit coordinate loss displacement as losses are re-paired across joint states. At budget $`\\rho`$, the stress problem is\n\n$$\nD_\\varepsilon(\\rho)\n=\\sup_{\\substack{P\\in\\Pi\\\\ W_\\gamma(P,P_0)\\le\\rho}}\n\\bigl\\{ES_p^P(B)-ES_p^P(A_\\varepsilon)\\bigr\\},\n\\qquad m_\\varepsilon=-D_\\varepsilon(0).\n$$\n\nUnder an exchangeable reference law, the traded positions have equal marginal ES contributions, so the nominal gain is of second order: $`m_\\varepsilon=(J/p)\\varepsilon^2+o(\\varepsilon^2)`$, with $`J>0`$. For example, when $`\\gamma_1\\le\\min\\{\\gamma_2,\\gamma_3/\\beta\\}`$, the local exchange in Proposition [2.2](#dep-prop-nominal-benchmark) attains an initial stress derivative $`K_\\varepsilon`$ of order one, which suggests the reversal budget $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon=O(\\varepsilon^2)`$. But the derivative at zero budget does not tell us whether this efficiency persists until the nominal gain is exhausted. Since the predicted reversal already occurs at scale $`\\rho=s\\varepsilon^2`$, the tangent at zero is not enough, and we need the risk profile over a nonvanishing interval of the rescaled budget $`s`$.\n\nThree effects contribute at the critical scale. Moving the $`O(\\varepsilon)`$ mass in an $`O(\\varepsilon)`$-layer around the reference tail boundary by distances of order $`\\varepsilon`$ changes both transport cost and tail expectation by $`O(\\varepsilon^2)`$, matching the nominal gain. This local movement need not preserve the marginals. For a fixed regular local plan, the marginal CDF discrepancies are $`O(\\varepsilon^2)`$; by Vallender's formula [[39](#ref-vallender1974)], their $`L^1`$ norms give the one-dimensional restoration costs, even when compensation occurs far from the tail boundary. An $`O(\\varepsilon)`$ shift of either ES threshold also contributes at order $`\\varepsilon^2`$ through the curvature of its objective. We must therefore optimize the local displacement, marginal restoration, and both thresholds together.\n\nWhen the reference law has a positive continuous density on a compact box $`I^2\\times I_3`$ and is exchangeable in its first two coordinates, Theorem [3.1](#dep-thm-critical-limit) gives a finite, continuous profile\n\n$$\nF(s)=\\lim_{\\varepsilon\\downarrow0}\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2),\n\\qquad F(0)=-J/p,\n$$\n\nwith convergence locally uniform in $`s\\ge0`$. The proof rests on a $`\\Gamma`$-limit [[11](#ref-dalmaso1993)] of the transport energies that charges both for the local displacement and for restoring its marginal projections. In the recovery argument, coordinatewise quantile maps attain the restoration cost for each fixed regular local plan. This correction moves losses throughout the box, but its leading effects on the two ES values cancel outside the threshold layer, and approximation then extends the paired-risk convergence to kernels with finite displacement moment.\n\nLet $`\\rho^-_\\varepsilon`$ and $`\\rho^+_\\varepsilon`$ be the infima of budgets permitting a tie and a strict reversal. The strict-growth theorem identifies their common leading coefficient as the unique zero $`s^*`$ of $`F`$:\n\n$$\n\\rho^-_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2),\\qquad\n\\rho^+_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2).\n$$\n\n*Initial efficiency and its limit.* The effective unit costs are $`\\gamma_1,\\gamma_2`$, and $`c_3=\\gamma_3/\\beta`$, with $`c_{\\min}=\\min\\{\\gamma_1,\\gamma_2,c_3\\}`$. If the first traded coordinate is cheapest, Proposition [3.2](#dep-prop-balanced-linear-profile) constructs a positive interval $`[0,s_{\\rm lin}]`$ on which $`F(s)=-J/p+s/(p\\gamma_1)`$. To reach zero at this rate, however, efficient displacement must remain possible throughout the required interval. Theorem [4.2](#dep-thm-zero-premium) characterizes attainment of the universal efficiency bound:\n\n$$\ns^*=Jc_{\\min}\\quad\\Longleftrightarrow\\quad\nc_3\\le\\min\\{\\gamma_1,\\gamma_2\\}.\n$$\n\nWe call the nonnegative gap $`s^*-Jc_{\\min}`$ the *marginal-restoration premium* (normalized by $`Jc_{\\min}`$ when reported as a percentage). When background displacement is cheapest, exchangeability permits opposite shifts to balance on every fixed-background slice. When only traded directions are cheapest, a corner region requires a displacement moment of order $`\\delta^2`$, but its coordinate projections admit only $`O(\\delta^4)`$ of compensating moment near the diagonal (Figure [2](#dep-fig-corner)). Under the attained-slope condition of Proposition [2.2](#dep-prop-nominal-benchmark), with $`c_{\\min}=\\gamma_11`$.\n\n*An analytical benchmark.* Consider independent uniform losses on $`[0,2]^3`$, with $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(1,1,2)`$. Here $`g=3/8`$ and $`J=5/32`$, so Proposition [2.2](#dep-prop-nominal-benchmark) gives $`m_\\varepsilon=(5/16)\\varepsilon^2+o(\\varepsilon^2)`$. By Corollary [5.2](#dep-cor-uniform-reference), the profile is exactly $`F(s)=2s-5/16`$ for $`0\\le s\\le1/96`$, yet its zero satisfies $`13/64\\le s^*\\le29/96`$. So even in this uniform model, where the initial slope is attained, the proved lower bound on reversal is already $`30\\%`$ above the local-efficiency prediction $`5/32`$.\n\n*Sensitivity at equal effective costs.* Along the cost family $`\\gamma_1=\\gamma_2=\\gamma`$, $`c_3=\\gamma(1+\\eta)`$ for $`\\eta\\ge0`$, the baseline $`Jc_{\\min}=\\gamma J`$ is constant, so $`s^*`$ and the marginal-restoration premium have the same right derivative $`\\gamma Ja_+`$ at $`\\eta=0`$. Here $`a_+\\in(0,1)`$ is the least background share among efficient stresses, characterized in Theorem [4.3](#dep-thm-projected-flow) by continuous maximum-flow/minimum-cut duality; Section 5 shows how this share and the budget sensitivity $`\\gamma Ja_+`$ can move in opposite directions as the reference dependence changes.\n\nOur question lies between extremal risk aggregation and local distributional sensitivity. On the aggregation side, subadditivity bounds aggregate risk [[1](#ref-artzner1999), [25](#ref-mcneil2015)], and rearrangements within a Fréchet class describe how diversification depends on the coupling [[33](#ref-ruschendorf2013), [15](#ref-embrechts2013), [5](#ref-bernard2014), [16](#ref-embrechts2015), [41](#ref-wang2015superadditive)]. Partial information leads to convex-order and mixability constraints [[22](#ref-li2018partial)] but can still permit comonotonic upper tails [[12](#ref-devecchinendelstreicher2026)], and fixed-marginal MES bounds quantify uncertainty in individual tail contributions [[9](#ref-chenfurmanlin2026)]. In allocation comparisons, the symmetry imposed on the candidate laws is decisive. Chen et al. [[10](#ref-chen2022)] and Liu and Liu [[23](#ref-liu2026)] use exchangeability and majorization to identify diversification protection, and Liu and Liu also analyze when complete dependence ambiguity favors concentration and how Wasserstein uncertainty modifies that conclusion. We impose exchangeability at the reference law alone: stressed couplings need not remain exchangeable, and the stationarity that symmetry gives at the reference does not by itself protect the reallocation once stress breaks the symmetry. We use this setting to measure the distance to a specified decision reversal, in the spirit of reverse sensitivity analysis [[28](#ref-pesenti2019), [29](#ref-pesenti2022)].\n\nTransport duality is a natural way to preserve the marginal information [[40](#ref-villani2003), [34](#ref-santambrogio2015)]. In Wasserstein robust optimization [[7](#ref-blanchet2019), [19](#ref-gao2023)], adding marginal constraints confines the perturbation to dependence changes [[18](#ref-gao2017), [13](#ref-eckstein2020)]; such constraints encode calibrated risk-neutral laws in multiasset robust pricing [[14](#ref-eckstein2021multiasset)], and Wasserstein and copula restrictions also arise in benchmark portfolio optimization [[30](#ref-pesentijaimungal2023)]. Here the constraint acts on a common stressed loss law for two allocations. After rescaling, its full coordinate projections remain as a transport penalty, and the efficient allocation of displacement reduces to a capacity-constrained flow problem [[20](#ref-korman2015)].\n\nLocal Wasserstein expansions relate the sensitivity of optimized decisions and AVaR criteria to changes in the reference law [[2](#ref-bartl2021)], and adapted transport brings in the information available in multiperiod decisions [[4](#ref-bartlwiesel2023)]. When a first-order term vanishes, one turns to higher-order expansions: Fiechtner and Blanchet [[17](#ref-fiechtner2026)] derive a second-order regret expansion under a unique empirical-risk minimizer in this situation, and Sauldubois [[35](#ref-sauldubois2026)] develops higher-order Wasserstein and adapted-Wasserstein expansions, including martingale constraints. Dependence restrictions also affect local sensitivity, as in Lam [[21](#ref-lam2018)], while Namkoong [[26](#ref-namkoong2026)] evaluates the first-order response over restricted transport directions. In our problem the trade size and stress radius vanish together, so locating a decision reversal at scale $`\\rho=s\\varepsilon^2`$ requires the full rescaled profile, with global marginal compensation and both optimized ES thresholds retained over the budget interval needed to reach zero.\n\nSection 2 derives the nominal gain and the initial stress slope, and Section 3 takes the variational limit at scale $`\\rho=s\\varepsilon^2`$ and identifies the unique zero of the limiting profile. Section 4 characterizes zero premium and represents the boundary derivative by continuous projected flow–cut duality. Section 5 works through analytical examples and what they imply for reverse stress testing. The analytic proofs are in Appendix A, and the potential estimates and section calculations for the examples are in Appendix B.\n\n**Notation and conventions** We write $`T_\\#\\mu`$ for a pushforward and measure displacements by $`|d|_\\gamma=\\sum_i\\gamma_i|d_i|`$. The constant in $`O_\\alpha(\\varepsilon^k)`$ may depend on the fixed kernel and model, but not on $`\\varepsilon`$. Scalar section integrands are extended by zero outside their domains, and projection identities are equalities of finite signed Borel measures. We use $`\\inf\\varnothing=+\\infty`$; allocation ratios are assigned the value zero where both numerator and denominator vanish.\n\n\n\n## 2. Model formulation and the local benchmark\n\nWe first formulate the comparison for arbitrary integrable marginal laws, which are enough for a worst stressed model to exist. The critical-scale analysis will use a positive continuous density and exchangeability of the two traded positions, both introduced in Section 2.2.\n\n### 2.1 Fixed marginals and a common stress law\n\nLet $`n\\ge2`$, let $`\\mu_1,\\ldots,\\mu_n`$ have finite first moments, and fix $`P_0`$ in the Fréchet class $`\\Pi(\\mu_1,\\ldots,\\mu_n)=\\{P\\in\\mathcal P(\\mathbb R^n):P_i=\\mu_i,\\ 1\\le i\\le n\\}`$.\n\nFor exposure vectors $`b,h\\in\\mathbb R^n`$, the losses before and after the proposed reallocation are $`A_\\varepsilon=B+\\varepsilon H`$ and $`B`$, with $`B=b\\cdot X`$ and $`H=h\\cdot X`$. We use the upper-tail mass $`p\\in(0,1)`$ and the Expected Shortfall representation of Rockafellar and Uryasev [[32](#ref-rockafellar2002)],\n\n\n\n$$\nES_p^P(L)=\\min_{t\\in\\mathbb R}\n\\left\\{t+\\frac1pE_P(L-t)_+\\right\\}.\n\\tag{2.1}\n$$\n\nWe write $`\\mathscr E_L(P;t)=t+p^{-1}E_P(L-t)_+`$ for the threshold objective. For an integrable loss, its minimum is attained on the nonempty compact interval characterized by\n\n\n\n$$\nt\\in\\arg\\min_{u}\\mathscr E_L(P;u)\n\\quad\\Longleftrightarrow\\quad P(L>t)\\le p\\le P(L\\ge t).\n\\tag{2.2}\n$$\n\nWithout a continuity assumption on the stressed loss law, support gaps can leave a nondegenerate interval of minimizing thresholds. We therefore keep the set-valued characterization [(2.2)](#dep-eq-2-2), so that subsequent estimates apply to every minimizer, including for laws with atoms.\n\nUnder a common candidate law $`P\\in\\Pi`$, we compare $`\\Phi_\\varepsilon(P)=ES_p^P(B)-ES_p^P(A_\\varepsilon)`$, minimizing each loss over its own threshold. Maximizing this paired difference tests whether a single dependence scenario can erase the nominal gain, whereas separate maximizations of the two ES values may select different joint laws. We measure changes in dependence with positive weights $`\\gamma_i`$ and ground cost\n\n$$\nc_\\gamma(x,y)=\\sum_{i=1}^n\\gamma_i|y_i-x_i|,\n\\qquad W_\\gamma(P,P_0)=\\min_{\\pi\\in\\Gamma(P_0,P)}\\int c_\\gamma\\,d\\pi.\n$$\n\nWithin $`\\Pi=\\Pi(\\mu_1,\\ldots,\\mu_n)`$, transport changes how component losses are paired across joint states. With the weights and all marginal laws held fixed, we define the worst comparison at budget $`\\rho`$ and the nominal gain by\n\n\n\n$$\n\\mathcal A_\\rho=\\{P\\in\\Pi:W_\\gamma(P,P_0)\\le\\rho\\},\\quad\nD_\\varepsilon(\\rho)=\\max_{P\\in\\mathcal A_\\rho}\\Phi_\\varepsilon(P),\\quad\nm_\\varepsilon=-\\Phi_\\varepsilon(P_0).\n\\tag{2.3}\n$$\n\nProposition [2.1](#dep-prop-general-wellposedness) guarantees attainment in [(2.3)](#dep-eq-2-3) even with atoms, singular dependence, and unbounded marginal supports. A nondecreasing profile can vanish on a whole interval, so we distinguish the infimal budgets for a tie and a strict reversal,\n\n\n\n$$\n\\rho^-_\\varepsilon=\\inf\\{\\rho\\ge0:D_\\varepsilon(\\rho)\\ge0\\},\n\\qquad\n\\rho^+_\\varepsilon=\\inf\\{\\rho\\ge0:D_\\varepsilon(\\rho)>0\\},\n\\tag{2.4}\n$$\n\nwhich, by Theorem [3.3](#dep-thm-strict-crossing), have the same leading asymptotics in the exchangeable box model of Section 2.2.\n\n\n\n**Proposition 2.1 (Existence under integrable marginal laws).** Let $`n\\ge2`$, let $`\\mu_1,\\ldots,\\mu_n`$ have finite first moments, and fix $`P_0\\in\\Pi(\\mu_1,\\ldots,\\mu_n)`$, positive weights $`\\gamma_i`$, exposure vectors $`b,h\\in\\mathbb R^n`$, and $`p\\in(0,1)`$. Then $`\\Pi`$ and every $`\\mathcal A_\\rho`$ are compact for $`W_\\gamma`$, and the maximum in [(2.3)](#dep-eq-2-3) is attained. For each fixed $`\\varepsilon`$, the profile $`D_\\varepsilon`$ is nondecreasing and Lipschitz on $`[0,\\infty)`$.\n\n*Proof.* The fixed-marginal class is tight and weakly closed. Moreover, along any weakly convergent sequence $`P_k\\to P`$ in $`\\Pi`$, the first moments satisfy $`\\int|x|_\\gamma\\,dP_k=\\sum_i\\gamma_i\\int|x_i|\\,d\\mu_i=\\int|x|_\\gamma\\,dP`$, so the Wasserstein convergence criterion [[40](#ref-villani2003)] implies $`W_\\gamma`$-compactness of $`\\Pi`$ and its closed subsets $`\\mathcal A_\\rho`$. For a linear loss $`L=a\\cdot X`$, coupling the laws and minimizing the uniformly Lipschitz hinges over their thresholds gives\n\n$$\n|ES_p^P(L)-ES_p^Q(L)|\\le p^{-1}\\|a\\|_{\\gamma,*}W_\\gamma(P,Q),\n\\qquad \\|a\\|_{\\gamma,*}=\\max_i|a_i|/\\gamma_i.\n$$\n\nApplying this bound to $`B=b\\cdot X`$ and $`A_\\varepsilon=(b+\\varepsilon h)\\cdot X`$, we find that their ES difference $`\\Phi_\\varepsilon`$ is Lipschitz with constant $`C_\\varepsilon=p^{-1}(\\|b\\|_{\\gamma,*}+\\|b+\\varepsilon h\\|_{\\gamma,*})`$. Compactness of $`\\mathcal A_\\rho`$ then ensures attainment. To compare two budgets $`\\rho_2>\\rho_1`$, take a maximizer $`P`$ at $`\\rho_2`$. The mixture $`P_a=(1-a)P_0+aP`$, $`a=\\rho_1/\\rho_2`$, satisfies $`W_\\gamma(P_a,P_0)\\le a\\rho_2`$ and $`W_\\gamma(P_a,P)\\le(1-a)\\rho_2`$, by mixing an optimal plan with the corresponding identity plans. Comparing $`\\Phi_\\varepsilon(P_a)`$ and $`\\Phi_\\varepsilon(P)`$ yields $`0\\le D_\\varepsilon(\\rho_2)-D_\\varepsilon(\\rho_1)\\le C_\\varepsilon(\\rho_2-\\rho_1)`$. ∎\n\nFor a transfer $`H=X_i-X_j`$ between positions with the same marginal law, replacing $`X_j`$ by $`X_i`$ preserves every marginal and gives $`A_\\varepsilon=B`$ at cost $`\\gamma_jE_{P_0}|X_i-X_j|`$; reversing their roles yields\n\n$$\n0\\le\\rho^-_\\varepsilon\n\\le\\min\\{\\gamma_i,\\gamma_j\\}\\,E_{P_0}|X_i-X_j|<\\infty.\n$$\n\nSo the tie budget is finite, but this bound leaves its small-trade scale open. Under the exchangeable box assumptions introduced next, Section 3 shows that both critical budgets are of order $`\\varepsilon^2`$.\n\n\n\n### 2.2 The exchangeable box model\n\nTo isolate how dependence with the rest of a portfolio affects a small reallocation, we keep the two traded losses separate and collect all unchanged positions into their aggregate loss $`\\beta X_3`$. The marginal law of $`X_3`$ counts as retained information, and we specify the transport metric directly on $`(X_1,X_2,X_3)`$. This aggregate model fixes the information used in the stress test: the retained marginal and displacement cost are those of the aggregate loss, a different specification from fixing separately the marginals and transport costs of every underlying background position.\n\n\n\n**Standing assumptions (exchangeable box model)** Let\n\n$$\nI=[\\ell,r],\\qquad I_3=[\\ell_3,r_3],\\qquad\nS=I^2\\times I_3,\\qquad \\Delta_I=r-\\ell>0,\\quad \\Delta_3=r_3-\\ell_3>0.\n$$\n\nWe also write $`I_1=I_2=I`$ when treating coordinate projections together. The reference law has a continuous density $`f>0`$ on the closed box $`S`$, normalized to integrate to one and satisfying $`f(x,y,z)=f(y,x,z)`$. We take\n\n$$\nB=X_1+X_2+\\beta X_3,\\qquad H=X_1-X_2,\\qquad\nA_\\varepsilon=B+\\varepsilon H,\\qquad\n\\beta>0,\\quad 0<\\varepsilon<1,\n$$\n\nwith arbitrary $`p\\in(0,1)`$ and $`\\gamma_i>0`$, so that $`b=(1,1,\\beta)`$ and $`h=(1,-1,0)`$ in the preceding notation. All model parameters remain fixed as $`\\varepsilon\\downarrow0`$. The first two unit positions are expressed at equal initial cost, so the trade changes their holdings from $`(1+\\varepsilon,1-\\varepsilon)`$ to $`(1,1)`$ without changing the initial outlay. The position $`\\beta X_3`$ represents the background book, whose holding remains unchanged during the trade.\n\nThe stressed law may be any coupling in $`\\Pi((P_0)_1,(P_0)_2,(P_0)_3)`$; exchangeability and absolute continuity are assumptions on the reference law only. Reference symmetry makes the equal allocation stationary under $`P_0`$, and equality of the first two marginals gives $`E_P[A_\\varepsilon]-E_P[B]=\\varepsilon(E_P[X_1]-E_P[X_2])=0`$ under every candidate, so before implementation costs the ES comparison distinguishes tail risks between portfolios with the same expected loss.\n\nWe work on a compact box with a positive continuous density because that lets us study this dependence mechanism with uniform control of marginal restoration. Positivity on the closed box gives both the interior mass that local exchanges need and lower bounds on the marginal densities, hence Lipschitz bounds for their inverse CDFs, and these bounds enter the quantile recovery in Section 3.2 and Appendix A.5. Section 5.3 takes up the additional difficulties posed by unbounded marginals.\n\n\n\n**Loss units and effective stress costs** A shift $`\\Delta X_3`$ changes the portfolio loss by $`\\beta\\Delta X_3`$ at cost $`\\gamma_3|\\Delta X_3|=(\\gamma_3/\\beta)|\\beta\\Delta X_3|`$. So $`c_1=\\gamma_1`$, $`c_2=\\gamma_2`$, and $`c_3=\\gamma_3/\\beta`$ are the costs per unit change of $`B`$ along the three coordinate axes, and the ratio $`c_3/\\gamma_1`$ compares a background displacement with a first-position displacement having the same effect on portfolio loss.\n\n**Canonical box normalization** Writing the original coordinates as $`X=T_{\\rm box}(u)=(\\ell+\\kappa u_1,\\ell+\\kappa u_2,\\ell_3+\\kappa_3u_3)`$, with $`\\kappa=\\Delta_I/2`$ and $`\\kappa_3=\\Delta_3/2`$, gives a normalized model on $`[0,2]^3`$ with $`\\widehat\\beta=\\beta\\kappa_3/\\kappa`$ and $`\\widehat\\gamma=(\\kappa\\gamma_1,\\kappa\\gamma_2,\\kappa_3\\gamma_3)`$. By cash invariance and positive homogeneity of ES, the original paired risk is $`\\kappa`$ times its normalized counterpart, while corresponding transport plans have equal cost; hence the critical budgets agree and $`p`$ is unchanged. A common rescaling of all transport weights rescales those budgets.\n\n\n\n### 2.3 Nominal curvature and the exact initial slope\n\nUnder the reference model of Section 2.2, let $`v`$ be the upper-tail $`p`$ threshold of $`B`$. The marginal expected shortfall (MES) of a unit position $`X_i`$, namely $`E_{P_0}[X_i\\mid B>v]`$, is the ES derivative with respect to its holding, and multiplying by that holding gives its Euler contribution [[36](#ref-scaillet2004), [38](#ref-tasche2001)]. The first two holdings in $`B`$ both equal one, so the directional derivative along $`H=X_1-X_2`$ is their MES difference: with $`T_H=E_{P_0}[H1_{\\{B>v\\}}]`$, it equals $`p^{-1}T_H=E_{P_0}[X_1-X_2\\mid B>v]=0`$ by exchangeability. To find the gain from a nonzero trade we then have to differentiate the moving tail boundary, which leads to a conditional variance on the tail section.\n\nWe record the classical section differentiation formula under our continuity assumptions in Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4).\n\nThe tail boundary can be parameterized by\n\n$$\n\\Omega=\\{(x,y)\\in(\\ell,r)^2:\\ell_3\n\nSince $`p\\in(0,1)`$ and $`f>0`$, the threshold satisfies $`2\\ell+\\beta\\ell_3\n\n$$\ng=\\sigma(\\Omega)>0,\\qquad J=\\frac12\\int_\\Omega H^2\\,d\\sigma>0.\n\\tag{2.5}\n$$\n\nHere $`g=f_B(v)`$ is the loss density at the threshold, and $`d\\sigma/g`$ is its normalized conditional section law. Exchangeability makes the conditional mean of $`H`$ zero, giving $`2J/g=\\operatorname{Var}_{P_0}(H\\mid B=v)`$, so the second-order diversification gain is governed by the conditional variance of the trade payoff on $`\\{B=v\\}`$, weighted by the threshold density $`g`$.\n\n\n\n**Proposition 2.2 (Nominal benchmark and exact initial efficiency).** In the exchangeable box model of Section 2.2:\n\n(i) *Nominal gain and threshold curvature.* The reference thresholds and nominal gain satisfy\n\n\n\n$$\nt_\\varepsilon-v=o(\\varepsilon),\\qquad\nm_\\varepsilon=\\frac Jp\\varepsilon^2+o(\\varepsilon^2).\n\\tag{2.6}\n$$\n\nThe threshold of $`B+\\eta H`$ is unique for every $`|\\eta|<1`$, and $`m_\\varepsilon>0`$ for $`0<\\varepsilon<1`$. With $`L_{c,\\varepsilon}=B+c\\varepsilon H`$, the unoptimized objective satisfies, uniformly for $`c=0,1`$ and $`|\\tau|\\le T<\\infty`$,\n\n\n\n$$\n\\mathscr E_{L_{c,\\varepsilon}}(P_0;v+\\varepsilon\\tau)\n=ES_p^{P_0}(B)+\\frac{\\varepsilon^2}{p}\n\\left(c^2J+\\frac g2\\tau^2\\right)+o(\\varepsilon^2).\n\\tag{2.7}\n$$\n\n(ii) *Fixed-marginal Lipschitz bound and initial slope.* For every $`P,Q\\in\\Pi`$,\n\n\n\n$$\n|\\Phi_\\varepsilon(P)-\\Phi_\\varepsilon(Q)|\\le \\kappa_\\varepsilon W_\\gamma(P,Q),\\quad\n\\kappa_\\varepsilon=\\frac1p\\max\\left\\{\\frac{1+\\varepsilon/2}{\\gamma_1},\n\\frac{1-\\varepsilon/2}{\\gamma_2},\\frac\\beta{\\gamma_3}\\right\\}.\n\\tag{2.8}\n$$\n\nFor fixed $`0<\\varepsilon<1`$, if\n\n\n\n$$\n\\frac{1+\\varepsilon/2}{\\gamma_1}\n\\ge\\max\\left\\{\\frac{1-\\varepsilon/2}{\\gamma_2},\\frac\\beta{\\gamma_3}\\right\\},\n\\tag{2.9}\n$$\n\nthen there exists $`r_\\varepsilon>0`$ such that\n\n\n\n$$\nK_\\varepsilon=\\frac{1+\\varepsilon/2}{p\\gamma_1},\\qquad\nD_\\varepsilon(\\rho)=-m_\\varepsilon+K_\\varepsilon\\rho\n\\quad(0\\le\\rho\\le r_\\varepsilon).\n\\tag{2.10}\n$$\n\n*Proof.* For part (i), the coefficients of $`B+\\eta H`$ are positive whenever $`|\\eta|<1`$, and its support is $`[2\\ell+\\beta\\ell_3,2r+\\beta r_3]`$. Since every interior level meets $`\\operatorname{int}S`$, Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4) gives a positive continuous loss density and a unique threshold. The gradients of $`H`$ and $`B+\\eta H`$ are linearly independent, so $`H`$ is nonconstant on each such section. Its conditional variance is therefore strictly positive, and the same lemma implies strict convexity of $`\\eta\\mapsto ES_p^{P_0}(B+\\eta H)`$ on $`(-1,1)`$.\n\nAt $`\\eta=0`$, exchanging the first two coordinates preserves both the reference tail and its boundary measure and reverses $`H`$, giving $`T_H=m_1=0`$. In the notation of Lemma [A.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4), the remaining section moments at $`(0,v)`$ are $`m_0=g`$ and $`m_2=2J`$. Thus the threshold derivative vanishes and the optimized second derivative is $`2J/p`$, which yield [(2.6)](#dep-eq-2-6) by Taylor expansion at zero. Strict convexity makes this stationary point the strict minimum, so $`m_\\varepsilon>0`$ for all $`0<\\varepsilon<1`$. For the unoptimized objective $`\\Psi`$, the Hessian at $`(0,v)`$ is $`\\operatorname{diag}(2J/p,g/p)`$. Expanding at the trade size and threshold displacement $`(c\\varepsilon,\\varepsilon\\tau)`$ gives [(2.7)](#dep-eq-2-7), uniformly on bounded $`\\tau`$-intervals.\n\n*The fixed-marginal Lipschitz bound.* We next establish the uniform Lipschitz bound over $`\\Pi`$. Given any minimizing $`u`$ for $`B`$ under $`Q`$ and any minimizing $`t`$ for $`A_\\varepsilon`$ under $`P`$, let\n\n$$\n\\begin{aligned}\nh_{u,t}&=(B-u)_+-(A_\\varepsilon-t)_+,\\\\\n\\Delta_B&=\\mathscr E_B(P;u)-ES_p^P(B),&\n\\Delta_A&=\\mathscr E_{A_\\varepsilon}(Q;t)-ES_p^Q(A_\\varepsilon).\n\\end{aligned}\n$$\n\nBoth gaps are nonnegative, and direct subtraction gives\n\n\n\n$$\n\\Phi_\\varepsilon(P)-\\Phi_\\varepsilon(Q)\n=p^{-1}\\int h_{u,t}\\,d(P-Q)-\\Delta_B-\\Delta_A.\n\\tag{2.11}\n$$\n\nA direct bound on $`h_{u,t}`$ would use its largest absolute $`x_1`$-slope, $`1+\\varepsilon`$. Because $`P`$ and $`Q`$ share their marginals, adding an integrable separable potential does not change the integral against $`P-Q`$. We can thus choose a representative of $`h_{u,t}`$ modulo marginal potentials with a smaller Lipschitz bound. Here the required adjustment is linear: since $`\\int H\\,d(P-Q)=0`$, we may add any scalar multiple of $`H=X_1-X_2`$. The extreme $`x_1`$-slopes, $`-(1+\\varepsilon)`$ and $`1`$, have midpoint $`-\\varepsilon/2`$, so adding $`\\varepsilon H/2`$ minimizes their largest absolute value among constant shifts and simultaneously centers the $`x_2`$-slopes. Away from the two tail planes, the four regime slopes are\n\n$$\n\\begin{array}{c|rrrr}\n(1_{\\{B>u\\}},1_{\\{A_\\varepsilon>t\\}})&(0,0)&(1,0)&(0,1)&(1,1)\\\\\n\\partial_{x_1}h_{u,t}&0&1&-(1+\\varepsilon)&-\\varepsilon\\\\\n\\partial_{x_1}(h_{u,t}+\\varepsilon H/2)&\\varepsilon/2&1+\\varepsilon/2&-1-\\varepsilon/2&-\\varepsilon/2.\n\\end{array}\n$$\n\nWith the extreme $`x_1`$-slopes centered at zero, their largest absolute value drops to $`1+\\varepsilon/2`$, while the corresponding bounds for $`x_2`$ and $`x_3`$ are $`1-\\varepsilon/2`$ and $`\\beta`$. Setting $`G_{u,t}=h_{u,t}+\\varepsilon H/2`$, which is absolutely continuous along coordinate lines, we obtain\n\n$$\n|G_{u,t}(Y)-G_{u,t}(X)|\n\\le(1+\\varepsilon/2)|Y_1-X_1|\n+(1-\\varepsilon/2)|Y_2-X_2|+\\beta|Y_3-X_3|.\n$$\n\nIntegrating against an optimal transport plan and then interchanging $`P,Q`$ proves [(2.8)](#dep-eq-2-8). The mixing argument in Proposition [2.1](#dep-prop-general-wellposedness), now with this sharp constant, also gives\n\n\n\n$$\n0\\le D_\\varepsilon(\\rho_2)-D_\\varepsilon(\\rho_1)\n\\le \\kappa_\\varepsilon(\\rho_2-\\rho_1)\\qquad(0\\le\\rho_1<\\rho_2).\n\\tag{2.12}\n$$\n\n\n\n*Attaining the initial slope.* Under [(2.9)](#dep-eq-2-9), we construct an attaining exchange near an interior intersection of the reference tail planes $`\\{B=v\\}`$ and $`\\{A_\\varepsilon=t_\\varepsilon\\}`$. To locate such an intersection, fix $`\\varepsilon\\in(0,1)`$ and let $`[m_F,M_F]`$ be the nondegenerate range of $`A_\\varepsilon`$ on $`F_v=\\{B=v\\}\\cap S`$. Both losses increase along each nonconstant segment from the lower vertex $`x_{\\min}=(\\ell,\\ell,\\ell_3)`$. Any point with $`B>v`$ lies beyond an intersection of that segment with $`F_v`$, so $`t_\\varepsilon\\le m_F`$ would imply $`\\{B>v\\}\\subset\\{A_\\varepsilon>t_\\varepsilon\\}`$. Since $`A_\\varepsilon`$ is nonconstant on $`F_v`$, the second tail would also contain an open set below $`B=v`$, contradicting the equal tail probabilities and $`f>0`$. Rays toward the upper corner similarly exclude $`t_\\varepsilon\\ge M_F`$, which leaves $`t_\\varepsilon\\in(m_F,M_F)`$. The relative-interior property of a linear map on a convex section gives an intersection $`c=(x_*,y_*,z_*)\\in\\operatorname{int}S`$.\n\nTo exchange the first-coordinate marginal intervals of two small source cubes in opposite disagreement regions along $`e_1`$, we first choose reference points with the same coordinate $`x_*`$ and seek equal and opposite changes in the two losses. From the interior intersection $`c`$, a displacement $`(0,-\\zeta_0,a\\zeta_0)`$ changes $`B`$ by $`(-1+\\beta a)\\zeta_0`$ and $`A_\\varepsilon`$ by $`(-(1-\\varepsilon)+\\beta a)\\zeta_0`$. Equating their sum to zero determines $`a=(1-\\varepsilon/2)/\\beta`$. Thus the points\n\n$$\nc_\\pm=(x_*,y_*\\mp \\zeta_0,z_*\\pm(1-\\varepsilon/2)\\zeta_0/\\beta)\n$$\n\nhave opposite strict tail memberships for small $`\\zeta_0>0`$: at $`c_+`$, the losses differ from their thresholds by $`-\\varepsilon \\zeta_0/2`$ for $`B`$ and $`\\varepsilon \\zeta_0/2`$ for $`A_\\varepsilon`$, with both signs reversed at $`c_-`$. Choosing $`0<\\delta_x<\\varepsilon\\zeta_0/(1+\\varepsilon)`$ bounds the loss change along either half-translation by the strict tail margin. With this margin fixed, we can take $`\\eta_0>0`$ small enough that the cubes $`E_\\pm`$ of half-width $`\\eta_0`$, centered at $`c_\\pm\\pm(\\delta_x/2)e_1`$, have disjoint second-coordinate intervals and remain inside $`S`$ and their respective tail regions throughout translation by $`\\mp\\delta_xe_1`$.\n\nThe uniform subdensities $`\\nu_\\pm=\\lambda1_{E_\\pm}\\,dx\\,dy\\,dz`$, where $`\\lambda=\\min_Sf`$, have equal mass $`\\mu=\\lambda(2\\eta_0)^3`$ and sum at most $`P_0`$. The translations $`T_\\pm(x)=x\\mp\\delta_xe_1`$ therefore define the admissible perturbation\n\n\n\n$$\nP_\\theta=P_0-\\theta(\\nu_++\\nu_-)\n+\\theta(T_+)_\\#\\nu_++\\theta(T_-)_\\#\\nu_-,\\qquad 0\\le\\theta\\le1.\n\\tag{2.13}\n$$\n\nThe translations exchange the first-coordinate marginal intervals, fix the other coordinates, and cross neither tail plane. Thus $`P_\\theta`$ preserves all marginals and both reference tail probabilities, leaving $`v`$ and $`t_\\varepsilon`$ optimal. Integrating the affine hinges on the two cubes gives\n\n\n\n$$\n\\Phi_\\varepsilon(P_\\theta)-\\Phi_\\varepsilon(P_0)\n=\\frac{(2+\\varepsilon)\\theta\\mu\\delta_x}{p}\n=K_\\varepsilon(2\\gamma_1\\theta\\mu\\delta_x).\n\\tag{2.14}\n$$\n\nThe plan has cost $`2\\gamma_1\\theta\\mu\\delta_x`$, and the sharp bound [(2.8)](#dep-eq-2-8) forces this to equal $`W_\\gamma(P_\\theta,P_0)`$. Varying $`\\theta\\in[0,1]`$ proves [(2.10)](#dep-eq-2-10) with $`r_\\varepsilon=2\\gamma_1\\mu\\delta_x>0`$. ∎\n\nIf $`\\gamma_2\\ge\\gamma_1`$ and $`\\gamma_3\\ge\\beta\\gamma_1`$, condition [(2.9)](#dep-eq-2-9) holds for every $`\\varepsilon\\in(0,1)`$, which suggests the extrapolated reversal budget $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon`$. The construction gives $`r_\\varepsilon>0`$ for each fixed trade size, but the disagreement regions narrow with $`\\varepsilon`$, so it leaves $`r_\\varepsilon/\\varepsilon^2`$ uncontrolled. Indeed, the unscaled bound $`|D_\\varepsilon(\\rho)|\\le\\varepsilon\\|H\\|_\\infty`$ makes both iterated limits in $`(\\rho,\\varepsilon)`$ vanish even though $`K_\\varepsilon\\to1/(p\\gamma_1)>0`$, which is why Section 3 tracks the profile under the joint scaling $`\\rho=s\\varepsilon^2`$.\n\n\n\n**Reference symmetry and stressed dependence** Our stress class allows asymmetric dependence between the two traded positions. If every candidate law were exchangeable, the losses $`B-\\varepsilon H`$ and $`B+\\varepsilon H`$ would have the same distribution, and convexity of ES would give\n\n\n\n$$\nES_p^P(B)\\le\\tfrac12 ES_p^P(B-\\varepsilon H)+\\tfrac12 ES_p^P(B+\\varepsilon H)\n=ES_p^P(A_\\varepsilon).\n\\tag{2.15}\n$$\n\nRequiring exchangeability of every stressed law would thus protect this reallocation [[10](#ref-chen2022), [23](#ref-liu2026)]. We impose it only at the reference, which gives stationarity there while permitting dependence stresses that test the persistence of the gain. Identical marginal laws alone do not imply stationarity. For example, on $`S=[0,2]^3`$ with $`\\beta=1`$ and $`p=1/2`$, the perturbation $`f_\\delta=[1+\\delta(x-y)(z-1)]/8`$, with $`0<|\\delta|<1/2`$, preserves uniform marginals but gives a nonzero tail expectation $`T_H`$. Antisymmetry in $`x,y`$ preserves the law of $`B=x+y+z`$, hence $`v=3`$. Pairing $`z=1\\pm t`$ leaves the strip $`2-t\n\n## 3. Variational limit and the critical profile\n\nUnder the traded-coordinate condition of Proposition [2.2](#dep-prop-nominal-benchmark), a local exchange attains the initial slope. To find out how far that efficiency persists toward reversal, we now hold $`s=\\rho/\\varepsilon^2`$ fixed. The limiting objects introduced below connect the stress profile to its reversal coefficient:\n\n$$\n\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2)\n\\xrightarrow{\\text{locally uniformly}}\nF(s)\n\\xrightarrow{\\text{unique zero}}\ns^*.\n$$\n\nThe first passage comes from Theorem [3.1](#dep-thm-critical-limit), which combines convergence of the transport energies with exact marginal recovery and convergence of the paired risk. Strict growth in Theorem [3.3](#dep-thm-strict-crossing) then identifies the common critical coefficient.\n\n\n\n### 3.1 Local transport and marginal compensation\n\nAn $`O(\\varepsilon)`$-layer around $`\\{B=v\\}`$ contains $`O(\\varepsilon)`$ reference mass, so displacements of order $`\\varepsilon`$ change both the transport cost and the paired hinge expectation at order $`\\varepsilon^2`$. We resolve this layer through the normal coordinate $`u=(B(X)-v)/\\varepsilon`$ and displacement $`d=(Y-X)/\\varepsilon`$, while keeping the cost of any transport outside bounded windows that is needed to restore the marginals.\n\n**Definition (Marginal flux and effective energy)** On bounded $`u`$-windows the rescaled source converges to $`d\\nu=d\\sigma\\,du`$. We disintegrate a limiting plan with this source into a probability kernel $`\\alpha_{\\xi,u}(dd)`$, specifying the rescaled displacement at surface point $`\\xi`$ and normal coordinate $`u`$; stationary sources have kernel $`\\delta_0`$. Writing $`d\\Pi_\\alpha(\\xi,u,d)=\\alpha_{\\xi,u}(dd)\\,d\\nu(\\xi,u)`$, we require a finite weighted displacement moment even though the source measure $`\\nu=\\sigma\\otimes du`$ has infinite total mass:\n\n$$\nI(\\alpha)=\\int|d|_\\gamma\\,d\\Pi_\\alpha<\\infty,\n\\qquad |d|_\\gamma=\\sum_i\\gamma_i|d_i|.\n$$\n\nAcross a coordinate cut $`\\{X_i\\le q\\}`$, upward movements remove mass from below the cut and downward movements add mass to it. The resulting signed flux has a Lebesgue density $`j_i\\in L^1(I_i)`$, as shown by the section formulas in Appendix A.3, and we define\n\n\n\n$$\nj_i(q)dq=(\\xi_i)_\\#(d_i\\,d\\Pi_\\alpha),\\qquad\n\\mathcal C(\\alpha)=I(\\alpha)+\\sum_{i=1}^3\\gamma_i\\|j_i\\|_{L^1(I_i)}.\n\\tag{3.1}\n$$\n\nIn words, $`j_i`$ is the signed displacement moment projected onto the $`i`$-th coordinate, and its negative is the leading marginal CDF discrepancy produced by a regular local plan. The full plan preserves each marginal, so transport outside a bounded normal and displacement window has to compensate the discrepancy within it, and the one-dimensional CDF formula for $`W_1`$ [[39](#ref-vallender1974)] turns that compensation into a cost lower bound. The energy $`\\mathcal C`$ adds this limiting restoration cost to $`I`$ (Figure [1](#dep-fig-tail-layer)). Projection contracts total variation, giving $`I\\le\\mathcal C\\le2I`$, with $`\\mathcal C=I`$ when all projected fluxes vanish and $`\\mathcal C=2I`$ when each $`d_i`$ has a fixed sign $`\\Pi_\\alpha`$-almost everywhere, so that $`\\|j_i\\|_1=\\int|d_i|\\,d\\Pi_\\alpha`$.\n\n\n\n\n\n\n\n### 3.2 The limiting risk and convergence theorem\n\nUnder the same local plan, each loss has its own minimizing threshold. At threshold $`v+\\varepsilon\\tau`$, a displacement $`d`$ changes the leading centered and rescaled loss of $`L_{c,\\varepsilon}=B+c\\varepsilon H`$ from $`u+cH`$ to $`u+b\\cdot d+cH`$. Denote the corresponding hinge increment by $`\\mathfrak q_c(\\tau)=(u+b\\cdot d+cH-\\tau)_+-(u+cH-\\tau)_+`$. We write $`T_c^\\alpha(\\tau)=\\int\\mathfrak q_c(\\tau)\\,d\\Pi_\\alpha`$ for its integral.\n\n**Definition (Limiting paired risk)** Moving the threshold away from $`v`$ contributes the curvature term $`g\\tau^2/2`$ from [(2.7)](#dep-eq-2-7). Adding the local transport increment gives the two limiting threshold objectives, which must still be minimized separately:\n\n\n\n$$\nG_c^\\alpha(\\tau)=\\frac g2\\tau^2+T_c^\\alpha(\\tau),\\qquad\n\\mathcal V(\\alpha)=-\\frac Jp+\n\\frac{\\min_{\\tau_B}G_0^\\alpha(\\tau_B)-\\min_{\\tau_A}G_1^\\alpha(\\tau_A)}p.\n\\tag{3.2}\n$$\n\nThe energy $`\\mathcal C`$ depends only on the absolute displacement moment and the signed coordinate projections, and it is convex under mixtures of kernels. The paired risk also depends on how the displacements are distributed relative to the normal coordinate $`u`$: each of the two hinge objectives optimizes its own threshold, so their difference need not be concave in the common kernel.\n\nWe omit the superscript $`\\alpha`$ when the kernel is clear, and write $`\\kappa_B=\\max_i b_i/\\gamma_i`$, $`\\kappa_H=\\max\\{\\gamma_1^{-1},\\gamma_2^{-1}\\}`$. Since $`|T_c^\\alpha|\\le\\kappa_BI`$, the quadratic term gives coercivity and confines every minimizer to $`g\\tau^2/2\\le2\\kappa_BI`$. Together with convexity from Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101), these properties justify the threshold minimizations in $`\\mathcal V`$. We then optimize the paired value over the energy budget, with attainment proved in Theorem [3.1](#dep-thm-critical-limit):\n\n\n\n$$\nF(s)=\\max_{\\mathcal C(\\alpha)\\le s}\\mathcal V(\\alpha),\\qquad s\\ge0.\n\\tag{3.3}\n$$\n\nWe use vague convergence on $`\\mathscr X=\\mathcal M^+_{\\rm loc}(Z)`$, testing against compactly supported continuous functions on $`Z=I^2\\times\\mathbb R\\times\\mathbb R^3`$, so that compensation may occur outside bounded windows. We lift a plan $`\\pi`$ from $`P_0`$ to a law in $`\\Pi`$ by\n\n$$\n\\mathscr T_\\varepsilon\\pi=\\varepsilon^{-1}\n\\left(X_1,X_2,\\frac{B(X)-v}{\\varepsilon},\\frac{Y-X}{\\varepsilon}\\right)_\\#\\pi.\n$$\n\nHere the factor $`\\varepsilon^{-1}`$ compensates for the thickness of the source layer. The lifted coordinates determine $`(X,Y)`$ for fixed $`\\varepsilon`$, so the rescaled energy can retain the cost of the entire original plan:\n\n$$\n\\begin{aligned}\n\\mathcal E_\\varepsilon(\\Lambda)\n&=\\begin{cases}\n\\displaystyle\\int |d|_\\gamma\\,d\\Lambda\n=\\varepsilon^{-2}\\int c_\\gamma\\,d\\pi,\n&\\begin{gathered}\\Lambda=\\mathscr T_\\varepsilon\\pi,\\\\ \\pi\\in\\Gamma(P_0,P),\\ P\\in\\Pi,\\end{gathered}\\\\\n+\\infty,&\\text{otherwise},\n\\end{cases}\\\\\n\\mathcal E(\\Lambda)\n&=\\begin{cases}\n\\mathcal C(\\alpha),&\\Lambda=\\Pi_\\alpha,\\ I(\\alpha)<\\infty,\\\\\n+\\infty,&\\text{otherwise}.\n\\end{cases}\n\\end{aligned}\n$$\n\nHere we identify $`\\Pi_\\alpha`$ with a measure on $`Z`$ through $`(\\xi,u,d)\\mapsto(\\xi_1,\\xi_2,u,d)`$, using $`\\xi=(x,y,z_v(x,y))`$ on $`\\Omega`$ and assigning zero measure to $`(I^2\\setminus\\Omega)\\times\\mathbb R\\times\\mathbb R^3`$. Because $`\\mathcal E_\\varepsilon`$ includes the whole plan, including transport outside bounded normal windows, minimizing it over plans with target law $`P`$ gives $`\\varepsilon^{-2}W_\\gamma(P,P_0)`$. We formulate the following $`\\Gamma`$-convergence statement in this vague topology [[11](#ref-dalmaso1993)].\n\n\n\n**Theorem 3.1 ($`\\Gamma`$-convergence and the critical risk profile).** Consider the exchangeable box model of [Section 2.2](#dep-model-critical-box).\n\n(i) *Transport plans and compactness.* $`\\mathcal E_\\varepsilon`$ $`\\Gamma`$-converges to $`\\mathcal E`$ as $`\\varepsilon\\downarrow0`$ in the vague topology on $`\\mathscr X`$. Every sequence $`\\varepsilon_n\\downarrow0`$, $`\\Pi_n\\in\\mathscr X`$, with bounded $`\\mathcal E_{\\varepsilon_n}(\\Pi_n)`$ has a vaguely convergent subsequence. If $`\\Pi_n=\\mathscr T_{\\varepsilon_n}\\pi_n\\to\\Pi_\\alpha`$ has bounded energy and $`P_n`$ is the target law of $`\\pi_n`$, then $`\\Phi_{\\varepsilon_n}(P_n)/\\varepsilon_n^2\\to\\mathcal V(\\alpha)`$.\n\n(ii) *Target laws and exact recovery.* For every sequence $`\\varepsilon_n\\downarrow0`$ and every feasible sequence satisfying $`W_\\gamma(P_n,P_0)=O(\\varepsilon_n^2)`$, there are a subsequence and a kernel with finite $`I`$ such that\n\n\n\n$$\n\\mathcal C(\\alpha)\\le\\liminf_n\\frac{W_\\gamma(P_n,P_0)}{\\varepsilon_n^2},\\qquad\n\\frac{\\Phi_{\\varepsilon_n}(P_n)}{\\varepsilon_n^2}\\longrightarrow\\mathcal V(\\alpha).\n\\tag{3.4}\n$$\n\nConversely, every kernel with finite $`I`$ admits a recovery sequence of distributions preserving the original three full marginals, with\n\n\n\n$$\n\\limsup_{\\varepsilon\\downarrow0}\\frac{W_\\gamma(P_\\varepsilon,P_0)}{\\varepsilon^2}\\le\\mathcal C(\\alpha),\\qquad\n\\frac{\\Phi_\\varepsilon(P_\\varepsilon)}{\\varepsilon^2}\\longrightarrow\\mathcal V(\\alpha).\n\\tag{3.5}\n$$\n\n(iii) *The stress profile.* The maximum in [(3.3)](#dep-eq-3-3) is attained. The function $`F`$ is finite, continuous, and nondecreasing, with $`F(0)=-J/p`$ and $`F>0`$ at some finite budget. For every finite $`S_0`$,\n\n\n\n$$\n\\sup_{0\\le s\\le S_0}\\left|\\frac{D_\\varepsilon(s\\varepsilon^2)}{\\varepsilon^2}-F(s)\\right|\\longrightarrow0.\n\\tag{3.6}\n$$\n\n**Proof strategy** Theorem [3.1](#dep-thm-critical-limit)(i) gives $`\\Gamma`$-convergence of the transport energy and convergence of the paired risk along bounded-energy plans, and the exact marginal recovery in part (ii) makes the limiting kernel values asymptotically attainable in $`\\Pi`$. Part (iii) gives uniform convergence of the optimized profiles on compact budget intervals, which, together with the strict crossing in Theorem [3.3](#dep-thm-strict-crossing), locates the reversal budget $`s^*`$.\n\n*Step 1: Localization and the energy lower bound.* Because the two ES thresholds are optimized independently, comparison of the objectives requires a common threshold window for stresses with bounded rescaled energy. In Appendix [A.2](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L1), Lemma [A.2](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-threshold-statement.tex#L1) places every minimizing threshold in an $`O_M(\\varepsilon)`$ window. On bounded windows the lifted plans are locally compact; outside them, marginal conservation requires compensation for the projected local flux. The compensation bound above then gives the energy lower bound for arbitrary admissible plans. Choosing optimal plans yields the corresponding Wasserstein bound for their target laws.\n\n*Step 2: Convergence of the paired risk.* Bounded first moments make each hinge increment integrable, but need not give uniform integrability along convergent plans. The two objectives share the same large-displacement contribution, so we subtract $`\\mathfrak q_0(0)`$ from both. The relative increments are bounded on the common threshold window. Outside a normal window enlarged to account for the bounded reallocation $`H`$, a nonzero relative increment requires a sufficiently large displacement, so the first moment controls the far-field contributions. The baseline integral is finite but need not converge. Since it translates both threshold objectives by the same amount independently of the threshold, it cancels exactly from their difference of minima.\n\n*Step 3: Exact marginal recovery.* Let $`\\mathcal K_{\\rm fin}^c`$ denote probability kernels with compactly supported continuous weights on finitely many fixed displacement vectors and the remaining mass at zero, as in [(A.13)](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L134). For a fixed kernel in this class, local translations yield a preliminary law $`Q_\\varepsilon`$ with uniform marginal CDF expansion $`F_{i,\\varepsilon}-F_i=-\\varepsilon^2j_i+o(\\varepsilon^2)`$, where $`F_i`$ is the reference CDF. The quantile maps $`F_i^{-1}\\circ F_{i,\\varepsilon}`$ restore all three marginals at cost $`\\varepsilon^2\\sum_i\\gamma_i\\|j_i\\|_1+o(\\varepsilon^2)`$, with $`O_\\alpha(\\varepsilon^2)`$ displacements.\n\nTo see why this correction does not alter the leading paired risk, compare its displacement scale with the $`O_\\alpha(\\varepsilon)`$ threshold band. Outside a sufficiently wide band, the paired hinge has gradient $`-\\varepsilon h`$ above both thresholds and zero below both, so multiplying by the restoration displacement gives an $`O_\\alpha(\\varepsilon^3)`$ contribution. Inside the band the gradient is bounded, but the $`Q_\\varepsilon`$-mass is only $`O_\\alpha(\\varepsilon)`$, giving the same order after integration. The restoration error is therefore $`O_\\alpha(\\varepsilon^3)`$, uniformly over all minimizing threshold pairs.\n\nApproximation by $`\\mathcal K_{\\rm fin}^c`$ in cost, paired risk, and the vague plan topology (Appendix A.4), followed by diagonal recovery, gives the general $`o(\\varepsilon^2)`$ error. Compactness, contraction toward the identity, and the common budget-Lipschitz bound then give locally uniform convergence of the optimized profiles.\n\n\n\n### 3.3 Initial linearity and the critical budget\n\nAt zero threshold shifts, the portfolios disagree on tail membership precisely on $`W=\\{u(u+H)<0\\}`$. The displacement $`\\zeta=-(u+H)1_W`$ moves a source in this wedge to the limiting $`A_\\varepsilon`$-boundary and attains the maximal paired hinge increment per unit of normal movement. When $`c_{\\min}=\\gamma_1`$, we can assign this movement along $`e_1`$ provided its downward and upward moments balance on each $`x`$-fiber. On active sources in $`W`$, $`\\zeta<0`$ for $`y0`$ for $`y>x`$, with $`\\int\\zeta\\,du=-H|H|/2`$. Writing $`k(x,y)=f(x,y,(v-x-y)/\\beta)1_{\\{\\beta\\ell_3\n\n$$\n\\ell_-(x)=\\frac1{2\\beta}\\int_I(x-y)_+^2k(x,y)\\,dy,\\qquad\n\\ell_+(x)=\\frac1{2\\beta}\\int_I(y-x)_+^2k(x,y)\\,dy.\n\\tag{3.7}\n$$\n\nExchangeability gives $`\\int_I\\ell_-=\\int_I\\ell_+=J/2`$, but the two moments need not agree on each fiber. Matching their pointwise minimum cancels the projected flux and gives the budget\n\n$$\n\\begin{aligned}\nq_{\\rm bal}(x)&=\\min\\{\\ell_-(x),\\ell_+(x)\\},&\nQ_{\\rm bal}&=\\int_I q_{\\rm bal}(x)\\,dx,\\\\\ns_{\\rm lin}&=2\\gamma_1Q_{\\rm bal}=\\gamma_1J(1-\\chi)>0,&\n\\chi&=\\|\\ell_--\\ell_+\\|_{L^1(I)}/J.\n\\end{aligned}\n$$\n\n\n\n**Proposition 3.2 (Capacity of the initial linear segment).** In the exchangeable box model of Section 2.2, suppose $`\\gamma_1=\\min\\{\\gamma_1,\\gamma_2,\\gamma_3/\\beta\\}`$. With $`\\ell_\\pm`$ as in [(3.7)](#dep-eq-3-7) and $`s_{\\rm lin}=\\gamma_1J(1-\\chi)>0`$ as above,\n\n$$\nF(s)=-J/p+s/(p\\gamma_1),\\qquad 0\\le s\\le s_{\\rm lin}.\n$$\n\n*Proof.* For $`W=\\{u(u+H)<0\\}`$ and $`\\zeta=-(u+H)1_W`$, select the proportion\n\n$$\na(\\xi)=\n\\begin{cases}\nq_{\\rm bal}(x)/\\ell_-(x),&H>0,\\ \\ell_-(x)>0,\\\\\nq_{\\rm bal}(x)/\\ell_+(x),&H<0,\\ \\ell_+(x)>0,\\\\\n0,&\\text{otherwise},\n\\end{cases}\n$$\n\nand define $`\\alpha^\\theta_{\\xi,u}=\\theta a(\\xi)\\delta_{\\zeta e_1}+(1-\\theta a(\\xi))\\delta_0`$ for $`0\\le\\theta\\le1`$. Because $`0\\le a\\le1`$ and the active sources satisfy $`|u|,|\\zeta|\\le\\Delta_I`$, this is a finite-first-moment probability kernel. Its opposite first-coordinate fluxes equal $`-\\theta q_{\\rm bal}(x)`$ and $`\\theta q_{\\rm bal}(x)`$, while the other coordinates remain fixed, so $`j_i=0`$ for all $`i`$ and $`I(\\alpha^\\theta)=\\mathcal C(\\alpha^\\theta)=2\\theta\\gamma_1Q_{\\rm bal}=\\theta s_{\\rm lin}`$.\n\nUnder this displacement, an active source has target normal losses $`-H`$ for $`B`$ and zero for $`A_\\varepsilon`$. Because $`\\operatorname{sgn}u=\\operatorname{sgn}(-H)\\ne0`$ on $`W`$, its strict $`B`$-tail membership is unchanged and no mass is placed at the $`B`$-threshold. For $`A_\\varepsilon`$, however, the active sources from the two signs of $`H`$ arrive at the kink itself. The resulting atom accounts for the jump between the one-sided derivatives: with $`\\mathfrak m_\\pm=\\int_{\\{\\pm H>0\\}}a(\\xi)|H|\\,d\\sigma`$,\n\n$$\nG'_{0,-}(0)=G'_{0,+}(0)=0,\\qquad\nG'_{1,-}(0)=-\\theta \\mathfrak m_-\\le0,\\qquad\nG'_{1,+}(0)=\\theta \\mathfrak m_+\\ge0.\n$$\n\nBy Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101), zero is globally optimal for both objectives. Evaluating the nonzero wedge integrals, $`H^2/2`$ for $`B`$ on $`H<0`$ and $`-H^2/2`$ for $`A_\\varepsilon`$ on $`H>0`$, we obtain $`T_0(0)=\\theta Q_{\\rm bal}`$ and $`T_1(0)=-\\theta Q_{\\rm bal}`$, which yield $`\\mathcal V(\\alpha^\\theta)=-J/p+\\theta s_{\\rm lin}/(p\\gamma_1)`$. Since $`0<\\chi<1`$ by Lemma [A.3](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L52), we may choose $`\\theta=s/s_{\\rm lin}`$ to attain the upper bound inherited from Proposition [2.2](#dep-prop-nominal-benchmark). ∎\n\n**Strict growth** To identify a common leading budget for a tie and a strict reversal, we improve every bounded-first-moment kernel at arbitrarily small additional cost. If the paired risk were concave in the kernel, mixing with a kernel of larger paired risk would do this directly, but the two threshold objectives are minimized separately and their difference need not be concave along mixtures. We also cannot assume that a kernel leaves stationary mass available for an added displacement, since an everywhere nonzero shift proportional to $`e^{-u^2}e_1`$ already has finite first moment.\n\nWe instead replace a small fraction of the existing kernel on two exchanged normal bands beyond the common threshold window. Reflecting $`u`$ to $`-u`$ crosses that window. Pairing the bands cancels their $`B`$-increments, while opposite signs of $`H`$ produce a positive paired gain per unit replacement fraction, independent of the distance $`U`$. After subtraction of the common baseline the old kernel contributes only an $`O_M(U^{-1})`$ remainder, because it takes a large displacement to affect the threshold window from those bands. So we first choose $`U`$ to make this remainder smaller than the reflection gain, and then reduce the replacement fraction to meet the additional budget. Appendix A.6 verifies the mixing obstruction and carries out this comparison uniformly over the original and modified minimizing thresholds.\n\n\n\n**Theorem 3.3 (Strict improvement by local replacement and a common critical coefficient).** Consider the exchangeable box model of [Section 2.2](#dep-model-critical-box). The limiting profile $`F`$ is strictly increasing and has a unique zero $`s^*>0`$, which gives the common leading coefficient of both pre-limit critical budgets at scale $`\\varepsilon^2`$. More precisely:\n\n(i) *Quantitative local improvement.* For every $`0\\le M<\\infty`$ there exists $`c_M>0`$ such that every kernel satisfying $`I(\\alpha)\\le M`$ and every $`0<\\delta\\le1`$ admit a feasible kernel $`\\widehat\\alpha`$ with\n\n\n\n$$\n\\mathcal C(\\widehat\\alpha)\\le\\mathcal C(\\alpha)+\\delta,\\qquad\n\\mathcal V(\\widehat\\alpha)\\ge\\mathcal V(\\alpha)+c_M\\delta.\n\\tag{3.8}\n$$\n\nThese kernels can be chosen as a family $`(\\widehat\\alpha_\\delta)_{0<\\delta\\le1}`$ satisfying\n\n$$\n\\int |d|_\\gamma\\,|\\widehat\\alpha_\\delta-\\alpha|(dd)\\,d\\nu\n\\xrightarrow[\\delta\\downarrow0]{}0,\n\\qquad\n\\lim_{\\delta\\downarrow0}\n\\frac{\\mathcal V(\\widehat\\alpha_\\delta)-\\mathcal V(\\alpha)}{\\delta}\n\\ge c_M>0.\n$$\n\nFor every finite $`S_0>0`$,\n\n\n\n$$\nF(t)-F(s)\\ge c_{S_0}(t-s)>0\\qquad(0\\le s0`$. Over the finite-first-moment kernel class,\n\n\n\n$$\ns^*=\\min_{\\alpha:\\,\\mathcal V(\\alpha)\\ge0}\\mathcal C(\\alpha)\n=\\min_{\\alpha:\\,\\mathcal V(\\alpha)=0}\\mathcal C(\\alpha)\n=\\inf_{\\alpha:\\,\\mathcal V(\\alpha)>0}\\mathcal C(\\alpha).\n\\tag{3.10}\n$$\n\nThe corresponding pre-limit critical budgets satisfy\n\n\n\n$$\n\\rho^-_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2),\\qquad\n\\rho^+_\\varepsilon=s^*\\varepsilon^2+o(\\varepsilon^2).\n\\tag{3.11}\n$$\n\nThe first two minima are attained, whereas the infimum over strictly positive risk values is not.\n\n\n\n**Execution costs** A positive execution cost proportional to $`\\varepsilon`$ exceeds the second-order nominal gain for small trades. To retain a critical-budget comparison, suppose instead that the deterministic cost paid to implement the new allocation $`B`$ satisfies $`c_{\\mathrm{exec},\\varepsilon}\\ge0`$ and $`c_{\\mathrm{exec},\\varepsilon}/\\varepsilon^2\\to\\theta\\in[0,J/p)`$. With the stress class and metric fixed, cash invariance gives $`D^{\\rm net}_\\varepsilon=D_\\varepsilon+c_{\\mathrm{exec},\\varepsilon}`$, whose rescaled profile converges locally uniformly to $`F+\\theta`$.\n\nApplying the tie and strict-reversal definitions in [(2.4)](#dep-eq-2-4) to the net profile, we obtain by strict growth and local uniform convergence $`\\rho^{\\pm,\\rm net}_\\varepsilon=s_\\theta\\varepsilon^2+o(\\varepsilon^2)`$, where $`s_\\theta`$ is the unique solution of $`F(s_\\theta)=-\\theta`$. As the rescaled execution cost $`\\theta`$ increases to $`J/p`$, this solution decreases continuously from $`s^*`$ to zero. For example, in the cheap-background case of Section 4.1, the explicit linear profile gives $`s_\\theta=c_3(J-p\\theta)`$.\n\n\n\n## 4. Efficiency premiums and projected flow duality\n\nThe critical coefficient reaches the local-efficiency lower bound exactly when background displacement is cheapest per unit change in portfolio loss, and in every other case a strict marginal-restoration premium remains. This section proves that classification and then characterizes the right derivative of the critical budget in the background penalty at equal effective costs. A locally efficient displacement need not preserve the marginal projections, so any unbalanced moment has to be compensated elsewhere. We measure this obstruction by testing each projected flux against a bounded measurable coordinate potential, using the $`L^1`$ dual representation of the restoration cost:\n\n$$\n\\mathcal C(\\alpha)-I(\\alpha)\n=\\sup_{|\\phi_i|\\le\\gamma_i}\\sum_i\\int_{I_i}\\phi_i(q)j_i(q)\\,dq\n=\\sup_{|\\phi_i|\\le\\gamma_i}\\int\\sum_i\\phi_i(\\xi_i)d_i\\,d\\Pi_\\alpha.\n$$\n\nHere $`\\phi_i(q)`$ acts as a dual shadow price for the projected displacement moment at coordinate level $`q`$. For the projections to vanish, opposite signed displacements have to balance at each coordinate level, and then the leading restoration penalty disappears. We first identify the pointwise displacement required for equality in the efficiency bound, and then test its projections near an off-diagonal corner to rule out equality when only traded coordinates are cheapest. At equal effective costs, continuous potentials lead to the flow–cut duality of Section 4.2, and allowing nonzero projected fluxes in the restoration-cost bound leads to the piecewise-constant potential estimates of Section 5.1.\n\n\n\n### 4.1 Local efficiency and the zero-premium classification\n\nWith $`c_1=\\gamma_1`$, $`c_2=\\gamma_2`$, $`c_3=\\gamma_3/\\beta`$, and $`c_{\\min}=\\min\\{c_1,c_2,c_3\\}`$, Proposition [2.2](#dep-prop-nominal-benchmark) and Theorem [3.1](#dep-thm-critical-limit) already give $`F(s)\\le-J/p+s/(pc_{\\min})`$. The issue at zero risk difference is whether a kernel can satisfy the pointwise efficiency conditions while conserving every marginal projection, and we characterize those simultaneous equality conditions below.\n\nProposition [2.2](#dep-prop-nominal-benchmark) gives initial efficiency at each fixed trade size under [(2.9)](#dep-eq-2-9), and when $`\\gamma_1=c_{\\min}`$, Proposition [3.2](#dep-prop-balanced-linear-profile) separately realizes the limiting slope by matching opposite first-coordinate moments on each fiber. When only traded coordinates are cheapest, that matching is enough for the initial linear segment but cannot cover the whole tail section. If the background coordinate is cheapest, exchanging $`x,y`$ and reversing $`u`$ instead balances the required displacement on every background slice, which sustains the efficiency bound up to zero risk.\n\n\n\n**Lemma 4.1 (Local-efficiency bound and critical-level optimality).** In the exchangeable box model of Section 2.2, every kernel with $`I(\\alpha)<\\infty`$ satisfies $`p\\mathcal V(\\alpha)+J\\le\\mathcal C(\\alpha)/c_{\\min}`$, so $`s^*\\ge c_{\\min}J`$. Set $`b=(1,1,\\beta)`$ and $`\\zeta=-(u+H)1_{\\{u(u+H)<0\\}}`$. Then $`s^*=c_{\\min}J`$ if and only if there is a measurable probability kernel satisfying\n\n\n\n$$\n\\begin{aligned}\nb\\cdot d=\\zeta,\\qquad |d|_\\gamma=c_{\\min}|\\zeta|\n&\\quad\\Pi_\\alpha\\text{-almost everywhere},\\\\\n(\\xi_i)_\\#(d_i\\,d\\Pi_\\alpha)=0\n&\\quad i=1,2,3.\n\\end{aligned}\n\\tag{4.1}\n$$\n\nSuch a kernel has $`\\mathcal C=I=c_{\\min}J`$ and $`\\mathcal V=0`$; every kernel with $`\\mathcal V\\ge0`$ and $`\\mathcal C\\le c_{\\min}J`$ satisfies these conditions.\n\n*Proof.* Set $`w=b\\cdot d`$, $`\\Delta_+(t,w)=(t+w)_+-t_+`$, and choose any $`\\tau_A\\in\\arg\\min G_1`$. To compare the risk gain with its transport cost, we evaluate the $`B`$-objective at zero, bound the difference of hinge increments by $`|w|`$, and use $`c_{\\min}|w|\\le |d|_\\gamma`$. Recording the loss in each comparison gives\n\n\n\n$$\n\\begin{aligned}\n\\frac{\\mathcal C}{c_{\\min}}-(p\\mathcal V+J)\n={}&\\underbrace{\\frac{\\mathcal C-I}{c_{\\min}}}_{\\mathrm{I}}\n+\\underbrace{\\left(\\frac I {c_{\\min}}-\\int |w|\\,d\\Pi_\\alpha\\right)}_{\\mathrm{II}}\n+\\underbrace{\\frac g2\\tau_A^2}_{\\mathrm{III}}\\\\\n&+\\underbrace{\\int[|w|-\\Delta_+(u,w)+\\Delta_+(u+H-\\tau_A,w)]\\,d\\Pi_\\alpha}_{\\mathrm{IV}}\\\\\n&+\\underbrace{G_0(0)-\\min_{\\tau_B}G_0(\\tau_B)}_{\\mathrm{V}}.\n\\end{aligned}\n\\tag{4.2}\n$$\n\nThe five terms (I)–(V) in [(4.2)](#dep-eq-4-2) have distinct roles: (I) the marginal-restoration penalty; (II) the excess transport cost over the cheapest normal movement, including cancellation between coordinate displacements; (III) the curvature cost of shifting the $`A_\\varepsilon`$-threshold; (IV) the unsaturated hinge gain; and (V) the gap from evaluating the $`B`$-objective at zero instead of its minimum. For the fourth term, $`\\Delta_+(t,w)\\in[\\min\\{0,w\\},\\max\\{0,w\\}]`$, an interval of length $`|w|`$, so the difference of the two hinge increments cannot exceed $`|w|`$. All five terms are consequently nonnegative, proving the efficiency bound. Since $`\\tau_A`$ was an arbitrary minimizer, equality forces each defect to vanish for every optimal $`A_\\varepsilon`$-threshold.\n\nTo reach zero risk at the local-efficiency budget, none of these defects can remain. Indeed, if $`\\mathcal V\\ge0`$ and $`\\mathcal C\\le c_{\\min}J`$, the left side of [(4.2)](#dep-eq-4-2) is nonpositive, so every nonnegative term must vanish and\n\n\n\n$$\n\\mathcal V=0,\\qquad \\mathcal C=I=c_{\\min}J,\\qquad\n\\int|w|\\,d\\Pi_\\alpha=J,\\qquad j_i=0\\quad(i=1,2,3).\n\\tag{4.3}\n$$\n\nIn addition, $`\\arg\\min G_1=\\{0\\}`$ and $`0\\in\\arg\\min G_0`$. The vanishing metric defect\n\n\n\n$$\n|d|_\\gamma-c_{\\min}|b\\cdot d|\n=\\sum_i(\\gamma_i-c_{\\min}b_i)|d_i|\n+c_{\\min}\\left(\\sum_i b_i|d_i|-\\left|\\sum_i b_id_i\\right|\\right)\n\\tag{4.4}\n$$\n\nallows only cheapest coordinates, with all nonzero components having the sign of $`w`$. At $`\\tau_A=0`$, equality in the hinge bound requires $`u\\ge0`$, $`u+H\\le-w`$ if $`w>0`$, and $`u\\le0`$, $`u+H\\ge-w`$ if $`w<0`$. Thus $`|w|\\le K:=|u+H|1_{\\{u(u+H)<0\\}}`$, with the sign of $`\\zeta`$ wherever $`w\\ne0`$. Since\n\n\n\n$$\n\\int_{\\mathbb R}K\\,du=H^2/2,\\qquad\n\\int_{\\mathbb R}\\zeta\\,du=-H|H|/2,\n\\tag{4.5}\n$$\n\nand every conditional kernel has mass one, $`\\int K\\,d\\Pi_\\alpha=J=\\int|w|\\,d\\Pi_\\alpha`$. Saturation of this nonnegative bound forces\n\n\n\n$$\nb\\cdot d=\\zeta\\quad\\Pi_\\alpha\\text{-almost everywhere},\n\\tag{4.6}\n$$\n\nwhich, with the metric and flux equalities, proves necessity.\n\nFor the converse, the threshold calculation depends only on the normal displacement $`w=\\zeta`$. We integrate over the disagreement wedge and use exchangeability to obtain\n\n\n\n$$\nT_0(0)=J/2,\\qquad T_1(0)=-J/2,\\qquad\n\\int_{H>0}H^2\\,d\\sigma=\\int_{H<0}H^2\\,d\\sigma=J.\n\\tag{4.7}\n$$\n\nBecause the transformed $`B`$-loss retains its strict tail sign and places no atom at zero, we have $`G_0'(0)=0`$. With $`m_\\pm=\\int(\\pm H)_+\\,d\\sigma>0`$, the mass placed at the $`A_\\varepsilon`$-threshold kink gives\n\n\n\n$$\nG_{1,-}'(0)=-m_-<0,\\qquad G_{1,+}'(0)=m_+>0.\n\\tag{4.8}\n$$\n\nBy convexity, these strict one-sided inequalities make zero the unique minimizer of $`G_1`$. For $`G_0`$, the stationary regions $`\\{H>0,u>0\\}`$ and $`\\{H<0,u<0\\}`$ each have section mass $`g/2`$, and their unchanged normal coordinates cover the positive and negative half-lines, respectively. Hence $`\\lambda_0\\ge(g/2)\\,dt`$ on $`\\mathbb R`$, and Lemma [A.4](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101) yields $`G_0(\\tau)-G_0(0)\\ge g\\tau^2/4`$, so zero is also the unique minimizer of $`G_0`$. Combining these minima with [(4.7)](#dep-eq-4-7) gives $`\\mathcal V=0`$. Under [(4.1)](#dep-eq-4-1), we have $`I=c_{\\min}J`$ and vanishing fluxes, so $`\\mathcal C=c_{\\min}J`$. Together with attainment of the critical cost in Theorem [3.3](#dep-thm-strict-crossing), this proves the equivalence. ∎\n\n\n\n**Theorem 4.2 (Zero-premium classification).** For the exchangeable box model of [Section 2.2](#dep-model-critical-box) and every positive weight vector,\n\n\n\n$$\ns^*=Jc_{\\min}\\quad\\Longleftrightarrow\\quad\nc_3\\le\\min\\{\\gamma_1,\\gamma_2\\}.\n\\tag{4.9}\n$$\n\nIf $`c_3=c_{\\min}`$, the profile is explicitly $`F(s)=-J/p+s/(pc_3)`$ for $`0\\le s\\le c_3J`$.\n\nIf $`\\gamma_2\\ge\\gamma_1`$ and $`c_3>\\gamma_1`$, then\n\n\n\n$$\n1<\\frac{s^*}{\\gamma_1J}\\le\\min\\{1+\\chi,c_3/\\gamma_1\\}<2,\n\\tag{4.10}\n$$\n\nwhere $`\\chi=\\|\\ell_--\\ell_+\\|_{L^1(I)}/J\\in(0,1)`$ is the projected imbalance of Section 3.3.\n\n*Proof of Theorem [4.2](#dep-thm-zero-premium).*\n\nBy Lemma [4.1](#dep-lem-efficiency-bound), it suffices to allocate the prescribed displacement $`\\zeta`$ among cheapest coordinates while preserving every marginal projection. The background coordinate admits such a balance because exchanging the traded losses and reversing the normal coordinate leaves each background slice fixed and reverses the required displacement. For the traded coordinates, the corresponding balance must instead be found along each $`x`$- or $`y`$-fiber. Near an off-diagonal corner, the required moment is bounded below by a positive multiple of $`\\delta^2`$, whereas the opposite-side sources available at those coordinate levels have capacity only $`O(\\delta^4)`$. We prove the background construction first and then use this discrepancy to exclude even randomized splitting among the traded directions.\n\n\n\n**Comparison kernels and the sufficient direction** Assigning the prescribed displacement to one coordinate gives\n\n\n\n$$\n\\alpha^{(1)}=\\delta_{\\zeta e_1},\\qquad\n\\alpha^{(2)}=\\delta_{\\zeta e_2},\\qquad\n\\alpha^{(3)}=\\delta_{(\\zeta/\\beta)e_3}\n\\tag{4.11}\n$$\n\nwith finite first moments and zero risk differences by the normal-displacement calculation [(4.7)](#dep-eq-4-7)–[(4.8)](#dep-eq-4-8) in Lemma [4.1](#dep-lem-efficiency-bound). The first two kernels have respective fluxes $`j_1=\\ell_+-\\ell_-`$ and $`j_2=\\ell_--\\ell_+`$, all other fluxes being zero. To verify the background projection, we use the exchange $`(x,y,u)\\mapsto(y,x,-u)`$, which preserves $`\\nu,\\xi_3`$ and reverses $`\\zeta`$. Absolute integrability then allows us to test against every bounded measurable $`\\varphi`$, obtaining\n\n\n\n$$\n\\int\\varphi(\\xi_3)d_3^{(3)}d\\nu\n=-\\int\\varphi(\\xi_3)d_3^{(3)}d\\nu=0.\n\\tag{4.12}\n$$\n\nAll three projections of $`\\alpha^{(3)}`$ vanish, yielding\n\n\n\n$$\n\\mathcal C(\\alpha^{(1)})=\\gamma_1J(1+\\chi),\\quad\n\\mathcal C(\\alpha^{(2)})=\\gamma_2J(1+\\chi),\\quad\n\\mathcal C(\\alpha^{(3)})=c_3J.\n\\tag{4.13}\n$$\n\nThe background kernel proves sufficiency, including ties, and the least of these costs gives $`s^*\\le J\\min\\{\\gamma_1(1+\\chi),\\gamma_2(1+\\chi),c_3\\}`$, which yields the upper bound in [(4.10)](#dep-eq-4-10) when $`\\gamma_2\\ge\\gamma_1`$ and $`c_3>\\gamma_1`$.\n\nWhen $`c_3=c_{\\min}`$, mix the background kernel $`\\alpha^{(3)}`$ with the identity using weight $`\\theta=s/(c_3J)\\in[0,1]`$. Both kernels have zero projected flux and both threshold objectives attain their minima at zero, by [(4.7)](#dep-eq-4-7)–[(4.8)](#dep-eq-4-8) and $`G_c^{\\delta_0}(\\tau)=g\\tau^2/2`$, so cost and paired risk are affine along the mixture. Its cost is $`\\theta c_3J=s`$, and [(4.7)](#dep-eq-4-7) gives\n\n\n\n$$\n\\mathcal V=-(1-\\theta)J/p.\n\\tag{4.14}\n$$\n\nThus the mixture attains the efficiency bound on the stated interval.\n\n\n\n**The obstruction for arbitrary randomized kernels** For necessity, suppose $`c_3>c_{\\min}`$ and a kernel with $`\\mathcal V\\ge0`$ attains cost $`c_{\\min}J`$. Lemma [4.1](#dep-lem-efficiency-bound) then requires $`d_3=0`$, $`d_1+d_2=\\zeta`$, and both traded components to have the sign of $`\\zeta`$. Since the kernel has finite first moment, its averaged displacement fields $`m_i(x,y)=\\int_{\\mathbb R}\\int d_i\\,\\alpha_{\\xi,u}(dd)\\,du`$, $`i=1,2`$, are integrable on the section and satisfy, by [(4.1)](#dep-eq-4-1),\n\n\n\n$$\n\\begin{aligned}\n&m_1+m_2=-H|H|/2,\\qquad |m_i|\\le H^2/2,\\\\\n&m_i\\ge0\\quad\\text{for}\\ xy.\n\\end{aligned}\n\\tag{4.15}\n$$\n\nProjected conservation gives, for every pair of Borel sets $`A,B\\subset I`$,\n\n\n\n$$\n\\int_{\\Omega\\cap\\{x\\in A\\}}m_1d\\sigma=0,\\qquad\n\\int_{\\Omega\\cap\\{y\\in B\\}}m_2d\\sigma=0.\n\\tag{4.16}\n$$\n\nWe test [(4.15)](#dep-eq-4-15)–[(4.16)](#dep-eq-4-16) near the upper-left off-diagonal corner of $`\\Omega`$, where $`x`$ is small, $`y`$ is large, and the positive moment $`(y-x)^2/2`$ is bounded away from zero. Compensating this moment at the same coordinate levels requires negative moment from $`\\{x>y\\}`$, where the small value of $`x`$ or large value of $`y`$ confines available sources to a small triangle near the diagonal (Figure [2](#dep-fig-corner)).\n\n\n\nThe section base is the intersection of an open square with a diagonal strip, $`\\Omega=(\\ell,r)^2\\cap\\{q_-0.\n\\end{gathered}\n$$\n\nBoth coordinate projections of $`\\Omega`$ equal $`(\\xi_-,\\xi_+)`$, and the interior diagonal point from [Section 2.3](#dep-geometry-interior-diagonal) satisfies $`\\xi_-0`$. Moreover, $`q_*:=\\xi_-+\\xi_+\\in[q_-,q_+]`$, so the corner $`(\\xi_-,\\xi_+)`$ lies in the closed strip, possibly on its boundary. The square diagonal through that corner runs parallel to the strip boundaries. On the corner square $`(x,y)=(\\xi_-+t,\\xi_+-z)`$, $`00`$, independent of $`\\delta`$, such that\n\n\n\n$$\nM_\\delta:=\\int_{E_\\delta}(m_1+m_2)d\\sigma\n\\ge c_\\Omega\\delta^2.\n\\tag{4.17}\n$$\n\nTesting [(4.16)](#dep-eq-4-16) on $`A_\\delta=(\\xi_-,\\xi_-+\\delta)`$ and $`B_\\delta=(\\xi_+-\\delta,\\xi_+)`$ requires $`M_\\delta\\le N_1+N_2`$, where\n\n$$\nN_1=\\int_{\\Omega\\cap\\{x\\in A_\\delta,\\ x>y\\}}(-m_1)d\\sigma,\n\\qquad\nN_2=\\int_{\\Omega\\cap\\{y\\in B_\\delta,\\ x>y\\}}(-m_2)d\\sigma.\n$$\n\nIn $`N_1`$, the conditions $`x\\in A_\\delta`$ and $`y\n\n$$\nN_1+N_2\\le C_\\Omega\\int_0^\\delta\\int_0^x(x-y)^2\\,dy\\,dx\n=\\frac{C_\\Omega}{12}\\delta^4.\n\\tag{4.18}\n$$\n\nHere $`C_\\Omega<\\infty`$ is independent of $`\\delta`$. For small $`\\delta`$, the bound contradicts $`c_\\Omega\\delta^2\\le M_\\delta\\le N_1+N_2`$. Attainment of the critical cost then implies $`s^*>Jc_{\\min}`$ when $`c_3>c_{\\min}`$, and [(4.13)](#dep-eq-4-13) gives [(4.10)](#dep-eq-4-10). ∎\n\n\n\n![Figure 2. Corner demand and return capacity, illustrated on $`I=I_3=[0,2]`$ for $`\\beta=7/4`$, $`v=15/4`$, $`\\delta=0.35`$. The blue region $`E_\\delta`$ stays away from the diagonal and has demand at least a positive multiple of $`\\delta^2`$. Return sources lie in the hatched triangles, which may extend beyond $`\\Omega`$; their area and squared displacement are both $`O(\\delta^2)`$, giving capacity $`O(\\delta^4)`$. The bounds shown, with $`\\lambda=\\min f`$, $`\\Lambda=\\max f`$, and $`\\Delta=\\xi_+-\\xi_-`$, violate $`M_\\delta\\le N_1+N_2`$ for sufficiently small positive $`\\delta`$.](/assets/research/dependence/corner-capacity.png)\n\n\n\n### 4.2 Boundary sensitivity and the flow–cut representation\n\nAt equal effective costs, every efficient kernel has total required displacement $`J`$. At this critical level, the equality conditions and total-moment saturation in Lemma [4.1](#dep-lem-efficiency-bound) fix $`b\\cdot d=\\zeta`$ pointwise almost everywhere, so normal integration and exchange symmetrization reduce the remaining minimum-background allocation problem to shares on the two-dimensional half-section. Raising only the background penalty selects, to first order, the kernels that use the least background displacement, and we can calculate that amount by assigning as much displacement as possible to the traded coordinates under their projected conservation law. On $`D=\\Omega\\cap\\{x\n\n**Theorem 4.3 (Boundary derivative and projected flow duality).** Fix the reference model and $`p,\\beta,\\gamma`$, and set\n\n$$\n\\boldsymbol\\gamma(\\eta)=(\\gamma,\\gamma,\\beta\\gamma(1+\\eta)),\\quad\n\\pi(\\eta)=s^*(\\boldsymbol\\gamma(\\eta))/(\\gamma J)-1,\\quad \\eta\\ge0.\n$$\n\nLet $`\\mathcal C_0`$ be the cost at $`\\eta=0`$, and let $`\\mathcal Z_0=\\{\\alpha:I(\\alpha)<\\infty,\\ \\mathcal V(\\alpha)=0,\\ \\mathcal C_0(\\alpha)=\\gamma J\\}`$. When kernels are identified with their measures $`\\Pi_\\alpha`$, the set $`\\mathcal Z_0`$ is nonempty and compact in the vague topology on $`\\mathscr X`$ from Section 3.2. The following minimum is attained, and\n\n\n\n$$\n\\pi(\\eta)=a_+\\eta+o(\\eta),\\qquad\na_+=\\frac\\beta J\\min_{\\alpha\\in\\mathcal Z_0}\n\\int|d_3|\\,d\\Pi_\\alpha\\in[0,1].\n\\tag{4.19}\n$$\n\nOn $`D=\\Omega\\cap\\{x\n\n$$\na_+=1-\\frac{2\\mathfrak F_\\mu}{J}\\in(0,1),\\qquad\n\\mathfrak F_\\mu=\\inf_{\\substack{\\varphi\\text{ measurable}\\\\|\\varphi|\\le1}}\n\\int_D\\max\\{0,1-\\varphi(x),1+\\varphi(y)\\}\\,d\\mu.\\tag{4.20}\n$$\n\nEquivalently, $`a_+`$ is the supremum of the projected cut deficits\n\n\n\n$$\na_+=\\frac2J\\sup_{U\\cap V=\\varnothing}\n\\left[\\mu(D\\cap(U\\times V))-\n\\mu\\bigl(D\\cap((V\\times I)\\cup(I\\times U))\\bigr)\\right].\\tag{4.21}\n$$\n\nHere $`U,V\\subset I`$ are disjoint measurable sets.\n\nIn words, $`2\\mathfrak F_\\mu`$ is the largest normal-displacement moment that the traded coordinates can accommodate while balancing their marginal projections. The cut formula [(4.21)](#dep-eq-4-21) writes $`a_+`$ as $`2/J`$ times the maximal deficit between the moment required on $`D\\cap(U\\times V)`$ and the compensating capacity on $`D\\cap((V\\times I)\\cup(I\\times U))`$.\n\n*Proof sketch.* For $`\\alpha\\in\\mathcal Z_0`$, normal integration and exchange symmetrization yield shares satisfying $`(x)_\\#(a\\mu)=(y)_\\#(b\\mu)`$ and background amount $`\\beta\\int|d_3|\\,d\\Pi_\\alpha=J-2\\int_D(a+b)\\,d\\mu`$. Conversely, every feasible pair of shares lifts to a kernel in $`\\mathcal Z_0`$, so the minimum background amount is $`J-2\\mathfrak F_\\mu`$ (Appendix [A.8](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-flow-proof.tex#L1)). Dividing by $`J`$ gives $`a_+`$, and Appendix [A.7](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-cost-proof.tex#L1) establishes [(4.19)](#dep-eq-4-19) by comparing perturbed minimizers with a least-background kernel. Sion's theorem with continuous coordinate potentials gives [(4.20)](#dep-eq-4-20); signed layer-cake reduction then gives [(4.21)](#dep-eq-4-21).\n\n\n\n**Dependence on the transport weights** Varying the positive transport weights leaves the finite-first-moment kernel class and $`\\mathcal V`$ unchanged. With $`A_i(\\alpha)=\\int|d_i|\\,d\\Pi_\\alpha+\\|j_i\\|_1`$, the coordinate cost including marginal restoration, we may therefore write\n\n\n\n$$\ns^*(\\gamma)=\\inf_{\\mathcal V(\\alpha)\\ge0}\\sum_i\\gamma_i A_i(\\alpha).\n\\tag{4.22}\n$$\n\nBecause the feasible class is independent of $`\\gamma`$, this infimum of nonnegative linear functions is positively homogeneous, coordinatewise nondecreasing, and concave. In particular, $`r_-\\gamma\\le\\widetilde\\gamma\\le r_+\\gamma`$, with $`0\\max\\{\\widetilde\\gamma_1,\\widetilde\\gamma_2\\}`$ and therefore retains a strict premium.\n\n\n\n## 5. Analytical illustrations and financial implications\n\n\n\n### 5.1 Projection potentials and a uniform reference\n\nAssigning a bounded shadow price $`\\phi_i`$ to each marginal flux lets us bound the strict premium without solving the full risk optimization. For a prescribed normal displacement we first minimize the adjusted coordinate costs $`\\gamma_i\\pm\\phi_i(\\xi_i)`$, and then we compare that minimum with the paired hinge gain.\n\n\n\n**Proposition 5.1 (A projection-potential bound).** In the exchangeable box model, let $`b=(1,1,\\beta)`$ and let $`|\\phi_i|\\le\\gamma_i`$ be bounded measurable potentials on the coordinate intervals such that $`k_\\pm(\\xi)=\\min_{1\\le i\\le3}(\\gamma_i\\pm\\phi_i(\\xi_i))/b_i`$ are bounded below by a positive constant. For every finite-first-moment kernel, every $`\\tau\\in\\arg\\min G_1`$, and every $`\\lambda>0`$,\n\n\n\n$$\n\\mathcal C(\\alpha)\\ge\n\\lambda\\left(J+p\\mathcal V+\\frac g2\\tau^2\\right)\n-\\frac{\\lambda}{2}\\int_\\Omega(H-\\tau)^2\n\\left(\\frac{\\lambda}{k(\\xi,H-\\tau)}-1\\right)_+\\,d\\sigma,\n\\tag{5.1}\n$$\n\nwhere $`k(\\xi,h)=k_+(\\xi)`$ for $`h<0`$ and $`k_-(\\xi)`$ for $`h>0`$. The integrand is zero when $`h=0`$.\n\nThe proof in [Appendix B.1](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L3) optimizes the hinge gain against the adjusted cost $`k_+w_++k_-(-w)_+`$, where $`w=b\\cdot d`$. To quantify how far the reversal budget can exceed the initial-slope prediction, we use an analytically tractable reference with independent uniform losses on $`[0,2]`$, $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(\\gamma_0,\\gamma_0,2\\gamma_0)`$. Equal exposures and the median tail make the section invariant under central reflection, so $`v=3`$, $`d\\sigma=dx\\,dy/8`$, $`g=3/8`$, and $`J=5/32`$.\n\nIn the cut representation of Theorem [4.3](#dep-thm-projected-flow), the midpoint partition corresponds to $`\\varphi(q)=1_{[0,1)}(q)-1_{(1,2]}(q)`$. Scaling this potential by $`\\gamma_0/2`$ and using opposite signs on the two traded coordinates carries the same partition over to the marginal-flux penalty in Proposition [5.1](#dep-prop-projection-potential-bound), giving the admissible two-level potentials\n\n$$\n\\phi_1(x)=\\frac{\\gamma_0}{2}\\operatorname{sgn}(1-x),\\qquad\n\\phi_2(y)=\\frac{\\gamma_0}{2}\\operatorname{sgn}(y-1),\\qquad \\phi_3=0.\n$$\n\nTheir signs distinguish low-loss from high-loss coordinate halves. At zero threshold shift, a positive paired hinge gain requires normal movement $`w=b\\cdot d>0`$ when $`xy`$. With amplitude $`\\gamma_0/2`$, the effective cost in this direction across the midpoint is $`k(\\xi,H)=3\\gamma_0/2`$, and both directional costs on either same-side region are $`\\gamma_0/2`$. Choosing $`\\lambda=3\\gamma_0/2`$ in [(5.1)](#dep-eq-5-1) therefore removes the cross-region penalty at zero threshold. The proof below keeps the correction for shifted minimizing thresholds, which makes the bound valid without prescribing either ES threshold.\n\n\n\n**Corollary 5.2 (Uniform reference with costly background displacement).** For this uniform reference,\n\n\n\n$$\n\\frac{13}{64}\\le\\frac{s^*}{\\gamma_0}\\le\\frac{29}{96},\n\\qquad \\frac{s^*}{\\gamma_0J}\\ge\\frac{13}{10}.\n\\tag{5.2}\n$$\n\nThe initial linear profile holds on $`0\\le s\\le\\gamma_0/96`$.\n\n*Proof.* For $`1/2\\le x\\le1`$, the negative-moment section has $`1-x\n\nSetting $`\\gamma_0=1`$, we obtain analytical envelopes for $`F(s)`$ on $`[0,U]`$, where $`U=29/96`$. The balanced kernel has cost $`s_0=1/96`$, zero projected flux, and risk $`-7/24`$, and the first-coordinate comparison kernel has cost $`U`$ and risk zero. Their convex mixtures have affine paired risk because zero minimizes both threshold objectives for each kernel, and affine cost because the balanced kernel has zero flux. Joining these mixtures to the exact initial linear segment realizes the lower envelope\n\n$$\n\\underline F(s)=\n\\begin{cases}\n2s-5/16,&0\\le s\\le1/96,\\\\\ns-29/96,&1/96\\le s\\le29/96.\n\\end{cases}\n$$\n\nTo bound $`F`$ from above, we invert the two threshold estimates and combine them with the efficiency bound:\n\n$$\n\\begin{aligned}\nF(s)&\\le\\min\\left\\{2s-\\frac5{16},\n\\max\\left\\{\\frac43s-\\frac{13}{48},2s-\\frac{11}{16}\\right\\}\\right\\}\n\\\\\n&=\\overline F(s):=\\min\\left\\{2s-\\frac5{16},\\frac43s-\\frac{13}{48}\\right\\}\n\\qquad(0\\le s\\le U).\n\\end{aligned}\n$$\n\nThe last equality holds because the branches inside the maximum meet only at $`s=5/8>U`$. Figure [3](#dep-fig-budget) plots the analytical envelopes $`\\underline F\\le F\\le\\overline F`$ on $`[0,U]`$. The guaranteed linear segment $`[0,s_{\\rm lin}]`$ covers one fifteenth of the tangent prediction $`J=5/32`$, whereas the critical coefficient $`s^*`$ lies between $`13/10`$ and $`29/15`$ times $`J`$. So the initial slope is attained on a positive interval, and the required reversal budget is still at least $`30\\%`$ higher than its tangent prediction.\n\n\n\n![Figure 3. Proved risk envelopes for the independent uniform model with $`\\beta=1`$, $`p=1/2`$, and $`\\gamma=(1,1,2)`$. Explicit mixtures attain the lower envelope; the upper envelope follows from midpoint potentials and the efficiency bound. The inset shows the exact initial interval $`[0,1/96]`$ in green. The dashed line extrapolates its slope to $`J=5/32`$, whereas the critical budget lies in $`[13/64,29/96]`$. Both plotted envelopes are piecewise linear analytical formulas.](/assets/research/dependence/budget-comparison.png)\n\n**Numerical approximation** The computational companion also discretizes a symmetric family of local kernels, using pure-coordinate moves to the two hinge breakpoints and retaining the absolute projected-flux penalty in [(3.1)](#dep-eq-3-1). With both thresholds centered, surface grids of 8, 12, 16, 24 and 32 points per coordinate place the zero of this family near $`0.23`$; the individual estimates increase with the grid from $`0.2292`$ to $`0.2324`$. The calculation reproduces the proved initial linear segment, and the paired risk of each computed kernel, evaluated again at its own minimizing thresholds, vanishes to rounding error. Because the family lies inside the admissible class, its zero is an upper estimate of $`s^*`$ at the stated grid resolution. On this evidence $`s^*`$ falls in the lower third of the interval in [(5.2)](#dep-eq-5-2), and the reversal premium over the tangent prediction lies between the proved $`30\\%`$ and a computed figure of about $`49\\%`$. The companion's numerical study records the normal-window checks and solver residuals behind that figure.\n\nFor finite trade sizes, the companion solves the discretized problem globally on cubic midpoint lattices with up to 24 points per marginal, for trade sizes from $`\\varepsilon=0.2`$ down to $`\\varepsilon=0.025`$. As the trade size decreases, the quadratic scale falls below the lattice spacing, so the recorded values measure discretization sensitivity; a convergence study of [(3.6)](#dep-eq-3-6) refines the lattice with the trade size.\n\n\n\n### 5.2 A dependence family with fixed marginal laws\n\nSection 5.1 used an independent uniform reference with background displacement twice as costly as either traded direction. We now move to the equal-effective-cost boundary to see how changing the reference dependence affects the required background share $`a_+`$ and the budget derivative $`\\gamma_0Ja_+`$. Keeping uniform marginals, $`\\beta=1`$, and $`p=1/2`$, we perturb the independent density by the product of the centered losses $`x+y-2`$ and $`z-1`$:\n\n$$\nf_\\vartheta(x,y,z)=\\frac{1+\\vartheta(x+y-2)(z-1)}8,\\qquad\n|\\vartheta|<\\frac12,\\qquad\nE_\\vartheta[(X_1+X_2-2)(X_3-1)]=\\frac{2\\vartheta}{9}.\n$$\n\nAfter rescaling the losses to $`[0,1]`$, this is a one-parameter subfamily of the trivariate Farlie–Gumbel–Morgenstern family [[27](#ref-otakimura2021), Section 2]: the interactions between each traded coordinate and the background have parameter $`\\vartheta`$, and the other interaction parameters are zero. Centering preserves every marginal, and the parameter restriction keeps the density strictly positive on the cube. Central reflection fixes $`v=3`$, where the section density is $`[1-\\vartheta(x+y-2)^2]/8`$, so increasing $`\\vartheta`$ reduces the section mass away from $`x+y=2`$ even as covariance increases. On the unchanged half-section $`D`$, let $`d\\mu_\\vartheta=w_{\\sigma,\\vartheta}(x,y)\\,dx\\,dy`$, where $`w_{\\sigma,\\vartheta}(x,y)=(y-x)^2\\{1-\\vartheta(x+y-2)^2\\}/16`$. Set $`D_-=D\\cap\\{y<1\\}`$, $`D_+=D\\cap\\{x>1\\}`$, and denote the same-side masses by $`M_{\\vartheta,\\pm}=\\mu_\\vartheta(D_\\pm)`$.\n\n\n\n**Proposition 5.3 (Background share along the dependence family).** For $`|\\vartheta|<1/2`$, the boundary flow is attained and\n\n\n\n$$\nJ_\\vartheta=\\frac{75-14\\vartheta}{480},\\qquad\nM_{\\vartheta,-}=M_{\\vartheta,+}=\\frac{3-2\\vartheta}{1152},\n\\qquad\na_{+,\\vartheta}\n=1-\\frac{4(M_{\\vartheta,-}+M_{\\vartheta,+})}{J_\\vartheta}.\n\\tag{5.3}\n$$\n\nIn particular,\n\n$$\nJ_\\vartheta a_{+,\\vartheta}=\\frac{195-22\\vartheta}{1440},\\qquad\n\\frac{d a_{+,\\vartheta}}{d\\vartheta}\n=\\frac{360}{(75-14\\vartheta)^2}>0,\\qquad\n\\frac{d(J_\\vartheta a_{+,\\vartheta})}{d\\vartheta}\n=-\\frac{11}{720}.\n$$\n\nTo obtain the explicit value, we use all the mass in $`D_-`$ and $`D_+`$ and match their projections in two disjoint parts of the cross-region. The section inequalities in [Appendix B.3](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L49) give the required capacity throughout $`|\\vartheta|<1/2`$, and the midpoint potential gives the same value in the dual problem, so this feasible flow is optimal.\n\nAs reference dependence varies through $`0<\\vartheta<1/2`$, the minimum background share and the corresponding unnormalized displacement moment move in opposite directions:\n\n\n\n$$\na_{+,\\vartheta}-a_{+,0}\n=\\frac{24\\vartheta}{5(75-14\\vartheta)}>0,\\qquad\nJ_\\vartheta a_{+,\\vartheta}-J_0a_{+,0}\n=-\\frac{11\\vartheta}{720}<0.\n\\tag{5.4}\n$$\n\nThe contrast comes from the median plane $`x+y+z=3`$, on which the centered factors satisfy $`z-1=-(x+y-2)`$. Increasing $`\\vartheta`$ thus favors equal signs in the full box but reduces the section density away from $`x+y=2`$ by the factor $`1-\\vartheta(x+y-2)^2`$. Across both exchanged half-sections, the traded-coordinate matching capacity is $`4(M_{\\vartheta,-}+M_{\\vartheta,+})`$. By [(5.3)](#dep-eq-5-3), $`J_\\vartheta`$ decreases faster in absolute units than this capacity, so their difference $`J_\\vartheta a_{+,\\vartheta}`$ declines. In proportion, though, the matching capacity falls faster than $`J_\\vartheta`$ throughout $`0<\\vartheta<1/2`$, which leaves a larger relative share for the background coordinate.\n\nFigure [4](#dep-fig-flow-shares) compares two examples: independence ($`\\vartheta=0`$, model U) and $`\\vartheta=1/4`$ (model D). The triples $`(J_\\vartheta,a_{+,\\vartheta},J_\\vartheta a_{+,\\vartheta})`$ are $`(5/32,13/15,13/96)`$ and $`(143/960,379/429,379/2880)`$, so the minimum background share rises from about $`0.8667`$ to $`0.8834`$ while its unnormalized displacement moment falls from about $`0.1354`$ to $`0.1316`$.\n\n\n\n; all use the same scale from zero to one. On each same-side section we allocate displacement to a rebalanced coordinate and match its projection using the cross-region. We allocate the remainder to the background coordinate. Blank regions lie outside $`D`$.](/assets/research/dependence/optimal-flow-shares.png)\n\n\n\n### 5.3 Implications for reverse stress testing and further questions\n\nAn exact initial sensitivity need not predict how much dependence stress is required to reverse a diversification trade. This bears on reverse sensitivity analysis [[28](#ref-pesenti2019), [29](#ref-pesenti2022)], where the quantity of interest is the distance to a decision reversal. Under the traded-coordinate dominance condition [(2.9)](#dep-eq-2-9), $`r_{0,\\varepsilon}=m_\\varepsilon/K_\\varepsilon`$ extrapolates an attained initial slope, yet if $`\\gamma_1=c_{\\min}1.\n$$\n\nProposition [3.2](#dep-prop-balanced-linear-profile) attains the initial slope by balancing opposite movements on selected fibers. Equality in the efficiency bound at reversal would require balancing cheapest-coordinate projections across the entire tail section, and under the stated cost condition the corner obstruction of Theorem [4.2](#dep-thm-zero-premium) excludes that. Unlike optimal allocation across full Fréchet or Wasserstein ambiguity sets [[23](#ref-liu2026)], our reverse stress test fixes a proposed trade and measures the least perturbation of a common reference law that reverses its paired ES ranking.\n\nA whole-book comparison across reference dependence structures also has to account for the size of the nominal advantage. In the family of Section 5.2, consider the boundary perturbation $`\\boldsymbol\\gamma(\\eta)=(\\gamma_0,\\gamma_0,\\gamma_0(1+\\eta))`$. The gain-normalized critical budget and its right derivative at equal effective costs satisfy\n\n$$\n\\mathcal R_\\vartheta(\\eta):=\\lim_{\\varepsilon\\downarrow0}\n\\frac{\\rho^\\pm_{\\varepsilon,\\vartheta}(\\eta)}{m_{\\varepsilon,\\vartheta}}\n=\\frac{p\\,s^*_\\vartheta(\\eta)}{J_\\vartheta},\\qquad\n\\mathcal R_\\vartheta'(0+)=p\\gamma_0a_{+,\\vartheta}.\n$$\n\nTheorem [4.3](#dep-thm-projected-flow) gives the absolute right derivative $`(s^*_\\vartheta)'(0+)=\\gamma_0J_\\vartheta a_{+,\\vartheta}`$. By [(5.4)](#dep-eq-5-4), increasing $`\\vartheta`$ raises the gain-normalized response $`\\mathcal R_\\vartheta'(0+)`$ but lowers the absolute response: $`\\frac{d}{d\\vartheta}(\\gamma_0J_\\vartheta a_{+,\\vartheta})=\\gamma_0(J_\\vartheta'a_{+,\\vartheta}+J_\\vartheta a_{+,\\vartheta}')<0`$, so the decline in the nominal-gain scale outweighs the increase in background share. Reporting both boundary sensitivities to a risk committee separates a greater relative need for background displacement from a smaller absolute response in budget units.\n\nCharacterizing the boundaries between the active and stationary regions of an optimal kernel could help determine the full profile $`F`$ between the analytical envelopes, beyond the guaranteed initial linear interval. A uniform rate for $`\\sup_{s\\le S_0}|\\varepsilon^{-2}D_\\varepsilon(s\\varepsilon^2)-F(s)|`$, combined with the lower growth estimate in Theorem [3.3](#dep-thm-strict-crossing), would yield finite-trade error bounds for the critical budgets.\n\nIn the compact-box benchmark, a restoration premium can arise from dependence geometry and full marginal conservation even with bounded marginal tails. For unbounded marginals such as integrable Pareto or lognormal laws, the inverse-CDF derivative is unbounded near probability one, so the uniform displacement estimate in the recovery argument would need a tail-weighted substitute. Since one-dimensional $`W_1`$ remains the $`L^1`$ integral of the CDF discrepancy, this loss of a uniform quantile bound does not by itself imply a different restoration-cost exponent. The open problem is to control that integral jointly with the paired-risk error and determine whether the $`\\varepsilon^2`$ critical scale persists under suitable tail assumptions. Higher-dimensional projected-flow representations and estimation of critical budgets from loss data raise related questions, alongside recent work on ES estimation under weak tail assumptions [[3](#ref-bartleckstein2026)].\n\n## Appendices and supporting material\n\nThis page includes the complete abstract, five main sections, all four main-text figures, and references. Technical appendices and declarations are available in the [full manuscript](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/paper.pdf) and the [GitHub repository](https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals).\n\n## References\n\n\n\n1. P. Artzner, F. Delbaen, J.-M. Eber, and D. Heath, *Coherent measures of risk*, Math. Finance, 9(3) (1999), pp. 203–228, [doi:10.1111/1467-9965.00068](https://doi.org/10.1111/1467-9965.00068).\n\n\n\n2. D. Bartl, S. Drapeau, J. Obłój, and J. Wiesel, *Sensitivity analysis of Wasserstein distributionally robust optimization problems*, Proc. R. Soc. A, 477(2256) (2021), article 20210176, [doi:10.1098/rspa.2021.0176](https://doi.org/10.1098/rspa.2021.0176).\n\n\n\n3. D. Bartl and S. Eckstein, *Optimal nonparametric estimation of the expected shortfall risk*, preprint (2026), to appear in SIAM J. Financial Math., [arXiv:2405.00357v2](https://arxiv.org/abs/2405.00357v2).\n\n\n\n4. D. Bartl and J. Wiesel, *Sensitivity of multiperiod optimization problems with respect to the adapted Wasserstein distance*, SIAM J. Financial Math., 14(2) (2023), pp. 704–720, [doi:10.1137/22M1537746](https://doi.org/10.1137/22M1537746).\n\n\n\n5. C. Bernard, X. Jiang, and R. Wang, *Risk aggregation with dependence uncertainty*, Insur. Math. Econ., 54 (2014), pp. 93–108, [doi:10.1016/j.insmatheco.2013.11.005](https://doi.org/10.1016/j.insmatheco.2013.11.005).\n\n\n\n6. D. P. Bertsekas, *Convex Optimization Theory*, Athena Scientific, Belmont, MA, 2009, [author's online edition](https://web.mit.edu/dimitrib/www/Convex_Theory_Entire_Book.pdf).\n\n\n\n7. J. Blanchet and K. Murthy, *Quantifying distributional model risk via optimal transport*, Math. Oper. Res., 44(2) (2019), pp. 565–600, [doi:10.1287/moor.2018.0936](https://doi.org/10.1287/moor.2018.0936).\n\n\n\n8. M. Brandtner, *Expected shortfall, spectral risk measures, and the aggravating effect of background risk, or: risk vulnerability and the problem of subadditivity*, J. Bank. Financ., 89 (2018), pp. 138–149, [doi:10.1016/j.jbankfin.2018.02.002](https://doi.org/10.1016/j.jbankfin.2018.02.002).\n\n\n\n9. J. Chen, E. Furman, and X. S. Lin, *Static marginal expected shortfall: Systemic risk measurement under dependence uncertainty*, Insur. Math. Econ., 130 (2026), article 103285, [doi:10.1016/j.insmatheco.2026.103285](https://doi.org/10.1016/j.insmatheco.2026.103285).\n\n\n\n10. Y. Chen, P. Liu, Y. Liu, and R. Wang, *Ordering and inequalities for mixtures on risk aggregation*, Math. Finance, 32(1) (2022), pp. 421–451, [doi:10.1111/mafi.12323](https://doi.org/10.1111/mafi.12323).\n\n\n\n11. G. Dal Maso, *An Introduction to $`\\Gamma`$-Convergence*, Progress in Nonlinear Differential Equations and Their Applications 8, Birkhäuser, Boston, 1993, [doi:10.1007/978-1-4612-0327-8](https://doi.org/10.1007/978-1-4612-0327-8).\n\n\n\n12. C. De Vecchi, M. Nendel, and J. Streicher, *Upper comonotonicity and risk aggregation under dependence uncertainty*, Math. Finance, 36(1) (2026), pp. 118–139, [doi:10.1111/mafi.70008](https://doi.org/10.1111/mafi.70008).\n\n\n\n13. S. Eckstein, M. Kupper, and M. Pohl, *Robust risk aggregation with neural networks*, Math. Finance, 30(4) (2020), pp. 1229–1272, [doi:10.1111/mafi.12280](https://doi.org/10.1111/mafi.12280).\n\n\n\n14. S. Eckstein, G. Guo, T. Lim, and J. Obłój, *Robust pricing and hedging of options on multiple assets and its numerics*, SIAM J. Financial Math., 12(1) (2021), pp. 158–188, [doi:10.1137/19M1286256](https://doi.org/10.1137/19M1286256).\n\n\n\n15. P. Embrechts, G. Puccetti, and L. Rüschendorf, *Model uncertainty and VaR aggregation*, J. Bank. Financ., 37(8) (2013), pp. 2750–2764, [doi:10.1016/j.jbankfin.2013.03.014](https://doi.org/10.1016/j.jbankfin.2013.03.014).\n\n\n\n16. P. Embrechts, B. Wang, and R. Wang, *Aggregation-robustness and model uncertainty of regulatory risk measures*, Finance Stoch., 19(4) (2015), pp. 763–790, [doi:10.1007/s00780-015-0273-z](https://doi.org/10.1007/s00780-015-0273-z).\n\n\n\n17. L.-B. Fiechtner and J. Blanchet, *Wasserstein distributionally robust regret optimization*, preprint, 2025, revised 2026, [arXiv:2504.10796v5](https://arxiv.org/abs/2504.10796v5).\n\n\n\n18. R. Gao and A. J. Kleywegt, *Data-driven robust optimization with known marginal distributions*, working paper, 2017, [https://faculty.mccombs.utexas.edu/rui.gao/copula.pdf](https://faculty.mccombs.utexas.edu/rui.gao/copula.pdf).\n\n\n\n19. R. Gao and A. J. Kleywegt, *Distributionally robust stochastic optimization with Wasserstein distance*, Math. Oper. Res., 48(2) (2023), pp. 603–655, [doi:10.1287/moor.2022.1275](https://doi.org/10.1287/moor.2022.1275).\n\n\n\n20. J. Korman, R. J. McCann, and C. Seis, *Dual potentials for capacity constrained optimal transport*, Calc. Var. Partial Differential Equations, 54 (2015), pp. 573–584, [doi:10.1007/s00526-014-0795-9](https://doi.org/10.1007/s00526-014-0795-9).\n\n\n\n21. H. Lam, *Sensitivity to serial dependency of input processes: A robust approach*, Manage. Sci., 64(3) (2018), pp. 1311–1327, [doi:10.1287/mnsc.2016.2667](https://doi.org/10.1287/mnsc.2016.2667).\n\n\n\n22. L. Li, H. Shao, R. Wang, and J. Yang, *Worst-case Range Value-at-Risk with partial information*, SIAM J. Financial Math., 9(1) (2018), pp. 190–218, [doi:10.1137/17M1126138](https://doi.org/10.1137/17M1126138).\n\n\n\n23. P. Liu and Y. Liu, *Portfolio diversification and concentration under dependence uncertainty: A majorization approach*, preprint, 2026, [arXiv:2609.04496v1](https://arxiv.org/abs/2609.04496v1).\n\n\n\n24. R. J. Martin, *Saddlepoint methods in portfolio theory*, in The Oxford Handbook of Credit Derivatives, A. Lipton and A. Rennie, eds., Oxford University Press, Oxford (2011), pp. 533–569, [doi:10.1093/oxfordhb/9780199546787.013.0015](https://doi.org/10.1093/oxfordhb/9780199546787.013.0015).\n\n\n\n25. A. J. McNeil, R. Frey, and P. Embrechts, *Quantitative Risk Management: Concepts, Techniques and Tools*, revised ed., Princeton University Press, Princeton, NJ, 2015.\n\n\n\n26. H. Namkoong, *Local sensitivity under transport restrictions*, preprint, 2026, [arXiv:2606.04276v1](https://arxiv.org/abs/2606.04276v1).\n\n\n\n27. S. Ota and M. Kimura, *Effective estimation algorithm for parameters of multivariate Farlie–Gumbel–Morgenstern copula*, Jpn. J. Stat. Data Sci., 4 (2021), pp. 1049–1078, [doi:10.1007/s42081-021-00118-y](https://doi.org/10.1007/s42081-021-00118-y).\n\n\n\n28. S. M. Pesenti, P. Millossovich, and A. Tsanakas, *Reverse sensitivity testing: What does it take to break the model?*, European J. Oper. Res., 274(2) (2019), pp. 654–670, [doi:10.1016/j.ejor.2018.10.003](https://doi.org/10.1016/j.ejor.2018.10.003).\n\n\n\n29. S. M. Pesenti, *Reverse sensitivity analysis for risk modelling*, Risks, 10(7) (2022), article 141, [doi:10.3390/risks10070141](https://doi.org/10.3390/risks10070141).\n\n\n\n30. S. M. Pesenti and S. Jaimungal, *Portfolio optimization within a Wasserstein ball*, SIAM J. Financial Math., 14(4) (2023), pp. 1175–1214, [doi:10.1137/22M1496803](https://doi.org/10.1137/22M1496803).\n\n\n\n31. G. Ch. Pflug and M. Pohl, *A review on ambiguity in stochastic portfolio optimization*, Set-Valued Var. Anal., 26(4) (2018), pp. 733–757, [doi:10.1007/s11228-017-0458-z](https://doi.org/10.1007/s11228-017-0458-z).\n\n\n\n32. R. T. Rockafellar and S. Uryasev, *Conditional value-at-risk for general loss distributions*, J. Bank. Financ., 26(7) (2002), pp. 1443–1471, [doi:10.1016/S0378-4266(02)00271-6](https://doi.org/10.1016/S0378-4266%2802%2900271-6).\n\n\n\n33. L. Rüschendorf, *Mathematical Risk Analysis: Dependence, Risk Bounds, Optimal Allocations and Portfolios*, Springer Series in Operations Research and Financial Engineering, Springer, Berlin, Heidelberg, 2013, [doi:10.1007/978-3-642-33590-7](https://doi.org/10.1007/978-3-642-33590-7).\n\n\n\n34. F. Santambrogio, *Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling*, Progress in Nonlinear Differential Equations and Their Applications 87, Birkhäuser, Cham, 2015, [doi:10.1007/978-3-319-20828-2](https://doi.org/10.1007/978-3-319-20828-2).\n\n\n\n35. N. Sauldubois, *Higher-order sensitivity analysis of Wasserstein and adapted Wasserstein distributionally robust optimization problems*, preprint, 2026, [arXiv:2609.25345v1](https://arxiv.org/abs/2609.25345v1).\n\n\n\n36. O. Scaillet, *Nonparametric estimation and sensitivity analysis of expected shortfall*, Math. Finance, 14(1) (2004), pp. 115–129, [doi:10.1111/j.0960-1627.2004.00184.x](https://doi.org/10.1111/j.0960-1627.2004.00184.x).\n\n\n\n37. M. Sion, *On general minimax theorems*, Pacific J. Math., 8(1) (1958), pp. 171–176, [doi:10.2140/pjm.1958.8.171](https://doi.org/10.2140/pjm.1958.8.171).\n\n\n\n38. D. Tasche, *Conditional expectation as quantile derivative*, preprint, 2001, [arXiv:math/0104190v1](https://arxiv.org/abs/math/0104190v1).\n\n\n\n39. S. S. Vallender, *Calculation of the Wasserstein distance between probability distributions on the line*, Theory Probab. Appl., 18(4) (1974), pp. 784–786, [doi:10.1137/1118101](https://doi.org/10.1137/1118101).\n\n\n\n40. C. Villani, *Topics in Optimal Transportation*, Graduate Studies in Mathematics 58, American Mathematical Society, Providence, RI, 2003.\n\n\n\n41. R. Wang, V. Bignozzi, and A. Tsanakas, *How superadditive can a risk measure be?*, SIAM J. Financial Math., 6(1) (2015), pp. 776–803, [doi:10.1137/140981046](https://doi.org/10.1137/140981046).\n"
}
diff --git a/website/content/sources/ai-portfolio/ai-research-and-engineering.md b/website/content/sources/ai-portfolio/ai-research-and-engineering.md
index ce6befc..004fb53 100644
--- a/website/content/sources/ai-portfolio/ai-research-and-engineering.md
+++ b/website/content/sources/ai-portfolio/ai-research-and-engineering.md
@@ -229,10 +229,10 @@ $$
The hard capacity $`H_j^{\text{review}}`$ goes to remediation actions, and reference audits are scheduled independently on frozen original submissions.
-**The joint rule.** Attach a price $`\lambda`$ to the budget, $`\nu`$ to the precision ceiling (in units of $`N^2 V`$) and $`\kappa_j`$ to each specialist's hours. At those prices the program separates into one small problem per task, and each has a closed form:
+**The joint rule.** Attach a price $`\lambda`$ to the budget, $`\nu`$ to the precision ceiling (in units of $`N^2 V`$) and $`\gamma_j`$ to each specialist's hours. At those prices the program separates into one small problem per task, and each has a closed form:
$$
-\pi_i^* = \mathrm{clip}_{[\pi_{\min}, 1]} \left( \sqrt{\frac{\nu\, m_i}{\big( \lambda c_i - r_{i a_i} \big)_+}} \right), \qquad a_i^* = \arg\min_a \; (1 - \pi_i)\, r_{ia} + \lambda d_{ia} + \sum_j \kappa_j t_{iaj} \tag{5.6}
+\pi_i^* = \mathrm{clip}_{[\pi_{\min}, 1]} \left( \sqrt{\frac{\nu\, m_i}{\big( \lambda c_i - r_{i a_i} \big)_+}} \right), \qquad a_i^* = \arg\min_a \; (1 - \pi_i)\, r_{ia} + \lambda d_{ia} + \sum_j \gamma_j t_{iaj} \tag{5.6}
$$
with $`\pi_i^* = 1`$ when $`r_{i a_i} \ge \lambda c_i`$. Three readings of the audit rule carry the argument of this article.
@@ -243,13 +243,13 @@ with $`\pi_i^* = 1`$ when $`r_{i a_i} \ge \lambda c_i`$. Three readings of the a
Initial risk, residual uncertainty per unit cost and removable loss now sit in one pair of formulas, and design-unbiasedness is untouched: the probabilities depend only on quantities fixed before the labels are revealed, and $`\hat{\mu}_{\text{raw}}`$ is still computed on frozen originals.
-The prices also make the program cheap to solve. For fixed $`(\lambda, \nu, \kappa)`$ every task's audit probability and action follow from the two formulas above in one pass, so a weekly batch of ten thousand tasks is a search over $`J + 2`$ multipliers with a linear-time step. The linear relaxation of the assignment splits at most one task per binding budget or capacity constraint, so rounding to whole assignments moves at most $`J + 1`$ tasks, and the gap between the priced bound and the rounded plan certifies the solution. When the program is infeasible, the price that diverges names what is short: the precision target, the specialist hours or the budget.
+The prices also make the program cheap to solve. For fixed $`(\lambda, \nu, \gamma)`$ every task's audit probability and action follow from the two formulas above in one pass, so a weekly batch of ten thousand tasks is a search over $`\lambda`$, $`\nu`$ and one price per specialist, with a linear-time step. The linear relaxation of the assignment splits at most one task per binding budget or capacity constraint, so rounding to whole assignments moves at most that many tasks, and the gap between the priced bound and the rounded plan certifies the solution. When the program is infeasible, the price that diverges names what is short: the precision target, the specialist hours or the budget.
The ratio $`\nu / \lambda`$ is the exchange rate between the two products: the number of delivered defects the program gives up to tighten the estimate by one unit of variance. Section 6 reads it off the reference batch.
## 6 Results on a reference batch
-**Reference batch.** Production records and operating figures are withheld under the nondisclosure agreement. The results below are stated on a reference batch at the program's scale: one week of work, $`N = 10,000`$ submissions in three strata by difficulty, most of them routine and a tenth of them hard, with defect rates and audit costs that rise with difficulty. Every result is given twice: at the reference batch, and as its range over an envelope in which each entry of the batch moves independently by up to 20% in either direction. A conclusion stated in this section holds across the whole envelope; none of them rests on a single operating point.
+**Reference batch.** Production records and operating figures are withheld under the nondisclosure agreement. The results below are stated on a reference batch at the program's scale: one week of work, $`N = 10{,}000`$ submissions in three strata by difficulty, most of them routine and a tenth of them hard, with defect rates and audit costs that rise with difficulty. Each headline result is given twice: at the reference batch, and as its range over an envelope in which each entry of the batch moves independently by up to 20% in either direction.
**Table 1. The reference batch in three strata by difficulty.**
@@ -261,7 +261,7 @@ The ratio $`\nu / \lambda`$ is the exchange rate between the two products: the n
| Cost of an independent reference audit ($`c`$, routine = 1) | 1 | 1.5 | 3 |
| Share of defects removed by one additional review | 80% | 80% | 30% |
-The batch defect rate is $`\mu = 7.10\%`$. One review costs half of a reference audit in the same stratum. The audit budget is $`2,000`$ cost units, 14.8% of the cost of auditing everything, with a sampling floor of $`\pi_{\min} = 3\%`$.
+The batch defect rate is $`\mu = 7.10\%`$. One review costs half of a reference audit in the same stratum. The audit budget is 2,000 cost units, 14.8% of the cost of auditing everything, with a sampling floor of $`\pi_{\min} = 3\%`$.
**Measurement at equal audit budget.** Every policy below uses the estimator of Section 2 and spends the same expected audit budget.
@@ -299,7 +299,7 @@ Three results hold everywhere in the envelope. The uncorrected error rate in the
The stress case is a residual ordering reversed across the strata, which sends the allocation to 23.00% / 11.77% / 3.00% and raises the variance to 199.7% above uniform. With a weight of $`\rho = 0.7`$ on the uniform design the same wrong estimates cost 17.1%, and the bound of Section 4 guarantees no more than 50.3% for any estimates.
-**One budget, two products.** The total budget is $`5,000`$ cost units, 37% of the cost of auditing everything, to be divided between audits and reviews. With no handling the batch ships 710 defects.
+**One budget, two products.** The total budget is 5,000 cost units, 37% of the cost of auditing everything, to be divided between audits and reviews. With no handling the batch ships 710 defects.
**Table 5. One budget of 5,000 cost units divided between audits and reviews.**
@@ -313,7 +313,7 @@ The stress case is a residual ordering reversed across the strata, which sends t
At the same budget and the same precision, the joint rule ships **38.6%** fewer defects than risk-first routing, and 19% to 67% fewer across the envelope. Held to the tighter precision of uniform sampling it still ships 34.6% fewer.
-Table 5 also shows why the two jobs cannot be planned one after the other. Choosing audits for measurement and then spending the rest on repair gives the best estimate in the table and ships more defects than the legacy policy, because it ignores the repair that audits perform. The joint rule sends every item of the hardest stratum to an independent reference solution: there a review removes 30% of the defects and a reference audit removes all of them, at a risk per unit cost of $`0.35 / 3 = 0.117`$ against a budget price $`\lambda`$ of 0.093 to 0.096. In the two easier strata the audit probabilities follow the closed form of Section 5: the $`\nu`$ implied by the routine stratum and the $`\nu`$ implied by the intermediate stratum agree to six digits in both joint plans.
+Table 5 also shows why the two jobs cannot be planned one after the other. Choosing audits for measurement and then spending the rest on repair gives the best estimate in the table and ships more defects than the legacy policy, because it ignores the repair that audits perform. Held to the sequential plan's own precision of 0.515 points, the joint rule ships 34.7% fewer defects than that plan, and 12.4% to 69.3% fewer across the whole envelope. The joint rule sends every item of the hardest stratum to an independent reference solution: there a review removes 30% of the defects and a reference audit removes all of them, at a risk per unit cost of $`0.35 / 3 = 0.117`$ against a budget price $`\lambda`$ of 0.093 to 0.096. In the two easier strata the audit probabilities follow the closed form of Section 5: the $`\nu`$ implied by the routine stratum and the $`\nu`$ implied by the intermediate stratum agree to six digits in both joint plans.
The last three rows are the exchange rate. Holding the estimate at 0.657 points costs 22 shipped defects relative to no ceiling, and tightening it to 0.568 costs 13 more. A program that knows this number can decide how much it wants to know.
@@ -408,7 +408,7 @@ The team averaged roughly 500 completed financial tasks a week. Each task combin
## 1 Distributional fit and task validity
-Modern tabular generators address different distributional problems. [CTGAN](https://arxiv.org/abs/1907.00503) handles mixed discrete and continuous data, multimodality, and class imbalance through mode-specific normalization and conditional vectors, and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) runs Gaussian diffusion on numerical fields and multinomial diffusion on categorical ones ([CTGAN](https://arxiv.org/abs/1907.00503); [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html)). Both learn empirical distributions effectively. A financial task can match its distribution and still be invalid, because part of what makes it valid is structural. Accounting identities have to hold exactly, fiscal periods have to run in chronological order, covenant applicability conditions have to be respected, and the numerical inputs have to agree logically with the reference judgment. Global distributional similarity metrics enforce none of this.
+Modern tabular generators address different distributional problems. [CTGAN](https://arxiv.org/abs/1907.00503) handles mixed discrete and continuous data, multimodality and class imbalance through mode-specific normalization and conditional vectors, and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) runs Gaussian diffusion on numerical fields and multinomial diffusion on categorical ones. Both learn empirical distributions effectively. A financial task can match its distribution and still be invalid, because part of what makes it valid is structural. Accounting identities have to hold exactly, fiscal periods have to run in chronological order, covenant applicability conditions have to be respected, and the numerical inputs have to agree logically with the reference judgment. Global distributional similarity metrics enforce none of this.
A simple construction shows how far apart the two can drift. Suppose valid financial inputs $`x \in \mathbb{R}^d`$ have to satisfy $`m`$ linear accounting or ratio identities $`Ax = b`$, where $`A \in \mathbb{R}^{m \times d}`$ has full row rank, so that the valid population distribution $`P`$ is supported entirely on the affine subspace $`M = \{x \in \mathbb{R}^d : Ax = b\}`$. Assume this distribution has finite second moments, and let the columns of $`N \in \mathbb{R}^{d \times m}`$ form an orthonormal basis for the normal space orthogonal to $`M`$. Now take a generator that captures $`P`$ along the manifold but adds a small amount of independent noise in the normal directions: $`X_\sigma = X + \sigma N \varepsilon`$, where $`\varepsilon \sim \mathcal{N}(0, I_m)`$ is independent of $`X \sim P`$. Writing $`Q_\sigma`$ for the generated distribution, I obtain, for every $`\sigma > 0`$,
@@ -429,9 +429,9 @@ The first line comes from coupling $`(X, X_\sigma)`$ on the same probability spa
The construction is one instance of a general gap. Synthetic data is evaluated on three axes, fidelity to the source distribution, utility for a downstream model and privacy, and the sample-level refinements of those metrics still score the inputs ([Alaa et al.](https://proceedings.mlr.press/v162/alaa22a.html)). None of the three has a place for the question a training team has to answer about supervision: is the label attached to this generated case still the right label? Data augmentation has the same blind spot from the other side. Counterfactual augmentation and contrast sets showed that a minimal edit can flip the gold answer, and they restore validity by having a person relabel every edited example ([Kaushik et al.](https://arxiv.org/abs/1909.12434); [Gardner et al.](https://aclanthology.org/2020.findings-emnlp.117/)), which does not scale to generated variants. I add the missing axis, supervision validity, and the remaining sections build the machinery that keeps it high at scale.
The counterexample also sets how I divided responsibilities in the synthetic data architecture ([SDV constraint-augmented generation](https://docs.sdv.dev/sdv/modeling/constraint-augmented-generation-cag)):
-• Copula models and conditional sampling capture cross-sectional rank and tail dependencies among continuous corporate indicators, including company size, operating history, leverage, and profit margins. They preserve the dependencies that independent column generation would discard.
+• Copula models and conditional sampling capture cross-sectional rank and tail dependencies among continuous corporate indicators, including company size, operating history, leverage and profit margins. They preserve the dependencies that independent column generation would discard.
• [CTGAN](https://arxiv.org/abs/1907.00503) and [TabDDPM](https://proceedings.mlr.press/v202/kotelnikov23a.html) model mixed-type and multimodal empirical distributions, and explicit business-rule checks follow them.
-• Deterministic variable transformations and manifold projections enforce accounting equalities, ratio bounds, and chronological order.
+• Deterministic variable transformations and manifold projections enforce accounting equalities, ratio bounds and chronological order.
## 2 Preserving the conditional distribution under business constraints
@@ -469,7 +469,7 @@ $$
Affine transformations preserve Gaussianity, so these moments identify the exact regular conditional Gaussian distribution of $`Z \mid AZ = b`$ ([Majumdar and Majumdar](https://arxiv.org/abs/1710.09285)). The update has a long lineage. It is the sampler for a normal distribution truncated on hyperplanes ([Cong, Chen and Zhou](https://doi.org/10.1214/17-BA1052)) and Matheron's rule for pathwise conditioning on a noise-free linear observation ([Wilson et al.](https://arxiv.org/abs/2011.04026)). What I did with it is put accounting identities in the place of the hyperplanes and state where exactness holds, in the latent coordinates of the copula. That also marks the division of labor with constraint layers for deep tabular generators, which enforce linear inequalities in any generator and make no statement about the law of the corrected sample ([Stoian et al.](https://arxiv.org/abs/2402.04823)). Inequalities such as ratio ranges go to transformations and layers; identities go to the projection, which returns the right distribution as well as a feasible point.
-On a reference run of the constrained generator, **200,000 twelve-dimensional balance-sheet vectors subject to 4 simultaneous linear identities** (assets sum to total assets, liabilities sum to total liabilities, assets equal liabilities plus equity, and receivables stand at 60 days of revenue), the maximum absolute constraint residual was **$`3.0 \times 10^{-15}`$**, and the sample covariance matched the theoretical conditional covariance to **0.359%** in relative Frobenius norm. Every draw is used; the 395,442 candidate draws per valid input in the table above are what rejection would have spent. Fields that accounting definitions fully determine are computed analytically. Range bounds and date ordering are enforced in the latent generation space when they admit bijective transformations, such as logarithmic transforms or positive increments $`t_{k} = t_{k-1} + \exp(\xi_k)`$.
+On a reference run of the constrained generator, **200,000 twelve-dimensional balance-sheet and income vectors subject to 4 simultaneous linear identities** (assets sum to total assets, liabilities sum to total liabilities, assets equal liabilities plus equity, and receivables stand at 60 days of revenue), the maximum absolute constraint residual was **$`3.0 \times 10^{-15}`$**, and the sample covariance matched the theoretical conditional covariance to **0.359%** in relative Frobenius norm. Every draw is used; the expected 395,442.59 candidate draws per valid input in Table 1 are what rejection would have spent. Fields that accounting definitions fully determine are computed analytically. Range bounds and date ordering are enforced in the latent generation space when they admit bijective transformations, such as logarithmic transforms or positive increments $`t_{k} = t_{k-1} + \exp(\xi_k)`$.
**The correction is also a monitor.** The squared Mahalanobis distance by which the projection moves a draw has a known law:
@@ -481,7 +481,7 @@ When the upstream generator is consistent with the identities, $`\delta = 0`$ an
## 3 Using decision margins to determine when reference judgments need recomputation
-A synthetic variant of a financial reference case often keeps the business narrative and changes decisive financial figures, covenant thresholds, or macroeconomic inputs. If the original reference answer stays in place after those values change, systematic label errors enter post-training, even when all accounting constraints hold. [FinQA](https://aclanthology.org/2021.emnlp-main.300/) pairs financial questions with executable reasoning programs, and synthetic preference data similarly require a maintained link between inputs, reference conclusions, and the expert's explanation of a preference ([FinQA](https://aclanthology.org/2021.emnlp-main.300/)).
+A synthetic variant of a financial reference case often keeps the business narrative and changes decisive financial figures, covenant thresholds or macroeconomic inputs. If the original reference answer stays in place after those values change, systematic label errors enter post-training, even when all accounting constraints hold. [FinQA](https://aclanthology.org/2021.emnlp-main.300/) pairs financial questions with executable reasoning programs, and synthetic preference data similarly require a maintained link between inputs, reference conclusions and the expert's explanation of a preference.
Neighboring work links data generation to feedback in other ways. [WizardMath](https://arxiv.org/pdf/2308.09583v1) couples the evolution of mathematical instructions with reinforcement feedback, and [FinGPT](https://arxiv.org/pdf/2307.10485v1) fine-tunes financial language models on feedback from the market. Templated reasoning sets recompute an answer from the template's variables for every instance ([GSM-Symbolic](https://arxiv.org/abs/2410.05229)), which works when the answer is one formula and leaves evidence and preference untouched. I make the dependencies among a financial scenario, its reference conclusion, the evidence behind that conclusion and the expert preference built on it the object of design, so that a change to the scenario reaches each of them, and only the ones it has to.
@@ -516,7 +516,7 @@ The thresholds are 1.645 at $`\alpha = 5\%`$ and 2.326 at $`\alpha = 1\%`$, and
I therefore tie case expansion to reference-answer recomputation and renewed five-level pairwise preference grading when perturbations approach or cross a decision boundary. In that event the workflow recomputes the reference solution and samples new candidate model reasoning trajectories. Domain experts then assess the pair on the five-level preference scale and identify which trajectory recognized the decisive variable whose change reversed the decision and supported its judgment with primary evidence. [HelpSteer2-Preference](https://arxiv.org/abs/2410.01257) gives a comparison point from the methods literature.
-**What a graded preference is worth.** A five-level judgment takes an expert a few seconds longer than a binary one, on a task that has already cost about three specialist-hours. Model the judgment as an ordered response to the latent quality gap $`g`$ between two responses, measured in units of judgment noise, with cutpoints $`c_1 < \dots < c_4`$. The Fisher information one response carries about the gap is
+**What a graded preference is worth.** A five-level judgment takes an expert a few seconds longer than a binary one, on a task that has already cost about three specialist-hours. Model the judgment as an ordered response to the latent quality gap $`g`$ between two responses, measured in units of judgment noise, with cutpoints $`c_1 < \dots < c_4`$, $`c_0 = -\infty`$ and $`c_5 = +\infty`$. The Fisher information one response carries about the gap is
$$
I(g) = \sum_{k} \frac{\big( \partial_g\, p_k(g) \big)^2}{p_k(g)}, \qquad p_k(g) = \Phi(c_k - g) - \Phi(c_{k-1} - g) \tag{3.4}
@@ -534,7 +534,7 @@ One graded judgment is worth 1.4 to 2 binary judgments, and the advantage grows
## 4 Specifying the information available to each temporal task
-A cross-sectional company comparison does not require continuous-time stochastic processes. Dynamic tasks such as derivatives valuation, interest-rate term-structure interpolation, and liquidity stress tests over multiple periods do require path generators. For those I use SDEs and neural SDEs to model path distributions beyond single-time marginals ([Neural SDEs as Infinite-Dimensional GANs](https://proceedings.mlr.press/v139/kidger21b.html)). Brownian bridges and stochastic interpolation support conditioning on known intermediate observations or endpoints, following the constructions in [Glasserman, *Monte Carlo Methods in Financial Engineering*, Chapter 3](https://doi.org/10.1007/978-0-387-21617-1). A bridge is the conditioning of Section 2 in one dimension: draw the free path, then correct it onto the known endpoint.
+A cross-sectional company comparison does not require continuous-time stochastic processes. Dynamic tasks such as derivatives valuation, interest-rate term-structure interpolation and liquidity stress tests over multiple periods do require path generators. For those I use SDEs and neural SDEs to model path distributions beyond single-time marginals ([Neural SDEs as Infinite-Dimensional GANs](https://proceedings.mlr.press/v139/kidger21b.html)). Brownian bridges and stochastic interpolation support conditioning on known intermediate observations or endpoints, following the constructions in [Glasserman, *Monte Carlo Methods in Financial Engineering*, Chapter 3](https://doi.org/10.1007/978-0-387-21617-1). A bridge is the conditioning of Section 2 in one dimension: draw the free path, then correct it onto the known endpoint.
That conditioning can introduce temporal leakage into evaluation, and I derive the size of the effect in closed form. Consider scaled Brownian motion $`X_u = \sigma W_u`$, where $`s < t < T`$. Under squared-error loss, the optimal predictor at time $`s`$, given only the historical filtration $`\mathcal{F}_s`$, has conditional mean squared error $`\sigma^2(t - s)`$. If the generator conditions paths on the future endpoint $`X_T`$ and inadvertently exposes that conditioning to a forecasting task, Gaussian conditioning gives ([Pitman on the Brownian bridge](https://www.stat.berkeley.edu/~pitman/s205s03/b_bridge.pdf)):
@@ -563,15 +563,15 @@ $$
Mean reversion discounts the leak, because the endpoint tells less about the middle of a path that forgets. The financial reading is direct: endpoint leakage is most damaging for series close to a unit root, prices and long rates, and nearly harmless for fast-reverting spreads. The information-set rule is therefore enforced with a severity that follows the persistence of the series.
The task specification keeps three information sets apart:
-1. **Construction information ($`\mathcal{F}_{\text{gen}}`$):** the complete path, endpoint, or latent parameters used to generate a scenario.
+1. **Construction information ($`\mathcal{F}_{\text{gen}}`$):** the complete path, endpoint or latent parameters used to generate a scenario.
2. **Permitted predictor filtration ($`\mathcal{F}_s`$):** the historical information the task allows the model to observe at decision time $`s`$.
3. **Evaluation and oracle information ($`\mathcal{F}_{\text{eval}}`$):** outcomes realized after $`s`$, reserved for scoring or for explicitly disclosed oracle baselines.
## 5 Batch Zero: independent execution and calibration for post-training data strategy
-I created **Batch Zero** around the capability objectives and data requirements of post-training research teams. Before anything scales, it brings task design, synthetic data strategy, and preference supervision under a common set of acceptance criteria. It tests whether domain experts can independently execute a research specification, reproduce reference solutions, and calibrate their preference judgments. The resulting specifications support subsequent data construction.
-• **Independent execution:** before a task family enters large-scale data construction (*Batch One*), doctoral-level domain experts independently write seed cases, generate synthetic variants, verify reference solutions, and assign five-level pairwise preference scores. Implicit assumptions, missing information, and ambiguities in the specification come to light this way.
-• **Calibration and feedback to data strategy:** failures, expert disagreements, and boundary-case errors inform revisions to task instructions, reference solutions, grading rules, and Batch One data acceptance criteria. Task design, execution, revision, and recalibration form an iterative process.
+I created **Batch Zero** around the capability objectives and data requirements of post-training research teams. Before anything scales, it brings task design, synthetic data strategy and preference supervision under a common set of acceptance criteria. It tests whether domain experts can independently execute a research specification, reproduce reference solutions and calibrate their preference judgments. The resulting specifications support subsequent data construction.
+• **Independent execution:** before a task family enters large-scale data construction (*Batch One*), doctoral-level domain experts independently write seed cases, generate synthetic variants, verify reference solutions and assign five-level pairwise preference scores. Implicit assumptions, missing information and ambiguities in the specification come to light this way.
+• **Calibration and feedback to data strategy:** failures, expert disagreements and boundary-case errors inform revisions to task instructions, reference solutions, grading rules and Batch One data acceptance criteria. Task design, execution, revision and recalibration form an iterative process.
• **Organizational adoption:** the partner foundation model company subsequently adopted **Batch Zero** across other post-training programs.
**Why the check belongs at the seed.** Expert-written references are less reliable than their reputation. When FinanceReasoning re-examined four public financial reasoning sets, it updated 15.6% of their questions, between 9.72% and 30.00% per set ([Tang et al.](https://arxiv.org/abs/2506.05828)). A synthetic pipeline multiplies whatever error its seeds carry. If a seed's reference solution is wrong with probability $`\varepsilon`$ and the seed yields $`k`$ variants, a family carries $`k\varepsilon`$ wrong records in expectation, and they are wrong together. If an independent re-solve catches a wrong reference with probability $`d`$, then $`r`$ re-solves leave
@@ -588,7 +588,7 @@ $$
\mathrm{Var}(\bar{Y}) = \frac{Gk\sigma_Y^2 + Gk(k - 1)\rho\sigma_Y^2}{(Gk)^2} = \frac{\sigma_Y^2}{Gk}[1 + (k - 1)\rho] \implies N_{\text{eff}} = \frac{Gk}{1 + (k - 1)\rho} \tag{5.2}
$$
-With $`G = 100`$ independent families, $`k = 20`$ variants per family ($`2,000`$ nominal tasks), and within-family correlation $`\rho = 0.5`$, the 2,000 records provide the same precision for estimating the mean as about 190 independent records. Ignoring this correlation understates standard errors by a factor of **$`3.24\times`$**. My data workflow therefore separates training and evaluation sets at the task-family level, and all confidence intervals account for hierarchical clustering at both the family and expert levels.
+With $`G = 100`$ independent families, $`k = 20`$ variants per family (2,000 nominal tasks) and within-family correlation $`\rho = 0.5`$, the 2,000 records provide the same precision for estimating the mean as about 190 independent records. Ignoring this correlation understates standard errors by a factor of **$`3.24\times`$**. My data workflow therefore separates training and evaluation sets at the task-family level, and all confidence intervals account for hierarchical clustering at both the family and expert levels.
**How many variants a seed should yield.** The design effect is a diagnosis; the same algebra gives the prescription. A new family costs $`c_G`$, mostly expert time on the seed and its reference, and each further variant costs $`c_v`$. At a fixed budget the variance of the family-clustered mean is proportional to $`(c_G + k c_v)[1 + (k - 1)\rho] / k`$, and it is smallest at the optimum cluster size of two-stage sampling:
@@ -600,7 +600,7 @@ With a seed that costs twenty variants and $`\rho = 0.5`$, $`k^* \approx 4.5`$,
## 6 Staged evaluation design and coverage efficiency
-Staged, blinded comparisons across independent task families test whether constraint-preserving synthesis and preference supervision meet the data requirements of post-training research teams. The task mix, generator training budget, expert review budget, and final acceptance criteria stay fixed as each method is introduced:
+Staged, blinded comparisons across independent task families test whether constraint-preserving synthesis and preference supervision meet the data requirements of post-training research teams. The task mix, generator training budget, expert review budget and final acceptance criteria stay fixed as each method is introduced:
**Table 4. The staged evaluation design.**
@@ -609,7 +609,7 @@ Staged, blinded comparisons across independent task families test whether constr
| **Current data construction baseline** | Unconstrained generation plus expert review and rework | Fully loaded cost of each valid case that is ultimately accepted |
| **Add manifold-constrained generation** | Hold the generator fixed; apply affine projections and variable transformations | Change in invalid candidates and manual repairs, with the conditional distribution preserved |
| **Add decision-margin checks and reference-answer recomputation** | Recompute reference judgments below the margin threshold; use $`\gamma/s`$ to quantify Gaussian label-flip risk | Change in stale-label errors, with valid supervision on both sides of the decision boundary |
-| **Apply Batch Zero acceptance criteria and filtration rules** | Jointly check the temporal information boundaries in $`\mathcal{F}_s`$, independent execution, reproducibility of reference solutions, and calibration of preference standards | Post-training gains that persist once leakage is excluded and clustering by source task family is accounted for |
+| **Apply Batch Zero acceptance criteria and filtration rules** | Jointly check the temporal information boundaries in $`\mathcal{F}_s`$, independent execution, reproducibility of reference solutions and calibration of preference standards | Post-training gains that persist once leakage is excluded and clustering by source task family is accounted for |
Raw acceptance rates favor trivial variants of easy templates, and generation speed alone omits the cost of expert repairs, so I allocate data construction and review resources by scenario-weighted task-family coverage efficiency. Let $`j`$ index predefined cells in the financial capability matrix, with strategic weights $`w_j \ge 0`$ ($`\sum_j w_j = 1`$) and required counts of independent valid task families $`n_j^* > 0`$:
@@ -617,7 +617,7 @@ $$
\mathrm{Coverage} = \sum_j w_j \min\left(1, \frac{n_j^{\text{valid}}}{n_j^*}\right), \qquad \mathrm{Efficiency} = \frac{\mathrm{Coverage}}{C_{\text{generation}} + C_{\text{review}} + C_{\text{rework}} + C_{\text{adjudication}}} \tag{6.1}
$$
-Once a cell reaches its target of $`n_j^*`$ independent task families, further variants of simple templates add no coverage. Coverage is concave and piecewise linear in the family counts, so the allocation that maximizes it per unit cost fills cells in decreasing order of $`w_j / (n_j^* c_j)`$, with $`c_j`$ the cost of one valid family in cell $`j`$. Four controls enter the efficiency: manifold constraint enforcement, decision-margin label recomputation, separation of temporal information sets, and Batch Zero calibration.
+Once a cell reaches its target of $`n_j^*`$ independent task families, further variants of simple templates add no coverage. Coverage is concave and piecewise linear in the family counts, so the allocation that maximizes it per unit cost fills cells in decreasing order of $`w_j / (n_j^* c_j)`$, with $`c_j`$ the cost of one valid family in cell $`j`$. Four controls enter the efficiency: manifold constraint enforcement, decision-margin label recomputation, separation of temporal information sets and Batch Zero calibration.
The team averaged roughly 500 completed financial tasks a week under these rules, and the partner company carried Batch Zero into its other post-training programs.
@@ -656,7 +656,7 @@ A model can reach the right answer through an unsupported step, and a grader tha
The question I took up in late 2023 was what should entitle a model's reasoning to be accepted. A correct answer can coexist with unsupported assumptions, misapplied rules or tool use that violates the task specification, and a fixed reference path will reject sound work that takes another route. I designed an acceptance rule around proof obligations: a conclusion stands when its assumptions, derivations and execution evidence support it, whichever valid route the model took, and a failed obligation removes support from everything that depends on it. Programmatic certificates discharge checkable obligations, LLM-assisted screening identifies possible gaps, and specialist scholars adjudicate disputed semantic questions.
-Two things follow from defining acceptance this way. A certificate bounds how far a submission is from the truth using only a residual the submission exposes and a constant the task designer has verified, so the check never consults a reference solution and inherits none of its errors. And once a grader's two error rates are named, they can be priced: I derived the false-acceptance rate below which a leaderboard lead is certified, the level of true validity a policy can reach when the grader is its reward, and the screening accuracy at which an LLM judge may decide instead of screen. I kept refining the design through September 2026. Sections 1 through 7 set out the acceptance criteria, the certificates, the thresholds, the measurements and the related literature.
+Two things follow from defining acceptance this way. A certificate bounds how far a submission is from the truth using only a residual the submission exposes and a constant the task designer has verified, so the check never consults a reference solution and inherits none of its errors. And once a grader's two error rates are named, they can be priced: I derived the false-acceptance rate below which a leaderboard lead is certified, the level of true validity a policy can reach when the grader is its reward, and the screening sensitivity at which an LLM judge may decide instead of screen. I kept refining the design through September 2026. Sections 1 through 7 set out the acceptance criteria, the certificates, the thresholds, the measurements and the related literature.
The team produced roughly 500 scientific reasoning tasks a week. Each task included a detailed, step-by-step reference solution and repeated verification of the argument.
@@ -666,7 +666,7 @@ The team produced roughly 500 scientific reasoning tasks a week. Each task inclu
* **I developed checkable error certificates and bounds on solution quality for programmatically verifiable task families.** The projected-residual optimality bound ($`0 \le f(x) - v^* \le \|P_{\ker A} r\|_2^2 / (2\mu)`$) splits feasibility, stationarity and the accuracy of a conclusion into separate verification obligations, and different proofs can pass by satisfying the same ones. The same template, a residual the submission exposes, a constant the designer verifies and an inequality between them, covers convex programs, linear systems, eigenpairs and fixed points. None of these checks needs a reference solution.
* **I established complementary forms of verification across mathematical certificates, code and rule checks, LLM-assisted screening and specialist adjudication, and later programs reused the layered design.** Code checks directly verifiable answers, required deliverables, tool permissions and other requirements specified in advance. LLM-as-a-Judge helps flag missing evidence, misapplied rules and gaps in reasoning, and disputed domain judgments go to scholars with relevant expertise. I derived when a screen may decide: at screening sensitivity $`\text{se}`$, a share $`1 - \alpha^*/(1 - \text{se})`$ of unflagged work still has to reach specialists to hold the system's false acceptance at $`\alpha^*`$.
* **I evaluated answers, visible reasoning and execution evidence together.** I cross-checked the model's displayed CoT against its actual tool calls, return values and intermediate outputs to see whether what it claimed had observable support. From elapsed time, attempt counts and changes in intermediate results I identified repeated operations, continued use of failed assumptions and attempts that added no evidence. Correctness, compliance and efficiency were recorded as separate results, and a resource limit counted toward acceptance only when the task specification set it beforehand.
-* **I derived how false acceptance ($`\alpha`$) and false rejection ($`\beta`$) reverse model rankings, and the condition under which a ranking is certified.** In the worked example, models with true validity rates of 72% and 64% score 73.12% and 76.96%, so the weaker one ranks first. A lead of $`g`$ points over a rival scoring $`s_B`$ certifies the order of true validity exactly when the grader's false-acceptance rate is below $`g / (1 - s_B)`$.
+* **I derived how false acceptance ($`\alpha`$) and false rejection ($`\beta`$) reverse model rankings, and the condition under which a ranking is certified.** In the worked example, models with true validity rates of 72% and 64% score 73.12% and 76.96%, so the weaker one ranks first. A lead of $`g`$ over a rival scoring $`s_B`$ certifies the order of true validity exactly when the grader's false-acceptance rate is below $`g / (1 - s_B)`$.
* **I showed that a policy optimized against a grader converges to the grader's positive predictive value.** The observed score goes to 100% while true validity stops at $`\theta(1-\beta) / [\theta(1-\beta) + (1-\theta)\alpha]`$. False acceptance sets the ceiling, and false rejection removes valid routes from what the policy produces. The same analysis gives the bias in the binary cross-entropy gradient used to train process reward models ($`E[\nabla_s \ell(s,V) - \nabla_s \ell(s,Z) \mid h] = p(h) - \tilde{p}(h)`$).
* **I gave the tolerance-propagation rule ($`a\delta + 2\varepsilon`$) for bijective task transformations and its composition law for chains of transformations.** Together they say when a transformed task still supports a valid comparison and what tolerance its solutions are owed.
* **On a test set of 3,600 exact-arithmetic submissions from 600 strongly convex quadratic task families, every verdict known, final-answer matching accepted 75.0% of the invalid submissions and strict reference-path matching rejected 50.0% of the valid proofs.** Certificate checking accepted all 1,200 valid proofs and rejected all 2,400 invalid submissions, so strict error detection and acceptance of valid alternative routes held together on the same set.
@@ -792,7 +792,7 @@ $$
J(u) - v \le \delta \implies J'(\Phi(u)) - v' \le a\delta + 2\varepsilon \tag{4.3}
$$
-Taking infima in the first line gives the second, and subtracting the lower bound on $`v'`$ from the upper bound on $`J'(\Phi(u))`$ gives the third. So after a unit conversion, parameter rescaling or numerical transformation with error bounded by $`\varepsilon`$, the result determines the acceptance tolerance for a previously $`\delta`$-optimal solution. Keeping the original tolerance $`\delta`$ can misclassify a valid transformed solution as a model failure, while relaxing it arbitrarily can conceal a real capability gap. When $`\varepsilon = 0`$, the bijection and positive scaling $`a > 0`$ also preserve whether an optimum is attained. Changes to boundary compactness or admissible control classes require a new attainment proof.
+Taking infima in (4.1) gives (4.2), and subtracting the lower bound on $`v'`$ from the upper bound on $`J'(\Phi(u))`$ gives (4.3). So after a unit conversion, parameter rescaling or numerical transformation with error bounded by $`\varepsilon`$, the result determines the acceptance tolerance for a previously $`\delta`$-optimal solution. Keeping the original tolerance $`\delta`$ can misclassify a valid transformed solution as a model failure, while relaxing it arbitrarily can conceal a real capability gap. When $`\varepsilon = 0`$, the bijection and positive scaling $`a > 0`$ also preserve whether an optimum is attained. Changes to boundary compactness or admissible control classes require a new attainment proof.
Variants are usually built by several transformations in a row. Applying the bound step by step to a chain with parameters $`(a_1, \varepsilon_1), \dots, (a_k, \varepsilon_k)`$ gives the composition law:
@@ -838,7 +838,7 @@ $$
The limit is the grader's positive predictive value at the starting policy, and the observed score goes to 100% along the way. For Model B of Section 1 under final-answer matching:
-**Table 2. Observed score of Model B under final-answer matching as optimization strengthens.**
+**Table 2. Observed score and true validity of Model B under final-answer matching as optimization strengthens.**
| Optimization strength $`\eta`$ | 0 | 1 | 2 | 4 | Limit |
| --- | --- | --- | --- | --- | --- |
@@ -856,31 +856,31 @@ The cost of false acceptance is the ceiling and the cost of false rejection is d
Measuring a grader's two error rates requires submissions whose validity is known without appeal to any grader. I built such a test set from 600 strongly convex quadratic task families. Using exact rational arithmetic, each family yields 6 algebraic submissions, for 3,600 in total:
• 1,200 valid certificates, 2 per family: one derived using Lagrange multipliers and one using equality elimination.
-• 2,400 submissions with controlled defects, 4 per family: (1) an infeasible construction, $`Ax \neq b`$, that yields the exact optimal value $`v^*`$; (2) failure of stationarity along a feasible direction in $`\ker A`$ while reporting $`v^*`$; (3) a claimed multiplier that fails the stationarity equation while reporting $`v^*`$; and (4) an incorrect final objective value. The first three classes, comprising 1,800 submissions or 75% of all invalid submissions, report the correct final answer $`v^*`$.
+• 2,400 submissions with controlled defects, 4 per family: (1) an infeasible construction, $`Ax \neq b`$, that reports the exact optimal value $`v^*`$; (2) failure of stationarity along a feasible direction in $`\ker A`$ while reporting $`v^*`$; (3) a claimed multiplier that fails the stationarity equation while reporting $`v^*`$; and (4) an incorrect final objective value. The first three classes, comprising 1,800 submissions or 75% of all invalid submissions, report the correct final answer $`v^*`$.
The composition is the instrument's design: three ways of being wrong with the right answer, one way of being wrong with the wrong answer, and two ways of being right. I compared three deterministic acceptance protocols on all 3,600 submissions:
**Table 3. Three acceptance protocols on 3,600 submissions with known verdicts.**
-| Acceptance protocol | False acceptance rate on invalid submissions ($`2,400`$) | False rejection rate on valid certificates ($`1,200`$) | Valid share of accepted submissions |
+| Acceptance protocol | False acceptance rate on invalid submissions (2,400) | False rejection rate on valid certificates (1,200) | Valid share of accepted submissions |
| --- | --- | --- | --- |
| Final-answer matching only | 75.0% (accepts 1,800 flawed solutions) | 0.0% | 40.0% |
| Strict reference-path matching | 0.0% | 50.0% (rejects 600 valid elimination proofs) | 100.0% |
| Certificate verification | 0.0% | 0.0% | 100.0% |
-Certificate verification accepted 1,800 fewer invalid proofs than final-answer matching and retained 600 more valid alternative solutions than strict reference-path matching. Strict error detection and acceptance of valid alternative proofs held together on the same submissions, which neither baseline managed. The verifier also passed all 600 feasible-perturbation checks on the projected-residual bound in Equation (3.2), and it rejected every submission of defect class (3), the multiplier that a partial witness check would have let through.
+Certificate verification accepted 1,800 fewer invalid proofs than final-answer matching and retained 600 more valid alternative solutions than strict reference-path matching. Strict error detection and acceptance of valid alternative proofs held together on the same submissions, which neither baseline managed. The projected-residual bound in equation (3.4) also held at all 600 feasible perturbations, one per family, and the verifier rejected every submission of defect class (3), the multiplier that a partial witness check would have let through.
The last column is the ceiling of Section 5. A policy trained against final-answer matching on this mix moves from one third valid work to 40% and stops. Trained against the reference path it reaches 100% and loses the elimination route. Trained against certificate checking it reaches 100% with both routes intact.
## 7 Dependence within task families and the economics of certificates
-Surface variations of one problem template do not provide independent observations of general reasoning ability. Suppose there are $`F`$ independent task families, each with $`K`$ variants, common marginal variance $`\sigma^2`$, and common pairwise within-family correlation $`\rho`$. The sample mean then carries the Kish design effect $`1 + (K - 1)\rho`$, and its variance and effective sample size $`n_{\text{eff}}`$ are:
+Surface variations of one problem template do not provide independent observations of general reasoning ability. Suppose there are $`F`$ independent task families, each with $`K`$ variants, common marginal variance $`\sigma^2`$, and common pairwise within-family correlation $`\rho_w`$. The sample mean then carries the Kish design effect $`1 + (K - 1)\rho_w`$, and its variance and effective sample size $`n_{\text{eff}}`$ are:
$$
-\mathrm{Var}(\bar{Z}) = \frac{\sigma^2}{FK}[1 + (K - 1)\rho], \qquad n_{\text{eff}} = \frac{FK}{1 + (K - 1)\rho} \tag{7.1}
+\mathrm{Var}(\bar{Z}) = \frac{\sigma^2}{FK}[1 + (K - 1)\rho_w], \qquad n_{\text{eff}} = \frac{FK}{1 + (K - 1)\rho_w} \tag{7.1}
$$
-For $`F = 50`$ families with $`K = 8`$ variants each, there are $`400`$ responses, and if $`\rho = 0.6`$, those $`400`$ responses provide the same precision as $`n_{\text{eff}} \approx 77`$ independent observations. Training and test sets are therefore split by task family, and new tasks vary assumptions, evidentiary relationships or the structure of a derivation, beyond substituting numbers.
+For $`F = 50`$ families with $`K = 8`$ variants each, there are $`400`$ responses, and if $`\rho_w = 0.6`$, those $`400`$ responses provide the same precision as $`n_{\text{eff}} \approx 77`$ independent observations. Training and test sets are therefore split by task family, and new tasks vary assumptions, evidentiary relationships or the structure of a derivation, beyond substituting numbers.
Certificates cost engineering once and almost nothing afterwards. Let $`C_0`$ be the upfront engineering cost of the proof-obligation certificates and verifier, $`c_e`$ the expert review cost per submission, $`c_v`$ the automated verifier's unit execution cost, and $`u`$ the fraction of submissions escalated to human adjudication at the same expert unit cost. The certificate workflow is the cheaper one when submission volume $`N`$ exceeds:
@@ -943,9 +943,9 @@ The architecture I developed takes a traceable evidence state as its input and b
## 1 Feature attribution and the value of verification
-[SHAP](https://arxiv.org/abs/1705.07874) splits a model prediction into additive feature contributions through its local accuracy property. Conditional Shapley methods also account for dependence among features, and Rashomon-set methods check how stable explanations are across near-optimal models ([SHAP](https://arxiv.org/abs/1705.07874); [Aas et al.](https://arxiv.org/html/1903.10464v3); [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html)). All of them take the corporate inputs $`x`$ as given. In investment diligence the inputs are the open question, because self-reported disclosures can be polished well past what the underlying facts support, and an analyst needs to know which corrections to the evidence could change a decision. Attribution was not built for that question, and there is now a theorem to the effect: for expressive model classes, attribution methods that are complete and linear, SHAP among them, can do no better than random guessing at inferring how the model behaves under changed inputs ([Bilodeau et al.](https://arxiv.org/abs/2212.11870)). The remedy those authors recommend is to evaluate the model directly on the changes of interest, which is what the bounds of Sections 2 to 4 do.
+[SHAP](https://arxiv.org/abs/1705.07874) splits a model prediction into additive feature contributions through its local accuracy property. Conditional Shapley methods also account for dependence among features, and Rashomon-set methods check how stable explanations are across near-optimal models ([SHAP](https://arxiv.org/abs/1705.07874); [Aas et al.](https://arxiv.org/html/1903.10464v3); [Marx et al.](https://proceedings.mlr.press/v206/marx23a.html)). All of them take the corporate inputs $`x`$ as given. In investment diligence the inputs are the open question, because self-reported disclosures can be polished well past what the underlying facts support, and an analyst needs to know which corrections to the evidence could change a decision. Attribution was not built for that question, and there is now a theorem to that effect: for expressive model classes, attribution methods that are complete and linear, SHAP among them, can do no better than random guessing at inferring how the model behaves under changed inputs ([Bilodeau et al.](https://arxiv.org/abs/2212.11870)). The remedy those authors recommend is to evaluate the model directly on the changes of interest, which is what the bounds of Sections 2 to 4 do.
-Neighboring work bears on other parts of the problem. [Aggregate Confusion](https://mitsloan.mit.edu/sustainability-initiative/aggregate-confusion-project) documents how widely ESG ratings diverge across agencies and traces most of the divergence to measurement. [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899) selects pairwise comparison queries under set-based uncertainty, with worst-case utility and regret as objectives. My framework works in institutional research where the facts are still being checked, and it connects model disagreement, evidence changes, ranking stability and the choice of what to verify next.
+Neighboring work bears on other parts of the problem. [Aggregate Confusion](https://academic.oup.com/rof/article/26/6/1315/6590670) documents how widely ESG ratings diverge across agencies and traces most of the divergence to measurement. [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899) selects pairwise comparison queries under set-based uncertainty, with worst-case utility and regret as objectives. My framework works in institutional research where the facts are still being checked, and it connects model disagreement, evidence changes, ranking stability and the choice of what to verify next.
The first failure case concerns verification priorities based on raw attribution magnitudes. Take a linear scoring model with a fixed marginal background distribution $`Q`$. The [SHAP](https://arxiv.org/abs/1705.07874) attribution for feature $`j`$ is $`\phi_j = \beta_j(x_j - E_Q X_j)`$. If the feature has unresolved factual uncertainty bounded by $`\left\vert\delta_j\right\vert \le d_j`$, its sensitivity to verification is $`\left\vert\beta_j\right\vert d_j`$, and over a box of admissible evidence errors the exact worst-case change in the company score is
@@ -1009,7 +1009,7 @@ If $`\underline{\Delta}_{ik} > 0`$, the ordering $`i \succ k`$ holds for every r
• The output of diligence is a certified partial order over the companies, and verification can only extend it: resolving evidence shrinks $`\mathcal{X}(E)`$, so every lower bound stays where it is or rises.
• Progress has a measure, the share of company pairs that are certified, and a diligence plan can be judged by how far it moves that share per analyst hour.
-• Selecting the top $`K`$ of $`n`$ companies needs only the $`K(n - K)`$ certificates that put each selected company above each rejected one. For five of twenty that is 75 certificates in place of all 190 pairs.
+• Selecting the top $`K`$ of $`N`$ companies needs only the $`K(N - K)`$ certificates that put each selected company above each rejected one. For five of twenty that is 75 certificates in place of all 190 pairs.
Three neighboring constructions come close and stop short of this. Robust ordinal regression reports the "necessary" preferences that hold for every value function compatible with a decision maker's stated judgments, and shows that the necessary relation grows with information ([Greco et al.](https://ideas.repec.org/a/eee/ejores/v191y2008i2p416-436.html)); its uncertainty is entirely on the model side and the inputs are taken as known. Certain predictions hold for every classifier trained on every possible completion of an incomplete data set, and the companion cleaning rule chooses which row a person should fix next ([Karlaš et al.](https://arxiv.org/abs/2005.05117)); its possible worlds are independent across rows, so an error shared by two units cannot cancel, and its selection rule counts worlds, which amounts to a uniform prior over them. Confidence sets for ranks give simultaneous statistical statements about league tables ([Mogstad et al.](https://www.nber.org/papers/w26883)); they pay for simultaneity and treat each unit's estimate as a noisy draw, with no account of unverified inputs. The joint set $`\Omega(E)`$ carries model multiplicity and cross-company evidence error together, without a prior.
@@ -1051,10 +1051,10 @@ Nonsmooth tree ensembles get their bounds from leaf-region partitioning or mixed
When $`\underline{\Delta}_{ik} \le 0 < \overline{\Delta}_{ik}`$, the ranking is unresolved, since a factual correction or a resolution of ambiguous evidence could change the decision. The system then prioritizes the next verification query $`q`$, such as reviewing a supplier contract, requesting original utility bills, or checking a regulatory filing.
-Let $`A`$ be a finite set of institutional actions, such as advancing company $`i`$ or company $`k`$ to final diligence, or deferring a rating. Let $`U(a, \omega)`$ be the prespecified utility of an action in scenario $`\omega \in \Omega`$, where $`\Omega`$ is nonempty and utilities are uniformly bounded. The baseline minimax regret over deterministic actions is
+Let $`A`$ be a finite set of institutional actions, such as advancing company $`i`$ or company $`k`$ to final diligence, or deferring a rating. Let $`W(a, \omega)`$ be the prespecified utility of an action in scenario $`\omega \in \Omega`$, where $`\Omega`$ is nonempty and utilities are uniformly bounded. The baseline minimax regret over deterministic actions is
$$
-R^*(\Omega) = \min_{a \in A} \sup_{\omega \in \Omega} \left[ \max_{b \in A} U(b, \omega) - U(a, \omega) \right] \tag{4.1}
+R^*(\Omega) = \min_{a \in A} \sup_{\omega \in \Omega} \left[ \max_{b \in A} W(b, \omega) - W(a, \omega) \right] \tag{4.1}
$$
A query $`q`$ returns an observation $`o \in O_q`$ and restricts the compatible scenarios to a nonempty subset $`\Omega_{q,o} \subseteq \Omega`$. I score the query by its guaranteed reduction in worst-case decision regret:
@@ -1063,7 +1063,7 @@ $$
V(q) = R^*(\Omega) - \sup_{o \in O_q} R^*(\Omega_{q,o}) \ge 0 \tag{4.2}
$$
-Nonnegativity follows from $`\Omega_{q,o} \subseteq \Omega`$: restricting the scenario set cannot increase worst-case regret for any fixed action, and minimizing over $`a \in A`$ preserves that inequality. Calculating $`V(q)`$ uses the set of compatible scenarios and requires no prior probability that a disclosure is exaggerated. Minimax regret is Savage's criterion, and choosing queries by it comes from utility elicitation ([Boutilier et al.](https://www.cs.toronto.edu/~cebly/Papers/BPPS-aij06-abs.html); [Robust Active Preference Elicitation](https://arxiv.org/html/2003.01899)). Those methods elicit a decision maker's preferences. Here the queries are facts about companies, their answers move the inputs of every model in $`\mathcal{F}_\alpha`$ at once, and they cost analyst hours.
+Nonnegativity follows from $`\Omega_{q,o} \subseteq \Omega`$: restricting the scenario set cannot increase worst-case regret for any fixed action, and minimizing over $`a \in A`$ preserves that inequality. Calculating $`V(q)`$ uses the set of compatible scenarios and requires no prior probability that a disclosure is exaggerated. Minimax regret is Savage's criterion, and choosing queries by it comes from utility elicitation ([Boutilier et al.](https://www.cs.toronto.edu/~cebly/Papers/BPPS-aij06-abs.html); [Robust Active Preference Elicitation](https://arxiv.org/abs/2003.01899)). Those methods elicit a decision maker's preferences. Here the queries are facts about companies, their answers move the inputs of every model in $`\mathcal{F}_\alpha`$ at once, and they cost analyst hours.
**The exact plan for affine scores.** For a choice between two companies under an affine score with interval evidence, the whole problem has a closed form. Write $`m`$ for the current margin, $`h = \sum_j s_j`$ for the total swing, and $`s_J = \sum_{j \in J} s_j`$ for the swing of a set $`J`$ of fields. Verify the fields in $`J`$, let the action be chosen again once their values are known, and take the worst values they could turn out to have. The regret that remains, and the guaranteed value of the set, are
@@ -1086,7 +1086,7 @@ With two-point evidence in place of intervals the worst observation cannot land
## 5 Propagating factual corrections through scores and attributions
-When diligence resolves a query and the verified evidence vector changes from $`x`$ to $`x'`$, the system follows versioned dependencies to recompute the affected scores, peer rankings and [SHAP](https://arxiv.org/abs/1705.07874) explanations. With the scoring function $`f`$, background reference $`Q`$ and rubric version held fixed, Shapley efficiency gives the exact identity ([SHAP local accuracy](https://arxiv.org/pdf/1705.07874)):
+When diligence resolves a query and the verified evidence vector changes from $`x`$ to $`x'`$, the system follows versioned dependencies to recompute the affected scores, peer rankings and [SHAP](https://arxiv.org/abs/1705.07874) explanations. With the scoring function $`f`$, background reference $`Q`$ and rubric version held fixed, Shapley efficiency gives the exact identity ([SHAP local accuracy](https://arxiv.org/abs/1705.07874)):
$$
f(x) = \phi_0 + \sum_j \phi_j(f, Q, x), \qquad \sum_j [\phi_j(f, Q, x') - \phi_j(f, Q, x)] = f(x') - f(x) \tag{5.1}
@@ -1101,11 +1101,11 @@ Every database update records `(field version, evidence document ID, model versi
Five worked examples fix the size of each effect under stated inputs.
1. **Confidence interval coverage under common reporting bias.**
-• Set the shared bias to $`b = 0.2`$ and independent noise to $`\sigma = 1`$. If the common bias $`b`$ is ignored, a nominal 95% interval has actual coverage of 48.40% at $`n = 100`$ records and 0.000637% at $`n = 1,000`$. Once targeted verification shows $`\left\vert b\right\vert \le 0.2`$, the half-width $`0.2 + 1.96/\sqrt{n}`$ gives 97.50% coverage at $`n = 100`$, and the bias-aware half-width of Section 1 is $`0.3645`$ against $`0.396`$: 8.0% shorter at exactly 95%. The threshold is $`n^* \approx 68`$ records. Past it, verification is the better use of the next analyst hour.
+• Set the shared bias to $`b = 0.2`$ and independent noise to $`\sigma = 1`$. If the common bias $`b`$ is ignored, a nominal 95% interval has actual coverage of 48.40% at $`n = 100`$ records and 0.000637% at $`n = 1{,}000`$. Once targeted verification shows $`\left\vert b\right\vert \le 0.2`$, the half-width $`0.2 + 1.96/\sqrt{n}`$ gives 97.50% coverage at $`n = 100`$, and the bias-aware half-width of Section 1 is $`0.3645`$ against $`0.396`$: 8.0% shorter at exactly 95%. The threshold is $`n^* \approx 68`$ records. Past it, verification is the better use of the next analyst hour.
2. **Cancellation of shared errors in pairwise rankings.**
• Two companies have a nominal score difference of $`0.6`$ points and share an additive sector-benchmark error in $`[-1, 1]`$. Subtracting separately constructed company intervals gives $`[-1.4, 2.6]`$, which cannot certify the ordering, whereas the joint support-function calculation cancels the shared error and gives $`[0.6, 0.6]`$, which certifies the ranking.
3. **Minimax-regret verification compared with [SHAP](https://arxiv.org/abs/1705.07874) magnitude.**
-• Consider a single query, with the same cost for each query option. The example has 8 scenarios, which come from 2 candidate models in the Rashomon set ($`s \in \{-1, 1\}`$) and 4 unresolved evidence configurations ($`z_1 \in \{-0.4, 0.4\}`$, $`z_2 \in \{-0.2, 0.2\}`$). The pairwise score difference is $`\Delta_s = 0.15s + z_1 + 0.3s z_2`$. A further shared feature has a large [SHAP](https://arxiv.org/abs/1705.07874) attribution but is already verified ($`d_0 = 0`$). The action is to select company $`i`$ or company $`k`$, and regret is measured in model-score units.
+• Consider a single query, with the same cost for each query option. The example has eight scenarios, which come from two candidate models in the Rashomon set ($`\theta \in \{-1, 1\}`$) and four unresolved evidence configurations ($`z_1 \in \{-0.4, 0.4\}`$, $`z_2 \in \{-0.2, 0.2\}`$). The pairwise score difference is $`\Delta_\theta = 0.15\theta + z_1 + 0.3\theta z_2`$. A further shared feature has a large [SHAP](https://arxiv.org/abs/1705.07874) attribution but is already verified ($`d_0 = 0`$). The action is to select company $`i`$ or company $`k`$, and regret is measured in model-score units.
**Table 1. Guaranteed regret reduction of each verification query.**
diff --git a/website/content/sources/selected-research/asian-source.md b/website/content/sources/selected-research/asian-source.md
index 6f573b1..73fd58c 100644
--- a/website/content/sources/selected-research/asian-source.md
+++ b/website/content/sources/selected-research/asian-source.md
@@ -1,10 +1,8 @@
-# Certified Valuation of Arithmetic Asian Options
+# Certified Valuation of Arithmetic Asian Options: The Cost of Exercising on the Geometric Average
[English PDF](../paper/paper.pdf) · [Repository](../README.md) · [Code and evidence](../docs/EVIDENCE.md)
-The Cost of Exercising on the Geometric Average
-
-**Theodore Ouyang**
+**Letao (Theodore) Ouyang**
[theodore.oy2025@gmail.com](mailto:theodore.oy2025@gmail.com) · [10@alumni.duke.edu](mailto:10@alumni.duke.edu)
@@ -496,7 +494,7 @@ The crude oil rows are options on a calendar month of 21 daily settlements of a
Four observations follow.
-For the monthly contracts the proxy value is the price for every commercial purpose. The exchange quotes these options in dollars and cents per barrel; the proved bound is between a 120th and a 440th of a cent. A market maker who exercises on the geometric average gives up less than ten dollars on a hundred contracts of a thousand barrels, whose premium exceeds three hundred thousand.
+For the monthly contracts the proxy value is the price for every commercial purpose. The exchange quotes these options in dollars and cents per barrel; the proved bound is between a 124th and a 446th of a cent. A market maker who exercises on the geometric average gives up less than ten dollars on a hundred contracts of a thousand barrels, whose premium exceeds three hundred thousand.
The bound is nearly attained. Its ratio to the simulated cost is 1.54, 1.19, 1.10 and 1.07 at one, three, six and twelve months, as the factor in (3.12) predicts: with a long lead the strike is within a small multiple of $`\sigma`$ of both averages on almost every path. The classical first-order estimate is 2.6 cents at every lead time, between 320 and 1,140 times larger.
@@ -518,7 +516,7 @@ Theorem 3.3 holds for every model that satisfies (CF), and under stochastic vola
| Crude oil call, 12 months ahead | 8.891 | 0.0000496 | 0.0000576 | 1.16 | 3.41 |
| Annual-average oil put at 70 | 2.053 | 0.00965 | 0.0296 | 3.06 | 5.05 |
-Each row uses $`2^{19}`$ paths ($`2^{18}`$ for the annual put) of absorbed Euler with four steps per trading day (two for the annual put). Standard errors are below 0.8 percent of each cost and bound, except 1.1 and 1.6 percent for the cost and bound of the annual put, and about 0.003 for the option values (0.009 for the annual put). The bound is the smaller of (3.10) at $`c_*`$ and (3.11) at $`c_\dagger`$.
+Each row uses $`2^{19}`$ paths ($`2^{18}`$ for the annual put) of absorbed Euler with four steps per trading day (two for the annual put). Standard errors are below 0.8 percent of each cost and bound, except 1.2 and 1.6 percent for the cost and bound of the annual put, and about 0.003 for the option values (0.009 for the annual put). The bound is the smaller of (3.10) at $`c_*`$ and (3.11) at $`c_\dagger`$.
Volatility of volatility roughly doubles the cost of the kink at three months and beyond, and the bound stays within a factor of 1.3 of it. For the monthly options the cost stays below a hundredth of a cent. For the annual put it is 0.96 cents per barrel and the bound 2.96 cents: 3.2 and 9.8 million dollars on 330 million barrels, against a value of 677 million. A long window is where a model with random volatility makes exercising on the geometric average expensive enough to measure, and where the interval of the next section earns its place.
@@ -783,7 +781,7 @@ Reported widths are computed from exact rational endpoints and rounded outward,
### 5.5 What the interval contains
-The interval for $`p_c`$ is $`.0105`$ wide, 16 basis points of the price. Its center is the proxy value, known to $`\pm.00021`$, and almost all of its width is the allowance for the two costs of the kink, $`.005407`$ above and $`.004670`$ below.
+The interval for $`p_c`$ is $`.0105`$ wide, 16 basis points of the price. It is built around the proxy value, known to $`\pm.00021`$, and almost all of its width is the allowance for the two costs of the kink, $`.005407`$ above and $`.004670`$ below.
A simulation locates the price inside the interval. On $`2^{20}`$ paths with the first-period factor integrated in closed form, the discounted cost of the kink is $`.000813`$ at $`K_1`$ and $`.000711`$ at $`K_2`$, with standard errors below $`4\cdot10^{-6}`$. The two enter the spread with opposite signs, so the spread exceeds its proxy value by $`.000102`$ with standard error $`.000004`$, and the continuous price is close to $`6.51335`$. The certificate bounds each cost on its own and from one side, so its allowances add. They are about 6.6 times the simulated costs, in line with the Black rows of Table 1, where the ratio is 4.7 before volatility of volatility is added. The lower bound (3.12), evaluated on the same paths, is $`.000777`$ and $`.000682`$, within five percent of the two costs. The distance between the upper bound and the cost is therefore the Gaussian factor in (3.12): on most paths that carry dispersion the strike is several $`\sigma`$ away from both averages.
@@ -859,7 +857,7 @@ The certificate also holds at a second parameter point, $`v_0=.04`$ with everyth
| Euler gap | $`[-0.011798506840,\;0.011383739259]`$ |
| Width of the gap interval | 0.023182246099 |
-The gap interval is 7.0 percent wider than at $`v_0=.045`$, and 83 percent of the increase comes from the discrete frequency tail, which rises from $`.000368746601`$ to $`.001629880668`$: a smaller $`v_0`$ lowers the floor in (5.8) and raises the first-month Laplace bound in (C.7). Theorem 5.1 itself holds for any spread, fixing schedule and parameter point; each certificate is one run of the validated computation that supplies its inputs.
+The gap interval is 7.0 percent wider than at $`v_0=.045`$, and 83 percent of the increase comes from the discrete frequency tail, which rises from $`.000368746602`$ to $`.001629880669`$: a smaller $`v_0`$ lowers the floor in (5.8) and raises the first-month Laplace bound in (C.7). Theorem 5.1 itself holds for any spread, fixing schedule and parameter point; each certificate is one run of the validated computation that supplies its inputs.
## 6. What discretization does to the price
@@ -1160,7 +1158,7 @@ A put is $`K`$-Lipschitz in the log price, and the spread is $`K_2`$-Lipschitz i
-the resulting bound on the Euler gap is $`.020948`$ for the Asian spread and $`.016702`$ for each of the nine puts with maturities $`1/4,1/2,1`$ and strikes $`90,100,110`$.
+the resulting bound on the Euler gap is $`.020920`$ for the Asian spread and $`.016701`$ for each of the nine puts with maturities $`1/4,1/2,1`$ and strikes $`90,100,110`$.
### 7.2 The posterior
@@ -1248,7 +1246,7 @@ Partition $`[.03,.06]`$ into 4096 cells. On each cell and over the whole $`\zeta
-In volatility terms the fitted initial variance moves by at most $`1.5\cdot10^{-5}`$ and $`3.7\cdot10^{-5}`$, less than a hundredth of a volatility point. The transport term is 39 percent of the first bound and 62 percent of the second: with nine quotes the envelopes of the posterior are wider and the coupling takes more steps. Table 11 refines the cells.
+The fitted initial variance moves by at most $`1.5\cdot10^{-5}`$ and $`3.7\cdot10^{-5}`$; in volatility terms that is at most 0.43 and 1.06 hundredths of a volatility point. The transport term is 39 percent of the first bound and 62 percent of the second: with nine quotes the envelopes of the posterior are wider and the coupling takes more steps. Table 11 refines the cells.
**Table 11. Bound on the displacement of every quantile of the Asian price.**
@@ -1258,7 +1256,7 @@ In volatility terms the fitted initial variance moves by at most $`1.5\cdot10^{-
| 8192 | 0.006452960804 | 0.010367612938 | 2 / 7 |
| 16384 | 0.006061495591 | 0.009584682511 | 3 / 12 |
-At 16384 cells the bounds are 24.4 and 24.6 percent smaller than at 4096, with more coupling steps at each refinement. For the continuous-model posterior at 4096 cells, the 2.5, 50 and 97.5 percent quantiles of the initial variance lie in
+At 16384 cells the bounds are 24.4 and 24.6 percent smaller than at 4096, with couplings of 3 and 12 steps instead of 2 and 5. For the continuous-model posterior at 4096 cells, the 2.5, 50 and 97.5 percent quantiles of the initial variance lie in
@@ -1522,7 +1520,7 @@ Set $`s_b=\lambda+h\left\vert b\right\vert`$. The Gaussian moment formula reduce
-Take $`\lambda=12`$. Suppose $`p`$ lies in the specified loading range, $`\eta_p\ge\eta_*`$, $`\gamma_p\le\gamma_*`$, and $`\left\vert b\right\vert\le b_*`$. Checking lower bounds for the concave quadratic at the two endpoints $`s_b=12`$ and $`s_b=12+b_*h`$ suffices to ensure that the expression in parentheses is at least 8.
+Take $`\lambda=12`$. Suppose $`p`$ lies in the specified loading range, $`\eta_p\ge\eta_*`$, $`\gamma_p\le\gamma_*`$, and $`\left\vert b\right\vert\le b_*`$. Checking lower bounds for the concave quadratic at the two endpoints $`s_b=12`$ and $`s_b=12+b_*h`$ suffices to ensure that the expression in parentheses exceeds 8.
### B.2 From local error bounds to true prefixes
@@ -2222,7 +2220,7 @@ The target radii from Appendix E, together with (F.3) and (F.7), yield (7.8). Th
## Appendix G. Programs and evidence
-The publication is the `V_20261101` tag of [github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/tree/V_20261101). It has two computational parts. The directory `code/` is the certified package: every number in Corollary 5.2, Tables 3–7, 9 and 11 and Appendix H is an outward-rounded enclosure produced there, and its file `code/MANIFEST.json` identifies the sources. The directory `studies/` holds the calculations of Sections 4 and 6 and the simulations quoted in Section 5: closed forms in ball arithmetic for Black's model, and binary64 estimates that are labeled as such wherever they appear. The root `PUBLICATION-MANIFEST.json` lists every file with its hash.
+The publication is the `V_202611` tag of [github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options](https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/tree/V_202611). It has two computational parts. The directory `code/` is the certified package: every number in Corollary 5.2, Tables 3–7, 9 and 11 and Appendix H is an outward-rounded enclosure produced there, and its file `code/MANIFEST.json` identifies the sources. The directory `studies/` holds the calculations of Sections 4 and 6 and the simulations quoted in Section 5: closed forms in ball arithmetic for Black's model, and binary64 estimates that are labeled as such wherever they appear. The root `PUBLICATION-MANIFEST.json` lists every file with its hash.
**Table G.1. Modules of the certified package.**
@@ -2463,9 +2461,9 @@ Define
-
-
-
+
+
+
@@ -2473,9 +2471,9 @@ For the deterministic $`M_K,N_K`$ in the coupling proposition, the pointwise bou
-
-
-
+
+
+
diff --git a/website/content/sources/selected-research/dependence/declarations.tex b/website/content/sources/selected-research/dependence/declarations.tex
index 89f6017..af7fa00 100644
--- a/website/content/sources/selected-research/dependence/declarations.tex
+++ b/website/content/sources/selected-research/dependence/declarations.tex
@@ -7,5 +7,5 @@ \subsection*{Use of artificial intelligence}
\subsection*{Funding and competing interests}
The author received no external funding for this research and declares no competing interests.
-\subsection*{Data and code availability}\label{computational-supplement}
+\subsection*{Data and code availability}\label{sec:code-availability}
No external datasets were used in this study. The reproducible Python scripts used to verify the analytical polynomial integrals and to generate Figures 1--4 are available in the accompanying code archive.
diff --git a/website/content/sources/selected-research/dependence/main-classification-flow.tex b/website/content/sources/selected-research/dependence/main-classification-flow.tex
index 5d10d0e..10cd64c 100644
--- a/website/content/sources/selected-research/dependence/main-classification-flow.tex
+++ b/website/content/sources/selected-research/dependence/main-classification-flow.tex
@@ -1,4 +1,4 @@
-\section{4. Efficiency premiums and projected flow duality}\label{background-substitutability-and-minimum-cost-feasibility}
+\section{4. Efficiency premiums and projected flow duality}\label{sec:efficiency-premiums}
The critical coefficient reaches the local-efficiency lower bound exactly when background displacement is cheapest per unit change in portfolio loss, and in every other case a strict marginal-restoration premium remains. This section proves that classification and then characterizes the right derivative of the critical budget in the background penalty at equal effective costs. A locally efficient displacement need not preserve the marginal projections, so any unbalanced moment has to be compensated elsewhere. We measure this obstruction by testing each projected flux against a bounded measurable coordinate potential, using the \(L^1\) dual representation of the restoration cost:
\[
@@ -8,13 +8,13 @@ \section{4. Efficiency premiums and projected flow duality}\label{background-sub
\]
Here \(\phi_i(q)\) acts as a dual shadow price for the projected displacement moment at coordinate level \(q\). For the projections to vanish, opposite signed displacements have to balance at each coordinate level, and then the leading restoration penalty disappears. We first identify the pointwise displacement required for equality in the efficiency bound, and then test its projections near an off-diagonal corner to rule out equality when only traded coordinates are cheapest. At equal effective costs, continuous potentials lead to the flow--cut duality of Section~4.2, and allowing nonzero projected fluxes in the restoration-cost bound leads to the piecewise-constant potential estimates of Section~5.1.
-\subsection{4.1 Local efficiency and the zero-premium classification}\label{a-necessary-and-sufficient-condition-for-zero-premium}\label{v21-classification}
+\subsection{4.1 Local efficiency and the zero-premium classification}\label{sec:zero-premium}
With \(c_1=\gamma_1\), \(c_2=\gamma_2\), \(c_3=\gamma_3/\beta\), and \(c_{\min}=\min\{c_1,c_2,c_3\}\), Proposition~\ref{prop:nominal-benchmark} and Theorem~\ref{thm:critical-limit} already give \(F(s)\le-J/p+s/(pc_{\min})\). The issue at zero risk difference is whether a kernel can satisfy the pointwise efficiency conditions while conserving every marginal projection, and we characterize those simultaneous equality conditions below.
Proposition~\ref{prop:nominal-benchmark} gives initial efficiency at each fixed trade size under \eqref{eq:2.9}, and when \(\gamma_1=c_{\min}\), Proposition~\ref{prop:balanced-linear-profile} separately realizes the limiting slope by matching opposite first-coordinate moments on each fiber. When only traded coordinates are cheapest, that matching is enough for the initial linear segment but cannot cover the whole tail section. If the background coordinate is cheapest, exchanging \(x,y\) and reversing \(u\) instead balances the required displacement on every background slice, which sustains the efficiency bound up to zero risk.
-\resultnumber{4.1}\begin{lemma}[Local-efficiency bound and critical-level optimality]\label{lem:efficiency-bound}\label{lem:v21-equality}\label{g.3-a-necessary-and-sufficient-condition-for-minimum-cost-feasibility}
+\resultnumber{4.1}\begin{lemma}[Local-efficiency bound and critical-level optimality]\label{lem:efficiency-bound}
In the exchangeable box model of Section~2.2, every kernel with \(I(\alpha)<\infty\) satisfies \(p\mathcal V(\alpha)+J\le\mathcal C(\alpha)/c_{\min}\), so \(s^*\ge c_{\min}J\). Set \(b=(1,1,\beta)\) and \(\zeta=-(u+H)1_{\{u(u+H)<0\}}\). Then \(s^*=c_{\min}J\) if and only if there is a measurable probability kernel satisfying
\begin{equation}
\begin{aligned}
@@ -92,7 +92,7 @@ \subsection{4.1 Local efficiency and the zero-premium classification}\label{a-ne
By Lemma~\ref{lem:efficiency-bound}, it suffices to allocate the prescribed displacement \(\zeta\) among cheapest coordinates while preserving every marginal projection. The background coordinate admits such a balance because exchanging the traded losses and reversing the normal coordinate leaves each background slice fixed and reverses the required displacement. For the traded coordinates, the corresponding balance must instead be found along each \(x\)- or \(y\)-fiber. Near an off-diagonal corner, the required moment is bounded below by a positive multiple of \(\delta^2\), whereas the opposite-side sources available at those coordinate levels have capacity only \(O(\delta^4)\). We prove the background construction first and then use this discrepancy to exclude even randomized splitting among the traded directions.
-\paragraph{Comparison kernels and the sufficient direction}\label{d.1-three-zero-risk-difference-kernels-scores-and-optimal-thresholds}\label{d.2-net-fluxes-total-costs-and-general-bounds}
+\paragraph{Comparison kernels and the sufficient direction}\label{par:comparison-kernels}
Assigning the prescribed displacement to one coordinate gives
\begin{equation}
\alpha^{(1)}=\delta_{\zeta e_1},\qquad
@@ -122,7 +122,7 @@ \subsection{4.1 Local efficiency and the zero-premium classification}\label{a-ne
\end{equation}
Thus the mixture attains the efficiency bound on the stated interval.
-\paragraph{The obstruction for arbitrary randomized kernels}\label{e.3-a-local-obstruction-to-zero-flux-in-the-two-asset-directions}
+\paragraph{The obstruction for arbitrary randomized kernels}\label{par:corner-obstruction}
For necessity, suppose \(c_3>c_{\min}\) and a kernel with \(\mathcal V\ge0\) attains cost \(c_{\min}J\). Lemma~\ref{lem:efficiency-bound} then requires \(d_3=0\), \(d_1+d_2=\zeta\), and both traded components to have the sign of \(\zeta\). Since the kernel has finite first moment, its averaged displacement fields \(m_i(x,y)=\int_{\mathbb R}\int d_i\,\alpha_{\xi,u}(dd)\,du\), \(i=1,2\), are integrable on the section and satisfy, by \eqref{eq:4.1},
\begin{equation}
\begin{aligned}
@@ -140,7 +140,7 @@ \subsection{4.1 Local efficiency and the zero-premium classification}\label{a-ne
\end{equation}
We test \eqref{eq:4.15}--\eqref{eq:4.16} near the upper-left off-diagonal corner of \(\Omega\), where \(x\) is small, \(y\) is large, and the positive moment \((y-x)^2/2\) is bounded away from zero. Compensating this moment at the same coordinate levels requires negative moment from \(\{x>y\}\), where the small value of \(x\) or large value of \(y\) confines available sources to a small triangle near the diagonal (Figure~\ref{fig:corner}).
-\leavevmode\phantomsection\label{lem:v21-corner-geometry}
+\leavevmode\phantomsection\label{par:corner-geometry}
The section base is the intersection of an open square with a diagonal strip, \(\Omega=(\ell,r)^2\cap\{q_-Jc_{\min}\) when \(c_3>c_{\min}\), and \eqref{eq:4.13} gives \eqref{eq:4.10}. \unskip\nobreak\hfill\proofbox
-% V21_CORNER_FIGURE
\input{figure-corner.tex}
-\subsection{4.2 Boundary sensitivity and the flow--cut representation}\label{required-background-share-and-the-boundary-response}\label{v21-flow}\label{variational-characterization-and-bounds-for-the-background-share}
+\subsection{4.2 Boundary sensitivity and the flow--cut representation}\label{sec:boundary-sensitivity}
At equal effective costs, every efficient kernel has total required displacement \(J\). At this critical level, the equality conditions and total-moment saturation in Lemma~\ref{lem:efficiency-bound} fix \(b\cdot d=\zeta\) pointwise almost everywhere, so normal integration and exchange symmetrization reduce the remaining minimum-background allocation problem to shares on the two-dimensional half-section. Raising only the background penalty selects, to first order, the kernels that use the least background displacement, and we can calculate that amount by assigning as much displacement as possible to the traded coordinates under their projected conservation law. On \(D=\Omega\cap\{xa_T\).
By \eqref{eq:A.45}--\eqref{eq:A.46} with \(R=1\), the reflection \(d^N=-2ue_1\) has moment \(I_N=4\gamma_1m(U+1/2)\) and hinge contribution \(N_c(\tau)=-c\Delta_E\) throughout \([-T,T]\). For any old kernel with \(I\le M\), denote its moment and hinge contribution on \(E\) by \(I_O\le M\) and \(O_c(\tau)\). Subtracting the common baseline \(k_O=O_0(0)\) leaves \(R_c(\tau)=O_c(\tau)-k_O\), whose integrand is bounded by \(a_T\) and can be nonzero only if \(|d|_\gamma\ge(U-a_T)/\kappa_B\). Thus the finite displacement moment gives the uniform estimate
diff --git a/website/content/sources/selected-research/dependence/main-crossing.tex b/website/content/sources/selected-research/dependence/main-crossing.tex
index 8bcfa49..b132bc7 100644
--- a/website/content/sources/selected-research/dependence/main-crossing.tex
+++ b/website/content/sources/selected-research/dependence/main-crossing.tex
@@ -1,5 +1,4 @@
\subsection{3.3 Initial linearity and the critical budget}\label{sec:balanced-profile}
-\phantomsection\label{g.4-balanced-source-selection-and-the-limiting-linear-segment}
At zero threshold shifts, the portfolios disagree on tail membership precisely on \(W=\{u(u+H)<0\}\). The displacement \(\zeta=-(u+H)1_W\) moves a source in this wedge to the limiting \(A_\varepsilon\)-boundary and attains the maximal paired hinge increment per unit of normal movement. When \(c_{\min}=\gamma_1\), we can assign this movement along \(e_1\) provided its downward and upward moments balance on each \(x\)-fiber. On active sources in \(W\), \(\zeta<0\) for \(y0\) for \(y>x\), with \(\int\zeta\,du=-H|H|/2\). Writing \(k(x,y)=f(x,y,(v-x-y)/\beta)1_{\{\beta\ell_31\).
-\emph{An analytical benchmark.} Consider independent uniform losses on \([0,2]^3\), with \(\beta=1\), \(p=1/2\), and \(\gamma=(1,1,2)\). Here \(g=3/8\) and \(J=5/32\), so Proposition~\ref{prop:nominal-benchmark} gives \(m_\varepsilon=(5/16)\varepsilon^2+o(\varepsilon^2)\). By Corollary~\ref{prop:uniform-variable-cost}, the profile is exactly \(F(s)=2s-5/16\) for \(0\le s\le1/96\), yet its zero satisfies \(13/64\le s^*\le29/96\). So even in this uniform model, where the initial slope is attained, the proved lower bound on reversal is already \(30\%\) above the local-efficiency prediction \(5/32\).
+\emph{An analytical benchmark.} Consider independent uniform losses on \([0,2]^3\), with \(\beta=1\), \(p=1/2\), and \(\gamma=(1,1,2)\). Here \(g=3/8\) and \(J=5/32\), so Proposition~\ref{prop:nominal-benchmark} gives \(m_\varepsilon=(5/16)\varepsilon^2+o(\varepsilon^2)\). By Corollary~\ref{cor:uniform-reference}, the profile is exactly \(F(s)=2s-5/16\) for \(0\le s\le1/96\), yet its zero satisfies \(13/64\le s^*\le29/96\). So even in this uniform model, where the initial slope is attained, the proved lower bound on reversal is already \(30\%\) above the local-efficiency prediction \(5/32\).
\emph{Sensitivity at equal effective costs.} Along the cost family \(\gamma_1=\gamma_2=\gamma\), \(c_3=\gamma(1+\eta)\) for \(\eta\ge0\), the baseline \(Jc_{\min}=\gamma J\) is constant, so \(s^*\) and the marginal-restoration premium have the same right derivative \(\gamma Ja_+\) at \(\eta=0\). Here \(a_+\in(0,1)\) is the least background share among efficient stresses, characterized in Theorem~\ref{thm:projected-flow} by continuous maximum-flow/minimum-cut duality; Section 5 shows how this share and the budget sensitivity \(\gamma Ja_+\) can move in opposite directions as the reference dependence changes.
diff --git a/website/content/sources/selected-research/dependence/main-model-benchmark.tex b/website/content/sources/selected-research/dependence/main-model-benchmark.tex
index d01f26b..45f274b 100644
--- a/website/content/sources/selected-research/dependence/main-model-benchmark.tex
+++ b/website/content/sources/selected-research/dependence/main-model-benchmark.tex
@@ -1,4 +1,4 @@
-\section{2. Model formulation and the local benchmark}\label{the-model-and-ranking-thresholds}
+\section{2. Model formulation and the local benchmark}\label{sec:model}
We first formulate the comparison for arbitrary integrable marginal laws, which are enough for a worst stressed model to exist. The critical-scale analysis will use a positive continuous density and exchangeability of the two traded positions, both introduced in Section~2.2.
@@ -86,16 +86,16 @@ \subsection{2.2 The exchangeable box model}\label{model:critical-box}
We work on a compact box with a positive continuous density because that lets us study this dependence mechanism with uniform control of marginal restoration. Positivity on the closed box gives both the interior mass that local exchanges need and lower bounds on the marginal densities, hence Lipschitz bounds for their inverse CDFs, and these bounds enter the quantile recovery in Section~3.2 and Appendix A.5. Section~5.3 takes up the additional difficulties posed by unbounded marginals.
-\paragraph{Loss units and effective stress costs}\label{m-units}
+\paragraph{Loss units and effective stress costs}\label{par:effective-costs}
A shift \(\Delta X_3\) changes the portfolio loss by \(\beta\Delta X_3\) at cost \(\gamma_3|\Delta X_3|=(\gamma_3/\beta)|\beta\Delta X_3|\). So \(c_1=\gamma_1\), \(c_2=\gamma_2\), and \(c_3=\gamma_3/\beta\) are the costs per unit change of \(B\) along the three coordinate axes, and the ratio \(c_3/\gamma_1\) compares a background displacement with a first-position displacement having the same effect on portfolio loss.
\paragraph{Canonical box normalization} Writing the original coordinates as \(X=T_{\rm box}(u)=(\ell+\kappa u_1,\ell+\kappa u_2,\ell_3+\kappa_3u_3)\), with \(\kappa=\Delta_I/2\) and \(\kappa_3=\Delta_3/2\), gives a normalized model on \([0,2]^3\) with \(\widehat\beta=\beta\kappa_3/\kappa\) and \(\widehat\gamma=(\kappa\gamma_1,\kappa\gamma_2,\kappa_3\gamma_3)\). By cash invariance and positive homogeneity of ES, the original paired risk is \(\kappa\) times its normalized counterpart, while corresponding transport plans have equal cost; hence the critical budgets agree and \(p\) is unchanged. A common rescaling of all transport weights rescales those budgets.
-\subsection{2.3 Nominal curvature and the exact initial slope}\label{nominal-improvement-initial-slope-and-the-protection-benchmark}
+\subsection{2.3 Nominal curvature and the exact initial slope}\label{sec:nominal-curvature}
Under the reference model of Section~2.2, let \(v\) be the upper-tail \(p\) threshold of \(B\). The marginal expected shortfall (MES) of a unit position \(X_i\), namely \(E_{P_0}[X_i\mid B>v]\), is the ES derivative with respect to its holding, and multiplying by that holding gives its Euler contribution \cite{scaillet2004,tasche2001}. The first two holdings in \(B\) both equal one, so the directional derivative along \(H=X_1-X_2\) is their MES difference: with \(T_H=E_{P_0}[H1_{\{B>v\}}]\), it equals \(p^{-1}T_H=E_{P_0}[X_1-X_2\mid B>v]=0\) by exchangeability. To find the gain from a nonzero trade we then have to differentiate the moving tail boundary, which leads to a conditional variance on the tail section.
-We record the classical section differentiation formula under our continuity assumptions in Lemma~\ref{lemma:general-variation}.
+We record the classical section differentiation formula under our continuity assumptions in Lemma~\ref{lem:general-variation}.
The tail boundary can be parameterized by
\[
@@ -153,9 +153,9 @@ \subsection{2.3 Nominal curvature and the exact initial slope}\label{nominal-imp
\end{resultparts}
\end{proposition}
\begin{proof}
-For part~(i), the coefficients of \(B+\eta H\) are positive whenever \(|\eta|<1\), and its support is \([2\ell+\beta\ell_3,2r+\beta r_3]\). Since every interior level meets \(\operatorname{int}S\), Lemma~\ref{lemma:general-variation} gives a positive continuous loss density and a unique threshold. The gradients of \(H\) and \(B+\eta H\) are linearly independent, so \(H\) is nonconstant on each such section. Its conditional variance is therefore strictly positive, and the same lemma implies strict convexity of \(\eta\mapsto ES_p^{P_0}(B+\eta H)\) on \((-1,1)\).
+For part~(i), the coefficients of \(B+\eta H\) are positive whenever \(|\eta|<1\), and its support is \([2\ell+\beta\ell_3,2r+\beta r_3]\). Since every interior level meets \(\operatorname{int}S\), Lemma~\ref{lem:general-variation} gives a positive continuous loss density and a unique threshold. The gradients of \(H\) and \(B+\eta H\) are linearly independent, so \(H\) is nonconstant on each such section. Its conditional variance is therefore strictly positive, and the same lemma implies strict convexity of \(\eta\mapsto ES_p^{P_0}(B+\eta H)\) on \((-1,1)\).
-At \(\eta=0\), exchanging the first two coordinates preserves both the reference tail and its boundary measure and reverses \(H\), giving \(T_H=m_1=0\). In the notation of Lemma~\ref{lemma:general-variation}, the remaining section moments at \((0,v)\) are \(m_0=g\) and \(m_2=2J\). Thus the threshold derivative vanishes and the optimized second derivative is \(2J/p\), which yield \eqref{eq:2.6} by Taylor expansion at zero. Strict convexity makes this stationary point the strict minimum, so \(m_\varepsilon>0\) for all \(0<\varepsilon<1\). For the unoptimized objective \(\Psi\), the Hessian at \((0,v)\) is \(\operatorname{diag}(2J/p,g/p)\). Expanding at the trade size and threshold displacement \((c\varepsilon,\varepsilon\tau)\) gives \eqref{eq:2.7}, uniformly on bounded \(\tau\)-intervals.
+At \(\eta=0\), exchanging the first two coordinates preserves both the reference tail and its boundary measure and reverses \(H\), giving \(T_H=m_1=0\). In the notation of Lemma~\ref{lem:general-variation}, the remaining section moments at \((0,v)\) are \(m_0=g\) and \(m_2=2J\). Thus the threshold derivative vanishes and the optimized second derivative is \(2J/p\), which yield \eqref{eq:2.6} by Taylor expansion at zero. Strict convexity makes this stationary point the strict minimum, so \(m_\varepsilon>0\) for all \(0<\varepsilon<1\). For the unoptimized objective \(\Psi\), the Hessian at \((0,v)\) is \(\operatorname{diag}(2J/p,g/p)\). Expanding at the trade size and threshold displacement \((c\varepsilon,\varepsilon\tau)\) gives \eqref{eq:2.7}, uniformly on bounded \(\tau\)-intervals.
\emph{The fixed-marginal Lipschitz bound.} We next establish the uniform Lipschitz bound over \(\Pi\). Given any minimizing \(u\) for \(B\) under \(Q\) and any minimizing \(t\) for \(A_\varepsilon\) under \(P\), let
\[
@@ -192,7 +192,7 @@ \subsection{2.3 Nominal curvature and the exact initial slope}\label{nominal-imp
\tag{2.12}\label{eq:sharp-budget-continuity}
\end{equation}
-\label{a.3-an-interior-intersection-and-attainment-of-the-initial-slope}
+\label{par:initial-slope-attainment}
\emph{Attaining the initial slope.} Under \eqref{eq:2.9}, we construct an attaining exchange near an interior intersection of the reference tail planes \(\{B=v\}\) and \(\{A_\varepsilon=t_\varepsilon\}\). To locate such an intersection, fix \(\varepsilon\in(0,1)\) and let \([m_F,M_F]\) be the nondegenerate range of \(A_\varepsilon\) on \(F_v=\{B=v\}\cap S\). Both losses increase along each nonconstant segment from the lower vertex \(x_{\min}=(\ell,\ell,\ell_3)\). Any point with \(B>v\) lies beyond an intersection of that segment with \(F_v\), so \(t_\varepsilon\le m_F\) would imply \(\{B>v\}\subset\{A_\varepsilon>t_\varepsilon\}\). Since \(A_\varepsilon\) is nonconstant on \(F_v\), the second tail would also contain an open set below \(B=v\), contradicting the equal tail probabilities and \(f>0\). Rays toward the upper corner similarly exclude \(t_\varepsilon\ge M_F\), which leaves \(t_\varepsilon\in(m_F,M_F)\). The relative-interior property of a linear map on a convex section gives an intersection \(c=(x_*,y_*,z_*)\in\operatorname{int}S\).
To exchange the first-coordinate marginal intervals of two small source cubes in opposite disagreement regions along \(e_1\), we first choose reference points with the same coordinate \(x_*\) and seek equal and opposite changes in the two losses. From the interior intersection \(c\), a displacement \((0,-\zeta_0,a\zeta_0)\) changes \(B\) by \((-1+\beta a)\zeta_0\) and \(A_\varepsilon\) by \((-(1-\varepsilon)+\beta a)\zeta_0\). Equating their sum to zero determines \(a=(1-\varepsilon/2)/\beta\). Thus the points
@@ -221,7 +221,7 @@ \subsection{2.3 Nominal curvature and the exact initial slope}\label{nominal-imp
If \(\gamma_2\ge\gamma_1\) and \(\gamma_3\ge\beta\gamma_1\), condition~\eqref{eq:2.9} holds for every \(\varepsilon\in(0,1)\), which suggests the extrapolated reversal budget \(r_{0,\varepsilon}=m_\varepsilon/K_\varepsilon\). The construction gives \(r_\varepsilon>0\) for each fixed trade size, but the disagreement regions narrow with \(\varepsilon\), so it leaves \(r_\varepsilon/\varepsilon^2\) uncontrolled. Indeed, the unscaled bound \(|D_\varepsilon(\rho)|\le\varepsilon\|H\|_\infty\) makes both iterated limits in \((\rho,\varepsilon)\) vanish even though \(K_\varepsilon\to1/(p\gamma_1)>0\), which is why Section~3 tracks the profile under the joint scaling \(\rho=s\varepsilon^2\).
-\paragraph{Reference symmetry and stressed dependence}\label{prop:exchangeable-protection}\label{rem:reference-symmetry}
+\paragraph{Reference symmetry and stressed dependence}\label{par:reference-symmetry}
Our stress class allows asymmetric dependence between the two traded positions. If every candidate law were exchangeable, the losses \(B-\varepsilon H\) and \(B+\varepsilon H\) would have the same distribution, and convexity of ES would give
\begin{equation}
ES_p^P(B)\le\tfrac12 ES_p^P(B-\varepsilon H)+\tfrac12 ES_p^P(B+\varepsilon H)
diff --git a/website/content/sources/selected-research/dependence/main-quantitative.tex b/website/content/sources/selected-research/dependence/main-quantitative.tex
index 169003c..c29019e 100644
--- a/website/content/sources/selected-research/dependence/main-quantitative.tex
+++ b/website/content/sources/selected-research/dependence/main-quantitative.tex
@@ -1,6 +1,6 @@
\keepproofstart{6\baselineskip}
-\section{5. Analytical illustrations and financial implications}\label{critical-cost-premia-background-dependence-and-finite-reallocations}\label{v21-quantitative}
-\subsection{5.1 Projection potentials and a uniform reference}\label{strict-premium-despite-equal-efficiency}
+\section{5. Analytical illustrations and financial implications}\label{sec:illustrations}
+\subsection{5.1 Projection potentials and a uniform reference}\label{sec:uniform-reference}
Assigning a bounded shadow price \(\phi_i\) to each marginal flux lets us bound the strict premium without solving the full risk optimization. For a prescribed normal displacement we first minimize the adjusted coordinate costs \(\gamma_i\pm\phi_i(\xi_i)\), and then we compare that minimum with the paired hinge gain.
@@ -27,7 +27,7 @@ \subsection{5.1 Projection potentials and a uniform reference}\label{strict-prem
\]
Their signs distinguish low-loss from high-loss coordinate halves. At zero threshold shift, a positive paired hinge gain requires normal movement \(w=b\cdot d>0\) when \(xy\). With amplitude \(\gamma_0/2\), the effective cost in this direction across the midpoint is \(k(\xi,H)=3\gamma_0/2\), and both directional costs on either same-side region are \(\gamma_0/2\). Choosing \(\lambda=3\gamma_0/2\) in \eqref{eq:5.1} therefore removes the cross-region penalty at zero threshold. The proof below keeps the correction for shifted minimizing thresholds, which makes the bound valid without prescribing either ES threshold.
-\resultnumber{5.2}\begin{corollary}[Uniform reference with costly background displacement]\label{prop:uniform-variable-cost}
+\resultnumber{5.2}\begin{corollary}[Uniform reference with costly background displacement]\label{cor:uniform-reference}
For this uniform reference,
\begin{equation}
\frac{13}{64}\le\frac{s^*}{\gamma_0}\le\frac{29}{96},
@@ -50,7 +50,7 @@ \subsection{5.1 Projection potentials and a uniform reference}\label{strict-prem
For \(|\tau|\le1\), the section integrations in \hyperref[app:uniform-bound-proof]{Appendix B.2} give the first line. For \(|\tau|\ge1\), the efficiency decomposition \eqref{eq:4.2} gives the second. Every kernel with \(\mathcal V\ge0\) therefore costs at least \(13\gamma_0/64\), which is \(30\%\) above \(\gamma_0J\).
\end{proof}
-\phantomsection\label{app:quantitative-benchmarks}
+\phantomsection\label{par:uniform-envelopes}
Setting \(\gamma_0=1\), we obtain analytical envelopes for \(F(s)\) on \([0,U]\), where \(U=29/96\). The balanced kernel has cost \(s_0=1/96\), zero projected flux, and risk \(-7/24\), and the first-coordinate comparison kernel has cost \(U\) and risk zero. Their convex mixtures have affine paired risk because zero minimizes both threshold objectives for each kernel, and affine cost because the balanced kernel has zero flux. Joining these mixtures to the exact initial linear segment realizes the lower envelope
\[
\underline F(s)=
@@ -72,7 +72,7 @@ \subsection{5.1 Projection potentials and a uniform reference}\label{strict-prem
The last equality holds because the branches inside the maximum meet only at \(s=5/8>U\). Figure~\ref{fig:budget} plots the analytical envelopes \(\underline F\le F\le\overline F\) on \([0,U]\). The guaranteed linear segment \([0,s_{\rm lin}]\) covers one fifteenth of the tangent prediction \(J=5/32\), whereas the critical coefficient \(s^*\) lies between \(13/10\) and \(29/15\) times \(J\). So the initial slope is attained on a positive interval, and the required reversal budget is still at least \(30\%\) higher than its tangent prediction.
\input{figure-budget.tex}
-\paragraph{Numerical approximation.}
+\paragraph{Numerical approximation}
The computational companion also discretizes a symmetric family of local
kernels, using pure-coordinate moves to the two hinge breakpoints and
retaining the absolute projected-flux penalty in \eqref{eq:3.1}.
@@ -82,21 +82,21 @@ \subsection{5.1 Projection potentials and a uniform reference}\label{strict-prem
The calculation reproduces the proved initial linear segment, and the paired
risk of each computed kernel, evaluated again at its own minimizing
thresholds, vanishes to rounding error. Because the family lies inside the
-admissible class, its zero is an upper estimate of \(s^*\), subject to
-discretization error. On this evidence \(s^*\) falls in the lower third of
+admissible class, its zero is an upper estimate of \(s^*\) at the stated grid
+resolution. On this evidence \(s^*\) falls in the lower third of
the interval in \eqref{eq:5.2}, and the reversal premium over the tangent
-prediction lies between the proved \(30\%\) and about \(49\%\). The upper
-figure is numerical and is not a proof; normal-window checks and solver
-residuals are recorded in the companion's numerical study.
+prediction lies between the proved \(30\%\) and a computed figure of about
+\(49\%\). The companion's numerical study records the normal-window checks
+and solver residuals behind that figure.
For finite trade sizes, the companion solves the discretized problem
globally on cubic midpoint lattices with up to 24 points per marginal, for
-trade sizes from \(\varepsilon=0.2\) down to \(\varepsilon=0.025\). These
-lattices do not resolve the quadratic scale as the trade size decreases, so
-the recorded values measure discretization sensitivity, and a convergence
-study of \eqref{eq:3.6} requires finer grids.
+trade sizes from \(\varepsilon=0.2\) down to \(\varepsilon=0.025\). As the
+trade size decreases, the quadratic scale falls below the lattice spacing, so
+the recorded values measure discretization sensitivity; a convergence study
+of \eqref{eq:3.6} refines the lattice with the trade size.
-\subsection{5.2 A dependence family with fixed marginal laws}\label{background-dependence-changes-substitutability-even-with-identical-marginals}
+\subsection{5.2 A dependence family with fixed marginal laws}\label{sec:dependence-family}
Section~5.1 used an independent uniform reference with background displacement twice as costly as either traded direction. We now move to the equal-effective-cost boundary to see how changing the reference dependence affects the required background share \(a_+\) and the budget derivative \(\gamma_0Ja_+\). Keeping uniform marginals, \(\beta=1\), and \(p=1/2\), we perturb the independent density by the product of the centered losses \(x+y-2\) and \(z-1\):
\[
diff --git a/website/content/sources/selected-research/dependence/main-scope.tex b/website/content/sources/selected-research/dependence/main-scope.tex
index 2b87f6d..ccdaef8 100644
--- a/website/content/sources/selected-research/dependence/main-scope.tex
+++ b/website/content/sources/selected-research/dependence/main-scope.tex
@@ -1,4 +1,4 @@
-\subsection{5.3 Implications for reverse stress testing and further questions}\label{conclusions}\label{v29-concluding-discussion}\label{stress-test-implications}
+\subsection{5.3 Implications for reverse stress testing and further questions}\label{sec:implications}
An exact initial sensitivity need not predict how much dependence stress is required to reverse a diversification trade. This bears on reverse sensitivity analysis \cite{pesenti2019,pesenti2022}, where the quantity of interest is the distance to a decision reversal. Under the traded-coordinate dominance condition \eqref{eq:2.9}, \(r_{0,\varepsilon}=m_\varepsilon/K_\varepsilon\) extrapolates an attained initial slope, yet if \(\gamma_1=c_{\min}1.
diff --git a/website/content/sources/selected-research/dependence/main-variational-proof.tex b/website/content/sources/selected-research/dependence/main-variational-proof.tex
index c588e8c..4214352 100644
--- a/website/content/sources/selected-research/dependence/main-variational-proof.tex
+++ b/website/content/sources/selected-research/dependence/main-variational-proof.tex
@@ -1,6 +1,6 @@
\paragraph{Proof strategy} Theorem~\ref{thm:critical-limit}(i) gives \(\Gamma\)-convergence of the transport energy and convergence of the paired risk along bounded-energy plans, and the exact marginal recovery in part~(ii) makes the limiting kernel values asymptotically attainable in \(\Pi\). Part~(iii) gives uniform convergence of the optimized profiles on compact budget intervals, which, together with the strict crossing in Theorem~\ref{thm:strict-crossing}, locates the reversal budget \(s^*\).
-\emph{Step 1: Localization and the energy lower bound.} Because the two ES thresholds are optimized independently, comparison of the objectives requires a common threshold window for stresses with bounded rescaled energy. In Appendix~\hyperref[proof:threshold-localization]{A.2}, Lemma~\ref{b-l2-all-optimal-thresholds-under-a-budget-of-order-varepsilon2} places every minimizing threshold in an \(O_M(\varepsilon)\) window. On bounded windows the lifted plans are locally compact; outside them, marginal conservation requires compensation for the projected local flux. The compensation bound above then gives the energy lower bound for arbitrary admissible plans. Choosing optimal plans yields the corresponding Wasserstein bound for their target laws.
+\emph{Step 1: Localization and the energy lower bound.} Because the two ES thresholds are optimized independently, comparison of the objectives requires a common threshold window for stresses with bounded rescaled energy. In Appendix~\hyperref[proof:threshold-localization]{A.2}, Lemma~\ref{lem:threshold-localization} places every minimizing threshold in an \(O_M(\varepsilon)\) window. On bounded windows the lifted plans are locally compact; outside them, marginal conservation requires compensation for the projected local flux. The compensation bound above then gives the energy lower bound for arbitrary admissible plans. Choosing optimal plans yields the corresponding Wasserstein bound for their target laws.
\emph{Step 2: Convergence of the paired risk.} Bounded first moments make each hinge increment integrable, but need not give uniform integrability along convergent plans. The two objectives share the same large-displacement contribution, so we subtract \(\mathfrak q_0(0)\) from both. The relative increments are bounded on the common threshold window. Outside a normal window enlarged to account for the bounded reallocation \(H\), a nonzero relative increment requires a sufficiently large displacement, so the first moment controls the far-field contributions. The baseline integral is finite but need not converge. Since it translates both threshold objectives by the same amount independently of the threshold, it cancels exactly from their difference of minima.
diff --git a/website/content/sources/selected-research/dependence/main-variational.tex b/website/content/sources/selected-research/dependence/main-variational.tex
index d828d74..9a96093 100644
--- a/website/content/sources/selected-research/dependence/main-variational.tex
+++ b/website/content/sources/selected-research/dependence/main-variational.tex
@@ -1,4 +1,4 @@
-\section{3. Variational limit and the critical profile}\label{critical-cost-and-recovery-under-full-marginal-constraints}
+\section{3. Variational limit and the critical profile}\label{sec:variational-limit}
Under the traded-coordinate condition of Proposition~\ref{prop:nominal-benchmark}, a local exchange attains the initial slope. To find out how far that efficiency persists toward reversal, we now hold \(s=\rho/\varepsilon^2\) fixed. The limiting objects introduced below connect the stress profile to its reversal coefficient:
\[
diff --git a/website/content/sources/selected-research/dependence/paper.tex b/website/content/sources/selected-research/dependence/paper.tex
index e9efa28..72d5cf9 100644
--- a/website/content/sources/selected-research/dependence/paper.tex
+++ b/website/content/sources/selected-research/dependence/paper.tex
@@ -1,4 +1,3 @@
-% Public reading edition. Mathematical body retained from the author's manuscript.
\documentclass[11pt,a4paper]{article}
\usepackage[margin=1in,headheight=14pt,footskip=30pt]{geometry}
\usepackage{fontspec}
@@ -11,9 +10,9 @@
bookmarksopen=false,bookmarksnumbered=false]{hyperref}
\urlstyle{same}
\newcommand{\proofbox}{\ensuremath{\square}}
-\hypersetup{pdfauthor={Letao Ouyang},
+\hypersetup{pdfauthor={Letao (Theodore) Ouyang},
pdftitle={Critical Dependence Stress for Diversification Gains under Fixed Marginals},
- pdfsubject={Author manuscript - public reading edition},
+ pdfsubject={Author manuscript},
pdfcreator={XeLaTeX}}
\makeatletter
\renewtheoremstyle{plain}
@@ -64,14 +63,9 @@
{\LARGE Critical Dependence Stress for\\[5pt]
Diversification Gains under Fixed Marginals\par}
\vspace{18pt}
-{\large Letao Ouyang\par}
-\vspace{6pt}
-{\small Independent Researcher, Nashua, NH 03062, USA\par}
-\vspace{5pt}
-{\small \href{mailto:theodore.oy2025@gmail.com}{\textcolor{black}{theodore.oy2025@gmail.com}}
-\quad \href{mailto:10@alumni.duke.edu}{\textcolor{black}{10@alumni.duke.edu}}\par}
-\vspace{10pt}
-{\small Author manuscript\par}
+{\large Letao (Theodore) Ouyang{\hypersetup{linkcolor=black}\footnote{\raggedright Independent Researcher, Nashua, NH 03062, USA. \href{mailto:theodore.oy2025@gmail.com}{\textcolor{black}{theodore.oy2025@gmail.com}} \textperiodcentered\ \href{mailto:10@alumni.duke.edu}{\textcolor{black}{10@alumni.duke.edu}}. Code, data and verification records: \href{https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals}{\textcolor{black}{\mbox{\nolinkurl{https://github.com/130U/}}\allowbreak\nolinkurl{RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals}}}.}}\par}
+\vspace{8pt}
+{\normalsize November 2026\par}
\end{center}
\vspace{8pt}
\begin{abstract}
diff --git a/website/content/sources/selected-research/dependence/provenance.json b/website/content/sources/selected-research/dependence/provenance.json
index cacf45f..80e13b1 100644
--- a/website/content/sources/selected-research/dependence/provenance.json
+++ b/website/content/sources/selected-research/dependence/provenance.json
@@ -1,18 +1,18 @@
{
"repositoryUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals",
- "sourceUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/paper.pdf",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/paper.pdf",
"sourceTitle": "Critical Dependence Stress for Diversification Gains under Fixed Marginals",
- "sourceCommit": "a62cc41106e58e5d2df1387a1cf39ebfa65e36e8",
- "sourceBlob": "e9efa283fc4b8efe2cac0ca483fa3e5da81e01e5",
- "sourceEditedAt": "2026-11-19T03:29:40-05:00",
+ "sourceCommit": "c0782609a95f9887c248631c56174615233c7ae4",
+ "sourceBlob": "72d5cf95d6812da8198cca12f313c59d254ef3d0",
+ "sourceEditedAt": "2026-11-19T11:43:47-05:00",
"fetchedWith": "Authenticated Git checkout at the exact commit; git show bytes checked against Git blob SHA-1",
"sourceFiles": [
{
"path": "paper/source/declarations.tex",
"local": "declarations.tex",
- "gitBlob": "89f601719f90503b89a8a062ffe8f83a36f16ccb",
- "sha256": "9d9dea327e9a340b93ea9e470a2a52238c0e6dd92b39fbc7c59cf8033deb790f",
- "bytes": 932
+ "gitBlob": "af7fa007c564ae03964b4ff6794298fde311774b",
+ "sha256": "8c7cd1e1ebbd5e23f977cbc877c07b38b8c387ccba446e2559ca56d5f739a552",
+ "bytes": 929
},
{
"path": "paper/source/figure-budget.tex",
@@ -38,72 +38,72 @@
{
"path": "paper/source/main-classification-flow.tex",
"local": "main-classification-flow.tex",
- "gitBlob": "5d10d0e4f87b084bc57560693156903f7e8eb48d",
- "sha256": "79be7f50ffeb003bfed2249b00950b5e23e7056dd814cd0c1a6d8f71a484955d",
- "bytes": 20134
+ "gitBlob": "10cd64c31c40954ccfb78c304a026574722f7c21",
+ "sha256": "61374c8ae5dc4f3801d8c63c573f2d2fca70e9fe184a3a8b558ec526d9db125f",
+ "bytes": 19601
},
{
"path": "paper/source/main-crossing-proof.tex",
"local": "main-crossing-proof.tex",
- "gitBlob": "394d003c1bdea0f52c5e4ea99559c7fce43d83e8",
- "sha256": "9b13fb1ddb3b081ef5ffc7c7df9a64fc0f36b39468367cc14f28aaee86dfa3d4",
- "bytes": 4308
+ "gitBlob": "c45aaa4c98e0af2f8ccb004bcd6e5a884b2d07a2",
+ "sha256": "31b44ccdf5ee40aa6637608a482eb3a3f46a703f36399dd6cc8970e870cfae75",
+ "bytes": 4201
},
{
"path": "paper/source/main-crossing.tex",
"local": "main-crossing.tex",
- "gitBlob": "8bcfa49229edb4320a2fbc077d74906267a42d07",
- "sha256": "b40839ba26652ff83fe52e5d2c1b660c2f574e4177e7eebb41580c7d1607d13a",
- "bytes": 8585
+ "gitBlob": "b132bc78e96f5be64b8963ea08e59e78715d783a",
+ "sha256": "e5b8bb86b3eb6bc96ec6d655cffb4c63fe51b1db14354972e1ec87470183e8fd",
+ "bytes": 8510
},
{
"path": "paper/source/main-introduction.tex",
"local": "main-introduction.tex",
- "gitBlob": "637cf0d47fd21be707f7f7821f35c72d281191db",
- "sha256": "6de76a53ce9cb48c99df3b44649d08b3c9f08701295b5ae9814c3d603d29f519",
- "bytes": 12283
+ "gitBlob": "fcea6eafa21b135bcdbe060cb2499ae60c2eda9f",
+ "sha256": "ad8a1733e9a4eeef931a4c0a345884936763f421b325d869c240ee7783fd27bd",
+ "bytes": 12229
},
{
"path": "paper/source/main-model-benchmark.tex",
"local": "main-model-benchmark.tex",
- "gitBlob": "d01f26b9fc7f72d36aaaf92b3642046643c64e2e",
- "sha256": "b3eb1da3b723f6fbe19b805a1532b140fe0a7bb1d26feb988e4267ea3e151716",
- "bytes": 23212
+ "gitBlob": "45f274bad2930d2a417c8717ab2d44db4f2c12b3",
+ "sha256": "3cdd22399e671b53fd5f10f5ee0d097a68e07a1f219c07f71e47611af7a09a60",
+ "bytes": 23082
},
{
"path": "paper/source/main-quantitative.tex",
"local": "main-quantitative.tex",
- "gitBlob": "169003cf20576b49ff370332fe15e428844e4687",
- "sha256": "86544582aeb27420b676364656e8cd470dbfbfeb854182c7cf9b79c2c51b5140",
- "bytes": 13038
+ "gitBlob": "c29019e57b7308207ddcdc01d9dd8a4308a2de38",
+ "sha256": "0fa7cf11258e4593070cf3742dd9bc404730c9a07c5cbddd48fa4f3282fa7321",
+ "bytes": 12880
},
{
"path": "paper/source/main-scope.tex",
"local": "main-scope.tex",
- "gitBlob": "2b87f6db400a97cd97ec3ac0e51a4d76e3de66fd",
- "sha256": "7d86109df3511118ab08215d74cf4e300db01164367c3596c3bf0efd76ccabc1",
- "bytes": 4014
+ "gitBlob": "ccdaef8cda3df6c3d5c4217e4c9fc19381770a15",
+ "sha256": "b857c903798bf0e9c6cf24f68e533eaa00ae7632b8513a283b73041eae486b40",
+ "bytes": 3954
},
{
"path": "paper/source/main-variational-proof.tex",
"local": "main-variational-proof.tex",
- "gitBlob": "c588e8c6021270004c6a84046c35c4a66bfd9ae1",
- "sha256": "1e347e0efeb8b5614737d4f9e74e35e6c4ca2d5d06eb4d580a1f0558245d2966",
- "bytes": 3676
+ "gitBlob": "4214352cc0a9f158ddc87ec37b4fd3126abe02fd",
+ "sha256": "5931642da24a3984b289baabb3fd0c19faeb81ecf946c17991f38cc0e30e39d3",
+ "bytes": 3639
},
{
"path": "paper/source/main-variational.tex",
"local": "main-variational.tex",
- "gitBlob": "d828d74fb7dac3b40780bb300927765307b23faa",
- "sha256": "bb099b84d0c395f3a94fc50d75ffeebc850211d025e3104ff5c070b0d8d45fff",
- "bytes": 9246
+ "gitBlob": "9a9609383ae8bb36e4b77150fbdfc346540511de",
+ "sha256": "d3a32bd24a49173b292064e79a7ebd6b64ded4461140c28c7f9d99d1dff91d1f",
+ "bytes": 9209
},
{
"path": "paper/source/paper.tex",
"local": "paper.tex",
- "gitBlob": "e9efa283fc4b8efe2cac0ca483fa3e5da81e01e5",
- "sha256": "18176ac22f5e98d4e13b30d054eb6369639e5d86bb7f7f52abc3946777533aa5",
- "bytes": 5755
+ "gitBlob": "72d5cf95d6812da8198cca12f313c59d254ef3d0",
+ "sha256": "047d09bab900538f6f7dddc1680a355aa174b5cdb54456afca76b311d565e0c4",
+ "bytes": 6004
},
{
"path": "paper/source/references.tex",
@@ -115,9 +115,9 @@
{
"path": "paper/source/technical-cost-proof.tex",
"local": "technical-cost-proof.tex",
- "gitBlob": "971784fbb37c00ffbfd8244d5f4c941cb33db2a3",
- "sha256": "6cd543bc80cc0705c214a64463ecfd00cac00c647e36de0399d19957a37c4b30",
- "bytes": 2773
+ "gitBlob": "ea715352f8899c7bc66d3cc9ff605c801064775a",
+ "sha256": "02d33b0e6e12834d7cdcaa40ea90f18d86c50b3921ab119dab26c7a24265e05c",
+ "bytes": 2740
},
{
"path": "paper/source/technical-efficiency-identity.tex",
@@ -129,51 +129,51 @@
{
"path": "paper/source/technical-flow-proof.tex",
"local": "technical-flow-proof.tex",
- "gitBlob": "734c3b233033bbedf3015185685a227a14edb945",
- "sha256": "a8d5111e9db4c4152a47f693e8e24830af267f4b3083108e749e727a73a52987",
- "bytes": 7787
+ "gitBlob": "e95bd18d25be9329ec063c12d1e5e46befe9362d",
+ "sha256": "6a36b874257df4cdff79cd98be329dd5945ecb7b6b85e85c6f465412b2a725e8",
+ "bytes": 7728
},
{
"path": "paper/source/technical-limit.tex",
"local": "technical-limit.tex",
- "gitBlob": "fb04ad1228419c5affc10a3d266734345d3f2bc3",
- "sha256": "cc827adbab6dc383d0e85d9840da0fab9c7e11daf64ee045b74424229b09643a",
- "bytes": 12423
+ "gitBlob": "2d6b7883cb179a58f9d0bed5c664624188ce0e99",
+ "sha256": "63fccd3e3fdbb894d4ed60bbe77525b131089aa6d6c115a07e509097ea9af956",
+ "bytes": 12160
},
{
"path": "paper/source/technical-mixture-obstruction.tex",
"local": "technical-mixture-obstruction.tex",
- "gitBlob": "7172e4ab665902cd3ab060a47ffea3ed4954f818",
- "sha256": "1ad7643a5aa542ba8492b6dbf02da666b224612116703387e3b9b2712cb8db3c",
- "bytes": 1416
+ "gitBlob": "0b82578a62446d1c75fe7176c156dceee12bdff3",
+ "sha256": "ca664dedc995a0417c2d6ddd1384704d2ecf3a73e22b15ed79cb1b17f1886492",
+ "bytes": 1420
},
{
"path": "paper/source/technical-quantitative.tex",
"local": "technical-quantitative.tex",
- "gitBlob": "9c1c919e4bb9709defe13fb8873d27cb77618991",
- "sha256": "5ad429e9c4ed97a2453900d3c8fd9cdb620fc65853b8dcbafa95f66ef66026ba",
- "bytes": 6520
+ "gitBlob": "2008838b2847da91fe1bf496147cacf3b91644ed",
+ "sha256": "2b5f129af4df6424e62ff7abed959d107bfefab1cf1197020cdaaff2391fc32f",
+ "bytes": 6494
},
{
"path": "paper/source/technical-reference.tex",
"local": "technical-reference.tex",
- "gitBlob": "44c7620f618c98ed96da2343ed6bba1c8126afcd",
- "sha256": "8c63a00a731b208d87a08340569d4017c09895e1d6f457e74d1c6c321cd69785",
- "bytes": 3244
+ "gitBlob": "0ee665753aee525c8b898fcf4d0ff4f5955c84d8",
+ "sha256": "24d84f944ef1225492b90850691033cf7d6eb944aa49ee6dff077e33e74f7eec",
+ "bytes": 3207
},
{
"path": "paper/source/technical-threshold-statement.tex",
"local": "technical-threshold-statement.tex",
- "gitBlob": "4955c4f0e239365098d4c2c71d53ff6c56fad2ed",
- "sha256": "1c16a52b68e35157c650be7e17e65654afc73283031df94677dd3ea559bd0c85",
- "bytes": 474
+ "gitBlob": "5e71c0fd445b34ce0ae2a17515b142dc7ad1f382",
+ "sha256": "cb6f203a7fc9b71ff3038e35eb6304126adc9707b170d339dcc7ad1cddc6c717",
+ "bytes": 437
},
{
"path": "paper/source/technical-variational-proof.tex",
"local": "technical-variational-proof.tex",
- "gitBlob": "f60f6fa47576f16977f548a3b95bffb8c32a1286",
- "sha256": "69b4d9df40d226ec6d71915653422bc9e2eba4e920fef7b57f4e7981e5406273",
- "bytes": 18761
+ "gitBlob": "8b4987f3cb404513f2a0fdbc29850412ef224aa9",
+ "sha256": "d2d627e542ffa793f9debe9bf5e7f9edf24843d6eac0666770cdf9830ea6e314",
+ "bytes": 18024
}
],
"includedFiles": [
@@ -347,7 +347,7 @@
"crossReferences": [
{
"label": "eq:B.2",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L80",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L80",
"number": "B.2"
},
{
@@ -386,13 +386,13 @@
"number": "2.2"
},
{
- "label": "prop:uniform-variable-cost",
- "href": "#dep-prop-uniform-variable-cost",
+ "label": "cor:uniform-reference",
+ "href": "#dep-cor-uniform-reference",
"number": "5.2"
},
{
"label": "thm:projected-flow",
- "href": "#dep-prop-boundary-sensitivity",
+ "href": "#dep-thm-projected-flow",
"number": "4.3"
},
{
@@ -421,18 +421,18 @@
"number": "2.3"
},
{
- "label": "lemma:general-variation",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4",
+ "label": "lem:general-variation",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4",
"number": "A.1"
},
{
- "label": "lemma:general-variation",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4",
+ "label": "lem:general-variation",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4",
"number": "A.1"
},
{
- "label": "lemma:general-variation",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-reference.tex#L4",
+ "label": "lem:general-variation",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-reference.tex#L4",
"number": "A.1"
},
{
@@ -502,7 +502,7 @@
},
{
"label": "lem:threshold-convexity",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101",
"number": "A.4"
},
{
@@ -532,17 +532,17 @@
},
{
"label": "proof:threshold-localization",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L2",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L1",
"number": "A.2"
},
{
- "label": "b-l2-all-optimal-thresholds-under-a-budget-of-order-varepsilon2",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-threshold-statement.tex#L1",
+ "label": "lem:threshold-localization",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-threshold-statement.tex#L1",
"number": "A.2"
},
{
"label": "eq:A.13",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L136",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L134",
"number": "A.13"
},
{
@@ -552,12 +552,12 @@
},
{
"label": "lem:threshold-convexity",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101",
"number": "A.4"
},
{
"label": "lem:projected-imbalance",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L54",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L52",
"number": "A.3"
},
{
@@ -617,7 +617,7 @@
},
{
"label": "lem:threshold-convexity",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-limit.tex#L103",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-limit.tex#L101",
"number": "A.4"
},
{
@@ -747,12 +747,12 @@
},
{
"label": "app:flow-proof",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-flow-proof.tex#L1",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-flow-proof.tex#L1",
"number": "A.8"
},
{
- "label": "app:cost-proofs",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-cost-proof.tex#L1",
+ "label": "app:boundary-sensitivity",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-cost-proof.tex#L1",
"number": "A.7"
},
{
@@ -772,12 +772,12 @@
},
{
"label": "app:potential-proof",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L3",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L3",
"number": "B.1"
},
{
"label": "thm:projected-flow",
- "href": "#dep-prop-boundary-sensitivity",
+ "href": "#dep-thm-projected-flow",
"number": "4.3"
},
{
@@ -802,7 +802,7 @@
},
{
"label": "app:uniform-bound-proof",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L32",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L32",
"number": "B.2"
},
{
@@ -832,7 +832,7 @@
},
{
"label": "app:dependence-family-proof",
- "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/a62cc41106e58e5d2df1387a1cf39ebfa65e36e8/paper/source/technical-quantitative.tex#L49",
+ "href": "https://github.com/130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals/blob/c0782609a95f9887c248631c56174615233c7ae4/paper/source/technical-quantitative.tex#L49",
"number": "B.3"
},
{
@@ -862,7 +862,7 @@
},
{
"label": "thm:projected-flow",
- "href": "#dep-prop-boundary-sensitivity",
+ "href": "#dep-thm-projected-flow",
"number": "4.3"
},
{
@@ -6844,14 +6844,13 @@
"labels": []
}
],
- "importedMarkdownSha256": "a38068886bffca76b7bc32cd070c324e22d0cd3b15634824990af864de316404",
+ "importedMarkdownSha256": "1da898ae2988ae838cbb9120cc7b38a3b47cf082ea681b627eb106665d5fe772",
"transformations": [
"Seven main TeX input files contain five numbered sections and eleven subsections; recursively included main-text proofs are retained.",
"TeX line wrapping, layout commands, font styling, and text accents are converted to equivalent web markup.",
"Every inline/display mathematical body is preserved, apart from moving TeX label commands to HTML anchors; explicit equation tags are retained.",
"Original theorem numbers, proof endings, result-part numbering, bibliography order, citation targets, and complete figure captions are retained.",
"Aliases on one original target share one web anchor; references to omitted appendices point to their exact pinned TeX source line.",
- "The four original PNG siblings of the PDF figures are referenced at local asset paths, with their original Git blob identities recorded.",
- "The public page title differs from the original paper title, which is recorded separately as sourceTitle."
+ "The four figures are drawn natively from the repository's formulas; their image paths key the native drawings and the complete captions are retained."
]
}
diff --git a/website/content/sources/selected-research/dependence/source-manifest.json b/website/content/sources/selected-research/dependence/source-manifest.json
index d010b77..a412e4d 100644
--- a/website/content/sources/selected-research/dependence/source-manifest.json
+++ b/website/content/sources/selected-research/dependence/source-manifest.json
@@ -1,13 +1,13 @@
{
"repository": "130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals",
- "commit": "a62cc41106e58e5d2df1387a1cf39ebfa65e36e8",
- "commitDate": "2026-11-19T03:29:40-05:00",
+ "commit": "c0782609a95f9887c248631c56174615233c7ae4",
+ "commitDate": "2026-11-19T11:43:47-05:00",
"fetchedWith": "Authenticated Git checkout at the exact commit; git show bytes checked against Git blob SHA-1",
"files": [
{
"path": "paper/source/declarations.tex",
- "sha": "89f601719f90503b89a8a062ffe8f83a36f16ccb",
- "size": 932
+ "sha": "af7fa007c564ae03964b4ff6794298fde311774b",
+ "size": 929
},
{
"path": "paper/source/figure-budget.tex",
@@ -26,53 +26,53 @@
},
{
"path": "paper/source/main-classification-flow.tex",
- "sha": "5d10d0e4f87b084bc57560693156903f7e8eb48d",
- "size": 20134
+ "sha": "10cd64c31c40954ccfb78c304a026574722f7c21",
+ "size": 19601
},
{
"path": "paper/source/main-crossing-proof.tex",
- "sha": "394d003c1bdea0f52c5e4ea99559c7fce43d83e8",
- "size": 4308
+ "sha": "c45aaa4c98e0af2f8ccb004bcd6e5a884b2d07a2",
+ "size": 4201
},
{
"path": "paper/source/main-crossing.tex",
- "sha": "8bcfa49229edb4320a2fbc077d74906267a42d07",
- "size": 8585
+ "sha": "b132bc78e96f5be64b8963ea08e59e78715d783a",
+ "size": 8510
},
{
"path": "paper/source/main-introduction.tex",
- "sha": "637cf0d47fd21be707f7f7821f35c72d281191db",
- "size": 12283
+ "sha": "fcea6eafa21b135bcdbe060cb2499ae60c2eda9f",
+ "size": 12229
},
{
"path": "paper/source/main-model-benchmark.tex",
- "sha": "d01f26b9fc7f72d36aaaf92b3642046643c64e2e",
- "size": 23212
+ "sha": "45f274bad2930d2a417c8717ab2d44db4f2c12b3",
+ "size": 23082
},
{
"path": "paper/source/main-quantitative.tex",
- "sha": "169003cf20576b49ff370332fe15e428844e4687",
- "size": 13038
+ "sha": "c29019e57b7308207ddcdc01d9dd8a4308a2de38",
+ "size": 12880
},
{
"path": "paper/source/main-scope.tex",
- "sha": "2b87f6db400a97cd97ec3ac0e51a4d76e3de66fd",
- "size": 4014
+ "sha": "ccdaef8cda3df6c3d5c4217e4c9fc19381770a15",
+ "size": 3954
},
{
"path": "paper/source/main-variational-proof.tex",
- "sha": "c588e8c6021270004c6a84046c35c4a66bfd9ae1",
- "size": 3676
+ "sha": "4214352cc0a9f158ddc87ec37b4fd3126abe02fd",
+ "size": 3639
},
{
"path": "paper/source/main-variational.tex",
- "sha": "d828d74fb7dac3b40780bb300927765307b23faa",
- "size": 9246
+ "sha": "9a9609383ae8bb36e4b77150fbdfc346540511de",
+ "size": 9209
},
{
"path": "paper/source/paper.tex",
- "sha": "e9efa283fc4b8efe2cac0ca483fa3e5da81e01e5",
- "size": 5755
+ "sha": "72d5cf95d6812da8198cca12f313c59d254ef3d0",
+ "size": 6004
},
{
"path": "paper/source/references.tex",
@@ -81,8 +81,8 @@
},
{
"path": "paper/source/technical-cost-proof.tex",
- "sha": "971784fbb37c00ffbfd8244d5f4c941cb33db2a3",
- "size": 2773
+ "sha": "ea715352f8899c7bc66d3cc9ff605c801064775a",
+ "size": 2740
},
{
"path": "paper/source/technical-efficiency-identity.tex",
@@ -91,38 +91,38 @@
},
{
"path": "paper/source/technical-flow-proof.tex",
- "sha": "734c3b233033bbedf3015185685a227a14edb945",
- "size": 7787
+ "sha": "e95bd18d25be9329ec063c12d1e5e46befe9362d",
+ "size": 7728
},
{
"path": "paper/source/technical-limit.tex",
- "sha": "fb04ad1228419c5affc10a3d266734345d3f2bc3",
- "size": 12423
+ "sha": "2d6b7883cb179a58f9d0bed5c664624188ce0e99",
+ "size": 12160
},
{
"path": "paper/source/technical-mixture-obstruction.tex",
- "sha": "7172e4ab665902cd3ab060a47ffea3ed4954f818",
- "size": 1416
+ "sha": "0b82578a62446d1c75fe7176c156dceee12bdff3",
+ "size": 1420
},
{
"path": "paper/source/technical-quantitative.tex",
- "sha": "9c1c919e4bb9709defe13fb8873d27cb77618991",
- "size": 6520
+ "sha": "2008838b2847da91fe1bf496147cacf3b91644ed",
+ "size": 6494
},
{
"path": "paper/source/technical-reference.tex",
- "sha": "44c7620f618c98ed96da2343ed6bba1c8126afcd",
- "size": 3244
+ "sha": "0ee665753aee525c8b898fcf4d0ff4f5955c84d8",
+ "size": 3207
},
{
"path": "paper/source/technical-threshold-statement.tex",
- "sha": "4955c4f0e239365098d4c2c71d53ff6c56fad2ed",
- "size": 474
+ "sha": "5e71c0fd445b34ce0ae2a17515b142dc7ad1f382",
+ "size": 437
},
{
"path": "paper/source/technical-variational-proof.tex",
- "sha": "f60f6fa47576f16977f548a3b95bffb8c32a1286",
- "size": 18761
+ "sha": "8b4987f3cb404513f2a0fdbc29850412ef224aa9",
+ "size": 18024
}
]
}
diff --git a/website/content/sources/selected-research/dependence/technical-cost-proof.tex b/website/content/sources/selected-research/dependence/technical-cost-proof.tex
index 971784f..ea71535 100644
--- a/website/content/sources/selected-research/dependence/technical-cost-proof.tex
+++ b/website/content/sources/selected-research/dependence/technical-cost-proof.tex
@@ -1,7 +1,7 @@
-\subsection{A.7 Boundary sensitivity}\label{app:cost-proofs}
+\subsection{A.7 Boundary sensitivity}\label{app:boundary-sensitivity}
-\leavevmode\phantomsection\label{l2.-the-baseline-minimizing-set-and-the-right-difference-quotient}
-\begin{proof}[Proof of the boundary expansion in Theorem~\ref{prop:boundary-sensitivity}] We first justify minimization over the boundary equality kernels. On the measures \(\Pi_\alpha\), the coordinate functional \(A_i(\alpha)=\int|d_i|d\Pi_\alpha+\|j_i\|_1\) is lower semicontinuous for vague convergence by \mbox{\hyperref[v21-limit-proof]{Appendix~A.5}}. On \(\{\mathcal V=0\}\), the efficiency inequality gives \(\mathcal C_0\ge\gamma J\), so \(\mathcal C_0=\gamma J\) is equivalent to \(\mathcal C_0\le\gamma J\). Hence
+\leavevmode\phantomsection\label{proof:boundary-expansion}
+\begin{proof}[Proof of the boundary expansion in Theorem~\ref{thm:projected-flow}] We first justify minimization over the boundary equality kernels. On the measures \(\Pi_\alpha\), the coordinate functional \(A_i(\alpha)=\int|d_i|d\Pi_\alpha+\|j_i\|_1\) is lower semicontinuous for vague convergence by \mbox{\hyperref[app:variational-proof]{Appendix~A.5}}. On \(\{\mathcal V=0\}\), the efficiency inequality gives \(\mathcal C_0\ge\gamma J\), so \(\mathcal C_0=\gamma J\) is equivalent to \(\mathcal C_0\le\gamma J\). Hence
\[
\mathcal Z_0=\{\alpha:\mathcal V(\alpha)=0,\ \mathcal C_0(\alpha)\le\gamma J\}.
\]
diff --git a/website/content/sources/selected-research/dependence/technical-flow-proof.tex b/website/content/sources/selected-research/dependence/technical-flow-proof.tex
index 734c3b2..e95bd18 100644
--- a/website/content/sources/selected-research/dependence/technical-flow-proof.tex
+++ b/website/content/sources/selected-research/dependence/technical-flow-proof.tex
@@ -1,6 +1,6 @@
\subsection{A.8 Equality kernels and continuous flow duality}\label{app:flow-proof}
-\leavevmode\phantomsection\label{i.2.-from-arbitrary-randomized-kernels-in-mathcal-z_0-to-linear-fields-and-flows}
+\leavevmode\phantomsection\label{proof:flow-representation}
\textit{Proof of the flow representation in Theorem~\ref{thm:projected-flow}.} We reduce the background-cost minimization over \(\mathcal Z_0\) to a flow on \(D=\Omega\cap\{x0\) on \(D\), the shares also define a simplex-valued allocation \(\sigma\)-almost everywhere. The kernel
\begin{equation}
\alpha_{\xi,u}=\delta_{\zeta(\theta_1,\theta_2,\theta_3/\beta)}
diff --git a/website/content/sources/selected-research/dependence/technical-limit.tex b/website/content/sources/selected-research/dependence/technical-limit.tex
index fb04ad1..2d6b788 100644
--- a/website/content/sources/selected-research/dependence/technical-limit.tex
+++ b/website/content/sources/selected-research/dependence/technical-limit.tex
@@ -1,8 +1,7 @@
-\label{appendix-a.-initial-slope-and-tail-surface-capacity}
\subsection{A.2 Threshold localization}\label{proof:threshold-localization}
\input{technical-threshold-statement.tex}
\textit{Proof.}
-For every \(P\) in Lemma~\ref{b-l2-all-optimal-thresholds-under-a-budget-of-order-varepsilon2}, the Lipschitz bound for the portfolio loss gives, uniformly in \(t\),
+For every \(P\) in Lemma~\ref{lem:threshold-localization}, the Lipschitz bound for the portfolio loss gives, uniformly in \(t\),
\begin{equation}
|\mathscr E_{L_{c,\varepsilon}}(P;t)
@@ -11,7 +10,7 @@ \subsection{A.2 Threshold localization}\label{proof:threshold-localization}
\tag{A.2}\label{eq:A.2}
\end{equation}
-Evaluating \eqref{eq:A.2} at any minimizer \(t\) under \(P\) and at a reference minimizer bounds the reference suboptimality at \(t\) by twice its right-hand side. By continuity of the section density in Lemma~\ref{lemma:general-variation}, \(m_0(c\varepsilon,t)\ge g/2\) on a fixed neighborhood of \(v\) for all sufficiently small \(\varepsilon\). The unique reference minimizers \(v_{c,\varepsilon}\), with \(v_{0,\varepsilon}=v\) and \(v_{1,\varepsilon}=t_\varepsilon\), converge to \(v\). Integrating the curvature bound within this neighborhood, we obtain
+Evaluating \eqref{eq:A.2} at any minimizer \(t\) under \(P\) and at a reference minimizer bounds the reference suboptimality at \(t\) by twice its right-hand side. By continuity of the section density in Lemma~\ref{lem:general-variation}, \(m_0(c\varepsilon,t)\ge g/2\) on a fixed neighborhood of \(v\) for all sufficiently small \(\varepsilon\). The unique reference minimizers \(v_{c,\varepsilon}\), with \(v_{0,\varepsilon}=v\) and \(v_{1,\varepsilon}=t_\varepsilon\), converge to \(v\). Integrating the curvature bound within this neighborhood, we obtain
\begin{equation}
\mathscr E_{L_{c,\varepsilon}}(P_0;t)
-ES_p^{P_0}(L_{c,\varepsilon})
@@ -22,8 +21,7 @@ \subsection{A.2 Threshold localization}\label{proof:threshold-localization}
Every candidate minimizer lies in the compact loss range. Since \(v_{c,\varepsilon}\to v\), \eqref{eq:A.3} gives a positive reference suboptimality gap at the endpoints of a smaller fixed neighborhood of \(v\), uniformly for small \(\varepsilon\). Convexity extends this gap outside the neighborhood, contradicting the uniform \(O(\varepsilon^2)\) bound from \eqref{eq:A.2}. Applying \eqref{eq:A.3} inside gives \(|t-v_{c,\varepsilon}|=O_M(\varepsilon)\), and \eqref{eq:2.6} yields \eqref{eq:A.1} for every minimizing threshold.\unskip\nobreak\hfill\proofbox
\subsection{A.3 Projections and threshold convexity}
-\label{a.4-source-fluxes-and-moving-optimal-thresholds}
-\label{a.5-strict-source-imbalance-on-an-exchangeable-tail-surface}
+\label{app:projections}
For a finite-\(I\) kernel, the conditional mean \(e_i(\xi,u)=\int d_i\alpha_{\xi,u}(dd)\) belongs to \(L^1(\nu)\), so Fubini's theorem gives
\begin{equation}
@@ -97,7 +95,7 @@ \subsection{A.3 Projections and threshold convexity}
\keepproofstart{9\baselineskip}
\paragraph{Convexity of limiting threshold objectives}
-\label{b.0-model-objects-and-main-statement}
+\label{par:threshold-convexity}
For finite-cost kernels, convexity of \(G_c\) follows by identifying its distributional second derivative with the measure of transported rescaled losses.
\resultnumber{A.4}\begin{lemma}[Threshold objectives of finite-cost kernels]\label{lem:threshold-convexity} For every kernel with finite \(I\) and every \(c\in\{0,1\}\), the transported rescaled loss \(u+b\cdot d+cH\) has a locally finite pushforward measure \(\lambda_c\), and
@@ -159,7 +157,7 @@ \subsection{A.4 Preliminary laws and kernel approximation}\label{proof:prelimina
with weight \(\alpha_k(v-y-\beta w,y,u)f(v-y-\beta w,y,w+\varepsilon u/\beta)\), after cancellation of the two Jacobian factors. Compact interior support and uniform continuity give \(a_{ik,\varepsilon}\to a_{ik,0}\) uniformly, including the zero extensions at the endpoints. Since \(j_i=\sum_kd_{ki}a_{ik,0}\), \eqref{eq:A.15} proves \eqref{eq:A.14} for all three coordinates.
\paragraph{Approximation by finite-branch kernels}
-\label{b-l8-bounded-measurable-kernels-arbitrary-finite-i-kernels-and-diagonal-recovery}
+\label{par:finite-branch-approximation}
To approximate a general finite-\(I\) kernel by kernels in \(\mathcal K_{\rm fin}^c\), we first retain activity on \(A_n=\{\xi\in\Omega_n,\ |u|\le n,\ |d|_\gamma\le n\}\), where \(\Omega_n\uparrow\Omega\) is a compact exhaustion, and return the omitted conditional probability to \(d=0\). With \(I_{\mathrm{tail},n}=\int_{A_n^c}|d|_\gamma\,d\Pi_\alpha\to0\), projection contraction and the displacement-Lipschitz bound for hinge increments give
\begin{equation}
|I_n-I|\le I_{\mathrm{tail},n},\qquad
diff --git a/website/content/sources/selected-research/dependence/technical-mixture-obstruction.tex b/website/content/sources/selected-research/dependence/technical-mixture-obstruction.tex
index 7172e4a..0b82578 100644
--- a/website/content/sources/selected-research/dependence/technical-mixture-obstruction.tex
+++ b/website/content/sources/selected-research/dependence/technical-mixture-obstruction.tex
@@ -1,4 +1,4 @@
-\label{rem:kernel-mixtures}
+\label{par:mixture-obstruction}
Finite cost does not ensure stationary mass: the kernel
\(\alpha_{\xi,u}=\delta_{\eta e^{-u^2}e_1}\), \(\eta>0\), moves every source and has
\(I=\eta\gamma_1g\sqrt\pi<\infty\).
diff --git a/website/content/sources/selected-research/dependence/technical-quantitative.tex b/website/content/sources/selected-research/dependence/technical-quantitative.tex
index 9c1c919..2008838 100644
--- a/website/content/sources/selected-research/dependence/technical-quantitative.tex
+++ b/website/content/sources/selected-research/dependence/technical-quantitative.tex
@@ -1,5 +1,5 @@
\keepproofstart{8\baselineskip}
-\section{Appendix B. Proofs for the analytical illustrations}\label{app:technical-quantitative}
+\section{Appendix B. Proofs for the analytical illustrations}\label{app:illustration-proofs}
\subsection{B.1 The projection-potential estimate}\label{app:potential-proof}
\begin{proof}[Proof of Proposition~\ref{prop:projection-potential-bound}]
Fix any minimizing threshold \(\tau\in\arg\min G_1\). With \(w=b\cdot d\), \(\Delta_+(u,w)=(u+w)_+-u_+\), and \(B_h(u,w)=\Delta_+(u,w)-\Delta_+(u+h,w)\), evaluating \(G_0\) at zero and \(G_1\) at \(\tau\) gives
@@ -31,7 +31,7 @@ \subsection{B.1 The projection-potential estimate}\label{app:potential-proof}
\subsection{B.2 The uniform midpoint-potential bound}\label{app:uniform-bound-proof}
-To establish the two threshold branches used for the lower bound in Corollary~\ref{prop:uniform-variable-cost}, we normalize costs and potentials by \(\gamma_0\), substitute the midpoint potentials of \hyperref[strict-premium-despite-equal-efficiency]{Section~5.1}, and take \(\lambda=3/2\) in these normalized units. We first evaluate the penalty on the same-side regions \(\{x0\) such that, for \(0<\varepsilon<\varepsilon_M\), any \(P\in\mathcal P(S)\) with \(W_\gamma(P,P_0)\le M\varepsilon^2\) has every ES-minimizing threshold of either portfolio loss in the window
+\resultnumber{A.2}\begin{lemma}[Threshold localization]\label{lem:threshold-localization} For every finite \(M\), there are \(K_M,\varepsilon_M>0\) such that, for \(0<\varepsilon<\varepsilon_M\), any \(P\in\mathcal P(S)\) with \(W_\gamma(P,P_0)\le M\varepsilon^2\) has every ES-minimizing threshold of either portfolio loss in the window
\begin{equation}
|t-v|\le K_M\varepsilon.
\tag{A.1}\label{eq:A.1}
diff --git a/website/content/sources/selected-research/dependence/technical-variational-proof.tex b/website/content/sources/selected-research/dependence/technical-variational-proof.tex
index f60f6fa..8b4987f 100644
--- a/website/content/sources/selected-research/dependence/technical-variational-proof.tex
+++ b/website/content/sources/selected-research/dependence/technical-variational-proof.tex
@@ -1,7 +1,7 @@
-\subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{appendix-b-proof-of-the-critical-variational-limit-under-full-marginal-constraints}
+\subsection{A.5 Proof of the variational limit}\label{app:variational-proof}
\begin{proof}[Proof of Theorem~\ref{thm:critical-limit}]
-\phantomsection\label{b-l3-local-compactness-of-arbitrary-transport-plans-and-the-three-fluxes}
+\phantomsection\label{step:compactness}
We first establish vague compactness and the energy lower bound in part~(i). Let \(\pi_n\) be admissible plans from \(P_0\) to \(P_n\in\Pi\) along a subsequence realizing a finite energy liminf, with rescaled coordinates and lifts
\begin{equation}
u=\frac{B(X)-v}{\varepsilon_n},\qquad
@@ -27,7 +27,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\tag{A.22}\label{eq:A.22}
\end{equation}
By Appendix A.3, the projected moments have densities \(j_i\in L^1(I_i)\). To include their marginal-restoration cost in the lower bound, we choose continuous cutoffs \(0\le\chi_R\uparrow1\), compactly supported in the interior surface and bounded \(u,d\)-windows.
-\phantomsection\label{b-l4-the-compensation-cost-forced-by-full-marginal-constraints}
+\phantomsection\label{step:compensation-cost}
The coordinate CDF discrepancy of \(\pi_n^R=\chi_R\pi_n\) is
\begin{equation}
\Delta_{i,n}^R(q)=\int[1_{\{Y_i\le q\}}-1_{\{X_i\le q\}}]\,d\pi_n^R.
@@ -68,7 +68,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\end{equation}
Thus any vaguely convergent sequence \(\Lambda_n\to\Lambda\) with finite energy liminf has \(\Lambda=\Pi_\alpha\) and \(\mathcal C(\alpha)\le\liminf_n\mathcal E_{\varepsilon_n}(\Lambda_n)\). In particular, the liminf must be infinite when \(\mathcal E(\Lambda)=+\infty\), proving the \(\Gamma\)-lower bound on all of \(\mathscr X\). Choosing optimal plans gives the Wasserstein bound in \eqref{eq:3.4}.
-\phantomsection\label{lem:v21-coordinate-lsc}\label{l1.-lower-semicontinuity-of-the-combined-cost-of-one-coordinate}
+\phantomsection\label{step:coordinate-lsc}
For later boundary sensitivity, we also need lower semicontinuity of \(A_i(\alpha)=M_i(\alpha)+\|j_i\|_1\), where \(M_i(\alpha)=\int|d_i|\,d\Pi_\alpha\). Retaining activity weighted by \(\chi_R\) and assigning the remainder to \(d=0\) defines \(\alpha^R\), for which projection contraction gives
\begin{equation}
\|j_i-j_i^R\|_1\le M_i-M_i^R,\qquad
@@ -83,7 +83,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\end{equation}
The same extraction with fixed source \(\nu\) makes cost sublevels compact in the vague topology, that is, convergence against \(C_c(Z)\). A countable compact exhaustion metrizes this topology on the sublevels.
-\paragraph{The risk limit}\label{b-l5-renormalization-of-the-common-term-and-identification-of-the-risk-difference}
+\paragraph{The risk limit}\label{par:risk-limit}
Although \(|\mathfrak q_c|\le\kappa_B|d|_\gamma\) makes each hinge integral finite, bounded moments alone do not imply uniform integrability. We therefore subtract the common increment and estimate \(r_c(\tau)=\mathfrak q_c(\tau)-\mathfrak q_0(0)\). For \(w=b\cdot d\), the function \(z\mapsto(z+w)_+-z_+\) is 1-Lipschitz, bounded in magnitude by \(|w|\), and constant on each exterior half-line determined by its breakpoints \(0\) and \(-w\). The two arguments defining \(r_c\), namely \(u+cH-\tau\) and \(u\), differ by at most \(A_T=\|H\|_\infty+T\) when \(|\tau|\le T\). If \(|u|>A_T+|w|\), both arguments lie on the same exterior half-line, so their increments agree. These observations give
\begin{equation}
@@ -113,14 +113,14 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
=\frac g2\tau^2+c^2J+\int r_c(\tau)d\Pi_n+o(1).
\tag{A.33}\label{eq:A.33}
\end{equation}
-Lemma~\ref{b-l2-all-optimal-thresholds-under-a-budget-of-order-varepsilon2} and \(g\tau^2/2\le2\kappa_BI\) place all original and limiting minimizers in a common compact interval. The finite scalar \(Q_n\), with \(|Q_n|\le\kappa_BM\), translates both objectives equally and cancels from their difference of minima without needing to converge itself. Uniform convergence on that interval therefore yields
+Lemma~\ref{lem:threshold-localization} and \(g\tau^2/2\le2\kappa_BI\) place all original and limiting minimizers in a common compact interval. The finite scalar \(Q_n\), with \(|Q_n|\le\kappa_BM\), translates both objectives equally and cancels from their difference of minima without needing to converge itself. Uniform convergence on that interval therefore yields
\begin{equation}
\Phi_{\varepsilon_n}(P_n)/\varepsilon_n^2\to\mathcal V(\alpha).
\tag{A.34}\label{eq:A.34}
\end{equation}
Since plan cost bounds target distance, localization applies to every bounded-energy sequence in part (i). Optimal plans give the corresponding law-level limits in \eqref{eq:3.4}.
-\paragraph{Exact recovery by quantile restoration}\label{b-l6-nonnegative-preliminary-recovery-for-finite-smooth-kernels-and-a-uniform-cdf-expansion}
+\paragraph{Exact recovery by quantile restoration}\label{par:quantile-recovery}
For a fixed kernel \eqref{eq:A.13}, the preliminary plan of Appendix A.4 modifies \(O_\alpha(\varepsilon)\) source mass and has cost
\begin{equation}
@@ -146,7 +146,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\tag{A.37}\label{eq:A.37}
\end{equation}
-\phantomsection\label{b-l7-exact-repair-of-all-three-marginals-repair-cost-and-cancellation-in-the-risk-difference}
+\phantomsection\label{step:marginal-restoration}
To restore the marginals, observe that their reference densities are bounded below by \(m_f=(\min_Sf)\min\{\Delta_I\Delta_3,\Delta_I^2\}>0\). Thus \(F_i^{-1}\) is \(m_f^{-1}\)-Lipschitz. Set
\begin{equation}
R_{i,\varepsilon}(q)=F_i^{-1}(F_{i,\varepsilon}(q)),\quad
@@ -168,7 +168,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\end{aligned}
\tag{A.40}\label{eq:A.40}
\end{equation}
-To show that restoration has no leading paired-risk effect, we estimate its effect uniformly over both minimizing thresholds. The preliminary and restored laws \(Q_\varepsilon,P_\varepsilon\) are within \(O_\alpha(\varepsilon^2)\) of \(P_0\), so Lemma~\ref{b-l2-all-optimal-thresholds-under-a-budget-of-order-varepsilon2} places all their minimizing thresholds in \([v-T_\alpha\varepsilon,v+T_\alpha\varepsilon]\). For any pair in this window, put \(\psi_{t_B,t_A}(z)=(B(z)-t_B)_+-(B(z)+\varepsilon H(z)-t_A)_+\). Away from a sufficiently wide band \(\mathcal B_\varepsilon=\{|B-v|\le K\varepsilon\}\), both hinges stay on the same affine branch along the restoration segment: the gradient of their difference is \(-\varepsilon h\) above both thresholds and zero below both. Inside the band we retain the uniform gradient bound, but only \(O_\alpha(\varepsilon)\) mass is affected. Indeed, the reference density of \(B\) is bounded by \(\Delta_I^2\max_Sf/\beta\) and the preliminary plan changes \(O_\alpha(\varepsilon)\) mass, so \(Q_\varepsilon(\mathcal B_\varepsilon)=O_\alpha(\varepsilon)\). Combining this small band mass and the extra factor \(\varepsilon\) outside the band with the \(O_\alpha(\varepsilon^2)\) restoration displacement in \eqref{eq:A.39} gives
+To show that restoration has no leading paired-risk effect, we estimate its effect uniformly over both minimizing thresholds. The preliminary and restored laws \(Q_\varepsilon,P_\varepsilon\) are within \(O_\alpha(\varepsilon^2)\) of \(P_0\), so Lemma~\ref{lem:threshold-localization} places all their minimizing thresholds in \([v-T_\alpha\varepsilon,v+T_\alpha\varepsilon]\). For any pair in this window, put \(\psi_{t_B,t_A}(z)=(B(z)-t_B)_+-(B(z)+\varepsilon H(z)-t_A)_+\). Away from a sufficiently wide band \(\mathcal B_\varepsilon=\{|B-v|\le K\varepsilon\}\), both hinges stay on the same affine branch along the restoration segment: the gradient of their difference is \(-\varepsilon h\) above both thresholds and zero below both. Inside the band we retain the uniform gradient bound, but only \(O_\alpha(\varepsilon)\) mass is affected. Indeed, the reference density of \(B\) is bounded by \(\Delta_I^2\max_Sf/\beta\) and the preliminary plan changes \(O_\alpha(\varepsilon)\) mass, so \(Q_\varepsilon(\mathcal B_\varepsilon)=O_\alpha(\varepsilon)\). Combining this small band mass and the extra factor \(\varepsilon\) outside the band with the \(O_\alpha(\varepsilon^2)\) restoration displacement in \eqref{eq:A.39} gives
\begin{equation}
\begin{aligned}
|(E_{P_\varepsilon}-E_{Q_\varepsilon})\psi_{t_B,t_A}|
@@ -209,9 +209,8 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\]
Using the \(m\)-th recovery for \(\varepsilon\in(\varepsilon_{m+1},\varepsilon_m]\) gives a full recovery family with exact marginals, vague limit \(\Pi_\alpha\), and the energy and risk limits in \eqref{eq:3.5}. Its risk error is \(o(\varepsilon^2)\), because the fixed-approximant cubic constants need not be uniform in \(m\).
-\paragraph{Optimizing over the budget}\label{completion-of-b-t-original-model-budgets-budget-boundaries-and-local-uniform-convergence}
+\paragraph{Optimizing over the budget}\label{par:budget-optimization}
-\label{b-l9-compact-cost-sublevel-sets-continuity-of-the-risk-difference-and-budget-attainment}
Parts (i)--(ii) are now proved. For kernels with fixed source \(\nu\) and \(I\le M\), the same estimates \eqref{eq:A.30}--\eqref{eq:A.32} give uniform convergence of the relative objectives on a common compact interval containing every minimizer. Their common term cancels, so \(\mathcal V\) is continuous on compact cost sublevels and \(F\) is finite and attained, with \(F(0)=-J/p\). Mixing with the identity, \(\alpha^\theta=\theta\alpha+(1-\theta)\delta_0\), scales \(I,j_i,T_c\) linearly and gives
\begin{equation}
\mathcal C(\alpha^\theta)=\theta\mathcal C(\alpha),\quad
@@ -225,7 +224,7 @@ \subsection{A.5 Proof of the variational limit}\label{v21-limit-proof}\label{app
\tag{A.44}\label{eq:A.44}
\end{equation}
-\paragraph{Reflection at finite cost}\label{sec:reflection-construction}\label{proof:positive-witness}
+\paragraph{Reflection at finite cost}\label{par:reflection}
To obtain a positive value of \(F\) at finite cost, we reflect two normal bands over exchanged surface patches. The \(B\)-increments then cancel, and reversal of \(H\) makes the \(A_\varepsilon\)-increments add with negative sign without shifting the minimizing thresholds. Choose a compact positive-area patch \(E_+\Subset\Omega\cap\{H>h_0\}\), \(h_0>0\), and its exchanged image \(E_-\), of common mass \(m>0\). Set \(\Delta_E=\int_{E_+}H\,d\sigma-\int_{E_-}H\,d\sigma>0\). With \(U>\Delta_I\) and \(R>0\), use \(d=-2ue_1\) on \(E_+\times[U,U+R]\) and \(E_-\times[-U-R,-U]\), leaving other sources fixed. Integration gives
\begin{equation}
I=4\gamma_1m(UR+R^2/2),\qquad \mathcal C\le2I<\infty.
diff --git a/website/content/sources/selected-research/provenance.json b/website/content/sources/selected-research/provenance.json
index 36464d7..f9ec13b 100644
--- a/website/content/sources/selected-research/provenance.json
+++ b/website/content/sources/selected-research/provenance.json
@@ -4,11 +4,11 @@
{
"slug": "certified-valuation-of-arithmetic-asian-options",
"fetchedAt": "2026-11-19",
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/manuscript/report.md",
- "sourceCommit": "41a814f6e5aa1eddddb4412e57ce68ec506c323e",
- "sourceBlob": "6f573b14c4c172e394688333b22726438bbed461",
- "sourceSha256": "cb4de5e9741183427b9406ae64e4654b3d4084b60ed2abff5111e40a2d137fe0",
- "importedMarkdownSha256": "8fea27c95d7e1fe9c5ef24204155b469bc58ad2c2bb1d3d5c531a89b29da5177",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/manuscript/report.md",
+ "sourceCommit": "dd4b147d822d45a4ca12a66294fae2a32cd252bb",
+ "sourceBlob": "73fd58c109bca144a8809e3fc9832e8ad258cae6",
+ "sourceSha256": "9390024a6e2cd57e036fbcac6aac05ef1fa8ef5580c5637f8403931472d1b055",
+ "importedMarkdownSha256": "337d0d026733919fc9219e0aa36fe7624201b5186957fa5fac5b41a5e3a4c11c",
"includedHeadings": [
"Abstract",
"1. Introduction",
@@ -88,12 +88,12 @@
],
"studies": [
{
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/studies/results/gap-profile.json",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/studies/results/gap-profile.json",
"sourceBlob": "30e1676f4cc0c72d0b0dbdb1cac384c41a70370b",
"sourceSha256": "c90aa118de02bff2d85bf816639e9b9883d47de60f829dc7085e0321931d8bac"
},
{
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/41a814f6e5aa1eddddb4412e57ce68ec506c323e/studies/results/euler-ladder.json",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options/blob/dd4b147d822d45a4ca12a66294fae2a32cd252bb/studies/results/euler-ladder.json",
"sourceBlob": "5359e2b42d32881df6fd678301feb0051cf4095c",
"sourceSha256": "8df582b58273fe949af1e892fcfe987c5dafe69425bd0ab76b3a219907d05e1f"
}
@@ -103,11 +103,11 @@
{
"slug": "certified-joint-pricing-errors-in-rough-heston",
"fetchedAt": "2026-11-19",
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/manuscript/merged-heston-en.md",
- "sourceCommit": "46c85b4b256dfbc597ed56b3b2b526681334e60f",
- "sourceBlob": "27bd396d79cfe3d99072f23d9799641c898442e7",
- "sourceSha256": "4659940d5ab3d468b7b555ebec25fd22ae76df20992d3944521aefbb75adee78",
- "importedMarkdownSha256": "0ba7ea3267c18bc4f125865916a1a76957871e94159b40230944f6abd8fb4435",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/manuscript/merged-heston-en.md",
+ "sourceCommit": "038db24a3aadf9600edef385dfd7508e612d53c8",
+ "sourceBlob": "12b074bc25984ea11f9fc7d75768800c267a43eb",
+ "sourceSha256": "bbb79f3b21e37fde92b02fb45a49b3a4404492e7f37819c5c1a7a4ee4db32642",
+ "importedMarkdownSha256": "6e103cb9b05c6548920b617517d5e39002ffaddc5de5928be14bf9f1e371d36c",
"includedHeadings": [
"Abstract",
"1. Introduction",
@@ -176,17 +176,17 @@
],
"studies": [
{
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/science/baseline/finite-history/finite-history-nodes-alpha-13_25.json",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/science/baseline/finite-history/finite-history-nodes-alpha-13_25.json",
"sourceBlob": "c6ae3634bd49abcf1ce00f1abc9cd1e9648acdb6",
"sourceSha256": "a0b2df22264e111b177541383eecf01bba63f47aac6296c3428f832a0c5ceca7"
},
{
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/science/baseline/frontier/transfer-time-local-nodes.json",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/science/baseline/frontier/transfer-time-local-nodes.json",
"sourceBlob": "085f5093feb64a2d51bfea8e25e5090875b78345",
"sourceSha256": "60e0cc82268facc9af3523847a707e1fda5ed7142060f4bf02a648db18e839f8"
},
{
- "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/46c85b4b256dfbc597ed56b3b2b526681334e60f/audit-experiments/portfolio-budget-counts.json",
+ "sourceUrl": "https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/038db24a3aadf9600edef385dfd7508e612d53c8/audit-experiments/portfolio-budget-counts.json",
"sourceBlob": "c54490b525090394fec3773dc497a5cc6ff463a0",
"sourceSha256": "8655215cef79ab6995f4472bd64fcd2e21b6daf5824d30d4b27c819520cfaf20"
}
diff --git a/website/content/sources/selected-research/rough-source.md b/website/content/sources/selected-research/rough-source.md
index 27bd396..12b074b 100644
--- a/website/content/sources/selected-research/rough-source.md
+++ b/website/content/sources/selected-research/rough-source.md
@@ -51,7 +51,7 @@ Once each \(z_n\) lies in a disk of radius \(\epsilon_n\), the set of price erro
\]
where \(k_i=\log m_i\) is log-moneyness, \(h_*\) is the grid step and \(\omega_n\) are the trapezoidal weights. The function \(\widehat w\) is the Fourier transform of the payoff of the portfolio. A portfolio of options with a common expiry is one payoff, and (1.4) prices its error as one payoff. Bounding each leg separately replaces \(|\widehat w(u)|\) by \(\sum_i|w_i|\sqrt{m_i}\), and the ratio of the two, which we call the netting ratio, measures what is lost.
-The netting ratio has a simple anatomy. For a vertical spread with log-strike gap \(\Delta\) it is \(|\sin(u\Delta/2)|\) up to terms of order \(\Delta\): the spread is hedged against transform error of wavelength longer than twice its strike spacing and fully exposed at the Nyquist frequency \(\pi/\Delta\). For a butterfly the ratio is squared. For any portfolio its mean square over a long band of frequencies is the Herfindahl index of the position (Propositions 5.2 and 5.3). Transform error therefore behaves like a factor model in frequency. Low-frequency error is systematic across strikes and is removed by hedging. High-frequency error is idiosyncratic across strikes and is removed by diversification. On the S&P 500 example the eleven adjacent call spreads keep on average 0.708 of their leg-by-leg node error, against a Herfindahl benchmark of \(1/\sqrt2\approx0.707\); an equal-weight basket of all twelve calls keeps 0.076.
+The netting ratio has a simple anatomy. For a vertical spread with log-strike gap \(\Delta\) it is \(|\sin(u\Delta/2)|\) up to terms of order \(\Delta\): the spread is hedged against transform error whose wavelength is long compared with its strike spacing and fully exposed at the Nyquist frequency \(\pi/\Delta\). For a butterfly the ratio is squared. For any portfolio its mean square over a long band of frequencies is the Herfindahl index of the position (Propositions 5.2 and 5.3). Transform error therefore behaves like a factor model in frequency. Low-frequency error is systematic across strikes and is removed by hedging. High-frequency error is idiosyncratic across strikes and is removed by diversification. On the S&P 500 example the eleven adjacent call spreads keep on average 0.708 of their leg-by-leg node error, against a Herfindahl benchmark of \(1/\sqrt2\approx0.707\); an equal-weight basket of all twelve calls keeps 0.076.
This part of the paper is model-free. It applies to any pricer that obtains a strip of options from one transform, whatever the model and whatever the method that produced the transform.
@@ -227,13 +227,13 @@ Four remarks place the statement.
**Remark 3.3 (the classical model).** At \(\alpha=1\) the resolvent is \(k_\lambda(t)=e^{-\lambda t}\) and \(\xi^\lambda(t)=V_0e^{-\lambda t}+\int_0^te^{-\lambda(t-s)}\theta(s)\,ds\). Inequality (3.5) is then variation of constants for the Heston Riccati equation, in which the error is transported by a flow that contracts at rate \(\lambda\). For \(\alpha<1\) there is no flow to transport it. Convexity of the Caputo derivative and positivity of \(k_\lambda\) recover the same conclusion with the same constant.
-**Remark 3.4 (the constant).** The constant one in (3.5) is sharp. At \(a=0\), the point of zero frequency on the undamped contour, the exact solution vanishes, \(\lambda=0\), and a residual of constant sign produces \(\widehat L_T-L_T=\int_0^T\xi(T-t)r(t)\,dt\) up to terms of second order in \(r\). On the Lewis contour the exact solution adds damping of its own through \(\operatorname{Re}h<0\), growing with the frequency, and Theorem 3.1 uses the part that is uniform in the frequency.
+**Remark 3.4 (the constant).** The constant one in (3.5) is sharp. Write \(a\) for the argument of \(\phi_T\), so that the Lewis contour is \(a=u-\tfrac i2\). At \(a=0\), the point of zero frequency on the undamped contour, the exact solution vanishes, \(\lambda=0\), and a residual of constant sign produces \(\widehat L_T-L_T=\int_0^T\xi(T-t)r(t)\,dt\) up to terms of second order in \(r\). On the Lewis contour the exact solution adds damping of its own through \(\operatorname{Re}h<0\), growing with the frequency, and Theorem 3.1 uses the part that is uniform in the frequency.
**Remark 3.5 (other contours and positive correlation).** On a contour \(\operatorname{Im}a=-\eta\) with \(0<\eta<1\) the linear coefficient of the Riccati polynomial has real part \(\eta\rho\nu\), the left half-plane is again invariant, and the proof gives (3.5) with \(\lambda=\eta|\rho|\nu\) and the measure \(\mathbb Q^\eta\). The damping available to an error bound thus rises linearly from nothing at the pricing measure to \(|\rho|\nu\) at the share measure. For \(\rho>0\) the same argument gives (3.4) and the first inequality of (3.5) with a negative \(\lambda\), the resolvent being positive term by term; the kernel \(\xi^\lambda\) then lies above \(\xi\), and residuals compound at the rate at which variance runs away under the tilted measure.
**Remark 3.6 (other kernels).** The fractional kernel enters the proof through three facts: \(g_\alpha*g_{1-\alpha}=1\) with \(g_{1-\alpha}\) nonnegative and nonincreasing, which gives the convexity inequality; positivity of the resolvent \(k_\lambda\); and \(\xi=g_\alpha*q\) with \(q\ge0\), which is the second face. Every completely monotone kernel \(K\) has a resolvent of the first kind with the same monotonicity and a completely monotone resolvent (Gripenberg, Londen and Staffans 1990, Chapter 5), its admissible curves are again of the form \(\xi=K*q\) with \(q\) a nonnegative measure (Abi Jaber and El Euch 2019a, Example 2.2), and the left half-plane is invariant for its Riccati–Volterra equation (Abi Jaber, Larsson and Pulido 2019, Theorem 7.1). The argument therefore gives (3.5) for Volterra Heston models with completely monotone kernels, with \(\xi^\lambda\) defined by \(\xi^\lambda+\lambda K*\xi^\lambda=\xi\): the classical Heston model, where \(K\equiv1\) and the statement is Remark 3.3; the lifted Heston model of Abi Jaber (2019), where \(K\) is a finite sum of exponentials; and gamma kernels. Appendix C writes out the regularity for the fractional kernel.
-Figure 1 shows the two weights for the curve calibrated in Section 7. At six months the half-share curve lies below the risk-neutral one by at most a few per cent, since \(\lambda_*T^\alpha\approx0.075\). The large gain in the experiments comes from the first inequality in (3.5), which weights each residual by the variance still to come, in place of a uniform bound on the state.
+Figure 1 shows the two weights for the curve calibrated in Section 7. At six months the half-share curve lies below the risk-neutral one by at most 7.5 per cent, the order of \(\lambda_*T^\alpha\approx0.075\). The large gain in the experiments comes from the first inequality in (3.5), which weights each residual by the variance still to come, in place of a uniform bound on the state.

@@ -423,9 +423,9 @@ The proof in Appendix B reduces (6.1) to the signs of nine real functions on a c
### 7.1. Data and market scale
-The quotes are the bid and ask implied volatilities of twelve SPX options with six months to expiry, strikes 3700 to 4800 in steps of 100, recorded on 18 June 2021 and published by Jeng and Kiliçman (2021). We use the index level 4221.86 of that data set as the forward, with zero carry, and convert each implied volatility to a call price by the Black–Scholes formula in exact arithmetic. Table 1 gives the result in index points. One point is 100 dollars per contract.
+The quotes are the bid and ask implied volatilities of twelve SPX options with six months to expiry, strikes 3700 to 4800 in steps of 100, recorded on 18 June 2021 and published by Jeng and Kiliçman (2021). We use the index level 4221.86 of that data set as the forward, with zero carry, and convert each implied volatility to a call price by the Black–Scholes formula in rigorous interval arithmetic. Table 1 gives the result in index points. One point is 100 dollars per contract.
-**Table 1. Six-month SPX call quotes of 18 June 2021 in index points: bid, ask and quoted width.**
+**Table 1. Six-month SPX call quotes of 18 June 2021 in index points: bid, ask and quoted width, each width computed from the unrounded quotes.**
| Strike | Bid | Ask | Width | Strike | Bid | Ask | Width |
|---:|---:|---:|---:|---:|---:|---:|---:|
@@ -541,7 +541,7 @@ The six-month experiment fixes 28 portfolio directions on the twelve strikes bef
| Wide spreads | \(e_0-e_3,\ e_3-e_6,\ e_6-e_9,\ e_0-e_{11}\) | 4 | 0.474–0.570–0.651 | 0.707 | 0.235 |
| Equal-weight baskets | twelve, lower six, upper six strikes | 3 | 0.076–0.156–0.221 | 0.289, 0.408 | 0.344 |
-The adjacent spreads keep 0.708 of their leg-by-leg node error on average, against the Herfindahl level 0.707 of a two-leg position. The butterflies keep 0.524. The baskets keep least of all: an equal-weight position in all twelve calls keeps 0.076 of its leg-by-leg node error, because three quarters of the node charge of this bank sits between \(u=64\) and \(u=96\), where twelve strikes spread over a quarter of a unit of log-moneyness respond out of phase. The centre runs the other way. The stored rational output lies above the reference at every strike, by 0.04 points at the ends of the ladder and by up to 0.65 points near the money. A displacement of one sign cancels in part in spreads and butterflies and accumulates in baskets, where it makes up 0.344 of a complete bound of 0.356 points. The approximation error of the stored formula is systematic across strikes and the uncertainty of the certificate is idiosyncratic, as Section 5.2 anticipated, and the joint bound accounts for each correctly.
+The adjacent spreads keep 0.708 of their leg-by-leg node error on average, against the Herfindahl level 0.707 of a two-leg position. The butterflies keep 0.524. The baskets keep least of all: an equal-weight position in all twelve calls keeps 0.076 of its leg-by-leg node error, because three quarters of the node charge of this bank sits between \(u=64\) and \(u=96\), where twelve strikes spread over a quarter of a unit of log-moneyness respond out of phase. The centre runs the other way. The stored rational output lies above the reference at every strike, by 0.04 and 0.07 points at the two ends of the ladder and by up to 0.65 points near the money. A displacement of one sign cancels in part in spreads and butterflies and accumulates in baskets, where it makes up 0.344 of a complete bound of 0.356 points. The approximation error of the stored formula is systematic across strikes and the uncertainty of the certificate is idiosyncratic, as Section 5.2 anticipated, and the joint bound accounts for each correctly.
Table 8 counts the positions certified at three tolerances. Within a row the two methods use identical radii.
@@ -565,7 +565,7 @@ At half a point and \(\alpha=0.52\) the joint bound certifies every position and
### 7.5. Is the roughness ranking numerical?
-The roughness fitted to this surface is \(H=0.0286\), and candidates a few thousandths apart in \(H\) produce prices that differ by fractions of an index point. Whether a calibration can tell them apart depends on the accuracy of the pricer. We test five candidates, \(\alpha\in\{0.520,0.525,0.530,0.540,0.550\}\), with the curve (7.1), the quotes and all other parameters fixed. The design was set in advance: every candidate receives a freshly generated reference field and a complete residual bank at \(N_t=1024\), and all five are regenerated at \(N_t=2048\) if any pair remains unresolved. The objective is the normalised least-squares distance to the mid prices, enclosed by Proposition 5.4.
+The roughness fitted to this surface is \(H=0.0286\), and candidates a few thousandths apart in \(H\) produce prices that differ by fractions of an index point. Whether a calibration can tell them apart depends on the accuracy of the pricer. We test five candidates, \(\alpha\in\{0.520,0.525,0.530,0.540,0.550\}\), with the curve (7.1), the quotes and all other parameters fixed. The design was set in advance: every candidate receives a freshly generated reference field and a complete residual bank at \(N_t=1024\), and all five are regenerated at \(N_t=2048\) if any pair remains unresolved. The objective is the normalised least-squares distance to the mid prices, enclosed by Proposition 5.4; Table 9 gives the enclosures at both levels.
**Table 9. Objective enclosures for five roughness candidates, as \(J\times10^8\), and the number of the ten pairs strictly separated.**
@@ -583,7 +583,7 @@ At the finer level the five enclosures are disjoint and ordered, and \(\alpha=0.
### 7.6. Model misfit without solver doubt
-A certificate also settles the opposite question. Each of the five candidates is inconsistent with at least one quoted bid–ask band, and for a certified price this is a statement about the model. At \(\alpha=0.52\), on the coarsest bank H0, the exact model price of the 4400 call lies in an interval whose lower end is 92.887 points, while the quoted ask is 92.396: the model is above the ask by at least 0.491 points. For the 4500 call the excess is at least 0.068 points. At the other ten strikes the certified interval meets the quoted band. A desk that sees its rough Heston calibration miss the 4400 and 4500 strikes on this date can rule out the pricer as the cause.
+A certificate also settles the opposite question. Each of the five candidates is inconsistent with at least one quoted bid–ask band, and for a certified price this is a statement about the model. At \(\alpha=0.52\), on the coarsest bank H0, the exact model price of the 4400 call lies in an interval whose lower end is 92.886 points, while the quoted ask is 92.396: the model is above the ask by at least 0.490 points. For the 4500 call the excess is at least 0.067 points. At the other ten strikes the certified interval meets the quoted band. A desk that sees its rough Heston calibration miss the 4400 and 4500 strikes on this date can rule out the pricer as the cause.
### 7.7. Auditing an output or replacing it
@@ -663,7 +663,7 @@ The curve (7.1) is nondecreasing, has a nonnegative initial value, and is locall
b_\alpha(t)=D_C^\alpha\xi_*(t)=I^{1-\alpha}\xi_*'(t)\ge0,\quad
b_\alpha(t)=
\frac{(\theta-V_0)\lambda_\xi}{\Gamma(1+\alpha_0-\alpha)}
-t^{\alpha_0-\alpha}+O(t^{2\alpha_0-\alpha})
+t^{\alpha_0-\alpha}+O(t^{2\alpha_0-\alpha}).
\tag{A.7}
\]
Its worst exponent is \(-.3714>-.5\), so it belongs to \(L^2_{\rm loc}\); also \(\xi_*=V_0+K_\alpha*b_\alpha\). The nonnegative input measure meets Example 2.2(ii). This verifies the model specification with a varying kernel and fixed forward variance curve.
@@ -858,7 +858,7 @@ The repository contains the interval implementation, all leaf rectangles and the
\tag{B.16}
\]
-where primes denote real \(\rho\) derivatives. The identity \(|R|\le1\) follows from \(|A+S_b|^2\ge1\): the terms \(\operatorname{Re} A\,\operatorname{Re} S_b\) and \(\operatorname{Im} A\,\operatorname{Im} S_b\) are nonnegative, while \(|A|^2=|1+S_b^2|\ge1-|S_b|^2\). Also \(\operatorname{Re} R>0\), since \((\operatorname{Re} A)^2-(\operatorname{Re} S_b)^2=(|1+S_b^2|+1-|S_b|^2)/2>0\).
+where primes denote real \(\rho\) derivatives. The bound \(|R|\le1\) follows from \(|A+S_b|^2\ge1\): the terms \(\operatorname{Re} A\,\operatorname{Re} S_b\) and \(\operatorname{Im} A\,\operatorname{Im} S_b\) are nonnegative, while \(|A|^2=|1+S_b^2|\ge1-|S_b|^2\). Also \(\operatorname{Re} R>0\), since \((\operatorname{Re} A)^2-(\operatorname{Re} S_b)^2=(|1+S_b^2|+1-|S_b|^2)/2>0\).
For the \(\alpha\)-dependent scalars of Proposition 6.1, exact outward endpoint enclosures give
@@ -890,7 +890,7 @@ m_{D,j}(C)-10^{-6}L_{D,j}>0
hold. By the mean-value theorem the test is sufficient. It certifies 182002 original leaves and 79758 finer leaves. For the remaining cells, directly evaluate the original complex polynomial formulas with \(\rho\), \(\alpha\) and \(\eta\) all interval valued. Exact outward dyadic arithmetic certifies strict \(B_j>0\) and \(D_j<0\) on 19808 additional closed cells.
-The original complete midpoint tree and the exact local splice trees verify that these 281568 cells cover the entire \((\alpha,\eta)\) rectangle without gaps or interior overlap. Their exact area is \(2/25\); each cell carries the entire correlation interval, giving exact three-dimensional volume \(1/6250000\). Hence the raw signs hold for every point of the stated domain. In particular \(\Delta\ne0\), the normalised denominator coefficients have positive real part, and all real numerator coefficients of the half-plane test are negative. The algebraic implication of B.1 gives the conclusions of Proposition 6.1, including \(\eta=1\), the infinite-frequency limit. Conjugacy supplies negative frequencies. \(\square\)
+The original complete midpoint tree and the exact local splice trees verify that these 281568 cells cover the entire \((\alpha,\eta)\) rectangle without gaps or interior overlap. Their exact area is \(2/25\); each cell carries the entire correlation interval, giving exact three-dimensional volume \(1/6250000\). Hence the raw signs hold for every point of the stated domain. In particular \(\Delta\ne0\), the normalised denominator coefficients have positive real part, and all real numerator coefficients of the half-plane test are negative. The algebraic implication of Appendix B.1 gives the conclusions of Proposition 6.1, including \(\eta=1\), the infinite-frequency limit. Conjugacy supplies negative frequencies. \(\square\)
A separate reader checks every saved sign, the 422481-node partition tree, the local refinements, the derivative majorants and all 19808 direct cells through separately assembled Cramer matrices and numerator convolutions; it shares the dyadic primitives with the generator. The interval is the one covered by the stored computation, which confirms stability at the calibrated value.
@@ -908,7 +908,7 @@ H'(x)=\frac{q(x)x^{\alpha-1}}{\Gamma(\alpha)}
The bound
\[
-|H'(x)|\le\frac{\|q\|_\infty x^{\alpha-1}}{\Gamma(\alpha)}+\frac{(1-\alpha)[q]_{C^\alpha}x^{2\alpha-1}}{\Gamma(\alpha)(2\alpha-1)}.
+|H'(x)|\le\frac{\|q\|_\infty x^{\alpha-1}}{\Gamma(\alpha)}+\frac{(1-\alpha)[q]_{C^\alpha}x^{2\alpha-1}}{\Gamma(\alpha)(2\alpha-1)}
\tag{C.2}
\]
@@ -956,7 +956,7 @@ In the variables of (2.3) this is Proposition 2.1.
**Theorem C.2 (dissipative residual bound).** Suppose \(s_0>0\), \(\widehat H(0)=0\), and \(\widehat H\in AC\), with local Lipschitz regularity at positive times. Assume, for every \(x\in(0,X]\), that
\[
\operatorname{Re}\widehat H\le\epsilon_R<2s_0,\qquad
-|D_C^\alpha\widehat H-F(\widehat H)|\le\delta,
+|D_C^\alpha\widehat H-F(\widehat H)|\le\delta.
\]
Set \(\sigma=s_0-\epsilon_R/2>0\). Then
\[
@@ -1154,7 +1154,7 @@ q_\alpha(t)=\frac{V_0t^{-\alpha}}{\Gamma(1-\alpha)}
\[
A_\alpha(t)=\frac{\theta t^{1-\alpha}}{\Gamma(2-\alpha)}
+(V_0-\theta)t^{1-\alpha}
- E_{\alpha_0,2-\alpha}(-\lambda_\xi t^{\alpha_0}),\quad A_\alpha(0)=0.
+ E_{\alpha_0,2-\alpha}(-\lambda_\xi t^{\alpha_0}),\quad A_\alpha(0)=0,
\tag{C.25}
\]
where \(E_{a,b}(z)=\sum_{n=0}^\infty z^n/\Gamma(an+b)\).
@@ -1162,7 +1162,7 @@ where \(E_{a,b}(z)=\sum_{n=0}^\infty z^n/\Gamma(an+b)\).
**Theorem C.7 (uniform-state exponent bound).** For a zero-initial-value AC trajectory \(Z,\widehat Z\), use the derivative-based exponent
\[
L_T=\nu^{-1}\int_0^T\xi_*(T-t)D_t^\alpha Z(t)\,dt,\quad
-\bar L_T=\nu^{-1}\int_0^T\xi_*(T-t)D_t^\alpha\widehat Z(t)\,dt,
+\bar L_T=\nu^{-1}\int_0^T\xi_*(T-t)D_t^\alpha\widehat Z(t)\,dt.
\]
Then
\[
@@ -1206,8 +1206,8 @@ and the convexity bound \(e^{(1-\tau)a+\tau b}\le(1-\tau)e^a+\tau e^b\). Since (
Write \(K_\lambda=q_\alpha*k_\lambda=\xi^\lambda\); at \(\lambda=0\) the resolvent is \(g_\alpha\) and \(K_0=\xi\).
**Proposition C.8 (cumulative-weight series).** Suppose
-\(0<\alpha,A_0<1\), \(0\le V_0\le\theta\), and the fixed curve is
-\(\xi(t)=\theta+(V_0-\theta)E_{A_0}(-\lambda_\xi t^{A_0})\), with \(\lambda_\xi\ge0\). Assume \(q_\alpha\ge0\). For \(\lambda\ge0\) and \(\lambda T^\alpha<1\), (3.9) holds on \([0,T]\). If it is truncated before index \(M\), the entire omitted outer series satisfies
+\(0<\alpha,\alpha_0<1\), \(0\le V_0\le\theta\), and the fixed curve is
+\(\xi(t)=\theta+(V_0-\theta)E_{\alpha_0}(-\lambda_\xi t^{\alpha_0})\), with \(\lambda_\xi\ge0\). Assume \(q_\alpha\ge0\). For \(\lambda\ge0\) and \(\lambda T^\alpha<1\), (3.9) holds on \([0,T]\). If it is truncated before index \(M\), the entire omitted outer series satisfies
\[
\left|W_\lambda(t)-\sum_{n=0}^{M-1}(-\lambda)^nW_n(t)\right|
\le
@@ -1240,17 +1240,17 @@ W_n(t)=t^{1+n\alpha}
\left[
\frac{\theta}{\Gamma(2+n\alpha)}
+(V_0-\theta)\sum_{m=0}^{\infty}
-\frac{(-\lambda_\xi t^{A_0})^m}
-{\Gamma(2+n\alpha+mA_0)}
+\frac{(-\lambda_\xi t^{\alpha_0})^m}
+{\Gamma(2+n\alpha+m\alpha_0)}
\right].
\tag{C.32}
\]
-When \(w=\lambda_\xi t^{A_0}<1\), every inner denominator also has argument at least two, and its tail before index \(L\) is bounded by
+When \(w=\lambda_\xi t^{\alpha_0}<1\), every inner denominator also has argument at least two, and its tail before index \(L\) is bounded by
\[
\left|\sum_{m=L}^{\infty}
-\frac{(-w)^m}{\Gamma(2+n\alpha+mA_0)}\right|
+\frac{(-w)^m}{\Gamma(2+n\alpha+m\alpha_0)}\right|
\le \frac{w^L}{1-w}.
\tag{C.33}
\]
@@ -1284,13 +1284,13 @@ The producer evaluates \(W_n\) through the strict power-field moment divided by
This appendix works in the variables of A.1. The stored function \(\bar G\) approximates \(D^\alpha Z=\nu g\), so its coefficients are \(\nu\) times those of (4.1)–(4.2).
-For each candidate \(\alpha=\beta\in\{.52,.6,.9\}\) and each \(u=n/8,\ n=0,\ldots,512\), we store
+For each candidate \(\alpha=\beta\in\{.52,.6,.9\}\) and each stored frequency \(u=n/8\), we store
\[
\bar G(t)=\nu c_0+A_1t^\beta+A_2t^{2\beta}+\bar L(t),\quad
c_0=-(u^2+1/4)/2,\quad \widehat Z=I^\alpha\bar G,\quad T=1/2.
\tag{D.1}
\]
-Here \(\bar L\) is the continuous piecewise-linear interpolation in physical time at nodes \((t_j,L_j)\), with \(L_0=0\). Nodes and complex coefficients are interpreted as the exact dyadic rationals represented by their binary64 values. The following independent residual controls the continuous-time error of this reference trajectory:
+The fields for \(\alpha=.52\) and \(.6\) store \(n\le1024\) on 2049 time knots and the field for \(\alpha=.9\) stores \(n\le640\) on 1025 knots; residual banks certify \(n\le512\) for all three and, for \(\alpha=.52\), also \(512a^+\), then \(c
Displayed endpoints are outward decimal roundings of the exact rational bounds. The unique minimiser of the model objective is \(\alpha=0.52\), with a gap of at least 0.0000002702052 to its competitors, and the stored rational output has the same unique minimiser.
-**Table E.3. Three-candidate quote-band objective and consistency decisions, using the original twelve quotes.**
+**Table E.3. Three-candidate quote-band objective and consistency decisions, using the twelve quotes of Section 7.1.**
| α | Rigorous model J_band interval | Rows strictly excluding quote consistency | All-row quote consistency certified |
|---|---|---|---|
@@ -1780,13 +1780,13 @@ Table E.13 lists the proof obligations, the reader of each, what it recomputes a
| ID | Obligation | Verification command | Recomputed | Inherited / shared |
| --- | --- | --- | --- | --- |
-| S1 | Padé coefficient signs on the original parameter rectangle | `full_structure_verify.py` ([S1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#s1)) | 211241 structural sign leaves and rectangle cover | model/parameter definitions; original strict interval algebra and generalised-power primitives |
-| S2 | Startup Caputo residual and left-halfplane enclosure | `startup_independent.py` ([S2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#s2)) | all 513 frequencies for each of the three original startup cells | stored field dyadics and later-time continuous residuals; Gamma/power/dyadic outward primitives |
-| R1 | Full original alpha=.52 low-frequency continuous residual cover | `check-time-envelope.py` ([R1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#r1)) | 8189 closed cells ×513 entries: cover, saved dyadic values and per-node maxima; startup audited separately | derivative inequalities in identified continuous generator; NumPy enclosure with proved dot-product error and strict scalar primitives |
-| R2 | Full alpha=.52 high-frequency continuous residual cover | `omission-verify.py` ([R2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#r2)) | 8189 closed cells ×512 entries: exact bank, startup/halfplane, saved-node maxima | identified Caputo derivative-generator inequalities; strict scalar primitives and original continuous generator |
+| S1 | Padé coefficient signs on the parameter rectangle of Proposition 6.1 | `full_structure_verify.py` ([S1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#s1)) | 211241 structural sign leaves and rectangle cover | model/parameter definitions; strict interval algebra and generalised-power primitives |
+| S2 | Startup Caputo residual and left half-plane enclosure | `startup_independent.py` ([S2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#s2)) | all 513 frequencies for each of the three startup cells | stored field dyadics and later-time continuous residuals; Gamma/power/dyadic outward primitives |
+| R1 | Full alpha=.52 low-frequency continuous residual cover | `check-time-envelope.py` ([R1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#r1)) | 8189 closed cells ×513 entries: cover, saved dyadic values and per-node maxima; startup audited separately | derivative inequalities in identified continuous generator; NumPy enclosure with proved dot-product error and strict scalar primitives |
+| R2 | Full alpha=.52 high-frequency continuous residual cover | `omission-verify.py` ([R2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#r2)) | 8189 closed cells ×512 entries: exact bank, startup/half-plane, saved-node maxima | identified Caputo derivative-generator inequalities; strict scalar primitives and continuous generator |
| P1 | Global finite-history and low-frequency 128-bin price accounts | `verify-finite-history-independent.py`; `independent-time-local.py` ([P1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#p1)) | forward masses, complete cell intersections, 513 tightened radii, all 1025 finite terms and complete spread remainder | continuous residual theorem/validity and identified stored field; strict forward-moment, exponent and true-tail primitives |
-| P2 | Expanded-reference frozen half-year output account | `independent-omission.py` ([P2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#p2)) | changed reference centre, all 1025 support terms, old/new complete interval intersection, global/local high-frequency accounts | same actual fast output and continuous residual banks; original strict moment/scalar primitives |
-| Q1 | Quarter-maturity used 64 certificate and actual fast output | `independent-transfer.py` ([Q1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#q1)) | all three exact terminal interpolations, 1539 restricted exponents, three original fast-output calculations, all 3075 true finite CF/tail/price terms | larger-horizon continuous residuals and halfplane certificate; strict moment/trig/dyadic/true-tail libraries; original fast implementation used only for output replay |
+| P2 | Expanded-reference frozen half-year output account | `independent-omission.py` ([P2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#p2)) | changed reference centre, all 1025 support terms, intersection of both complete intervals, global/local high-frequency accounts | same actual fast output and continuous residual banks; strict moment/scalar primitives |
+| Q1 | Quarter-maturity used 64 certificate and actual fast output | `independent-transfer.py` ([Q1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#q1)) | all three exact terminal interpolations, 1539 restricted exponents, three fast-output calculations, all 3075 true finite CF/tail/price terms | larger-horizon continuous residuals and half-plane certificate; strict moment/trig/dyadic/true-tail libraries; fast implementation used only for output replay |
| Q2 | Quarter full 128 global and 128-bin complete certificate | `independent-transfer-supplement.py` ([Q2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#q2)) | all 1025 restricted exponents, all 1025 CF/tail/coefficient/radius terms, all high bank entries, all 128 local weights and exact account | low/high continuous derivative validity; actual fast replay is separately performed by Q1; strict moment/trig/dyadic/true-tail libraries; path helpers of Q1 |
| N1 | Fresh fields/residuals at the five nearby alpha points | `nearby_independent.py` ([N1](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#n1)) | all cells, startup, halfplane, 513 exponents and 1025 CF/coefficient terms per point; all objective/pair decisions | strict proof of continuous residual generator and final propagation theorem; declared sdk outward primitives, stored reference data |
| N2 | Descriptive 128-bin reused-bank control, outside nearby objective grid | `local_output_independent.py` ([N2](https://github.com/130U/RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston/blob/main/COMMANDS.md#n2)) | 128 exact positive weights, every intersecting closed-bank maximum, 513 tightened radii, all 12 prices and all actual output rounding | complete identity-matched N2048 point proof and continuous residual generator; same strict moment/trig/dyadic primitives |
diff --git a/website/content/visible-copy-manifest.json b/website/content/visible-copy-manifest.json
index f23551c..035e5c6 100644
--- a/website/content/visible-copy-manifest.json
+++ b/website/content/visible-copy-manifest.json
@@ -6,9 +6,9 @@
],
"routes": {
"/": {
- "textSha256": "47931fd0032d04b503c560475b9646dec87cffdba3d187b6cedc325dbd2aac2a",
- "characterCount": 4049,
- "wordCount": 622,
+ "textSha256": "403ababdf4e9e92a4a2b9944c2b5dba17abbd85ea008483b8a265aa6af669da4",
+ "characterCount": 4063,
+ "wordCount": 623,
"preview": "Skip to Main Content Theodore Ouyang Duke Alumnus Sequoia Fellow Theodore Ouyang About Experience Blog Thinking Mathematically Working in AI Mathematics is how ",
"altTexts": [],
"ariaLabels": [
@@ -17,8 +17,7 @@
"Read more about Critical Dependence Stress for Diversification Gains under Fixed Marginals",
"Read more about Certified Valuation of Arithmetic Asian Options: The Cost of Exercising on the Geometric Average",
"Read more about Certified Joint Pricing Errors in Rough Heston",
- "Read more about coursework",
- "Contact"
+ "Read more about coursework"
],
"metadata": {
"title": "Theodore Ouyang",
@@ -54,9 +53,9 @@
}
},
"/critical-dependence-stress-for-diversification-gains-under-fixed-marginals/": {
- "textSha256": "dce47af3082d6d59fbac301503a9bf724bc4c1f89ff356848ff4312d9618b4df",
- "characterCount": 77443,
- "wordCount": 11711,
+ "textSha256": "a2529a73260af1ab27c2f8049a1958867237a58397f47546bb60809f8b7ddd28",
+ "characterCount": 77456,
+ "wordCount": 11710,
"preview": "Skip to Main Content Theodore Ouyang Back to Selected Research Critical Dependence Stress for Diversification Gains under Fixed Marginals GitHub Full Manuscript",
"altTexts": [],
"ariaLabels": [
@@ -432,9 +431,9 @@
"formulaCount": 1025
},
"/certified-valuation-of-arithmetic-asian-options/": {
- "textSha256": "dffbd2ad109d48dff82c091362e7044b7a6f6b8b93d234ee029b4a41d9afd909",
- "characterCount": 69624,
- "wordCount": 11483,
+ "textSha256": "635872832a204fc7c8d23fa0a5bea1a838d410ced91bfee782410c317e48cf12",
+ "characterCount": 69689,
+ "wordCount": 11493,
"preview": "Skip to Main Content Theodore Ouyang Back to Selected Research Certified Valuation of Arithmetic Asian Options: The Cost of Exercising on the Geometric Average ",
"altTexts": [],
"ariaLabels": [
@@ -702,13 +701,13 @@
"twitterTitle": "Certified Valuation of Arithmetic Asian Options: The Cost of Exercising on the Geometric Average",
"twitterDescription": "The cost of exercising an arithmetic Asian option on the geometric average: a proved bound, its cubic law in Black's model, and a certified Heston price interval with its Euler discretization error."
},
- "mathSourceSha256": "03f6d5bc1411109aef5f3063d00088f9b30fb3faaf48bf5044dae5b0b1a7da90",
+ "mathSourceSha256": "6b27445e28276f9adfcb080776f1c1a8497f69399741a22969195b159b0d7f05",
"formulaCount": 751
},
"/certified-joint-pricing-errors-in-rough-heston/": {
- "textSha256": "43229e7783d999a704b9176d6c2e76bf50c18d05c7cf3cda8b71c258c51d62f4",
- "characterCount": 65363,
- "wordCount": 10578,
+ "textSha256": "48aad8fd769e9aabe90491e89fe4320328ac6bfc02549d4deaef091c0955efd6",
+ "characterCount": 65645,
+ "wordCount": 10615,
"preview": "Skip to Main Content Theodore Ouyang Back to Selected Research Certified Joint Pricing Errors in Rough Heston GitHub Full Manuscript Abstract Every option on on",
"altTexts": [],
"ariaLabels": [
@@ -791,7 +790,7 @@
"Scrollable mathematical expression",
"Scrollable mathematical expression",
"Zoom in",
- "The weight on a unit Riccati residual against the time to maturity at which the residual occurs, over a six-month horizon. Both weights are forward variance curves: the curve under the pricing measure falls from about 0.040 six months before maturity to 0.0262 at maturity, and the curve under the half-share measure lies below it, by about 0.003 at six months.",
+ "The weight on a unit Riccati residual against the time to maturity at which the residual occurs, over a six-month horizon. Both weights are forward variance curves: the curve under the pricing measure rises from 0.0262 for a residual six months before maturity to about 0.040 for a residual at maturity, and the curve under the half-share measure lies below it, by about 0.003 at maturity.",
"Mathematical equation",
"Scrollable mathematical expression",
"Scrollable mathematical expression",
@@ -908,13 +907,13 @@
"twitterTitle": "Certified Joint Pricing Errors in Rough Heston",
"twitterDescription": "Numerical pricing error as a common factor across strikes: a Riccati residual priced by forward variance in rough Heston, the netting of portfolio errors in strike space, and certified tolerances for S&P 500 option positions."
},
- "mathSourceSha256": "5230e304025818d361f0b5c67d1c055ff1606d65b3741dd15b41a5baadb1e65a",
- "formulaCount": 457
+ "mathSourceSha256": "ee3203993173a45a329f04da7271a14d1efce66641690ac0e2197fefcd698dce",
+ "formulaCount": 460
},
"/experience/": {
- "textSha256": "945b4c95b04abe74dc1308e7a19aa424a8c562f73f41bcd31033017295bf9cfa",
- "characterCount": 10270,
- "wordCount": 1515,
+ "textSha256": "8b76b2d82047a27707a4b5ee80f88ef01d7018c7548b37c7aaefbe42f920e3b0",
+ "characterCount": 10388,
+ "wordCount": 1536,
"preview": "Skip to Main Content Theodore Ouyang About Experience Blog Experience: Areas of Focus AI Research and Engineering Data Science Legal Research and Policy Analysi",
"altTexts": [],
"ariaLabels": [
@@ -952,9 +951,9 @@
}
},
"/experience/large-scale-domain-expert-feedback-governance-and-adaptive-review/": {
- "textSha256": "a787e056f3f3efaf74652e6ede9b13132c6e0763e9fe649c715a51df8829aa58",
- "characterCount": 36199,
- "wordCount": 5726,
+ "textSha256": "d4871101056f03c2211d856816f6c25e9ec6d497b02f5526dec5b5f009795615",
+ "characterCount": 36373,
+ "wordCount": 5766,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Large-Scale Domain Expert Feedback Governance and Adaptive Review Organization Duke University × Leading",
"altTexts": [],
"ariaLabels": [
@@ -965,6 +964,7 @@
"Scrollable mathematical expression",
"Scrollable mathematical expression",
"Scrollable mathematical expression",
+ "Scrollable mathematical expression",
"Article contents",
"Scrollable mathematical expression",
"Mathematical equation",
@@ -1011,7 +1011,6 @@
"Mathematical equation",
"Mathematical equation",
"Mathematical equation",
- "Scrollable mathematical expression",
"Zoom in",
"Research data table",
"Zoom in",
@@ -1043,13 +1042,13 @@
"twitterTitle": "Duke University × Leading Foundation Model Developer: Large-Scale Domain Expert Feedback Governance and Adaptive Review",
"twitterDescription": "One pool of expert hours has to do two jobs in a post-training program: tell the team how good a batch is, and fix the data that ships."
},
- "mathSourceSha256": "927bfca124668d7ba957f9597b45f5a43746778c69cfd223660d38e52c1ab0bc",
- "formulaCount": 196
+ "mathSourceSha256": "ca24e7be57fa920da39b71863ba66b48d9aaea29f83f003e9c88617a4c3da6ef",
+ "formulaCount": 195
},
"/experience/financial-task-design-and-synthetic-data-engineering/": {
- "textSha256": "ff48998f283491592177734f230b8e740cc4460e67268fae4a683179d3bdad46",
- "characterCount": 34092,
- "wordCount": 5020,
+ "textSha256": "01c0d49a5ee859bd813d8ee328af4c6edd4df00a9fe1f8d15e73e99a711db4d0",
+ "characterCount": 34124,
+ "wordCount": 5023,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Financial Task Design and Synthetic Data Engineering Organization Duke University × Leading Foundation M",
"altTexts": [],
"ariaLabels": [
@@ -1118,13 +1117,13 @@
"twitterTitle": "Duke University × Leading Foundation Model Developer: Financial Task Design and Synthetic Data Engineering",
"twitterDescription": "When a synthetic financial case changes one number, its reference answer, the evidence behind that answer and the expert's preference can all go stale together."
},
- "mathSourceSha256": "654eaabaa94a670c74308bd0061b863602223aea3be555923a63f3dda7772454",
- "formulaCount": 187
+ "mathSourceSha256": "9ac614bce86276bf0e66de1376e49662e28b96c4699acfb79d1eb18c9dfbc35f",
+ "formulaCount": 189
},
"/experience/general-reasoning-evaluation-and-layered-verification/": {
- "textSha256": "f9f938c48eef1ba61e1e583e98ebb39457414a1fd2895a91b0ed7c30129c7c0a",
- "characterCount": 33444,
- "wordCount": 5072,
+ "textSha256": "fa1eb0864a9b2717dc26f3430dec7bff463575358886660f93a5e5cb6fe90426",
+ "characterCount": 33426,
+ "wordCount": 5067,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience General Reasoning Evaluation and Layered Verification Organization Duke University × Frontier AI Researc",
"altTexts": [],
"ariaLabels": [
@@ -1199,7 +1198,7 @@
"Zoom in",
"Research data table",
"Zoom in",
- "Final-answer matching: 1,200 valid accepted, 1,800 invalid accepted, 600 invalid rejected; 40.0% of accepted submissions are valid. Strict reference-path matching: 600 valid accepted, 600 valid rejected, 2,400 invalid rejected. Semantic certificate verification: 1,200 valid accepted, 2,400 invalid rejected.",
+ "Final-answer matching: 1,200 valid accepted, 1,800 invalid accepted, 600 invalid rejected; 40.0% of accepted submissions are valid. Strict reference-path matching: 600 valid accepted, 600 valid rejected, 2,400 invalid rejected. Certificate verification: 1,200 valid accepted, 2,400 invalid rejected.",
"Scrollable mathematical expression",
"Mathematical equation",
"Mathematical equation"
@@ -1214,13 +1213,13 @@
"twitterTitle": "Duke University × Frontier AI Research Laboratory: General Reasoning Evaluation and Layered Verification",
"twitterDescription": "A model can reach the right answer through an unsupported step, and a grader that matches a reference solution will reject a correct proof that takes another route."
},
- "mathSourceSha256": "2dd747e1fe938074f86aa7fd48b4f064fdddb77c187af5871b94fc2ab2c21345",
- "formulaCount": 205
+ "mathSourceSha256": "c5b5770fbd85e3ce69f1f8a28ecd0192c581aabbfa5f7640168ee12f5fde3aae",
+ "formulaCount": 203
},
"/experience/evidence-verification-and-robust-comparison-in-investment-research/": {
- "textSha256": "19b4fba1f4ca8abc7f30a9b8e6c9c50eb567edbf624da00a20660c28b9494f90",
- "characterCount": 29858,
- "wordCount": 4633,
+ "textSha256": "ea2e2e88b40c336c7fbe39f6ec249304df69f22bd9bd17b241d9a791fc16efd4",
+ "characterCount": 29891,
+ "wordCount": 4638,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Evidence Verification and Robust Comparison in Investment Research Organization Duke University × Leadin",
"altTexts": [],
"ariaLabels": [
@@ -1228,6 +1227,7 @@
"Parent navigation",
"Scrollable mathematical expression",
"Scrollable mathematical expression",
+ "Scrollable mathematical expression",
"Article contents",
"Scrollable mathematical expression",
"Mathematical equation",
@@ -1299,13 +1299,13 @@
"twitterTitle": "Duke University × Leading Global Alternative Asset Manager: Evidence Verification and Robust Comparison in Investment Research",
"twitterDescription": "Due diligence has more facts to check than analyst time to check them."
},
- "mathSourceSha256": "8b2ba1f0a5a5ea95169b8861f4ef6e2d236814e9d4b0e3de6866a1187f8cbc6a",
+ "mathSourceSha256": "9b8be759a972114f3f850b35bb9966a2d52a5a25424d78e5661e0174a3a8e7a4",
"formulaCount": 214
},
"/experience/autonomous-authority-in-space/": {
- "textSha256": "b18750c059b98fb1765f2d9f985488e2561f2af9dcb05cb940737c26e3bba157",
- "characterCount": 78817,
- "wordCount": 11134,
+ "textSha256": "80649e4992bc0665e3968d9cc351a849da5e491f93e654d68de5bb024d7e831b",
+ "characterCount": 79482,
+ "wordCount": 11245,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Autonomous Authority in Space: Risk Tradeoffs and the Law of Delegation Organization Duke Law School Pos",
"altTexts": [],
"ariaLabels": [
@@ -1422,9 +1422,9 @@
}
},
"/experience/small-states-and-the-governance-of-strategic-space-dependence/": {
- "textSha256": "1218e6d95582fc5077f161f0331bf1a4d54710fe0cb6d0e0db247abf6892d7d6",
- "characterCount": 62882,
- "wordCount": 8720,
+ "textSha256": "90661d6fd0d2d8534588a2f3cf6e333dbbb8350a3a48bfd674c18f5448937cf4",
+ "characterCount": 63113,
+ "wordCount": 8779,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Small States and the Governance of Strategic Space Dependence Organization Duke Law School Position Auth",
"altTexts": [],
"ariaLabels": [
@@ -1540,9 +1540,9 @@
}
},
"/experience/who-may-choose-the-lesser-risk/": {
- "textSha256": "e8b1fd47670da6b890b3dccb34811782e87666a169914024e32e4fd86c0a9a34",
- "characterCount": 61695,
- "wordCount": 8657,
+ "textSha256": "cf73b434706937c92806e99a95b0237ea90011dd2d9f975259bb53a0cef81736",
+ "characterCount": 62183,
+ "wordCount": 8731,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Who May Choose the Lesser Risk: Solar Geoengineering and the Legal Duties of Comparison and Continuity O",
"altTexts": [],
"ariaLabels": [
@@ -1667,9 +1667,9 @@
}
},
"/experience/mangrove-restoration-and-the-limits-of-compensatory-mitigation/": {
- "textSha256": "92d8dc08bdbcec4fae150c3dbfc598f516438736b5fc6b308a39fe73f46c9eeb",
- "characterCount": 77124,
- "wordCount": 10848,
+ "textSha256": "1e3dc43948e467e9e4a3c771cbefba006e04d8fb745667f7480c9623bd1e4c71",
+ "characterCount": 78724,
+ "wordCount": 11100,
"preview": "Skip to Main Content Theodore Ouyang Back to Experience Mangrove Restoration and the Limits of Compensatory Mitigation: Lessons from Florida for the Greater Bay",
"altTexts": [],
"ariaLabels": [
@@ -2070,9 +2070,9 @@
}
},
"/blog/ten-states-of-water/": {
- "textSha256": "4fd58afd8f94557b477596d805ea0db30c5828cf46cf8752f9da80dae3119535",
- "characterCount": 2691,
- "wordCount": 496,
+ "textSha256": "725ec816a23bc7a4d0e47f86ee22eea3e147c786d403145e34f349d38197032c",
+ "characterCount": 2646,
+ "wordCount": 486,
"preview": "Skip to Main Content Theodore Ouyang Back to Blog Ten States of Water By Theodore Ouyang November 8, 2026 Ten States of Water Ten landscapes of Alaska and Yello",
"altTexts": [
"The cover of Ten States of Water: the name Theodore Ouyang written across the sky above a snowy ridge, a sunset horizon, and winding tidewater",
diff --git a/website/public/assets/blog/the-etude-sky/chart.jpg b/website/public/assets/blog/the-etude-sky/chart.jpg
index 6960b9d..859d1c9 100644
Binary files a/website/public/assets/blog/the-etude-sky/chart.jpg and b/website/public/assets/blog/the-etude-sky/chart.jpg differ
diff --git a/website/scripts/check-browser-layout.mjs b/website/scripts/check-browser-layout.mjs
index 8dc98a1..88dd7b7 100644
--- a/website/scripts/check-browser-layout.mjs
+++ b/website/scripts/check-browser-layout.mjs
@@ -55,8 +55,8 @@ function inspectFigure(figure) {
const bounds = rect(frame);
const issues = [];
const shown = element => {
- if (!element.checkVisibility({ checkVisibilityCSS: true, checkOpacity: true }) || element.closest('[hidden], mjx-assistive-mml, script, style')) return false;
- // Accessibility MathML remains in the layout tree even when CSS clips its
+ if (!element.checkVisibility({ checkVisibilityCSS: true, checkOpacity: true }) || element.closest('[hidden], math, script, style')) return false;
+ // Visually hidden text remains in the layout tree even when CSS clips its
// entire box. Range rectangles from that hidden copy are not visible ink.
for (let ancestor = element; ancestor && ancestor !== frame.parentElement; ancestor = ancestor.parentElement) {
const style = getComputedStyle(ancestor);
@@ -104,7 +104,7 @@ function inspectFigure(figure) {
const walker = document.createTreeWalker(frame, NodeFilter.SHOW_TEXT);
for (let node = walker.nextNode(); node; node = walker.nextNode()) {
const element = node.parentElement;
- if (!node.textContent.trim() || !shown(element) || element.closest('svg, mjx-assistive-mml')) continue;
+ if (!node.textContent.trim() || !shown(element) || element.closest('svg, math')) continue;
const range = document.createRange();
range.selectNodeContents(node);
for (const fragment of range.getClientRects()) {
diff --git a/website/scripts/import-dependence-research.mjs b/website/scripts/import-dependence-research.mjs
index ec40ff8..cc1870b 100644
--- a/website/scripts/import-dependence-research.mjs
+++ b/website/scripts/import-dependence-research.mjs
@@ -13,8 +13,8 @@ async function writeOrCheck(url, content) {
else await writeFile(url, content);
}
const manifest = JSON.parse(await readFile(new URL("source-manifest.json", sourceDirectory), "utf8"));
-const commit = "a62cc41106e58e5d2df1387a1cf39ebfa65e36e8";
-const paperBlob = "e9efa283fc4b8efe2cac0ca483fa3e5da81e01e5";
+const commit = "c0782609a95f9887c248631c56174615233c7ae4";
+const paperBlob = "72d5cf95d6812da8198cca12f313c59d254ef3d0";
assert.equal(manifest.commit, commit);
assert.equal(manifest.repository, "130U/RESEARCH-Critical-Dependence-Stress-for-Diversification-Gains-under-Fixed-Marginals");
const repositoryUrl = `https://github.com/${manifest.repository}`;
@@ -60,7 +60,6 @@ function braced(text, start, open = "{", close = "}") {
if (text[index] === "\\") { index += 1; continue; }
if (text[index] === open) depth += 1;
if (text[index] === close && --depth === 0) return { value: text.slice(start + 1, index), end: index + 1 };
- if (text[index] !== close) continue;
}
throw new Error(`Unclosed ${open}: ${text.slice(start, start + 90)}`);
}
@@ -251,7 +250,7 @@ for (const [index, match] of [...references.matchAll(/\\bibitem\{([^}]+)\}([\s\S
referenceParagraphs.push(`\n\n${index + 1}. ${value}`);
}
assert.equal(referenceParagraphs.length, bibliographyKeys.length);
-converted += `\n\n## Appendices and Supporting Material\n\nThis page includes the complete abstract, five main sections, all four main-text figures, and references. Technical appendices and declarations are available in the [full manuscript](${sourceUrl}) and the [GitHub repository](${repositoryUrl}).\n\n## References\n\n${referenceParagraphs.join("\n\n")}\n`;
+converted += `\n\n## Appendices and supporting material\n\nThis page includes the complete abstract, five main sections, all four main-text figures, and references. Technical appendices and declarations are available in the [full manuscript](${sourceUrl}) and the [GitHub repository](${repositoryUrl}).\n\n## References\n\n${referenceParagraphs.join("\n\n")}\n`;
converted = converted.replace(/@@DEPENDENCE-BLOCK-(\d+)@@/gu, (_, index) => blocks[Number(index)]);
const mathUses = mathematics.map(() => 0);
converted = converted.replace(/@@DEPENDENCE-MATH-(\d+)@@/gu, (_, rawIndex) => {
@@ -291,8 +290,7 @@ await writeOrCheck(new URL("provenance.json", sourceDirectory), JSON.stringify({
"Every inline/display mathematical body is preserved, apart from moving TeX label commands to HTML anchors; explicit equation tags are retained.",
"Original theorem numbers, proof endings, result-part numbering, bibliography order, citation targets, and complete figure captions are retained.",
"Aliases on one original target share one web anchor; references to omitted appendices point to their exact pinned TeX source line.",
- "The four original PNG siblings of the PDF figures are referenced at local asset paths, with their original Git blob identities recorded.",
- "The public page title differs from the original paper title, which is recorded separately as sourceTitle.",
+ "The four figures are drawn natively from the repository's formulas; their image paths key the native drawings and the complete captions are retained.",
],
}, null, 2) + "\n");
console.log(`${checkOnly ? "Verified" : "Imported"} ${headings.length} main headings, ${resultReceipts.length} numbered results, ${mathematics.length} formulas, ${emittedFigure} figures and ${bibliographyKeys.length} references.`);
diff --git a/website/scripts/import-selected-research.mjs b/website/scripts/import-selected-research.mjs
index cd5d70e..5d16f9d 100644
--- a/website/scripts/import-selected-research.mjs
+++ b/website/scripts/import-selected-research.mjs
@@ -7,8 +7,8 @@ const sourceDirectory = new URL("../content/sources/selected-research/", import.
const sources = [
{
key: "asian", slug: "certified-valuation-of-arithmetic-asian-options", repository: "RESEARCH-Certified-Valuation-of-Arithmetic-Asian-Options", file: "manuscript/report.md",
- sourceCommit: "41a814f6e5aa1eddddb4412e57ce68ec506c323e", sourceBlob: "6f573b14c4c172e394688333b22726438bbed461",
- sourceEditedAt: "2026-11-19T03:29:58-05:00", fetchedAt: "2026-11-19", mainSections: 9,
+ sourceCommit: "dd4b147d822d45a4ca12a66294fae2a32cd252bb", sourceBlob: "73fd58c109bca144a8809e3fc9832e8ad258cae6",
+ sourceEditedAt: "2026-11-19T11:43:56-05:00", fetchedAt: "2026-11-19", mainSections: 9,
appendixAnchor: "appendix-a-finite-certification-of-the-weighted-second-moment",
description: "The cost of exercising an arithmetic Asian option on the geometric average: a proved bound, its cubic law in Black's model, and a certified Heston price interval with its Euler discretization error.",
figures: {
@@ -22,9 +22,10 @@ const sources = [
},
{
key: "rough", slug: "certified-joint-pricing-errors-in-rough-heston", repository: "RESEARCH-Certified-Joint-Pricing-Errors-in-Rough-Heston", file: "manuscript/merged-heston-en.md",
- sourceCommit: "46c85b4b256dfbc597ed56b3b2b526681334e60f", sourceBlob: "27bd396d79cfe3d99072f23d9799641c898442e7",
- sourceEditedAt: "2026-11-19T03:44:46-05:00", fetchedAt: "2026-11-19", mainSections: 8,
- appendixAnchor: "appendix-a-rescaled-variables-model-and-transform",
+ sourceCommit: "038db24a3aadf9600edef385dfd7508e612d53c8", sourceBlob: "12b074bc25984ea11f9fc7d75768800c267a43eb",
+ sourceEditedAt: "2026-11-19T11:45:08-05:00", fetchedAt: "2026-11-19", mainSections: 8,
+ // The manuscript is the TeX build source; the appendix link opens the GitHub-form article.
+ appendixFile: "ARTICLE.md", appendixAnchor: "appendix-a-rescaled-variables-model-and-transform",
description: "Numerical pricing error as a common factor across strikes: a Riccati residual priced by forward variance in rough Heston, the netting of portfolio errors in strike space, and certified tolerances for S&P 500 option positions.",
figures: {
directory: "rough", count: 4,
@@ -173,10 +174,11 @@ for (const config of sources) {
if (figures.length !== (config.figures?.count ?? 0) || /^!\[/mu.test(captioned.replace(/^!\[[^\n]*\]\(\/assets\/research\/[a-z0-9/-]+\.png\)$/gmu, ""))) {
throw new Error(`The ${config.key} manuscript's figures do not match their reviewed form.`);
}
+ const appendixUrl = `${repositoryUrl}/blob/${config.sourceCommit}/${config.appendixFile ?? config.file}#${config.appendixAnchor}`;
const markdown = [
captioned,
- "## Appendices and Supporting Material",
- `This page includes the complete main text and references. Appendices, supplementary proofs, code, and verification material are available in the [GitHub repository](${repositoryUrl}) and the [full manuscript](${sourceUrl}#${config.appendixAnchor}).`,
+ "## Appendices and supporting material",
+ `This page includes the complete main text and references. Appendices, supplementary proofs, code and verification material are available in the [GitHub repository](${repositoryUrl}) and the [full manuscript](${appendixUrl}).`,
refs,
].join("\n\n").replace(/\]\((\.\.\/[^\s)]+)\)/gu, (_, relative) => `](${repositoryUrl}/blob/${config.sourceCommit}/${relative.slice(3)})`) + "\n";
const content = {
diff --git a/website/scripts/integrate-architecture.mjs b/website/scripts/integrate-architecture.mjs
index ddb6097..db4831f 100644
--- a/website/scripts/integrate-architecture.mjs
+++ b/website/scripts/integrate-architecture.mjs
@@ -36,6 +36,29 @@ const readerRules = ` `;
+// The viewer returns to the site the way every page does: the Return Home link
+// with the site's return arrow, above the title.
+const returnLink = ' Return Home\n';
+
+function applyReturnLink(viewer) {
+ const cleared = viewer.replace(/ \n/u, "");
+ assert.ok(/