From e95ed83f6ae7d41105cc778f9e3c15a524f7e8bf Mon Sep 17 00:00:00 2001
From: Johan Larsson <johan@jolars.co>
Date: Sun, 30 Aug 2026 12:46:09 +0200
Subject: [PATCH] refactor: use min tolerance for sortedl1

---
 solvers/sortedl1.py    | 22 +++++++++++++++++++++-
 tests/test_sortedl1.py | 30 ++++++++++++++++++++++++++++++
 2 files changed, 51 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_sortedl1.py

diff --git a/solvers/sortedl1.py b/solvers/sortedl1.py
index ed5f326..189e480 100644
--- a/solvers/sortedl1.py
+++ b/solvers/sortedl1.py
@@ -1,14 +1,34 @@
 from benchopt import BaseSolver, safe_import_context
-from benchopt.stopping_criterion import INFINITY
+from benchopt.stopping_criterion import INFINITY, SufficientProgressCriterion
 
 with safe_import_context() as import_ctx:
     import numpy as np
     from sortedl1 import Slope
 
 
+MIN_TOLERANCE = 1e-6
+
+
+class MinimumToleranceCriterion(SufficientProgressCriterion):
+    def __init__(self, min_tol=MIN_TOLERANCE, **kwargs):
+        super().__init__(**kwargs)
+        self.min_tol = min_tol
+        self.kwargs["min_tol"] = min_tol
+
+    def should_stop(self, stop_val, objective_list):
+        stop, status, next_stop_val = super().should_stop(
+            stop_val,
+            objective_list,
+        )
+        if not stop and next_stop_val < self.min_tol:
+            return True, "done", stop_val
+        return stop, status, next_stop_val
+
+
 class Solver(BaseSolver):
     name = "sortedl1"
     sampling_strategy = "tolerance"
+    stopping_criterion = MinimumToleranceCriterion()
     install_cmd = "conda"
     requirements = ["pip::sortedl1"]
     parameters = {
diff --git a/tests/test_sortedl1.py b/tests/test_sortedl1.py
new file mode 100644
index 0000000..c98733c
--- /dev/null
+++ b/tests/test_sortedl1.py
@@ -0,0 +1,30 @@
+import pytest
+
+from solvers.sortedl1 import MIN_TOLERANCE, Solver
+
+
+def test_tolerance_sampling_stops_at_minimum_tolerance():
+    solver = Solver.get_instance()
+    criterion = solver.stopping_criterion.get_runner_instance(
+        solver=solver,
+        max_runs=100,
+    )
+    stop_val = criterion.init_stop_val()
+    evaluated_tolerances = []
+    objective_values = []
+
+    while True:
+        evaluated_tolerances.append(stop_val)
+        objective_values.append(
+            {"objective_value": -float(len(evaluated_tolerances))}
+        )
+        stop, status, stop_val = criterion.should_stop(
+            stop_val,
+            objective_values,
+        )
+        if stop:
+            break
+
+    assert status == "done"
+    assert evaluated_tolerances[-1] == pytest.approx(MIN_TOLERANCE)
+    assert min(evaluated_tolerances) >= MIN_TOLERANCE