generate_e uses np.random.uniform to sample a continuous float e_i, checks the filter condition against that float, then stores int(e_i). When e_i lands in (-1, 1), int() truncates it to 0 and silently drops the error.
With concealment_rate=1 and default params (tau=39, filterthresh=39), the accepted range for e_i always contains (-1, 1), so about 2.6% of samples can end up with error=0 due to int() truncation. Confirmed with m=1,000,000: current implementation produced 25,423 zero errors (actual concealment 0.9746). For comparison, the expected number of zero errors is ~12,658 (actual concealment rate ~0.987 or 2tau/2tau+1).
Fix: Change error sampling from e_i = np.random.uniform(-4*tau, 4*tau) to e_i = np.random.randint(-4 * tau, 4 * tau + 1)
generate_euses np.random.uniform to sample a continuous float e_i, checks the filter condition against that float, then stores int(e_i). When e_i lands in (-1, 1), int() truncates it to 0 and silently drops the error.With concealment_rate=1 and default params (tau=39, filterthresh=39), the accepted range for e_i always contains (-1, 1), so about 2.6% of samples can end up with error=0 due to int() truncation. Confirmed with m=1,000,000: current implementation produced 25,423 zero errors (actual concealment 0.9746). For comparison, the expected number of zero errors is ~12,658 (actual concealment rate ~0.987 or 2tau/2tau+1).
Fix: Change error sampling from
e_i = np.random.uniform(-4*tau, 4*tau)toe_i = np.random.randint(-4 * tau, 4 * tau + 1)