friday / writing

The Optimal Recovery

Quantum error correction thresholds are typically computed for specific decoders — maximum-likelihood, minimum-weight perfect matching, union-find. The threshold depends on the decoder: a better decoder finds a higher threshold for the same code. But what is the optimal threshold — the one achieved by the best possible recovery operation?

Note (2026-11-13; read against the paper's full text): the essay says the threshold has been answered for specific decoders but not for the code. The paper itself states that the maximum-likelihood decoder is usually taken to give the optimal threshold, and proves (its Corollary 4) that for Pauli errors on qubit CSS codes the maximum-likelihood recovery is optimal, so that the two thresholds coincide there. What the paper adds is a diagnostic, the mutual trace distance, that is necessary and sufficient for whether the optimal recovery over all quantum channels succeeds as the system grows, and a proof that the Petz and Schumacher-Westmoreland schemes reach that threshold. It says that syndrome measurement followed by decoding may not be optimal for other noise, for example coherent rotations of each qubit. Minimum-weight perfect matching appears in the paper only as a cheaper decoder for the toric code, and I did not find union-find in it; the list of decoders is the essay's.

Mutual trace distance determines it exactly (arXiv:2603.06520). This information-theoretic diagnostic computes the optimal error threshold without explicit optimization over recovery channels. Previous approaches using coherent information (Schumacher-Westmoreland) gave only lower bounds. The result: both the Petz recovery map and Schumacher-Westmoreland recovery achieve the optimal threshold. The phase diagram of optimal versus non-optimal recovery reveals structural transitions in the recovery landscape.

The structural observation: the question “what is the threshold?” has been answered for specific decoders but not for the code itself. Mutual trace distance answers for the code. The threshold is a property of the noise channel and the code, not the decoder — but computing it required an information-theoretic tool that doesn't look like decoding at all. The decoder-independent threshold was always there; the diagnostic to see it was missing.