no. 16 · Tuesday 17 November 2026
Correcting quantum errors one gradient step at a time
Abstract
Probabilistic approximate quantum error correction (PAQEC) lets a recovery fail with some probability, in exchange for higher fidelity when it succeeds. Prior PAQEC code searches keep the codewords orthonormal with penalty terms and tune the code without its best recovery. We propose a gradient method with three new parts. First, the codewords move on the complex Stiefel manifold, so they are orthonormal at every step without penalties. Second, every code is trained against its best recovery, which we solve for at each step rather than learn. Third, a fast Petz recovery that we can differentiate even for symmetric codes gets the search close, and the exact semidefinite program finishes it. We train a single five-qubit code over generalized amplitude damping noise. At weak noise ($γ= 0.1$, $p = 0.05$) and equal success probability, it has $23$ times lower infidelity than the permutation-invariant code of prior work, and $9$ times lower than that work's numerically optimized code. At equal fidelity it succeeds $91\%$ of the time, against $76\%$ for the permutation-invariant code.
- fields
- Numerical Analysis