Curator's Take
AI Commentary
This article shows that the number of physical Pauli recovery gates needed for repetition‑code process tomography can be dramatically reduced—down to zero for three qubits and just five for seven—by exhaustively searching optimal decoder designs rather than relying on the conventional minimum‑weight rule. The resulting gate‑optimal decoders deliver up to an order‑of‑magnitude improvement in tomographic precision while cutting experimental runtime, a benefit that directly supports current efforts to benchmark and calibrate noisy intermediate‑scale quantum processors more efficiently. However, the advantage diminishes for larger codes where full decoding can even lose rank, underscoring that scaling these techniques will require careful conditioning of hardware and possibly new decoder heuristics.
— Mark Eatherly
Summary
Process tomography built on stabilizer codes compares the logical channel of a repetition code with a reference channel, and the decoder that produces that channel is almost always fixed to the standard minimum-weight one. We ask how many physical single-qubit Pauli gates the decoder actually needs when the measurement basis is fixed---the constraint under which the count is a resource rather than a bookkeeping artefact. Exhaustive enumeration of every recovery design for $n=3$, $5$ and $7$ qubits, with rank decided by exact modular arithmetic at two primes, gives minimal counts of $0$, $1$ and $5$ gates, against $6$, $10$ and $14$ for first-order decoding and $6$, $50$ and $308$ for full minimum-weight decoding. The count depends only on how many sites are corrected and never on which, so the two halves of the design problem separate. Correcting more sites is not monotonically better: at $n=7$ full decoding loses three dimensions of rank, and under a calibration-residual floor the precision curve at $n=3$ is U-shaped. The three-gate design is $9.5$ times more precise than the standard six-gate decoder at the median of $40$ sampled working points, and more precise at every one of them ($1.8$ to $88$ times); the size of that factor belongs to the working point, the ordering does not. Reaching a $10$\,mrad rms error then costs $1.2\times10^{7}$ shots in $65$\,s on a current superconducting device at a well-conditioned point, while $n\ge5$ stays out of reach for conditioning reasons rather than gate-count reasons. Finally, the optimal designs are not covariant under chain reversal: they separate the mirror pair of arXiv:2610.19779 by $1.3\times10^{-3}$ while leaving the Jacobian at full rank.