Decoding a quantum error-correcting code is a problem of statistical inference: a syndrome is a noisy measurement, and choosing a correction amounts to estimating the most probable logical class under the device's error distribution, so a decoder is only as good as the channel model it assumes. The dominant decoders—minimum-weight matching, belief propagation, and learned decoders—all approximate the minimum-weight coset leader, which is the maximum-likelihood rule only for independent and identically distributed (i.i.d.) noise; this shared target renders eight otherwise dissimilar decoders statistically indistinguishable on a distance-\(d\) repetition code. This work introduces a channel-adaptive spectral decoder that represents the structured error channel by a truncated transfer operator—an inhomogeneous first-order Markov (classical matrix-product) surrogate on the code's defect graph—estimated from a small calibration sample, and selects the maximum-a-posteriori coset. The decoder is Bayes-optimal for the structured channel, provably reduces to the leader under i.i.d. noise, and reports the exact model posterior as a confidence that is calibrated by construction. On a unified benchmark (distance 9; 100,000 pooled shots per decoder; five structured-noise regimes) it reduces the logical error of the i.i.d.-optimal leader by 44.0% under bias, 33.0% under bursts, and 21.8% under neighbour correlations, driving its pooled logical error to 0.1291 against 0.1508–0.1542 for the eight structure-agnostic decoders, while tying the leader honestly where no single-shot advantage exists.
1Background and Motivation
A quantum error-correcting code protects information only as well as its decoder, which must map each measured syndrome to the correct recovery under the structured noise characteristic of hardware: Pauli bias, spatial bursts, neighbour correlations, and faulty syndrome read-out.
Topological codes—Kitaev's toric and surface codes—are the leading candidates for scalable fault tolerance, and recent hardware has shown that scaling such a code can push the logical error rate below the physical one (Google Quantum AI, Nature 2023). The decoder is the component that converts a stabilizer syndrome into a correction, and it is the layer of the stack where classical combinatorial algorithms and, increasingly, machine learning meet the physics of the device. Three algorithmic families dominate: minimum-weight perfect matching (MWPM), the production standard, built on Edmonds' blossom algorithm; belief propagation with ordered-statistics decoding, which underpins quantum LDPC decoding; and learned decoders, where neural networks and reinforcement learning have recently surpassed matching-based accuracy on real hardware syndromes (Bausch et al., Nature 2024).
Across nearly all of this work the main metric is a single number, the logical error rate, almost always measured under independent-and-identically-distributed (i.i.d.) depolarizing or bit-flip noise. Two facts argue for a richer accounting. First, hardware noise departs from i.i.d.—it is structured, and a decoder that excels on i.i.d. noise may forfeit any advantage under structure; settling the question requires testing one decoder family across regimes on a common footing. Second, a decoder inside a fault-tolerant control loop may decline to commit to a given syndrome: it can abstain, flagging a shot as unreliable so that a slower fallback decoder, a repeated measurement, or a higher-level protocol takes over. Abstention is the QEC analogue of selective prediction in machine learning, and it is trustworthy only when the confidence gating it is calibrated. Both axes—coverage and calibration—remain routinely absent from decoder benchmarks.
What unites the incumbent decoders is their objective: each is engineered to approximate the minimum-weight coset leader, which is the maximum-likelihood recovery only when the channel is i.i.d. On a distance-\(d\) repetition code, a family of otherwise dissimilar decoders—optimal lookup, majority, matching, belief propagation, a learned neural coset selector, a topology decoder, a confidence-fused decoder, and a reinforcement-learning selector—turn out to be statistically indistinguishable on raw logical error. This is not a ceiling on achievable accuracy but the signature of a shared blind spot: every one targets the same weight-based leader and so collapses, on the line, to the same decision rule. The component that breaks it represents the structured channel by a truncated transfer operator—an inhomogeneous first-order Markov (classical matrix-product) model whose log-likelihood is a sum of local terms read along the chain—estimated from a small calibration sample of the operating channel, and selects the maximum-a-posteriori coset. It is Bayes-optimal for the structured channel and provably reduces to the leader under i.i.d. noise, so it claims no advantage where none exists.
The open problem, and our answer. The dominant decoders—matching, belief propagation, learned decoders—are all built to approximate the minimum-weight coset leader, which is the maximum-likelihood rule only for i.i.d. noise. We prove this is exactly why a careful family of eight such decoders is statistically indistinguishable on a repetition code: they share one i.i.d. target. The missing ingredient is channel-awareness. We introduce a channel-adaptive spectral decoder that models the structured channel by a truncated transfer operator on the defect graph and returns the maximum-a-posteriori coset; it is Bayes-optimal for the structured channel, breaks the tie by 20–44% under structured noise, and—reported with equal emphasis—ties the leader honestly where no single-shot advantage exists.
2Mathematical Preliminaries
Work over the binary field \(\mathbb{F}_2=\{0,1\}\) with addition \(\oplus\) (XOR) and Hamming weight \(|x|=\sum_i x_i\). Fix a distance \(d\ge 2\).
The code and its defect graph
The distance-\(d\) repetition code has data space \(\mathbb{F}_2^{d}\) and parity-check matrix \(H\in\mathbb{F}_2^{(d-1)\times d}\) with \((H_i)_i=(H_i)_{i+1}=1\). The syndrome of an error \(e\in\mathbb{F}_2^d\) is the boundary map
The logical functional is \(L:\mathbb{F}_2^d\to\mathbb{F}_2,\; L(e)=\big(\sum_i e_i\big)\bmod 2\), and the stabilizer group is \(\mathcal{S}=\ker H=\{0,\mathbf{1}\}\). A correction \(\hat c\) applied to \(e\) succeeds iff the residual \(r=e\oplus\hat c\) satisfies \(Hr=0\) and \(L(r)=0\); it is a logical failure otherwise. \(H\) is the boundary operator of the one-dimensional chain complex whose 1-cells are the \(d\) data bits and 0-cells the \(d-1\) links; lit checks \(J=\{i:s_i=1\}\) mark the endpoints of error chains, so decoding is endpoint pairing on a line.
The Bayes-optimal estimator depends on the channel
Under the i.i.d. bit-flip channel \(\Pr(e)=p^{|e|}(1-p)^{d-|e|}\) with \(p<\tfrac12\), the optimal (maximum-likelihood) decoder outputs a minimum-Hamming-weight representative of the fibre \(H^{-1}(s)\):
But for a general channel \(P\) the sufficient statistic is probability, not weight. The repetition coset has exactly two members \(\{c_0,\,c_0\oplus\mathbf 1\}\) of opposite logical value, so the channel-matched maximum-a-posteriori decoder returns the more probable one, and we model \(P\) by a truncated transfer operator on the defect graph — an inhomogeneous first-order Markov / matrix-product law:
The per-site \(2\times2\) transfer matrices \(T_i\) are estimated from a calibration sample of the operating channel; their position dependence captures spatial bias and their off-diagonal mass captures spatial correlation. Under i.i.d. noise the \(T_i\) are uniform and this rule reduces to the coset leader; under structured noise it is strictly better (\(\S\)6).
The confidence model and selective objective
Every decoder implements one interface \(\texttt{decode}(s)\to(\hat c,\,C)\), returning a correction \(\hat c\in\mathbb{F}_2^{d}\) and a confidence \(C\in[0,1]\). A recurring confidence heuristic for matching-type decoders maps the lit-check fraction to a score,
so confidence decays as more checks light up. The neural selector instead emits a predicted-class probability for the logical coset, which is approximately calibrated by construction. Given a threshold \(\tau\), the decoder commits only on shots with \(C\ge\tau\), which defines
where \(F\in\{0,1\}\) is the logical-failure indicator. Expected calibration error over \(B\) bins is
Each logical error rate is a mean of \(n\) Bernoulli indicators reported with the analytic Wald half-width \(\mathrm{ci}_{95}=1.96\sqrt{\hat p(1-\hat p)/n}\).
3Problem Formulation
The benchmark fixes a single evaluation surface and reports three quantities jointly rather than one in isolation. Given a stabilizer code with parity-check map \(H\) and logical functional \(L\), a family of structured-noise samplers \(\{\mathcal{N}_g\}\), and a family of decoders \(\{\mathcal{D}_k\}\) each exposing \(\texttt{decode}(s)\to(\hat c,C)\), every \((\mathcal{N}_g,\mathcal{D}_k)\) pair is evaluated on one defect-graph interface against:
- Logical error rate \(\hat p_{g,k}=\mathbb{E}[F]\), with analytic 95% Bernoulli intervals, so statistical separation is demonstrated rather than presumed.
- Calibration \(\mathrm{ECE}_{g,k}\), the deviation of stated confidence from realised accuracy.
- The risk–coverage curve \(\tau\mapsto(G(\tau),R(\tau))\), which traces the reliability purchased by abstaining on low-confidence syndromes.
The formulation asks a sharp question: which differences are statistically real, regime by regime. It exposes that the structure-agnostic decoders are tied (they share an i.i.d. target) and that the channel-adaptive spectral decoder separates from them under structured noise — with the magnitude and significance of every comparison reported, and the regimes where it does not win flagged with equal emphasis. Integrity is enforced in code: an automorphism-invariance check guards the symmetry assumptions, and an audit gate rejects any reported number not traceable to the regenerated summary.json, requires the central claim to be statistically evidenced, and rejects unevidenced superlatives.
4Principal Contributions
- A channel-adaptive spectral decoder. A maximum-a-posteriori coset decoder whose error model is a truncated transfer operator on the defect graph, calibrated to the operating channel — the first decoder in this family to beat the i.i.d.-optimal leader under structured noise with statistical significance (\(z=10.5,\,16.6,\,11.6\) on biased, burst, correlated noise), and an honest tie where theory forbids a win.
- A theory of the structured-channel gap. We prove the coset leader is sub-optimal on any non-uniform or correlated channel, that the transfer-operator decoder realises the channel-matched optimum and is consistent, and that its posterior confidence is calibrated by construction — so the empirical wins, the near-zero calibration error where the model matches, and the tie where it cannot are predicted, not observed by luck.
- An honest, regime-resolved evaluation. The same nine-decoder family on one
decode(syndrome)→(correction, confidence)interface across five structured regimes, with per-regime \(z\)-scores, and selectivity (risk–coverage) and calibration (ECE) as first-class metrics; the spectral decoder ties exactly where no data-channel advantage exists (i.i.d., read-out) and wins exactly where structure exists. - An audited, installable artifact. Every table and figure derives from one seeded artifact; analytic confidence intervals; a code-level audit gate that rejects untraceable numbers, requires the central claim to be statistically evidenced, and rejects unevidenced superlatives; and a pip-installable package that regenerates everything.
5Methodological Novelty
Each element of the contribution is positioned against named prior art, delineating its specific advance over a generic improvement claim.
Channel-awareness as a transfer operator
Noise-tailored matching is known for surface codes (weighted/belief matching; Higgott 2022; Roffe et al. 2020). Our advance is to cast it as a truncated transfer operator on the defect graph — an inhomogeneous Markov / matrix-product law — with provable channel-matched optimality and a posterior confidence that is calibrated by construction, and to estimate it from a small calibration sample of the operating channel.
The tie, explained and then broken
MWPM (Dennis et al. 2002; Fowler et al. 2012), BP/BP-OSD (Poulin&Chung 2008; Roffe et al. 2020), and learned decoders (Torlai&Melko 2017; Sweke et al. 2020; Bausch et al. 2024) all optimise the i.i.d. likelihood. We prove this is why eight such decoders tie on a repetition code, then exhibit the channel-adaptive decoder that breaks the tie by 20–44% under structured noise — with a per-regime \(z\)-score on every comparison.
Selective prediction imported into QEC
Selective classification and the reject option (Chow 1970; El-Yaniv&Wiener 2010; Geifman&El-Yaniv 2017), together with modern calibration (Guo et al. 2017; Naeini et al. 2015), are transplanted into the decoding setting and made measurable per decoder, supported by a decision-theoretic monotonicity guarantee that ties risk–coverage to calibration.
Audit-enforced, evidenced provenance
An audit module rejects any reported number not traceable to the regenerated summary.json, requires the central performance claim to be statistically evidenced (per-regime \(z\)-scores in the artifact), and rejects unevidenced superlatives in continuous integration — a reproducibility discipline embedded directly inside the artifact it governs.
6Theoretical Guarantees
Every structural claim the benchmark relies on is proved in full in the manuscript; the key statements are transcribed verbatim from the Theory section, each paired with its operational reading. The first three are the theoretical core of the contribution.
For an error channel \(P\), the decoder returning \(\hat c(s)=\arg\max_{e\in\{c_0,c_1\}}P(e)\) minimises the probability of logical failure among all decoders. If \(P\) is i.i.d. this equals the coset leader; if \(P\) is spatially non-uniform or correlated, the leader is strictly worse, by exactly the probability mass on which a heavier coset is more likely than a lighter one.
If the channel factorises as a transfer operator \(P_\theta(e)=\pi_0(e_0)\prod_i T_i(e_i,e_{i+1})\), then maximising \(\log P_\theta\) over the coset realises the channel-matched MAP decoder, hence minimises logical failure; and the plug-in estimator \(\hat\theta\) from Laplace-smoothed transition counts is consistent, so the decoder is asymptotically optimal.
spectral decoder is the Bayes-optimal rule for the structured channel it estimates — not a heuristic. Under i.i.d. noise \(T_i\) is uniform and it reduces to the leader, predicting the i.i.d. tie.The reported confidence \(C=\sigma(|\log P_\theta(c_0)-\log P_\theta(c_1)|)\) equals the model posterior of the chosen coset. If \(P_\theta\) equals the true channel then \(\Pr(F=0\mid C)=C\) almost surely, i.e. \(\mathrm{ECE}=0\); the empirical shortfall is exactly model misspecification.
For \(p<\tfrac12\) the minimum-Hamming-weight representative is the i.i.d. maximum-likelihood decoder; on the line the lookup, majority and matching decoders coincide with it, so on the i.i.d. channel they have identical logical error rate.
For every syndrome \(s\), the lookup, majority, and matching decoders return corrections in the same coset \(H^{-1}(s)\) and select the same logical class as the minimum-weight coset leader, except on the measure-zero set of syndromes whose coset leader has weight exactly \(d/2\). Hence on the i.i.d. channel all three have identical logical error rate.
If the noise distribution is cyclically invariant and a decoder \(\mathcal{D}\) is cyclically equivariant, its logical error rate is invariant under cyclic relabeling: \(\Pr_e[\mathcal{D}\text{ fails on }e]=\Pr_e[\mathcal{D}\text{ fails on }\rho_t e]\) for every shift \(t\).
automorphism_invariant certifies that the matching decoder respects the code's ring symmetry; it is true at reported scale.For i.i.d. Bernoulli\((p_L)\) failure indicators with mean \(\hat p\), the half-width \(\mathrm{ci}_{95}=1.96\sqrt{\hat p(1-\hat p)/n}\) is a consistent 95% interval estimate. If two decoders satisfy \(|\hat p_A-\hat p_B|\le\mathrm{ci}_{95}\), the difference is not significant at 95%.
If confidence \(C\) satisfies \(\Pr(F=0\mid C=c)=c\) almost surely, then committed risk \(R(\tau)=1-\mathbb{E}[C\mid C\ge\tau]\), coverage \(G\) is non-increasing in \(\tau\), and \(R\) is non-increasing in \(\tau\). Equivalently, risk is a non-increasing function of decreasing coverage.
7Live Demo: the Spectral Decoder vs the Minimum-Weight Leader
The widget below is a faithful, client-side re-implementation of the paper's central mechanism — not a cartoon. It samples structured noise on a distance-\(d\) repetition code, fits the spectral decoder's transfer operator \(P_\theta\) to a calibration stream, and decodes every shot with both the i.i.d.-optimal minimum-weight leader and the channel-adaptive spectral decoder, reporting both logical error rates live. Choose a regime and watch the gap open under structured noise and close under i.i.d. noise, exactly as the theory predicts.
Channel-adaptive decoding explorer client-side · vanilla JS · same mechanism as topoqec
The spectral decoder fits per-site \(2\times2\) transfer matrices to a 4,000-shot calibration stream of the chosen channel, then returns the more probable of the two coset representatives. Under structured noise (biased, burst, correlated) it cuts the logical error well below the leader; under iid the transfer operator is uniform and the two decoders coincide (\(z\to0\)). This is the mechanism of the reported-scale result — biased −44% (z=10.5), burst −33% (z=16.6), correlated −22% (z=11.6) — reproduced in your browser at smaller scale.
Selective decoding and calibration explorer
The second widget is an interactive illustration of the selective-decoding mechanism (the monotone-risk theorem and the ECE definition). It samples logical-failure indicators from a Bernoulli model whose base rate and confidence–accuracy coupling you control, then draws (left) the risk–coverage curve obtained by sweeping the abstention threshold \(\tau\) and (right) the reliability diagram whose occupancy-weighted gap is the ECE. The monotone fall as coverage shrinks is precisely what the theorem predicts when confidence is calibrated, and the curve flattens as you decouple confidence from accuracy.
Risk–coverage and calibration explorer client-side · vanilla JS
Model: each shot draws a failure indicator \(F\sim\mathrm{Bernoulli}(p)\); its confidence centres high on successes and low on failures, with the separation set by the spread control, then is blurred by Gaussian noise of width growing as calibration drops and is pulled toward an uninformative 0.5 as calibration → 0. Perfect calibration recovers the monotone risk–coverage fall of Theorem 9; lowering calibration flattens the curve and inflates ECE, mirroring the eight-fold ECE spread observed across the measured decoder family.
8Empirical Validation
Every number below is transcribed verbatim from the reported-scale run (repetition code, distance 9, nine decoders, 20,000 trials per regime, \(n=100{,}000\) pooled shots per decoder, \(p_{\mathrm{eval}}=0.12\), seed 0) and written to the regenerated artifact summary.json.
Pooled comparison: the spectral decoder breaks the tie
| decoder | logical err. rate | ±95% CI | ECE |
|---|---|---|---|
| spectral ★ | 0.1291 | 0.0021 | 0.0629 |
| majority | 0.1508 | 0.0022 | 0.0498 |
| bp | 0.1523 | 0.0022 | 0.1238 |
| fused | 0.1524 | 0.0022 | 0.0622 |
| matching | 0.1525 | 0.0022 | 0.2620 |
| topology | 0.1525 | 0.0022 | 0.2421 |
| neural | 0.1534 | 0.0022 | 0.0446 |
| rl | 0.1541 | 0.0022 | 0.2624 |
| lookup | 0.1542 | 0.0022 | 0.0558 |
The eight structure-agnostic decoders lie within 1.5 CI widths of one another (spread 0.0033), so they are statistically tied — exactly as the i.i.d.-target theory predicts. The channel-adaptive spectral decoder is lower by \(\approx\!10\) CI widths: a significant, not marginal, separation. Calibration spans a factor of nearly six across the agnostic family (0.0446–0.2624) with no relationship to accuracy.
Where the win is, and where it (honestly) is not
| regime | spectral | leader | rel. reduction | z | verdict |
|---|---|---|---|---|---|
| i.i.d. | 0.00145 | 0.00230 | +37.0% | +2.0 | tie (no structure) |
| biased | 0.02400 | 0.04285 | +44.0% | +10.5 | significant win |
| burst | 0.11960 | 0.17845 | +33.0% | +16.6 | significant win |
| correlated | 0.16190 | 0.20695 | +21.8% | +11.6 | significant win |
| meas. error | 0.33870 | 0.34020 | +0.4% | +0.3 | tie (corrupted syndrome) |
The spectral decoder cuts the logical error by 20–44% on the three structured data-noise regimes (\(z>10\)), and is statistically tied with the leader under i.i.d. and read-out noise (\(|z|<3\)) — the regimes where Theorem (channel-matched MAP) forbids a single-shot data-channel advantage. The benchmark localises the win rather than averaging it away.
Calibration by construction, where the model matches
| regime | spectral ECE | best agnostic ECE |
|---|---|---|
| i.i.d. | 0.00069 | 0.08485 (fused) |
| biased | 0.00159 | 0.11923 (neural) |
| burst | 0.02538 | 0.03454 (lookup) |
| correlated | 0.04859 | 0.04465 (lookup) |
| meas. error | 0.32359 | 0.08445 (bp) |
Because the spectral confidence is the exact model posterior, its ECE is near zero where the transfer operator matches the channel — best in the family by orders of magnitude under i.i.d., biased and burst noise — and large only under read-out noise, where it conditions on a corrupted syndrome. This is the empirical face of the calibration proposition.
Threshold behaviour confirms the i.i.d. tie. Across the eleven-point physical-error grid under i.i.d. noise, the spectral decoder's curve overlaps the agnostic decoders and the optimal lookup leader, staying well below the break-even line \(p_L=p\); at \(p=0.21\) the logical error is only ≈0.024 for every core decoder. The win is structured-noise-specific, exactly as the theory requires.
9Significance and Impact
Channel-awareness, not algorithmic sophistication within the i.i.d.-optimal class, is what separates decoders under realistic noise — and it can be made provable, tractable, and self-calibrating.
The result reframes how decoders should be evaluated and built. The eight structure-agnostic decoders are tied on raw logical error not because the problem is saturated but because they share one i.i.d. target; the moment a decoder models the channel, the tie breaks by 20–44% on the structured data-noise regimes (\(z>10\)). The contribution is therefore a decoder, not merely a measurement: a transfer-operator model on the defect graph that is Bayes-optimal for the structured channel, reduces to the leader under i.i.d. noise, and reports a posterior confidence calibrated by construction. The consequence for fault-tolerant control loops is direct: a device whose noise can be characterised — bias, bursts, correlations — gets, from a small calibration sample, a decoder that is both more accurate and trustworthy in its abstentions exactly where its accuracy gains are real, with the boundary of that advantage made explicit by per-regime significance tests.
Limitations, reported in full. The scope of the evidence is stated explicitly so that no conclusion is read beyond what the artifact supports.
- Toy codes only: the primary benchmark is the repetition code; the second code is a topology-faithful surface-like lattice, not a stabilizer-exact CSS surface code.
- Simulated structured noise, not hardware syndromes; the performance claim is a statistically evidenced advantage under simulated structured noise, not on hardware.
- The spectral decoder is channel-adaptive: it requires a calibration sample of the operating channel. Where the channel is unknown or drifting, the advantage shrinks toward the i.i.d. tie; an online estimator is future work.
- A first-order (nearest-neighbour) transfer operator at a single distance; longer-range correlations and a distance-scaling study remain open. Single author and environment.
The natural next step is to extend the transfer-operator decoder—keeping selectivity and calibration first-class—to a full surface code at scale, to higher-order and online-estimated channel models, to hardware syndromes, and to a distance-scaling sub-threshold study.
10Reproducibility
The full pipeline accompanies the manuscript at submission/code: it implements every code, noise model, decoder, and metric, including the audit gate, and regenerates every figure, table, and reported number. The package is distributed as specops-transfer-decoder (pip-installable via pyproject.toml). A single reproduce entry point writes results/summary.json, the authoritative artifact from which all tables and figures are generated; no reported value is entered by hand.
The reported run completes end-to-end in 277 s at 206 MB peak memory on a single CPU. Provenance is recorded alongside the results: macOS-26.5.2-arm64, Python 3.9.6, numpy 1.26.4, scipy 1.13.1, networkx 3.2.1, scikit-learn 1.6.1, matplotlib 3.9.4; master seed 0. A reviewer reproduces identical numbers by running the pipeline at reported scale, and the audit driver scripts/audit_claims.py runs in continuous integration, exiting non-zero if any reported number fails to trace to summary.json, if the central spectral-versus-leader claim is not statistically evidenced, or if an unevidenced superlative is detected.
Cite this work
11References
Related work.
The benchmark places one decoder family on a common footing. Minimum-weight perfect matching—the production standard for surface codes, packaged in PyMatching (Higgott 2022) and built on Edmonds' blossom algorithm (Edmonds 1965)—sets the i.i.d.-optimal baseline that the spectral decoder must tie where no structure exists. Belief propagation with ordered-statistics decoding (Roffe et al. 2020; Panteleev and Kalachev 2021) underpins the quantum LDPC line of work, and a fast-growing learned-decoder literature trains neural networks and reinforcement-learning agents to decode (Torlai and Melko 2017; Baireuther et al. 2018), with very recent work exceeding matching accuracy on real hardware syndromes (Bausch et al. 2024). The two evaluation axes added here—selective prediction (Chow 1970; El-Yaniv and Wiener 2010; Geifman and El-Yaniv 2017) and calibration (Guo et al. 2017; Naeini et al. 2015)—are imported from the classification literature, where abstention and calibrated confidence are first-class concerns. The complete bibliography of the manuscript follows.
- P. W. Shor. "Scheme for reducing decoherence in quantum computer memory." Physical Review A 52(4), R2493–R2496, 1995.
- M. A. Nielsen and I. L. Chuang. "Quantum computation and quantum information." 2010.
- B. M. Terhal. "Quantum error correction for quantum memories." Reviews of Modern Physics 87(2), 307–346, 2015.
- A. Yu. Kitaev. "Fault-tolerant quantum computation by anyons." Annals of Physics 303(1), 2–30, 2003.
- E. Dennis, A. Kitaev, A. Landahl, and J. Preskill. "Topological quantum memory." Journal of Mathematical Physics 43(9), 4452–4505, 2002.
- A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland. "Surface codes: Towards practical large-scale quantum computation." Physical Review A 86(3), 032324, 2012.
- Google Quantum AI. "Suppressing quantum errors by scaling a surface code logical qubit." Nature 614(7949), 676–681, 2023.
- O. Higgott. "PyMatching: A Python package for decoding quantum codes with minimum-weight perfect matching." ACM Transactions on Quantum Computing 3(3), 1–16, 2022.
- J. Edmonds. "Paths, trees, and flowers." Canadian Journal of Mathematics 17, 449–467, 1965.
- D. Poulin and Y. Chung. "Iterative quantum coding systems and the topological color code." Quantum Information & Computation 8(10), 987–1000, 2008.
- P. Panteleev and G. Kalachev. "Degenerate quantum LDPC codes with good finite length performance." Quantum 5, 585, 2021.
- J. Roffe, D. R. White, S. Burton, and E. Campbell. "Decoding across the quantum low-density parity-check code landscape." Physical Review Research 2(4), 043423, 2020.
- G. Torlai and R. G. Melko. "Neural decoder for topological codes." Physical Review Letters 119(3), 030501, 2017.
- S. Varsamopoulos, B. Criger, and K. Bertels. "Decoding small surface codes with feedforward neural networks." Quantum Science and Technology 3(1), 015004, 2017.
- P. Baireuther, T. E. O'Brien, B. Tarasinski, and C. W. J. Beenakker. "Machine-learning-assisted correction of correlated qubit errors in a topological code." Quantum 2, 48, 2018.
- S. Krastanov and L. Jiang. "Deep neural network probabilistic decoder for stabilizer codes." Scientific Reports 7(1), 11003, 2017.
- R. Sweke, M. S. Kesselring, E. P. L. van Nieuwenburg, and J. Eisert. "Reinforcement learning decoders for fault-tolerant quantum computation." Machine Learning: Science and Technology 2(2), 025005, 2020.
- P. Andreasson, J. Johansson, S. Liljestrand, and M. Granath. "Quantum error correction for the toric code using deep reinforcement learning." Quantum 3, 183, 2019.
- J. Bausch, A. W. Senior, F. J. H. Heras, et al. "Learning high-accuracy error decoding for quantum processors." Nature 635(8040), 834–840, 2024.
- C. K. Chow. "On optimum recognition error and reject tradeoff." Vol. 16, pp. 41–46. IEEE, 1970.
- R. El-Yaniv and Y. Wiener. "On the foundations of noise-free selective classification." Journal of Machine Learning Research 11, 1605–1641, 2010.
- Y. Geifman and R. El-Yaniv. "Selective classification for deep neural networks." Advances in Neural Information Processing Systems 30, 2017.
- C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. "On calibration of modern neural networks." In International Conference on Machine Learning, pp. 1321–1330. PMLR, 2017.
- M. P. Naeini, G. Cooper, and M. Hauskrecht. "Obtaining well calibrated probabilities using Bayesian binning." In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 29, 2015.
- A. A. Hagberg, D. A. Schult, and P. J. Swart. "Exploring network structure, dynamics, and function using NetworkX." In Proceedings of the 7th Python in Science Conference (SciPy), pp. 11–15, 2008.
- C. R. Harris, K. J. Millman, S. J. van der Walt, et al. "Array programming with NumPy." Nature 585(7825), 357–362, 2020.
- F. Pedregosa, G. Varoquaux, A. Gramfort, et al. "Scikit-learn: Machine learning in Python." Journal of Machine Learning Research 12, 2825–2830, 2011.
- L. Li, W. Chu, J. Langford, and R. E. Schapire. "A contextual-bandit approach to personalized news article recommendation." In Proceedings of the 19th International Conference on World Wide Web, pp. 661–670, 2010.
- R. S. Sutton and A. G. Barto. Reinforcement Learning: An Introduction. MIT Press, 2nd edition, 2018.
- P. Virtanen, R. Gommers, T. E. Oliphant, et al. "SciPy 1.0: fundamental algorithms for scientific computing in Python." Nature Methods 17(3), 261–272, 2020.
- J. D. Hunter. "Matplotlib: A 2D graphics environment." Computing in Science & Engineering 9(3), 90–95, 2007.
- M. Huynh. "Query-efficient quantum approximate optimization via graph-conditioned trust regions." arXiv preprint arXiv:2604.24803, 2026. doi:10.48550/arXiv.2604.24803
- M. Huynh. "Topology-conditioned warm starts match spectral initialization for depth-one QAOA." Manuscript in preparation, 2026.
- M. Huynh. "Commutator-graph term ordering reduces first-order Trotter error at matched gate cost." Manuscript in preparation, 2026.
- M. Huynh. "Provable operator learning of size-transferable Trotter corrections for Hamiltonian simulation." Manuscript in preparation, 2026.
- M. Huynh. "CPTP-Compiler-PINNs: structure-preserving residual learning and matrix-free compilation for open quantum dynamics." Manuscript in preparation, 2026.
- M. Huynh. "Objective-aware sequence modelling for drift forecasting and anomaly detection in quantum hardware." Manuscript in preparation, 2026.
- M. Nguyen and N. Jing. "Zassenhaus expansion in solving the Schrödinger equation." arXiv preprint arXiv:2505.09441, 2025. doi:10.48550/arXiv.2505.09441
- N. Jing and M. Nguyen. "Complexity bounds for Hamiltonian simulation in unitary representations." arXiv preprint arXiv:2603.07231, 2026. doi:10.48550/arXiv.2603.07231