acceptodds
Under review as a conference paper at ICLR 2027

From Noise Beliefs to Logical Decisions: Surface-Code Decoding with Temporal Side Information

Abstract

Noise in quantum processors has memory: defects drift, error rates switch between regimes, and bursts persist over many syndrome rounds. A decoder that tracks this hidden process should make better corrections, but in continuous operation it must commit each correction before the full record is available. We ask when temporal memory actually changes a logical correction under this constraint. Our decoder couples a causal two-state hidden Markov filter over auxiliary observations to rolling minimum-weight matching: the filter's belief sets the edge weights, and a carried residual-syndrome frontier keeps every committed correction consistent with the detector record. Because the filter, the physical hypotheses, and the commit rule are separate modules, each can be compared against matched controls. Across 45 noise conditions and 92,160 trajectories, filtering reduces state-estimation error by up to 93% and reduces exactly to the current-observation rule under IID noise. In a held-out confirmation with 4,096 rolling streams per code distance, the filter lowers logical failures from 507 to 433 at and from 158 to 131 at relative to current-only calibration of the same decoder (Holm-adjusted and ). Three independently trained GRUs match these error counts but spend 13–15 longer per belief update. On shared terminal records the belief reduces failures by 6-9% for each of four solver families, and fixed-budget candidate interventions show that better physical hypotheses can lower the error rate even while support coverage shrinks. A replay under frozen calibrations from two cloud quantum processors (6,144 returned shots) preserves the same ordering.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Error Correction, Surface-Code Decoding, Temporal Side Information, Hidden Markov Models, Belief-State Tracking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QAxXOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies an interesting and practically relevant problem: how to exploit temporal information about time-varying hardware noise to improve online surface-code decoding. The proposed framework models the noise environment as a two-state hidden Markov process, uses auxiliary observations to maintain a causal belief about the current noise regime, and uses this belief to adapt the fault probabilities/edge weights supplied to existing decoders. To support continuous operation, the paper develops a rolling decoding procedure that commits corrections from older rounds while carrying a residual-syndrome frontier forward. The authors carefully compare static, current-only, EMA, GRU, and Markov estimators and evaluate several decoder families. In the main held-out rolling experiment, the Markov approach reduces failures from 507 to 433 at d=3 and from 158 to 131 at d=5, compared with current-only calibration of the same MWPM decoder.

Critical strengths
  • Interesting and important problem. Temporal variation and correlations in hardware noise are relevant challenges for practical quantum error correction. Understanding whether knowledge of such temporal structure can actually improve logical decoding decisions is worthwhile.

  • Multiple decoder families and controls. The evaluation includes MWPM, BP+OSD, BP+LSD, and BP+UF, as well as GRU-based temporal estimation, which helps show that the observed phenomenon is not entirely tied to one particular decoder or estimator.

Critical weaknesses
  • Limited algorithmic novelty. This is my primary concern. Once the two-state temporal-noise model is assumed, the core approach consists of a standard HMM/Bayesian filter followed by existing QEC decoders whose priors or edge weights are updated according to the inferred belief. Neither the temporal inference technique nor the underlying decoders are new. Indeed, the paper explicitly states that the Bayesian filter itself is standard. The rolling-frontier construction and associated analysis add value, but I am not convinced that the overall methodological contribution is sufficiently substantial for a top-tier venue.

  • Theoretical novelty is also relatively modest. The memory-decay analysis, Brier-score identity, decision-boundary observation, and candidate-support decomposition provide useful explanations of the experiments, but most are relatively elementary consequences of the assumed probabilistic model. The rolling-decoder correctness result is more substantive, but applies to the particular short-range graph setting considered here.

  • The strongest evaluation assumes a favorable model match. The primary experiments generate noise using a two-state Markov environment and then use a two-state Markov filter to infer that environment. Prior work motivates temporally correlated noise, but the paper provides limited evidence that this specific latent-state model accurately describes realistic hardware noise. It is therefore unclear how much of the observed gain would survive substantial model mismatch.

  • Limited scale of the main surface-code experiments. The main confirmation experiments use d=3 and d=5, while d=7 appears only in additional scale diagnostics. It would be valuable to understand whether the benefit persists at substantially larger code distances.

  • Limited real-hardware evidence. The hardware component is a frozen-calibration/classical replay rather than live surface-code error correction with naturally occurring temporal noise and online adaptation. The authors appropriately acknowledge this limitation. Consequently, the main logical-error improvements remain simulation results.

  • The benefit is not uniformly robust across experimental regimes. In the long-horizon experiments, the Markov-vs-current-only differences are small and not statistically significant, and in the 30%-missing condition the Markov method actually has more failures (460 versus 450). This raises questions about the practical operating regimes in which temporal modeling provides a meaningful logical advantage.

Questions for authors
  1. What is the main technical novelty beyond combining a standard HMM estimator with existing QEC decoders? Which component should be viewed as the central new algorithmic contribution of the paper?

  2. How robust is the method to model mismatch? For example, what happens if the true noise has three or more latent states, continuous drift, non-Markovian dwell times, changing transition probabilities, or spatially heterogeneous temporal behavior?

  3. Can the authors provide results at larger code distances? The main confirmation focuses on d=3 and d=5. Is there a reason that the logical-error advantage was not evaluated systematically at d=7 or beyond?

  4. What realistic hardware signal would provide the auxiliary observations Y_t? How frequently can such information be obtained, and what latency and measurement overhead would it introduce in an actual QEC system?

  5. Why does the long-horizon evaluation show substantially weaker benefits? In particular, can the authors explain when improvements in local decoding decisions should be expected to translate into improvements in long-horizon logical-memory performance?

  6. Can syndrome information itself be used to estimate the hidden noise state? The current framework estimates Z_t from auxiliary observations. It would be interesting to consider P(Z_t\mid Y_{0:t},D_{0:t}), since the syndrome itself may contain substantial information about the current noise regime.

  7. Can the hardware evaluation be strengthened beyond classical replay? A particularly convincing experiment would estimate temporal noise from real device telemetry and prospectively demonstrate improved decoding on subsequently collected syndrome data.

3 comments
AuthorsOpenReview ↗

Response to Reviewer QAxX (1/3): Novelty

Dear Reviewer,

Thank you for an expert and thorough review. We took each of your seven questions as a concrete task, and together they account for most of the new experiments. The changes made for you carry the margin tag R3 in the revised PDF.

Summary of changes for you

  1. Novelty: the filter and the excess-risk identity are presented as classical; the contributions are a decision-level account of when memory pays, commit-time beliefs, one recursion for all evidence sources, and a prospective hardware test (Sec. 1, 2, 4, 5.4, 5.5).
  2. All five mismatch cases you named, plus eight more, and a spatial control that isolates memory (Sec. 5.3, Table 3, App. B.2).
  3. d=3,5,7d=3, 5, 7 and 99 (Sec. 5.2, Table 5); a syndrome-based belief (Sec. 3.1); an explanation of the long-horizon behaviour (Sec. 4, 5.2); prospective decoding with native errors (Sec. 5.5).

You named limited algorithmic novelty as your primary concern; we address it first.

Q1: What is new beyond a standard HMM plus existing decoders?

Short answer: not the filter, and not the excess-risk identity itself. We agree that the forward filter is textbook, and the revision says so (Positioning, L137). Theorem 2(b) is the classical excess-risk identity for binary decisions (Devroye et al., 1996; App. A.2), and we agree that the original propositions were elementary; they are now auxiliary results in App. A.4. What we believe is new is what these tools yield for streaming QEC:

  • (a) When temporal memory changes a logical decision (Theorem 2(c)–(e), Corollary 1, Eq. 7): a checkable strictness criterion, equality under IID noise, a headroom bound that predicts the fading long-horizon gains, and a transfer identity that measures how much of the Bayes gain matching keeps (79%). An exact d=3d=3 Bayes decoder confirms each part (Fig. 4).
  • (b) Commit-time beliefs in a checked streaming construction (L234). Every windowed decoder already waits W−BW-B rounds before committing; we use that wait to smooth the belief, so each committed weight is the optimal estimate given everything known at commit time, at no extra latency. Its gain grows with the lookahead (0.17, 0.32, 0.68 and 0.73 pp for W−B=3,9,21,33W-B = 3, 9, 21, 33; the last two significant), and it is never significantly worse in any of 31 studies. The frontier invariant is checked over 2.1M commits; you are right that its proof covers the short-range graphs considered here.
  • (c) One recursion for every evidence source. Telemetry, syndrome statistics (an approximate Poisson likelihood, Q6) and parameter uncertainty (Eq. 3) enter the same filter.
  • (d) A prospective hardware test. To our knowledge, Sec. 5.5 is the first test on superconducting processors in which temporal beliefs, learned from telemetry without labels and frozen before the test data, drive a decoder's decisions (Q7).

Takeaway, from the comparison that tests (c): against executed DGR-style reweighting on identical records, our syndrome-only belief ties DGR (all p≥.18p \ge .18); only when an extra regime signal exists, which DGR has no input for, does our decoder make 13–19% fewer errors, in all four settings (Table 2).

Our answers to Q2–Q3 and Q4–Q7 follow in the next two comments.

AuthorsOpenReview ↗

Response to Reviewer QAxX (2/3): Model Mismatch and Code Distance

Q2: Robustness to three or more states, continuous drift, non-Markovian dwell times, changing transition rates and spatial heterogeneity?

Short answer: as suggested, we ran all five; the gain survives each, and continuous drift is the weakest case. The decoder still assumes two symmetric states; 4,096 streams each:

your casecurrent-only − ours (pp, 95% CI)
three latent levels+1.47 [0.66, 2.27]
four latent levels+1.42 [0.66, 2.20]
continuous OU drift+0.85 [0.05, 1.66]
semi-Markov dwell times+2.39 [1.59, 3.17]
switching rate drifting 25×25\times+2.93 [2.20, 3.69]
two spatial regions$^\dagger$+3.15 [2.32, 3.96]

†^\dagger Ours keeps one belief per region, while this current-only control sees the averaged sensor. Giving the memoryless control one belief per region as well, on the same streams, isolates memory: +2.42 [1.66, 3.17] (App. B.2).

Takeaway: significant in all 12 non-IID environments (Table R2); under IID noise the decisions equal the memoryless ones.

Q3: Larger code distances?

Short answer: the advantage persists and grows in relative terms. In the submission d=7d=7 appeared only in scale diagnostics; the revision evaluates d=3d=3–99 systematically, at a common channel scale chosen so that d=9d=9 still fails measurably, 4,096 streams each (Table 5):

failuresd=3d=3d=5d=5d=7d=7d=9d=9
current-only997499242103
ours86039218772
true hidden state85238117969
relative reduction14%21%23%30%

Takeaway: all four 95% intervals exclude zero, and our decoder stays within 11 failures of MWPM given the true hidden state at every distance.

Q4–Q7 follow in the next comment.

AuthorsOpenReview ↗

Response to Reviewer QAxX (3/3): Telemetry, Long Horizons, Syndrome Information and Hardware

Q4: What signal provides YtY_t, how often, and at what cost?

Short answer: any signal with a known per-regime likelihood; sparse telemetry is enough; the filter costs a few scalar operations per round. Candidates are short Ramsey or relaxation probes on spectator qubits, leakage or readout-resonator monitors, TLS spectroscopy and burst detectors (L194). The filter skips missing rounds exactly, so one value every 2, 4 or 8 rounds keeps a 1.5–1.6 pp gain (Fig. 5c).

  • Measurement overhead. In our device study the probe occupies five spectator qubits and one telemetry circuit per epoch, of which the decoder uses only 8–64 of 1,000 shots; the syndrome-aware belief (Q6) needs no probe at all.
  • Latency. Eq. (2) is a constant number of scalar operations per round, and commit-time smoothing costs O(W)O(W) per commit inside the wait the window already imposes. Our CPU research implementation measures 11.2 μ\mus per belief update at d=5d=5 (App. G.4); this is not a hardware deadline, and per-round service is dominated by graph construction and matching, not by the filter (Fig. 9).

Q5: Why are long-horizon benefits weaker, and when should local improvements carry over?

Short answer: because the terminal parity saturates. Corollary 1 (L307) bounds any gain over the Bayes decoder by 1−2R(δ⋆)1 - 2\mathcal{R}(\delta^\star), so the gain must vanish as the memory approaches 50% failure, even while decisions keep differing. Exact d=3d=3 computations show the gain first rising and then following this bound downward (Fig. 4c, Fig. 7c, Table 4). The original long-horizon conditions sat at 37–48% failure, where even the Bayes-optimal gain is only 0.24–0.35 pp. Re-run with four times as many streams, both give positive reductions (1.56 pp, p=.032p=.032; 1.27 pp, p=.11p=.11), and the 460-versus-450 missing-data cell does not replicate. With headroom, the long-horizon gain is large: 3.9–4.8 pp at 257 rounds with the frozen operating point, p<10−4p \lt 10^{-4} (Table 1).

Q6: Can the syndrome itself estimate ZtZ_t?

Short answer: yes, and we added it; its value is decoding without a sensor. The belief approximates P(Zt∣Y0:t,D0:t)P(Z_t \mid Y_{0:t}, D_{0:t}) with a Poisson likelihood of time-like coincidences versus other detection events, which track the ratio of measurement to data errors (L185). Without any sensor it lowers failures relative to the static decoder in five of six settings (d=5d=5: 172 → 126) but not with fast switching at d=3d=3 (533 vs. 520), and it ties executed DGR (Table 2). Added to telemetry, it changes failures by at most 8 in either direction (Tables 1–2), so once telemetry is present the syndrome adds little regime information in our noise model.

Q7: Can the hardware evaluation go beyond classical replay?

Short answer: yes. We ran the experiment you suggested, prospectively and with native errors (Sec. 5.5, Fig. 6, App. C; General Response, Table R1). Temporal parameters are learned without labels from spectator telemetry and frozen before the 360 test epochs; a hidden two-state idle-time schedule is never shown to the decoder, so every error is the device's own.

Takeaway: on three disjoint qubit sets of two processors from different vendors, all 48 paired comparisons favour our decoder with 95% intervals above zero (WK_C180 set A: 3.75%, against 4.07–5.14% for the baselines and 3.56% for the oracle).

You are right that this is not yet closed-loop surface-code correction: the devices run repetition memories, and the regime is engineered, although every error is native (L535). Without engineering, 11.5 hours of drift (3.11M shots, Table 7) give a weaker result for our model: temporal beliefs still beat current-only decoding, but the drift is continuous, the two-regime filter selects a single regime for most qubits and gains only 0.02–0.06 pp over frozen calibration, and a Kalman filter is best (4.615% vs. 4.663% at 16 shots). We report this as it is.

Presentation. The introduction now states the two gaps explicitly (L54–69), each method section opens by naming the gap it addresses (L147, L201, L257), and Sec. 5 opens with one protocol paragraph for all studies (L345–353).

We hope these answers address your concerns, in particular the question of novelty. Please let us know if anything remains unclear; we would welcome further questions.

Reviewer_P1vROpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies real-time surface-code decoding under temporally correlated noise. A causal two-state HMM uses auxiliary observations to update edge weights for rolling MWPM. A residual-syndrome frontier is carried across commits, and the paper also studies terminal-candidate selection. The main question is when temporal memory improves irreversible logical decisions, rather than just hidden-state estimates. The evaluation includes state-estimation maps, held-out decoding, ablations, latency measurements, and a replay using frozen device calibrations.

Critical strengths
  1. The estimator comparisons are well controlled. Current-only, static, EMA, Markov and GRU estimators all use the same inputs, and a matched-marginal candidate control separates the effect of temporal correlation from marginal quality (Sections 5.3–5.4; Figures 5–6).

  2. The rolling interface has explicit residual-syndrome invariants, verified across all 2,107,136 commit checks (Proposition 3; Table 24). On held-out streams, Markov filtering cuts failures from 507 to 433 at distance 3 and from 158 to 131 at distance 5, both significant after Holm correction (Table 1). GRUs achieve similar accuracy, but their belief updates are 13–15 times slower (Table 3).

Critical weaknesses
  1. The main logical gain is narrow. The operating point was chosen on development data, and the setting assumes a known two-state transition law, engineered Gaussian auxiliary observations and idealized Pauli noise. At horizon 257, none of the memory contrasts is significant, and Markov filtering is worse in one missing-data condition (Table 11). State estimation also deteriorates substantially under IID or shuffled model mismatch (Table 30). The device study uses frozen calibration probabilities with synthetic histories, rather than native temporal telemetry or closed-loop correction (Section 5.5; Appendix H). So outside the selected, model-matched regime, the paper does not yet show robustness, a threshold benefit or a consistent logical advantage.

  2. The terminal-decision results are less convincing. With 512 records, neither the gains across the four solver families nor the joint-candidate intervention is significant after correction (Figure 6; Table 15). The paper discusses DGR (arXiv:2311.16214), Libra, AlphaQubit, AlphaQubit 2 and neural MWPM but does not compare against them experimentally (Appendices F and S). The public prefix is also not an official benchmark (Table 34). It remains unclear how much of the improvement goes beyond earlier adaptive graph reweighting and recent correlated-noise decoders.

Questions for authors
  1. If the filter parameters and the operating point are fixed in advance, can you report a paired confidence interval for the reduction in terminal logical failures at horizon 257?
  2. Could you add an information-matched comparison with at least one adaptive or correlated-noise baseline that is actually run, such as DGR?
2 comments
AuthorsOpenReview ↗

Response to Reviewer P1vR (1/2): Paired CI at 257 Rounds and an Executed DGR Baseline

Dear Reviewer,

Thank you for a careful review and for two very concrete questions; we ran both as you specified. The changes made for you carry the margin tag R2 in the revised PDF.

Summary of changes for you

  1. Q1 done: frozen operating point and filter parameters at H=257H=257, with paired CIs, next to a re-run of the original stress conditions (Sec. 5.2, Table 1, App. B.3).
  2. Q2 done: an executed DGR-style baseline on identical records, with and without matched information (Sec. 5.4, Table 2).
  3. The terminal comparison re-run with 16,384 instead of 512 records, the IID and shuffled controls, and mismatch, distance and native-hardware results (Sec. 5.2–5.5, App. B.2).

We understand your main concern to be that the logical gain is narrow: shown only at an operating point selected on development data and matched to the model, null at 257 rounds, underpowered at 512 records, and without an executed adaptive baseline. We address it first.

Q1: With filter parameters and operating point fixed in advance, what is the paired CI at H=257H=257?

Short answer: 4.76 pp at d=3d=3 and 3.93 pp at d=5d=5, both p<10−4p \lt 10^{-4}. We kept the operating point frozen for the original confirmation (W=12W=12, B=3B=3, a=.01a=.01, σ=1.5\sigma=1.5, 30% missing; filter parameters unchanged), extended it to H=257H=257 and drew fresh seeds. For comparison, the last two rows re-run the submission's own long-horizon stress conditions (ς=2\varsigma=2) with four times as many streams:

H=257H=257, 4,096 streamscurrent-onlyoursreduction (pp)pp
frozen operating point, d=3d=31,3771,1824.76 [3.44, 6.10]<10−4\lt 10^{-4}
frozen operating point, d=5d=55854243.93 [3.03, 4.83]<10−4\lt 10^{-4}
stress, persistent, d=5d=51,7671,7031.56.032
stress, 30% missing, d=5d=51,7561,7041.27.11

Takeaway: at the frozen operating point the gain at 257 rounds is larger than at 65 rounds (1.49 and 1.07 pp), because a persistent environment rewards evidence accumulated over a longer stream; the causal filter alone already gives 3.66 and 3.25 pp (1,377 → 1,227 and 585 → 452, Table 1). Under the stress conditions the gain is positive but small, and significant in one of the two.

Why the stress conditions give small gains. They leave a d=5d=5 memory failing 37–48% of the time, and Corollary 1 bounds any gain over the Bayes decoder by 1−2R(δ⋆)1 - 2\mathcal{R}(\delta^\star). An exact d=3d=3 analogue shows that even the Bayes-optimal gain is only 0.24–0.35 pp there (App. B.3, Table 4), below what 1,024 streams can resolve; the negative missing-data cell of the submission (old Table 11, now Table 15) does not replicate.

Q2: An information-matched comparison with an executed adaptive baseline such as DGR?

Short answer: as suggested, we implemented DGR-style reweighting for rolling decoding and ran it on identical records. Information-matched, with the detector record only, our syndrome-only belief and DGR are statistically tied. Our decoder is better only when an additional regime signal exists, which DGR has no input for. The estimator follows the two-point correlation formula of Wang et al. over a 32-layer window with shrinkage (16 and 64 layers give nearly identical counts); "fast" is switching probability a=.05a=.05 instead of .01.01:

failures, 4,096 streamsd=3d=3d=5d=5d=3d=3, fastd=5d=5, fast
DGR-style reweighting (syndrome)522132517167
ours, syndrome only544126533161
ours, with auxiliary signal454108439136

Takeaway: information-matched, the two are tied (all p≥.18p \ge .18); with an extra signal, our decoder makes 13–19% fewer errors than DGR in all four settings (p≤.0023p \le .0023). We therefore claim the second, not the first: what the HMM adds over reweighting is a principled way to combine syndrome and telemetry evidence.

AuthorsOpenReview ↗

Response to Reviewer P1vR (2/2): Terminal Power, Model Assumptions, Hardware and Learned Decoders

Terminal decisions with adequate power. Short answer: each of the four solver families now improves significantly. With 16,384 complete records (8,192 profiles, d=5d=5) instead of 512:

failures / 16,384MWPMBP+OSDBP+LSDBP+UF
current-only616621623920
ours468514514719
Holm pp.0004.0004.0004.0004

The four identical values are the resolution limit of the test, not a rounding artefact: every raw pp-value is 1/10,0011/10{,}001, the smallest that 10,000 sign randomizations can return, and Holm over four comparisons gives 4/10,0014/10{,}001 (Table 1 caption).

You are right that the joint-candidate intervention is not significant (512 records, Holm p=.46p=.46, old Table 15, now Table 19); the main text now presents it as a mechanism rather than a result (L321).

Operating point chosen on development data. No decoder or filter setting was re-tuned: every new rolling study reuses the operating point frozen before the original confirmation (Q1). Where a study changes the environment itself (channel scale ς=1.6\varsigma=1.6 for d=3d=3–99, a=.05a=.05 for "fast"), that setting is part of the study's specification, fixed before its first run (Sec. 3.3, L248; App. B.1).

Known two-state law and engineered Gaussian observations. The parameters are now learned from unlabeled streams (Eq. 3), with no loss in the well-specified case (430 vs. 430 failures). With Student-tt, biased and missing observations, three or four latent levels, continuous drift and the other cases of Table R2 (General Response), failures fall significantly in all 12 non-IID environments; continuous drift is the weakest case. The syndrome-only belief needs no auxiliary signal at all (Q2).

State estimation under IID or shuffled mismatch (old Table 30, now Table 34). That study keeps the filter's switching rate fixed by design (App. R.1), which is where the degradation came from. We re-ran it with the learned filter of Eq. (3): under IID and shuffled data its posterior concentrates on a=1/2a=1/2 and the state Brier score equals the current-only score exactly (.096 and .095, against .134 and .127 with the fixed rate), while in the Markov, missing and semi-Markov conditions it keeps the full gain (App. B.2). In decoding, it makes exactly the memoryless decisions under IID noise (559 vs. 559).

Idealized Pauli noise and a threshold benefit. We agree that the simulations remain phenomenological, and the conclusion now says so (L535); the device experiments supply native errors, although on repetition memories. We have not estimated a threshold. What we can show is that at a fixed channel scale the relative reduction grows with distance: 14%, 21%, 23% and 30% at d=3,5,7d=3, 5, 7 and 99, each 95% interval above zero (Table 5).

Device study. The frozen-calibration replay (old App. H, now App. K) is now complemented by prospective decoding with native errors on two processors from different vendors (General Response, Table R1); all 48 paired comparisons favour our decoder. It is not closed-loop surface-code correction: the devices run repetition memories, and the two-state regime is engineered through the idle time; on unengineered drift our two-regime filter gains little over frozen calibration and a Kalman filter is best (Table 7).

Libra, AlphaQubit, neural MWPM, and the public prefix. These methods replace the decoder rather than its noise model, and our package does not reproduce their training interfaces, so the information-matched comparison for our contribution is DGR (Q2). Our belief could be supplied to such decoders as an input, which Related Work now states. We agree that the public prefix is not an official benchmark; App. V reports it as a descriptive transfer check only.

We hope these results address your concern about the narrowness of the gain. Please let us know if anything remains unclear; we would be glad to provide further analyses.

Reviewer_X9hbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies sequential decoding in quantum error correction, in particular regarding the impact of sequential hidden state estimation on the actual decoding performance. A hidden Markov modeling has been employed to design the sequential estimation of hidden state which then affects the sequential fault modeling hence affects the actual decoding process. Experimental results show that the proposed approach outperforms existing baselines most of the time.

Critical strengths

The idea of using hidden Markov modeling to sequentially estimate the hidden state, to enhance sequential decoding process is sensible.

Critical weaknesses

While I believe that this work has a great potential, lack of theoretical justifications hide the value of this work. Specifically, it is not clear how the filter prediction and update in (2) have been designed; whether such update is optimal (e.g., follows Bayesian rule) or not. Furthermore, it is not clear whether the hidden Markov modeling is suitable for practical scenarios including the iid noise assumptions; and whether the experimental results are designed in favor of such modeling or not (i.e., whether the experimental results are carried out in real-world quantum hardwares -- it seems to me that the experimental results are carried out in a simulation-based modeling designed in favor of the authors' Markov modeling). Finally, despite the authors' emphasis in the introduction that "when a better estimate of the hidden noise state actually changes a logical correction", it is hard to parse from the current text whether this has been answered or not. It seems that there are some theoretical results in Sec. 4, but they are not actually answering the question directly.

In summary, weaknesses of this paper are: (i) lack of theoretical justification of whether the proposed sequential decoding rule can indeed lead to some better error correction; (ii) lack of experimental justification on real-world quantum hardware that goes beyond the synthetic experiments (I may have missed this point, but it seems to me that real-world experiment is missing).

Questions for authors

If the authors can clarify:

(i) Why the sequential state update rule (2) is a sensible choice? Is (2) optimal? For instance, one may aim at showing: First, consider a well-specified case where everything behaves as per the assumed hidden Markov model. Further suppose that the parameters of hidden Markov model follows some (prior) distribution. Now one can consider sequential Bayesian posterior update based on the sequential observations. Does (2) satisfy such property? If so, please clarify this aspect to justify (2); if not, why (2) is a good choice?

(ii) Suppose that (i) has been well addressed, i.e., we agree that (2) is a good sequential hidden state estimation. The authors nicely pointed out at the beginning of the paper that: better hidden state estimation may not lead to better decoding as the decoding is in discrete nature. But it is not clear whether this point has been answered in the paper or not. I think there should be some results like: "under some well-specified scenarios (e.g., scenarios assumed in Sec. 4), after T rounds, the proposed scheme achieves decoding error rate that is strictly smaller than all the baselines (static and current-only)". But it seems such theoretical results are missing in the current form.

(iii) It seems that the experimental results are designed in favor of the assumed hidden Markov model. At the end of the day, all the modelings are designed to mimic the real world; to better design the algorithm for the real world. That said, would the proposed algorithm work well under the misspecified case, including on real-world quantum hardware?

If all the above points are well addressed, then I would be happy to adjust my assessments.

2 comments
AuthorsOpenReview ↗

Response to Reviewer X9hb (1/2): Optimality of the Filter and a Decision-Level Theorem

Dear Reviewer,

Thank you for a clear and constructive review, and for telling us what would change your assessment. We took your three questions as our checklist; the changes made for you carry the margin tag R1 in the revised PDF.

Summary of changes for you

  1. (i) done. Theorem 1: Eq. (2) is exactly Bayes' rule for the HMM and the optimal causal estimator. Eq. (3) is the exact Bayesian filter when the HMM parameters follow a prior, learned without labels (Sec. 3.1, L165–173; App. A.1).
  2. (ii) done. Theorem 2 proves when, and by how much, the Bayes decoder with Markov beliefs is strictly better than the static and current-only Bayes decoders after HH rounds; Eq. (7) measures how much of that gain matching keeps; an exact Bayes-optimal decoder verifies both (Sec. 4, L288; Sec. 5.1, Fig. 4).
  3. (iii) done. 13 misspecified environments and prospective experiments with native errors on two real processors (Sec. 5.3, 5.5; General Response, Tables R1–R2).

We understand your main concern to be the missing theoretical justification, (i) and (ii); we address it first.

(i) Is Eq. (2) optimal? What if the HMM parameters follow a prior?

Short answer: yes. You are right that the original text never said so. The prediction step of Eq. (2) is the Chapman–Kolmogorov step of the chain; the update adds the log-likelihood ratio 2Yt/σ22Y_t/\sigma^2 in log-odds, which is Bayes' rule; a missing observation adds nothing. Theorem 1: in the well-specified model, bt=P(Zt=1∣Y0:t,M0:t)b_t = P(Z_t=1 \mid Y_{0:t}, M_{0:t}) exactly, so Eq. (2) has the smallest squared error and log loss of all causal estimators (current-only, any EMA, any recurrent network).

For the setting you describe, with a prior π0\pi_0 over θ=(a,σ)\theta=(a,\sigma), the exact sequential posterior is Eq. (3): P(Zt=1∣Ft)=∑θπt(θ)btθP(Z_t=1 \mid \mathcal{F}_t) = \sum_\theta \pi_t(\theta) b_t^\theta, a mixture of Eq. (2) filters weighted by their one-step predictive likelihoods. It is Bayes-optimal under the joint law, reduces to Eq. (2) for a point prior and converges to it as πt\pi_t concentrates (Proposition 1). Because θ\theta is a property of the device, unlabeled calibration streams sharpen the prior.

Takeaway: in the well-specified study the learned filter makes exactly as many errors as one given the true parameters (430 vs. 430, Table 3), because its posterior puts all but 10−1710^{-17} of its mass on them (App. B.1); under IID noise it reproduces the memoryless decoder decision for decision (559 vs. 559).

(ii) After HH rounds, is the decoding error provably and strictly smaller than the baselines?

Short answer: yes for the Bayes-optimal decoder, with an exact criterion. Let P1=P(L=1∣data)P_1 = P(L=1 \mid \text{data}) and let δ⋆\delta^\star be the Bayes decoder with the Markov environment. Theorem 2: for every horizon HH and every decoder δ\delta,

  • (a) R(δ⋆)≤R(δ)\mathcal{R}(\delta^\star) \le \mathcal{R}(\delta); in particular δ⋆\delta^\star is no worse than the static and current-only Bayes decoders;
  • (b) R(δ)−R(δ⋆)=E[∣2P1−1∣⋅1[δ≠δ⋆]]\mathcal{R}(\delta) - \mathcal{R}(\delta^\star) = \mathbb{E}\big[ |2P_1 - 1| \cdot \mathbf{1}[\delta \ne \delta^\star] \big] exactly;
  • (c) the gain over current-only is strict if and only if their decisions differ with positive probability where P1≠1/2P_1 \ne 1/2; for the probe of Fig. 3 this holds for every 0<a<1/20 \lt a \lt 1/2;
  • (d) under IID noise the two risks are equal;
  • (e) the gain is at most 1−2R(δ⋆)1 - 2\mathcal{R}(\delta^\star) (Corollary 1).

Part (b) is the classical excess-risk identity for binary decisions (App. A.2); applied here, it answers the question of our introduction: a better belief changes a logical correction exactly when it moves the logical posterior across 1/21/2, and the value of that change is the posterior confidence of those records.

To test this without approximation, we implemented the exact Bayes-optimal decoder for d=3d=3 with a hidden Markov environment, checked against brute-force enumeration:

exact Bayes decoder, d=3d=3, 4,096 held-out streamsstaticcurrent-onlyMarkovtrue state
failures548502417406

Takeaway: the strict gain over current-only is 2.08 pp in failure counts, 95% CI [1.34, 2.83], p<10−4p \lt 10^{-4}; the two sides of identity (b), computed from exact posteriors on these streams, agree to machine precision (both 2.2767%); over an 18-point parameter grid the gain is positive for every persistent environment and exactly zero under IID noise.

For matching decoders a general monotonicity theorem cannot hold, and the revision says so (L312). Instead, Eq. (7) gives an exact transfer identity: rolling MWPM with Markov beliefs recovers 79% of the expected-risk gap between current-only MWPM and the Bayes decoder (Sec. 5.1).

In the paper: Sec. 3.1 (Theorem 1, Eq. 3), Sec. 4 (Theorem 2, Corollary 1, Eq. 7), Sec. 5.1 and Fig. 4; all proofs in App. A.

AuthorsOpenReview ↗

Response to Reviewer X9hb (2/2): Model Mismatch, IID Noise and Real-Hardware Experiments

(iii) Are the experiments designed in favour of the HMM? Misspecification, IID noise, real hardware?

Short answer: the original evidence was indeed almost entirely well-specified, and you were right to press on it. We have added both tests you asked for; the gain survives misspecification, and on real hardware it holds where the device has regime structure and becomes small where it does not.

Misspecification (Sec. 5.3, Fig. 5a, Table 3; General Response, Table R2). The decoder always assumes the symmetric two-state Gaussian HMM; only the data generator changes, in 13 ways: faster switching, wrong σ\sigma, Student-tt or biased observations, state-dependent missingness, three and four latent levels, semi-Markov dwell times, rare bursts, a switching rate drifting 25×25\times, continuous OU drift, two spatial regions, and IID noise.

Takeaway: with parameters learned from unlabeled streams, failures fall significantly in all 12 non-IID environments (by 0.85–3.15 pp, every 95% interval above zero, 4,096 streams each). The weakest case is continuous drift; in the spatial case memory accounts for 2.42 of the 3.15 pp (Table R2, note).

IID noise. Here memory cannot help, and Theorem 2(d) says the gain must be zero. The learned parameter posterior concentrates on a=1/2a=1/2, and the decoder reproduces the memoryless decisions exactly (559 vs. 559 errors), so the method costs nothing when the Markov assumption fails in this way.

Real hardware (Sec. 5.5, Fig. 6, App. C; General Response, Table R1). This is a prospective test with native errors, not a replay:

  • ten code qubits of a superconducting processor store an XX-basis bit through an idle period; five separate spectator qubits run the same probe as telemetry, of which the decoder sees only 8–64 of 1,000 shots per epoch, and 30% of epochs carry none;
  • a two-state schedule of the idle time, never shown to the decoder, is the hidden environment, so every bit flip is the device's own decoherence or readout error;
  • likelihoods come from a short known-regime calibration, the switching probability is learned without labels, and everything is frozen before the 360 test epochs.
WK_C180 set A, 16 shotsfrozen calib.Kalmancurrent-onlyEMAoursoracle
decision error (%)5.144.754.074.073.753.56

Takeaway: our decoder closes 88% of the gap between frozen calibration and the regime oracle. On two further qubit sets, including a second vendor's processor (tianyan176), the result replicates, and all 48 paired comparisons favour our decoder with 95% intervals above zero.

Because the regime is engineered, this study tests whether beliefs learned from a few native shots improve decisions, not whether real drift is two-state. To test the latter, we also recorded 11.5 hours of unengineered drift (3.11M shots, Table 7). There temporal beliefs still beat current-only decoding, but the drift is continuous: the two-regime filter selects a single regime for most qubits and gains only 0.02–0.06 pp over frozen calibration, and a Kalman filter, the continuous-state analogue, is best (4.615% vs. 4.663% at 16 shots). We report this as it is.

Presentation. Following your comment that the text requires effort, the introduction now states the two gaps explicitly (L54–69), each method section opens by naming the gap it addresses (L147, L201, L257), and Sec. 5 opens with one protocol paragraph that defines the units and tests used throughout (L345–353).

We hope these results address (i)–(iii). Please let us know if any part remains unclear; we would be glad to run further checks.

Comments on the paper

6 comments
AuthorsOpenReview ↗

General Response (1/2): Summary of the Revision

Dear Reviewers and Area Chair,

Thank you for three careful reviews. They pointed at real gaps, and we have tried to answer each with a proof, an experiment or a corrected claim. This note summarizes the revision and, in the next comment, reports once for everyone the two results that all three reviews asked about; our replies to each review then go through every point.

What changed (in brackets: the reviews addressed)

  • Theory (Sec. 3.1, 4; App. A) [R1, R3]. Theorem 1: Eq. (2) is the exact Bayesian posterior and optimal among causal estimators; Eq. (3) is the exact filter when the HMM parameters have a prior and are learned from unlabeled data. Theorem 2 applies the classical excess-risk identity to streaming decoding: a strictness criterion, equality under IID noise and a headroom bound (Corollary 1). An exact Bayes-optimal decoder confirms it (Sec. 5.1).
  • Simulation (Sec. 5.2–5.4; App. B) [R1, R2, R3]: 41 new decoding studies with fresh seeds: the frozen operating point at 257 rounds, d=3d=3–99, four solver families on 16,384 records, 13 misspecified environments, sparse telemetry, a commit-window ablation and an executed DGR-style baseline.
  • Hardware (Sec. 5.5; App. C) [R1, R2, R3]: prospective decoding with native errors on two superconducting processors from different vendors (4.75M shots), and 11.5 hours of unengineered drift (3.11M shots).
  • Scoped claims [R2, R3]: the filter is presented as textbook (L137); the 512-record candidate intervention (Holm p=.46p=.46) is a mechanism, not a result (L321); the conclusion states that our simulations are phenomenological and our devices run repetition memories (L535).

The original decoder, Eq. (2), and its frozen confirmation (507 → 433 at d=3d=3, 158 → 131 at d=5d=5) are unchanged (Sec. 5.2). What the revision adds is what these classical tools yield for streaming QEC: when memory changes an irreversible decision (Theorem 2, Corollary 1, Eq. 7), commit-time beliefs, and, to our knowledge, the first prospective test on superconducting hardware of a decoder driven by temporal beliefs learned from telemetry and frozen before the test data (Table R1, next comment).

Numbering. Because the reviews cite the submission's numbers: old main Tables 1–3 and Figs. 4–6 are replaced by the new Tables 1–2 and Figs. 4–6; appendix tables moved by four (old 11, 15, 24, 30, 34 are now 15, 19, 28, 34, 38), appendix figures by one, appendix letters by three (old F, H, S are now I, K, V), and old Proposition 3 is now Proposition 5.

Reading the revised PDF. Red marks the text that answers the reviews, including the new numbers in the abstract and conclusion; for new theorems, figures and tables only the heading is red. Margin tags R1–R3 point each review to the paragraphs revised for it. Anonymized code for all new experiments is in the supplementary material.

The two results that all three reviews asked about follow in the next comment.

AuthorsOpenReview ↗

General Response (2/2): Hardware and Model-Mismatch Results Shared by All Reviews

Results shared by all three reviews (continued from the previous comment)

Table R1 — Prospective decoding with native errors (decision error in %, 16 telemetry shots per epoch, 360 test epochs per set, parameters frozen in advance; Table 6). The two-state regime is an idle-time schedule hidden from the decoder; every bit flip is the device's own.

qubit setfrozenKalmancurrent-onlyEMAoursoracleours vs. best baseline (pp)
WK_C180, A5.144.754.074.073.753.560.321 [0.224, 0.411]
WK_C180, B5.785.395.475.385.355.330.026 [0.013, 0.042]
tianyan17611.7711.4611.5311.4511.4411.420.017 [0.009, 0.026]

Takeaway: all 48 paired comparisons (3 sets × 4 telemetry budgets × 4 baselines) favour our decoder with 95% intervals above zero; at 16 shots it removes 52–96% of each baseline's excess error over the oracle. Margins are small where a device leaves little headroom. On unengineered drift (Table 7) the result is weaker: the drift is continuous, our two-regime filter gains only 0.02–0.06 pp over frozen calibration, and a Kalman filter is best.

Table R2 — Model mismatch (decoder always assumes two symmetric Gaussian states; learned parameters, commit-time beliefs; current-only minus ours in pp, 95% CI, 4,096 streams each; Table 3)

true environmentgaintrue environmentgain
5×5\times faster switching+1.73 [1.05, 2.42]semi-Markov dwell times+2.39 [1.59, 3.17]
observation noise σ=2.5\sigma=2.5+2.71 [1.90, 3.54]rare asymmetric bursts+1.20 [0.68, 1.71]
Student-t3t_3 observations+2.76 [2.05, 3.49]switching rate drifting 25×25\times+2.93 [2.20, 3.69]
biased observations+2.05 [1.32, 2.76]continuous OU drift+0.85 [0.05, 1.66]
state-dependent missingness+2.27 [1.51, 3.03]two spatial regions$^\dagger$+3.15 [2.32, 3.96]
three latent levels+1.47 [0.66, 2.27]IID noise0 (identical decisions)
four latent levels+1.42 [0.66, 2.20]

†^\dagger Ours keeps one belief per region; with one belief per region for the memoryless control too, memory alone gives +2.42 [1.66, 3.17] (App. B.2).

Takeaway: significant in all 12 non-IID environments, weakest under continuous drift; under IID noise the learned filter reproduces the memoryless decisions exactly.

We would be glad to discuss any point further.

AuthorsOpenReview ↗

Thank you: Theorem 3 on the risk of the proposed decoder vs. the current-only and static baselines

Dear Reviewer X9hb,

Thank you for reading our revision so closely and for spelling out what you hoped to see; we are glad the misspecification results were useful. Your point is fair. Our reply answered question (ii) for the optimal decoder, while you asked about the decoder we actually propose. We now answer it directly, along the lines you suggested, as Theorem 3 in the revised PDF (App. A.5; pointer after Theorem 2 at L309).

How is δ⋆\delta^\star related to our method? Your guess is right: δ⋆\delta^\star is the well-specified decoder with known switching probability aa and noise level σ\sigma. We should have said plainly that it is our own decoder with θ=(a,σ)\theta=(a,\sigma) set to the true θ0\theta_0. The proposed decoder δ^n\hat\delta_n runs the same filter but learns θ\theta: it updates a prior with nn unlabeled calibration streams of HH rounds and with the record being decoded (Eq. 3), and decides with P^1=∑θπn(θ∣x)P1θ\hat P_1=\sum_\theta \pi_n(\theta \mid x) P_1^\theta. So δ⋆\delta^\star is the target our method learns toward, and Theorem 3 measures how close it gets.

Extending Theorem 2 to the proposed decoder.

(a) Theorem 2(b) holds for any decoder, so for ours: R(δ^n)−R(δ⋆)=E[∣2P1−1∣⋅1[δ^n≠δ⋆]]≤2E∣P^1−P1∣\mathcal{R}(\hat\delta_n)-\mathcal{R}(\delta^\star)=\mathbb{E}\big[|2P_1-1| \cdot \mathbf{1}[\hat\delta_n\ne\delta^\star]\big]\le 2\mathbb{E}|\hat P_1-P_1|, since the two decisions differ only where P^1\hat P_1 and P1P_1 lie on opposite sides of 1/2. So our risk is the optimal risk plus a term set by how well θ\theta is learned.

(b) Your expectation that it eventually performs as well as the optimal decoder holds, with a rate. For a finite prior containing θ0\theta_0 (our grid) and identifiable θ0\theta_0, ER(δ^n)−R(δ⋆)≤2∑θ≠θ0π0(θ)/π0(θ0)ρθn\mathbb{E}\mathcal{R}(\hat\delta_n)-\mathcal{R}(\delta^\star)\le 2\sum_{\theta\ne\theta_0}\sqrt{\pi_0(\theta)/\pi_0(\theta_0)}\rho_\theta^{n}, where ρθ<1\rho_\theta\lt 1 is the Bhattacharyya affinity between one-stream laws: the gap closes exponentially after nHnH rounds of unlabeled data.

(c) This is the provable gain over the baselines that ignore the Markov structure. Theorem 2(c) says exactly when the current-only decoder (or the static one) is strictly worse than δ⋆\delta^\star; whenever it is, there is an n0n_0 beyond which our learned decoder is strictly better than that baseline. Without memory (a=1/2a=1/2) it converges to the current-only decoder, which is then optimal.

Bernstein–von Mises: Markov and hidden Markov. Thank you for pointing us to Borwanker et al. (1971); it is now cited in App. A.5. Their result is for Markov processes whose states are observed; our regime ZtZ_t is seen only through noisy readings YtY_t, so our model is a hidden Markov model. Your idea applies in two ways. (i) The calibration streams are independent replicates, so the stream-level likelihood is i.i.d. and the classical Bernstein–von Mises theorem applies (van der Vaart, 1998, Thm. 10.1). This is the route "similar to the static case" you describe: for a continuous prior it gives contraction at rate n−1/2n^{-1/2} and, under a margin condition, excess risk O(1/n)O(1/n) (Audibert & Tsybakov, 2007). (ii) Learning within one long stream needs the hidden-Markov counterpart of Borwanker et al., given by de Gunst & Shcherbakova (2008) building on Bickel et al. (1998); we cite it without deriving rates for that case.

Your "after H rounds" case, computed exactly. Using the exact Bayes decoder of Sec. 5.1 (d=3d=3, H=65H=65), 8,192 held-out streams, a uniform prior on 98 grid points and 20 independent calibration draws per nn:

unlabeled calibration streams nn0864256
excess risk over δ⋆\delta^\star (pp)0.0370.0030.00070.0002
gain over current-only (pp)2.182.212.212.21

(Bayes risks: δ⋆\delta^\star 10.36%, current-only 12.57%, static 13.59%.) At n=0n=0 the decoder learns θ\theta only from the 65 rounds of the record it is decoding, yet it is within 0.04 pp of δ⋆\delta^\star: 864 errors, against 860 with the true parameters and 1,049 with current-only beliefs (paired reduction 2.26 pp, 95% CI [1.72, 2.78], p<10−4p\lt 10^{-4}). It beats both baselines in all 200 calibration draws. With the truth off the grid (a=.013a=.013, σ=1.37\sigma=1.37) the excess levels off at 0.008 pp, the grid's own error, and the gain stays at 2.0 pp.

Scope. Theorem 3 concerns the Bayes decoder; for matching no general theorem of this kind can hold (Sec. 4). There, learned and true parameters give identical failures in every column of Table 1 (397, 111, 1,182, 424 for both), and Eq. (7) measures the remaining gap.

Thank you again for pressing on this. It made clear what Theorem 2 is for, and the paper now gives a guarantee for the decoder we actually propose. If you had a different formulation in mind, we would be glad to adapt it.

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.