{"id":"990c75fe-4ff9-4a1f-a2c2-06c4ace81fc2","arxiv_id":"2504.14643","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Detector error model probabilities can be recovered from syndrome histories through a Walsh-Hadamard transform, giving a closed-form inverse for individual and aggregated DEM events.","lead":"This paper derives a mathematical method for extracting error event probabilities in quantum error correction from recorded syndrome data, using a Fourier-style transform. It also shows how to estimate aggregated error classes and sketches sparse algorithms for large systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inversion is exact only within the DEM model; the paper does not test what Eq. 31 returns for real non-DEM noise, so the abstract estimation claim lacks robustness evidence.","rationale":"The mathematical derivation of Eq. 31 appears sound: the Walsh-Hadamard diagonalization of the commuting stochastic matrices is standard, and the pseudoinverse inversion on the support of W is correctly handled. I independently spot-checked the N=2 case and the sign conventions, and the algebra is consistent. The reader's weakest assumption, that real QEC noise may violate the DEM independence/stationarity axioms, is exactly the condition on which the central claim's practical validity rests. The concern is not an internal inconsistency but a correctness risk for the abstract's estimation promise: under model misspecification, the closed-form inversion can return nonphysical probabilities, as my N=2 counterexample shows. The paper explicitly defers robustness studies, sparse-algorithm demonstration, and detailed statistics to future work, and supplies no numerical experiments. These gaps justify the CONDITIONAL verdict: the core derivation is a genuine contribution, but the promise of estimating individual DEM event probabilities from experimental syndrome data is not yet supported. No change to the reader's verdict is needed.","tokens_in":72479,"tokens_out":20924,"duration_ms":182031,"concrete_test":"Simulate a non-DEM syndrome-data generator for a small repetition code, e.g., a two-cycle code with a Markovian error process where the probability of an error at cycle t depends on whether an error occurred at cycle t-1, or with p_s drifting linearly across cycles. Generate K detector histories, compute empirical polarizations, and apply Eqs. 31-34 to recover per-event probabilities. Then compare the recovered values to the true time-averaged event probabilities and check for nonphysical outputs (negative probabilities, probabilities > 1/2, or estimates outside the error bars). If such mismatches occur, the abstract's estimation claim needs an explicit robustness qualifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inversion (Eqs. 31-34) relies on the Section 1.2 DEM axioms: errors are independent, each event has a fixed detector support, identical supports are merged, and probabilities are constant across cycles. These axioms are what justify the factorization in Eq. 23, ⟨z_y⟩ = ∏_s d_s^{y·s}, and hence the linear relation ω = W a. If syndrome data are generated by time-correlated errors, crosstalk, leakage, or slowly drifting rates, the empirical polarizations are not of this product form, and applying Eq. 31 yields a vector a that need not correspond to any physical DEM. A concrete N=2 example shows this is not merely abstract: the valid probability distribution with ⟨z1⟩=⟨z2⟩=0.5 and ⟨z12⟩=0.1 (outcome probabilities 0.525, 0.225, 0.225, 0.025) is not generated by any DEM, yet Eq. 34 returns a negative value for p_{1,2}. The paper explicitly assumes an unknown DEM in Section 1.2 and never tests robustness to model violation, nor does it provide code or numerical experiments that would expose such failures. Since the abstract promises estimation of DEM event probabilities from syndrome data, this gap is load-bearing for the practical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a closed-form, decoder-free method for estimating detector error model (DEM) event probabilities from syndrome-extraction data. It defines parities z_y = (-1)^{x·y}, polarizations <z_y>, and depolarizations ω_y = -ln<z_y>; under the DEM axioms of Section 1.2 (independent errors with fixed detector supports, merged identical supports, constant probabilities), the vector of depolarizations is shown to be a linear function of event attenuations a_s = -ln(1-2p_s), namely ω = W a (Eq. 25). The authors derive the Moore-Penrose inverse W^+ and obtain the main inversion (Eqs. 31-34): event attenuations, decay factors, and probabilities are given explicitly by a scaled Walsh-Hadamard transform of the depolarization vector, involving only empirical polarizations. They show that reduced-DEM aggregations reproduce and clarify the p_ij coefficients of Refs. [5,7], give explicit low-weight aggregation formulas (Eqs. 44-48), propose Monte Carlo averaging for high-weight quantities such as total attenuation, and sketch two sparse-DEM algorithms based on submatrix inversion and lattice pruning (Section 4).","tokens_in":72739,"tokens_out":20093,"duration_ms":154270,"significance":"The central derivation is clean and parameter-free: the inversion from empirical polarizations to DEM event probabilities involves no fitted constants and no decoder, and the key identities (Eqs. 25-34) are derived step by step rather than asserted. The paper gives explicit, falsifiable closed-form predictions, and it makes a genuine conceptual contribution by explaining what the widely used p_ij coefficients actually estimate (aggregated probabilities over event classes) and by unifying earlier partial results [5,7,11] in a single algebraic framework. If the results hold up, the formalism is likely to become a standard reference for DEM estimation and calibration. The main weaknesses are equally clear: the paper is purely algebraic, with no numerical experiments, no code, and the sparse algorithms that carry the realistic-N claims are explicitly deferred to future work. The practical claim that Eq. 31 can be used to infer DEMs from data therefore rests on unvalidated finite-sample and model-robustness behavior, and the uncertainty formulas (Eqs. 15-16) contain a mathematical error.","major_comments":[{"comment":"The standard error formulas are wrong by a square root. Since Eq. 14 correctly gives the variance σ^2(z_y) = 1 - <z_y>^2, the standard error of the mean is sqrt(1-<z_y>^2)/sqrt(K), but Eq. 15 states σ_<z_y> = (1-<z_y>^2)/sqrt(K), and Eq. 16 states σ_ωy = (1-<z_y>^2)/(<z_y> sqrt(K)). The correct form of the latter is sqrt(1-<z_y>^2)/(<z_y> sqrt(K)). As written, the formulas overstate the error bars by a factor that diverges as <z_y> approaches zero, which is exactly the regime the paper later identifies as problematic in Sec. 3.4.1. Because the paper recommends these error bars as the practical uncertainty quantification for the whole estimation pipeline, this needs to be corrected before acceptance.","section":"Sec. 2.1.1 (Eqs. 15-16)"},{"comment":"The abstract promises that the paper shows how to estimate probabilities of individual DEM events from data, but for realistic N the individual-event estimators require the Section 4 algorithms, which the paper explicitly defers: 'We defer the work of demonstrating and analyzing these algorithms to future work' (Sec. 4). There are no numerical experiments anywhere in the manuscript, no finite-sample study, and no discussion of how to choose the threshold wmax that the sparse estimates depend on, even though wmax is a free input to the proposed procedures. Consequently, the central practical claim is not demonstrated in the regime the paper itself identifies as relevant (large N, local low-weight DEMs). I request simulation-based tests, for example sampling from the N=4 DEM of Sec. 3.1.1 and from a local weight-<=2 DEM with N = 20-100, comparing Eq. 31 and the Section 4 algorithms against ground truth with proper error bars.","section":"Secs. 3.4 and 4; abstract"},{"comment":"The inversion is exact only for distributions of the product form (Eq. 23) generated by a DEM; for data outside the DEM family, Eq. 31 does not return a physical DEM, and the paper offers no analysis, warning, or test. A concrete example: the valid N=2 distribution with outcome probabilities (0.525, 0.225, 0.225, 0.025) has <z1>=<z2>=0.5 and <z12>=0.1, yet no DEM generates it (the implied decay factors would require d12^2 = 2.5, impossible since |d12| <= 1), and Eq. 34 returns p_{1,2} = 1/2 - (1/2)sqrt(0.25/0.1) ≈ -0.29. Real syndrome data will contain time-correlated errors, crosstalk, and drift that violate the Section 1.2 axioms, so the statement at the end of Sec. 2.3 that the linear relationship 'can be inverted to learn or infer attenuations from empirical depolarizations' needs either a robustness bound with respect to distance from the DEM family or an explicit demonstration on simulated non-DEM noise that the estimator fails gracefully. As it stands, this is a load-bearing gap for the estimation claim.","section":"Sec. 1.2 axioms; Eqs. 31-34"}],"minor_comments":[{"comment":"Both equations contain the typo '1 - 2<xi> - 2<xi> + 4<xixj>' with the second expectation equal to the first; Eq. 9 should read '1 - 2<xi> - 2<xj> + 4<xixj>' and Eq. 37 should use <x1> and <x2> consistently.","section":"Eqs. 9 and 37"},{"comment":"The Monte Carlo estimator expression 'a_s = 2<[(-1)^{y·s}+1 ω_y]>_y' is ambiguous and, as typeset, incorrect; the intended formula is a_s = 2<(-1)^{y·s+1} ω_y>_y, with the +1 inside the exponent. Please fix the typesetting.","section":"Sec. 3.4"},{"comment":"The stated count of additional events, C(N + wmax - 2|E| - 1, wmax - |E|), does not equal the sum over w = 1 to wmax - |E| of C(N - |E|, w) in general. For example, with N=5, |E|=1, and wmax=4, the sum is 4+6+4=14 events, while the formula gives C(6,3)=20. The counting should be corrected or qualified as an upper bound.","section":"Sec. 4.1 (footnote 5)"},{"comment":"The claim that Eq. 36 is 'Equation (11) from the Supplementary Material to Ref. [7], mod a little algebra' would be easier to verify if the intermediate algebra were shown; as written, the reader must obtain the supplement to check the match.","section":"Sec. 3.1.1"},{"comment":"The paper never mentions that Eq. 34 can produce estimates outside [0,1] even for data drawn from a DEM when K is finite, and no guidance is given on how to handle such estimates (e.g., clipping, reframing as a model-violation diagnostic, or reporting the unconstrained attenuation instead). A brief remark would substantially improve practical usability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a theory-only paper: the central linear-algebra derivation (Eqs. 25-34) is sound and I could not find an error in the main inversion, but there is no numerical validation at all, and the sparse algorithms that carry the realistic-N claims are explicitly deferred by the authors. I have recommended major revision with a request for simulation-based tests, including at least one non-DEM robustness experiment of the kind described in my third major comment. If the journal accepts theory-only papers, the editor may wish to weigh whether the 'estimation from data' framing in the abstract is proportionate to what is actually proven; the algebraic result and the aggregation interpretation of p_ij are the solid contributions. Novelty is moderate, as the authors themselves acknowledge that their results may be unsurprising to the p_ij literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central result is Eq. 31: attenuations are recovered from depolarizations by a scaled Walsh-Hadamard transform. That is clean, correct, and useful. The paper also explains what the pij coefficients actually measure — aggregated attenuations over event classes — and gives a lattice-pruning sketch for sparse DEMs. The derivation is careful, and the treatment of prior work is honest; Refs [5,7,11] are handled fairly rather than buried.\n\nThe soft spots are real but mostly not fatal. Eq. 15 is missing a square root: the standard error of a polarization is sqrt(1−⟨z⟩²)/√K, not (1−⟨z⟩²)/√K. Eq. 16 inherits the same error. It is a typo in an aside, but error bars are where readers will trust the paper, so it should be fixed.\n\nThe bigger caveat is model violation. The inversion is exact only if the data are actually generated by a DEM. For an arbitrary empirical distribution, Eq. 34 can return negative probabilities. A concrete N=2 case with ⟨z1⟩=⟨z2⟩=0.5 and ⟨z12⟩=0.1 (outcome probabilities 0.525, 0.225, 0.225, 0.025) gives p_{1,2} ≈ −0.29. The paper assumes the DEM model and never tests what happens when it is wrong. That is not a flaw in the math, but it is a gap between the abstract's estimation promise and the tool actually delivered. A referee should ask for a short numerical sanity check on synthetic non-DEM data, or at least an explicit statement about what the inversion returns under model mismatch.\n\nThe sparse algorithms are sketches, not implementations, and the paper says so. That limits practical reach but not the value of the framework.\n\nOverall: a solid theoretical contribution, not groundbreaking — the pij formula was known — but the unified inversion, the class semantics, and the lattice view are new and worth having. The math is sound, the writing is clear, and limitations are mostly admitted. I would send it to a serious referee, and I would tell the authors to fix the error-bar formula and add a paragraph about model mismatch before I would rely on it for hardware characterization.","headline":"Clean Fourier inversion for DEM estimation, with honest prior-work context, plus an error-bar typo and an unaddressed model-violation caveat.","tokens_in":73231,"tokens_out":4181,"would_cite":true,"duration_ms":38009,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that every detector error model event probability is identifiable in closed form from syndrome data, via a scaled Walsh-Hadamard transform of empirical polarizations.","keywords":["detector error model","syndrome data","Walsh-Hadamard transform","polarization","attenuation","quantum error correction","noise estimation","decoder graph"],"falsifier":"Simulate many syndrome histories from a known DEM, estimate all polarizations, and apply the paper's inversion formulas; if the reconstructed event probabilities do not match the input within finite-sample error, the identity itself is wrong. Then repeat with time-correlated or crosstalk errors: mismatches there would show the DEM assumptions, not the algebra, are the limiting factor.","tokens_in":1671,"feed_emoji":"⚛️","tokens_out":3968,"duration_ms":76132,"temperature":0.7,"pith_summary":"The paper tries to show that a detector error model (DEM) describing syndrome extraction noise can be learned directly and exactly from syndrome data, without running a decoder and without iterative fitting. It introduces the attenuation of a DEM event as the key variable and proves that the vector of all attenuations is obtained by a scaled Walsh-Hadamard transform of the vector of empirical depolarizations. If this is correct, every DEM event probability, and every aggregated class of event probabilities, has an explicit closed-form estimator. That matters because accurate DEMs improve decoding and logical error rates, and current methods are limited by degeneracy or require numerical Bayesian inference.","feed_headline":"Every DEM event probability is computable in closed form from syndrome data","feed_subtitle":"No decoder or fitting is needed: a scaled Walsh-Hadamard transform inverts the noise model.","key_machinery":"The central object is the linear relation between depolarizations and attenuations, where the matrix entry records whether a DEM event flips a given parity. Because all DEM transition matrices commute and become diagonal in the Walsh-Hadamard (polarization) basis, this matrix can be inverted by its Moore-Penrose pseudoinverse, which is proportional to the Walsh-Hadamard matrix with its first row and column removed. The same diagonalization shows that attenuations, unlike probabilities, add linearly under aggregation, which is what makes reduced DEMs and the lattice-pruning algorithm work.","core_discovery":"The paper's first main result is Eq. 31, which states that the vector of all event attenuations is obtained by a scaled Walsh-Hadamard transform of the vector of depolarizations. Unpacked, this gives explicit formulas for the attenuation, decay factor, and probability of any DEM event in terms of observable polarizations of syndrome data. The paper also shows how to estimate aggregated classes of events using only low-weight depolarizations, and how to recover sparse DEMs by pruning a lattice of event classes. A central conceptual payoff is an answer to what the widely used p_ij coefficients quantify: they are aggregated probabilities of all DEM events that flip both detectors i and j.","pith_inferences":["The paper leaves implicit that the same closed-form inversion could be applied to time-windowed syndrome data, offering a direct route to detecting time-varying or nonstationary DEMs by comparing attenuation estimates across windows.","Because the inversion is exact under the DEM model, a natural benchmark is to compare its estimates against decoder-counting frequencies in regimes where degeneracy is absent; agreement there would validate the algebra, and disagreement would localize the failure to the DEM assumptions.","The lattice-pruning algorithm suggests a concrete testable extension: in a surface-code or repetition-code experiment, clutter of nonzero weight-2 classes that cannot be explained by genuine weight-2 events should appear as higher-order hypercliques, and the paper's method would distinguish those cases.","The use of sparse Walsh-Hadamard transform ideas points toward a fully blind sparse-DEM estimator that does not require a candidate event list, though the paper only sketches the candidate-based version."],"forward_implications":["Each DEM event probability can be estimated from the full set of empirical polarizations with no decoder and no fitting, in closed form.","Aggregated attenuations of event classes can be computed from low-weight depolarizations, so useful DEM estimates remain possible when N is large enough that high-weight polarizations are statistically zero.","The total DEM attenuation, a scalar measure of all noise in syndrome extraction, can be estimated efficiently via a Monte Carlo average over depolarizations.","If the DEM is sparse and low-weight, lattice pruning can identify which individual events have nonzero probability and estimate them, extending the p_ij method into a data-driven decoder-graph construction with higher-order hyperedges.","The derivation of the p_ij coefficients reinterprets existing calibration results and tells experimenters what those numbers actually mean."],"supporting_citations":[{"why":"Proposes decoding syndrome data into events and estimating event frequencies, the baseline decoder-based approach that degeneracy limits and that the paper's closed-form inversion replaces.","marker":"[2]"},{"why":"Derives the p_ij relations from empirical detector moments, which the paper rederives and interprets as aggregated DEM event probabilities.","marker":"[5]"},{"why":"Uses empirical moments to estimate DEM properties and presents the p_ij formula; the paper reproduces and explains that formula.","marker":"[7]"},{"why":"Presents formulas closely related to the paper's main result without derivation, supplying the motivation for the derivation given here.","marker":"[11]"},{"why":"Investigates learning Pauli channels from syndrome statistics, the closest prior approach that the paper contrasts with DEM-based identifiability.","marker":"[13]"},{"why":"Shows Pauli channels can be estimated from syndrome measurements, establishing the context that DEMs are the uniquely identifiable objects.","marker":"[14]"},{"why":"Provides the generalized Fourier-analysis perspective, including the Walsh-Hadamard transform, that carries the paper's derivation.","marker":"[24]"}],"fun_headline_variants":["Closed-form DEM estimation from syndrome data","Walsh-Hadamard transform unlocks detector error models","No fitting required: estimate DEM events directly","Syndrome data directly yields all DEM event probabilities"],"cache_read_input_tokens":75392,"weakest_assumption_plain":"The inversion is only as good as the DEM model's assumption that errors are independent, each error always flips the same fixed set of detector bits, and every event's probability is constant across cycles.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form DEM estimation from syndrome data","Walsh-Hadamard transform unlocks detector error models","No fitting required: estimate DEM events directly","Syndrome data directly yields all DEM event probabilities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1466,"prompt_tokens":729,"completion_tokens":737,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":345,"completion_tokens_details":{"reasoning_tokens":679}},"tokens_in":345,"tokens_out":737,"duration_ms":6622,"temperature":1.0,"reasoning_tokens":679,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:44:02.265317+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate many syndrome histories from a known DEM, estimate all polarizations, and apply the paper's inversion formulas; if the reconstructed event probabilities do not match the input within finite-sample error, the identity itself is wrong. Then repeat with time-correlated or crosstalk errors: mismatches there would show the DEM assumptions, not the algebra, are the limiting factor.","supporting_citations":[{"cited_title":"Adaptive weight estimator for quantum error correction in a time-dependent environment","cited_arxiv_id":null,"evidence_quote":"Derives the p_ij relations from empirical detector moments, which the paper rederives and interprets as aggregated DEM event probabilities."},{"cited_title":"Optimal noise estimation from syndrome statistics of quantum codes","cited_arxiv_id":"2010.02243","evidence_quote":"Investigates learning Pauli channels from syndrome statistics, the closest prior approach that the paper contrasts with DEM-based identifiability."}],"review_version":1}