{"id":"6828fe78-6264-4aea-9d8d-a37c343eda4d","arxiv_id":"2412.13739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tensor-network-based maximum likelihood decoder is constructed for quantum error correction under process-tensor noise, with an MPS approximation demonstrated on the five-qubit and Steane codes.","lead":"This paper constructs maximum likelihood decoders for quantum error correction codes when the noise is correlated in time and space, described by a process tensor. It demonstrates the method on small codes and proposes a matrix product state approximation to make the calculations scalable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (45) lacks an explicit normalization for chi_HS, so p_fail is not established as a probability.","rationale":"The reader's weakest_assumption identifies exactly the same point: chi_HS is asserted to be a success probability but never proven normalized, so p_fail in Eq. (45) is not a demonstrated logical failure rate. This is the most load-bearing concern because the central claim is that the tensor-network contraction computes the ML decoder and its logical failure rate; if the reported quantity is not a probability, the headline claim is undermined. The argmax decoder itself is likely unaffected, since maximizing chi_HS over L for a fixed s is equivalent to maximizing the normalised process fidelity up to a syndrome-independent factor, so the decoder choice is probably still correct. However, the absolute values of p_fail, the plots, and any threshold or performance statement drawn from them would be wrong without the missing normalization. This is a fixable but essential gap, consistent with a conditional verdict. I do not see a stronger competing concern: the MPS approximation claims are explicitly regime-dependent and supported by Table 2 in the low-noise case, and the paper itself flags the distance metric limitation. The normalization issue is therefore the primary check needed before the central claim can be accepted.","tokens_in":21139,"tokens_out":11239,"duration_ms":103418,"concrete_test":"Analytically evaluate Eq. (43) for the five-qubit code in the ideal case of zero noise and perfect syndrome measurements, with recovery L = I for every syndrome. Compute S = sum_s chi_HS(I,s). If S is not equal to 1, then Eq. (45) requires the replacement p_fail = 1 - (1/S) * sum_s chi_HS(Lbar(s),s). Then recompute Fig. 9 with this normalization and check whether, at J_NM = J_CT = 0, p_fail tends to 0 as perr -> 0. A non-trivially positive floor would indicate that the inner-product metric has a normalization or interpretation error beyond a simple constant factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.2, chi_HS(L,s) is defined in Eq. (43) as the Hilbert-Schmidt inner product between the strategic-code process for syndrome s and recovery L, and the vectorized Choi state of the identity channel. The Choi states throughout the paper are built from the unnormalised maximally entangled state |Phi+> = sum_i |ii> (Eq. 9), whose norm-squared is d^2 = 4 for a logical qubit. For a subnormalised operation describing syndrome s and recovery L, chi_HS(L,s) is therefore proportional to d^2 * p(s) * F(L,s), where p(s) is the syndrome probability and F(L,s) is the normalised process fidelity of the conditional channel. Eq. (45) simply sums chi_HS(Lbar(s),s) over syndromes and sets p_fail = 1 - sum_s chi_HS. Unless an explicit 1/d^2 normalization is inserted, the sum equals d^2 times the true total success probability, so p_fail is not a valid logical failure rate. The paper never states such a normalization, does not prove that sum_s chi_HS = 1 in the noiseless limit, and in the same section concedes that the alternate distance metric of Eq. (49) does not satisfy probability axioms (text after Eq. 49). Because the central claim is that the single-shot tensor-network contraction yields the logical failure rate, this missing normalization is load-bearing: the reported p_fail values (Fig. 9, and the framework in general) are not established as probabilities.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper integrates the process-tensor framework with stabilizer quantum error correction, following the recently proposed 'strategic code' idea, and constructs a decoder by maximizing a closeness metric between the effective logical process and the identity channel. The authors define two metrics, the Hilbert-Schmidt inner product in Eq. (43) and a channel-distance metric in Eq. (47), and use them to define the logical failure rate in Eqs. (45) and (49). They implement the resulting tensor-network contractions numerically for the five-qubit code under depolarizing noise, non-Markovian Heisenberg interactions, and ZZ crosstalk, and they propose a matrix-product-state (MPS) approximation of the process tensor and tester for larger codes such as the Steane code. Numerical results compare exact tensor-network contraction with MPS approximations at various bond dimensions and report timing and fidelity data.","tokens_in":21438,"tokens_out":9636,"duration_ms":96826,"significance":"If the central claim is correct, the paper provides a general tensor-network recipe for constructing decoders and evaluating logical failure rates for quantum error correction under non-iid, spatiotemporally correlated noise, going beyond the usual iid Pauli-error assumptions. This would be a useful step toward realistic noise-aware decoding and code evaluation. The paper has clear strengths: it gives explicit tensor-network diagrams and equations for the full contraction, it implements exact and approximate numerical experiments with reproducible open-source tensor libraries, and it carefully identifies limitations such as the non-probabilistic nature of the distance metric and the regime-dependence of the MPS approximation. The main reservation is that the quantity called the logical failure rate in Eq. (45) is not proven to be a normalized probability, so the reported numerical failure rates are not yet established as probabilities.","major_comments":[{"comment":"The central output of the paper, the logical failure rate p_fail in Eq. (45), is not established as a probability. The Choi states used throughout are built from the unnormalized maximally entangled state in Eq. (9), so the Hilbert-Schmidt inner product in Eq. (43) carries a dimension-dependent normalization factor. For example, for a logical qubit the identity Choi state has norm-squared d^2 = 4, and an identity strategic-code process would not yield sum_s chi_HS = 1 without an explicit normalization. The paper neither inserts such a factor nor proves that sum_s chi_HS(Lbar(s), s) equals 1 in the noiseless limit. This is load-bearing because the single-shot contraction is claimed to yield the logical failure rate, and the values in Figs. 9 and Table 2 depend on this identification. Please add an explicit normalization argument, or re-label chi_HS and p_fail as an unnormalized score and the corresponding heuristic failure estimate.","section":"Sec. 3.2, Eqs. (43)-(45)"},{"comment":"The paper acknowledges after Eq. (49) that the channel-distance metric does not satisfy the axioms of probability, yet the dashed curves in Fig. 9 are still plotted and described as logical failure rates. As a heuristic performance indicator this is acceptable, but the text should either relabel these curves as a non-probabilistic score or provide a quantitative statement of how they relate to a true failure probability, such as an upper or lower bound. Without this, the comparison between the solid and dashed curves in Fig. 9 is ambiguous.","section":"Sec. 3.2, Eq. (49) and Fig. 9"},{"comment":"The 'exact' reference values in Table 2 are compared with the MPS estimates p_est and p_perf, but the text does not state explicitly which of the two metrics, Eq. (45) or Eq. (49)/(54), is used to generate the exact column. Since the two metrics are not interchangeable, this should be stated in the table caption or in the surrounding text to allow the reader to verify the 'accuracy comparable to exact contraction' claim.","section":"Sec. 4.3.2, Table 2"}],"minor_comments":[{"comment":"The notation for k is inconsistent with the earlier use of k for the number of logical qubits. In Sec. 4.3.1 the classical system is said to consist of k bits and the Steane code is described with k = 6, whereas earlier [[n,k,d]] notation uses k = 1 for the Steane code. Please rename one of these quantities (e.g., use m for the number of syndrome bits).","section":"Sec. 4.3.1"},{"comment":"There are several typos and infelicities, including 'concrete the better construction' and 'algoirhtms'. A careful language pass is needed.","section":"Conclusion"},{"comment":"The manuscript would benefit from a data and code availability statement, since the tensor-network contractions are implemented with quimb and cotengra and the numerical claims are central to the paper.","section":"General"},{"comment":"The description of Fig. 8(b) says the contraction returns the success probability 'as a matrix of dangling legs on the first four-round syndrome measurements and logical recovery operation.' This wording is unclear; the figure captions should more precisely specify which indices are left open and which are contracted.","section":"Fig. 8 and Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core tensor-network construction appears original and useful. The normalization issue in Eq. (45) is the main technical obstacle; it is likely fixable with a short proof or a clear normalization convention, so I recommend major revision rather than rejection. I do not see grounds for questioning the novelty or the authors' good faith. Please also ask the authors to clarify which metric is used for the exact reference in Table 2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful methods paper with a real but modest contribution, and one load-bearing gap in the presentation. The genuinely new part is the single-shot tensor-network contraction that computes the ML decoder and the logical failure rate for stabilizer codes under process-tensor noise, plus the MPS approximation with truncation based on syndrome-outcome bias. The numerical work is honest: five-qubit and Steane codes, exact contraction benchmark, convergence in bond dimension, clear statements about regime dependence. I believe the central idea is right.\n\nThe problem is Eqs. (43)-(45). chi_HS is defined as a Hilbert-Schmidt inner product with the vectorized Choi state of the identity channel, and all Choi states in the paper use the unnormalized maximally entangled state |Phi+> = sum_i |ii>. The identity Choi state then has operator norm squared d^2 (4 for a logical qubit). As written, chi_HS carries a d^2 factor and Eq. (45) does not define a normalized probability unless that factor is divided out. The paper never supplies the normalization, never checks the noiseless limit, and concedes in the same section that the alternative metric (49) is not a probability. This matters for the paper's central claim that the contraction yields the logical failure rate. The decoder itself, which only ranks chi_HS over L for fixed s, may survive if the normalization is a global constant, but the reported absolute p_fail values are not established.\n\nMinor soft spots: no code or data artifact list, so the numerical claims are not independently checkable; the exact contraction benchmark is not validated against Monte Carlo or an independent decoder; and the low-noise floor discussed in Sec. 4.2 is explained but would benefit from a quantitative statement as p_err goes to zero. None of these are fatal. The citations look appropriate, and the MPS approximation is a fair extension of existing PT-MPO work rather than an overclaim.\n\nWho should read this: people working on decoders for correlated/non-Markovian noise, and anyone benchmarking tensor-network methods for QEC. It deserves a serious referee. I would send it out, with the normalization question as the clear revision priority.","headline":"A genuinely new tensor-network construction for ML decoding under process-tensor noise, but the success metric in Eq. (45) is missing a normalization that the authors need to fix before the reported failure rates can be trusted.","tokens_in":21984,"tokens_out":10605,"would_cite":false,"duration_ms":102725,"reading_group":"maybe","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 shows that for a stabiliser code driven by process-tensor noise, the maximum-likelihood decoder and its logical failure rate follow from one tensor-network contraction, and that an MPS approximation reproduces exact results in…","keywords":["quantum error correction","process tensor","non-Markovian noise","maximum likelihood decoder","tensor network","matrix product state","strategic code","crosstalk noise"],"falsifier":"Simulate the same five-qubit code under the same process-tensor noise model twice: once by contracting the paper's tensor network for $p_{\\mathrm{fail}}$, and once by a direct state-vector Monte Carlo simulation that applies the recovered correction and counts how often the logical state returns to the codespace. If the two logical failure rates disagree, the interpretation of $\\chi_{HS}$ as a success probability is wrong. A cheaper check is to compute $\\sum_{\\vec{s}} \\chi_{HS}(\\bar{L}(\\vec{s}),\\vec{s})$ for increasing noise strengths and test whether it stays in $[0,1]$ and decreases monotonically as expected of a probability.","tokens_in":20925,"feed_emoji":"⚛️","tokens_out":9115,"duration_ms":75038,"temperature":0.7,"pith_summary":"Quantum error correction usually assumes errors arrive independently and identically distributed, but real hardware suffers correlated, non-Markovian, and crosstalk noise. This paper claims that when the device noise is represented by a process tensor—a complete multi-time description of the noise's correlations—the optimal (maximum-likelihood) decoder for a stabiliser code can be obtained by a single contraction of the tensor network formed from the process tensor, the syndrome-measurement tester, and the recovery channel; the same contraction directly yields the logical failure rate. The authors implement this exactly for the five-qubit and Steane codes under a noise model combining depolarising errors, non-Markovian bath couplings, and ZZ crosstalk. For the Steane code they also approximate the process tensor and tester as matrix product states and find that, in low-noise regimes, the resulting decoder matches exact contraction results in substantially less time. If correct, the method gives a general, code-agnostic way to evaluate and design QEC codes under realistic correlated noise.","feed_headline":"One tensor contraction computes decoders for correlated quantum noise","feed_subtitle":"The method turns arbitrary process-tensor noise plus syndrome data into a maximum-likelihood recovery rule in one pass.","key_machinery":"The load-bearing objects are the process tensor and the tester. The process tensor is a multi-time generalisation of a quantum channel: a single Choi state over a chain of input and output Hilbert spaces that encodes the device's full spatiotemporal noise correlations, including memory and crosstalk. The tester is the corresponding generalisation of a measurement instrument: a sequence of completely positive maps with a classical memory, which is what lets the recovery operation depend on the earlier syndrome outcomes. The argument combines these with the encoder and decoder through the link product, producing one tensor network whose contraction equals $\\chi_{HS}(L,\\vec{s})$; the maximum-likelihood decoder is obtained by reading off the $L$ that maximises this quantity for each syndrome, and $p_{\\mathrm{fail}}$ is obtained from the same contraction. The efficient variant approximates the process tensor and tester by matrix product states, truncating small singular values in canonical form, so that the computational cost is controlled by a bond dimension rather than by the full Hilbert space.","core_discovery":"On the paper's own terms, the central claim is that decoding can be posed as a single tensor-network contraction. For a stabiliser code $C$ with pure errors $P(\\vec{s})$ and logical operators $L$, the recovery operation is $R(\\vec{s}) = \\bar{L} P(\\vec{s})$, where $\\bar{L} = \\operatorname{argmax}_L \\chi_{HS}(L,\\vec{s})$ and $\\chi_{HS}(L,\\vec{s})$ is the Hilbert-Schmidt inner product between the encoded, measured, recovered process (built from the process tensor $\\Upsilon$, the syndrome tester $C_{L,\\vec{s}}$, the encoder $\\Pi_{\\mathrm{enc}}$, and the logical identity channel) and the Choi state of the logical identity. The logical failure rate is then $p_{\\mathrm{fail}} = 1 - \\sum_{\\vec{s}} \\chi_{HS}(\\bar{L}(\\vec{s}),\\vec{s})$. Replacing the inner product by a 2-norm channel distance gives a second metric that the paper reports numerically but notes does not obey probability axioms. When the process tensor and tester are written as matrix product states, the same contraction can be approximated by truncating small Schmidt values; in low-noise regimes the approximation reproduces the exact logical failure rates while running in much less time, because unlikely syndrome outcomes carry little entanglement and are automatically discarded.","pith_inferences":["If the normalisation issue is resolved, the tester formalism already carries classical memory, so the same single-contraction prescription should extend to multi-round and circuit-level syndrome decoding, where the parity of syndrome histories matters.","The MPS speed-up is tied to syndrome-outcome bias: in low noise, rare syndromes carry little Schmidt weight and are truncated. The approach should therefore be most efficient for codes whose syndrome distribution is strongly peaked, and index ordering could be tuned per noise model.","The single-contraction formula turns logical failure rate into a differentiable function of the process tensor, so it could serve as a training objective for optimising the seed tensors of tensor-network codes against measured hardware noise.","One could test the decoder on process tensors obtained by process-tensor tomography of real devices, rather than the synthetic bath-and-crosstalk model used here, giving an experimentally anchored comparison against lookup-table or matching decoders."],"forward_implications":["For any stabiliser code whose noise is characterised by a process tensor, the optimal recovery map and the logical failure rate are available from one contraction, with no Monte Carlo sampling of error histories.","The numerical results show that non-Markovian bath coupling and ZZ crosstalk degrade the five-qubit code's logical failure rate, with bath coupling adding effective noise beyond information scrambling.","For the Steane code, matrix-product-state approximation with moderate bond dimension reproduces the exact logical failure rate and decoder performance in low-noise regimes, with a substantial speed-up; at high noise the approximation can become slower than exact contraction.","Because the decoder is defined from the measured process tensor, the same construction applies to whatever noise the device actually has, and it gives a direct metric for optimising tensor-network code blueprints."],"supporting_citations":[{"why":"Introduces the strategic-code framework that this paper builds on, combining QEC with process tensors via a semidefinite program.","marker":"[21]"},{"why":"Supplies the process-tensor formalism used to represent multi-time correlated noise.","marker":"[28]"},{"why":"Defines testers, the correlated-instrument objects into which the syndrome measurements and feed-forward recovery are packaged.","marker":"[20]"},{"why":"Provides the link product used to compose Choi states of the encoder, recovery, process tensor, and identity channel in $\\chi_{HS}$.","marker":"[42]"},{"why":"Earlier maximum-likelihood decoder methods for QEC, whose objective-function viewpoint the process-tensor ML decoder extends.","marker":"[36, 37, 38]"},{"why":"Give the matrix product state representation and canonical truncation scheme used to approximate the process tensor and tester.","marker":"[24, 25]"},{"why":"Source for the 2-norm channel-distance metric used as the alternative objective and for the caveat that it is not a strict probability.","marker":"[43]"},{"why":"Tensor-network code frameworks whose logical failure rates the proposed method can evaluate and optimise.","marker":"[40, 41]"}],"fun_headline_variants":["One tensor contraction decodes correlated quantum noise","Tensor networks decode non-Markovian noise in a single step","Strategic codes: maximum-likelihood decoding via tensor contraction","From process tensor to decoder: one contraction does it all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole reported failure rate rests on treating the Hilbert-Schmidt inner product $\\chi_{HS}(L,\\vec{s})$ as a true success probability, so that subtracting its syndrome-summed maximum from one gives a legitimate logical failure probability; the paper does not prove this normalisation, and its own alternative distance metric is admitted not to be a probability.","fun_headline_variants_meta":{"raw":{"variants":["One tensor contraction decodes correlated quantum noise","Tensor networks decode non-Markovian noise in a single step","Strategic codes: maximum-likelihood decoding via tensor contraction","From process tensor to decoder: one contraction does it all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3492,"prompt_tokens":1042,"completion_tokens":2450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":2384}},"tokens_in":658,"tokens_out":2450,"duration_ms":16275,"temperature":1.0,"reasoning_tokens":2384,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:50:03.561449+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the same five-qubit code under the same process-tensor noise model twice: once by contracting the paper's tensor network for $p_{\\mathrm{fail}}$, and once by a direct state-vector Monte Carlo simulation that applies the recovered correction and counts how often the logical state returns to the codespace. If the two logical failure rates disagree, the interpretation of $\\chi_{HS}$ as a success probability is wrong. A cheaper check is to compute $\\sum_{\\vec{s}} \\chi_{HS}(\\bar{L}(\\vec{s}),\\vec{s})$ for increasing noise strengths and test whether it stays in $[0,1]$ and decreases monotonically as expected of a probability.","supporting_citations":[{"cited_title":"Quantum stochastic processes and quantum non-Markovian phenomena","cited_arxiv_id":"2012.01894","evidence_quote":"Supplies the process-tensor formalism used to represent multi-time correlated noise."}],"review_version":1}