{"id":"060495d5-3ec8-4ce7-a6e3-3625758052d6","arxiv_id":"2509.06756","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"An iterative reweighting decoder for surface codes uses X-Z error correlations from circuit-level noise, improving logical error rates and raising the threshold from about 1% to 1.16%.","lead":"Quantum error correction relies on a decoder that reads measurements and guesses where errors occurred; the standard decoder treats two error types independently. This paper adds an iterative reweighting step that exploits correlations between the two error types, reporting lower logical error rates and a higher error threshold under realistic circuit noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3 proves monotone lattice matching cost, not Pauli weight; Theorem 5's distance guarantee is unsupported.","rationale":"The reader's weakest assumption is the same as the most load-bearing concern: Lemma 3 conflates lattice matching cost with Pauli weight, and Theorem 5 depends on this conflation. This is not a stylistic issue: the proof of Theorem 5 explicitly uses the nonincreasing Pauli weight to bound wt(ÊÊ') below the code distance. If reweighting can reduce matching cost while increasing Pauli support, the distance guarantee may fail. The numerical improvement is plausible and the paper provides a code link, so the empirical part may survive; however, the formal claim 'preserves the distance guarantee of MWPM' is a headline contribution and is currently unsupported. I do not see a reason to move beyond the reader's CONDITIONAL verdict: the authors should either repair the proof with a direct argument that Pauli weight is nonincreasing, or temper/remove the distance-preservation theorem and reposition the paper as an empirical decoder study. The reader already recommended this, so no verdict change is needed.","tokens_in":20196,"tokens_out":7688,"duration_ms":77558,"concrete_test":"Using the provided GitHub code (or a fresh Stim/PyMatching implementation), enumerate or sample for d=5 and d=7 under the paper's circuit-level noise model with a perfect final round: all Pauli errors with wt(E) ≤ ⌊(d−1)/2⌋ that standard MWPM corrects. For each such E, run IRMWPM and (i) record the per-iteration Pauli weights wt(Ê^(j)) and the matching costs on L_X^(j), L_Z^(j); (ii) check whether the final logical outcome equals MWPM's. If wt(Ê^(j)) ever increases while lattice matching cost decreases, Lemma 3's identification fails. If any MWPM-correctable E yields a different logical outcome under IRMWPM, Theorem 5 is false. Absence of a violation after 10^5 such low-weight errors would support the theorem despite the proof gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof gap identified by the reader is real and load-bearing. In Section IV.B.3, Lemma 3 claims that wt(Ê^(j)) (the support size of the joint Pauli correction) is nonincreasing, but its proof switches to wt(L_X^(j), M_X^(k)) and wt(L_Z^(j), M_Z^(k))—the MWPM costs on reweighted lattices. Equations (3)–(6) are inconsistent: e.g., W_{j+0.5} is written both with M_X^(j) on L_X^(j+1) and with M_Z^(j+1) on L_Z^(j+1), and W_j is expanded using L_Z^(0)/M_Z^(j) plus L_X^(j+1)/M_X^(j), none of which equals the Pauli weight of the correction. In circuit-level decoding, edge weights are −ln P, not unit Pauli weights, so a matching can have lower lattice cost while the corresponding Pauli correction has larger support; correlated edges are precisely reweighted to be cheaper. Theorem 5 needs wt(Ê') ≤ wt(Ê) ≤ wt(E) to conclude ÊÊ' has weight < d and is a stabilizer. That chain uses the unproven Lemma 3 and also assumes MWPM's optimality minimizes Pauli weight, whereas MWPM minimizes lattice matching cost. The distance-preservation claim therefore does not follow from the arguments given.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Iterative Reweighting Minimum-Weight Perfect Matching (IRMWPM), a decoder that alternates between the X and Z decoding lattices of a surface code, reweighting edges by conditional probabilities of correlated X/Z detection events derived from circuit-level depolarizing noise. The authors claim finite-time convergence (Theorem 1), preservation of the MWPM decoding radius (Theorems 2 and 5), and report numerical improvements: an increased threshold (about 1.16% vs. 1%), over 20x logical-error reduction at low physical error rates for d ≥ 17, and an extrapolated resource saving of d = 31 vs. d = 50 at p_L = 10^-16. The central theoretical result is Lemma 3 in Section IV.B.3, which is used to prove convergence and the distance guarantee.","tokens_in":20533,"tokens_out":7487,"duration_ms":74242,"significance":"If the empirical results are reproducible, IRMWPM is a practically valuable enhancement of MWPM for circuit-level noise: it uses a systematic catalog of fault-path correlations (Table III), converges in a few iterations in the simulated regime, and is supported by public code. The improvement in threshold and logical error rate at large distance is interesting and potentially useful for real-time decoding. However, the theoretical guarantees have a load-bearing gap: Lemma 3 conflates lattice matching cost with Pauli correction weight, and Equations (3)-(6) are internally inconsistent. The headline d=31 vs. d=50 qubit-overhead claim is an unvalidated extrapolation of a six-parameter fit. These issues must be repaired before the central claims can be relied upon.","major_comments":[{"comment":"The proof defines W_j as a sum of matching costs on reweighted lattices, not as wt(Ê^(j)). Equations (3)-(6) are internally inconsistent: Eq. (3) uses M_Z^(j) on L_Z^(0) while Eq. (4) uses M_Z^(j) on L_Z^(j); Eq. (5) pairs M_Z^(j+1) on L_Z^(0) with M_X^(j) on L_X^(j+1), while Eq. (6) pairs M_X^(j) on L_X^(0) with M_Z^(j+1) on L_Z^(j+1). In circuit-level decoding, edge weights are -ln P, so a minimum-cost matching on a reweighted lattice is not a minimum Pauli-support correction; reweighted correlated edges are cheaper by design. Thus monotonicity of W_j does not imply wt(Ê^(j)) is nonincreasing. Since Theorem 5 requires wt(Ê') ≤ wt(Ê) ≤ wt(E), the distance guarantee is unsupported. The initial inequality wt(Ê) ≤ wt(E) also assumes MWPM minimizes Pauli weight, which is not what the decoder minimizes for logarithmic edge weights.","section":"Section IV.B.3, Lemma 3, Eqs. (3)-(6)"},{"comment":"The headline claim that IRMWPM requires d=31 while MWPM requires d=50 to reach p_L = 10^-16 is obtained by extrapolating a six-parameter fit to logical error rates for L ≤ 17 and p ≥ 0.001, evaluated at p=0.001, L=31 and L=50. This is many orders of magnitude outside the simulated range. No held-out validation at intermediate L (e.g., L=23) or uncertainty quantification is provided. Because the abstract advertises this as a major qubit-overhead reduction, the claim needs either validation with additional data/error bars or an explicit downgrade to a rough, speculative estimate.","section":"Section II.B, Eq. (1), Figs. 3-5"},{"comment":"The text and Figure 1 define a half-step in which M_X^(j) reweights L_Z to produce M_Z^(j+1), then M_Z^(j+1) reweights L_X to produce M_X^(j+1). Algorithm 1 instead initializes Ê_Z^(0)=∅ and in each loop first runs Reweight(L_X, Ê_Z^(k)), then Reweight(L_Z, Ê_X^(k+1)). Consequently the objects Ê^(j) and Ê^(j+0.5) used in Lemma 3 and Theorem 5 do not correspond to the outputs of Algorithm 1. The theoretical guarantees therefore apply to a different iterative schedule from the one presented in the pseudocode. Please align the pseudocode with the analyzed schedule and state explicitly which schedule was used in the numerical simulations.","section":"Algorithm 1 vs. Section II.A and Fig. 1"},{"comment":"The manuscript states that the MWPM distance guarantee “is expected to extend to circuit-level noise ... although no formal proof currently exists.” Theorem 5 nevertheless presents a circuit-level decoding-radius result for IRMWPM, conditional on this unproved property of MWPM. In addition, Lemma 4 gives per-component conditions (wt(E_X) ≤ ... and wt(E_Z) ≤ ...), while Theorem 5 assumes only total weight wt(E) ≤ ...; MWPM decodes X and Z separately and does not jointly minimize the total Pauli weight. These statements must be reconciled; as written, the theorem does not establish a circuit-level decoding radius for IRMWPM.","section":"Section IV.B.3, Lemma 4 and Theorem 5"}],"minor_comments":[{"comment":"The abstract states “over 20x” improvement for d ≥ 17 and p ≤ 0.001, but Figure 4 shows data up to L=17 and p ≥ 0.001; please specify the exact p range and note that the largest simulated distance is L=17.","section":"Abstract and Section II.B"},{"comment":"Typo: “horozontal” should be “horizontal”.","section":"Section IV.B.1"},{"comment":"The sentence “IRMWPM, which incurs lower overhead than MWPM” appears to contradict the algorithm’s additional iterations; please clarify whether the intended meaning is lower logical qubit overhead or comparable computational overhead.","section":"Section III, Discussion"},{"comment":"The stopping condition breaks when either the X or Z estimate repeats, not necessarily both; this should be clarified, since convergence of one estimate does not guarantee convergence of the other.","section":"Algorithm 1, stopping criterion"}],"recommendation":"major_revision","confidential_remarks":"The empirical study appears carefully done and the proposed method is plausible, but the theoretical guarantees—a major advertised selling point—are not supported by the proof as written. The proof of Lemma 3 must either be repaired or the theorems must be weakened to empirical claims. The resource-overhead extrapolation should also be toned down unless it can be validated at intermediate distances. I do not see grounds for rejection if the authors can fix these points, but the revision needs substantive work, not just copy-editing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you my take on 2509.06756. The headline: the empirical decoder improvement looks plausible and the circuit-level conditional probability table (Table III) is a genuinely useful addition, but the formal guarantees, especially Theorem 5, do not hold up as written.\n\nWhat's new: the paper credits the iterative reweighting idea to Fowler and others, which is honest. The new material is the systematic calculation of X/Z conditional probabilities for CNOT fault paths in circuit-level noise, plus the claimed finite-time convergence and distance preservation. They also ship code, which is good practice.\n\nThe simulations show a threshold increase from about 1% to 1.16% and a 20x logical error reduction for d≥17, p≤0.001. That is plausible and worth testing. But there are no error bars or statistical details, and the headline d=31 vs d=50 resource claim comes from a six-parameter extrapolation fit to data up to d=17. That's a long extrapolation; I wouldn't trust it without more checks.\n\nThe soft spots are in the proofs. Lemma 3 is supposed to show the Pauli weight of the correction is nonincreasing. Instead, the proof tracks the lattice matching cost on reweighted graphs. Those are not the same object, especially under circuit-level noise where edge weights are −ln P. Equations (3)-(6) use inconsistent indexing, and inequality (a) only shows MWPM optimality for lattice cost, not for Pauli weight. Theorem 5 then relies on the unproven Lemma 3 and also assumes MWPM minimizes Pauli weight, which it does not in general. So the distance guarantee does not follow. This is not a minor typo; the central theoretical claim is unsupported.\n\nWho is this for? People working on real-time decoders and wanting to exploit X/Z correlations will find Table III and the empirical results interesting. But they should treat Theorem 5 as unproven and the resource reduction as a speculative extrapolation.\n\nMy recommendation: send it to peer review, but the referee should demand a corrected proof of Lemma 3 and Theorem 5, or a clear statement that these are only conjectures. The numerical part alone could survive as a useful contribution if the code is solid.","headline":"Empirical decoder improvement looks plausible, but the distance-preservation theorem is not proven as written — send to review if the theory is fixable, otherwise treat as a numerical advance.","tokens_in":20987,"tokens_out":6167,"would_cite":false,"duration_ms":55499,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P70","81P68"],"pacs":["03.67.Pp"],"model":"deepseek-v4-flash","headline":"Iterative lattice reweighting preserves MWPM's distance guarantee while exploiting X/Z correlations to lower logical error rates.","keywords":["surface codes","minimum-weight perfect matching","iterative reweighting decoding","circuit-level noise","X/Z error correlations","decoding threshold","quantum error correction","fault tolerance"],"falsifier":"Track both the matching cost W_j and the true Pauli weight of the correction at every IRMWPM iteration on injected syndromes. If any run shows W_j decreasing while the actual Pauli weight of the correction increases, the identification used in the convergence and distance proofs fails. A second check is to enumerate all errors of weight at most (d-1)/2 that MWPM corrects and verify that IRMWPM always returns a correction differing from MWPM's by a stabilizer; any counterexample would falsify the radius-preservation claim.","tokens_in":1656,"feed_emoji":"⚛️","tokens_out":5163,"duration_ms":100452,"temperature":0.7,"pith_summary":"The paper introduces IRMWPM, an iterative version of minimum-weight perfect matching decoding for surface codes that deliberately exploits correlations between X-type and Z-type errors under circuit-level noise. It claims the reweighting procedure converges in finite time and preserves the decoding radius of standard MWPM: any error MWPM corrects, IRMWPM also corrects. Simulations under circuit-level depolarizing noise show over 20x lower logical error rates at low physical error rates for distances 17 and above, a threshold increase from 1% to 1.16%, and extrapolations suggesting distance 31 instead of 50 to reach logical error rate 10^-16. If correct, this makes the decoder a near drop-in improvement for near-term fault-tolerant quantum computers, with only a few added iterations and no loss of MWPM protection.","feed_headline":"Reweighted matching decoder slashes surface-code logical errors 20x","feed_subtitle":"Iterative reweighting exploits X-Z error correlations, lifting the threshold from 1% to 1.16%.","key_machinery":"The load-bearing object is a pair of dual 3D space-time decoding lattices for X and Z errors, initially weighted by single-fault probabilities. The reweighting uses conditional probabilities (Table III) of a dual-lattice matching given a primal-lattice matching; each edge is replaced by -ln of that conditional probability. The proof machinery is the monotone sequence of total correction weights W_j >= W_j+0.5 >= W_j+1, which yields finite-time convergence, and MWPM optimality plus the code distance yields the decoding-radius guarantee.","core_discovery":"The central discovery is that surface-code decoding graphs for X and Z errors can be reweighted against each other using conditional fault probabilities, and doing so iteratively does not weaken the MWPM distance guarantee. The paper classifies single-fault correlations into six paired types on the 3D decoding lattices, uses them to compute conditional probabilities (e.g., P(d1|a)=3/31), and alternates MWPM on reweighted lattices. The resulting sequence of joint correction weights is nonincreasing, implying finite-time convergence and, because the initial MWPM estimate is already within the code's correction radius, the algorithm terminates at a correction that differs from MWPM's by a stabi","pith_inferences":["This editor would probe the gap between the matching cost W_j and the actual Pauli weight of the correction: Lemma 3 proves the former is nonincreasing, while Theorem 5 needs the latter. Tracking both quantities in simulation would show whether they can diverge on adversarial syndromes.","The same reweighting table should transfer to non-depolarizing biased or amplitude-damping noise, where X/Z correlations are stronger; the paper notes this extensibility but does not simulate it.","An FPGA or ASIC implementation should be straightforward because the added steps are edge-weight updates plus repeated MWPM; each iteration has the same per-round cost as standard MWPM.","The distance theorem relies on exact MWPM optimality, so replacing the inner decoder with approximate matching methods such as union-find or sparse blossom would require a separate radius analysis."],"forward_implications":["Circuit-level surface-code decoders can exploit X/Z correlations without leaving MWPM's guaranteed decoding radius, so the reweighting strategy can be layered onto existing MWPM decoder implementations.","A threshold of about 1.16% rather than 1% relaxes the physical error rate required for fault-tolerant operation, easing hardware demands.","To reach a logical error rate of 10^-16 at p=0.001, IRMWPM needs distance 31 versus 50 for standard MWPM, implying more than 60% fewer physical qubits per logical qubit.","In practice only two to four iterations are needed at realistic error rates, keeping the runtime overhead modest for real-time decoding.","Because the conditional-probability tables can be recomputed for other noise models, the decoder becomes noise-aware without per-hardware training."],"supporting_citations":[{"why":"Classifies the elementary error matchings and fault-path types on surface-code decoding lattices, supplying the correlated detection-event patterns from which IRMWPM's conditional reweighting probabilities are computed.","marker":"[27]"},{"why":"Introduces the single reweighting step for correlated X/Z errors under code-capacity noise that IRMWPM generalizes to the circuit-level noise model.","marker":"[52]"},{"why":"Proves nonincreasing correction weight for iterative MWPM under the code-capacity model; the paper extends that monotonicity argument to circuit-level noise.","marker":"[54]"},{"why":"Establishes the MWPM decoding framework for topological memories, including syndrome rounds T=O(L), the baseline used for the decoding-radius comparison.","marker":"[21]"},{"why":"Shows an iterative union-find decoder can lose its distance guarantee, motivating the proof that IRMWPM preserves MWPM's decoding radius.","marker":"[56]"},{"why":"Supplies the logical-lifetime simulation protocol used to benchmark IRMWPM against belief-propagation decoding under circuit-level noise.","marker":"[36]"},{"why":"Provides the circuit-level 3D space-time decoding lattice construction whose edge weights IRMWPM rewrites.","marker":"[58]"}],"fun_headline_variants":["Iterative reweighting decoder cuts surface-code errors 20x","Surface-code decoder uses error correlations for 20x gains","Reweighted matching lifts surface-code threshold to 1.16%","Quantum decoding made 20x better with iterative reweighting","New decoder reduces qubit overhead via reweighted matching"],"cache_read_input_tokens":22784,"weakest_assumption_plain":"The load-bearing premise is that a cheaper matching on the reweighted lattices also means a physically lighter correction; the proof shows the matching cost falls each round but does not show that the actual Pauli error count falls, so if the two ever separate, the distance guarantee would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Iterative reweighting decoder cuts surface-code errors 20x","Surface-code decoder uses error correlations for 20x gains","Reweighted matching lifts surface-code threshold to 1.16%","Quantum decoding made 20x better with iterative reweighting","New decoder reduces qubit overhead via reweighted matching"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1355,"prompt_tokens":806,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":462}},"tokens_in":550,"tokens_out":549,"duration_ms":6115,"temperature":1.0,"reasoning_tokens":462,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:09:17.745165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Track both the matching cost W_j and the true Pauli weight of the correction at every IRMWPM iteration on injected syndromes. If any run shows W_j decreasing while the actual Pauli weight of the correction increases, the identification used in the convergence and distance proofs fails. A second check is to enumerate all errors of weight at most (d-1)/2 that MWPM corrects and verify that IRMWPM always returns a correction differing from MWPM's by a stabilizer; any counterexample would falsify the radius-preservation claim.","supporting_citations":[{"cited_title":"Sweke, M","cited_arxiv_id":null,"evidence_quote":"Introduces the single reweighting step for correlated X/Z errors under code-capacity noise that IRMWPM generalizes to the circuit-level noise model."},{"cited_title":"Chamberland, L","cited_arxiv_id":null,"evidence_quote":"Proves nonincreasing correction weight for iterative MWPM under the code-capacity model; the paper extends that monotonicity argument to circuit-level noise."},{"cited_title":"deMarti iOlius, P","cited_arxiv_id":null,"evidence_quote":"Shows an iterative union-find decoder can lose its distance guarantee, motivating the proof that IRMWPM preserves MWPM's decoding radius."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the logical-lifetime simulation protocol used to benchmark IRMWPM against belief-propagation decoding under circuit-level noise."},{"cited_title":"Liyanage, Y","cited_arxiv_id":null,"evidence_quote":"Provides the circuit-level 3D space-time decoding lattice construction whose edge weights IRMWPM rewrites."}],"review_version":1}