{"id":"824a5604-235f-4bca-9639-1a3e93ee9fc8","arxiv_id":"2608.00171","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Under bit-flip-dominated circuit-level noise, an N-qubit repetition code achieves Heisenberg-limited sensing for interrogation times up to O(1/p^{⌊(N−1)/2⌋+1}).","lead":"A new protocol shows that a quantum sensor built from an N-qubit repetition code can hold Heisenberg-limited precision for interrogation times up to about 1/p^{(N+1)/2}, instead of 1/p, provided all circuit noise is strongly biased toward bit-flip errors. This is a step toward practical fault-tolerant quantum sensing under a specific, heavily biased noise model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The result is internally consistent, but its central advantage is entirely contingent on an extreme noise-bias assumption: p_z must be O(p^{t'}). The repetition code is blind to phase-flip errors, so any fixed physical bias fails for large N.","rationale":"The paper's mathematical argument is sound under its stated assumptions: the logical-error scaling, the unbiased-estimator construction, and the worst-case CFI bound all cohere, and the authors are transparent about the restrictive noise model. The single most load-bearing condition is the extreme X-bias requirement, because if p_z is not suppressed to O(p^{t'}), the repetition code provides no phase-flip protection and the claimed Heisenberg extension disappears completely. This is the same assumption the reader identified as weakest, and it is properly flagged by the authors themselves in Appendix B.7 and Appendix C. It is therefore not a fatal flaw, but it is the condition on which the central claim's practical significance rests, and it should remain prominent in any public summary. The even-N abstract exponent mismatch is a real but secondary issue, already noted in the reader's rationale; it does not change the verdict. Hence the reader's CONDITIONAL assessment stands unchanged.","tokens_in":37671,"tokens_out":29715,"duration_ms":328184,"concrete_test":"Analyze the N=3 protocol of Sec. IV.D with an added independent phase-flip (Z) error of probability p_z at every physical location, not just as a logical channel. Compute the CFI at the claimed optimal M_1 = 0.2/p^2 for p = 10^{-3} while varying p_z. The prediction is that M_1^opt = min(O(1/p^{t'}), O(1/p_z)): at p_z = p the advantage collapses to the O(1/p) single-qubit timescale, while at p_z = p^2 it survives. This directly tests the bias requirement stated in Appendix B.7 and would quantify how rapidly the advantage degrades with finite bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a theorem about a specific, explicitly defined X-only circuit-level noise model (Sec. IV.C). Under that model the derivation of Eq. (46) and the worst-case bound in Appendix D are coherent. The load-bearing condition is the assumption that phase-flip errors are negligible at the level p_z = O(p^{t'}), t' = floor((N-1)/2)+1. Because the repetition-code stabilizers Z_iZ_{i+1} detect only X-type errors, any Z fault acts as a logical phase flip on the GHZ sensor. As the paper itself states in Appendix B.7, the CFI is then suppressed by a factor (1-2p_z)^{2M_1}; with the claimed M_1 = O(1/p^{t'}), retaining Heisenberg scaling requires p_z M_1 << 1, i.e. p_z = O(p^{t'}). This is exponentially small in N. For any finite bias ratio epsilon = p_z/p, the condition epsilon << p^{t'-1} fails once N is sufficiently large, even if the absolute error rate p is small. The authors acknowledge this in Appendix B.7 and describe the all-X restriction as 'quite artificial' in Appendix C. This is not an internal inconsistency, but it sharply limits the practical reach of 'fault-tolerant': the protocol is fault-tolerant only against a noise model whose bias must improve exponentially with code size.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes and analyzes a fault-tolerant quantum sensing protocol that uses an N-qubit repetition code to estimate a Z-rotation under a circuit-level noise model in which every operation fails with X-type (bit-flip) errors with probability p and phase-flip errors are absent (or, in App. B.7, suppressed as O(p^{t'})). The central result (Sec. IV, Eq. (46)) is a lower bound on the classical Fisher information J ≥ M1^2 (1−2p)^6 (2(1−˜p_{L,p,N})(1−˜p_{L,N})^{M1}−1)^2, where ˜p_{L,N}=O(p^{t'}) with t'=⌊(N−1)/2⌋+1. This extends the coherent interrogation time for Heisenberg scaling from M1=O(1/p) (no encoding) to M1=O(1/p^{t'}), at the cost of a noise bias that must improve exponentially with N. The authors provide fault-tolerant state preparation, syndrome extraction with minimum-weight perfect matching, a final uncomputed measurement, an unbiased estimator based on syndrome histories, and a worst-case bound over logical phase histories (Appendix D). The result is compared with the parallel protocol of Ref. [35].","tokens_in":37985,"tokens_out":12481,"duration_ms":123780,"significance":"If the claimed scaling holds, this is a conceptual advance: it shows that circuit-level faults at every location need not destroy Heisenberg scaling for a restricted, but explicitly characterized, biased-noise model, and it gives a concrete sequential sensing advantage over parallel approaches. The manuscript's strengths include a parameter-free derivation of the logical-error scaling (with only an optimization constant a=0.2), a rigorous worst-case CFI bound in Appendix D, explicit fault-tolerant circuits, STIM-based numerical threshold data, and an honest discussion of the artificiality of the X-only noise model (App. C.3) and of the required bias (App. B.7). The main limitation is the noise-bias requirement: the practical scope is narrower than the title and abstract might suggest, and the abstract contains a technically incorrect exponent for even N.","major_comments":[{"comment":"The abstract states T ∝ 1/p^{(N+1)/2} for an N-qubit code. This is correct only for odd N. The definition t'=⌊(N−1)/2⌋+1 (Eq. (45)) gives t'=(N+1)/2 for odd N but t'=N/2 for even N. For even N the correct exponent is ceil(N/2), so the abstract's formula is strictly wrong (e.g., for N=2 it claims 1/p^{1.5} rather than 1/p). The same error appears in Sec. IV.E where p^{2t'}=p^{(N+1)} is asserted. Please correct the abstract and all derived statements, or explicitly restrict the headline claim to odd N.","section":"Abstract and Sec. IV.E"},{"comment":"The practical reach of 'fault-tolerant' is limited by the noise-bias assumption. The repetition code is blind to Z errors; a phase-flip with probability p_z suppresses the GHZ coherence by (1−2p_z)^{M1}. With M1=O(1/p^{t'}), retaining Heisenberg scaling requires p_z=O(p^{t'}), i.e. a bias ratio p_z/p that must shrink exponentially with N. The authors acknowledge this in App. B.7 and call the X-only model 'quite artificial' in App. C.3, but these caveats are absent from the abstract and Introduction. Any fixed physical bias ratio fails for large N; the title and abstract should state this explicitly, for example by saying the advantage holds for noise whose bias improves exponentially with code size. This is not an internal inconsistency, but it is load-bearing for the practical interpretation of the result.","section":"App. B.7 and main text Sec. IV.C"}],"minor_comments":[{"comment":"The inequality P_mv(r) ≤ C(r,(r+1)/2) η^{(r+1)/2} is not correct as stated; e.g., for r=3, η=0.1, P_mv ≈ 0.244 but the RHS is 0.03. The conclusion r≥2t+1 is standard, but the proof should be replaced with a valid Chernoff/Hoeffding bound.","section":"Appendix B.6.a, Eq. (B39)"},{"comment":"The statement 'p^{2t'}/ν = p^{(N+1)}/ν' is only valid for odd N. Use p^{2t'} or ceil(N/2) to cover even N.","section":"Sec. IV.E"},{"comment":"The author name 'Lorc´ an' appears with a misplaced combining character; it should be 'Lorcán'. Also 'deMarti iOlius' in Ref. [92] should be 'de Martí iOlius'.","section":"Author list and Refs."},{"comment":"The threshold at p≈0.02 is reported without specifying the exact syndrome-extraction circuits and the number of syndrome repetitions used in the STIM simulation. Please provide the simulation parameters so the threshold can be reproduced.","section":"Fig. 8"},{"comment":"The paper's specialized definition of 'fault-tolerant quantum sensing' (handling a restricted class of errors at every location) is introduced in the Introduction but would benefit from being stated in the abstract or at the start of Sec. II.C to avoid confusion with the standard fault-tolerance notion against arbitrary errors.","section":"Introduction / Sec. II.C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically sound under its stated noise model, but the abstract overstates the exponent and the bias requirement severely limits practical relevance. The incorrect inequality in Eq. (B39) is localized. With the abstract corrected and the bias caveat promoted, the paper would be a valuable contribution. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one if you care about fault-tolerant quantum metrology. The genuinely new thing is sequential sensing: an N-qubit repetition code that keeps Heisenberg scaling in interrogation time up to M_1 = O(1/p^{t'}) with t' = floor((N-1)/2)+1, under a circuit-level noise model in which every location suffers bit-flip errors with probability p and phase-flip errors are negligible. Ref. [35] did the parallel version; the sequential extension matters because with a fixed number of qubits it converts the short O(1/p) coherent window into a polynomial one that grows with code size. The central bound, Eq. (46), follows from a standard binomial-tail logical error rate and a worst-case classical Fisher information argument in Appendix D. I checked that argument and it is correct. The fault-tolerant state preparation, syndrome extraction, and decoding details in the appendices are also careful. The STIM-based logical error rates in Fig. 8 are not reproducible—no code or data is provided—but that is an addressable issue, not a conceptual one.\n\nThe real soft spot is the noise bias. The repetition code sees X errors and is blind to Z errors, so the claimed advantage requires p_z = O(p^{t'}), i.e. phase-flip probability that is exponentially small in N relative to p. Any fixed physical bias will eventually fail as N grows. The authors are candid about this: Appendix B.7 states that a phase-flip channel suppresses the GHZ coherence by (1-2p_z)^{M_1}, and Appendix C.3 calls the all-X restriction \"quite artificial.\" My complaint is not that they hide it; it is that the abstract gives no hint that the practical regime is one of exponentially improving noise bias. That should be front and center in the public version.\n\nOne more concrete fix: the abstract's (N+1)/2 exponent is exact only for odd N. The correct general statement is ceil(N/2). Minor, but it should be corrected.\n\nOverall: the central theorem holds under its stated assumptions, the assumptions are restrictive and somewhat artificial, and the authors know it. That is a fair combination for a publishable theory paper. Send it to a referee who knows quantum metrology and noise models; the verdict should be a revise with the exponent and the bias caveat addressed. I would not desk-reject it.","headline":"Genuine sequential extension of fault-tolerant sensing with a sound central bound, but the exponential noise-bias assumption should be much more prominent and the abstract exponent fixed.","tokens_in":38515,"tokens_out":3039,"would_cite":true,"duration_ms":34288,"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":"This paper argues that a fault-tolerant N-qubit repetition code can keep Heisenberg-limited quantum sensing for a coherent time M1 = O(1/p^{⌊(N−1)/2⌋+1}) even when every circuit operation fails with bit-flip probability p, provided phase-fl","keywords":["fault-tolerant quantum sensing","Heisenberg limit","repetition code","circuit-level noise","biased noise","quantum error correction","quantum metrology","Fisher information"],"falsifier":"Simulate or implement the full protocol with an extra tunable Z-error rate p_z alongside the X-error rate p, and measure the parity fringe (2(1−p̃_{L,p,N})(1−p̃_{L,N})^{M1}−1) that appears in the Fisher-information bound. If p_z is held at a fixed fraction of p rather than being suppressed as O(p^{⌊(N−1)/2⌋+1}), the fringe visibility should decay as (1−2p_z)^{M1} and the Heisenberg extension M1 ∝ 1/p^{t'} should disappear; in particular, running the same circuit with depolarising noise (p_z = p) should make the variance cross over from 1/M1^2 to 1/M1 scaling at M1 ≈ 1/p.","tokens_in":37512,"feed_emoji":"📡","tokens_out":7273,"duration_ms":73413,"temperature":0.7,"pith_summary":"The paper tries to show that fault-tolerant quantum error correction can be imported into quantum sensing: even when state preparation, gates, idling, syndrome extraction, and readout all fail with probability p, a simple repetition code can extend the time over which Heisenberg scaling holds. In the ideal setting, a single qubit accumulating phase over M channel uses reaches error ~1/M; under bit-flip noise without encoding, coherent accumulation is limited to M ~ 1/p. The paper claims that, under a noise model where only X-type bit-flip errors occur and Z-type phase-flip errors are negligibly rare, an N-qubit repetition code pushes the coherent accumulation time to M1 = O(1/p^{⌊(N−1)/2⌋+1}). This matters because it removes the usual assumption that all control operations are perfect, moving error-corrected sensing closer to realistic hardware — if the required noise bias can be engineered.","feed_headline":"Repetition code stretches quantum-sensing time past the 1/p barrier","feed_subtitle":"Fault-tolerant sensing with bit-flip-only noise reaches Heisenberg scaling for M ~ 1/p^{(N+1)/2}, not just 1/p.","key_machinery":"The central object is the N-qubit repetition code, whose logical states are GHZ states (|0⟩^N + |1⟩^N)/√2, used with a signal rotation acting on only the first qubit. For X-type noise the code has distance N and corrects t = ⌊(N−1)/2⌋ errors, so an uncorrectable logical error per round occurs only with probability p̃_{L,N} = c p^{t'} + O(p^{t'+1}). The argument is carried by the space-time decoding graph: each physical X fault in state preparation, gates, idling, syndrome measurement, or the sensing channel creates at most one local detector edge, and repeating each syndrome measurement 2t+1 times gives a decoding distance 2t+1, so t+1 faults are needed to cause a logical error. This error-l","core_discovery":"The central claim is the Fisher-information bound of Eq. (46): after M1 channel uses with a fault-tolerant N-qubit repetition-code protocol, J ≥ M1^2 (1−2p)^6 (2(1−p̃_{L,p,N})(1−p̃_{L,N})^{M1}−1)^2, so choosing M1 ≈ 0.2/p̃_{L,N} gives variance O(p̃_{L,N}^2) and Heisenberg scaling in M1. Because the per-round logical error rate is p̃_{L,N} = O(p^{t'}) with t' = ⌊(N−1)/2⌋+1, the protocol reaches Heisenberg-limited precision for M1 = O(1/p^{t'}), a polynomial extension beyond the 1/p limit of an unprotected sensor. The authors construct the full fault-tolerant circuit: verified preparation of the logical |+⟩ state, repeated stabilizer measurements decoded through a space-time detector graph, co","pith_inferences":["The practical reach of the protocol hinges entirely on engineered noise bias: because the repetition code does not correct Z errors, any residual phase-flip probability p_z must scale as O(p^{⌊(N−1)/2⌋+1}), and a fixed physical bias will eventually be insufficient as N grows. A testable target for hardware is therefore to demonstrate a suppression of phase-flip versus bit-flip rates that grows wit","Sequential fault-tolerant sensing appears to extract more advantage per qubit than parallel fault-tolerant sensing: for a fixed qubit budget and fixed total channel uses, the sequential protocol's error scales as p^{⌊(N−1)/2⌋+1}/√ν, while a parallel protocol with the same qubit number is limited by N^2. This ordering could be checked empirically by comparing the two protocols with identical total ","The analysis of ambiguous ±ϕ phase histories suggests a general design rule for fault-tolerant sensors: a circuit fault hurts only if it flips the sign of the accumulated phase, and faults that merely randomise the phase at quadratic order in p are harmless. Applying this rule to codes other than the repetition code could identify which stabilizer codes can support fault-tolerant sensing.","Combining this time-encoded protocol with a qubit-number-encoded fault-tolerant protocol would, if possible, yield Heisenberg scaling in both total channel uses and qubit number; the paper leaves that combination open, but it is the natural next benchmark."],"forward_implications":["Error-corrected sensing need not assume perfect control: state preparation, gates, syndrome extraction, and measurement can all suffer bit-flip errors at rate p, and Heisenberg scaling in time still survives under the biased-noise model.","The coherent interrogation window grows from M1 ∝ 1/p without encoding to M1 ∝ 1/p^{⌊(N−1)/2⌋+1} with an N-qubit repetition code, so each additional pair of qubits adds roughly another factor of 1/p to the achievable coherent time.","A threshold behaviour appears: for physical error rates below roughly p ≈ 0.02 (as computed from the simulated logical error rates), increasing the code distance reduces the estimation error, so near-term operation below threshold, rather than p → 0, is the operating principle.","The protocol comes with an explicit unbiased estimator that uses the full syndrome history and avoids the systematic bias that would otherwise erase the quantum advantage.","The same construction generalises to qudit systems where errors are X and X^2, so the fault-tolerant advantage is not specific to qubits."],"fun_headline_variants":["Fault-tolerant codes push quantum sensing to Heisenberg limit longer","Repetition code yields Heisenberg scaling beyond 1/p limit","Fault tolerance extends quantum sensing time polynomially beyond 1/p"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole advantage rests on the noise being almost perfectly biased: every operation fails only with X-type bit-flip errors at probability p, while the probability p_z of any Z-type phase-flip error must be O(p^{⌊(N−1)/2⌋+1}); if this bias fails, the GHZ coherence decays as (1−2p_z)^{M1} and the Heisenberg extension collapses — the authors themselves call the X-only gate assumption 'quite artificial'.","fun_headline_variants_meta":{"raw":{"variants":["Fault-tolerant codes push quantum sensing to Heisenberg limit longer","Repetition code yields Heisenberg scaling beyond 1/p limit","Fault tolerance extends quantum sensing time polynomially beyond 1/p"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2347,"prompt_tokens":772,"completion_tokens":1575,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":1516}},"tokens_in":516,"tokens_out":1575,"duration_ms":14142,"temperature":1.0,"reasoning_tokens":1516,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T01:05:25.470846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or implement the full protocol with an extra tunable Z-error rate p_z alongside the X-error rate p, and measure the parity fringe (2(1−p̃_{L,p,N})(1−p̃_{L,N})^{M1}−1) that appears in the Fisher-information bound. If p_z is held at a fixed fraction of p rather than being suppressed as O(p^{⌊(N−1)/2⌋+1}), the fringe visibility should decay as (1−2p_z)^{M1} and the Heisenberg extension M1 ∝ 1/p^{t'} should disappear; in particular, running the same circuit with depolarising noise (p_z = p) should make the variance cross over from 1/M1^2 to 1/M1 scaling at M1 ≈ 1/p.","supporting_citations":[],"review_version":1}