{"id":"d072acc4-8fd6-4c88-a172-3be30e750dd8","arxiv_id":"2608.05686","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Syndrome detection rates are locally convex in drifted control parameters after Pauli twirling, giving O(1/ε²) self-calibration that is code-distance-independent for local LDPC codes.","lead":"This paper proves that the rate of error-detection events during quantum error correction is a locally convex function of drifting analog control parameters, so syndrome measurements alone can steer hardware back into calibration. The result provides formal convergence guarantees for continuous self-calibration, including a code-size-independent bound for LDPC codes.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convexity is proven only for a Gaussian-perturbed operating point θ*+σg, not for the fixed physical θ*; a non-gauge zero-Hessian hardware instance falls outside Theorem 1, so the claimed provable efficiency is not established for that machine.","rationale":"The strongest claim is that the syndrome detection rate is locally strongly convex for a fixed processor, enabling SPSA. The proof of that claim in Appendix A5 does not establish it for a fixed processor: the Gaussian perturbation changes the object being differentiated. The reader's weakest assumption is exactly this, and it is load-bearing because all subsequent theorems inherit the convexity radius. Without a positive radius at the actual θ*, the optimization guarantees have no domain of validity. I agree with the reader's conditional verdict: the flaw is specific and potentially fixable by rephrasing the result as a smoothed-objective guarantee or by adding a verifiable non-degeneracy condition, but it is not merely cosmetic. I also note the quadratic-truncation issue in Lemma 11 as a secondary proof gap, but it does not displace the convexity concern. The numerical validations are useful evidence for the generic case, but they do not test the degenerate case, because Eq. (10) prescribes random PSD sensitivity matrices rather than computing the actual Hessian at a fixed miscalibrated point. The proposed concrete test—computing det M at the physical θ* in the paper's own small simulation—would settle whether the degeneracy actually occurs for a representative hardware model.","tokens_in":73193,"tokens_out":12517,"duration_ms":140280,"concrete_test":"Compute the detector-event Hessian ∇²C(0) at the actual physical calibrated point θ* for the pulse-level [[4,1,2]] simulation of Appendix E (or a small surface code with the same control parametrization), without the Gaussian replacement θ*+σg. For each gate i, evaluate det M^(i)(θ*) (Eq. A17) and the minimum eigenvalue of the full Hessian. If a zero-curvature direction is found and is not a gauge (moving along it changes the logical channel), run Algorithm 1 from a small drift along that direction: a failure to reduce C with T=O(ε^-2) would falsify the theorem for that machine. Independently, symbolically check whether det M^(i)(θ*) is identically zero for a rotation-angle parameter around a Pauli axis absent from all detector-flip sets; an identically zero determinant would show Assumption 6 fails for a standard parametrization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix A5 (Lemma 9, Corollary 2, Lemma 10) proves positivity of the CPTP and leakage Hessian blocks at θ~* = θ* + σg, not at the physical calibrated point θ*. The physical objective is C(δθ) with V_i(δθ_i) = U_i(θ*_i + δθ_i)U_i(θ*_i)^†, so the Hessian coefficients h_P,j and ℓ_α,P,j are evaluated at θ*. Replacing θ* by θ~* changes the ideal gates and thereby changes the objective; the Gaussian randomness is over a quantity the controller cannot sample. Assumptions 6 and 7 only require det M not identically zero, so the bad set has measure zero, but a fixed processor can lie in that set. If at the actual θ* there is a zero-curvature direction that is not a structural gauge, then λ_min^C = 0, the convexity radius θ_C^(th) = 2λ_min^C / L_C vanishes, and Theorems 2–4 have no positive-radius regime. The phrase 'generically strictly convex' is a statement about random circuits, not about a specific hardware instance. The Clifford simulations (Eq. 10) use random PSD sensitivity matrices Ω_i and therefore assume away exactly this degeneracy. A secondary internal gap: Lemma 11 in Appendix B treats C as exactly quadratic although Eq. (B1) has O(||δθ||^3), so the claimed unbiasedness of the SPSA estimator is not exact; this is likely fixable with a bias term, whereas the convexity issue is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for self-calibrating quantum fault tolerance in which the syndrome stream generated during QEC is used as a zeroth-order feedback signal for online calibration of drifting analog control parameters. The central theoretical claim is that, after Pauli twirling, the average detector-event rate is locally strongly convex around the calibrated point for coherent, CPTP, and leakage errors, and that this convexity enables provably efficient calibration: Theorem 2 gives O(ε^{-2}) epochs for one-time drift, Theorem 3 gives sublinear dynamic regret for slowly drifting hardware, and Theorem 4 removes the code-distance dependence for local LDPC codes by exploiting detector locality. The paper also reports pulse-level neutral-atom simulations and circuit-level Clifford simulations supporting these claims. The appendices contain detailed derivations of the convexity results, SPSA convergence proofs, dynamic regret bounds, nonconvex extensions, and simulation details.","tokens_in":73575,"tokens_out":4048,"duration_ms":49621,"significance":"If the central convexity claim holds for fixed physical hardware, the paper would establish an important conceptual and practical result: syndrome measurements can serve not only for error correction but also as a provably efficient calibration signal, with rigorous rates and a locality-based scalability result for LDPC codes. The strengths of the manuscript are its explicit and careful appendix derivations, the concrete SPSA-based algorithms with stated convergence and regret bounds, the honest treatment of nonconvex global-calibration hardness, and the two complementary numerical validations. The main weakness is that the strict-convexity guarantee is obtained only for a Gaussian-perturbed calibration point, not for the fixed physical operating point of a given processor, and the simulations use random positive-semidefinite sensitivity matrices that assume away the degeneracy that the smoothed-analysis argument is meant to remove. These issues affect the load-bearing claim that the syndrome stream is provably efficient for a specific hardware instance.","major_comments":[{"comment":"The strict-convexity guarantee is proven for the Gaussian-perturbed calibration point θ~* = θ* + σg, not for the physical calibrated point θ*. The objective in Eq. (1) is defined with V_i(δθ_i) = U_i(θ*_i + δθ_i)U_i(θ*_i)^†, so the coefficients h_P,j and ℓ_α,P,j are evaluated at the actual θ*. Replacing θ* by θ~* changes the ideal gates implemented by the QEC circuit and hence changes the objective. For a fixed processor whose actual θ* has a non-gauge zero-curvature direction, Assumptions 6 and 7 only say that the bad set has measure zero; they do not imply that this particular processor is outside the bad set. In that case λ_min^C = 0, the convexity radius θ_C^(th) = 2λ_min^C / L_C in Theorem 5 vanishes, and Theorems 2-4 have no positive-radius regime. The phrase \"generically strictly convex\" is a statement about random circuits, not about a specific hardware instance, so the paper should either prove convexity at any fixed θ* under physically verifiable conditions or explicitly state the additional non-degeneracy assumption needed for a given processor.","section":"Appendix A5, Theorem 1, Lemmas 9-10"},{"comment":"Lemma 11 claims that the SPSA gradient estimator is exactly unbiased because C(δθ) is treated as exactly quadratic, but Eq. (B1) explicitly contains an O(||δθ||^3) remainder. The proof of Theorem 8 relies on this exact unbiasedness, and a similar reliance appears in the dynamic regret proof of Theorem 9. With the cubic remainder present, the two-point SPSA estimator has a bias whose size depends on λ and on the third-order term; the convergence and regret bounds as stated are therefore not justified. This gap is likely fixable by adding a standard bias term and including it in the step-size choice, but as written it is a load-bearing inconsistency in the optimization proofs.","section":"Appendix B, Lemma 11 and Eq. (B1)"},{"comment":"The Clifford simulations model control-induced error strength as ε_i = ε_i^(0) + δθ_i^T Ω_i δθ_i with random positive-semidefinite sensitivity matrices Ω_i. This assumes, by construction, that every control direction has positive second-order curvature at the calibrated point. The simulations therefore do not test the accidental-degeneracy regime that Appendix A5 is designed to address; they verify the algorithms under the strong-convexity assumption rather than testing whether the assumption holds generically for realistic gate parametrizations. To support the paper's generic-position claim, the numerical section would need to include either parameter sweeps showing that the Hessian is positive definite at physically relevant operating points or an explicit example of a non-gauge degeneracy and a demonstration that it is removed by an allowed perturbation.","section":"Section III.B, Eq. (10)"}],"minor_comments":[{"comment":"The abstract says the detection rate is locally strongly convex \"with high probability\", but Theorem 1 states generic strict convexity under Assumptions 6 and 7. The probability statement is not quantified in the main text; the appendix gives probability one for a continuous Gaussian perturbation, but no finite-sample or dimension-dependent high-probability bound is stated.","section":"Abstract and Section II.A"},{"comment":"The smoothed-analysis argument replaces the calibrated point θ* by θ* + σg, which changes the ideal gates. The paper should discuss whether such a perturbation preserves the desired logical operation and whether the controller can choose a nearby target operating point without degrading the fault-tolerance threshold.","section":"Appendix A5 after Lemma 9"},{"comment":"The locality-aware estimator description says \"masking out all other detectors\" and uses graph coloring to avoid interference, but the main-text proof of Theorem 4 is deferred to Appendix B3 and C4. A short statement in the main text about why the masked estimator remains unbiased would improve readability.","section":"Section II.D, Algorithm 2"},{"comment":"Figure 7 shows a steady-state detection rate plateau after control switches, but the main text does not reference this figure; adding a cross-reference or moving the figure to the appropriate simulation subsection would clarify its role.","section":"Appendix E, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important question, and the appendix work is substantial. The main issue is not the quality of the optimization analysis under strong convexity, but the fact that the central convexity guarantee is proven only for a Gaussian-perturbed calibration point. This is a load-bearing point that needs to be either fixed by a stronger argument or explicitly acknowledged as a conditional assumption. The exact-quadratic assumption in Lemma 11 is also a correctable but nontrivial gap. I recommend major revision rather than rejection because the framework is promising and the optimization machinery is mostly sound once the convexity premise is supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read on Gong–Hu 2608.05686. The headline: the core geometric observation is real and valuable. After Pauli twirling, the detection-rate landscape is locally convex around the calibrated point, and turning that into an online-optimization problem with O(1/ε²) convergence, sublinear dynamic regret, and code-distance independence for LDPC codes is exactly the kind of statement the QEC-control community needs. The appendices are careful, the pulse-level neutral-atom simulations are physical, and the theory genuinely is the first to make syndrome-based calibration a provably efficient task rather than a heuristic proxy.\n\nWhat is actually new: the local convexity proof for the twirled detection rate under coherent, CPTP, and leakage errors, and the locality-aware SPSA that removes explicit dimension dependence for local codes. The paper is honest about the nonconvex regime and cites the relevant online-optimization and control literature. Credit is earned there.\n\nNow the soft spots, in proportion. The stress-test concern is accurate and load-bearing. Appendix A5 proves strict convexity at the Gaussian-perturbed point θ*+σg, not at the fixed physical θ*. 'Generically strictly convex' is a statement over random circuit parameters; a particular processor can sit at a non-gauge degeneracy where the convexity radius vanishes. The paper does not close that gap, so the 'provably efficient' claim is currently a generic guarantee, not an instance-wise one. This needs an explicit caveat about generic processors or a mechanism to randomize the operating point without changing the ideal gates.\n\nSecond, Lemma 11 in Appendix B drops the O(||δθ||³) remainder and treats C as exactly quadratic; the SPSA unbiasedness is therefore only for the truncated model. This is minor relative to the convexity issue—standard bias arguments should absorb the remainder with a small perturbation radius—but it should be written down.\n\nThird, the Clifford simulations use Eq. (10), which prescribes exactly the quadratic landscape the theory predicts, so they are consistency checks rather than independent falsification tests. The pulse-level simulations are more independent but only exercise a small [[4,1,2]] code.\n\nBottom line: this paper deserves a serious referee. The core idea is likely right, the contribution is significant, and the gaps are technical rather than fatal. I would send it to review expecting major revision focused on Theorem 1's scope and the SPSA lemma.","headline":"Solid, important step for syndrome-based self-calibration; the advertised provable efficiency currently outruns the proof because strict convexity is shown for a Gaussian-perturbed operating point rather than the fixed hardware point.","tokens_in":74084,"tokens_out":2967,"would_cite":true,"duration_ms":37988,"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 proves that the average syndrome-detection rate is a locally strictly convex function of drifting analog control parameters after Pauli twirling, making the syndrome stream a provably efficient self-calibration signal with…","keywords":["quantum error correction","self-calibration","syndrome measurements","online convex optimization","stochastic approximation","quantum LDPC codes","control drift","Pauli twirling"],"falsifier":"Take a single control parameter on a real device, scan it through its calibrated value while measuring the average detector-event rate, and check for a local maximum or flat region at the calibrated point; Theorem 1 predicts a positive curvature valley for every detector-visible direction, so a measured local maximum at θ* for a non-gauge direction would refute the generic strict-convexity claim for that hardware instance.","tokens_in":72990,"feed_emoji":"🎛️","tokens_out":6636,"duration_ms":65984,"temperature":0.7,"pith_summary":"This paper argues that the syndrome measurements already collected during quantum error correction contain enough information to keep a drifting quantum processor calibrated, without interrupting computation for dedicated recalibration runs. The central claim is that, after randomized compiling (Pauli twirling), the average detector-event rate is locally strictly convex in the analog control parameters for coherent, general CPTP, and leakage errors, so the syndrome stream defines a well-behaved optimization landscape. Building on this geometry, the paper proves convergence to an ε detection rate in O(1/ε²) syndrome epochs for one-time drifts, sublinear dynamic regret for slowly time-dependent drifts, and convergence rates independent of code distance for local quantum LDPC codes. If correct, this means the same data stream that protects logical qubits can also serve as the feedback signal that stabilizes the hardware, removing the need for separate calibration phases in long computations.","feed_headline":"Syndrome data alone recalibrates a drifting quantum processor","feed_subtitle":"A proof that detection-rate landscapes are locally convex lets the syndrome stream double as a calibration signal with O(1/ε²) convergence.","key_machinery":"The central object is the average detector-event rate C(θ) = (1/N_D) Σ_k DR_k(θ), the expected fraction of detectors that fire per error-correction round. Its local geometry is generated by the detector-visible Pauli overlap vectors $h^{{(i)}}$_{P,j} = Tr(P G_{i,j})/d_i, which enter the Hessian as rank-one positive-semidefinite contributions 2 Σ_P h_P h_P^T. Strict convexity is obtained by combining this PSD structure with a detectability condition and a smoothed-analysis argument that perturbs the calibrated point by a small Gaussian vector to remove accidental zero-curvature directions almost surely, leaving only gauge directions which are quotiented out. On top of this landscape, the proofs use zeroth-order online convex optimization: a projected two-point SPSA estimator that is unbiased for quadratic surrogates, with variance bounds that depend either on the total control dimension d or, for local codes, only on the locality constants s and c via graph-coloring of compatible control coordinates.","core_discovery":"After Pauli twirling, a small control displacement on any gate produces a quadratic response in the detector-relevant Pauli error probabilities; the Hessian of each detector's event rate is a sum of rank-one positive-semidefinite matrices, and averaging over detectors preserves positive-semidefiniteness. Directions in which the Hessian vanishes are either gauge directions, which can be quotiented out, or accidental degeneracies that a generic infinitesimal perturbation of the calibrated point removes with probability one. The paper therefore claims that the average detection rate C(θ) is generically locally strongly convex around the calibrated operating point for coherent unitary errors, general CPTP errors, and leakage or atom loss with block-diagonal readout. On this landscape, a projected simultaneous perturbation stochastic approximation (SPSA) controller, using only noisy estimates of C from finite syndrome batches, provably drives the detection rate to within ε of its calibrated value in O($ε^{{-2}}$) epochs, tracks slowly drifting optima with vanishing average dynamic regret, and, using the locality of detector regions in surface and quantum LDPC codes, removes the explicit dependence on code distance from the convergence rate.","pith_inferences":["If the smoothed-analysis assumption is not met on a given device, deliberately injecting small calibration noise could make the landscape genuinely strictly convex, at the price of slightly altering the implemented gates.","The convexity guarantee depends on Pauli twirling, so platforms that already use randomized compiling inherit self-calibration; platforms that do not must add twirling before the bound applies.","The per-epoch window length m sets a trade-off between noise suppression and tracking speed; adapting m to the observed drift rate could improve the dynamic-regret bound beyond the fixed-window analysis.","Estimating the detector-event Hessian from the syndrome stream could identify which control directions are actually visible to the detectors, allowing calibration to focus on a lower-dimensional subspace."],"forward_implications":["A fault-tolerant processor can run for extended periods without dedicated recalibration, because the syndrome stream continuously corrects slow control drifts.","Reaching a target detection-rate excess ε needs only O(ε^{-2}) syndrome epochs, matching the optimal zeroth-order scaling.","For surface codes and quantum LDPC codes, calibration complexity stays constant as the code grows, since detector locality removes the code-distance dependence.","The detection-rate objective gives a rigorous, convex training signal for more expressive controllers such as neural-network-based calibration agents."],"supporting_citations":[{"why":"Motivates the framework by demonstrating experimentally that syndrome measurements can serve as a calibration signal.","marker":"[1]"},{"why":"Supplies the surface-code threshold demonstrations that define the operating regime this self-calibration protects.","marker":"[5]"},{"why":"Defines the bivariate-bicycle LDPC codes used to test code-distance-independent convergence.","marker":"[6]"},{"why":"Provides randomized compiling, converting coherent control errors into Pauli channels and producing the rank-one Hessian structure.","marker":"[17]"},{"why":"Supplies the online convex optimization framework and dynamic-regret definitions used for time-dependent drift.","marker":"[20]"},{"why":"Provides the circuit-level Clifford simulator used to validate the scaling predictions.","marker":"[24]"},{"why":"Supplies the smoothed-analysis technique that removes accidental zero-curvature directions via generic Gaussian perturbations.","marker":"[28]"},{"why":"Introduces the simultaneous perturbation stochastic approximation gradient estimator used in the calibration loop.","marker":"[32]"},{"why":"Provides the optimal zeroth-order convergence rates that the O(ε^{-2}) guarantee matches.","marker":"[33]"}],"fun_headline_variants":["Self-calibrating quantum fault tolerance proven efficient with syndrome data","Quantum self-calibration: local convexity guarantees O(1/ε²) convergence","Provably efficient self-calibration for quantum error correction","Syndrome-based self-calibration: now with a proof of efficiency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that the landscape is strictly convex uses a smoothed-analysis step that perturbs the assumed calibrated operating point by a small random vector; a real fixed processor has a specific, non-random operating point, and if that point happens to sit at an accidental zero-curvature direction, the 'with high probability' guarantee does not certify convexity for that particular machine.","fun_headline_variants_meta":{"raw":{"variants":["Self-calibrating quantum fault tolerance proven efficient with syndrome data","Quantum self-calibration: local convexity guarantees O(1/ε²) convergence","Provably efficient self-calibration for quantum error correction","Syndrome-based self-calibration: now with a proof of efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001172,"raw_usage":{"total_tokens":4899,"prompt_tokens":1047,"completion_tokens":3852,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":3777}},"tokens_in":663,"tokens_out":3852,"duration_ms":30808,"temperature":1.0,"reasoning_tokens":3777,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:59:17.262720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a single control parameter on a real device, scan it through its calibrated value while measuring the average detector-event rate, and check for a local maximum or flat region at the calibrated point; Theorem 1 predicts a positive curvature valley for every detector-visible direction, so a measured local maximum at θ* for a non-gauge direction would refute the generic strict-convexity claim for that hardware instance.","supporting_citations":[{"cited_title":"Recall from Eq","cited_arxiv_id":null,"evidence_quote":"Motivates the framework by demonstrating experimentally that syndrome measurements can serve as a calibration signal."},{"cited_title":"The setting is the same as for Theorem 9","cited_arxiv_id":null,"evidence_quote":"Supplies the surface-code threshold demonstrations that define the operating regime this self-calibration protects."},{"cited_title":"However, now we do not assume that we remain inside the convex ball of Appendix A","cited_arxiv_id":null,"evidence_quote":"Defines the bivariate-bicycle LDPC codes used to test code-distance-independent convergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides randomized compiling, converting coherent control errors into Pauli channels and producing the rank-one Hessian structure."},{"cited_title":"Its objective is the mean detection rate DR(p) = 1 ND NDX k=1 DRk(p), DR k(p) = Pr[detectorkfires],(F4) estimated frommMonte Carlo shots of the compiled detector sampler","cited_arxiv_id":null,"evidence_quote":"Supplies the online convex optimization framework and dynamic-regret definitions used for time-dependent drift."}],"review_version":1}