{"id":"6867814c-58b9-4c54-9cae-1228a14d1d31","arxiv_id":"1908.04416","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Variational quantum compiling's optimal parameters are provably unchanged by a broad class of incoherent noise, so noisy devices can still train the correct short-depth circuit.","lead":"This paper proves that variational quantum compiling still finds the correct gate sequence even when the hardware is noisy. The result could make a key near-term quantum computing tool more reliable in practice.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's weak-OPR is proven over the full unitary group; for a parameterized ansatz, even a complete ansatz may contain no noisy global optimizer, so the claim that noisy training finds correct parameters does not follow.","rationale":"The reader's weakest assumption correctly identifies the positivity and completeness assumptions. My concern is distinct and more specific: even granting strict positivity and a complete ansatz, weak-OPR over the full unitary group does not guarantee noise resilience for the restricted parameterized family used in practice. This is because weak-OPR only gives containment of noisy global optimizers in the noiseless set; it does not ensure the ansatz contains any noisy global optimizer. The paper explicitly notes that noise can break degeneracy for FISC, making this gap relevant. The theorems themselves appear sound under their stated assumptions, so I do not recommend changing the conditional verdict; however, the authors should either prove an ansatz-level statement or soften the claim that Theorem 3 implies correct variational parameters are learned.","tokens_in":36595,"tokens_out":37879,"duration_ms":398275,"concrete_test":"Construct a 2-qubit FISC instance with Noise Model 3 in which the Pauli-noise output distribution q has ordering q00 > q10 > q01 > q11 while the measurement distribution p has ordering p00 > p01 > p10 > p11, both with strictly positive coefficients. Take a complete ansatz containing W = 11 (i.e., V = U) and a second unitary W_eps that leaves |00> with small amplitude but nearly realizes the permutation mapping q's sorted basis to p's sorted basis, so W_eps is not a noiseless optimizer. Evaluate noiseless CLET and noisy ~CLET at both points. If ~CLET(W_eps) < ~CLET(11) while CLET(11) < CLET(W_eps), then weak-OPR over V_d does not imply ansatz-level parameter resilience, confirming the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 6 defines OPR on the full unitary group V_d, not on the parameterized family V(alpha). Theorem 3 proves weak-OPR over V_d: every global optimizer of the noisy cost over all unitaries is a noiseless global optimizer. But variational quantum compiling optimizes over an ansatz A = {V(alpha)}. Weak-OPR over V_d does not imply that argmin_A ~C is contained in argmin_A C. For strong-OPR, completeness of the ansatz is enough: if A contains a noiseless optimizer, it contains the same unitary as a noisy optimizer. For weak-OPR, however, noise can break the degeneracy of Vopt_d, selecting a subset S (Appendix F, Eqs. F6-F10). If the ansatz contains only noiseless optimizers outside S, the noisy global optimum over A may be a unitary that is not noiseless-optimal. The paper's own discussion says weak-OPR arises because noise can break degeneracy (Section V.B), and Section VII.C restricts the theorems to the complete-ansatz case. But completeness only ensures A intersects Vopt_d, not S. Hence the statement in Section V.B that weak-OPR 'implies that training in the presence of noise will lead one to find the correct optimal parameters' is not justified by Theorem 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Optimal Parameter Resilience (OPR) for variational quantum compiling (VQC). It defines strong-OPR as equality of the noisy and noiseless global-optimizer sets over the full unitary group V_d, and weak-OPR as inclusion of the noisy optimizers in the noiseless optimizers. The main analytic results are Theorem 1 and Theorem 2, which state that the Hilbert-Schmidt Test and Local Hilbert-Schmidt Test costs exhibit strong-OPR under two composite incoherent noise models (depolarizing, Pauli, non-unital Pauli, Pauli gate noise, and measurement noise), and Theorem 3, which states that the Loschmidt Echo Test and Local Loschmidt Echo Test costs exhibit weak-OPR under a simpler noise model. Corollaries extend some results to noise acting during the implementation of W = V^dagger U for Clifford and tensor-product cases. The paper also reports numerical implementations on IBM's noisy simulator for the Toffoli gate, the three-qubit QFT, and W-state preparation, observing that noisy training reduces the noiseless costs. The central mathematical statements are proven in Appendices D-G with explicit assumptions, and the core Cauchy-Schwarz and rearrangement arguments are valid.","tokens_in":36870,"tokens_out":8899,"duration_ms":102488,"significance":"If the results hold as stated, they identify a genuinely surprising property that is potentially useful for NISQ algorithms: for certain VQC cost circuits, the global optimum of the noisy cost over all unitaries coincides with, or is contained in, the noiseless global optimum, so the noise does not shift the target unitary. The proof strategy is transparent and the assumptions about positive Pauli eigenvalues and measurement-error probabilities are explicit. The strong-OPR results for FUMC are especially clean because equality of the full optimizer sets transfers to any complete ansatz containing an exact compilation. The numerical demonstrations are a useful proof-of-principle, although they consist of individual optimization runs without statistical error bars. The main caveat, discussed in the major comments, is that the weak-OPR theorem for FISC does not by itself justify the paper's parameter-training claims, because variational training optimizes over a parameterized ansatz rather than over the full unitary group.","major_comments":[{"comment":"The inference from weak-OPR over the full unitary group to parameter-level noise resilience is not justified. Definition 6 and Theorem 3 concern the sets Vopt_d and ~Vopt_d defined over all d-by-d unitaries in Eqs. (16)-(17), but variational quantum compiling optimizes a cost over an ansatz A = {V(alpha)}. For strong-OPR this gap is harmless: if the ansatz contains a noiseless optimizer and the noisy optimizer set equals the noiseless optimizer set, then the same element optimizes both costs over A. For weak-OPR, however, Theorem 3 only shows ~Vopt_d is a subset of Vopt_d; the proof in Appendix F shows noise can break the degeneracy and select the subset S defined in Eq. (F10). A complete ansatz contains some V with W = V^dagger U in Vopt_d, but it need not contain any W in S. If A intersects Vopt_d but not S, the global optimum of the noisy cost over A is not a noiseless global optimum. Therefore the statement in Section V.B that weak-OPR 'implies that training in the presence of noise will lead one to find the correct optimal parameters for V(alpha)' is unsupported, and the remark in Section VII.C that the theorems are restricted to the complete-ansatz case does not repair the gap, since completeness is sufficient for strong-OPR but not for weak-OPR. The same issue affects Corollaries 6-8, which invoke Theorem 3. The authors should either prove a version of Theorem 3 for parameterized ansatze, for example under a condition ensuring A intersects ~Vopt_d, or explicitly restrict the training-level conclusion and state Theorem 3 as a property of the cost landscape over unitaries.","section":"Definition 6, Section V.B, Appendix F"},{"comment":"The saturation arguments establishing strong-OPR in Theorems 1 and 2 require strict positivity of the coefficients kappa, xi, theta, and tau. These coefficients are positive only under the assumptions in Definitions 2, 3, and 5, namely all Pauli eigenvalues c_lk > 0 and measurement probabilities satisfying p00 > p01 and p11 > p10. Physical Pauli channels with a non-positive eigenvalue, such as a dephasing channel with error probability above 1/2, are therefore outside the theorems' scope. This is an explicit restriction rather than an internal inconsistency, but the abstract's phrase 'broad class of noise models, such as ... Pauli channel noise' should be qualified in the main text so readers do not conclude that all Pauli channels are covered.","section":"Appendices D-E, Eqs. (D20), (D26), (E9), (E18)"}],"minor_comments":[{"comment":"The numerical evidence consists of single optimization runs without statistical uncertainties. The text already uses cautious language such as 'appear to exhibit OPR', but it would be helpful to state explicitly that each curve is one run and that no error bars or repeated-initialization statistics are provided.","section":"Section VI and Figures 7-8"},{"comment":"The figure captions contain placeholder text 'Lorem ipsum'. These placeholders must be replaced with the intended descriptions before publication.","section":"Figure 1-4 captions"},{"comment":"The color labels are inconsistent between the text and the caption: the text refers to 'blue and red' curves while the caption says 'blue and green', and the pair 'green and pink' is used for two different roles. Please harmonize the color names.","section":"Section VI.B and Figure 7 caption"},{"comment":"The statement reads 'The cost functions CHST exhibits strong-OPR'; since only CHST is discussed, it should read 'The cost function CHST exhibits strong-OPR'.","section":"Corollary 5 statement"}],"recommendation":"major_revision","confidential_remarks":"The core theorems appear mathematically correct, and the strong-OPR results for FUMC are a solid contribution. The main issue is that the weak-OPR result for FISC is stated as if it directly guarantees correct variational parameters under noisy training, but the proof is over the full unitary group and completeness of the ansatz does not bridge that gap. This is fixable either by adding an ansatz-intersection condition or by repositioning Theorem 3 as a cost-landscape property. I would support publication after that point is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main new thing here is Optimal Parameter Resilience (OPR), and the paper earns it. The distinction between strong and weak OPR is useful, and the proofs of Theorems 1 and 2 for the Hilbert-Schmidt cost functions are honest math: the Cauchy-Schwarz and rearrangement arguments are valid, and the positivity assumptions are stated explicitly. This is not a toy model. The FUMC results in particular are solid and give a clean sufficient condition for when the global optimum over all unitaries is unchanged by a broad class of incoherent noise. I also appreciated the corollaries for Clifford and tensor-product unitaries; they extend the result to practically relevant cases without much extra machinery.\n\nWhere the paper gets soft is the step from Theorem 3 to the verbal claim that “training in the presence of noise will lead one to find the correct optimal parameters.” The second-pass stress test lands. Weak-OPR is proven over the full unitary group V_d, not over a parameterized ansatz A. Completeness of the ansatz only guarantees that A intersects the noiseless optimizer set Vopt_d, but noise can break degeneracy and select a subset S of Vopt_d. If A contains no element of S, the noisy global optimum over A may be outside Vopt_d. That is not a flaw in the theorem, but it is a gap in the interpretation, and the text in Section V.B overstates what follows. This matters most for FISC; the FUMC strong-OPR result does not have this problem.\n\nThe other soft spots are minor but real. The positivity assumptions (Pauli eigenvalues c_lk > 0, p00 > p01) exclude some physical channels; the authors are clear about this, so it is a scope restriction rather than a hidden flaw. The numerics show single trajectories with no error bars, and the figures still contain placeholder “Lorem ipsum” captions, so the experimental section is not submission-ready. That said, the numerics are consistent with the theory and go beyond what the theorems cover, especially for incomplete ansatzes, so they are a useful complement rather than the core evidence.\n\nWho should read this? Anyone working on variational quantum algorithms or noise-aware compilation. The OPR concept will be cited, and the strong-OPR theorem is a good starting point for proving robustness of other hybrid algorithms. The paper deserves serious peer review. I would want the authors to fix the ansatz interpretation, add error bars or multiple seeds, and replace the placeholder figures, but the central mathematical contribution is sound.\n\nRecommendation: send it to review. It is a genuine result with a recognized limitation, not a desk reject.","headline":"OPR is a real, well-proved concept for VQC, but the weak-OPR theorem does not by itself justify the claim that noisy training on a fixed ansatz finds correct parameters.","tokens_in":37333,"tokens_out":2682,"would_cite":true,"duration_ms":31250,"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 incoherent noise during cost evaluation leaves the optimal parameters of variational quantum compiling unchanged for broad classes of depolarizing, Pauli, gate, and measurement noise.","keywords":["variational quantum compiling","optimal parameter resilience","noise resilience","Hilbert-Schmidt test","Loschmidt echo","Pauli channels","measurement noise","NISQ"],"falsifier":"Compile a known two-qubit unitary with a complete ansatz using the HST cost, first with no noise and then with a Pauli channel that has one of its Pauli eigenvalues set to zero while preserving complete positivity, and compare the optimal parameters; the strong-OPR theorem predicts identical optima only when the positivity condition holds, so any shift would falsify the claim.","tokens_in":1975,"feed_emoji":"⚛️","tokens_out":2926,"duration_ms":113611,"temperature":0.7,"pith_summary":"Variational quantum compiling tries to train a short gate sequence $V(\\alpha)$ to mimic a target unitary $U$ by minimizing a quantum-evaluated cost. This paper's central claim is that for several standard cost functions the optimal parameters are the same whether or not the cost-evaluation circuit is subject to incoherent noise: measurement error, depolarizing noise, Pauli noise, non-unital Pauli noise, and Pauli gate noise. The paper formalizes this as Optimal Parameter Resilience and proves strong-OPR for the Hilbert-Schmidt and local Hilbert-Schmidt costs (full unitary compiling) under two composite noise models, and weak-OPR for the Loschmidt Echo and local Loschmidt Echo costs (fixed-input compiling) under a third. This matters because noise resilience means the circuit-depth compression that VQC promises is not undone by the noise present when evaluating the cost; the learned $V$ really is the correct short circuit, so replacing $U$ by $V$ should reduce noise. Numerical experiments on a realistic noisy-device simulator for the Toffoli gate, quantum Fourier transform, and W-state preparation show the noiseless cost decreasing to about $10^{-4}$ to $10^{-5}$ during noisy training.","feed_headline":"Incoherent noise leaves the best compiled circuit unchanged","feed_subtitle":"Measurement, gate, and Pauli noise leave the optimal compiled parameters intact, keeping depth compression viable.","key_machinery":"The central objects are four cost-evaluation circuits: the Hilbert-Schmidt Test (HST) and Local Hilbert-Schmidt Test (LHST) for full unitary compiling, and the Loschmidt Echo Test (LET) and Local Loschmidt Echo Test (LLET) for fixed-input compiling. The mechanism carrying the argument is the expansion of noisy circuits in the Pauli basis, where Pauli channels are diagonal, non-unital Pauli channels add a Pauli tail to the identity, and measurement noise becomes a weighted bit-flip POVM. In the HST and LHST proofs, the noisy cost decomposes into a sum over Pauli strings, and each term is bounded by a Cauchy-Schwarz inequality; strict positivity of the noise coefficients $\\kappa$, $\\xi$, $\\vartheta$, and $\\tau$ makes the bound tight exactly when $W = V^\\dagger U$ is proportional to the identity. For the LET and LLET proofs, a rearrangement inequality on the doubly stochastic matrix $w_{il} = |\\langle i|W|l\\rangle|^2$ bounds the noisy fidelity by the ordered inner product $p^\\downarrow \\cdot q^\\downarrow$, saturated by permutations that map the noisy input basis to the noisy measurement basis, a subset of the noiseless optima.","core_discovery":"On its own terms, the paper's central discovery is that Optimal Parameter Resilience is a genuine property of variational quantum compiling. For Full Unitary Matrix Compiling, the cost functions $C_{\\mathrm{HST}} = 1 - |\\mathrm{Tr}(V^\\dagger U)|^2/d^2$ and its local counterpart $C_{\\mathrm{LHST}}$ have exactly the same set of global optimizers in the presence of Noise Model 1 or Noise Model 2 as in the absence of noise (strong-OPR, Theorems 1 and 2). For Fixed Input State Compiling, the Loschmidt Echo cost $C_{\\mathrm{LET}}$ and its local version $C_{\\mathrm{LLET}}$ have the property that every optimizer under Noise Model 3 is also a noiseless optimizer (weak-OPR, Theorem 3). The proofs show that the noisy cost is bounded by a noise-dependent constant through Cauchy-Schwarz or rearrangement inequalities, with equality holding exactly at the noiseless optima when the noise coefficients are strictly positive. Numerical simulations with a realistic hardware noise model show that training in the noisy setting drives the noiseless cost to about $10^{-4}$ to $10^{-5}$, matching and even extending the theorems.","pith_inferences":["A testable extension: the proof mechanism suggests resilience should hold when the noise channel is any unital or non-unital Pauli map with strictly positive Pauli coefficients, even if the noise acts at times not covered by Theorems 1–3; the numerics already hint at this because the simulator noise includes non-unital terms throughout $W$.","The strict positivity condition draws a sharp boundary: as any coefficient approaches zero, the Cauchy-Schwarz bound stops being tight at the noiseless optimum, so one might see a transition where noise resilience fails; scanning a dephasing strength from small to large would map where the optimal parameters start to move.","Weak-OPR could be turned into a design tool: since noise prunes the degenerate set of noiseless optima, one could engineer the measurement noise to select an optimum with desirable properties such as fewer entangling gates, though the paper does not explore this."],"forward_implications":["Circuit-depth compression with VQC remains meaningful on noisy hardware: because the parameters learned in noise match the noiseless optimum, the shorter compiled circuit $V$ should genuinely incur less noise than the longer target $U$.","Readout errors do not need to be corrected or calibrated for the purpose of finding the optimal compilation; the optimal parameters are unaffected by the measurement noise model of Definition 5.","Pauli gate noise during the entangling and disentangling stages, depolarizing noise throughout, and non-unital (T1-like) noise at specified times are all tolerated together in FUMC; Clifford or tensor-product structure of $W$ extends this resilience to noise acting during $W$ itself.","For fixed-input compiling, weak-OPR means noisy training may select a subset of the already degenerate noiseless optima, so any circuit it finds still compiles $U|0\\rangle$ correctly, but the optimizer might not explore all equally good compilations.","Because each VQC cost is an expectation value of an effective Hamiltonian, the paper's results show that VQE exhibits the same optimal parameter resilience for those specific Hamiltonians."],"supporting_citations":[{"why":"Introduced the HST and LHST cost functions and circuits for full unitary matrix compiling that Theorems 1 and 2 are about.","marker":"[19]"},{"why":"Introduced the fixed-input-state compiling approach that underlies the LET and LLET costs treated in Theorem 3.","marker":"[20]"},{"why":"Supplied the prior notion of cost value resilience against coherent errors, which the paper contrasts with the new optimal parameter resilience.","marker":"[7]"},{"why":"Established the VQE framework used in the warm-up example and in the claim that VQC is a special case of VQE.","marker":"[8]"},{"why":"Provided the noisy-device simulator and realistic noise model used for the numerical demonstrations of resilience.","marker":"[28]"},{"why":"Supplied the rearrangement inequality used in the proof of weak-OPR for the LET and LLET costs.","marker":"[48]"}],"fun_headline_variants":["Noise can't shift the optimal compiled circuit","Quantum compiling stays optimal despite noise","Optimal parameters resist gate and measurement noise","Noise leaves best quantum compiled circuit unchanged"],"cache_read_input_tokens":39552,"weakest_assumption_plain":"The theorems require that every noise coefficient entering the Cauchy-Schwarz and rearrangement bounds be strictly positive, which for Pauli channels means all Pauli eigenvalues are positive and for measurement noise means the correct outcome is more likely than the incorrect one, and they further assume the variational ansatz is complete enough to contain an exact compilation of the target unitary.","fun_headline_variants_meta":{"raw":{"variants":["Noise can't shift the optimal compiled circuit","Quantum compiling stays optimal despite noise","Optimal parameters resist gate and measurement noise","Noise leaves best quantum compiled circuit unchanged"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1747,"prompt_tokens":1002,"completion_tokens":745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":691}},"tokens_in":618,"tokens_out":745,"duration_ms":7835,"temperature":1.0,"reasoning_tokens":691,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:43:44.568967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compile a known two-qubit unitary with a complete ansatz using the HST cost, first with no noise and then with a Pauli channel that has one of its Pauli eigenvalues set to zero while preserving complete positivity, and compare the optimal parameters; the strong-OPR theorem predicts identical optima only when the positivity condition holds, so any shift would falsify the claim.","supporting_citations":[{"cited_title":"Quantum- assisted quantum compiling,","cited_arxiv_id":null,"evidence_quote":"Introduced the HST and LHST cost functions and circuits for full unitary matrix compiling that Theorems 1 and 2 are about."},{"cited_title":"The theory of variational hybrid quantum-classical algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplied the prior notion of cost value resilience against coherent errors, which the paper contrasts with the new optimal parameter resilience."},{"cited_title":"A variational eigenvalue solver on a photonic quantum processor,","cited_arxiv_id":null,"evidence_quote":"Established the VQE framework used in the warm-up example and in the claim that VQC is a special case of VQE."},{"cited_title":"Hardy, J.E","cited_arxiv_id":null,"evidence_quote":"Supplied the rearrangement inequality used in the proof of weak-OPR for the LET and LLET costs."}],"review_version":1}