{"id":"e35d177b-921c-4db9-9e4e-8effc1247d27","arxiv_id":"2411.16653","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Two new feature maps for Clifford Data Regression, geometric and insertion, with the insertion variants reducing RMSE in small noisy-circuit simulations.","lead":"This paper extends Clifford Data Regression, a quantum error mitigation method, with two new ways to build the feature vectors used for learning: one that repeats the noisy circuit multiple times, and one that inserts a small rotation circuit inside it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The insertion feature map's theoretical motivation is unsound: Eq. (18) coefficients depend on U1 and U2, contradicting the paper's claim of U-independence used to justify learning a single alpha.","rationale":"The reader's weakest_assumption correctly identifies the false U-independence claim in Section 3.3 as the most important theoretical defect. The claim is used to justify why one learned coefficient vector α can serve all circuits in the training distribution, and it is demonstrably wrong because the coefficients in (18) contain overlaps with U1 and U2. I agree this is a load-bearing error in the paper's argument, though it does not invalidate the formally correct Theorem 1 nor the numerical RMSE results. The paper's actual protocol retrains α for each target U, which mitigates the practical impact, but the motivation as written is unsound and requires correction. Reproducibility issues (missing split point, no code/data) are secondary but reinforce the need for revision. The verdict should remain CONDITIONAL as the reader concluded; my stress-test does not move it further.","tokens_in":6,"tokens_out":11594,"duration_ms":289405,"concrete_test":"For a single-qubit example, set U=RY(θ), V=RX(π/8), split U1=I, U2=U. Evaluate the exact coefficients in (18) for several θ values. They equal <0|U†|v_p><v_p|O|v_q><v_q|U|0> and hence vary with θ, directly refuting the independence claim. Then train a ridge-regression α on noiseless insertion features for near-Clifford Wi generated from U(θ0) and measure the error on U(θ) for θ away from θ0. If the error grows with |θ−θ0|, the transferability of a single α is not automatic, confirming the concern is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 claims that the coefficients of the exponential expansion for the insertion feature map are independent of U because the frequencies belong to Ω(V). This is false. Expanding (18) in the eigenbasis of V gives coefficients <0|U1†|v_p><v_p|U2† O U2|v_q><v_q|U1|0>, which explicitly depend on both U1 and U2. The paper's stated justification for why a single learned α can be transferred from Clifford training circuits Wi to the target U therefore collapses. Theorem 1 does provide a formal generalization bound relating |f(U)−α^T φ∞(U)| to the training error and a Rademacher term, and this proof does not rely on the independence claim. However, the paper's narrative presents the independence as the reason the method is valid, and without it the reader is left with no theoretical argument that the training error obtained on near-Clifford circuits with different U1,U2 transfers to U. The numerical experiments may still demonstrate empirical improvement, but the theoretical underpinning as written is incorrect and must be repaired.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two new feature maps for Clifford Data Regression (CDR) quantum error mitigation: a geometric feature map that uses powers of the noisy circuit, and an insertion feature map that inserts a known unitary V^t into the target circuit, together with a noise-scaled variant (insertion-ZNE). The authors analyze the computational complexity of the resulting estimators, prove a generalization bound based on Rademacher complexity (Theorem 1), and derive an information-theoretic lower bound on the required resources (Theorem 2). They support the proposals with numerical experiments on random 3-qubit circuits and on QFT circuits, reporting reduced root mean square error relative to classical CDR and ZNE-based CDR.","tokens_in":26155,"tokens_out":6137,"duration_ms":57940,"significance":"The empirical finding that the insertion and insertion-ZNE feature maps reduce RMSE relative to classical CDR and ZNE-based CDR, especially in the finite-sample regime, is potentially useful for practical error mitigation. The paper also contributes formal results: Theorem 1 provides a concrete generalization bound with explicit scaling in S, J, and N, and Theorem 2 gives a resource trade-off that goes beyond generic impossibility bounds. The complexity analysis in Table 1 is a useful practical guide. However, the theoretical motivation for the insertion method contains a false claim about U-independence of the expansion coefficients, and a similar conflation appears in the justification of the geometric feature map, so the theoretical narrative needs repair before the results can be accepted as presented.","major_comments":[{"comment":"The claim that the coefficients in the linear combination for the insertion feature map are independent of U because the frequencies belong to Ω(V) is false. Expanding Eq. (18) in the eigenbasis of V gives coefficients of the form ⟨0|U1†|v_q⟩⟨v_q|U2†OU2|v_p⟩⟨v_p|U1|0⟩, which explicitly depend on both U1 and U2. This invalidates the stated justification for learning a single coefficient vector α on training circuits with different U1 and U2 and applying it to the target U. The paper should either provide a correct invariance argument or replace this motivation with the generalization bound of Theorem 1, which does not rely on U-independence of the coefficients.","section":"Section 3.3, Eq. (18)"},{"comment":"The same conflation of frequency independence with coefficient independence occurs in the support for the geometric feature map. Proposition 1 guarantees that g(t,U) can be approximated by a Fourier series with frequencies {±q/p}, but the optimal coefficients c_q depend on U through the eigenbasis overlaps ⟨0|u_i⟩. Therefore the statement that replacing Ω(U) by a U-independent frequency set implies 'there exists a set of coefficients such that f(U) ≈ ... for all circuits on the training set' is not justified. The theoretical support for the geometric feature map should be restated to avoid claiming a single coefficient set works for all U.","section":"Section 3.2, after Proposition 1"}],"minor_comments":[{"comment":"There are several typographical errors, including 'fullfill' in Section 2.1, 'an the' in Section 3, and missing commas between the blocks of the vector in Eq. (19).","section":"Throughout"},{"comment":"The use of '≈O' is nonstandard; the asymptotic statement should be written as '= O(...)' or the approximation sign alone.","section":"Section 4.2, Eq. (28)"},{"comment":"The caption refers to 'Figure 1b' when describing the outlier threshold; this should be 'Figure 9b'.","section":"Appendix E, Figure 9 caption"},{"comment":"The quantity N(θ) is used in the bound before its definition in Eq. (27); the definition should be introduced before or with the theorem statement.","section":"Theorem 1"}],"recommendation":"major_revision","confidential_remarks":"The false U-independence claim appears twice and is a clear mathematical error, but it does not invalidate the core formal results: Theorem 1, Theorem 2, and the complexity analysis do not depend on this claim. The numerical evidence for the insertion methods appears solid. I recommend requiring the authors to correct the motivation, either by removing the false claim and relying on the generalization bound, or by providing a different theoretical rationale. The paper should also carefully re-examine the analogous claim in Section 3.2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a decent incremental QEM paper with a real empirical finding and an unsound motivational claim that needs fixing. The insertion and insertion-ZNE feature maps are new, simple, and show consistent RMSE reduction over CDR and ZNE-CDR in the 3-qubit simulations. The geometric map is a sensible idea and honestly reported as failing, which I credit. The complexity table and the Rademacher generalization bound are useful; Theorem 1 does not rely on the false independence claim, and the appendix checks the scaling numerically. That is real work and should be recognized.\n\nThe soft spot is Section 3.3. After Eq. (18) the paper says the coefficients of the exponential expansion are independent of U because the frequencies belong to Omega(V). That is false: expanding in the eigenbasis of V gives coefficients <0|U1^dag|v_p><v_p|U2^dag O U2|v_q><v_q|U1|0>, which depend on both halves of U. The stated reason for transferring a single alpha from Clifford training circuits to the target circuit collapses. Theorem 1 offers an alternative route through a Rademacher bound over the training distribution, but the paper does not explicitly make that substitution in the motivation. So the theoretical narrative needs repair, not a footnote. This does not invalidate the numerics, but it means the method currently rests on empirical evidence plus a bound that is not connected to the stated motivation.\n\nReproducibility is also weaker than it should be: no code or data release, no specification of where U is split, and no error bars on the RMSE plots. The experiments are 3 qubits, 30 gates, depolarizing noise on CNOTs only—fine for a first look, not enough to claim practical advantage.\n\nWho this is for: people actively building CDR variants. They will want to see whether insertion survives on larger circuits and with better noise models. It deserves a serious referee—conditional accept after major revision. The authors should fix Sec 3.3, add explicit split-point details, release code and data, and add error bars. I would bring it to a reading group to discuss whether the generalization bound can stand in for the missing independence argument.","headline":"Useful empirical extension of CDR with a broken motivational argument in Sec 3.3 that a careful revision can fix.","tokens_in":26739,"tokens_out":3240,"would_cite":true,"duration_ms":32158,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes two new feature maps for Clifford Data Regression—inserting an extra unitary between circuit halves, and adding noise-scaled variants—and shows they reduce root mean square error in numerical simulations.","keywords":["Clifford data regression","quantum error mitigation","zero-noise extrapolation","feature maps","Rademacher complexity","near-Clifford circuits","expectation value estimation","gate folding"],"falsifier":"Train $\\alpha$ on near-Clifford circuits, then compute the noiseless interpolation error $|f(U) - \\alpha^T \\phi_\\infty(U)|$ for many target circuits with the same gate layout but widely separated rotation angles; if this error grows systematically as the angles move away from the training distribution, the claimed $U$-independence of the learned coefficients is refuted.","tokens_in":25725,"feed_emoji":"📉","tokens_out":4768,"duration_ms":48639,"temperature":0.7,"pith_summary":"The paper aims to make Clifford Data Regression (CDR) more accurate by feeding it richer feature vectors, not just the noisy expectation value of the original circuit. It introduces two feature maps: one that evaluates powers of the target circuit, and one that splits the circuit into two halves and inserts a known unitary between them. The insertion map, alone or combined with zero-noise extrapolation, is shown numerically to reduce root mean square error relative to classical CDR and ZNE-based CDR in three-qubit experiments. The paper also derives a Rademacher-complexity bound on the estimation bias and information-theoretic lower bounds on the required resources.","feed_headline":"Two feature maps lower Clifford Data Regression error","feed_subtitle":"By inserting a known rotation circuit and noise-scaled copies, the method beats standard CDR on 3-qubit simulations.","key_machinery":"The central objects are the feature vectors themselves: the geometric map $\\phi(U) = [1, \\phi(\\tilde U), \\phi(\\tilde U^2), \\ldots, \\phi(\\tilde U^J)]^T$ and the insertion map $\\phi(U) = [1, \\phi(\\tilde U_2 \\tilde V^{t_1} \\tilde U_1), \\ldots, \\phi(\\tilde U_2 \\tilde V^{t_J} \\tilde U_1)]^T$, where $\\tilde{\\cdot}$ denotes the noisy implementation. These feature maps turn the estimator into an affine combination of noisy measurements of perturbed circuits. The theoretical analysis rides on two tools: the decomposition of single-qubit rotations into Clifford channels, which lets the noisy feature values of $U$ be expressed as a weighted sum of feature values of near-Clifford circuits, and a Rademacher-complexity generalization bound that controls the difference between expectation over the training distribution and the finite training sample.","core_discovery":"The central claim is that the insertion feature map and its noise-scaled variant, insertion-ZNE, improve the accuracy of Clifford Data Regression. For a target unitary $U$, the circuit is split as $U = U_2 U_1$ and perturbed circuits $U_2 V^{t} U_1$ are measured, with $V$ a fixed tensor product of single-qubit rotations and $t$ a real parameter; because $V^t$ is implemented by changing rotation angles, the enlarged circuits do not grow with $t$. The coefficients of the linear estimator are learned by ridge regression on near-Clifford circuits with the same gate layout. The paper reports that these insertion-based feature maps outperform classical CDR, the geometric feature map, and a ZNE-based feature map in the tested 3-qubit circuits, and that the bias term is bounded by $N(\\theta)\\sqrt{J+1}/\\sqrt{S}$ plus a training-error term, so the expected error decays roughly as $1/\\sqrt{S}$ and $1/\\sqrt{N}$.","pith_inferences":["The paper's stated justification for transferability—that the expansion coefficients are independent of $U$ because the frequency set $\\Omega(V)$ depends only on $V$—does not account for the $U$-dependent Fourier coefficients that appear in the expansion of Eq. (18); if those coefficients matter, the learned $\\alpha$ may be sensitive to which target circuit is being estimated.","The performance of the insertion method likely depends on where $U$ is split and on the choice of inserted unitary $V$; the paper fixes one split and $V = R_X(\\pi/8)\\otimes I\\otimes I$, so testing other splits and rotations is a natural next step.","Because the method learns a linear combination of noisy expectation values, its success on circuits far outside the near-Clifford training distribution remains an open question; a stress test with circuits whose rotation angles are deliberately far from those used in training would clarify the limits of transfer.","The same insertion idea could be extended to non-Clifford training circuits if classical simulation remains tractable, or to nonlinear post-processing of the feature vector, as the paper itself suggests as future work."],"forward_implications":["In the 3-qubit random-circuit experiments, the insertion and insertion-ZNE feature maps reduce root mean square error compared with classical CDR and ZNE-based CDR for the tested noise levels and feature-map sizes.","The bias of the learned estimator is bounded by $N(\\theta)\\sqrt{J+1}/\\sqrt{S}$ plus the training error, so increasing the training set size $S$ or the number of samples $N$ lowers the expected error, with an optimal intermediate size for the feature map.","The straightforward implementation of the insertion method has complexity $O(SNJ\\ell + n^2\\ell S + J^3)$, about a factor $J$ more than classical CDR; a kernel-trick variant reduces this when $J \\gg S$.","The information-theoretic analysis implies a trade-off: at least one of $J$, $S$, and $N$ must grow as $(1-p)^{-1/3}$ (or $(1-p)^{-2d/3}$ for depth-$d$ circuits) as depolarizing noise increases.","The insertion-ZNE variant is reported to be more resilient to finite sampling than the ZNE-based method, which degrades noticeably when the number of samples $N$ is limited."],"supporting_citations":[{"why":"Defines the original Clifford Data Regression method with the simple feature map $[1, \\phi(\\tilde U)]$, which serves as the baseline the paper extends and compares against.","marker":"[13]"},{"why":"Introduces the ZNE-based CDR feature map that the paper compares with its insertion-ZNE variant, and which supplies the noise-scaling approach it modifies.","marker":"[16]"},{"why":"Provides the decomposition of rotations into Clifford channels that the paper uses to express noisy feature values of $U$ as weighted sums of near-Clifford circuits in Theorem 1.","marker":"[18]"},{"why":"Supplies the gate-folding technique used to implement the noise-scaled perturbations in the ZNE and insertion-ZNE feature maps.","marker":"[19]"},{"why":"Gives the stabilizer simulation algorithm whose cost $O(n^2\\ell)$ drives the training-set complexity analysis.","marker":"[21]"},{"why":"Provides the Rademacher-complexity generalization bound used in Lemma 4 and Theorem 1 to control the difference between population and empirical training error.","marker":"[34]"},{"why":"Supplies the Jackson-type approximation theorem used in Proposition 1 to replace the true frequency set by a fixed set independent of $U$.","marker":"[33]"},{"why":"Establishes the information-theoretic lower-bound framework that Theorem 2 adapts to derive trade-offs among $J$, $S$, and $N$ as functions of noise strength.","marker":"[26]"}],"fun_headline_variants":["Insertion feature maps cut CDR error in simulations","Two CDR variants reduce error via inserted rotations","New feature maps improve Clifford data regression","Insertion-ZNE beats standard CDR on 3-qubit tests","Rotated circuits lower error in quantum mitigation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on one learned coefficient vector transferring from near-Clifford training circuits to the target circuit $U$, on the grounds that the frequencies of the inserted gate $V$ do not depend on $U$; the $U$-dependent overlap coefficients that also appear in the expansion are what would break this transfer.","fun_headline_variants_meta":{"raw":{"variants":["Insertion feature maps cut CDR error in simulations","Two CDR variants reduce error via inserted rotations","New feature maps improve Clifford data regression","Insertion-ZNE beats standard CDR on 3-qubit tests","Rotated circuits lower error in quantum mitigation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1372,"prompt_tokens":891,"completion_tokens":481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":507,"tokens_out":481,"duration_ms":5233,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:56:18.775270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train $\\alpha$ on near-Clifford circuits, then compute the noiseless interpolation error $|f(U) - \\alpha^T \\phi_\\infty(U)|$ for many target circuits with the same gate layout but widely separated rotation angles; if this error grows systematically as the angles move away from the training distribution, the claimed $U$-independence of the learned coefficients is refuted.","supporting_citations":[{"cited_title":"Error miti- gation with clifford quantum-circuit data,","cited_arxiv_id":null,"evidence_quote":"Defines the original Clifford Data Regression method with the simple feature map $[1, \\phi(\\tilde U)]$, which serves as the baseline the paper extends and compares against."},{"cited_title":"Unified approach to data-driven quantum error mitiga- tion,","cited_arxiv_id":null,"evidence_quote":"Introduces the ZNE-based CDR feature map that the paper compares with its insertion-ZNE variant, and which supplies the noise-scaling approach it modifies."},{"cited_title":"Unbiased simulation of near-clifford quantum circuits,","cited_arxiv_id":null,"evidence_quote":"Provides the decomposition of rotations into Clifford channels that the paper uses to express noisy feature values of $U$ as weighted sums of near-Clifford circuits in Theorem 1."},{"cited_title":"Digital zero noise extrapolation for quantum error mitigation,","cited_arxiv_id":null,"evidence_quote":"Supplies the gate-folding technique used to implement the noise-scaled perturbations in the ZNE and insertion-ZNE feature maps."},{"cited_title":"Improved simulation of stabilizer circuits,","cited_arxiv_id":null,"evidence_quote":"Gives the stabilizer simulation algorithm whose cost $O(n^2\\ell)$ drives the training-set complexity analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Rademacher-complexity generalization bound used in Lemma 4 and Theorem 1 to control the difference between population and empirical training error."},{"cited_title":"11, American Mathematical Soc., 1930","cited_arxiv_id":null,"evidence_quote":"Supplies the Jackson-type approximation theorem used in Proposition 1 to replace the true frequency set by a fixed set independent of $U$."},{"cited_title":"Information-theoretic bounds on quantum advantage in machine learning,","cited_arxiv_id":null,"evidence_quote":"Establishes the information-theoretic lower-bound framework that Theorem 2 adapts to derive trade-offs among $J$, $S$, and $N$ as functions of noise strength."}],"review_version":1}