{"id":"8f0a3e2f-830b-4423-a17f-f3f6a4b64594","arxiv_id":"2608.02325","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new resource measure, dynamical stabilizer entropy, provably governs the sample complexity of classical surrogates for quantum circuit expectation-value functions, and separates learnability from simulability under BQP ⊄ BPP.","lead":"The paper introduces a new quantum \"resource\" measure — dynamical stabilizer entropy (DSE) — that measures how spread out a quantum circuit's output function is across trigonometric frequency modes, and proves it sharply controls how many quantum measurements a classical surrogate needs for training. It also exhibits circuit families that are efficiently learnable from quantum-generated data but cannot be efficiently emulated from their circuit descriptions, unless BQP ⊆ BPP.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DSE-learnability lower bound is only for relative error; superlog-DSE families may be trivially learnable at fixed target accuracy.","rationale":"The reader's weakest assumption concerned the per-instance DSE ≤ OSE phase-diagram region; that is a real presentation overclaim but is secondary to the central learnability claim. The more load-bearing issue is that the central lower bound is only proved for relative accuracy: the hard family's squared norm ν_f decays as 2^{-2Mλ}, so for any fixed constant ε the zero surrogate already succeeds. This does not invalidate the formal lemma, but it means the paper has not shown that superlogarithmic DSE implies intractability for constant target error. The upper bound is also stated with ε in the exponent, so the 'polynomial when Mλ = O(log N)' claim requires constant accuracy. Together these make the advertised DSE-based phase transition an overstatement of the formal results. Since the formal theorems remain correct and the issue is addressable by clarifying the accuracy convention or finding a constant-ν_f hard family, the conditional verdict stands without change.","tokens_in":62710,"tokens_out":26745,"duration_ms":244521,"concrete_test":"Compute, for the SI E lower-bound family with Mλ = 20 and fixed ε = 0.01, the zero-predictor risk R(0) = ν_f = 2^{-40}. Since R(0) < ε, the family is learned with zero samples, while the Lemma 1 lower-bound condition ε ≤ ν_f/4 fails. This single check exposes the accuracy-convention gap: the claimed superlog-DSE intractability does not apply to fixed target accuracy, and the abstract/main-text phrasing would need to be restricted to relative accuracy or replaced by a construction with ν_f = Ω(1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal lower bound in Lemma 1/Theorem 5 (SI E, Eq. E19) applies only under the condition ε ≤ ν_f/4. The hard family constructed in SI E has ν_f = 2^{-2⌊Mλ⌋} (Eq. E6), so for any fixed target accuracy ε > 0 and all sufficiently large Mλ, the condition fails. For such instances the zero predictor h ≡ 0 already achieves R(0) = ν_f < ε with zero samples. Consequently, the main-text claim in Sec. II.C that 'when Mλ grows superlogarithmically, there exist circuit families whose expectation-value functions cannot be learned sample-efficiently' is not established for the standard constant-accuracy setting; it holds only for relative accuracy ε ≤ ν_f/4, which for the constructed family is exponentially small. The upper bound (Theorem 4) is stated for arbitrary ε but has ε in the exponent, so 'polynomial for Mλ = O(log N)' also implicitly requires constant ε. The formal inequalities are internally consistent, but the paper's headline interpretation of DSE as governing constant-error learnability is stronger than what is proved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a resource measure called dynamical stabilizer entropy (DSE) for families of parametrized quantum circuits, defined as the maximum Rényi entropy of the frequency-mode distribution induced by the circuit's expectation-value function. The authors claim that DSE tightly governs the classical learnability of such circuit families: they prove sample-complexity upper and lower bounds in terms of DSE (Lemma 1 / Theorems 4 and 5), construct a DSE-guided classical surrogate with a quantum feature-identification subroutine (Theorem 2), and prove, under BQP ⊄ BPP, that some low-DSE families are efficiently learnable from quantum-generated data but not efficiently simulable from circuit descriptions alone (Theorem 3). Numerical experiments with up to 80 qubits are reported as supporting the predicted phase diagram. The main text is accompanied by a detailed Supplementary Information containing the formal statements and proofs.","tokens_in":62948,"tokens_out":6315,"duration_ms":60864,"significance":"If the claims as stated were fully established, the paper would provide a resource-theoretic characterization of a practically relevant regime: when quantum-data-assisted classical surrogates can efficiently predict expectation values of tunable quantum circuits even when individual instances are hard to simulate. The paper has real strengths: the information-theoretic upper bound uses a concrete truncation argument controlled by DSE; the lower bound is a Fano-type argument built on an explicit hard family; the quantum subroutine for frequency sampling is described in enough detail to count gates; and the BQP↛BPP separation gives a formal sense in which learning from quantum data can outperform description-only simulation. The numerical section is extensive and largely consistent with the qualitative picture. However, a central advertised consequence—that superlogarithmic DSE obstructs sample-efficient learning at a fixed target accuracy—is not supported by the lower bound as proved, because the hard family has exponentially small squared norm ν_f and therefore becomes trivially learnable by the zero predictor once ε is fixed. This issue affects the main interpretation of the paper","major_comments":[{"comment":"The lower-bound statement requires ε ≤ ν_f/4, and the hard family constructed in SI E has ν_f = 2^{-2⌊M_λ⌋} (Eq. E6). For any fixed constant target accuracy ε > 0 and all sufficiently large M_λ, the condition ε ≤ ν_f/4 fails; in that regime the zero predictor h ≡ 0 achieves R(0) = ν_f < ε with zero samples. Thus the lower bound does not establish the main-text claim that superlogarithmic M_λ makes expectation-value functions sample-inefficient to learn at standard constant accuracy. It only establishes hardness for relative accuracy ε ≤ ν_f/4, which for the constructed family is exponentially small in M_λ. This is not a cosmetic caveat: for fixed ε the constructed family is actually trivially learnable as soon as ν_f < ε, so the claimed 'exact dividing line' for constant-error learnability is not proved. The authors should either construct a superlog-DSE hard family with ν_f a constant (","section":"§II.C, Lemma 1 (formal: SI C, Eq. C1; SI E, Eq. E6)"},{"comment":"The infeasible region of the phase diagram, where DSE > OSE, is asserted to be excluded because 'DSE is upper-bounded by a magic-resource measure in the worst case.' But Theorem 1 proves only worst-case inequalities: max_U DSE ≲ d ≲ max_{x,U} OSE. The two maxima are taken, respectively, over the circuit family for DSE and over instances for OSE, and the bound does not imply DSE(U) ≤ OSE(U) for the same circuit U. The only per-instance evidence is the numerical check in Fig. 2d on 446 instances of one structured family. The phase-diagram completeness claim therefore needs either a genuine per-instance proof or a substantially softened statement that distinguishes the worst-case relation from an instance-wise one.","section":"§II.B, Theorem 1 and Fig. 1c (formal: SI B3, Eq. B16)"},{"comment":"The separation theorem is stated for a 'fixed sufficiently small constant target error ε > 0' with d ≥ 1 and dε ≤ 1/288. This forces d = O(1/ε) = O(1), and Lemma 15 then gives M^(1)[O(x;U_ℓ)] ≤ log(2d) = O(1), not O(log N) as claimed in the main text. If instead d is taken to grow with N so that M = O(log N), then ε must be inverse-polynomial in d, which is not the fixed-constant-accuracy setting used elsewhere in the paper. The construction proves a valid separation for constant DSE, but the advertised logarithmic-DSE scaling requires reconciliation between the error parameter and the number of rotation gates.","section":"§II.D and SI H (Lemma 15 and Theorem 3)"}],"minor_comments":[{"comment":"The notation M^(α)[O(x;U)] leaves implicit the dependence on the input distribution D and the reference state ρ0. Since the induced probability p(ω) depends on both, the definition should state explicitly where D and ρ0 enter; the current text may confuse readers who take the resource measure to be a property of the circuit alone.","section":"Eq. (4) and SI B1"},{"comment":"The text says 'For both settings of d={2,8}' but the figure panels show d=2 and d=6. Check that the labels and description match.","section":"§III / Fig. 2a"},{"comment":"The numerical surrogate uses ridge regression with α=1, while Theorem 2 analyzes an unregularized kernel surrogate. The relation between the practical implementation and the theoretical predictor should be clarified; otherwise a reader may wonder whether the regularization changes the sample-complexity guarantees.","section":"SI I1"},{"comment":"References [59] and [134] appear to be the same paper (Sweke et al., 'Potential and limitations of random Fourier features...'). Please deduplicate.","section":"References"},{"comment":"The definition p(ω)=2^{-∥ω∥0}Tr(ρ0Qω)^2/ν_f is written in a form that assumes the uniform input distribution; this is fine, but it should explicitly state that this is the specialization of Eq. (B7) to D = Unif[−π,π]^d.","section":"SI D, Eq. (D4)"}],"recommendation":"major_revision","confidential_remarks":"The main reason for major revision is not the internal logic of the proofs—which is generally careful—but the mismatch between the constant-accuracy interpretation advertised in the main text and the relative-accuracy condition under which the lower bound is valid. The paper would be materially strengthened if the authors either (a) prove a lower bound for a superlog-DSE family with constant ν_f, or (b) explicitly reframe the impossibility claims as relative-accuracy results and add a discussion of the zero-predictor triviality that arises at fixed ε for their constructed family. The phase diagram issue and the DSE-scaling gap in Theorem 3 are additional claims that need to be made precise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious paper. The DSE measure is new, the matching sample-complexity bounds in terms of it are new, and the quantum mode-sampling subroutine (three-angle grid with exact quadrature, Lemma 8) plus the BQP-conditional separation (Theorem 3) are genuine advances over the Hamming-weight heuristics of Refs. [60, 61]. I read the SI fairly carefully, and the core logic is coherent: the upper bound uses DSE to control the truncation tail, the lower bound builds a hard family with uniform mode distribution and applies Fano, and the reduction in Lemma 16 checks out. The paper is not cargo-cult theory; it is a real step toward a resource-theoretic account of classical learnability.\n\nNow the soft spots, in proportion.\n\n1. The stress-test note lands. The formal lower bound (Lemma 1 / Theorem 5, SI E) requires ε ≤ ν_f/4, and the constructed hard family has ν_f = 2^{-2⌊M_λ⌋}. So for any fixed target accuracy ε > 0 and sufficiently large M_λ, the condition fails, and the zero predictor already achieves error < ε. The main-text claim that superlog-DSE families are not sample-efficiently learnable is therefore not established for constant-accuracy learning; it holds only for relative accuracy. The authors do include a remark and the inequality is internally consistent, but the headline interpretation overshoots what is proved. This is addressable—state the relative-error regime honestly or construct a family with ν_f not exponentially small—but it should be fixed before publication.\n\n2. The phase diagram's forbidden upper triangle is justified by worst-case bounds: max_U DSE ≲ d ≲ max_U OSE, where the maxima may be over different circuits. The diagram is per-instance. The numerical check on 446 instances of one family is suggestive, not a proof. This is an overclaim in presentation, not an error in the central DSE-learnability theorems, but it should be corrected or explicitly recast as conjectural.\n\n3. No code or data is released, and the numerics use ideal mode selection rather than the proposed quantum subroutine. That limits reproducibility claims, though it does not affect the theory.\n\n4. The polynomial-sample claims implicitly assume constant target accuracy ε, since ε appears in the exponent. Worth saying explicitly.\n\nWho is this for? Researchers in quantum learning theory, classical surrogates, and resource-theoretic simulation. It deserves a serious referee. My recommendation: send it to peer review, with instructions to stress-test the relative-error issue and the OSE comparison. Conditional accept after revision is a plausible outcome.","headline":"Solid resource-theoretic contribution; the DSE learnability bounds mostly hold, but the headline superlog-DSE intractability claim is only proved for relative error, and the phase-diagram forbidden region rests on a worst-case comparison—both fixable, but they need referee scrutiny.","tokens_in":63547,"tokens_out":1625,"would_cite":true,"duration_ms":17634,"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 classical learnability of quantum circuits is governed by a single entropy measure, the dynamical stabilizer entropy (DSE), which the paper defines, bounds, and makes operational.","keywords":["dynamical stabilizer entropy","quantum resource theory","classical learnability","classical surrogates","sample complexity","operator stabilizer entropy","parameterized quantum circuits","quantum-classical separation"],"falsifier":"Find one circuit instance in the paper's allowed class with DSE > OSE at the same x and observable; that single counterexample would break the claimed forbidden region. Alternatively, exhibit a description-only randomized classical poly-time algorithm that emulates the constructed low-DSE BQP-based family, which would refute the computational separation (and imply BQP⊆BPP).","tokens_in":62471,"feed_emoji":"⚛️","tokens_out":4655,"duration_ms":38797,"temperature":0.7,"pith_summary":"This paper tries to establish that classical learnability, not just simulability, of parametrized quantum circuits has a sharp resource-theoretic characterization. The resource is a new quantity, the dynamical stabilizer entropy (DSE), which measures how broadly a circuit's expectation-value function spreads over the exponential set of trigonometric frequency modes generated by its tunable rotation gates. The paper proves matching upper and lower sample-complexity bounds that are exponential in DSE: families with logarithmic DSE are efficiently learnable by classical surrogates trained on quantum data, and families with superlogarithmic DSE are not. It also constructs a DSE-guided surrogate that achieves these bounds and proves, under a standard complexity assumption, that no description-only classical simulator can match it. If right, the work explains why quantum-data-assisted learning can remain efficient where classical simulation fails, and it gives a practical criterion for when a one-time training cost replaces repeated quantum queries.","feed_headline":"Classical learning of quantum circuits splits at one entropy threshold","feed_subtitle":"The dynamical stabilizer entropy predicts when a classical model trained on quantum data beats direct simulation—and when it cannot.","key_machinery":"The load-bearing object is the trigonometric (Pauli-transfer-matrix) expansion f(x)=Σ_ω Φ_ω(x) Tr(ρ0 Q_ω), with modes ω∈{−1,0,1}^d. DSE is the maximum Rényi entropy of the induced mode distribution p(ω)∝ E_x Φ_ω(x)^2 Tr(ρ0 Q_ω)^2 over all fixed-gate conjugations; it quantifies how concentrated the function is on a few dominant frequencies. The machinery does three jobs: a Markov-inequality argument converts small DSE into a small tail probability, so truncating to modes with p(ω)≥τ costs little; the quantum subroutine prepares a state with amplitudes √p(ω) and samples the dominant modes; and the surrogate combines those modes into a partial trigonometric kernel, with sample complexity govern","core_discovery":"The central claim is that DSE is the resource that determines how many quantum-generated training samples a classical model needs to predict the expectation values of a tunable circuit family. Each such function is expanded over 3^d frequency modes; DSE is the maximum order-α Rényi entropy of the normalized squared mode weights, maximized over conjugation by fixed gates, and it acts as a resource monotone. Lemma 1 shows the sample complexity is both lower- and upper-bounded by quantities exponential in the DSE bound M_λ, so M_λ = O(log N) yields polynomial sample complexity and superlogarithmic M_λ yields intractability. Theorem 2 provides a DSE-guided surrogate—a quantum subroutine samples","pith_inferences":["Editorial inference: DSE could be adapted to other tunable gate sets and non-uniform input distributions; the same Fourier-truncation mechanism would likely yield analogous resource measures.","Editorial inference: real devices suffer noise, and noise tends to damp high-frequency modes; this could lower effective DSE and make surrogates more robust—a testable prediction.","Editorial inference: the phase diagram's 'forbidden' region rests on a worst-case relation; proving a per-instance inequality would remove the main structural gap left open."],"forward_implications":["If DSE is O(log N), polynomially many quantum-labeled samples and a polynomial-size classical model suffice to predict expectation values across the whole circuit family.","If DSE grows superlogarithmically, any classical learner requires exponentially many samples, so the boundary is genuine and not an artifact of a specific algorithm.","Low-DSE families can be efficiently learned even when individual instances are high-magic and highly entangled, placing them beyond tensor-network and Pauli-path simulators.","Under BQP ⊄ BPP, no description-only classical algorithm can match the surrogate on these low-DSE families, so the separation between learning and simulation is computational, not merely statistical.","For families evaluated many times, the surrogate converts repeated quantum queries into a one-time training cost, which the paper argues benefits Hamiltonian simulation, variational algorithms, and certification."],"fun_headline_variants":["Entropy threshold decides when quantum data beats simulation","Learnable yet not simulable: entropy draws the line","One quantum entropy measure controls classical learnability","Circuits learnable via quantum data if entropy stays low","Dynamical stabilizer entropy sets the learning–simulation split"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The phase diagram's infeasible upper triangle assumes that DSE is pointwise no larger than operator stabilizer entropy for the same circuit; the theorem proves this only for worst cases over possibly different circuits, with per-instance support limited to a numerical check on one structured family.","fun_headline_variants_meta":{"raw":{"variants":["Entropy threshold decides when quantum data beats simulation","Learnable yet not simulable: entropy draws the line","One quantum entropy measure controls classical learnability","Circuits learnable via quantum data if entropy stays low","Dynamical stabilizer entropy sets the learning–simulation split"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1285,"prompt_tokens":812,"completion_tokens":473,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":409}},"tokens_in":556,"tokens_out":473,"duration_ms":5011,"temperature":1.0,"reasoning_tokens":409,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:17:45.175128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find one circuit instance in the paper's allowed class with DSE > OSE at the same x and observable; that single counterexample would break the claimed forbidden region. Alternatively, exhibit a description-only randomized classical poly-time algorithm that emulates the constructed low-DSE BQP-based family, which would refute the computational separation (and imply BQP⊆BPP).","supporting_citations":[],"review_version":1}