{"id":"bd90827d-2978-41ba-a769-efdde4d6a0b5","arxiv_id":"2502.05264","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Quantum automated learning trains a quantum classifier by imaginary-time-like dissipation, converging to the global minimum of a data-encoded Hamiltonian with a logarithmic generalization bound.","lead":"This paper introduces quantum automated learning, a training method that replaces variational parameters with iterative quantum state preparation. It proves convergence to the global minimum of a natural loss and bounds the generalization error, but the practical cost and the heavy-tail assumption limit the result.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The heavy-tail assumption (Definition S2) is the unproven hinge for constant post-selection success; the paper provides no evidence that the low-energy eigenstate fraction stays constant as n or the target accuracy ε varies.","rationale":"I read the proof of Theorem S2 and the imaginary-time evolution derivation in Supplementary Sec. III A–B carefully. The algebra up to the stated error bounds is internally consistent, and the conditioned-loss estimate is a legitimate averaging over training-sample sequences. The theorem does what it claims under its explicit assumptions. The load-bearing question is therefore whether the heavy-tail assumption (Definition S2) is actually satisfied in the parameter regimes that make QAL practical. The reader identified this same assumption as the weakest point; I agree. My stress test sharpens it in one way: the assumption is not a single qualitative property but a quantitative one, and the paper gives no scaling evidence. To achieve near-optimal loss, ε must be small, and c2(ε) — the fraction of eigenstates within ε of the ground energy — is what controls the constant success probability c4. The paper's numerical spectra are only for n = 10 and do not show how c2(ε) behaves as n grows or as ε approaches 0. Without that, the claim of constant success probability is an unverified scaling assumption, not a proven property. I also note the step-complexity issue in Theorem S2 when g > 0, reinforcing that 'provable convergence' is an existence statement rather than an efficiency guarantee. These concerns do not overturn the paper's internal mathematics, but they justify the CONDITIONAL verdict already given; hence no verdict adjustment is needed.","tokens_in":25746,"tokens_out":23316,"duration_ms":258306,"concrete_test":"Estimate c2(ε) = 2^{-n} Tr(Π_{g+ε}) for the same Fashion-MNIST encoding used in Fig. 2, with n = 10, 12, 14, 16 and ε = 0.1, 0.01, 0.001, using the kernel polynomial method or stochastic Lanczos on H_S (a sum of N projectors, so sparse or low-rank approximations are feasible). If the fraction c2(ε) decays with n or drops sharply as ε shrinks, then the heavy-tail assumption is not robust in the regime claimed and Theorem 2's constant-success guarantee does not scale.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central practical claim of QAL is not just convergence in expectation (Theorem S2), but convergence with constant post-selection success probability (Theorem 2, proven as Theorem S4). Theorem S4 is conditional on Definition S2: the averaged Hamiltonian H_S must have a constant proportion c2 of eigenstates in [g, g+ε]. The paper validates this only for the ten-qubit datasets in Fig. 2d and Figs. S3–S5, and the accompanying heuristic ('similar data ⇒ similar Hamiltonians') is not a proof. The fraction c2(ε) is not shown to remain constant as n grows with the data dimension, nor as ε shrinks toward the accuracy one actually wants. If c2(ε) decays with n, or decays with ε faster than the theorem's required constant behavior, then Theorem 2 fails: the post-selection success probability need not be constant and the resource cost of QAL becomes exponential. A related quantitative gap is that Sec. III B (Theorem S2) requires η ≲ σ_g e^{-2βg}/β; when the ground energy g is positive, this can force an exponentially large number T = β/η of steps, so 'provable convergence' does not by itself imply practical trainability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces quantum automated learning (QAL), a supervised quantum classification protocol without variational parameters. Training data x are encoded into unitaries U(x), and a random state is repeatedly updated by the non-unitary map U(x)^\\dagger M_y U(x), where M_y is a label-dependent perturbation implemented by block encoding and post-selection. In the unnormalized density-matrix formalism the update is (I-\\eta H_x)\\rho(I-\\eta H_x), and averaging over the training set gives an imaginary-time evolution under H_S = E_{x\\sim S} H_x. The central analytic results are: (i) Theorem 1 / Theorem S2, exponential convergence of the averaged conditional loss to the ground energy of H_S provided the initial state has nonzero ground-state overlap; (ii) Theorem 2 / Theorems S3-S4, constant post-selection success probability with near-optimal loss under a 'heavy-tailed Hamiltonian' assumption; and (iii) Theorem 3 / Theorem S5, a generalization gap bound of order \\sqrt{n/N} obtained from matrix Bernstein inequalities. The paper also reports numerical simulations on Fashion MNIST, MNIST, Aubry-Andr\\'e Hamiltonians, and cluster-Ising ground states, together with noise-robustness and state-reusability experiments.","tokens_in":25996,"tokens_out":15543,"duration_ms":155048,"significance":"If the central claims are correct, QAL is a genuinely different training paradigm: it avoids variational parameters and gradient evaluation, and the convergence to the global minimum of the empirical risk is proved rather than assumed. The connection to imaginary time evolution gives a transparent physical picture, and the quadratic form of the loss yields a clean generalization bound. The paper is also careful to state several limitations, such as the open question of universal representation power. The numerical demonstrations on several datasets and the explicit noise-robustness results are valuable. However, the practical significance of the main theorem is substantially weakened by the issues below, especially the gap in the proof of Theorem S3 and the unsupported scaling of the heavy-tail assumption. The paper deserves serious consideration, but it needs major revision before the central practical claims can be accepted.","major_comments":[{"comment":"The equality in Eq. (S27) is algebraically incorrect. With beta chosen as beta = 3 ln(1+c2)/(c3 epsilon), the term ln(1/c2)/(2 beta) equals c3 epsilon ln(1/c2)/(6 ln(1+c2)), which is strictly larger than c3 epsilon/6 for every c2 in (0,1/10). Therefore the displayed bound leading to 'g + epsilon + c3' does not follow from the preceding inequalities. Since Theorem S4 and hence Theorem 2 rely directly on Theorem S3, the proof of the constant-success-probability claim is incomplete. The argument is likely repairable by taking beta proportional to ln(1/c2)/(c3 epsilon) rather than ln(1+c2)/(c3 epsilon), but the theorem statement and its constants need to be reworked.","section":"Supplementary Sec. III C (Theorem S3), Eq. (S27)"},{"comment":"The heavy-tail assumption (Definition S2) is the only bridge between generic convergence in expectation and the constant post-selection success probability claimed in Theorem 2. The assumption is not derived from any data model; the heuristic in the main text about dogs and cats as a mixture of two random projectors is an illustration, not a proof. The numerical support in Fig. 2d and Figs. S3-S5 is restricted to ten-qubit datasets and a fixed accuracy parameter, and no evidence is given that the low-energy eigenstate fraction c2 stays constant as n grows with the data dimension or as epsilon shrinks toward the target accuracy. If c2 decays with n or epsilon, the success probability lower bound in Theorem S3 decays correspondingly and the resource cost becomes exponential. The authors should either prove heavy-tailedness for a concrete class of data-encoding Hamiltonians or explicitly state this as a limitation and provide scaling experiments.","section":"Definition S2 and Theorem S4"},{"comment":"Theorem S2 is existential and the proof gives no bound on the required number of steps T. From the proof conditions e^{-2 beta delta} < sigma_g c/4 and beta eta e^{2 beta g} < sigma_g c/16, one needs eta of order sigma_g e^{-2 beta g}/beta, so T = beta/eta grows at least like e^{2 beta g}/sigma_g. For the generic case g > 0 this is exponential in beta g, and beta itself must be about (1/delta) ln(1/(sigma_g c)) to reach error c. Thus the main-text statement that 'the training process will converge exponentially as T increases' does not by itself imply practical trainability. The paper should report the dependence of T on g, delta, and sigma_g, and clarify whether the claim is convergence for sufficiently large T or efficient trainability in the system size.","section":"Supplementary Sec. III B (Theorem S2)"}],"minor_comments":[{"comment":"The notation '2n+1' should be '2^{n+1}' in the logarithm; as written it does not match the Hilbert-space dimension used in the proof.","section":"Main text Eq. (3) and Supplementary Eq. (S28)"},{"comment":"Theorem 2 says 'with a random initial state in the computational basis', but the proof of Theorem S4 uses the maximally mixed initial state I/2^n. The relation should be clarified: the averaged state over random computational-basis inputs is maximally mixed, but a single random pure state is not covered by the stated proof.","section":"Theorem 2 and Theorem S4"},{"comment":"The caption states that the success probability is calculated analytically from Eq. (2) with the O(T eta^2) term omitted; for eta = 0.1 and the large step counts shown, T eta^2 is not obviously small, so the plotted trade-off should be validated against the exact unnormalized-state simulation.","section":"Fig. 2c"},{"comment":"There are minor typographical issues: 'ploted' in Methods, 'Supplimentary' in Supplementary Sec. IV, and reference [57] duplicates reference [50].","section":"Various"},{"comment":"The statement that QAL 'escapes the barren plateau problem inherently' is too strong without a complexity analysis of the training circuit; the absence of variational parameters removes gradient-vanishing concerns, but the circuit depth and post-selection overhead have not been analyzed as a function of n.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely question, and the overall framework could be a valuable contribution to the quantum machine learning literature. The main obstacles are the algebraic gap in the proof of Theorem S3, on which the constant-success-probability claim depends, and the lack of scaling evidence for the heavy-tail assumption. If the authors can repair the proof and either prove the assumption for a natural data model or substantially temper the practical claims, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe short version: this is a serious, interesting paper, and the core math holds up. QAL is a genuinely new framework—no variational parameters, data encoded into unitaries, training as state preparation via imaginary time evolution. The convergence proof (Theorem 1) is correct under the stated assumptions, the generalization bound is clean, and the imaginary-time interpretation is more than a metaphor: the update rule really does approximate imaginary time evolution. The numerics on multiple datasets are credible, and the spectra they show do look heavy-tailed.\n\nThe soft spots are real but not fatal. The heavy-tail assumption (Definition S2) is the hinge for constant post-selection success, and it is only checked numerically at 10 qubits. The stress-test worry is fair: the fraction c2(ε) could decay with n or as ε shrinks, and the heuristic about similar data producing similar Hamiltonians is plausible but not a proof. If the assumption fails, success probability can decay exponentially. The paper is explicit that this is an assumption, and the numerical evidence is at least consistent, but the scaling behavior is unproven. The bigger practical gap is step complexity: Theorem 1 guarantees convergence but the required T can grow like (1/ε)e^{O(βg)} when the ground energy g is positive, so 'provable convergence' does not obviously mean 'practical trainability.' The paper doesn't quantify this. Also, no code or data is currently available, which limits reproducibility.\n\nI disagree with the idea that the convergence theorem is circular. The loss is defined via H_S and the update uses the same H_x, but that is a coherent optimization problem; the proof is a legitimate imaginary-time analysis with error bounds. The matrix Bernstein bound in the generalization proof is standard and correctly applied.\n\nThe paper is for anyone working on gradient-free quantum machine learning, state preparation, or dissipative learning dynamics. It deserves a serious referee. My recommendation: send it to peer review. The referee should push on the heavy-tail scaling and ask for explicit step-complexity bounds, but the framework is sound enough to warrant publication after revision.","headline":"Serious, mostly rigorous proposal for gradient-free quantum learning; the heavy-tail assumption is the main unproven hinge, and step-complexity could be exponential for positive ground energy.","tokens_in":26508,"tokens_out":2729,"would_cite":true,"duration_ms":25737,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12","68T05"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"This paper introduces quantum automated learning, a gradient-free scheme in which data-encoded unitaries and label-guided perturbations drive a quantum state to the global minimum of the training loss, and proves exponential convergence…","keywords":["quantum automated learning","gradient-free quantum machine learning","imaginary time evolution","trainability","generalization bound","heavy-tailed Hamiltonian","post-selection","quantum state preparation"],"falsifier":"Construct a dataset with no heavy tail—for instance, random labels attached to a generic state-encoding circuit, or a random-matrix-like Hamiltonian whose spectrum concentrates in the middle—and run QAL from a maximally mixed state. If the post-selection success probability decays exponentially with the number of steps while the conditional loss stays bounded away from the ground energy, then Theorem 2's practical claim fails for that dataset; Theorem 1 alone would still hold but with negligible success probability.","tokens_in":25583,"feed_emoji":"⚛️","tokens_out":10459,"duration_ms":96431,"temperature":0.7,"pith_summary":"This paper proposes quantum automated learning (QAL), a way to train a quantum classifier with no variational parameters and no gradient computation. Training is recast as quantum state preparation: a random initial state is repeatedly acted on by unitaries that encode data samples, with a label-dependent perturbation that suppresses wrong predictions. Averaged over samples, the update is exactly imaginary time evolution under the data-averaged Hamiltonian, so the loss—the energy of the state—decays exponentially to its global minimum. The paper proves a generalization bound of order $\\sqrt{\\log D/N}$ and demonstrates on images and quantum many-body data that the protocol reaches near-perfect accuracy with constant post-selection success probability. If correct, this sidesteps the three main scaling obstacles of variational quantum machine learning: local minima, barren plateaus, and costly per-parameter gradient estimation.","feed_headline":"Gradient-free quantum learning provably reaches the global minimum","feed_subtitle":"No parameters, no gradients—just data-encoded unitaries cooling a state to the loss minimum.","key_machinery":"The load-bearing object is the non-unitary update $(I-\\eta H_x)|\\psi\\rangle/\\|(I-\\eta H_x)|\\psi\\rangle\\|$, physically realized by sandwiching a block-encoded label perturbation $M_y$ between $U(x)$ and $U(x)^\\dagger$ and post-selecting on the ancilla. Averaged over the training set, one step becomes imaginary time evolution under $H_S$, so the training trajectory is a cooling process whose fixed point is the ground state—the global minimum of the loss. The heavy-tail assumption (Definition S2) is what turns exponential-in-$\\beta$ convergence into convergence with constant success probability, and the quadratic form of the loss is what yields the logarithmic-dimension generalization bound through matrix concentration.","core_discovery":"The central claim is that a supervised learning task can be solved by preparing the ground state of the data-averaged Hamiltonian $H_S = \\mathbb{E}_{x\\sim S} H_x$, where $H_x = I - U(x)^\\dagger \\Pi_{y(x)} U(x)$. One training step applies $U(x)$, then the perturbation $M_y = |y\\rangle\\langle y| + (1-\\eta)(I - |y\\rangle\\langle y|)$, then $U(x)^\\dagger$, which updates the state as $|\\psi\\rangle \\to (I-\\eta H_x)|\\psi\\rangle / \\|(I-\\eta H_x)|\\psi\\rangle\\|$. At the ensemble level this is $\\rho \\to e^{-\\eta H_S}\\rho e^{-\\eta H_S} + O(\\eta^2)$, i.e. imaginary time evolution. Theorem 1 proves that for any small constant $c$ one can choose the learning rate and number of steps so that the averaged final loss is at most the ground energy $E_g$ plus $c$, with the excess decaying like $e^{-2\\beta\\delta}$. Theorem 2 proves that when $H_S$ has a heavy-tailed spectrum—a constant fraction of eigenstates near $E_g$—the post-selection success probability stays constant while the loss becomes near-optimal. Theorem 3 bounds the generalization gap by $\\sqrt{4\\ln(2^{n+1}/\\delta)/N}$ with probability at least $1-\\delta$.","pith_inferences":["Our inference: the heavy-tail assumption is checkable classically before any quantum run—compute the empirical spectrum of $H_S$ from the training set—so one could filter datasets that are predictably bad for QAL.","Our inference: the imaginary-time-equivalence picture suggests that decoherence acts like a finite-temperature heat bath; QAL's dissipation may therefore be self-correcting under realistic noise, a property the paper demonstrates numerically but does not prove.","Our inference: replacing the label projector $\\Pi_y$ with task-dependent reward operators could extend the same automated-cooling mechanism to reinforcement or unsupervised learning, but convergence in those settings is not established here.","Our inference: since representation power is delegated to the encoding, testing encodings with provably universal feature maps is the natural next step; without such a result QAL's expressivity is only as good as the chosen circuit family."],"forward_implications":["Training a QAL model is provably free of local minima and barren-plateau obstructions, because the loss is quadratic and no variational parameters enter the circuit.","On near-term hardware the protocol needs only shallow data-encoding circuits and about $\\mathcal{O}(\\log k)$-gate perturbation unitaries; the simulations reach about 0.99 accuracy on Fashion MNIST with roughly 290 post-selected runs.","The generalization bound means a training set of size $N = \\Omega(n)$ suffices to control the gap between training and true loss, so sample complexity grows only logarithmically in the Hilbert-space dimension.","The trained state can be reused: gentle label measurements do not destroy it, a few additional training steps recover high accuracy, and shadow tomography can supply the few copies needed for many inference queries.","The same protocol handles classical images, Hamiltonian data, and quantum state data by choosing the appropriate unitary encoding for each data type."],"supporting_citations":[{"why":"Establishes the barren-plateau phenomenon in variational quantum circuits, the scaling obstacle QAL is designed to bypass.","marker":"[37]"},{"why":"Introduces Hamiltonian-echo backpropagation, the self-learning mechanism whose error injection the target-oriented perturbation $M_y$ adapts.","marker":"[54]"},{"why":"Supplies the repeated real-time-evolution protocol used to encode quantum state data into unitaries for QAL experiments.","marker":"[65]"},{"why":"Gives the CNOT-efficient decomposition of multi-controlled single-qubit gates used to compile the perturbation unitary $U_y$.","marker":"supp. [20]"},{"why":"Provides the matrix Bernstein inequality that yields the paper's generalization-gap bound for the quadratic loss.","marker":"supp. [22]"}],"fun_headline_variants":["Quantum auto-learning: gradient-free with provable global minimum","No parameters, no gradients: quantum learning provably converges","Auto-learn quantum: no gradients, provable optimum","Quantum training without gradients: provable global min","Quantum auto-learn: exponential convergence to global min"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical guarantee that training succeeds with constant probability rests on the heavy-tail assumption: the data-averaged Hamiltonian must have a constant fraction of eigenstates with energy close to its ground energy, and the paper only verifies this numerically on the datasets it tests rather than proving it for general datasets.","fun_headline_variants_meta":{"raw":{"variants":["Quantum auto-learning: gradient-free with provable global minimum","No parameters, no gradients: quantum learning provably converges","Auto-learn quantum: no gradients, provable optimum","Quantum training without gradients: provable global min","Quantum auto-learn: exponential convergence to global min"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001451,"raw_usage":{"total_tokens":5932,"prompt_tokens":1120,"completion_tokens":4812,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":736,"completion_tokens_details":{"reasoning_tokens":4736}},"tokens_in":736,"tokens_out":4812,"duration_ms":34848,"temperature":1.0,"reasoning_tokens":4736,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:58:10.437219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a dataset with no heavy tail—for instance, random labels attached to a generic state-encoding circuit, or a random-matrix-like Hamiltonian whose spectrum concentrates in the middle—and run QAL from a maximally mixed state. If the post-selection success probability decays exponentially with the number of steps while the conditional loss stays bounded away from the ground energy, then Theorem 2's practical claim fails for that dataset; Theorem 1 alone would still hold but with negligible success probability.","supporting_citations":[{"cited_title":"L ´opez-Pastor and F","cited_arxiv_id":null,"evidence_quote":"Introduces Hamiltonian-echo backpropagation, the self-learning mechanism whose error injection the target-oriented perturbation $M_y$ adapts."}],"review_version":1}