{"id":"6545c3ad-c4ba-4818-87d7-04c58ca97b21","arxiv_id":"2608.05405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single re-uploading quantum circuit, with depth as the chaos expansion order, learns 8-cell calorimeter shower images and certifies that its cross-cell dependence is generated by entangling gates.","lead":"This paper introduces a quantum generative model in which a single re-uploading circuit maps random inputs directly to calorimeter shower images, with no classical network after the circuit. It reports a noiseless fit to Geant4 data, a hardware execution on an IBM processor, and a proposed certification that the learned correlations come from entanglement.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3 assumes shot noise is independent across cells, but the single-setting readout estimates all <Z_i> from the same shots on entangled states; correlated shot errors can inflate correlations, so the hardware 'floor' and deconvolution are unsupported.","rationale":"The paper's central claim is that a single re-uploading circuit with 154 trained angles generates 8-cell showers whose dependence matches Geant4, with the dependence provably produced by entangling gates, and that the ibm fez execution transports this dependence with D=0.873 as a lower bound. The simulation-side attribution (Prop. 1, strip test, CHSH capability) is carefully controlled and internally consistent. The hardware floor, however, is the load-bearing link between the simulation claim and the deployed demonstration. The reader's weakest assumption flags unmodeled coherent error; my stress-test finds a more direct problem: Proposition 3's assertion that shot noise is independent across cells is false whenever the state is entangled, which is exactly the regime the model is designed to occupy. Because the protocol reads all cells from the same shots, the finite-sample estimation errors inherit the state's cross-cell covariance; such correlated noise can inflate rank correlations relative to the noiseless values. The measured D=0.873 therefore cannot be read as a floor, and the deconvolution attributing only 0.02 of mean |\\rho| to genuine device error is not justified. This does not overturn the central conceptual contribution — the simulation results, the verification protocol, and the CHSH separation stand — but it does mean the hardware demonstration's quantitative claims need revision. The p=0.24/statistically-distinguishable contradiction in Sec. V C and the abstract's overstatement of Prop. 4 are real but secondary. The reader's CONDITIONAL verdict remains appropriate, with the additional condition that Prop. 3 be corrected or its use in the hardware interpretation removed.","tokens_in":26626,"tokens_out":10072,"duration_ms":92624,"concrete_test":"Simulate the trained checkpoint at the deployed shot count: for each of 20000 fresh germs, draw S=1152 bitstrings from the exact state |psi(eps,eps0)>, compute the eight \\hat m_i, then the Spearman matrix and D. Compare with (a) exact noiseless values and (b) the Prop. 3 prediction c_i c_j applied pairwise. If the attenuation factor for mean |\\rho| differs from 0.986 by more than ~0.001, or if any pairwise rho increases under shot noise, Prop. 3's independence assumption is falsified. Additionally compute the empirical covariance of the estimation errors across cells; nonzero off-diagonal entries directly disprove the 'independent across cells' assertion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 3 (Sec. IV F) states Var(\\hat m_i) = (1-m_i^2)/S for the S-shot estimator, 'independent across germs and cells,' and concludes shot noise can only attenuate cross-cell dependence. This independence is false for the deployed protocol. Generate (Algorithm 1) measures all n qubits in one computational-basis setting, so the estimators \\hat m_i and \\hat m_j are formed from the same S bitstrings. For a fixed germ the outcome vector (Z_1,...,Z_n) has covariance Cov_state, with off-diagonal entries generally nonzero because the circuit's RZZ gates entangle the register. Hence the estimation errors e_i = \\hat m_i - m_i satisfy Cov(e_i,e_j) = (1/S) E_eps[Cov_state(i,j)] \\neq 0. The regression-dilution formula (12) is derived under independent noise; with correlated noise the finite-S correlation is (C_signal + N_ij)/sqrt((Var(m_i)+Var(e_i))(Var(m_j)+Var(e_j))), which can exceed the noiseless correlation when signal and noise covariances share sign. The predicted mean attenuation 0.986, the implied noiseless mean |\\rho_S|=0.5315, and the statement that device error 'cannot inflate' the measured dependence (Sec. VI C/D) all rest on this false premise. The hardware D=0.873 is therefore not established as a floor on the dependence the device genuinely transports; correlated shot noise could raise it above the true noiseless value. This is not a speculative device effect: it follows from the entangled states the model itself prepares.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the quantum polynomial chaos expansion (QPCE), a generative model in which a single re-uploading quantum circuit, driven only by i.i.d. uniform latent germs and a set of trained gate angles, produces continuous calorimeter cell intensities via Pauli-Z expectation values. The authors argue that re-uploading makes the circuit depth equal to the chaos expansion order, that cross-cell dependence is generated only by RZZ entanglers and one shared germ wire, and that the architecture admits built-in verification: with entanglers removed and the shared germ frozen, the cells are provably independent. They train the model on an 8-cell CLIC Geant4 dataset, report D=0.991 and mean absolute Spearman correlation 0.566 in noiseless simulation, deploy the identical circuit on the ibm fez processor with 500 germs at 1152 shots each, and further prove a no-go theorem on asymptotic tail dependence for smooth expectation-value readouts. The paper explicitly disclaims computational quantum advantage and states that the n=8 instance is exactly classically simulable.","tokens_in":26951,"tokens_out":9494,"duration_ms":88236,"significance":"If the central claims hold, the paper is a significant step for fully quantum generative models: it provides a model class in which the circuit is the entire generator, the learned dependence can be attributed to specific circuit elements by measurement rather than by argument, and a hardware deployment is accompanied by a pre-registered germ batch and an error budget. The manuscript is unusually careful in several respects: it validates on fresh germs against a held-out split, checks gradients against the parameter-shift rule to 1e-14, ships code and data, verifies the depth-order identity numerically, and states the classical simulability of the deployed instance without overclaiming. The main caveat, detailed below, is that the hardware floor reading of D relies on an independence assumption for shot noise that is violated by the single-setting readout protocol.","major_comments":[{"comment":"Proposition 3 assumes that the S-shot estimators \\hat m_i and \\hat m_j have errors that are independent across cells, and concludes that shot noise can only attenuate cross-cell dependence. For the deployed protocol in Algorithm 1, all n expectation values are estimated from the same S computational-basis bitstrings, so for a fixed germ the estimation errors satisfy Cov(e_i,e_j) = (1/S) E_eps[Cov_state(Z_i,Z_j)], which is generically nonzero for the entangled states this circuit prepares. The regression-dilution formula (12) is therefore not applicable, and correlated shot noise can inflate rather than only contract rank correlations when the signal and noise covariances share a sign. Consequently the predicted mean attenuation of 0.986, the deconvolved noiseless mean |rho_S|=0.5315, and the statement in Sec. VI D that device error 'cannot inflate' the measured dependence are unsupported. The hardware D=0.873 cannot be read as a lower bound on the dependence the device genuinely transports without either deriving the correct correlated-noise correction or changing the protocol to estimate each cell from independent shots.","section":"Sec. IV F, Eq. (12); Sec. VI C/D"},{"comment":"The claim that the reported dependence statistics survive coherent error because 'rank statistics are insensitive to a smooth germ-independent bias field' is not justified. Proposition 2 covers only strictly monotone per-cell maps, which are rank-invariant; it does not cover the coherent over-rotation of fractional RZZ pulses identified in the reproduction test. A two-qubit coherent error of this kind is not a per-cell map and can change the rank structure in either direction, potentially inflating the measured dependence. Since this is precisely the channel that the reproduction test detects at chi^2/dof=2.42, the argument that D=0.873 is a floor on the transported dependence is not established. A numerical or analytical bound showing that the observed coherent error cannot increase the rank statistics would be needed to restore the floor reading.","section":"Sec. VI D"},{"comment":"The reported confidence intervals are internally inconsistent: D=0.991 is quoted with interval [0.830,0.952], which does not contain the point estimate, and mean |rho_S|=0.566 is quoted with interval [0.552,0.558], which also excludes the point estimate. These appear to be swapped or mistyped bounds and should be corrected, since the reader cannot otherwise assess the precision of the headline simulation results.","section":"Sec. V B"}],"minor_comments":[{"comment":"The permutation test is reported as 'E=0.0137 at p=0.24' while the text says 'the model is statistically distinguishable from Geant4.' With p=0.24 the test does not detect a difference at conventional significance levels; if the intended p-value is 0.024, it should be corrected, and if the intended statement is that the model is not distinguishable, the wording should be changed.","section":"Sec. V C and Sec. IX"},{"comment":"There is a duplicated phrase: 'we report in Sec. IV C we report the measured decomposition they produce' should be reduced to a single 'we report.'","section":"Sec. IV C"},{"comment":"The sentence 'The price is stated because the alternative that keeps the attribution unconditional...' is grammatically incomplete; 'the price is stated' needs a complement, e.g., 'The price is stated explicitly.'","section":"Sec. III C 5"},{"comment":"The sentence beginning 'One bookkeeping rule is mandatory and easy to violate silently, is that...' should be rewritten; the current construction is not grammatical.","section":"Sec. IV C"}],"recommendation":"major_revision","confidential_remarks":"The central simulation-side construction appears sound and the verification protocol is a genuine strength. The main concern is that the hardware interpretation, which is a headline contribution, rests on the false premise that single-setting shot noise is independent across cells. This is not a speculative device effect; it follows from the entangled states the model itself prepares. I would send the manuscript back for a substantive revision of Sec. IV F and Sec. VI, rather than reject, because the noiseless-simulation results and the architectural ideas are valuable and the hardware claims could in principle be repaired by a correct correlated-noise analysis or a modified measurement protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is the most complete fully quantum generative model for physics data I have seen. One circuit, no classical decoder, 154 gate angles as the entire model, and a verification protocol that actually tests where the dependence comes from. The shared germ wire, the entangler-off null, the CHSH certification of the trained circuit, and the tail no-go theorem are all genuinely new relative to the Barthe et al. re-uploading framework. They also ship code and data, and every numerical claim is backed by a serious pipeline audit. Give credit where it is due: this is careful, honest work, and they are explicit that at n=8 the model is classically simulable and they claim no computational advantage. That is the right posture.\n\nThe central simulation results appear sound. The verification protocol is the real contribution: switching off the couplers and freezing the shared wire provably returns independence (Proposition 1), and the measured strip test sits at the sampling floor. The CHSH capability of 2.76 is a nice emergent property with no Bell term in the loss. The no-go theorem (Proposition 4) is correctly stated with its non-degenerate critical point condition, even if the abstract overstates it as \"every smooth generator.\"\n\nThe main problem is the hardware interpretation. Proposition 3 assumes shot noise is independent across cells. But the deployed protocol estimates all eight <Z_i> from the same S bitstrings. The estimators are therefore correlated, because the state is entangled. The regression-dilution formula (12) does not apply, and correlated shot noise can inflate the estimated correlation above the noiseless value, not just attenuate it. So the claim that \"device error cannot inflate the measured dependence\" and the whole reading of D=0.873 as a floor are unsupported. This is not speculative; it follows from the model's own entangled states. The paper even acknowledges coherent error from untwirled fractional RZZ pulses, which is exactly the channel that could produce non-monotone, germ-dependent effects. The authors need to either measure the shot-noise covariance directly, use a protocol that estimates each cell from independent shots, or substantially soften the hardware claims.\n\nMinor: Sec. V C says p=0.24 and \"statistically distinguishable\" in the same sentence. That is wrong. p=0.24 means the difference is within sampling noise. And the abstract's \"every smooth generator\" should be qualified by Proposition 4's assumptions.\n\nWho is this for? Anyone working on quantum generative models, variational quantum algorithms, or fast calorimeter simulation. It deserves a serious referee, but the referee should require a fix to the shot-noise analysis before publication. I would take it, and I would push for revision rather than rejection.","headline":"A genuinely fully quantum generative model with a sound verification protocol, but the hardware floor claim does not survive contact with the actual shot-noise structure.","tokens_in":27533,"tokens_out":2440,"would_cite":true,"duration_ms":24107,"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":"A single re-uploading circuit with 154 trained angles, no fitted classical parameters, generates calorimeter showers matching Geant4's densities and correlations, and proves by measurement that the entangling gates produce those…","keywords":["quantum polynomial chaos expansion","re-uploading circuit","calorimeter shower simulation","generative model","entanglement dependence","energy distance","tail dependence","expectation-value readout"],"falsifier":"Run the entangler-strip test on hardware: set the 42 RZZ phases to zero, freeze the shared germ at its median, and execute the deployed circuit; if the maximum pairwise rank correlation departs from the 0.019–0.025 sampling floor, or if the CHSH capability computed with two measurement bases fails to exceed 2, the claim that the dependence is generated by the entangling gates and that device error cannot inflate it would be directly refuted.","tokens_in":26404,"feed_emoji":"⚛️","tokens_out":8646,"duration_ms":72243,"temperature":0.7,"pith_summary":"This paper claims that one small quantum circuit can be the entire generative model for calorimeter shower images, with random uniform germs as the only input and the gate angles as the only fitted parameters. Because the germs are re-uploaded at every one of the L blocks, the circuit depth becomes the order of a polynomial chaos expansion, so expressivity grows with depth rather than with classical coefficients. Trained on Geant4 shower data, the 8-qubit, 154-angle instance reproduces per-cell marginals and cross-cell dependencies with a rank-dependence effect size D = 0.991, and its verification protocol turns the model's own settings into measurements: with the entanglers off and the shared germ wire frozen, the cells become exactly independent, while the full model's conditional laws violate the CHSH bound at |S| = 2.76. Run unchanged on a superconducting processor, the circuit retains D = 0.873 with a shot-noise contraction predicted in advance. If right, this establishes that a quantum generator can carry the full joint structure of real detector data and can certify, by measurement rather than argument, which of its gates produced the learned correlations.","feed_headline":"One quantum circuit, 154 angles, matches Geant4 shower dependence","feed_subtitle":"Depth becomes expansion order; switching off entanglement provably removes all cross-cell correlations.","key_machinery":"The QPCE interferometric ansatz of Eq. (2): an n-qubit circuit of L blocks in which the same uniform random germs (epsilon, epsilon_0) are re-uploaded at the head of every block as R_y rotations, followed by diagonal layers of trainable R_z and R_ZZ phases separated by trainable R_y walls. Re-uploading makes each measured <Z_i> an exact degree-L trigonometric polynomial in every germ, so circuit depth plays the role of the truncation order of a polynomial chaos expansion. The two designed dependence channels—the R_ZZ couplers on a hardware-native path and the shared germ wire read by every qubit—are what the verification protocol strips away to certify where the correlations come from; the energy distance (equivalently 2 $MMD^{2}$ with the distance kernel) is the single loss whose exact gradients (three state evolutions via the adjoint) drive the fit.","core_discovery":"The central claim is that the QPCE ansatz—a single re-uploading circuit driven solely by uniform random germs—is the entire generative model: the observed intensities are degree-L trigonometric chaos expansions of those germs, with L the circuit depth, and the gate angles are the only fitted parameters. Trained on 2600 Geant4 showers, the deployed 8-qubit, 154-angle instance reproduces the per-cell Wasserstein distances at 2.27×$10^{-3}$ and the rank-correlation structure at mean |rho_S| = 0.566 against 0.576 in the data, with dependence effect size D = 0.991 relative to a provable independence floor. The paper's distinctive result is attribution: setting the 42 RZZ phases to zero and freezing the shared germ wire leaves the cells exactly independent (Proposition 1), releasing the wire isolates the shared-latent channel, and the trained conditional laws violate the CHSH bound at |S| = 2.76, certifying that the entangling gates—not a classical shared input—produce the dependence. Executed unchanged on a superconducting processor, the circuit reaches D = 0.873 with a shot-noise contraction predicted in advance, while Proposition 4 states a no-go: no smooth expectation-value readout of absolutely continuous latents can produce asymptotic tail dependence strictly between 0 and 1, locating the architectural change needed for tail-heavy distributions.","pith_inferences":["The depth-to-order mechanism is not specific to calorimetry: any re-uploading observable in this ansatz class inherits exact finite chaos expansions, so the same 'depth is the truncation order' identity could serve other continuous-output generative tasks, with the entangler-strip test traveling with it.","The measured CHSH capability of 2.76 in a classically simulable model suggests that the classical/quantum distinction here is about the reachable set of conditional laws, not computational hardness; scaling to non-simulable circuits would make this a practical resource, but would also remove the exact reference that the current verification harness relies on.","If the shared-germ-wire design generalizes, the number of shared wires should follow the rank of the data's dependence structure (here essentially one global mode), so the model's resource count would scale with the physical collective modes of the target rather than with the number of cells.","The no-go theorem implies a concrete design rule for the next generation: to capture tail-heavy joint extremes, one should introduce a discrete common latent that makes the extreme-dominating branch comonotone; the paper's engineered branch example (weight 0.15, plateau at lambda_L ≈ 0.22) is a proof of concept that this rule works."],"forward_implications":["Increasing circuit depth L raises the chaos expansion order covered by a fixed set of angles, so the model's expressivity can be tuned without adding any classical parameter.","The entangler-strip test provides a built-in, backend-agnostic control: every trained QPCE model can be verified, by running its own circuit at phi=0, to have zero cross-cell dependence when the entanglers are off—a certification hybrid classical-quantum models cannot supply.","The hardware protocol predicts the shot-noise contraction of every correlation from the trained model alone, so the error budget for a deployment can be set before any device runs.","Proposition 4 and its corollary imply that any smooth expectation-value generator of this family, at any depth and on any graph, has asymptotic tail dependence in {0,1}; reproducing measured intermediate tail asymmetry at finite quantile is possible, but reproducing it asymptotically requires a discrete common latent with comonotone branches."],"supporting_citations":[{"why":"Introduces expectation-value readout as a generative mechanism, which QPCE realizes with the circuit itself as the model.","marker":"[38]"},{"why":"Provides the general analysis of re-uploading circuits as universal generative models with finite chaos expansions, defining the model class QPCE realizes.","marker":"[39]"},{"why":"Gives the Fourier structure of re-uploading encodings, the basis for the claim that depth equals chaos order.","marker":"[41]"},{"why":"Defines the energy distance used as the single training loss.","marker":"[42]"},{"why":"Proves the equivalence of the energy distance to 2 MMD^2 with the distance kernel, grounding the independence floor and effect-size statistic D.","marker":"[44]"},{"why":"Surveys classical polynomial chaos expansions whose separable-basis limitation QPCE is built to avoid.","marker":"[33]"},{"why":"Supplies the 8-cell CLIC calorimeter shower dataset, the sole training and testing data.","marker":"[46]"},{"why":"Describes the heavy-hex lattice into which the deployed 8-qubit path fits with zero SWAPs.","marker":"[47]"}],"fun_headline_variants":["Single quantum circuit simulates calorimeter showers with depth as order","Quantum chaos expansion: one circuit, 154 angles, matches Geant4","Entanglement off provably kills correlations in quantum generative model","No-go for tail dependence in smooth readouts of quantum generators","Circuit's 154 angles fit Geant4; Bell violation certifies dependence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The hardware claims assume that device error beyond shot noise acts as a smooth, germ-independent, per-cell distortion that leaves rank statistics unchanged; if non-monotone, cross-cell, germ-dependent coherent error (e.g., crosstalk from untwirled fractional RZZ pulses) is present, the measured D=0.873 would not be a reliable floor on the dependence the device transports.","fun_headline_variants_meta":{"raw":{"variants":["Single quantum circuit simulates calorimeter showers with depth as order","Quantum chaos expansion: one circuit, 154 angles, matches Geant4","Entanglement off provably kills correlations in quantum generative model","No-go for tail dependence in smooth readouts of quantum generators","Circuit's 154 angles fit Geant4; Bell violation certifies dependence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1725,"prompt_tokens":1085,"completion_tokens":640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":701,"tokens_out":640,"duration_ms":6064,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:46:26.997675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the entangler-strip test on hardware: set the 42 RZZ phases to zero, freeze the shared germ at its median, and execute the deployed circuit; if the maximum pairwise rank correlation departs from the 0.019–0.025 sampling floor, or if the CHSH capability computed with two measurement bases fails to exceed 2, the claim that the dependence is generated by the entangling gates and that device error cannot inflate it would be directly refuted.","supporting_citations":[{"cited_title":"Xiu and G","cited_arxiv_id":null,"evidence_quote":"Introduces expectation-value readout as a generative mechanism, which QPCE realizes with the circuit itself as the model."},{"cited_title":"Xiu,Numerical Methods for Stochastic Computations (Princeton University Press, 2010)","cited_arxiv_id":null,"evidence_quote":"Provides the general analysis of re-uploading circuits as universal generative models with finite chaos expansions, defining the model class QPCE realizes."},{"cited_title":"Schuldet al., Phys","cited_arxiv_id":null,"evidence_quote":"Gives the Fourier structure of re-uploading encodings, the basis for the claim that depth equals chaos order."},{"cited_title":"Grettonet al., J","cited_arxiv_id":null,"evidence_quote":"Defines the energy distance used as the single training loss."},{"cited_title":"Schuld, R","cited_arxiv_id":null,"evidence_quote":"Supplies the 8-cell CLIC calorimeter shower dataset, the sole training and testing data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the heavy-hex lattice into which the deployed 8-qubit path fits with zero SWAPs."}],"review_version":1}