{"id":"0159b3bf-4cbb-484e-b07a-0de69312e691","arxiv_id":"2507.21036","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A Hong-Ou-Mandel interferometer fed with a mixed single-photon state is proposed as a shallow neural network with constant photon cost per inference.","lead":"This paper proposes an optical setup in which two single photons meet at a beam splitter, and the rate of simultaneous clicks computes the output of a shallow neural network whose hidden layer is stored as a mixture of photon states. The authors claim the photon cost per classification is constant, independent of input size and neuron count, and support this with image classification simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The family in Eq. (5) is PSD quadratic forms, so Hornik-Stinchcombe-White universality does not apply; the central claim of a universal approximator is unsupported.","rationale":"The strongest claim and the weakest assumption identified by the reader coincide: the proof of universality is the load-bearing step. The HOM identity in Appendix A is derived carefully and I see no inconsistency there; the mixture formula Eq. (A17) follows from linearity of trace. But the step from 'this has the algebraic form of a one-hidden-layer network' to 'therefore it is a universal approximator' is where the argument fails. The constraints w_i≥0, Σw_i=1, ||W_i||=1, no hidden biases are not harmless: they restrict the model to even, homogeneous PSD quadratic forms. No matter how many neurons M are used, the closure of the family is the cone of PSD quadratic forms, which is closed and not dense. This is not a matter of the approximation being hard; it is a structural impossibility, as the odd linear function example shows. The paper's own conclusion lists a rigorous characterization as future work, which is an explicit admission that the advertised theorem is not proved. I therefore agree with the reader's REJECT verdict. The constant-resource claim is also overstated because state preparation of N-feature input and M hidden configurations is not counted, but the expressivity failure is the more fundamental problem: even with infinite photons and zero post-processing cost, the device cannot approximate general continuous functions. The numerical results are not evidence against this concern; they show only that three binary benchmarks fall inside the PSD quadratic class. A useful remedy would be to either drop the universality claim and reframe the contribution as a constant-photon quadratic classifier, or add hidden biases and sign-indefinite output weights and prove the corresponding density result.","tokens_in":17259,"tokens_out":5831,"duration_ms":70035,"concrete_test":"Analytic falsification: take K=[-1,1]^N and target f(x)=x_1. For every θ, F_θ(x)=σ(x^T A x+β) is even, so at x=e_1 and x=-e_1 the model outputs are equal while the targets are 1 and -1; hence sup_{x∈K}|F_θ(x)-f(x)| ≥ 1 for all θ. This proves the family is not dense in C(K) and directly contradicts the universality claim. A corroborating computational check is to use the released GitHub code to regress the model on f(x)=x_1; the best achievable test error on a symmetric test set should remain near the trivial value, matching the analytic bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised central claim is that the coincidence rate f_wW(I)=Σ_i w_i |⟨I,W_i⟩|^2 (Eq. (5), Section II) is a classical shallow network of M neurons and hence a universal approximator by the Hornik-Stinchcombe-White theorem. This inference is invalid. For the pixel encoding used in Section II and the numerics, f_wW(x)=Σ_i w_i |W_i·x|^2 = x^T A x with A=Σ_i w_i W_i W_i^T, which is positive semidefinite because w_i≥0. Hence every model output F_θ(x)=σ(f_wW(x)+β) is even in x and has a PSD quadratic pre-activation. The representable family is therefore contained in the finite-dimensional set of PSD quadratic forms (dimension N(N+1)/2) regardless of M; it is not dense in C(K). For example, on K=[-1,1]^N the linear function f(x)=x_1 cannot be approximated below uniform error 1, since g(x)=g(-x) for every model output. HSW universality requires hidden biases and cannot be invoked for square-modulus activations with positive normalized output weights. The paper itself defers a rigorous characterization to future work in the Conclusions, confirming the gap. The correct HOM derivation and the numerical demonstrations of three binary classification tasks do not repair this: they establish that PSD quadratic forms can classify those datasets, not that the model is a universal approximator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Hong-Ou-Mandel interferometer in which one single-photon state encodes an N-feature input and a mixed single-photon state encodes M hidden neurons with probabilities w_i. The authors derive, in Appendix A, that the coincidence probability yields f_wW(I)=Σ_i w_i |⟨I,W_i⟩|^2, which they identify with a classical shallow network and, invoking Hornik-Stinchcombe-White, claim is a universal approximator with O(ε^{-2} log(2/δ)) photons per inference independent of N and M. They support this with constrained gradient-descent training and numerical classification experiments on MNIST, Fashion-MNIST, and CIFAR-10, and conclude that the protocol gives a superexponential speedup over classical shallow networks.","tokens_in":17570,"tokens_out":6150,"duration_ms":73895,"significance":"The paper has genuine strengths: Appendix A gives a self-contained derivation of the coincidence rate from the beam-splitter unitary, Appendix B provides explicit training gradients, Appendix C lays out a sample-complexity argument, the code is openly available, and the numerical comparison with constrained classical networks is a useful sanity check. If the central claims were correct, a constant-photon classifier whose output is a trainable function of the input would be a significant contribution to quantum optical inference. However, the central universal-approximation claim is not merely unproved but false as stated: the constrained family is finite-dimensional and even, so the Hornik-Stinchcombe-White theorem does not apply. The resource claim also omits the cost of encoding the input and hidden parameters. As it stands, the contribution is a constant-shot estimator of a PSD quadratic form, not a universal approximator.","major_comments":[{"comment":"The identification of f_wW(I)=Σ_i w_i |⟨I,W_i⟩|^2 with a universal approximator following Hornik-Stinchcombe-White is invalid for the constrained family. Under the SLM discretization used in Section II, f_wW(x)=Σ_i w_i |Σ_j (W_i)_j x_j|^2 = x^T A x with A=Σ_i w_i W_i W_i^T, a positive semidefinite matrix because w_i≥0. Thus every representable pre-activation is a PSD quadratic form, the model output F_θ(x)=σ(f_wW(x)+β) is even under x→−x, and the representable functions lie in a finite-dimensional subset of C(K) (dimension N(N+1)/2), so they cannot be dense. For instance, on K=[-1,1]^N the linear function x_1 is at uniform distance at least 1 from every model output. Hornik-Stinchcombe-White requires hidden biases and unconstrained output weights and therefore does not apply. The Conclusions explicitly defer a rigorous characterization, so the central universal-approximation claim is unsupported.","section":"Section II, Eq. (5); Appendix A"},{"comment":"The claimed constant resource cost counts only the number of photons needed to estimate p(1_a∩1_b) and the post-processing operations. It does not include the resources needed to encode the input: for a digital image encoded on an N-pixel SLM (Section II and Section III), writing the N pixel amplitudes requires O(N) resources, and preparing the hidden mixture with M components of N parameters requires O(MN) parameters and operations. Hence the abstract's statement that 'once trained, our model requires constant optical resources regardless of the number of input features and neurons' is not supported by the analysis as stated. If the claim is only that the number of measurement shots is O(ε^{-2} log(2/δ)) independent of N and M, this should be stated explicitly and the comparison with the classical O(MN) inference cost adjusted accordingly.","section":"Section II, 'constant resources'; Section C"},{"comment":"The numerical benchmarks show that the constrained model achieves good accuracy on three binary tasks and that accuracy increases with M, but this does not provide evidence for universal approximation. Since the model family is contained in the set of PSD quadratic forms (see the first comment), any dataset that is separable by such forms can show this scaling; the experiments do not address approximation of arbitrary continuous functions or of functions outside the even/PSD class.","section":"Section III, Table I and Fig. 3"}],"minor_comments":[{"comment":"The Hoeffding inequality in Theorem 1 is stated for random variables bounded in [0,1], but it is applied to coincidence outcomes bounded in [0,1/2]. The correct bound is δ ≤ 2 exp(-8 n ε²), not 2 exp(-2 n ε²); the asymptotic O(ε^{-2} log(2/δ)) scaling is unchanged.","section":"Section C, Eqs. (C3)-(C5)"},{"comment":"The text says that Eq. (5) is 'precisely the response function of a classical shallow neural network' and that 'we obtain Eq. (1)' for a square absolute value activation with no biases, but Eq. (1) as written includes hidden biases β_i and a general activation g. Since the model has no hidden biases, the claimed identification with Eq. (1) is misleading and should be stated as a constrained specialization.","section":"Section II, Eq. (5)"},{"comment":"The hyperparameter C is introduced as accounting for optical losses, but in the subsequent definitions f(I)=C-2p(1_a∩1_b), C is fully determined by C=∥I∥²Tr[ρ_Uλ]. Please clarify whether C is a fixed physical normalization or a trainable/free parameter, since the text is ambiguous.","section":"Appendix A, Eq. (A11)"}],"recommendation":"reject","confidential_remarks":"The central universal-approximation claim is load-bearing and, as shown in the report, cannot be repaired within the manuscript's scope because the representable family is finite-dimensional and even. The paper may have value as a constant-shot estimator of PSD quadratic forms with numerical evidence on specific datasets; a substantially reframed resubmission along those lines could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The Hong-Ou-Mandel derivation in Appendix A is correct, and the mixture extension to M hidden neurons is a real, natural step beyond the group's single-neuron classifier [38]. The coherent-superposition negative result is also a genuinely useful comparison. The catch: the central claim that this apparatus is a universal approximator is not supported. The representable response functions are PSD quadratic forms, a finite-dimensional set, so the Hornik-Stinchcombe-White theorem cannot be invoked. The paper itself defers the proof to future work, and that gap is load-bearing because the abstract and conclusions lean on it.\n\nWhat the paper does well: the Wick's theorem derivation is self-contained, the O(epsilon^-2) sample-complexity bound via Hoeffding is correctly applied to the mixture estimation, and the training gradients in Appendix B are explicit. The numerical results show the constrained quadratic model reaches accuracy comparable to unconstrained classical networks on the three binary tasks, and the code is on GitHub. That is reproducible evidence.\n\nThe soft spots, in proportion. The universality claim is the main one. For the pixel encoding f_wW(x) = sum_i w_i |W_i dot x|^2 = x^T A x with A PSD, every output is even in x, and the function class has dimension N(N+1)/2 regardless of M. It cannot approximate a linear function like x_1 on [-1,1]^N to uniform error below 1. HSW requires hidden biases and an appropriate activation; square modulus with positive normalized output weights and no biases is too constrained. The 'superexponential speedup' framing also overreaches: it counts photons but not the N-mode state preparation on both sides, and the classical baseline for one inference is O(MN) operations. The constant-photon claim itself is fine; the speedup comparison is inflated.\n\nA minor note: the speedup framing cites the group's prior work [38], but the derivation here does not depend on it, so that is not a real problem.\n\nBottom line: a solid optical identity and a legitimate empirical study of a constrained classifier, but the advertised universality is unjustified. The intended reader is someone working on photon-starved inference or HOM-based classifiers. It deserves a serious referee; the right outcome is major revision. My recommendation: engage with it, but send it back for a rewrite of the claims or a real approximation proof for the PSD quadratic family.","headline":"Correct HOM identity and a useful M-neuron extension, but the universality claim is not supported—the model is a PSD quadratic form, not a universal approximator.","tokens_in":18083,"tokens_out":3561,"would_cite":false,"duration_ms":37396,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a Hong-Ou-Mandel interferometer evaluates a shallow neural network's hidden layer with a constant number of photons, independent of input size and neuron count.","keywords":["quantum optical neural network","Hong-Ou-Mandel effect","universal approximation","shallow neural network","single-photon interference","image classification","photon resource scaling","quantum machine learning"],"falsifier":"Use the paper's own training setup to fit the indefinite quadratic target $f(x,y)=x^2-y^2$ on the unit square with increasing $M$: if the best test error does not go to zero, the constrained family is not a universal approximator; separately, an experiment comparing measured coincidence rates with Eq. (5) for prepared input and mixture states would test the optical identity itself.","tokens_in":17018,"feed_emoji":"⚛️","tokens_out":6971,"duration_ms":78949,"temperature":0.7,"pith_summary":"This paper aims to show that a single Hong-Ou-Mandel interferometer, using two single photons, one beam splitter, and two bucket detectors, can evaluate the entire hidden layer of a classical shallow neural network. The output is read from the coincidence rate: for an input state $|I\\rangle$ and a mixture of $M$ single-photon states with weights $w_i$, that rate encodes $f_{wW}(I)=\\sum_i w_i |\\langle I,W_{\\lambda_i}\\rangle|^2$, which the authors identify with a one-hidden-layer network using square-modulus activation, no biases, and positivity/normalization constraints. Once trained, inference costs $O(\\varepsilon^{-2}\\log(1/\\delta))$ photons independent of input size $N$ and neuron count $M$, which the paper calls a superexponential speedup over the classical $O(MN)$ cost. Numerical simulations on MNIST, Fashion-MNIST, and CIFAR-10 show accuracy that improves with $M$ and matches an unconstrained classical network. The strongest intended consequence is that a fixed optical apparatus can classify images with constant optical resources per query.","feed_headline":"Hong-Ou-Mandel interferometer runs a neural net with constant photons","feed_subtitle":"A single coincidence-rate measurement computes all M hidden neurons; photons stay flat as image size and neuron count grow.","key_machinery":"The load-bearing object is Eq. (5), the Hong-Ou-Mandel coincidence identity $f_{wW}(I)=\\sum_i w_i |\\langle I,W_{\\lambda_i}\\rangle|^2$. The hidden layer of $M$ neurons is encoded as a density operator: a convex mixture of $M$ single-photon states, each carrying $N$ trainable spectral amplitudes, while the input features are encoded in another single-photon state. The beam splitter turns photon indistinguishability into the nonlinear square-modulus operation, and bucket-detector coincidence counting reads out the weighted sum. This object does the work of collapsing an $O(MN)$ matrix-vector computation into a single optical measurement, with the sample-complexity bound $O(\\varepsilon^{-2}\\log(2/\\delta))$ from Hoeffding's inequality supplying the constant number of photons.","core_discovery":"On the paper's own terms, the central claim is that the Hong-Ou-Mandel coincidence probability is not merely an overlap measurement but a neural-network evaluation. Feeding a pure single-photon input and a mixed single-photon hidden state $\\rho_U = \\sum_i w_i |W_{\\lambda_i}\\rangle\\langle W_{\\lambda_i}|$ through a balanced beam splitter gives $p(1_a\\cap 1_b)=[1-f_{wW}(I)]/2$, so measuring the coincidence rate is equivalent to computing the hidden layer of a shallow network. The model is Eq. (5): $f_{wW}(I)=\\sum_i w_i |\\langle I,W_{\\lambda_i}\\rangle|^2$, with each $|W_{\\lambda_i}\\rangle$ an $N$-parameter neuron and $w_i$ positive output weights summing to one. The authors treat this as a universal approximator in the Hornik-Stinchcombe-White sense, with square-modulus activation and no biases, subject to $L^2$ normalization of the hidden rows and $L^1$ positivity of the output weights. They additionally show that the alternative coherent superposition encoding collapses to a single quadratic neuron and does not gain accuracy with $M$.","pith_inferences":["A direct test of the universality claim would be to train the constrained family to approximate an indefinite quadratic form such as $x^2-y^2$ on the unit square; if the error has a positive floor as $M$ grows, the constrained square-modulus family is not dense, contradicting the asserted universality.","The constant-resource guarantee should carry over to other scalar losses of $f_{wW}(I)$, so the same interferometer could serve as an optical kernel evaluator for kernel methods, not only for binary classification.","If the universality gap is real, the practical claim narrows from approximating any continuous function to approximating the learnable positive-semidefinite quadratic family, which may still suffice for the low-photon binary imaging tasks the paper targets."],"forward_implications":["A trained network can classify a new image with $O(\\varepsilon^{-2}\\log(2/\\delta))$ photons, independent of how many pixels or neurons the network has.","The numerical benchmarks show accuracy increasing with $M$: CIFAR-10 planes-versus-dogs rises from 71.2% at $M=2$ to 78.7% at $M=256$, close to the unconstrained classical network's 74.7-78.4% range.","The coherent superposition encoding does not scale with $M$; it is equivalent to a single quadratic neuron, so mixtures are the essential ingredient for expressivity.","Photon loss can be absorbed as a renormalization hyperparameter $C=\\|I\\|^2 \\operatorname{Tr}[\\rho_U]$, and a bias plus sigmoid in post-processing completes the classifier.","Because the mixture can be generated by sampling pure states agnostically, training can be done classically and only the final inference needs the optical apparatus."],"supporting_citations":[{"why":"Supplies the Hornik-Stinchcombe-White universality theorem that Eq. (5) is claimed to satisfy.","marker":"[3]"},{"why":"Provides the Hong-Ou-Mandel effect whose coincidence rate is the network output.","marker":"[35]"},{"why":"Prior single-neuron quantum optical classifier protocol that this paper extends to $M$ neurons.","marker":"[38]"},{"why":"Gives the Hoeffding bound used for the $O(\\varepsilon^{-2}\\log(2/\\delta))$ photon sample complexity.","marker":"[41]"},{"why":"MNIST dataset used for the numerical accuracy benchmarks.","marker":"[44]"},{"why":"Fashion-MNIST dataset used for the numerical accuracy benchmarks.","marker":"[45]"},{"why":"CIFAR-10 dataset used to show accuracy scaling with the number of hidden neurons.","marker":"[46]"}],"fun_headline_variants":["Hong-Ou-Mandel effect runs a shallow network with constant photons","Single coincidence measurement evaluates all hidden neurons","Constant optical resources for any shallow network size","HOM interference computes every neuron in one shot","Quantum optical shallow net uses constant photons at any scale"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the constrained family of networks, with positive output weights summing to one, unit-norm hidden rows, square-modulus activation, and no hidden biases, still satisfies the Hornik-Stinchcombe-White universality theorem, a proof of which the paper defers to future work.","fun_headline_variants_meta":{"raw":{"variants":["Hong-Ou-Mandel effect runs a shallow network with constant photons","Single coincidence measurement evaluates all hidden neurons","Constant optical resources for any shallow network size","HOM interference computes every neuron in one shot","Quantum optical shallow net uses constant photons at any scale"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4285,"prompt_tokens":908,"completion_tokens":3377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":3304}},"tokens_in":524,"tokens_out":3377,"duration_ms":25092,"temperature":1.0,"reasoning_tokens":3304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:02:42.634433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the paper's own training setup to fit the indefinite quadratic target $f(x,y)=x^2-y^2$ on the unit square with increasing $M$: if the best test error does not go to zero, the constrained family is not a universal approximator; separately, an experiment comparing measured coincidence rates with Eq. (5) for prepared input and mixture states would test the optical identity itself.","supporting_citations":[{"cited_title":"Hornik, M","cited_arxiv_id":null,"evidence_quote":"Supplies the Hornik-Stinchcombe-White universality theorem that Eq. (5) is claimed to satisfy."},{"cited_title":"Hong, Z.-Y","cited_arxiv_id":null,"evidence_quote":"Provides the Hong-Ou-Mandel effect whose coincidence rate is the network output."},{"cited_title":"Hoeffding, Probability inequalities for sums of bounded random variables, J","cited_arxiv_id":null,"evidence_quote":"Gives the Hoeffding bound used for the $O(\\varepsilon^{-2}\\log(2/\\delta))$ photon sample complexity."},{"cited_title":"LeCun, C","cited_arxiv_id":null,"evidence_quote":"MNIST dataset used for the numerical accuracy benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fashion-MNIST dataset used for the numerical accuracy benchmarks."},{"cited_title":"Krizhevsky,https://www.tensorflow.org/ datasets/catalog/cifar10","cited_arxiv_id":null,"evidence_quote":"CIFAR-10 dataset used to show accuracy scaling with the number of hidden neurons."}],"review_version":1}