{"id":"67c83038-567c-48e8-a682-e7c50fbba8f4","arxiv_id":"2505.10331","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Gibbs-weighted vote of random perceptrons has an optimal inverse temperature beta* = pi*sqrt(d-2) for Gaussian teacher-perceptron data, independent of the teacher vector and ensemble size.","lead":"Random neural networks with no training can classify data when their per-classifier votes are exponentially weighted by their individual errors. The authors derive and numerically test a formula for the optimal weighting temperature, which they say depends only on the data dimension.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unconditional CLT in Appendix A ignores the shared test point x; the mixture variance from x is not reduced by n and dominates in the tested regime, so the proof of β*=π√(d-2) does not follow.","rationale":"The reader's identification of the CLT step as the weakest assumption is correct and is the most load-bearing threat to the central claim. The variance decomposition above is concrete and testable. The flaw matters because the paper uses the word 'prove' for a formula that currently rests on an approximation that fails in the parameter regime where the numerical confirmation is presented. Independent support (reproducible code, matching numerics across d and MNIST) means the phenomenon is plausible, so the correct disposition is to keep CONDITIONAL: tighten the derivation or soften 'prove' to 'conjecture supported by simulations.' No verdict change is needed.","tokens_in":15613,"tokens_out":20889,"duration_ms":195706,"concrete_test":"For d=500, n=20000 and β=π√(d-2), estimate m(x) and v(x) by Monte Carlo over w for many draws of x, and compute Var_x[m(x)] and n^{-1}E_x[v(x)]. If Var_x[m(x)] is not negligible relative to n^{-1}E_x[v(x)] (say, >10%), the Appendix A CLT is invalid in the tested regime. Then locate the minimizer of the conditional-CLT loss E_x[Φ(-√n m(x)/√v(x))] over β; if it differs from π√(d-2) by more than 20%, the claimed β* is an artifact of the unconditional approximation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix A obtains β* by declaring n^{-1}ΣY_i normally distributed with mean E_{x,w}Y and variance n^{-1}Var_{x,w}Y, invoking the CLT over the joint distribution of (x,W). But in the ensemble loss x is fixed, so Y_i are only conditionally i.i.d. given x. Writing m(x)=E_w[Y|x] and v(x)=Var_w(Y|x), the exact variance of the average is Var_x[m(x)] + n^{-1}E_x[v(x)]. The first term does not decay with n; in the regime of Figure 4 (d=500, n=20000) it is O(1/d) while the second is O(1/n), so the x-induced variance is roughly n/d ≈ 40 times larger than the sampling variance the CLT uses. The unconditional normal approximation therefore understates the fluctuations of the ensemble average, and the function ξ(β) whose maximizer is claimed to be π√(d-2) is not the correct loss profile. The analytic proof does not go through; the formula remains a numerically supported conjecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a Gibbs-weighted ensemble of random perceptrons, with classifier weights proportional to exp(-β times the 0-1 loss), and claims that for Gaussian teacher-perceptron data the optimal inverse temperature is β* = π√(d-2), independent of the teacher vector, the number of classifiers n, and the dataset size N. The claim is supported by loss-profile experiments in Section 3, a heuristic analytic derivation in Appendix A, and MNIST experiments in Section 4, and a minimal working example is provided.","tokens_in":15848,"tokens_out":20996,"duration_ms":203623,"significance":"The empirical observation of a finite optimal temperature is clean and clearly documented, and the apparent independence of the optimum from the teacher and from n and N is striking. The manuscript ships a minimal working example and the core numerical results appear reproducible. However, the analytic proof in Appendix A relies on an unjustified central-limit step and on uncontrolled asymptotic approximations; as it stands, the closed form β* = π√(d-2) is a numerically supported conjecture rather than a theorem.","major_comments":[{"comment":"The line \"We can therefore study the sum as a normally distributed random variable, thanks to the Central Limit Theorem\" is not justified, because the variables Y_i share the same test point x and are only conditionally i.i.d. given x. Writing m(x) = E_w[Y|x] and v(x) = Var_w(Y|x), the exact variance of n^{-1}Σ_i Y_i is Var_x[m(x)] + n^{-1}E_x[v(x)], whereas the unconditional CLT used in the paper gives n^{-1}Var_{x,w}Y = n^{-1}(Var_x[m(x)] + E_x[v(x)]). The term Var_x[m(x)] does not decay with n, and in the regime of Figure 4 (d=500, n=20000) it is of order 1/d while the sampling term is of order 1/n; the omitted term is numerically dominant. Consequently the derivation of ξ(β) and of β* = π√(d-2) in Eqs. (2)-(3) does not go through as written.","section":"Appendix A, CLT step"},{"comment":"Because of the issue in Appendix A, the statement in the Abstract and Section 3 that the optimal temperature is \"analytically prove[n]\" is not supported. The closed form β* = π√(d-2) should be presented as a numerically verified conjecture unless a correct derivation is supplied; the current proof is load-bearing for the paper's central theoretical contribution.","section":"Abstract and Section 3"},{"comment":"The approximations arcsin(A) ≈ A and 1/||w|| ≈ E[1/||w||] are used without error bounds. The claim following Eq. (4) is supported only by numerical evidence, and these approximations enter the exponential arguments defining μ and σ² in Appendix A. Since the maximizer of ξ(β) depends on those arguments, the closed form β* is not rigorously controlled by the presented analysis.","section":"Appendix C"},{"comment":"The empirical claim that β* is independent of n is checked at only three values (n=100, 5000, 20000) for a single d. Under the correct variance decomposition of the ensemble average, the relative contribution of Var_x[m(x)] versus n^{-1}E_x[v(x)] changes with n, so these three points are too sparse to establish n-independence, especially because the present theory incorrectly predicts n-independence through cancellation of the √n factor in ξ(β).","section":"Section 3, Figure 4b"}],"minor_comments":[{"comment":"There are several typos and grammatical issues: \"does not depend neither\" should be \"does not depend on either,\" \"wether\" should be \"whether,\" and the Figure 3 caption contains \"minimmum\" for \"minimum.\"","section":"Abstract and Section 3"},{"comment":"The notation for the weights is inconsistent: α_w(β) is defined for a single classifier, but later α_W(β) is used as a vector, and the normalization by the partition function Z_n(β) is not made explicit in Eq. (1). Please clarify the notation.","section":"Section 1, Eq. (1)"},{"comment":"The phrase \"an affine transformation x ↦ 2(-x + 1/2)\" is not a well-defined affine map on R^d as written, and the claim that affine transformations cannot alter the number or position of critical points is confusing; the change of variables should be rewritten carefully.","section":"Appendix A"},{"comment":"The identity used for E_{x} sign(w_*^T x x^T w) is attributed to \"Grothendieck's Identity\" via reference [32]; a more specific citation or statement of the identity would help the reader verify the step.","section":"Appendix C"},{"comment":"The axis labels in Figure 4 are minimal, and it is not immediately clear which curve corresponds to which fixed parameter in each panel; please add legend entries or annotate the panels with the fixed values of the other parameters.","section":"Figure 4"},{"comment":"The Python code in Code Snippet 1 contains inconsistent spacing in variable names such as \"t r a i n _ s e e d\" and \"te st _s ee d\"; this is likely a formatting artifact but should be cleaned up for reproducibility.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The numerical experiments are interesting and the formula β* = π√(d-2) appears to match the Gaussian simulations, but the analytical derivation in Appendix A is invalid at a load-bearing point. I recommend major revision rather than rejection: the authors should either supply a correct conditional-CLT derivation or explicitly downgrade the analytic claim. If the corrected derivation yields a different β*, the manuscript's central claim would need to be reassessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth reading for the finding, not for the proof. The observation that a Gibbs-weighted vote over random perceptrons has a finite optimal inverse temperature, with beta* = pi*sqrt(d-2) independent of teacher, n, and N, is new and numerically convincing for Gaussian data. The experiments are careful, the minimal working example is reproducible, and the universality checks (varying d, n, N, teacher) are exactly the right ones.\n\nWhat is genuinely new: I do not know of a prior derivation of an optimal Gibbs temperature for random classifier ensembles. The general exponential-weighting idea is classical (Hedge, PAC-Bayes), but the closed-form dependence on sqrt(d) is specific to this Gaussian setup and worth having.\n\nWhere it gets soft: the analytic derivation in Appendix A does not do what the authors claim. They treat the Y_i as i.i.d. over the joint law of (x,W) and invoke the CLT to get variance (1/n) Var_{x,w} Y. But x is the same test point for every classifier. Conditional on x, the Y_i are i.i.d. with mean m(x) and variance v(x), so the unconditional variance of the average is Var_x[m(x)] + (1/n) E_x[v(x)]. The first term does not shrink with n. In the regime of Figure 4 (d=500, n=20000) it is about 40 times the second. So the normal approximation used to define xi(beta) is not the variance of the actual ensemble average. The derivation of beta* = pi sqrt(d-2) does not follow from the written argument. This is not a minor typo; it is the central step of the \"proof\". The paper also linearizes arcsin and replaces the inverse norm by its expectation with an appeal to concentration; each step may be defensible at large d, but together they are heuristics, not a proof. That is why the abstract's \"we analytically prove\" should be softened to \"we show numerically and support with an asymptotic heuristic\".\n\nThe MNIST experiments are suggestive but thin: one dataset, no error bars, and the universal claim across data distributions is a stretch from one case.\n\nVerdict: the empirical phenomenon and the formula are likely to survive closer scrutiny, and the paper is a useful contribution to the random-features/untrained-networks literature. It deserves a serious referee, but the authors should either make the CLT step rigorous (e.g., conditional CLT plus control of Var_x[m(x)]) or re-frame the analytic part as a conjecture. I would send it to review with the expectation of major revision.","headline":"A clean, novel empirical finding with an overclaimed proof: the optimal Gibbs temperature for random ensembles is real, but the CLT in Appendix A does not establish beta* = pi*sqrt(d-2).","tokens_in":16345,"tokens_out":2625,"would_cite":true,"duration_ms":25450,"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":"A Gibbs-weighted ensemble of random perceptrons has a finite optimal inverse temperature, $\\beta^* = \\pi\\sqrt{d-2}$, which the paper proves is independent of the teacher vector, the number of classifiers, and the dataset size.","keywords":["random neural networks","Gibbs measure","inverse temperature","teacher-student","0-1 loss","central limit theorem","untrained classification","perceptron ensemble"],"falsifier":"Fix $d$ and sample many independent weight matrices $W$ and test points $x$. For a given teacher $w_*$, compute the empirical ensemble 0-1 loss at $\\beta = \\pi\\sqrt{d-2}$ and at, say, $0.5\\,\\pi\\sqrt{d-2}$. If, averaged over realizations, the loss at the predicted $\\beta$ is not below the loss at neighboring values, or if the location of the empirical minimum shifts systematically with $n$ or with the choice of $w_*$, the universality claim fails. A sharper test is to repeat the measurement with two structurally different teachers and check whether the empirical minimizer coincides.","tokens_in":15407,"feed_emoji":"🧠","tokens_out":9494,"duration_ms":86897,"temperature":0.7,"pith_summary":"The paper argues that a collection of randomly initialized perceptrons (single linear threshold classifiers) can perform nontrivial classification if each perceptron's vote is weighted by a Gibbs factor $\\exp(-\\beta \\cdot \\text{loss})$, with $\\beta$ a free inverse temperature. The central claim is that for Gaussian isotropic data labeled by a teacher perceptron, this weighting has a finite optimal temperature, $\\beta^* = \\pi\\sqrt{d-2}$, at which the ensemble's expected misclassification rate is minimized. The paper proves this through an analytic large-$n$, large-$d$ approximation and verifies it numerically, finding that the optimum does not depend on the unknown teacher vector, the number of classifiers, or the dataset size. If true, the result means an untrained random network can be tuned for optimal classification by a dimension-dependent constant alone, and the same finite-temperature structure is observed on the MNIST benchmark.","feed_headline":"An ensemble of random perceptrons votes best at beta = pi sqrt(d-2)","feed_subtitle":"The optimal voting temperature depends only on input dimension, not on the data, the teacher, or the ensemble size.","key_machinery":"The load-bearing object is the Gibbs measure over the random classifiers: each perceptron with weight vector $w_i$ receives a vote weight $\\alpha_i(\\beta) \\propto \\exp(-\\beta L(\\phi(w_i)))$, so the loss acts as an energy and $\\beta$ as an inverse temperature. The argument then replaces the empirical average over the $n$ sampled classifiers by a Gaussian random variable via the Central Limit Theorem, applied to $Y_i = \\exp\\bigl(\\tilde{\\beta}\\,\\mathbb{E}_{x'}[\\operatorname{sign}(w_*^T x'\\, x'^T w_i)]\\bigr)\\, \\operatorname{sign}(w_*^T x\\, x^T w_i)$, where $\\tilde{\\beta} = \\beta/2$. The mean and variance of $Y$ are evaluated using Grothendieck's identity, $\\mathbb{E}_{x'}[\\operatorname{sign}(w_*^T x'\\, x'^T w)] = \\tfrac{2}{\\pi}\\arcsin\\bigl(w_*^T w / \\lVert w\\rVert\\bigr)$, together with a concentration result that the inverse norm $1/\\lVert w\\rVert$ concentrates around $1/\\sqrt{d-2}$, allowing the arcsin to be linearized. These steps produce the function $\\xi(\\beta)$ whose maximizer is $\\beta^* = \\pi\\sqrt{d-2}$.","core_discovery":"On the paper's own terms, the discovery is that the Gibbs-weighted random perceptron ensemble has a single optimal inverse temperature, $\\beta^* = \\pi\\sqrt{d-2}$, and that this value is universal: it depends only on the ambient dimension $d$ of the input space. The ensemble classifier is $\\operatorname{sign}\\bigl(\\sum_i \\exp(-\\beta L(\\phi(w_i)))\\, \\phi(w_i)(x)\\bigr)$, with $w_i \\sim \\mathcal{N}(0, I_d)$, $L$ the 0-1 loss, and $\\phi(w)(x) = \\operatorname{sign}(w^T x)$. For data $x \\sim \\mathcal{N}(0, I_d)$ with labels $y = \\operatorname{sign}(w_*^T x)$, the paper derives the loss profile through a Central Limit Theorem approximation, identifies its minimizer in closed form, and confirms numerically that the minimum is unchanged when the teacher $w_*$, the number of classifiers $n$, or the number of samples $N$ are varied. The same qualitative finite-temperature minimum is observed on MNIST for three different label tasks at essentially the same $\\beta^*$, suggesting the phenomenon extends beyond isotropic Gaussians.","pith_inferences":["Editorial inference: if the universality of $\\beta^*$ survives moderately non-Gaussian data, the temperature could serve as a default hyperparameter in random-feature and reservoir models, replacing cross-validated temperature searches.","Editorial inference: the i.i.d. Central Limit Theorem step suggests the approximation is safest when $n$ is large relative to the correlation length induced by $x$; an explicit finite-$n$ correction or a conditional CLT would show whether $\\beta^* = \\pi\\sqrt{d-2}$ is only the leading term of a series in $1/\\sqrt{n}$.","Editorial inference: on low-intrinsic-dimension data like MNIST, the relevant $d$ in $\\beta^* = \\pi\\sqrt{d-2}$ may be the intrinsic dimension rather than the ambient dimension; studying synthetic data with known intrinsic dimension would test whether the formula transfers."],"forward_implications":["At $\\beta^* = \\pi\\sqrt{d-2}$, an untrained random perceptron ensemble classifies Gaussian teacher data near its best attainable performance, so no validation-set tuning of the temperature is needed on such data.","The value $\\beta^*$ is invariant under changes of teacher, ensemble size $n$, and sample size $N$, so the optimal weighting is a function of the data geometry alone in the isotropic Gaussian case.","The same finite-temperature optimum appears on MNIST for three distinct binary tasks, indicating that the mechanism is not an artifact of Gaussian symmetry and may apply to structured real-world data.","Because $\\beta^*$ grows as $\\sqrt{d}$, higher-dimensional inputs call for sharper weighting of low-loss classifiers even when $n$ and $N$ are held fixed.","The perceptron-merging approximation collapses the ensemble to a single weight vector $w_{\\mathrm{ens}} = \\sum_i \\alpha_i w_i$ with nearly unchanged accuracy, reducing inference cost to one dot product."],"supporting_citations":[{"why":"It supplies Grothendieck's identity and the concentration-of-measure bounds that convert the expected loss into a closed-form function of beta and d.","marker":"[32]"},{"why":"It provides the teacher-vector label model that defines the Gaussian classification task whose optimum the paper computes.","marker":"[26]"},{"why":"It supplies the Gibbs measure and canonical-ensemble formalism that defines the exponential weighting of classifiers by their loss.","marker":"[17]"},{"why":"It supplies the softmin limit used to show that beta tending to infinity selects the best sampled classifier, motivating a finite-beta optimum.","marker":"[27]"},{"why":"It provides the MNIST handwritten-digit data used to test whether the Gaussian phenomenon transfers to structured real-world inputs.","marker":"[18]"}],"fun_headline_variants":["Random perceptrons vote best at a single universal temperature","Optimal voting temperature in random ensembles depends only on input dimension","Finite temperature turns random classifiers into a sharp ensemble","Universal beta* = pi sqrt(d-2) governs random classifier ensembles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's central approximation is that the average over classifiers of $Y_i$ is Gaussian with the independent-sample variance $(1/n)\\operatorname{Var}(Y)$, although every $Y_i$ shares the same test point $x$; if the correlations induced by $x$ are not negligible at the studied $d$ and $n$, the predicted $\\beta^*$ need not minimize the actual loss.","fun_headline_variants_meta":{"raw":{"variants":["Random perceptrons vote best at a single universal temperature","Optimal voting temperature in random ensembles depends only on input dimension","Finite temperature turns random classifiers into a sharp ensemble","Universal beta* = pi sqrt(d-2) governs random classifier ensembles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2597,"prompt_tokens":1000,"completion_tokens":1597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1528}},"tokens_in":616,"tokens_out":1597,"duration_ms":12044,"temperature":1.0,"reasoning_tokens":1528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:12:13.922831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $d$ and sample many independent weight matrices $W$ and test points $x$. For a given teacher $w_*$, compute the empirical ensemble 0-1 loss at $\\beta = \\pi\\sqrt{d-2}$ and at, say, $0.5\\,\\pi\\sqrt{d-2}$. If, averaged over realizations, the loss at the predicted $\\beta$ is not below the loss at neighboring values, or if the location of the empirical minimum shifts systematically with $n$ or with the choice of $w_*$, the universality claim fails. A sharper test is to repeat the measurement with two structurally different teachers and check whether the empirical minimizer coincides.","supporting_citations":[{"cited_title":"sign nX i exp ˜βEx′ h sign(wT ∗ x′x′T wi) i sign(wT ∗ xxT wi) !# , which can be rewritten, employing the identity sign(x) = 2· 1 ≥(x)− 1 that holds for everyx̸= 0, as −¯L∝ Ex,W1","cited_arxiv_id":null,"evidence_quote":"It supplies Grothendieck's identity and the concentration-of-measure bounds that convert the expected loss into a closed-form function of beta and d."},{"cited_title":"Huang, Statistical mechanics","cited_arxiv_id":null,"evidence_quote":"It supplies the Gibbs measure and canonical-ensemble formalism that defines the exponential weighting of classifiers by their loss."}],"review_version":1}