{"id":"02cc953f-08de-4c69-8b6e-913068f59b8e","arxiv_id":"2502.01032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Closed-form polynomial approximations of MLPs and GLUs under Gaussian mixture inputs reveal a training phase where networks shift from linear to quadratic behavior.","lead":"Researchers derived closed-form formulas for the best polynomial approximations of neural network layers, assuming the inputs follow a Gaussian mixture distribution. This lets them inspect what MLPs and GLUs compute by looking at the polynomials' eigendecompositions, and they show the approximations transfer to adversarial-style attacks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The closed-form derivation is internally incorrect: Eq. 16's E[g(X)] coefficient is α1α2Cov(ε1,ε2) but should be α1α2+Cov(ε1,ε2), and Eq. 6 omits the E[xj]E[ϕ(yi)] term in Stein's lemma.","rationale":"The reader's weakest_assumption emphasized the Gaussian-mixture input assumption. I agree that is a limitation, but the most load-bearing problem is internal: the algebraic identities supporting the closed-form claim contain errors. Eq. 6 is wrong for non-centered inputs and Eq. 16 is wrong even for centered inputs, so the theoretical derivation does not, as written, yield the least-squares coefficients. This is not a disagreement with the community's consensus; it is a correctness risk in the paper's own equations. The intended framework may be repairable—the Master Theorem's general coefficient formula (Eq. 14) is plausibly correct, and the code may already implement corrected moments—but the text must be fixed and the experiments re-verified before the central claim can be taken as stated. The reader did identify Eq. 6 in the rationale, though not as the weakest assumption, so agreement is partial. The transparent Appendix B and released code are creditworthy, but the SGD-acquired quadratic approximant on MNIST does not test the closed-form theory. Verdict stays conditional: correct the algebra, re-run the FVU and attack experiments, and the paper could be accepted; as is, the derivation needs revision.","tokens_in":11029,"tokens_out":13276,"duration_ms":119951,"concrete_test":"Check Eq. 16 with g(x)=1, X~N(0,1), Y1=X+ε1, Y2=X+ε2, where ε1,ε2 are mean-zero with Cov(ε1,ε2)=1 and independent of X. Then β1=β2=1, α1=α2=0, so Eq. 16 predicts E[Y1Y2]=β1β2E[X^2]=1, while the true value is E[X^2]+Cov(ε1,ε2)=2. Separately, verify Eq. 6 with x~N(μ,1), y=x, ϕ(y)=y: the formula gives LHS=1, but E[x^2]=μ^2+1, so the missing E[xj]E[ϕ(yi)] term is confirmed. Recompute the linear and quadratic FVU curves on MNIST with the corrected identities; if the curves shift materially, the reported phase transition and 95% variance claim need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that polynomial coefficients can be computed analytically as the exact least-squares solution for Gaussian mixture inputs. Two algebraic errors in the derivation break that claim as written. First, Eq. 6 states E[ϕ(y_i)x_j] = Cov(y_i,x_j)E[ϕ'(y_i)]. For jointly Gaussian y_i,x_j, the correct identity is E[ϕ(y_i)x_j] = E[x_j]E[ϕ(y_i)] + Cov(y_i,x_j)E[ϕ'(y_i)]. The missing first term does not vanish in the subsequent subtraction: substituting Eq. 6 into Eq. 4 produces Cov[f(x),x] = W2 C D - W2 E[ϕ(y)]E[x]^T instead of W2 C D, so the displayed β is wrong whenever means are nonzero, as they are for every MNIST class. Second, the n=2 specialization of Theorem 3.1 (Eq. 16) is also mis-derived. Expanding (α1+β1X+ε1)(α2+β2X+ε2) and using independence of ε_i from X gives E[g(X)Y1Y2] = β1β2E[g(X)X^2] + (α1β2+α2β1)E[g(X)X] + (α1α2 + Cov(ε1,ε2))E[g(X)], not α1α2Cov(ε1,ε2)E[g(X)]. The product form is wrong even for zero-mean inputs when residuals are correlated. Since these formulas feed the linear and quadratic coefficients (Secs. 3.1, 3.2, 3.4), the paper's derivation does not currently establish the claimed closed-form least-squares approximants. The MNIST quadratic FVU plot cannot rescue the theory, because Appendix B obtains those coefficients by SGD finetuning rather than the closed form.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for converting one-hidden-layer MLPs and gated linear units (GLUs) into least-squares-optimal polynomial approximants in closed form, assuming the input distribution is Gaussian or a Gaussian mixture. The derivation combines Stein's lemma with a 'master theorem' that reduces multivariate Gaussian integrals of the form E[g(X)Y1...Yn] to univariate expectations. The authors apply the linear approximants to an MNIST-trained MLP, track the fraction of variance unexplained (FVU) over training time as evidence for the distributional simplicity bias, visualize top eigenvectors of the quadratic approximants, and show that SVD-based ablations of the linear approximant transfer as adversarial interventions to the original network. They also release a code library. The manuscript is transparent about its main limitations, notably the Gaussian-mixture assumption and the fact that the quadratic MNIST approximants are obtained by SGD finetuning rather than the closed-form formulas.","tokens_in":11424,"tokens_out":12529,"duration_ms":118508,"significance":"If the derivation is corrected, the paper offers a useful and reusable interpretability tool: analytic polynomial surrogates for common feedforward modules, with explicit integral formulas for ReLU and GELU, a clean way to quantify the order of statistics a network uses, and falsifiable predictions via adversarial transfer. The master-theorem reduction and the appendices with Gaussian integrals are valuable contributions, and the code release is a strength. The paper does not appear to assume its conclusions: the FVU measurements are not used to fit the polynomial coefficients beyond the disclosed finetuning. However, the headline quadratic results are currently produced by SGD finetuning, not the closed-form derivation, so the theoretical contribution and the strongest empirical validation are less tightly coupled than the abstract suggests.","major_comments":[{"comment":"The displayed identity E[phi(y_i)x_j] = Cov(y_i,x_j)E[phi'(y_i)] is not correct for nonzero-mean jointly Gaussian variables; the general identity is E[phi(y_i)x_j] = E[x_j]E[phi(y_i)] + Cov(y_i,x_j)E[phi'(y_i)]. If Eq. (6) is used literally in the computation of E[f(x)x^T], the omitted E[x_j]E[phi(y_i)] term does not cancel in the subsequent subtraction defining Cov[f(x),x], because that subtraction is not shown; the final beta in Eq. (4) is therefore not justified as written. The derivation should either state and use the covariance identity Cov(phi(y_i),x_j) = Cov(y_i,x_j)E[phi'(y_i)] and show the cancellation, or carry the extra term through the computation.","section":"Section 3.1, Eq. (6)"},{"comment":"The n=2 specialization of the master theorem is mis-derived. Expanding (alpha1+beta1X+epsilon1)(alpha2+beta2X+epsilon2) and using the independence of epsilon1,epsilon2 from X gives E[g(X)Y1Y2] = beta1beta2 E[g(X)X^2] + (alpha1beta2+alpha2beta1)E[g(X)X] + (alpha1alpha2 + Cov(epsilon1,epsilon2))E[g(X)]. Equation (16) writes the constant coefficient as alpha1alpha2 Cov(epsilon1,epsilon2), omitting the alpha1alpha2 E[g(X)] term. Since this identity feeds the quadratic coefficient derivation in Sections 3.2 and 3.5, the closed-form quadratic formulas are not established as written and must be corrected and re-derived downstream.","section":"Section 3.3, Eq. (16)"},{"comment":"The quadratic FVU curves that support the abstract's claim that 'quadratic approximants explain over 95% of the variance' are not computed with the closed-form Gaussian-mixture solution. As the footnote and Appendix B state, that solution is intractable on MNIST, so the authors initialize with the standard-Gaussian closed form and finetune with SGD on Gaussian-mixture samples. This is disclosed, but it means the headline quadratic result and the quadratic branch of the simplicity-bias experiment validate a numerical fit to a convex objective, not the analytic derivation. Please qualify the abstract/conclusion accordingly, or supply convergence or error evidence that the SGD solution is close to the true least-squares coefficients.","section":"Appendix B / Section 4.1, footnote 3"}],"minor_comments":[{"comment":"The upper limit of the sum in Eq. (9) should be n, not n-1; the product of n linear factors has degree n, and the subsequent n=2 formula is consistent with k=0,1,2.","section":"Theorem 3.1, Eq. (9)"},{"comment":"The caption reads 'Fraction of variance explained (FVU)', but FVU is the fraction of variance unexplained; the label should be corrected.","section":"Figure 2 caption"},{"comment":"The abstract and main text state the >95% quadratic FVU result without noting that the quadratic MNIST coefficients were SGD-finetuned rather than computed by the closed-form formulas; this should be flagged where the number is advertised.","section":"Abstract and Section 4.2"},{"comment":"The phrase 'the the simple' contains a duplicated article and should read 'the simple'.","section":"Section 2"},{"comment":"The feature map phi_2(x) is defined with n as the input dimension, but the rest of the paper uses d; the notation should be unified.","section":"Eq. (7)"},{"comment":"The caption says 'at step (y)' without defining y; it should reference the log-spaced step index used in the text, e.g., step 2^12 or a named axis variable.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The algebra errors in Eqs. (6) and (16) are local and fixable, and the underlying method appears sound, so I do not recommend rejection. The more substantive issue is that the paper's main empirical demonstration of the quadratic approximant bypasses the closed-form formula; I would ask the authors to either run the closed-form computation in a reduced-dimensional setting where it is tractable, or to add a quantitative convergence check for the SGD finetuning. There is no circularity concern: the FVU measurements are not used to fit the polynomial coefficients beyond the disclosed finetuning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this one because the intended theory is likely correct but the paper as written doesn't establish it. The core contribution is genuinely new: analytic formulas for least-squares polynomial approximants of MLPs and GLUs under Gaussian mixture inputs, extending Pearce et al.'s bilinear-layer work. Releasing code and tracing FVU across training is a nice application of the simplicity-bias idea. The paper is also transparent about its limitations and about the SGD finetuning used for the quadratic approximants on MNIST.\n\nThe algebra, however, has two load-bearing errors. Equation 6 omits the E[x_j]E[phi(y_i)] term in Stein's lemma for nonzero-mean Gaussians. On MNIST every class mean is nonzero, so the displayed linear coefficient is not the least-squares solution. Equation 16's coefficient on E[g(X)] should be alpha1*alpha2 + Cov(eps1, eps2), not alpha1*alpha2*Cov(eps1, eps2). Even for zero-mean inputs the quadratic coefficient is wrong when residuals are correlated. These aren't appendix typos; they feed the formulas used for the linear and quadratic approximants. The main theoretical claim as stated is therefore incorrect.\n\nThe empirical quadratic result also doesn't exercise the closed form. The footnote in Section 4.1 says the analytic solution for Gaussian mixtures is intractable on MNIST, so they finetune under N(0,1) with SGD. The \"over 95% variance explained\" plot is a property of the SGD-finetuned approximant, not the closed-form derivation. The adversarial attack lacks a baseline — no comparison to random projections or other ablations — and the curves are single-run, so the lockstep claim is weaker than presented.\n\nThese are serious but fixable issues. The intended approach — using Stein's lemma and Gaussian-moment identities to derive polynomial approximants — is sound in principle, and the errors are the kind a careful revision can correct. The paper is not incoherent or circular; it's a promising method with a flawed write-up.\n\nThis is for mechanistic interpretability readers who want a closed-form alternative to SGD-fitted proxies. It deserves a serious referee, but only with the promise of a heavy revision. I'd send it to review rather than desk-reject, then require corrections to Eq. 6 and Eq. 16, and ask for a closed-form quadratic demonstration on a lower-dimensional problem or a clear separation of the SGD-based empirical claims.","headline":"Useful idea, honest paper, but the closed-form derivation has two fixable algebraic errors and the headline quadratic result isn't actually computed in closed form.","tokens_in":11959,"tokens_out":3203,"would_cite":false,"duration_ms":31169,"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 derives closed-form least-squares polynomial approximations of MLPs and GLUs under Gaussian-mixture inputs, making their linear and quadratic structure directly computable from weights and moments.","keywords":["closed-form polynomial approximation","Gaussian mixture inputs","mechanistic interpretability","simplicity bias","gated linear units","multilayer perceptrons","SVD adversarial examples","fraction of variance unexplained"],"falsifier":"Train a network on a non-Gaussian dataset (e.g., natural images with heavy-tailed pixel statistics or raw audio), fit the closed-form linear and quadratic approximants using a Gaussian mixture model of the inputs, and compare their $R^2$ against polynomials of the same degree fitted by SGD on the true samples; a large drop in $R^2$ for the Gaussian-derived approximant would show the closed-form optimality premise fails on that distribution.","tokens_in":10781,"feed_emoji":"🧠","tokens_out":7640,"duration_ms":66536,"temperature":0.7,"pith_summary":"The paper tries to establish that feedforward networks need not be treated as black boxes: under a Gaussian-mixture input assumption, the best least-squares polynomial approximation of any degree to an MLP or a gated linear unit can be computed exactly, with coefficients read off from the network weights and input moments. This gives access to the network's linear and quadratic structure without training an approximator or sampling from the model. The authors show that a quadratic approximant captures over 95% of an MNIST-trained MLP's output variance, and that directions found in the approximant's SVD transfer as effective adversarial attacks on the original network. They also use the approximants to watch how the network's effective complexity grows during training, finding a phase where the linear fit worsens while the quadratic fit holds steady—consistent with the idea that networks learn low-order statistics first.","feed_headline":"Closed-form polynomials replicate MLPs at 95% output variance","feed_subtitle":"No training needed: the best polynomial fit to a network follows from weights and input moments.","key_machinery":"The load-bearing object is the paper's Master Theorem (Theorem 3.1), which reduces any expectation of the form $\\mathbb{E}[g(X)\\prod_{i=1}^n Y_i]$ over jointly Gaussian variables to a linear combination of univariate expectations $\\mathbb{E}[g(X)X^k]$. The reduction works by regressing each $Y_i$ on $X$ to get $Y_i = \\alpha_i+\\beta_i X+\\epsilon_i$, expanding the product combinatorially, and using Isserlis' theorem to evaluate products of the residuals, which are Gaussian and independent of $X$. This theorem turns the feature-map/OLS strategy for polynomial approximation into a closed-form computation: for Gaussian inputs the preactivations of an MLP or GLU are Gaussian, Stein's lemma gives the linear cross-covariances, and the theorem supplies the higher moments, while the law of total covariance extends the result to Gaussian mixtures.","core_discovery":"The paper's central claim is that the coefficients of the least-squares optimal polynomial approximant of an MLP or GLU, of any chosen degree, can be written analytically from the network's weights and biases together with the mean and covariance of a Gaussian input distribution (or the component means and covariances of a Gaussian mixture). The derivation reduces every required expectation to univariate Gaussian integrals via a master theorem, so the approximant is produced without stochastic optimization. The authors report that on MNIST, the quadratic approximant explains over 95% of the variance of the trained MLP's outputs, and the same approximant's top singular vectors, when ablated from the input, drive the original network's accuracy to near-chance—evidence that the polynomial captures causally relevant structure.","pith_inferences":["Editorial: the Master Theorem's structure suggests the method extends to any activation whose Gaussian moments are analytically or numerically tractable, including Swish, softplus, and possibly learned activations, without new theory.","Editorial: the sharp rise in linear FVU combined with flat quadratic FVU could be used as a diagnostic for when a network stops learning mean/covariance structure and starts exploiting higher moments; testing it on transformers or with different optimizers would reveal whether the 'quadratic phase' is a general inductive bias.","Editorial: because the linear approximant's SVD directions are computed analytically from weights alone, they could support zero-shot concept steering or erasure in feedforward modules, complementing approaches that require training a probe on model activations.","Editorial: the overcomplete eigenbasis of the quadratic approximant ($d^2$ vectors for residual-stream dimension $d$) may function as a fixed, data-independent analogue of sparse autoencoder features, potentially helping separate features the model computes from features that merely exist in the data."],"forward_implications":["For any single-hidden-layer MLP or GLU with Gaussian mixture inputs, a linear approximant is computed by inverting the input covariance and applying Stein's lemma; no data fitting is needed, only the network weights and the input moments.","Higher-degree approximants follow from the same feature-map trick, with the Master Theorem supplying all cross-moments; quadratic approximants are feasible for input dimensions up to thousands because the feature covariance has closed form.","On MNIST, the quadratic approximant explains more than 95% of output variance, and ablating the top four SVD directions of the linear approximant drops the original MLP's accuracy below 50%, so the approximants expose input directions that are causally load-bearing for the network.","Training-time $R^2$ trajectories show the linear fit degrading sharply after a phase transition while the quadratic fit stays tight, evidence that the network first learns linear statistics and then quadratic ones, consistent with a simplicity bias.","The same analytic machinery applies to GLUs because their preactivations are jointly Gaussian, so gated linear units inherit the closed-form polynomial approximation and its interpretations."],"supporting_citations":[{"why":"Shows that each output unit of a bilinear layer is a quadratic form in the input, motivating the use of quadratic approximants and eigendecomposition for interpretability.","marker":"Pearce et al. (2024)"},{"why":"Supplies the formula for the expected product of centered jointly Gaussian variables used to evaluate residual expectations in the Master Theorem.","marker":"Isserlis (1918)"},{"why":"Provides the normal integral tables used to compute closed-form expectations for GELU activations.","marker":"Owen (1980)"},{"why":"Introduces the GELU activation function, whose moments the paper evaluates analytically for the polynomial approximants.","marker":"Hendrycks & Gimpel (2016)"},{"why":"Introduces GLU variants, whose closed-form polynomial approximants are derived in Section 3.5.","marker":"Shazeer (2020)"},{"why":"Provides the MNIST dataset, which the experiments use as a Gaussian-mixture-modeled input distribution.","marker":"LeCun et al. (1998)"},{"why":"Advances the 'statistics of increasing complexity' hypothesis that the paper tests by tracing approximant fit across training.","marker":"Belrose et al. (2024)"},{"why":"Supports the prior claim that networks learn functions of increasing complexity, which motivates the training-time FVU analysis.","marker":"Nakkiran et al. (2019)"}],"fun_headline_variants":["Analytic polynomials replicate MLP outputs at 95% variance","Closed-form polynomial fits for MLPs need no optimization","MLPs become polynomials: weights plus moments give fit","Optimal polynomial approximation of MLPs in closed form"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the input genuinely follows (or is well-approximated by) a Gaussian mixture with known component means and covariances; the closed-form coefficients are least-squares optimal only under that distribution, and if the true data distribution is far from Gaussian, the interpretability and attack conclusions will not follow.","fun_headline_variants_meta":{"raw":{"variants":["Analytic polynomials replicate MLP outputs at 95% variance","Closed-form polynomial fits for MLPs need no optimization","MLPs become polynomials: weights plus moments give fit","Optimal polynomial approximation of MLPs in closed form"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1644,"prompt_tokens":846,"completion_tokens":798,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":733}},"tokens_in":462,"tokens_out":798,"duration_ms":7841,"temperature":1.0,"reasoning_tokens":733,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:51:43.872956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a network on a non-Gaussian dataset (e.g., natural images with heavy-tailed pixel statistics or raw audio), fit the closed-form linear and quadratic approximants using a Gaussian mixture model of the inputs, and compare their $R^2$ against polynomials of the same degree fitted by SGD on the true samples; a large drop in $R^2$ for the Gaussian-derived approximant would show the closed-form optimality premise fails on that distribution.","supporting_citations":[{"cited_title":"On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables","cited_arxiv_id":null,"evidence_quote":"Supplies the formula for the expected product of centered jointly Gaussian variables used to evaluate residual expectations in the Master Theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the normal integral tables used to compute closed-form expectations for GELU activations."},{"cited_title":"T., and Fern, X","cited_arxiv_id":null,"evidence_quote":"Advances the 'statistics of increasing complexity' hypothesis that the paper tests by tracing approximant fit across training."},{"cited_title":"L., Zhang, F., and Barak, B","cited_arxiv_id":null,"evidence_quote":"Supports the prior claim that networks learn functions of increasing complexity, which motivates the training-time FVU analysis."}],"review_version":1}