{"id":"7ab04478-fed9-4b95-b6a8-bb22c106da7f","arxiv_id":"2507.01044","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A short note arguing via Minkowski sums that infinite-width shallow network objectives become convex in the limit, while leaving the finite-width local landscape unproved.","lead":"This paper claims that for a simple wide shallow neural network, the epigraph of the input-output map becomes nearly the epigraph of a convex function as width grows, so the limiting training problem is convex. The argument relies on known Minkowski-sum convexification, but the paper does not bridge the gap between the epigraph of averaged parameters and the actual nonconvex training objective.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5 does not follow from the epigraph-convexification result: Φ_N is minimized over the product space B_N and never forms the averaged parameter that Corollary 3.3 and Lemma 4.1 actually control.","rationale":"The reader identified the load-bearing weakness as the mismatch between the single-parameter epigraph convexification and the product parameter space of the trained objective. This is indeed the critical gap. Corollary 3.3 proves that averaged points from one truncated epigraph fill out the epigraph of the convex minorant; it has no content about minimizers of a separable sum over independent parameters. The proof of Theorem 4.5 cites Lemma 4.1 after establishing only scalar Cesàro/Abel limits, and never constructs the averaged parameter point to which the epigraph convergence would apply. The theorem is therefore either ill-typed or, on the natural product reading, does not express the claimed convexification. The paper's introduction explicitly draws the stronger conclusion 'all local minima are global minima' from the convexity of the limiting problem, and this conclusion fails already for finite N on any q with a non-global local minimum, since componentwise tuples give local minima of Φ_N. The LMS section does not repair the gap: it again replaces a weighted sum of component errors by a Cesàro average before squaring, without justifying the interchange or connecting the averaged output to a convex minorant in the parameter actually trained. I find no independent support that would rescue the main claim: the Minkowski-sum theorem is standard, but the paper's use of it does not match the structure of the neural-network objective. The verdict should remain REJECT, as the central theorem and the interpretation built on it are not established.","tokens_in":6664,"tokens_out":15702,"duration_ms":181759,"concrete_test":"Test the claimed transfer on a scalar example. Take p=1, D=[-2,2], and q(β)=min((β+1)^2, (β-1)^2) (equal global minima at ±1, convex minorant q* with Argmin(q*)=[-1,1]). For every N, both (1,...,1) and (-1,...,-1) are global minimizers of Φ_N. Check whether Theorem 4.5's conclusion is meaningful: under the product topology the limit set is {−1,1}^∞, not Argmin(q*)=[-1,1]; under the weighted-average projection the limit is ±1, which is in Argmin(q*) but is not the object minimized by Φ_N. Then perturb the two wells to make them unequal, e.g. q(β)=(β^2-1)^2+0.1β, and compute all stationary points of Φ_N for N=1,5,20 by grid refinement plus Newton polish. If the tuple with every coordinate at the higher local minimum is a local minimum of Φ_N with positive-definite Hessian but non-global value, the paper's 'all local minima are global' claim is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 3.3 and Lemma 4.1 are statements about the Minkowski average of points (β_i, y_i) drawn from a single truncated epigraph fepi(q) ⊂ R^{p+1}: the point ((1/(2N+1))Σβ_i, (1/(2N+1))Σy_i) has lower envelope converging to fepi(q*). The trained objective (II.3), Φ_N^{α,n}(B_N) = Σ_{i=-N}^N K_N^α(i) g(X_n, Y_n^i; β_i), is instead minimized over the product B_N = D^{2N+1}. Because the sum is coordinatewise separable, a minimizer is a tuple of minimizers of the individual terms, and the argument of Φ_N is this tuple, not its weighted average parameter. The proof of Theorem 4.5 uses (IV.1)–(IV.2) only to compare scalar loss values; the sentence “the claim follows by … Lemma 4.1” supplies no map from a minimizer B̂_N to a point of fepi(q*) whose coordinate is an average of the β_i. Without such an embedding, “β̂_N → Argmin(g*)” is not well-typed: the natural product topology gives limit points in the hypercube ∏_i Argmin(g_i), while Argmin(g*) is a convex set in R^p. Moreover, if q has a non-global local minimum, the tuple placing every coordinate in that local minimum is a local minimum of Φ_N, although a convex function has none. Thus the central claim that the infinite-width optimization problem is effectively convex is unsupported; the epigraph result at best concerns an averaged parameter, not the trained parameter tuple.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers a shallow-and-wide network whose output is a convex combination \\sum_{i=-N}^N K_N^\\alpha(i) f_{\\beta_i}(x) of component outputs, with kernel weights K_N^\\alpha(i) proportional to \\alpha^{|i|}. For fixed data, the training objective (II.3) is a weighted sum of per-component errors over the product parameter space D^{2N+1}. The paper recalls the Shapley-Folkman-Starr result that Minkowski averages of truncated epigraphs converge to the epigraph of the convex minorant, and it claims that this implies asymptotic convexity of the neural-network map, culminating in Theorem 4.5, which asserts convergence of minimizers of Phi_N to the Argmin of the convex minorant. Section V applies the same idea to the LMS criterion and sketches an SGD convergence argument.","tokens_in":7048,"tokens_out":30221,"duration_ms":301408,"significance":"If true, the paper would provide a compact geometric explanation for the empirical success of wide shallow networks and would illustrate a known convexification mechanism in a new setting. It is concise, transparent about the geometric ingredients, and it correctly identifies the relevant literature (Fradelizi-Madiman-Marsiglietti-Zvavitch; Starr; Shapley-Folkman). The isolated epigraph statements in Section III are correct as far as they go. However, the central transfer from Minkowski averages of epigraphs to the actual product-space optimization problem is not made; Theorem 4.5 has a type mismatch between the averaged parameter and the trained parameter tuple, and Section V inherits this gap. I therefore view the main claim as unsupported rather than merely insufficiently detailed.","major_comments":[{"comment":"Lemma 4.1 and Corollary 3.3 are statements about averages (1/n)\\sum_{i=1}^n(\\beta_i,y_i) of points chosen from a single truncated epigraph fepi(q) in R^{p+1}. The objective (II.3), however, is minimized over the product space D^{2N+1}, and its argument is the tuple B_N=(\\beta_{-N},...,\\beta_N), not the averaged parameter \\bar\\beta_N=(2N+1)^{-1}\\sum_i\\beta_i. The proof of Theorem 4.5 uses (IV.1) and (IV.2) only to compare scalar loss values; it then says that 'the claim follows by Lemma 4.1', but it never constructs a point of fepi(g_*) whose first coordinate is obtained from a minimizer \\hat\\beta_N by any averaging map. In addition, lim_{N\\to\\infty}\\hat\\beta_N is not well-typed, since \\hat\\beta_N lives in a different product space for each N; the natural limit object is a point of \\prod_i Argmin(g_i), not of the convex set Argmin(g_*) in R^p. This is the load-bearing step of the paper, and it is missing.","section":"Section IV, Theorem 4.5 and its proof"},{"comment":"Even under a charitable reading that replaces \\hat\\beta_N by its componentwise average, the conclusion would not establish the claimed convexity of the training landscape. For finite N, Phi_N^{\\alpha,n} is coordinatewise separable, so if q has a non-global local minimum \\beta_loc, the constant tuple (\\beta_loc,...,\\beta_loc) is a non-global local minimum of Phi_N^{\\alpha,n}: each coordinate is locally minimal and the total value exceeds the value at the constant global-minimizer tuple. A convex function has no non-global local minima, but Phi_N^{\\alpha,n} is not the convex function g_* on R^p, and no argument in the paper transfers the convexity of g_* to the function that is actually minimized. Consequently, the statement in Section I that all local minima are global minima is not supported by the preceding lemmas.","section":"Section I and Section IV (local-minima and convexity claims)"},{"comment":"The LMS section applies 'the theory developed above' to Cesaro averages of (f_{\\beta_*}(x)-f_{\\beta_i}(x))^2, but it does not identify the resulting convex function \\tilde f_\\beta(x) with the value function of the network output \\sum_i K_N^\\alpha(i)f_{\\beta_i}(x); without this identification, the replacement of the trained objective by \\tilde f is unjustified. The claim that the squared sign-indefinite convex function 'will not be convex, but it will nevertheless have all its local minima = global minima' is asserted without proof. Finally, the SGD convergence argument assumes that the limiting objective is the one being optimized, that gradient norms are controlled outside the minimizer set, and that unstable equilibria are almost surely avoided; these are nontrivial conditions that are not derived from the lemmas.","section":"Section V (LMS criterion and SGD)"}],"minor_comments":[{"comment":"The function g(x,y;\\beta) is defined as the vector y-f_\\beta(x), but Phi_N^{\\alpha,n} is stated to take values in R_+; please state explicitly that g in (II.3) is the squared Euclidean norm or otherwise clarify the codomain.","section":"Section II, Eq. (II.3)"},{"comment":"The output is written as Y=\\sum_{i=-N}^N K_N^\\alpha(i)\\alpha^i Y_i^n, which mixes the kernel K_N^\\alpha(i) with an additional \\alpha^i; this appears to be a typo and should be reconciled with (II.3).","section":"Section V, first display"},{"comment":"The word 'Haussdorff' should be 'Hausdorff'.","section":"Section III, Corollary 3.3"},{"comment":"The citation to Radhakrishnan, Belkin and Uhler gives the year 2013, but the PNAS article cited is from 2023; please verify the year.","section":"References"},{"comment":"Theorem 4.5 should state the assumption, used in (IV.2), that the Cesaro limit in Lemma 4.4 exists for the parameter sequences under consideration; otherwise the argument is conditional on an unstated hypothesis.","section":"Section IV, Lemma 4.4 and Theorem 4.5"}],"recommendation":"reject","confidential_remarks":"The paper's geometric lemmas are fine, but the advertised neural-network conclusion is not derived; the mismatch between the product-space objective and the averaged-parameter epigraph is fundamental, not a local fix. A revised version could perhaps claim convergence of the Cesaro-averaged parameter or of the minimal values, but that would be a different statement from Theorem 4.5. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know one thing about this paper up front: the main theorem, Theorem 4.5, is not supported by the preceding material. The epigraph convexification result (Corollary 3.3 and Lemma 4.1) lives in a fixed parameter space R^{p+1} and controls averages of points drawn from a single truncated epigraph. The objective Φ_N is minimized over the product space D^{2N+1}, one parameter vector per network copy, and the proof never connects a minimizer tuple to an averaged parameter. As a result, \"β̂_N → Argmin(g*)\" is not well-typed in the natural topology: a tuple could sit in a product of individual local minima. That is a load-bearing gap, not a technicality.\n\nThe paper does deserve some credit. The idea of using Shapley-Folkman/Starr convexification of Minkowski sums as an explanation for the trainability of wide shallow networks is new in framing, and the writing is direct and honest—the abstract's 'plausible explanation' is the right register. The citations to Starr and to Fradelizi et al. are apt, and the Tauberian argument in Lemma 4.4 appears sound.\n\nThe soft spots, in proportion. The product-space gap is fatal to the stated theorem. A weaker, salvageable statement would be convergence of the averaged parameter (or the empirical distribution of the tuple) to Argmin(g*). The LMS section has a second, related overreach: the claim that all local minima are global for the squared error goes beyond what epigraph convergence can show, and the square of a sign-indefinite convex minorant is generally not convex. The local-minima issue is not rescued by the observation that unstable maxima are avoided.\n\nIf you want a quick geometric intuition for why wide networks might behave well, this note gives it. But as a rigorous proof, it doesn't hold up. I would send it to a referee who knows convex geometry and stochastic approximation, with the expectation of heavy revision or a recast conclusion. If the authors can prove the averaged-parameter version, that would be a publishable short note; as written, it is not there yet.","headline":"The minimizer convergence claim doesn't follow from the epigraph result, but the Shapley-Folkman framing is a fresh heuristic that could become a publishable note after recasting to averaged parameters.","tokens_in":7538,"tokens_out":4741,"would_cite":false,"duration_ms":48800,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","52A41","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a shallow wide network whose output averages many parameterized component maps, the paper proves that as width grows the network objective's epigraph converges to that of a convex function, so the limiting training problem has no…","keywords":["shallow and wide networks","truncated epigraphs","Minkowski sums","convex minorant","stochastic gradient descent","asymptotic convexity","convexification","nonconvex optimization"],"falsifier":"Take a one-dimensional nonconvex error with two separated global minima, e.g. $q(\\beta)=(\\beta^2-1)^2$, form the finite-width objective (II.3), and enumerate all local minima for growing $N$ with $\\alpha$ close to 1. If any sequence of local minima stays a positive distance away from $\\operatorname{co}(\\operatorname{Argmin}(q))=\\{-1,1\\}$ as $N\\to\\infty$ and $\\alpha\\to 1$, the conclusion of Theorem 4.5 is false.","tokens_in":6411,"feed_emoji":"🧠","tokens_out":8377,"duration_ms":77142,"temperature":0.7,"pith_summary":"This paper asks why wide neural networks train so well despite nonconvex objectives. For a simple architecture—a shallow network whose output is a weighted average of many copies of a parameterized map, with weights decaying geometrically away from the central copy—it proves an asymptotic convexity result: the epigraph of the finite-width error function converges in Hausdorff distance to the epigraph of the convex minorant of the single-component error. Consequently the minimizers of the finite-width objective converge, as width grows and then the averaging weight flattens, to minimizers of that convex minorant, whose argmin is the closed convex hull of the original problem's global minimizers. If correct, this gives a structural reason why infinitely wide shallow networks are effectively convex: all local minima are global, and stochastic gradient descent can find them.","feed_headline":"Wide networks become convex in the infinite-width limit","feed_subtitle":"Averaging many parameter copies makes every local minimum global, explaining why wide nets train well.","key_machinery":"The load-bearing device is the Minkowski sum of truncated epigraphs. For a continuous function $q$ on a compact convex domain, its truncated epigraph is the compact set $\\{(\\beta,y): \\beta\\in D,\\, q(\\beta)\\le y\\le M\\}$. A classical convexification theorem, presented as Theorem 3.1 in the paper and drawn from reference [4], says that $(1/n)$ times the $n$-fold Minkowski sum of a compact set converges in Hausdorff metric to its closed convex hull, at rate $O(1/n)$. Corollary 3.3 applies this to $q$: the normalized Minkowski sum of truncated epigraphs converges to the truncated epigraph of the convex minorant of $q$. The paper then observes that the network objective $\\Phi^{\\alpha,n}_N$ is exactly such a weighted average of component errors over the product parameter space, and uses equicontinuity plus a Tauberian theorem from reference [15] to replace the geometric kernel average by a Cesàro average, turning the geometric-mean structure into the Minkowski-averaged epigraph whose limit is convex.","core_discovery":"The central claim is Theorem 4.5: for the weighted error $\\Phi^{\\alpha,n}_N(B_N)=\\sum_{i=-N}^N K^\\alpha_N(i)\\,g(X_n,Y_i^n;\\beta_i)$ with $K^\\alpha_N(i)\\propto \\alpha^{|i|}$, any sequence of minimizers $\\hat{\\beta}_N$ satisfies $\\lim_{\\alpha\\uparrow 1}\\lim_{N\\to\\infty}\\hat{\\beta}_N\\in\\operatorname{Argmin}(g^*)$, where $g^*$ is the convex minorant of the single-component error $g$ and $\\operatorname{Argmin}(g^*)=\\operatorname{co}(\\operatorname{Argmin}(g))$. In words, minimizing the finite, randomly weighted network error over all component parameters drives the parameters into the closed convex hull of the global minimizers of one component's error. Because that limit is a convex minimization problem, the paper concludes that the infinite-width training landscape is effectively convex and that every local minimum is global. The same mechanism is applied to the least mean square criterion: conditioned on the input, the Cesàro-averaged error epigraphs converge to the epigraph of a convex function, and for the squared error the gradient of the asymptotic stochastic gradient descent tends to zero in mean square, with iterates converging in probability to the argmin set.","pith_inferences":["The paper leaves implicit that the convexification is a property of ensembling many parameter copies, not of the network architecture itself; the same Minkowski-sum argument should apply to any averaged ensemble of uniformly continuous nonlinear maps with a common input, giving a general explanation for why ensembling smooths loss landscapes.","A natural extension is to deep networks: if layers or residual blocks are treated as the averaged components, the same reasoning would predict that sufficiently wide deep networks also develop convex-minorant landscapes, though the paper only treats shallow networks.","A testable consequence is that for finite width the set of local minima of the finite objective should cluster around $\\operatorname{co}(\\operatorname{Argmin}(q))$ with diameter $O(1/N)$; measuring this clustering for a simple two-well error would directly probe the mechanism.","The sign-indefinite case in the LMS section shows that convexity of the epigraph can be lost after squaring, yet all local minima remain global along the zero level set; this suggests that the robust signature of width may be the weaker property of having no spurious local minima rather than full convexity."],"forward_implications":["The infinite-width limiting objective is convex: its minimizers form a convex set, so every local minimum is a global minimum.","Finite-width minimizers approach the closed convex hull of the original nonconvex problem's global minimizers, meaning descent methods that succeed on the limit are seeking global minima of the original problem.","For the least mean square criterion, the asymptotic stochastic gradient descent has gradient norm tending to zero in mean square and iterates converging in probability to the argmin set, with unstable local maxima avoided almost surely.","The $O(1/N)$ rate in the Minkowski-sum convergence suggests that finite-width minimizers deviate from the convex-minorant argmin by at most $O(1/N)$ in Hausdorff distance.","The convexification mechanism applies to any family of parameterized maps with a uniformly bounded modulus of continuity, not to a specific activation or architecture beyond the averaging structure."],"supporting_citations":[{"why":"Supplies the convexification theorem for Minkowski sums of compact sets with O(1/n) rate, which is the core mechanism behind Corollary 3.3.","marker":"[4]"},{"why":"Gives the convergence-of-integrals-with-varying-measures result used to pass from the finite weighted sum to the infinite weighted sum in the proof of Theorem 4.5.","marker":"[11]"},{"why":"Provides the Tauberian theorem equating the geometric-kernel weighted limit with the Cesàro average limit, enabling identification of the limit objective with the convex minorant.","marker":"[15]"},{"why":"Supplies the stochastic-approximation background used in the LMS section to argue that unstable local maxima are avoided almost surely by the SGD recursion.","marker":"[1]"}],"fun_headline_variants":["Wide nets: local minima become global","Infinite width makes network loss convex","Wide shallow nets: convexity in the limit","Wide nets: every local minimum becomes global"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that Minkowski-averaging the epigraphs of the individual component errors accurately represents the landscape of the actual combined network objective over the whole product parameter space; if that identification fails, the convex minorant $g^*$ would not describe the trained network's minimizers.","fun_headline_variants_meta":{"raw":{"variants":["Wide nets: local minima become global","Infinite width makes network loss convex","Wide shallow nets: convexity in the limit","Wide nets: every local minimum becomes global"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001119,"raw_usage":{"total_tokens":4603,"prompt_tokens":839,"completion_tokens":3764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":3707}},"tokens_in":455,"tokens_out":3764,"duration_ms":29774,"temperature":1.0,"reasoning_tokens":3707,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:38:52.235870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a one-dimensional nonconvex error with two separated global minima, e.g. $q(\\beta)=(\\beta^2-1)^2$, form the finite-width objective (II.3), and enumerate all local minima for growing $N$ with $\\alpha$ close to 1. If any sequence of local minima stays a positive distance away from $\\operatorname{co}(\\operatorname{Argmin}(q))=\\{-1,1\\}$ as $N\\to\\infty$ and $\\alpha\\to 1$, the conclusion of Theorem 4.5 is false.","supporting_citations":[{"cited_title":"The convexification effect of Minkowski summation","cited_arxiv_id":null,"evidence_quote":"Supplies the convexification theorem for Minkowski sums of compact sets with O(1/n) rate, which is the core mechanism behind Corollary 3.3."},{"cited_title":"Convergence of Lebesgue integrals with varying measures","cited_arxiv_id":null,"evidence_quote":"Gives the convergence-of-integrals-with-varying-measures result used to pass from the finite weighted sum to the infinite weighted sum in the proof of Theorem 4.5."},{"cited_title":"Some comments on a theorem of Hardy and Littlewood","cited_arxiv_id":null,"evidence_quote":"Provides the Tauberian theorem equating the geometric-kernel weighted limit with the Cesàro average limit, enabling identification of the limit objective with the convex minorant."},{"cited_title":"Borkar, Stochastic Approximation: A Dynamical Systems Viewpoint (second edition), Hindustan Publishing Agency and Springer Nature, 2022/24","cited_arxiv_id":null,"evidence_quote":"Supplies the stochastic-approximation background used in the LMS section to argue that unstable local maxima are avoided almost surely by the SGD recursion."}],"review_version":2}