{"id":"fb8e5583-1e04-4525-8292-a0f3f2341f65","arxiv_id":"2607.08370","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Tube-volume bounds for smooth Pfaffian hypersurfaces yield condition-number tails for Pfaffian neural classifiers, with polynomial-in-width control for single-layer rational-weight sigmoids.","lead":"The paper bounds the volume of thin tubes around smooth Pfaffian hypersurfaces in terms of their format, then turns those bounds into tail estimates on a condition number that measures how easily a neural classifier can be flipped by small input noise. For single-hidden-layer sigmoid nets with rational first-layer weights the tube volume (and thus the tail) becomes only polynomial in width rather than exponential.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the ball-containment/smoothness hypotheses already identified; the BKK and reduced-chain arguments for the poly-in-w claims check out under those hyps.","rationale":"The reader correctly isolates the two modelling hypotheses that keep the exponential Khovanskii factor out of the single-layer rational-weight regime. The remainder of the argument—rational regular values + non-degeneracy via the implicit-function chart (Lemma 4.8), Bernstein mixed-volume on the zonotope of lattice vectors of height ≤ L, and the multiplicative-chart reduction that freezes the Pfaffian chain length at O(n) independent of width—is self-contained and free of circularity or hidden parameters. Because those steps hold whenever the hypotheses are met, and the paper already flags the sphere obstruction and the multi-layer conjecture, no further adjustment to the CONDITIONAL verdict is warranted. The concrete numerical check on the sharpness example would either corroborate the degree bounds or expose an arithmetic error in the volume estimate, settling the only remaining practical doubt.","tokens_in":35106,"tokens_out":678,"duration_ms":31793,"concrete_test":"Take the explicit compact grid construction of Prop. 4.11 (n=2, even m≥4, L=1, w=n(m+2)) whose mdeg is known to be Θ(w^n). Numerically estimate vol(T(V,ε) ∩ B(0,ρ)) for a sequence of widths and small ε/ρ; if the observed volume scales faster than any fixed power of w (or exceeds the explicit constant 2K(n,1)w^{2n}[(1+ε/ρ)^n-1]), the section-degree bound of Prop. 4.14 is false. Conversely, agreement within a moderate factor confirms the poly claim under the stated hyps.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (Prop. 4.9/4.14 + Thm. 4.15 + Cor. 5.8) delivers mdeg(V) = O(w^n) and mdeg_i(V) = O(w^{2n}) only when V is a smooth compact hypersurface contained in the sampling ball B(p, ρ). The containment is used to invoke the pure interior tube formula of Thm. 3.2 and avoid the codimension-2 sphere section ∂M = V ∩ S^{n-1}(p, ρ+ε). As Rem. 4.19 notes, the general complete-intersection bound (3.3) on that section re-introduces the full Khovanskii factor 2^{w(w-1)/2}. Smoothness (∇f ≠ 0 on V) is likewise indispensable for the Gauss-map degree to be well-defined and for Lemma 4.8. Both hypotheses are stated explicitly and the multi-layer/singular cases are correctly left open; the BKK zonotope count and the width-independent Pfaffian chain of length ≤ 2n in the multiplicative chart appear free of internal gaps.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper derives volume bounds for tubular neighbourhoods of smooth bounded Pfaffian hypersurfaces in terms of the Pfaffian format (α, β, s), via Khovanskii bounds on the degrees of the generalised Gauss map of generic affine sections (Prop. 3.4, Thm. 3.6). These are applied to neural-network classifiers with Pfaffian activations to obtain uniform and Gaussian tail bounds on a local condition number C_p(X) = ∥X−p∥/dist(X, Σ) measuring distance to the decision boundary (Thms. 5.3–5.4, Cor. 5.7). For single-hidden-layer logistic networks with rational first-layer weights of lattice constant L, a Bernstein–Kushnirenko–Khovanskii count after an exponential chart yields mdeg(V) ≤ C(n,L) w^n (Prop. 4.9); a multiplicative-chart argument with a width-independent Pfaffian chain of length O(n) then bounds all section degrees by K(n,L) w^{2n} (Prop. 4.14), giving a polynomial-in-width tube formula (Thm. 4.15) and an O(w^n/t) condition-number tail (Cor. 5.8, Rem. 5.9). Multi-layer and singular cases are left as conjectures.","tokens_in":35360,"tokens_out":1238,"duration_ms":10473,"significance":"The work cleanly extends the algebraic tube-volume programme of Lotz and of Basu–Lerario to the Pfaffian setting that naturally contains sigmoid, tanh and related activations, and converts those bounds into explicit condition-number tails for neural classifiers. The single-layer rational-weight results are the strongest contribution: they replace the exponential Khovanskii factor 2^{w(w−1)/2} by a polynomial of degree 2n (with sharp leading order w^n for the top Gauss degree, Prop. 4.11), under explicitly stated hypotheses. All constants are expressed in terms of format or lattice data; the derivations are fully written and the limitations (ball containment, smoothness) are recorded honestly. This is a solid, technically careful contribution at the interface of real algebraic geometry and the geometric analysis of neural networks.","major_comments":[{"comment":"Thm. 4.15 and Cor. 5.8 require every pairwise decision boundary V_ij to lie inside the sampling ball B(p, ρ). Rem. 4.19 correctly notes that without this inclusion the sphere-boundary term ∂M = V ∩ S^{n−1}(p, ρ+ε) re-introduces the full Khovanskii factor 2^{w(w−1)/2} via the complete-intersection bound (3.3). The hypothesis is load-bearing for the polynomial-width claim that is the paper’s main selling point; the abstract and introduction should state it as prominently as the polynomial bound itself, and the Gaussian hybrid (Prop. 5.10) should be flagged as recovering only a hybrid (not fully polynomial) rate at fixed Gaussian scale (Rem. 5.11).","section":null},{"comment":"Smoothness (∇f never vanishes on V) is indispensable for the Gauss-map degree to be well-defined and for the non-degeneracy lemma (Lem. 4.8) that justifies the BKK count. Decision boundaries of sigmoid networks can develop singularities for generic weights; the paper correctly excludes them by hypothesis and leaves the singular case open (§6.2). A short discussion of how restrictive this is in practice (or a pointer to the algebraic deformation strategy of Basu–Lerario and why it does not transfer) would strengthen the claims of applicability.","section":null}],"minor_comments":[{"comment":"Abstract and first paragraph of the introduction: the phrase “polynomial-in-width bounds for tubular neighbourhoods of the decision boundary” should be qualified by the ball-containment and smoothness hypotheses that make Thm. 4.15 possible.","section":null},{"comment":"Notation: the same symbol σ is used for the logistic sigmoid and for the Gaussian standard deviation (e.g. Thm. 5.4, Prop. 5.10). A brief local reminder or a different letter for the variance would avoid momentary confusion.","section":null},{"comment":"Prop. 4.9: the constant is stated as C(n,L) ≤ 2·n!(2L)^n, yet the BKK count already yields n!(2L)^n w^n; the extra factor 2 is harmless but could be tightened or explained.","section":null},{"comment":"Example 5.12: the explicit multi-layer constant 6·2^{h(h−1)/2}(n(4ℓ+1)+2)^h is useful; a one-line comparison with the single-layer polynomial of Cor. 5.8 would help the reader appreciate the improvement.","section":null},{"comment":"Typographical: “Khovanskii’s theorem [Kho91], Theorem 2.10” (p. 1) and a few similar double citations; also “the (1−δ)-quantile” in Rem. 5.2 could be written more carefully.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a natural fit for a geometry or applied-algebra journal that already publishes tube formulae and condition-number analysis. The single-layer BKK argument is the part most likely to be cited; the multi-layer conjectures are appropriately left open. No citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new pieces are a clean Khovanskii-based tube formula for smooth Pfaffian hypersurfaces (Thm 3.6) and, more interestingly, a BKK-plus-multiplicative-chart argument that replaces the exponential Khovanskii factor by O(w^{2n}) for single-hidden-layer logistic networks with rational first-layer weights (Props 4.9/4.14, Thm 4.15, Cor 5.8). That is the result worth remembering: mdeg(V) = O(w^n) and section degrees O(w^{2n}), giving an O(w^n/t) condition-number tail once the decision boundary sits inside the sampling ball.\n\nThey do the work carefully. The non-degeneracy lemma for the rational Gauss-map fibre (Lem 4.8) is needed and is proved; the exponential chart turns the sigmoids into Laurent polynomials whose Newton polytope is a zonotope of volume O((2L w)^n); the lower section degrees are handled by a short Pfaffian chain of length O(n) independent of width. The multi-layer and singular cases are left as conjectures rather than papered over. Citations to Lotz, Khovanskii, BKK, and the neural-complexity literature are accurate and not padded.\n\nThe soft spots are exactly the ones the authors name. Thm 4.15 and Cor 5.8 require V subset B(p,rho) so that the sphere-boundary term never appears; without it the complete-intersection bound re-introduces 2^{w(w-1)/2} (Rem 4.19). Smoothness (nabla f never zero on V) is also load-bearing for the Gauss map. Both are stated up front, and the Gaussian hybrid (Prop 5.10) still carries a residual Khovanskii piece multiplied by a concentration factor. These are modelling restrictions, not hidden gaps in the proofs that are given.\n\nThis is for people who already care about geometric complexity of classifiers or fewnomial bounds; it will not move the average ML practitioner. The math is self-contained and the strongest claim holds under the stated hypotheses. I would send it to referees without hesitation.","headline":"Solid Pfaffian tube formula plus a genuine poly-in-width improvement for shallow rational sigmoids; the ball-containment and smoothness hyps are real but already flagged by the authors.","tokens_in":36034,"tokens_out":563,"would_cite":true,"duration_ms":7104,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["14P10","68T07","52A39","14P15"],"pacs":[],"model":"grok-4.5","headline":"Tube volumes around single-layer sigmoid decision boundaries grow only polynomially in network width, giving O(w^n/t) condition-number tails.","keywords":["Pfaffian functions","tubular neighbourhoods","Gauss map degree","neural network classifiers","condition number","sigmoid networks","Bernstein–Kushnirenko–Khovanskii","fewnomials"],"falsifier":"Construct a single-hidden-layer logistic network of width w with rational weights of fixed lattice constant L whose zero set is a smooth compact hypersurface whose Gauss-map degree grows faster than any constant times w^n, or whose ε-tube volume inside a containing ball exceeds the claimed polynomial bound for large w.","tokens_in":35965,"feed_emoji":"📏","tokens_out":809,"duration_ms":7174,"temperature":0.7,"pith_summary":"The paper bounds how much volume sits inside an ε-tube around a smooth hypersurface defined by Pfaffian functions—functions that obey triangular first-order PDEs with polynomial coefficients, including common neural activations such as sigmoid and tanh. Those volume bounds are stated purely in terms of the Pfaffian format of the defining function and recover classical algebraic tube estimates when the format is purely polynomial. Applied to neural classifiers, the same bounds become tail estimates for a condition number that measures relative distance to the decision boundary: large condition number means a small relative perturbation can flip the predicted class. For generic multi-layer Pfaffian nets the resulting constants still carry an exponential factor in the number of hidden units. The main improvement is for single-hidden-layer logistic networks whose first-layer weights are rational with bounded denominator: after an exponential substitution the Gauss-map system becomes a Laurent system whose Bernstein volume is polynomial in width, and the tube probability and condition-number tails therefore become polynomial rather than exponential in width.","feed_headline":"Sigmoid decision tubes grow only as width to the n","feed_subtitle":"Polynomial tube volumes give O(w^n/t) robustness tails for single-layer rational nets","key_machinery":"The maximal degree of the generalised Gauss map of a hypersurface (and of its generic affine sections). Bounding that degree by Khovanskii’s fewnomial count (general Pfaffian case) or by a Bernstein–Kushnirenko–Khovanskii volume after an exponential substitution (rational-weight single-layer case) converts classical integral-geometry tube formulae into explicit volume and condition-number estimates.","core_discovery":"For a smooth compact hypersurface V = Z(f) given by a single-hidden-layer logistic network with rational first-layer weights of lattice constant L, every section degree of the Gauss map is at most K(n,L) w^{2n}. Consequently the uniform probability that a random point in a ball lies within distance ε of V is at most 2 K(n,L) w^{2n} [(1+ε/ρ)^n − 1], and the local condition-number tail decays as O(w^n / t).","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Sigmoid decision tubes scale polynomially as width to n","Pfaffian tube volumes yield O(w^n/t) net robustness tails","Logistic hypersurface Gauss degrees bounded by K w^{2n}","Single-layer rational nets: decision tubes grow as w^{2n}","Polynomial tube bounds control sigmoid classifier condition tails"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"Every pairwise decision boundary must sit entirely inside the ball on which the data are sampled, and the defining gradient must never vanish on that boundary; otherwise an exponential factor reappears or the surface is no longer smooth.","fun_headline_variants_meta":{"raw":{"variants":["Sigmoid decision tubes scale polynomially as width to n","Pfaffian tube volumes yield O(w^n/t) net robustness tails","Logistic hypersurface Gauss degrees bounded by K w^{2n}","Single-layer rational nets: decision tubes grow as w^{2n}","Polynomial tube bounds control sigmoid classifier condition tails"]},"model":"grok-4.5","effort":"low","cost_usd":0.00395,"raw_usage":{"total_tokens":1177,"prompt_tokens":682,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":39500000,"prompt_tokens_details":{"text_tokens":682,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":420,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":682,"tokens_out":75,"duration_ms":4827,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T08:34:32.122277+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct a single-hidden-layer logistic network of width w with rational weights of fixed lattice constant L whose zero set is a smooth compact hypersurface whose Gauss-map degree grows faster than any constant times w^n, or whose ε-tube volume inside a containing ball exceeds the claimed polynomial bound for large w.","supporting_citations":[],"review_version":1}