{"id":"1e5588f2-88da-4fe8-a2d9-70d6e179cac1","arxiv_id":"2608.03482","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors introduce a q-Hermite I polynomial SVM kernel, prove it is a valid Mercer kernel, and show it matches Gegenbauer accuracy with lower training time, but the no-scaling claim is weakened by a clipped weight.","lead":"This paper builds a new SVM kernel from discrete q-Hermite I polynomials, a q-deformed relative of Hermite polynomials, and tests it on 20 datasets. The kernel is competitive with classical and other polynomial kernels, and faster to train than the Gegenbauer kernel, though the boundedness proof has gaps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's uniform bound is false as stated (n=1, q near 1 violates it), so the no-explosion premise fails; the weight's exponentially small lower bound also forces the very clipping the paper claims to avoid.","rationale":"Read in good faith: the Mercer proof for the unclipped kernel is valid (a finite sum of squares with a positive product weight), and the 20-dataset comparison is extensive and honestly reports accuracy close to Gegenbauer with faster training. The fragile point is the advertised numerical-stability/no-scaling advantage. The stated uniform-bound theorem is demonstrably false; it cannot be the basis for the claim. Even if a correct sup-norm bound for H_n were supplied, the weight lower bound is exponentially small near the endpoints for q close to 1, so the unclipped kernel suffers exactly the annihilation the paper says it avoids. The paper's own clipping acknowledges this. The appropriate fix is to amend or remove the no-scaling claim, prove the clipped kernel is PSD (or treat clipping as part of the kernel definition), and resubmit; the empirical contribution alone may survive. This matches the reader's CONDITIONAL verdict, so no adjustment is needed.","tokens_in":21751,"tokens_out":18193,"duration_ms":208314,"concrete_test":"Evaluate Proposition 1 at n=1 for q=0.99: LHS=1, RHS=2√q·∏_{m≥2}(1-q^m); verify RHS<1. Then compute the unclipped kernel (Eq. 14) at x=z=1 with degree N=1 and q=0.99; if the value is below 1e-10 while the clipped implementation returns ≈0.01, the no-scaling/no-annihilation claim fails without the clipping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that K_Q-HERMITE needs no scaling rests on Propositions 1 and 2, but Proposition 1 is false as stated. For n=1 the LHS sup_{[-1,1]} |H_1| = 1 (since H_1(x)=x), while the RHS of Eq. 15 is 2·((q;q)_∞/(q;q)_1)·q^{1/2} = 2√q ∏_{m≥2}(1-q^m), which is <1 for every q∈(0,1); at q=0.99 it is <10^{-50}. The proof's line 'C(n) ≤ sup |H_n|' has the inequality reversed, and M_1=2 is wrong for H_1. Thus Eq. 15 cannot support the 'no explosion' part of the claim. Proposition 2 only proves 0<w≤1, not a useful lower bound: at x=z=1, q=0.99, w_{-1,q}(1)=(q^2;q^2)_∞≈10^{-22}, so an unclipped kernel entry is ~10^{-44}, squarely in the annihilation regime the paper claims to avoid. Section 3.2.1 then clips w<0.1 to 0.1, reintroducing the epsilon correction used by Padierna et al. Hence the advertised 'without explicit scaling' advantage is unsupported and contradicted by the implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new SVM kernel, K_Q-HERMITE (Eq. 14), built from discrete q-Hermite I polynomials with a=-1 and q in (0,1). It claims Mercer validity, inherent boundedness that makes explicit scaling unnecessary, and reports experiments on 20 datasets showing accuracy competitive with Gegenbauer and faster training times. The Mercer/PSD argument in Section 4 is correct, and the experimental pipeline is substantial. However, the theoretical anchor for the no-scaling claim, Proposition 1, is false as stated, and the weight's lack of a useful lower bound plus the paper's own clipping undermines the advertised advantage. These are fixable by revising the proofs and claims; the kernel construction and empirical comparison remain of interest.","tokens_in":22122,"tokens_out":8424,"duration_ms":104339,"significance":"If the kernel construction and experimental results are taken at face value, the paper contributes a new orthogonal-polynomial kernel with a tunable deformation parameter q and a clean factorization that makes positive semidefiniteness immediate. The implementation is reproducible, and the comparison methodology is more careful than many kernel papers (35 repetitions, 10-fold CV during optimization, Bayesian hyperparameter search, open-source pipeline). The measured training-time advantage over Gegenbauer is a genuine empirical finding. But the advertised absence of scaling is load-bearing and is not established: Proposition 1 is false, and Section 3.2.1 reintroduces clipping. With a corrected argument and a more modest statement of the stability claim, the contribution would be publishable; with the current claims it is not.","major_comments":[{"comment":"The stated uniform bound is false. For n=1, the LHS is sup_{[-1,1]} |H_1(x;q)| = sup_{[-1,1]} |x| = 1, while the RHS is 2*sqrt(q)*(q;q)_inf/(q;q)_1 = 2*sqrt(q)*prod_{m>=2}(1-q^m), which is <1 for every q in (0,1) and is astronomically small for q=.99. The proof's line 'C(n) <= sup |H_n|' has the inequality reversed: the recurrence gives sup |H_n| <= C(n), not >=. Moreover M_1=2 is wrong for H_1(x)=x, whose sup is 1. The final normalization identity does not repair the bound. Since the no-explosion claim in Sections 3.2.1, the abstract, and the conclusions rests on this proposition, this is a load-bearing error.","section":"Section 3.2, Proposition 1 (Eq. 15)"},{"comment":"Proposition 2 establishes only 0<w<=1; it gives no positive lower bound. At x=z=1 and q=.99, w_{-1,q}(1,1) = ((q^2;q^2)_inf)^2 is about 10^-44, so raw kernel entries are exponentially small. The paper then clips w<0.1 to 0.1, which is exactly the epsilon-regularization used for the Gegenbauer kernel in Eq. (3). This contradicts the sentence 'such a correction is not required' and the abstract's claim that no explicit scaling mechanisms are needed. Even if clipping is described as purely practical, it is part of the implementation, and the no-scaling advantage is not supported by the stated results.","section":"Section 3.2.1, Proposition 2 and clipping"},{"comment":"There is a mismatch between the mathematical definition and the implementation. Section 3.2.1 says that whenever w<0.1, it is clipped to 0.1, but Eq. (14) and Algorithm 1 contain no clipping step. If the experiments use the raw weight, boundary Gram entries can be of order 10^-44, raising numerical concerns. If they use the clipped weight, then the evaluated kernel is not the kernel of Eq. (14) and the no-scaling statement is void. The paper should state exactly which kernel is used in the experiments and report sensitivity to the clipping threshold and to the Pochhammer truncation bound theta.","section":"Section 3.2.1 vs. Algorithm 1"}],"minor_comments":[{"comment":"Proposition 2 appears twice with two different proofs; consolidate into one statement and one proof.","section":"Section 3.2"},{"comment":"Typo: 'ilustrates' should be 'illustrates'. The algorithm also does not show the clipping described in Section 3.2.1.","section":"Algorithm 1"},{"comment":"The weight function is cited to [34], which is a q-difference-equations paper, not a kernel paper. Since the author lists overlap, please identify the exact formula being borrowed and state explicitly that no kernel result is imported from [34].","section":"Eq. (11) and references"},{"comment":"The paper acknowledges that epsilon=0.1 and theta=10 are not optimized and that no sensitivity analysis is given. A short experiment varying these two parameters would strengthen the numerical-stability claims.","section":"Section 6.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the provenance of the weight and polynomial family is reference [34], whose authors overlap with the present paper. I do not see circularity in the Mercer proof, but the paper should make the provenance explicit. The main risk is the overclaiming in the abstract and conclusions about eliminating scaling; after revision I would want the no-scaling claim either removed or replaced by a concrete lower-bound condition, and the clipping implementation clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked for a quick read of the q-Hermite SVM kernel paper. The short version: give the authors a chance to revise, but don't accept the current boundedness story. The kernel itself is a genuine novelty — I don't know of a prior SVM kernel built from discrete q-Hermite I polynomials — and the Mercer proof in Section 4 is correct: the scalar kernel factorizes as w(x)w(z) times a finite sum of polynomial products, so the double integral is a sum of squares. The weight function is taken from the authors' own earlier math paper, but that paper isn't about kernels, and the PSD argument here is independently derived. The code and experimental pipeline are public, and the empirical work is unusually honest: accuracy differences vs. Gegenbauer are not significant (Wilcoxon p=0.87), training time is genuinely faster than Gegenbauer on most datasets, and the authors openly discuss the support-vector-proportion results that don't match Padierna's claims.\n\nThe weak spot is Proposition 1. The claimed uniform bound on the polynomials is false as stated. The proof uses M1=2, but H1(x)=x has sup 1 on [-1,1], and the line 'C(n) ≤ sup |H_n|' is backwards. The stress-test computation for n=1 gives the RHS below 1 for every q in (0,1), so the proposition simply cannot support the \"no explosion\" claim. What's worse, the paper's own Section 3.2.1 clips w below 0.1 to 0.1 — exactly the epsilon correction the paper claims to avoid. At x=z=1, q=0.99, the unclipped weight is ~10^-22, so kernel values would annihilate. The abstract and conclusions repeat the \"inherent boundedness prevents annihilation and explosion\" claim, but it is unsupported.\n\nThat said, the kernel is still a valid Mercer kernel, and it may be practically useful. The PSD proof does not depend on Proposition 1. The experiments suggest competitive accuracy with a training-time advantage. The paper just needs to either fix the bound (if a correct one exists) or drop the no-scaling claim and treat the clipping as part of the kernel definition. The duplicated Proposition 2 should also be merged, and the second proof's domain statement (\"0≤x,z≤1\") should be corrected to [-1,1].\n\nThis paper deserves a serious referee. The kernel is novel, reproducible, and the empirical design is careful. I'd bring it to a reading group, and I'd cite it in my own kernel-methods work. It just isn't ready without revision of the theoretical claims.","headline":"A useful new kernel with honest experiments, but the no-scaling claim rests on a false bound and the paper's own clipping; fix Proposition 1 and the framing.","tokens_in":22634,"tokens_out":2839,"would_cite":true,"duration_ms":31447,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["33D45","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that kernels built from discrete q-Hermite I polynomials need no scaling to stay bounded on [-1,1], making them valid Mercer kernels that match Gegenbauer accuracy while training faster.","keywords":["support vector machines","kernel methods","orthogonal polynomial kernels","q-Hermite polynomials","Mercer kernels","q-calculus","numerical stability","kernel design"],"falsifier":"Compute sup_{x in [-1,1]} |H_n(x;q)| directly from the recurrence for q = 0.99 and q = 0.999 and n up to a few hundred; if these suprema grow with n instead of staying bounded, the uniform-boundedness premise fails and the no-scaling claim collapses. A one-line spot check: the proof asserts sup|H_1| = 2 on [-1,1], but H_1(x) = x has supremum 1, so the stated bound cannot be correct as written.","tokens_in":21678,"feed_emoji":"📈","tokens_out":13623,"duration_ms":135753,"temperature":0.7,"pith_summary":"This paper claims that a kernel built from discrete q-Hermite I polynomials, a q-deformed relative of Hermite polynomials indexed by a deformation parameter q in (0,1), is a valid Mercer kernel—symmetric, continuous, and positive semidefinite—whose terms stay bounded on [-1,1] by themselves. If true, that boundedness removes the need for the explicit scaling function that the Gegenbauer kernel construction requires to stop kernel values from vanishing (annihilation) or blowing up (explosion). The paper proves the weight function lies strictly between 0 and 1, proves positive semidefiniteness by writing the kernel as a sum of squared integrals, and reports experiments on 20 datasets where the new kernel ranks third in accuracy behind Gegenbauer and polynomial kernels, beats Gegenbauer on 9 of 20 datasets, and trains faster than Gegenbauer with the speed advantage growing with dimensionality. The payoff, if the claim holds, is a numerically stable SVM kernel whose single parameter q tunes behavior between a classical-Hermite-like regime and strongly deformed regimes.","feed_headline":"Tame SVM blow-up with a q-deformed Hermite kernel","feed_subtitle":"It rivals the best orthogonal-polynomial kernels on 20 datasets, and its speed edge grows with dimensionality.","key_machinery":"The central object is the discrete q-Hermite I polynomial family H_n(x;q) with parameter a = -1, generated by the three-term recurrence xH_n = H_{n+1} + q^{n-1}(1 - q^n)H_{n-1} and supported on [-1,1]. Two properties carry the argument: a claimed uniform bound on these polynomials over [-1,1] (a Fibonacci-type growth constant times (q;q)_∞/(q;q)_n q^{n/2}), and the bivariate q-Pochhammer weight w_{-1,q}(x,z), which is confined to (0,1] through the identity (qx,-qx;q)_∞ = (q²x²;q²)_∞, a product of factors 1 - q^{2m}x². Together they make the kernel expansion bounded term by term, which is what lets the construction skip the scaling function that earlier Gegenbauer-based kernels require. The k","core_discovery":"On the paper's own terms, the discovery is that the discrete q-Hermite I polynomials specialized at a = -1 are uniformly bounded on [-1,1], and their associated weight function w_{-1,q}(x,z) = (qx,-qx;q)_∞ (qz,-qz;q)_∞ takes values strictly inside (0,1]. Together these two facts guarantee that the kernel K_Q-HERMITE(x,z) = ∏_j ∑_i H_i(x_j;q)H_i(z_j;q) w_{-1,q}(x_j,z_j) neither annihilates nor explodes, eliminating the scaling function the Gegenbauer formulation needs. The kernel is shown to be a Mercer kernel: its scalar form is a sum of squares of integrals ∫ H_i(x;q)(qx,-qx;q)_∞ g(x) dx, so it is positive semidefinite, and products of positive semidefinite kernels extend the construction t","pith_inferences":["The no-scaling claim rests on the uniform bound in Proposition 1, and as printed that proof contains a reversed inequality (it states C(n) ≤ sup|H_n| where the comparison yields the opposite direction) and sets sup|H_1| = 2 although H_1(x) = x has supremum 1 on [-1,1]; my inference is that uniform boundedness is not actually established for q near 1, and the stabilization claim needs a corrected p","The paper does not analyze how q affects generalization; a natural testable hypothesis, my own, is that q acts like a bandwidth or smoothness parameter analogous to the RBF kernel's γ, with a monotone effect on margin and accuracy that a simple sweep over q could reveal.","Because the kernel factorizes per dimension and its feature map is explicit, the mapped features could be plugged into linear models or used without forming the Gram matrix in moderate dimension—a use the paper does not explore.","The paper reports a dissociation between support-vector proportion and accuracy (e.g., high accuracy with high PSV on ionosphere and fertility); my inference is that support-vector counts should not be treated as a reliable complexity or performance proxy for this kernel family, despite their use in earlier orthogonal-polynomial kernel studies."],"forward_implications":["SVM users gain a valid Mercer kernel that needs no scaling function, removing the annihilation and explosion failure modes that the Gegenbauer construction had to patch with a scaling map and a weight offset.","The single parameter q tunes the polynomial family continuously from a classical-Hermite-like regime (q approaching 1) to strongly deformed regimes, giving one formula a tunable kernel family.","The training-time advantage over the Gegenbauer kernel grows with feature count (the paper reports r = 0.479, p = 0.0325), so high-dimensional tabular problems are where the kernel pays off most.","The weight function can be truncated to a small number of q-Pochhammer factors at negligible cost, keeping the kernel cheap to evaluate while preserving its stability properties.","As a Mercer kernel with an explicit feature map, the construction is offered by the authors as a natural basis for quantum support-vector machines, where q-structures already arise."],"supporting_citations":[{"why":"Supplies Mercer's theorem, the positive-semidefiniteness criterion the q-Hermite kernel must satisfy.","marker":"[3]"},{"why":"Provides the earlier Hermite orthogonal-polynomial kernel that this work generalizes.","marker":"[20]"},{"why":"Defines the Gegenbauer kernel framework being extended, including the scaling function and weight offset the q-Hermite construction claims to remove, and supplies the experimental baseline.","marker":"[21]"},{"why":"Provides the Askey-scheme classification of q-orthogonal polynomials and the limiting relations used to place q-Hermite polynomials.","marker":"[23]"},{"why":"Supplies the basic-hypergeometric identities (q-Pochhammer, q-binomial) used throughout the construction.","marker":"[32]"},{"why":"The source of the q-weight function w_{a,q}(x) adopted for the kernel.","marker":"[34]"},{"why":"Establishes the discrete q-Hermite I (Al-Salam-Carlitz) polynomial family used as the basis.","marker":"[35]"},{"why":"Gives the asymptotic result that the polynomials converge to classical Hermite polynomials as q approaches 1.","marker":"[36]"},{"why":"Justifies forming the bivariate weight as a product of univariate weights.","marker":"[37]"}],"fun_headline_variants":["q-deformed Hermite kernel fixes SVM blow-up","New kernel kills SVM explosion, keeps accuracy","SVM kernel with built-in stability: q-Hermite","q-Hermite kernel: stable SVMs without scaling","Replace Gegenbauer with q-Hermite for robust SVMs"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the discrete q-Hermite I polynomials are uniformly bounded on [-1,1] for every q in (0,1); the proof of that bound contains a reversed inequality and a wrong supremum for the first polynomial, so the premise is not established as printed for q close to 1, and without it the kernel's no-scaling stability claim unravels.","fun_headline_variants_meta":{"raw":{"variants":["q-deformed Hermite kernel fixes SVM blow-up","New kernel kills SVM explosion, keeps accuracy","SVM kernel with built-in stability: q-Hermite","q-Hermite kernel: stable SVMs without scaling","Replace Gegenbauer with q-Hermite for robust SVMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3185,"prompt_tokens":816,"completion_tokens":2369,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2290}},"tokens_in":560,"tokens_out":2369,"duration_ms":18395,"temperature":1.0,"reasoning_tokens":2290,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:15:47.211590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute sup_{x in [-1,1]} |H_n(x;q)| directly from the recurrence for q = 0.99 and q = 0.999 and n up to a few hundred; if these suprema grow with n instead of staying bounded, the uniform-boundedness premise fails and the no-scaling claim collapses. A one-line spot check: the proof asserts sup|H_1| = 2 on [-1,1], but H_1(x) = x has supremum 1, so the stated bound cannot be correct as written.","supporting_citations":[{"cited_title":"On discrete orthogonal polynomials of several variables.Advances in Applied Mathematics, 33(3):615–632, 2004","cited_arxiv_id":null,"evidence_quote":"Justifies forming the bivariate weight as a product of univariate weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Mercer's theorem, the positive-semidefiniteness criterion the q-Hermite kernel must satisfy."},{"cited_title":"New hermite orthogonal poly- nomial kernel and combined kernels in support vector machine classifier.Pattern Recog- nition, 60:921–935, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the earlier Hermite orthogonal-polynomial kernel that this work generalizes."},{"cited_title":"A novel formulation of orthogonal polynomial kernel functions for svm classifiers: The gegenbauer family.Pattern Recognition, 84:211–225, 2018","cited_arxiv_id":null,"evidence_quote":"Defines the Gegenbauer kernel framework being extended, including the scaling function and weight offset the q-Hermite construction claims to remove, and supplies the experimental baseline."},{"cited_title":"Hypergeometric orthogonal polynomials","cited_arxiv_id":null,"evidence_quote":"Provides the Askey-scheme classification of q-orthogonal polynomials and the limiting relations used to place q-Hermite polynomials."},{"cited_title":"Cambridge university press, 2011","cited_arxiv_id":null,"evidence_quote":"Supplies the basic-hypergeometric identities (q-Pochhammer, q-binomial) used throughout the construction."},{"cited_title":"Onsecond order q-difference equations satisfied by al-salam–carlitz i-sobolev type polynomials of higher order.Mathematics, 8(8):1300, 2020","cited_arxiv_id":null,"evidence_quote":"The source of the q-weight function w_{a,q}(x) adopted for the kernel."},{"cited_title":"On combinatorics of al-salam carlitz polynomials.European Journal of Combinatorics, 18(3):295–302, 1997","cited_arxiv_id":null,"evidence_quote":"Establishes the discrete q-Hermite I (Al-Salam-Carlitz) polynomial family used as the basis."},{"cited_title":"Multivariable al–salam & carlitz polynomials associated with the type a q–dunkl kernel.Mathematische Nachrichten, 212(1):5–35, 2000","cited_arxiv_id":null,"evidence_quote":"Gives the asymptotic result that the polynomials converge to classical Hermite polynomials as q approaches 1."}],"review_version":1}