Pith. sign in

REVIEW 2 major objections 4 minor 28 references

This paper proves that the excess risk of the diagonal Wiener-Hermite cross-correlation rule under non-Gaussian input is exactly the quadratic form ||(I−C^T D^{-1}C)b||_2^2, computable in O(s^3) from raw moments to order 2s, with no simulat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 01:20 UTC pith:GV52Y64X

load-bearing objection Correct and honest: the exact mismatch penalty formula holds up; the real caveat is high-order moment estimation, which the paper discloses. the 2 major comments →

arxiv 2607.14699 v1 pith:GV52Y64X submitted 2026-07-16 stat.ME

Exact Computation of Non-Gaussian Mismatch Penalties in Wiener-Hermite Cross-Correlation Identification

Serhii Zabolotnii This is my paper
classification stat.ME MSC 65F3565C6093E1242C05
keywords Wiener-Hermite identificationnon-Gaussian inputmismatch penaltyHankel-Cholesky algorithmcumulant formsbasis selectiondiagonal cross-correlationprojection normal system
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper establishes that when a Wiener-Hermite polynomial basis is used to identify a nonlinear system under non-Gaussian excitation, the price of not matching the basis to the true input law is an exact, finite-order quantity. For a response in the span of polynomials up to degree s, that excess risk is a quadratic form in the response coefficients involving the cross-Gram matrix between the Gaussian Wiener basis and the law-matched orthonormal basis. The formula can be evaluated from raw moments up to order 2s with two Cholesky factorizations and one diagonal solve, in O(s^3) operations, with no simulation, quadrature, or asymptotics. Closed cumulant expressions at orders 3 and 4 show which non-Gaussian features trigger the penalty, and a bootstrap-based plug-in diagnostic tells a practitioner when the Gaussian basis is adequate. The practical caveat is that the exactness is conditional on knowing the moments: under heavy contamination the empirical plug-in has high variance, so moment estimability, not the algebra, is the load-bearing premise for applications.

Core claim

The central claim is that the excess L2(P) risk of the diagonal Wiener-Hermite rule equals ||(I−C^T D^{-1}C)b||_2^2, where C is the matrix of inner products of the Wiener basis with the P-matched orthonormal polynomial basis, D is its diagonal, and b are the signal coefficients in the matched basis. Because the signal lies in the polynomial span, this is not truncation error; it is the exact cost of using a Gaussian basis under a non-Gaussian law. Symmetric inputs have zero penalty at order 2 but a positive penalty at order 3 once excess kurtosis is present; skewness activates the penalty already at order 2. The paper verifies the formula across laws and orders by independent numerical paths

What carries the argument

The mismatch penalty W(s;b,P) = ||(I−C^T D^{-1}C)b||_2^2, obtained from the cross-Gram matrix C_{ki} = <η_k, ψ_i>_P between the variance-matched Gaussian Hermite basis η and the input-law-matched orthonormal polynomial basis ψ; it is evaluated from two Hankel moment matrices through Cholesky factors and a diagonal solve. An off-diagonal Gram certificate, the spectral mass ε of the correlation-normalized Gram matrix, bounds the penalty per unit signal energy by (1+ε)/(1−ε)ε^2 when the Gaussian basis is nearly P-orthogonal. Closed block forms at orders 3 and 4, expressed through standardized cumulants, identify exactly which cumulant defects the Gaussian basis fails to annihilate.

Load-bearing premise

The central formula is exact only given the raw moments of the input law through order 2s; in data applications those moments are estimated, and for heavy-tailed laws the estimated penalty is unstable—the paper itself reports that at n=2000 the contaminated-normal plug-in is within 10% of the population value in only 12% of replications, so moment estimability, not the algebra, is the load-bearing premise.

What would settle it

Generate data from a heavy-tailed law with known population moments, compute the plug-in penalty and its bootstrap interval, and compare with the formula evaluated at the true moments; if the intervals rarely contain the population value or the width does not flag the instability, the diagnostic premise fails. In the other direction, a symmetric three-point input law at order 3 has singular Hankel moment matrix, making W(3) undefined rather than positive—a direct test of the support hypothesis behind the order-threshold corollary.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The diagonal rule's finite-sample identification error converges to the population penalty as n grows, so the gap between diagonal and matched rules is an unavoidable floor, not finite-sample noise.
  • A practitioner with reliable moments to order 2s can decide without simulation whether to build a matched basis: the penalty is small whenever the Wiener family's off-diagonal Gram correlations are small.
  • For symmetric inputs, Gaussian orthogonality protects only up to order two; any nonzero excess kurtosis makes the mismatch penalty positive at order three, so basis matching is needed even for symmetric non-Gaussian laws.
  • The plug-in diagnostic's bootstrap intervals separate resolvable from unresolvable penalties: at n≈2000, moderate-tail laws give reliable intervals, while heavy-tailed laws produce wide intervals that warn against trusting the point estimate.
  • The same projection core reproduces estimation, detection, classification, characteristic-function distance, and identification functionals, so the normal-system-to-projection correspondence is one shared object across those branches.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit: because the penalty is a smooth function of the moments, it could be used as a data-driven basis-selection criterion, with matching triggered when the estimated penalty crosses a user-specified cost threshold; the paper sketches this as future sequential-design work.
  • For multivariate inputs with independent coordinates, the paper notes the penalty remains a single quadratic form in the coefficient tensor and does not decompose into a sum of per-coordinate penalties, so practitioners should not budget basis-matching effort coordinate by coordinate.
  • For dependent coordinates, mixed cumulants such as co-skewness and co-kurtosis can create a penalty at order 3 even when every marginal is symmetric; the paper flags this but leaves the multivariate cumulant bookkeeping to future work.
  • The exact cumulant forms imply that non-Gaussianity is not one scalar: penalties depend on ratios such as γ3^2/(2+γ4) and γ4^2/(6+9γ4+γ6), so diagnostics built on a single skewness or kurtosis summary can miss the relevant defect.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper derives an exact finite-order expression for the excess L2(P) risk incurred by the diagonal Wiener–Hermite cross-correlation rule when the input law P is non-Gaussian. The main result, Theorem 2.1, states that for f in the polynomial span, this penalty equals ||(I - C^T D^{-1} C)b||_2^2, where C is the cross-Gram matrix between the Gaussian Wiener basis and the P-matched orthonormal polynomial basis, and D is the diagonal of C C^T. Algorithm 1 computes this in O(s^3) from raw moments up to order 2s via two Hankel–Cholesky factorizations. The paper also gives closed cumulant forms at orders 3 and 4, a plug-in basis-selection diagnostic with bootstrap calibration, a unifying projection calculus covering PMM, GMM, detection, classification, and characteristic-function methods, and a Lean 4 machine-checked proof of the underlying normal-system–projection correspondence.

Significance. If the results are correct, the paper provides a useful exact substitute for simulation-based assessments of Wiener–Hermite basis mismatch: the penalty is a computable moment quadratic form rather than an asymptotic or Monte Carlo estimate. Strengths include the self-contained proof of Theorem 2.1, the reproducible three-path numerical verification, the honest disclosure of heavy-tail plug-in instability and the single real-data chain, and the machine-checked Lean module for the elementary projection core. The main reservation is that the displayed even-block closed form in Theorem 2.4(ii) is, as stated, incorrect; this affects an advertised contribution and the derivations that rely on it, although it does not undermine the central matrix formula.

major comments (2)
  1. [Theorem 2.4(ii) and Appendix A.3] The displayed even-block formula is incorrect. For the uniform law on [-sqrt(3), sqrt(3)] and signal b=(1,0,0), direct evaluation of (2.2) gives W_even(4) = c0^2 / ||h4||_P^2 = 1.44 / (1656/35) ≈ 0.127, while the displayed formula gives c0^2 / d2 = 1.8. The correct expression is W_even = [(c0 b0 + c2 b2 / sqrt(d2))^2 + b4^2 (c0^2 + c2^2/d2)] / ||h4||_P^2 (equivalently, in monic {1,h2} coordinates, [(c0 b0 + c2 b2)^2 + b4^2 (c0^2 + c2^2/d2)] / ||h4||_P^2). The proof's claim that eta0 = psi0 = 1 and h2 proportional to psi2 make the constant and quadratic coefficients exact is true only for the scalar coefficient estimates; the h4 term contaminates the reconstruction through c0 and c2. The printed formula should be corrected or removed.
  2. [Appendix A.3 and Section 5 verification] The paper states that the structured form of Theorem 2.4(ii) was verified symbolically and that the blocks are checked against (2.2) to near machine precision in Section 5. However, Section 5 verifies only total penalties W(s), not the algebraic decomposition, and the displayed even-block formula fails the one-line check above. The verification protocol as described does not cover the printed closed form. Please update the theorem and re-run the symbolic/numerical checks, or state that the closed forms are superseded by the matrix formula (2.2).
minor comments (4)
  1. [Abstract and Section 3] The abstract's claim that the core normal-system correspondence is machine-checked in Lean 4 could be misread as covering Theorem 2.1. The paper later clarifies that the Lean module covers the projection correspondence of Section 3, not the closed forms; consider making this scope explicit in the abstract.
  2. [Theorem 2.4] The notation b0,b2,b4 is ambiguous: are these coefficients in the matched orthonormal basis or in the monic basis {1,h2,...}? Clarifying this would also help avoid scaling errors such as the one in the even-block formula.
  3. [Section 4.3] The proposed order-selection rule uses 'a few standard errors' of the increment W(s)-W(s-1), but the standard error of that increment is not defined. Provide an explicit formula or cite the relevant component of Proposition 4.3.
  4. [Section 1.3] The s=2 base case and several recovered results are taken from companion preprints with future arXiv numbers. Since these are not yet published, consider proving the needed s=2 case in an appendix to make the paper self-contained.

Circularity Check

0 steps flagged

No significant circularity: Theorem 2.1 follows from the definitions of the bases and the diagonal rule; self-citations are contextual or special-case references, not load-bearing for the central formula.

full rationale

Theorem 2.1 is the exact matrix-coordinate expression of the definitional excess risk W = ||f - \hat f||_P^2 after expanding f in the matched basis and \hat f through the diagonal rule; no fitted constant or data-dependent parameter is introduced. The algorithm is an identity in the moments and the signal-coordinate vector b, and the paper's own proof shows the formula is the residual of the diagonal projection, not a prediction fitted to any target. The companion-work citations (Zabolotnii 2026c for the s=2 penalty, Zabolotnii 2026a for Lean context, Zabolotnii 2026b for moment-free developments) do not carry the central argument: the s=2 case is a special case of Theorem 2.1's proof, and the general theorem is proved directly from the definitions of C and D. Section 3 is explicitly an organizing layer ('We use that geometry rather than claiming it' and 'Proposition 3.5 consolidates known facts'), so the dictionary presentation is not a disguised derivation. Verification is independent of the construction: moment algebra, adaptive quadrature, randomized QMC, and finite-sample simulation agree to machine precision, and the paper reports the one place sampling evidence degrades (lognormal tail truncation) and the plug-in's heavy-tail instability (Table 2: contaminated law at 12% within 10% at n=2000). These disclosed limitations concern moment estimability, not a reduction of the claimed formula to its inputs. The Lean 4 module is scoped to the projection correspondence and is machine-checked; the paper explicitly says extending it to the cumulant forms would be needed to mechanize the main results end-to-end. No circular step is exhibited in the derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

No free parameters are fitted: all quantities are moments, cumulant transforms, or user-supplied signal coordinates. The paper postulates no new physical entities, forces, or dimensions. The central claim rests on explicit finite-moment and support assumptions plus standard linear algebra and asymptotic statistics.

axioms (6)
  • domain assumption Standing non-degeneracy: the Hankel moment matrix H_s of P through order 2s is positive definite, equivalently the support contains at least s+1 points.
    Section 2.1. Needed for C square/nonsingular and D>0, so the diagonal rule and formula (2.2) are well-defined.
  • domain assumption The input law P has finite moments to order 2s (and to order 4s for the asymptotic normality of the plug-in diagnostic).
    Section 2.1 and Proposition 4.3. Hankel matrices and the delta-method covariance require these finite moments.
  • domain assumption The response f lies exactly in the polynomial span Pi_s; the penalty is an excess risk, not a truncation residual.
    Section 1 states 'we keep f in Pi_s throughout'. The analysis deliberately excludes approximation error from higher-degree signal content.
  • domain assumption For the closed forms of Theorem 2.4, P is symmetric with finite moments through order 2s, including an eighth moment for s=4.
    Theorem 2.4 and Remark 2.7. Symmetry zeroes odd moments and supports the parity-block decomposition.
  • standard math Standard mathematical machinery: Cholesky factorization, orthogonal projection in Hilbert space, the delta method, bootstrap consistency, and soundness of the Lean 4/Mathlib kernel.
    Used throughout Sections 2, 3, and 4; the Lean module reports a sorry-free build with audited axioms.
  • domain assumption The Wiener-Hammerstein benchmark records are treated as realizations from a stationary regime whose marginal moments represent the relevant input law.
    Section 4.3. The real-data illustration computes marginal penalties and block-bootstrap intervals, which presumes that level of ergodicity.

pith-pipeline@v1.3.0-alltime-deepseek · 19121 in / 33462 out tokens · 292601 ms · 2026-08-02T01:20:53.022640+00:00 · methodology

0 comments
read the original abstract

Wiener-Hermite cross-correlation identification represents a polynomial response in the Hermite basis. Under Gaussian excitation the basis is orthogonal and a diagonal rule recovers it exactly; under non-Gaussian excitation the same basis is kept, but its Gram matrix gains off-diagonal terms and the diagonal rule is no longer the population projection. We give the exact finite-order excess $L^2(P)$ risk of this mismatch: a moment quadratic form from two Hankel-Cholesky factorizations and one diagonal solve, at $O(s^3)$ cost from moments to order $2s$. Closed cumulant forms at orders three and four expose which non-Gaussian features drive it; symmetry protects the Gaussian basis only through order two. A bootstrap decides, from data, whether a matched basis is worth building; on a Wiener-Hammerstein benchmark it separates a near-Gaussian channel (penalty $\approx 10^{-4}$) from a skewed output (penalty $0.05$). The computation is a weighted-$L^2$ projection whose core normal-system correspondence is machine-checked in Lean 4.

Figures

Figures reproduced from arXiv: 2607.14699 by Serhii Zabolotnii.

Figure 1
Figure 1. Figure 1: Penalty W(s; b, P) for the canonical signal across six input laws (symlog scale). Symmetric kurtotic laws (uniform, contaminated normal) have W(2) = 0 exactly and join at s = 3 (Corollary 2.6); skewed laws pay from s = 2; the Gaussian control stays at zero. 3 The generating-element projection calculus 3.1 Setting and standing assumptions Throughout, (U, A, µ) is a measure space and L 2 (µ) = L 2 (µ)(µ; K),… view at source ↗
Figure 2
Figure 2. Figure 2: One projection core reproduces all five dictionary functionals on a single law (centered [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    1982 , journal =

    Hansen, Lars Peter , title =. 1982 , journal =

    1982

  2. [2]

    1981 , journal =

    Feuerverger, Andrey and McDunnough, Philip , title =. 1981 , journal =

    1981

  3. [3]

    , title =

    Koutrouvelis, Ioannis A. , title =. 1980 , journal =

    1980

  4. [4]

    2000 , journal =

    Carrasco, Marine and Florens, Jean-Pierre , title =. 2000 , journal =

    2000

  5. [5]

    1999 , journal =

    Breusch, Trevor and Qian, Hailong and Schmidt, Peter and Wyhowski, Donald , title =. 1999 , journal =

    1999

  6. [6]

    and McLeish, D

    Small, Christopher G. and McLeish, D. L. , title =. 1994 , publisher =

    1994

  7. [7]

    , title =

    Heyde, Christopher C. , title =. 1997 , publisher =

    1997

  8. [8]

    2007 , booktitle =

    Chen, Xiaohong , title =. 2007 , booktitle =

    2007

  9. [9]

    Godambe, V. P. and Heyde, C. C. , title =. 1987 , journal =

    1987

  10. [10]

    , title =

    Merkouris, T. , title =. 2007 , journal =

    2007

  11. [11]

    and Yu, Jun , title =

    Knight, John L. and Yu, Jun , title =. 2002 , journal =

    2002

  12. [12]

    2008 , journal =

    Wang, Liqun and Leblanc, Alexandre , title =. 2008 , journal =

    2008

  13. [13]

    Lee, Y. W. and Schetzen, M. , title =. 1965 , journal =

    1965

  14. [14]

    Kunchenko, Y. P. , title =. 2002 , publisher =

    2002

  15. [15]

    Kunchenko, Y. P. , title =. 2006 , publisher =

    2006

  16. [16]

    1980 , publisher =

    Schetzen, Martin , title =. 1980 , publisher =

    1980

  17. [17]

    2018 , booktitle =

    Zabolotnii, Serhii and Warsza, Zygmunt Lech and Tkachenko, Oleksandr , title =. 2018 , booktitle =

    2018

  18. [18]

    2026 , howpublished =

    Zabolotnii, Serhii , title =. 2026 , howpublished =. doi:10.48550/arXiv.2605.23419 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2605.23419 2026

  19. [19]

    A Closed-Form Skew Penalty for Volterra Cross-Correlation Identification under Non-Gaussian Input

    Zabolotnii, Serhii , title =. 2026 , howpublished =. doi:10.48550/arXiv.2606.12884 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.12884 2026

  20. [20]

    2026 , howpublished =

    Zabolotnii, Serhii , title =. 2026 , howpublished =. doi:10.48550/arXiv.2606.16289 , url =

    work page doi:10.48550/arxiv.2606.16289 2026

  21. [21]

    1938 , journal =

    Wiener, Norbert , title =. 1938 , journal =

    1938

  22. [22]

    Cameron, R. H. and Martin, W. T. , title =. 1947 , journal =

    1947

  23. [23]

    and Spanos, Pol D

    Ghanem, Roger G. and Spanos, Pol D. , title =. 1991 , publisher =

    1991

  24. [24]

    2002 , journal =

    Xiu, Dongbin and Karniadakis, George Em , title =. 2002 , journal =

    2002

  25. [25]

    and Mugler, Antje and Starkloff, Hans-J

    Ernst, Oliver G. and Mugler, Antje and Starkloff, Hans-J. On the Convergence of Generalized Polynomial Chaos Expansions , year =. ESAIM: Mathematical Modelling and Numerical Analysis , volume =

  26. [26]

    1979 , journal =

    Yasui, Syozo , title =. 1979 , journal =

    1979

  27. [27]

    and Hunter, Ian W

    Korenberg, Michael J. and Hunter, Ian W. , title =. 1996 , journal =

    1996

  28. [28]

    15th IFAC Symposium on System Identification (SYSID 2009) , address =

    Schoukens, Johan and Suykens, Johan and Ljung, Lennart , title =. 15th IFAC Symposium on System Identification (SYSID 2009) , address =. 2009 , url =

    2009