pith. sign in

arxiv: 2606.17665 · v1 · pith:EJBAPNESnew · submitted 2026-06-16 · 🧮 math.ST · math.PR· stat.ML· stat.TH

Non-asymptotic Tail Bounds for the Kostlan--Shub--Smale Field: Tensor PCA and Spherical k-Spin Complexity

Jean-Marc Aza\"is (IMT) , Federico Dalmao (UDELAR) , Yohann De Castro (ICJ , ECL , PSPM , IUF) This is my paper

Pith reviewed 2026-06-26 22:15 UTC · model grok-4.3

classification 🧮 math.ST math.PRstat.MLstat.TH
keywords Kostlan-Shub-Smale fieldtensor PCAspherical k-spin modeltail boundsKac-Rice formulanon-asymptotic boundsprofile maximum likelihood
0
0 comments X

The pith

Explicit non-asymptotic tail bounds for the Kostlan-Shub-Smale field recover the asymptotically optimal rate for rank-R tensor PCA.

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

The paper derives a hierarchy of four explicit tail bounds on the supremum of the Kostlan-Shub-Smale random field on the sphere. It applies these bounds to spiked tensor PCA by showing that the error of the profile maximum likelihood estimator for a rank-R signal is controlled by this supremum. The control follows from a deterministic geometric inequality and a rank-reduction argument that converts the estimation problem into a Gaussian integral against the expected absolute characteristic polynomial of a shifted GOE matrix. The resulting finite-(k,d) error bound recovers the rate sqrt(d log k) with explicit dependence on rank and coherence. The same bounds produce a two-sided non-asymptotic bracketing of the annealed complexity of the spherical k-spin Hamiltonian that converges to the known limiting expression.

Core claim

The estimation error for the profile MLE of a rank-R symmetric signal tensor of order k and dimension d is bounded above by the supremum of the canonical Kostlan-Shub-Smale field via the Tube Method and rank reduction; this supremum admits four explicit non-asymptotic tail bounds obtained from the Kac-Rice formula, a Mehta-Fyodorov representation, and a Ben Arous-Dembo-Guionnet large-deviation principle, yielding an error of order sqrt(d log k) with explicit constants that depend on R and kappa.

What carries the argument

The Tube Method, a deterministic geometric inequality that, together with a rank-reduction step, bounds the tensor PCA estimation error by the supremum of the canonical Kostlan-Shub-Smale field.

If this is right

  • A finite-(k,d) error bound holds for the profile maximum likelihood estimator in tensor PCA with explicit dependence on rank R and coherence kappa.
  • The bound recovers the asymptotically optimal rate sqrt(d log k) of Perry, Wein and Bandeira.
  • A two-sided non-asymptotic bracketing holds for the annealed complexity of the spherical k-spin Hamiltonian.
  • The bracketing recovers the Auffinger-Ben Arous-Cerny complexity function in the high-dimensional limit.

Where Pith is reading between the lines

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

  • The explicit constants allow direct numerical computation of the minimal signal-to-noise ratio guaranteeing a target error level for given finite k, d, R and kappa.
  • The reduction via the Tube Method may be reusable for other manifold-constrained estimation problems whose error can be expressed as a supremum of a smooth random field.
  • The hierarchy of tail bounds supplies a template for obtaining non-asymptotic control on extrema of other Gaussian random fields defined on spheres.

Load-bearing premise

The deterministic geometric inequality of the Tube Method together with the rank-reduction step accurately reduces the tensor PCA estimation error to the supremum of the canonical Kostlan-Shub-Smale field.

What would settle it

A concrete counter-example in which the actual estimation error of the profile MLE on a rank-R spiked tensor exceeds the upper bound implied by the derived tail probabilities for the KSS supremum at the given signal-to-noise ratio.

Figures

Figures reproduced from arXiv: 2606.17665 by ECL, Federico Dalmao (UDELAR), IUF), Jean-Marc Aza\"is (IMT), PSPM, Yohann de Castro (ICJ.

Figure 1
Figure 1. Figure 1: Comparison of the tail bounds δexact, δIMF, δSMF, δSM against the asymptotic baseline δbl, on logarithmic scale, plotted against the level u for (k, d) = (3, 5) (left, uSM = 73.90) and (k, d) = (4, 6) (right, uSM = 87.64), both well outside the displayed range. The five curves are: (i) the strictly exact closed-form bound δexact(u) of Theorem 2 (rigorous, sharpest closed form, and the reference for δ0 sinc… view at source ↗
Figure 2
Figure 2. Figure 2: Numerical inversion of the two-sided master tail bound at significance levels α ∈ {10−3 , 5×10−2} for (k, d) ∈ {(3, 5), (4, 6)}. In each panel: the two-sided IMF bound δmin(u) = 2δIMF(u) (blue), the two￾sided exact bound 2δexact(u) of Theorem 2 (black dashed), the horizontal target level α (red dotted), the threshold uα obtained by bisecting δmin(u)−α (blue dotted vertical), and the exact threshold u exact… view at source ↗
read the original abstract

This paper builds a hierarchy of explicit, non-asymptotic tail bounds for the supremum of the Kostlan--Shub--Smale (KSS) random field on the sphere, and applies it to two problems: Spiked Tensor PCA and the landscape of the spherical $k$-spin model. For Tensor PCA, we study the non-asymptotic statistical limits of estimating a rank-$R$ symmetric signal tensor of order~$k\ge 3$ and dimension~$d\ge 3$ from a single Gaussian observation at signal-to-noise ratio~$\lambda$, through the \emph{profile maximum likelihood estimator}, the MLE restricted to normalized rank-$R$ tensors of coherence at least~$\kappa$. Our analysis uses a single reduction: a deterministic geometric inequality (the Tube Method) and a rank-reduction step bound the estimation error by the supremum of the canonical KSS field, which the Kac--Rice formula turns into a Gaussian integral against the expected absolute characteristic polynomial of a shifted Gaussian Orthogonal Ensemble, controlled in turn by the four explicit tail bounds of our hierarchy (three from a Mehta--Fyodorov representation, one from a Ben Arous--Dembo--Guionnet large deviation). The same reduction yields two results, each with explicit constants. For estimation, a finite-$(k,d)$ error bound recovers the asymptotically optimal rate~$\sqrt{d\log k}$ of Perry, Wein and Bandeira, with explicit dependence on the rank~$R$ and the coherence~$\kappa$. For the landscape, a two-sided non-asymptotic bracketing of the annealed complexity of the spherical $k$-spin Hamiltonian recovers the Auffinger--Ben Arous--\v{C}ern\'y complexity function in the high-dimensional limit.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper develops a hierarchy of four explicit non-asymptotic tail bounds for the supremum of the Kostlan-Shub-Smale (KSS) random field on the sphere, obtained via Mehta-Fyodorov representations and Ben Arous-Dembo-Guionnet large deviations, then controlled through Kac-Rice integrals against the expected absolute characteristic polynomial of a shifted GOE. These bounds are applied to two problems via a single deterministic reduction: for spiked tensor PCA of order k, a Tube Method geometric inequality plus rank-reduction step bounds the estimation error of the profile MLE (restricted to normalized rank-R tensors of coherence at least kappa) by the KSS supremum, yielding finite-(k,d) error bounds that recover the Perry-Wein-Bandeira rate sqrt(d log k) with explicit R and kappa dependence; the same reduction produces two-sided non-asymptotic bracketing of the annealed complexity of the spherical k-spin Hamiltonian that recovers the Auffinger-Ben Arous-Černý function in the high-dimensional limit.

Significance. If the central reduction is valid with the stated constants, the explicit finite-(k,d) tail bounds and their application to tensor PCA and spin-glass complexity constitute a useful contribution, supplying concrete non-asymptotic rates and constants where only asymptotic statements existed. The deterministic character of the reduction (no fitted parameters) and the explicit hierarchy derived from standard random-matrix tools are strengths that support reproducibility.

major comments (1)
  1. [Abstract] Abstract (paragraph on the single reduction): the claim that the deterministic Tube Method geometric inequality together with the rank-reduction step bounds the profile MLE estimation error exactly by the supremum of the canonical KSS field on the constrained manifold of normalized rank-R tensors with coherence >= kappa is load-bearing for every subsequent explicit rate with R and kappa dependence. The manuscript must supply a self-contained verification that the inequality holds with the stated constants on this manifold and that the rank-reduction introduces no uncontrolled additive error term; without this, the non-asymptotic claim does not follow even if the KSS tail hierarchy is correct.
minor comments (1)
  1. Clarify the precise statement of the four tail bounds (which three come from Mehta-Fyodorov and which from Ben Arous-Dembo-Guionnet) and ensure every constant appearing in the final error bound is tracked explicitly from the KSS supremum through the Kac-Rice integral.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying the load-bearing nature of the reduction. We address the single major comment below and commit to revisions that strengthen the self-contained presentation of the argument.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on the single reduction): the claim that the deterministic Tube Method geometric inequality together with the rank-reduction step bounds the profile MLE estimation error exactly by the supremum of the canonical KSS field on the constrained manifold of normalized rank-R tensors with coherence >= kappa is load-bearing for every subsequent explicit rate with R and kappa dependence. The manuscript must supply a self-contained verification that the inequality holds with the stated constants on this manifold and that the rank-reduction introduces no uncontrolled additive error term; without this, the non-asymptotic claim does not follow even if the KSS tail hierarchy is correct.

    Authors: We agree that the reduction must be fully verifiable. The full manuscript derives the Tube Method inequality in Lemma 2.1 (Section 2) and the rank-reduction step in Proposition 3.3 (Section 3), where the additive error is explicitly bounded by a term depending only on R, κ, k and d that is absorbed into the final non-asymptotic rates; the bound is one-sided (error ≤ KSS supremum + controlled term) rather than equality. To make the argument self-contained on the constrained manifold, we will add a consolidated lemma (new Lemma A.1 in the appendix) that restates the geometric inequality and rank-reduction together with the explicit constants, without cross-references. This addresses the referee’s concern directly while preserving the deterministic character of the reduction. revision: yes

Circularity Check

0 steps flagged

Derivation chain self-contained via deterministic reduction and standard tail bounds

full rationale

The paper's central reduction (abstract) is a deterministic geometric inequality via the Tube Method plus rank-reduction that bounds profile MLE error exactly by the KSS supremum; this step is presented as an inequality without embedding the final rate or coherence dependence inside fitted parameters. Subsequent tail bounds are obtained from Kac-Rice formula applied to the GOE characteristic polynomial, with three bounds from Mehta-Fyodorov representation and one from Ben Arous-Dembo-Guionnet LDP, all independent of the tensor-PCA target quantities. The finite-(k,d) error bound recovers the Perry-Wein-Bandeira rate asymptotically but is derived from the explicit hierarchy rather than by construction or self-citation load-bearing. No self-definitional, fitted-input, or ansatz-smuggling steps appear.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper rests on standard tools from random field theory and random matrix theory; no free parameters, ad-hoc axioms, or invented entities are introduced in the abstract.

axioms (2)
  • standard math Kac-Rice formula converting the expected number of critical points or the distribution of the supremum into an integral involving the characteristic polynomial of a GOE matrix
    Invoked to turn the KSS supremum into the Gaussian integral against the expected absolute characteristic polynomial.
  • standard math Mehta-Fyodorov representation of the characteristic polynomial of a shifted GOE
    Used as one of the three routes to the explicit tail bounds.

pith-pipeline@v0.9.1-grok · 5891 in / 1624 out tokens · 49073 ms · 2026-06-26T22:15:52.154563+00:00 · methodology

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

23 extracted references · 4 canonical work pages

  1. [1]

    M., and Telgarsky, M

    Anandkumar, A., Ge, R., Hsu, D., Kakade, S. M., and Telgarsky, M. (2014). Tensor decompositions for learning latent variable models. Journal of Machine Learning Research , 15(1):2773--2832

    2014

  2. [2]

    W., Guionnet, A., and Zeitouni, O

    Anderson, G. W., Guionnet, A., and Zeitouni, O. (2010). An Introduction to Random Matrices , volume 118 of Cambridge Studies in Advanced Mathematics . Cambridge University Press

    2010

  3. [3]

    Auffinger, A., Ben Arous , G., and C ern \`y , J. (2013). Random matrices and complexity of spin glasses. Communications on Pure and Applied Mathematics , 66(2):165--201

    2013

  4. [4]

    Aza \" s, J.-M., Dalmao, F., and De Castro, Y. (2024). Second maximum of a G aussian random field and exact ( t -)spacing test. ArXiv preprint , abs/2406.18397

    work page arXiv 2024

  5. [5]

    and Wschebor, M

    Aza \" s, J.-M. and Wschebor, M. (2009). Level sets and extrema of random processes and fields . John Wiley & Sons

    2009

  6. [6]

    Baik, J., Ben Arous , G., and P \'e ch \'e , S. (2005). Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. The Annals of Probability , 33(5):1643--1697

    2005

  7. [7]

    Ben Arous , G., Dembo, A., and Guionnet, A. (2001). Aging of spherical spin glasses. Probab. Theory Relat. Fields , 120:1--67

    2001

  8. [8]

    Ben Arous, G., Gerbelot, C., and Piccolo, V. (2024a). Langevin dynamics for high-dimensional optimization: the case of multi-spiked tensor PCA . ArXiv preprint , abs/2408.06401

    work page arXiv

  9. [9]

    Ben Arous, G., Gerbelot, C., and Piccolo, V. (2024b). Stochastic gradient descent in high dimensions for multi-spiked tensor PCA . ArXiv preprint , abs/2410.18162

    work page arXiv

  10. [10]

    Ben Arous , G., Mei, S., Montanari, A., and Nica, M. (2019). The landscape of the spiked tensor model. Communications on Pure and Applied Mathematics , 72(11):2282--2330

    2019

  11. [11]

    Fyodorov, Y. V. (2004). Complexity of random energy landscapes, glass transition, and absolute value of the spectral determinant of random matrices. Physical Review Letters , 92(24):240601

    2004

  12. [12]

    and Husson, J

    Guionnet, A. and Husson, J. (2022). Asymptotics of k dimensional spherical integrals and applications. ALEA Lat. Am. J. Probab. Math. Stat. , 19(1):769--797

    2022

  13. [13]

    and Ma\" da, M

    Guionnet, A. and Ma\" da, M. (2005). A F ourier view on the R -transform and related asymptotics of spherical integrals. Journal of Functional Analysis , 222(2):435--490

    2005

  14. [14]

    Kostlan, E. (1993). On the distribution of roots of random polynomials. In From Topology to Computation: Proceedings of the Smalefest , pages 419--431. Springer

    1993

  15. [15]

    Lifshits, M. A. (1983). On the absolute continuity of distributions of functionals of random processes . Theory of Probability & Its Applications , 27(3):600--607

    1983

  16. [16]

    Ma \"i da, M. (2007). Large deviations for the largest eigenvalue of rank one deformations of G aussian ensembles. Electronic Journal of Probability , 12:1131--1150

    2007

  17. [17]

    Mehta, M. L. (2004). Random matrices , volume 142 of Pure and Applied Mathematics . Elsevier/Academic Press, 3rd edition

    2004

  18. [18]

    S., and Bandeira, A

    Perry, A., Wein, A. S., and Bandeira, A. S. (2020). Statistical limits of spiked tensor models. In Annales de l’Institut Henri Poincar \'e . Probabilit \'e s et Statistiques , volume 56, pages 230--264. Association des Publications de l’Institut Henri Poincar \'e

    2020

  19. [19]

    Piccolo, V. (2023). Topological complexity of spiked random polynomials and finite-rank spherical integrals. ArXiv preprint , abs/2312.12323

    work page arXiv 2023

  20. [20]

    and Montanari, A

    Richard, E. and Montanari, A. (2014). A statistical model for tensor PCA . In Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D., and Weinberger, K. Q., editors, Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada , pages 2897--2905

    2014

  21. [21]

    and Smale, S

    Shub, M. and Smale, S. (1993). Complexity of b \'e zout's theorem. II . volumes and probabilities. In Computational Algebraic Geometry , pages 267--285. Birkh \"a user Boston

    1993

  22. [22]

    Szeg o , G. (1975). Orthogonal Polynomials , volume 23 of American Mathematical Society Colloquium Publications . American Mathematical Society, 4th edition

    1975

  23. [23]

    Tsirelson, V. S. (1976). The density of the distribution of the maximum of a Gaussian process . Theory of Probability & Its Applications , 20(4):847--856

    1976