pith. sign in

arxiv: 2605.22760 · v1 · pith:MYIGCFOYnew · submitted 2026-05-21 · 🧮 math.PR

Extremes of Gaussian fields with a product term in the variance

Svyatoslav Novikov This is my paper

Pith reviewed 2026-05-22 03:15 UTC · model grok-4.3

classification 🧮 math.PR
keywords Gaussian random fieldshigh excursionsextreme value asymptoticsvariance losslocalization domainscorrelation functions
0
0 comments X

The pith

When the variance loss includes a product term with exponent a less than beta over 2, high excursions of the Gaussian field localize along the sides rather than inside a rectangle.

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

The paper studies the probability that a centered Gaussian field on a square reaches unusually high values near a corner where its standard deviation peaks. It assumes the loss in standard deviation takes the form t1 to the beta plus t2 to the beta plus a product term t1 to a times t2 to a, together with power-law decay in the correlation function. For a smaller than beta over 2, the usual rectangular region that captures the leading contribution in additive cases no longer works; instead the dominant regions attach to the sides and in some regimes behave as if one-dimensional. This produces new asymptotic expressions for the excursion probability that include logarithmic corrections and side-dominated terms absent from the classical locally additive setting.

Core claim

Under the local expansion 1 minus sigma of t equals t1 to the beta plus t2 to the beta plus t1 to a times t2 to a as t approaches zero in the positive quadrant, with a less than beta over 2, and with correlation 1 minus r of t and s approximately the sum of absolute differences to the alpha in each coordinate, the leading high-level excursion probability is governed by side-attached localization sets rather than the classical essential rectangle at the variance-loss scale; the resulting asymptotics include previously unseen logarithmic and effectively one-dimensional regimes.

What carries the argument

The side-attached localization domain, whose geometry is determined by balancing the product term in the variance loss against the correlation decay, replaces the classical essential rectangle and determines the scaling of the high-excursion probability.

If this is right

  • The high-level asymptotics acquire logarithmic factors when the side attachment becomes dominant.
  • In a sub-range of parameters the problem reduces to an effectively one-dimensional excursion calculation along the boundary.
  • The classical essential-rectangle method fails to capture the leading term once the product contribution exceeds the additive terms in the variance loss.
  • New scaling regimes appear that have no counterpart under the usual locally additive assumptions on the variance.

Where Pith is reading between the lines

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

  • The same change in localization geometry may appear in other Gaussian models that incorporate multiplicative variance effects, such as certain random fields on domains with corners.
  • Simulation algorithms for extreme-value sampling could be improved by concentrating samples along the sides once a is known to be below beta over 2.
  • The transition point a equals beta over 2 may mark a phase boundary separating rectangle-dominated from side-dominated regimes, suggesting a natural next calculation at the critical value.

Load-bearing premise

The local expansion of one minus the standard deviation must include the product term t1 to the a times t2 to the a with a strictly less than beta over 2 and must hold uniformly as the two coordinates approach zero.

What would settle it

Numerical Monte Carlo estimation of the tail probability for a discretized field on a fine grid near the corner, with the product exponent a set below beta over 2, should deviate from the classical rectangle-based prediction and instead match the side-attached asymptotic formula.

read the original abstract

We study the high excursion probability of a centered Gaussian field on a square. Writing \(\sigma\) and \(r\) for its standard deviation and correlation function, we assume that \(\sigma\) has a unique maximum at the corner \(\boldsymbol{0}=(0,0)\) and \[ 1-\sigma(\boldsymbol{t}) \sim t_1^\beta+t_2^\beta+t_1^a t_2^a , \qquad \boldsymbol{t}=(t_1,t_2)\to\boldsymbol{0} \] in \(\mathbb R_+^2\). The local correlation is assumed to satisfy \[ 1-r(\boldsymbol{t},\boldsymbol{s})\sim |t_1-s_1|^\alpha+|t_2-s_2|^\alpha, \qquad 0<\alpha<\beta . \] This product form of the standard-deviation loss is not covered by the usual locally additive assumptions. In the range \(a<\beta/2\), the classical essential rectangle at the variance-loss scale no longer captures the leading contribution; the relevant localization becomes side-attached and, in one regime, effectively one-dimensional. We determine the corresponding high-level asymptotics, including the logarithmic and side-dominated regimes which do not arise in the locally additive case.

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

2 major / 2 minor

Summary. The paper studies high excursion probabilities for a centered Gaussian field on the unit square with unique maximum of the standard deviation at the origin. It assumes the local expansion 1-σ(t) ∼ t₁^β + t₂^β + t₁^a t₂^a as t→0 in R₊² and the separable correlation 1-r(t,s) ∼ |t₁-s₁|^α + |t₂-s₂|^α with 0<α<β. In the regime a<β/2 the classical essential-rectangle localization at the variance-loss scale is shown to miss the leading contribution; the paper derives the correct side-attached (and in one sub-regime effectively one-dimensional) asymptotics, including previously unseen logarithmic and side-dominated regimes.

Significance. If the asymptotic analysis is rigorous, the work meaningfully extends the theory of extremes of Gaussian fields to non-additive variance losses. The identification of qualitatively new localization geometries and the explicit treatment of the logarithmic and side-dominated regimes are the main contributions; these regimes do not appear under the usual locally additive assumptions and may be relevant for applications involving product-type variance structures.

major comments (2)
  1. [§3.2] §3.2, display (3.8) and the subsequent integral representation: the change-of-variables argument that reduces the side-attached contribution to a one-dimensional integral when a<β/2 appears to absorb the product term into the boundary measure, but the error control between the original two-dimensional integral and this reduced form is only sketched; a quantitative bound on the remainder that is uniform in the level u would be needed to justify that the logarithmic correction is indeed leading.
  2. [Theorem 4.1] Theorem 4.1 (side-dominated regime): the statement claims that the probability is asymptotically equivalent to a constant times u^{-γ} exp(-u²/2) for an explicit γ depending on a,β,α. The proof outline invokes a localization to a thin strip along one axis, yet the justification that the contribution from the opposite side and from the interior of the square are o(1) relative to this term relies on a comparison that is not fully detailed in the displayed estimates.
minor comments (2)
  1. [§2] The notation for the essential rectangle versus the side-attached region is introduced only in the text of §2; a small diagram or explicit definition in a preliminary subsection would improve readability.
  2. Several constants (e.g., the prefactor C in the logarithmic regime) are left in implicit integral form; writing the explicit integral expression or at least stating its dependence on α,β,a would make the result easier to use.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading of the manuscript and for identifying points where the error estimates and localization arguments can be strengthened. We address each major comment below, indicating the revisions we will make to improve the rigor of the presentation.

read point-by-point responses

  1. Referee: [§3.2] §3.2, display (3.8) and the subsequent integral representation: the change-of-variables argument that reduces the side-attached contribution to a one-dimensional integral when a<β/2 appears to absorb the product term into the boundary measure, but the error control between the original two-dimensional integral and this reduced form is only sketched; a quantitative bound on the remainder that is uniform in the level u would be needed to justify that the logarithmic correction is indeed leading.

    Authors: We agree that a more quantitative control on the remainder after the change of variables is desirable to confirm uniformity in u. In the revised version we will insert explicit bounds showing that the difference between the original two-dimensional integral and the reduced one-dimensional expression is o(1) times the leading term, uniformly for large u. The argument will exploit the monotonicity properties of the variance loss and the Hölder regularity of the correlation to obtain the required uniform estimate. revision: yes

  2. Referee: [Theorem 4.1] Theorem 4.1 (side-dominated regime): the statement claims that the probability is asymptotically equivalent to a constant times u^{-γ} exp(-u²/2) for an explicit γ depending on a,β,α. The proof outline invokes a localization to a thin strip along one axis, yet the justification that the contribution from the opposite side and from the interior of the square are o(1) relative to this term relies on a comparison that is not fully detailed in the displayed estimates.

    Authors: We acknowledge that the comparison arguments establishing negligibility of the opposite side and the interior could be presented with greater detail. In the revision we will expand the relevant estimates, providing explicit upper bounds on the Gaussian tail probabilities in those regions that are o(u^{-γ} exp(-u²/2)) and that rely only on the assumed form of the variance loss away from the chosen side. This will make the localization step fully rigorous. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained under explicit assumptions

full rationale

The paper states two modeling assumptions—the local expansion 1-σ(t) ∼ t₁^β + t₂^β + t₁^a t₂^a and the separable correlation form 1-r(t,s) ∼ |t₁-s₁|^α + |t₂-s₂|^α with 0<α<β—as the inputs that define the regime a<β/2. It then performs direct asymptotic analysis to obtain the high-excursion probabilities, showing that the essential rectangle fails and new side-attached or one-dimensional localizations appear. No equation reduces a derived quantity to a fitted parameter by construction, no self-citation is invoked as a uniqueness theorem, and no ansatz is smuggled in; the central claims follow from the stated local behaviors without circular reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard regularity properties of centered Gaussian fields and on the two local asymptotic expansions supplied as modeling assumptions. The exponents α, β, a function as regime parameters rather than fitted constants. No new entities are postulated.

free parameters (1)
  • exponents a, β, α
    These define the local expansions and the regime a < β/2; they are chosen to delineate the new localization behavior rather than estimated from data.
axioms (2)
  • domain assumption The random field is centered Gaussian with sufficient regularity for excursion probabilities to be well-defined.
    Invoked implicitly throughout the study of high-level asymptotics for Gaussian fields.
  • domain assumption The stated local expansions for 1-σ(t) and 1-r(t,s) hold in a neighborhood of the corner maximum.
    These expansions are the load-bearing modeling assumptions that enable the change in localization geometry.

pith-pipeline@v0.9.0 · 5775 in / 1759 out tokens · 66467 ms · 2026-05-22T03:15:34.591116+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    and Taylor, Jonathan E

    Adler, Robert J. and Taylor, Jonathan E. , TITLE =. 2007 , PAGES =

    work page 2007

  2. [2]

    Extremes of threshold-dependent

    Bai, Long and D. Extremes of threshold-dependent. Sci. China Math. , FJOURNAL =. 2018 , NUMBER =. doi:10.1007/s11425-017-9225-7 , URL =

    work page doi:10.1007/s11425-017-9225-7 2018

  3. [3]

    Advances in Applied Probability , author=

    Extremes of Gaussian random fields with nonadditive dependence structure , DOI=. Advances in Applied Probability , author=. 2025 , pages=

    work page 2025

  4. [4]

    usler, J\

    Hashorva, Enkelejd and H\"usler, J\"urg , TITLE =. Methodol. Comput. Appl. Probab. , FJOURNAL =. 2000 , NUMBER =. doi:10.1023/A:1010029228490 , URL =

    work page doi:10.1023/a:1010029228490 2000

  5. [5]

    Electron

    Ievlev, Pavel and Kriukov, Nikolai , TITLE =. Electron. J. Probab. , FJOURNAL =. 2025 , PAGES =. doi:10.1214/25-ejp1288 , URL =

    work page doi:10.1214/25-ejp1288 2025

  6. [6]

    Leadbetter, M. R. and Lindgren, Georg and Rootz\'en, Holger , TITLE =. 1983 , PAGES =

    work page 1983

  7. [7]

    Pickands, III, James , TITLE =. Trans. Amer. Math. Soc. , FJOURNAL =. 1969 , PAGES =. doi:10.2307/1995058 , URL =

    work page doi:10.2307/1995058 1969

  8. [8]

    , TITLE =

    Piterbarg, Vladimir I. , TITLE =. 1996 , PAGES =. doi:10.1090/mmono/148 , URL =

    work page doi:10.1090/mmono/148 1996

  9. [9]

    Extremes of

    D. Extremes of. ESAIM Probab. Stat. , FJOURNAL =. 2017 , PAGES =. doi:10.1051/ps/2017019 , URL =

    work page doi:10.1051/ps/2017019 2017