pith. sign in

arxiv: 2604.12405 · v1 · submitted 2026-04-14 · 📊 stat.ME

A sub-asymptotic model for bivariate threshold exceedances

Mirco Lescart , Anna Kiriliouk , Philippe Naveau This is my paper

Pith reviewed 2026-05-10 15:29 UTC · model grok-4.3

classification 📊 stat.ME
keywords bivariate extremesthreshold exceedancessub-asymptotic modelstail dependencegeneralized Pareto distributionneural Bayes estimationrainfall extremes
0
0 comments X

The pith

A new parametric model for bivariate threshold exceedances captures a wide range of tail dependence while reducing to the multivariate generalized Pareto distribution in the limit.

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

The authors develop a sub-asymptotic parametric family specifically for pairs of variables that exceed high thresholds together. The model is constructed so that its margins approach standard univariate generalized Pareto tails as thresholds rise, yet the strength of dependence between the two variables can still change with those thresholds on the original scale. It includes the standardized multivariate GP distribution as a limiting special case when dependence becomes fully asymptotic. Parameter estimation uses a neural Bayes method that sidesteps direct likelihood evaluation and incorporates custom priors. The approach is tested in simulations and applied to Belgian rainfall records to show it can represent both strongly and weakly dependent extremes.

Core claim

We propose a flexible parametric class for modeling bivariate threshold exceedances that accommodates various tail dependence behaviors, includes the standardized multivariate GP distribution as a special limiting case, and preserves margins that converge to univariate GP tails, with extremal dependence evolving naturally with the marginal parameters on the original data scale.

What carries the argument

The sub-asymptotic parametric family for bivariate exceedances, which lets the dependence function vary with the marginal scale and shape parameters.

If this is right

  • Joint failure probabilities can be evaluated directly on the original measurement scale without additional transformation steps.
  • Extremal dependence strength can increase or decrease as thresholds are raised, matching observed behavior in many environmental series.
  • The model recovers the multivariate GP as a boundary case, so it nests standard asymptotic dependence models.
  • Likelihood-free neural estimation supplies posterior distributions for all parameters, including uncertainty in predicted return levels.

Where Pith is reading between the lines

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

  • The same construction could be extended to three or more variables by replacing the bivariate dependence function with a suitable multivariate analogue.
  • In climate applications the ability to let dependence evolve with severity would improve estimates of compound extremes such as simultaneous high rainfall and wind.
  • Comparative studies on other environmental datasets would clarify when the extra flexibility of the sub-asymptotic form is required versus simpler asymptotic models.

Load-bearing premise

The chosen parametric form is flexible enough to match the dependence patterns that actually occur in real bivariate extreme data, and the neural Bayes procedure recovers the parameters with acceptable accuracy.

What would settle it

A large simulation study in which the neural Bayes estimates show persistent bias in the dependence parameters or produce joint exceedance probabilities that deviate systematically from the empirical frequencies.

Figures

Figures reproduced from arXiv: 2604.12405 by Anna Kiriliouk, Mirco Lescart, Philippe Naveau.

Figure 1
Figure 1. Figure 1: Probability density functions of Yj in Definition 3.1 with ξj = 0.2 and βj = 1. In the left panel, the parameter σT = 0.25 is fixed and the weight w ∈ {0.5, 0.8, 1} (blue, red, yellow lines), while w = 0.8 and σT ∈ {0, 0.25, 0.5} (blue, red, yellow lines) in the right panel. for some non-negative random vector S 0 = (S 0 1 , S0 2 ) T with min(S 0 1 , S0 2 ) = 0 a.s. Then (Y1 , Y2 ) p −→ [PITH_FULL_IMAGE:f… view at source ↗
Figure 2
Figure 2. Figure 2: The left panel shows that varying w leads to large changes in χ, whereas changes in the marginal tail index ξ = 1/α have a comparatively minor effect. The right panel illustrates how quickly η increases to 1 as a function of the shared coefficient α. 3.2 Sub-asymptotic model features While Section 3.1 focused on asymptotic properties, convergence to the limiting coefficients (χ, η) can be slow, so that dep… view at source ↗
Figure 2
Figure 2. Figure 2: Limiting tail dependence coefficients χ (left panel) and η (right panel) as functions of the weight w in Definition 3.1, for the special case of identical tail indices ξ1 = ξ2 = ξ. markedly different dependence strengths at moderately high quantile levels. This phenomenon is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Scatterplots of samples of size n = 1000 simulated from the sBGP model (Defini￾tion 3.1) with identical coefficients (χ, η) = (0, 0.9) and parameters (α, α1 , α2 , β1 , β2 , σT ) = (4.44, 0.56, 0.56, 1, 1, 0.1). From left to right: w = 0.1, w = 0.5, w = 0.9. No closed-form expression is available for χ(q) when q < 1. We approximate it empirically using a large simulated sample Y1 , . . . ,YN from the model… view at source ↗
Figure 4
Figure 4. Figure 4: Empirical χ(q) curves for the sBGP model (3.1). The parameters (α, α1 , α2 ) are chosen such that ξ1 = ξ2 = 0.2 and η ∈ {0.5, 0.6, . . . , 1}, with σT = 0.1. Left: w = 0.1; Middle: w = 0.5; Right: w = 0.9. For each pair (w, η), χ(q) is estimated using (3.5), based on a single synthetic sample of size N = 106 , with q ranging up to 0.999. In case of asymptotic dependence, [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 5
Figure 5. Figure 5: Residual tail-dependence curves η(q) for w = 0 in the asymptotically dependent case with α1 = α2 = 0 and α = 5 (so that ξ1 = ξ2 = 0.2 and χ = 1/32). Blue: theoretical curve with σT = 0; Purple and yellow: empirical curves with respectively σT = 0.2 and σT = 0.5. 4 Neural Bayes Estimation Inference for complex extreme value models is generally challenging, as their likelihood functions often lack closed-for… view at source ↗
Figure 6
Figure 6. Figure 6: Boxplots for each parameter (η, ξ1 , ξ2 , β1 , β2 , σT , w) under three dependence settings: θ (1) (green), θ (2) (orange), and θ (3) (red), as defined in (5.1), based on K = 1000 samples of size N = 1000. Diamond markers indicate true values. To assess whether estimation errors in the asymptotic dependence parameters propagate to sub-asymptotic dependence measures, we also examine recovery of the function… view at source ↗
Figure 7
Figure 7. Figure 7: Top: samples from the sBGP distribution (Definition 3.1) with parameters [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Spatial fields of estimated dependence parameters for bivariate precipitation extremes [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Bivariate scatterplots of threshold exceedances of weekly precipitation maxima for [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Estimated χZ˜ 0.7 (q) curves for Brussels–Nivelles and Ostend–Spa, where χ(q) is estimated using (3.5). Quantiles q are defined relative to the exceedance subset Z˜ 0.7 (i.e., within the exceedance region, uj = F −1 j (0.7)). The black line shows the empirical χZ˜ 0.7 (q) from observed exceedances; the green and red lines are the model estimates; the shaded area denotes the 95% bootstrap confidence region… view at source ↗
Figure 11
Figure 11. Figure 11: QQ-plots comparing empirical marginal exceedances to fitted marginal distributions [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Empirical histograms (gray bars) and fitted marginal densities (solid lines) for [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison between the sBGP and the bivariate GP model for Brussels–Nivelles. [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
read the original abstract

Extreme value theory offers a statistical framework for quantifying the risk of rare events, with the generalized Pareto (GP) distribution providing the canonical limit model for univariate threshold exceedances. In many applications, however, extremes are intrinsically multivariate, requiring models that capture both marginal tail behaviours and joint extremal dependencies. Under asymptotic dependence, the multivariate GP distribution represents a suitable modelling family, but when asymptotic independence arises, sub-asymptotic models are needed. In this work, we propose and study a flexible sub-asymptotic parametric class to model bivariate threshold exceedances. Our new model accommodates a broad range of tail dependence behaviours and contains the standardised multivariate GP distribution as a limiting case while retaining margins that converge to univariate GP tails. Our formulation allows extremal dependence to evolve naturally with the marginal parameters on the original data scale, facilitating direct computation and interpretation of failure probabilities. Model inference is done via a likelihood-free neural Bayes estimation approach, with tailored prior specifications. An extensive simulation study and an application to Belgian rainfall extremes illustrate the estimation framework and the flexibility of the model.

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 proposes a new flexible parametric class for modeling bivariate threshold exceedances under sub-asymptotic regimes in extreme value theory. The model accommodates a broad range of tail dependence behaviors, contains the standardized multivariate generalized Pareto distribution as a limiting case, and retains margins that converge to univariate GP tails, with extremal dependence evolving naturally via the marginal parameters on the original scale. Inference is performed via a likelihood-free neural Bayes estimator with tailored priors, illustrated through an extensive simulation study and an application to Belgian rainfall extremes.

Significance. If the central claims on the limiting behavior and estimator performance hold, the work provides a practical parametric tool for bivariate extremes in the asymptotically independent regime, with the advantage of direct computation of joint failure probabilities without transformation to a standardized scale. The linkage to the multivariate GP and the neural Bayes approach for tractable inference represent strengths that could facilitate broader adoption in environmental risk modeling, building on the simulation and real-data validation provided.

major comments (2)
  1. [Model formulation section] The abstract and model section claim that the proposed class 'contains the standardised multivariate GP distribution as a limiting case'; an explicit derivation or parameter restriction (e.g., specific values or limits of the dependence parameters) is needed to verify this reduction, as it is load-bearing for positioning the model as sub-asymptotic.
  2. [Simulation study] §4 (simulation study): while the neural Bayes estimator is shown to recover parameters in simulations, the absence of reported bias, RMSE, or coverage probabilities across varying tail dependence strengths leaves the reliability of the likelihood-free approach unsubstantiated, particularly for the weakest assumption of stable recovery without substantial bias.
minor comments (2)
  1. The description of how dependence evolves with marginal parameters could include a brief illustrative example or plot to aid interpretation on the original data scale.
  2. [Application] In the rainfall application, a direct comparison of fitted joint exceedance probabilities against a standard multivariate GP fit would better demonstrate the practical benefit of the sub-asymptotic extension.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the positioning and validation of our sub-asymptotic bivariate threshold exceedance model. We address each major point below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Model formulation section] The abstract and model section claim that the proposed class 'contains the standardised multivariate GP distribution as a limiting case'; an explicit derivation or parameter restriction (e.g., specific values or limits of the dependence parameters) is needed to verify this reduction, as it is load-bearing for positioning the model as sub-asymptotic.

    Authors: We agree that an explicit derivation is necessary to substantiate this central claim. In the revised manuscript, we will add a new subsection (or appendix) in the model formulation section that derives the limiting case step by step. Specifically, we will show the parameter restrictions (e.g., the dependence function approaching the boundary values that recover the standardized multivariate GP margins and dependence structure) under which our model reduces exactly to the standardized multivariate GP distribution. This will verify the reduction rigorously and reinforce the sub-asymptotic interpretation. revision: yes

  2. Referee: [Simulation study] §4 (simulation study): while the neural Bayes estimator is shown to recover parameters in simulations, the absence of reported bias, RMSE, or coverage probabilities across varying tail dependence strengths leaves the reliability of the likelihood-free approach unsubstantiated, particularly for the weakest assumption of stable recovery without substantial bias.

    Authors: We acknowledge that reporting bias, RMSE, and coverage probabilities would provide more comprehensive evidence of estimator performance, especially across varying tail dependence strengths. In the revised version, we will expand §4 to include these metrics in additional tables or supplementary figures, computed over the simulation replicates for different dependence regimes. This will directly address the concern and better substantiate the stability and reliability of the neural Bayes estimator. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper defines a new parametric family for bivariate sub-asymptotic exceedances whose limiting case is the standardized multivariate GP distribution and whose margins converge to univariate GP tails. The central construction is introduced directly via its functional form and dependence structure on the original scale; inference proceeds via an external likelihood-free neural Bayes estimator trained on simulations. No equation reduces a claimed prediction or uniqueness result to a fitted parameter or to a self-citation whose content is itself the target claim. The simulation study and rainfall application supply independent checks outside the fitted values, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on the existence of a new parametric family whose dependence structure can be parameterized flexibly while satisfying the GP limiting behavior; no explicit free parameters, axioms, or invented entities are named.

pith-pipeline@v0.9.0 · 5481 in / 1104 out tokens · 42968 ms · 2026-05-10T15:29:10.814408+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

5 extracted references · 5 canonical work pages

  1. [1]

    André, L. M., J. L. Wadsworth, and R. Huser (2025). Neural Bayes estimation and selection of complex bivariate extremal dependence models.Extremes, 1–40. Andrieu, C. and G. O. Roberts (2009). The pseudo-marginal approach for efficient Monte Carlo computations.The Annals of Statistics 37(2), 697–725. Bacro, J.-N., C. Gaetan, T. Opitz, and G. Toulemonde (20...

    work page 2025

  2. [2]

    Engelke, S., T

    Springer Science & Business Media. Engelke, S., T. Opitz, and J. Wadsworth (2019). Extremal dependence of random scale constructions.Extremes 22(4), 623–666. 43 Hazra, A., R. Huser, and D. Bolin (2025). Efficient modeling of spatial extremes over large geographical domains.Journal of Computational and Graphical Statistics 34(3), 795–811. He, T., S. Yu, Z....

    work page 2019

  3. [3]

    Kiriliouk, A. and C. Zhou (2022). Estimating probabilities of multivariate failure sets based on pairwise tail dependence coefficients. Available athttps://arxiv.org/abs/2210.12618. Krupskii, P., R. Huser, and M. G. Genton (2018). Factor copula models for replicated spatial data.Journal of the American Statistical Association 113(521), 467–479. Ledford, A...

    work page arXiv 2022

  4. [4]

    Resnick, S

    Springer. Resnick, S. (2007).Heavy-Tail Phenomena. Springer. 44 Richards, J., M. Sainsbury-Dale, A. Zammit-Mangion, and R. Huser (2024). Neural Bayes estimators for censored inference with peaks-over-threshold models.Journal of Machine Learning Research 25(390), 1–49. Rödder, A., M. Hentschel, and S. Engelke (2025). Theoretical guarantees for neural estim...

    work page arXiv 2007

  5. [5]

    Rohrbeck, and T

    45 Zhang, L., C. Rohrbeck, and T. Opitz (2025). Subasymptotic models for spatial extremes. In P. N. M. de Carvalho, R. Huser and B. Reich (Eds.),Handbook on Statistics of Extremes. CRC Press. 46

    work page 2025