pith. sign in

arxiv: 2604.16206 · v1 · submitted 2026-04-17 · 🧮 math.PR · math.ST· stat.TH

Extrapolation of max-stable random fields with Fr\'echet marginals

Vitalii Makogin , Evgeny Spodarev , Ilja Sukhanov This is my paper

Pith reviewed 2026-05-10 07:17 UTC · model grok-4.3

classification 🧮 math.PR math.STstat.TH
keywords max-stable random fieldsFréchet marginalsextrapolationlevel setsheavy tailspredictionDavis-Resnick distanceextreme value theory
0
0 comments X

The pith

A level-set extrapolation method predicts stationary max-stable random fields with α-Fréchet marginals without requiring moment assumptions.

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

The paper introduces a prediction method for stationary max-stable random fields that follow α-Fréchet marginal distributions. This technique handles heavy tails for α between 0 and 2 and aims to be approximately exact in the marginal distributions. It relies on a level-set based extrapolation that avoids any moment conditions. The authors establish a link between the excursion metric and the Davis-Resnick distance, prove the predictor's existence, and show through examples that forecasts are not always unique. They validate the approach with simulations and real annual maximum precipitation data.

Core claim

We propose a method for the prediction of stationary max-stable random fields with α-Fréchet marginal distribution H_α. The method is suitable to cope with heavy tails for α∈(0,2) and is (approximately) exact in marginal distributions. It is based on a recent extrapolation approach via level sets which requires no moment assumptions. An explicit connection between the excursion metric and the Davis-Resnick distance is established. The existence of the predictor is proven. The non-uniqueness of the forecast is demonstrated on several examples. The method is tested on multiple simulated time series and random fields as well as applied to real data of annual maximum precipitation.

What carries the argument

The level-set extrapolation approach for max-stable fields, which links the excursion metric to the Davis-Resnick distance to construct predictions that respect the Fréchet marginals.

If this is right

  • The predictor exists under the stated conditions of stationarity and max-stability.
  • The forecasts are approximately exact in the marginal distributions.
  • The method applies to both time series and spatial random fields.
  • Non-uniqueness of the predictor is possible, as shown in examples.
  • It works on real data such as annual maximum precipitation records.

Where Pith is reading between the lines

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

  • If the level-set method generalizes well, it could extend to non-stationary fields or other marginal distributions in extreme value modeling.
  • The metric connection might simplify computations in other dependence structures for max-stable processes.
  • Further tests could check performance when the max-stability assumption is only approximate.
  • The approach avoids moment assumptions, which may make it robust in fields with very heavy tails where other predictors fail.

Load-bearing premise

The random field is stationary and max-stable with the given Fréchet marginals, allowing the level-set extrapolation to apply directly.

What would settle it

A stationary max-stable random field with α-Fréchet marginals for which the proposed level-set predictor either does not exist or fails to match the marginal distributions in simulations would falsify the method.

Figures

Figures reproduced from arXiv: 2604.16206 by Evgeny Spodarev, Ilja Sukhanov, Vitalii Makogin.

Figure 1
Figure 1. Figure 1: The derivative Ψ′ 1 (y) of the function (33) with γ = 1 for ζ(y) = y + 1 (left) and ζ(y) = y ∨ 1 (right), compare the proof of Theorem 12 and Example 2. for y ∈ [0, 1) and positive for y > 1. Moreover, it holds Ψ′ 1 (0) = ζ ′ (0)/2 − 1 − 3/2γ < 0 for all γ > 0, since 0 ≤ ζ ′ (0) ≤ 1 by upper bound (34). In addition, it can be shown ex adverso from the inequality (34) that limy→+∞ ζ ′ (y) ≤ 1. Hence, we get… view at source ↗
Figure 2
Figure 2. Figure 2: The left plot shows the function Φ associated to the Brown-Resnick process B from Example 1 with volatility σB = 1.68, a forecast sample size of n = 2, t1 = 201, t2 = 202, t0 = 203, γ = 100, N = 100, and all even shifts kj from 2 to 198. The minimum value of the grid is located at (λ1, λ2) = (0.08, 0.83) with Φ((λ1, λ2)) ≈ 0.4265. The right plot shows the iterations of the Adam algorithm, which achieved a … view at source ↗
Figure 3
Figure 3. Figure 3: The red line indicates the empirical excursion metric EˆH1 (Xt0 , Xˆ λˆ) and the blue line represents MSE [(γ) as functions of γ for the three max–stable processes X ∈ {B, S, G}. After having determined optimal penalty weights γ, we perform L = 20-step prediction for each of the three types of processes X using the non-bootstrap formulation (29). To do so, we use a forecast sample size of n = 21 as well as… view at source ↗
Figure 4
Figure 4. Figure 4: A 20 time steps’ forecast (dashed blue line), together with its corresponding theoretical (black line) and empirical (red line) excursion metric, of a Brown-Resnick process B (blue line). The underlying Gaussian process Y of the Brown-Resnick process is a standard Brownian motion with variance parameter σB = 0.771 (left plot) and σB = 2.073 (right plot). For the left plot, γopt = 3 and for the right plot, … view at source ↗
Figure 5
Figure 5. Figure 5: A 20 time steps’ forecast (dashed blue line), together with its corresponding theoretical (black line) and empirical (red line) excursion metric, of a Smith process S (blue line) constructed with a standard deviation of σS = 1.298 (left plot) and σS = 0.482 (right plot), see (20). For the left plot, γopt = 0 and for the right plot, γopt = 2, cf [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A 20 time steps’ forecast (dashed blue line), together with its corresponding theoretical (black line) and empirical (red line) excursion metric, of an extremal Gaussian process G (blue line). The underlying process Y of G in (21) is Gaussian with Ornstein– Uhlenbeck covariance function C(t) = exp(−|t|/σG), where σG = 5.039 (left plot) and σG = 0.256 (right plot). For the left plot, γopt = 0 and for the ri… view at source ↗
Figure 7
Figure 7. Figure 7: A ten steps’ forecast of random fields X ∈ {B, S, G} in both directions t1 and t2. After observing true values Xt at locations t ∈ {1, . . . , 50} × {1, . . . , 50}, the predictor extended the random surface X to t ∈ {1, . . . , 60} × {1, . . . , 60}. The Brown-Resnick field B (left) is simulated using σB = 1.683, the Smith field S (center) is simulated using Σ = σ 2 S I2 with σS = 0.594 and the extremal G… view at source ↗
Figure 8
Figure 8. Figure 8: Left: the yearly maximum of daily rainfall in Munich, Germany from 1879 to 2025. Right: Q-Q plot of the empirical rainfall data against the theoretical Fréchet quan￾tiles with quasi-ML-estimated parameters αˆ ≈ 7.5263, µˆ ≈ −51.4312 and σˆ ≈ 92.7826. 10 Summary We propose a simple method to predict stationary heavy tailed max-stable random fields. The advantages lie in its implementational ease, fast compu… view at source ↗
Figure 9
Figure 9. Figure 9: Left: Correlation function for the time series of annual daily rainfall maxima in Munich. The red dashed lines represent the approximate 95% confidence region under the null hypothesis of zero autocorrelation. Since most bars are inside this region, there is no strong evidence of significant autocorrelation. Right: Forecasts for the annual daily rainfall maxima in the years 2023 to 2025. Data of the years … view at source ↗
read the original abstract

We propose a method for the prediction of stationary max--stable random fields with $\alpha$-Fr\'echet marginal distribution $H_\alpha$. The method is suitable to cope with heavy tails for $\alpha\in(0,2)$ and is (approximately) exact in marginal distributions. It is based on a recent extrapolation approach via level sets which requires no moment assumptions. An explicit connection between the excursion metric and the Davis-Resnick distance is established. The existence of the predictor is proven. The non-uniqueness of the forecast is demonstrated on several examples. The method is tested on multiple simulated time series and random fields as well as applied to real data of annual maximum precipitation.

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

0 major / 2 minor

Summary. The paper proposes a prediction method for stationary max-stable random fields with α-Fréchet marginals (α ∈ (0,2)) that is based on level-set extrapolation, requires no moment assumptions, is approximately marginal-exact, proves existence of the predictor, demonstrates non-uniqueness via examples, establishes an explicit link between the excursion metric and the Davis-Resnick distance, and validates the approach on simulated time series/fields plus annual maximum precipitation data.

Significance. If the central claims hold, the work advances extrapolation techniques for heavy-tailed max-stable processes by removing moment restrictions that limit many existing methods. The proven existence result, explicit metric connection, and demonstration of non-uniqueness supply useful theoretical structure, while the simulation studies and real-data application to precipitation provide concrete evidence of practical relevance in environmental extremes modeling.

minor comments (2)
  1. Abstract and §1: the phrase '(approximately) exact in marginal distributions' is used without an immediate quantitative statement of the approximation error or the sense in which exactness holds; a short clarifying sentence or reference to the relevant theorem would improve readability.
  2. The manuscript would benefit from an explicit statement (perhaps in the simulation section) of the number of Monte Carlo replications and the precise dependence structures used to generate the test fields, so that the reported performance metrics can be reproduced.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our manuscript, the recognition of its contributions to extrapolation methods for max-stable fields without moment assumptions, and the recommendation for minor revision. No specific major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation builds on external technique

full rationale

The paper's central method for extrapolating stationary max-stable fields with α-Fréchet margins is explicitly presented as an application of a recent external level-set extrapolation approach (requiring no moment assumptions). The abstract and skeptic summary confirm the technique is imported rather than internally derived, with the paper supplying independent technical content: existence proofs, non-uniqueness examples, an explicit excursion-metric to Davis-Resnick distance link, and validation on simulations plus real data. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citation chains appear; the assumption that the external method applies directly is addressed by construction without circularity. This is the common case of a self-contained development against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no free parameters, axioms, or invented entities are identifiable; the method is described at a high level without explicit fitting constants or new postulated objects.

pith-pipeline@v0.9.0 · 5417 in / 1172 out tokens · 45044 ms · 2026-05-10T07:17:14.862852+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

48 extracted references · 48 canonical work pages · 1 internal anchor

  1. [1]

    Online learning and stochastic approximations

    Léon Bottou. Online learning and stochastic approximations. In David Saad, editor, On-line Learning in Neural Networks , pages 9–42. Cambridge University Press, New York, 1998

    work page 1998

  2. [2]

    Towards practical adam: Non- convexity, convergence theory, and mini-batch acceleration

    Congliang Chen, Li Shen, Fangyu Zou, and Wei Liu. Towards practical adam: Non- convexity, convergence theory, and mini-batch acceleration. Journal of Machine Learning Research, 23(229):1–47, 2022. A vailable at: http://jmlr.org/papers/ v23/20-1438.html (accessed 2026-04-02)

    work page 2022

  3. [3]

    Geostatistics: modeling spatial uncertainty

    Jean-Paul Chiles and Pierre Delfiner. Geostatistics: modeling spatial uncertainty . John Wiley & Sons, 2012

    work page 2012

  4. [4]

    Variograms for spatial max-stable random fields

    Dan Cooley, Philippe Naveau, and Paul Poncet. Variograms for spatial max-stable random fields. In Patrice Bertail, Philippe Soulier, and Paul Doukhan, editors, De- pendence in Probability and Statistics , pages 373–390. Springer New York, New York, NY, 2006

    work page 2006

  5. [5]

    Extrapolation of stationary random fields via level sets

    Abhinav Das, Vitalii Makogin, and Evgeny Spodarev. Extrapolation of stationary random fields via level sets. Theory of Probability and Mathematical Statistics, 106:85– 103, 2022

    work page 2022

  6. [6]

    R. A. Davis and S. I. Resnick. Basic properties and prediction of max-ARMA pro- cesses. Adv. in Appl. Probab. , 21(4):781–803, 1989

    work page 1989

  7. [7]

    R. A. Davis and S. I. Resnick. Prediction of stationary max-stable processes. Ann. Appl. Probab., 3(2):497–525, 1993

    work page 1993

  8. [8]

    A characterization of multidimensional extreme-value distributions

    Laurens de Haan. A characterization of multidimensional extreme-value distributions. Sankhyā Ser. A , 40(1):85–88, 1978. A vailable at: https://www.jstor.org/stable/ 25050137 (accessed 2026-03-27)

    work page 1978

  9. [9]

    Extreme value theory: an introduction

    Laurens De Haan and Ana Ferreira. Extreme value theory: an introduction . Springer, New York, 2006

    work page 2006

  10. [10]

    Dombry, S

    C. Dombry, S. Engelke, and M. Oesting. Exact simulation of max-stable processes. Biometrika, 103(2):303–317, 2016

    work page 2016

  11. [11]

    Dombry, M

    C. Dombry, M. Oesting, and M. Ribatet. Conditional simulation of max-stable pro- cesses. In Extreme value modeling and risk analysis , pages 215–237. CRC Press, Boca Raton, FL, 2016

    work page 2016

  12. [12]

    Conditional simulation of max-stable processes

    Clément Dombry, Frédéric Eyi-Minko, and Mathieu Ribatet. Conditional simulation of max-stable processes. Biometrika, 100(1):111–124, 2013

    work page 2013

  13. [13]

    Modelling Extremal Events: for Insurance and Finance , volume 33 of Stochastic Modelling and Applied Probability

    Paul Embrechts, Claudia Klüppelberg, and Thomas Mikosch. Modelling Extremal Events: for Insurance and Finance , volume 33 of Stochastic Modelling and Applied Probability. Springer Berlin Heidelberg, Berlin, Heidelberg, 1997

    work page 1997

  14. [14]

    Multivariate extreme value theory and D-Norms

    Michael Falk. Multivariate extreme value theory and D-Norms . Springer, Cham, Switzerland, 2019

    work page 2019

  15. [15]

    Guillaume Garrigos and Robert M. Gower. Handbook of convergence theorems for (stochastic) gradient methods, 2024. 30

    work page 2024

  16. [16]

    Gneiting

    T. Gneiting. Quantiles as optimal point forecasts. International Journal of forecasting, 27(2):197–207, 2011

    work page 2011

  17. [17]

    Extreme-value copulas

    Gordon Gudendorf and Johan Segers. Extreme-value copulas. In Piotr Jaworski, Fabrizio Durante, Wolfgang Karl Härdle, and Tomasz Rychlik, editors, Copula theory and its applications , pages 127–145. Springer, Germany, 2010

    work page 2010

  18. [18]

    Families of min-stable multivariate exponential and multivariate extreme value distributions

    Harry Joe. Families of min-stable multivariate exponential and multivariate extreme value distributions. Statistics & Probability Letters , 9(1):75–81, 1990

    work page 1990

  19. [19]

    Multivariate distributions from mixtures of max- infinitely divisible distributions

    Harry Joe and Taizhong Hu. Multivariate distributions from mixtures of max- infinitely divisible distributions. Journal of Multivariate Analysis , 57(2):240–265, 1996

    work page 1996

  20. [20]

    Ergodic properties of max-infinitely divisi- ble processes

    Zakhar Kabluchko and Martin Schlather. Ergodic properties of max-infinitely divisi- ble processes. Stochastic Process. Appl., 120(3):281–295, 2010

    work page 2010

  21. [21]

    Stationary max-stable fields associated to negative definite functions

    Zakhar Kabluchko, Martin Schlather, and Laurens De Haan. Stationary max-stable fields associated to negative definite functions. The Annals of Probability , 37(5):2042– 2065, 2009

    work page 2042

  22. [22]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  23. [23]

    Komunjer

    I. Komunjer. Quantile prediction. In Handbook of economic forecasting , volume 2, pages 961–994. Elsevier, Amsterdam, 2013

    work page 2013

  24. [24]

    Komunjer and Q

    I. Komunjer and Q. Vuong. Semiparametric efficiency bound in time-series models for conditional quantiles. Econometric Theory, 26(2):383–405, 2010

    work page 2010

  25. [25]

    Modelling the epidemic extremities of dengue transmissions in Thailand

    Jue Tao Lim, Borame Sue Lee Dickens, and Alex R Cook. Modelling the epidemic extremities of dengue transmissions in Thailand. Epidemics, 33:100402, 2020

    work page 2020

  26. [26]

    Prediction of random variables by excursion metric projections

    Vitalii Makogin and Evgeny Spodarev. Prediction of random variables by excursion metric projections. Bernoulli, 31(4):3187–3212, 2025

    work page 2025

  27. [27]

    Estimation of max-stable processes using Monte Carlo methods with applications to financial risk assessment

    Francisco Chamú Morales. Estimation of max-stable processes using Monte Carlo methods with applications to financial risk assessment . PhD thesis, The University of North Carolina at Chapel Hill, Chapel Hill, 2005

    work page 2005

  28. [28]

    Stochastic modeling of flood peaks using the generalized extreme value distribution

    Julia E Morrison and James A Smith. Stochastic modeling of flood peaks using the generalized extreme value distribution. Water Resources Research, 38(12):41–1, 2002

    work page 2002

  29. [29]

    Noaa national centers for environmental information (ncei)

    NOAA National Centers for Environmental Information. Noaa national centers for environmental information (ncei). https://www.ncei.noaa.gov/. Accessed: 2026- 01-13

    work page 2026

  30. [30]

    Conditional sampling for max-stable processes with a mixed moving maxima representation

    Marco Oesting and Martin Schlather. Conditional sampling for max-stable processes with a mixed moving maxima representation. Extremes, 17(1):157–192, 2014

    work page 2014

  31. [31]

    Battle of extreme value distributions: A global survey on extreme daily rainfall

    Simon Michael Papalexiou and Demetris Koutsoyiannis. Battle of extreme value distributions: A global survey on extreme daily rainfall. Water Resources Research, 49(1):187–201, 2013. 31

    work page 2013

  32. [32]

    Mapping extreme rainfall in a mountain- ous region using geostatistical techniques: a case study in Scotland

    Christel Prudhomme and Duncan W Reed. Mapping extreme rainfall in a mountain- ous region using geostatistical techniques: a case study in Scotland. International Journal of Climatology: A Journal of the Royal Meteorological Society , 19(12):1337– 1356, 1999

    work page 1999

  33. [33]

    Using kriging with a heterogeneous measurement error to improve the accuracy of extreme precipitation return level estimation

    Shui qing Yin, Zhonglei Wang, Zhengyuan Zhu, Xu-Kai Zou, and Wen-Ting Wang. Using kriging with a heterogeneous measurement error to improve the accuracy of extreme precipitation return level estimation. Journal of Hydrology , 562:518–529, 2018

    work page 2018

  34. [34]

    Reddi, Satyen Kale, and Sanjiv Kumar

    Sashank J. Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond, 2019

    work page 2019

  35. [35]

    Extreme values, regular variation, and point processes

    Sidney I Resnick. Extreme values, regular variation, and point processes . Springer Science & Business Media, New York, 1987

    work page 1987

  36. [36]

    Spatial extremes: Max-stable processes at work

    Mathieu Ribatet. Spatial extremes: Max-stable processes at work. Journal de la Société Française de Statistique , 154(2):156–177, 2013. A vailable at: https: //statistique-et-societe.fr/index.php/J-SFdS/article/view/184 (accessed 2026-03-27)

    work page 2013

  37. [37]

    Models for stationary max-stable random fields

    Martin Schlather. Models for stationary max-stable random fields. Extremes, 5(1):33– 44, 2002

    work page 2002

  38. [38]

    Max-stable processes and spatial extremes

    Richard L Smith. Max-stable processes and spatial extremes. Unpublished manuscript,

    work page

  39. [39]

    A vailable at: https://www.rls.sites.oasis.unc.edu/postscript/rs/ spatex.pdf (accessed 2026-03-27)

    work page 2026

  40. [40]

    Interpolation of spatial data: some theory for kriging

    Michael L Stein. Interpolation of spatial data: some theory for kriging . Springer Science & Business Media, New York, 1999

    work page 1999

  41. [41]

    Code for ”extrapolation of max-stable random fields”, 2026

    Ilja Sukhanov. Code for ”extrapolation of max-stable random fields”, 2026. GitHub repository

    work page 2026

  42. [42]

    Multivariate geostatistics: an introduction with applications

    Hans Wackernagel. Multivariate geostatistics: an introduction with applications . Springer Science & Business Media, Germany, 2003

    work page 2003

  43. [43]

    Conditional sampling for spectrally discrete max- stable random fields

    Yizao Wang and Stilian A Stoev. Conditional sampling for spectrally discrete max- stable random fields. Advances in Applied Probability , 43(2):461–483, 2011

    work page 2011

  44. [44]

    Detection of non-stationarity in precipitation ex- tremes using a max-stable process model

    Seth Westra and Scott A Sisson. Detection of non-stationarity in precipitation ex- tremes using a max-stable process model. Journal of Hydrology , 406(1-2):119–128, 2011

    work page 2011

  45. [45]

    Evidence that coronavirus superspreading is fat- tailed

    Felix Wong and James J Collins. Evidence that coronavirus superspreading is fat- tailed. Proceedings of the National Academy of Sciences , 117(47):29416–29418, 2020

    work page 2020

  46. [46]

    Non-existence of continuous convex functions on certain Riemannian manifolds

    Shing-Tung Yau. Non-existence of continuous convex functions on certain Riemannian manifolds. Mathematische Annalen , 207(4):269–270, 1974

    work page 1974

  47. [47]

    Upper bounds on value-at-risk for the maximum portfolio loss

    Robert Yuen and Stilian Stoev. Upper bounds on value-at-risk for the maximum portfolio loss. Extremes, 17(4):585–614, 2014

    work page 2014

  48. [48]

    Multivariate extremes, max-stable process estimation and dynamic financial modeling

    Zhengjun Zhang. Multivariate extremes, max-stable process estimation and dynamic financial modeling . PhD thesis, The University of North Carolina at Chapel Hill, Chapel Hill, 2002. 32

    work page 2002