pith. sign in

arxiv: 2604.21620 · v1 · submitted 2026-04-23 · 📊 stat.ME

Causal Discovery in Multivariate Extremes via Tail Asymmetry

Mengran Li , Daniela Castro-Camilo This is my paper

Pith reviewed 2026-05-09 21:18 UTC · model grok-4.3

classification 📊 stat.ME
keywords causal discoverymultivariate extremestail asymmetrymax-linear modelsstructure learninghigh-dimensional inferencelatent confounding
0
0 comments X

The pith

Tail asymmetry in extreme events identifies causal directions under max-linear models.

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

The paper seeks to establish that heavy-tailed data contain a systematic asymmetry: forward prediction of extremes along a causal path is easier than backward prediction. This asymmetry is shown to make causal orientations identifiable when the data follow a canonical max-linear model. If the claim holds, it supplies a score that can orient edges in a graph recovered from tail observations alone, without needing a prespecified skeleton or assuming no latent shocks. The approach therefore offers a route to causal discovery in settings where extremes are rare, dependent, and possibly confounded, such as river networks or financial tail risks.

Core claim

Under a canonical max-linear model, extreme events propagate asymmetrically so that the risk of predicting one variable's tail from another's is lower in the forward causal direction than in the reverse. This tail-induced asymmetry is identifiable and can be used as a score to orient edges. The paper builds a two-stage procedure that first screens a sparse undirected skeleton via proxy-adjusted penalized neighborhood selection and then orients the edges by minimizing tail prediction risk with max-linear envelope models, proving high-dimensional consistency under population separation conditions.

What carries the argument

Tail-induced asymmetry, the property that forward tail prediction risk is systematically lower than backward risk in a directed max-linear model.

If this is right

  • Causal directions become identifiable from tail data without restricting the graph in advance.
  • A sparse candidate skeleton can be recovered consistently in high dimensions even with latent confounding.
  • The score-based orientation step is consistent when population separation holds.
  • The procedure scales to larger systems than prior extremal causal methods while remaining robust to hidden common shocks.

Where Pith is reading between the lines

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

  • The same asymmetry principle might be testable in non-max-linear extreme-value models if analogous directional prediction gaps appear.
  • Applications could extend to temporal extremes by replacing static neighborhoods with lagged versions.
  • If the asymmetry is detectable in climate or environmental extremes, the framework could inform early-warning graphs for cascading failures.

Load-bearing premise

The data are generated from a canonical max-linear model that exhibits the stated directional tail asymmetry, and the population risks for the true orientations are strictly separated from those of the reversed orientations.

What would settle it

A controlled simulation or real dataset with known ground-truth directions in which the tail prediction risk minimizer orients at least one edge opposite to the true direction would falsify the identifiability claim.

Figures

Figures reproduced from arXiv: 2604.21620 by Daniela Castro-Camilo, Mengran Li.

Figure 1
Figure 1. Figure 1: High-dimensional scaling results (m = 1, n = 1000, 50 replicates). Boxplots summarize F1 (left) and SHD (right) across p ∈ {20, 50, 100, 200} for S3ME skeleton, EASE, FGES, and S3ME. All methods degrade as p grows, but the rate of degradation differs markedly. FGES shows a pronounced deterioration in overall recovery quality, with lower F1 and larger SHD as dimension increases, because its Gaussian BIC sco… view at source ↗
Figure 2
Figure 2. Figure 2: Upper Danube comparison. Left Panel shows the physical river-flow graph, while [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Threshold sensitivity analysis (n = 1000, p = 50, 50 replicates). Left: mean F1 vs. β for the skeleton step (blue) and the full DAG (red). Right: mean SHD vs. β for the skeleton (blue) and the full DAG (red). DAG-level metrics are stable across all thresholds; skeleton metrics show mild degradation only at large β where fewer tail observations are retained. H Robustness to Graph Structure We examine whethe… view at source ↗
Figure 6
Figure 6. Figure 6: Regularization parameter selection for the S&P 500 application. Left panel shows [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
read the original abstract

Causal discovery in multivariate extremes is challenging because extreme observations are sparse, dependent, and often affected by latent common shocks. Existing approaches focus on undirected extremal dependence, require prior graph restriction, and do not scale beyond small systems. We introduce tail-induced asymmetry as a principle for causal directionality in heavy-tailed systems, where extreme events propagate asymmetrically so that forward tail prediction is systematically easier than backward prediction. We show that this asymmetry yields identifiable causal direction under a canonical max-linear model and provides a basis for score-based structure learning in the tail regime. Building on this, we propose Sparse Structure diScovery in Multivariate Extremes (S3ME), a two-stage data-driven framework for causal discovery. The first stage performs proxy-adjusted penalized neighbourhood selection to recover a sparse candidate skeleton under latent confounding. The second stage orients edges by minimizing tail prediction risk based on max-linear envelope models, exploiting directional asymmetry. We establish high-dimensional guarantees for skeleton screening and consistency of the score-based estimator under population separation conditions. Simulations demonstrate robustness to latent confounding and favourable scaling relative to existing extremal methods. Applications to river network data and financial tail-risk networks show that the approach recovers sparse, interpretable propagation structures without prespecified graph structure.

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

3 major / 1 minor

Summary. The paper introduces tail-induced asymmetry as a principle for causal directionality in heavy-tailed multivariate systems under a canonical max-linear model. It proposes the S3ME framework: a first stage of proxy-adjusted penalized neighborhood selection to recover a sparse skeleton under latent confounding, followed by a second stage that orients edges by minimizing tail prediction risk using max-linear envelope models. The manuscript claims high-dimensional guarantees for skeleton screening and consistency of the score-based estimator under population separation conditions, with supporting simulations and applications to river networks and financial tail-risk data.

Significance. If the identifiability result and consistency guarantees hold, the work would represent a meaningful advance in causal discovery for extremes, where existing methods are limited by sparsity, dependence, and lack of scalability. The exploitation of directional tail asymmetry for orientation without prespecified graph structure addresses a genuine gap and could enable interpretable propagation modeling in environmental and financial applications.

major comments (3)
  1. [Abstract] Abstract: the central consistency claim for the score-based orientation estimator is stated to hold only under population separation conditions (true orientations yield strictly lower tail prediction risk than alternatives), yet the manuscript supplies neither an explicit definition of these conditions, a diagnostic for their violation, nor a robustness analysis when they fail due to latent confounding, finite tails, or perturbations in max-linear coefficients.
  2. [Abstract] Abstract: the orientation step minimizes tail prediction risk based on max-linear envelope models that exploit the same tail asymmetry used to establish identifiability; without the full equations it is unclear whether the score has independent grounding or reduces to a fitted quantity defined from the data, raising a potential circularity concern for the consistency guarantee.
  3. [Simulations] Simulations section: robustness to latent confounding and favorable scaling are claimed, but the manuscript provides neither the simulation code, data, nor error-bar information, preventing assessment of whether the reported performance actually supports the high-dimensional guarantees.
minor comments (1)
  1. [Abstract] The abstract and introduction would benefit from a brief statement of the precise form of the canonical max-linear model and the tail asymmetry property to make the identifiability claim immediately accessible.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment below and indicate the revisions we will make to improve clarity and reproducibility.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central consistency claim for the score-based orientation estimator is stated to hold only under population separation conditions (true orientations yield strictly lower tail prediction risk than alternatives), yet the manuscript supplies neither an explicit definition of these conditions, a diagnostic for their violation, nor a robustness analysis when they fail due to latent confounding, finite tails, or perturbations in max-linear coefficients.

    Authors: The population separation conditions are formally defined in Section 3.2 as the requirement that the true orientation yields strictly lower tail prediction risk than alternatives under the max-linear model, following from Theorem 1. In the revision we will add a concise definition to the abstract, introduce an empirical diagnostic based on the observed risk gap, and include additional simulation results examining performance under perturbations to max-linear coefficients and moderate latent confounding. revision: yes

  2. Referee: [Abstract] Abstract: the orientation step minimizes tail prediction risk based on max-linear envelope models that exploit the same tail asymmetry used to establish identifiability; without the full equations it is unclear whether the score has independent grounding or reduces to a fitted quantity defined from the data, raising a potential circularity concern for the consistency guarantee.

    Authors: The tail prediction risk is derived directly from the max-linear envelope model and the identifiability result in Theorem 2; it is evaluated on held-out extreme observations via cross-validation and is therefore independent of the in-sample fitting used for skeleton recovery. We will revise the abstract to state this distinction explicitly and reference the out-of-sample evaluation procedure. revision: yes

  3. Referee: [Simulations] Simulations section: robustness to latent confounding and favorable scaling are claimed, but the manuscript provides neither the simulation code, data, nor error-bar information, preventing assessment of whether the reported performance actually supports the high-dimensional guarantees.

    Authors: We agree that error bars and reproducibility materials are needed. The revised manuscript will add standard error bars (computed over 50 replications) to all simulation figures. We will also release the simulation code and datasets via a public GitHub repository linked in the paper. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained under model assumptions

full rationale

The paper derives identifiability of causal directions from tail asymmetry under the canonical max-linear model as a first-principles result, then builds a two-stage estimator (proxy-adjusted skeleton screening followed by score-based orientation via tail prediction risk on max-linear envelopes) that operationalizes the same property. Consistency guarantees are stated to hold under explicitly listed population separation conditions that follow from the identifiability theorem rather than being presupposed. No quoted step reduces a claimed prediction or theorem to its inputs by construction, self-definition, or a load-bearing self-citation chain. The framework supplies independent content (high-dimensional screening rates, score minimization procedure) beyond restating the input asymmetry or model.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim depends on the max-linear model and separation conditions being true; these are domain assumptions not independently evidenced in the abstract. No free parameters or invented entities are explicitly listed, but the penalized selection step implies at least one tuning parameter.

free parameters (1)
  • penalty parameter for neighbourhood selection
    Used in the proxy-adjusted penalized step to recover the sparse skeleton; its value is chosen to achieve the claimed high-dimensional screening properties.
axioms (2)
  • domain assumption Data follows a canonical max-linear model
    Invoked to establish that tail asymmetry yields identifiable causal directions.
  • ad hoc to paper Population separation conditions hold for the score-based estimator
    Required for consistency of edge orientation; stated as a condition rather than derived.

pith-pipeline@v0.9.0 · 5512 in / 1458 out tokens · 40493 ms · 2026-05-09T21:18:48.287723+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

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

  1. [1]

    On random graphs

    Erd. On random graphs. Publicationes Mathematicae Debrecen , year =

  2. [2]

    Emergence of scaling in random networks , journal =

    Barab. Emergence of scaling in random networks , journal =. 1999 , volume =

    1999

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    Rina Foygel and Mathias Drton , title =. Advances in Neural Information Processing Systems , volume =. 2010 , pages =

    2010

  4. [4]

    Wadsworth , title =

    Raphaël Huser and Jennifer L. Wadsworth , title =. WIREs Computational Statistics , year =

  5. [5]

    Huber , title =

    Peter J. Huber , title =. Annals of Mathematical Statistics , year =

  6. [6]

    Hampel and Elvezio M

    Frank R. Hampel and Elvezio M. Ronchetti and Peter J. Rousseeuw and Werner A. Stahel , title =. 1986 , address =

    1986

  7. [7]

    2009 , edition =

    Trevor Hastie and Robert Tibshirani and Jerome Friedman , title =. 2009 , edition =

    2009

  8. [8]

    Wainwright , title =

    Martin J. Wainwright , title =. 2019 , series =

    2019

  9. [9]

    Journal of the American Statistical Association , year =

    Daniela Castro‑Camilo and Raphaël Huser , title =. Journal of the American Statistical Association , year =

  10. [10]

    Davison and Sebastian Engelke , title =

    Peiman Asadi and Anthony C. Davison and Sebastian Engelke , title =. Annals of Applied Statistics , year =

  11. [11]

    Advances in Neural Information Processing Systems , series =

    Yuning You and Tianlong Chen and Yongduo Sui and Ting Chen and Zhangyang Wang and Yang Shen , title =. Advances in Neural Information Processing Systems , series =. 2020 , volume =

    2020

  12. [12]

    Hitz , title =

    Sebastian Engelke and Adrien S. Hitz , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

  13. [13]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

    Sebastian Engelke and Stanislav Volgushev , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , year =

  14. [14]

    Journal of Machine Learning Research , year =

    Sebastian Engelke and Armeen Taeb , title =. Journal of Machine Learning Research , year =

  15. [15]

    Journal of the American Statistical Association , year =

    Manuel Hentschel and Sebastian Engelke and Johan Segers , title =. Journal of the American Statistical Association , year =

  16. [16]

    Journal of Multivariate Analysis , year =

    Claudia Klüppelberg and Mario Krali , title =. Journal of Multivariate Analysis , year =

  17. [17]

    Econometrics and Statistics , year =

    Nadine Gissibl and Claudia Kl\"uppelberg and Moritz Otto , title =. Econometrics and Statistics , year =

  18. [18]

    Dupuis , title =

    Linda Mhalla and Valérie Chavez‑Demoulin and Debbie J. Dupuis , title =. Journal of the Royal Statistical Society: Series C (Applied Statistics) , year =

  19. [19]

    Advances in Applied Probability , year =

    Stefka Asenova and Johan Segers , title =. Advances in Applied Probability , year =

  20. [20]

    Estimating a directed tree for extremes , journal =

    Ngoc Mai Tran and Johannes Buck and Claudia Kl. Estimating a directed tree for extremes , journal =. 2024 , volume =

    2024

  21. [21]

    Pasche and Val\'erie Chavez‑Demoulin and Anthony C

    Olivier C. Pasche and Val\'erie Chavez‑Demoulin and Anthony C. Davison , title =. Extremes , year =

  22. [22]

    Identifiability and estimation of recursive max-linear models , journal =

    Nadine Gissibl and Claudia Kl. Identifiability and estimation of recursive max-linear models , journal =. 2021 , volume =

    2021

  23. [23]

    High-dimensional graphs and variable selection with the

    Meinshausen, Nicolai and B. High-dimensional graphs and variable selection with the. The Annals of Statistics , volume=. 2006 , publisher=

    2006

  24. [24]

    Journal of Machine Learning Research , year =

    Chickering, David Maxwell , title =. Journal of Machine Learning Research , year =

  25. [25]

    , title =

    Wainwright, Martin J. , title =. IEEE Transactions on Information Theory , year =

  26. [26]

    The Annals of Statistics , year =

    Learning extremal graphical structures in high dimensions , author =. The Annals of Statistics , year =

  27. [27]

    2000 , edition =

    Causation, Prediction, and Search , author =. 2000 , edition =

    2000

  28. [28]

    A linear non-

    Shimizu, Shohei and Hoyer, Patrik O and Hyv. A linear non-. Journal of Machine Learning Research , volume=

  29. [29]

    and Xing, Eric P

    Zheng, Xun and Aragam, Bryon and Ravikumar, Pradeep K. and Xing, Eric P. , booktitle =. 2018 , publisher =

    2018

  30. [30]

    2007 , publisher=

    Heavy-Tail Phenomena: Probabilistic and Statistical Modeling , author=. 2007 , publisher=

    2007

  31. [31]

    2004 , publisher =

    Statistics of Extremes: Theory and Applications , author =. 2004 , publisher =

    2004

  32. [32]

    arXiv preprint arXiv:2508.00223 , year =

    Structural Causal Models for Extremes: an Approach Based on Exponent Measures , author =. arXiv preprint arXiv:2508.00223 , year =. 2508.00223 , archivePrefix =

    work page internal anchor Pith review arXiv

  33. [33]

    The Annals of Statistics , volume =

    Causal discovery in heavy-tailed models , author =. The Annals of Statistics , volume =. 2021 , doi =

    2021

  34. [34]

    Network Science: An Aerial View , editor =

    Bayesian Networks for Max-Linear Models , author =. Network Science: An Aerial View , editor =. 2019 , pages =

    2019

  35. [35]

    Scandinavian Journal of Statistics , year =

    Causal discovery in heavy-tailed linear structural equation models via scalings , author =. Scandinavian Journal of Statistics , year =

  36. [36]

    Extremes , volume =

    Causal modelling of heavy-tailed variables and confounders with application to river flow , author =. Extremes , volume =

  37. [37]

    2011 , publisher=

    Statistics for High-Dimensional Data: Methods, Theory and Applications , author=. 2011 , publisher=

    2011