Structural Causal Models for Extremes: an Approach Based on Exponent Measures

Fei Fang; Shuyang Bai; Tiandong Wang

arxiv: 2508.00223 · v3 · submitted 2025-08-01 · 🧮 math.ST · stat.ME· stat.TH

Structural Causal Models for Extremes: an Approach Based on Exponent Measures

Shuyang Bai , Fei Fang , Tiandong Wang This is my paper

Pith reviewed 2026-05-19 02:06 UTC · model grok-4.3

classification 🧮 math.ST stat.MEstat.TH

keywords extremal structural causal modelsexponent measuresextreme value theorycausal inferenceextremal conditional independencecausal identifiabilitymultivariate extremesactivation variables

0 comments

The pith

Extremal structural causal models based on exponent measures identify causal directions via inherent asymmetry.

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

The paper introduces extremal structural causal models that replace ordinary probability distributions with exponent measures to describe dependence structures among extreme events. These models incorporate activation variables that capture the single-big-jump principle together with extra randomization to generate a rich family of behaviors. The construction is shown to include every possible law of a directed graphical model that satisfies extremal conditional independence. Under natural conditions the models display an asymmetry that makes causal directions identifiable from data, and the authors supply a method that exploits this asymmetry.

Core claim

The authors define the extremal structural causal model (eSCM) using an exponent measure and activation variables. This model encompasses all possible laws of directed graphical models under the notion of extremal conditional independence. Under natural assumptions, eSCMs exhibit an inherent asymmetry that enables the identification of causal directions.

What carries the argument

The exponent measure, an infinite-mass law arising in multivariate extremes, together with activation variables that abstract the single-big-jump principle.

If this is right

Every directed graphical model consistent with extremal conditional independence can be represented as an eSCM.
The inherent asymmetry under natural conditions makes causal directions identifiable.
A discovery procedure that exploits the asymmetry recovers causal structure from extreme data.
The procedure recovers the correct directions on both simulated and real datasets.

Where Pith is reading between the lines

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

The framework could be used to improve causal analysis of tail risks in finance, climate, or network reliability.
One could examine how the asymmetry behaves when the threshold defining extremes is varied continuously.
Applying the method to additional real-world extreme-event records would test its robustness beyond the reported examples.

Load-bearing premise

Natural conditions exist under which eSCMs possess an inherent asymmetry that permits identification of causal directions.

What would settle it

A counterexample consisting of an eSCM that meets the natural assumptions yet displays no identifiable asymmetry in its exponent measure or activation structure would refute the identifiability result.

read the original abstract

We introduce a new formulation of structural causal models for extremes, called the extremal structural causal model (eSCM). Unlike conventional structural causal models, where randomness is governed by a probability distribution, eSCMs use an exponent measure, an infinite-mass law that naturally arises in the analysis of multivariate extremes. Central to this framework are activation variables, which abstract the single-big-jump principle, along with additional randomization that enriches the class of eSCM laws. This formulation encompasses all possible laws of directed graphical models under the recently introduced notion of extremal conditional independence. We also identify an inherent asymmetry in eSCMs under natural assumptions, enabling the identifiability of causal directions, a central challenge in causal inference. Finally, we propose a method that utilizes this causal asymmetry and demonstrate its effectiveness in both simulated and real datasets.

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.

Desk Editor's Note private letter to a colleague

The paper defines extremal SCMs via exponent measures and activation variables, claims full coverage of extremal conditional independence, and uses an asymmetry for causal identifiability under natural assumptions.

read the letter

The core idea here is replacing ordinary probability distributions in structural causal models with exponent measures to handle multivariate extremes, plus activation variables that encode the single-big-jump principle. This setup is meant to represent every possible law of a directed graph under the notion of extremal conditional independence, and the authors point to an inherent asymmetry that lets you recover causal directions once some natural conditions hold. They also give a practical method that exploits the asymmetry and test it on simulated and real data. That combination of a new modeling language for tails and an identifiability route is the main novelty. It sits at the intersection of extreme-value theory and causal graphical models, and the data checks are a plus because they show the method is not purely formal. The soft spot is the reliance on those natural assumptions for the asymmetry result. The abstract does not list them explicitly, so the paper needs to make clear whether they are mild regularity conditions on the exponent measure or something stronger that rules out interesting cases. If the assumptions are broad and the proofs close the gaps, the identifiability claim holds up; if they turn out narrow, the result applies only to a subclass. The coverage statement likewise needs the technical steps to be fully spelled out and checked. Overall this is aimed at people working on tail dependence in climate, finance, or reliability who also care about causal structure. A reader already comfortable with exponent measures and graphical models will get the most out of it. I would bring it to a reading group to walk through the assumptions and the proofs. It deserves peer review because the problem is real and the framework is a coherent attempt to solve it, even if revisions on the conditions and derivations are likely.

Referee Report

2 major / 2 minor

Summary. The paper introduces extremal structural causal models (eSCMs) that replace standard probability distributions with exponent measures to represent causal structures in multivariate extremes. Central elements include activation variables that encode the single-big-jump principle together with additional randomization. The framework is claimed to encompass all laws of directed graphical models compatible with extremal conditional independence. Under natural assumptions the models exhibit an inherent asymmetry that permits identifiability of causal directions; a practical method exploiting this asymmetry is proposed and evaluated on simulated and real data.

Significance. If the technical claims are established, the work would supply a principled extension of structural causal models to the infinite-mass setting of extreme-value theory and address a recognized identifiability obstacle in causal inference for tails. The explicit use of exponent measures and the introduction of activation variables constitute a technically coherent adaptation. Empirical demonstrations on both synthetic and real datasets provide concrete evidence of applicability.

major comments (2)

[Abstract and identifiability section] Abstract (final paragraph) and the section defining the identifiability result: the claim that eSCMs possess an 'inherent asymmetry' under 'natural assumptions' that enables causal-direction identifiability is load-bearing. The precise content of these assumptions (regularity conditions on the exponent measure, properties of activation variables, or restrictions on the directed graph) is not stated explicitly. Without an enumerated list and a discussion of necessity or counter-examples, it is impossible to assess whether the identifiability result holds for the full class of multivariate extreme laws or only for a narrower subclass.
[Encompassing theorem section] Section establishing the encompassing property: the assertion that the eSCM formulation 'encompasses all possible laws of directed graphical models under extremal conditional independence' requires a complete proof. The abstract supplies no derivations; the manuscript must contain a self-contained argument showing that every law satisfying the extremal conditional independence axioms arises from some eSCM, including verification that the additional randomization does not inadvertently restrict the attainable class.

minor comments (2)

[Preliminaries] Early sections: introduce the precise definition and distributional properties of activation variables before they are used in the main constructions, to improve readability for readers unfamiliar with the single-big-jump principle.
[Numerical experiments] Simulation and real-data sections: report the exact parameter values or tail indices used in the simulated examples and the preprocessing steps applied to the real datasets so that the experiments are fully reproducible.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive report. The two major comments identify areas where greater explicitness and rigor will strengthen the manuscript. We address each point below and will incorporate the suggested clarifications and expansions in the revised version.

read point-by-point responses

Referee: [Abstract and identifiability section] Abstract (final paragraph) and the section defining the identifiability result: the claim that eSCMs possess an 'inherent asymmetry' under 'natural assumptions' that enables causal-direction identifiability is load-bearing. The precise content of these assumptions (regularity conditions on the exponent measure, properties of activation variables, or restrictions on the directed graph) is not stated explicitly. Without an enumerated list and a discussion of necessity or counter-examples, it is impossible to assess whether the identifiability result holds for the full class of multivariate extreme laws or only for a narrower subclass.

Authors: We agree that the assumptions underlying the identifiability result should be stated explicitly. In the revised manuscript we will add a dedicated subsection that enumerates the natural assumptions, specifying the required regularity conditions on the exponent measure, the properties of the activation variables, and any restrictions imposed on the directed graph. We will also include a discussion of necessity together with counter-examples showing when identifiability fails if the assumptions are dropped. These additions will clarify the precise scope of the result. revision: yes
Referee: [Encompassing theorem section] Section establishing the encompassing property: the assertion that the eSCM formulation 'encompasses all possible laws of directed graphical models under extremal conditional independence' requires a complete proof. The abstract supplies no derivations; the manuscript must contain a self-contained argument showing that every law satisfying the extremal conditional independence axioms arises from some eSCM, including verification that the additional randomization does not inadvertently restrict the attainable class.

Authors: We accept that a fully self-contained proof is required. The present manuscript contains an outline of the argument; we will expand the section to supply a complete, detailed proof that every law obeying the extremal conditional independence axioms is realizable by an eSCM. The proof will also verify explicitly that the additional randomization does not restrict the attainable class but instead allows all laws compatible with the axioms to be represented. This revision will render the encompassing result rigorous and self-contained. revision: yes

Circularity Check

0 steps flagged

No circularity: eSCM asymmetry derived from model definition without reduction to inputs or self-citation.

full rationale

The paper introduces eSCM via exponent measures and activation variables as a new formulation that encompasses laws under extremal conditional independence. The inherent asymmetry is presented as identified under natural assumptions, enabling identifiability as a derived consequence rather than a definitional equivalence or fitted input renamed as prediction. No load-bearing self-citation, uniqueness theorem from prior work, or ansatz smuggling is evident in the abstract or description; the construction appears mathematically independent and self-contained against external benchmarks for multivariate extremes.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The framework rests on standard properties of exponent measures and the single-big-jump principle; new entities are introduced without external falsifiable evidence.

axioms (2)

standard math Exponent measures govern the tail behavior of multivariate extremes
Invoked as the replacement for probability distributions in the eSCM definition.
domain assumption Extremal conditional independence is well-defined for the new model class
Required for the claim that eSCMs encompass all directed graphical models under this notion.

invented entities (2)

extremal structural causal model (eSCM) no independent evidence
purpose: Structural causal model adapted to extremes via exponent measures
Core new object introduced to unify graphical models for extremes.
activation variables no independent evidence
purpose: Abstract the single-big-jump principle in the causal mechanism
Introduced to enrich the class of eSCM laws.

pith-pipeline@v0.9.0 · 5672 in / 1269 out tokens · 35754 ms · 2026-05-19T02:06:58.273274+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We introduce a new formulation of structural causal models for extremes, called the extremal structural causal model (eSCM). ... randomness is governed by an exponent measure... activation variables, which abstract the single-big-jump principle
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Assumption 1. (Nonzero Activation.) ... Assumption 2. (Nonzero Parent Effect.) ... Proposition 2. ... Λ{u,v}(yu > 0, yv = 0) > 0 when u ∉ an(v)

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.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Causal Discovery in Multivariate Extremes via Tail Asymmetry
stat.ME 2026-04 unverdicted novelty 7.0

S3ME recovers sparse causal skeletons in multivariate extremes via proxy-adjusted penalized selection and orients edges by minimizing tail prediction risk under max-linear models, with high-dimensional consistency guarantees.
Extrapolation in Statistical Learning with Extreme Value Theory
stat.ML 2026-05 unverdicted novelty 2.0

A survey of recent methods that apply extreme value theory to enable extrapolation in statistical learning and machine learning.