A Vector Space Approach to Heavy Tailed Analysis

Daniel Cooley; Kenneth Broadhead

REVIEW 3 major objections 6 minor 25 references

This paper builds a normed vector space of heavy-tailed random variables and shows that at tail index 2, the tail pairwise dependence measure is exactly the inner product inducing the norm.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 15:13 UTC pith:EBTLYQGK

load-bearing objection New vector space construction for tail equivalence classes is a real conceptual step, and the TPDM-as-inner-product result at α=2 is new; the main gaps are the triangle inequality proof for general elements and the joint-regular-variation assumption in the modeling section. the 3 major comments →

arxiv 2607.18505 v1 pith:EBTLYQGK submitted 2026-07-20 math.PR math.STstat.TH

A Vector Space Approach to Heavy Tailed Analysis

Kenneth Broadhead , Daniel Cooley This is my paper

classification math.PR math.STstat.TH MSC 60G7062G3246B2026A12

keywords regular variationheavy tailstail equivalence classestail pairwise dependence measureinner productnormed vector spaceextreme value analysislinear prediction

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's aim is to give extreme value analysis a linear-algebraic foundation rooted in univariate regular variation. It defines a vector space W_b whose elements are equivalence classes of random variables with identical asymptotic tail behavior under a fixed normalizing function b(s)=s^αL(s), and proves that this space carries a natural norm when α>1. On any subspace where all pairs are jointly regularly varying, the norm satisfies the parallelogram law at α=2, and the polarization identity shows the inducing inner product is exactly the tail pairwise dependence measure. The payoff, if the construction holds, is that familiar Hilbert-space machinery—projection, orthogonal decomposition, linear prediction—becomes available for modeling extremes, in direct analogy to classical covariance modeling.

Core claim

The central discovery is that tail behavior alone can define a normed vector space. Random variables are grouped into equivalence classes when their scaled tail probabilities coincide asymptotically—more precisely, when their difference is 'trivially normalized' so that the scaled tail limits are zero. The norm of a class is ||X||_b = (lim_{s→∞} b(s)P(|X|>s))^{1/α} = κ^{1/α}, the α-root of the total α-scale. For α>1 the triangle inequality is proved by passing to comonotone versions, where the scale of a sum is exactly additive and any other dependence yields a smaller scale via convex order. When α=2 and the two variables are jointly regularly varying, ||·||_b satisfies the parallelogram la

What carries the argument

The central machinery is the quotient space W_b = V_b/N_b: V_b collects random variables for which b(s)P(X>sx) and b(s)P(X<-sx) have finite limits, N_b is the subspace of variables with zero such limits (trivially normalized), and W_b consists of tail-equivalence classes. The norm is ||X||_b = κ_X^{1/α} where κ_X is the sum of the left and right limiting tail probabilities. The triangle inequality is established by two lemmas: convex order shows the comonotone sum dominates any other sum's α-scale, and quantile additivity of comonotone sums gives exact additivity of scales. For α=2, the angular-measure decomposition of the exponent measure yields the parallelogram law, and polarization ident

Load-bearing premise

The Hilbert-space conclusions for modeling depend on the observed variables being jointly regularly varying, not merely individually heavy-tailed; without joint regular variation the tail pairwise dependence measure need not be the inner product.

What would settle it

Take two random variables with identical Pareto marginal tails but a dependence structure that creates asymptotic dependence without multivariate regular variation; numerically evaluate ||X_1+X_2||^2 + ||X_1-X_2||^2 and 2||X_1||^2 + 2||X_2||^2. If they differ, the parallelogram law, and therefore the inner-product identification, requires joint regular variation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

On a finite-dimensional subspace of W_b spanned by jointly regularly varying variables with α=2, the tail pairwise dependence matrix is a Gram matrix, so existing TPDM estimators inherit inner-product geometry.
The projection theorem applies to these subspaces, giving a principled way to perform linear prediction of one extreme variable from others and to define partial tail dependence, mirroring partial correlation.
The restriction α=2 used in previous TPDM work is not merely convenient but necessary: for α≠2 the norm cannot be induced by an inner product, even on well-behaved subspaces.
The space D_b of absolutely summable infinite linear combinations of i.i.d. regularly varying innovations provides a Hilbert-space framework for heavy-tailed time series, enabling linear prediction for extremes of such series.

Where Pith is reading between the lines

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

If the normed space W_b is complete (an open question here), it would be a Banach space of tail equivalence classes; the outcome likely hinges on whether every tail-equivalent Cauchy sequence can be represented by a single regularly varying random variable.
The inner-product identification suggests a direct estimator for the TPDM on a subspace: the Gram matrix formed from norms of pairwise sums and differences, computed from tail empirical measures, should be positive semidefinite.
One could empirically test the construction by simulating jointly regularly varying variables with α=2 and checking whether the sample TPDM matrix is positive semidefinite and whether the parallelogram law holds; failure would signal that the joint regular variation assumption is genuinely load-bearing.
Because the construction treats left and right tails separately, the inner product uses the signed angular measure while the norm uses the total α-scale, suggesting that tail balance plays a subtle role in the geometry of extreme dependence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

New vector space construction for tail equivalence classes is a real conceptual step, and the TPDM-as-inner-product result at α=2 is new; the main gaps are the triangle inequality proof for general elements and the joint-regular-variation assumption in the modeling section.

read the letter

Worth your time. The core construction—V_b/N_b with the κ^{1/α} norm—is genuinely new, and the identification of the TPDM as the inner product induced by that norm at α=2 is a real explanation for why the α=2 restriction kept showing up in prior work. The authors also get credit for being upfront that completeness of W_b is unresolved and that the inner product only lives on jointly regularly varying subspaces.

What is new and good: the quotient space gives a clean language for tail equivalence; the norm theorem is plausible; the polar-coordinate proof of the parallelogram law (Theorem 4) and the polarization computation that lands on ∫ω1ω2 dH are elegant. The de Bruijn conjugacy machinery in Lemmas 3 and Proposition 4 is solid. The paper is honest that the full W_b inner product question remains open and that the Hilbert-space modeling framework depends on joint regular variation of the observed variables.

The soft spots, in order of importance. First, Theorem 3's proof of the triangle inequality (Theorem 2) invokes Lemma 2, which assumes regularly varying marginals, but the triangle inequality must hold for all X1,X2 in V_b, including trivially normalized elements. The reduction to positive variables and then to comonotone versions needs regular variation of |X1| and |X2| to apply Lemma 2; the zero-scale case is skipped. The claim probably survives—trivial elements should behave like zero—but as written the proof has a gap. Second, Corollary 2 asserts by inspection that α≠2 fails the parallelogram law; a concrete counterexample would make it convincing. Third, the modeling section's reach is narrower than the abstract suggests: Theorem 4 needs pairwise joint regular variation, and W_b membership alone does not supply it. The paper flags this with 'it seems reasonable to assume,' but the projection theorem for linear prediction applies only when that assumption holds; the discussion could overstate the generality. Minor: the proof of Theorem 2 contains a copy-paste slip in the indices of the second random variable.

Who it's for: extreme value theorists and statisticians who care about why TPDM works at α=2 and want a Hilbert-space framework for linear prediction of extremes. The central argument holds up; the gaps are addressable. I'd send it to a serious referee, expecting major or minor revision on the triangle inequality and scope statements.

Referee Report

3 major / 6 minor

Summary. The paper constructs a quotient vector space W_b from random variables whose tail probabilities are finite when normalized by a regularly varying function b(s)=s^α L(s), with equivalence determined by differences whose normalized tails vanish. It claims that W_b is a normed vector space for α>1, with norm given by the α-scale κ_X^{1/α} = [lim_s b(s)P(|X|>s)]^{1/α}. The paper further shows that for α=2 and for pairs that are jointly regularly varying, the norm satisfies the parallelogram law and the inducing inner product is the tail pairwise dependence measure, thereby providing a Hilbert-space framework for extreme-value modeling through finite-dimensional subspaces of jointly regularly varying variables.

Significance. If the main theorems are correct, the paper gives a novel and useful algebraic structure for heavy-tailed analysis: a normed space of tail-equivalence classes, an inner-product interpretation of an existing dependence measure, and a principled route to projection-based prediction for extremes. The work is largely self-contained, uses standard tools (Karamata theory, convex order, de Bruijn conjugacy, vague convergence), and the proofs of the parallelogram law and the TPDM identification are concrete and checkable rather than heuristic. The paper also makes the α=2 restriction in earlier TPDM work more transparent. However, the applicability of the Hilbert-space methods is explicitly conditional on joint regular variation, and the manuscript's presentation in places overstates the generality of the modeling framework.

major comments (3)

[§2.1, Definition 4 and Proposition 1] The proof of closure under addition does not establish the defining property of V_b. Definition 4 requires that lim_{s→∞} b(s)P(X+Y>sx) exists and is finite, but the proof only shows that the limit superior is bounded by the sum of two finite limits. For dependent random variables with regularly varying marginals, the normalized tail of a sum need not have a limit; joint regular variation is what guarantees such limits. Since V_b being a vector space is the foundation for W_b and for all subsequent claims, this is a load-bearing gap. The authors should either prove that the limit exists for X+Y (for example by classifying V_b elements and proving closure for sums) or modify the definition so that the construction is sound.
[§3.1, Theorem 2] The proof of the triangle inequality invokes Lemma 2, which is stated only for X1,X2 ≥ 0 that are regularly varying with normalizing function b. In Theorem 2, the positive variables |X1| and |X2| may be trivially normalized, i.e., have zero α-scale, in which case Lemma 2 does not apply. A separate argument is needed for this case, e.g., showing that if ||X||_b=0 then ||X+Y||_b ≤ ||Y||_b by splitting the tail event at thresholds (1−ε)s and εs. In addition, Lemma 2's proof assumes that the tail of X1+X2 is regularly varying; this is not automatic from the marginal regular-variation hypothesis and needs justification (possibly via Theorem 1) or a separate proof. Because Theorem 3 rests on Theorem 2, this is a central technical issue.
[§4.2 and §5] The Hilbert-space modeling claim is proved only under an additional joint-regular-variation assumption on the observed basis variables. Membership in W_b gives only univariate regular variation (or trivial normalization); it does not imply that the p-dimensional vector is multivariate regularly varying, so the exponent measure and angular measure used in Theorem 4 and Corollary 1 need not exist for a general heavy-tailed data vector. The text introduces this as 'it seems reasonable to assume,' but the abstract and Section 5 state more broadly that all Hilbert-space tools are 'now available for modeling extremes.' This overstates the reach of the framework. The joint-regular-variation condition should be stated explicitly as a hypothesis in the abstract and in the main modeling theorem, and the practical implications should be discussed.

minor comments (6)

[§2.2, Proposition 3 proof] There is an algebraic typo in the translation argument: S(aX+b) = |a|S(X+b/a), and the following line should have b/a, not a/b. The intended conclusion is clear but the notation should be corrected.
[§3.1, Theorem 2 proof] Near the end of the proof, the displayed inequality contains a repeated index: κ_{|X2|}^{1/α} appears twice; it should be κ_{|X1|}^{1/α} + κ_{|X2|}^{1/α}.
[§4.1, Corollary 1 proof] In the penultimate displayed equation, the integral over the complement of P3∪P4 is written as \int_{\backslash P3∪P4}; the intended symbol is \int_{S^1\setminus(P_3\cup P_4)}.
[§4.1, Corollary 2] The proof says only that the parallelogram expression 'will not simplify' for α≠2. Since this is a claim about all α≠2, it would be preferable to exhibit a concrete jointly regularly varying pair and a compactly supported angular measure for which the parallelogram law fails.
[References] The reference to Kiriliouk & Zhou (2024) is incomplete ('arXive Preprint x, xx–xx'). Please provide a complete citation or remove the placeholder.
[§3.2, Example 1] The statement that the partial sums 'converge to any element in N_b' is slightly misleading: the partial sums are themselves elements of N_b, so they form the zero sequence in W_b. The example is fine, but the wording should be tightened.

Circularity Check

0 steps flagged

No significant circularity: the norm and inner-product results are derived from the paper's own definitions, with the joint-regular-variation assumption an explicit scope condition rather than a hidden input.

full rationale

Walking the derivation chain: V_b is defined by finite b-normalized tail limits; N_b is the subspace of trivially normalized variables; W_b = V_b/N_b. Theorem 1 derives that nonzero classes are regularly varying from these definitions rather than assuming it. The norm ||X||_b = lim [b(s)P(|X|>s)]^{1/α} is shown to satisfy the triangle inequality for α>1 via convex order, comonotonic upper bounds, and quantile additivity (Theorems 2–3). The inner-product claim is then proved, not assumed: for α=2 and jointly regularly varying X1,X2, Theorem 4 computes ||X1±X2||^2 as integrals against the exponent/angular measure and verifies the parallelogram law; Corollary 1 obtains the polarization identity and shows it equals ∫_{S^1} ω1 ω2 dH_X, i.e., exactly the TPDM defined in (1.3.2). This is a direct reduction of the new claim to the standing definitions, not a circular restoration of an input. The only places the paper leans on the authors' previous TPDM work are terminological/contextual (Cooley & Thibaud 2019; Mhatre & Cooley 2024), and the definition is reproduced in the paper; the load-bearing span property in §4.2 cites the external Basrak et al. (2002, Prop. A.1). The modeling section explicitly flags “it seems reasonable to assume that they are jointly regularly varying” as an assumption restricting the modeling framework; this limits scope but is not circularity because the paper does not claim W_b membership alone implies joint regular variation. The open completeness question is explicitly acknowledged as unresolved, and the α≠2 no-inner-product corollary is not used to support the main identification. No fitted parameters are renamed as predictions. Hence score 0.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No empirical parameters are fitted. The normalizing function b(s)=s^α L(s) and tail index α are inputs chosen by the modeler, not numbers derived from data. The vector space W_b is a new mathematical object but not a new physical entity requiring external falsifiable evidence. The main unproved inputs are standard theorems from regular variation and convex order, plus the domain assumption that observed variables are jointly regularly varying.

axioms (4)

standard math Bingham/Karamata theory of regular variation, including de Bruijn conjugacy and tail-quantile asymptotic representation.
Invoked in Lemma 3 and Proposition 4; cited to Beirlant et al. (2005) but not proved in the paper.
standard math Comonotonicity gives the convex-largest sum when marginals have finite means.
Used in Lemma 2 to bound the scale of a sum by that of the comonotone sum; cited to Kaas et al. (2002) and Shaked & Shanthikumar (2007).
standard math Linear combinations of jointly regularly varying random vectors are jointly regularly varying (in non-degenerate cases).
Used in §4.2 for the subspace S_p and the infinite series space D_b; cited to Basrak et al. (2002) proposition A.1 and Kulik & Soulier (2020).
domain assumption The observed basis variables X_1,...,X_p are jointly regularly varying.
This is the modeling assumption that makes the TPDM an inner product on S_p; stated explicitly in §4.2 as 'it seems reasonable to assume'.

pith-pipeline@v1.3.0-alltime-deepseek · 18722 in / 25159 out tokens · 251796 ms · 2026-08-01T15:13:09.133922+00:00 · methodology

0 comments

read the original abstract

We construct a vector space whose defining characteristics are rooted in univariate regular variation of random variables. Specifically, the base vector space $\mathbb{V}_b$ consists of random variables whose limiting tail probabilities, when scaled by regularly varying functions of the form $b(s)=s^\alpha L(s)$, are finite. Defining a subspace ${\cal N}_b$ corresponding to random variables in $\mathbb{V}_b$ whose limiting tail probabilities are zero when normalized by $b(s)$ allows the base space $\mathbb{V}_b$ to be partitioned into equivalence classes. We define a vector space $\mathbb{W}_b$ consisting of these equivalence classes, and show its nonzero elements are equivalence classes of regularly varying random variables. We show that a natural norm exists for $\mathbb{W}_b$ if $\alpha > 1$. We show that the equivalence classes and convergence in norm are different than more familiar vector spaces of random variables. Turning our attention to extreme value modeling, we consider finite-dimensional subspaces of $\mathbb{W}_b$ whose basis vectors are jointly regularly varying. We show that in the case $\alpha = 2$, the previously defined tail pairwise dependence measure serves as an inner product. As any finite-dimensional space is complete, we can use the projection theorem to perform linear prediction.

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

[1]

2020 , publisher =

Heavy Tailed Time Series , author =. 2020 , publisher =

2020
[2]

Regular Variation for Measures on Metric Spaces , author=. Publ. Inst. Math. (Beograd) (N.S.) , year=
[3]

Point processes, Regular Variation and Weak Convergence , author=. Publ. Inst. Math. (Beograd) (N.S.) , year=
[4]

2007 , publisher =

Heavy-Tail Phenomena , author =. 2007 , publisher =

2007
[5]

Statistics and Probability Letters , year=

A Note on Vague Convergence of Measures , author=. Statistics and Probability Letters , year=
[6]

2023 , publisher =

Closure Properties for Heavy-Tailed and Related Distributions , author =. 2023 , publisher =

2023
[7]

2005 , publisher =

Statistics of Extremes Theory and Applications , author =. 2005 , publisher =

2005
[8]

1986 , publisher =

Extreme Values, Regular Variation, and Point Processes , author =. 1986 , publisher =

1986
[9]

1997 , publisher =

Modeling Extreme Events for Insurance and Finance , author =. 1997 , publisher =

1997
[10]

2007 , publisher =

Stochastic Orders , author =. 2007 , publisher =

2007
[11]

ASTIN Bulletin , year=

A Simple Geometric Proof that Comonotonic Risks Have the Convex-Largest Sum , author=. ASTIN Bulletin , year=
[12]

Stochastic Processes and their Applications , year=

Regular Variation of GARCH Processes , author=. Stochastic Processes and their Applications , year=
[13]

Biometrika , year=

Decompositions of dependence for high-dimensional extremes , author=. Biometrika , year=
[14]

Extremes , year=

Extremal dependence measure and extremogram: the regularly varying case , author=. Extremes , year=
[15]

arXive Preprint , year=

Estimating probabilities of multivariate failure sets based on pairwise tail dependence coefficients , author=. arXive Preprint , year=
[16]

Stochastic Models , year=

The Extremal Dependence Measure and Asymptotic Independence , author=. Stochastic Models , year=
[17]

Environmetrics , volume=

Simultaneous autoregressive models for spatial extremes , author=. Environmetrics , volume=. 2021 , publisher=

2021
[18]

and Cooley, D

Rohrbeck, C. and Cooley, D. , year =. Simulating Flood Event Sets Using Extremal Principal Components , journal =
[19]

Journal of Time Series Analysis , volume=

Transformed-Linear Models for Time Series Extremes , author=. Journal of Time Series Analysis , volume=. 2024 , publisher=

2024
[20]

Transformed-linear prediction for extremes , publisher =

Lee, Jeongjin and Cooley, Daniel , keywords =. Transformed-linear prediction for extremes , publisher =. 2022 , copyright =. doi:10.48550/ARXIV.2111.03754 , note =

work page doi:10.48550/arxiv.2111.03754 2022
[21]

Electronic Journal of Statistics , volume=

Principal component analysis for multivariate extremes , author=. Electronic Journal of Statistics , volume=
[22]

2009 , publisher =

Robust Statistics , author =. 2009 , publisher =

2009
[23]

Annals of Actuarial Science , volume=

Diversification in Heavy-Tailed Portfolios: Properties and Pitfalls , author=. Annals of Actuarial Science , volume=
[24]

Publications de I'institut mathematique , volume=

Regularly Varying Functions , author=. Publications de I'institut mathematique , volume=
[25]

Bernoulli , volume=

Tail probabilities for infinite series of regularly varying random vectors , author=. Bernoulli , volume=

[1] [1]

2020 , publisher =

Heavy Tailed Time Series , author =. 2020 , publisher =

2020

[2] [2]

Regular Variation for Measures on Metric Spaces , author=. Publ. Inst. Math. (Beograd) (N.S.) , year=

[3] [3]

Point processes, Regular Variation and Weak Convergence , author=. Publ. Inst. Math. (Beograd) (N.S.) , year=

[4] [4]

2007 , publisher =

Heavy-Tail Phenomena , author =. 2007 , publisher =

2007

[5] [5]

Statistics and Probability Letters , year=

A Note on Vague Convergence of Measures , author=. Statistics and Probability Letters , year=

[6] [6]

2023 , publisher =

Closure Properties for Heavy-Tailed and Related Distributions , author =. 2023 , publisher =

2023

[7] [7]

2005 , publisher =

Statistics of Extremes Theory and Applications , author =. 2005 , publisher =

2005

[8] [8]

1986 , publisher =

Extreme Values, Regular Variation, and Point Processes , author =. 1986 , publisher =

1986

[9] [9]

1997 , publisher =

Modeling Extreme Events for Insurance and Finance , author =. 1997 , publisher =

1997

[10] [10]

2007 , publisher =

Stochastic Orders , author =. 2007 , publisher =

2007

[11] [11]

ASTIN Bulletin , year=

A Simple Geometric Proof that Comonotonic Risks Have the Convex-Largest Sum , author=. ASTIN Bulletin , year=

[12] [12]

Stochastic Processes and their Applications , year=

Regular Variation of GARCH Processes , author=. Stochastic Processes and their Applications , year=

[13] [13]

Biometrika , year=

Decompositions of dependence for high-dimensional extremes , author=. Biometrika , year=

[14] [14]

Extremes , year=

Extremal dependence measure and extremogram: the regularly varying case , author=. Extremes , year=

[15] [15]

arXive Preprint , year=

Estimating probabilities of multivariate failure sets based on pairwise tail dependence coefficients , author=. arXive Preprint , year=

[16] [16]

Stochastic Models , year=

The Extremal Dependence Measure and Asymptotic Independence , author=. Stochastic Models , year=

[17] [17]

Environmetrics , volume=

Simultaneous autoregressive models for spatial extremes , author=. Environmetrics , volume=. 2021 , publisher=

2021

[18] [18]

and Cooley, D

Rohrbeck, C. and Cooley, D. , year =. Simulating Flood Event Sets Using Extremal Principal Components , journal =

[19] [19]

Journal of Time Series Analysis , volume=

Transformed-Linear Models for Time Series Extremes , author=. Journal of Time Series Analysis , volume=. 2024 , publisher=

2024

[20] [20]

Transformed-linear prediction for extremes , publisher =

Lee, Jeongjin and Cooley, Daniel , keywords =. Transformed-linear prediction for extremes , publisher =. 2022 , copyright =. doi:10.48550/ARXIV.2111.03754 , note =

work page doi:10.48550/arxiv.2111.03754 2022

[21] [21]

Electronic Journal of Statistics , volume=

Principal component analysis for multivariate extremes , author=. Electronic Journal of Statistics , volume=

[22] [22]

2009 , publisher =

Robust Statistics , author =. 2009 , publisher =

2009

[23] [23]

Annals of Actuarial Science , volume=

Diversification in Heavy-Tailed Portfolios: Properties and Pitfalls , author=. Annals of Actuarial Science , volume=

[24] [24]

Publications de I'institut mathematique , volume=

Regularly Varying Functions , author=. Publications de I'institut mathematique , volume=

[25] [25]

Bernoulli , volume=

Tail probabilities for infinite series of regularly varying random vectors , author=. Bernoulli , volume=