Morillas-type transformations of copulas and stable tail dependence functions

Klaus Herrmann; Marius Hofert; M\'elina Mailhot; Nahid Sadr

Only power-law distortions preserve every tail-dependence function, and exactly one family of distortions preserves every extreme-value copula.

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 13:32 UTC pith:ZAOTFEPO

load-bearing objection The stdf characterization in Theorem 3.5 is genuine and the sampling construction is useful; the paper needs a proof repair for Corollary 3.10 and a domain fix in Example 3.6, not a change of results. the 2 major comments →

arxiv 2607.19079 v1 pith:ZAOTFEPO submitted 2026-07-21 stat.ME math.PR

Morillas-type transformations of copulas and stable tail dependence functions

Klaus Herrmann , Marius Hofert , M\'elina Mailhot , Nahid Sadr This is my paper

classification stat.ME math.PR MSC 62H0562G3260G7039B22

keywords Morillas-type transformationsstable tail dependence functionsextreme value copulasPexider functional equationmaximum domain of attractionArchimax copulasstochastic representationabsolute monotonicity

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.

This paper asks which 'reshape-the-copula' operations are safe to apply when the dependence structure will later be used for extremes. The operations studied are Morillas-type distortions: write a copula (or its tail-dependence function), apply a fixed increasing function to the result, and pre-invert each coordinate using that same function. The main result is a complete answer: a continuous increasing inner function preserves every stable tail dependence function if and only if it is a power g(x)=c x^γ with exponent at most one, and a distortion preserves every extreme-value copula if and only if it has the form D(u)=e^{-λ(-log u)^γ}. Along the way the paper supplies a stochastic representation and sampling algorithm for distorted copulas, and it proves that if the original copula converges to an extreme-value limit and the distortion is in a Weibull domain of attraction, the distorted copula converges to a distorted version of that same limit — with Gumbel–Hougaard parameters shifting from θ to θ/γ. The upshot is a parameter-free rule for when a distortion can be trusted in extremal dependence modeling.

Core claim

The central discovery is a complete, parameter-free characterization of admissible Morillas-type distortions in the extreme-value world. Theorem 3.5: for continuous strictly increasing g with g(x)→∞, the transformed function ℓ^g(x)=g(ℓ(g^{-1}(x_1),...,g^{-1}(x_d))) is a stable tail dependence function for every such ℓ iff g(x)=c x^γ, c>0, γ∈(0,1]. Corollary 3.10 translates this to copulas: a continuous strictly increasing D maps every extreme-value copula to an extreme-value copula iff D(u)=e^{-λ(-log u)^γ}, λ>0, γ∈(0,1]. The proof uses a domain-restricted Pexider functional equation whose continuous solutions are powers; the scale c is a nuisance parameter, and γ>1 is excluded by the indepe

What carries the argument

The machinery has four pieces: (1) the Morillas transform itself, a composition D(C(D^{-1}(u_1),...,D^{-1}(u_d))) for copulas and its analog g(ℓ(g^{-1}(x_1),...)) for stable tail dependence functions; (2) the stable tail dependence function ℓ, a homogeneous, unit-margin, fully alternating function whose D-norm representation ℓ(x)=E[max_j |x_j| W_j] ties it to extreme-value copulas via C(u)=e^{-ℓ(-log u)}; (3) the domain-restricted Pexider functional equation g(xy)=g(x)g(g^{-1}(1)y), whose continuous solutions are powers and which is the load-bearing step in Theorem 3.5; and (4) the representation of absolutely monotone distortions as probability generating functions D(u)=Σ p_n u^n, which yie

Load-bearing premise

The load-bearing premise is that the distortion must work for every stable tail dependence function (or every extreme-value copula); without that universality requirement non-power transformations can preserve specific structures, and the printed proof of the EVC corollary additionally depends on an absolute-monotonicity claim that is false for γ<1.

What would settle it

Evaluate ℓ^g at the all-ones vector for g(x)=x^2 and the independence stdf ℓ(x)=Σx_i: Theorem 3.5 predicts ℓ^g(1,...,1)=d^2>d, violating the universal upper bound ℓ(x)≤Σx_i; if a non-power g passed every stdf axiom on every test stdf, the theorem would be false. Separately, computing the second finite difference of e^{-(-log u)^{1/2}} near u=0 yields negative values, which disproves the absolute-monotonicity assertion used in the printed sufficiency proof of the EVC corollary.

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

If this is right

Any universal Morillas distortion of a stable tail dependence function must be a power; fitting or simulating extremal dependence with a non-power inner–outer transform is guaranteed to leave the class of stable tail dependence functions for some input.
The preserved extreme-value copulas form a closed parametric family under distortion: D_{λ,γ}(u)=e^{-λ(-log u)^γ} maps EVCs to EVCs, with λ only an irrelevant scaling, so the effective tuning parameter is γ∈(0,1].
If C is in the MDA of a Gumbel–Hougaard copula with parameter θ and D is an absolutely monotone distortion in the Weibull MDA with index γ≤θ, then C^D is in the MDA of the Gumbel–Hougaard copula with parameter θ/γ; in particular, asymptotic independence becomes Gumbel–Hougaard tail dependence with parameter 1/γ.
The zeta distortion example gives a family D_s with MDA index min{s−1,1}, so by choosing s∈(1,2] one can continuously tune the limiting tail dependence coefficient of a distorted copula.
The stochastic representation replaces sampling from a frailty distribution by sampling a discrete generator N and then evaluating D on maxima of N iid samples from C, which yields a novel sampling route for Archimedean and Archimax copulas when D is absolutely monotone.

Where Pith is reading between the lines

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

The printed proof of the EVC-preservation corollary relies on an auxiliary absolute-monotonicity lemma for D_{λ,γ}(u)=e^{-λ(-log u)^γ}; that lemma is not valid for γ<1 (the second finite difference near 0 turns negative), so as written the sufficiency proof needs repair — though the corollary is correct and follows directly from Theorem 3.5.
Example 3.6(1) asserts that g(x)=e^x preserves the comonotone stdf for all x≥0; because g^{-1}(y)=log y is undefined for y<1, that identity is only valid after restricting the domain to x≥1 (or after replacing g by a function invertible on all of [0,∞)).
Because the characterization is universal, it can be read as an identifiability test: if an estimated extremal model is produced by a Morillas-type distortion of a known stdf, any fitted non-power distortion signals that the model has left the class, while any power distortion is parameter-free up to the exponent γ.
Theorem 4.2 suggests a constructive recipe for tail engineering: choosing an absolutely monotone distortion with a heavy-tailed discrete generator (infinite mean, like the zeta example with s≤2) moves the limiting tail-dependence index γ below 1 and can convert asymptotic independence into Gumbel–Hougaard dependence; bounded-density distortions sit at γ=1 and leave the limiting EVC unchanged.

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

The stdf characterization in Theorem 3.5 is genuine and the sampling construction is useful; the paper needs a proof repair for Corollary 3.10 and a domain fix in Example 3.6, not a change of results.

read the letter

This one is worth reading. The genuinely new part is Theorem 3.5: a Pexider-style argument showing the only Morillas-type stdf-to-stdf transforms are monomials g(x)=c x^γ with γ∈(0,1]. It is a clean, first-principles, parameter-free characterization. The sampling representation in Prop. 2.7 — D(W_N) with N from the probability generating function of D — is simple and seems genuinely missing from the literature; it closes a practical gap and, as a byproduct, gives a frailty-free way to sample some Archimedean and Archimax copulas. Section 4's MDA analysis is coherent, and the zeta distortion example is a nice concrete illustration. I do not see circularity in the central derivation; it is an independent functional-equation argument.

The reader's report is right about the soft spots. The sufficiency proof of Corollary 3.10, as written, cites the authors' working paper for the claim that D_{λ,γ}(u)=e^{-λ(-log u)^γ} is absolutely monotone for γ∈(0,1]. That is false for γ<1 — γ=1/2 fails 2-absolute monotonicity — so the proof route through Morillas/Ressel is broken. The corollary still survives: taking g(x)=x^γ in Theorem 3.5 gives a valid stdf and hence CD as an EVC directly. So this is a proof repair, not a result change. Example 3.6(1) with g(x)=e^x also has a boundary issue at x=0 because g^{-1}(0) is undefined; it needs a domain convention. These are minor in proportion to the rest of the paper. The overlap with the authors' working paper (Herrmann et al. 2026) is worth clarifying — the EVC-preserving result may already be there — but the stdf characterization itself is this paper's own.

Verdict: the central result holds up, the paper is formally grounded, and there are no data or fitting games. I would send it to a serious EVT/copula referee and recommend conditional acceptance: fix the sufficiency proof of Corollary 3.10, patch the example, and clarify the relation to the working paper. I would cite Theorem 3.5 and Prop. 2.7 in my own work.

Referee Report

2 major / 3 minor

Summary. The paper studies Morillas-type copula transformations C^D(u)=D(C(D^{-1}(u))) and their analogues for stable tail dependence functions (stdfs). Section 2 provides a stochastic representation and sampling algorithm for absolutely monotone distortions via probability generating functions, together with uniform bounds for distorted copulas. Section 3 introduces the stdf distortion ℓ^g(x)=g(ℓ(g^{-1}(x_1),...,g^{-1}(x_d))) and proves Theorem 3.5: for a continuous strictly increasing g with g(x)→∞, ℓ^g is a stdf for every stdf ℓ if and only if g(x)=c x^γ with γ∈(0,1] and c>0. This yields Corollary 3.10, which characterizes EVC-to-EVC Morillas distortions as D(u)=e^{-λ(-log u)^γ}. Section 4 analyzes the maximum domain of attraction of C^D, showing that under D∈MDA(Ψ_γ) and d-absolute monotonicity, the limiting EVC is C^* distorted by e^{-(-log u)^γ}; applications to Gumbel–Hougaard and Archimedean/Archimax copulas are also given. The main functional-equation proof of Theorem 3.5 is self-contained and appears sound, but the sufficiency proof of Corollary 3.10 as written relies on a false absolute-monotonicity claim and needs repair.

Significance. If correct, Theorem 3.5 is a sharp, parameter-free characterization of admissible Morillas-type distortions in the extreme-value setting, and Corollary 3.10 gives the corresponding EVC-preserving distortions. The stochastic representation of Section 2 is simple and likely useful for simulation, and the MDA results provide explicit limiting EVCs under mild conditions. The central derivation is an independent functional-equation argument and is not circular. However, the submitted proof of Corollary 3.10 is invalid in its present form because D_{λ,γ} is not absolutely monotone for γ<1; the statement is nevertheless recoverable directly from Theorem 3.5. Once this proof is repaired, the central claims of the paper appear sound and the contribution is significant.

major comments (2)

[Appendix B, proof of Corollary 3.10] The sufficiency step asserts that D_{λ,γ}(u)=e^{-λ(-log u)^γ} is absolutely monotone of every order for γ∈(0,1], citing Herrmann et al. (2026, Corollary A.3). This is false for γ<1. For example, when γ=1/2, D''(u)<0 for all sufficiently small u, so D is not convex and hence not 2-absolutely monotone. Consequently Morillas/Ressel Theorem 2.1 cannot be invoked, and the proof as written does not establish sufficiency for the stated parameter range. The corollary's statement is nevertheless correct: with g(x)=-log D(e^{-x})=λx^γ, Theorem 3.5 shows that ℓ^g(x)=(ℓ(x^{1/γ}))^γ is a stdf, and a direct calculation gives C^D(u)=e^{-ℓ^g(-log u_1,...,-log u_d)}, which is an EVC. This direct proof should replace the invalid absolute-monotonicity argument and also removes the reliance on the self-cited working paper.
[Example 3.7] Example 3.7 claims that the convex combination ℓ=Σ_{k=1}^K α_k ℓ_k of stdfs is a stdf 'even without the condition Σα_k=1'. This is false: for a unit vector e_j, ℓ(e_j)=Σα_k, and the stdf marginal condition requires ℓ(e_j)=1. Thus the normalization Σα_k=1 (or an explicit rescaling) is necessary for C_ℓ(u)=e^{-ℓ(-log u)} to have uniform margins. The formulas in Example 3.7 and the proposed flexibility construction should be corrected accordingly.

minor comments (3)

[Example 3.6(1)] The function g(x)=e^x maps [0,∞) to [1,∞), so g^{-1} is undefined at 0. The displayed identity ℓ_M^g=ℓ_M therefore holds only on (0,∞)^d, or with an explicit limiting convention at the boundary. Please state the domain convention used.
[Throughout] Minor typos: 'stabe tail dependence functions' in the Section 3 heading; 'direct calcuation' before Algorithm 2.8; 'disucsses' in Remark 4.12.
[Section 4.2] The analysis of alternative rate functions is thorough but somewhat lengthy. Consider condensing or moving some of the details to the Supplementary Material to improve readability.

Circularity Check

0 steps flagged

Central Theorem 3.5 is an independent Pexider-equation derivation; the only self-citation is a technical monotonicity lemma in Corollary 3.10's sufficiency proof, which is a proof gap rather than a circular reduction.

full rationale

The central claim (Theorem 3.5) is an independent functional-equation derivation, not a fit: the paper starts with an arbitrary continuous increasing g, uses the independence stdf to force g(0)=0 and homogeneity, reduces the problem to a domain-restricted Pexider functional equation, and invokes Aczél's external classification; the power form g(x)=cx^γ is the output. Sufficiency is supported by external results (Falk 2019; Chatelain et al. 2020). Corollary 3.10's necessity is a direct reduction to Theorem 3.5 via g(x)=-log D(e^{-x}), so it is self-contained. The only load-bearing self-citation is in the sufficiency proof of Corollary 3.10 (Supplementary Material B): 'by the proof of Herrmann et al. (2026, Corollary A.3), D_{λ,γ} is absolutely montone of any order.' This is a working paper by overlapping authors and the lemma is not proved in the present paper; moreover it is false for γ<1 (e.g., D_{1,1/2} has negative second divided differences near 0), so the written proof has a real gap. This is a correctness/proof-support problem, however, not a circular reduction: the lemma is a parameter-free analytic claim, not a restatement of the target result, and the corollary can be proved directly from Theorem 3.5 via g(x)=x^γ. Example 3.6(1)'s boundary issue with g(x)=e^x is a domain convention glitch. Thus no prediction is equal to its input by construction; circularity is minimal (score 2).

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted. The main machinery is standard copula/extreme-value theory plus a functional equation theorem; the only non-standard input is a self-cited working-paper lemma used in one proof, which we flag.

axioms (4)

standard math Pexider functional equation solutions on (0,∞)×(1,d) are power functions (Aczél 1987, Cor 5)
Used in Theorem 3.5 necessity proof to solve g(xy)=g(x)g(g^{-1}(1)y).
standard math Ressel/Morillas theorem: d-absolutely monotone distortion D maps every copula to a copula (Theorem 2.1)
Basis for C^D being a copula.
ad hoc to paper D_{λ,γ}(u)=e^{-λ(-log u)^γ} is absolutely monotone of every order (Herrmann et al. 2026, Cor A.3)
Used in Cor 3.10 sufficiency proof; unproved in this paper and false for γ<1 as stated.
standard math Power distortions of stdfs are stdfs (Falk 2019; Chatelain et al. 2020)
Cited for the sufficiency direction of Theorem 3.5.

pith-pipeline@v1.3.0-alltime-deepseek · 37031 in / 33053 out tokens · 272173 ms · 2026-08-01T13:32:59.332610+00:00 · methodology

0 comments

read the original abstract

A stochastic representation and sampling algorithm for Morillas-type copula-to-copula transformations and related distortions of multivariate distribution functions is derived, resulting as a byproduct in a novel sampling scheme for Archimedean and Archimax copulas. This closes a methodological gap and facilitates simulation-based applications of distorted copulas. For stable tail dependence functions (stdfs), a Morillas-type distortion framework is introduced, where monomial distortions with exponents below 1 are shown to preserve stdfs via a domain-restricted Pexider equation analysis. This characterization is leveraged to identify distortions preserving extreme value copulas, and convex combinations of distorted stdfs are proposed to increase flexibility in extremal dependence modeling. The impact of distortions on maximum domain of attraction limits is also analyzed. Explicit limiting EVC distortions are identified under non-restrictive regular variation assumptions. Examples of absolutely monotone distortions allowing to fine-tune the extreme value behavior after distortion, but also of non-regularly varying 2-absolutely monotone distortions are given.

Morillas-type transformations of copulas and stable tail dependence functions

Core claim

What carries the argument

Load-bearing premise

What would settle it

If this is right

Where Pith is reading between the lines

discussion (0)