Abstract
A bivariate random vector can exhibit either asymptotic independence or dependence between the largest values of its components. When used as a statistical model for risk assessment in fields such as finance, insurance or meteorology, it is crucial to understand which of the two asymptotic regimes occurs. Motivated by their ubiquity and flexibility, we consider the extremal dependence properties of vectors with a random scale construction (X1,X2) = R(W1,W2), with non-degenerate R > 0 independent of (W1,W2). Focusing on the presence and strength of asymptotic tail dependence, as expressed through commonly-used summary parameters, broad factors that affect the results are: the heaviness of the tails of R and (W1,W2), the shape of the support of (W1,W2), and dependence between (W1,W2). When R is distinctly lighter tailed than (W1,W2), the extremal dependence of (X1,X2) is typically the same as that of (W1,W2), whereas similar or heavier tails for R compared to (W1,W2) typically result in increased extremal dependence. Similar tail heavinesses represent the most interesting and technical cases, and we find both asymptotic independence and dependence of (X1,X2) possible in such cases when (W1,W2) exhibit asymptotic independence. The bivariate case often directly extends to higher-dimensional vectors and spatial processes, where the dependence is mainly analyzed in terms of summaries of bivariate sub-vectors. The results unify and extend many existing examples, and we use them to propose new models that encompass both dependence classes.
Article PDF
Similar content being viewed by others
Avoid common mistakes on your manuscript.
References
Abdous, B., Fougères, A. -L., Ghoudi, K.: Extreme behaviour for bivariate elliptical distributions. Can J Stat 33(3), 317–334 (2005)
Asmussen, S., Hashorva, E., Laub, P.J., Taimre, T.: Tail asymptotics of light-tailed Weibull-like sums. Probab. Math. Stat. 37(2), 235–256 (2017). https://doi.org/10.19195/0208-4147.37.2.3
Balkema, G., Embrechts, P.: High risk scenarios and extremes: A geometric approach. Zürich Lectures in Advanced Mathematics. European Mathematical Society, Zürich (2007)
Balkema, G., Nolde, N.: Asymptotic independence for unimodal densities. Adv. Appl. Probab. 42(2), 411–432 (2010)
Beirlant, J., Goegebeur, Y., Segers, J., Teugels, J.L: Statistics of extremes. Wiley, New York (2004)
Belzile, L., Nešlehová, J.: Extremal attractors of Liouville copulas. J. Multivar. Anal. 160, 68–92 (2018)
Breiman, L.: On some limit theorems similar to the arc-sin law. Theory of Probability & Its Applications 10(2), 323–331 (1965)
Capéraà, P., Fougères, A. -L., Genest, C.: Bivariate distributions with given extreme value attractor. J. Multivar. Anal. 72(1), 30–49 (2000)
Charpentier, A., Fougères, A. -L., Genest, C., Nešlehová, J.: Multivariate archimax copulas. J. Multivar. Anal. 126, 118–136 (2014)
Cline, D.B.: Convolution tails, product tails and domains of attraction. Probab. Theory Relat. Fields 72(4), 529–557 (1986)
Cline, D.B., Samorodnitsky, G.: Subexponentiality of the product of independent random variables. Stoch. Process. Appl. 49(1), 75–98 (1994)
Coles, S.G., Tawn, J.A.: Modelling extreme multivariate events. Journal of the Royal Statistical Society: Series B (Statistical Methodology), pp. 377–392 (1991)
Davis, R.A., Mikosch, T.: Extreme value theory for space–time processes with heavy-tailed distributions. Stoch. Process. Appl. 118(4), 560–584 (2008)
Dȩbicki, K., Farkas, J., Hashorva, E.: Extremes of randomly scaled Gumbel risks. J. Math. Anal. Appl. 458(1), 30–42 (2018)
de Haan, L., Zhou, C.: Extreme residual dependence for random vectors and processes. Adv. Appl. Probab. 43, 217–242 (2011)
Dombry, C., Ribatet, M.: Functional regular variations, Pareto processes and peaks over threshold. Statistics and its Interface 8, 9–17 (2015)
Embrechts, P., Klüppelberg, C., Mikosch, T.: Modelling extremal events: for insurance and finance, volume 33. Springer Science & Business Media (2013)
Ferreira, A., de Haan, L.: The generalized Pareto process; with a view towards application and simulation. Bernoulli 20(4), 1717–1737 (2014)
Foss, S., Korshunov, D., Zachary, S.: Convolutions of long-tailed and subexponential distributions. J. Appl. Probab. 46(3), 756–767 (2009)
Fougères, A. -L., Soulier, P.: Limit conditional distributions for bivariate vectors with polar representation. Stoch Model 26(1), 54–77 (2010)
Hashorva, E.: On the residual dependence index of elliptical distributions. Statist. Probab. Lett. 80(13), 1070–1078 (2010)
Hashorva, E.: Exact tail asymptotics in bivariate scale mixture models. Extremes 15(1), 109–128 (2012)
Hashorva, E., Hüsler, J.: On multivariate Gaussian tails. Ann. Inst. Stat. Math. 55(3), 507–522 (2003)
Hashorva, E., Pakes, A.G., Tang, Q.: Asymptotics of random contractions. Insurance Math. Econom. 47(3), 405–414 (2010)
Heffernan, J.E., Resnick, S.I.: Limit laws for random vectors with an extreme component. Ann. Appl. Probab. 17(2), 537–571 (2007)
Heffernan, J.E., Tawn, J.A.: A conditional approach for multivariate extreme values (with discussion). J. R. Stat. Soc. Ser. B (Stat Methodol.) 66(3), 497–546 (2004)
Hult, H., Lindskog, F.: Multivariate extremes, aggregation and dependence in elliptical distributions. Adv. Appl. Probab. 34(03), 587–608 (2002)
Huser, R.G., Opitz, T., Thibaud, E.: Bridging asymptotic independence and dependence in spatial extremes using Gaussian scale mixtures. Spatial Statistics 21(A), 166–186 (2017)
Huser, R.G., Wadsworth, J.L: Modeling spatial processes with unknown extremal dependence class. J. Am. Statist. Assoc 114(525), 434–444 (2019). https://doi.org/10.1080/01621459.2017.1411813
Johnson, N.L., Kotz, S., Balakrishnan, N.: Continuous univariate distributions, 2nd edn., vol. 1. Wiley, Chichester (1994)
Johnson, N.L., Kotz, S., Balakrishnan, N.: Continuous univariate distributions, 2nd edn., vol. 2. Wiley, New York (1995)
Krupskii, P., Huser, R., Genton, M.: Factor copula models for replicated spatial data. J. Am. Statist. Assoc. 113(521), 467–479 (2018). https://doi.org/10.1080/01621459.2016.1261712
Ledford, A.W., Tawn, J.A.: Modelling dependence within joint tail regions. J. R. Stat. Soc. Ser. B (Stat Methodol.) 59(2), 475–499 (1997)
McNeil, A.J., Nešlehová, J.: Multivariate Archimedean copulas, d-monotone functions and ℓ 1 −norm symmetric distributions. Ann. Stat. 37(5B), 3059–3097 (2009)
Nolde, N.: Geometric interpretation of the residual dependence coefficient. J. Multivar. Anal. 123, 85–95 (2014)
Opitz, T.: Extremal t processes: Elliptical domain of attraction and a spectral representation. J. Multivar. Anal. 122, 409–413 (2013)
Pakes, A.G.: Convolution equivalence and infinite divisibility. J. Appl. Probab. 41(02), 407–424 (2004)
Resnick, S.I.: Extreme values, regular variation and point processes. Springer, Berlin (1987)
Resnick, S.I.: Heavy-tail phenomena: probabilistic and statistical modeling. Springer Science & Business Media, Berlin (2007)
Rootzén, H., Segers, J., Wadsworth, J.L.: Multivariate peaks over thresholds models. Extremes 21, 115–145 (2018)
Rootzén, H., Tajvidi, N.: Multivariate generalized Pareto distributions. Bernoulli, pp. 917–930 (2006)
Seifert, M.I.: On conditional extreme values of random vectors with polar representation. Extremes 17(2), 193 (2014)
Wadsworth, J.L., Tawn, J.A., Davison, A.C., Elton, D.M.: Modelling across extremal dependence classes. J. R. Stat. Soc. Ser. B (Stat Methodol.) 79(1), 149–175 (2017)
Watanabe, T.: Convolution equivalence and distributions of random sums. Probab. Theory Relat. Fields 142(3), 367–397 (2008)
Acknowledgments
Financial support from the Swiss National Science Foundation grant 200021-166274 (Sebastian Engelke) and UK Engineering and Physical Sciences Research Council grant EP/P002838/1 (Jennifer Wadsworth) is gratefully acknowledged. Thomas Opitz was partially funded by the French national programme LEFE/INSU.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A: Lemmas and proofs
1.1 A.1 Additional lemmas
The following Lemma 8 is widely known as Breiman’s lemma and is useful in several contexts throughout Section 7.
Lemma 8 (Breiman’s lemma, see Breiman (1965), Cline and Samorodnitsky (1994), and Pakes (2004), Lemma 2.1)
SupposeX ∼ F,Y ∼ Gare independentrandom variables. If\(\overline {F}\in \text {RV}_{-\alpha }^{\infty }\)withα ≥ 0 andY ≥ 0 with E(Yα+ε) < ∞forsomeε > 0,then
Equivalently, if F ∈ETα and E(e(α+𝜖)Y) < ∞, then \(\bar {F}_{X+Y}(x)=\overline {F\star G}(x)\sim \mathrm {E}\left (e^{\alpha Y}\right ){\kern 1.7pt} \overline {F}(x)\).
The following Lemma 9 provides some additional detail on the function τ, as defined in Section 2.
Lemma 9
If\(1-\tau _{1}(b_{1}-\cdot ) \in \text {RV}_{1/\gamma _{1}}^{0}\),forγ1 > 0,then\(b_{1} - \tau _{1}^{-1}(1-\cdot ) \in \text {RV}_{1/\gamma _{1}}^{0}\),andγ1 ≤ 1.Similarly, if\(1-\tau _{2}(b_{2}+\cdot ) \in \text {RV}_{1/\gamma _{2}}^{0}\),forγ2 > 0,then\(\tau _{2}^{-1}(1-\cdot )-b_{2} \in \text {RV}_{\gamma _{2}}^{0}\)andγ2 ≤ 1.
Proof
The argument is similar for both cases, so we focus on the first one. Let g(t) = 1 − τ1(b1 − t), which is decreasing as t → 0 and invertible, with \(g^{-1}(s)=b_{1} - \tau _{1}^{-1}(1-s)\). Since \(g \in \text {RV}^{0}_{1/\gamma _{1}}\), then \(g^{-1} \in \text {RV}^{0}_{\gamma _{1}}\) (Resnick (2007), Proposition 2.6(v), adapting to regular variation at zero). Now make the left-sided Taylor expansion
and note that τ1(b1) = 1. Consequently, \(1-\tau _{1}(b_{1}-t) = t \tau _{1}^{\prime }(b_{1-}) + O(t^{2})\), and so finiteness of \(\tau _{1}^{\prime }(b_{1-})\) will imply that the index of regular variation of g is at least 1. Define the convex function μ : (0,∞) → (0,∞) by μ(x) = ν(1,x), so that τ(x) = 1/μ(1/x − 1). Since τ1 is increasing on (0,b1), μ is increasing on ((1 − b1)/b1,∞). We have
and so \(\tau _{1}^{\prime }(b_{1-}) = b_{1}^{-2} \mu ^{\prime }([(1-b_{1})/b_{1}]_{+}) / \mu ((1-b_{1})/b_{1})^{2}\). For h ∈ (0, 1), convexity entails
and so
Hence μ′([(1 − b1)/b1]+) < ∞, giving \(\tau _{1}^{\prime }(b_{1-})<\infty \). Thus the index of regular variation of g is at least 1. □
The following Lemma 10 clarifies the influence of negative values in convolutions of exponential-tailed distributions. It allows us to extend certain results from the literature formulated for nonnegative random variables to the real line.
Lemma 3 (Convolutions of exponential-tailed distributions with negative values)
Fori = 1, 2 and probabilitydistributionsFi ∈ETαdefined over\(\mathbb {R}\)withα > 0,denote\(p_{i}=\bar {F}_{i}(0)\in [0,1]\)theprobability of nonnegative values. Using the convention0/0 = 0,let\(F_{i}^{+}\)with\(1-F_{i}^{+}(x)=\bar {F}_{i}(x)/p_{i}\),x ≥ 0, denote the conditionaldistribution ofFiovernonnegative values, and let\(F_{i}^{-}(x)=F_{i}(x)/(1-p_{i})\),x < 0, denote the conditionaldistribution ofFiover negativevalues. We use the notation\(M_{H}(\alpha )={\int }_{-\infty }^{\infty }\exp (\alpha y)H(\mathrm {d}y)\)fora given distribution H. Then:
- 1.
If\(\bar {F_{i}^{+}}(x)/\bar {F_{1}^{+}\star F_{2}^{+}}(x)\rightarrow 0\)whenx →∞fori = 1, 2,then
$$ \bar{F_{1}\star F_{2}} \sim p_{1}p_{2}\bar{F_{1}^{+}\star F_{2}^{+}}. $$(1) - 2.
If \(\bar {F_{1}^{+}\star F_{2}^{+}} \sim c_{1}\bar {F_{1}^{+}}+c_{2}\bar {F_{2}^{+}}\) with constants 0 ≤ c1,c2 < ∞, then
$$ \bar{F_{1}\star F_{2}}(x) \sim \bar{F}_{1}(x) \left( p_{2}c_{1}+(1-p_{2})M_{F_{2}^{-}}(\alpha)\right)+\bar{F}_{2}(x) \left( p_{1}c_{2}+(1-p_{1})M_{F_{1}^{-}}(\alpha)\right). $$(2)Specifically, if \(c_{1}=M_{F_{2}^{+}}(\alpha )\) and \(c_{2}=M_{F_{1}^{+}}(\alpha )\), then
$$ \bar{F_{1}\star F_{2}}(x)\sim M_{F_{2}}(\alpha) \bar{F}_{1}(x)+M_{F_{1}}(\alpha) \bar{F}_{2}(x), $$and if \(c_{1}=M_{F_{2}^{+}}(\alpha )\) and c2 = 0, then
$$ \bar{F_{1}\star F_{2}}(x)\sim M_{F_{2}}(\alpha) \bar{F}_{1}(x)+\bar{F}_{2}(x)(1-p_{1})m_{F_{1}^{-}}(\alpha). $$
Proof
We start with the mixture representation
We can then use the equation \(p_{i}\bar {F_{i}^{+}}(x)=\bar {F_{i}}(x)\), x ≥ 0, and the following inequalities for {i1,i2} = {1, 2},
Moreover, Lemma 8 can be applied for mixed terms, yielding \(\overline {F_{i_{1}}^{+}\star F_{i_{2}}^{-}}\sim M_{F_{i_{2}}^{-}}(\alpha )\overline {F_{i_{1}}^{+}}\), and then Eqs. 38 and 39 follow from straightforward calculations. To determine the behavior for special cases of c1 and c2, observe that \((1-p_{i})M_{F_{i}^{-}}(\alpha )+p_{i}M_{F_{i}^{+}}(\alpha )=M_{F_{i}}(\alpha )\). □
1.2 A.2 Proof of lemmas in Section 7
Proof of Lemma 2
- 1.
If b1 = b2 = 1 then P(W = 1) = 0 and \(\bar {F}_{W}(1-s)=\bar {F}_{Z}(\tau ^{-1}(1-s))\). By assumption, \(\bar {F}_{Z}(\tau ^{-1}(1-s)) = \ell _{Z}(1-\tau ^{-1}(1-s))(1-\tau ^{-1}(1-s))^{\alpha _{Z}}\), where \(\ell _{Z} \in {\text {RV}_{0}^{0}}\). Since \(1-\tau ^{-1}(1-s) \in \text {RV}_{\gamma }^{0}\) (Lemma 9) with limit zero, results on composition of regularly varying functions (Resnick (2007), Proposition 2.6 (iv)) implies the result, with αW = αZγ.
- 2.a)
If b1 < 1 then
$$ \bar{F}_{W}(1-s) - \P(W=1) = \P(Z\in[\tau_{1}^{-1}(1-s),b_{1}))+\P(Z\in (b_{2},\tau_{2}^{-1}(1-s)]), $$(3)and since Z has a Lebesgue density,
$$ \begin{array}{@{}rcl@{}} &&\P(Z\in[\tau_{1}^{-1}(1-s),b_{1}))+\P(Z\in (b_{2},\tau_{2}^{-1}(1-s)])\\ &\sim& f_{Z}(b_{1})(b_{1} - \tau_{1}^{-1}(1-s)) + f_{Z}(b_{2})(\tau_{2}^{-1}(1-s) - b_{2}) \in \text{RV}_{\gamma}^{0}, \end{array} $$(4)again using Lemma 9.
- 2.b)
A left-sided Taylor expansion of \(\tau _{1}^{-1}\) about 1 gives
$$ \tau_{1}^{-1}(1-s) = b_{1} - (\tau_{1}^{-1})'(1_{-}) s + O(s^{2}), $$where O(s2)/s uniformly tends to 0 as s → 0, and similarly we can make a right-sided expansion for \(\tau _{2}^{-1}(1-s)\). Hence, using Eqs. 40 and 41, \( \bar {F}_{W}(1-s) - \P (W=1) = s \ell (s)\) with \(\lim _{s\to 0}\ell (s) = f_{Z}(b_{1}) (\tau _{1}^{-1})'(1_{-}) - f_{Z}(b_{2})(\tau _{2}^{-1})'(1_{-})\). Noting the link \(1/\tau _{1}^{\prime }(b_{1-}) = (\tau _{1}^{-1})'(1_{-})\), similarly for \(1/\tau _{2}^{\prime }(b_{2+})\), gives Eq. 24.
- 2.
For the final part, we have
$$ \begin{array}{@{}rcl@{}}\bar{F}_{W_{\wedge}}(\zeta(1-s)) &=& \bar{F}_{Z}(\tau_{1}^{-1}(\zeta(1-s))) - \bar{F}_{Z}(1-\tau_{1}^{-1}(\zeta(1-s)))\\ &\sim& f_{Z}(1/2)\left\{ 1 - 2\tau_{1}^{-1}(\zeta(1-s))\right\}. \end{array} $$Again by left-sided Taylor expansion of \(\tau _{1}^{-1}\) about ζ = τ(1/2), we have
$$\tau_{1}^{-1}(\zeta(1-s)) = 1/2 - (\tau_{1}^{-1})'(\zeta_{-}) \zeta s + O(s^{2}),$$and so we obtain \(\bar {F}_{Z}(\tau _{1}^{-1}(\zeta (1-s))) - \bar {F}_{Z}(1-\tau _{1}^{-1}(\zeta (1-s))) = s \ell _{\wedge }(s)\) with \(\lim _{s\to 0}\ell _{\wedge }(s) = 2 f_{Z}(1/2)\zeta (\tau _{1}^{-1})'(\zeta _{-})\). Noting again that \(1/\tau _{1}^{\prime }(1/2_{-}) = (\tau _{1}^{-1})'(\zeta _{-})\), we arrive at Eq. 25.
□
Proof of Lemma 5
The result for nonnegative S and V is found in Theorem 4(v) of Cline (1986). The extension to negative values then follows from Lemma 10(1). □
Proof
of Lemma 6 The result is given in Theorem 6(ii,iii) of Cline (1986) for \(\bar {F}(0)=1\) and \(\bar {G}_{1}(0)=1\). For point 1, the extension to negative values in F and G1 follows from observing that Theorem 6(iii) of Cline (1986) implies
where F+ is obtained from F by setting F+(0) = F(0) and \(\bar {F^{+}}(0)=1\), and the same construction is taken for \(G_{1}^{+}\); i.e., F+ and \(G_{1}^{+}\) arise from projecting negative values to 0. For point 2, the extension to negative values can be shown using Lemma 10(2). Indeed, the same limit \(M_{G_{1}}(\alpha )\) arises for \(\bar {F\star G_{1}}(x)/\bar {F}(x)\) if we project negative values in F and G1 to 0 or not. □
Appendix B: Tail classes and examples
Definitions of tail classes are given in Section 1.1. The following lemma summarizes important relationships between such tail classes. In this section, we refer to the class of heavy-tailed distributions by HT, and to superheavy-tailed distributions by SHT.
Lemma 4 (Relationships between tail classes)
The following relationships between distribution classes hold:
-
1.
\(\text {RV}_{\alpha }^{\infty }\subset \text {CE}_{0}\)forα > 0,
-
2.
\(\text {ET}_{0}\subsetneq \text {HT}\) .
-
3.
For ETαwithα > 0,we have:
\(F(\exp (\cdot ))\in \text {ET}_{\alpha } \Leftrightarrow F \in \text {RV}_{\alpha }^{\infty }\),
CEα ⊂ETα,
ETα,β>− 1 ∩CE = ∅.
-
4.
For WTβ,we have:
\(\text {WT}_{1}\subset \bigcup _{\alpha >0}\text {ET}_{\alpha }\),
WTβ ⊂CE0forβ < 1,
LWTβ ⊂SHT forβ < 1.
-
5.
By denotingF1 ≺ F2if\(\overline {F}_{1}(x)/\overline {F}_{2}(x)\rightarrow 0\)forx →∞,we have:
If\(\tilde {\alpha }<\alpha \),then\(\text {WT}_{\beta >1}\prec \text {ET}_{\alpha } \prec \text {ET}_{\tilde {\alpha }} \prec \text {WT}_{\beta <1}\prec \text {LWT}_{\beta >1}\prec \text {RV}_{\alpha >0}^{\infty }\prec \text {RV}_{\tilde {\alpha }}^{\infty }\prec \text {SHT}\).
\(\text {CE}_{\alpha >0}\prec \text {ET}_{\tilde {\alpha },\beta }\)for\(\tilde {\alpha }\leq \alpha \)andanyβ > 0.
We recall the membership in tail classes for well-known parametric distribution families in Table 1, see Johnson et al. (1994, 1995) for reference about parameters. Here we abstract away from the usual parameter symbols of these distributions to avoid conflicting notations with general tail parameters. We refer parameters as scl and loc if scl × X + loc has scale scl and location loc, where X has scale 1 and location 0. Another parameter shp may be related to shape for some distributions.
Appendix C: Additional illustrations
Figure 4 illustrates further examples of norms ν and related functions τ(z) and τ(1 − z), as defined in Section 2.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Engelke, S., Opitz, T. & Wadsworth, J. Extremal dependence of random scale constructions. Extremes 22, 623–666 (2019). https://doi.org/10.1007/s10687-019-00353-3
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10687-019-00353-3