no. 16 · Tuesday 17 November 2026
Dependence measures as spectral risk measures
Abstract
For a bivariate copula $C$ with $(U,V)\sim C$, the conditional distribution functions $L_t=\partial_1C(U,t)$ are random variables with means $t$, so that measures of directed dependence become risk functionals. The absolute Lorenz curves of the variables $L_t$ form the rearranged copula $C^\uparrow$. In this picture Chatterjee's rank correlation $ξ$ is a squared Sobolev seminorm and the rearranged Spearman correlation an $L^1$ norm of $C^\uparrow-Π$, and both arise from one spectral risk measure, applied either to the conditional laws of $V$ or to the laws of $L_t$. The resulting classes $ω_φ$ and $κ_φ$ of spectral dependence measures have mirror-image axiomatic properties. For every spectrum we determine the exact upper bound of $κ_φ$ in terms of $ξ$, and we prove sharp inequalities between $ξ$, $κ_φ$ and the values of $C^\uparrow$, which show how little global measures say about the tails. For plug-in estimators based on the empirical checkerboard copula we obtain consistency, rates and a resolution limit, and a simulation study shows that tail-localised measures separate copulas with identical $ξ$. An application to large credit portfolios illustrates the results.
- fields
- Statistics Theory