Bounds for randomly shared risk of heavy-tailed loss factors (Q347153)

Let \(\{V_j: 1 \leq j \leq d\}\) be \(d\) possibly dependent risks. It is assumed that \(V\) is multivariate regularly varying with index \(\alpha\), that is \[ n P[n^{-1/\alpha} V \in \cdot] \to \nu(\cdot) \] as \(n \to \infty\), where \(\to\) stands for vague convergence. These risks are shared by \(q\) agents, such that \(F = A V\) denotes the shares. Here, \(A\) is a (stochastic) \(q \times d\) matrix. That means, that the sharing is possibly done in a stochastic way. The matrix \(A\) is assumed to be independent of \(V\). The (quasi-) norm of \(F\) is defined as \[ \|F\| = \Bigl(\sum_{j=1}^q F_j^r\Bigr)^{1/r} \] for an \(r > 0\). Two risk measures are now considered. The value-at-risk at confidence level \(1-\gamma\) is defined as \[ \text{VaR}_{1-\gamma} (X) = \inf\{t : P[X > t] \leq \gamma\}\;, \] the tail-value-at-risk at confidence level \(1-\gamma\) is \[ \text{TVaR}_{1-\gamma} (X) = E[X \mid X > \text{VaR}_{1-\gamma} (X)]\;. \] It is already known by previous work of the authors that \[ \text{VaR}_{1-\gamma} (F_i) \sim C_i^{1/\alpha} \gamma^{-1/\alpha}\;,\qquad \text{TVaR}_{1-\gamma} (F_i) \sim \frac\alpha{\alpha-1} C_i^{1/\alpha} \gamma^{-1/\alpha}\;, \] and \[ \text{VaR}_{1-\gamma} (\|F\|) \sim C_S^{1/\alpha} \gamma^{-1/\alpha}\;,\qquad \text{TVaR}_{1-\gamma} (\|F\|) \sim \frac\alpha{\alpha-1} C_S^{1/\alpha} \gamma^{-1/\alpha}\;, \] as \(\gamma \to 0\) for a constant \(C\) depending on the distribution of \(V\) and \(A\). Denote for aymptotically independent risks the constant \(C^{\text{ind}}\), and for completely dependent risks \(C^{\text{dep}}\). Then it turns out that \(C_i^{\text{ind}} \leq C_i \leq C_i^{\text{dep}}\) for \(\alpha \geq 1\) and \(C_i^{\text{dep}} \leq C_i \leq C_i^{\text{ind}}\) for \(\alpha < 1\). Then it is shown that the analogous inequalities hold for \(C_S\), \(C_S^{\text{ind}}\) and \(C_S^{\text{dep}}\) in the case of equal marginal distributions of \(V\).