On regular variation of probability densities (Q579734): Difference between revisions

Let F be the d.f. of a probability measure on \({\mathbb{R}}^ d_+\). The tail of F varies regularly of index \(\alpha\) if there is g: (0,\(\infty)\to (0,\infty)\), varying regularly at \(\infty\) of index \(\alpha\), such that \(\{1-F(tx)\}/g(t)\to \gamma (x)>0\), \(t\to \infty\), \(x\in {\mathbb{R}}^ d\), \(x>0\), i.e. \(x_ i>0\), \(i=1,...,d\). Then \(\gamma (tx)=t^{\alpha}\gamma (x)\). This is an important property for establishing stable attraction and limiting behaviour of extremes. Let F have a density f such that \[ (*)\quad t^ df(tx)/V(t)\quad \to \quad \lambda (x)\quad >\quad 0,\quad t\to \infty,\quad x\in {\mathbb{R}}^ d_+,\quad x\neq 0, \] where V varies regularly at \(\infty\) of index \(\rho <0.\) The paper derives conditions under which \[ (**)\quad \{1- F(tx)\}/V(t)\quad \to \quad \nu ([0,x]^ c)\quad <\quad \infty,\quad t\to \infty,\quad x>0, \] with \(\nu\) a measure on \({\mathbb{R}}^ d_+\) having density \(\lambda\). Here \([a,b]=\{x:\) \(a_ i\leq x_ i\leq b_ i\), \(i=1,...,d\}\). A sufficient condition is componentwise monotonicity of f, see the first author and \textit{E. Omey}, Stochastic Processes Appl. 16, 157-170 (1984; Zbl 0528.60021). Let \(\| \|\) be any norm on \({\mathbb{R}}^ d\) and \(Q=\{x\in {\mathbb{R}}^ d_+:\| x\| =1\}\). It is proved that if \(\lambda\) is bounded on Q and (*) holds uniformly with respect to x on Q, then (**) holds. Then also (*) holds uniformly on \(\{x\in {\mathbb{R}}^ d_+:\| x\| \geq \epsilon \}\). Application of the condition is facilitated by the free choice of the norm. It also leads to a Karamata-type representation of f. Monotonicity of f and continuity of \(\lambda\) on Q are shown to imply the uniformity condition.