Asymptotics of the maximal radius of an \(L^{r}\)-optimal sequence of quantizers (Q408108)

Let \(X\) be a random \(\mathbb R^d\)-valued vector with \(\operatorname{E}|X|^r<\infty\) for some \(r>0\). The quantization problem at level \(n\) for \(X\) consists in finding a subset \(\alpha\subset \mathbb R^d\) with \(\mathrm{card}(\alpha)\leq n\) such that \[ e_{r}(X;\alpha):=\left[\operatorname{E}\min_{a\in\alpha}|X-a|^r\right]^{1/r} \] is small. A set \(\alpha\) of cardinality less than \(n\) is called optimal \(n\) quantizer provided that the above approximation error \(e_{r}(X;\alpha)\) becomes minimal for \(\alpha\). Suppose now that for each \(n\in\mathbb N\), we have an optimal \(n\)-quantizer \(\alpha_n\). The aim of the presented paper is to investigate the radii of the \(\alpha_n\), i.e., the behavior of the sequence \((\rho_n)_{n\geq 1}\) with \[ \rho_n:= \max\{ |a| : a\in\alpha_n\}. \] It is not surprising that \(\rho_n\to\infty\) provided that \(X\) is not mapping into a compact subset of \(\mathbb R^d\). Thus it is natural to ask how tail properties of the law of \(X\) influence the speed of convergence of the \(\rho_n\). In the present paper, two typical examples are treated: Firstly, the law of \(X\) possesses a polynomial tail, and, secondly, this tail behaves hyper-exponentially. The authors prove very precise results about the asymptotic of the \(\rho_n\) in these cases. For example, in the polynomial case it follows that \[ \lim_{n\to\infty}\frac{\log \rho_n}{\log n}= c \] where \(c>0\) is precisely calculated and depends on the order of the tail, on \(r>0\) and on the dimension \(d\) of the underlying space. Similarly, in the hyper-exponential case, the \(\rho_n\) behave as some power of \(\log n\).