The local quantization behavior of absolutely continuous probabilities (Q439886)

From MaRDI portal
scientific article
Language Label Description Also known as
English
The local quantization behavior of absolutely continuous probabilities
scientific article

    Statements

    The local quantization behavior of absolutely continuous probabilities (English)
    0 references
    0 references
    0 references
    0 references
    0 references
    17 August 2012
    0 references
    Let \(\mathbb R^d\) be endowed with a norm \(\|\cdot\|\) and let \(\operatorname{P}\) be a Borel probability measure on \(\mathbb R^d\) with finite \(r\)-th moment for some \(r>0\). Then its \(n\)-th quantization error w.r.t.~the \(L_r\)-norm is defined by \[ e_{n,r}= e_{n,r}(\operatorname{P}):= \inf\left\{\left(\int_{\mathbb R^d} d(x,\alpha)^r \mathrm{d} \operatorname{P}(x)\right)^{1/r} : \alpha\subset\mathbb R^d,\;\mathrm{card}(\alpha)\leq n\right\}. \tag{1} \] Here, as usual, \[ d(x,\alpha)=\min_{a\in\alpha}\|x-a\| \] is the distance of \(x\in\mathbb R^d\) to the set \(\alpha\subset\mathbb R^d\). The finite sets \(\alpha\) appearing in (1) are called code books of order \(n\). Code books generate in a natural way a partition \((V_{n,a})_{a\in\alpha}\) of \(\mathbb R^d\) (called Voronoi partition) where the Voronoi cells \(V_{n,a}\) have to obey the nearest neighbor rule. As it is known, for each \(\operatorname{P}\) and each \(n\geq 1\), there are optimal code books \(\alpha_n\) for which the infimum in (1) is attained. The aim of the present paper is to investigate the behavior of the Voronoi cells \(V_{n,a}\) for \(a\in\alpha_n\). Of particular interest are questions about the behavior of \[ \operatorname{P}(V_{n,a})\quad\text{and}\quad \int_{V_{n,a}}\|x-a\|^r\mathrm{d} \operatorname{P}(x),\;a\in\alpha_n \] as \(n\to \infty\). For example, an open question for several years is whether or not \[ \int_{V_{n,a}}\|x-a\|^r\mathrm{d} \operatorname{P}(x)\sim \frac 1 n e_{n,r}^r,\;a\in\alpha_n, \] as \(n\to\infty\). The aim of the present paper is to prove ``weak'' asymptotics of \(\operatorname{P}(V_{n,a})\) as well as of the above integral whenever \(a\) is chosen in the optimal code book \(\alpha_n\). The probability measure \(\operatorname{P}\) is assumed to be absolutely continuous with respect to the Lebesque measure on \(\mathbb R^d\) and its density and/or its support have to satisfy some additional regularity assumptions. Thereby, the cases in which \(\operatorname{P}\) has compact or non-compact support are treated differently. Then the authors prove that \[ \mathbf{P}(V_{n,a})\approx n^{-1}\quad\text{and}\quad \int_{V_{n,a}}\|x-a\|^r\mathrm{d} \operatorname{P}(x)\approx \frac{e_{n,r}^r}{n}\approx n^{-(1+r/d)},\;a\in\alpha_n, \] as \(n\to\infty\) as long as the Voronoi cells \(V_{n,a}\) intersect a fixed compact set in the interior of the support of \(\operatorname{P}\). Although this is a big step to the solution of the general conjecture, many questions about the behavior of the cells \(V_{n,a}\) with \(a\in\alpha_n\) remain unanswered.
    0 references
    0 references
    0 references
    vector quantization
    0 references
    probability of Voronoi cells
    0 references
    inertia of Voronoi cells
    0 references
    0 references