Linear \(n\)-widths of diagonal matrices in the average and probabilistic settings (Q418714): Difference between revisions

Let \(D\) be a diagonal matrix mapping \(\mathbb R^m\) into \(l_q^m\). The aim of the present paper is to investigate the average linear \(n\)-widths of \(D\) defined by \[ \lambda_n^{(a)}(D : \mathbb R^m\to l_q^n ;\gamma_m)_p:= \inf_{T_n}\left(\int_{\mathbb R^m} \|D x - T_n x\|_q^p \,d \gamma_m(x)\right)^{1/p}, \] where \(\gamma_m\) denotes the standard \(m\)-dimensional Gaussian measure and the infimum is taken over all operators \(T_n\) of rank \(\leq n\). The number \(p\) may be chosen arbitrarily in \((0,\infty)\) since a result of \textit{X. Fernique} [``Intégrabilité des vecteurs gaussiens'', C. R. Acad. Sci., Paris, Sér. A 270, 1698--1699 (1970; Zbl 0206.19002)] implies that there are universal constants such that the quantities for different \(p\) may be estimated mutually. The main result of the present paper gives exact formulae for the average width \(\lambda_n^{(a)}(D : \mathbb R^m\to l_q^n ;\gamma_m)_q\) provided that \(1\leq q<\infty\). In contrast to some earlier results, for example by \textit{P. Mathé} [``Random approximation of Sobolev embeddings'', J. Complexity 7, No. 3, 261--281 (1991; Zbl 0735.65004)], these formulae are valid for all possible pairs \((n,m)\) with \(n < m\). Perhaps unsurprisingly, the exact expressions differ significantly for the cases \(1\leq q\leq 2\) and \(2<q<\infty\), respectively. While in the first case, the best approximation of \(D\) is given by cutting off the smallest \(m-n\) terms of \(D\), the case \(2<q<\infty\) is more involved and leads to a very surprising result. The case \(q=\infty\) is more complicated and only an upper estimate for \(\lambda_n^{(a)}(D : \mathbb R^m\to l_\infty^n ;\gamma_m)_p\) is proved, still better than the formerly known estimates and valid for all \(n< m\). Finally, the authors calculate the probabilistic linear \((n,\delta)\)-widths for \(D:\mathbb R^m\to l_q^n\) with \(1\leq q\leq 2\). Reviewer's remark. For \(p=2\), the average linear \(n\)-widths coincide with the \(l_n\)-numbers investigated in the book of \textit{G. Pisier} [The volume of convex bodies and Banach space geometry. Cambridge Tracts in Mathematics 94 (1989; Zbl 0698.46008)]. For example, there one also finds a proof of Theorem 3.2 in the case \(p=2\), but Pisier's proof easily extends to all \(p>0\). We also refer to a paper by \textit{T. Kühn} and the reviewer [``Optimal series representation of fractional Brownian sheets'', Bernoulli 8, No. 5, 669--696 (2002; Zbl 1012.60074)] for further properties of the \(l_n\)-numbers and their probabilistic interpretation.