On the approximation of quantile processes by Kiefer processes (Q1272499)

Let \(\{\alpha_n:n\geq 1\}\) be a sequence of uniform empirical processes. According to a deep result of \textit{J. Komlós}, \textit{P. Major} and \textit{G. Tusná}dy [Z. Wahrscheinlichkeitstheor. verw. Geb. 32, 111-131 (1975; Zbl 0308.60029)] a version of the sequence \(\{\alpha_n:n\geq 1\}\) and a sequence \(\{B_{n,1}:n\geq 1\}\) of Brownian bridges can be constructed on a single (appropriate) probability space such that \[ (1)\qquad \|\alpha_n-B_{n,1}\|=O(n^{-1/2}\log n)\;a.s.\quad \text{as} n\to\infty. \] It is not possible to improve the rate of approximation in (1) to \(o(n^{-1/2}\log n)\). Moreover, there exists a probability space carrying a version of \(\{\alpha_n:n\geq 1\}\) and a Kiefer process \(K(n,\cdot)\) such that \[ (2)\qquad \|\alpha_n-n^{-1/2}K(n,\cdot)\|=O(n^{-1/2}\log^2n)\;a.s.\quad \text{as} n\to\infty. \] \textit{M. Csörgő} and \textit{P. Révész} [Colloq. Math. Soc. Janos Bolyai 11, 59-71 (1975; Zbl 0315.62013)] obtained an approximation result corresponding to (1) for a sequence \(\{\beta_n:n\geq 1\}\) of uniform quantile processes: There exist a version of \(\{\beta_n:n\geq 1\}\) and a sequence \(\{B_{n,2}:n\geq 1\}\) of Brownian bridges on a single probability space such that \[ (3)\qquad \|\beta_n-B_{n,2}\|=O(n^{-1/2}\log n)\;a.s.\quad \text{as} n\to\infty. \] Again, the rate of approximation cannot be improved to \(o(n^{-1/2}\log n)\). By combining (2) with the Bahadur-Kiefer bound for the uniform empirical difference process \(\alpha_n+\beta_n\) a version of (2) can be established: For the sequence \(\{\beta_n:n\geq 1\}\) pertaining to \(\{\alpha_n:n\geq 1\}\) from (2) and for \(K'(n,t)=-K(n,t)\) one has: \[ (4)\qquad \limsup_{n\to\infty}n^{1/4}(\log n)^{-1/2}(\log\log n)^{-1/4} \|\beta_n-n^{-1/2}K'(n,\cdot)\|=2^{-1/4} a.s. \] A comparison of (1) - (4) has led to the conjecture that the \(n^{-1/4}(\log n)^{1/2}(\log\log n)^{1/4}\)-rate of approximation in (4) is far from being optimal. In Ann. Probab. 25, No. 4, 2007-2054 (1997; Zbl 0902.60027), the author of the paper under review disproved this conjecture by showing that the approximation rate in (4) cannot be better than \(O(n^{-1/4}(\log n)^{-\epsilon})\) for any \(\epsilon>0\). In this paper he solves the problem completely by proving the following deep result: There exists a universal constant \(D\) with \(1/11<D\leq 2^{-1/4}\) such that for any probability space which carries a version of \(\{\beta_n:n\geq 1\}\) and a Kiefer process \(K(n,\cdot)\), one has \[ \limsup_{n\to\infty}n^{1/4}(\log n)^{-1/2}(\log\log n)^{-1/4} \|\beta_n-n^{-1/2}K(n,\cdot)\|\geq D\quad a.s. \]