On small deviations of stationary Gaussian processes and related analytic inequalities

Abstract: Let $ {X_j, j\in \Z}$ be a Gaussian stationary sequence having a spectral function $F$ of infinite type. Then for all $n$ and $z\ge 0$,$$ \P\Big{\sup_{j=1}^n |X_j|\le z \Big}\le \Big(\int_{-z/\sqrt{G(f)}}^{z/\sqrt{G(f)}} e^{-x^2/2}\frac{\dd x}{\sqrt{2\pi}} \Big)^n,$$ where $ G(f)$ is the geometric mean of the Radon Nycodim derivative of the absolutely continuous part $f$ of $F$. The proof uses properties of finite Toeplitz forms. Let $ {X(t), t\in \R}$ be a sample continuous stationary Gaussian process with covariance function $\g(u) $. We also show that there exists an absolute constant $K$ such that for all $T>0$, $a>0$ with $T\ge \e(a)$, $$\P\Big{\sup_{0\le s,t\le T} |X(s)-X(t)|\le a\Big} \le \exp \Big {-{KT \over \e(a) p(\e(a))}\Big} ,$$ where $\e (a)= \min\big{b>0: \d (b)\ge a\big}$, $\d (b)=\min_{u\ge 1}{\sqrt{2(1-\g((ub))}, u\ge 1}$, and $ p(b) = 1+\sum_{j=2}^\infty {|2\g (jb)-\g ((j-1)b)-\g ((j+1)b)| \over 2(1-\g(b))}$. The proof is based on some decoupling inequalities arising from Brascamp-Lieb inequality. Both approaches are developed and compared on examples. Several other related results are established.