A generalization of Chernoff inequality via stochastic analysis (Q1077065)

Let \(\mu\) be a probability measure on \(R^ d\) with density c(exp(- 2U(x)), where \(U\in C^ 2(R^ d)\), \(| \nabla U(x)|^ 2-\Delta U(x)\to \infty\) and U(x)\(\to \infty\) as \(| x| \to \infty\). By using stochastic analysis and theorems in Schrödinger operators we have the following result: there exists a constant \(c>0\) such that \[ (1)\quad Var_{\mu}f\leq cE_{\mu}| \nabla f|^ 2 \] for any \(f\in L^ 1(\mu)\) with a well-defined distributional gradient \(\nabla f\). Under our condition, the operator \(-\Delta +\nabla U\cdot \nabla\) in \(L^ 2(\mu)\) has discrete spectrum \(0=\lambda_ 1<\lambda_ 2=...=\lambda_ m<\lambda_{m+1}\leq...\) with corresponding eigenfunctions \(\{\phi_ n\}\) which form a C.O.N.S (complete orthonormal system). If the R.H.S. of (1) is finite then equality holds iff \(f=\sum^{m}_{i=1}b_ i\phi_ i\) for some \(b_ 1,...,b_ m\in R.\) Moreover, the constant c can be taken as \((2\lambda_ 2)^{-1}.\) When U is a quadratic form, (1) is the Chernoff inequality [\textit{H. Chernoff}, Ann. Probab. 9, 533-535 (1981; Zbl 0457.60014), \textit{L. H. Y. Chen}, J. Multivariate Anal. 12, 306-315 (1982; Zbl 0483.60011)]. The approach here can be generalized to infinite dimensional Gaussian measures, or the case with \(\mu\) being a probability measure in a bounded domain of \(R^ d\) or some discrete cases.