Annealing diffusions in a potential function with a slow growth (Q2469491): Difference between revisions

The author is concerned with the solution of the stochastic differential equation \[ \mathrm{d}X_{t}=\sqrt{\sigma \left(t\right) }\mathrm{d}B_{t}-\nabla V\left( X_{t}\right) \mathrm{d}t, \] a continuous analogue of the simulated annealing algorithm in \(^{d}\). Here, \(V\) is a function called potential. He proves a convergence result, similar to one in \textit{L. Miclo} [Thèse de doctorat, Université Paris VI, 1991], under weaker hypotheses on the potential. In particular, cases where the gradient of the potential converges to zero at infinity are also covered. The main idea is to replace the Poincaré and log-Sobolev inequalities used by Miclo by the weak Poincaré inequalities introduced by \textit{M. Röckner} and \textit{F. Y. Wang} [J. Funct. Anal. 185, No. 2, 564--603 (2001; Zbl 1009.47028)], and to estimate constants with measure-capacity criteria. It is shown that the convergence still holds for the `classical' schedule \(\sigma \left( t\right) =c/\ln \left( t\right) \), where \(c\) is greater than a constant related to \(V\), namely, the height of the largest potential barrier.