Séminaire du LPTMS: Alexei Chepelianskii (LPS)

Quand

11/01/2022    
11:00 - 12:30

Petit amphi, bâtiment Pascal n° 530
rue André Rivière, Orsay, 91405

Type d’évènement

Carte non disponible

Monte Carlo sampling: convergence, localization transition and optimality

Alexei Chepelianskii  (Laboratoire de Physique des Solides)

Onsite + zoom.

Zoom link: https://cnrs.zoom.us/j/96861999262?pwd=bzlkQThNenNyYVdoMXpKUlF0VXloQT09

Meeting ID: 968 6199 9262
Passcode: 5C7NBW

Among random sampling methods, Markov Chain Monte Carlo algorithms are foremost. From a combination of analytical and numerical approaches, we study their convergence properties towards the steady state. We show that the deviations from the target steady-state distribution features a localization transition as a function of the characteristic length of the attempted jumps defining the random walk in a Metropolis scheme. This transition changes drastically the error which is introduced by incomplete convergence. Remarkably, the localization transition occurs for parameters that also provide the optimal Monte Carlo convergence speed. We show that the relaxation of the error in the localized regime has some similarities with relaxation in a greedy Monte Carlo algorithm where jumps always  occur to lower energy sites (zero temperature limit). In both cases relaxation is described by a self-similar ansatz instead of being dominated by the  eigenmode with the slowest relaxation rate as in diffusion problems.

 

Retour en haut