Eric Moulines
Wednesday 24th February 2016
Time: 4.00pm
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Sampling from log-concave non-smooth densities, when Moreau meets Langevin
In this talk, a new algorithm to sample from possibly non-smooth
log-concave probability measures is introduced. This algorithm uses
Moreau-Yosida envelope combined with the Euler-Maruyama discretization
of Langevin diffusions. They are applied to a deconvolution problem in
image processing, which shows that they can be practically used in a
high dimensional setting. Finally, non-asymptotic convergence bounds (in
total variation and wasserstein distances) are derived. These bounds
follow from non-asymptotic results for ULA applied to probability
measures with a continuously differentiable log-concave density. [A
paper will be arxived soon (updating http://arxiv.org/abs/1507.05021).]