When: 5 April 2017

Where: Room 402, USI Main Building

12:30 - 13:30 ** Joris Bierkens, Delft University of Technology.**

**Zig-Zag Sampling for Doubly Intractable Distributions **

In important models in Bayesian statistics the computation of the likelihood function is intractable. The corresponding posterior distributions are referred to as doubly intractable distributions. For example this situation occurs when inferring the temperature in an Ising model, or the parameters in an exponential random graph model (ERGM).

Existing methodology to deal with such models rely on the exchange algorithm of Møller (2006) and variations thereof. In order for this class of algorithms to be asymptotically exact it is necessary to draw perfect samples from the forward model using e.g. the Propp & Wilson (1996) methodology.

It turns out that, when applying the Zig-Zag Sampler (Bierkens et al., 2016) to problems with intractable likelihood, it suffices to obtain unbiased estimates from the forward model, which greatly enlarges the scope of possible models which can be dealt with, as well provides significant gains in computational efficiency.

This is currently work in progress in collaboration with Antonietta Mira (Lugano) and Gareth Roberts (Warwick).