Title: Efficient Bernoulli factory MCMC for intractable likelihoods
Authors: Dootika Vats - Indian Institute of Technology, Kanpur (India) [presenting]
Abstract: Accept-reject based Markov chain Monte Carlo (MCMC) algorithms have traditionally been a function of the ratio of the target density at two contested points. We note that this feature is rendered almost useless in Bayesian MCMC problems within tractable likelihoods. We introduce a new acceptance probability that has the distinguishing feature of not being as a function of the ratio of the target density at two points. We show that such a structure allows for the construction of an efficient and stable Bernoulli factory. The resulting ``Portkey Barker's'' algorithm is exact and is computationally more efficient that the current state-of-the-art.