Target-aware Bayesian inference via generalized thermodynamic integration

Abstract In Bayesian inference, we are usually interested in the numerical approximation of integrals that posterior expectations or marginal likelihoods (a.k.a., evidence). this paper, focus on computation expectation a function $$f(\textbf{x})$$ f ( x ) . We consider target-aware scenario where is known advance and can be exploited order to improve estimation expectation. scenario, task reduced perform several independent likelihood tasks. The idea using path tempered distributions has been widely applied literature for likelihoods. Thermodynamic integration, sampling annealing importance well-known examples algorithms belonging family methods. work, introduce generalized thermodynamic integration (GTI) scheme which able i.e., GTI approximate given function. Several scenarios application discussed different simulations provided.