Sequential Bayesian Experimental Design with Variable Cost Structure

The main paper provides a brief overview of nested Monte Carlo bounds. We give a more detailed derivation of those specific bounds here and discuss alternative bounds in Sec. 3. We begin with notation elements that are used throughout. Let Dr be the outcomes of past experiments at stage r. We further assume samples of the latent variables X = {xn}n=1:N , drawn from updated beliefs, xn ∼ p(x | Dr), and data samples Y = {yn}n=1:N , drawn from the forward model of the candidate experiment, yn ∼ pa(y | xn). We can obtain upper and lower bounds at subsequent stage r as follows.