Abstract
We present a method for conditional sampling for pre-trained normalizingflows when only part of an observation is available. We derive a lower bound tothe conditioning variable log-probability using Schur complement properties inthe spirit of Gaussian conditional sampling. Our derivation relies onpartitioning flow's domain in such a way that the flow restrictions tosubdomains remain bijective, which is crucial for the Schur complementapplication. Simulation from the variational conditional flow then amends tosolving an equality constraint. Our contribution is three-fold: a) we providedetailed insights on the choice of variational distributions; b) we discuss howto partition the input space of the flow to preserve bijectivity property; c)we propose a set of methods to optimise the variational distribution. Ournumerical results indicate that our sampling method can be successfully appliedto invertible residual networks for inference and classification.