Likelihood-Coin Poisson Sampling for Bayesian Inverse Problems: Exact Sampling and Sharp Complexity

RESUMO

We develop an exact direct-sampling framework for Bayesian inverse problems whose bounded forward observables can be queried through Bernoulli events rather than numerically evaluated. A Bernstein--Poisson construction converts these events into scaled Gaussian likelihood coins. Thinning a prior-based Poisson point process with these coins yields a posterior point process whose locations, conditional on its cardinality, are independent posterior draws; the cardinality also provides an unbiased estimator of the reduced evidence. The main theoretical contribution is a sharp analysis of the actual early-stopped computational work. In the small-noise regime, the work is governed by the local prior-predictive mass near the exact-fit set, revealing a predictive-dimension effect distinct from the nominal dimension of the unknown. A multivariate extension accommodates correlated Gaussian observation errors and gives a precision-weighted allocation of factory effort. For a bounded class of elliptic resolvent problems, Feynman--Kac representations, Poisson killing, and lazy random-series evaluation provide exact continuum Bernoulli oracles without introducing a spatial discretization or a fixed parameter truncation in the posterior target. Numerical experiments illustrate posterior thinning in parameter space, confirm the sharp work regimes, and validate the function-valued PDE construction.