Total: 1
Generalized Bayesian inference uses weights of the form $\exp\{-β_n\ell_{n}(θ)\}$, even when the loss is available only through simulation, numerical integration, or subsampling. Exponentiating an unbiased loss estimate changes the target, and when $β_n\asymp n$ an ordinary Monte Carlo (MC) loss estimate with $M^{-1}$ variance needs a per-proposal budget of order $n^2$ to keep the leading log-weight variance bounded. We introduce the Sign-Corrected Bessel Debiasing (SCBD) algorithm, a signed pseudo-marginal method based on $K$ independent block estimates of the loss, and study its independently randomized Randomised Quasi-Monte-Carlo (RQMC) implementation. Under an iid Gaussian block model, a Bessel factor constructed from the block sample variance exactly removes the Gaussian exponential bias. For general finite MC or RQMC blocks, signed averages instead target a density proportional to $π_n(θ)w_M(θ)$, where $w_M$ is the finite-block target factor remaining after the signed correction; its posterior variation is assessed separately from sign efficiency. If an RQMC block estimator, based on $d$-dimensional randomized inputs, has variance $O\{B^{-α}(\log B)^{d-1}\}$, a sufficient budget for bounded leading log-weight variance has order $n^{2/α}$, up to logarithmic factors. We give a finite-block total-variation bound on the approximation and show the separate improvements from the debiasing correction and RQMC. The numerical examples show that, when the RQMC representation is favorable, the selected budget for stabilizing the estimated log weight can inherit this order, and that the Bessel correction can matter even when the RQMC variance rate is close to that of ordinary Monte Carlo.