Generalized posteriors update a prior through empirical loss, with a learning rate controlling posterior concentration. Their credible intervals can misrepresent sampling uncertainty. We propose infinitesimal-jackknife variance matching (IJVM), which selects the learning rate for a prespecified scalar functional by matching its posterior variance to an IJ estimate of the sampling variance of its posterior mean. The IJ variance is computed from posterior covariances with observation-level losses. A one-step update uses a pilot posterior fit and one final fit. IJVM shares the functional-specific first-order target of rank-one weighted score matching but requires neither loss Hessians nor gradient covariance estimates for calibration. This facilitates applications with regular non-smooth convex losses, including check and hinge losses, without constructing problem-specific curvature estimators. Under suitable regularity and Monte Carlo conditions, we establish consistency of the one-step rate and, with additional assumptions, asymptotically nominal credible-interval coverage. In simulations of quadratic-loss regression, quantile regression, and hinge-loss classification, IJVM generally achieves coverage near 95%, although finite-sample undercoverage remains in quantile regression. A second update can improve finite-sample calibration in some settings, particularly at smaller sample sizes, but the benefits of further iterations are task-dependent and not necessarily monotone. Compared with bootstrap-based calibration, IJVM yields comparable interval lengths with substantially fewer posterior fits.
翻译:暂无翻译