Climate-field generation stops treating the transport map as known: a Laplace approximation adds hyperparameter uncertainty at near empirical-Bayes cost
Synopsis
Generative modeling of non-Gaussian climate fields has mostly fixed the hyperparameters of a triangular transport map at an empirical Bayes point estimate, so later inference treats the map as known and can produce overconfident prediction intervals. A fully Bayesian view is adopted instead, with a Laplace approximation giving a Gaussian approximation to the hyperparameter posterior at its mode, at negligible extra cost relative to empirical Bayes. In simulations and on CESM global log-precipitation fields, the approximation reproduced most posterior summaries of MCMC for moderate-to-large training sets, with predictive log scores comparable to or higher than MCMC.
Interpretation
The hyperparameters of a triangular transport map are no longer fixed at a point estimate but carry a samplable approximate posterior, so parameter uncertainty propagates into generated fields. Earlier Bayesian transport maps optimized a small set of Gaussian-process hyperparameters by empirical Bayes and then treated them as known, ignoring hyperparameter uncertainty, while the fully Bayesian route relies on MCMC, which is costly and leaves no closed-form map. In the simulated NR scenario, with few training samples the Laplace approximation and MCMC both scored slightly above empirical Bayes on held-out log scores, the gain shrank quickly as training samples grew, and by larger sample sizes the three methods scored closely.
The method finds the posterior mode and then uses the Hessian of the negative log-posterior there as the precision matrix of the approximate posterior, yielding a Gaussian approximation to the hyperparameters. Finding the mode is identical to empirical Bayes, so the only addition is one Hessian evaluation at the mode, computed by automatic differentiation accumulated over batches of locations to avoid holding the full computational graph in memory. Across all reported fits on simulated data, the Newton decrement, which measures the remaining distance to the mode in posterior standard deviations, was less than one posterior standard deviation for every parameter, and the Hessian was positive definite at each estimate; on the CESM log-precipitation fields the decrement was small at 10 training fields but larger at 20, 35 and 50, at 4.90 and 2.78 among them.
On the real climate field, the Laplace marginals agree well with the MCMC posterior for all hyperparameters except the conditioning-set size, and predictive scores match MCMC. The conditioning-set size is an integer-valued step function, so the posterior is discontinuous in it; gradient-based optimization and the Hessian evaluated on the smooth part are blind to the jumps, so that marginal is misrepresented. On CESM global-grid log-precipitation fields fitted with 10, 20, 35 and 50 training fields, held-out log scores from the Laplace approximation and MCMC were similar and both continued the simulated-data trend of beating empirical Bayes.
Perspective
The demonstrated scope is the simulated NR scenario and univariate log-precipitation-rate anomaly fields on a global grid from the CESM Large Ensemble, with training sets of 10, 20, 35 and 50 fields; the method targets triangular Bayesian transport maps and their small hyperparameter vector.
The Laplace approximation assumes the posterior is near-normal around the mode, and it misrepresents uncertainty when the posterior is multimodal, strongly skewed, heavy-tailed or discontinuous; the conditioning-set size makes the posterior discontinuous, some real-data training sizes did not fully converge, and the authors suggest reparameterizing that size and adding an optimization stage over the remaining parameters with it held fixed.
