A limited-memory acceleration strategy for MCMC sampling in hierarchical Bayesian calibration of hydrological models

Hydrological calibration and prediction using conceptual models is affected by forcing/response data uncertainty and structural model error. The Bayesian Total Error Analysis methodology uses a hierarchical representation of individual sources of uncertainty. However, it is shown that standard multiblock Metropolis-within-Gibbs Markov chain Monte Carlo (MCMC) samplers commonly used in Bayesian hierarchical inference are exceedingly computationally expensive when applied to hydrologic models, which use recursive numerical solutions of coupled nonlinear differential equations to describe the evolution of catchment states such as soil and groundwater storages. This note develops a limited-memory algorithm for accelerating multiblock MCMC sampling from the posterior distributions of such models using low-dimensional jump distributions. The new algorithm exploits the decaying memory of hydrological systems to provide accurate tolerance-based approximations of traditional full-memory MCMC methods and is orders of magnitude more efficient than the latter.