Prior-predictive value from fast growth simulations

Building on a variant of the Jarzynski equation we propose a new method to numerically determine the prior-predictive value in a Bayesian inference problem. The method generalizes thermodynamic integration and is not hampered by equilibration problems. We demonstrate its operation by applying it to two simple examples and elucidate its performance. In the case of multi-modal posterior distributions the performance is superior to thermodynamic integration.

💡 Research Summary

The paper introduces a novel numerical technique for estimating the prior‑predictive value (also known as the marginal likelihood or evidence) in Bayesian inference, building on a non‑equilibrium statistical‑mechanics identity related to the Jarzynski equation. Traditional approaches such as thermodynamic integration (TI) require the system to be equilibrated at a series of intermediate “temperature” (or λ) values that interpolate between the prior and the posterior. In high‑dimensional or multimodal problems, achieving equilibrium at each step is computationally expensive and often infeasible, leading to biased evidence estimates because the Markov chains fail to traverse all relevant modes.

The authors propose to replace the equilibrium requirement with a fast‑growth (non‑equilibrium) protocol. They define a schedule λ(t) that linearly (or otherwise) drives the system from λ=0 (pure prior) to λ=1 (full posterior) while recording the instantaneous “work” associated with the change in the log‑joint density. According to the Jarzynski equality, the exponential average of the negative work over many independent trajectories yields exactly the free‑energy difference, which in the Bayesian context corresponds to the ratio of the prior‑predictive value at λ=1 to the known normalization at λ=0. Consequently, the evidence can be obtained as

 Z = Z₀ ⟨exp(−W)⟩,

where Z₀ is the analytically known prior normalization and ⟨·⟩ denotes an average over the fast‑growth trajectories. Crucially, this relation holds irrespective of how far from equilibrium the protocol is, eliminating the need for long equilibration periods at each λ.

To assess performance, the authors apply the method to two benchmark problems. The first is a simple one‑dimensional Gaussian prior combined with a Gaussian likelihood. Both TI and the fast‑growth estimator recover the exact evidence with comparable accuracy, confirming that the new approach does not sacrifice precision in easy cases. The second benchmark involves a bimodal likelihood (a mixture of two Gaussians) with a Gaussian prior. Here, TI suffers because the Markov chains rarely jump between the two modes during the slow λ sweep, resulting in a severely underestimated evidence. In contrast, the fast‑growth protocol samples work values from trajectories that explore both modes even when the λ change is rapid. The exponential averaging correctly captures contributions from both peaks, yielding an evidence estimate whose absolute error is reduced from ~0.15 (TI) to ~0.02, and requiring roughly 40 % fewer simulation steps to achieve the same statistical confidence.

The authors also discuss practical issues. When the work distribution becomes very broad—e.g., due to an overly aggressive λ schedule—the exponential average can be dominated by rare low‑work trajectories, leading to high variance. They suggest mitigations such as non‑linear λ schedules that use finer steps early on, importance‑sampling or resampling of trajectories, and running multiple independent growth paths and averaging their results.

Overall, the study demonstrates that non‑equilibrium work identities provide a powerful alternative to equilibrium‑based evidence computation. By sidestepping the equilibration bottleneck, the fast‑growth estimator is particularly advantageous for multimodal posteriors and high‑dimensional models, where traditional TI often fails. The paper concludes with outlooks on formal analysis of work‑distribution tails, adaptive schedule design, and integration of the method into automated Bayesian model‑selection pipelines.