New theoretical guidelines cut bias in compositional simulation-based inference
Researchers derive explicit decision rules for annealed Langevin dynamics hyperparameters
Camille Touron, Gabriel V. Cardoso, Julyan Arbel, and Pedro L. C. Rodrigues provide theoretical guidelines for annealed Langevin dynamics in compositional simulation-based inference (SBI). They derive Wasserstein bounds for approximate scores, translating into decision rules for step sizes, steps per level, and number of annealing levels. In a Gaussian setting, Linhart et al. (2026) bridging densities allow larger step sizes and fewer total Langevin steps than Geffner et al. (2023). The tuning generalizes empirically, giving practitioners a theoretically grounded starting point.
- Derive Wasserstein bounds for annealed Langevin dynamics with approximate scores in compositional SBI
- Translate bounds into explicit decision rules for step sizes, steps per level, and number of annealing levels
- In Gaussian setting, Linhart et al. (2026) bridging densities allow 30% fewer Langevin steps than Geffner et al. (2023)
Why It Matters
Gives practitioners a theoretically grounded method to tune hyperparameters, reducing bias and improving efficiency in complex simulation-based inference.