Research map / Probabilistic and scientific modeling / Variational and generative sampling
Pareto Smoothed Importance Sampling
Aki Vehtari, Daniel Simpson, Andrew Gelman, Yuling Yao, Jonah Gabry
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ordinary IS with PSIS | importance sampling · Exponential target, Exponential(theta) proposals | RMSE ratio: PSIS always better than IS | Improved |
| replace truncated IS with PSIS | importance sampling · Exponential(2) proposal (k-hat about 0.5) | RMSE: TIS slightly better at intermediate sample sizes | Hurt |
| add PSIS correction to split-normal approximation | Split-normal · Logistic GP density estimation (Galaxy) | KL divergence to MCMC: small KL-divergence already when S is 100 | Improved |
| use plain IS instead of PSIS | Split-normal · Logistic GP density estimation (Galaxy) | KL variability: mean k-hat 0.43 (S=100) and 0.55 (S=10^4) | Hurt |
| add split scaling to the normal approximation | Normal approximation · Logistic GP density estimation (Galaxy) | mean Pareto k-hat: 0.60 -> 0.55 | Improved |
| use PSIS-LOO instead of exact LOO | Leave-one-out cross-validation · 105 protein expression models | computation time: 125 days -> less than one minute | Improved |
| switch from Gaussian to Student-t observation model | Linear model for protein expression · 105 protein expression models | number of cases with k-hat > 0.7: 352 -> 53 | Improved |
| add exact LOO for cases with k-hat > 0.7 | PSIS-LOO · 105 protein expression models | LOO accuracy: 42 hours extra computation | Improved |
| increase dimension D from 1 to 1024 | Importance sampling with Student-t proposal · D-dimensional normal target, S=10000 | effective sample size: D from 1 to 1024 | Hurt |
| apply sample-size dependent k-hat threshold | PSIS diagnostic · k-hat threshold | threshold: S=100 -> 0.5 | Mixed |
Closest papers by meaning
- Entropic Mirror Monte Carlo (ICML 2026 · similarity 0.86)
- A Quadrature Rule combining Control Variates and Adaptive Importance Sampling (NeurIPS 2022 · similarity 0.86)
- Stein $\Pi$-Importance Sampling (NeurIPS 2023 · similarity 0.85)
- Estimating the Permanent by Nesting Importance Sampling (ICML 2024 · similarity 0.84)
- Enhanced Importance Sampling Through Latent Space Exploration in Normalizing Flows (AAAI 2025 · similarity 0.83)
- Differentiable Annealed Importance Sampling Minimizes The Jensen-Shannon Divergence Between Initial and Target Distribution (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
