Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.27442

Geometry-informed importance sampling for rare events

Qinyi Wang·Anya Katsevich·Shanyin Tong
View original source ↗Read source document ↗Export metadata ↓

Abstract

We propose GeoIS, a new importance sampling method for estimating rare event probabilities for an underlying Gaussian distribution that exploits asymptotic theory and local event geometry. The proposal distribution is a mixture of two components. The first component draws a Gaussian along the boundary near the instanton and adds to it an exponentially distributed shift into the rare event. This component is derived from the known asymptotic approximation to the optimal zero-variance proposal distribution. The second mixture component is designed to cover missing event mass due to boundary-approximation mismatch, thereby preserving unbiasedness. We rigorously analyze this sampling strategy's performance as a function of event geometry, dimension, rarity level, and number $N$ of importance samples. Under explicit and relatively mild conditions on the former three features, we prove that the squared relative error of GeoIS is bounded by $c/N$ with high probability, for an absolute constant $c$. We implement our method numerically in several examples, including a diffusion PDE problem. The numerical experiments exhibit the qualitative dependence on geometry, dimension, rarity and sample size predicted by the theory, and show that GeoIS markedly outperforms importance sampling methods that use a fully Gaussian proposal distribution.

Explore related subjects

stat.COmath.PRmath.STstat.TH

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Qinyi Wang, Anya Katsevich, Shanyin Tong. 2026-10-20. Geometry-informed importance sampling for rare events. https://arxiv.org/abs/2610.27442

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

When Scatter Plots Hide the Data: Diagnosing Ties and Choosing the Right Display in R

Scatter plots draw one mark per observation, so observations that share a location are hidden. In a two-item Likert survey of 2,757 respondents, 98.7% of the data lie under other points and the plot shows only 36 dots. The density-distribution sunflower plot of Dupont and Plummer addresses this by drawing raw points in sparse regions and sunflowers, whose petals count observations, in dense ones. It is not available in R, and its single petal weight for dense cells fails when cell counts span more than an order of magnitude: small dense cells collapse to one petal, and the count a reader decodes can be wrong by 150% or more. We propose geometric petal tiers, in which each cell uses the smallest weight among 1, b, b2, . . . observations per petal that keeps its flower within P petals. Every flower then has between about P/b and P petals, and the relative error of every decoded count is below b/(2P), 17.9% at the defaults b = 5 and P = 14. We also define tie diagnostics and decision rules that use them to choose between an ordinary scatter plot, jitter or transparency, a classic sunflower plot and geometric tiers. Both are implemented in the ggplot2 extension ggsunflower. We illustrate the methods on survey, automotive, historical and large commercial data, and in a simulation over 35 conditions geometric tiers keep the worst-cell error at or below 14% where a single weight reaches a median of 251%.

stat.CO↗
arXiv2026

Bayesian Treed Vecchia Gaussian Process Emulation via Sequential Monte Carlo

Computer models (or simulators) are used in many scientific domains to understand and predict real-world processes, but they are expensive to run. One way to address this hurdle is to use Gaussian processes (GPs). For complex physical systems such as storm surge, simulator outputs often exhibit heterogeneous structures, and many nonstationary GP methods become impractical with several inputs and tens of thousands of runs. To address these issues, we combine the divide-and-conquer idea of treed GP models with Vecchia approximations in a Bayesian treed Vecchia GP (TVGP) model. The model uses the Bayesian classification and regression tree prior for random recursive partitioning of the input domain. Within each leaf, a Vecchia approximation allows efficient likelihood evaluation. For Bayesian inference, we develop a sequential Monte Carlo algorithm that explores tree space as an alternative to reversible-jump Markov chain Monte Carlo. The method is demonstrated on synthetic examples with two and ten inputs and on a storm-surge simulator with seven inputs and up to 50,649 training runs. TVGP has the smallest prediction error in both synthetic examples and with the larger storm-surge training set, although its predictive intervals undercover in the example with ten inputs and the storm-surge application.

stat.CO↗
arXiv2026

Reducing Complexity of Pseudo-Marginal Sequential Monte Carlo via Hierarchical Tempering

The Bayesian approach to inverse problems provides a powerful framework for inferring unknown parameters from noisy observations. Among the computationally most challenging statistical inference problems are problems with intractable likelihoods, especially if no unbiased likelihood estimator is available. We focus on a class of such problems, where a biased likelihood estimator is obtained through a stochastic approximation of a forward map (stochastic solver). Obtaining an empirical posterior approximation in applications is very expensive due to the repeated stochastic solver evaluations required by standard sampling methods. To address this computational bottleneck, we develop a novel algorithm based on a density-tempered marginalized Sequential Monte Carlo (SMC) sampler. Our approach (hierarchical SMC) accelerates inference by leveraging a hierarchy of stochastic solvers with increasing accuracy both along the tempering sequence and within mutation stages of the SMC sampler. We establish, under additional assumptions, that the asymptotic speedup is proportional to the number of tempering stages $T$. We demonstrate computational gains on several examples, including a nonlinear, high-contrast inverse problem arising in photoacoustic medical imaging. On this problem, we demonstrate a speedup by a factor of 14.3, reducing the inference wall-clock time from about 4.6 days to 7.7 hours on an NVIDIA A40 GPU.

stat.CO↗