[Paper Review] On the relation between Gaussian process quadratures and sigma-point methods

Simo Särkkä, Jouni Hartikainen+2 more|arXiv (Cornell University)|Apr 22, 2015

Target Tracking and Data Fusion in Sensor Networks18 references15 citations

TL;DR

This paper establishes a theoretical connection between Gaussian process quadrature and sigma-point methods, showing that many sigma-point filters and smoothers—such as unscented, cubature, and Gauss–Hermite filters—can be interpreted as special cases of Gaussian process quadrature with specific covariance functions. The key contribution is a unified framework that interprets these methods through Gaussian process regression, enabling improved numerical integration via flexible covariance selection and superior performance in nonlinear filtering and smoothing tasks.

ABSTRACT

This article is concerned with Gaussian process quadratures, which are numerical integration methods based on Gaussian process regression methods, and sigma-point methods, which are used in advanced non-linear Kalman filtering and smoothing algorithms. We show that many sigma-point methods can be interpreted as Gaussian quadrature based methods with suitably selected covariance functions. We show that this interpretation also extends to more general multivariate Gauss--Hermite integration methods and related spherical cubature rules. Additionally, we discuss different criteria for selecting the sigma-point locations: exactness for multivariate polynomials up to a given order, minimum average error, and quasi-random point sets. The performance of the different methods is tested in numerical experiments.

Motivation & Objective

• To establish a theoretical and practical connection between Gaussian process quadrature and sigma-point methods used in nonlinear filtering and smoothing.
• To demonstrate that widely used sigma-point filters (e.g., unscented, cubature, Gauss–Hermite) are special cases of Gaussian process quadrature with appropriately chosen covariance functions.
• To unify diverse numerical integration techniques under a common framework based on Gaussian process regression.
• To evaluate and compare different sigma-point selection criteria—exactness for polynomials, minimum average error, and quasi-random sets—for improved integration accuracy.

Proposed method

• The paper formulates numerical integration as a Bayesian inference problem using Gaussian process regression, where the integral of the function is approximated by integrating the posterior mean of the GP.
• It shows that by selecting a squared exponential covariance function and a Gaussian weight function, the integral of the GP posterior can be computed in closed form, linking to Bayes–Hermite quadrature.
• The authors derive conditions under which sigma-point methods (e.g., unscented transform, Gauss–Hermite rules) emerge as special cases of GP quadrature by matching moment conditions and covariance structure.
• Different sigma-point selection strategies are analyzed: exactness for multivariate polynomials up to a given degree, minimization of average integration error, and use of quasi-random sequences.
• The framework is extended to spherical cubature rules and multivariate Gauss–Hermite integration by leveraging the structure of spherically symmetric integration formulas.
• Numerical experiments are conducted to compare performance across different sigma-point configurations and integration rules, using both synthetic and real-world filtering problems.

Experimental results

Research questions

• RQ1 Can sigma-point methods such as the unscented Kalman filter and Gauss–Hermite quadrature be interpreted as instances of Gaussian process quadrature with specific covariance functions?
• RQ2 How do different criteria for sigma-point selection—exactness for polynomials, minimum average error, and quasi-random sequences—affect integration accuracy?
• RQ3 To what extent do Gaussian process quadrature rules outperform traditional polynomial-based integration rules in nonlinear filtering and smoothing?
• RQ4 What is the theoretical relationship between multivariate Gauss–Hermite rules and spherically symmetric integration formulas in the context of GP quadrature?
• RQ5 How can the Fourier–Hermite series representation of functions and Gaussian processes be used to derive and analyze the structure of GP quadrature rules?

Key findings

• Many sigma-point methods, including unscented, cubature, and Gauss–Hermite filters, are shown to be special cases of Gaussian process quadrature when using specific polynomial covariance functions.
• The proposed GP quadrature framework with squared exponential covariance functions enables closed-form computation of integrals, linking to Bayes–Hermite quadrature and extending to multivariate settings.
• Numerical experiments demonstrate that GP quadrature rules often outperform classical polynomial-based integration rules and traditional sigma-point filters in terms of accuracy and robustness.
• Sigma-point selection based on minimizing average integration error or using quasi-random sequences leads to improved performance over standard methods, especially in high-dimensional or non-polynomial regimes.
• Theoretical analysis confirms that the method generalizes to symmetric integration rules and spherical cubature, providing a unified view of numerical integration techniques.
• The use of Fourier–Hermite series and Hilbert space representations enables a deeper understanding of the functional structure of GP-based integration and its convergence properties.