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.
Introduction
In non-linear state estimation, such as tracking a drone's position or a robot's path, a recurring challenge is solving "Gaussian integrals." These integrals calculate the expected value of a non-linear function when its input is a Gaussian random variable. Traditionally, engineers rely on sigma-point methods, such as the Unscented Transform (UT). These methods select a small set of "sigma points," propagate them through the non-linear function, and compute a weighted average.
While effective, these methods are typically derived under the assumption that the non-linear function is a polynomial. When the function is highly non-linear or "wiggly," these classical approaches can falter. This paper bridges the gap between heuristic sigma-point methods and the more rigorous probabilistic framework of Gaussian Process Quadrature (GPQ), which treats the integrand as a realization of a Gaussian Process (GP).
The Gaussian Process Transform (GPT)
The core contribution of this work is the formalization of the Gaussian Process Transform (GPT). The authors prove that any GPQ can be expressed as a sigma-point rule: $$\int g(x) N(x|m,P) dx \approx \sum W_i g(x_i)$$ Unlike standard methods where weights $W_i$ are fixed, GPQ weights are derived from the GP's covariance function (kernel).
A key finding is the kernel equivalence: using a specific polynomial kernel exactly recovers the weights used in the Unscented Kalman Filter (UKF) and Gauss-Hermite quadrature. This reveals that the popular UKF is actually a special case of Bayesian quadrature with a 3rd-order polynomial kernel.
Point Selection and Flexibility
Beyond weight derivation, the GPQ framework allows for greater flexibility in selecting sigma-point locations ($x_i$). While classical methods are "exact" for polynomials of a certain degree, GPQ methods aim to minimize the "average error" across a class of functions defined by a kernel.
The paper explores several point selection strategies:
- Minimum Variance Points: Optimizing locations to minimize expected integration error.
- Quasi-Random Sets: Utilizing Hammersley point sets to better cover high-dimensional spaces.
- Kernel Choice: Using a Squared Exponential (SE) kernel allows the filter to "ignore" points far from the mean, providing a localized approximation that is often more accurate than global polynomial fits.
Experimental Results
The proposed GPQ-Filter and GPQ-Smoother were evaluated against synthetic benchmarks:
- High-Dimensional Moment Computation: GPQ methods using Hammersley points significantly outperformed standard cubature rules as dimensionality increased.
- Univariate Non-linear Growth Model (UNGM): GPQ-based filters achieved significantly lower Root Mean Square Error (RMSE) than EKF, UKF, and standard Cubature Kalman Filters.
- Bearings-Only Target Tracking: GPQ methods performed competitively with high-order Gauss-Hermite filters, demonstrating viability for complex real-world tasks.
Conclusion
By reframing sigma-point methods as Gaussian Process Quadrature, this work provides a more flexible and robust framework for non-linear filtering. While calculating GPQ weights involves matrix inversion—adding some computational overhead—the gains in accuracy and stability, particularly in high-dimensional or highly non-linear scenarios, are substantial. Future work will explore automated hyperparameter tuning and the application of more diverse kernels, such as Matern kernels, to specific physical systems.
Summary updated