Every Gaussian filter for a nonlinear model needs the mean and covariance of y = g(x) when x ~ N(μ, sΣ). The Taylor transform (red, used by the EKF) linearises g at μ; the unscented transform (blue, UKF) pushes 2n + 1 sigma points through g; Gauss–Hermite (green, GHKF) uses a grid of p2 points. The white ellipse is the Gaussian with the exact mean and covariance of y: that is the best any Gaussian filter can do (moment matching). Grey dots are samples of x and of y.
g(x) = (x1 + a x22 − a, x2 + a sin 2x1), μ = 0, Σ = [[0.6, −0.55], [−0.55, 1.6]]. "Exact" moments use a 30 × 30 Gauss–Hermite grid. KL is KL(exact Gaussian ‖ approximation); ellipses are 2 standard deviations.
ch08_gaussian_filtering_smoothing.py (taylor_transform, sigma_points, gauss_hermite_points, moments_by_points, Figure 8.5).