1. How is the measurement update performed in non-Gaussian filtering?
In non-Gaussian filtering, the measurement update is performed by multiplying two densities. Although the densities are non-Gaussian, the measurement update (3) can be easily executed. However, obtaining an explicit form of the one-step prediction in (4) can be challenging when the densities are not Gaussian. The problem then becomes approximating r xt+1|Yt. The power moments of r xt+1|Yt are easy to obtain, and by using the method of moments, truncated power moments can be used to estimate r xt+1|Yt. This approach differs from previous research, which focused on density approximation. The proposed filter admits treating the non-Gaussian state estimation problem and provides an analytic error analysis to measure performance. The density estimate for the one-step prediction at time t + 1 is chosen as the density surrogate, rxt+1|Yt, which is an order-2n density surrogate. This problem is treated as a Hamburger moment problem, and a solution is provided in the next section.
read more
2. What is the concept of sub-Gaussian distributions and how are they related to the moment error propagation in light-tailed density surrogate?
Sub-Gaussian distributions are distributions whose tails decay at least as fast as a Gaussian distribution. They are denoted by the space SG. The moment error propagation in light-tailed density surrogate is related to the concept of sub-Gaussian distributions as it ensures the existence and boundedness of all power moments of the estimated densities. This allows for the approximation of the true densities using truncated power moments without introducing uncontrollable cumulative errors. Theorem 4.2 proves that the first 2n moment terms of the estimated densities with the density surrogate are approximately the true ones throughout the whole filtering process for sufficiently large n, given that all power moments of the true system states exist and are finite.
read more
3. What is the relationship between power moments errors and L norm of density surrogate error?
The errors of power moments of rxt+1|Yt+1 are bounded by a value proportional to the L norm of the error of the density surrogate rxt+1|Yt. This relationship is established in Theorem 4.3. It indicates that a satisfactory performance of density estimation, characterized by a relatively small max x r xt+1|Yt, results in a small error of the estimated moments of the density. The theorem highlights the importance of accurate density estimation in minimizing the error of power moments. However, it is essential to note that in real applications, the infinite-dimensional estimation problem cannot be fully addressed, and the density estimate may not always match the true density for either the Stochastic Gradient (SG) or the Stochastic Gradient with Hinge (SG/Hinge) methods. The next section aims to analyze the error upper bounds of r to determine its maximum deviation from the true density, considering the first 2n terms of power moments.
read more
4. What is the error upper bound for the state estimate in Bayesian filtering with non-Gaussian distributions?
The error upper bound for the state estimate in Bayesian filtering with non-Gaussian distributions has not been established. However, in this section, an error upper bound of r(x) in the sense of total variation distance is proposed. This upper bound distinguishes the proposed filter from other Bayesian filters. The total variation distance between the density estimate r and the true density r is defined as EQUATION. The Shannon-entropy maximizing distribution F r, which has the same moments as the density r, is used to calculate the upper bound. By following the KL distance and applying Theorem 4.2, the error upper bound is obtained as EQUATION. This upper bound provides a measure of the error in the state estimate and is useful for conservative decision-making in practical situations such as financial applications. The proposed algorithm ensures that the power moments of the density estimates are approximately the true ones, leading to bounded and small estimation errors. The upper bounds for the probability of subsets of the real line, given the power moments, are also derived, providing further insights into the error analysis of the state estimates.
read more