Gaussian Integral based Bayesian Smoother

Rohit Kumar Singh, Kundan Kumar, Shovan Bhaumik · 2025

Read Gaussian Integral based Bayesian Smoother by Rohit Kumar Singh, Kundan Kumar, Shovan Bhaumik (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-06881.txt.

Introduction

Rohit Kumar Singh, Kundan Kumar, and Shovan Bhaumik smooth noisy state estimates with exact Gaussian integrals in this 2025 paper. Smoothing revisits past states of a system using later data, a staple of tracking and control. Nonlinear systems force approximations, and common smoothers cut corners in different ways.

Their smoother assumes bell-shaped state distributions and evaluates the needed averages with Gaussian integration. That integral is exact for polynomial shapes, which suits polynomial nonlinearities perfectly. A classic benchmark, the Van der Pol oscillator, supplies the test case. Simulations show the new smoother beating extended, cubature, and unscented rivals. The edge grows where polynomial terms dominate the dynamics.

Catalog record

CREATORRohit Kumar Singh, Kundan Kumar, Shovan Bhaumik
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-06881

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Gaussian Integral based Bayesian Smoother Rohit Kumar Singh, Kundan Kumar, Shovan Bhaumik Abstract This work introduces the Gaussian integration to address a smoothing problem of a nonlinear stochastic state space model. The probability densities of states at each time instant are assumed to be Gaussian, and their means and covariances are evaluated by utilizing the odd-even properties of Gaussian integral, which are further utilized to realize Rauch-Tung-Striebel (RTS) smoothing expressions. Given that the Gaussian integration provides an exact solution for the integral of a polynomial function over a Gaussian probability density function, it is anticipated to provide more accurate results than other existing Gaussian approximation-based smoothers such as extended Kalman, cubature Kalman, and unscented Kalman smoothers, especially when polynomial types of nonlinearity are present in the state space models. The developed smoothing algorithm is applied to the Van der Pol oscillator, where the nonlinearity associated with their dynamics is represented using polynomial functions. Simulation results are provided to demonstrate the superiority of the proposed algorithm. I Introduction In this paper, we developed an approximated Gaussian integral-based smoothing algorithm for the nonlinear stochastic state space models. The Smoothing algorithms find application in various areas like target tracking [ 1 ] , navigation [ 2 ] , robotics [ 3 ] , image analysis [ 4 ] , statistical modelling [ 5 ] , and finances [ 6 ] . Here, we consider a discrete-time stochastic state space model in the following form [ 7 ] \displaystyle x_{k}= \displaystyle f(x_{k-1})+\mu_{k-1}, (1) \displaystyle y_{k}= \displaystyle h(x_{k})+\nu_{k}, (2) where x_{k}\in\mathbb{R}^{n} and y_{k}\in\mathbb{R}^{m} are the state of the system and measurement, respectively, at k^{th} instant, k\in 1,2,\ldots,T . The state evaluation function, f(x):\mathbb{R}^{n}\rightarrow\mathbb{R}^{n} and measurement mapping h(x):\mathbb{R}^{n}\rightarrow\mathbb{R}^{m} are known nonlinear functions. The process and measurement noises are uncorrelated, and they follow a Gaussian probability density function (pdf) given as \mu_{k-1}\sim\mathcal{N}(0,Q_{k-1}) and \nu_{k}\sim\mathcal{N}(0,R_{k}) , respectively, where Q_{k-1} and R_{k} are the process and measurement noise covariance matrix, respectively. Smoothing in the context of estimation refers to Bayesian methodology for estimating the past state’s history of a system using the sensor noisy measurements available up to the final time step. It enhances the estimation performance by refining earlier state estimates obtained from the filtering [ 7 ] . The smoother computes the marginal posterior distribution of the state ( x_{k} ) given measurement up to time step T i.e. p(x_{k}|y_{1:T}) in two steps: (i) forward filtering and (ii) backward filtering. The forward filter step involves recursively computation of the posterior state estimate p(x_{k}|y_{k}) [ 8 , 7 ] \displaystyle p(x_{k}|y_{k-1}) \displaystyle=\int p(x_{k}|x_{k-1})p(x_{k-1}|y_{1:k-1})dx_{k-1} (3) \displaystyle p(x_{k}|y_{k}) \displaystyle\propto p(y_{k}|x_{k})p(x_{k}|y_{1:k-1}). (4) The Backward filtering step computes p(x_{k}|y_{1:T}) recursively backwards starting from the final time step ( T ) p(x_{k}|y_{1:T})=p(x_{k}|y_{1:k})\int\dfrac{p(x_{k+1}|x_{k})p(x_{k+1}|y_{1:T})}{p(x_{k+1}|y_{1:k})}dx_{k+1}. (5) For the linear systems, the distribution remains Gaussian, and the Rauch-Tung-Striebel smoother (RTSS) provide a closed-form solution [ 9 , 7 ] . However, for nonlinear systems, the distributions no longer remain Gaussian, and the closed-form solutions are not available [ 10 ] . In literature, many times, using moment matching, these distributions are approximated to follow the Gaussian distribution [ 7 , 10 , 11 , 12 , 13 ] , and subsequently mean and their covariances are being computed. In this work, we use these Gaussian approximation-based methods. Other approximation methods also exist, such as the sequential Monte Carlo [ 14 ] and other related approaches. Various Gaussian approximation-based smoothing algorithms exist in the literature [ 7 ] , such as Taylor series approximation-based extended RTS smoother (ERTSS) [ 15 ] and numerical approximation-based smoother, including the unscented RTSS (URTSS) [ 12 ] , cubature RTSS (CRTSS) [ 16 ] , Gauss-Hermite RTSS [ 17 ] , Fourier-Hermite RTSS [ 18 ] , among others. In these Gaussian approximation-based RTS smoother, to realize them, the integrals in ( 3 )-( 5 ) need to be evaluated. If we look closely, these integrals are of the form \int nonlinearfunction\times Gaussian\;pdf . In this work, we propose to solve the smoothing problems using the Gaussian integral as developed in [ 19 ] . The Gaussian integral provides an exact solution for the integration mentioned in RTS smoothing for a non-linear polynomial function. So, if we use Gaussian integral to develop a smoother for a nonlinear system having polynomial type of nonlinearity, it is expected to perform better in terms of estimation accuracy than existing Gaussian approximation smoothing algorithms [ 7 ] . When the nonlinear function is not of polynomial type, the developed method can still be applied after approximating the function using the Taylor series expansion. However, the integration with such expansion will no longer be exact as we will have to use the truncated Taylor series expansion. To demonstrate the effectiveness of the proposed smoothing algorithm, we apply it to the Van der Pol (VDP) oscillator and compare the results with existing methods. The VDP oscillator, known for its non-linear damping and complex dynamics, is widely used in fields such as electrical engineering, biology [ 20 ] , and physics [ 21 ] . II Backward filtering in Bayesian smoother Non-linear forward filtering is the process of estimating the hidden variable of a time-varying system using noisy sensor measurements, conditioned that the system is observable [ 22 ] . The integrations involved in calculating the posterior state estimate and covariance matrix during the forward filtering are provided in [ 23 , pp.2] or [ 13 , pp.3] . The backward filtering procedure for the RTS smoother in the Bayesian framework with Gaussian assumptions is derived below. The joint distribution of x_{k} and x_{k+1} , given the measurements available up to instant k is given as p(x_{k},x_{k+1}|y_{1:k})=p(x_{k+1}|x_{1:k})p(x_{k}|y_{1:k}), (6) where p(x_{k+1}|x_{1:k})\sim\mathcal{N}(x_{k+1};\hat{x}_{k+1|k},P_{xx,k+1|k}) and p(x_{k}|y_{1:k})\sim\mathcal{N}(x_{k};\hat{x}_{k|k},P_{xx,k|k}) are pdf for predicted states and posterior state estimate, obtained from the forward filtering. Hence, the joint pdf of x_{k} and x_{k+1} conditioned over measurement y_{1:k} using [ 7 , lemma A.2] , is given as p(\begin{bmatrix}x_{k}\\ x_{k+1}\end{bmatrix}|y_{1:k})\sim\mathcal{N}(\begin{bmatrix}\hat{x}_{k|k}\\ \hat{x}_{k+1|k}\end{bmatrix},\begin{bmatrix}P_{xx,k|k}&P_{x_{k},x_{k+1}}\\ P_{x_{k},x_{k+1}}^{\top}&P_{xx,k+1|k}\end{bmatrix}), (7) where the cross-covariance matrix ( P_{x_{k},x_{k+1}} ) is [ 7 , p. 279] \begin{split}&P_{x_{k},x_{k+1}}=\mathbb{E}[(x_{k}-\hat{x}_{k|k})(x_{k+1}-\hat{x}_{k+1|k})^{\top}],\\ &=\int x_{k}x_{k+1}^{\top}\mathcal{N}(x_{k};\hat{x}_{k|k},P_{xx,k|k})dx_{k}-\hat{x}_{k|k}\hat{x}_{k+1|k}^{\top},\\ &=\int x_{k}f(x_{k})^{\top}\mathcal{N}(x_{k};\hat{x}_{k|k},P_{xx,k|k})dx_{k}-\hat{x}_{k|k}\hat{x}_{k+1|k}^{\top}.\end{split} (8) The smoothing pdf for time step (k+1) is p(x_{k+1}|y_{1:T}) , which is assumed to be known and Gaussian, is given as p(x_{k+1}|y_{1:T})\sim\mathcal{N}(x_{k+1};\hat{x}_{k+1|T}^{s},P_{xx,k+1|T}^{s}). (9) Following [ 7 , pp. 255–257] , we obtain the smoothing distribution at time step k , p(x_{k}|y_{1:T})\sim\mathcal{N}(x_{k};\hat{x}_{k|T}^{s},P_{xx,k|T}^{s}) , where \begin{split}\hat{x}_{k|T}^{s}=&\hat{x}_{k|k}+P_{x_{k},x_{k+1}}P_{xx,k+1|k}^{-1}(\hat{x}_{k+1|T}^{s}-\hat{x}_{k+1|k}),\end{split} (10) \begin{split}P_{xx,k|T}^{s}=&P_{xx,k|k}+P_{x_{k},x_{k+1}}P_{xx,k+1|k}^{-1}(P_{xx,k+1|T}^{s}-\\ &P_{xx,k+1|k})(P_{x_{k}.x_{k+1}}P_{xx,k+1|k}^{-1})^{\top}.\end{split} (11) From the above expressions, we see that the backward filtering is also a recursive process proceeding backwards in time starting from the final time step T , such that the smoothing distribution of states at time k is derived from the smoothing distribution at (k+1) . The parameters such as \hat{x}_{k|k} , P_{xx,k|k} , \hat{x}_{k+1|k} , and P_{xx,k+1|k} , in the ( 10 ) and ( 11 ) are fetched from the forward filtering. \hat{x}_{k+1|T}^{s} and P_{xx,k+1|T}^{s} is known and initialized with the posterior estimate at the final time step T , i.e \hat{x}_{T|T}^{s}=\hat{x}_{T|T},P_{xx,T|T}^{s}=P_{xx,T|T} . P_{x_{k},x_{k+1}} is calculated using ( 8 ). III Smoothing using Gaussian Integral In this section, we propose a Gaussian-integral RTS smoother (GIRTSS) for the stochastic state-space model from ( 1 ) and ( 2 ). The realisation of the smoothing algorithm demands the evaluation of integrals from forward filtering mentioned in [ 13 ,

Read the complete text