Does the final estimate change during the backward initialization?

In the standard full-interval RTS construction, the final smoothed estimate equals the final filtered estimate.

Does it require derivatives?

Sigma-point moment calculations avoid the usual explicit Jacobians.

Can it recover an incorrect motion model?

It may expose discrepancies, but later data do not automatically correct the model assumptions.

AI Glossary / Probabilistic machine learning / Advanced

Unscented Kalman smoother

Definition

An unscented Kalman smoother estimates past hidden states using observations from an entire interval, applying sigma-point approximations to nonlinear state relationships. A common form is the unscented Rauch - Tung - Striebel smoother, which starts from forward-filtered estimates and propagates corrections backward. It retains approximate state means and covariances rather than the full nonlinear smoothing distribution.

What it means

A live filter can only use measurements that have arrived. After a journey or experiment is complete, later measurements may clarify earlier uncertainty. Smoothing uses that additional evidence to revise historical state estimates.

The word unscented indicates how nonlinear moment calculations are approximated: deterministic sigma points are passed through the dynamics. It does not mean that future states are predicted with certainty, nor that arbitrary nonlinear posteriors become Gaussian. The common unscented RTS construction depends on cross-covariance between successive states, which determines how much a later correction should influence an earlier estimate.

How it works

Run an appropriate forward filter and retain the filtered means and covariances. For each transition, use sigma points from the filtered state to approximate the next-state predicted mean, covariance and cross-covariance with the current state. Independent additive process-noise covariance contributes to the predicted next-state covariance.

At the final time, initialize the smoothed mean and covariance to the final filtered values. Then move backward through time. The smoother gain is the current-to-next cross-covariance multiplied by the inverse predicted next-state covariance.

Correct the current filtered mean by this gain times the difference between the next smoothed mean and its earlier prediction. Correct the covariance using the next smoothed covariance minus its predicted covariance, multiplied by the gain on both sides.

This is a probabilistic correction, not an ordinary backward average of a plotted trajectory. Store or consistently recompute the quantities needed by the selected algorithm. Numerical solves are preferable to explicitly forming matrix inverses, and covariance checks remain important because finite precision and approximate moments can undermine a theoretically valid calculation.

Three practical examples

1. A robotics team reconstructs a completed route using later position observations to improve earlier estimates during temporary sensor uncertainty.

2. An engineering analyst estimates historical equipment states from a complete nonlinear sensor sequence after a test has finished.

3. A model-development team obtains approximate smoothed state moments for diagnosing a state-space model or supporting parameter estimation.

Think of it this way

Think of revising a travel diary after discovering later photographs. A photograph from the end of a road helps reinterpret where an earlier ambiguous picture was taken. The transition model determines how strongly that later evidence connects to the earlier moment. A weak connection should produce a small correction, not an equally weighted rewrite.

Worked example

Use a one-dimensional identity transition with independent process-noise variance one. At time t, suppose the forward filter has mean 10 and variance four. Use the scalar sigma-point rule with alpha = 1, kappa = 0 and beta = 2.

The sigma points are 10, 12 and 8. Their mean weights are zero, one half and one half; the central covariance weight is two. Under the identity transition, the transformed points are unchanged. Their predicted mean is 10 and state-induced variance is four. Adding process noise gives predicted next-state variance five, while current-to-next cross-covariance remains four.

Suppose later evidence produces next-state smoothed mean 12 and variance two. The smoother gain is 4/5 = 0.8. The current smoothed mean becomes 10 + 0.8 times (12 minus 10) = 11.6.

Its variance becomes 4 + 0.8 squared times (2 minus 5) = 2.08. The later evidence shifts the earlier estimate and reduces its uncertainty in this calculation.

This deliberately linear transition is an arithmetic check: the unscented moments agree with ordinary Kalman moments here. A nonlinear transition requires propagating the actual sigma points through that nonlinear function. The example is one backward correction using stipulated valid next-state information, not a complete fitted dataset.

Illustrative case study

Fictional case: Maple Robotics reviews a completed inspection run after a positioning sensor briefly loses quality. Its live UKF estimates remain uncertain during the interruption, while observations after recovery provide additional information about the route.

An unscented RTS smoother uses the later evidence to revise earlier positions. Analysts retain the live estimates for evaluating real-time navigation and the smoothed estimates for retrospective mapping. They also inspect residuals around abrupt turns, where the motion model may be inaccurate. A more attractive reconstructed path is not accepted without checking against independent references.

Common misconceptions and mistakes

A smoother is not simply a filter with a smaller noise setting. It uses a different information set, including observations later than the state being estimated.

The smoother gain is not generally the measurement-update Kalman gain. It depends on cross-covariance between consecutive states. Reusing an unrelated gain can produce incorrect corrections.

A single Gaussian representation may fail when several separate trajectories remain plausible. Avoid claiming exact nonlinear Bayesian inference or guaranteed error reduction for every individual case. Using smoothed states as if they were available at the original decision time also introduces future-information leakage in backtests.

Business use

The business value is improved retrospective reconstruction for analysis, diagnostics and mapping when delayed processing is acceptable. Costs include storing forward results, backward computation and maintaining consistent model assumptions. Teams should label the observation horizon used for every estimate, compare reconstruction against independent data, and separate retrospective reporting from real-time decision evaluation. Fixed-lag variants offer a different delay-versus-information tradeoff.

Check your understanding

Why is the gain 0.8 in the example?

Answer: Cross-covariance four is divided by predicted next-state variance five.

Why does process noise not add to that cross-covariance?

Answer: The example assumes it is independent of the current state.

Related terms

Unscented Kalman filter
Unscented transform
Rauch–Tung–Striebel smoother
cross-covariance
fixed-interval smoothing
fixed-lag smoothing and state-space model

Learn next

Next, repeat the correction with different process-noise variances to see how the connection between adjacent states changes. Then study nonlinear sigma-point propagation and compare approximate smoothing against a reference solution on a small problem.

Sources

Primary source: Simo Särkkä, Bayesian Estimation of Time-Varying Systems, Section 4.2.3 and the additive unscented Rauch–Tung–Striebel smoother equations: https://users.aalto.fi/~ssarkka/course_k2009/full_course_booklet.pdf

The original worked example uses a linear special case to check the backward correction. The organizational case is explicitly fictional.

Independent review checked Algorithm 4.4.

Previous: Unscented Kalman filter Next: Unscented transform