Does it predict the future?

Its defining purpose is to estimate earlier states using a longer observation record, although it relies on forward predictions internally.

Is it exact for nonlinear systems?

Generally no. The extended version uses local approximations.

Can it operate with a delay?

Yes. A fixed-lag design can revise a recent window, with an explicit delay and computational tradeoff.

AI Glossary / Probabilistic machine learning / Intermediate

Extended Kalman smoother

Definition

An extended Kalman smoother estimates past hidden states of a nonlinear dynamic system using observations collected both before and after those states. It combines approximate Gaussian state estimates with local linearization. A common version runs an extended Kalman filter forward, then an extended Rauch - Tung - Striebel smoothing pass backward through the recorded sequence.

What it means

A filter answers what the state probably was using information available up to that time. A smoother can revise that answer after later evidence arrives. For example, a later location reading may clarify an earlier uncertain movement. Smoothing is therefore particularly useful when reconstructing completed trajectories or reviewing recorded sensor data.

Extended means that nonlinear transition or measurement relationships are approximated locally, typically using first derivatives collected in Jacobian matrices. It does not mean the method exactly solves any nonlinear estimation problem. The Gaussian approximation can miss ambiguity, multiple plausible trajectories or strong curvature in the system model.

How it works

First run the extended Kalman filter, retaining filtered means and covariances together with the information needed for transition predictions. For additive process noise, the transition function predicts the next mean, while its Jacobian propagates uncertainty and the process-noise covariance adds uncertainty.

The backward pass starts from the final filtered estimate. At each earlier step, a smoothing gain relates uncertainty in that state to uncertainty in the next predicted state. The difference between the next smoothed mean and its forward prediction is carried backward through this gain. The covariance is adjusted consistently with the same local approximation.

The basic fixed-interval method uses the completed observation interval. Fixed-lag methods revise only a recent window; iterated extended smoothers additionally revisit linearization points. These are related choices, not interchangeable names. Large state dimensions, poor Jacobians and ill-conditioned covariance calculations can make implementation difficult.

Three practical examples

1. Mobile robotics: reconstruct a recorded robot route using later sensor readings to refine earlier uncertain positions. The smoothed route is useful for analysis, but was not necessarily available to the robot when it made decisions.

2. Industrial monitoring: estimate a past temperature trajectory from noisy measurements and a nonlinear process model. Check whether unmodeled operating changes could explain residual discrepancies.

3. Motion analysis: combine recorded position and motion measurements to reconstruct a movement sequence. Compare uncertainty with reference measurements rather than judging accuracy solely by how visually smooth the path looks.

Think of it this way

Think of reconstructing a journey after reading the final pages of a travel diary. A later landmark can clarify where an earlier ambiguous note was written. The analogy captures the value of later evidence. The algorithm instead performs probabilistic calculations through a specified dynamic model; it does not understand a story or freely rewrite the past.

Worked example

Consider one scalar backward step, with dimensionless state and nonlinear transition f(x) = x squared. Suppose the filtered mean is 2 and its variance is 0.25. The local derivative at 2 is 4. With process-noise variance 1, the extended-filter prediction is mean 4 and variance 4 squared × 0.25 + 1 = 5.

The smoothing gain is 0.25 × 4 / 5 = 0.2. Suppose the next state's smoothed mean is 4.5 and variance is 2. The earlier smoothed mean becomes 2 + 0.2 × (4.5 − 4) = 2.1. Its smoothed variance becomes 0.25 + 0.2 squared × (2 − 5) = 0.13.

Later evidence has shifted the earlier estimate upward and reduced its approximate uncertainty. These stipulated inputs demonstrate one recursion, not an entire fitted sequence. They also expose the approximation: for a Gaussian variable with mean 2 and variance 0.25, the exact mean of its square is 4.25, whereas this first-order prediction used 4. Local linearization need not reproduce exact nonlinear moments.

Illustrative case study

Fictional case: Birch Robotics reviews a warehouse trial after a localization dropout. Engineers compare the original online position estimates with a trajectory smoothed using later observations. The revised route supports their investigation, but they preserve the online record because that is what the controller actually knew.

They inspect the dropout interval, compare alternative noise assumptions and check against independent markers. A plausible-looking reconstruction is treated as an estimate with uncertainty. This fictional example illustrates responsible use of retrospective evidence, not a claim that smoothing always recovers the true route.

Common misconceptions and mistakes

Smoothing is not simply applying a moving average to sensor readings. It estimates hidden states through dynamics, observations and uncertainty relationships. A smoother can produce abrupt changes when the model and evidence support them.

Smoothed estimates are not automatically valid real-time inputs: using future observations during evaluation can create information leakage. Reduced approximate covariance also does not prove improved actual accuracy. Incorrect dynamics, correlated errors or an inadequate Gaussian approximation can make a confident estimate wrong. Matrix implementations should preserve numerical stability rather than naively invert poorly conditioned matrices.

Business use

Use extended smoothing when completed records contain later evidence that can improve a model-based reconstruction. Define whether the business task needs an online estimate, a delayed estimate or a retrospective analysis before choosing the method.

Evaluate against trustworthy reference data where possible, and retain both the data available at decision time and the reconstructed history. Test sensitivity to noise assumptions, missing intervals and operating changes. If strongly nonlinear or multimodal behavior matters, compare other inference methods instead of assuming more smoothing will repair the underlying approximation.

Check your understanding

Why does the worked example predict the next mean as 4 rather than the exact Gaussian square mean 4.25?

Answer: It applies the nonlinear transition to the filtered mean as part of a first-order approximation. It does not exactly integrate the squared transition over the state distribution.

Related terms

Extended Kalman filter
Jacobian
Fixed-lag smoothing
Unscented Kalman smoother
Rauch–Tung–Striebel smoother
state-space model
process noise
posterior inference

Learn next

Compare filtering and smoothing on a short simulated sequence with a known trajectory. Then inspect how increasing transition curvature changes the discrepancy between local Gaussian approximations and numerical or sampling-based reference inference.

Sources

Särkkä, Bayesian Filtering and Smoothing, Chapter 9, extended RTS smoother: https://users.aalto.fi/~ssarkka/pub/cup_book_online_20131111.pdf

Independent review verified Algorithm 9.1 and its Gaussian linearization assumptions. The scalar nonlinear calculation and fictional operational scenarios are original illustrations, not reported experimental results.

Previous: Extended Kalman filter Next: Extendible hashing