Is it always better than the extended Kalman filter?

No. Relative performance depends on the problem and implementation.

Does it estimate future states?

Its prediction step does; its measurement update then incorporates current evidence.

Can it represent two separate hypotheses well?

A single Gaussian UKF generally cannot preserve a strongly multimodal posterior.

AI Glossary / Probabilistic machine learning / Advanced

Unscented Kalman filter

Definition

The unscented Kalman filter, or UKF, is a recursive state estimator that uses carefully selected sigma points to approximate how uncertainty passes through nonlinear dynamics and measurement functions. It maintains a state mean and covariance and updates them as observations arrive. It avoids requiring the Jacobians used by a conventional extended Kalman filter, while remaining an approximate nonlinear estimator.

What it means

Many hidden states, such as position, velocity or equipment condition, are observed indirectly. A nonlinear measurement can transform uncertainty in ways that evaluating the function only at the mean misses. The UKF evaluates the function at several representative points and combines their transformed values.

These sigma points are deterministic numerical representatives, not randomly sampled particles. They are regenerated from the current mean and covariance. The filter usually summarizes its result with a Gaussian approximation, so complicated posterior shapes, including separated modes, are not preserved merely because the transformation is nonlinear. Better moment approximation does not make every nonlinear filtering problem exact.

How it works

Start from an estimated state mean and covariance. Construct sigma points and associated weights according to a specified unscented-transform rule. Propagate the points through the state-transition function, then combine them to approximate the predicted mean and covariance.

For independent additive process noise, include its covariance in the prediction. Apply a consistent sigma-point construction for the measurement calculation, propagate through the measurement function, and estimate the predicted observation, its covariance and the state-observation cross-covariance.

The Kalman gain is the cross-covariance multiplied by the inverse observation-prediction covariance. Correct the predicted state mean using the observed-minus-predicted measurement residual, and update its covariance correspondingly.

Additive and augmented-noise variants differ in how noise enters the sigma-point calculations. Scaling parameters and mean-versus-covariance weights must match the chosen variant. Numerical implementation should maintain covariance symmetry and positive semidefiniteness; special state spaces, such as wrapped angles, also require appropriate means and residuals rather than ordinary subtraction everywhere.

Three practical examples

1. A robot estimates position and velocity from nonlinear range measurements while combining a motion model with sensor uncertainty.

2. An engineering prototype estimates a hidden physical condition from a nonlinear calibrated sensor response.

3. A tracking researcher compares UKF and extended-Kalman estimates on simulated maneuvers where linearization around one point may be inadequate.

Think of it this way

Think of testing how a flexible shape changes by following several strategically placed markers. Following only the center misses how the shape bends; following the markers captures more information about its spread. The UKF then compresses that information back into a mean and covariance, so the complete transformed shape is still not retained.

Worked example

Consider a measurement-update example with predicted state X distributed normally with mean two and variance one. The measurement is Y = X squared plus independent zero-mean noise with variance one. Observe y = 6.

Use the scalar scaled unscented transform with alpha = 1, kappa = 0 and beta = 2. The sigma points are 2, 3 and 1. Their mean weights are 0, one half and one half; their covariance weights are 2, one half and one half.

The transformed points are 4, 9 and 1. The predicted measurement mean is five. Its transformed-state variance is 2 times (4 minus 5) squared + one half times (9 minus 5) squared + one half times (1 minus 5) squared = 18. Adding measurement noise gives innovation variance 19.

The state-measurement cross-covariance is four, giving gain 4/19. The updated state mean is 2 + (4/19)(6 minus 5) = 42/19, approximately 2.2105. The updated variance is 1 minus 16/19 = 3/19, approximately 0.1579.

The transformed moments happen to match the exact Gaussian quadratic moments here. The resulting conditional state update is still a Gaussian approximation to a nonlinear posterior, not the exact posterior distribution. This is one measurement update, not a full filtering trajectory.

Illustrative case study

Fictional case: Aspen Sensors uses a nonlinear sensor curve to estimate a changing machine condition. Engineers compare a UKF with an extended Kalman filter on recorded calibration sequences and synthetic trajectories with known states.

The UKF improves some estimates, but an ambiguous sensor region produces multiple plausible conditions that one Gaussian summary cannot describe well. The team restricts the operating range or evaluates a richer estimator for that region. It judges the filter by accuracy, uncertainty behavior and numerical reliability, rather than assuming that derivative-free computation guarantees success.

Common misconceptions and mistakes

The UKF does not require random Monte Carlo samples, and its sigma-point weights are not always ordinary positive probabilities. Mean and covariance weights can differ.

Avoiding Jacobians does not eliminate the need for a sound transition model, measurement model or noise assumptions. A UKF can diverge when these are poor, and tuning noise solely to make a trajectory look smooth can conceal systematic error.

Claims that it is exact for all nonlinear functions are incorrect. Accuracy depends on the transform, function, input distribution and repeated Gaussian approximation. Process noise should not be omitted simply because sigma points already represent state uncertainty.

Business use

Business applications include tracking, robotics and industrial estimation where nonlinear models must run repeatedly with manageable computation. Evaluate estimation error, innovation behavior, latency and covariance consistency under expected disturbances. Keep the sigma-point convention, noise treatment and initialization documented. Compare against simpler and richer alternatives where model ambiguity or nonlinear effects materially affect decisions.

Check your understanding

Why is measurement variance 19 in the example?

Answer: The transformed state contributes 18 and independent measurement noise contributes one.

Why is the final posterior still approximate?

Answer: Accurate selected moments do not make a nonlinear joint distribution Gaussian.

Related terms

Unscented transform
Kalman filter
Extended Kalman filter
sigma point
innovation
state-space model and unscented Kalman smoother

Learn next

Next, reproduce the same update with a first-order linearization and compare its predicted measurement moments. Then investigate additive versus augmented noise and test covariance consistency on simulated data with known states.

Sources

Primary source: Simo Särkkä, Bayesian Estimation of Time-Varying Systems, sections on the unscented transform and unscented Kalman filter: https://users.aalto.fi/~ssarkka/course_k2009/full_course_booklet.pdf

The quadratic-sensor calculation and fictional case are original teaching material. The calculation states its sigma-point parameters and separate covariance weights explicitly.

Independent review consulted the scaled-transform weights and additive UKF Algorithm 3.12, including separate noise covariances and measurement cross-covariance.

Previous: Unrolling Next: Unscented Kalman smoother