How many points does it use?

A common rule uses 2n plus one, but other sigma-point constructions exist.

Does it require a Gaussian input?

Its usual filtering use assumes a Gaussian approximation; moment matching alone can be applied more broadly, with accuracy depending on distributional information.

Does it recover the full output distribution?

No; it principally approximates moments.

AI Glossary / Probabilistic machine learning / Advanced

Unscented transform

Definition

The unscented transform is a deterministic method for approximating the mean and covariance of a random quantity after a nonlinear transformation. It selects weighted sigma points that represent the input mean and covariance, passes those points through the nonlinear function, and combines the transformed values. It is a building block of the unscented Kalman filter and related estimators.

What it means

For a nonlinear function, applying the function to an average generally differs from averaging the transformed values. Uncertainty therefore matters even when the only output of interest is a mean. The unscented transform represents some of that uncertainty with a small, structured set of points.

It is a numerical moment approximation, not a claim that those points are the only possible input values. It is also not random Monte Carlo sampling. The resulting mean and covariance summarize selected properties of the transformed distribution; they do not determine its full shape, especially when nonlinear transformations create skewness, bounded support or multiple modes.

How it works

For an n-dimensional input with mean m and covariance P, a common scaled rule uses 2n plus one sigma points. One is at m; the others lie in positive and negative directions given by columns of a matrix square root of P, scaled according to the chosen parameters.

In the scaled convention, lambda = alpha squared times (n plus kappa) minus n. The point spread uses the square root of n plus lambda. The central mean weight is lambda divided by (n plus lambda); the other mean weights are one divided by [2 times (n plus lambda)]. The central covariance weight additionally includes one minus alpha squared plus beta.

Evaluate the nonlinear function at every sigma point. Compute the weighted transformed mean, then weighted outer products of deviations for covariance. Input-output cross-covariance can be calculated similarly.

Weights and point construction must be used consistently. Some valid parameter choices involve negative weights, so treating every weight as a probability is misleading. Accuracy depends on the chosen rule, the function and the input distribution. Maintaining a valid numerical covariance and handling constrained or angular variables requires further care.

Three practical examples

1. A tracking estimator approximates the mean and spread of a nonlinear range measurement from an uncertain position.

2. An engineering model propagates uncertain calibration parameters through a nonlinear response curve before comparing predictions with observations.

3. A teaching exercise compares transforming only the mean with transforming sigma points, demonstrating why curvature changes expected outputs.

Think of it this way

Think of testing a bridge's response at several deliberately chosen load configurations instead of testing only the average load. The configurations represent uncertainty and reveal curvature, but cannot describe every possible extreme.

Worked example

Let X be standard normal, with mean zero and variance one. Use a scalar scaled unscented transform with alpha = 1, kappa = 0 and beta = 2. Then lambda is zero, and the sigma points are 0, 1 and minus 1.

Their mean weights are zero, one half and one half. Their covariance weights are two, one half and one half. Notice that the covariance weights do not have to sum to one like probability weights.

For Y = X squared, transformed points are 0, 1 and 1. The approximate mean is one. The approximate variance is 2 times (0 minus 1) squared plus one half times (1 minus 1) squared plus one half times (1 minus 1) squared = 2.

These match the exact moments: E[X squared] is one, E[X to the fourth] is three, and variance of X squared is three minus one squared = two. Transforming only the mean would instead give zero for the predicted mean.

Now keep the same sigma-point rule but change the function to Y = X to the fourth. The transformed points are again 0, 1 and 1, giving approximate mean one, while the exact mean is three. The rule succeeds for the first calculation but misses the second. This illustrates the limits of this specific rule.

Illustrative case study

Fictional case: Rowan Measurement estimates a nonlinear sensor's expected output from an uncertain input condition. A mean-only calculation disagrees with repeated calibration results, so analysts compare it with an unscented-transform calculation.

The sigma points improve agreement in a moderate operating range. Near a sharply curved region, the team compares against higher-resolution numerical integration and finds remaining error. It documents the region where the approximation is adequate instead of treating the transform as an exact substitute for uncertainty propagation. The final choice depends on both accuracy requirements and computation cost.

Common misconceptions and mistakes

Sigma points are not random observations, and their weights need not be positive probabilities. Mean and covariance weights can differ, so substituting one set for the other can change the result.

Avoiding derivatives does not mean avoiding approximation. A finite collection of function evaluations cannot reproduce arbitrary transformed distributions. Matching the input mean and covariance also does not preserve all input information when the true distribution is non-Gaussian.

The transform is not itself a complete filter. A filtering algorithm additionally needs time dynamics, observations, noise handling and an update procedure. Likewise, a Gaussian summary with correct moments can still assign probability to outputs that are physically impossible.

Business use

Business value comes from relatively economical uncertainty propagation in nonlinear estimation and engineering analysis. Teams should benchmark it against simulation, numerical integration or measured references on representative inputs. Record the parameter choices, matrix-square-root convention and domain constraints. The method is most useful when its approximation error is acceptable for the decision and a small deterministic evaluation budget is advantageous.

Check your understanding

Why is the central point relevant when its mean weight is zero?

Answer: Its covariance weight can still be nonzero, as the example shows.

What demonstrates the limitation?

Answer: The fourth-power transformation has approximate mean one but exact mean three under the stated rule.

Related terms

Unscented Kalman filter
sigma point
Gaussian quadrature
moment matching
covariance propagation
extended Kalman filter and Monte Carlo integration

Learn next

Next, compare this rule with numerical integration for several nonlinear functions. Then study how scaling parameters affect point placement and why appropriate residual definitions matter for angles and constrained states.

Sources

Primary source: Simo Särkkä, Bayesian Estimation of Time-Varying Systems, Section 3.2.5 on the unscented transform and its scaled weights: https://users.aalto.fi/~ssarkka/course_k2009/full_course_booklet.pdf

The paired square and fourth-power calculations and fictional case are original teaching material. The parameter convention is stated explicitly.

Independent review checked Equations 3.87–3.88.

Previous: Unscented Kalman smoother Next: Unsharp mask