Can a measurement update an unmeasured state component?

Yes, through the prior cross-covariance, as in the position-and-velocity example.

Does larger scalar measurement variance reduce the scalar gain when H=1 and the prior variance is fixed?

Yes: K=P/(P+R) decreases as R increases.

Is the gain always constant?

No. It can change as covariances or models change, though some systems use a justified steady-state gain.

AI Glossary / Mathematics and optimization / Intermediate

Kalman gain matrix

Definition

The Kalman gain matrix determines how a Kalman filter converts a new measurement's prediction error into a correction of its estimated state. It is calculated from the predicted state uncertainty, measurement model, and measurement-noise uncertainty. Rather than being a universal confidence percentage, it maps discrepancies in measurement coordinates into adjustments in state coordinates.

What it means

A state estimator first predicts quantities such as position and velocity. A sensor then supplies a measurement that may observe only part of that state. The difference between the actual measurement and the predicted measurement is called the innovation.
The gain controls how that innovation changes each state component. Correlations matter: a position measurement can change a velocity estimate when their prediction errors are correlated. Consequently, the matrix can contain off-diagonal or differently scaled effects. The familiar description of blending a prediction and measurement is most direct in a scalar model where both measure the same quantity.

How it works

For the standard linear measurement model z=Hx+v, let P-minus be the predicted state-error covariance and R the measurement-noise covariance. With the usual uncorrelated-noise assumptions, the innovation covariance is S=H P-minus H-transpose+R. If S is invertible, the gain is K=P-minus H-transpose S⁻¹.
Update the estimate by x-plus=x-minus+K(z−H x-minus). In the linear Gaussian model this gives the conditional mean update. Practical implementations usually compute the needed linear solve rather than explicitly invert S. Different assumptions, correlated noises, or nonlinear measurement models require appropriate modifications.

Three practical examples

A tracking system can adjust an estimated object's position and velocity after receiving a noisy position measurement.
A sensor-fusion application can combine model predictions with measurements whose uncertainty varies by sensor.
A process-monitoring system can revise an estimated internal state when a sensor observes only one related quantity.

Think of it this way

Think of an editor deciding how a new observation should revise several linked estimates. A reliable observation deserves attention, but it must be translated into the right quantities and units. The analogy is limited because Kalman gain follows a covariance model and linear algebra, not a subjective trust score attached to a person or device.

Worked example

Consider an original two-component state: position in meters and velocity in meters per second. The predicted state is x-minus=(10,2). Let its covariance matrix have rows (4,1) and (1,1), using the corresponding squared and cross-product units. A sensor measures position only, so H=(1,0), and its measurement-noise variance is R=1 square meter.
The innovation covariance is S=4+1=5. Multiplying P-minus by H-transpose selects the column (4,1), so K=(4/5,1/5) as a two-by-one matrix. Suppose the measured position is 12 meters. The innovation is 12−10=2 meters, and the state correction is (1.6,0.4). The updated state is (11.6 meters,2.4 meters per second).
The velocity changes even though the sensor did not measure velocity directly. That happens because the prior covariance connects position and velocity errors. If the off-diagonal covariance were zero instead, the velocity component of the gain would be zero for this measurement.
The updated covariance in exact arithmetic is P-plus=P-minus−K S K-transpose. Here K S K-transpose has rows (3.2,0.8) and (0.8,0.2), leaving P-plus with rows (0.8,0.2) and (0.2,0.8).
Now increase the sensor variance to 16 while keeping the same prior and observed value. S becomes 20, K becomes (0.2,0.05), and the estimate changes to (10.4,2.1). The less precise measurement produces a smaller correction under this model. These numbers assume the covariance specification is meaningful; declaring a sensor precise does not make it precise.

Illustrative case study

In a fictional example, Aspen Tracking notices that its estimated motion jumps after an occasional unreliable position reading. The team discovers that the filter's measurement-noise setting represents ideal laboratory conditions rather than the actual operating environment.
Engineers compare innovations with the assumed uncertainty and evaluate a revised measurement model. They also investigate outlier handling rather than expecting the ordinary gain formula to recognize every faulty reading. This invented case illustrates the importance of uncertainty assumptions; it is not a claim that changing one covariance setting solves all tracking failures.

Common misconceptions and mistakes

The gain matrix is not generally a collection of numbers between zero and one. That intuition comes from the special scalar case H=1 with positive variances. Different measurement scales and coupled states can produce other values and units.
Do not use the posterior covariance where the formula requires the predicted covariance. The update must be based on uncertainty before incorporating the new measurement. Also avoid treating measurement noise and process noise as interchangeable: process noise influences the prediction, while measurement noise enters the observation update.
Under the basic fixed linear model with predetermined covariance inputs, K depends on those covariances and H, not directly on the observed innovation value. A large residual alone does not automatically make the standard filter distrust the sensor.

Business use

Kalman gain is useful to understand when AI products incorporate tracking, estimation, or sensor fusion. A review should ask how uncertainty was estimated, whether state and measurement units align, and whether the model covers the operating conditions.
Evaluate state accuracy and uncertainty calibration together. A filter that appears smooth may simply underreact to genuine changes, while one that follows every measurement may amplify noise. Record the measurement assumptions and update timing so that apparent gain changes can be interpreted. The matrix is a consequence of the estimation model, not a standalone setting that guarantees reliability.

Check your understanding

In the worked example, why is the second gain entry nonzero?

Answer: The predicted position and velocity errors have nonzero cross-covariance, allowing a position innovation to inform the velocity correction.

Related terms

Kalman filter
Covariance matrix
Measurement model
Extended Kalman filter
innovation
state estimation
process noise
sensor fusion

Learn next

Recalculate the example with zero position-velocity covariance. Then change the measurement unit from meters to centimeters and transform H and R consistently. Verify that the resulting physical state correction is unchanged even though the gain's numerical entries change.

Sources

Primary teaching source: Stephen Boyd, EE363, The Kalman Filter, measurement-update and statistical-assumptions sections: https://ee363.stanford.edu/archive/lectures/kf.pdf . The original instructor slides were read. The numerical state, covariance matrices, and fictional tracking case above are original.

Previous: Kalman filter Next: Kaniadakis distribution