State estimation: the system acts on an estimate, never on the world

How it works

The classical treatment makes the trade-off explicit. Welch and Bishop describe the Kalman filter as a cycle: a time update projects the state forward through a model of the dynamics, producing an a priori estimate; a measurement update then corrects it into an a posteriori estimate. The weighting between the two is a gain computed from the process-noise covariance Q, which says how much the model is trusted, and the measurement-noise covariance R, which says how much the sensor is trusted. Their limiting cases are the clearest statement of what the parameters mean: as R approaches zero the gain weights the residual more heavily, tending to the inverse of the measurement matrix, and as the a priori error covariance approaches zero the gain weights it less heavily, tending to zero. Put plainly, the filter trusts the measurement more as sensor noise falls, and trusts its own prediction more as it becomes confident. That confidence is computed from the assumed noise, not from whether the model is right. A filter given an over-optimistic Q becomes confident, stops listening to its sensors and drifts — and it does so quietly, because its reported covariance is small exactly when it should not be.

A concrete case

A tracking system reports a tight uncertainty and a smooth trajectory while the object it is following has already left the modelled path. Nothing in the filter is broken. It was told the process noise was small, so it believed its own prediction, and a small reported covariance is the expected output of that assumption rather than evidence that the estimate is good.

Questions worth asking

• Ask where Q and R came from — measured from the hardware, inherited from another project, or tuned until the output looked smooth.
• Ask whether the reported covariance has ever been compared against actual error on held-out data.
• Ask what the estimator does when a sensor drops out, and for how long its prediction is allowed to stand alone.
• Treat a smooth trajectory as evidence about the filter, not evidence about the world.

Sources

• Welch and Bishop — An Introduction to the Kalman Filter (UNC-Chapel Hill TR 95-041) ↗

  The filter alternates a time update that projects the state forward to produce an a priori estimate with a measurement update that corrects it into an a posteriori estimate, weighting the two by a gain computed from the process-noise covariance Q and the measurement-noise covariance R. The document states that as R approaches zero the gain weights the residual more heavily, with the limit of the gain equal to the inverse of the measurement matrix, and that as the a priori error covariance approaches zero the gain weights the residual less heavily, with the limit equal to zero — equivalently, the measurement is trusted more as R falls and less as the prediction becomes confident.

  Boundary: A tutorial on a linear estimator under assumed Gaussian noise with known Q and R. It does not tell you the real noise of any sensor, does not cover the nonlinear or non-Gaussian case beyond an introduction, and offers no guarantee about a robot that uses it.

  Locator, anchor and reuse basis

  Read at: §1 The Discrete Kalman Filter — “The Computational Origins of the Filter” and the discussion of equation (1.8); the time-update / measurement-update cycle. Inspected 2026-09-20. Paraphrase and link to the publicly posted technical report; equations and text not reproduced.

  Verify by searching the source for: the gain K weights the residual more heavily. If that phrase is not there, or does not carry the meaning stated above, this citation is wrong and we want to know.

Continue

• The loop is the system: perception, state, action, consequence
• Uncertainty estimates and the inputs a model has never seen
• Planning and feedback control do different jobs

Elsewhere on this site

• Robotics: the calibration records an estimate depends on
• Mathematics: the formal layer under estimation