Kalman Filter with Measurement Error
Also known as: Kalman filter measurement noise, state-space model with observation error, KF with measurement uncertainty, linear quadratic estimation with observation noise
The Kalman filter with measurement error is a recursive Bayesian state-space algorithm that estimates the true hidden state of a dynamic system from noisy observations. It explicitly separates process noise (system dynamics uncertainty) from measurement noise (observation uncertainty), propagating both sources of error through a two-step predict-update cycle to yield optimal filtered state estimates and their associated uncertainty.
Key highlights
- Optimal linear estimator under Gaussian noise: minimises the mean squared error of state estimates.
- Explicitly and separately models process noise and measurement error, making uncertainty decomposition transparent.
- Computationally efficient: O(n) recursive updates rather than reprocessing the entire time series.
- Provides a full posterior mean and covariance for the state at every time step, supporting uncertainty quantification.
- The Kalman gain automatically adjusts how much to trust the data versus the model at each step.
- Extends naturally to smoothing (retrospective re-estimation) via the Kalman smoother.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use the Kalman filter with explicit measurement error when your data are time-ordered noisy observations of a process whose true state you cannot observe directly, and when both the system dynamics and the observation noise are approximately linear and Gaussian. It is ideal for tracking, navigation, signal processing, econometrics (e.g., estimating an unobserved trend from noisy GDP data), and engineering control systems. Do not use it when the process or observation model is strongly nonlinear (prefer extended or unscented Kalman filters, or particle filters), when the noise distributions are heavy-tailed or multimodal, or when the measurement error structure is poorly known and R cannot be meaningfully specified.
Strengths & limitations
- Optimal linear estimator under Gaussian noise: minimises the mean squared error of state estimates.
- Explicitly and separately models process noise and measurement error, making uncertainty decomposition transparent.
- Computationally efficient: O(n) recursive updates rather than reprocessing the entire time series.
- Provides a full posterior mean and covariance for the state at every time step, supporting uncertainty quantification.
- The Kalman gain automatically adjusts how much to trust the data versus the model at each step.
- Extends naturally to smoothing (retrospective re-estimation) via the Kalman smoother.
- Assumes linear state transition and observation equations; violated in most real physical systems.
- Requires correct specification of R (measurement noise) and Q (process noise); misspecification degrades performance significantly.
- Optimality guarantees hold only under Gaussian noise; heavy-tailed or skewed noise violates the assumptions.
- The filter diverges if the covariance matrices are poorly conditioned or if the model is not identifiable.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is the measurement error covariance R and how do I choose it?
R is a matrix (diagonal when measurement errors are independent) encoding the variance of the observation noise for each measured variable. In practice, R can be derived from instrument calibration data, estimated jointly with Q by maximising the log-likelihood of the observations (using the prediction-error decomposition), or specified from domain knowledge. Sensitivity analysis over plausible R values is strongly recommended.
What is the difference between process noise Q and measurement error R?
Q captures uncertainty in the system dynamics — how unpredictably the true state changes between steps. R captures uncertainty in the sensor or instrument — how much noise contaminates each reading. A large Q means the state moves erratically; a large R means the measurements are imprecise. Both affect the Kalman gain and therefore how much weight the filter assigns to new observations.
When should I use a particle filter instead?
Use a particle filter when the state transition or observation equation is nonlinear and non-Gaussian, and a linearisation (as in the extended Kalman filter) is inadequate. Particle filters are more flexible but computationally much heavier. For mildly nonlinear systems the unscented Kalman filter (UKF) is often a practical middle ground.
Can I estimate R from the data rather than fixing it?
Yes. The prediction-error decomposition of the Kalman filter yields the log-likelihood of the observations as a function of Q and R. Maximising this likelihood (or using Bayesian priors in a full Bayesian state-space framework) gives data-driven estimates of both covariance matrices, which is preferable to ad-hoc specification when calibration information is unavailable.
What is the Kalman smoother and when do I need it?
The Kalman filter is a forward-pass algorithm that estimates the state using only past and current observations. The smoother (Rauch-Tung-Striebel) runs a backward pass and produces retrospective estimates using all observations, yielding lower uncertainty. Use the smoother when you need the best possible state estimates for historical analysis rather than real-time tracking.
Sources
- 1.Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1), 35–45.
- 2.Durbin, J. & Koopman, S. J. (2012). Time Series Analysis by State Space Methods (2nd ed.). Oxford University Press.ISBN 978-0199641178
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Kalman Filter with Measurement Error. ScholarGate. https://scholargate.app/bayesian/kalman-filter-with-measurement-error