kalman filter
**Kalman filter is a recursive state estimator that combines a dynamic model with noisy measurements using their uncertainty.** Kalman-family filters support navigation, tracking, sensor fusion, SLAM, control, communications, finance, process estimation, and any system where important state is not measured directly. The classical filter is optimal in the minimum-mean-square sense for a linear model with Gaussian noise and correct covariances. It remains useful more broadly, but consistency and optimality claims depend on observability, modeling, initialization, and noise assumptions. An engineering definition states variables, units, assumptions, domains, initial and boundary conditions, sampling or update rate, uncertainty, stability or error objective, and implementation constraints. Mathematical guarantees apply to the stated model; they do not automatically cover unmodeled dynamics, finite precision, sensor faults, saturation, delay, concurrency, or hostile inputs.
**Architecture, representation, and operating mechanism.** State transition and control models predict state and covariance; a measurement model predicts observations; process covariance represents unmodeled evolution and measurement covariance represents sensor uncertainty. The Kalman gain balances prediction and innovation, then updates state and covariance. Prediction propagates the prior through dynamics. Update computes innovation, innovation covariance, gain, corrected state, and corrected covariance. The cycle repeats without storing all history. Information filters use inverse covariance; smoothers refine past states using later observations. State RMSE, bias, covariance consistency, normalized innovation squared, normalized estimation error squared, innovation whiteness, convergence, divergence rate, outlier rejection, latency, numerical stability, memory, update rate, and robustness to dropout matter. Sensors, actuators, sampling clocks, quantizers, communication, memory, processors, power, thermal behavior, software scheduling, safety interlocks, and operators affect the delivered result. End-to-end design allocates error and latency budgets to named components instead of assuming ideal data and unlimited compute. Results report accuracy or error, stability and robustness margins where applicable, convergence, latency, throughput, memory, numerical conditioning, precision, energy, coverage, false alarms, and behavior at operating limits. Reference models, analytic cases, independent implementations, and confidence bounds make numerical or test evidence interpretable.
**Implementation, hardware, and failure modes.** EKF linearizes nonlinear models with Jacobians, UKF propagates sigma points, particle filters represent non-Gaussian distributions with weighted samples, square-root filters improve numerical behavior, and factor graphs optimize many states over a window or batch. Matrix multiply, factorization, solve, and covariance updates run on MCUs, DSPs, CPUs, GPUs, or FPGA linear-algebra blocks. State dimension, sparsity, rate, precision, cache, and sensor asynchronous scheduling determine cost. Wrong process/measurement covariance makes the filter overconfident or noisy; unobservable states drift; outliers dominate Gaussian updates; linearization fails under strong nonlinearity; covariance loses symmetry/positive definiteness; timestamp/frame errors masquerade as model error. Engineering must include data movement, finite precision, resource contention, numerical or physical limits, error propagation, and deterministic behavior when assumptions are violated. Requirements, mathematical model, discretization, algorithm, numerical format, implementation, calibration, verification, deployment, monitoring, update, and incident response form one lifecycle. Versions of coefficients, transforms, test corpora, compiler settings, hardware kernels, tolerances, and assumptions remain linked to measurements.
**Evaluation, verification, and deployment.** Use simulated truth and recorded trajectories, innovation consistency and whiteness, covariance calibration, initial-condition sweeps, outliers/dropouts/bias, delayed and out-of-sequence data, nonlinear maneuvers, numerical stress, and comparison with independent estimators. Sensors require calibration, frames, timing, lever arms, latency, and health status. The filter output feeds control or decisions, so confidence gates, reset, mode switching, map updates, and safe degradation are part of design. Navigation and tracking can expose location and identity. Access, retention, integrity, spoof detection, calibration custody, tuning authorization, and event logs apply to estimator data and parameters. Verification uses analytic identities, invariants, dimensional checks, deterministic unit cases, randomized and property tests, Monte Carlo uncertainty, worst-case boundaries, high-precision references, formal reasoning where tractable, extracted or hardware models, fault injection, and closed-loop or production replay. Independent evidence is essential when one model is used to validate itself. Requirements, mathematical model, discretization, algorithm, numerical format, implementation, calibration, verification, deployment, monitoring, update, and incident response form one lifecycle. Versions of coefficients, transforms, test corpora, compiler settings, hardware kernels, tolerances, and assumptions remain linked to measurements. Results report accuracy or error, stability and robustness margins where applicable, convergence, latency, throughput, memory, numerical conditioning, precision, energy, coverage, false alarms, and behavior at operating limits. Reference models, analytic cases, independent implementations, and confidence bounds make numerical or test evidence interpretable.
| Estimator | Nonlinearity handling | Distribution | Compute | Best fit |
|---|---|---|---|---|
| Linear KF | Exact linear model | Gaussian | Low-medium | Linear tracking/control |
| EKF | Jacobian linearization | Approximate Gaussian | Medium | Navigation and robotics |
| UKF | Sigma-point transform | Approximate Gaussian | Medium-high | Stronger smooth nonlinearities |
| Particle filter | Sample propagation | General/multimodal | High | Ambiguous localization |
| Factor graph/smoother | Nonlinear batch/window optimization | Model dependent | High | SLAM and trajectory refinement |
```svg
```
**Selection and practical application.** Use linear KF for linear-Gaussian systems, EKF for smooth nonlinear models with reliable derivatives, UKF when nonlinear propagation matters at modest state size, particle filters for multimodal/non-Gaussian belief, and graphs for smoothing and SLAM. GPS/INS, target tracking, battery state, motor observers, process monitoring, radar, visual-inertial odometry, finance, sensor fusion, and robotics use recursive estimation. Sensors, actuators, sampling clocks, quantizers, communication, memory, processors, power, thermal behavior, software scheduling, safety interlocks, and operators affect the delivered result. End-to-end design allocates error and latency budgets to named components instead of assuming ideal data and unlimited compute. An engineering definition states variables, units, assumptions, domains, initial and boundary conditions, sampling or update rate, uncertainty, stability or error objective, and implementation constraints. Mathematical guarantees apply to the stated model; they do not automatically cover unmodeled dynamics, finite precision, sensor faults, saturation, delay, concurrency, or hostile inputs. CFS connects this topic to semiconductor architecture, implementation, verification, manufacturing, packaging, test, and deployed AI-system tradeoffs across the platform.