What is the Kalman gain, and what does it actually represent mathematically?

The Kalman gain (K) is the weighting factor the filter computes at every update step to decide how much to trust the new sensor measurement versus how much to trust the model's own prediction, and it is derived, not guessed or hand-tuned, directly from the ratio of the prediction's uncertainty (its estimated variance) to the total uncertainty of the prediction and measurement combined: K = P_predicted / (P_predicted + R), where P_predicted is the predicted state's variance and R is the measurement noise variance. When the prediction is relatively uncertain (large P_predicted) compared to the measurement noise, K approaches 1 and the filter leans heavily on the new measurement; when the measurement is relatively noisy (large R) compared to the prediction's uncertainty, K approaches 0 and the filter leans heavily on its own prediction instead. This is exactly what makes the Kalman filter statistically optimal for a linear system with Gaussian noise, rather than an arbitrary or hand-tuned blend: at every single step it computes the mathematically best-possible weighting given the actual, current uncertainty of both sources, not a fixed ratio decided in advance.

What is the difference between process noise and measurement noise in a Kalman filter?

Process noise (denoted Q) represents uncertainty in the motion/prediction model itself, how much the actual system state can deviate from what the model predicts due to unmodeled effects like wind gusts on a drone, wheel slip on a mobile robot, or simply the model being a simplification of more complex real physics. Measurement noise (denoted R) represents uncertainty in the sensor reading itself, the sensor's own inherent noise, quantization error, or environmental interference, independent of how good or bad the motion model is. These play opposite roles in tuning filter behavior: increasing Q (telling the filter the model is less trustworthy) makes the filter lean more on incoming measurements and respond faster to real changes, at the cost of more measurement noise passing through into the estimate, while increasing R (telling the filter the sensor is less trustworthy) makes the filter lean more on its own model prediction and produce a smoother estimate, at the cost of responding more sluggishly to real, sudden changes in the actual state. Tuning a Kalman filter in practice is largely about setting realistic Q and R values (from sensor datasheets and known model limitations) rather than the state estimation math itself, which is fixed once the state-space model is defined.

Why can't you just average a sensor reading and a model prediction instead of using a full Kalman filter?

A simple average implicitly assumes both sources are equally trustworthy at all times, which is rarely true and, more importantly, doesn't stay true: a motion model's prediction uncertainty naturally grows the longer it goes without a correcting measurement (since small errors compound each prediction step), while a given sensor's measurement noise is typically fixed or only slowly varying. A Kalman filter's gain adapts to this automatically and continuously, weighting the prediction more heavily right after a correcting measurement (when the prediction is still fresh and relatively certain) and weighting the next measurement more heavily as more time passes since the last update (as the prediction's uncertainty grows). It also correctly propagates uncertainty through the actual system dynamics rather than treating the prediction as a static, equally-weighted number, and extends naturally to fusing more than two sources or estimating multiple correlated state variables (e.g., position and velocity together) in a single consistent framework, something a simple averaging or fixed-weight approach has no principled way to do.

When is a complementary filter genuinely preferable to a Kalman filter in a real mechatronic design?

A complementary filter is preferable whenever the Kalman filter's added statistical rigor doesn't translate into a measurable improvement worth its implementation and computational cost, which in practice means: simple two-sensor fusion problems with roughly stable, well-characterized noise properties (e.g., fusing a gyroscope and accelerometer for basic tilt estimation on a self-balancing robot or a small drone), tight computational or memory budgets (an 8-bit microcontroller with no floating-point hardware, where the matrix operations a full Kalman filter needs become a genuine performance burden), and time-constrained projects where a complementary filter's few lines of code can be implemented and correctly verified in minutes, while a full Kalman filter's covariance matrix tuning can consume real engineering time to get right. The practical decision rule: start with a complementary filter, since it's simple to implement correctly and good enough for a large share of practical fusion needs, and move to a Kalman filter specifically when a measured, concrete shortcoming of the complementary filter's fixed-weighting approach, not a general assumption that more sophisticated must be better, justifies the added complexity.

Engineering·14 min read·July 12, 2026

📐 Kalman Filtering & Sensor Fusion Mathematics

A technical deep dive into the mathematics of Kalman filtering for mechatronic sensor fusion: state estimation, the predict/update cycle, how a Kalman filter optimally combines a noisy sensor reading with a motion model, complementary filters as a simpler alternative, and real IMU and encoder+accelerometer fusion use cases.

📖Mechatronics Engineering Handbook (Full Access)

View →

From System-Level Sensor Fusion to the Underlying Mathematics

This studio's article on mechatronic system design introduces sensor fusion at a system level, why an accelerometer and gyroscope are combined rather than used alone, and briefly contrasts complementary and Kalman filtering as two fusion approaches. This article goes deeper into the actual mathematics behind Kalman filtering specifically: the state estimation framework it operates within, the predict/update cycle that runs at every time step, how the Kalman gain is derived and what it represents, and a closer technical look at the complementary filter as a simpler alternative, aimed at an engineer who needs to actually implement, not just conceptually understand, a fusion algorithm.

Topics covered

Kalman filter mathematicssensor fusion mathstate estimation Kalmanpredict update cycle Kalman filterIMU sensor fusionKalman gain explainedprocess noise measurement noisecomplementary filter vs Kalman filterextended Kalman filterencoder accelerometer fusionKalman filter covariancestate space model KalmanKalman filter tutorialoptimal state estimator