Kalman Filter Interview Questions: EKF, UKF & Q/R Tuning Interview Prep
Kalman filter interview prep — predict-update cycle, EKF vs UKF vs particle filters, Q/R tuning, square-root forms, and sensor-fusion applications.
Quick answer
The Kalman filter is a recursive Bayesian estimator that fuses noisy measurements with a dynamic model to produce optimal estimates of an unobserved state.
Kalman filter questions test three competencies at once: stochastic-systems literacy (do you understand state-space models, covariance propagation, and Bayesian recursion?), engineering judgment (can you tune Q and R, detect divergence, and know when to escalate from KF to EKF to UKF to particle filter?), and implementation discipline (square-root forms, joseph-form covariance updates, fixed-point arithmetic, real-time scheduling).
Editorial review
Written by
CompoundLearn editorial team
Wireless / RF / hardware engineering
Reviewed by
CompoundLearn editorial team
Wireless / RF / hardware engineering
Last reviewed
Built from curated topic maps, editorial validation, and subject-matter review so the page stays aligned with the interview intent and the current content pipeline.
What it is
The Kalman filter is a recursive Bayesian estimator that fuses noisy measurements with a dynamic model to produce optimal estimates of an unobserved state. It alternates two steps: predict, where the previous state estimate is propagated forward through the state-transition model and the covariance grows by the process noise; and update, where a new measurement is compared to the prediction, the residual (innovation) is weighted by the Kalman gain, and both the state estimate and the covariance are corrected. Under linear dynamics, linear measurements, and zero-mean Gaussian process and measurement noise, the Kalman filter is the minimum-mean-square-error estimator — provably optimal in the class of linear estimators, and equal to the optimal Bayesian estimator within that class. Real systems are rarely linear, so the practical toolbox includes the Extended Kalman Filter (Jacobian linearization), the Unscented Kalman Filter (sigma-point propagation through the non-linearity), the Iterated EKF (relinearizes around the updated estimate), and the Particle Filter (sample-based, for multi-modal posteriors). Applications span aerospace navigation (GPS/IMU fusion), automotive ADAS (radar/camera/lidar tracking), robotics (SLAM, state estimation), economics (regime-switching models), control systems, and recently sensor-rich consumer electronics. The filter is also the canonical example of recursive Bayesian inference and shows up in any role at the intersection of signal processing, control, and machine learning.
Why interviewers ask
Kalman filter questions test three competencies at once: stochastic-systems literacy (do you understand state-space models, covariance propagation, and Bayesian recursion?), engineering judgment (can you tune Q and R, detect divergence, and know when to escalate from KF to EKF to UKF to particle filter?), and implementation discipline (square-root forms, joseph-form covariance updates, fixed-point arithmetic, real-time scheduling). The filter sits at the intersection of EE, control, robotics, autonomous systems, and quantitative finance, so it is asked in interviews for radar/sonar, ADAS, GNSS/INS, robotics SLAM, sensor-fusion, and any role that touches state estimation. Strong candidates can write the predict and update equations from memory, sketch the geometric intuition (the gain is the projection of model uncertainty onto the measurement subspace, scaled by total-uncertainty inverse), and quote tuning recipes (Q from process model uncertainty, R from sensor characterization, NIS-based consistency tests for online tuning). Interviewers also probe failure modes — what happens when Q is too small (the filter over-trusts its model, stops listening to measurements, and diverges), R is too small (the filter over-trusts noisy measurements, so the estimate chases measurement noise), the model is wrong (bias that grows with time), or sensors fault (innovation gating, multi-hypothesis tracking). The Kalman filter is one of those rare topics where theoretical understanding and production engineering are equally important, and the interview expects both.
Common mistakes
The most common mistake is treating the Kalman filter as a black-box smoother — running it without thinking about whether the underlying linear-Gaussian assumptions hold. Real systems have non-linearities and non-Gaussian noise; a textbook KF applied directly will be biased or divergent, and the candidate who does not flag this loses credibility. A second frequent error is confusing the EKF and UKF: candidates know "EKF linearizes" but cannot articulate that it linearizes via Jacobians around the current estimate, which means EKF accuracy degrades when the function has high curvature near the operating point — exactly where UKF wins. A third mistake is sloppy covariance management: forgetting Joseph form for the covariance update (which preserves symmetry and positive-semidefiniteness under round-off), failing to symmetrize after each step, or assuming Q and R are scalar when they should be full matrices for coupled state and measurement models. Fourth, candidates often cannot distinguish observability (is your state recoverable from the measurements?) from controllability — observability failure causes the filter to put unbounded uncertainty on unobservable subspaces, and the filter will appear to converge while the unobservable error grows silently. Fifth, on tuning, candidates default to "tweak Q until it works" instead of using the innovation sequence (NIS or normalized estimation error squared) as a principled diagnostic. Finally, candidates often miss the connection between the Kalman filter and modern probabilistic models — it is the same recursive Bayesian update that drives factor graphs, Gaussian belief propagation, and the prediction step in many sequential models — and that connection is exactly what senior interviewers probe to distinguish someone who memorized the equations from someone who actually understands estimation.
Frequently asked questions
- What assumptions does the standard Kalman filter make, and what happens when they are violated?
- The standard linear Kalman filter assumes (1) linear state-transition and observation models, (2) Gaussian, zero-mean process and measurement noise, (3) noise sequences that are white and mutually uncorrelated, and (4) correctly-specified covariance matrices Q and R. Under these assumptions it is the minimum-mean-square-error linear estimator. Violations show up as biased estimates (model mismatch), filter divergence (Q too small for the actual process variability), or over-confident covariance (R too small). Practical filters spend much more engineering effort on tuning Q and R and on detecting innovation outliers than on the propagation math itself.
- When do you choose an Extended Kalman Filter (EKF) vs an Unscented Kalman Filter (UKF) vs a Particle Filter?
- EKF linearizes the non-linear models around the current estimate via Jacobians and propagates Gaussian moments — cheap, well-understood, and the default for mildly non-linear systems (vehicle dynamics, GNSS/INS fusion). UKF uses the unscented transform: it propagates a deterministic set of sigma points through the non-linear function and reconstructs mean and covariance — typically more accurate than EKF when curvature is significant, with comparable cost. Particle filters represent the posterior with weighted samples and handle multi-modal, highly non-linear, non-Gaussian problems (robot localization with kidnapping, target tracking with clutter) at much higher computational cost.
- What is the role of the Kalman gain, and why does it converge to a steady-state value?
- The Kalman gain K balances trust between the model prediction and the new measurement: K = P_pred * H^T * (H * P_pred * H^T + R)^-1. When measurement noise R is small relative to predicted uncertainty, K is large — the filter trusts the measurement and pulls the estimate toward it. When R is large, K is small and the filter trusts its prediction. For time-invariant linear systems with stationary noise, the covariance update equation has a fixed point (the solution of the discrete algebraic Riccati equation), and the filter converges to a steady-state gain that can be precomputed offline — this is why steady-state Kalman filters are common in embedded implementations.
- How do you tune the process noise covariance Q and measurement noise covariance R?
- R is usually estimated from sensor characterization: take stationary measurements, compute the empirical covariance, and use it directly. Q is harder because it captures unmodeled dynamics — the parts of reality your state-space model misses. A common pragmatic approach is to start with Q proportional to the squared expected acceleration noise (for kinematic models), then check the innovation sequence: it should be zero-mean, white, and have covariance matching H * P_pred * H^T + R (the normalized innovation squared, or NIS, follows chi-squared in steady state). If NIS is consistently high, Q or R is too small; if low, too large. Adaptive filters (IMM, residual-based scaling) automate this when conditions change.
- How does the Kalman filter relate to least-squares and Bayesian estimation?
- The Kalman filter is exactly the recursive Bayesian solution under the linear-Gaussian assumptions: posterior at time k equals likelihood times prior, both Gaussian, so the result stays Gaussian and only the mean and covariance need to be propagated. Equivalently, it is the recursive form of weighted least squares where the weighting is the inverse covariance and the prior is the previous posterior. This dual interpretation matters at interview time: it explains why the filter is optimal under its assumptions, why information-form variants (Information Filter) are more numerically stable in some conditions, and why MAP estimation in linear-Gaussian models reduces to the Kalman filter.
- Why is the square-root or UD-factored Kalman filter used in production?
- The standard covariance-form Kalman filter can lose positive-semidefiniteness on the covariance matrix P due to floating-point round-off, especially in long-running embedded systems with poorly-conditioned dynamics. Square-root forms propagate the Cholesky factor S of P (so P = S * S^T) and never compute P explicitly — every update is structured to keep S triangular. UD factorization (P = U * D * U^T) is similar and avoids square roots entirely, which was important on early flight computers without hardware sqrt. The numerical robustness gain is significant: a square-root filter can run for months without re-initialization, whereas a naive covariance filter can diverge in hours.
Related topics
Essential AI-Native Skills for Kalman Filter Interview Questions: EKF, UKF & Q/R Tuning
Modern engineering work increasingly uses AI tools for design and code review, debugging, documentation, test and testbench generation, and workflow automation. The goal is not to let AI replace engineering judgment — it is to move faster while keeping verification discipline.
- Use AI to explain unfamiliar code, logs, waveforms, datasheets, or test failures.
- Break large problems into small, reviewable steps you can verify independently.
- Ask AI for hypotheses, then validate them against tests, measurements, simulations, or lab data.
- Version-control your analysis scripts, testbenches, and configs — keep changes small and reviewable.
- Document your assumptions, design tradeoffs, and debugging decisions.
- Verify AI output before trusting it: run the checks that fit the domain — unit tests, linters, simulations, or bench/lab measurements.
- Review AI output for correctness, edge cases, and real-world consequences.
Kalman Filter Interview Questions: EKF, UKF & Q/R Tuning — coming to the question bank
The adaptive practice engine is already live for core wireless, RF, and ML systems. Kalman Filter Interview Questions: EKF, UKF & Q/R Tuning isn't covered in the question bank yet — get notified when it's added.