Kalman Filter, Live
Recap · Reading

Lesson 3 recap: the matrix Kalman filter

The matrix Kalman filter on one page: how to describe linked uncertainty, how it moves through a model, and the five equations that track a whole state vector.

Lesson 3 · slide 7Recap
x⁻ = F x + B u    P⁻ = F P Fᵀ + Q    K = P⁻Hᵀ(H P⁻ Hᵀ + R)⁻¹    x = x⁻ + K(z − H x⁻)    P = (I − K H) P⁻
Key ideas
  1. Covariance is the average product of deviations. Its sign is the tilt of the cloud.
  2. The covariance matrix: variances on the diagonal, covariances off it, symmetric. Its picture is the 2σ ellipse.
  3. y = A x + b moves the covariance to A P Aᵀ. The shift b never changes it.
  4. A state-space model is xₖ₊₁ = F xₖ + B uₖ and zₖ = H xₖ. The physics lives in F, the inputs in B, the sensors in H.
  5. The matrix filter is the 1D filter with matrices, and a position sensor also corrects velocity.
  • Correlation and covariance. Ice cream rises with temperature, hot chocolate falls, coffee does not care. cov(x, y) = average of (x − μₓ)(y − μᵧ): positive, negative or about zero.
  • The covariance matrix. P holds σₓ², σᵧ² and σₓᵧ. The correlation ρ = σₓᵧ / (σₓσᵧ) lies between −1 and 1. The 2σ ellipse is wide where the variance is big and tilts with the covariance.
  • Transforming covariance. The scalar rule a²σ² becomes A P Aᵀ. One tick of motion turns P with 4 and 1 on the diagonal into F P Fᵀ = 5111: motion links position and velocity.
  • State space. The state is what you need to predict the future: position and velocity. F = 1Δt01 moves it, friction and a motor push change the velocity row, H = 10 is a GPS.
  • The matrix Kalman filter. Predict with F and F P Fᵀ + Q. Correct with the gain K = P⁻Hᵀ(H P⁻ Hᵀ + R)⁻¹ and the innovation z − H x⁻. The worked cycle took the position variance from 6 to 2.4 and left an off-diagonal of 0.4.
  • Rocket tracking. State (x, y, vₓ, vᵧ), F as two constant-velocity blocks, H keeps the positions. Big q follows the GPS, small q smooths it, too small misses the bend.

Free play in the lab

An optional goal for the lab below: see how close you can get the printed velocity estimate to the true speed (2 m/s) after 40 s. Try a smaller q, then a bigger R, then change the true speed, and watch how the ±2σ band reacts.

Next lesson: the extended Kalman filter

Keep reading with the full course

Finish lesson 3 and you will be able to: track position and velocity in 2D with the matrix Kalman filter.

Still locked here:

  • The rest of this explanation
  • 1 more example with step-by-step solutions
  • Five extra members-only problems with worked solutions

$20 once, lifetime access. Lessons 1 and 2 stay free.

Examples

Example 1

P = 16−4−44. Find σₓ, σᵧ and ρ, and say which way the ellipse tilts.

Show solutionHide solution

σₓ = 4, σᵧ = 2, ρ = −4 / (4 · 2) = −0.5. Wider than tall, tilting down to the right.

Example 2

P = 1001, F = 1101, Q = 0. Find P⁻. Then, with H = 10 and R = 2, find K.

Show solutionHide solution

P⁻ = F Fᵀ = 2111.

S = 2 + 2 = 4, K = (2, 1) / 4 = (0.5, 0.25).

Exercise Free play: change q, R or the true speed and watch the ±2σ band and the velocity estimate react. Edit the code and press Run.
slide_7.py Loading Python…
1
Output · Ctrl+Enter runs
Press Run to execute this slide's code. Ctrl+Enter works too.

Questions and comments

Stuck on something, or found a mistake? Ask here and I'll answer. You sign in with GitHub.

The full course

From noisy sensors to the EKF, by running it

Lessons 1 and 2 are free. Finish all four and you will be able to:

You get: all 4 lessons with Python labs · every worked example and solution · 5 extra problems per lesson · your progress tracked · Steppe, your AI tutor inside every lesson, day or night · every future update.

$20 once · lifetime access · 14-day money-back guarantee
Continue to checkout →

Already a member? Log in to the course