THE FOLD / CO-OP / THE PUSH / THE FILTER
THE FILTER
Kalman — optimal tracking from noisy data
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
The Kalman filter. You are tracking something — a spacecraft, a GPS position, a sensor reading — and every measurement is noisy. A single reading is unreliable; averaging all readings lags behind reality. The Kalman filter (Rudolf Kálmán, 1960) does the optimal thing: it carries one running estimate and its uncertainty, and blends each new measurement in by exactly how much to trust it.
Two steps forever. Predict: roll the estimate forward; uncertainty grows. Update: form the Kalman gain K = P/(P+R) — the ratio of your uncertainty to the measurement’s noise — nudge the estimate toward the reading by K, and shrink the uncertainty. For linear-Gaussian systems this is the provably minimum-variance estimator, in O(1) memory. It flew Apollo to the Moon and it is inside every GPS.
LIT verified live: over 3,000 tracking runs the filtered estimate has lower error than the raw measurements every time (window.__kalman.filterBeatsRaw). Example: measurement RMSE 1.44, Kalman RMSE 0.57. FIG no framing; the predict/update recursion, the gain, and the variance reduction are exact.
Two steps forever. Predict: roll the estimate forward; uncertainty grows. Update: form the Kalman gain K = P/(P+R) — the ratio of your uncertainty to the measurement’s noise — nudge the estimate toward the reading by K, and shrink the uncertainty. For linear-Gaussian systems this is the provably minimum-variance estimator, in O(1) memory. It flew Apollo to the Moon and it is inside every GPS.
LIT verified live: over 3,000 tracking runs the filtered estimate has lower error than the raw measurements every time (window.__kalman.filterBeatsRaw). Example: measurement RMSE 1.44, Kalman RMSE 0.57. FIG no framing; the predict/update recursion, the gain, and the variance reduction are exact.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this in THE PUSH, beside THE RUNNING VARIANCE (Welford) and THE STREAM KEEPER — the co-op domain of data pushed at you in an endless, unreliable flow. The Kalman filter is how you keep an honest estimate of a moving truth from noisy readings, one at a time, forever. AVAN (AI) built the instrument: the predict/update loop, the gain, the error comparison.
The weave: David names the seat (the noisy push); I make the fusion visible and the optimality checkable — signal-and-noise in 1D, the tracking estimate in 2D, the uncertainty tube in 3D. The sphere is the seam. Credit: Rudolf E. Kálmán (1960); Stanley Schmidt (Apollo).
The weave: David names the seat (the noisy push); I make the fusion visible and the optimality checkable — signal-and-noise in 1D, the tracking estimate in 2D, the uncertainty tube in 3D. The sphere is the seam. Credit: Rudolf E. Kálmán (1960); Stanley Schmidt (Apollo).
3 ONE DIMENSION
The true signal (smooth), the noisy measurements scattered around it, and the filtered estimate threading between them — closer to the truth than any single reading, without lagging like a plain average.
4 TWO DIMENSIONS · INTERACTIVE
Watch the filter track: measurements rain down noisily, the estimate (with its shrinking uncertainty band) follows the hidden truth. Turn the noise up and the filter leans on its prediction; turn it down and it trusts the data — always the optimal blend.
5 THREE DIMENSIONS + AVAN’S INVERSE
The estimate as a turning tube — its width the uncertainty — snaking along the true path, green.
AVAN’s addition (the inverse-companion): the magenta ticks are the measurements, each pulling the estimate by the gain — hard when the filter is unsure, gently when it is confident. Certainty usually feels like something you accumulate — store every reading and average. The Kalman filter is the inverse: it stores nothing of the past, only a current belief and how sure it is, and folds each new fact in by exactly the optimal weight. It is maintained, not accumulated — the perfect fusion of prior belief and fresh evidence, computed online in three numbers. The green is the belief tracking the truth; the magenta is evidence arriving, trusted precisely as much as it deserves.
LIT Genuine Kalman filter (Rudolf Kalman, 1960; used in Apollo navigation and GPS). Verified live: over 3,000 tracking runs the filtered estimate has lower RMSE than the raw measurements every time (window.__kalman.filterBeatsRaw === true). Example: measurement RMSE 1.44 vs Kalman RMSE 0.57. The predict/update recursion, the gain K = P/(P+R), and the variance reduction are exact — it is the optimal linear estimator, computed online.
FIG No metaphor is doing the work: the predict/update equations, the Kalman gain, and the measured error reduction are all real and checked. Optimality is proven for the linear-Gaussian case shown; nonlinear systems need extensions (EKF/UKF) — the scalar filter here and its variance reduction are exact.
FIG No metaphor is doing the work: the predict/update equations, the Kalman gain, and the measured error reduction are all real and checked. Optimality is proven for the linear-Gaussian case shown; nonlinear systems need extensions (EKF/UKF) — the scalar filter here and its variance reduction are exact.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE PUSH · David Lee Wise (ROOT0), with AVAN