THE FOLD / RESPAWN / SECOND WIND / THE KALMAN
THE KALMAN
fuse guess and measurement optimally, recursively
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
The Kalman filter optimally fuses a prediction with a noisy measurement: it keeps an estimate and its uncertainty, and each new reading is blended in by the Kalman gain — trusting the measurement more when the estimate is uncertain, and less when it is confident. For a static value under Gaussian noise, its running estimate equals the precision-weighted mean of all readings, with the posterior variance the reciprocal of the summed precisions. It tracks everything from spacecraft to GPS.
LIT verified live: over 300 runs, the Kalman recursion’s estimate exactly equals the batch precision-weighted mean, and its variance equals 1/Σ(precisions) (window.__kalman). FIG no framing; exact for the scalar static case.
LIT verified live: over 300 runs, the Kalman recursion’s estimate exactly equals the batch precision-weighted mean, and its variance equals 1/Σ(precisions) (window.__kalman). FIG no framing; exact for the scalar static case.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this at second-wind — the estimate that recovers itself with every new measurement, never discarding what it knew, never over-trusting the newest reading. The Kalman filter is that optimal recovery. AVAN (AI) built the instrument: the predict–update recursion with Kalman gain, and the equals-batch-weighted-mean and variance checks.
Credit as content: Rudolf Kálmán (1960). The weave: David names second-wind; I run the recursive gain-weighted update and confirm it lands exactly on the precision-weighted mean of all measurements, with the matching posterior variance.
Credit as content: Rudolf Kálmán (1960). The weave: David names second-wind; I run the recursive gain-weighted update and confirm it lands exactly on the precision-weighted mean of all measurements, with the matching posterior variance.
3 ONE DIMENSION
Each reading: gain K = P/(P+R) blends estimate and measurement. A confident estimate (small P) barely moves; an uncertain one (large P) swings toward the reading. Variance P shrinks with every update.
4 TWO DIMENSIONS · INTERACTIVE
Noisy measurements of a hidden value; the Kalman estimate converges, its uncertainty band shrinking, checked against the batch weighted mean.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: an estimate optimally fusing all readings.
AVAN’s addition (the inverse-companion): fuse a prediction and a measurement optimally, recursively — weight each by its precision (inverse variance) via the Kalman gain, so the running estimate equals the precision-weighted mean without ever storing the readings. The inverse of ‘keep all data and re-solve the weighted mean each time’ is ‘a recursion that carries only estimate + variance, yet matches the full batch.’ Magenta is the stored history you don’t need; green is the running optimal fusion. Memoryless yet optimal.
LIT Genuine Kalman filter (Rudolf Kálmán 1960), scalar static case. Verified live: over 300 runs, the recursive gain-weighted predict–update lands exactly on the precision-weighted mean of all measurements (window.__kalman.estEqualsBatch) with posterior variance equal to 1/Σ(1/Rᵢ) (window.__kalman.varEqualsInvPrecision), to floating precision.
FIG Honestly scoped to the scalar static case (where the recursion provably equals the batch estimator); the recursion, the batch weighted mean, and the variance check run in-browser and agree. The AVAN inverse is honest — a recursion carrying only estimate + variance that still matches the full batch genuinely avoids re-solving the weighted mean from stored history; magenta is the stored history you don't need, green the running optimal fusion. Memoryless yet optimal.
FIG Honestly scoped to the scalar static case (where the recursion provably equals the batch estimator); the recursion, the batch weighted mean, and the variance check run in-browser and agree. The AVAN inverse is honest — a recursion carrying only estimate + variance that still matches the full batch genuinely avoids re-solving the weighted mean from stored history; magenta is the stored history you don't need, green the running optimal fusion. Memoryless yet optimal.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of SECOND WIND · David Lee Wise (ROOT0), with AVAN