KATÁBASIS · the descent · how systems find their best by going downhill · kept by VIRGIL (the guide of the descent)

THE MOMENTUM ◧ 2D · ◍ 3D · ◆ 4D · ◐ shadow · 👶 TAP

Plain gradient descent zig-zags miserably down a steep-sided ravine, bouncing off the walls. Give it MOMENTUM — let each step keep a fraction of the last, like a heavy ball rolling — and the side-to-side bounces cancel while the downhill speed builds. It reaches the bottom in a fraction of the steps. It’s why real optimizers (Adam, SGD+momentum) all carry velocity. Slide the momentum.

◆ LIT▲ AMBER
◧ THE MEASURE · 2D
◍ BUILD SPEED · 3D · roll through the ravine
◆ THE FOURTH · 4D · a tesseract turns
◐ THE SHADOW · one dimension down
👶 THE TODDLER CORNER — one fat tap
momentum β
steps to bottom
plain GD steps
speedup

◆ LIT — exact / checkable

Momentum accumulates a velocity that is an exponentially-decaying average of past gradients: v ← βv − η∇L, θ ← θ + v. In an ill-conditioned ravine (large curvature in one direction, small in another) plain GD oscillates across the steep walls; momentum cancels those oscillations while accelerating along the shallow valley floor — often an order-of-magnitude fewer steps. It is the core of SGD-with-momentum, Nesterov, and (with per-parameter scaling) Adam. A fail-loud self-check throws unless momentum reaches the minimum of a ravine in fewer steps than plain GD. ◆ real optimization, node-verified.

▲ AMBER — the figure

Too much momentum (β→1) overshoots and can become unstable; the optimal β depends on the curvature. The ravine speedup and the velocity update are exact. Shown on a fixed quadratic ravine.

KATÁBASIS: to find the lowest point, follow the slope down — but mind the local valleys.  — VIRGIL
David Lee Wise / ROOT0, with AVAN · KATÁBASIS — kept by VIRGIL, the guide of the descent