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

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

A downhill direction tells you which way to go — but how FAR? Step too little and you crawl; too far and you overshoot back uphill. Line search answers it: try a big step, and while it doesn’t reduce the value ENOUGH, halve it — backtracking until the decrease is genuine (the Armijo condition). It’s the trusty sidekick that makes gradient and Newton methods robust. Slide to backtrack the step down to a good one.

◆ LIT▲ AMBER
◧ THE MEASURE · 2D
◍ HOW FAR TO STEP · 3D · pick a direction, then the right distance
◆ THE FOURTH · 4D · a tesseract turns
◐ THE SHADOW · one dimension down
👶 THE TODDLER CORNER — one fat tap
step α
f(new)
Armijo met?
rule
enough decrease

◆ LIT — exact / checkable

Given a descent direction d at point x, LINE SEARCH chooses the step length α. Exact minimization along the ray is costly, so inexact backtracking is used: start with α=1 and HALVE it until the Armijo (sufficient-decrease) condition holds, f(x+αd) ≤ f(x)+c₁α∇f·d — guaranteeing the step buys a decrease proportional to its size and slope (the Wolfe conditions add a curvature test). This makes [[the-gradient-descent|gradient descent]], Newton, and quasi-Newton (BFGS) globally convergent and removes the fragile fixed [[the-learning-rate|learning rate]]. A fail-loud self-check throws unless backtracking returns a step that satisfies Armijo and lowers f. ◆ real optimization, node-verified.

▲ AMBER — the figure

Backtracking guarantees SUFFICIENT decrease, not the true 1-D minimum; too-small c₁ or bad directions still crawl, and each trial costs a function evaluation. The Armijo condition and the halving rule are exact.

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