fullseye

LinkedIn 投稿キット(数学の絵を定理で採点する話)

投稿するのは人間です。ここにあるのは下書きと、添える画像の指定と、出す前に 確かめることの一覧。確かめずに貼らないこと —— 数字は全部「その時点の実測」です。

切り口

作ったものを先に、新しい結果を次に、教訓は記事側で読ませる。 この話の芯は 「数学の絵はきれいなので誰も確かめない」ではなく、「絵の外に真値を置く方法が ちゃんとある」というほうです。失敗談から入ると何を作ったのか伝わりません。

数字は 1 つだけ選ぶこと。全部並べると読まれません。いちばん効くのは 対照群のある数字(極小曲面の |H| 0.0001 に対し球 1.00004・円柱 0.50000)です —— 「門が素通しでない」ことまで 1 行で言えるのはこれだけ。

本文(English)

Mathematical pictures never get checked, because they already look right. A
slightly wrong implementation still packs circles, still spirals, and a Lorenz
plot is the same butterfly whether the integrator is 1st or 4th order.

So I built 32 drawing operators under one rule: ship it only if a theorem can
act as the gate, or an existing independent implementation supplies the truth.
Checking a formula against itself is not allowed.

What that buys you:
• "This is a minimal surface" is scored by an op that does not know how it was
  built — median |H| of 1e-4, against controls of 1.00004 (unit sphere) and
  0.50000 (unit cylinder). The gate is provably not a pass-through.
• The sum of the Lyapunov exponents matches the trace identity to 2.6e-06,
  independently of the tangent-flow/QR procedure that produced them.
• A print moiré and a two-slit fringe are the same beat, so the period is
  predicted in closed form before anything is drawn, then measured back by FFT.

The gates found four real defects in my own work — including a hatching routine
whose docstring said "follows the edges" while the code crossed them. You cannot
see that in the picture.

https://qiita.com/furuse-kazufumi/items/dc100e7ea90e9575c40b
#ComputerVision #ScientificComputing #Python

出す前に確かめること