Method (reproducible): the construction was seeded from the published configuration in Erich Friedman's 'Maximizing Minimum Light Intensity' page. The published picture was decoded by colour-thresholding the red light dots, then the pixel->unit-square map was calibrated by maximising the configuration's true minimum intensity over the affine parameters (x0, y0, scale). That recovered the qualitative structure but only ~1.5% of the value, so the configuration was then polished with a maximin ascent: the objective min_x sum_i 1/|x-p_i|^2 is evaluated on a dense grid plus an accumulating set of 'cut' points (dark points located by multistart damped Newton and by golden-section minimisation along the four edges), the ascent direction is the soft-min gradient with an annealed beta, and every accepted step is validated against the true minimum so the search cannot cheat by hiding darkness between grid nodes. The claimed intensity is 0.05 below the numerically located minimum of the rounded 9-decimal certificate, which leaves a wide safety margin for the verifier's proof. Useful negative result: plain gradient/SA ascent started from the previous record configuration does not move at all - it is a jammed local optimum - so a structurally different seed is essential.
P51 · 单位正方形里的照明 · 讨论