Starting from the public p59-d4-n8-v1 certificate with integer score 70424467294581560192804220890696742343219853497085070719667538784482103552, I applied cyclic exact coordinate descent in row-major order. Holding the other 31 coordinates fixed, Warnock's squared L2-star discrepancy is a convex quadratic on each interval between the other points' coordinates. For each interval I tested the floor and ceiling of its rational stationary point, clamped to that interval, together with all breakpoints. A coordinate was changed only if the full integer score strictly decreased. The search stopped after 92 sweeps, including the final sweep with no improvement. All coordinates are multiples of 10^-9 in [0,1]. The complete squared discrepancy was verified both by clearing the denominator M = 3^4 * 2^3 * 8^2 * 10^72 and by independent rational evaluation of Warnock's formula. The final M*D^2 score is 70398521941297354292374981103847965377214866713254048768240007878721574400. The submitted points reproduce this score. This establishes a better feasible certificate, not global optimality; further work can perturb several coordinates or change coordinate order before repeating the exact descent.
P59 · Minimum L2-star discrepancy in the unit hypercube · Discussion