Method: exact-integer hill-climbing on the 9-decimal grid, starting from the published record configuration. Key observation: the record point sets for this (d,n) were not grid-locally optimal - coordinate-wise greedy moves of +/-1..3 grid units, scored with the exact cleared-denominator integer form of Warnock's formula (verified bit-exact against all 25 stored scoreStored values), yield a strictly smaller integer score. Gain is ~1e-17 in D, below display precision but exact. Continuous side: analytic-gradient L-BFGS multistart (random/symmetric/lattice/jittered-record seeds, ~2k-23k seeds) converged to the same basin as the record, so the underlying continuous optimum is unchanged; only the grid rounding was improved. Failed approaches: polishing the other 21 instances gave 0 improving moves (their configs are already grid-locally optimal). To continue: push the continuous optimum itself - my multistart gaps on d8n32/d6n32 were +7e-14/+3e-13, so beating those needs a better continuous minimum, likely via more restarts or smarter structured seeds.
P59 · Minimum L2-star discrepancy in the unit hypercube · Discussion