You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bench #478] Scale every canonical operation across samples, features, classes and model size #490
The 19-row canonical suite is excellent for reproducibility but one fixed shape can confuse interpreter/setup overhead with algorithmic throughput. Every canonical operation needs a scaling surface.
Build deterministic scaled fixtures covering relevant axes: n_samples, n_features, n_classes/clusters, n_estimators/depth, support vectors/kernel matrix size and batch size. Use logarithmic sizes from tiny through realistically large, with explicit memory-feasibility caps for O(n^2)/O(n^3) estimators.
Acceptance: each canonical operation has at least three meaningful sizes; slopes/complexity regimes are machine-readable; Pages can show crossover points; a Flow win at tiny size is not considered complete if it becomes a material loss at larger supported sizes; all scaled losses automatically link into #478.
Worked out a concrete benchmark-grid rule for #490 and archived the derivation + Flow witness in godofecht/OpenProjects.
For fixed endpoints n_min < n_max and L benchmark points, if r_i = n_{i+1}/n_i, then prod r_i = n_max/n_min. Therefore R = max r_i obeys
R >= (n_max / n_min)^(1/(L-1)),
with equality iff all adjacent ratios are equal. So geometric spacing is minimax-optimal for the worst unresolved multiplicative scale interval.
This gives a clean default for the scaled suite: geometric base points over each feasible axis, then refine only observed status transitions at geometric midpoints. If the initial transition bracket has multiplicative width rho, after k refinements the width is rho^(1/2^k).
Example: 7 points over a 64x range give worst-case factor-2 crossover localization immediately. A 5-point exact fixture [81, 324, 1296, 5184, 20736] attains the 4x minimax lower bound; an uneven grid over the same endpoints has an 8x hole.
The note also records the limitation: speedup need not be monotone, so same-status intervals can still hide double crossovers; slope/curvature changes should be allowed to trigger extra refinement.
Activity
godofecht commentedon Aug 26, 2026
Worked out a concrete benchmark-grid rule for #490 and archived the derivation + Flow witness in
godofecht/OpenProjects.For fixed endpoints
n_min < n_maxandLbenchmark points, ifr_i = n_{i+1}/n_i, thenprod r_i = n_max/n_min. ThereforeR = max r_iobeysR >= (n_max / n_min)^(1/(L-1)),with equality iff all adjacent ratios are equal. So geometric spacing is minimax-optimal for the worst unresolved multiplicative scale interval.
This gives a clean default for the scaled suite: geometric base points over each feasible axis, then refine only observed status transitions at geometric midpoints. If the initial transition bracket has multiplicative width
rho, afterkrefinements the width isrho^(1/2^k).Example: 7 points over a 64x range give worst-case factor-2 crossover localization immediately. A 5-point exact fixture
[81, 324, 1296, 5184, 20736]attains the 4x minimax lower bound; an uneven grid over the same endpoints has an 8x hole.The note also records the limitation: speedup need not be monotone, so same-status intervals can still hide double crossovers; slope/curvature changes should be allowed to trigger extra refinement.
Files:
ml/minimax-geometric-scaling-surfaces.mdml/minimax-geometric-scaling-surfaces.flowNo hosted CI/workflow was touched.