Experiments, notes, and small studies on statistical learning, geometry, optimization, and the gap between formal guarantees and practical interpretation.
A note on why MMD and other distribution discrepancies are not, by themselves, classifier-risk statements.
A small robustness note on weak distributional perturbations, tail-sensitive learners, and downstream model instability.
A diagnostic on when nonlinear heads stop adding much after a frozen representation becomes linearly readable.
A note on why raw Wasserstein loss measures displacement in the chosen ground geometry, not automatically scale-relative statistical error.
A note on why kernel power functions do not become pointwise error estimates without a meaningful RKHS norm bound.
A note on empirical assignment, population transport, and missing tail mass.
A note on conformal validity, score geometry, and efficient prediction sets.
A small edge-case experiment for natural gradient, Fisher preconditioning, and non-convex basin selection.
A note on when Euclidean distance stops meaning relevance.
A geometric reading of James–Stein estimation, bias, and high-dimensional risk.
A geometric reading of Firth’s adjustment, Jeffreys volume, and why separation breaks ordinary GLMs.
A note on why high-dimensional structure can disappear in 2D.