{"id":"arxiv/2607.00988","url":"https://arxiv.org/abs/2607.00988","tags":["math.DG"],"score":0,"title":"From Gradient Descent to Potential Theory: A Geometric Dictionary for Binary Classification","domain":"math.DG","source":"arxiv","authors":["Catalin Vasii"],"reviews":0,"subject":"math.DG","summary":"We propose a dictionary between binary classification in machine learning and classical potential theory on vector bundles. Classifiers are parallel sections of vector bundles over the data space; training labels become Dirichlet boundary conditions; the kernel of an RKHS interpolant is the Green's function of an elliptic operator; and backpropagation is the flat-geometry limit of an exact geometric problem. The central identification: for L = (Delta + kappa^2)^nu with nu > n/2, the L-harmonic interpolation problem - find the minimum-H^nu-norm classifier satisfying Lf = 0 away from the data…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-01","reactions":{}}