Weighted Score Matching on Bounded Domains: Nonparametric Density Estimation and Interaction Graph Recovery

We study nonparametric density estimation and interaction-graph recovery on bounded product domains using boundary-weighted score matching. In a canonically centered pairwise log-density model, graph edges are determined by bivariate interactions, but penalizing the univariate main effects can induce false edges. We derive the resulting shift in interaction scores and study partially penalized weighted score matching (P-WSM), which leaves main effects unpenalized. At the population level, P-WSM excludes edges between independent subsystems even when the marginal log densities are not represented exactly by the sieve. A nonparametric construction separates partial and joint penalization: throughout a common deterministic tuning interval, P-WSM recovers the graph while joint penalization selects a false edge. We also give growing-dimensional recovery conditions that do not require full-Gram invertibility. For density estimation at fixed dimension, we establish weighted Fisher-divergence rates and, under stronger basis-specific conditions, Kullback-Leibler and Hellinger rates for the normalized estimator. Simulations examine the benefits and limitations of partial penalization. An analysis of bone-marrow signaling data illustrates nonlinear conditional-association estimation.