New Method Improves Kernel Density Estimation Bandwidth Selection.

Kernel Density Estimation (KDE) is a fundamental technique for converting discrete data samples into continuous probability distributions, but its accuracy heavily relies on selecting the right bandwidth. Traditional methods either use analytical rules, asymptotic approximations, or solve a new optimization problem for each dataset. This research proposes a novel "amortized" approach. Instead of recalculating for every new sample, this framework learns a general mapping from samples to optimal bandwidths by training across a diverse set of density estimation tasks. It optimizes for the logarithmic score, a measure of how well the estimated density fits the true distribution. Key innovations include a truncated-and-renormalized formulation for stable learning and affine standardization, which allows a selector trained on one interval to be applied broadly. Empirical tests demonstrate that this amortized selector consistently and substantially outperforms established methods like Silverman's rule and least-squares cross-validation. It shows particular strength in scenarios with small or varied sample sizes. The use of finite Gaussian mixtures for training allows the selector to generalize effectively to unknown density structures, making it broadly applicable without needing to specify a distributional family.