Scalable Minimum-Volume Simplex Estimation with Non-asymptotic Analysis
En palabras de los autores
We study the estimation of a -dimensional simplex from i.i.d.\ points sampled uniformly from its interior; the observations are convex combinations of unknown prototypes. Existing polynomial-time estimators need cubic per-sample work or storage and are impractical at --. We propose DeepMVSA, which re-expresses the minimum-volume principle in neural implicit form: a lightweight coordinate network generates the mixing weights and a triangular LU-type parameterization the dual simplex matrix, reducing the trainable-state memory to , independent of , and the cost per data pass to . We prove a non-asymptotic sample-complexity bound of the polynomial-time benchmark order for a localized surrogate estimator; an oracle inequality for every global minimizer of the neural objective, with volume-inflation control and an explicit shrinkage bias; a conditional end-to-end error budget separating statistical, approximation, optimization, and enclosure-residual terms on an explicit envelope event; and two-point lower bounds: at any noise level fixed independently of , the scaling is unimprovable in its -exponent. Experiments with up to synthetic observations are consistent with the predicted accuracy and scaling, and feasibility on real scenes of pixels is demonstrated.
Apareció: miércoles, 23 de septiembre. arXiv. Preprint, todavía sin revisión por pares.