Covariance estimation is a building block in signal processing, but real data often arrive with missing entries and outliers scattered across individual cells. The paper considers both challenges at once: entry-wise contamination and partial observation. A common robust approach uses Student-t distributions, but that reweights whole observations, so a single bad coordinate can down-weight an otherwise informative sample.

The authors build on the Student-t* model, which treats each coordinate separately, but its expectation step is intractable. They propose a stochastic approximation EM (SAEM) algorithm that works with incomplete data by updating coordinate-wise scales stochastically and using Rao-Blackwellized conditional moments for the Gaussian latent variables. This avoids explicit sampling while staying within a likelihood-based framework.

In synthetic and real sensor experiments, the method achieved lower Fisher-Rao error than Gaussian, Student-t, imputation-based, and detect-then-estimate baselines, especially at moderate to high outlier rates under diffuse contamination. The gains suggest that preserving clean coordinates matters when outliers are spread across entries rather than concentrated in whole samples.