Optimal prediction in the linearly transformed spiked model (Q2176630)

The authors consider the linearly transformed spiked model \(Y_i=A_iX_i+\epsilon_i\), \(i=1,\dots,n\). Here, the observations \(Y_i\) are noisy linear transforms of unobserved signals of interest \(X_i\). The observed transform matrices \(A_i\) reduce the dimension of the signal \(X_i\in \mathbb{R}^p\) to a possibly observation-dependent dimension \(q_i\leq p,\) thus \(A_i\in \mathbb{R}^{q_i\times p}\). The unobserved signals (or regression coefficients) \(X_i\) are vectors lying on an unknown low-dimensional space. Given only \(Y_i\) and \(A_i\), to predict or recover \(X_i\) values, the authors develop optimal methods by ``borrowing strength'' across the different samples. They use linear empirical Bayes methods scale to large datasets and weak moment assumptions. \newline The model has wide-ranging applications in signal processing, deconvolution, cryo-electron microscopy, and missing data with noise. For missing data, the authors demonstrate in simulations that the proposed methods are more robust to noise and to unequal sampling than well-known matrix completion methods.