Adaptive Estimation in Multivariate Response Regression with Hidden Variables
This paper studies the estimation of the coefficient matrix $\Ttheta$ in multivariate regression with hidden variables, $Y = (\Ttheta)^TX + (B^*)^TZ + E$, where $Y$ is a $m$-dimensional response vector, $X$ is a $p$-dimensional vector of observable features, $Z$ represents a $K$-dimensional vector of unobserved hidden variables, possibly correlated with $X$, and $E$ is an independent error. The number of hidden variables $K$ is unknown and both $m$ and $p$ are allowed but not required to grow with the sample size $n$. Since only $Y$ and $X$ are observable, we provide necessary conditions for the identifiability of $\Ttheta$. The same set of conditions are shown to be sufficient when the error $E$ is homoscedastic. Our identifiability proof is constructive and leads to a novel and computationally efficient estimation algorithm, called HIVE. The first step of the algorithm is to estimate the best linear prediction of $Y$ given $X$ in which the unknown coefficient matrix exhibits an additive decomposition of $\Ttheta$ and a dense matrix originated from the correlation between $X$ and the hidden variable $Z$. Under the row sparsity assumption on $\Ttheta$, we propose to minimize a penalized least squares loss by regularizing $\Ttheta$ via a group-lasso penalty and regularizing the dense matrix via a multivariate ridge penalty. Non-asymptotic deviation bounds of the in-sample prediction error are established. Our second step is to estimate the row space of $B^*$ by leveraging the covariance structure of the residual vector from the first step. In the last step, we remove the effect of hidden variable by projecting $Y$ onto the complement of the estimated row space of $B^*$. Non-asymptotic error bounds of our final estimator are established. The model identifiability, parameter estimation and statistical guarantees are further extended to the setting with heteroscedastic errors.