On conditional moments of high-dimensional random vectors given lower-dimensional projections (Q2405194)

Consider a random \(d\)-vector \(Z\) that has a Lebesgue density, and a \(d \times p\) matrix \(B\) with \(d\) large, \(p\) small. If \(Z\) is Gaussian, then the conditional expectation of \(Z\) given \(B' Z\) is a linear function of \(B'Z\) and the conditional variance of \(Z\) given \(B' Z\) is constant. It can be shown that the multivariate Gaussian distribution is the only distribution that satisfies these properties for all choices of \(B\). This paper provides sufficient conditions on the distribution of \(Z\) such that these properties hold approximately, in the sense that the distance between the actual conditional expectation/variance and the linear/constant expression that would hold if \(Z\) were Gaussian, tends to zero in probability as \(d \to \infty\), uniformly over a growing set of \(d \times p\)-matrices. Also non-asymptotic bounds are given. This extends earlier results by the second author in [Ann. Stat. 41, No. 2, 464--483 (2013; Zbl 1360.62371)], where the case \(p=1\) was considered.