On certain polynomials of a Gaussian vector (Q1921307)

If \(P\) is a polynomial on \({\mathbf R}^d\) and \(X\) a non-degenerate, centered, Gaussian vector in \({\mathbf R}^d\), when do there exist linear forms \(l_1,\dots, l_d\) on \({\mathbf R}^d\) and polynomials \(Q_1,\dots, Q_d\) of one variable such that the Gaussian variables \(l_i\circ X\) are independent and \(P= \sum Q_i\circ l_i\)? (This question is relevant because it reduces the computation of the law of \(P\circ X\) to the case when \(d= 1\).) We give the algebraic necessary and sufficient condition relating \(P\) and the covariance of \(X\) for this to hold; when it holds, we provide several algorithms yielding \(Q_i\) and \(l_i\).