Two short proofs of incompatibility for correlation matrices

Suppose first that such a \mathbf{V} exists, and let \mathbf{X} = A\mathbf{V}, the random vector in \mathbb{R}^d with coordinates X_j = \langle \mathbf{a}_j,\mathbf{V}\rangle. By assumption, each X_j has law F, so \mathbf{X} has all its margins equal to F; in particular, each X_j has mean 0 and variance 1, so the correlation matrix of \mathbf{X} is simply \mathbb{E}[\mathbf{X}\mathbf{X}^{\!\top}]. Since \mathbb{E} \mathbf{V} = 0, we have \mathbb{E}[\mathbf{V}\mathbf{V}^{\!\top}] = \operatorname{Cov} \mathbf{V} = I_k (the k \times k identity matrix), and therefore \mathbb{E}[\mathbf{X}\mathbf{X}^{\!\top}] = A \, \mathbb{E}[\mathbf{V}\mathbf{V}^{\!\top}]A^{\!\top} = AA^{\!\top} = R. Thus R \in S_d^F.

Conversely, suppose R \in S_d^F, and let \mathbf{Y} be a random vector with margins F and correlation matrix R. Since F is standardised, \mathbb{E} \mathbf{Y} = 0 and \mathbb{E}[\mathbf{Y}\mathbf{Y}^{\!\top}] = R. The matrix A has full column rank, so A^{\!\top}A is invertible, and A^{+} = (A^{\!\top}A)^{-1}A^{\!\top} (the Moore–Penrose inverse of A) satisfies A^{+}A = I_k. The matrix P = AA^{+} is the orthogonal projection of \mathbb{R}^d onto the column space of A: it is symmetric, P^2 = P, and PA = A. Set \mathbf{V} = A^{+}\mathbf{Y}. Then \mathbb{E} \mathbf{V} = 0 and

Moreover \mathbf{Y}-A\mathbf{V} = (I-P)\mathbf{Y}, and (I-P)^{\!\top}(I-P) = I-P, so

because (I-P)A = 0. Hence \mathbf{Y} = A\mathbf{V} a.s., and \langle \mathbf{a}_j,\mathbf{V}\rangle = Y_j has law F for every j.

The matrix M\oplus I_{d-m} lies in P_d. If it were the correlation matrix of some \mathbf{X} with margins F, then (X_1,\dots,X_m) would have margins F and correlation matrix M, which would contradict M\notin S_m^F.

Write \mathbf{b}_1,\dots,\mathbf{b}_{11} for the coefficient vectors of the eleven linear forms in (?), with multiplicities c_1 = 6 and c_2 = \dots = c_{11} = 1. Put \mathbf{a}_j = \mathbf{b}_j/|\mathbf{b}_j|. The first vector has |\mathbf{b}_1|^2 = 4 and the rest have |\mathbf{b}_j|^2 = 12, so (?) reads 192|\mathbf{x}|^4 = \sum_jc_j|\mathbf{b}_j|^4\langle \mathbf{a}_j,\mathbf{x}\rangle^4, with

Dividing by 1536 gives us

so that \sum_jw_j = \tfrac1{16}+10\cdot\tfrac3{32} = 1.

Let A be the 11\times4 matrix with rows \mathbf{a}_1^{\!\top},\dots,\mathbf{a}_{11}^{\!\top}, and let M = AA^{\!\top}, the matrix with entries M_{ij} = \langle \mathbf{a}_i,\mathbf{a}_j\rangle. Then M is symmetric and positive semidefinite with unit diagonal, so M \in P_{11}. The \mathbf{a}_j must span \mathbb{R}^4 (otherwise they would all be orthogonal to some \mathbf{b}\neq 0, and (?) at \mathbf{x} = \mathbf{b} would give 0 = 192|\mathbf{b}|^4). Hence A has rank 4, so does M = AA^{\!\top}, and A is a rank decomposition of M.

Suppose M \in S_{11}, and let \mathbf{V} be as in Lemma 2.1, so that \langle \mathbf{a}_j,\mathbf{V}\rangle\sim u for every j. Substituting \mathbf{x} = \mathbf{V} in (Proof) and taking expectations, using \mathbb{E}[\langle \mathbf{a}_j,\mathbf{V}\rangle^4] = \mathbb{E}[u^4] = 9/5,

But \mathbb{E}|\mathbf{V}|^2 = \operatorname{tr}\operatorname{Cov} \mathbf{V} = 4, so Jensen’s inequality gives \mathbb{E}|\mathbf{V}|^4\ge(\mathbb{E}|\mathbf{V}|^2)^2 = 16, a contradiction. Hence M \in P_{11}\setminus S_{11}. We extend this to all d \ge 11 using Lemma 2.2.

(\Leftarrow) This is a result of Devroye and Letac [DL2015, §4]: for every k\ge\tfrac12 and every d\le5, each d\times d correlation matrix is the correlation matrix of a random vector whose margins are all \mathrm{Beta}(k,k). The arcsine law is the case k = \tfrac12. We can standardise using the map map t\mapsto\sqrt2 \, (2t-1) and change no correlation (since it is affine). Then P_d\subseteq S_d^F for d\le5. Since the opposite inclusion S_d^F\subseteq P_d holds for any F, S_d^F = P_d.

(\Rightarrow) We show that S_6^F\neq P_6; by Lemma 2.2 it follows that S_d^F\neq P_d for every d\ge6. Let \varphi = \tfrac{1+\sqrt5}{2}, so that \varphi^2 = \varphi+1. Consider the six vectors

that is, one vertex from each pair of opposite vertices of a regular icosahedron. Using (p+q)^4+(p-q)^4 = 2 \, (p^4+6p^2q^2+q^4) on the pairs \{\mathbf{u}_1,\mathbf{u}_2\}, \{\mathbf{u}_3,\mathbf{u}_4\} and \{\mathbf{u}_5,\mathbf{u}_6\},

Summing up, the coefficient of each x_i^4 is 2(1 + \varphi^4), and that of each x_i^2 x_j^2 is 12 \varphi^2. Since \varphi^4 = (\varphi + 1)^2 = 3 \varphi+2, we find the sum to be 6 \varphi^2 \bigl(\sum_ix_i^4 + 2\sum_{i<j} x_i^2 x_j^2 \bigr) = 6 \varphi^2 |\mathbf{x}|^4. Each \mathbf{u}_j has |\mathbf{u}_j|^2 = 1 + \varphi^2 = \varphi+2, and (\varphi+2)^2 = 5 \varphi^2. Therefore, the unit vectors \mathbf{a}_j = \mathbf{u}_j/|\mathbf{u}_j| satisfy

Let A be the 6\times3 matrix with rows \mathbf{a}_1^{\!\top},\dots,\mathbf{a}_6^{\!\top}, and let M = AA^{\!\top}, the matrix with entries M_{ij} = \langle \mathbf{a}_i,\mathbf{a}_j\rangle. Then M is symmetric, positive semidefinite, and has unit diagonal, so M \in P_6. The \mathbf{a}_j must span \mathbb{R}^3 (otherwise they would all be orthogonal to some \mathbf{b}\neq 0, and (?) at \mathbf{x} = \mathbf{b} would give 0 = \tfrac65|\mathbf{b}|^4). So A has rank 3, as does M = AA^{\!\top}, meaning that A is a rank decomposition of M.

Suppose now that M \in S_6^F. Let \mathbf{V} be as in Lemma 2.1, so that \langle \mathbf{a}_j,\mathbf{V}\rangle\sim F for every j. Substituting \mathbf{x} = \mathbf{V} in (?) and taking expectations, using \mathbb{E}[\langle \mathbf{a}_j,\mathbf{V}\rangle^4] = \mathbb{E}[F^4] = \tfrac32,

But \mathbb{E}|\mathbf{V}|^2 = \operatorname{tr}\operatorname{Cov} \mathbf{V} = 3, so Jensen’s inequality gives \mathbb{E}|\mathbf{V}|^4\ge(\mathbb{E}|\mathbf{V}|^2)^2 = 9, a contradiction. Hence M \in P_6 \setminus S_6^F.