You are on page 1of 3
The Determinant of Real Symmetric Positive-Definite Matrices is Log-Concave A self-adjoint linear mapping H : Rn → Rn is called positive if x, Hx ≥ 0 for all x ∈ Rn \ {0}. We begin with the following elementary property of positive mappings. Lemma 1. H is positive if and only if its eigenvalues are positive. Proof. If v is an eigenvector of H with corresponding eigenvalue λ, then Hv = λv and so v, Hv = λ v, v . Since the left-hand side is strictly positive, it follows that λ > 0 as desired. Conversely, by the spectral theorem, since H is self-adjoint, there exists an orthonormal basis {xi } of eigenvectors of H and so any vector x ∈ Rn can be n n written as x = n i=1 ci xi . Hence Hx = i=1 ci λi xi and by orthonormality of the xi , we have x, Hx = i=1 λi c2 i . Hence, if λi > 0 for all i = 1, . . . , n, then H is positive. The proof of the above lemma also provides the following useful estimate. Corollary 1. If H is positive, then x, Hx ≥ m||x||2 , where m = min {λi } > 0. It follows that a positive mapping has trivial kernel and so is invertible. We recall that a function ϕ : Rn×n → R is concave if for all X, Y ∈ Rn×n and t ∈ [0, 1], we have ϕ(tX + (1 − t)Y ) ≥ tϕ(X ) + (1 − t)ϕ(Y ). Since the determinant of a positive matrix is always positive, the logarithm of the determinant is well-defined and so we may consider the function ϕ(X ) = log(det(X )). Our goal is to show that ϕ(X ) is concave on the subset of positive symmetric matrices. That is, for X and Y symmetric positive matrices and t ∈ [0, 1], log(det(tX + (1 − t)Y )) ≥ t log(det(X )) + (1 − t) log(det(Y )). We first observe that this is equivalent to showing det(tX + (1 − t)Y ) ≥ (det(X ))t (det(Y ))1−t . Our approach to proving the above inequality will be based on the following integral formula. 1 Theorem 1. If H ∈ Rn×n is symmetric positive, then π n/2 det(H ) = Rn e− x,Hx dx. Proof. The integral on the right hand side converges because of the estimate from the corollary above. By the spectral theorem, there exists a unitary matrix M such that Λ = M ∗ HM , where Λ is the diagonal matrix of eigenvalues of H . Consider the change of coordinates x = M y . Then x, Hx = M y, HM y = y, M ∗ HM y = n i=1 2 . Since λ i yi M is unitary, | det(M )| = 1 and so by the change of coordinates formula, we have n e Rn − x,Hx − dx = Rn e n i=1 2 λi yi dy = Rn i=1 n e−λi yi dy 2 2 = i=1 R n e−λi yi dyi 1 √ λi i=1 n = e−zi dzi R 2 √ π √ = λi i=1 = π n/2 n λi i=1 Since the determinant of a matrix is the product of its eigenvalues, the desired identity follows immediately. We use the above result to prove that the determinant is a log-concave function on the space of real symmetric positive matrices. Theorem 2. For M and N real symmetric positive matrices and t ∈ [0, 1], we have det(tM + (1 − t)N ) ≥ (det(M ))t (det(N ))1−t . 2 Proof. Let H = tM + (1 − t)N . Then H is real symmetric positive and so by the integral formula above, we have π n/2 det(tM + (1 − t)N ) = Rn e− e−t Rn x,(tM +(1−t)N )x dx dx = 1 t x,M x e−(1−t) x,N x Applying Holder’s inequality with p = and q = 1 1−t to the right hand side, we obtain t  1−t e− x,N x  e−t Rn x,M x e−(1−t) x,N x dx ≤  Rn e− x,M x dx  Rn dx . Once again, by the integral formula, we have the right hand side is equal to the quantity n/2 π tn/2 π (1−t)n/2 = (det(M ))t/2π , which gives us the desired inequality. (det(M ))t/2 (det(N ))(1−t)/2 (det(N ))(1−t)/2 3