In summary, Log concavity of determinants refers to the property of a determinant that states that the natural logarithm of the determinant is a concave function of its entries. This allows us to use convex optimization techniques to solve problems and has practical applications in fields such as economics, engineering, and statistics. It is closely related to the concept of positive definiteness and can be tested using methods such as the Hessian matrix or numerical methods like the Newton-Raphson method.
The case ##t=0## is considered separately. Obviously
$$
\det B \geq (\det A)^0 \det B .
$$
We now consider the case ##0 < t \leq 1##. I start with
$$
\det (t A + (1-t) B) \geq (\det A)^t (\det B)^{1-t}
$$
and work backwards to an inequality that can be seen to hold. Take the log of both sides
$$
\ln \det ( t A + (1-t) B) \geq t \ln \det A + (1-t) \ln \det B
$$
where we have used that ##\det A > 0## and ##\det B > 0##. As ##\det A \not= 0## the matrix ##A## has an inverse and the above inequality becomes
$$
\ln \det [tA ( \mathbb{1} + t^{-1} (1-t) A^{-1} B)] \geq t \ln \det A + (1-t) \ln \det B
$$
or
$$
\ln \det (t A) + \ln \det ( \mathbb{1} + t^{-1} (1-t) A^{-1} B) \geq t \ln \det A + (1-t) \ln \det B
$$
or
$$
nt + \ln \det ( \mathbb{1} + t^{-1} (1-t) A^{-1} B) \geq (1-t) (\ln \det B - \ln \det A)
$$
or
$$
nt + \ln \det ( \mathbb{1} + t^{-1} (1-t) A^{-1} B) \geq (1-t) \ln \det A^{-1} B
$$
The matrix ##A^{-1} B## is not necessarily symmetric and so can't in general be diagonalised, but there is a similarity transformation that puts it into Jordan normal form. So that
$$
nt + \ln \det [S^{-1} ( \mathbb{1} + t^{-1} (1-t) A^{-1} B) S] \geq (1-t) \ln \det (S^{-1} A^{-1} B S)
$$
where ##\lambda_i## are the eigenvalues of ##A## and ##\mu_i## are the eigenvalues of ##B##. These eigenvalues are positive and so the logarithm makes sense. The above inequality can be rewritten as
I think I spotted an error in the argument of my previous post, but I think I know how to fix it.
EDIT: on 9th Sept I have added further explanation.
In my previous post I demonstrated that
$$
\det (t A + (1-t) B) \geq (\det A)^t (\det B)^{1-t}
$$
reduces to
$$
nt + \ln \det ( \mathbb{1} + t^{-1} (1-t) A^{-1} B) \geq (1-t) \ln \det A^{-1} B \qquad (*)
$$
A positive definite matrix has a positive definite square-root matrix. Proof:
The matrix ##B## is positive definite and so there is a similarity transformation such that ##S^{-1} B S = D## where ##D## is the diagonal matrix. Where the matrix ##S## is chosen to be the matrix with the ##n## eigenvectors as columns. The elements of ##D## are eigenvalues of ##B## and so are positive. When a matrix is symmetric, the diagonalizing matrix ##S## can be made an orthogonal matrix by suitably choosing the eigenvectors (then ##S^{-1} = S^T##). As such we can define a positive definite square-root matrix ##B^{1/2} = S D^{1/2} S^T## (##B^{1/2}## has the same eigenvectors as ##B## but with eigenvalues that are the square-root of the eigenvalues of ##B##).
Q.E.D.
Now, the matrix ##A^{-1} B## can be written as ##B^{-1/2} (B^{1/2} A^{-1} B^{1/2}) B^{1/2}##. As such the matrix ##A^{-1} B## has the same eigenvalues as the positive definite matrix ##B^{1/2} A^{-1} B^{1/2}## because:
As such the matrix ##A^{-1} B## has positive eigenvalues, let us denote them as ##\nu_i##.
The matrix ##A^{-1} B## is not necessarily symmetric and so can't in general be diagonalised, but there is a similarity transformation that puts it into Jordan normal form. So that ##(*)## becomes
I have made a mistake in the previous attempts (inattention!). I think this time I have done the problem.
I start with
$$
\det (t A + (1-t) B) \geq (\det A)^t (\det B)^{1-t}
$$
and work backwards to an inequality that can be proven to hold. Take the log of both sides
$$
\ln \det ( t A + (1-t) B) \geq t \ln \det A + (1-t) \ln \det B
$$
where we have used that ##\det A > 0## and ##\det B > 0##. As ##\det A \not= 0## the matrix ##A## has an inverse and the above inequality becomes
$$
\ln \det [A ( t\mathbb{1} + (1-t) A^{-1} B)] \geq t \ln \det A + (1-t) \ln \det B
$$
or
$$
\ln \det A + \ln \det ( t \mathbb{1} + (1-t) A^{-1} B) \geq t \ln \det A + (1-t) \ln \det B
$$
or
$$
\ln \det ( t \mathbb{1} + (1-t) A^{-1} B) \geq (1-t) (\ln \det B - \ln \det A)
$$
or
$$
\ln \det ( t \mathbb{1} + (1-t) A^{-1} B) \geq (1-t) \ln \det A^{-1} B \qquad (*)
$$
A positive definite matrix has a positive definite square-root matrix. Proof:
The matrix ##B## is positive definite and so there is a similarity transformation such that ##S^{-1} B S = D## where ##D## is the diagonal matrix. Where the matrix ##S## is chosen to be the matrix with the ##n## eigenvectors as columns. The elements of ##D## are eigenvalues of ##B## and so are positive. When a matrix is symmetric, the diagonalizing matrix ##S## can be made an orthogonal matrix by suitably choosing the eigenvectors (then ##S^{-1} = S^T##). As such we can define a positive definite square-root matrix ##B^{1/2} = S D^{1/2} S^T## (##B^{1/2}## has the same eigenvectors as ##B## but with eigenvalues that are the square-root of the eigenvalues of ##B##).
Q.E.D.
Now, the matrix ##A^{-1} B## can be written as ##B^{-1/2} (B^{1/2} A^{-1} B^{1/2}) B^{1/2}##. As such the matrix ##A^{-1} B## has the same eigenvalues as the positive definite matrix ##B^{1/2} A^{-1} B^{1/2}## because:
As such the matrix ##A^{-1} B## has positive eigenvalues, let us denote them as ##\nu_i##.
The matrix ##A^{-1} B## is not necessarily symmetric and so can't in general be diagonalised, but there is a similarity transformation that puts it into Jordan normal form. So that ##(*)## becomes