Let $Z$ be a positive semidefinite matrix with nonnegative entries, and define $X=\log(1+Z)$, where the $\log$ is taken entrywise, i.e., $X_{ij}=\log(1+Z_{ij})$. Are there some simple sufficient conditions that guarantee that $X$ is positive semidefinite?

Note 1 Intuitively, this should work when $Z$ is small enough, because then $X \simeq Z \succeq 0$.

Note 2 A simple counter-example is given by $Z=\left[ \begin{array}{cc} 3 & 5\\ 5 & 9 \end{array} \right]$.

Note 3 It is easy to verify that the converse relation allways holds, that is, $$X\succeq 0 \Longrightarrow Z:=exp(X)-1\succeq 0,$$ where the above operations are elementwise: $Z_{ij} = e^{X_{ij}} -1$. Indeed, we have the relation $Z=\sum_{k=1}^\infty X^{\circ k}/k!$, where $X^{\circ k}$ is the $k$th power of $X$ with respect to the Hadamard (elementwise) product, and it is well-known that the Hadamard product of 2 positive semidefinite matrices is positive semidefinite.

Note 4 Some background on this question: I ask myself when a random variable $Y$ with mean vector $\mu\in\mathbb{R}^n$ and variance-covariance matrix $\Sigma \succeq 0 \in \mathbb{R}^{n\times n}$ can be fitted to a multivariate log-normal distribution using the method of moments. Applying the formulas, we find that the elements of variance-covariance matrix $V$ of $\log Y$ must satisfy $$V_{ij}= \log(\frac{\Sigma_{ij}}{\mu_i \mu_j} +1),$$ but in general this matrix is not guaranteed to be positive semidefinite.


1 Answer 1


It's not true that it works for $Z$ small enough. Consider the $2 \times 2$ case $$ Z = \pmatrix{t & 2t\cr 2t & 4t\cr} $$ which is positive semidefinite for $t \ge 0$. Then $$\det(X) = \log(1+t)\log(1+4t) - \log(1+2t)^2 $$ which appears to be negative for all $t > 0$, and certainly is negative for small $t > 0$: its Maclaurin series is $ \det(X) = -2 t^3 + O(t^4)$.

What is true is that if $Z$ is positive definite, $\log(1+tZ)$ will be positive definite for sufficiently small $t > 0$. This can be obtained using the Maclaurin series for $\log(1+z)$.


Let $g(z) = z - \log(1+z)$, and write $X = Z - G$, where $G_{ij} = g(Z_{ij})$. Note that $g$ is increasing on $[0,1]$. If all $|Z_{ij}| \le b$, then $|W_{ij}| \le g(b)$. If your matrices are $N \times N$, for any vector $x$ we have by Cauchy-Schwarz $x^T G x \le N g(b) \|x\|^2$. Thus to have $X \succeq 0$ it suffices for $Z$ to be positive definite with least eigenvalue $\lambda$ and elementwise bound $b$ where $$ \lambda \ge N g(b) $$

  • 1
    $\begingroup$ Thank you, this is a very enlightening answer. But now the question is: How do I recognize if a positive definite matrix $Z$ is small enough (so that $X\succeq 0$)? Maybe we can find a criterion relying on the eigenvalues of $Z$ ? $\endgroup$
    – guigux
    Aug 19, 2015 at 9:29
  • $\begingroup$ Very nice, his is exactly the kind of criterion I was looking for ! $\endgroup$
    – guigux
    Aug 20, 2015 at 12:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.