Let $A$ be a real, symmetric and positive definite matrix and suppose $B$ is a real symmetric matrix such that $\exp(B) = A$.

Is $B$ unique?

The solution of my homework sheet says that $B$ is uniquely determined by $A$ since, for every eigenvalue $t \in \mathbb{R}$ of $B$, the $t$-eigenspace of $B$ equals the $e^t$-eigenspace of $A$.

So far I was only able to show one inclusion of this equality. How do I prove that if $r>0$ is an eigenvalue of $A$ and $v$ a vector such that $Av = rv$, then $Bv = \log(r)v$?

Is this even true?

  • 1
    $\begingroup$ Use a dimensional argument. $B$ is real symmetric, so it has a full basis of eigenvectors. So all of $\exp(B)$'s eigenvectors are accounted for, and (since this is in finite dimension) $A$ can't have any "extra" eigenvectors that didn't arise from $B$ (note that we're using the injectiveness of $e^x$ here: the analogous argument doesn't hold for $B^2 = A$). $\endgroup$
    – Erick Wong
    May 22, 2016 at 19:20

1 Answer 1


Thanks for the hint, Erick.

I now came up with the following argument. Suppose $A$ is a real positive definite symmetric Matrix and that $B$ is a real symmetric logarithm of $A$, i.e. $B$ is real and symmetric and $\exp(B) = A$.

We wish to show that $B$ is uniquely determined by $A$. Since $B$ is diagonalizable, it suffices to show that $\sigma(B)$, the set of eigenvalues of $B$, is uniquely determined by $A$, and that every eigenspace $E_t(B) = \lbrace v \in \mathbb{R}^n \, : \, Bv = tv \rbrace$, $t \in \sigma(B)$ is uniquely determined by $A$.

Since $B$ is symmetric, $B$ is similar to a diagonal matrix $D$, which has entries from $\sigma(B)$. Hence $\exp(B) $ is similar to $\exp(D)$. It follows that

$$\sigma(\exp(B)) = \lbrace e^t \, : \, t \in \sigma(B) \rbrace \Leftrightarrow \sigma(B ) = \lbrace \log(s) \,: \, s \in \sigma(A) \rbrace$$ Since $t \mapsto e^t$ is a bijection from $\mathbb{R}$ to $(0, \infty)$.

We next show that for every eigenvalue $t$ of $B$ we have

$$E_{t}(B) = E_{e^t}(\exp(B))$$

Let $v \in E_{t}(B)$. Then

$$\exp(B)v = \sum_{k\geq 0}(\frac{B^k}{k!})v = \sum_{k\geq 0}(\frac{B^kv}{k!}) = \sum_{k\geq 0}(\frac{t^kv}{k!}) = e^t v$$

Hence $E_t(B) \subset E_{e^t}(\exp(B))$. In order to prove equality, we prove that the dimensions of both spaces are equal. Since both $B$ and $\exp(B)$ are diagonalizable, we have the direct sum decompositions

$$ \bigoplus_{t \in \sigma(B)} {E_t(B)} = \mathbb{R}^n = \bigoplus_{t \in \sigma(B)} {E_{e^t}(\exp(B))}$$


$$\sum_{t \in \sigma(B)} {\dim(E_t(B))} = n = \sum_{t \in \sigma(B)} {\dim(E_{e^t}(\exp(B)))}$$

This implies $\dim(E_t(B)) = \dim(E_{e^t}(\exp(B)))$ for every $t \in \sigma(B)$, as desired.

  • $\begingroup$ If the argument is correct, this gives an alternative for the proof given here $\endgroup$
    – m.s
    May 23, 2016 at 19:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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