Can all real diagonalizable matrices with real eigenvalues be diagonalized by a symmetric matrix? That is, if $A$ is a real diagonalizable matrix with real eigenvalues, can we write

$$ A = S D S^{-1} $$

for some real symmetric matrix $S$ and some real diagonal matrix $D$?

  • $\begingroup$ An observation: if $S$ and $A$ are symmetric, then $$ A^T = [SDS^{-1}]^T = S^{-1}DS = A = SDS^{-1} $$ Which means that we have $S^{-1}DS = SDS^{-1}$, which implies that $$ S^4D = DS^4 $$ If $A$ has distinct eigenvalues, then the above implies that $S^4$ is a diagonal matrix. $\endgroup$ – Omnomnomnom Jun 28 '19 at 17:25

Let us consider the more general setting in which $A$ is an $n\times n$ diagonalisable matrix over a field $\mathbb F$. By assumption, $A$ admits a diagonalisation $A=VD_1V^{-1}$. We are asking whether there exist a symmetric matrix $S$ and a diagonal matrix $D$ over $\mathbb F$ such that $VD_1V^{-1} = SDS^{-1}$. Since the two sides have the same spectrum, $D_1=PDP^T$ for some permutation $P$. Thus the equation can be rewritten as $$ (VP)D(VP)^{-1} = SDS^{-1}.\tag{1} $$ We now consider two cases:

  1. $n=2$. Then $(1)$ is always solvable. This is obvious if $A$ is a scalar matrix (so that $D=\lambda I$). Suppose $A$ has two distinct eigenvalues. Then all eigenspaces of $A$ are one-dimensional. Therefore, $(1)$ is solvable if and only if $$ S = VP\Lambda\tag{2} $$ for some nonsingular diagonal matrix $\Lambda$. Since $V$ is nonsingular, it has at most two zero entries but not any zero row/column. Therefore, there always exists a permutation $P$ such that $VP$ is in one of the following forms, where $a,b,c,d\ne0$: $$ \pmatrix{a&b\\ c&d}, \ \pmatrix{0&b\\ c&d}, \ \pmatrix{a&b\\ c&0}, \ \pmatrix{0&b\\ c&0}. $$ It follows that $VP\Lambda$ is symmetric when $\Lambda=\operatorname{diag}\left(1,\frac{c}{b}\right)$.
  2. $n\ge3$ and $\mathbb F$ has at least $n$ elements. Then $(1)$ is not always solvable. First, we claim that there exists an invertible and entrywise nonzero matrix of every order $n$. We can prove it by mathematical induction. The base case $n=1$ is trivial. In the inductive step, suppose $B\in GL_{n-1}(\mathbb F)$ is entrywise nonzero. Since $\mathbb F$ has at least $n\ge3$ elements, some two of them, say $b_1$ and $b_2$, are nonzero. Therefore $$ \det\pmatrix{B&\mathbf1^T\\ \mathbf1&b}=b\det(B)+\text{constant} $$ is nonzero for some $b\in\{b_1,b_2\}$ and our claim is proved. Now consider a diagonalisable matrix $A$ with distinct eigenvalues (such an $A$ always exists because $\mathbb F$ has at least $n$ elements) and suppose its eigenvectors are given by the columns of $$ V=\pmatrix{1&0\\ \mathbf1&B} $$ where $B$ is an invertible and entrywise nonzero matrix of order $n-1$. As $A$ has distinct eigenvalues, the problem again boils down to solving $(2)$ for $S,P$ and $\Lambda$. However, should $(2)$ be solvable, $VP$ must be sign-symmetric (by a sign, we mean whether an entry is nonzero). Yet, this is impossible, because $VP$ always contains an off-diagonal zero entry on the first row but none in any row below.

The remaining case where $n\ge3$ and $\mathbb F$ has fewer than $n$ elements is more intricate. Since $A$ must contain some repeated eigenvalues, the above counterexample does not apply.

| cite | improve this answer | |
  • 2
    $\begingroup$ But in this case, we can say $$ A = \pmatrix{0&1\\1&1} \pmatrix{0&0\\0&1}\pmatrix{0&1\\1&1}^{-1} $$ $\endgroup$ – Omnomnomnom Jul 1 '19 at 22:35
  • $\begingroup$ @Omnomnomnom You are right. I hope the loophole is fixed now. $\endgroup$ – user1551 Jul 2 '19 at 0:32
  • $\begingroup$ I think so. $\phantom{1}$ $\endgroup$ – Omnomnomnom Jul 3 '19 at 15:20

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.