(a) Assume $A\in\mathbb{C}^{n\times n}$ has $n$ distinct eigenvalues. Prove that there are exactly $2^n$ distinct matrices $B$ such that $B^2 = A$ (i.e., in particular, there are no more than $2^n$ matrices with this property).

(b) How many such matrices $B\in\mathbb{C}^{3\times3}$ exist if $A=\begin{pmatrix}2&0&0\\0&2&0\\0&0&1\end{pmatrix}$? Why?

(a) It is clear that if $\lambda_i$ is an eigenvalue of $A$, then $\pm\sqrt\lambda_i$ is an eigenvalue of $B$, therefore $$B = \begin{pmatrix}\pm\sqrt\lambda_1&\cdots&0\\\vdots&\ddots&\vdots\\0&\cdots&\pm\sqrt\lambda_n\end{pmatrix}$$ satisfies $B^2=A$ and there are $2^n$ such matrices $B$ because of $2^n$ possibilities of rearranging different numbers of plus and minus signs on the diagonal.

(b) We can construct $2^3=8$ such matrices $B$ using approach of (a), but $B$ is not necessarily diagonal. Consider $$B=\begin{pmatrix}\sqrt2&-1&0\\0&-\sqrt2&0\\0&0&1\end{pmatrix}$$ which satisfies $B^2=A$ as well. So, there exist more than $2^3$ such matrices.

I can't give an explanation why there are no more than $2^n$ possibilities in the first case and more than $2^3$ in the second. It definitely follows from the fact that Jordan normal form is unique in (a) and isn't unique in (b) because of an eigenvalue with algebraic multiplicity $2$, however, I can't formulate it into a self-contained statement.

  • 1
    $\begingroup$ Did you mean $$B=\begin{pmatrix}\sqrt2&-1&0\\0&-\sqrt2&0\\0&0&1\end{pmatrix}?$$ $\endgroup$ – Lord Shark the Unknown Jun 29 at 16:50
  • $\begingroup$ @LordSharktheUnknown, yes, it was a typo and I edited the question. Thank you. $\endgroup$ – Hasek Jun 29 at 17:06

See my answer in

Can we reduce finding matrix roots to finding roots of Jordan blocks?

In (b), your matrix $A$ has $2$ Jordan blocks associated to the same eigenvalue $2$. Then $A$ admits an infinity of square roots.

Here the infinity of matrices $B\in M_2(\mathbb{C})$ s.t. $tr(B)=0,\det(B)=-2$ are among the square roots of $2I_2$.

EDIT. I forgot to write that, in case (a), there are no more than $2^n$ square roots because, necessarily $AB=BA$.


You showed that $A$ has at least $2^n$ square roots. Assume $B^2 = A$. Since $\sigma(B)^2 =\sigma(B^2) = \sigma(A)$, it follows that $$\sigma(B) = \{\varepsilon_1\sqrt{\lambda_1}, \ldots, \varepsilon_n\sqrt{\lambda_n}\}$$ for some $\varepsilon_1, \ldots, \varepsilon_n \in \{-1,1\}$ where $\sigma(A) = \{\lambda_1, \ldots \lambda_n\}$. Set $D = \operatorname{diag}(\varepsilon_1\sqrt{\lambda_1}, \ldots, \varepsilon_n\sqrt{\lambda_n})$. $A$ has $n$ distinct eigenvalues so it is diagonalizable, i.e. there exists an invertible matrix $P$ such that $P^{-1}AP = \operatorname{diag}(\lambda_1, \ldots, \lambda_n) = D^2$.

Assume that $C^2 = A$ and that $\sigma(C) = \sigma(B)$. $B$ and $C$ both have $n$ distinct eingenvalues so there exist invertible matrices $Q,R$ such that $$Q^{-1}BQ = D = R^{-1}CR$$ It follows $$Q^{-1}AQ = Q^{-1}B^2Q = (Q^{-1}BQ)^2 = D^2 = (R^{-1}CR)^2 = R^{-1}C^2R = R^{-1}AR$$

If $e_1, \ldots, e_n$ is the standard basis, it follows that $A(Pe_i) = \lambda_iPe_i, A(Qe_i) = \lambda_iQe_i, A(Re_i) = \lambda_iRe_i$.

Since $\lambda_1, \ldots, \lambda_n$ are all distinct, the eigenvectors are unique up to scalar multiplication so it follows that $Qe_i = \sigma_i Pe_i, Re_i = \pi_i Pe_i$ for some nonzero scalars $\sigma_i, \pi_i$, or $Q = P\Sigma, R = P\Pi$ for some invertible diagonal matrices $\Sigma, \Pi$.

Finally, $$B = QDQ^{-1} = (P\Sigma)D(P\Sigma)^{-1} = P(\Sigma D\Sigma^{-1})P^{-1} =\\ PDP^{-1} = P(\Pi D\Pi^{-1})P^{-1} = (P\Pi)D(P\Pi)^{-1} = R^{-1}DR = C$$

Hence $B$ is uniquely determined by its spectrum $\sigma(B)$, which is uniquely defined by choosing $\varepsilon_i \in \{-1,1\}$ which can be done it $2^n$ ways. We conclude that there are $2^n$ such matrices $B$.


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.