Enumerating the square roots of a $2 \times 2$ complex matrix using its eigenvalues and diagonalizability.
My answer:
Cases:
If the eigenvalues are equal then infinitely many.
If all of it's eigenvalues are distinct it has 4 unless one of them is zero and then 2.
If it is not diagonalizable then it doesn't have any square roots.
My questions:
I'm very unsure about case 3, and how to prove it.
In the first two cases all the options are diagonalizable, so does it mean all the matrices who square to a diagonalizable matrix are diagonalizable?
I suspect that the first case is supposed to refer to diagonalizable matrices with equal eigenvalues. If so, then first two statements regarding diagonaliziable matrices are correct. A correct version of the third case is as follows.
Proof: For the case that $A$ has zero as an eigenvalue, suppose for the sake of contradiction that there exists a matrix $B$ such that $B^2 = A$. Because all eigenvalues of $A$ must be zero, all eigenvalues of $B$ must also be zero. By the Cayley-Hamilton theorem, this implies that $B^2 = 0$. Thus, we have $A = B^2 = 0$, contradicting the fact that $A$ was non-diagonalizable.
For the case that $A$ has an eigenvalue $\lambda \neq 0$: the existence of Jordan form implies that there exists a matrix $N\neq 0$ such that $$ A = \lambda I + N $$ where $I$ denotes the identity matrix. Let $B$ be a matrix such that $B^2 = A$. It follows that $$ BA = BB^2 = B^3 = B^2B = AB. $$ It can be shown that if $BA = AB$, there must exist $p,q\in \Bbb C$ such that $$ B = pI + qN. $$ With that, we have $$ B^2 = A \implies p^2 I + 2pq N = \lambda I + N \implies \begin{cases} p^2 = \lambda \\ 2pq = 1 \end{cases} \implies\\ p = \pm \sqrt{\lambda}, \quad q = \frac 1{2 p}. $$ Thus, $A$ indeed has exactly $2$ square roots.
Proof: Because $N$ is non-zero nilpotent, there exists a basis $\{v_1,v_2\}$ of $\Bbb C$ such that $Nv_1 = 0$ and $Nv_2 = v_1$ (the Jordan basis of $N$). From the fact that $AB = BA$, conclude that $NB = BN$.
First, write $Nv_1 = pv_1 + kv_2$ for some $p,k \in \Bbb C$. We have $$ (NB)v_1 = N(pv_1 + kv_2) = kv_1,\\ (BN)v_1 = B(0) = 0. $$ Because $NB = BN$, we have $k = 0$.
Similarly, write $N v_2 = qv_1 + (p+k) v_2$. Note that $$ (NB)v_2 = N(qv_1 + (p+k)v_2) = (p+k)v_1,\\ (BN)v_2 = Bv_1 = pv_1. $$ Thus, we must have $k = 0$. That is, there exist $p,q \in \Bbb C$ such that $Bv_1 = pv_1$ and $Bv_2 = qv_1 + p v_2$.
It follows that $B = p I + qN$.