1: \begin{proof} The idea is that determinant zero is a codimension-one condition. To show it rigorously,

2: we prove the following stronger result:  for fixed $m$, we will show that 

3: $\Pbar(\det M_\ell^{(k)}=0)=0$ for each $1\le k\le m$.  For $k=1$, we note that $M_\ell^{(1)}=\hat A_{\ell,i}$, 

4: so the statement follows from Lemma~\ref{c}.  Let's show that if it is true for $M_\ell^{(k-1)}$, then 

5: it is true for $M=M_\ell^{(k)}$.

6: Let us write $q_\ell$ to denote the lower-right entry of $M_\ell^{(k)}$ and 

7: $\mu_{\ell}$ to denote the list of the other  $k^2-1$ entries $(M_{1,1}, \dots, M_{k,k-1})$.  

8: The induction hypothesis tells us the probability that $\det N=0$ tends to zero

9: for $N$ the upper left-hand  $k-1 \times k-1$ minor.

10: Assuming that minor is nonsingular, there is exactly one value 

11: of $q_\ell$ making $\det M_\ell^{(k)}=0$ for each $\mu$; call it $q(\mu)$.  

12: But, recalling that $0$ is the most likely

13: value for $q_\ell$ and that the different $\mu_\ell=\mu$ are disjoint events, we have

14: $$\Pr(\det M_\ell^{(k)}=0)=\sum_\mu \Pr(q_\ell=q(\mu)) \le \sum_\mu \Pr(q_\ell=0)=\Pr(q_\ell=0).$$

15: But $q_\ell$ is distributed like $\hat A_{\ell,i}$, so by Lemma~\ref{c}, this tends to zero. 

16: \end{proof}

17: