This question originates from an engineering problem, which I am solving. Any related references are highly appreciated.

Let $M_k(T)=\prod_{t=1}^T P_t S_k$ over some field (finite or reals), where $P_t$ is a random (uniform, i.i.d. over $t$) $n\times n$ permutation matrix, and $S_k$ is a fixed idempotent matrix of rank $n-1$ of the following form

$$ S_1=\begin{pmatrix} 1 & 0 & 0 \cdots \\ 1 & 0 & 0 \cdots \\ 0 & 0 & \\ \vdots & \vdots & \huge{I_{n-2}} \\ 0 & 0 & \end{pmatrix}, S_2=\begin{pmatrix} 1 & 0 &0 & 0 \cdots \\ 0 & 1 &0& 0 \cdots \\ 1 & 1 &0& 0 \cdots \\ 0 & 0 &0& \\ \vdots &\vdots & \vdots & \huge{I_{n-3}} \\ 0 & 0 &0& \end{pmatrix}, \\ S_k=\begin{pmatrix} & &&0 & \\ & \huge{I_k} &&\vdots& \huge{0} \\ & &&0& \\ 1 & \cdots &1&0& 0 \cdots \\ & & &0& \\ & \huge{0}& & \vdots & \huge{I_{n-k-1}} \\ &&&0& \end{pmatrix}.$$

Essentially, each additional multiplication by $P_t S_k$ replaces some row of $M_k(t-1)$ with a linear combination of $k$ other rows. It is not hard to see that $\mbox{rank } M_k(t)$ is non-increasing in $t$, and $\lim_{T\to \infty} \mbox{rank } M_k(T) \leq k$ almost surely. I am looking into the rate of this convergence.

Question: find $E[\mbox{rank } M_k(T)]$ or the CDF of $\mbox{rank } M_k(T)$ for $T\gg n$.

One approach to this is to track the number $N_0(M(t))$ of zero-columns or the number $N(M(t))$ of rows $r$, s.t. $r$ contains the first row of $M(1)$ in its linear decomposition. These numbers follow the Markov property, and one can construct corresponding linear Markov chains with $n+1$ states to find their long-term behavior and bound

$$\mbox{rank } M(T) \leq n-N_0(M(T)) \\ E[\mbox{rank } M(T)] \leq n \Pr[N(M(T))>0] $$

Although the bounds are tight for $k=1$, they are very loose for $k>1, T\gg n$. In fact for $k=1$ the rank drops to $1$ rather quickly in $T = O(n^2)$, but for $k>1$, the decay seems to be exponentially slow, e.g. $E[\mbox{rank } M_3(10^5)]\approx E[\mbox{rank } M_3(10^7)] \approx 5$ for $n=20$ (numerically).

Another numerical observation shows that for $n\ge 10$ the resulting $E[\mbox{rank } M(T)]$ does not depend much on the field size once it is prime and larger than $7$.

Any reference to studying the long products like $M(t)$ of random permutations with a fixed matrix $S$ would be also appreciated.