Set of Matrices Over Finite Field Whose Pairwise Differences are Invertible Given the full matrix ring over a finite field, $M := M_{n \times n}(\mathbb{F}_q)$ for prime $q$ and integer $n$, what can one say about subsets $S$ of $M$ satisfying the condition that:
$A,B \in S$ implies $A - B$ is invertible unless $A = B$ (n.b. I'm not making any assumptions on the invertibility of elements of $S$ themselves). 
Is it possible to construct, or show the existence of, relatively large subsets satisfying this condition? I'm particularly interested in the case where $q$ scales asymptotically and $n$ is a fixed, small constant, and by large subsets I'm attempting to construct $S$ with $\vert S \vert \approx O(q^n)$, but suspect this is infeasible.
As for my own attempts, I thought initially that given the collection of all subspaces of $\mathbb{F}_q^n$ of dimension $n-1$, one ought to be able to choose one element from each subspace so that the difference of any two elements is full rank, and thus invertible. This would lead to a set $S$ of size ${n \choose n-1}_q = 1+ \dots + q^{n-1}$, but I am unable to prove that this is either achievable or impossible. 
Asymptotically, I've only been able to come up with obvious naive sets of size $q$, but playing around with small cases suggests that larger sets are possible in some cases, just they do not exhibit an obvious pattern (at least not obvious to me).
 A: The largest possible such a set $S$ has exactly $q^n$ elements:


*

*The extension field $L=\Bbb{F}_{q^n}$ has $q^n$ elements and can be embedded into $M_{n\times n}(\Bbb{F}_q)$ as a subring. The difference between any two distinct elements of $L$ then has an inverse in $L$.

*On the other hand, if $|S|\ge q^n+1$ then, by the pigeonhole principle, some two elements of $S$, say $A$ and $B$, will have the same first row. Their difference $A-B$ thus has all zeros on the first row, and cannot be invertible.



In small cases it is easy to describe an embedding (there are several). For example, if $n=2$ and $q$ is an odd prime we can find a quadratic non-residue $\epsilon\in\Bbb{F}_q$. Then the collection of matrices of the form
$$
M(a,b)=\left(\begin{array}{cc} a& \epsilon b\\ b&a\end{array}\right)
$$
forms a subfield $L=\Bbb{F}_{q^2}$. The determinant $\det M(a,b)=a^2-\epsilon b^2$ vanishes only when $a=b=0$. Furthermore, the difference of two such matrices is of the same form. This is a generalization of the well known way of representing complex numbers by $2\times 2$ real matrices:
$$a+ib\mapsto\left(\begin{array}{cc}a&-b\\ b&a\end{array}\right),$$
where we use $-1$ as a non-square.
When $n=3$, $p\neq3$ we can similarly use a non-cube $\epsilon$ and matrices of the form
$$\left(\begin{array}{ccc}
a&\epsilon c&\epsilon b\\
b&a&\epsilon c\\
c&b&a\end{array}\right).$$
For larger $n$ the method of describing an explicit set of matrices becomes a bit more complicated. If $m(x)$ is the minimal polynomial of a generator of the extension field, we can use $\Bbb{F}_q$-linear combinations of the powers $A^i$, $i=0,1,2\ldots,n-1$, of the companion matrix $A$ of $m(x)$.
