Lecture 18: Elementary Matrices
Recall three types of elementary row operations
- \(R_i \leftrightarrow R_j\)
- \(R_i \rightarrow \lambda R_i\) where \(\lambda \neq 0\)
- \(R_i \rightarrow R_i + \lambda R_j\)
Now that we’ve studied matrix multiplication we can state the fact that performing an elementary row operation on \(A \in M_{2 \times 2}\) can actually be described using matrix multiplication.
Elementary Matrices
Example
Applying the three types of elementary row operations results in the following matrices
This leads to the next theorem
Example
Let’s apply the elementary matrices on the following given matrix
RREF by Matrix Multiplication
Since we can perform elementary row operations by matrix multiplication, then we can possibly see how we can put a matrix in reduced row echelon by multiplication. But first, there is an observation
Proof: apply the elementary row operation by multiplying by \(E(\mathcal{R})\) and then apply the inverse again by multiplying by \(E(\mathcal{R}^{-1})\). The result is the identity matrix. In other words, \(E(\mathcal{R})E(\mathcal{R}^{-1}) = I_n\).
Note here that from the expression above we can see that \(A^{-1} = E_k...E_1\) and \(A = E_1^{-1}E_2^{-1}...E^{-1}_{k-1}E^{-1}_{k}\)
The Rank of a Matrix
Recall that the rank of \(T: V \rightarrow W\) is \(\dim(R(T))\).
This definition is kind of awkward and instead we want to find an expression for rank(\(A\)) in terms of \(A\) itself and not have to rely on \(L_A\). To figure this out, we need the following result
So multiplication by \(B\) doesn’t change the rank.
Proof:
By definition \(rank(BA)\) is the rank of the linear map \(L_{BA}\). But the rank of a linear map is the dimension of its range and so
By definition, to get the range, we just apply the linear map on \(v \in \mathbf{R}^n\). This range is a subset of \(\mathbf{R}^m\). Applying the linear map is just multiplying \(v\) by the matrices \(A\) and \(B\). So we can see this below:
This is equivalent to applying the linear map \(L_{B}\) on the set that we get from applying \(A\) as follows
But this internal set is the range of the linear map \(L_A\), \(R(L_A)\). So what we want to show is that applying \(L_B\) doesn’t change the dinemsion of \((R(L_A))\) or multiplying by the matrix \(B\) above, doesn’t change anything about the dimension of the internal set.
The idea is simple but subtle. We know that \(L_B\) is a map from \(\mathbf{R}^m\) to \(\mathbf{R}^m\). But here, \(L_B\) is not acting on \(\mathbf{R}^m\) but rather \(R(L_A)\) (the internal set)\(. The range of\)L_A\(is a subset of\)\mathbf{R}^m$$. So define the map
We claim this new map is invertible. It’s onto because we restricted the target to the image \(R(L_A)\) and it’s one-to-one because \(B\) is invertible. Therefore, the dimension of the domain and the target are the same (It’s an invertible map!). In other words,
But we know that \(\dim(R(L_A))\) is the rank of \(A\). and \(\dim(L_B(R(L_A))\) is the rank of \(L_{BA}\) so \(rank(A) = rank(BA) \ \blacksquare\).
The Rank of a Matrix
So now we can go back to our original goal of finding an expression for finding the rank of a matrix \(A\) without going back to the rank of the linear map \(L_A\). We just proved that \(\text{rank}(BA) = \text{rank}(A)\). for any invertible matrix \(B\). We further studied earlier that elementary matrices are invertible. So multiplying \(A\) by a set of elementary matrices will not change its rank. So we can get \(A\) in RREF without its rank changing. Based on this we have the following corollaries:
and
Why do we want RREF? because it’s easy to read off and we can easily figure out the dimension easily from seeing a matrix in its RREF.
Example
What is the range of the following matrix?
We know the rank of \(A\). It is the dimension of the range of \(L_A\) or \(\dim(R(L_A))\). Notice for the above matrix, if we multiply \(A\) by a vector \(v\), the result is a linear combination of the columns of \(A\). Recall
In other words, the columns span the range. Here, we see that we have 3 non-zero columns and they are linearly independent. so \(rank(A) = \dim(R(L_A)) = 3\). This works here but it’s not generally true!! we’re missing something … so let’s clarify with another example
Example
What is the range of the following matrix?
It’s still true that the range of this is still spanned by the columns (always true). But the rank of \(A\) in general is the number of columns with leading entries in \(REF\) of \(A\).
Nullity and The Dimension Theorem
So we know that \(\text{nullity}(A) = \dim(N(A))\). We found a basis for the null space of \(A\) by solving \(Ax = 0\) and finding all the solutions and then writing a set that spans that solution set. From the basis we knew the dimension of the null space. Specifically when we solved \(Ax = 0\), it was the number of columns without leading entries. So we can write \(\text{nullity}(A) = \dim(N(A)) =\) # of columns without leading entries.
So now if we put together the number of columns without leading entries (nullity of \(A\)) and the number of columns with leading entries (rank of \(A\)), then we get \(n\). This is basically the dimension theorem.
References
- Math416 by Ely Kerman