2.8 Subspaces of R

Column Space

Theorem 1 (Theorem 13) The pivot columns of a matrix \(A\) form a basis for the column space of \(A\).

NoteProof (Proof of Theorem 1)

Proof (Proof of Theorem 1). Suppose \(B\) is the RREF of \(A\), then there’s a sequence of elementary matrices, \(E_1,\dots,E_p\), where \(E_p\cdots E_1 A=B\). Let \(E=E_p\cdots E_1\), so we have \(EA=B\).

Denote the columns of \(A\) and \(B\) as \(\mathbf{a}_k\) and \(\mathbf{b}_k\), respectively. The linear relationship in the columns of \(A\) is preserved in the columns of \(B\) because \[\begin{align} EA &= E\begin{bmatrix}\mathbf{a}_1 & \cdots & \mathbf{a}_n\end{bmatrix} \\ &= \begin{bmatrix}E\mathbf{a}_1 & \cdots & E\mathbf{a}_n\end{bmatrix} \\ &= \begin{bmatrix}\mathbf{b}_1 & \cdots & \mathbf{b}_n\end{bmatrix} \end{align}\] Thus, if \(\mathbf{a}_k = \sum_{j\neq k} c_j\mathbf{a}_j\), \(\mathbf{b}_k=E\mathbf{a}_k=E\sum_{j\neq k} c_j \mathbf{a}_j=\sum_{j\neq k} c_j E\mathbf{a}_j=\sum_{j\neq k} c_j\mathbf{b}_j\).

Conversely, since \(E\) is invertible, we can prove the linear relationship in the columns of \(B\) is preserved in the columns of \(A\) by replacing \(E\) in the above proof with \(E^{-1}\) and swapping \(A\) and \(B\).

Therefore, we have proved the linear relationship is the same in both \(A\) and \(B\).

TipRemark

Remark. It’s easy to show that every non-pivot column can be written as a linear combination of the prior pivot columns by the definition of RREF, so we can safely remove the non-pivot columns from the set of the columns without changing the span. Thus, we have completely proved Theorem 1.

Although the pivot columns of \(A\) form a basis for \(\operatorname{Col} A\), it’s not the only basis that can be formed from the columns. If we do row reduction in a different order of the columns, the pivot columns may be different.

TipRemark

Remark. It may be a bit confusing why we need to prove both directions here. We can formalize this proof using sets. Let \(U\) be the set that represents the linear relationship in \(A\), then \(U=\{(k,c_1,\dots,c_n):\mathbf{a}_k=c_1\mathbf{a}_1+\dots+c_{k-1}\mathbf{a}_{k-1}+c_{k+1}\mathbf{a}_{k+1}+\dots+c_n\mathbf{a}_n\}\). In the same way, we define \(V\) to be a set containing the linear relationship in \(B\).

The direction that is directly written in the above proof shows that for every combination \(t\in U\), \(t\in V\), so \(U\subseteq V\). The opposite direction shows that \(V\subseteq U\), so together we show that \(U=V\).

Null Space

NoteDefinition (Null Space)

Definition 1 (Null Space) The null space of a matrix \(A\) is the set \(\operatorname{Nul} A\) of all solutions of the homogeneous equation \(A\mathbf{x}=\mathbf{0}\).

Proposition 1 Row operations don’t change null space.

NoteProof (Proof of Proposition 1)

Proof (Proof of Proposition 1). This is how we solve equations.

Proposition 2 Row operations don’t change row space, and the non-zero rows of the REF of a matrix form a basis of the row space.

Proposition 3 By some observation, \(\operatorname{Nul} A=\bigcap_{i=1}^m \operatorname{Nul}\operatorname{row}_i(A)\).

NoteProof (Proof to Proposition 3 (Informal))

Proof (Proof to Proposition 3 (Informal)). The solution to the system of linear equations is also the solution to each linear equation.

Proposition 4 Each row is orthogonal to the null space of the row. The row space is orthogonal to the null space.

NoteProof

Proof. Since \(A\mathbf{x}=\mathbf{0}\), by the rule of matrix multiplication, \(\operatorname{row}_i(A)\mathbf{x}=0\) for all \(1\leq i\leq m\). Thus, each vector \(x\) in \(\operatorname{Nul} A\) is orthogonal to the row space.

The first sentence is a special case of the second.

TipRemark

Remark. Let’s construct a basis for the null space of a row of \(A\).

The null space of the \(i\)-th row is \(\{(x_1,\dots,x_n):r_1x_1+\dots+r_nx_n=0\}\), where \(r_j\) is the \(j\)-th column of the \(i\)-th row of \(A\).

Set \(x_2=1\) and \(x_3,\dots,x_n=0\), we have \((-\frac{r_2}{r_1},1,0,\dots,0)\) as one vector in the basis.

Similarly, we can obtain the other \(n-2\) vectors in the basis.

\[(-\frac{r_3}{r_1},0,1,0,0,\dots,0)\] \[(-\frac{r_4}{r_1},0,0,1,0,\dots,0)\] \[\cdots\] \[(-\frac{r_n}{r_1},0,0,0,0,\dots,1)\]

Linear Transformation

TipRemark

Remark. Consider a linear transformation \(T: \mathbb{R}^n \to \mathbb{R}^m\), where \(T(\mathbf{x})=A\mathbf{x}\). \(\operatorname{Col}A\) is the range of the transformation, and \(\operatorname{Nul}A\) is the set of \(\mathbf{x}\in\mathbb{R}^n\) whose image under transformation \(T\) is \(\mathbf{0}\). Thus, for every \(T(\mathbf{u})=\mathbf{v}\) and \(\mathbf{p}\in\operatorname{Nul}A\), \(T(\mathbf{u}+\mathbf{p})=\mathbf{v}\).