Let $n \in \mathbb{N}$, and $U,V$ two linear subspaces of $\mathbb{R}^n$ of the same dimension. Could one always make a matrix $A \in \mathbb{M}^{n \times n}(\mathbb{R})$ such that $spanA = U$ and $spanA^T = V$?

I find it hard to find a counterexample. I know that for any $n \in \mathbb{N}$, there is such a matrix if $\dim U = \dim V \in \{0,n\}$, we could just take $0,I$ as matrices. Moreover, if $U,V$ are lines, that means $U = \mathbb{R}v$ and $V = \mathbb{R}w$ for some nontrivial vectors $v,w$, then the matrix below works.

$$ \left( \ w_1v \ | \ w_2v \ | \ \cdots \ | \ w_nv \ \right) $$ Here $w_i$ are coordinates of the vector $w$, and each entry represents a column this way.

Then I took two subspaces of dimension two in $\mathbb{R}^3$, namely $$ U \quad = \quad span \left\{ \left( \begin{array}{ccc} 1 \\ 0 \\ 0 \end{array} \right) , \left( \begin{array}{ccc} 0 \\ 1\\ 0 \end{array} \right) \right\} \qquad V \quad = \quad \left( \begin{array}{ccc} 0 \\ 1\\ 0 \end{array} \right)^\perp $$ The matrix we need here is $$ \left( \begin{array}{ccc} 1 & 0 & 0 \\ -1 & 1 & 0 \\ 0 & -1 & 0 \end{array} \right) $$

I find it hard to deal with the general case. Could you give me some help to establish the statement, or give a counterexemple?


Let $m$ be the dimension of $U$ and $V$. Let $M_U$ be an $n\times m$ matrix whose columns are a basis of $U$ and $M_V$ be an $n \times m$ matrix whose columns are a basis of $V$. Now $U = \{M_U \vec{x} : \vec{x} \in \mathbb{R}^m \}$ and $V = \{M_V \vec{x} : \vec{x} \in \mathbb{R}^m \}$.

We wish to find a matrix $A$ such that $$ \begin{align*} U &= \{M_U \vec{x} : \vec{x} \in \mathbb{R}^m \} = \{A \vec{x} : \vec{x} \in \mathbb{R}^n \}, \\ V &= \{M_V \vec{x} : \vec{x} \in \mathbb{R}^m \} = \{A^T \vec{x} : \vec{x} \in \mathbb{R}^n \}. \end{align*} $$

We note that $\{ M_V^T \vec{x} : \vec{x} \in \mathbb{R}^n\} = \mathbb{R}^m$, because the rank of $M_V^T$ is $m$ and it has $m$ rows. Thus $$ U = \{M_U \vec{x} : \vec{x} \in \mathbb{R}^m \} = \{M_U M_V^T \vec{x} : \vec{x} \in \mathbb{R}^n \}. $$ But this immediately suggests that $A=M_U M_V^T$, because $\mathrm{span}(M_U M_V^T) = U$. By symmetry, also $\mathrm{span}(M_V M_U^T) = V$, and because $A^T=M_V M_U^T$, we are done.

  • $\begingroup$ I don't understand what you mean by $M_U^T$, could you please tell me? $\endgroup$ Aug 11 '14 at 21:02
  • $\begingroup$ @KoenraadvanDuin Transpose of the matrix $M_U$. $\endgroup$
    – JiK
    Aug 11 '14 at 21:06
  • $\begingroup$ right, I forgot about that. $\endgroup$ Aug 11 '14 at 21:18

This is indeed always possible. Let $U$ be some $r$-dimensional subspace of $\mathbb{R}^n$. Let $R$ denote any matrix with rowspace $U$. Since $U$ is $r$-dimensional, it follows that the rank of $R$ is $r$ and so the image of $R$ is also $r$-dimensional.

Let $\mathcal{B}'$ denote a basis for the image of $R$ and extend $\mathcal{B}'$ to a basis $\mathcal{B}$ of $\mathbb{R}^n$. Form the matrix $P$ with columns given by $\mathcal{B}$, with the first $r$ columns consisting of the vectors of $\mathcal{B}'$. It follows that $P$ maps the subspace $S$ spanned by $\left\{\mathbf{e}_1,\ \cdots, \mathbf{e}_r\right\}$ to $\mathrm{im}(R)$ and therefore $P^{-1}$ maps $\mathrm{im}(R)$ to $S$.

Now let $\mathcal{C}'$ denote a basis for $V$ and extend $\mathcal{C}'$ to $\mathcal{C}$. Form the matrix $Q$ with the basis as the columns, with the first $r$ columns consisting of the vectors of $\mathcal{C}'$. It follows that $Q$ takes $S$ to $V$.

Then the matrix $QP^{-1}R$ is your desired matrix. Since $QP^{-1}$ is invertible, the rowspace of $R$ is unchanged, therefore $\mathrm{row}(QP^{-1}R) = U$. Also, $P^{-1}$ maps $\mathrm{im}(R)$ to $S$ and $Q$ maps $S$ to $V$. Therefore $\mathrm{im}(QP^{-1}R) = V$, as required.

  • $\begingroup$ I understand that a map $X \mapsto AX$ does not change the kernel of $X$ if $A$ is invertible, but I don't see how that should work for the rowspace of $X$. Could you explain me? $\endgroup$ Aug 11 '14 at 20:37
  • $\begingroup$ @KoenraadvanDuin There are two immediate ways to see this. First you note that the kernel is invariant under $X\mapsto AX$. The rowspace is the orthogonal complement of the kernel, hence it is also left invariant. Alternatively, if $A$ is invertible then it can be written as a product of elementary matrices, so the overall effect of $A$ is to perform a sequence of elementary row operations on $X$, which does not change the rowspace. $\endgroup$
    – EuYu
    Aug 11 '14 at 23:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.