Rank-one matrices: a representation theorem

We prove the theorem mentioned here:

Theorem: outer product representation of a rank-one matrix

Every rank-one matrix A \in \mathbb{R}^{m \times n} can be written as an ‘‘outer product’’, or dyad

    \[A=pq^{T},\]

where p \in \mathbb{R}^{m}, q \in \mathbb{R}^{n}.

Proof: For any non-zero vectors p \in \mathbf{R}^m, q \in \mathbf{R}^n, the matrix p q^T is indeed of rank one: if x \in \mathbf{R}^n, then

    \[p q^T x=\left(q^T x\right) p .\]

When x spans \mathbf{R}^n, the scalar q^T x spans the entire real line (since q \neq 0), and the vector \left(q^T x\right) p spans the subspace of vectors proportional to p. Hence, the range of p q^T is the line

    \[\mathbf{R}\left(p q^T\right)=\{t p: t \in \mathbf{R}\},\]

which is of dimension 1.
Conversely, if A is of rank one, then its range is of dimension one, hence it must be a line passing through 0 . Hence for any x there exist a function t: \mathbf{R} \rightarrow \mathbf{R} such that

    \[A x=t(x) p .\]

Using x=e_i, where e_i is the i-th vector of the standard basis, we obtain that there exist numbers t_1, \ldots, t_n such that for every i :

    \[A e_i=t_i p, \quad i=1, \ldots, p .\]

We can write the above in a single matrix equation:

    \[A\left(\begin{array}{ccc} e_1 & \ldots & e_n \end{array}\right)=p\left(\begin{array}{ccc} t_1 & \ldots & t_n \end{array}\right) .\]

(Indeed, the previous equation is the above, written column by column.)
Now letting q=\left(t_1, \ldots, t_n\right) \in \mathbf{R}^n, and realizing that the matrix \left(e_1, \ldots, e_n\right) is simply the identity matrix, we obtain

    \[A=p q^T,\]

as desired.

License

Hyper-Textbook: Optimization Models and Applications Copyright © by L. El Ghaoui. All Rights Reserved.

Share This Book