What is the rank of $\sum_{i=1}^k \vec{u}_i \vec{v}_i^T$ with $\{ \vec{u}_i \}$ and $\{ \vec{v}_i \}$ each linearly independent.

67 Views Asked by At

Let $A = \sum_{i=1}^k \vec{u}_i \vec{v}_i^T$, where $\vec{u}_i \in \mathbb{R}^m$ and $\vec{v}_i \in \mathbb{R}^n$. Assume that $\{ \vec{u}_i \}_{i=1}^k$ are linearly independent and $\{ \vec{v}_i \}_{i=1}^k$ are linearly independent. Assume $m > n > k$.

What is the rank of $A$?

Let $A_i = \vec{u}_i \vec{v}_i^T$ so that $A = \sum_{i=1}^k A_i$.

Column $j$ of $A_i$ is $(\vec{v}_i)_j \vec{u}_i$. The rank of each $A_i$ is one since all columns are multiples of a single vector $\vec{u}_i$.

Column $j$ of $A$ is $\sum_{i=1}^k (\vec{v}_i)_j \vec{u}_i$. It seems the columns of $A$ may or may not be linearly independent. Is that correct? Then the rank can be from $1$ to $k$. The wording of the question suggests that the rank of $A$ is $k$ but I don't see why this is necessarily the case.

2

There are 2 best solutions below

0
On BEST ANSWER

On the one hand, for any $v \in \mathbf{R}^m$, we have that $$ Av = \sum_{i=1}^k (v_i^\mathrm{T}v) u_i \in \mathrm{span}\{u_1, \dotsc, u_k\}, $$ so that $\mathrm{Im}(A) \subseteq \mathrm{span}\{u_1, \dotsc, u_k\}$

On the other hand, let $u = \sum_{i=1}^k a_i u_i$ for $a_i \in \mathbf{R}$. Let $V$ be the $k\times n$ matrix whose $i$th row is $v_i^{\mathrm{T}}$, Since the $v_i$ are linearly independent, this matrix has rank $k$ and so is surjective. In particular, there is a $w\in \mathbf{R}^m$ that solves $Vw = a$, where $a$ is the vector whose $i$th entry is $a_i$. Equating entries, this means that $(Vw)_i = v_i^{\mathrm{T}} w = a_i$

But then, we have that $$ Aw = \sum_{i=1}^k (v_i^\mathrm{T}w) u_i = \sum_{i=1}^k a_i u = u, $$ so that $\mathrm{span}\{u_1, \dotsc, u_k\} \subseteq \mathrm{Im}(A)$.

Hence, the rank of $A$ is $$ \mathrm{\dim}\, \mathrm{Im}(A) = \mathrm{\dim}\, \mathrm{span}\{u_1, \dotsc, u_k\}, $$ which is $k$ since th $u_i$ are linearly independent.

0
On

The outer product $(u,v)\mapsto uv^T$ makes $\mathbb{R}^{m\times n}$ a tensor product of $\mathbb{R}^m$ and $\mathbb{R}^n$. Tensor rank in this context is just matrix rank. But the tensor rank of a sum of tensor products of linearly independent vectors is just the number of summands, which in this case is $k$.