How to construct a matrix with one dimension nullspace?

22 Views Asked by At

Suppose we have a fat matrix $\boldsymbol{E} \in \mathbb{R}^{p \times t}$ and we conduct SVD as follows \begin{eqnarray} \boldsymbol{E} = \boldsymbol{U}\boldsymbol{\Sigma}\boldsymbol{V}^\top. \end{eqnarray} We extract the first q columns to build a basis matrix $\boldsymbol{B} \in \mathbb{R}^{p \times q}$. Based on this basis matrix, we want to obtain a non-zero solution of a homogeneous linear system $\boldsymbol{Dx} = \boldsymbol{0}$, \begin{eqnarray} \label{eq:solve_linear} \underbrace{ \left[ -\rm{diag}(\boldsymbol{m})\boldsymbol{B} | \boldsymbol{K} \right]}_{\boldsymbol{D} \in \mathbb{R}^{p \times (q + 3)}} \boldsymbol{x} = \boldsymbol{0}, \end{eqnarray} where $\rm{diag}(\cdot)$ denotes diagonal operator, $\boldsymbol{K} \in \mathbb{R}^{p \times 3}$ has full column rank, $p > q + 3$. Also $\boldsymbol{m} \in \mathbb{R}^{p}$ is a non-zero vector.

The necessary condition to obtain a non-trivial solution for $\boldsymbol{Dx} = \boldsymbol{0}$ is ${\rm null}(\boldsymbol{D}) = 1$, i.e. $\boldsymbol{D}$ has $1d$ nullspace.

Question:

Given $\boldsymbol{U}, \boldsymbol{L}, \boldsymbol{m}$, can we automaticaly determine q such that ${\rm null}(\boldsymbol{D}) = 1$?

Note

One direct method is gradually increasing q from $1$ to $p$ until ${\rm rank}(\boldsymbol{D}) = q+3-1$. However, this strategy is time-consuming, I wonder if there is a smart method to solve this problem?