A sufficient condition for non-negative solutions to non-negative linear systems
Here we are concerned with “non-negative” linear systems, that is, linear systems where $\mathbf{A}, \mathbf{b} \geq \mathbf{0}$ (elementwise). In particular, we give a sufficient condition for non-negativity of the solution $\mathbf{x}$ to $\mathbf{Ax} = \mathbf{b}$.
First, we discuss an intuition for the result. The columns of $\mathbf{A}$ form a linear cone; if a vector $\mathbf{c}$ (a right hand side to the linear system) is known to lie in the cone, then one can find a “subcone” containing $\mathbf{c}$. The sufficient condition essentially upper bounds the “radius” of this subcone.
Theorem: let $\mathbf{A} \in \mathbb{R}^{n \times n}$ and $\mathbf{b}$ be an invertible matrix and a vector, both non-negative. Futher, let $\mathbf{A}$ satisfy the following form of diagonal dominance by rows
$$ 0 \leq \alpha \leq |a_{ii}| - \sum_{j} |a_{ij}| \quad \forall i . $$
Then
$$ \exists \mathbf{c} : \mathbf{Ay} = \mathbf{c}, \; \mathbf{y} \geq \mathbf{0}, \; \mathbf{c} \geq \mathbf{0}, \; ||\mathbf{b} - \mathbf{c}||_\infty \leq \min\left( \alpha \min(|\mathbf{y}|), \min(\mathbf{|c|}) \right) \\ \quad \Rightarrow \quad \exists \mathbf{x} : \mathbf{x} \geq \mathbf{0}, \mathbf{Ax}=b $$
In particular, one could apply this theorem to the case when $\mathbf{c} = \mathbf{A}\mathbb{1}$, a point expected to be relatively deep in the interior of the cone (assuming all columns normalized). Contrast with the case when $\mathbf{c} = \mathbf{Az}$, where $\mathbf{z}$ has a 0 in one or more elements, meaning $\mathbf{c}$ would be on a face of the cone (and hence the above theorem would predict zero for the radius of the subcone).
Proof
We have $\mathbf{Ay}=\mathbf{c}$, now consider $\mathbf{A}(\mathbf{y} + \delta\mathbf{y}) = \mathbf{c} + \delta\mathbf{c}$ and define $\mathbf{x} = \mathbf{y} + \delta\mathbf{y}$ and $\mathbf{b} = \mathbf{c} + \delta\mathbf{c}$. We will show that given the assumptions above $\mathbf{x},\mathbf{b} \geq \mathbf{0}$.First, we consider $\mathbf{b}$. Since $||\mathbf{b} - \mathbf{c}||_\infty = ||\delta\mathbf{c}||_\infty \leq \min(|\mathbf{c}|)$ then $\mathbf{b} = \mathbf{c} + \delta\mathbf{c} \geq \mathbf{0}$.
Now, we consider $\mathbf{x}$. Note that $\mathbf{A} \delta\mathbf{y} = \delta\mathbf{c} $, or equivalently $\delta\mathbf{y} = \mathbf{A}^{-1} \delta\mathbf{c} $. By norm submultiplicativity, $||\delta\mathbf{y}||_\infty \leq ||\mathbf{A}^{-1}||_\infty ||\delta\mathbf{c}||_\infty$. Now since $||\mathbf{b} - \mathbf{c}||_\infty = ||\delta\mathbf{c}||_\infty \leq \alpha\min(|\mathbf{y}|)$, we have $||\delta\mathbf{y}||_\infty \leq \alpha ||\mathbf{A}^{-1}||_\infty \min(|\mathbf{y}|)$. Finally, via Varah's upper bound for the $\infty$-norm of the inverse of a row diagonal dominant matrix we have $\alpha ||\mathbf{A}^{-1}||_\infty \leq 1$. Therefore, $||\delta\mathbf{y}||_\infty \leq \min(|\mathbf{y}|)$, so $\mathbf{x} = \mathbf{y} + \delta\mathbf{y} \geq \mathbf{0}$.
Comments
Comment 1: There exist other upper bounds for the $\infty$-norm of the inverse of a diagonally dominant matrix. Specifically, if $\mathbf{A}$ instead satisfies $|a_{ii}| \geq \gamma \sum_{i \neq j} |a_{ij}| \; \forall i$, then $||\mathbf{A}^{-1}||_\infty \leq \frac{\gamma}{\gamma-1} \frac{1}{\min_i |a_{ii}|}$.
Comment 2: The proof does not actually require $\mathbf{A} \geq 0$, so perhaps this assumption could strengthen the conclusion. Likewise, for a symmetry assumption if one is added.