Upper bound on the norm of a solution to a linear system

768 Views Asked by At

I have a linear system $Ax = b$ where $A$ is an $m \times n$ integral matrix and I know there exists at least one solution. I want an upper bound on the $\ell_2$ norm $\|x\|$ such that $\|x\|$ is minimized over all possible solutions. To state the problem differently I would like to find an $x$ such that $\|x\|$ has a nice upper bound.

Preferably, I would like the bound to be in terms of $m,n,b$ and properties of $A$ such as its singular values or determinants of square submatrices. However, I'd be interested in seeing bounds involving other values.

1

There are 1 best solutions below

6
On BEST ANSWER

(I don't have enough reputation to comment, so here's one short example offered as an answer: Suppose A is a matrix that projects 3d space onto the x-axis. $\text{diag}\,(1,0,0)$, if you will. Say $b=(5,0,0)^T$. Then any $x$ of the form $(5,p,q)^T$ will satisfy your equation, no matter how huge $p$ and $q$ are. The solutions seem to be unbounded. You can probably formalize this with the rank-nullity theorem or something. If $A$'s range has the same dimension as the dimension of $x$, then there's precisely one solution, which is boring. Otherwise, if $A$ is 'lossy' in some sense, then there are some subspaces of the domain that $A$ is collapsing to zero, and $x$ can have as big of a component it wants in these subspaces.)

Aight, ignore all that, I think I got it!

So basically, if we want $x$ to be as small as possible, let's make sure it has no components in $A$'s nullspace (because any such component would add to $x$'s size unnecessarily). So, $x$ should be orthogonal to the nullspace of $A$, i.e. in the rowspace of $A$, i.e. i the columnspace of $A^T$.

Now, consider all $y$ that satisfy $Ay=b$. If we project such a $y$ down to the columnspace of $A^T$, we are done. Fortunately, we have a simple tool to project stuff to columnspaces of matrices - this is where the Moore-Penrose pseudoinverse comes in. The wikipedia page on projection matrices has some nice intuition on this.

Okay, so we get that $x=A^T(AA^T)^{-1}b$. How to find an upper bound on size? $$\|x\|^2=b^T(AA^T)^{-1^T}A\,A^T(AA^T)^{-1}b=b^T(AA^T)^{-1^T}b=\langle b, (AA^T)^{-1^T}b\rangle$$ So by Cauchy-Schwarz we get $$\|x\|^2\leq\|b\|\,\|(AA^T)^{-1^T}b\|=\|b\|^2\,\|(AA^T)^{-1^T}\hat{b}\|$$ Where $\hat{b}=b/\|b\|$. Now notice that in the end we're basically looking at the operator norm for $(AA^T)^{-1^T}$, which is bounded above by the biggest singular value. And we also know that transposing doesn't change the singular values, and taking the inverse also inverts the singular values, so $$\|x\|^2\leq\|b\|^2\cdot\frac{1}{\text{smallest singular value of }AA^T}$$