# Perturbation theory for least squares for very different A, b

Consider the least squares problem

$f(x;A,b) = \|Ax-b\|_2^2$

and define $x^*$ the minimizer of $f(x;\hat A,\hat b)$, and $\hat x$ the minimizer of $f(x; A_2, b_2)$.

I want to put some bound on $\|x^* - \hat x\|$.

Looking through Golub/Van Loan, I see a lot of stuff that is basically some function of $\epsilon = \max\{\|A-\hat A\|, \|b-\hat b\|\}$, but in some sense that's not the best you can do. For example, if

$A =\left[\begin{matrix} 0 \\ C\end{matrix}\right], \hat A =\left[\begin{matrix} C\\ 0\end{matrix}\right], b =\left[\begin{matrix} 0 \\ d\end{matrix}\right], \hat b =\left[\begin{matrix} d\\ 0\end{matrix}\right]$

then $\epsilon = \max\{2\|C\|, 2\|d\|\}$ which may be very large, but $\|x^*-\hat x\| = 0$.

Are there existing bounds that take this into account? It has to be some bound on some function that involves $a_i$ AND $b_i$ (for $A= [a_1^T, ...]$ and $b = [b_1;...]$) and not just some stuff on range space of $A$.

I suspect there is some result in machine learning, since this is basically about regression solutions if you get enough "important" samples. Can anyone point me to any known results?

Thank you!

• It seems like a "small" perturbation of $A$ could change $x^*$ significantly. eg take $A=\mathrm{diag}(1,\epsilon)$ and $b=(0,1)^T$ where $\epsilon$ is small. – stewbasic Aug 29 '16 at 4:15
• Definitely, adversarial changes in A and b can drastically change $x^*$. The full form of the Golub/van Loan type bounds depend also on the condition number of $A$, which in your example would be very large. I guess I'm looking for the opposite; what kind of assumptions can I impose on these perturbations so that the result is very small? I played around a bit with SVD terms, which is somewhat illuminating; by bounding $b^TU-\hat b^T\hat U$ and other terms you can get something. I just feel like there should be an established result somewhere! – Y. S. Aug 29 '16 at 4:37
• You may use componentwise perturbation bounds from Higham, Accuracy and Stability of numerical algorithms. These bounds are better, than given by Golub/van Loan and allow for imposing additional constraints on perturbations. However you consider huge modifications and I don't thing, that any perturbation results will be useful for this problem. – Pawel Kowal Aug 29 '16 at 6:51
• Hmm thank you for the reference! Perhaps I am asking the wrong question. Should be more related to subsampling from structured distributions... – Y. S. Aug 30 '16 at 15:25