8
$\begingroup$

I've got two inverse problems,

$$A_1 ~ x = b_1 \qquad A_2 ~ x = b_2$$

So far I've been solving them independently using Tikhonov Regularization and getting two estimates for $x$. However in my case $x$ represents the same solution in both equations. Is it possible to do a 'simultaneous' solve? Ideally I would be finding the answer for

$$\min \left( \lVert A_1 x - b_1 \rVert^2 + \lVert A_2 x - b_2 \rVert^2 + \lVert\Gamma x\lVert^2 \right)$$

Where $\Gamma = \alpha ~ I$ and $I$ is the identity matrix as in Tikhonov Regularization (aka ridge regression). I suppose I could just take the average of both solutions, wondering if there is a more statistically powerful way of approaching this however.

$\endgroup$
  • 1
    $\begingroup$ What's the relative accuracy of the measurements in $b_{1}$ and $b_{2}$? You may need to scale to adjust for this. Are all the measurements independent? Correlation can complicate things. $\endgroup$ – Brian Borchers Aug 6 '16 at 21:44
  • $\begingroup$ Right now it's all modeled so I know $b_1$ and $b_2$ perfectly, but in practice I will know $b_1$ with perhaps 10x more accuracy. However at this step I want to assume I know them both equally and that they are independent. $\endgroup$ – abnowack Aug 6 '16 at 21:53
  • $\begingroup$ So what is your question here? It's easy to solve the three term least squares problem you've given in your question. $\endgroup$ – Brian Borchers Aug 6 '16 at 22:06
  • $\begingroup$ Is it? If you explain in an answer I'll mark it correct. I just use basic routines like numpy's least square solver . I'm not from a CS background, so I could be missing something obvious. $\endgroup$ – abnowack Aug 6 '16 at 22:24
9
$\begingroup$

You can write your problem as

$\min \| Fm - g \|_{2}^{2}$

where

$F=\left[ \begin{array}{c} A_{1} \\ A_{2} \\ \alpha I \\ \end{array} \right] $

and

$g=\left[ \begin{array}{c} b_{1} \\ b_{2} \\ 0 \\ \end{array} \right]. $

You can use whatever linear least squares solver you want to solve this problem.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.