A few years ago, I came across the following minimization formulation

$$J = \arg \min_u \{\|s-u\|+\lambda\|u-L\left(u\right)\|\}$$

$s$ is the measurement, $\lambda$ is the regularization parameter, and $L\left(u\right)$ is a (local) convolution with adaptive coefficients in its kernel, which will be found iteratively. After these coefficients are updated, $u$ will also be updated by minimizing $J$. These steps will be looped until convergency is reached in both the coefficients of $L$ and the solution $u$.

Unfortunately, I can't find the right keywords to search for the method. Could someone help me find a paper, webpage, the method's name, or any hint so that I can read and study the method in more detail? Thank you.

  • 3
    $\begingroup$ This bears a faint similarity to a blind image deblurring problem, but it's strange that the regularization term encourages $u \approx L(u)$. $\endgroup$ – littleO Oct 13 '17 at 0:12
  • 2
    $\begingroup$ Agreed, that's a very interesting choice. I'm wondering if it perhaps it is meant to reduce overshoot. $\endgroup$ – Michael Grant Oct 13 '17 at 22:34
  • 2
    $\begingroup$ Well, you're still applying an operator to $u$, namely $I-L$, so it's not really so unusual $\endgroup$ – Jason Born Aug 8 '18 at 22:12

Consider \begin{align} Ax = u \end{align} Because if \begin{align} L = \frac{Axx^TA^T}{x^TA^TAx} \end{align} We see that \begin{align} \left\lVert \left(I-L\right)u \right\rVert &= \left\lVert \left(I-\frac{Axx^TA^T}{x^TA^TAx}\right)u \right\rVert\\ &=\left\lVert\left(I-\frac{uu^T}{\left\lVert u\right\rVert^2}\right)u \right\rVert\\ &=0 \end{align}

Here is where it would be useful: We wish to describe $u$ by a weighted superposition of model matrices $A_i$ via parameters $x$, i.e. \begin{align} u &= B_ax\\ &= \left(\sum a_i A_i\right) x \end{align} Then optimizing over both $x$ and $a_i$ means that you are asking the question:

"Given the data, what parameters $x$ do I have to estimate, and what model $B_a$ do I have to choose to describe the data the best"

If $s$ can be described by $B_ax$ completely, then we will also have $s=u$ in the end. I guess the regularization parameter doesn't really serve a purpose in this case. Correct me if I'm wrong.

But I guess this kind of objective function would be used in circumstances where the operators $a_i$ coefficients are not perfectly known. For example in inverse problems this would be useful, because approximate (wrong) models can yield formation of artifacts in the parameters $x$, which you want to correctly estimate.

I do not know the right keywords either to look for it, but would also appreciate any further comments hinting in this direction.

| cite | improve this answer | |

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.