1
$\begingroup$

I am reading this paper: http://dl.acm.org/citation.cfm?id=1390696

Following optimization problem is defined in section 2: \begin{align} \max_{\mathbf{X}>0} \log det(\mathbf{X})-trace(\mathbf{SX})-\lambda||\mathbf{X}||_1 \end{align} where $\mathbf{X}$ and $\mathbf{S}$ are $p\times p$ matrix and $\mathbf{X}$ is constrained to be positive definite.

In the next subsection, the author creates a dual representation by writing \begin{align} ||\mathbf{X}||_1 = \max_{||\mathbf{U}||_\inf \leq 1} trace(\mathbf{XU}) \end{align}

This is fine, as $||\mathbf{X}||_1 = trace(\mathbf{XU})$, if all entries of U are 1, and that is the maximum value possible. But in the next step, this is substituted in the original optimization problem to form: \begin{align} \max_{\mathbf{X}>0} \min_{||\mathbf{U}||_\inf \leq \lambda}\log det(\mathbf{X})-trace(\mathbf{(S+U)X}) \end{align} and the $\max_{||\mathbf{U}||_\inf \leq 1} $ is changed $\min_{||\mathbf{U}||_\inf \leq \lambda} $ . Changing 1 to $\lambda$ is obvious, but why is the max changed to min. Shouldn't a minus sign be placed before $\mathbf{U}$ in that case?
Is there some property of trace operator which allowed for this step?

$\endgroup$
1
$\begingroup$

$$a-\max_{x\in X} f(x)=\min_{x\in X}(a-f(x)),$$ this is just placing the minus sign inside the max which makes it a min.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ but then it should be $\min_{||\mathbf{U}||_\inf \leq \lambda}\log det(\mathbf{X})-trace(\mathbf{(S-U)X})$ instead of $\min_{||\mathbf{U}||_\inf \leq \lambda}\log det(\mathbf{X})-trace(\mathbf{(S+U)X})$ $\endgroup$ – user2329744 May 29 '14 at 17:13
  • $\begingroup$ No, because then you would have a double negation, $-tr(-UX)=tr(UX)$. $\endgroup$ – Lutz Lehmann May 29 '14 at 17:26
  • $\begingroup$ @LutzL is right; it's really that simple. Try breaking apart $\mathop{\textrm{Tr}}((S+U)X)$ into two terms, $\mathop{\textrm{Tr}}(UX)+\mathop{\textrm{Tr}}(SX)$, to help see it. $\endgroup$ – Michael Grant May 29 '14 at 19:17
  • $\begingroup$ Yes, I see it now. It is simple arithmetic. Thanks for your responses $\endgroup$ – user2329744 May 31 '14 at 11:13

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.