Does anyone know what new features have been added to CVX 3.0 beta??

Anything like new capabilities for sparsity or chordal sparsity patterns?

I’d like to give it a try, should it have new interesting features…

Does anyone know what new features have been added to CVX 3.0 beta??

Anything like new capabilities for sparsity or chordal sparsity patterns?

I’d like to give it a try, should it have new interesting features…

I don’t think there are new cap[abilities in CVX 3.0beta for sparsity or chordal sparsity patterns. I don’t know whether there have been any improvements inside solvers callable from CVX, such as Mosek. A Mosek,developer, @Joachim_Dahl is co-author of “Linear matrix inequalities with chordal sparsity patterns andapplications to robust quadratic optimization” http://www.seas.ucla.edu/~vandenbe/publications/cacsd10.pdf

For info on changes in CVX 3.0beta vs. CVX 2.1, read the top text box at http://cvxr.com/cvx/beta/ and the links provided there.

The most significant changes can be found in the DCP ruleset section, due to the introduction of

sign-sensitive convexity analysis.

Unfortunately, CVX 3.0beta is riddled with serious bugs, so use at your own peril.

One other thing CVX 3.0beta offers is the ability to use the native exponential cone capability of ECOS abd SCS. Unfortunately, the likelihood of encountering CXV 3.0beta bugs when trying to do so may negate this advantage. Also, CVXQUAD is now available for use under CVX, and that alows CVX;'s Successive Approximation method to be avoided - see CVXQUAD: How to use CVXQUAD's Pade Approximant instead of CVX's unreliable Successive Approximation for GP mode, log, exp, entr, rel_entr, kl_div, log_det, det_rootn, exponential cone. CVXQUAD's Quantum (Matrix) Entropy & Matrix Log related functions

There are no new features for semidefinite optimization in MOSEK 9. Perhaps we will investigate new possibilities in MOSEK 10.

The current state of exploiting chordal sparsity requires a good deal of knowledge and is not easy to fit into a modeling language such as CVX. The chordal conversion methods are most interesting for large very sparse SDPs, which is not what CVX was designed for. A few users (including Martin Andersen) have succesfully solved very large chordal SDPs using MOSEK, but it required careful implementation using the solver API.

Dear @Joachim_Dahl

thanks for that valuable information. By the way, you’re one of the coauthors of the solver conelp implemented in cvxopt, not written for matlab though, but which can be coupled with chompack developed by Andersen and Vanderberghe to solve problems with chordal patterns. My question to you then is: if I know that my problem is both convex and has chordal sparse structure for all blocks (SOCPs and SDP blocks), then is it enough to apply the chordal conversion implemented in chompack (theory by fukuda et al.) so that the solver conelp takes advantage of the structure?

CHOMPACK cannot be used directly with CVXOPT, but it is used in SMCP by Martin Andersen.

I think currently the most competitive method is to perform the chordal conversion manually and then solve the problem using MOSEK.

By manually you mean writing the problem in conic form, do the chordal conversion using Fukuda et al method and then passing it to MOSEK outside CVX or is there a way to use CVX’s modelling capabilities along the way?