qp() in octave-3.0.0 gets stuck on wrong answer to convex problem

[John W. Eaton]

On 22-Aug-2008, Joshua Redstone wrote:

| With respect to the bug I described, I don't think it matters if H is
| positive semi-definite or not.
| It should be impossible for qp(), given a *single* quadratic programming
| problem and multiple starting points,
| to return an answer claiming the problem is convex for some starting points,
| and for other starting points
| return an answer indicating it has gotten stuck in a local minimum.  A
| single quadratic programming problem can not be both convex
| and not-convex at the same time.  To put it another way, either the problem
| is convex, in which case all starting points should lead
| to the global optimum, or the problem is not-convex, in which case qp() is
| in error claiming that it is convex.  Either way there is a bug in
| either the qp() implementation or interface specification.
| Does my reasoning seem reasonable?

I probably missed some of the details of the discussion.  I was just
pointing out that your problem does not appear to have a positive
definite H.  If the qp function is sometimes saying that the problem
is convex and sometimes not, then that does seem to be a bug.  I'd be
happy to have a patch for that problem.

Thanks,

jwe