PSOR is iterative, which means it *can* be slow depending on the spectrum (eigenvalues) of A.
Just out of interest, did you try Brennan-Schwartz? It's fast and easy to implement.
Matty,
is this the method from Nocedal?
http://ziena.com/papers/case_options.pdf
Looking at the numbers, 2.1 secs is NOT fast. I vaguely remember ~0.2 seconds using ADE and Brennan Schwarz.
//
Some articles tend to have the adjective 'fast' as the first word in the title. Even if fast, it seems strange...
---------- Post added at 07:56 AM ---------- Previous post was at 07:46 AM ----------
Here is a post I made from another forum last year, hth
I have done a number of tests against published results and other methods (e.g. penalty). Until now, I reckon that ADE with domain transformation and with checking of the intrinsic value constraint is up there in the first few.
Example: CEV model with beta = 2/3.
machine is Window XP (Dell vostro 1710) hobby laptop 1.8 Ghz, 2GB RAM. Duo core (and I use them).
In release mode the ADE model prices to the penny (2, 3 decimal places) in interval [45, 130] milliseconds. Using OpenMP multithreading gives a speedup of roughly 60%.
The nice thing is I do not need to define any nonlinear penalty functions, front fixing or domain truncation tricks. These methods I tried were much slower.
I may have posted this article already, but here it is on SSRN (ADE methods and with early exercise as well).
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1552926