Hi Luc, it is exactly as you've described. After manually setting maxIters to 100, I was able to find eigenvectors for bigger matrices. Tomorrow I'll decide whether to open an issue or implement new constructors myself.
Thanks & BRs, Maciek Rubikowski > ---------- Wiadomość przekazana dalej ---------- > From: Luc Maisonobe <[email protected]> > To: Commons Users List <[email protected]> > Date: Tue, 29 Nov 2011 20:48:54 +0100 > Subject: Re: [math] EigenDecompositionImpl -- maxIters causing troubles? > Le 28/11/2011 22:41, Maciej Rubikowski a écrit : >> Hello, > > Hi Maciej, > >> >> before I fill in the bug report, I want to ask if anyone had similar >> issue before. Maybe I simply misunderstood the whole idea behind >> EigenDecomposition, but I think what I've come up with is in fact a >> serious limitation. >> >> I've tried to decompose a covariance matrix, approximately 1000x1000 >> big. After few seconds I've ended up with a >> MaxIterationsExceededException. I've investigated the cause and found >> the following field in EigenDecompositionImpl: >> >> /** Maximum number of iterations accepted in the implicit QL >> transformation */ >> private byte maxIter = 30; >> >> Does anyone know any reason for hard-coding such value? What are the >> true limitations of the EigenDecomposition? Is there any way to get >> around it? > > We could probably add new constructors that would allow users to choose > another setting for this value, and let 30 be only a default one. > > Did you try to temporarily change this limit in the code to see if this > solved your problem ? If so, you should open a Jira issue. Of course, if > you feel so inclined, you can also provide a patch for this, adding the > new constructors. > > best regards, > Luc > >> >> >> Best regards, >> Maciej Rubikowski >> Warsaw University of Technology / Institute of Computer Science >> >> --------------------------------------------------------------------- >> To unsubscribe, e-mail: [email protected] >> For additional commands, e-mail: [email protected] >> > > > > --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
