On Tuesday, 3 July 2012 07:46:48 UTC+8, Martin Albrecht wrote:
>
> Hi, 
>
> On Tuesday 03 Jul 2012, Simon King wrote: 
> > Hi! 
> > 
> > On 2012-07-02, Martin Albrecht  wrote: 
> > > Shouldn't both give the same distribution mod p? Since every 
> non-singular 
> > > matrix A has a LU decomposition we should be able to just sample L and 
> U 
> > > separately to produce A? 
> > 
> > Sorry for my ignorance, but is it really the case that an LU 
> > decomposition exists for all invertible matrices? I thought there may 
> > only be an LUP decomposition. 
>
> Argh, yes, you're right: should be LUP. 


> > If I am not mistaken, the LU decomposition is unique if one requires 
> > that L (or U) has only 1 on the diagonal. Because of the uniqueness, I'd 
> > expect that putting 1 on the diagonal of L and choosing the entries of U 
> > and the remaining of L randomly equally distributed yields a reasonable 
> > distribution of invertible matrices. 
> > 
> > However, if it is really the case that we must consider LUP 
> > decompositions, then I am not totally convinced that a nicely 
> distributed 
> > random choice of a permutation matrix P on top of the choice of L and U 
> as 
> > above yields a nice distribution of invertible matrices. 
>
> Mhh, why not? If A = LUP we just write AP^-1  = LU, hence for each LU we 
> construct there are as many As as there are permutation matrices, or am I 
> missing something (again :))? 
>

P can be taken to be identity iff the leading principal minors of A are 
non-0.
Thus LU (say, with diag(L)=(1,...,1)) gives you a random matrix from this 
class.
So if you are content with sampling from such subclass, LU suffices.


There is, in fact, a big industry in the theory of finite groups that 
handles these
kinds of questions.

Dima
 

>
> Cheers, 
> Martin 
>
>

-- 
To post to this group, send an email to [email protected]
To unsubscribe from this group, send an email to 
[email protected]
For more options, visit this group at http://groups.google.com/group/sage-devel
URL: http://www.sagemath.org

Reply via email to