R help - Jan 2005 - Memory Efficient Methods for Building Matrix

Dear List:

I am having to build a block-diagonal matrix (vl) and am currently using
the following code.

I<-diag(sample.size) 
vl<-kronecker(I,vl.mat)
 

This code works fine, but for large N, it is a huge memory hog. Is there
a more efficient method for constructing vl?

Thanks,

Harold


	[[alternative HTML version deleted]]

On Mon, 3 Jan 2005, Doran, Harold wrote:
> Dear List:
>
> I am having to build a block-diagonal matrix (vl) and am currently using
> the following code.
>
> I<-diag(sample.size)
> vl<-kronecker(I,vl.mat)
>
>
> This code works fine, but for large N, it is a huge memory hog. Is there
> a more efficient method for constructing vl?
>
Obvious alternatives such as

nr<-nrow(v1.mat)
nc<-ncol(v1.mat)
result<-matrix(0,nrow=sample.size*nr,ncol=sample.size*nc)
for(i in 1:sample.size){
     result[ (i-1)*nr+1:nr, (i-1)*nc+1:nc]<-v1.mat
}

or
nr<-nrow(v1.mat)
nc<-ncol(v1.mat)
iy<-as.vector(outer(rep(1:nc,each=nr),(0:(sample.size-1))*nc,"+"))
ix<-as.vector(outer(rep(1:nr, nc),(0:(sample.size-1))*nr,"+"))
result<-matrix(0,nrow=sample.size*nr,ncol=sample.size*nc)
result[cbind(ix,iy)]<-v1.mat

seem to take a little less memory and about the same time,
but constructing a large block-diagonal matrix is intrinsically an
inefficient thing to do.


 	-thomas

Thomas Lumley <tlumley <at> u.washington.edu> writes:

: 
: On Mon, 3 Jan 2005, Doran, Harold wrote:
: 
: > Dear List:
: >
: > I am having to build a block-diagonal matrix (vl) and am currently using
: > the following code.
: >
: > I<-diag(sample.size)
: > vl<-kronecker(I,vl.mat)
: >
: >
: > This code works fine, but for large N, it is a huge memory hog. Is there
: > a more efficient method for constructing vl?
: >
: 
: Obvious alternatives such as
: 
: nr<-nrow(v1.mat)
: nc<-ncol(v1.mat)
: result<-matrix(0,nrow=sample.size*nr,ncol=sample.size*nc)
: for(i in 1:sample.size){
:      result[ (i-1)*nr+1:nr, (i-1)*nc+1:nc]<-v1.mat
: }
: 
: or
: nr<-nrow(v1.mat)
: nc<-ncol(v1.mat)
:
iy<-as.vector(outer(rep(1:nc,each=nr),(0:(sample.size-1))*nc,"+"))
: ix<-as.vector(outer(rep(1:nr, nc),(0:(sample.size-1))*nr,"+"))
: result<-matrix(0,nrow=sample.size*nr,ncol=sample.size*nc)
: result[cbind(ix,iy)]<-v1.mat
: 
: seem to take a little less memory and about the same time,
: but constructing a large block-diagonal matrix is intrinsically an
: inefficient thing to do.

You could try combining some of those ideas with use of the SparseM
package.