[Rd] Catching errors from solve() with near-singular matrices

[David Sterratt]

Dear all,

many thanks to Jon & Ravi for their help on this, and apologies if
r-help would have been a more appropriate forum.

On Tue, 2012-12-11 at 15:43 +0000, Jon Clayden wrote:

>       Strategy 1: Some code like this:
>            if (det(X) < epsilon) {
>               warning("Near singular matrix")
>               return(NULL)
>            }
>            return(solve(X))
>
> This solution is probably the easiest one to take, but to match
> solve.default, the test should be
>
>   if (rcond(X) < .Machine$double.eps)
>
> Catching that case should avoid the error. I hope this helps.

Yes, that works well.

On Tue, 2012-12-11 at 15:46 +0000, Ravi Varadhan wrote:
> In any case,  you might want to try MASS::ginv instead of solve(), if
> you expect ill-conditioning.  Here is one possible solution:
> 
>         f <- function(X) {
>           invX <- tryCatch(ginv(X, tol=.Machine$double.eps),
> error=function(e) {
>             warning(e)
>             error.flag <- TRUE})  # you should avoid the global
> assignment `<<-' 
>           if (error.flag) return(NULL)
>           return(invX)
>         }

I'll try this if any problems do emerge with the above solution. One
small point: I used <<- because there seemed to be no other way of
returning from the function f if an error had been thrown by ginv() (or
solve()):

> flag <- FALSE
> tryCatch(stop(), error=function(e) {flag <- TRUE})
> flag
[1] FALSE
> tryCatch(stop(), error=function(e) {flag <<- TRUE})
> flag
[1] TRUE

I feel sure that there must be a more elegant way of doing this!

David.

-- 
David C Sterratt, Research Fellow http://homepages.inf.ed.ac.uk/sterratt
Institute for Adaptive and Neural Computation     tel: +44 131 651 1739
School of Informatics, University of Edinburgh    fax: +44 131 650 6899
Informatics Forum, 10 Crichton Street, Edinburgh EH8 9AB, Scotland  
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
NEW BOOK: Principles of Computational Modelling in Neuroscience
Sterratt, Graham, Gillies & Willshaw (CUP, 2011).
http://www.compneuroprinciples.org  


The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.