[Rd] using so-library involving Taucs

Douglas Bates dmbates at gmail.com
Fri Jun 3 21:45:40 CEST 2005


On 6/3/05, Susanne Heim <susanne.heim at stat.uni-muenchen.de> wrote:
> Dear R developers,
> 
> The trace of the hat matrix H~(n,n) is computed as follows:
> 
> tr(H) = tr(BS^-1B')  = tr(S^-1B'B) := tr(X) = sum(diag(X))
> 
> with B~(n,p), S~(p,p).
> Since p is of the order 10^3 but S is sparse I would like to employ
> Taucs linear solver ( http://www.tau.ac.il/~stoledo/taucs/ ) on
> 
> SX = B'B.
> 
> (Further improvement by implying a looping over i=1,...,p, calling
> taucs_linsolve(S, X[,i], (B'B)[,i]) and saving X[i,i] only is pending.)
> 
> For this purpose I compiled the C code "hattrace.c" to a shared object
> using:
> gcc -g -Wall -I/usr/local/taucs/src -I/usr/local/taucs/build/linux  -c
> hattrace.c -o hattrace.o
> gcc -g  -L/usr/local/taucs/external/lib/linux
> -L/usr/local/taucs/lib/linux -L/usr/local/lib -L/opt/gnome/lib
> -L/usr/lib/R/lib -shared -fpic -o hattrace.so hattrace.o -ltaucs
> -llapack -lf77blas -lcblas -latlas  -lmetis -lm -lg2c -lR
> 
> I tried the following test commands:
> library(splines)
> library(SparseM)
> B <- splineDesign(knots = 1:10, x = 4:7)
> D <- diff(diag(dim(B)[2]), differences = 1)
> BB <- t(B) %*% B
> S <- as.matrix.ssc(BB + t(D) %*% D)
> if (!is.loaded(symbol.C("hattrace"))) { dyn.load(paste("hattrace",
> .Platform$dynlib.ext, sep = "")) }
> out <- 0
> spur <- (.C("hattrace", as.double(as.vector(slot(S, "ra"))),
>             as.integer(as.vector(slot(S, "ja") - 1)),
>             as.integer(as.vector(slot(S, "ia") - 1)),
>             as.integer(dim(S)[1]), as.double(as.vector(BB)),
>             as.double(out), PACKAGE = "hattrace"))[[6]]
> 
> Unfortunately, I get an R process segmentation fault although the C Code
> outputs the correct trace value to /tmp/hattrace.log which I checked by
> a equivalent R routine. Since this segmentation fault does not occur
> every time, I assume a pointer problem. Any help on how to solve it is
> greatly appreciated.

Is S positive definite?  If so, it may be more effective to take the
Cholesky decomposition of S and solve the system S^(1/2)X = B then
take the sum of the squares of the elements of X.

If you wish to provide me off-list with examples of the matrices S and
B, I can check how best to do this with the Matrix package.



More information about the R-devel mailing list