Martin.Schlather@uni-bayreuth.de
2001-Mar-14 15:41 UTC
[Rd] segmentation fault of unknown cause (PR#877)
Dear all,
Paulo Ribeiro, p.ribeiro@lancaster.ac.uk, and myself noticed
that R sometimes breaks down with a segmentation fault of
(at least for us) unknown course. (We've read #411 and #671
in the bug report on "optim" -- the problem described
there sounds familiar; but we have not found any hint on a
segmentation fault.)
Please find at the very end of this email the file
strange.R and on
http://www.geo.uni-bayreuth.de/~martin/R/strange.tgz
an archiv of three files, namely strange.R, cov.data,
and rf.data.
The file strange.R uses only mva as library; it does
not contain any Fortran or C call.
On three machines in Lancaster and on my three
computers (2x R-1.2.2, 1x R-1.2.1) we get
`segmentation fault' for
/usr/bin/R --vanilla < strange.R
Furthermore, I get a segmentation fault for
/usr/bin/R
source("strange.R")
but not Paulo at Lancaster, neither Martin Maechler on
his preversion R-1.3.0 (I had asked Martin before Paulo
found out that `usr/bin/R --vanilla < strange.R' breaks
on his (and mine) computers.)
We would be happy for any comment on whether the
segmentation fault appears anywhere else (and why).
We noticed that R does not fail if /usr/bin/R is called
with other or further options but `--vanilla' (Not
extensively tested yet.)
Cheers,
Martin
PS: For example, a computer in Lancaster (all RedHat) and
one here (all SuSe) are both installed by compiling the
source code of R-1.2.2 and have the specification
_
platform i686-pc-linux-gnu
arch i686
os linux-gnu
system i686, linux-gnu
status
major 1
minor 2.2
year 2001
month 02
day 26
language R
PPS: If the file strange.R works well then it stops with an
error message different from "segmentation fault".
--
Martin Schlather email: Martin.Schlather@uni-bayreuth.de
Abteilung Bodenphysik phone: +49 (0)921 55 2193
Univ. Bayreuth Fax : +49 (0)921 55 2246
D -- 95440 Bayreuth, Germany http://www.geo.uni-bayreuth.de/~martin/
#### the comments in the file strange.R give a report on what happens
#### additionally on all *my* computers -- not tested elsewhere yet
######################################################
############ strange.R ###############
######################################################
# calling R (v1.2.1 or v1.2.2) without options and
# then `source("strange.R")' causes a segmentation fault
# Note that before failure MLEtarget is called with
# variab=c(1.0501141215486309848e-267, 1.0000000000000000208e-03)
# a value that is outside the given bounds, see lb and optim below
# (This is something that appears quite frequently, independently (?)
# of the segmentation fault I'm interested here!)
# The segmentation fault does not appear if R is called by
# /usr/bin/R --vsize 20M --nsize 250K --no-save
# (The algorithm will fail with "fn(par, ...) : subscript out of
bounds"
# then, what is OK since the matrix `cm' does not have enough
# components)
library(mva)
"mleRF" <-
function (data, coord)
{
ENVIR <- environment()
lc <- nrow(coord)
## if you comment out either line, the algorithm
## returns the correct error message
## "fn(par, ...) : subscript out of bounds"
distances <- as.double(dist(coord))
distances <- dist(coord)
## lower und upper bounds for optim
lb <- c(0.00235992765321886, 0.2)
ub <- c(39.6649897977748, 10)
assign("COUNTER", 1, envir = ENVIR)
CoVariate <- matrix(1, ncol = 1, nrow = lc)
MLEtarget <- function(variab) {
print(variab, dig = 20)
## the covariance matrix is calculated by an own C programme;
## therefore the results have been stored in `cm' to exclude
## my programme as course of the trouble
cov.matrix <- cm[, COUNTER]
cov.matrix <- try(chol(matrix(cov.matrix, ncol = lc)))
## if you replace the above two lines by
## cov.matrix <- matrix(cm[, COUNTER], ncol = lc)
## cov.matrix <- try(chol(cov.matrix))
## segmentation fault vanishes, and `optim' returns
## "Error in optim(...: bad value"
assign("COUNTER", COUNTER + 1, envir = ENVIR)
if (is.numeric(cov.matrix)) {
if (any(diag(cov.matrix) < 0)) {
stop("chol det<0!")
}
}
else {
return(1e+30)
}
logdet <- 2 * sum(log(diag(cov.matrix)))
cov.matrix <- chol2inv(cov.matrix)
dummy <- t(CoVariate) %*% cov.matrix
m <- (solve(dummy %*% CoVariate) %*% (dummy %*% data))
V <- data - CoVariate %*% m
quadratic <- (t(V) %*% cov.matrix %*% V)
res <- logdet + lc * log(quadratic)
print(as.vector(res), dig = 20)
return(res)
}
## if you outcomment the following `for'-loop
## the segmenatation fault vanishes, and optim returns
## "Error in optim(...: bad value"
for (W in 1:3) {
mlet <- MLEtarget(c(1,1))
}
## if the `for'-loop is active, the following line
## is not necessary as COUNTER has already the value 4
assign("COUNTER", 4, envir = ENVIR)
variab <- c(0.00235992765321890, 2.66384186745229)
## if you outcomment the following line, the algorithm
## returns the correct error message "fn ... out of bounds"
distances <- as.double(distances)
variab <- try(optim(variab, MLEtarget, method = "L-BFGS-B",
lower = lb, upper = ub)$par)
print("END")
}
## the random field given on a grid by x,y and f is
## calculated by an own C-programme. Therefore the results
## have been stored to exclude my programme as course of
## the trouble
#save(file="rf.data",x,y,f,.Random.seed)
load("rf.data")
## see `MLEtarget' for comments on `cm'
#save(file="cov.data",cm)
load("cov.data");
mleRF(coord=cbind(x,y),data=f)
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !) To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
Reasonably Related Threads
Wisdom of the Ancients
