R-1.7.1 regression test failure on alphaev68-dec-osf5.1

3 messages · Jeff Lewis, Peter Dalgaard, Spencer Graves

Tue, Jun 24, 2003 11:42 AM #

Thanks for the quick response.  The two sides of the equality are
definately different.  Here's what I'm seeing

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

[1] 3.141593

[1] 0+1i

[1] 0+3.141593i

[1] -1+1.224647e-16i

[1] 0+3.141593i

[1] 3.141593+0i

[1] 4.440892e-16+0i

[1] 4.440892e-16

[1] 2.220446e-16

[1] FALSE

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

I get the same thing from R 1.6.2, which I compiled about six months
ago.  Is there anything I can/should do to fix this?

Peter Dalgaard

Tue, Jun 24, 2003 12:02 PM #

"Jeff Lewis" <jlewis at genomecorp.com> writes:

Not really. It seems that your platform just has slightly less
accurate complex log/exp routines than the most common ones (Linux and
Sparc/Solaris both give exact zero). Probably the check is simply
overly stringent.

You might want to change the check to say  ... < 3*.Machine$double.eps
or so and rerun, to check whether the rest of the checks pass.

O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)             FAX: (+45) 35327907

Spencer Graves

Tue, Jun 24, 2003 12:13 PM #

I just tried this with both R 1.6.2 and S-Plus 6.1 under Windows 2000: 
I got the same .Machine$double.eps as reported below, but I did NOT get 
the problem:

pi - log(exp(pi*1i)) / 1i

# reported below:  [1] 4.440892e-16+0i
# R1.6.2 and S-Plus 6.1 both gave 0+oi

Mod(pi - log(exp(pi*1i)) / 1i)

# reported below: 4.440892e-16
# R1.6.2 and S-Plus 6.1 both gave 0

.Machine$double.eps

# Same as reported: 2.220446e-16

Mod(pi - log(exp(pi*1i)) / 1i) < .Machine$double.eps

# reported below:  [1] FALSE
# R1.6.2 and S-Plus 6.1 both gave T

hth.  spencer graves

Jeff Lewis wrote:

-----Original Message-----
From: Peter Dalgaard BSA [mailto:p.dalgaard at biostat.ku.dk] 
Sent: Tuesday, June 24, 2003 1:20 PM
To: Jeff Lewis
Cc: r-help at stat.math.ethz.ch
Subject: Re: [R] R-1.7.1 regression test failure on 
alphaev68-dec-osf5.1


"Jeff Lewis" <jlewis at genomecorp.com> writes:

I'm attempting to compile and install R version 1.7.1 for

my statistical

geneticists.  It seems to compile correctly -- that is, it compiles
without errors -- but the regression test is failing in the

following

manner:

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

## log
stopifnot(all.equal(log(1:10), log(1:10, exp(1))))
stopifnot(all.equal(log10(30), log(30, 10)))
stopifnot(all.equal(log2(2^pi), 2^log2(pi)))
stopifnot(Mod(pi - log(exp(pi*1i)) / 1i) < .Machine$double.eps)

Error: Mod(pi - log(exp(pi * (0+1i)))/(0+1i)) <

.Machine$double.eps is

not TRUE
Execution halted
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

I'm compiling on Tru64UNIX 5.1A using DECs C and Fortran

compilers and

perl 5.6.0.  I found the above error in a file named
'reg-tests-1.Rout.fail'.  Any help you can give me would be most
appreciated.

Well, it seems to be an accuracy issue, so the first question would be
what the values on both sides of the equality are (just start up R and
enter the expressions on te command line). You might have

1) Completely wrong results in complex arithmetic 
2) Slightly less than optimal accuracy
3) An underestimated .Machine$double.eps

Case 2) seems most likely, but case 3) has been observed with buggy
compilers that optimize calculations where they shouldn't.




Thanks for the quick response.  The two sides of the equality are
definately different.  Here's what I'm seeing

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

pi

[1] 3.141593

1i

[1] 0+1i

pi*1i

[1] 0+3.141593i

exp(pi*1i)

[1] -1+1.224647e-16i

log(exp(pi*1i))

[1] 0+3.141593i

log(exp(pi*1i)) / 1i


[1] 3.141593+0i

pi - log(exp(pi*1i)) / 1i

[1] 4.440892e-16+0i

Mod(pi - log(exp(pi*1i)) / 1i)

[1] 4.440892e-16

.Machine$double.eps

[1] 2.220446e-16

Mod(pi - log(exp(pi*1i)) / 1i) < .Machine$double.eps

[1] FALSE

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=

I get the same thing from R 1.6.2, which I compiled about six months
ago.  Is there anything I can/should do to fix this?

______________________________________________
R-help at stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help