[Rd] bug in L-BFGS-B? (PR#8099)

[Prof Brian Ripley]

Check your versions of MASS.  The Windows one appears to be an outdated 
version, and does different things.

As for the problem, yes it probably is a bug in L-BFGS-B.  Fancy debugging 
the code?

On Sat, 27 Aug 2005 berwin at maths.uwa.edu.au wrote:

>
> --WWm7B+u2U4
> Content-Type: text/plain; charset=us-ascii
> Content-Description: message body text
> Content-Transfer-Encoding: 7bit
>
> G'day all,
>
> I believe that this is related to PR#1717 (filed under
> not-reproducible) which was reported for a version of R that is a
> quite a bit older than the ones used in for this report.  But I
> noticed this behaviour under R 2.1.1 and R 2.2.0 on my linux box and
> under R 2.0.1 and R 2.1.1 on my Windows box.
>
> I recently ran the examples for fitdistr in the MASS library and on my
> linux machine I get the following results:
>
>> library(MASS)
>>      set.seed(123)
>>      x <- rgamma(100, shape = 5, rate = 0.1)
>>      fitdistr(x, "gamma")
>     shape         rate
>  6.45947303   0.13593172
> (0.89052006) (0.01948648)
>>      fitdistr(x, dgamma, list(shape = 1, rate = 0.1), lower = 0.01)
>      shape         rate
>  2.156761461   0.010000000
> (0.277607887) (0.001433778)
>
> I find the second result quite extraordinary.  O.k., the likelihood
> that is maximised is quite flat along a certain direction, but I don't
> think that it is so flat that it should fool the underlying
> optimisation routine into believing that it has converged.
>
> Moreover:
>
>>      fitdistr(x, dgamma, list(shape = 1, rate = 0.1), lower = 0.001)
>     shape         rate
>  6.48686551   0.13651012
> (0.89438949) (0.01956818)
>
> As far as I can tell, the problem is with L-BFGS-B hitting the lower
> bound and stopping, the same problem as reported in #1717.
>
> I tried the same comments on my Windows box and the result is as
> attached below (together with the system information that bug.report
> produces on that machine).  Looking at the output given on the Windows
> machine, I am actually amazed that my Linux box didn't give me any of
> those warning messages --- the R input/output given above is cut and
> paste.  I am not aware of any fiddling with ptions regarding warning
> messages by any configurations files anywhere on my linux system; and I
> ran the same commands using `R --vanilla' on the linux box and did not
> get any warning messages.
>
> Cheers,
>
>        Berwin
>
>
> --WWm7B+u2U4
> Content-Type: text/plain
> Content-Description: Manifestation of bug on a Windows machine
> Content-Disposition: inline;
> 	filename="Bug.txt"
> Content-Transfer-Encoding: 7bit
>
>> set.seed(123)
>>      x <- rgamma(100, shape = 5, rate = 0.1)
>>      fitdistr(x, "gamma")
>     shape         rate
>  6.48592894   0.13649147
> (0.89425721) (0.01956555)
> Warning messages:
> 1: NaNs produced in: dgamma(x, shape, scale, log)
> 2: NaNs produced in: dgamma(x, shape, scale, log)
> 3: NaNs produced in: dgamma(x, shape, scale, log)
>>      fitdistr(x, dgamma, list(shape = 1, rate = 0.1), lower = 0.01)
>      shape         rate
>  2.156761461   0.010000000
> (0.277607888) (0.001433778)
> Warning message:
> bounds can only be used with method L-BFGS-B in: optim(start, mylogfn, x = x, hessian = TRUE, ...)
>>      fitdistr(x, dgamma, list(shape = 1, rate = 0.1), lower = 0.001)
>     shape         rate
>  6.48686551   0.13651012
> (0.89438949) (0.01956818)
> Warning message:
> bounds can only be used with method L-BFGS-B in: optim(start, mylogfn, x = x, hessian = TRUE, ...)
>
> --please do not edit the information below--
>
> Version:
> platform = i386-pc-mingw32
> arch = i386
> os = mingw32
> system = i386, mingw32
> status =
> major = 2
> minor = 1.1
> year = 2005
> month = 06
> day = 20
> language = R
>
> Windows XP Professional (build 2600) Service Pack 2.0
>
> Locale:
> LC_COLLATE=English_Australia.1252;LC_CTYPE=English_Australia.1252;LC_MONETARY=English_Australia.1252;LC_NUMERIC=C;LC_TIME=English_Australia.1252
>
> Search Path:
> .GlobalEnv, package:MASS, package:methods, package:stats, package:graphics, package:grDevices, package:utils, package:datasets, Autoloads, package:base
>
> --WWm7B+u2U4
> Content-Type: text/plain; charset=us-ascii
> Content-Description: message body text
> Content-Transfer-Encoding: 7bit
>
>
> --please do not edit the information below--
>
> Version:
> platform = i686-pc-linux-gnu
> arch = i686
> os = linux-gnu
> system = i686, linux-gnu
> status = Under development (unstable)
> major = 2
> minor = 2.0
> year = 2005
> month = 08
> day = 26
> svn rev = 35424
> language = R
>
> Locale:
> C
>
> Search Path:
> .GlobalEnv, package:methods, package:stats, package:graphics, package:grDevices, package:utils, package:datasets, Autoloads, package:base
>
> --WWm7B+u2U4--
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>
>

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595