R help - Jun 2003 - optim with contraints

There seems to exist peculiar cases where optim does not take care
of constraints on the parameters to be optimized over.  The call to
optim is of the form

  opt <- optim(cp, fn=sn.dev, gr=sn.dev.gh, method="L-BFGS-B",
           lower=c(-Inf, 1e-10, -0.99527), 
           upper=c( Inf, Inf,    0.99527), 
           control=control, X=X, y=y, hessian=FALSE)

The code has worked fine many times, but I have come across cases 
(for suitable data X and y) where the constraint on the last component 
is ignored;  that means that a call is made to sn.dev with
       cp = -1.3546  0.4645  3.1741
so the third component exceeds 0.99527, and the program stops 
because of a check in the  function to be obtimised.

The call just before was to the gradient function
sn.dev.gh: gradient =    219013   -312643 441647332
which has rather large values.

To make the problem more interesting, it shows up on a Linux
(Debian) installation, but it works fine on MS-windows.
In both cases, R is 1.7.0.  

Perhaps this sort of question should not be directed to the R-help
list, rather to the developer(s) of optim. Please instruct me on this 
point. Also, I appreciate that the above is not a reproducible example;
this would be longish in text, and I though to ask first to whom 
it is appropriate that I direct my question.


Best wishes,

Adelchi Azzalini


-- 
Adelchi Azzalini  <azzalini at stat.unipd.it>
Dipart.Scienze Statistiche, Universit? di Padova, Italia
http://azzalini.stat.unipd.it/

Adelchi,

R is a volunteer project, and we would need a volunteer to look at this.  
`The developers of optim' (who are R-core) used published code for this
method. Since you are getting different answers on different platforms
this might be a bug in your compiler or run-time rather than R.

Please submit a reproducible bug report to R-bugs.  As you will see some
bugs don't get fixed very fast (including one I submitted from Padua in
2001).

Brian

On Sat, 21 Jun 2003, Adelchi Azzalini wrote:
> 
> There seems to exist peculiar cases where optim does not take care
> of constraints on the parameters to be optimized over.  The call to
> optim is of the form
> 
>   opt <- optim(cp, fn=sn.dev, gr=sn.dev.gh, method="L-BFGS-B",
>            lower=c(-Inf, 1e-10, -0.99527), 
>            upper=c( Inf, Inf,    0.99527), 
>            control=control, X=X, y=y, hessian=FALSE)
> 
> The code has worked fine many times, but I have come across cases 
> (for suitable data X and y) where the constraint on the last component 
> is ignored;  that means that a call is made to sn.dev with
>        cp = -1.3546  0.4645  3.1741
> so the third component exceeds 0.99527, and the program stops 
> because of a check in the  function to be obtimised.
> 
> The call just before was to the gradient function
> sn.dev.gh: gradient =    219013   -312643 441647332
> which has rather large values.
> 
> To make the problem more interesting, it shows up on a Linux
> (Debian) installation, but it works fine on MS-windows.
> In both cases, R is 1.7.0.  
> 
> Perhaps this sort of question should not be directed to the R-help
> list, rather to the developer(s) of optim. Please instruct me on this 
> point. Also, I appreciate that the above is not a reproducible example;
> this would be longish in text, and I though to ask first to whom 
> it is appropriate that I direct my question.
> 
> 
> Best wishes,
> 
> Adelchi Azzalini
> 
> 
> 
-- 
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