[R-SIG-Mac] [R] lme4 + R 2.11.0 + mac unavailable

[Prof Brian Ripley]

Let us be clear about what the issue seems to be.  On i386 Mac (and 
not on x86_64 Mac)

library(lme4)
y <- (1:20)*pi; x <- (1:20)^2;group <- gl(2,10)
M2. <- lmer (y ~ 1 + x + (1 + x | group))
M2 <- lmer (y ~ x + ( x | group))
identical(fixef(M2), fixef(M2.))

usually fails the first time but if you run it again it is usually 
true, e.g. following on gave

> fixef(M2); fixef(M2.)
(Intercept)           x
  15.2354383   0.1855865
(Intercept)           x
  15.2354486   0.1855864
> M2. <- lmer (y ~ 1 + x + (1 + x | group))
> fixef(M2.)
(Intercept)           x
  15.2354383   0.1855865

I think 'later version of MacOS' is most likely someone not 
understanding that x86_64 is the default for 'R' on Snow Leopard. 
I've just checked this on Leopard and Snow Leopard.  *But* it is not 
repeatable, as sometimes I get TRUE the first run in a session -- and 
once that happens it seems to keep happening until a reboot or I 
reinstall lme4.

I ran this under valgrind (on Leopard: there is now experimental 
valgrind support for SL, but I've not installed it as yet) and saw no 
reports of problems.  However, on one of the times I tried it under 
valgrind  the result was further away:
> fixef(M2.)
(Intercept)           x
  15.2381811   0.1856179
and persistent.

For completeness, I also saw this using R's reference BLAS rather than 
the (default on a CRAN build) Apple 'veclib' BLAS.

So the issue appears to be one of repeating the calculation and 
getting different answers.  That does look like a i386-Mac-specific 
bug, but whether this is worth not distributing a binary for is 
something for the lme4 authors to judge -- if not then a 
platform-specific test opt-out seems the best way forward.


On Sat, 31 Jul 2010, John Maindonald wrote:

> Is optimisation of code a possibility, i.e., the same algebraic 
> calculations done in a slightly different way depending on the state 
> of various registers? E.g., (ab)c vs a(bc), but it would almost 
> certainly be more subtle than that. That seems to me the sort of 
> thing that is likely to be Mac-specific.

But the Mac shares a compiler with lots of other platforms, and the 
gcc optimizer for i386 is pretty much shared across platforms (albeit 
the Mac gcc is now rather old, but is of similar vintage to that used 
on Windows for R 2.11.x).  In any case, this is not (in my 
experiments) an issue of numerical accuracy but of repeatability,

>
> John Maindonald             email: john.maindonald at anu.edu.au
> phone : +61 2 (6125)3473    fax  : +61 2(6125)5549
> Centre for Mathematics & Its Applications, Room 1194,
> John Dedman Mathematical Sciences Building (Building 27)
> Australian National University, Canberra ACT 0200.
> http://www.maths.anu.edu.au/~johnm
>
> On 31/07/2010, at 5:09 AM, Martin Maechler wrote:
>
>> On Fri, Jul 30, 2010 at 17:29, Simon Urbanek
>> <simon.urbanek at r-project.org> wrote:
>>>
>>> On Jul 29, 2010, at 7:51 PM, Martin Maechler wrote:
>>>
>>>>>>>>> "KB" == Ken Beath <ken at kjbeath.com.au>
>>>>>>>>>   on Wed, 28 Jul 2010 01:19:34 -0000 (GMT) writes:
>>>>
>>>>   KB> On Tue, July 27, 2010 11:07 pm, John Maindonald wrote:
>>>>>> Binaries for lme4 are not for the time being available either on
>>>>>> CRAN or on r-forge?
>>>>>>
>>>>
>>>>   KB> There was a discussion about this on R-Sig-ME. It fails one of the checks
>>>>   KB> on the machine used to build the mac packages. It may pass on machines
>>>>   KB> with a later version of MacOS, but there didn't seem to be a consensus of
>>>>   KB> whether failing the checks was a problem.
>>>>
>>>> The underlyign issue seems to remain unresolved.
>>>> My best guess is still to low-level bugs somewhere in the libraries
>>>> used.
>>>>
>>>> However, to help users,
>>>> I'm willing to *not* run those tests when *running* on the Mac.
>>>>
>>>> What does
>>>>    Sys.info()[["sysname"]]
>>>>
>>>> return on the Mac, i.e., different versions of R running on the
>>>> Mac?
>>>> From grepping around I'd expect it give  "Darwin"  in all cases.
>>>> Is that true?
>>>>
>>>
>>> ... why don't you just compare the values with tolerance? That 
>>> seems to make more sense that to disable the test altogether ...
>>>
>>
>> Doug and I have explained in several occasions (but maybe not on R-SIG-Mac)
>> that this is *NOT* the point here.
>> It's not about "basically the same computations" it's about
>> identical computations, i.e. from identical numerical matrices and so
>> the use of identical() has been very much on purpose, and is different
>> from the well justified typical use of
>> all.equal() which we use in many other places.
>>
>> And what people where reporting *is* a bit disturbing: The same
>> computations giving different answers for *exactly* the same
>> computation, e.g., depending if they had used
>> library(lme4)  or  require(lme4)  to load the package ...
>> and the difference was not just the last significant bit, but changes
>> at about the 7-th significant digit.
>>
>> Martin
>>
>>
>>> Cheers,
>>> Simon
>>>
>>
>> _______________________________________________
>> R-SIG-Mac mailing list
>> R-SIG-Mac at stat.math.ethz.ch
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mac
>
> _______________________________________________
> R-SIG-Mac mailing list
> R-SIG-Mac at stat.math.ethz.ch
> https://stat.ethz.ch/mailman/listinfo/r-sig-mac
>

-- 
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