bug (PR#13570)

Hi

Nice to hear from you Ryan. I also do not have the capability to  
debug on windows; however, there is a chance that the behavior you  
are seeing is caused by the following bug noted in my thesis  
(available on ProQuest; email me if you don't have access):

"When lambda = 0 there are no local slopes to aid the blending  
algorithm, yet the
interpolator would still assume they were available, and thus use  
arbitrary values
from memory. This had implications for both fit and tr[L]  
computation. In the
updated code these are set equal to zero which seems the best  
automatic rule when
lambda = 0." [lambda refers to degree]

I submitted a bug fix to Eric Grosse, the maintainer of the netlib  
routines; the fixed lines of fortran are identified in the comments  
at (just search for my email address):

http://www.netlib.org/a/loess

These fixes would be relatively simple to incorporate into R's  
version of loessf.f

Alternatively, a quick check would be for someone to compile the  
source package at https://centauri.stat.purdue.edu:98/loess/loess_0.4-1.tar.gz 
 and test it on windows. Though this package incorporates this and a  
few other fixes, please be aware that it the routines are converted  
to C and thus there is a slight performance hit compared to the  
fortran.

Hope this helps,
Ben

Ryan Hafen wrote:

That is true - good point.

lp1 <- predict(loess(y ~ x, degree=0))
lp2 <- predict(loess(y ~ x, degree=0,  
control=loess.control(surface="direct")))
sort(abs(lp1-lp2))

It appears that the interpolating fit is correct at the vertices.   
I know when degree>=1, the interpolation uses the slopes of the  
local fits to get a better approximation.  Perhaps it's still  
trying to do this with degree=0 but the slopes aren't available.   
And we have just been lucky in the past with uninitialized values?   
If this is the problem it would probably be very simple to fix and  
I'd love to see degree=0 stay.  I will see if I can figure it out.


On Mar 5, 2009, at 6:01 PM, Greg Snow wrote:

I see the same problem on Windows XP.

But if I run loess with surface='direct' then the results are  
correct.  So it looks like the problem comes from the smoothing/ 
interpolating, not the main loess algorithm.

-- 
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
greg.snow at imail.org
801.408.8111

-----Original Message-----
From: r-devel-bounces at r-project.org [mailto:r-devel-bounces at r-
project.org] On Behalf Of Ryan Hafen
Sent: Thursday, March 05, 2009 7:43 AM
To: Prof Brian Ripley
Cc: Uwe Ligges; Berwin A Turlach; r-devel at stat.math.ethz.ch; Peter
Dalgaard
Subject: Re: [Rd] bug (PR#13570)


On Mar 5, 2009, at 7:59 AM, Prof Brian Ripley wrote:

On Thu, 5 Mar 2009, Peter Dalgaard wrote:

Prof Brian Ripley wrote:

Undortunately the example is random, so not really reproducible
(and I
see nothing wrong on my Mac). However, Linux valgrind on R- 
devel is
showing a problem:

==3973== Conditional jump or move depends on uninitialised  
value(s)
==3973==    at 0xD76017B: ehg141_ (loessf.f:532)
==3973==    by 0xD761600: lowesa_ (loessf.f:769)
==3973==    by 0xD736E47: loess_raw (loessc.c:117)

(The uninitiialized value is in someone else's code and I  
suspect
it was
either never intended to work or never tested.)  No essential
change has
been made to the loess code for many years.

I would not have read the documentation to say that degree = 0  
was

a

reasonable value. It is not to my mind 'a polynomial surface',  
and
loess() is described as a 'local regression' for degree 1 or 2  
in
the
reference.  So unless anyone wants to bury their heads in that
code I
think a perfectly adequate fix would be to disallow degree = 0.
(I vaguely recall debating allowing in the code ca 10 years  
ago.)

The code itself has

if (!match(degree, 0:2, 0))
    stop("'degree' must be 0, 1 or 2")

though. "Local fitting of a constant" essentially becomes kernel
smoothing, right?

I do know the R code allows it: the question is whether it is  
worth
the effort of finding the problem(s) in the underlying c/dloess
code, whose manual (and our reference) is entirely about 1 or  
2.  I
am concerned that there may be other things lurking in the  
degree=0
case if it was never tested (in the netlib version: I am sure it  
was
only minmally tested through my R interface).

I checked the original documentation on netlib and that says

29      DIM     dimension of local regression
             1               constant
             d+1             linear   (default)
             (d+2)(d+1)/2    quadratic
             Modified by ehg127 if cdeg<tdeg.

which seems to confirm that degree = 0 was intended to be allowed,
and what I dimly recall from ca 1998 is debating whether the R  
code
should allow that or not.

If left to me I would say I did not wish to continue to support
degree = 0.

True.  There are plenty of reasons why one wouldn't want to use
degree=0 anyway.  And I'm sure there are plenty of other simple  
ways
to achieve the same effect.

I ran into the problem because some code I'm planning on  
distributing
as part of a paper submission "blends" partway down to degree 0
smoothing at the endpoints to reduce the variance.  The only bad
effect of disallowing degree 0 is for anyone with code depending on
it, although there are probably few that use it and better to  
disallow
than to give an incorrect computation.  I got around the problem by
installing a modified loess by one of Cleveland's former students:
https://centauri.stat.purdue.edu:98/loess/
(but don't want to require others who use my code to do so as  
well).

What is very strange to me is that it has been working fine in
previous R versions (tested on 2.7.1 and 2.6.1) and nothing has
changed in the loess source but yet it is having problems on 2.8.1.
Would this suggest it not being a problem with the netlib code?

Also strange that it reportedly works on Linux but not on Mac or
Windows.  On the mac, the effect was much smaller. With windows, it
was predicting values like 2e215 whereas on the mac, you would  
almost
believe the results were legitimate if you didn't think about the  
fact
that a weighted moving average involving half the data shouldn't
oscillate so much.

If the consensus is to keep degree=0, I'd be happy to help try to  
find
the problem or provide a test case or something.  Thanks for  
looking
into this.

Ryan

On Thu, 5 Mar 2009, Uwe Ligges wrote:

Berwin A Turlach wrote:

G'day Peter,

On Thu, 05 Mar 2009 09:09:27 +0100
Peter Dalgaard <p.dalgaard at biostat.ku.dk> wrote:

rhafen at stat.purdue.edu wrote:

<<insert bug report here>>

This is a CRITICAL bug!!!  I have verified it in R 2.8.1 for

mac

and for windows.  The problem is with loess degree=0  
smoothing.
For example, try the following:

x <- 1:100
y <- rnorm(100)
plot(x, y)
lines(predict(loess(y ~ x, degree=0, span=0.5)))

This is obviously wrong.

Obvious? How? I don't see anything particularly odd (on  
Linux).

Neither did I on linux; but the OP mentioned mac and  
windows. On
windows, on running that code, the lines() command added a  
lot of
vertical lines; most spanning the complete window but some  
only
part.
Executing the code a second time (or in steps) gave sensible
results. My guess would be that some memory is not correctly
allocated or
initialised.  Or is it something like an object with storage  
mode
"integer" being passed to a double?  But then, why doesn't it
show on
linux?

Happy bug hunting.  If my guess is correct, then I have no  
idea
how to
track down such things under windows.....

Cheers,

 Berwin

Please can you folks try under R-devel (to be R-2.9.0 in a  
couple
of
weeks) and report if you still see it. I do not under R-devel
(but do
under R-release), so my guess is that something called by  
loess()
has
been fixed in the meantime.

Moreover it is not the plot stuff that was wrong under R-2.8.1
(release) but the loess computations.

Uwe Ligges

-- 
O__  ---- Peter Dalgaard             ?ster Farimagsgade 5, Entr.B
c/ /'_ --- Dept. of Biostatistics     PO Box 2099, 1014 Cph. K
(*) \(*) -- University of Copenhagen   Denmark      Ph:  (+45)
35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)              FAX: (+45)
35327907

-- 
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