[R] Fwd: ts.intersect bug?

[Prof Brian Ripley]

The problem was that a tolerance in window.default was abs(start)*ts.eps, 
which is no tolerance at all if start = 0.

So the special feature here was that the intersection of the series 
started at zero. I had already tested and committed a fix.

window.default does use a tolerance for en, just as Baz suggests it 
should.  The code is there to be read ....  (That tolerance was inadequate 
if end = 0, but has been changed.)

On Wed, 21 Sep 2005, Barry Rowlingson wrote:

> Antonio, Fabio Di Narzo wrote:
>
>> Ooops, bad example.
>> Try this instead:
>> a <- ts(runif(6500), start=0, freq=10)
>> b <- lag(a, 1)
>> c <- ts.intersect(a, b)
>>
>> Gives an error from .cbind.ts
>
> or even:
>
>  > a=ts(numeric(2600),freq=10,start=0);b=lag(a,1);c=ts.intersect(a,b)
> Error in "[<-"(`*tmp*`, , (1 + cs[i]):cs[i + 1], value = c(0, 0, 0, 0,  :
>         number of items to replace is not a multiple of replacement length
>
>  seems to be sensitive to the lengths of the time series:
>
>  > a=ts(numeric(2601),freq=10,start=0);b=lag(a,1);c=ts.intersect(a,b)
>  > (works fine)
>
>
>  Digging into the code, the window() function is returning a different
> length time series for each one in these failing cases. I reckon its a
> floating-point precision situation, where the last time series point
> should be included but the arithmetic precision of a 2600-long series at
> separation of 1/10 is leaving it out.
>
>  > st;en
>  [1] 0
>  [1] 259.8
>
>  > a=ts(numeric(2600),freq=10,start=0);b=lag(a,1)
>  > length(window(a,st,en))
>  [1] 2599
>  > length(window(b,st,en))
>  [1] 2598
>
>  - ts.intersect is trying to put these two time series together, and so
> fails. But:
>
>  > a=ts(numeric(2601),freq=10,start=0);b=lag(a,1)
>  > length(window(a,st,en))
>  [1] 2599
>  > length(window(b,st,en))
>  [1] 2599
>
>  - works.
>
>  Note that en is not precisely 259.8:
>
>  > en == 259.8
>  [1] FALSE
>  > en-259.8
>  [1] -5.684342e-14
>
>  I've computed 'en' as the .cbind.ts function does, and its not exactly
> 259.8. If it were, then it would work... Perhaps .cbind.ts should round
> to the nearest true time point or something...
>
>  Note that it fails in plenty of smaller cases too:
>
>  > a=ts(numeric(13),freq=10,start=0);b=lag(a,1);c=ts.intersect(a,b)
> Error in "[<-"(`*tmp*`, , (1 + cs[i]):cs[i + 1], value = c(0, 0, 0, 0,  :
>         number of items to replace is not a multiple of replacement length
>
>  Seems to not like 13s and 10s and integer products thereof (2600,
> 6500). Are you superstitious?
>
> Baz
>
> PS without counting, how many letters are there in 'superstitious'?
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
>

-- 
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