bug in plot.acf (PR#8705)

2006/3/24, maechler at stat.math.ethz.ch <maechler at stat.math.ethz.ch>:

"Duncan" == Duncan Murdoch <murdoch at stats.uwo.ca>
    on Fri, 24 Mar 2006 13:55:03 +0100 (CET) writes:

    Duncan> (Moved from r-devel to r-bugs)
    Duncan> On 3/24/2006 5:03 AM, Antonio, Fabio Di Narzo wrote:

    >> Hi all.
    >> There's a bug in plot.acf, when plotting acf for multivariate time

series.

    >> Here a reproducible example:
    >>
    >> X <- rnorm(1000)
    >> Y <- -X + rnorm(1000, sd=0.6)
    >> Z <- cbind(X,Y)
    >>
    >> In
    >> acf(Z)
    >> cross-correlation plot y-axis is limited to 0-1. But:
    >> acf(Z, ylim=c(-1,1))
    >> shows that there was a negative correlation, that was cut away in

the

    >> previous plot.
    >>
    >> I've seen the error is trivial. There's something like:
    >>
    >> for(each pair of univariate time series) {
    >> if(is.null(ylim)) {
    >> ...#set ylim properly
    >> }
    >> ...
    >> }
    >>
    >> in plot.acf code, so that in the first iteration the ylim par is

properly

    >> set to about c(0,1), but in
    >> subsequent interations, ylim is no more NULL, and the old, unproper

ylim

    >> specification
    >> remains.

    Duncan> Thanks for noticing this.  It's easy to fix, but
    Duncan> before I do, I'd like an opinion on the proper fix.
    Duncan> Should all the plots use the same ylim, or in the
    Duncan> case where it is unspecified, should they each
    Duncan> choose their own?  I can see arguments for both
    Duncan> possibilities.

    Duncan> Duncan Murdoch

I'd vote for the first one, a common y-scale.
Only that is also consistent with the behavior of explicitly
speficied 'ylim'.

Martin

This sounds to me to be the best, also for visual comparison of multiple
plots at once (having them all on the same screen...).
Antonio, Fabio Di Narzo.