R devel - May 2006 - Noncentral dt() with tiny 'x' values (PR#8874)

Full_Name: Mike Meredith
Version: 2.3.0
OS: WinXP SP2
Submission from: (NULL) (210.195.228.29)



Using dt() with a non-centrality parameter and near-zero values for 'x'
results
in erratic output. Try this:

tst <- c(1e-12, 1e-13, 1e-14, 1e-15, 1e-16, 1e-17, 0)
dt(tst,16,1)

I get:  0.2381019 0.2385462 0.2296557 0.1851817 0.6288373 3.8163916 (!!)
0.2382217

The 0.238 values are okay, the others nonsense, and they cause confusing spikes
on plots of dt() vs 'x' if 'x' happens to include tiny values.
(Other values of
df and ncp also malfunction, but not all give results out by an order of
magnitude!)

I'm using the work-around dt(round(x,10),...), but dt() should really take
care
of this itself.

Regards,  Mike.

>>>>> "MikeMer" == meredith  <meredith at
easynet.co.uk>
>>>>>     on Thu, 18 May 2006 03:52:51 +0200 (CEST) writes:
    MikeMer> Full_Name: Mike Meredith
    MikeMer> Version: 2.3.0
    MikeMer> OS: WinXP SP2
    MikeMer> Submission from: (NULL) (210.195.228.29)



    MikeMer> Using dt() with a non-centrality parameter and near-zero values
for 'x' results
    MikeMer> in erratic output. Try this:

    MikeMer> tst <- c(1e-12, 1e-13, 1e-14, 1e-15, 1e-16, 1e-17, 0)
    MikeMer> dt(tst,16,1)

    MikeMer> I get:  0.2381019 0.2385462 0.2296557 0.1851817 0.6288373
3.8163916 (!!)
    MikeMer> 0.2382217

I get quite different values (several '0', BTW), which just
confirms the erratic nature.

As often, plots give even a clearer picture:

if(!require(sfsmisc)) # pkg 'sfsmisc'
    lseq <- function (from, to, length) 
         2^seq(log2(from), log2(to), length.out = length)
x <- lseq(1e-3, 1e-33, length= 301)

plot(x, dt(x, df=16, ncp=1),   type = "o", cex=.5, log =
"x")
plot(x, dt(x, df=16, ncp=0.1), type = "o", cex=.5, log =
"x")
plot(x, dt(x, df= 3, ncp=0.1), type = "o", cex=.5, log =
"x")



    MikeMer> The 0.238 values are okay, the others nonsense, and
    MikeMer> they cause confusing spikes on plots of dt() vs 'x'
    MikeMer> if 'x' happens to include tiny values. (Other
    MikeMer> values of df and ncp also malfunction, but not all
    MikeMer> give results out by an order of magnitude!)

I think almost all do, once you start looking at plots like the above.

    MikeMer> I'm using the work-around dt(round(x,10),...), but
    MikeMer> dt() should really take care of this itself.

or actually rather do something more smart; the cutoff at 1e-10
is quite crude.

Note that this is not a new bug at all; but rather as old as
we have dt(*, ncp= .) in R.

Thanks for reporting it!
Martin Maechler, ETH Zurich