R devel - Jan 2019 - Bug in dt() function in stats - accuracy only 6 dp.

Hi



I believe the dt() function in stats has small areas of inaccuracy.



SmallRangeMostDetaileddf150.png shows this occurring in the 10-11th decimal
place.



MyD4 is computed using Myt_NC(), at the bottom of BugL.R - my function is very
slow due to using mpfr, but i believe

accurate to the accuracy of genhypergeo() which it uses - 14-15 dp.



LargeRangeDiffdf150.png, which is difference between dt() and my function, shows
spikes across -1.5 < x < 1.5.



I assumed probably only a problem for large df, but SmallRangeCloseZerodf20.png,
shows actually more of a problem with

df = 20.? Accuracy seems to be only to 6 decimal places.



Using R version 3.5.2.



Cheers



Mark
-------------- next part --------------
A non-text attachment was scrubbed...
Name: SmallRangeMostDetaileddf150.png
Type: image/png
Size: 1262560 bytes
Desc: not available
URL:
<https://stat.ethz.ch/pipermail/r-devel/attachments/20190109/20c85cda/attachment-0003.png>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: LargeRangeDiffdf150.png
Type: image/png
Size: 1262560 bytes
Desc: not available
URL:
<https://stat.ethz.ch/pipermail/r-devel/attachments/20190109/20c85cda/attachment-0004.png>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: SmallRangeCloseZerodf20.png
Type: image/png
Size: 1262560 bytes
Desc: not available
URL:
<https://stat.ethz.ch/pipermail/r-devel/attachments/20190109/20c85cda/attachment-0005.png>

>>>>> mark chappell via R-devel 
>>>>>     on Wed, 9 Jan 2019 13:10:26 +1100 writes:
    > Hi I believe the dt() function in stats has small areas of
    > inaccuracy.



    > SmallRangeMostDetaileddf150.png shows this occurring in
    > the 10-11th decimal place.



    > MyD4 is computed using Myt_NC(), at the bottom of BugL.R -
    > my function is very slow due to using mpfr, but i believe

    > accurate to the accuracy of genhypergeo() which it uses -
    > 14-15 dp.



    > LargeRangeDiffdf150.png, which is difference between dt()
    > and my function, shows spikes across -1.5 < x < 1.5.



    > I assumed probably only a problem for large df, but
    > SmallRangeCloseZerodf20.png, shows actually more of a
    > problem with
    > df = 20.? Accuracy seems to be only to 6 decimal places.

We did not get a single piece of R code.
Maybe you've wanted to attach the 'BugL.R' R script mentioned
above, but failed to use 'text/plain'  MIME encoding?

Note: If this is about using *non*central t-distribution (as the
 above "NC" suggests), then this is *not* a bug per se,
as the help page *does* mention that the acuracy there is reduced.

Still, I have created an R bugzilla account for you to report
there, i.e.,  https://bugs.r-project.org/

Best,
Martin


    > Using R version 3.5.2.

    > Cheers

    > Mark

    > [DELETED ATTACHMENT SmallRangeMostDetaileddf150.png, PNG image]
    > [DELETED ATTACHMENT LargeRangeDiffdf150.png, PNG image]
    > [DELETED ATTACHMENT SmallRangeCloseZerodf20.png, PNG image]