[datatable-help] sorting on floating point column

[Arunkumar Srinivasan]

Matthew, 

So what's the resolution here? Is it okay to sort in the "proper" order on the key column but use *machine tolerance* for subset on key column?


Arun


On Tuesday, April 30, 2013 at 4:26 PM, Arunkumar Srinivasan wrote:

> Matthew, 
> 
> Precisely. That's what I was thinking as well. But was hesitant to tell as I dint know how complex it would be to implement / change it. Since the join requires tolerance, sorting could be still done in the "right" order (by disregarding tolerance during sort). 
> 
> Arun
> 
> 
> On Tuesday, April 30, 2013 at 4:22 PM, Matthew Dowle wrote:
> 
> >  
> > Maybe it doesn't actually need to sort within machine tolerance.   If it was precise, the sort would be faster, that's for sure.  But at the time,  I remember thinking that it should preserve the order of rows within a group of values within machine tolerance (e.g. 3.99999999, 4.00000001, 3.99999999 should be consider 4.0 and order of those 3 rows maintained).   But maybe sorting them to 3.99999999, 3.99999999, 4.00000001 is ok as it's just the join that should be within machine tolerance?
> > Interested in how fast order(y) is, though.  Compared to data.table sorting of doubles.
> > Matthew
> >  
> > On 30.04.2013 15:16, Arunkumar Srinivasan wrote:
> > > Matthew,
> > > I see. I din't think about tolerance. Although
> > > dt[with(dt, order(y)), ] 
> > > seems to do the task right (similar to data.frame). I'm glad that I don't have to convert to data.frame to perform the order. I am not keying by this column. Unless one needs this column for keying, I don't think a tolerance option is essential. Although, having it definitely would be only nicer.
> > > Arun
> > > 
> > > 
> > > On Tuesday, April 30, 2013 at 4:09 PM, Matthew Dowle wrote:
> > > 
> > > >  
> > > > Hi,
> > > > data.table sorts double within machine tolerance :
> > > > > sqrt(.Machine$double.eps)
> > > > [1] 1.490116e-08
> > > > > 
> > > >  
> > > > i.e. numbers closer than this are considered equal.
> > > >  
> > > > Otherwise we wouldn't be able to do things like DT[.(3.14)].
> > > >  
> > > > I had a quick look, see arguments of data.table:::ordernumtol which takes "tol" but there is no option provided (yet) to change this. Do we need one?
> > > >  
> > > > In the examples section of one of the help pages it has an example which generates a series of numers very close together using pi. Note that your numbers are both close together, and, very close to 0.
> > > >  
> > > > Matthew
> > > >  
> > > > On 30.04.2013 14:52, Arunkumar Srinivasan wrote:
> > > > > Hi there,
> > > > > I just saw something strange when I was sorting a column of p-values. I checked the data.table bug tracker for words "sort" and "floating point" and there were no hits for this case. There's a bug for "integer 64" sort on a column though.
> > > > > So, here's a reproducible example. I'd be glad to file a bug, if it is and be corrected if it's something I am doing wrong.
> > > > > set.seed(45)
> > > > > dt <- data.table(x=sample(50), y= sample(c(seq(0, 1, length.out=1000), 7000000:7000100), 50)/1e7)
> > > > > head(dt)
> > > > >     x            y
> > > > > 1: 32 5.395395e-08
> > > > > 2: 16 6.956957e-08
> > > > > 3: 12 2.142142e-08
> > > > > 4: 18 5.855856e-08
> > > > > 5: 17 6.216216e-08
> > > > > 6: 14 5.025025e-08
> > > > > setkey(dt, "y") # sort by column y
> > > > > head(dt, 10)
> > > > >      x            y
> > > > >  1: 47 1.401401e-09
> > > > >  2: 12 2.142142e-08
> > > > >  3: 24 1.391391e-08
> > > > >  4: 43 9.809810e-09 <~~~ obviously false
> > > > >  5:  1 2.932933e-08
> > > > >  6: 48 2.562563e-08
> > > > >  7: 49 1.891892e-08
> > > > >  8: 40 2.182182e-08
> > > > >  9:  9 7.307307e-09 <~~~ obviously false
> > > > > 10: 45 2.482482e-08
> > > > > 
> > > > > Best,
> > > > > Arun
> > > > > 
> > > > > 
> > > > > 
> > > > 
> > > >  
> > > >  
> > > > 
> > > > 
> > > 
> > > 
> > 
> >  
> >  
> > 
> > 
> > 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/datatable-help/attachments/20130502/ed261ea9/attachment.html>