[GRASS-dev] [bug #2790] (grass) r.bilinear result is rounded to .5

[Glynn Clements]

Maciek Sieczka via RT wrote:

> Indeed r.bilinear doesn't produce atrifact .5 and stuff anymore, but it's
> output is slightly different than the output of the new r.resamp.interp
> method=bilinear. 
> 
> See spearfish60 example:
> 
> $ g.region rast=slope -a
> $ g.region res=10.95 -a
> $ r.bilinear input=slope output=slope_rbil
> $ r.resamp.interp input=slope output=slope_rinterp method=bilinear
> 
> $ r.mapcalc 'diff=slope_rbil-slope_rinterp'
> 
> There is a minimal difference...
> 
> $ r.info -r diff
> min=-0.000000
> max=0.000000
> 
> ... which can be seen on display as a kind off a regular pattern, and if
> exagerated x10^15 it is noticable as:
> 
> $ r.mapcalc 'diff_exag=diff*10.0^15.0'
> $ r.info diff_exag -r
> min=-7.105427
> max=7.105427
> 
> ?

That level of error can easily be attributed to differences in
rounding. It's by no means clear which of the two (if either) is more
accurate.

For IEEE FP, DBL_EPSILON is ~2.2e-16, so a value of ~7.1e-15 is
roughly 32 * DBL_EPSILON. "r.info slope" gives the range as roughly
0-52.5, so I'd expect to see errors of +/- 1.2e-14 (i.e. around 63%
more than you're actually getting).

Differences at that scale are irrelevant; there's no way that your
source data is going to be accurate to 16 digits (2.2e-16 is
equivalent to measuring the equatorial circumference to an accuracy of
10 nanometres).

-- 
Glynn Clements <glynn at gclements.plus.com>