gnuplot-info

[Gnuplot-info] int() strange behaviour

[Simon G. <tar...@gm...>]

Hi all,
I'm using Gnuplot 4.6 patchlevel 0 on Linux x86_64
and I get a strange result when using int() function:
( I see that also in previous versions)

gnuplot> print 19.2/0.2
96.0
gnuplot> print int(19.2/0.2)
95
gnuplot> print int(96.0)
96

I don't know if it is required by some other feature.
I think it looks a little bit strange and mathematically wrong, isn't it?
I'm missing something?
Thank you

Re: [Gnuplot-info] int() strange behaviour

[Hans-Bernhard B. <HBB...@t-...>]

On 06.07.2012 12:10, Simon Garfunkel wrote:
> Hi all,
> I'm using Gnuplot 4.6 patchlevel 0 on Linux x86_64
> and I get a strange result when using int() function:
> ( I see that also in previous versions)
>
> gnuplot> print 19.2/0.2
> 96.0

Your problem is not actually with int(), it's with floating point 
numbers in general, and the fact that your 'print' command doesn't 
display all there is to know.  Try this instead:

gnuplot> print sprintf("%.16f", 19.2/0.2)
95.9999999999999858

or this:

gnuplot> print 19.2/0.2 - 96
-1.4210854715202e-014

As you see, your result is not quite as exactly equal to 96.0 as the 
simple `print' command makes it out to be.

> I don't know if it is required by some other feature.

It is.  The feature in question is most easily remembered like this:

	In computers 10.0 times 0.1 is hardly ever 1.0

Re: [Gnuplot-info] int() strange behaviour

[Simon G. <tar...@gm...>]

Ok, thank you for being kind with my ignorance.
Anyway in my particular situation these "missed approximations"
have a disastrous effect, since I'm creating classes for an hystogram:

gnuplot> print int(19.1/0.1)
191
gnuplot> print int(19.2/0.1)
191
gnuplot> print int(19.3/0.1)
193
gnuplot> print int(19.4/0.1)
193
gnuplot> print int(19.5/0.1)
195
gnuplot> print int(19.6/0.1)
196
gnuplot> print int(19.7/0.1)
196
gnuplot> print int(19.8/0.1)
198

Undoubtedly unconvincing classes.
I'm bypassing it using

bind(x, l) = int(x/l) * l
plot 'data' using ( bind($1 * 10**6, width * 10**6) * 10**(-6) ): [...]

I think that for my particular data (19.1, 19.5, 19,6, 19.2, ... ) can be
used also

plot 'data' using ( bind($1 +0.05, width) ): [...]

or even the suggested "+ eps".
Well, this was my really problem!
Thank you again for explanations

Re: [Gnuplot-info] int() strange behaviour

[Hans-Bernhard B. <HBB...@t-...>]

On 07.07.2012 01:45, Simon Garfunkel wrote:

> Anyway in my particular situation these "missed approximations"
> have a disastrous effect, since I'm creating classes for an hystogram:
>
> gnuplot> print int(19.1/0.1)
> 191
> gnuplot> print int(19.2/0.1)
> 191
[...]
> Undoubtedly unconvincing classes.

That's because you put your class boundaries exactly on those points 
where you want to check.  In effect you're comparing floating-point 
numbers for exact equality, which is never a good idea.  Hence that 
"hardly ever 1.0" line I quoted earlier.

Yes, you could use an offset.  But ultimately nothing can really help. 
Wherever you place your class boundaries, sometimes the inevitable 
floating-point round-off _will_ move data right across that boundary.