gnuplot-beta

cvs: log cb broken

[Petr M. <mi...@ph...>]

There is a difference between gnuplot 4.2.2 and cvs; 4.2.2 works OK for log 
cb axis, while cvs gives this error:

  "bug_logcb.gp", line 4: color axis has cb coord of -10; must be above 0 
  for log scale!

What has broken in cvs?


Script:

plot 'matrix.dat' matrix with image

set log cb
replot

File matrix.dat:
2 4 6 8
1 2 3 4
0 1 2 3
-2 -1 1 2
-4 -3 -2 -1
-6 -5 -4 -3

---
PM

Re: cvs: log cb broken

[Ethan M. <merritt@u.washington.edu>]

On Thursday 03 January 2008 04:18, Petr Mikulik wrote:
> There is a difference between gnuplot 4.2.2 and cvs; 4.2.2 works OK for log 
> cb axis, while cvs gives this error:
> 
>   "bug_logcb.gp", line 4: color axis has cb coord of -10; must be above 0 
>   for log scale!
> 
> What has broken in cvs?

But the error message is correct.  
The cb values go negative and therefore log scale cannot work.

I see that 4.2 does draw something, but I don't understand 
what it is drawing.  I'd say that 4.2 fails to detect the
error, while CVS is doing the right thing by reporting it.

 
> 
> Script:
> 
> plot 'matrix.dat' matrix with image
> 
> set log cb
> replot
> 
> File matrix.dat:
> 2 4 6 8
> 1 2 3 4
> 0 1 2 3
> -2 -1 1 2
> -4 -3 -2 -1
> -6 -5 -4 -3
> 
> ---
> PM

-- 
Ethan A Merritt

Re: cvs: log cb broken

[Petr M. <mi...@ph...>]

> >   "bug_logcb.gp", line 4: color axis has cb coord of -10; must be above 0 
> >   for log scale!
> > 
> > plot 'matrix.dat' matrix with image
> >
> > What has broken in cvs?
> 
> But the error message is correct.  
> The cb values go negative and therefore log scale cannot work.
> 
> I see that 4.2 does draw something, but I don't understand 
> what it is drawing.  I'd say that 4.2 fails to detect the
> error, while CVS is doing the right thing by reporting it.

No, the error message is not correct, it should autoscale the cb-value to a 
lower positive value, as it does in the same case for "with image" for 
logged y or z axes. There must have been some change in cvs which broke this 
for "matrix with image".

Gnuplot has been always excellent in handling negative ranges in log scales. 
For example,

	gnuplot> set xrange [-10:10]; set log y; plot 1/x

---
PM

Re: cvs: log cb broken

[Ethan M. <merritt@u.washington.edu>]

On Friday 04 January 2008 03:42, Petr Mikulik wrote:
> > >   "bug_logcb.gp", line 4: color axis has cb coord of -10; must be above 0 
> > >   for log scale!
> > > 
> > > plot 'matrix.dat' matrix with image
> > >
> > > What has broken in cvs?
> > 
> > But the error message is correct.  
> > The cb values go negative and therefore log scale cannot work.
> > 
> > I see that 4.2 does draw something, but I don't understand 
> > what it is drawing.  I'd say that 4.2 fails to detect the
> > error, while CVS is doing the right thing by reporting it.
> 
> No, the error message is not correct, it should autoscale the cb-value to a 
> lower positive value, as it does in the same case for "with image" for 
> logged y or z axes. There must have been some change in cvs which broke this 
> for "matrix with image".
> 
> Gnuplot has been always excellent in handling negative ranges in log scales. 

This is new to me, and does not seem to agree with the documentation:

help autoscale
    When autoscaling, the axis range is automatically computed and the dependent
    axis (y for a `plot` and z for `splot`) is scaled to include the range of the
    function or data being plotted.

You are saying that for log scales, only the range of the data greater than
some positive value epsilon is to be scaled?

I can see how this makes sense for functions, but it strikes me as being
incorrect for actual data.

Anyhow, I don't quite understand what the equivalent result would be for
plots "with image".  In the example you give below for log scale on y, only
points with f(x) > 0 are displayed on the plot.  Would the equivalent for
"with image" be that individual pixels are only drawn if the value of f(x,y)
is positive?   That is not what version 4.2 does.

And what would be the assignment of colors?  Would you limit the colors
used in the bar to the color range corresponding to positive Z values?
That would be the logical equivalent of drawing only the positive region
on y in the example below.  But that is not what 4.2 does.

> For example,
> 
> 	gnuplot> set xrange [-10:10]; set log y; plot 1/x


Anyhow, I don't have any immediate guesses as to what changed between
4.2 and current cvs.

-- 
Ethan A Merritt

Re: cvs: log cb broken

[Tait <gnu...@t4...>]

Petr Mikulik <mikulik_physics.muni.cz> said (on 2008/01/04):
> No, the error message is not correct, it should autoscale the cb-value to a
> lower positive value, as it does in the same case for "with image" for
> logged y or z axes. There must have been some change in cvs which broke this
> for "matrix with image".
>
> Gnuplot has been always excellent in handling negative ranges in log scales.
> For example,
>
>	gnuplot> set xrange [-10:10]; set log y; plot 1/x

I don't understand your example. The y axis range is auto-calculated
(correctly, it seems) as 0.1 to 1. There is no negative log range 
there, and the negative y-values associated with the x=[-10:0] range 
are ignored. If you try something like:

	gnuplot> set xrange [-10:10]; set yrange [-1:1]; set log y; plot 1/x;

You will, as expected, get the error "y range must be greater than 0 for
log scale." This seems to be the case in 4.2 as well.

Tait

Re: cvs: log cb broken

[Petr M. <mi...@ph...>]

> > Gnuplot has been always excellent in handling negative ranges in log scales.
> > For example,
> >
> >	gnuplot> set xrange [-10:10]; set log y; plot 1/x
> 
> I don't understand your example. The y axis range is auto-calculated
> (correctly, it seems) as 0.1 to 1. There is no negative log range 
> there, and the negative y-values associated with the x=[-10:0] range 
> are ignored. If you try something like:
> 
> 	gnuplot> set xrange [-10:10]; set yrange [-1:1]; set log y; plot 1/x;
> 
> You will, as expected, get the error "y range must be greater than 0 for
> log scale." This seems to be the case in 4.2 as well.

OK, I should have written "Gnuplot has been always excellent in autoscaling 
log axis ranges."

---
PM

Re: cvs: log cb broken

[Petr M. <mi...@ph...>]

> This is new to me, and does not seem to agree with the documentation:
> 
> help autoscale
>     When autoscaling, the axis range is automatically computed and the dependent
>     axis (y for a `plot` and z for `splot`) is scaled to include the range of the
>     function or data being plotted.
> 
> You are saying that for log scales, only the range of the data greater than
> some positive value epsilon is to be scaled?

Yes, it has always worked like that. Hitting "l" for log scale, for any plot 
or splot with at least few positive points, gnuplot produces a reasonable 
drawing.

> I can see how this makes sense for functions, but it strikes me as being
> incorrect for actual data.

Why? Negative values are e.g. electronic noise. The user is not interested 
in, but want to see the image in log scale of intensities.

> Anyhow, I don't quite understand what the equivalent result would be for
> plots "with image".  In the example you give below for log scale on y, only
> points with f(x) > 0 are displayed on the plot.  Would the equivalent for
> "with image" be that individual pixels are only drawn if the value of f(x,y)
> is positive?   That is not what version 4.2 does.
> 
> And what would be the assignment of colors?  Would you limit the colors
> used in the bar to the color range corresponding to positive Z values?
> That would be the logical equivalent of drawing only the positive region
> on y in the example below.  But that is not what 4.2 does.

It should work in 4.3 as 4.2 works: the cb min is 0.1 (the smallest positive 
data point value if "set autoscale fix"), and colors of smaller values are 
drawn same as those corresponding to this minimum.

It works everywhere like this for the cb axis. Try for example this:

	splot x*x-y*y*2 with pm3d
	set log cb
	replot

and the colors below the limit are black.
If you want to mask out some rectangles, then you need
	set log z; replot

Consequently, there is a bug in 4.3. 
Wasn't it introduced in your patch from 2007-12-18?
 
---
PM

Re: cvs: log cb broken

[Ethan M. <merritt@u.washington.edu>]

On Saturday 05 January 2008, Petr Mikulik wrote:
> Consequently, there is a bug in 4.3.
> Wasn't it introduced in your patch from 2007-12-18?

I don't think it is at all related. So far as I can determine,
the difference is due to a single line change from back in Aug 2007. 
The following patch would return it to 4.2 behavior.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
--- gnuplot-27aug2007/src/plot2d.c	2007-08-27 12:31:57.000000000 -0800
+++ gnuplot-cvs/src/plot2d.c	2008-01-05 13:41:58.000000000 -0800
@@ -223,7 +223,7 @@
     AXIS_INIT2D(SECOND_Y_AXIS, 1);
     AXIS_INIT2D(T_AXIS, 0);
     AXIS_INIT2D(R_AXIS, 1);
-    AXIS_INIT2D(COLOR_AXIS, 0);
+    AXIS_INIT2D(COLOR_AXIS, 1);
 
     t_axis = (parametric || polar) ? T_AXIS : FIRST_X_AXIS;
 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

However. I do not agree that in the case of image plots the 4.2 behavior is 
correct.  

The plot produced by 4.2 shows a color range in which black indicates a pixel 
value, in this case log(f(x,y)),  less than 0.1.
But these pixel values are not, in fact, less than 0.1.  They are entirely 
undefined because  log(x<0) is undefined.   I think that the correct behavior 
would be to not draw these pixels at all.  This is exactly analogous to your 
previous example of autoscaling    'plot [-10:10] log(x)'.   The y values 
corresponding to x<0  are not shown as zero;  they are not drawn at all.   

I take your point that one case in which pixel values are undefined is the 
case of taking the log of an experimentally measured value near 0 in the 
presence of finite experimental error.   But that is not the general case.  
For example, the matrix data file you posted at the start of this thread 
contained values evenly distributed on both sides of zero.  It was not 
obvious that the values were expected to be all positive.

-- 
Ethan Merritt
(on the road)