Plots.jl: Problem with :log10 scale for small numbers

[fedxa]

If you plot in :log10 scale, and output falls below 1e-16, some craziness happen:

x=logspace(-9,-7,20);plot(x,x.^2,xscale=:log10,yscale=:log10)

Correct plot is just a line, as you can see:

x=logspace(-9,-7,20);plot(log10.(x),log10.(x.^2))

The bug is very manifest in PyPlot backend. The GR backends just fails to produce the plot at all. Plotly backend works correctly.

[apalugniok]

For PyPlot
https://github.com/JuliaPlots/Plots.jl/blob/master/src/backends/pyplot.jl#L906

kw[Symbol(:linthresh,letter)] = NaNMath.max(1e-16, py_compute_axis_minval(axis))

seems fishy.

https://matplotlib.org/api/_as_gen/matplotlib.axes.Axes.set_xscale.html
matplotlib documentation for linthresh:
inthreshx/linthreshy:
The range (-x, x) within which the plot is linear (to avoid having the plot go to infinity around zero).

Hope this helps I'll try to look into this a bit more when I have time :D

[fedxa]

Oh, cool! "max"->"min" solves it. Will make a pull request. Let us see what is the problem with GR.

Anyway, a strange option for matplotlib...

[apalugniok]

I wonder if the whole line is even necessary at this point. Since if we take the min every time the whole scale will be a logscale which should be the case if the linthresh keyword isn't provided, although I'm no matplotlib expert so I don't know if that's how it'd behave.

But I think min() is a good solution until we get to the bottom of this :D

[fedxa]

That was my first idea -- just remove the line. Somehow (no idea why!) in this case the default for linthresh turns to be something like 0.1, which is completely crazy.

[apalugniok]

Oh wow matplotlib really throwing a curveball there... Seems like sticking with min() is the way then.

[fedxa]

It seems more complicated :) In plain PyPlot.jl I don't see any problems. I am lazy to check with python :) So, there is something in the way matplotlib is called in the backend.

using PyPlot
x=logspace(-9,-7,20);
plot(x,x.^2)
loglog()

[apalugniok]

So I tried a few things with GR. I get a load of error GKS: Rectangle definition is invalid in routine SET_WINDOW.
It seems like it comes back to the GR.setwindow(xmin,xmax,ymin,ymax) function.
I tried running this function on it's own

using GR
GR.setwindow(1e-9, 1e-7, 1e-18, 1e-14) # This gives the error
GR.setwindow(1e-9, 1e-7, 1e-18, 1e-11) # This doesn't

Similarly

using Plots
x=logspace(-9, -5.5, 20) 
plot(x, x.^2, xscale=:log10, yscale=:log10) 

This should call GR.setwindow() as is in the second case above which gave no error and it produces the expected plot for me with no trouble... Confusing

[mkborregaard]

@jheinen can you easily see what's the issue here?

[jheinen]

@mkborregaard : just fixed in https://github.com/jheinen/gr/commit/d2eb09b532ec7c7bb41a4e45d93ed1b92bcb4757

[mkborregaard]

Fantastic :heart: BTW the link doesn't work - where do you mean?

[jheinen]

... sorry for the broken link. Should be correct now.

[mkborregaard]

Ah! great! Closing :-)