fix colorbar corner cases (singleton)

[t-bltg]

Description

Fix #3470.
Fix #3174.

• add an explicit error message and test on invalid colorrange as suggested in Colorbar breaks with min limit == max limit #3174 (comment);
• rework Fix Colorbar with all zero heatmap #1739 (change arbitrary fallback limits, closer to PyPlot);
• add a similar fix for contourf and tricontourf (as for heatmap in Fix Colorbar with all zero heatmap #1739), add tests.

With this PR:

using CairoMakie

x, y = [0, 1], [0, 2]
z = zeros(2, 2)

fig, ax, pl = heatmap(x, y, z)
Colorbar(fig[1, 2], pl)
save("zero-heatmap.png", fig)
display(fig)

fig, ax, pl = contourf(x, y, z)
Colorbar(fig[1, 2], pl)
save("zero-contourf.png", fig)
display(fig)

x, y = randn(10), randn(10)
z = zero(x)
fig, ax, pl = tricontourf(x, y, z)
Colorbar(fig[1, 2], pl)
save("zero-tricontourf.png", fig)
display(fig)

heatmap

contourf

tricontourf

Type of change

Delete options that do not apply:

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)

Checklist

• Added an entry in NEWS.md (for new features and breaking changes)
• Added or changed relevant sections in the documentation
• Added unit tests for new algorithms, conversion methods, etc.
• Added reference image tests for new plotting functions, recipes, visual options, etc.

[jkrumbiegel]

Ah that's good to fix all of those. What do you think, my feeling is that the super small limits are a bit confusing. My strategy so far, whenever I've thought about it specifically, was to either do the 0 to x interval when x is not zero, or 0 to 1 when x is 0. This does mean that the color you get is the top one when your value is positive, or the bottom one when it's negative or zero. But then at least you don't think long about what kind of range you've gotten I think.

[t-bltg]

Ah that's good to fix all of those. What do you think, my feeling is that the super small limits are a bit confusing.

Yeah, the more I think about this, I'm neither satisfied with small values or the arbitrary (-0.5, 0.5) implemented in #1739.

either do the 0 to x interval when x is not zero, or 0 to 1 when x is 0

Good idea, here is the updated PR, IIUC:

using CairoMakie

for v ∈ (0, 1)
  x, y = [0, 1], [0, 2]
  z = fill(float(v), length(x), length(y))
  fig, ax, pl = heatmap(x, y, z)
  Colorbar(fig[1, 2], pl)
  save("$v-heatmap.png", fig)
  display(fig)

  fig, ax, pl = contourf(x, y, z)
  Colorbar(fig[1, 2], pl)
  save("$v-contourf.png", fig)
  display(fig)
end

for v ∈ (0, 1)
  x, y = randn(10), randn(10)
  z = fill(float(v), length(x))
  fig, ax, pl = tricontourf(x, y, z)
  Colorbar(fig[1, 2], pl)
  save("$v-tricontourf.png", fig)
  display(fig)
end

0-heatmap

0-contourf

0-tricontourf

1-heatmap

1-contourf

1-tricontourf

[jkrumbiegel]

Hm I don't really like the contourf and tricontourf behavior. We have these colorbars with only one color (because they are categorical with one category but you don't easily understand that) but a range of values mapped to that color. So we don't know which uniform value we're plotting. While with heatmap, we can still recover the value, it's just that the range has to be arbitrarily chosen.

So far, I've opted for behavior in Makie where you get an error when a singular value is supposed to be treated as a range. I find that more explicit (even though some error messages might be better) and it forces users to be clear what kind of bins they want. I do get that it's a bit inconvenient now and then. What do matplotlib and ggplot do here?

[t-bltg]

This is what pyplot does:

import pylab as plt
import numpy as np

x, y = [0, 1], [0, 2]
for v in (0, 1):
    z = np.full((len(x), len(y)), float(v))

    fig, ax = plt.subplots()
    ct = ax.contourf(x, y, z)
    fig.colorbar(ct)
    fig.savefig(f'{v}-contourf.png')

    fig, ax = plt.subplots()
    ct = ax.pcolormesh(x, y, z)
    fig.colorbar(ct)
    fig.savefig(f'{v}-heatmap.png')

    # fig, ax = plt.subplots()
    # ct = ax.tricontourf(x, y, z)
    # fig.colorbar(ct)
    # fig.savefig(f'{v}-tricontourf.png')

i.e. the singleton is extended from both side by an epsilon: (val - eps, val] ∪ (val, val + eps] to form the colormap.

Not sure what the right choice here is ...