Confirmed. I've updated the title since this is explicitly a Collection issue. Adding the patch to the axes doesn't have this problem (It works as expected).

Even more explicitly, I think the problem is with the match_original keyword to PathPatch. Changing the collection creation to:

p = PatchCollection([c], facecolor='none')

Also solves the problem.

Thanks for the workaround!

In case anyone else has this issue, I found a slightly improved workaround, if one wants to keep all the other attributes (edgecolor, etc.) which is just to keep match_original=True and then call:

p.set_facecolor('none')

Interestingly with the pdf backend the line's alpha is bound to the alpha of the face color, i.e.:

p = PatchCollection([c], facecolors=((0, 0.1, 0.1, 0.1), ), edgecolor='green')

produces a circle with a green line, but the line has an alpha of 0.1. This is not the behaviour with Agg.

@jkseppan & @mdboom - I'm out of my depth now, any advice?

FWIW, changing the following helped remove the circle, though I don't fully understand the impacts:

--- a/lib/matplotlib/backend_bases.py
+++ b/lib/matplotlib/backend_bases.py
@@ -395,15 +396,15 @@ class RendererBase:
                 if Nlinestyles:
                     gc0.set_dashes(*linestyles[i % Nlinestyles])
             if rgbFace is not None and len(rgbFace) == 4:
-                if rgbFace[3] == 0:
-                    rgbFace = None
-                else:
+#                if rgbFace[3] == 0:
+#                    rgbFace = None
+#                else:
                     gc0.set_alpha(rgbFace[3])
                     rgbFace = rgbFace[:3]
             gc0.set_antialiased(antialiaseds[i % Naa])
             if Nurls:
                 gc0.set_url(urls[i % Nurls])

An interesting data-point, this bug does not happen in v1.1.1.

This arises because:

1. Objects that are reused (which is what happens in a collection) in PDF must have their stroke and fill hard coded -- so there's no way to create an XObject which is sometimes filled and sometimes not.

2. There is code in the PDF backend to determine whether the reused object it creates should be filled or not, but it ignores the alpha-values (created when fill is set to 'none').

3. PDF doesn't consider alpha part of the color, but uses alpha pushing and popping, so we can't just "fill" it with an invisible color.

I'm not sure how to fix this yet, but I thought that might be illuminating.

There is a broader bug here which is that collections can not have different alphas for different instances of the same reused object.

Since 1.1.1 was working, I'll try a git bisect to see if anything simple pops up, but I suspect this is due to an important refactoring of the collections handling in PDF that otherwise had more problems than we have now...

FWIWI, git bisect points at 4dd3de1.

The fact that this can be worked around by doing p.set_facecolor('none') on the PatchCollection suggests to me that it's not really a backend problem - or am I mistaken?

That workaround only works if all of the collection items are to have no fill. If the collection mixes filled and unfilled, we'd still have this problem.

I believe the attached patch fixes the issue. @astrofrog: Can you confirm?

The patch by @mdboom looks good to me on a first read. The logic for handling transparency correctly seems to be pretty complicated. It's probably partially because the Agg and PDF rendering models are not quite the same, but I can't help wondering if there isn't something that we could do in the frontend to simplify it.

In a first pass, I tried to create a separate XObject for each combination of (path_shape, fill_alpha, stroke_alpha), and then use the correct one when generating the use object commands. However, that just got extremely complicated really fast, so I ultimately arrived at this which is to use XObjects and use when they are a natural fit to the data, and fall back to just drawing each path independently when not. It's suboptimal, for sure, but it seems like a pretty unusual case to support in the first place, and I decided to do the simplest thing that's less likely to be buggy than to be "clever" about it.

Out of interest, what does the output look like with a PDF generated with the cairo backend?

@dmcdougall : Good question. The Cairo output is correct, but that's mainly because it doesn't do any optimizations (reusing of path objects in a collection).

@mdboom - this patch works for me - thanks for the quick fix!

👍 from me. Looks like a good fix with a good test. All that is missing is some documentation for the change 😉

@pelson: as you wish 😉

😄 thanks @mdboom