Notice that if one does not set xlim and ylim, the same happens, where autoscaling would be an expected outcome, as is part of issue #10497.

If the clip_on argument is set to True,

plt.annotate(xy=(3, 3), s='test', clip_on=True)

the output is as expected.

A confusing bit is that the annotations have a annotation_clip argument in addition to the clip_on. This allows some extra tweaking on when to show or not to show the annotation. But it is ignored for the tight_bbox calculation.

So this is purely an issue of setting defaults? I would then suggest to go from clip_on=False, annotation_clip=True to either clip_on=True, annotation_clip=True or clip_on=False, annotation_clip=False, because there is no reason to make place on the canvas for an annotation that is not shown.

@ImportanceOfBeingErnest There is a catch: Why does this not occur when using plt.savefig?

import matplotlib.pyplot as plt
plt.xlim(0,1)
plt.ylim(0,1)
plt.annotate(xy=(3, 3), s='test')
plt.savefig("test.png")

No, it's not an issue of the correct defaults. The issue is that annotation_clip is considered only for finding out whether or not to draw the annotation, but not for finding out if it is to be taken into account for the calculation of the tight_bbox.

The issue only occurs if you use plt.savefig("test.png", bbox_inches="tight") (which is what the jupyter/IPython inline backend does by default).

In principle the defaults are fine: clip_on=False and annotation_clip=None. Those ensure that annotations that are defined in coordinates other than data coordinates are still shown outside the axes, which is what most people would expect.

So possible solutions are:

• In the tight_bbox calculations, check if an artist is an annotation, or has a _check_xy attribute, and include it or not depending on its return. Drawback of this: It makes drawing less efficient and is in principle bad style, because in all other cases we would rather ask the artist if it to be included or not through the clip_on property.
• Synchronize the clip_on with the annotation_clip. Meaning, clip_on might change its state depending on the current value of annotation_clip and the coordinate system in use. This sounds a lot better, but is harder to implement. Also it's not clear which one should take precedence in case both are defined.

As of my understanding while quickly looking through the source, lib/matplotlib/artist.py uses clip_on in https://github.com/matplotlib/matplotlib/blob/master/lib/matplotlib/artist.py#L268, so more or less any value for clip_on other than None or False is True. Therefore, wouldn't it be possible to introduce a new state to clip_on, so it stores clip_on=True, clip_on=False or clip_on=Annotation.ANNOTATION_CLIP, where the latter replaces annotation_clip=True? Then you could not have a weird combination of clip_on and annotation_clip, as their functionality is combined into one parameter.
This idea probably needs polishment, but it would help not having to synchronize multiple parameters.

Due to the drawbacks of both possible solutions to this, I don't think this will be solved anytime soon. Also, one may consider it as a documentation issue, clarifying the role of clip_on and annotation_clip.

Hence, relabeling and remilestoning.

In principle the defaults are fine: clip_on=False and annotation_clip=None. Those ensure that annotations that are defined in coordinates other than data coordinates are still shown outside the axes, which is what most people would expect.

Sorry for reacting so late to this, but in original example, the annotation is NOT shown outside the axes, although I do not change the defaults.

Yes, but you can easily get the annotation to show up using

plt.annotate(xy=(3, 3), s='test', annotation_clip=False)

So the question is, whether it's worth putting in a hacky solution with some heuristics about whether or not to show the annotation in case bbox_inches="tight" is used.

annotation_clip : bool or None, optional
    Whether to draw the annotation when the annotation point *xy* is
    outside the axes area.

    - If *True*, the annotation will only be drawn when *xy* is
      within the axes.
    - If *False*, the annotation will always be drawn.
    - If *None*, the annotation will only be drawn when *xy* is
      within the axes and *xycoords* is 'data'.

So, by default its none, and its not drawn. But because its not clipped using the usual clipping argument, its still included in the tight_layout calculation. I'm not sure I understand the subtleties of why annotations have multiple clip arguments, or what behaviour is particularly desired here.

@ImportanceOfBeingErnest I'm not quite sure I get this. When the default for annotation_clip is not to show the annotation outside plot borders, should not then the default for clip_on be to clip outside plot borders? Not to show the annotation outside but reserving space for it seems like a confusing default.
This would also improve consistency with every other plot item (like scatterplots, line plots etc.), that are clipped outside plot borders.

The point is, you do want to show the annotations outside the axes in cases the anchor point is in data coordinates and still inside the axes, or in cases where you have used coordinates, other than data coordinates. Maybe the below animation makes this clearer.

Thank you for the explanation and the work of doing a GIF!
I think I might understand the reasoning, and have a suggestion on how to fix this. To discuss it, I created PR #15096.

In #15096 @jklymak asked

Is there any use to annotation_clip and clip_on to ever be different?

As answer, here are all combinations:

OK, but I don't think we ever want the clip_on=True result. Is there a practical use for that?

I'm not sure. Maybe in cases where you don't have any arrow? What's the alternative, though? Giving clip_on a third state None that is only used by annotations? But that would still require to somehow special-casing annotations in the bbox_inches="tight"- algorithms, right?

Hi, just to give a feedback (I hope this is the appropriate place) I agree with @jklymak: the case clip_on=True is weird and, I would argue, actually harmful. Indeed, I thought that the behaviour of annotation_clip=True would be achieved with clip_on=True which I think is the reasonable expectation. I think it makes a lot of sense to consider the annotation as a whole.

I do not know what changed but I just tried this with Matplotlib v3.8.4 in a notebook and got the expected result, i.e. the figure is not resized to include the invisible annotation.

import matplotlib.pyplot as plt

plt.xlim(0,1)
plt.ylim(0,1)
plt.annotate('test', xy=(3, 3))

This was fixed by #18772.