Bug summary

Noted while investigating#21917 (comment):Collection.draw callsCollection._prepare_points, and in some cases (ifnot self.get_transform().is_affine and len(paths) == len(self.get_transforms()) == len(facecolors) == ... == 1 and ..., the latter being the check for single-path-optimization) this results in the paths being transformed in the following order:

• first by the non-affine part of self.get_transform() (in_prepare_points)
• then bytransforms.Affine2D(trans[0]) + transform, i.e.
  • byself.get_transforms()[0] (which is always affine) (in the single-path-optimization check)
  • by the affine part of self.get_transform() (also in the single-path optimization check)

i.e.self.get_transforms()[0] gets inserted in the middle ofself.get_transform(), between the affine and non-affine parts, which seems wrong (in the general case it doesn't commute with either).

Probably this doesn't cause problems in practice because we never(?) set bothself.get_transform() ANDself.get_transforms(), so one of them is identity, but still, this seems wrong from a spec PoV.

From looking at the implementation of RendererBase.draw_path_collection (and associated helpers) it seems like the intent is to first transform byself.get_transforms()[0] and then byself.get_transform() (cftransform + master_transform in_iter_collection_raw_paths.

Code for reproduction

Abitartificial,but...frommatplotlibimportpyplotaspltfrommatplotlib.transformsimport*# Like an affine transform, but pretend that the affine part is actually not affine.classPseudoNonAffineTransform(Transform):is_affine=Falsedef__init__(self,mtx):self._mtx=mtxdeftransform_non_affine(self,values):returnAffine2D(self._mtx).transform_affine(values)ax=plt.figure().add_subplot()ax.grid()ax.margins(.5)pc1=ax.scatter([0,1], [0,1])pc1.set_transform(Affine2D().translate(20,20))pc2=ax.scatter([0,1], [0,1])pc2.set_transform(PseudoNonAffineTransform(Affine2D().translate(20,20).get_matrix()))plt.show()

Actual outcome

Expected outcome

Both scatter plots should be the same.

Additional information

Probably long-standing misspec?

Operating system

No response

Matplotlib Version

3.5.1+

Matplotlib Backend

agg

Python version

39

Jupyter version

no

Installation

git checkout