Meta comments:

Is scipy really our only way to get this functionality?

I do wonder, in 2025, about our reluctance to make scipy a dependency. It would make most of mlab obsolete and has various standardized stats routines that we could use without duplicated effort. I think the view in the past was that scipy was a heavy dependency, and when folks had to compile locally was a pretty slow and tool intensive task. Now that most folks are covered by wheels, is this really the case?

I appreciate adding more dependencies is a slippery slope, but I bet some large fraction of folks using Matplotlib are installing it beside scipy in their environments.

I would be reluctant to add scipy it as a hard dependency. An optional dependency may be ok, in which case violinplot() would error out if scipy is not there. For the conda-world we would also have the option to do it analogous to pyqt: There is matplotlib-base which only has the minimal dependencies, and is typically good for CI and other packages to depend on matplotlib. And there is matplotlib, which includes the "heavy" dependencies on pyqt and tornado to give better out-of-the box experience.