Density-based jitter

[teunbrand]

It would be nice if we could jitter the points, where the jitter amount is proportional to the density. This would allow jittered points to better match violins and show distributions better. The proposal is not to go all-out beeswarm but do better than the current uniform jitter.

[ashishtiwari03]

This is a really nice idea the mismatch between uniform jitter and violin width has always bugged me.

A couple of things I was wondering about:

1. Would this live as a new position_density_jitter() or as a parameter on the existing position_jitter()? I could see arguments either way, but a separate function might be cleaner since it introduces a density estimation step that regular jitter doesn't need.

2. For the density estimate would it make sense to expose the bw (bandwidth) parameter so users can match it to a geom_violin() in the same plot? Otherwise you could end up with the jitter shape not quite lining up with the violin even though they're showing the same data.

3. ggforce::position_auto() already does something sina-plot-like when one axis is discrete. Is the idea here similar in spirit, just brought into ggplot2 core?

4. One edge case I was thinking about with very small groups or multimodal distributions, the density estimate can get a bit wobbly. Any thoughts on how to handle that gracefully?

Happy to help prototype if this moves forward!