Problem

• Violin plots for log scaled data currently require transforming the data by log-scaling and then modifying the ticks inversely to match. This becomes problematic if you are plotting multiple data forms, etcetera.
• Seaborn implemented alog_scale capability in version 0.13, and I think matplotlib could do the same

Proposed solution

I believe all that is needed is to allow for a log version of the KDE.

In log scale, that would look like:

def _kde_method(X, coords):        # Unpack in case of e.g. Pandas or xarray object        X = matplotlib.cbook._unpack_to_numpy(X)        # fallback gracefully if the vector contains only one value        if np.all(X[0] == X):            return (X[0] == coords).astype(float)        X_log = np.log(X)  # Transform to log space        kde = matplotlib.mlab.GaussianKDE(X_log, bw_method)        # Return KDE evaluated at log-transformed coordinates        coords_log = np.log(coords)  # Transform to log space        return kde.evaluate(coords_log)

so inmatplotlib/lib/matplotlib/axes/_axes.py:violinplot() , thescale could just be passed as an option to_kde_method to allow switching between the current and this form.

You would probably also want, inmatplotlib/lib/matplotlib/cbook.py:violin_stats() to also pass in thescale so that the coords can, rather than beingcoords = np.linspace(min_val, max_val, points), instead becoords = np.geomspace(min_val, max_val, points).

I've checked and this does create reasonable plots, and I think this is all that would need to be done.

All other stats like quantiles remain unchanged, though perhaps one could debate whethermean should be calculated in log scale or not.