Ahh, could have sworn I ran that. Long day.

Should I rebase or just leave the separate commits?

Up to you.

Tagged this as proposed next release as I would like to understand the stats issue here a bit better.

Not sure your background (new to Python community), but I can probably give a quick explanation. The violinplot requires a kernel density estimation--an estimation of the probability density function of the data. The relevant function is, of course, mlab.GaussianKDE (which, btw, is a reimplementation of the fully-featured scipy.kde....).

The algorithm needs to scale its results by some factor times the determinant of the sample covariance matrix by default (calculated on line 3764 of mlab.py). Since this is done once per "violin" in the plot, and each "violin" corresponds to a 1D dataset, we're really talking about simply calculating 1/the sample variance in this case.

In the case when the user provides a set of data whose sample variance is zero---as is often the case in statistical simulations were the initial conditions are often chosen to be a point mass---the variance of the data is zero and so we get a simple divide-by-zero error. It's simply disguised as a linear algebra error since the calculation is np.linalg.det([number]).

Now, since the violin plot is designed to show the distribution of the data along the "violin"'s axis, the natural thing to do when all the data is at a point seems to be to simply represent this by a perfectly flat line, which is what my change does.

Haven't read the code yet, but the explanation makes sense to me.

On Wed, Jul 29, 2015 at 12:41 AM, Bruno Beltran [email protected]
wrote:

Not sure your background (new to Python community), but I can probably
give a quick explanation. The violinplot requires a kernel density
estimation--an estimation of the probability density function of the data.
The relevant function is, of course, mlab.GaussianKDE (which, btw, is a
reimplementation of the fully-featured scipy.kde....).

The algorithm needs to scale its results by some factor times the
determinant of the sample covariance matrix by default (calculated on line
3764 of mlab.py). Since this is done once per "violin" in the plot, and
each "violin" corresponds to a 1D dataset, we're really talking about
simply calculating 1/the sample variance in this case.

In the case when the user provides a set of data whose sample variance is
zero---as is often the case in statistical simulations were the initial
conditions are often chosen to be a point mass---the variance of the data
is zero and so we get a simple divide-by-zero error. It's simply disguised
as a linear algebra error since the calculation is np.linalg.det([number]).

Now, since the violin plot is designed to show the distribution of the
data along the "violin"'s axis, the natural thing to do when all the data
is at a point seems to be to simply represent this by a perfectly flat
line, which is what my change does.

—
Reply to this email directly or view it on GitHub
#4816 (comment)
.

Any reason to not merge this?