Initialization of psis and lambdas

[rossfadely]

• Initializing psis to 0.5 * datavar seems to be ok (see other issue on logdet going negative)
• Initializing lambas to something close to the datavar is bad, the covariance initialization in this case is HUGE and the final covariance doesnt become sensible

[rossfadely]

• Final lambdas have values order ~ +/- unity, so using randn to initialize isnt crazy
• PCA is probably the most sensible way to go

[davidwhogg]

holy crap that sucks. Let's make sure we either:

1. figure out what initializations are legit, or
2. figure out why the code is broken and fails on some initializations

And either way, let's document it in the code and document.

[rossfadely]

Hogg and Fadely concluded:

• That initializing lambdas to zero may be sensible
• If we dont, there seem to be cases where psis get negative elements early during EM
• There may be some bug
• If there isn't this is important to document and get the word out

[davidwhogg]

we should initialize with a probabilsitic pca on each of the k-means. Will issue this.