Dtrace Estimators

[mnarayan]

This is uses the D-trace Loss (which is akin to a least squares type loss) instead of the log-likelihood loss to estimate sparse inverse covariance.

It most closely resembles the graphical lasso (produces a positive definite, symmetric, and sparse inverse covariance estimate) but has good & different incoherence/irrepresentability conditions for exact recovery (possibly but not yet provably weaker).

Probably not a priority for the next scikitquic milestone, but personally very useful/critical to me, as I need a variant of this for work. Also no good implementations exist yet in python for this.

[jasonlaska]

See http://arxiv.org/pdf/1511.09188v1.pdf