v0.1.0

Refactoring

ConceptEraser has been split into two separate classes, LeaceFitter and LeaceEraser. This makes it easy to save the fitted erasure function by itself in a compact format, without also saving the covariance and cross-covariance statistics used to create it.

Algorithmic changes

We now use the asymptotically optimal shrinkage formula from this paper to shrink the covariance matrix of X toward a multiple of the identity matrix. Under weak assumptions, this provably speeds up the convergence of the covariance matrix estimate toward the population covariance matrix. Prior versions had used the raw sample covariance matrix with no shrinkage, which can cause numerical instability and very suboptimal edits when the sample size is low.