MSE Comparison

Below are results obtained with a variety of data matrices of dimensions $n\times p$. For each pair of dimension, 50 covariance matrices are generated with associated sample data matrices. The covariance obtained with the different estimators are then compared to the ground-truth and the MSE is reported.

AbbreviationMethod
anshrinkanalytical nonlinear shrinkage
ccorLSE with constant correlation target
ccovLSE with constant covariance target
d1vLSE with identity target
dcvLSE with diagonal common variance target
duvLSE with diagonal unequal variance target
ppcLSE with perfect positive correlation target
sSimple estimator (baseline)
_lwuses ledoit-wolf shrinkage
_ssuses schaffer-strimmer shrinkage
_oasuses oracle approximating shrinkage
_rblwuses rao-blackwellised ledoit-wolf shrinkage

Fat matrices

Tall matrices