Condition Numbers and Minimax Ridge Regression Estimators
研究了极小化极大岭回归估计量在改善均方误差的同时，能否提升数值稳定性（用条件数衡量），并讨论了为极小化极大性牺牲数值稳定性的后果。
Abstract Ridge regression was originally formulated with two goals in mind: improvement in mean squared error and numerical stability of the coefficient estimates. Conditions are given under which a minimax ridge regression estimator can also improve numerical stability, a quantity that can be measured with the condition number of the matrix to be inverted. The consequences of trading numerical stability for minimaxity are also discussed.