@sp_monte_carlo Well for linear ridge regression it's pretty natural, going back to the original motivation (equivalent to (X^TX + λI)^{-1}X^TY, to stabilize OLS) and/or OLS augmenting design matrix with \sqrt{λ}I. Regularizing using ||•||_p^p also pretty…
@jon_mellon You could consider the bridge estimator, which generalizes the lasso and ridge (those are special cases of the bridge estimator with γ = 1 or 2, respectively). For γ in (0, 1), you get feature selection like the lasso, but also asymptotic norm…
@jwhandley17 @ColinJMcAuliffe (3) (whoops, sorry, there's going to be 4 total.) nonconvex sparse estimators. for example, the bridge estimator (penalizing ||β||_q^q) converges faster than lasso and is asymptotically normal for 0 < q < 1 under pretty…