Ridge Regression

What is ridge regression?

Answer

Ridge regression is a form of penalised (linear) regression in which the estimation of the coefficients is done by minimizing the sum of squared residuals (OLS) plus an \(\ell_2\) penalty on the coefficients:

\[ L_{\rm ridge}(\beta) = \sum_{i=1}^n (Y_i - x_i^\top\beta)^2 + \lambda \sum_{j=1}^p \beta_j^2. \]

Here \(\lambda\) is called the regularization parameter (also known as the shrinkage parameter), and it is a hyperparameter of this regression model.

The solution has the closed‐form

\[ \hat\beta_{\rm ridge} = \bigl(X^\top X + \lambda I\bigr)^{-1} X^\top Y, \]

The estimated slope coefficients from ridge regression are smaller in an \(\ell_2\) sense than those from OLS, i.e.

\[ \sqrt{\sum_{j=1}^p \hat\beta_{\mathrm{ridge},j}^2} \;\le\; \sqrt{\sum_{j=1}^p \hat\beta_{\mathrm{OLS},j}^2}. \]
Notes and comments

Note: The \(\ell_2\) norm \(\|\beta\|_2 = \sqrt{\sum_{j=1}^p \beta_j^2}\) is the Euclidean distance of the vector \(\beta\) from the origin (this comes from the Pythagorean Theorem).

Back to collection