Ex. 3.13

Ex. 3.13

Derive the expression (3.62), and show that \(\hat\beta^{\text{pcr}}(p) = \hat\beta^{\text{ls}}\).

Soln. 3.13

Since \(\bb{z}_m=\bb{X}v_m\), from (3.61) in the text we have

\[\begin{eqnarray} \hat{\by}^{\text{pcr}}_{(M)} &=& \bar y\bb{1} + \bX\sum_{m=1}^M\hat\theta_mv_m\nonumber\\ &=&\begin{pmatrix} \bb{1} & \bX \end{pmatrix} \begin{pmatrix} \bar y\\ \sum_{m=1}^M\hat\theta_mv_m \end{pmatrix}.\nonumber \end{eqnarray}\]

Therefore, we can represent

\[\begin{equation} \hat\beta^{\text{pcr}}(M) = \sum_{m=1}^M\hat\theta_mv_m,\nonumber \end{equation}\]

which is (3.62) in the text.

Recall the SVD decomposition of \(\bX=\bb{U}\bb{D}\bb{V}^T\). Here \(\bb{U}\) and \(\bb{V}\) are \(N\times p\) and \(p\times p\) orthogonal matrices, and \(\bb{D}\) is a \(p\times p\) diagonal matrix. We have \(\bX\bb{V} = \bb{U}\bb{D}\) since \(\bb{V}\) is orthogonal, so that

\[\begin{equation} \bb{z}_m = \bX v_m = d_mu_m,\nonumber \end{equation}\]

where \(d_m\in\mathbb{R}\) is the \(m\)-th diagonal element in \(\bb{D}\) and \(u_m\) is \(m\)-th column in \(\bb{U}\). By definition of \(\hat\theta_m=\langle \bb{z}_m, \by\rangle/\langle \bb{z}_m, \bb{z}_m \rangle\), we have \(\hat\theta_m = \langle u_m, \by\rangle/d_m\) so that

\[\begin{equation} \hat\beta^{\text{pcr}}(p) = \sum_{m=1}^M\hat\theta_mv_m =\bb{V}\begin{pmatrix} \hat\theta_1\\\hat\theta_2\\\vdots\\\hat\theta_p \end{pmatrix} = \bb{V}\bb{D}^{-1}\bb{U}^T\by.\nonumber \end{equation}\]

On the other hand, recall \(\bX=\bb{U}\bb{D}\bb{V}^T\), by simple algebra we have

\[\begin{eqnarray} \hat\beta^{ls} &=&(\bX^T\bX)^{-1}\bX^T\by\nonumber\\ %&=&(\bb{V}\bb{D}^2\bb{V}^T)^{-1}\bb{U}\bb{D}\bb{V}^T\by\nonumber\\ &=&\bb{V}\bb{D}^{-1}\bb{U}^T\by.\nonumber \end{eqnarray}\]

Therefore we have

\[\begin{equation} \hat\beta^{\text{pcr}}(p)=\hat\beta^{ls}.\nonumber \end{equation}\]