Ex. 5.14

Ex. 5.14

Derive the constraints on the \(\alpha_j\) in the thin-plate spline expansion (5.39) to guarantee that the penalty \(J(f)\) is finite. How else could one ensure that the penalty was finite?

Soln. 5.14

Consider

\[\begin{equation} f(x,y) = \beta_0 + \beta^T(x,y) + \sum_{j=1}^N\alpha_jh_j(x,y)\non \end{equation}\]

where

\[\begin{eqnarray} h_j(x,y) &=& [(x-x_j)^2 + (y-y_j)^2]\log\left(\sqrt{(x-x_j)^2 + (y-y_j)^2}\right)\non\\ &=&\frac{1}{2}[(x-x_j)^2 + (y-y_j)^2]\log((x-x_j)^2 + (y-y_j)^2).\non \end{eqnarray}\]

Without loss of generality, we will drop the constant and linear term \(\beta_0 + \beta^T(x,y)\) in \(f\) since they vanish after taking second derivatives below.

The penalty function \(J\) is

\[\begin{equation} J(f) = \int\int_{\mathbb{R}^2}\left[\left(\frac{\partial^2f(x,y)}{\partial x^2}\right)^2+2\left(\frac{\partial^2f(x,y)}{\partial x\partial y}\right)^2 + \left(\frac{\partial^2f(x,y)}{\partial y^2}\right)^2\right]dxdy.\non \end{equation}\]

Next we compute each integrand above. First, denote

\[\begin{eqnarray} r_{jx} &=& x-x_j\non\\ r_{jy} &=& y-y_j\non\\ r^2_j &=& r^2_{jx} + r^2_{jy}=(x-x_j)^2 + (y-y_j)^2.\non \end{eqnarray}\]

Then we have

\[\begin{eqnarray} \frac{\partial f(x,y)}{\partial x} &=& \sum_{j=1}^N\alpha_j[r_{jx}(\log(r_j^2) + 1)]\non\\ \frac{\partial^2 f(x,y)}{\partial^2 x} &=& \sum_{j=1}^N\alpha_j\left[\log(r_j^2) + 2\frac{r_{jx}^2}{r_j^2} + 1\right]\non\\ \frac{\partial^2 f(x,y)}{\partial x\partial y} &=& \sum_{j=1}^N\alpha_j\frac{2r_{jx}r_{jy}}{r_j^2}\non\\ \frac{\partial^2 f(x,y)}{\partial^2 y} &=& \sum_{j=1}^N\alpha_j\left[\log(r_j^2) + 2\frac{r_{jy}^2}{r_j^2} + 1\right].\non \end{eqnarray}\]

To get penalty \(J[f]\), we calculate the first integrand as

\[\begin{eqnarray} &&\left(\frac{\partial^2 f(x,y)}{\partial^2 x}\right)^2\non\\ &=&\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)^2 + \left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)^2 + \left(\sum_{j=1}^N\alpha_j\right)^2\non\\ &&+2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)\left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)\non\\ &&+2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)\left(\sum_{j=1}^N\alpha_j\right)\non\\ &&+2\left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)\left(\sum_{j=1}^N\alpha_j\right).\non \end{eqnarray}\]

At this point we see that \(\sum_{j=1}^N\alpha_j = 0\), otherwise the integral would be infinite. So that above integrand is simplified to

\[\begin{eqnarray} \label{eq:5-14a} &&\left(\frac{\partial^2 f(x,y)}{\partial^2 x}\right)^2\non\\ &=&\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)^2 + \left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)^2+2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)\left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right).\non \end{eqnarray}\]

Similarly we have

\[\begin{eqnarray} \label{eq:5-14b} \left(\frac{\partial^2 f(x,y)}{\partial x\partial y}\right)^2=\left(\sum_{j=1}^N\alpha_j\frac{2r_{jx}r_{jy}}{r_j^2}\right)^2.\non \end{eqnarray}\]

and

\[\begin{eqnarray} \label{eq:5-14c} &&\left(\frac{\partial^2 f(x,y)}{\partial^2 y}\right)^2\non\\ &=&\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)^2 + \left(\sum_{j=1}^N2\alpha_j\frac{r_{jy}^2}{r_j^2}\right)^2+2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)\left(\sum_{j=1}^N2\alpha_j\frac{r_{jy}^2}{r_j^2}\right).\non \end{eqnarray}\]

Sum them up we get

\[\begin{eqnarray} &&2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)^2 + \left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)^2+\left(\sum_{j=1}^N2\alpha_j\frac{r_{jy}^2}{r_j^2}\right)^2\non\\ && + \left(\sum_{j=1}^N\alpha_j\frac{2r_{jx}r_{jy}}{r_j^2}\right)^2\non\\ && + 2\left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)\left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2+r_{jy}^2}{r_j^2}\right).\non \end{eqnarray}\]

Note that \(\frac{r_{jx}^2+r_{jy}^2}{r_j^2}=1\) and \(\sum_{j=1}^N\alpha_j=0\), the last summand above vanishes and we are left with

\[\begin{equation} \left(\sum_{j=1}^N\alpha_j\log(r_j^2)\right)^2 + \left(\sum_{j=1}^N2\alpha_j\frac{r_{jx}^2}{r_j^2}\right)^2+\left(\sum_{j=1}^N2\alpha_j\frac{r_{jy}^2}{r_j^2}\right)^2+ \left(\sum_{j=1}^N\alpha_j\frac{2r_{jx}r_{jy}}{r_j^2}\right)^2.\non \end{equation}\]

Remark

It has been shown that \(\sum_{j=1}^N\alpha_jx_j=0\) (see, e.g., Thin-Plate Splines) , however I don't see how to arrive that from here.