Which curve has smaller squared error?

31 Views Asked by At

Consider two curves $\hat{g}_1$ and $\hat{g}_2$ defined by

$$ \hat{g}_1 = \operatorname*{argmin}_g \left((y_i - g(x_i))^2 + \lambda \int [g'(x)]^2 dx \right)$$

$$ \hat{g}_2 = \operatorname*{argmin}_g \left((y_i - g(x_i))^2 + \lambda \int [g''(x)]^2 dx \right)$$

As $\lambda$ gets larger, which curve $\hat{g}_i$ has the smaller training SSE? How about for testing SSE?

(Asked this also on the stats stackexchange, but no reply.)

1

There are 1 best solutions below

0
On

It's impossible to say in general, it depends on the way the data are generated. They both describe a form of spline regression of the observations and the parameter $\lambda$ tunes the bias-variance tradeoff.