= Solution
The <kernel ridge regression> estimator minimizes
$$
\sum_{i=1}^n\{y_i-f(x_i)\}^2+\lambda\lVert f\rVert_{\mathcal H}^2.
$$
By the <representer theorem>, its fitted-value vector is
$$
\widehat f=K(K+\lambda I)^{-1}y.
$$
Writing $f$ for the true value vector, the variance contribution to $\mathbb E\lVert f-\widehat f\rVert_2^2$ is
$$
\sigma^2\operatorname{tr}\{K^2(K+\lambda I)^{-2}\}
=\sigma^2\sum_{i=1}^n\frac{d_i^2}{(d_i+\lambda)^2}.
$$
The squared bias is $\lVert\lambda(K+\lambda I)^{-1}f\rVert_2^2$. In an orthonormal eigenbasis of $K$, the scalar inequality
$$
\frac{\lambda^2}{(d+\lambda)^2}\leq\frac{\lambda}{4d},
$$
which is equivalent to $4d\lambda\leq(d+\lambda)^2$, yields
$$
\boxed{
\mathbb E\sum_{i=1}^n\{f^0(x_i)-\widehat f_\lambda(x_i)\}^2
\leq\sigma^2\sum_{i=1}^n\frac{d_i^2}{(d_i+\lambda)^2}
+\frac\lambda4f^TK^{-1}f}.
$$
For $\lVert f\rVert_2=1$, only the last term varies. The <Rayleigh quotient> of $K^{-1}$ is maximized by a unit eigenvector associated with its largest eigenvalue $1/d_n$, equivalently an eigenvector of $K$ associated with $d_n$.
Back to article page
