= Solution
For prediction at covariate value $u$, the intercept contributes variance $\sigma^2/n$ in every case because $x$ is centered. The one-copy ridge shrinkage is $a_1=s/(s+\lambda)$, while the duplicated-design total has $a_2=2s/(2s+\lambda)$. Thus
$$
\operatorname{Bias}\widehat m_j(u)=u(a_j-1)\beta_0,
$$
and
$$
\operatorname{Var}\widehat m_1(u)=\frac{\sigma^2}{n}
+\frac{u^2\sigma^2s}{(s+\lambda)^2},
\qquad
\operatorname{Var}\widehat m_2(u)=\frac{\sigma^2}{n}
+\frac{4u^2\sigma^2s}{(2s+\lambda)^2}.
$$
Since $a_2>a_1$, duplication reduces ridge bias and increases variance.
For constrained Lasso, let $Z=z/s\sim N(\beta_0,\sigma^2/s)$ and $C_t(Z)=\max(-t,\min(Z,t))$. Both designs have the identical fitted total $C_t(Z)$, so both have
$$
\operatorname{Bias}\widehat m(u)=u\{\mathbb EC_t(Z)-\beta_0\},
\qquad
\operatorname{Var}\widehat m(u)=\frac{\sigma^2}{n}+u^2\operatorname{Var}\{C_t(Z)\}.
$$
Duplicating the predictor has no effect on Lasso predictions, despite making the coefficient vector nonunique.
Back to article page
