I need to give a counter-example against the following theorem:
Suppose $H \subset \operatorname{Gal}(\mathbb{Q}(\zeta_n)/\mathbb{Q})$ is a subgroup. Then we have $\mathbb{Q}(\zeta_n)^H = \mathbb{Q}(\eta_H)$, with $\eta_H = \sum_{\sigma \in H} \sigma(\zeta_n)$, the Gaussian-period.
This theorem is true for $n = p$ prime, but not for general $n$. So far, my attempt is the following. We take $n = 8$. Then $\operatorname{Gal}(\mathbb{Q}(\zeta_8)/\mathbb{Q}) \cong (\mathbb{Z}/8\mathbb{Z})^\ast = \{1,3,5,7\} \cong \mathbb{Z}/4\mathbb{Z}$.
Now we have the subfield $\mathbb{Q}(\zeta_4) = \mathbb{Q}(i)$, since $4 \mid 8$. We also have the subfield $\mathbb{Q}(\sqrt{2})$ since $\zeta_8 + \zeta_8^{-1} = 2 \cos(2\pi/8) = \sqrt{2}$. Hence, we have also the quadratic subfield $\mathbb{Q}(\sqrt{-2})$. Notice that $1,5$ keep $i$ fixed, so according to the above theorem, we have $$ \mathbb{Q}(i) = \mathbb{Q}(\zeta_8)^{\{1,5\}} = \mathbb{Q}(\zeta_8 + \zeta_8^5). $$ I am stuck at this point, how to derive a contradiction from this?