1: \begin{proof}

2: Assume that event $\m E:=\l\{\|\hat\mu-\mu\|>C_\alpha\eps\r\}$ occurs. 

3: Lemma \ref{lem:median} implies that there exists a subset $J\subseteq\{1,\ldots,k\}$ of cardinality $|J|\geq\alpha k$ such that 

4: $\|\mu_j-\mu\|>\eps$ for all $j\in J$, hence

5: \[

6: \Pr(\m E)\leq \Pr\l(\sum_{j=1}^k I\l\{\|\hat\mu_j-\mu\|>\eps\r\}>\alpha k \r).

7: \]

8: If $W$ has Binomial distribution $W\sim B(k,p)$, then 

9: \begin{align*}

10: &

11: \Pr\l(\sum_{j=1}^k I\l\{\|\hat\mu_j-\mu\|>\eps\r\}>\alpha k \r)\leq \Pr\l(W>\alpha k\r)

12: \end{align*}

13: (see Lemma 23 in \cite{lerasle2011robust} for a rigorous proof of this fact). 

14: Chernoff bound (e.g., Proposition A.6.1 in \cite{Vaart1996Weak-convergenc00}) implies that

15: \[

16: \Pr\l(W>\alpha k\r)\leq \exp\big(-k\psi\l(\alpha;p\r)\big).

17: \] 

18: \end{proof}

19: