1: \begin{proof}[Proof overview]

2: See appendix Section~\ref{app:beatrep} for a complete and rigorous proof. \textbf{Here we only provide the main intuition behind the proof.} From Theorem~\ref{thm:beat_uncoded}, we have

3: \begin{equation}\label{eqn:trace_der0}

4: \Ep_f \left[\twonorm{\Ebf_{\text{uncoded}} }^2\right]-\Ep_f \left[\twonorm{\Ebf_{\text{coded}} }^2\right]\ge\Ep_f[\trace(\J_2\J_4^{-1}\J_2^\top)].

5: \end{equation}

6: To prove \eqref{eqn:coded_and_replication}, the main technical difficulty is to simplify the term $\trace(\J_2\J_4^{-1}\J_2^\top)$. For a Fourier matrix $\Fbf$, we are able to show that the matrix $\Fbf \boldsymbol{\Lambda}\Fbf^\top=   \left[

7: \begin{matrix}

8:    {\J_1} & {\J_2} \\

9:    {\J_2^\top} & {\J_4} \\

10: \end{matrix}  \\

11: \right]$ (see Corollary~\ref{cor:upper_bound}) is a Toeplitz matrix, which provides a good structure for us to study its behavior. Then, we use the Gershgorin circle theorem \cite{golub2012matrix} (with some algebraic derivation) to show that the maximum eigenvalue of $\J_4$ satisfies $\sigma_\text{max}(\J_4)\approx \Ep_f[\trace(\C(l_1))]$, and separately using some algebraic manipulations, we show

12: \begin{equation}

13: \trace(\J_2\J_2^\top)\approx (n-k) \text{var}_f[\trace(\C(l_1))],

14: \end{equation}

15: for large matrix size $n$. Since

16: \begin{equation}

17: \begin{split}

18: \trace(\J_2\J_4^{-1}\J_2^\top)\ge \trace(\J_2(\sigma_\text{max}(\J_4))^{-1}\J_2^\top)\\

19: =\frac{1}{\sigma_\text{max}(\J_4)} \trace(\J_2\J_2^\top),

20: \end{split}

21: \end{equation}

22: we obtain

23: \begin{equation}\label{eqn:trace_der1}

24: \trace(\J_2\J_4^{-1}\J_2^\top)\ge \frac{(n-k)\text{var}_f[\trace(\C(l_1))]}{\Ep_f[\trace(\C(l_1))]},

25: \end{equation}

26: for large $n$. Then, \eqref{eqn:coded_and_replication} can be proved by plugging \eqref{eqn:trace_der1} into \eqref{eqn:trace_der0}. After that, we can combine \eqref{eqn:coded_and_replication}, \eqref{eqn:error_replication} and the variance heavy-tail property to prove \eqref{eqn:coded_beats_replication}.

27: \end{proof}

28: