1: \begin{proof}[Proof sketch]

2: First, consider $\Er_\text{e1}$. As in the proof of Marton's inner bound for the broadcast channel, the mutual covering lemma~\cite{ElGamalKim} implies $\P(\Er_\text{e1}) \to 0$ as $n\to \infty$ if~\eqref{eq:atLeastOne} holds.

3: 

4: Now consider $\Er_\text{e2}$, for which we need to control the number of typical pairs that can occur in the same ``row'' or ``column'' of the product set $\Lc_1(m_1) \times \Lc_2(m_2)$, i.e., for the same $l_{1}$ or $l_{2}$ coordinate.  The probability $\P(\Er_\text{e2})$ tends to zero provided that~\eqref{eq:noDoubles1} and~\eqref{eq:noDoubles2} hold. 

5: 

6: This is akin to the birthday problem~\cite{Mises1939}, where $k$ samples are drawn uniformly and independently from $\natSet N$, and the interest is in samples that have the same value (collisions). It is well-known that for the probability of collision to be $p_c$, the number of samples required is roughly $k \approx \sqrt{-2 N \ln(1-p_c)  }$, which scales with $\sqrt{N}$. %The classical example are birthdays ($N=365$): In a random group of $k\geq 23$ people, with probability $p_c\geq 1/2$, at least two of them were born on the same day of the year. 

7: %

8: In our case, the number of samples is the cardinality of the set $\Sc(m_0,m_1,m_2)$, which is roughly $k = 2^{n(r_{1} + r_{2}-I(Z_1; Z_2\cond U))}$. The samples are categorized into $N_1 = 2^{nr_{1}}$ and $N_2 = 2^{nr_{2}}$ classes along rows and columns, respectively. To achieve a probability of collision $p_c \to 0$ along both dimensions, we need $k \ll \min\{\sqrt{N_1},\sqrt{N_2}\}$, which yields exactly the conditions~\eqref{eq:noDoubles1} and~\eqref{eq:noDoubles2}. 

9: 

10: A rigorous proof is given in Appendix~\ref{sec:encodingErrorProof}.

11: \end{proof}

12: