1: \begin{proof}

2:   Here we highlight the key steps of the proof, and defer the detailed proof to

3:   Section~\ref{sec:proofs-phase-i-1}.

4: 

5:   % Note that given \pk{$W_k>\epsilon_0 e^{-k}$} and conditional on the high probability even that

6:   % $\ol W_k \le 2e^{-e^k d_0}$, we have $Nd_k^{max} = O(N \log M / M)$

7: 

8:   In Figure~\ref{fig:block-decomp-Bk}, the rows and the columns of $B_{\wh{\mc I}_k, \wh{\mc I}_k}$

9:   are sorted according to the exact marginal probabilities of the words in ascending order, with the

10:   rows and columns set to 0 by regularization shaded.

11:   %

12:   Consider the block decomposition according to the good words $\wh{\mc L}_k$ and the

13:   spillover words $\wh{\mc J}_k$.

14:   %

15:   We bound the spectral distance of the 4 blocks ($A_1, A_2, A_3, A_4$) separately. The bound for

16:   the entire matrix $\wt B_k$ is then an immediate result of triangle inequality.

17: 

18:   For block $A_1$ whose rows and columns all correspond to the ``good words'' with roughly

19:   uniform marginals, we show its concentration by applying the result in

20:   \cite{le2015concentration}.

21:   %

22:   For block $A_2$ and $A_3$, we show that after regularization the spectral norm of these two blocks

23:   are small. Intuitively, the expected row sums of block $A_2$ are bounded by $2d_k^{\max}$ and the

24:   expected column sums are bounded by $2d_k^{\max}{\ol W_k \over W_k}= O(1/N)$, as a result of the

25:   bound on $\ol W_k$ in Lemma~\ref{lem:small-spillover}. Thus the spectral norm of the block $A_2$

26:   is likely to be bounded by $O(\sqrt{d_k^{\max}/N})$. We show this rigorously with high probability

27:   arguments.

28:   %

29:   Lastly for block $A_4$, which rows and columns all correspond to the spillover words. We show that

30:   the spectral norm of this block is very small, as a result of the small spillover marginal $\ol

31:   W_k$.

32: \end{proof}

33: