1: \begin{proof} We give the detailed argument for $s=1$.
2: Let $p_\ell(x)=\Pr(\hat A_{\ell,i}=x)$.
3: This result will follow from monotonicity of the distribution of $\hat A_{\ell,i}$, i.e.,
4: $p_\ell(x)>p_\ell(x+1)$ for $x\ge 0$.
5: We proceed by induction on $\ell$.
6: For $\ell=1$, we have $p_1(0)=\frac{2m-1}m$ and $p_1(1)=\frac{1}{4m}$, which establishes the base
7: case. For $\ell>1$, we have
8: $$ p_\ell(x)=\frac{1}{2m} p_{\ell-1}(x-1)+\frac{m-1}{m} p_{\ell-1}(x) + \frac{1}{2m}p_{\ell-1}(x+1);$$
9: $$ p_\ell(x+1)=\frac{1}{2m} p_{\ell-1}(x)+\frac{m-1}{m} p_{\ell-1}(x+1) + \frac{1}{2m}p_{\ell-1}(x+2).$$
10: Now we know that $\frac{m-1}m>\frac{1}{2m}$ (since $m\ge 2$), and this means
11: $$\frac{m-1}m p_{\ell-1}(x) + \frac{1}{2m} p_{\ell-1}(x+1) >
12: \frac{1}{2m} p_{\ell-1}(x) + \frac{m-1}{m} p_{\ell-1}(x+1),$$
13: since the LHS has a larger coefficient on the larger term. This compares two of the terms of
14: $p_\ell(x)$ with two of the terms of $p_\ell(x+1)$, so it only remains to compare the remaining
15: terms.
16: Since $x\ge 0$, we have $|x-1|\le x+1$. Thus, by repeatedly applying the inductive hypothesis,
17: we have $p_{\ell-1}(x-1)>p_{\ell-1}(x+2)$, which completes the proof for all $\ell$.
18: In particular, we have shown: if the positive integers $\Z_{>0}$ are partitioned into intervals
19: $[kn+1,kn+n]$, then the farthest point in each interval from 0 (the value divisible by $n$) has the lowest
20: probability.
21:
22:
23: For NBSRW, we would need to inspect the LCLT bounds to establish monotonicity rigorously, though it is intuitively clear for $\ell\gg 1$.
24:
25: The argument for general $s$ runs along exactly the same lines: $\Z_{>0}^m$ is cut up into
26: $n\times \dots\times n$ boxes which are obtained as products of the intervals described above,
27: then in each box, the
28: point farthest from the origin (which satisfies the congruence condition in the statement of the lemma) has the lowest probability in the random walk.
29: \end{proof}
30: