arXiv++ Combinatorics

Let $P$ be a fixed finite poset whose maximal chains all have $h$ elements, and let $\mathcal W_{n,h}$ be the union of $h$ consecutive middle levels of $2^{[n]}$. We prove that the maximum size of a $P$-free subfamily of $\mathcal W_{n,h}$ is $(π(P)+o(1))M_n$ for a constant $π(P)$, where $M_n=\binom n{\lfloor n/2\rfloor}$, and that the number of $P$-free subfamilies is $2^{(π(P)+o(1))M_n}$. For connected $P$ we prove supersaturation above $(π(P)+ε)M_n$ and determine, in terms of the coordinate patterns of the copies of $P$ in the window, a range of $p$ in which the largest $P$-free subfamily of the random family $\mathcal W_{n,h}(p)$ has size $(π(P)+o(1))pM_n$ and a range in which it has size $(h+o(1))pM_n$; the two ranges meet whenever all copies have the same pattern. For the Boolean poset $Q_d$, which has height $d+1$, the threshold order in the corresponding window $\mathcal W_{n,d+1}$ is $n^{-d/(2^d-1)}$, and $π_d=π(Q_d)$ satisfies $(d+1)λ_d\leπ_d\le d+1-\binom d{\lfloor d/2\rfloor}^{-1}$, where $λ_d$ is the vertex Turán density of $d$-dimensional subcubes in the hypercube. Together with a construction of Ellis, Ivan and Leader this gives $d+1-π_d=\exp(-Θ(d))$. For the diamond, known bounds give $2.1479\leπ_2\le2.15121$.

We prove a multilevel non-shadow refinement of the Alon--Babai--Suzuki (ABS) nonuniform restricted-intersection theorem. Let $K=\{k_1,\dots,k_r\}$ and let $L$ be a set with $|L|=s$. If $\mathcal{F}\subseteq \bigcup_{k\in K}\binom{[n]}{k}$ is $L$-intersecting and $k_i>s-r$ for every $i$, then $|\mathcal{F}| + \sum_{j=s-r+1}^{s} |\mathcal{N}_j(\mathcal{F})| \le N(n,s,r),$ equivalently $|\mathcal{F}| \le \sum_{j=s-r+1}^{s} |\partial_j\mathcal{F}|.$ Thus the ABS bound is sharpened by the total non-shadow deficit on the top $r$ levels. In the modular setting, we take a coefficient-sensitive viewpoint: the polynomial method depends not just on the degree of the annihilator polynomial $P_L(t)=\prod_{\ell\in L}(t-\ell)\in\mathbb{F}_p[t]$, but on which binomial terms actually appear in it. This yields a gap-free modular bound depending only on the active support levels of $P_L$. For almost-initial residue patterns $L=\{0,1,\dots,s-m-1\}\cup R \pmod p$ we obtain the collapse $|\mathcal{F}|\le \sum_{i=0}^{m}\binom{n}{s-i}.$ In particular, for consecutive residues $L=\{0,1,\dots,s-1\}\pmod p$ we get the sharp bound $|\mathcal{F}|\le \binom{n}{s}$, giving a partial negative answer to a question of Alon--Babai--Suzuki: the modular ABS bound $N(n,s,r)$ is not attainable in the consecutive-residue regime whenever $r\ge 2$.