Wiki
Wiki

Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.

Updated


Source. Published p. 262, Theorem 1.7, and pp. 272–274, Section 4 (PDF). This is the Section 4 proof, separate from the later general counting theorem.

Statement. Given γ>0\gamma>0, there are ϵ,σ>0\epsilon,\sigma>0 and n0n_0 such that, for n≥n0n\ge n_0, a family F⊆2[n]\mathcal F\subseteq2^{[n]} with ∣F∣≥2ne−ϵn|\mathcal F|\ge2^ne^{-\epsilon n} and an integer ∣l−n/4∣≤σn|l-n/4|\le\sigma n satisfy

il(F,F)≥∣F∣2e−γn.(1)i_l(\mathcal F,\mathcal F)\ge|\mathcal F|^2e^{-\gamma n}. \tag{1}

Proof. First prove the following middle-layer form: for any δ>0\delta>0 there are ϵ0,σ0>0\epsilon_0,\sigma_0>0 such that, for large mm, H⊆Ω([4m];2m)\mathcal H\subseteq\Omega([4m];2m) of density f≥e−ϵ0mf\ge e^{-\epsilon_0m} and ∣l−m∣≤σ0m|l-m|\le\sigma_0m satisfy

il(H,H)≥(4m2m) ⁣2e−δm.(2)i_l(\mathcal H,\mathcal H)\ge\binom{4m}{2m}^{\!2}e^{-\delta m}. \tag{2}

Write B=(4m2m)B=\binom{4m}{2m} and L=(2mm)L=\binom{2m}m. For a 2m2m-set AA, let xAx_A count members F∈HF\in\mathcal H with ∣F∩A∣=m|F\cap A|=m. The incidence graph is regular of degree L2L^2 on both sides. Lemma 4.1 gives at least fB/2fB/2 choices of AA with xA≥fL2/2x_A\ge fL^2/2.

Choose a small α>0\alpha>0, set a=⌊αm⌋a=\lfloor\alpha m\rfloor, and let yAy_A count pairs (F,F′)(F,F') with

∣F∩A∣=∣F′∩A∣=m,∣F∩F′∣=l,∣F∩F′∩A∣=a.|F\cap A|=|F'\cap A|=m,\quad |F\cap F'|=l, \quad |F\cap F'\cap A|=a.

For a fixed pair with intersection ll, its four atoms have sizes l,2m−l,2m−l,ll,2m-l,2m-l,l. There are exactly

T=(la)2(2m−lm−a) ⁣2(3)T=\binom la^2\binom{2m-l}{m-a}^{\!2} \tag{3}

choices of AA; infeasible coefficients mean zero. As α,σ0→0\alpha,\sigma_0\to0, the entropy estimate gives log⁡T=oα(m)+oσ0(m)+O(log⁡(m+1))\log T=o_\alpha(m)+o_{\sigma_0}(m)+O(\log(m+1)). If (2) failed, ∑AyA<B2e−δmT\sum_Ay_A<B^2e^{-\delta m}T. Averaging over the at least fB/2fB/2 popular choices of AA gives one A0A_0 with

yA0xA0≤4Bf2L2e−δmT<14.(4)\frac{y_{A_0}}{x_{A_0}} \le\frac{4B}{f^2L^2}e^{-\delta m}T<\frac14. \tag{4}

Here log⁡(B/L2)=O(log⁡(m+1))\log(B/L^2)=O(\log(m+1)); choose α\alpha first so its loss is small compared with δ\delta, then σ0\sigma_0 smaller, then ϵ0\epsilon_0 smaller, and finally mm large.

Delete all endpoints of these yA0y_{A_0} ordered pairs from the local family. At most 2yA02y_{A_0} members are lost, so at least fL2/4fL^2/4 remain. Each remaining member has an mm-set as its intersection with A0A_0, and an mm-set in the complement. Let B\mathcal B consist of inner mm-sets with residual fiber of size at least fL/8fL/8. Nonpopular fibers contribute at most fL2/8fL^2/8 members; each popular fiber has at most LL members. Thus ∣B∣≥fL/8|\mathcal B|\ge fL/8.

Choose ϵ0\epsilon_0 sufficiently small relative to α\alpha. Theorem 1.1 on the 2m2m-set A0A_0 gives distinct B1,B2∈BB_1,B_2\in\mathcal B with ∣B1∩B2∣=a|B_1\cap B_2|=a. Their residual fibers each have size at least fL/8fL/8, so Theorem 1.4 on the complementary 2m2m-set gives residual intersection l−al-a. The required buffers are positive when σ0<α/2\sigma_0<\alpha/2 and α\alpha is small: l−al-a is bounded away from both zero and mm. The reconstructed pair satisfies every condition defining yA0y_{A_0}, contradicting its deletion. This proves (2).

Now start with the family in (1). Choose a size level kk containing at least ∣F∣/(n+1)|\mathcal F|/(n+1) members. For any fixed τ>0\tau>0, the entropy estimate forces ∣k−n/2∣≤τn|k-n/2|\le\tau n once ϵ\epsilon is small enough and nn is large. If 2k≤n2k\le n, average containment over all 2k2k-subsets to find a containing set on which the kk-set family has at least its original relative density. If 2k>n2k>n, enlarge the ground set to size 2k2k by unused points. In either case the new ambient size N=2kN=2k is within 2τn2\tau n of nn, all members have size N/2N/2, and their relative density in that layer is at least

exp⁡(−ϵn−2τnlog⁡2−O(log⁡(n+1))).(5)\exp(-\epsilon n-2\tau n\log2-O(\log(n+1))). \tag{5}

If N≡2(mod4)N\equiv2\pmod4, add two new points x,yx,y and adjoin xx to every member. The new ambient size N′=N+2=4mN'=N+2=4m and each member's size is 2m2m; the target intersection becomes l′=l+1l'=l+1. Otherwise set N′=N=4mN'=N=4m and l′=ll'=l. This operation is injective and preserves target pairs under the stated shift. It changes (5) only by an absolute factor.

Apply (2) with a sufficiently small output δ\delta relative to γ\gamma. After its constants ϵ0,σ0\epsilon_0,\sigma_0 are fixed, choose τ\tau small compared with ϵ0,σ0,γ\epsilon_0,\sigma_0,\gamma, then σ\sigma and ϵ\epsilon smaller still. Equations (5) and ∣l′−m∣≤∣l−n/4∣+∣N′−n∣/4+1|l'-m|\le|l-n/4|+|N'-n|/4+1 verify its hypotheses for large nn. The resulting pair count is at least

(N′N′/2) ⁣2e−δm≥4nexp⁡(−4τnlog⁡2−δm−O(log⁡(n+1)))≥∣F∣2e−γn.\binom{N'}{N'/2}^{\!2}e^{-\delta m} \ge4^n\exp(-4\tau n\log2-\delta m-O(\log(n+1))) \ge |\mathcal F|^2e^{-\gamma n}.

All these pairs were already in the original family, proving (1). □\square

Source precision. A reference to “Theorem 1.2” in the inner-family step on p. 274 is to Theorem 1.1; there is no such numbered theorem. The fiber threshold is taken consistently as fL/8fL/8 after deletion. The large-nn qualification used in the source proof is necessary in the statement: for fixed small nn and the full cube, the fraction of pairs with one prescribed intersection is less than one, whereas (1−δ)n→1(1-\delta)^n\to1 as δ→0\delta\to0. Decreasing the input tolerance cannot remove that obstruction. For example, at n=4,l=1n=4,l=1 the full cube has target-pair proportion 27/6427/64, less than (1−δ)4(1-\delta)^4 for sufficiently small δ>0\delta>0. The proof above includes its threshold.

Dependencies. lemma_4_1, theorem_1_1, theorem_1_4, entropy_estimates.