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. 265, Theorem 1.15, and pp. 279–281 (PDF).

Statement. Given η,γ>0\eta,\gamma>0, there is ϵ>0\epsilon>0 such that for a compatible s×ts\times t matrix M=(mij)M=(m_{ij}) with integer entries mij≥ηnm_{ij}\ge\eta n, and partition families A⊆Ω([n];l)\mathcal A\subseteq\Omega([n];\mathbf l), B⊆Ω([n];k)\mathcal B\subseteq\Omega([n];\mathbf k), density product at least e−ϵne^{-\epsilon n} implies

iM(A,B)≥e−γnN(M),N(M)=n!∏i,jmij!.(1)i_M(\mathcal A,\mathcal B)\ge e^{-\gamma n}N(M),\qquad N(M)=\frac{n!}{\prod_{i,j}m_{ij}!}. \tag{1}

Compatibility means exactly the stated row and column marginals. All pairs are ordered. Constants are uniform over the feasible matrix sizes and marginals.

Proof. If one matrix dimension is one, its corresponding partition family, being nonempty, contains its unique possible partition. Every member of the other family then has pattern MM, so choosing ϵ≤γ\epsilon\le\gamma proves (1). For s,t≥2s,t\ge2, induct on s+ts+t. The base s=t=2s=t=2 is Theorem 1.14: recording the first cell of each partition identifies its four atoms with the four entries of MM.

Transpose if necessary so that s≥3s\ge3. Write l=l1l=l_1, n′=n−ln'=n-l, and let AG\mathcal A_G be the fiber of A\mathcal A with first cell GG, an ordered (s−1)(s-1)-partition of [n]∖G[n]\setminus G. Put

A′=n′!∏i=2sli!,Z=(nl).A'=\frac{n'!}{\prod_{i=2}^s l_i!},\qquad Z=\binom nl.

Since ∣A∣≥e−ϵnZA′|\mathcal A|\ge e^{-\epsilon n}ZA', Lemma 4.1, or its direct fiber count, gives a family U\mathcal U of at least Ze−ϵn/2Ze^{-\epsilon n}/2 sets GG, each with ∣AG∣≥A′e−ϵn/2|\mathcal A_G|\ge A'e^{-\epsilon n}/2.

Collapse rows 2,…,s2,\ldots,s of MM into one row, obtaining the 2×t2\times t matrix M1M_1. Its entries remain at least ηn\eta n. The pair of families consisting of the two-cell partitions (G,[n]∖G)(G,[n]\setminus G) for G∈UG\in\mathcal U, and B\mathcal B, has density product at least e−2ϵn/2e^{-2\epsilon n}/2. Apply the induction hypothesis for M1M_1, with a small output tolerance τ>0\tau>0. It gives at least

e−τnZUVcompatible pairs (G,B),U=l!∏jm1j!,V=n′!∏j(kj−m1j)!.(2)e^{-\tau n}ZUV\quad\hbox{compatible pairs }(G,B), \qquad U=\frac{l!}{\prod_jm_{1j}!},\quad V=\frac{n'!}{\prod_j(k_j-m_{1j})!}. \tag{2}

This use is valid because 2+t<s+t2+t<s+t. For a fixed GG, there are at most UVUV compatible full-family partitions BB. Removing those GG with fewer than e−τnUV/2e^{-\tau n}UV/2 such members of B\mathcal B loses at most half the count in (2). There are therefore at least Ze−τn/2Ze^{-\tau n}/2 remaining G∈UG\in\mathcal U.

Fix one. For each ordered partition C=(C1,…,Ct)C=(C_1,\ldots,C_t) of GG with ∣Cj∣=m1j|C_j|=m_{1j}, let BG,C\mathcal B_{G,C} be the family of residual partitions (B1−G,…,Bt−G)(B_1-G,\ldots,B_t-G) of [n]∖G[n]\setminus G obtained from members of B\mathcal B with Bj∩G=CjB_j\cap G=C_j. There are UU choices of CC and at most VV residual partitions in each fiber. The count at this GG is at least e−τnUV/2e^{-\tau n}UV/2, so at least Ue−τn/4Ue^{-\tau n}/4 choices of CC satisfy

∣BG,C∣≥Ve−τn/4.(3)|\mathcal B_{G,C}|\ge Ve^{-\tau n}/4. \tag{3}

Let M2M_2 be MM with its first row removed. Its entries are at least ηn≥ηn′\eta n\ge\eta n'. Apply the induction hypothesis to AG,BG,C\mathcal A_G,\mathcal B_{G,C} on the actual n′n'-set, with output tolerance γ/3\gamma/3. Their density product is at least e−(ϵ+τ)n/8e^{-(\epsilon+\tau)n}/8. To justify the parameter order, first take the input tolerance supplied for this residual induction, then choose τ\tau small in comparison with it and γ\gamma. Since n′≥ηnn'\ge\eta n, choose ϵ\epsilon still smaller; for large nn the last density product exceeds the required threshold on n′n' coordinates. Finally decrease ϵ\epsilon to meet the earlier coarse induction (2). This proves at least e−(γ/3)n′N(M2)e^{-(\gamma/3)n'}N(M_2) residual pairs.

Joining the residual row partition to GG, and joining each residual column to its specified CjC_j, reconstructs one original pair with pattern MM. Conversely that original pair determines G,CG,C and both residual partitions, so no pair is counted twice. The total is at least

18e−2τn−(γ/3)n′ZUN(M2).\frac18 e^{-2\tau n-(\gamma/3)n'} ZU N(M_2).

The factorial identity ZUN(M2)=N(M)ZU N(M_2)=N(M) is exact. Choosing τ<γ/8\tau<\gamma/8 and then nn large proves (1).

For fixed η\eta only finitely many shapes occur, since stη≤1st\eta\le1. The same applies to all residual and collapsed shapes. The uniform Theorem 1.14 and the proportional bound n′≥ηnn'\ge\eta n therefore allow a common positive tolerance throughout this finite induction. For the finitely many excluded small nn, reduce ϵ\epsilon so that the density hypothesis forces both partition families to be full. This completes the proof with the stated uniformity. □\square

Precision. In the source's intermediate set-family notation, a single displayed column records a set together with its implicit complement. Here both cells, the residual column sizes, and all normalizing factors are retained explicitly; no one-column pattern count is substituted for a two-cell partition count.

Dependencies. theorem_1_14, lemma_4_1, definitions.