Wiki
Wiki

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

Updated

Hereditary span bound


Proved. Let a≥2a\ge2 and r≥64a2r\ge64a^2. There is no graph HH with

∣H∣=ar+1,α(H)≤a|H|=ar+1,\qquad \alpha(H)\le a

such that every W⊆V(H)W\subseteq V(H) satisfies

e(H[W])≤Dr(∣W∣),Dr(n)=(n2)−(r−1)pr(n),e(H[W])\le D_r(|W|),\qquad D_r(n)=\binom n2-(r-1)p_r(n),

where pr(rq+b)=r(q2)+qbp_r(rq+b)=r\binom q2+qb, 0≤b<r0\le b<r.

Consequently, in a hypothetical coloring counterexample, every induced color graph of independence number at most aa has at most arar vertices in this range. This is not a proof of Problem 617 for an infinite set of parameters: it has the additional small-independence hypothesis.

0. A self-contained finite partition lemma

The short induction in Section 3 of the independent input checks out. It proves the precise finite stability statement used here, so no external stability theorem is needed for the argument below.

Write τa(n)=(n2)−pa(n)\tau_a(n)=\binom n2-p_a(n). If FF is Ka+1K_{a+1}-free on nn vertices, it has a partition into aa possibly empty parts with at most τa(n)−e(F)\tau_a(n)-e(F) internal edges.

Induct on aa, with a=1a=1 immediate; an empty graph is also immediate. Choose a maximum-degree vertex vv, and put A=NF(v)A=N_F(v), d=∣A∣d=|A|, B=V(F)∖AB=V(F)\setminus A. The graph F[A]F[A] is KaK_a-free. Its inductive partition, together with BB, has at most

τa−1(d)−e(F[A])+e(F[B])\tau_{a-1}(d)-e(F[A])+e(F[B])

internal edges. Maximum degree gives

eF(A,B)+2e(F[B])=∑x∈BdF(x)≤d(n−d).e_F(A,B)+2e(F[B])=\sum_{x\in B}d_F(x)\le d(n-d).

Also τa−1(d)+d(n−d)≤τa(n)\tau_{a-1}(d)+d(n-d)\le\tau_a(n), by taking a balanced (a−1)(a-1)-partite graph on AA and joining it completely to the independent part BB. Combining the two inequalities proves the claim.

If T=τa(n)−e(F)T=\tau_a(n)-e(F) and KK is the complete multipartite graph of this partition, then its number of missing cross-edges is

e(K)−e(F)+∑ie(F[Vi])≤2T,e(K)-e(F)+\sum_i e(F[V_i])\le2T,

and e(K)≥τa(n)−2Te(K)\ge\tau_a(n)-2T. This is the finite content of Füredi's Theorem 1 and Corollary 2. The supplied theorem and proof were checked; the induction above is the same maximum-degree mechanism in recursive form.

1. Quadratic-scale extraction

The span theorem and its equality case can be proved together. Set n=ar+εn=ar+\varepsilon, where ε∈{0,1}\varepsilon\in\{0,1\}, and assume α(H)≤a\alpha(H)\le a and all inherited caps. Put F=H‾F=\overline H. Its Turán deficit is

T=e(H)−pa(n)≤Dr(n)−pa(n)≤a2r/2.T=e(H)-p_a(n)\le D_r(n)-p_a(n)\le a^2r/2.

For ε=1\varepsilon=1 the exact difference is ((a2)−1)r+a(\binom a2-1)r+a; for ε=0\varepsilon=0 it is (a2)r\binom a2r. The finite partition lemma gives aa parts ViV_i, at most TT internal FF-edges, and at most

C≤2T≤a2rC\le2T\le a^2r

target edges between parts. Comparing squared part sizes gives

∑i(∣Vi∣−n/a)2≤4T+1≤2a2r+1.\sum_i(|V_i|-n/a)^2\le4T+1\le2a^2r+1.

Since r≥64a2r\ge64a^2, every part has order at least 4r/54r/5.

Mark vertices whose target cross-degree is at least r/3r/3, obtaining XX with t=∣X∣≤6a2t=|X|\le6a^2. Suppose an internal FF-edge uvuv has two unmarked endpoints. In each other part, their common FF-neighborhood has at least 4r/5−2r/3=2r/154r/5-2r/3=2r/15 vertices. Independent uniform choices from these neighborhoods have expected target-edge count at most

225a24r≤225256<1.\frac{225a^2}{4r}\le\frac{225}{256}<1.

Some choice, together with u,vu,v, is a forbidden Ka+1K_{a+1} in FF. Thus Qi=Vi∖XQ_i=V_i\setminus X are nonempty target cliques. The local cap forbids Kr+1K_{r+1}, so write

∣Qi∣=r−di,di≥0,D=∑idi=t−ε<r.|Q_i|=r-d_i,\quad d_i\ge0,\qquad D=\sum_i d_i=t-\varepsilon<r.

If t=0t=0, the overloaded case is already impossible, and the balanced case already consists of aa full cliques. Hence assume t≥1t\ge1.

2. Assigning owners

For x∈Xx\in X, put mi(x)=∣Qi∖NH(x)∣m_i(x)=|Q_i\setminus N_H(x)| and cij=eH(Qi,Qj)c_{ij}=e_H(Q_i,Q_j). Applying the cap to Qi∪Qj∪{x}Q_i\cup Q_j\cup\{x\} and subtracting its clique and incident edges gives

cij≤mi(x)+mj(x)+(di−1)(dj−1)≤mi(x)+mj(x)+didjc_{ij}\le m_i(x)+m_j(x)+(d_i-1)(d_j-1) \le m_i(x)+m_j(x)+d_id_j

when di+dj≥1d_i+d_j\ge1. When both deficits vanish, the stronger bound cij≤mi(x)+mj(x)−r+2c_{ij}\le m_i(x)+m_j(x)-r+2 gives the same final inequality. The relevant set sizes lie between r+1r+1 and 2r+12r+1 since D<rD<r.

Set L=a2+DL=a^2+D. Every exception has an owner i(x)i(x) with mi(x)(x)≤Lm_{i(x)}(x)\le L. Otherwise random representatives from all its nonneighbor lists have expected target-edge count less than

a(a−1)L+D22L2<1.\frac{a(a-1)}L+\frac{D^2}{2L^2}<1.

For D>0D>0, setting u=D/(a2+D)u=D/(a^2+D) bounds this by 1−u+u2/2<11-u+u^2/2<1; for D=0D=0 the first term is 1−1/a1-1/a. An independent choice together with xx contradicts α(H)≤a\alpha(H)\le a.

3. Weighted completion without a concentration loss

For x∈Xx\in X, define

wi(x)=dH(x,Qi)+di−1r−1=1−mi(x)r−1,Si={x∈X:wi(x)≥1/3}.w_i(x)=\frac{d_H(x,Q_i)+d_i-1}{r-1} =1-\frac{m_i(x)}{r-1},\qquad S_i=\{x\in X:w_i(x)\ge1/3\}.

All weights are at most one. They can be negative outside SiS_i when di=0d_i=0; none of the following counts uses such negative incidences. Every owner incidence belongs to Si(x)S_{i(x)}, since L≤7a2L\le7a^2 and 1−L/(r−1)>1/31-L/(r-1)>1/3.

Call an exception active if it belongs to at least two of the sets SiS_i. An inactive exception belongs only to its owner's set. If xx is inactive with owner ii, then, for every j≠ij\ne i,

dH(x,Qj)<(r+2)/3−dj.(1)d_H(x,Q_j)<(r+2)/3-d_j. \tag{1}

Every inactive exception is complete to its owner core, and the inactive exceptions with a common owner form a clique. To prove both assertions, consider a nonadjacent pair consisting either of such an xx and u∈Qiu\in Q_i, or of two inactive exceptions with owner ii. In the first case uu has cross-degree below r/3r/3 by the marking rule. Equation (1) therefore leaves at least (r−4)/3≥r/4(r-4)/3\ge r/4 common target nonneighbors in every other core, in both cases. Uniform representatives from these a−1a-1 lists have expected target-edge count at most 16a2/r<116a^2/r<1. They extend the pair to an independent (a+1)(a+1)-set, a contradiction.

For every S⊆XS\subseteq X the row-plus-exception cap is

∑x∈Swi(x)≤di+e‾H(S)r−1,(2)\sum_{x\in S}w_i(x) \le d_i+\frac{\overline e_H(S)}{r-1}, \tag{2}

where e‾H(S)\overline e_H(S) counts nonedges within SS. For ∣S∣<di|S|<d_i, this follows from wi≤1w_i\le1. Otherwise Qi∪SQ_i\cup S has order between rr and 2r2r, and substitution in DrD_r gives (2) exactly.

Let BB be the sum of weights over nonowner incidences x∈Six\in S_i, and let KK count these incidences. If AA is the active set, then

K≤3B,∣A∣≤K.K\le3B,\qquad |A|\le K.

Put

M=∑x∈Xmi(x)(x),E=∑ie‾H(Si).M=\sum_{x\in X}m_{i(x)}(x),\qquad E=\sum_i\overline e_H(S_i).

Inactive exceptions have no owner holes, whence M≤L∣A∣≤3LBM\le L|A|\le3LB. Every nonedge within SiS_i meets an active exception, because its inactive vertices form a clique. Thus

E≤(t−1)∑i∣A∩Si∣=(t−1)(∣A∣+K)≤6(t−1)B.E\le(t-1)\sum_i|A\cap S_i| =(t-1)(|A|+K)\le6(t-1)B.

Summing (2), the owner weights sum to t−M/(r−1)t-M/(r-1), so

t−Mr−1+B≤D+Er−1.t-\frac M{r-1}+B\le D+\frac E{r-1}.

Consequently

ε+B≤3L+6(t−1)r−1B.(3)\boxed{\varepsilon+B\le \frac{3L+6(t-1)}{r-1}B.} \tag{3}

But

3L+6(t−1)=3a2+9t−3ε−6≤57a2−3ε−6<r−1.3L+6(t-1)=3a^2+9t-3\varepsilon-6 \le57a^2-3\varepsilon-6<r-1.

For ε=1\varepsilon=1, (3) is impossible, proving the span theorem. The argument pays for all owner holes and exceptional nonedges using the additional owner incidences, rather than bounding their product by O(a4)O(a^4).

4. Equality rigidity at the same quadratic threshold

Theorem. If r≥64a2r\ge64a^2, ∣H∣=ar|H|=ar, α(H)≤a\alpha(H)\le a, and all inherited caps hold, then HH partitions into aa cliques of order rr, with a matching between every pair.

For t=0t=0 this was already proved. Otherwise (3) with ε=0\varepsilon=0 forces B=0B=0. Every exception is inactive, so every core together with its assigned exceptions is a clique. These aa cliques partition arar vertices and each has order at most rr; therefore each has order exactly rr. The cap on a full clique plus an outsider gives the asserted matching property.

Checks, improvement, and remaining gap

Both quadratic theorems received a separate audit of extraction, random transversals, negative-weight boundary cases, and the final incidence accounting. No additional external theorem is used beyond the self-contained finite partition lemma above.

These results supersede the previous 64a464a^4 span and 256a4256a^4 equality thresholds. The eighth-power proof and the independent comparison remain as historical records of weaker methods and differing hypotheses.

A hypothetical overloaded core must now have

a>r/8.a>\sqrt r/8.

This still leaves an unbounded range in both main core reductions. Neither an unconditional reduction to this small-independence range nor a universal proof or counterexample has been obtained.