Wiki
Wiki

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

Updated


Statement

Let FF be a finite connected bipartite graph with a fixed two-coloring, let k≥1k\ge1 be an integer, and let 0≤α<20\le\alpha<2. If ex⁡(n,F)=O(nα)\operatorname{ex}(n,F)=O(n^\alpha), then its two-hub replacement satisfies

ex⁡(n,Hk(F))=O ⁣(n1+1/(k+3−α)).\operatorname{ex}(n,H_k(F)) =O\!\left(n^{1+1/(k+3-\alpha)}\right).

Here each old edge is replaced by a path of length k+1k+1, the hubs meet the two old color classes, and there is no hub-hub edge. This is a single-forbidden-graph upper bound, with the constant depending on F,kF,k and the old upper bound. It does not assert a matching lower bound for an arbitrary FF.

Proof

Make the old upper bound uniform for n≥1n\ge1 by increasing its constant to C≥1C\ge1; the finitely many smaller positive inputs permit this. Set

p=k+3,γ=1+1p−α>1.p=k+3,\qquad\gamma=1+\frac1{p-\alpha}>1.

Choose an integer L∗≥1L_*\ge1 with 4⋅2γ≤L∗γ−14\cdot2^\gamma\le L_*^{\gamma-1} and put R=8L∗R=8L_*. All these choices precede the host graph. For almost-regular hosts, use

ε=12p+1(k+4)(k+2)Rp>0.\varepsilon= \frac{1}{2^{p+1}(k+4)(k+2)R^p}>0.

The uniform pruning lemma with r=k+1r=k+1 supplies a threshold parameter BB, independent of the host, so that ∣Bj∣≤εNDj|\mathcal B_j|\le\varepsilon ND^j for every 2≤j≤k+12\le j\le k+1 in an Hk(F)H_k(F)-free graph of maximum degree at most DD.

Let such a host have N≥1N\ge1 vertices and an integer δ≥1\delta\ge1 with

δ≤dH(v)≤Rδ.\delta\le d_H(v)\le R\delta.

If δ<2p\delta<2p, then δ≤2pN1/(p−α)\delta\le2pN^{1/(p-\alpha)}. Otherwise put d=δ−p≥δ/2d=\delta-p\ge\delta/2 and D=RδD=R\delta. The light-path count applies and gives

N(δ−p)p≤UN2δα+2−(p+1)Nδp,U=ΛC2αRα,N(\delta-p)^p\le UN^2\delta^\alpha+2^{-(p+1)}N\delta^p, \qquad U=\Lambda C2^\alpha R^\alpha,

where Λ\Lambda is its fixed threshold-dependent loss. The choice of ε\varepsilon gives precisely the displayed error coefficient. Since (δ−p)p≥2−pδp(\delta-p)^p\ge2^{-p}\delta^p, subtraction yields

2−(p+1)Nδp≤UN2δα.2^{-(p+1)}N\delta^p\le UN^2\delta^\alpha.

Both NN and δ\delta are positive, and p−α>0p-\alpha>0, so

δp−α≤2p+1UN,δ≤(2p+1U)1/(p−α)N1/(p−α).\delta^{p-\alpha}\le2^{p+1}UN,\qquad \delta\le(2^{p+1}U)^{1/(p-\alpha)}N^{1/(p-\alpha)}.

Choose A=max⁡{2p,(2p+1U)1/(p−α)}A=\max\{2p,(2^{p+1}U)^{1/(p-\alpha)}\}. This covers both degree ranges and proves δ≤ANγ−1\delta\le AN^{\gamma-1} for every required almost-regular host. In particular it applies to all bipartite such hosts. The bipartite form of regularization then gives e(H)≤8Anγe(H)\le8An^\gamma for every nn-vertex Hk(F)H_k(F)-free graph. Avoidance is hereditary under all subgraph choices made there. At n=0n=0 the edge count is zero. Taking the maximum over hosts proves the stated asymptotic bound.

Source and scope

Exposition, Proposition 4.1, pp. 4–5. Its compressed path-pruning argument is completed by the linked same-source lemmas. The final formal deductions are HubPathDegree.almost_regular, BipartiteRegularization.bound, and HubPathBounds.edge_bound and upper_isBigO, pinned Lean lines 9966–10175. The source invokes a more general PowerErrorAbsorption interface with its error parameter set to zero; the division and positive power extraction above prove exactly that needed instance directly. No logarithmic term is hidden in the OO notation.

Used by. Proposition 4.2.

Bears on. #571.