Wiki
Wiki

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

Updated


Statement

Fix γ>1\gamma>1, A≥0A\ge0, and an integer L≥1L\ge1 such that 4⋅2γ≤Lγ−14\cdot2^\gamma\le L^{\gamma-1}. Suppose every nonempty subgraph JJ of a finite simple graph HH, with an integer d≥1d\ge1 satisfying

d≤dJ(v)≤8Ld(v∈V(J)),d\le d_J(v)\le8Ld\qquad(v\in V(J)),

also satisfies d≤A∣V(J)∣γ−1d\le A|V(J)|^{\gamma-1}. Then

e(H)≤4A∣V(H)∣γ.e(H)\le4A|V(H)|^\gamma.

If the degree hypothesis is available only for bipartite subgraphs, the conclusion is e(H)≤8A∣V(H)∣γe(H)\le8A|V(H)|^\gamma. Here the graphs under consideration may be required to avoid a fixed graph, since avoidance passes to every subgraph. A suitable LL exists for each fixed γ>1\gamma>1.

Proof

For a finite set of NN real weights, averaging over all ss-element subsets shows that some subset has weight at least s/Ns/N times the total. We first apply this to a graph GG on N>0N>0 vertices for which

e(G)=CNγ>0,e(G[U])≤C∣U∣γ(U⊆V(G)).e(G)=CN^\gamma>0,\qquad e(G[U])\le C|U|^\gamma\quad(U\subseteq V(G)).

For S⊆V(G)S\subseteq V(G) of size s>0s>0, weight a vertex yy by its number of neighbors in SS. The total weight is ∑v∈SdG(v)\sum_{v\in S}d_G(v). Choose an ss-element set TT carrying at least s/Ns/N of that weight. The ordered adjacencies from SS to TT are at most 2e(G[S∪T])2e(G[S\cup T]). This remains true if S,TS,T overlap, since each undirected edge contributes at most twice. Consequently

s∑v∈SdG(v)≤2Ne(G[S∪T])≤2NC(2s)γ,s\sum_{v\in S}d_G(v) \le2N e(G[S\cup T])\le2NC(2s)^\gamma,

and therefore

∑v∈SdG(v)≤2NC2γsγ−1.(1)\sum_{v\in S}d_G(v)\le2NC2^\gamma s^{\gamma-1}. \tag{1}

Put E=e(G)E=e(G) and let SS now consist of the vertices with dG(v)>2LE/Nd_G(v)>2LE/N. The degree sum 2E2E implies ∣S∣L≤N|S|L\le N. If S≠∅S\ne\varnothing, (1) and the choice of LL give

2∑v∈SdG(v)≤4NC2γ∣S∣γ−1≤CNγ=E.2\sum_{v\in S}d_G(v) \le4NC2^\gamma|S|^{\gamma-1} \le CN^\gamma=E.

The same inequality holds for S=∅S=\varnothing. Delete SS. The remaining induced graph G0G_0 has at least E/2E/2 edges and maximum degree at most 2LE/N2LE/N. Define

d=⌈E4N⌉≥1.d=\left\lceil\frac{E}{4N}\right\rceil\ge1.

Repeatedly delete a vertex of current degree less than dd. Each deletion loses at most d−1d-1 edges; all deletions together lose at most (d−1)N<E/4(d-1)N<E/4. Thus the final graph JJ is nonempty, indeed has more than E/4E/4 edges, and

d≤dJ(v)≤2LE/N≤8Ld,E≤4Nd.(2)d\le d_J(v)\le2LE/N\le8Ld,\qquad E\le4Nd. \tag{2}

This proves the needed degree trimming, including the ceiling case when E/(4N)E/(4N) is an integer.

If HH has no edges the theorem is immediate. Otherwise choose, among its nonempty induced subgraphs, one maximizing e(G)/∣V(G)∣γe(G)/|V(G)|^\gamma. Write that maximum as C>0C>0, and write N=∣V(G)∣N=|V(G)|. Then GG has the density property used above and e(H)≤C∣V(H)∣γe(H)\le C|V(H)|^\gamma. Apply (2) and the assumed degree bound to its subgraph JJ:

CNγ=E≤4Nd≤4NA∣V(J)∣γ−1≤4ANγ.CN^\gamma=E\le4Nd\le4NA|V(J)|^{\gamma-1} \le4AN^\gamma.

Since N>0N>0, C≤4AC\le4A, proving the assertion. The empty graph also satisfies it, with its edge count zero.

Finally, give each vertex an independent uniform color in {0,1}\{0,1\}. Each edge crosses the resulting cut with probability 1/21/2, so a bipartite spanning subgraph has at least half the edges of HH. Apply the first part to this subgraph. Its further subgraphs remain bipartite, and the extra factor is exactly two. Since Lγ−1→∞L^{\gamma-1}\to\infty, an integer LL satisfying the stated inequality exists.

Source and scope

Complete reconstruction of Regularization.weighted_subset, capture_degree_mass, small_set_degree_mass, high_degree_mass, minimum_degree_core, almost_regular_of_density_max, maximum_density, and edge_bound_of_almost_regular, pinned Lean lines 894–1367, together with MaxCut.exists_bipartite_subgraph, lines 1368–1474. The exposition, Proposition 4.1, pp. 4–5, only announces the regularization step.

Used by. Suspension upper bound; Proposition 4.1.

Bears on. #571.