Wiki
Wiki

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

Updated

Ajtai 1980 note ramsey numbers

../

theorem_2: The Ajtai–Komlós–Szemerédi independence bound α(G) ≥ 0.01 (n/t) ln t for a triangle-free graph on n vertices with average degree t, the case r = 3 of Problem 802 and the input of the Ramsey bounds the problem pages consume, with the paper's remark that it is best possible up to the constant when t < n^(1/3+o(1)).

theorem_3: The Ramsey bound R(3,x) < 100 x^2/ln x, from the independence bound by the degree step, with the elementary rewriting as the lower bound H(n) ≥ c √(n ln n) on the least independence number of a triangle-free graph on n vertices that Problems 151 and 610 consume.

theorem_6: The off-diagonal Ramsey bound R(k,x) ≤ 5000^k x^(k-1)/(ln x)^(k-2) for every fixed k ≥ 2 and x large depending on k, by induction on k from the triangle-free case through the few-triangles lemma; at k = 4 the upper bound of Problem 166, and for general k that of Problem 986.


Miklós Ajtai, János Komlós and Endre Szemerédi, A Note on Ramsey Numbers, Journal of Combinatorial Theory, Series A 29 (1980), no. 3, 354--360, DOI 10.1016/0097-3165(80)90030-8 (the running head reads "Series A 29, 354--360 (1980)"; the issue number is from the publisher's record); communicated by the Managing Editors, received June 10, 1980; the authors at the Math Institute, Reáltanoda u. 13--15, 1053 Budapest, Hungary (p. 354). The acknowledgment (p. 360) reads "We are indebted to Joel Spencer, who wrote this paper for us." Cited as [AKS80] on the problem pages. Its three references (p. 360) are the authors' own "A dense infinite Sidon sequence, to appear" (the paper's [1], the site's AKS81b, European J. Combin. 2 (1981), 1--11, not held; the introduction says "A quite different proof of (1) is given in our paper [1]"); Erdős, Graph theory and probability, II, Canad. J. Math. 13 (1961), 346--352 (the paper's [2], the lower bound cx2/(ln⁡x)2cx^2/(\ln x)^2 on R(3,x)R(3,x), filed as erdos_1961_graph_theory_probability); and Graver and Yackel, Some graph theoretic results associated with Ramsey's theorem, J. Combinatorial Theory 4 (1968), 125--175 (the paper's [3], the earlier upper bound cx2ln⁡ln⁡x/ln⁡xcx^2\ln\ln x/\ln x, filed as graver_yackel_1968_graph_theoretic_results_associated_ramsey_theorem; its Proposition 9, "There exists a constant BB so that R(3,y)≤By2log⁡log⁡y/log⁡yR(3,y)\le By^2\log\log y/\log y", is on printed p. 154 (PDF p. 30), located here on the text layer of that page on 2026-09-22 and paged on proposition_9). The edition cited is the publisher's version of record; no preprint or later version is known here. The same independence theorem is restated as Theorem 1 of Ajtai, Erdős, Komlós and Szemerédi 1981, filed as ajtai_1981_turan_s_theorem_sparse_graphs, and sharpened by Shearer 1983, filed as shearer_1983_note_independence_number_triangle_free_graphs.

The copy read for this card is the publisher's open-archive scan of the printed article: 7 pages, printed pp. 354--360 = PDF pp. 1--7 (printed p. nn is PDF p. n−353n-353), a 2003 scan (the scan's metadata names an Acrobat Capture source and a November 2003 creation date) with an OCR text layer that locates passages and garbles the displays, exponents, subscripts, inequality signs and the flow chart of Fig. 1. Provenance: the copy read was obtained from the publisher's open archive, a free download from the article's PDF endpoint on the publisher's site (https://www.sciencedirect.com/science/article/pii/0097316580900308), the DOI https://doi.org/10.1016/0097-3165(80)90030-8 resolving to the same article; 336,495 bytes. The scan prints "Copyright © 1980 by Academic Press, Inc. All rights of reproduction in any form reserved." in the footer of its first page (printed p. 354; the text layer renders the symbol as "0"), every other right reserved.

Read status: claims checked for the abstract, displays (1)--(2), the recalled bounds of Erdős and of Graver and Yackel and the notation (p. 354), the definition of a groupie, Lemma 1 and Theorem 2 with its Note (p. 355), the restatement of Theorem 2 and Remarks 1--3 (pp. 357--358), Theorem 3 and its proof, Lemma 4 (p. 358), Lemma 5, Remark 4 and Theorem 6 (p. 359), the proof of Theorem 6, Theorem 7, the acknowledgment and the references (p. 360), each read clause by clause on the page images of PDF pp. 1--7 on 2026-09-22. The proofs of Lemma 1 (p. 355) and Theorem 3 (p. 358) were read in full on the page images and followed; the proof of Theorem 2 (pp. 355--357, with the flow chart of Fig. 1 on p. 356) and the proofs of Lemmas 4--5 and Theorem 6 (pp. 358--360) were read on the page images for structure only, and the two calculations the paper omits ((11) to g(n′,t′)≥g(n,t)g(n',t')\ge g(n,t) and inequality (15)) were not reconstructed. Nothing here is independently reviewed.

Contents

  • Abstract and introduction (p. 354, page image). The abstract announces upper bounds for the Ramsey function: "We prove R(3,x)<cx2/ln⁡xR(3,x)<cx^2/\ln x and, for each k≥3k\ge3, R(k,x)<ckxk−1/(ln⁡x)k−2R(k,x)<c_kx^{k-1}/(\ln x)^{k-2} asymptotically in xx." The introduction defines R(k,x)R(k,x) as the least nn such that every graph on nn vertices has a clique of size kk or an independent set of size xx, states the two bounds as displays (1) R(3,x)≤cx2/ln⁡xR(3,x)\le cx^2/\ln x and (2) R(k,x)≤ckxk−1/(ln⁡x)k−2R(k,x)\le c_kx^{k-1}/(\ln x)^{k-2} for each kk, and recalls the earlier asymptotic bounds cx2/(ln⁡x)2<R(3,x)<cx2ln⁡ln⁡x/ln⁡xcx^2/(\ln x)^2<R(3,x)<cx^2\ln\ln x/\ln x, the lower one from Erdős [2] and the upper one from Graver and Yackel [3]; it adds that the authors' paper [1] proves (1) by a quite different method. Notation: all graphs finite; n=n(G)n=n(G) the number of vertices, e=e(G)e=e(G) the number of edges, t=t(G)=2e/nt=t(G)=2e/n the average degree, δ=δ(G)=2e/n(n−1)\delta=\delta(G)=2e/n(n-1) the edge density, ω(G)\omega(G) the clique number, α(G)\alpha(G) the independence number, deg⁡(P)\deg(P) the degree of the vertex PP.
  • Groupies and Lemma 1 (p. 355, page image). "Set r(P)r(P) equal to the summation of the degrees of the points QQ adjacent to PP. We call PP a groupie if r(P)≥tdeg⁡(P)r(P)\ge t\deg(P), where t=t(G)t=t(G)." Lemma 1 (quoted): "Every graph GG has a groupie." Proof: ∑Pr(P)=∑Qdeg⁡(Q)2\sum_Pr(P)=\sum_Q\deg(Q)^2 (4); if r(P)<tdeg⁡(P)r(P)<t\deg(P) for all PP then t2n=∑Ptdeg⁡(P)>∑Pdeg⁡(P)2t^2n=\sum_Pt\deg(P)>\sum_P\deg(P)^2 (5), contradicting the Cauchy--Schwarz inequality ∑Pdeg⁡(P)2≥(∑Pdeg⁡(P))2/n=t2n\sum_P\deg(P)^2\ge(\sum_P\deg(P))^2/n=t^2n (6). Followed here.
  • Theorem 2 (p. 355, quoted): "Let GG be a graph with n=n(G)n=n(G), t=t(G)t=t(G). Assume GG is trianglefree. Then (7) α(G)≥0.01(n/t)ln⁡t\alpha(G)\ge0.01(n/t)\ln t." The Note that follows says the paper does not try to optimize its constants. Turán's theorem gives (8) α(G)≥n/(t+1)\alpha(G)\ge n/(t+1), which implies (7) when t<e99t<e^{99}. The proof (pp. 355--357, structure only) is by induction on n(G)n(G) with g(n,t)=0.01(n/t)ln⁡tg(n,t)=0.01(n/t)\ln t (9) and the claim (10) α(G)≥g(G)\alpha(G)\ge g(G): for t<e99t<e^{99} apply (8); otherwise take a groupie PP of degree dd. Case 1, d≥10td\ge10t: delete PP; then t′≤t(n−20)/(n−1)t'\le t(n-20)/(n-1) (11), and a calculation the paper omits as simple gives g(n′,t′)≥g(n,t)g(n',t')\ge g(n,t), whence (12). Case 2, d<10td<10t: delete PP and its neighbors; because GG is triangle-free (the paper marks this as the essential point) exactly r(P)r(P) edges are lost, so e′≤e−tde'\le e-td and t′≤t(n−2d)/(n−1−d)t'\le t(n-2d)/(n-1-d) (14); a second calculation, which the paper refers to Remark 1, gives (15) g(n′,t′)>g(n,t)−1g(n',t')>g(n,t)-1, and α(G)≥α(G′)+1≥g(G′)+1≥g(G)\alpha(G)\ge\alpha(G')+1\ge g(G')+1\ge g(G) (16). Fig. 1 (p. 356) is the flow chart of this loop: while tt is large, find a groupie PP and either delete it alone or star it and delete it with its neighbors; when tt is small, star n/(t+1)n/(t+1) independent points by Turán's theorem. The chart departs from the proof in two places: its threshold reads "t<100t<100" where the proof uses e99e^{99}, and its test on whether deg⁡(P)>10t\deg(P)>10t sends "Yes" to starring PP and deleting it with its neighbors and "No" to deleting PP alone, the reverse of Cases 1 and 2 and of the text on p. 355, where groupies of very high degree are discarded. Paged at theorem_2.
  • Remark 1 (p. 357, page image): inequality (15) is, the paper says, no coincidence; deleting the neighbors of a groupie decreases the edge density, and if each loop produces a groupie of average degree with constant edge density, the number of remaining vertices decays exponentially in time, so about (n/t)ln⁡t(n/t)\ln t independent points are found before t(G)t(G) becomes small; the constant 0.010.01 leaves room to select groupies of moderate degree.
  • Theorem 2 (restatement) (p. 357, quoted): "Let GG be a trianglefree graph with n(G)≤nn(G)\le n and 1≤t(G)≤t1\le t(G)\le t. Then α(G)≥0.01(n/t)ln⁡t\alpha(G)\ge0.01(n/t)\ln t", from "The monotone behavior of g(n,t)g(n,t)". As printed, n(G)≤nn(G)\le n makes it false (a single edge against a large nn); the monotone form needs n(G)≥nn(G)\ge n.
  • Remark 2 (p. 357, page image): for t<n1/3+o(1)t<n^{1/3+o(1)} the paper calls Theorem 2 best possible, by this sketch: the random graph GG on nn vertices with nt/2nt/2 edges has independence number of order at most (n/t)ln⁡t(n/t)\ln t and about t3/6t^3/6 triangles, and removing every vertex that lies on a triangle leaves a triangle-free G′G' with n′=n(G′)∼nn'=n(G')\sim n, t′=t(G′)∼tt'=t(G')\sim t and α(G′)≤α(G)≲c(n′/t′)ln⁡t′\alpha(G')\le\alpha(G)\lesssim c(n'/t')\ln t'. No further argument is printed.
  • Remark 3 (pp. 357--358, quoted): "Erdös has asked if a result similar to Theorem 2 may be proven with the condition 'GG is trianglefree' replaced by 'ω(G)<4\omega(G)<4.' In particular, let f4(n,t)f_4(n,t) be the smallest value of α(G)\alpha(G) over all GG with n(G)≤nn(G)\le n, t(G)≤tt(G)\le t and ω(G)<4\omega(G)<4. We cannot decide if Lim⁡t→∞Lim⁡n→∞f4(n,t)/(n/t)=+∞\operatorname{Lim}_{t\to\infty}\operatorname{Lim}_{n\to\infty}f_4(n,t)/(n/t)=+\infty (?)." This is the K4K_4-free question of Problem 802 in a weaker form (growth faster than n/tn/t rather than the order (n/t)ln⁡t(n/t)\ln t); the 1981 paper's Theorem 2 answers the question as intended (the printed n(G)≤nn(G)\le n admits a one-vertex graph and makes f4≤1f_4\le1) for every fixed clique size and leaves the order open, and Shearer's Remark 4 asks the same question in 1983.
  • Theorem 3 (p. 358, quoted): "R(3,x)<100x2/ln⁡xR(3,x)<100x^2/\ln x." The printed proof is four sentences, followed here: in a triangle-free GG on nn vertices with α(G)<x\alpha(G)<x, each vertex's neighborhood is an independent set, so every degree, and hence t(G)t(G), is below xx; Theorem 2 then gives (17) x>α(G)>0.01(n/x)ln⁡xx>\alpha(G)>0.01(n/x)\ln x, that is (18) n<100x2/ln⁡xn<100x^2/\ln x. The step from t(G)<xt(G)<x to the bound with xx in place of tt is the restatement's monotonicity. Paged at theorem_3.
  • Lemma 4 (p. 358, quoted): with h=h(G)h=h(G) the number of triangles, "Let GG be a graph with n=n(G)n=n(G), e=e(G)e=e(G), h=h(G)h=h(G), t=t(G)t=t(G). Let 0<p<10<p<1 with pn≥3pn\ge3. There exists an induced subgraph G′G' with parameters n′n', e′e', h′h', t′t' satisfying (19) n′>np/2n'>np/2, e′<3ep2e'<3ep^2, h′<3hp3h'<3hp^3, t′<6tpt'<6tp." Proof (pp. 358--359, structure only): a random induced subgraph keeping each vertex independently with probability pp; Chebyshev for n(G′)n(G') (21) and Markov-type bounds for e(G′)e(G') and h(G′)h(G') (22)--(23), each failing with probability less than 1/31/3.
  • Lemma 5 (p. 359, quoted): "Let ε>0\varepsilon>0. Let GG be a graph with n=n(G)n=n(G), t=t(G)t=t(G), h=h(G)h=h(G) and h<nt2−εh<nt^{2-\varepsilon}. If tt is sufficiently large (dependent on ε\varepsilon) (25) α(G)>c′(n/t)ln⁡t\alpha(G)>c'(n/t)\ln t, where c′c' is a positive constant dependent on ε\varepsilon." Proved for c′=0.01ε/48c'=0.01\varepsilon/48 and t>122/εt>12^{2/\varepsilon} (structure only): Lemma 4 with p=tε/4−1p=t^{\varepsilon/4-1}, one vertex deleted from each triangle of G′G' (as h′<n′/2h'<n'/2), then Theorem 2 on the triangle-free G′′G'' with n′′>np/4n''>np/4 and t′′<12tpt''<12tp (27)--(28). Remark 4 states that the hypothesis h<nt2−εh<nt^{2-\varepsilon} of Lemma 5 cannot be relaxed: the Turán graph, n/(t+1)n/(t+1) disjoint cliques of size t+1t+1, has about nt2/6nt^2/6 triangles and α∼n/t\alpha\sim n/t.
  • Theorem 6 (p. 359, quoted): "For every k≥2k\ge2 (29) R(k,x)≤(5000)kxk−1/(ln⁡x)k−2R(k,x)\le(5000)^kx^{k-1}/(\ln x)^{k-2} for xx sufficiently large (dependent on kk)." Proof (pp. 359--360, structure only): trivial for k=2k=2, Theorem 3 for k=3k=3, then induction on kk. Fix ε\varepsilon with (30) 0.96(k−2)−1<ε<(k−2)−10.96(k-2)^{-1}<\varepsilon<(k-2)^{-1} (the paper notes that for (2) with an unspecified ckc_k any sufficiently small ε\varepsilon would do). Let n=n(G)>(5000)kxk−1/(ln⁡x)k−2n=n(G)>(5000)^kx^{k-1}/(\ln x)^{k-2} (31), set m=(5000)k−1xk−2/(ln⁡x)k−3m=(5000)^{k-1}x^{k-2}/(\ln x)^{k-3} and assume ω(G)<k\omega(G)<k; every PP has deg⁡(P)<R(k−1,x)≤m\deg(P)<R(k-1,x)\le m, so t(G)≤mt(G)\le m. Case 1, h(G)<nm2−εh(G)<nm^{2-\varepsilon}: Lemma 5 with c′=0.01ε/48c'=0.01\varepsilon/48 gives α(G)>c′(n/m)ln⁡m>x\alpha(G)>c'(n/m)\ln m>x (32). Case 2, h(G)>nm2−εh(G)>nm^{2-\varepsilon}: a vertex PP on at least m2−ε/3m^{2-\varepsilon}/3 triangles has a neighborhood G′G' with at most mm vertices and at least that many edges, hence a neighbor QQ whose degree inside G′G' is at least 2m1−ε/32m^{1-\varepsilon}/3; the common neighborhood G′′G'' of PP and QQ then has more than 2m1−ε/3>R(k−2,x)2m^{1-\varepsilon}/3>R(k-2,x) vertices (33), so it avoids Kk−2K_{k-2} (which would close a KkK_k with PP and QQ) and contains an independent set of xx points. Paged at theorem_6.
  • Theorem 7 (p. 360, quoted): "Fix ε>0\varepsilon>0. For every k≥2k\ge2 there exists ckc_k so that for xx sufficiently large either R(k,x)<ckR(k−1,x)x/ln⁡xR(k,x)<c_kR(k-1,x)x/\ln x or R(k−1,x)<R(k−2,x)xεR(k-1,x)<R(k-2,x)x^\varepsilon." The paper presents it as a slight alteration of the proof of Theorem 6 and prints no proof.
  • What the paper does not print. No explicit d0d_0 or "t≥t0t\ge t_0" qualifies Theorem 2: the printed statement is α(G)≥0.01(n/t)ln⁡t\alpha(G)\ge0.01(n/t)\ln t for every triangle-free GG, with Turán's bound covering t<e99t<e^{99} (and the bound trivial for t<1t<1, where ln⁡t<0\ln t<0); Shearer's introduction (p. 83) states the same bound as "α>nln⁡d/(100d)\alpha>n\ln d/(100d) for d≥d0d\ge d_0" but credits it to the authors' Sidon-sequence paper (his [1]), not to this note (his [2]); the 1981 paper (p. 314) restates it as "α>0.01(n/t)log⁡t\alpha>0.01(n/t)\log t", crediting both papers; both print strict inequalities. The paper states no bound of the form H(n)≫nlog⁡nH(n)\gg\sqrt{n\log n} on the least independence number of a triangle-free graph on nn vertices; that rewriting of Theorem 3, which Problems 151 and 610 consume, is recorded on the result page as an elementary step made here.

Compiled scope

The paper is compiled at statement depth for the three results the citing problems consume: Theorem 2 (p. 355), Theorem 3 (p. 358) and Theorem 6 (p. 359), read on the page images and paged on theorem_2, theorem_3 and theorem_6. Lemma 1 and Theorem 3 have their proofs followed; the proofs of Theorem 2, Lemmas 4--5 and Theorem 6 were read for structure only, Remarks 2--4 and Theorem 7 carry no printed argument, and nothing here is independently reviewed.

Bears on. #165: Theorem 3 (printed p. 358, PDF p. 5), "R(3,x)<100x2/ln⁡xR(3,x)<100x^2/\ln x", is the upper bound R(3,k)=O(k2/log⁡k)R(3,k)=O(k^2/\log k) the problem records as the one Shearer sharpened, from Theorem 2 (p. 355), "Assume GG is trianglefree. Then α(G)≥0.01(n/t)ln⁡t\alpha(G)\ge0.01(n/t)\ln t"; the introduction (p. 354) places it against Erdős's lower bound cx2/(ln⁡x)2cx^2/(\ln x)^2 and Graver and Yackel's upper bound cx2ln⁡ln⁡x/ln⁡xcx^2\ln\ln x/\ln x, the removal of the ln⁡ln⁡x\ln\ln x factor that the problem's origins describe. #553: Theorem 3 (p. 358) is the upper bound r(K3,Km)=O(m2/log⁡m)r(K_3,K_m)=O(m^2/\log m) that the resolving paper cites for its k=1k=1 base case, with Kim's lower bound; the site credits Shearer's sharper constant. #925: the same Theorem 3 (p. 358), the k=1k=1 input of the resolving paper's induction. #1182: Theorem 3 (p. 358) is the bound r(K3,Ks)<cs2/log⁡sr(K_3,K_s)<cs^2/\log s on which the 1980 lower bound f(n)>An3/2(log⁡n)1/2f(n)>An^{3/2}(\log n)^{1/2} (the site's f(n)f(n), their g(n)g(n); their Theorem 2, p. 198) of Burr, Erdős, Faudree, Rousseau and Schelp rests; they cite the bound from the authors' Sidon-sequence paper (their [1], p. 203), which the note says proves it by a quite different method. #802: Theorem 2 (p. 355) is the case r=3r=3 of the problem's statement, restated as Theorem 1 of the 1981 paper that poses the problem; Remark 2 (p. 357) states that it is best possible up to the constant for t<n1/3+o(1)t<n^{1/3+o(1)}, and Remark 3 (pp. 357--358) records Erdős's question for ω(G)<4\omega(G)<4 and the authors' inability to decide whether f4(n,t)/(n/t)→∞f_4(n,t)/(n/t)\to\infty, the problem's question at r=4r=4 in a weaker form. #151: Theorem 3 (p. 358), rewritten on its result page as H(n)≥cnln⁡nH(n)\ge c\sqrt{n\ln n} for the least independence number H(n)H(n) of a triangle-free graph on nn vertices, is the lower half of c1nlog⁡n≤H(n)≤c2nlog⁡nc_1\sqrt{n\log n}\le H(n)\le c_2\sqrt n\log n that the 1992 paper quotes from this paper and from Erdős 1961. #610: the same rewriting of Theorem 3 is the site's "H(n)≫nlog⁡nH(n)\gg\sqrt{n\log n}", used on that page as context for the transfer from Problem 151. #801: Theorem 2 (p. 355) is the Ajtai--Komlós--Szemerédi bound inside Alon's 1996 proof of the problem's statement, applied to a triangle-free graph on n0.6/4n^{0.6}/4 vertices with independence number below n\sqrt n to force its average degree up to c′n0.1log⁡nc'n^{0.1}\log n. #166: Theorem 6 (p. 359, PDF p. 6), "For every k≥2k\ge2 (29) R(k,x)≤(5000)kxk−1/(ln⁡x)k−2R(k,x)\le(5000)^kx^{k-1}/(\ln x)^{k-2} for xx sufficiently large (dependent on kk)", at k=4k=4 is the upper bound R(4,k)≪k3/(log⁡k)2R(4,k)\ll k^3/(\log k)^2 against which the problem's statement was posed and which Mattheus and Verstraete's theorem meets up to the power of the logarithm. #986: Theorem 6 (p. 359) is the upper bound R(s,k)≪sks−1/(log⁡k)s−2R(s,k)\ll_sk^{s-1}/(\log k)^{s-2} for every fixed s≥3s\ge3, in the site's letters, that the problem's lower bound matches up to the power of the logarithm; the paper's ckc_k is 5000k5000^k.

Results.

  • Theorem 2 (p. 355): a triangle-free graph with nn vertices and average degree tt has α(G)≥0.01(n/t)ln⁡t\alpha(G)\ge0.01(n/t)\ln t; restated on p. 357 for 1≤t(G)≤t1\le t(G)\le t and, as printed, n(G)≤nn(G)\le n, which makes the restatement false (the monotone form needs n(G)≥nn(G)\ge n); best possible up to the constant for t<n1/3+o(1)t<n^{1/3+o(1)} (Remark 2).
  • Theorem 3 (p. 358): R(3,x)<100x2/ln⁡xR(3,x)<100x^2/\ln x; hence, by an elementary step made here, every triangle-free graph on nn vertices has an independent set of cnln⁡nc\sqrt{n\ln n} vertices for large nn.
  • Theorem 6 (p. 359): R(k,x)≤(5000)kxk−1/(ln⁡x)k−2R(k,x)\le(5000)^kx^{k-1}/(\ln x)^{k-2} for every k≥2k\ge2 and xx large depending on kk.

No file of this source is held: no license on record permits its redistribution, and the card cites the edition it names above.