Wiki
Wiki

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

Updated

Independent investigation of Erdős Problem 617


Status, sources, and conventions

Unresolved. This investigation neither proves the assertion for every r≥3r\ge3 nor constructs a counterexample. It establishes parameterized structural restrictions and records concrete failed approaches. It does not claim another isolated fixed case.

The only supplied mathematical sources used were:

  • Problem 617, the canonical problem page.
  • source record, Erdős and Gyárfás, Split and balanced colorings of complete graphs (1999).

There was no web browsing and no consultation of later research notes. The notes linked below were generated during this investigation. Turán's theorem and, for the homogeneous-coloring observation, Brooks' theorem are used as in the supplied paper. The stability and other structural lemmas are proved here rather than attributed to outside sources.

“Hypothetical counterexample” means an rr-edge-coloring of Kr2+1K_{r^2+1} in which every (r+1)(r+1)-set sees every color. Write GiG_i for its color-ii graph. Thus

α(Gi)≤r(1≤i≤r).(C)\alpha(G_i)\le r\qquad(1\le i\le r). \tag{C}

For every (r+1)(r+1)-set XX,

e(Gi[X])≤(r+12)−(r−1)=(r2)+1.(L)e(G_i[X])\le \binom{r+1}{2}-(r-1)=\binom r2+1. \tag{L}

In particular, ω(Gi)≤r\omega(G_i)\le r. If AA is a monochromatic KrK_r of color ii, then

di(x,A)≤1(x∉A).(K)d_i(x,A)\le1\qquad(x\notin A). \tag{K}

Indeed, the rr edges from xx to AA must supply every other color.

1. The hereditary edge budget

Let Tr(m)T_r(m) be the minimum edge count of an mm-vertex graph with independence number at most rr. By Turán's theorem it is the sum of (bj2)\binom{b_j}{2} for an equitable partition into rr parts. In particular,

Tr(ar+1)=r(a2)+a.T_r(ar+1)=r\binom a2+a.

A necessary condition on each individual color graph is

e(Gi[S])≤(∣S∣2)−(r−1)Tr(∣S∣)for every S⊆V.(H)e(G_i[S])\le \binom{|S|}{2}-(r-1)T_r(|S|) \quad\text{for every }S\subseteq V. \tag{H}

This follows by counting the edges required by the other r−1r-1 colors on SS.

Condition (H) is essential. An individual graph can pass the minority-edge bound, (L), and several small-independence tests while failing (H). Such a graph is not a counterexample coloring.

For the whole graph set

T=Tr(r2+1)=r(r2)+r=r(r2−r+2)2.T=T_r(r^2+1)=r\binom r2+r =\frac{r(r^2-r+2)}2.

The total edge surplus over rTrT is

(r2+12)−rT=r2(r−1)2.(B)\binom{r^2+1}{2}-rT=\frac{r^2(r-1)}2. \tag{B}

2. Main checked result: uniform capacity for small independence number

The following complete argument was also preserved, with an independent proof audit including the numerical constants and empty-set cases, in companion notes that are not retained.

Theorem

Let a≥2a\ge2 be an integer. If

r≥256a6,r\ge256a^6,

then, for every color ii and every vertex subset HH,

\alpha(G_i[H])\le a\quad\Longrightarrow\quad |H|\le ar. \tag{P_a}

In particular, this gives (Pa)(P_a) simultaneously for all integers 2≤a≤⌊(r/256)1/6⌋2\le a\le\lfloor(r/256)^{1/6}\rfloor.

We first prove the elementary stability fact needed in the argument.

Lemma: a near-Turán partition

Let s≥1s\ge1, and let FF be a Ks+1K_{s+1}-free graph on nn vertices. Define

Qs(n)=s−12sn2,D=Qs(n)−e(F)≥0.Q_s(n)=\frac{s-1}{2s}n^2,\qquad D=Q_s(n)-e(F)\ge0.

There is a partition V(F)=V1∪˙⋯∪˙VsV(F)=V_1\dot\cup\cdots\dot\cup V_s such that the total number II of edges internal to the parts is at most DD. For this partition, writing ni=∣Vi∣n_i=|V_i| and WW for the number of missing cross-part pairs, one has

W≤2D,∣ni−n/s∣≤2D(1≤i≤s).(7)W\le2D, \qquad |n_i-n/s|\le2\sqrt D\quad(1\le i\le s). \tag{7}

Proof. Induct on ss. For s=1s=1, the graph is edgeless and the result is immediate. For s≥2s\ge2, choose a maximum-degree vertex vv, with degree dd, and put U=NF(v)U=N_F(v) and B=V(F)∖UB=V(F)\setminus U. The graph F[U]F[U] is KsK_s-free. Set

DU=Qs−1(d)−e(F[U])≥0.D_U=Q_{s-1}(d)-e(F[U])\ge0.

The maximum-degree condition gives

e(U,B)+2e(B)≤d∣B∣,e(U,B)+2e(B)\le d|B|,

so

e(F)≤e(F[U])+d(n−d)−e(B).e(F)\le e(F[U])+d(n-d)-e(B).

The identity

Qs(n)−Qs−1(d)−d(n−d)=s2(s−1)(d−s−1sn)2Q_s(n)-Q_{s-1}(d)-d(n-d) =\frac{s}{2(s-1)}\left(d-\frac{s-1}{s}n\right)^2

therefore implies D≥DU+e(B)D\ge D_U+e(B). By induction, partition UU into s−1s-1 parts with at most DUD_U internal edges and take BB as the last part. The resulting partition has I≤DI\le D.

The number of all possible cross-part pairs is

n2−∑ini22=Qs(n)−12∑i(ni−n/s)2.\frac{n^2-\sum_i n_i^2}{2} =Q_s(n)-\frac12\sum_i(n_i-n/s)^2.

Subtracting the actual number e(F)−Ie(F)-I of cross-part edges gives

W=D+I−12∑i(ni−n/s)2.(8)W=D+I-\frac12\sum_i(n_i-n/s)^2. \tag{8}

Since 0≤W0\le W and I≤DI\le D, this proves (7). □\square

Proof of the theorem

If (Pa)(P_a) fails, restrict to a subset HH of exactly

n=ar+1n=ar+1

vertices with α(Gi[H])≤a\alpha(G_i[H])\le a. Let FF be the graph on HH formed by edges of all colors other than ii. Then FF is Ka+1K_{a+1}-free.

For each other color jj, the graph Gj[H]G_j[H] has independence number at most rr. The complement form of Turán's theorem gives

e(Gj[H])≥(r−1)(a2)+(a+12)=r(a2)+a.e(G_j[H])\ge(r-1)\binom a2+\binom{a+1}{2} =r\binom a2+a.

Summing over the r−1r-1 colors different from ii yields

e(F)≥(r−1)(r(a2)+a).e(F)\ge(r-1)\left(r\binom a2+a\right).

Consequently its deficit from the real Turán bound satisfies

D=Qa(ar+1)−e(F)≤((a2)−1)r+a+a−12a≤a2r.(9)\begin{aligned} D&=Q_a(ar+1)-e(F)\\ &\le\left(\binom a2-1\right)r+a+\frac{a-1}{2a}\\ &\le a^2r. \tag{9} \end{aligned}

The last inequality holds for all a≥2,r≥1a\ge2,r\ge1.

Apply the lemma with s=as=a. Its partition has at most DD internal edges, at most

W≤2D≤2a2r(10)W\le2D\le2a^2r \tag{10}

missing cross-part pairs, and each part has size at least

n/a−2D≥r−2ar≥9r10.(11)n/a-2\sqrt D\ge r-2a\sqrt r\ge\frac{9r}{10}. \tag{11}

For the final inequality it suffices that r≥400a2r\ge400a^2, which follows from r≥256a6r\ge256a^6 and a≥2a\ge2.

For a vertex xx, let b(x)b(x) be the number of its nonneighbors in parts different from its own. Define

X={x:b(x)≥2r/5},k=∣X∣.X=\{x:b(x)\ge2r/5\},\qquad k=|X|.

Since ∑xb(x)=2W\sum_xb(x)=2W, we have

k≤2W2r/5≤10a2.(12)k\le\frac{2W}{2r/5}\le10a^2. \tag{12}

We claim that every set Ci=Vi∖XC_i=V_i\setminus X is independent in FF. Suppose that xyxy is an edge with x,y∈Cix,y\in C_i. In every other full part VjV_j, the common-neighbor set

Sj=NF(x)∩NF(y)∩VjS_j=N_F(x)\cap N_F(y)\cap V_j

has size strictly greater than

9r10−2r5−2r5=r10.\frac{9r}{10}-\frac{2r}{5}-\frac{2r}{5}=\frac r{10}.

Choose independently and uniformly one vertex from each of these a−1a-1 sets. The probability that some chosen pair is a nonedge is at most

∑j<ℓ, j,ℓ≠ieF‾(Sj,Sℓ)∣Sj∣∣Sℓ∣≤100Wr2≤200a2r<1.(13)\sum_{j<\ell,\,j,\ell\ne i} \frac{e_{\overline F}(S_j,S_\ell)}{|S_j||S_\ell|} \le\frac{100W}{r^2} \le\frac{200a^2}{r}<1. \tag{13}

For a=2a=2 the sum is empty, which is harmless. Thus there is an (a−1)(a-1)-clique of common neighbors of x,yx,y, giving a forbidden Ka+1K_{a+1}. This proves the claim. Notice that (13) uses the full parts VjV_j, not the smaller cores CjC_j; no subtraction of marked vertices is required here.

Next consider any x∈Xx\in X. It has at most W\sqrt W neighbors in at least one of the cores C1,…,CaC_1,\ldots,C_a. Otherwise choose uniformly one neighbor of xx in each core. The same union bound, now with all aa cores, gives failure probability at most

Wmin⁡j∣NF(x)∩Cj∣2<1.\frac{W}{\min_j |N_F(x)\cap C_j|^2}<1.

This would produce a KaK_a inside NF(x)N_F(x), hence a forbidden Ka+1K_{a+1}. The argument also covers W=0W=0: if every core contained a neighbor of xx, all cross pairs would be edges. If a core is empty, its zero neighbor count already proves the assertion.

Assign every x∈Xx\in X to a core in which it has at most W\sqrt W neighbors. Since the cores are independent, the total number of internal edges of this new aa-partition is at most

kW+(k2)≤102 a3r+50a4.(14)\begin{aligned} k\sqrt W+\binom k2 &\le10\sqrt2\,a^3\sqrt r+50a^4. \tag{14} \end{aligned}

Under r≥256a6r\ge256a^6, one has r≥16a3\sqrt r\ge16a^3 and a≥2a\ge2, so

102 a3r≤528r≤1516r,50a4≤25128a2r≤25512r.10\sqrt2\,a^3\sqrt r \le\frac{5\sqrt2}{8}r \le\frac{15}{16}r, \qquad 50a^4\le\frac{25}{128a^2}r\le\frac{25}{512}r.

Thus (14) is at most 505r/512<r−1505r/512<r-1; indeed r≥256⋅26=16384>512/7r\ge256\cdot2^6=16384>512/7.

Since ∣H∣=ar+1|H|=ar+1, one part of this new aa-partition has at least r+1r+1 vertices. Any (r+1)(r+1)-subset of that part spans fewer than r−1r-1 edges of FF. Yet every such set must have at least one edge in each of the r−1r-1 colors other than ii, requiring at least r−1r-1 edges of FF. This contradiction proves (Pa)(P_a). □\square

3. Stronger capacity statements at independence two and three

These arguments hold uniformly in rr; they do not reprove a single isolated value. Their full provenance and an independent check were recorded in companion notes that are not retained.

Independence two

The preceding template analysis is superseded by the following general theorem.

Theorem (two-slot capacity). In a hypothetical counterexample, every vertex set SS satisfying α(Gi[S])≤2\alpha(G_i[S])\le2 has ∣S∣≤2r|S|\le2r, for every color ii.

Proof. Put H=Gi[S]H=G_i[S] and m=∣S∣m=|S|, and suppose m>2rm>2r. The earlier local KrK_r lemma shows that HH cannot contain KrK_r: otherwise its complement outside that clique would itself be a clique and m≤2rm\le2r. Thus ω(H)≤r−1\omega(H)\le r-1. Since α(H)≤2\alpha(H)\le2, the nonneighbors in HH of any vertex form a clique of HH, of size at most r−1r-1. Consequently

e(H‾)≤m(r−1)2.e(\overline H)\le \frac{m(r-1)}2.

Every other color jj induces a graph on SS with independence number at most rr. Every graph on mm vertices and ee edges has an independent set of size at least m−em-e: choose at most one endpoint from each edge to obtain a vertex cover. Hence e(Gj[S])≥m−re(G_j[S])\ge m-r. Summing the r−1r-1 other colors gives

e(H‾)≥(r−1)(m−r)>m(r−1)2,e(\overline H)\ge(r-1)(m-r)>\frac{m(r-1)}2,

because m>2rm>2r, a contradiction. □\square

The proof only requires pairwise edge-disjoint spanning graphs G1,…,GrG_1,\ldots,G_r with α(Gi)≤r\alpha(G_i)\le r: one may first color any uncovered pairs arbitrarily, which preserves all these independence bounds. It therefore applies to proposed packings as well as complete edge partitions.

Corollary. In a hypothetical counterexample, every color graph has a connected component of independence number at least three. Indeed, if all its components had independence number at most two, a clique component has at most rr vertices and a component of independence number two has at most 2r2r vertices. Summing would give ∣V∣≤rα(Gi)≤r2|V|\le r\alpha(G_i)\le r^2.

This eliminates every construction assembled in each color from cliques and pentagon-type independence-two pieces, irrespective of their sizes, balance, overlap supports, or edge densities.

Equality in the independence-two bound

Here r≥6r\ge6.

Theorem. If SS has 2r2r vertices and α(Gi[S])≤2\alpha(G_i[S])\le2 in a hypothetical counterexample, then SS is the disjoint union of two monochromatic KrK_r's in color ii, and all color-ii edges between them form a matching.

Proof. Again write H=Gi[S]H=G_i[S]. If HH contains a Kr=AK_r=A, the local KrK_r lemma makes S∖AS\setminus A a clique, hence another KrK_r. Every vertex of either clique has at most one neighbor in the other by the local KrK_r restriction, so the edges between them form a matching.

Suppose instead that HH has no KrK_r. Its complement FF is triangle-free and has maximum degree at most r−1r-1. The proof of the capacity theorem gives

r(r−1)≤e(F)≤r(r−1),r(r-1)\le e(F)\le r(r-1),

so FF is (r−1)(r-1)-regular. (Each other color restricted to SS has exactly rr edges and independence number at most rr, and is therefore a perfect matching: if two of these edges met, a vertex cover with at most r−1r-1 vertices could be selected.)

We use the elementary fact that a nonbipartite triangle-free graph on NN vertices has minimum degree at most 2N/52N/5. To prove it, take a shortest odd cycle CC of length q≥5q\ge5. It has no chord, and any vertex outside CC has at most two neighbors on CC. For if it had at least three, the cyclic gaps between consecutive neighbors are at least two, and one is odd; that odd gap has length at most q−4q-4, producing an odd cycle of length at most q−2q-2 through the outside vertex. This contradicts minimality. Counting incidences into CC gives

qδ≤∑v∈Cd(v)≤2N,q\delta\le\sum_{v\in C}d(v)\le2N,

and q≥5q\ge5 proves the fact. Applied to FF, it shows that FF is bipartite, since

r−1>4r5(r≥6).r-1>\frac{4r}{5}\qquad(r\ge6).

One side of a bipartition of FF has at least rr vertices, giving a KrK_r in HH, a contradiction. □\square

Independence three

For r≥6r\ge6, the equality theorem yields another capacity statement: every SS with α(Gi[S])≤3\alpha(G_i[S])\le3 has ∣S∣≤3r|S|\le3r.

Indeed, suppose H=Gi[S]H=G_i[S] has m>3rm>3r. It cannot contain an independence-two induced subgraph on 2r2r vertices. Such a set would, by the equality theorem, split into two color-ii KrK_r's A,BA,B, with a matching between them. If two vertices x,yx,y outside A∪BA\cup B were nonadjacent in color ii, at most two vertices in each of A,BA,B could be adjacent to either xx or yy. Since r≥4r\ge4, there are at least two candidates in each clique, and the matching between AA and BB permits a nonadjacent choice a,ba,b. Then {a,b,x,y}\{a,b,x,y\} is independent in HH, impossible. Thus H∖(A∪B)H\setminus(A\cup B) is a clique, so it has at most rr vertices and m≤3rm\le3r, again impossible.

Every nonneighborhood in HH has independence number at most two, and therefore has size at most 2r−12r-1 (the size 2r2r case was just ruled out). Thus Δ(H‾)≤2r−1\Delta(\overline H)\le2r-1. The complement form of Turán's theorem supplies the universal supporting-line bound

e(Gj[S])≥3m−6r(j≠i).e(G_j[S])\ge3m-6r\qquad(j\ne i).

One way to see the supporting line is to write the Turán minimum as ∑a=1r(na2)\sum_{a=1}^r\binom{n_a}{2} with ∑ana=m\sum_a n_a=m and use (na2)≥3na−6\binom{n_a}{2}\ge3n_a-6 for every nonnegative integer nan_a; this last difference is (na−3)(na−4)/2≥0(n_a-3)(n_a-4)/2\ge0.

Consequently

0≥(r−1)(3m−6r)−m(2r−1)2.0\ge (r-1)(3m-6r)-\frac{m(2r-1)}2.

The right side is increasing in mm and at m=3r+1m=3r+1 equals (r−5)/2>0(r-5)/2>0, a contradiction.

4. Minimal unresolved subgraphs

For r≥6r\ge6, put

Ar=max⁡{3,⌊(r/256)1/6⌋}.A_r=\max\left\{3,\left\lfloor(r/256)^{1/6}\right\rfloor\right\}.

Every color of a hypothetical counterexample has a connected induced subgraph HH with

∣H∣=ar+1,α(H)=a>Ar,δ(H)≥r.(M)|H|=ar+1,\qquad \alpha(H)=a>A_r,\qquad \delta(H)\ge r. \tag{M}

To prove this, choose an inclusion-minimal induced set with ∣H∣>rα(H)|H|>r\alpha(H); the whole color graph is such a set. Write a=α(H)a=\alpha(H). Deleting one vertex and applying minimality gives

∣H∣−1≤rα(H−v)≤ra,|H|-1\le r\alpha(H-v)\le ra,

so ∣H∣=ar+1|H|=ar+1. If a vertex had degree at most r−1r-1, deleting its closed neighborhood would leave at least (a−1)r+1(a-1)r+1 vertices of independence number at most a−1a-1, a smaller violation. Disconnectedness would also contradict minimality by adding the capacity inequalities for the components. The proved capacity statements force a>Ara>A_r.

This is a precise remaining obstruction, not a contradiction. No argument establishes a≤Ara\le A_r. Applying the main theorem with a=ra=r is invalid, since its hypothesis would be r≥256r6r\ge256r^6.

5. Clique packings and interaction between colors

5.1 A strengthened lower bound for every color

Every graph satisfying (C) and (L) has

e(Gi)≥T+r−1.(E)e(G_i)\ge T+r-1. \tag{E}

Repeatedly choose a minimum-degree vertex of the remaining graph and delete its closed neighborhood BjB_j, of size bjb_j. The chosen vertices are independent, so there are at most rr batches. If qjq_j is the number of missing edges inside BjB_j, the number of edges removed at that step is at least

bj(bj−1)−e(Gi[Bj])=(bj2)+qj.b_j(b_j-1)-e(G_i[B_j])=\binom{b_j}{2}+q_j.

Thus

e(Gi)≥∑j(bj2)+∑jqj.e(G_i)\ge\sum_j\binom{b_j}{2}+\sum_jq_j.

Convexity gives the first sum at least TT. Some batch has at least r+1r+1 vertices, so (L) supplies qj≥r−1q_j\ge r-1. More information can be retained:

e(Gi)≥∑j(bj2)+∑bj≥r+1⌈(r−1)bj(bj−1)r(r+1)⌉.e(G_i)\ge \sum_j\binom{b_j}{2}+ \sum_{b_j\ge r+1} \left\lceil\frac{(r-1)b_j(b_j-1)}{r(r+1)}\right\rceil.

5.2 Matching transversals

Suppose kk disjoint parts each have at least k−1k-1 vertices, and the forbidden pairs between every two parts form a matching. If there is no independent transversal, all parts have size exactly k−1k-1, and the cross edges form k−1k-1 disjoint transversal KkK_k's.

A part of size at least kk could be left last in a greedy choice. A missing edge from a pairwise perfect matching can similarly be exploited by starting with an unmatched vertex and leaving its opposite part last. Hence all pairwise matchings are perfect. If two nonadjacent vertices in different parts share a matched neighbor in a third part, choose those two first and the third part last; the last part loses at most k−2k-2 choices. This proves coherence. When applying the lemma to monochromatic clique rows, discard within-row edges: they do not affect transversals.

If a color had rr disjoint KrK_r's, its remaining vertex would need one neighbor in each row. Avoiding those neighbors and applying this lemma forces r−1r-1 crossing KrK_r's. That color has at least

(2r−1)(r2)+r(2r-1)\binom r2+r

edges. Together with (E) for the other colors, this exceeds (r2+12)\binom{r^2+1}{2} by (r−1)(r−2)/2(r-1)(r-2)/2. Thus no color has such a packing.

5.3 Cross-color fibers

Let pp be the largest number of disjoint monochromatic KrK_r's in any one color. Fix a color-ii clique A=KrA=K_r. The star of an outside vertex into AA is either rainbow or omits ii and repeats exactly one other color.

For j≠ij\ne i, let Ej(A)E_j(A) consist of the vertices whose repeated color is jj, and let ej=∣Ej(A)∣e_j=|E_j(A)|. Every remaining outside vertex has a unique color-jj neighbor in AA. For each a∈Aa\in A, its fiber together with aa is a color-jj clique: otherwise (A∖{a})∪{x,y}(A\setminus\{a\})\cup\{x,y\} would omit jj. These rr disjoint cliques have total size r2+1−ejr^2+1-e_j, so their total deficit from size rr is ej−1e_j-1. At least r−ej+1r-e_j+1 are full KrK_r's. Consequently

ej≥r−p+1,e_j\ge r-p+1,

and, writing ti(A)t_i(A) for the outgoing color-ii edges of AA,

ti(A)≤(r−1)(p−1)+1.(F)t_i(A)\le (r-1)(p-1)+1. \tag{F}

5.4 Excluding a packing of r−1r-1 full cliques

The result is

p≤r−2.(P)p\le r-2. \tag{P}

Here are the main calculations; all steps were recorded and independently audited in companion notes that are not retained.

Suppose p=r−1p=r-1, with rows A1,…,Ar−1A_1,\ldots,A_{r-1} and residue RR of size r+1r+1. Let F=Gi[R]‾F=\overline{G_i[R]}, and let ss count its nonisolated vertices. Since e(F)≥r−1e(F)\ge r-1, s≥2s\ge2. Any edge xyxy of FF must have distinct color-ii neighbors in every row. The transversal lemma, after deleting those two neighbors, forces at least r−2r-2 color-ii edges between every pair of rows. Each of the ss residue vertices has a color-ii neighbor in every row. Thus

ti(Aa)≥(r−2)2+s.t_i(A_a)\ge(r-2)^2+s.

Comparing with (F) gives s≤r−1s\le r-1.

Each ej(Aa)≥2e_j(A_a)\ge2, while

∑j≠iej(Aa)≤3r−3−s.\sum_{j\ne i}e_j(A_a)\le3r-3-s.

Therefore at least ss other colors have ej=2e_j=2, and hence also have r−1r-1 disjoint full cliques. This is the crucial propagation to other colors.

For r≥5r\ge5, e(F)≥4e(F)\ge4 implies s≥4s\ge4. Any propagated color has at least

(r−1)(r2)+(r−2)(r−12)+4(r−1)(r-1)\binom r2+(r-2)\binom{r-1}{2}+4(r-1)

edges. Its surplus over TT is

A=r3−6r2+15r−122.A=\frac{r^3-6r^2+15r-12}{2}.

There are at least three such colors. Their surpluses and (E) for the rest exceed the budget (B), because

3A+(r−3)(r−1)−r2(r−1)2=(r−3)(r−2)(2r−5)2>0.3A+(r-3)(r-1)-\frac{r^2(r-1)}2 =\frac{(r-3)(r-2)(2r-5)}2>0.

The same structural statement at the lower parameters follows from the canonical residue calculation in the linked proof; no separate new fixed case is claimed here. Substituting (P) into (F) gives

ti(A)≤(r−2)2.t_i(A)\le(r-2)^2.

5.5 Last derived extension: p≤r−3p\le r-3 for r≥8r\ge8

The complete derivation was preserved in a companion note that is not retained. Unlike the preceding packing theorem, this extension did not complete a separate independent audit. Its proof was checked locally, including the following support lemma and budget calculation; it is kept distinct from the independently audited results.

A graph FF on 2r+12r+1 vertices containing a triangle, with all triangle vertices lying in a set SS of size s≤rs\le r, satisfies

e(F)≤r2+s−1.(TS)e(F)\le r^2+s-1. \tag{TS}

To see this, symmetrize Y=V(F)∖SY=V(F)\setminus S without creating a triangle meeting YY. Choose a maximum-degree y∈Yy\in Y, put A=NY(y)A=N_Y(y), B=Y∖AB=Y\setminus A, and replace every vertex of BB by a twin of yy. The edge gain is

∑b∈B(d(y)−d(b))+e(F[B])≥0.\sum_{b\in B}(d(y)-d(b))+e(F[B])\ge0.

Then make the independent set AA twins of a vertex with maximum number of neighbors in SS. The resulting YY is complete bipartite. Its two neighborhoods C,DC,D in SS are disjoint independent sets. With t=∣S∖(C∪D)∣≥1t=|S\setminus(C\cup D)|\ge1, counting yields

e(F)≤⌊(2r+1−t)24⌋+ts−t(t+1)2≤r2+s−1.e(F)\le\left\lfloor\frac{(2r+1-t)^2}{4}\right\rfloor +ts-\frac{t(t+1)}2 \le r^2+s-1.

Now suppose p=r−2p=r-2. The residue has 2r+12r+1 vertices, so the two-slot capacity theorem forces an independent triple in its color-ii graph. Let ss be the number of vertices participating in such triples. The other colors require

e(Gi[R]‾)≥(r−1)(r+2)=r2+r−2.e(\overline{G_i[R]})\ge(r-1)(r+2)=r^2+r-2.

Applying (TS), if necessary, gives s≥r−1s\ge r-1. Each independent triple and the matching-transversal lemma force r−3r-3 cross edges between every pair of rows and attachments from all ss supported vertices. The duplicate-color budget then propagates the same packing to at least three other colors.

Each of these at least four colors has surplus at least

A2=r3−8r2+15r−142A_2=\frac{r^3-8r^2+15r-14}{2}

over TT. The budget excess is

4A2+(r−4)(r−1)−r2(r−1)2=3r3−29r2+50r−482=(r−8)(3r2−5r+10)2+16>04A_2+(r-4)(r-1)-\frac{r^2(r-1)}2 =\frac{3r^3-29r^2+50r-48}{2} =\frac{(r-8)(3r^2-5r+10)}2+16>0

for r≥8r\ge8. This yields the stated extension. No argument yet forces any color to have a packing of size r−2r-2.

6. The affine-plane extension: exact and quantitative obstructions

Assume an affine plane of order rr exists. Start with the paper's coloring of its r2r^2 points: r−1r-1 parallel classes get separate colors, and the last two classes share the remaining color.

Unchanged extension fails

For a singleton-class color ii, the independent rr-sets are all transversals of its rr lines. The new vertex's color-ii neighbors must meet every such transversal, and therefore must contain an entire old line. Two singleton colors force two intersecting lines, assigning two colors to the edge from the new vertex to their intersection. Thus the unchanged affine coloring cannot be extended.

Recolored-edge count

Let AiA_i be the old points joined to the new vertex in color ii, ai=∣Ai∣a_i=|A_i|, and let ti,dit_i,d_i count old edges recolored into and out of color ii. Put t=∑iti=∑idit=\sum_i t_i=\sum_i d_i.

For a singleton color with no entire line in AiA_i, let b1≤b2≤⋯≤brb_1\le b_2\le\cdots\le b_r be the positive residual line sizes. Every transversal must contain an added color-ii edge. A random transversal gives

ti≥b1b2≥b1+b2−1≥2r−ai−1.t_i\ge b_1b_2\ge b_1+b_2-1\ge2r-a_i-1.

Thus ai+ti≥2r−1a_i+t_i\ge2r-1.

If AiA_i contains a whole old line, that line together with the new vertex needs at least r−1r-1 recolored edges to supply other colors. Turán's bound gives

ai+ti≥r+di≥2r−1.a_i+t_i\ge r+d_i\ge2r-1.

Equality would force a forbidden Kr+1K_{r+1}, so this case is strict.

For the merged color, independent rr-sets are permutation transversals of an r×rr\times r grid. The union bound gives

ar+trr−1≥r.a_r+\frac{t_r}{r-1}\ge r.

Summing proves

t≥(r−1)2+r−2r−1tr.t\ge(r-1)^2+\frac{r-2}{r-1}t_r.

Equality t=(r−1)2t=(r-1)^2 is impossible. In that equality case the merged star is an entire line; the nearly-full singleton star lines must all meet at one point on it. The remaining points force ai=r,ti=r−1a_i=r,t_i=r-1 and a specific added-edge star in every singleton color. Some added edge must have been removed from another singleton color. But deleting an edge from a singleton line requires at least rr additions to that color: take the two endpoints and one random point from every other old line, and apply the union bound. This contradicts ti=r−1t_i=r-1. Hence

t≥(r−1)2+1.(AR)t\ge(r-1)^2+1. \tag{AR}

Vertex-cover bound for the changed edges

Delete kk old points and keep the old coloring on the remainder; allow k+1k+1 new vertices with arbitrary incident edges. Define

q=min⁡{d≥1:(d+12)≥r−1}.q=\min\{d\ge1:\binom{d+1}{2}\ge r-1\}.

A hypothetical counterexample obtained this way must satisfy

k≥⌈(r−1)q2r+q−2⌉=2r−O(r).(AV)k\ge \left\lceil\frac{(r-1)q^2}{r+q-2}\right\rceil =2r-O(\sqrt r). \tag{AV}

For every singleton direction, some old line must lose at least qq points. Otherwise every new vertex must send that color to an entire residual line. If line LL loses dL<qd_L<q points, it can be chosen by at most dLd_L new vertices: dL+1d_L+1 such vertices and its r−dLr-d_L surviving points form an (r+1)(r+1)-set seeing at most 1+(dL+12)<r1+\binom{d_L+1}{2}<r colors. There are k+1k+1 new vertices but ∑LdL=k\sum_Ld_L=k, a contradiction.

Choose qq deleted points on one heavy line in each of the r−1r-1 directions. If cpc_p counts their incidences at a deleted point, then

∑pcp=(r−1)q,∑pcp(cp−1)≤(r−1)(r−2).\sum_pc_p=(r-1)q,\qquad \sum_pc_p(c_p-1)\le(r-1)(r-2).

Cauchy–Schwarz proves (AV). Deleting a vertex cover of the recolored-old-edge graph puts any one-vertex affine repair in this situation, proving the advertised vertex-cover restriction.

Full equality arguments and a separate exact small-replacement obstruction were recorded in a companion note that is not retained. These are necessary conditions on affine repairs, not a closeness theorem for arbitrary colorings.

7. Failed approaches and explicit one-color constructions

7.1 Why a minority graph alone is insufficient

For every r=5ur=5u, u≥3u\ge3, consider

G=(r−2)Kr ∪˙ J,G=(r-2)K_r\ \dot\cup\ J,

where J‾\overline J is the independent-bag blow-up of C5C_5 with bag sizes

(2u+1,2u,2u,2u,2u).(2u+1,2u,2u,2u,2u).

Then

∣G∣=r2+1,α(G)=ω(G)=r,e(J)=30u2+u,|G|=r^2+1,\quad \alpha(G)=\omega(G)=r,\quad e(J)=30u^2+u,

and

r(r2+1)2−e(G)=15u2−7u2>0.\frac{r(r^2+1)}2-e(G)=\frac{15u^2-7u}{2}>0.

It satisfies (L). An (r+1)(r+1)-set meeting an ordinary component has at least rr cross-component nonedges. Within JJ, minimize

∑ℓ=04xℓxℓ+1,∑ℓxℓ=5u+1\sum_{\ell=0}^4x_\ell x_{\ell+1},\qquad \sum_\ell x_\ell=5u+1

under the bag capacities. Moving two non-boundary coordinates while keeping their sum makes the objective affine or concave, so a minimum has at most one non-boundary coordinate. Two full bags and a third partial bag then yield at least 2u2≥5u−12u^2\ge5u-1 nonedges.

Thus a hypothetical lemma saying “minority edge count plus (L) forces an independent (r+1)(r+1)-set” is false. This graph cannot actually be a color graph of a counterexample: its exceptional JJ violates the proved two-slot capacity.

A sharper family, recorded in a companion note that is not retained, works for every r≥27r\ge27. Put

L=⌈r1/3⌉,b=⌈r−1L⌉,c=2L−1,a=r+1−b−L,L=\lceil r^{1/3}\rceil,\quad b=\left\lceil\frac{r-1}{L}\right\rceil,\quad c=2L-1,\quad a=r+1-b-L,

take J‾=C5[a,a,b,c,b]\overline J=C_5[a,a,b,c,b], and again set G=(r−2)Kr∪˙JG=(r-2)K_r\dot\cup J. A weighted five-bag minimization shows every (r+1)(r+1)-set in JJ has at least bL≥r−1bL\ge r-1 nonedges. Writing q=L−1q=L-1,

e(G)=T+D,D=r+2rq−q2+b2−2b(2q+1)=O(r4/3)<(r2).e(G)=T+D,\qquad D=r+2rq-q^2+b^2-2b(2q+1)=O(r^{4/3})<\binom r2.

An elementary complementary lower bound in the same note shows the r4/3r^{4/3} order is sharp for this particular independence-two exceptional-block model. This still does not produce compatible color graphs.

7.2 Growing independence and alpha-criticality do not rescue that shortcut

A more elaborate individual graph also passes all capacity tests below a growing parameter. For every integer a≥2a\ge2, let

r=16a3,g=4a,t=ag+1,B=r−t−g.r=16a^3,\qquad g=4a,\qquad t=ag+1,\qquad B=r-t-g.

Let HH be the clique blow-up of C2a+1C_{2a+1} with aa alternating heavy bags of size BB and a+1a+1 light bags of size tt; the two extra consecutive bags are light. Set

G=(r−a)Kr ∪˙ H.G=(r-a)K_r\ \dot\cup\ H.

Then

∣H∣=ar+1,α(H)=a,α(G)=r.|H|=ar+1,\quad \alpha(H)=a,\quad \alpha(G)=r.

Every edge is alpha-critical: deleting an intra-bag edge permits two selected vertices there and an independent set of size a−1a-1 on the remaining path; deleting an inter-bag edge permits its two endpoints and an independent set of size a−1a-1 on the complementary path.

All adjacent bag sums are at most rr. Any set whose support omits a cycle bag has path components; pairing adjacent bags proves ∣S∣≤rα(G[S])|S|\le r\alpha(G[S]). Hence all capacity statements with independence number less than aa hold. In fact the overcapacity sets are exactly the whole HH together with zero or more complete ordinary KrK_r components.

The minimum number of nonedges on an (r+1)(r+1)-set in HH is

t(g+1)=(4a2+1)(4a+1)>r.t(g+1)=(4a^2+1)(4a+1)>r.

The proof again pushes coordinates to bag boundaries. At most one heavy bag can be full; if one is full, the remaining selected vertices force at least t(g+1)t(g+1) nonedges. If no heavy bag is full, a partial heavy bag and at least two full light bags force the same bound. Equality is attained by a heavy bag and portions of its two neighboring light bags.

Its edge surplus is

e(G)−T=(a−1)(64a5−16a4+8a3−8a2−1)<64a6=r24<(r2).e(G)-T =(a-1)\bigl(64a^5-16a^4+8a^3-8a^2-1\bigr) <64a^6=\frac{r^2}{4}<\binom r2.

Full proofs and an independent audit were recorded in companion notes that are not retained.

Important limitation: this example obstructs only an inference from minority sparsity, (L), alpha-criticality, and the smaller capacity statements. It does not satisfy all necessary conditions for a color graph. In particular its surplus inside HH is of order r2r^2, while (H), applied to ∣H∣=ar+1|H|=ar+1, permits only order a2ra^2r, which here is order r5/3r^{5/3}. The hereditary edge budget already rejects it. It is therefore not evidence of a counterexample to the original problem.

7.3 Other construction attempts

  • If every color graph is regular, one has degree at most rr; Brooks' theorem then yields an independent (r+1)(r+1)-set or a forbidden Kr+1K_{r+1}. This rules out color-preserving vertex-transitive/difference colorings, but not actions that also permute colors.
  • The asymmetric C5C_5 clique blow-up with bags (r−1,1,r−1,1,1)(r-1,1,r-1,1,1), attached to r−2r-2 ordinary KrK_r's, has attractive edge and independence counts. It fails (K): an outside singleton has r−1r-1 neighbors in a full KrK_r.
  • Putting every color's independence-two defect on two affine lines plus one common added vertex fails the two-slot capacity theorem. Merely rearranging those two-line supports is not enough.
  • Lexicographic and strong-product constructions did not give a valid independence bound with the required number of colors. No product construction was promoted to a theorem.
  • A stronger geometric statement forcing all color graphs to be near affine clique factors was not proved.
  • The proposed shortcut “small total edge surplus forces a minimal overcapacity component of bounded independence” was not proved. The preceding individual-graph examples show why sparsity and local density alone are insufficient; (H) must also be retained.

8. Uncertainties and exact endpoint

The following are established in the record:

  1. The universal capacity theorem for r≥256a6r\ge256a^6, and the sharper a=2,3a=2,3 statements.
  2. The minimal-obstruction description (M).
  3. The checked clique-packing restriction p≤r−2p\le r-2, with a separately recorded, locally checked extension p≤r−3p\le r-3 for r≥8r\ge8.
  4. Quantitative obstructions to affine repairs.
  5. Explicit individual graphs refuting weaker one-color proof strategies.

What is not established:

  • A bound a≤Ara\le A_r for a minimal obstruction in (M).
  • Capacity for all a≤ra\le r.
  • A sufficiently large clique packing in any color.
  • A theorem placing every hypothetical coloring close to an affine construction.
  • An edge-disjoint packing of the proposed individual graphs.
  • A counterexample coloring.

A further proposed improvement concerning arbitrary odd-cycle clique blow-ups and a weighted-cycle inequality is not recorded in full. It is not needed for any proved result above and is a provisional calculation until its precise lemma and proof are checked. No conclusion of this investigation relies on it.

No substantial computation or numerical search was used as proof. The mathematical claims above rest on the displayed counting, extremal, and transversal arguments; the independent checks were proof audits, not certificates of the original conjecture.

The quantified assertion of Erdős Problem 617 remains unresolved in this independent record.