Wiki
Wiki

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

Updated

Large clique clusters


Status. The following is a proved structural consequence of the hereditary density inequalities and the localized independence inequality recorded in Packing tools. It does not resolve Problem 617 or produce a packing of full KrK_r's.

Let rr be an integer at least three, and put

N=r2+1,pr(rq+b)=r(q2)+qb(0≤b<r),Dr(n)=(n2)−(r−1)pr(n).N=r^2+1,\qquad p_r(rq+b)=r\binom q2+qb\quad(0\le b<r), \qquad D_r(n)=\binom n2-(r-1)p_r(n).

Suppose that a graph GG on NN vertices satisfies

α(G)≤r,2e(G)≤rN,e(G[W])≤Dr(∣W∣)(W⊆V(G)).(1)\alpha(G)\le r,\qquad 2e(G)\le rN, \qquad e(G[W])\le D_r(|W|)\quad(W\subseteq V(G)). \tag{1}

In particular, a least color in a hypothetical counterexample satisfies (1).

Write c(v)c(v) for the largest order of a clique containing vv. The (r+1)(r+1)-vertex case of (1) gives c(v)≤rc(v)\le r for r≥3r\ge3. The localized independence inequality, in the uniform-vector form already checked in the research notes, gives

2e(G)+N+∑vc(v)≥2N2r.(2)2e(G)+N+\sum_v c(v)\ge\frac{2N^2}{r}. \tag{2}

Consequently, putting D=∑v(r−c(v))D=\sum_v(r-c(v)), we have

D≤N(1−2r),∑v(d(v)−c(v)+1)≤(2−2r)N.(3)D\le N\left(1-\frac2r\right), \qquad \sum_v\bigl(d(v)-c(v)+1\bigr) \le \left(2-\frac2r\right)N. \tag{3}

The second inequality follows by combining (2) with 4e(G)≤2rN4e(G)\le2rN. Each summand in that inequality is nonnegative.

1. The clustering lemma

Fix an integer k≥1k\ge1 such that

r>(k+1)2+1,r−3k≥2,r−1>32k(k+1).(4)r>(k+1)^2+1,\qquad r-3k\ge2,\qquad r-1>\frac32k(k+1). \tag{4}

Let Q\mathcal Q consist of all cliques of order at least r−kr-k. Join two such cliques when their intersection has at least two vertices, and let C1,…,CqC_1,\ldots,C_q be the unions of the connected components of this clique-intersection graph. Then

  1. ∣Cj∣≤r|C_j|\le r for every jj;
  2. cliques belonging to distinct components intersect in at most one vertex;
  3. if qv=∣{j:v∈Cj}∣q_v=|\{j:v\in C_j\}| and
L=∑vmax⁡(qv−1,0),L=\sum_v\max(q_v-1,0),

then

L≤(2−2/r)Nr−k−1;(5)L\le\frac{(2-2/r)N}{r-k-1}; \tag{5}
  1. the uncovered set U=V(G)∖⋃jCjU=V(G)\setminus\bigcup_j C_j satisfies
∣U∣≤Dk+1.(6)|U|\le\frac{D}{k+1}. \tag{6}

Pairwise intersection and transitivity

Take cliques Q,RQ,R of orders r−t,r−ur-t,r-u, where 0≤t,u≤k0\le t,u\le k, and put h=∣Q∩R∣h=|Q\cap R|. Suppose that h≥2h\ge2 and ∣Q∪R∣≥r+1|Q\cup R|\ge r+1. Its order is

n=2r−t−u−h∈[r+1,2r].n=2r-t-u-h\in[r+1,2r].

All nonedges of G[Q∪R]G[Q\cup R] are between Q∖RQ\setminus R and R∖QR\setminus Q. Therefore the inherited density inequality requires

(r−t−h)(r−u−h)≥(r−1)(r−t−u−h).(r-t-h)(r-u-h)\ge(r-1)(r-t-u-h).

Equivalently,

f(h):=h2+h(t+u−r−1)+tu−t−u+r≥0.(7)f(h):=h^2+h(t+u-r-1)+tu-t-u+r\ge0. \tag{7}

The allowed interval is 2≤h≤r−t−u−12\le h\le r-t-u-1. At both endpoints,

f(2)=f(r−t−u−1)=tu+t+u+2−r≤(k+1)2+1−r<0.f(2)=f(r-t-u-1)=tu+t+u+2-r \le(k+1)^2+1-r<0.

Since ff is convex, (7) is impossible throughout the interval. Thus

∣Q∩R∣≥2⟹∣Q∪R∣≤r.(8)|Q\cap R|\ge2\quad\Longrightarrow\quad |Q\cup R|\le r. \tag{8}

If Q,R,S∈QQ,R,S\in\mathcal Q and both ∣Q∩R∣|Q\cap R| and ∣R∩S∣|R\cap S| are at least two, then (8) gives

∣Q∩S∣≥∣Q∩R∣+∣R∩S∣−∣R∣≥∣Q∣+∣R∣+∣S∣−2r≥r−3k≥2.\begin{aligned} |Q\cap S| &\ge |Q\cap R|+|R\cap S|-|R|\\ &\ge |Q|+|R|+|S|-2r\ge r-3k\ge2. \end{aligned}

Consequently the intersection-at-least-two relation is transitive (with its evident reflexive interpretation). Every two cliques in one connected component intersect in at least two vertices, and (8) applies to them.

Each cluster union has at most rr vertices

Fix an anchor clique QQ in one component, with ∣Q∣=r−t|Q|=r-t, 0≤t≤k0\le t\le k. For any vertex xx of its cluster union outside QQ, choose a clique RR of that component containing xx. The preceding transitivity argument and (8) give

∣Q∩R∣≥∣Q∣+∣R∣−r≥∣Q∣−k.|Q\cap R|\ge |Q|+|R|-r\ge |Q|-k.

The vertex xx is adjacent to every vertex of Q∩RQ\cap R, so it has at most kk nonneighbors in QQ.

If the cluster union had at least r+1r+1 vertices, choose t+1t+1 of these outside vertices. Their union with the entire anchor QQ has exactly r+1r+1 vertices. Its number of nonedges is at most

k(t+1)+(t+12)≤32k(k+1)<r−1.k(t+1)+\binom{t+1}{2} \le\frac32k(k+1)<r-1.

This contradicts Dr(r+1)=(r+12)−(r−1)D_r(r+1)=\binom{r+1}{2}-(r-1), proving ∣Cj∣≤r|C_j|\le r. Notice that this argument follows the transitivity proof; no bound on a cluster union was used in proving transitivity.

Overlaps and uncovered vertices

For a covered vertex vv, a maximum clique through vv belongs to one of its qvq_v clusters. Choose that clique, and choose one clique through vv from each of its other clusters. Any two chosen cliques intersect only in vv: an intersection of size at least two would put them in the same cluster. Hence their sets of neighbors of vv are disjoint and

d(v)≥c(v)−1+(qv−1)(r−k−1).d(v)\ge c(v)-1+(q_v-1)(r-k-1).

For an uncovered vertex, d(v)≥c(v)−1d(v)\ge c(v)-1 still holds. Summing and using (3) proves (5). An uncovered vertex has c(v)≤r−k−1c(v)\le r-k-1, hence r−c(v)≥k+1r-c(v)\ge k+1; summing proves (6).

2. Number and total deficits of maximum anchors

In each cluster choose a clique QjQ_j of maximum order among all cliques in that cluster, and write

∣Qj∣=r−tj,0≤tj≤k,T=∑j=1qtj.|Q_j|=r-t_j,\qquad 0\le t_j\le k, \qquad T=\sum_{j=1}^q t_j.

Then

∑j(r−tj)tj≤D+kL,(r−k)T≤D+kL.(9)\sum_j(r-t_j)t_j\le D+kL, \qquad (r-k)T\le D+kL. \tag{9}

Here is a vertexwise check of the first inequality, which avoids assuming that a globally maximum clique through a vertex is its fixed anchor. Put δv=r−c(v)\delta_v=r-c(v). If vv belongs to at least one anchor, choose a maximum clique through vv, and call its cluster j0j_0.

If v∈Qj0v\in Q_{j_0}, then ∣Qj0∣=c(v)|Q_{j_0}|=c(v): maximum order within its cluster gives ∣Qj0∣≥c(v)|Q_{j_0}|\ge c(v), and containment of vv gives the reverse inequality. Thus tj0=δvt_{j_0}=\delta_v, and all other anchors through vv contribute at most kk each.

If v∉Qj0v\notin Q_{j_0}, at most qv−1q_v-1 anchors contain vv, and their total contribution is at most k(qv−1)k(q_v-1). In either case,

∑j:v∈Qjtj≤δv+kmax⁡(qv−1,0).\sum_{j:v\in Q_j}t_j\le \delta_v+k\max(q_v-1,0).

For a vertex belonging to no anchor, the same inequality is immediate. Summing proves (9).

The maximum anchors from different clusters have edge-disjoint edge sets, so

∑j(r−tj2)≤e(G).\sum_j\binom{r-t_j}{2}\le e(G).

Expanding gives the useful bound

qr(r−1)≤rN+(2r−1)T−∑jtj2≤rN+(2r−1)T.(10)q r(r-1) \le rN+(2r-1)T-\sum_j t_j^2 \le rN+(2r-1)T. \tag{10}

3. Explicit square-root-scale form

Assume r≥25r\ge25 and take

k=⌊r2⌋.k=\left\lfloor\frac{\sqrt r}{2}\right\rfloor.

All three inequalities in (4) hold. Then

∣U∣<2r3/2,L≤2r+2k+3,T≤r+3k,(11)|U|<2r^{3/2},\qquad L\le2r+2k+3,\qquad T\le r+3k, \tag{11}

and

r−2r<q≤r+3.(12)r-2\sqrt r<q\le r+3. \tag{12}

For the uncovered bound, use D<r2D<r^2 and k+1>r/2k+1>\sqrt r/2. For the overlap bound, (5) equals

L≤2r+2rk+2−2/rr−k−1.L\le 2r+\frac{2rk+2-2/r}{r-k-1}.

The final fraction is at most 2k+32k+3, because

2k2+5k+5≤3r2k^2+5k+5\le3r

for r≥25r\ge25 and k≤r/2k\le\sqrt r/2.

Also D≤r2−2r+1D\le r^2-2r+1. Applying (9) and the overlap bound,

D+kL≤r2−2r+1+2kr+2k2+3k≤(r−k)(r+3k),\begin{aligned} D+kL &\le r^2-2r+1+2kr+2k^2+3k\\ &\le (r-k)(r+3k), \end{aligned}

where the second inequality is 5k2+3k+1≤2r5k^2+3k+1\le2r, valid in the stated range. This proves the bound on TT.

The lower bound in (12) follows from qr≥N−∣U∣qr\ge N-|U|. For the upper bound, (10) gives

q≤r+3+3r+(6r−3)kr(r−1)<r+4.q\le r+3+ \frac{3r+(6r-3)k}{r(r-1)}<r+4.

The strict final inequality follows from r>6k+4r>6k+4, valid for r≥25r\ge25. Since qq is an integer, q≤r+3q\le r+3.

4. A disjoint packing of near-full cliques

Assign every vertex lying in several anchors to just one of those anchors and delete it from the others. The total number of deletions, counted with multiplicity, is at most LL. Delete completely every resulting anchor that lost more than kk vertices. At most L/(k+1)L/(k+1) anchors are discarded, and every retained anchor is a clique of order at least r−2k≥r−rr-2k\ge r-\sqrt r.

For r≥25r\ge25 one can use

L≤2r+r+2.(13)L\le2r+\sqrt r+2. \tag{13}

Indeed, the preceding exact bound for LL and k≤r/2k\le\sqrt r/2 show that (13) follows from

rr+2r−r/2−1≤r+2,\frac{r\sqrt r+2}{r-\sqrt r/2-1}\le\sqrt r+2,

which reduces to 32r−2r−4≥0\tfrac32r-2\sqrt r-4\ge0. Therefore

Lk+1<4r+2+4r≤4r+3.\frac{L}{k+1}<4\sqrt r+2+\frac4{\sqrt r} \le4\sqrt r+3.

Combining this with (12), the number of retained pairwise disjoint cliques is greater than

r−6r−3.r-6\sqrt r-3.

Thus, in particular, there are at least

max⁡{0,⌊r−6r−3⌋}\max\{0,\lfloor r-6\sqrt r-3\rfloor\}

pairwise disjoint cliques of order at least r−rr-\sqrt r. They cover all but O(r3/2)O(r^{3/2}) vertices.

Limitation

This gives actual large cliques and exact upper capacity rr for every overlap cluster, with only O(r)O(r) excess incidences. It does not say that the clusters themselves are complete, or that all vertices are covered. In particular, the packing consists of cliques of order r−O(r)r-O(\sqrt r), not full KrK_r's. The inherited span theorem does not presently apply to the remaining square-root-scale independence range.

5. Balanced disjointification

All qq maximum anchors can be retained, instead of discarding some of them. Put ℓ=⌈L/2⌉\ell=\lceil\sqrt{L/2}\rceil. There are pairwise disjoint subcliques Qj′⊆QjQ'_j\subseteq Q_j with

∣Qj′∣≥∣Qj∣−ℓ.|Q'_j|\ge |Q_j|-\ell.

For a family of ss anchors its overlap loss is at most min⁡{L,(s2)}≤sL/2\min\{L,\binom s2\}\le s\sqrt{L/2}. The first bound uses global excess incidences; the second uses the fact that two distinct anchors meet in at most one point. Thus every family has enough distinct vertices to meet the demands ∣Qj∣−ℓ|Q_j|-\ell, and the capacitated form of Hall's theorem applies. Negative demands, if any, are replaced by zero. For the square-root choice of kk, every retained block has order r−(3/2+o(1))rr-(3/2+o(1))\sqrt r, and their number remains greater than r−2rr-2\sqrt r. They are still not full cliques.

6. A sharper cluster-deficit tradeoff

Write ∣Cj∣=r−bj|C_j|=r-b_j, so 0≤bj≤tj0\le b_j\le t_j, and put a=r−qa=r-q. Then

T+(k+1)(r−q)≤r−3.(14)\boxed{T+(k+1)(r-q)\le r-3.} \tag{14}

For a covered vertex vv, choose the cluster of a maximum clique through vv. Its anchor deficit is at most δv=r−c(v)\delta_v=r-c(v); hence

∑j:v∈Cjtj≤δv+k(qv−1).\sum_{j:v\in C_j}t_j\le\delta_v+k(q_v-1).

The uncovered set has order ar+∑jbj+L+1ar+\sum_jb_j+L+1, and every one of its vertices has deficit at least k+1k+1. Summing gives

D≥r[T+(k+1)a]+∑jbj(k+1−tj)+L+k+1.D\ge r[T+(k+1)a]+\sum_j b_j(k+1-t_j)+L+k+1.

Use D≤r2−2r+1−2/rD\le r^2-2r+1-2/r and integrality to obtain (14). When q≤rq\le r, the number of full anchors is therefore at least

q−T≥k(r−q)+3.q-T\ge k(r-q)+3.

This is a dichotomy between missing clusters and full anchors, not a proof that the near-full anchors complete.

7. The upper cluster count is r+2r+2 for large rr

For r≥625r\ge625, one can sharpen (12) to q≤r+2q\le r+2. If q=r+3q=r+3, the edge budget outside all anchor cliques is at most

B≤3r+(6r−3)k2≤32r(1+r).B\le\frac{3r+(6r-3)k}{2} \le\frac32r(1+\sqrt r).

Let hjh_j count vertices of QjQ_j shared with another anchor. The three largest hjh_j have sum at most L+3L+3: the excess over their contribution to LL consists only of vertices shared entirely within those three anchors, and there are at most three such vertices. Discard the two anchors of largest hjh_j. In each of the remaining r+1r+1 anchors, the vertices private to the full anchor family number at least r−k−(L+3)/3≥r/4r-k-(L+3)/3\ge r/4 for r≥144r\ge144.

All edges between these private parts lie outside the anchor cliques. Independent uniform representatives have expected conflict count at most 16B/r2≤24(1+r)/r<116B/r^2\le24(1+\sqrt r)/r<1 for r≥625r\ge625. An independent transversal contradicts α(G)≤r\alpha(G)\le r.

8. Joint deficit and degree-variance accounting

Proved. For r≥16384r\ge16384, the cluster count in Section 3 satisfies

q≤r+1.(15)\boxed{q\le r+1.} \tag{15}

This improves the upper cluster count, not the unrestricted coloring assertion. The case q=r+1q=r+1 and the completion of clusters remain open.

Put Av=d(v)+c(v)+1A_v=d(v)+c(v)+1, μ=2N/r\mu=2N/r, and

V=∑v(Av−μ)2Av.V=\sum_v\frac{(A_v-\mu)^2}{A_v}.

The localized independence inequality gives ∑v2/Av≤r\sum_v2/A_v\le r. Sum the exact identity

1A=1μ−A−μμ2+(A−μ)2μ2A\frac1A=\frac1\mu-\frac{A-\mu}{\mu^2} +\frac{(A-\mu)^2}{\mu^2 A}

and use N/μ=r/2N/\mu=r/2. It follows that V≤∑vAv−NμV\le\sum_v A_v-N\mu. Since ∑vAv=2e+(r+1)N−D\sum_v A_v=2e+(r+1)N-D, the edge budget gives

D+V≤(1−2/r)N<r2.(16)\boxed{D+V\le(1-2/r)N<r^2.} \tag{16}

Let hv=max⁡(qv−1,0)h_v=\max(q_v-1,0) and keep k=⌊r/2⌋k=\lfloor\sqrt r/2\rfloor. Clique geometry gives, when hv≥1h_v\ge1,

Av≥2(r−k)+hv(r−k−1)≥μ+hvz,z=r−3k−2>0.A_v\ge2(r-k)+h_v(r-k-1) \ge\mu+h_v z,\qquad z=r-3k-2>0.

The function (A−μ)2/A(A-\mu)^2/A increases for A≥μA\ge\mu. Hence

(Av−μ)2Av≥hv2z2μ+hvz≥hvz2μ+z≥hv(r/3−2k−2).\frac{(A_v-\mu)^2}{A_v} \ge\frac{h_v^2z^2}{\mu+h_vz} \ge\frac{h_vz^2}{\mu+z} \ge h_v(r/3-2k-2).

Here μ+z≤3r\mu+z\le3r, and expanding z2/(3r)z^2/(3r) gives the last inequality. Combining this with (9) and (16) yields

(r−k)T+γL<r2,γ=r/3−3k−2.(17)(r-k)T+\gamma L<r^2, \qquad \gamma=r/3-3k-2. \tag{17}

Section 7 already gives q≤r+2q\le r+2, so assume q=r+2q=r+2. Counting anchor incidences gives L≥2r−T−1L\ge2r-T-1, while (10) gives T≥(r−3)/2T\ge(r-3)/2. Substitute the first in (17) to obtain

(2r/3+2k+2)T≤r2/3+6kr+(13/3)r−3k−2.(2r/3+2k+2)T \le r^2/3+6kr+(13/3)r-3k-2.

Consequently

T≤r/2+(15/2)k+5.T\le r/2+(15/2)k+5.

Also γ≥r/4\gamma\ge r/4 for r≥625r\ge625. Substituting T≥(r−3)/2T\ge(r-3)/2 into (17) gives

L≤32r+(5k+9/2)r−3k/2γ≤32r+20k+18.(18)L\le\frac32r+\frac{(5k+9/2)r-3k/2}{\gamma} \le\frac32r+20k+18. \tag{18}

Let BB count edges outside the union of the anchor edge sets. Those edge sets are disjoint. The edge budget and the bound on TT give

2B≤−r2+3r+(2r−1)T−∑jtj2≤15rk+(25/2)r,B≤8r(k+1).(19)2B\le-r^2+3r+(2r-1)T-\sum_jt_j^2 \le15rk+(25/2)r, \qquad B\le8r(k+1). \tag{19}

Let hjh_j count vertices of QjQ_j that belong to another anchor. For any two anchors hi+hj≤L+1h_i+h_j\le L+1. Indeed each shared vertex contributes at most its cluster multiplicity minus one, except a vertex lying in both chosen anchors and in no other cluster; there is at most one such vertex. Drop an anchor with largest hjh_j. In each of the remaining r+1r+1 anchors, retain only vertices private to the entire anchor family. Each retained part has order at least

r−k−(L+1)/2≥r/4−11k−10≥r/5r-k-(L+1)/2\ge r/4-11k-10\ge r/5

for r≥16384r\ge16384. Their cross edges are all among the BB edges in (19). Uniform independent representatives have expected conflict count at most

25B/r2≤200(k+1)/r≤100/r+200/r<1.25B/r^2\le200(k+1)/r \le100/\sqrt r+200/r<1.

They therefore have an independent transversal, contrary to α(G)≤r\alpha(G)\le r. This proves (15).

The private-vertex restriction is essential here. Balanced disjointification alone does not ensure that all cross edges lie outside the original anchors: an edge of a third anchor may join vertices assigned to two different retained anchors.

9. Two thresholds in a minimal-ratio core

Proved refinement, with an unresolved extension step. Here work in the minimal-ratio core, not the original NN-vertex graph. Write

∣H∣=ar+1,h=r−a,S=∑v(dH(v)−r)≤rh,D=∑v(r−cH(v))≤r2−2r+1−2/a.|H|=ar+1,\quad h=r-a,\quad S=\sum_v(d_H(v)-r)\le rh,\quad D=\sum_v(r-c_H(v))\le r^2-2r+1-2/a.

Choose 1≤k<K1\le k<K, where KK satisfies the three clustering hypotheses (4), and assume r>2K+2r>2K+2. Form clusters from all cliques of size at least r−Kr-K, but retain only clusters whose largest clique has size at least r−kr-k. For their maximum anchors and cluster sizes put

∣Qj∣=r−tj,∣Cj∣=r−bj,1≤tj≤k,0≤bj≤tj.|Q_j|=r-t_j,\quad |C_j|=r-b_j,\quad 1\le t_j\le k,\quad0\le b_j\le t_j.

Let qq count retained clusters, T=∑jtjT=\sum_jt_j, and let LL be their excess vertex incidences. Then

L≤L0:=rhr−2K−2,(20)\boxed{L\le L_0:=\frac{rh}{r-2K-2},} \tag{20} T+(k+1)(a−q)≤r−3,a−q≤h−3k,(21)\boxed{T+(k+1)(a-q)\le r-3,\qquad a-q\le\frac{h-3}{k},} \tag{21} (r−k)T≤D+kL,q(r−k)≤ar+1+L.(22)\boxed{(r-k)T\le D+kL,\qquad q(r-k)\le ar+1+L.} \tag{22}

A vertex in m≥2m\ge2 retained clusters has mm cliques through it, each of size at least r−Kr-K, intersecting pairwise only at that vertex. Thus

dH(v)−r≥m(r−K−1)−r≥(m−1)(r−2K−2).d_H(v)-r\ge m(r-K-1)-r\ge(m-1)(r-2K-2).

Summing proves (20). Every uncovered vertex has clique deficit at least k+1k+1. At a covered vertex the smallest retained anchor deficit is at most its local clique deficit: if its maximum clique belongs to a retained cluster this is immediate; otherwise its deficit is at least k+1k+1. The same vertexwise counting as before therefore gives

D≥r[T+(k+1)(a−q)]+∑jbj(k+1−tj)+L+k+1.D\ge r[T+(k+1)(a-q)] +\sum_jb_j(k+1-t_j)+L+k+1.

The bound on DD and integrality prove the first inequality in (21). Since T≥qT\ge q, the second follows. The anchor-incidence arguments from Section 2 give (22).

Let PjP_j be the vertices of QjQ_j belonging to no other retained cluster. Then these parts are disjoint and

∣Pj∣≥r−k−L0.|P_j|\ge r-k-L_0.

For a vertex x∉Cjx\notin C_j, the broader cluster threshold gives

dH(x,Qj)≤r−K−2,∣Qj∖N(x)∣≥K+2−tj.(23)\boxed{d_H(x,Q_j)\le r-K-2,\qquad |Q_j\setminus N(x)|\ge K+2-t_j.} \tag{23}

Otherwise xx and its anchor neighbors form a clique of order at least r−Kr-K meeting QjQ_j in at least two vertices, putting xx in CjC_j.

If FF counts edges between distinct PjP_j, then

2F≤r(T+q)+S.(24)\boxed{2F\le r(T+q)+S.} \tag{24}

A private-part vertex in QjQ_j has tj+1+s(v)t_j+1+s(v) neighbors outside that anchor, where s(v)=dH(v)−r≥0s(v)=d_H(v)-r\ge0. Summing over the disjoint private parts proves (24). For fixed k,Kk,K these formulas imply T≤r+O(k)T\le r+O(k), q≤a+O(k)q\le a+O(k), and F≤r2+O(kr)F\le r^2+O(kr).

There is an exact way to discard large overlap losses. If hjh_j is the number of nonprivate vertices in QjQ_j, then ∑jhj≤2L\sum_jh_j\le2L: a vertex of cluster multiplicity m≥2m\ge2 contributes at most m≤2(m−1)m\le2(m-1). Discarding the at most 2L/(ηr)2L/(\eta r) anchors with hj>ηrh_j>\eta r leaves parts of size at least (1−η)r−k(1-\eta)r-k. For fixed small η>0\eta>0, their normalized cross-edge weight is at most 1+O(η+k/r)1+O(\eta+k/r). This is near the independent-transversal union-bound threshold, not strictly below it.

Why this does not justify suppression

The linear independent-outsider bound in Saturated-row suppression requires a nearly complete set of order rr. For a mere clique QQ of order r−tr-t and an independent outside set JJ of order t≤s≤r+tt\le s\le r+t, the inherited cap only gives

e(Q,J)≤t(r−s−1)+(s+12).(25)e(Q,J)\le t(r-s-1)+\binom{s+1}{2}. \tag{25}

This follows by subtracting (r−t2)\binom{r-t}{2} from Dr(r−t+s)=(r2)+(s−t+12)D_r(r-t+s)=\binom r2+\binom{s-t+1}{2}. The term of order trtr cannot be discarded.

In particular, when t≥2t\ge2, two independent outsiders can partition and cover QQ by their neighborhoods on a graph of order at most rr. Every hereditary cap is then vacuous, even if each outsider misses a linear number of clique vertices. Also the missing-neighbor margin in (23) concerns the full anchor, not its private part; removing shared vertices can remove all those missing neighbors.

No argument here proves q≤aq\le a, or proves that an independent (a−q+1)(a-q+1)-set in a residual extends across the retained clusters. The refinement separates useful counting and adjacency thresholds, but a quantitative transversal stability or a valid completion and extension argument remains missing.