Wiki
Wiki

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

Updated

Two supported private cliques


Proved reductions, not a resolution. A hypothetical balanced coloring has a color with a maximum independent set of order rr, disjoint private cliques of total deficit at most r−1r-1, and at most rr exceptional vertices, each adjacent to exactly two members of that independent set. The selected color need not be a least color. The reductions below force a private deficit of order at least r\sqrt r, but do not exclude this concentration.

After full-row deletion, the minimum-anchor normalization also removes hidden KrK_r's while retaining two-support structure, and ensures degree at least rr at every private vertex. It does not give the least-color budget or minimum degree on the exceptions.

Throughout, r≥4r\ge4, N=r2+1N=r^2+1, and the color graphs G1,…,GrG_1,\ldots,G_r partition E(KN)E(K_N) and satisfy α(Gj)≤r\alpha(G_j)\le r. We use the inherited density bound

e(Gj[W])≤Dr(∣W∣),Dr(n)=(n2)−(r−1)pr(n),e(G_j[W])\le D_r(|W|),\qquad D_r(n)=\binom n2-(r-1)p_r(n),

where pr(rq+b)=r(q2)+qbp_r(rq+b)=r\binom q2+qb for 0≤b<r0\le b<r. In particular, every (r+1)(r+1)-set spans at most (r2)+1\binom r2+1 edges of one color. There is no monochromatic Kr+1K_{r+1}, and an outside vertex has at most one neighbor in a monochromatic KrK_r in its color.

1. Selecting a two-supported private-clique system

Choose a color-ii clique QQ of order rr. Such a clique exists by the localized independence inequality: a least color graph has average degree at most rr, and clique number at most r−1r-1 would give

α(Gi)≥2Nd‾(Gi)+ω(Gi)+1≥Nr>r.\alpha(G_i)\ge \frac{2N}{\overline d(G_i)+\omega(G_i)+1} \ge\frac Nr>r.

This checked independence inequality is the only non-elementary input in the selection of QQ.

For every j≠ij\ne i, the set QQ is a maximum independent set in GjG_j. Every outside vertex has at least one color-jj neighbor in QQ; otherwise it augments QQ. Consequently, among its rr edges to QQ, either color ii occurs once and all other colors occur once, or color ii is absent and exactly one other color occurs twice.

Let BjB_j be the outside vertices having two color-jj neighbors in QQ, and write bj=∣Bj∣b_j=|B_j|. Then

∑j≠ibj=N−r−ei(Q,V∖Q)≤r2−r+1.(1)\sum_{j\ne i}b_j=N-r-e_i(Q,V\setminus Q)\le r^2-r+1. \tag{1}

For q∈Qq\in Q, put

Pj,q={q}∪{x∉Q:Nj(x)∩Q={q}}.P_{j,q}=\{q\}\cup \{x\notin Q:N_j(x)\cap Q=\{q\}\}.

Each Pj,qP_{j,q} is a clique: two nonadjacent private vertices could replace qq in QQ. The private cliques are disjoint and have order at most rr. Writing tj,q=r−∣Pj,q∣t_{j,q}=r-|P_{j,q}|, vertex counting gives

∑q∈Qtj,q=bj−1.(2)\sum_{q\in Q}t_{j,q}=b_j-1. \tag{2}

In particular bj≥1b_j\ge1. Since r≥3r\ge3, (1) ensures that some j≠ij\ne i has bj≤rb_j\le r: otherwise the left side is at least (r−1)(r+1)>r2−r+1(r-1)(r+1)>r^2-r+1.

Fix such a color and suppress jj. We obtain rr private cliques PqP_q, with

∣Pq∣=r−tq,T:=∑qtq=b−1≤r−1,B:=V∖⋃qPq,∣B∣=b.(3)|P_q|=r-t_q,\qquad T:=\sum_qt_q=b-1\le r-1, \qquad B:=V\setminus\bigcup_qP_q,\quad |B|=b. \tag{3}

Each x∈Bx\in B has a two-element support Kx=N(x)∩QK_x=N(x)\cap Q. Thus the supports form a multigraph on QQ, with one edge for each exception.

2. The exact two-support inequality

Fix an exception xx supported on {u,v}\{u,v\}, and put

Au=Pu∖N(x),Av=Pv∖N(x),mu=∣Au∣,mv=∣Av∣.A_u=P_u\setminus N(x),\qquad A_v=P_v\setminus N(x), \qquad m_u=|A_u|,\quad m_v=|A_v|.

The sets Au,AvA_u,A_v are complete to each other. Otherwise xx, two nonadjacent representatives, and Q∖{u,v}Q\setminus\{u,v\} are an independent (r+1)(r+1)-set. Therefore, with cuv=e(Pu,Pv)c_{uv}=e(P_u,P_v),

cuv≥mumv.(4)c_{uv}\ge m_um_v. \tag{4}

If tu+tv≥1t_u+t_v\ge1, then Pu∪Pv∪{x}P_u\cup P_v\cup\{x\} has between r+2r+2 and 2r2r vertices. Substituting its order in DrD_r and subtracting its internal clique edges and the edges incident with xx yields

cuv≤mu+mv+(tu−1)(tv−1).c_{uv}\le m_u+m_v+(t_u-1)(t_v-1).

Together with (4), this gives

(mu−1)(mv−1)≤(tu−1)(tv−1)+1.(5)(m_u-1)(m_v-1)\le(t_u-1)(t_v-1)+1. \tag{5}

Suppose tu,tv≥1t_u,t_v\ge1 and tu+tv≥3t_u+t_v\ge3. If both mu>tum_u>t_u and mv>tvm_v>t_v, the left side of (5) is at least tutvt_ut_v, while its right side is at most tutv−1t_ut_v-1. Thus xx misses at most tqt_q vertices in at least one supporting private clique. When tu=tv=1t_u=t_v=1, either some mq≤1m_q\le1, or mu=mv=2m_u=m_v=2.

If one supporting deficit is zero, Section 4 proves that xx is complete to its other supporting private clique, whose deficit is positive. Consequently every exception has an owner q∈Kxq\in K_x such that

tq≥1,∣Pq∖N(x)∣≤max⁡(tq,2).(6)t_q\ge1, \qquad |P_q\setminus N(x)|\le\max(t_q,2). \tag{6}

3. A necessary large private deficit

Assign each exception one owner satisfying (6). There are T+1T+1 exceptions and total capacity ∑qtq=T\sum_qt_q=T, so some qq owns at least tq+1t_q+1 exceptions. Adjoin any tq+1t_q+1 of them to the entire clique PqP_q. The resulting (r+1)(r+1)-set has at most

(tq+1)max⁡(tq,2)+(tq+12)(t_q+1)\max(t_q,2)+\binom{t_q+1}{2}

missing target edges. The local cap requires at least r−1r-1 missing edges, so

r−1≤(tq+1)max⁡(tq,2)+(tq+12).(7)r-1\le(t_q+1)\max(t_q,2)+\binom{t_q+1}{2}. \tag{7}

For r≥7r\ge7, one cannot have tq=1t_q=1, since the right side would be 55. Hence tq≥2t_q\ge2 and

tq(tq+1)≥2(r−1)3.(8)t_q(t_q+1)\ge\frac{2(r-1)}3. \tag{8}

Thus the near-full hypothesis ∣Pq∣∈{r−1,r}|P_q|\in\{r-1,r\} cannot hold for this two-supported system when r≥7r\ge7. The reduction instead forces a concentrated deficit of order at least r\sqrt r. There is no established upper bound excluding that concentration.

4. Exact clique absorption at full private rows

Suppose an exception xx has support {u,v}\{u,v\} and tv=0t_v=0. Its sole neighbor in the full clique PvP_v is vv. If some y∈Puy\in P_u were nonadjacent to xx, (4) would make yy adjacent to all r−1r-1 vertices of Pv∖{v}P_v\setminus\{v\}, contradicting the outside-neighbor bound for PvP_v. Thus xx is complete to PuP_u. Necessarily tu>0t_u>0, since xx cannot be complete to another KrK_r. In particular no support edge joins two zero-deficit anchors.

For each positive-deficit anchor uu, let HuH_u consist of exceptions whose support is {u,v}\{u,v\} for some zero-deficit anchor vv. The set HuH_u is a clique. Indeed, suppose x,y∈Hux,y\in H_u are nonadjacent, with other supporting anchors v,wv,w. If v=wv=w, choose any z∈Pv∖{v}z\in P_v\setminus\{v\}. Then

{x,y,z}∪(Q∖{u,v})\{x,y,z\}\cup(Q\setminus\{u,v\})

is an independent (r+1)(r+1)-set. If v≠wv\ne w, each list

Pv∖(N(x)∪N(y)),Pw∖(N(x)∪N(y))P_v\setminus(N(x)\cup N(y)),\qquad P_w\setminus(N(x)\cup N(y))

has at least r−2≥2r-2\ge2 vertices. The edges between the two full cliques form a matching, so there are nonadjacent representatives zv,zwz_v,z_w from these lists. They, x,yx,y, and Q∖{u,v,w}Q\setminus\{u,v,w\} again give an independent (r+1)(r+1)-set. The representatives are private vertices, so they are anticomplete to the remaining anchors.

Every member of HuH_u is complete to PuP_u. Consequently

Pu∪Hu is a clique,∣Hu∣≤tu.(9)P_u\cup H_u\text{ is a clique},\qquad |H_u|\le t_u. \tag{9}

These enlarged cliques are disjoint. Delete the original full rows and absorb each HuH_u into its owner row. The absorbed vertices are private relative to the remaining anchors, and each unabsorbed exception still has a two-element support among those anchors.

Ordinary greedy extension does not alone prove that deleting kk full rows reduces independence to r−kr-k: it only extends a set II when ∣I∣+k≤r|I|+k\le r. Absorbed exceptions also retain neighbors among the deleted original anchors. The two-support structure nevertheless resolves this boundary case when there are at least three full original rows, by the separate exact full-row deletion theorem.

For the selected system and r≥12r\ge12, (8) forces a deficit at least three. If zz is the number of zero-deficit anchors, then

T≥max⁡qtq+(r−z−1),T≤r−1,T\ge\max_qt_q+(r-z-1),\qquad T\le r-1,

so z≥max⁡qtq≥3z\ge\max_qt_q\ge3. Thus exact deletion applies. After the first drop the anchor count is less than rr, and subsequent full-row deletions and absorptions preserve the required independence bound by greedy extension; see the linked theorem. This still does not bring the core into the known small-independence range of the hereditary span bound.

5. Transfer to degree-sum-minimal independent sets

The original anchor set QQ has degree sum

∑q∈Qd(q)=∑q(∣Pq∣−1)+2b=N−r+b=r2−r+T+2≤r2+1.(10)\sum_{q\in Q}d(q) =\sum_q(|P_q|-1)+2b =N-r+b =r^2-r+T+2\le r^2+1. \tag{10}

Choose a maximum independent set II minimizing its degree sum in the same color. Let Pu′P'_u be its private cliques, let B′B' be the remaining vertices, and put kx=∣N(x)∩I∣≥2k_x=|N(x)\cap I|\ge2 for x∈B′x\in B'. Exact counting of the edges incident to II gives

∑u∈Id(u)=N−r+∣B′∣+∑x∈B′(kx−2).\sum_{u\in I}d(u) =N-r+|B'|+\sum_{x\in B'}(k_x-2).

Comparison with (10) yields the joint budget

∣B′∣+∑x∈B′(kx−2)≤b≤r.(11)|B'|+\sum_{x\in B'}(k_x-2)\le b\le r. \tag{11}

In particular, the new total private deficit satisfies

T′:=∑u∈I(r−∣Pu′∣)=∣B′∣−1≤T≤r−1,∑x∈B′(kx−2)≤r−T′−1.(12)T':=\sum_{u\in I}(r-|P'_u|)=|B'|-1\le T\le r-1, \qquad \sum_{x\in B'}(k_x-2)\le r-T'-1. \tag{12}

Also d(x)≥d(u)d(x)\ge d(u) for x∈Pu′x\in P'_u, since exchanging uu for xx preserves a maximum independent set. Thus degree-sum minimality and total private deficit at most r−1r-1 can be imposed simultaneously. The new exceptions need not be two-supported, and the selected color still need not be a least color.

6. A sufficient condition for absorption to lower independence

The elementary criterion below remains valid. The exact full-row deletion theorem is stronger when there are at least three full rows and applies to the selected system for r≥12r\ge12, irrespective of this matching criterion.

Return to the original two-supported system. Let ZZ be its zero-deficit anchors and k=∣Z∣k=|Z|. Perform the deletion and absorption of Section 4, obtaining an induced residual RR. If the bipartite support graph between ZZ and Q∖ZQ\setminus Z has no matching covering ZZ, then

α(R)≤r−k.(13)\alpha(R)\le r-k. \tag{13}

In particular, (13) holds when k>r/2k>r/2.

To prove this, suppose J⊆RJ\subseteq R is independent of order r−k+1r-k+1. For each v∈Zv\in Z, the list Lv=Pv∖N(J)L_v=P_v\setminus N(J) has at least k−1k-1 vertices. Unless JJ contains an absorbed exception whose original support includes vv, the original anchor vv lies in LvL_v. It is anticomplete to every other deleted private row. Choose it first and greedily choose representatives from the remaining k−1k-1 lists, each of size at least k−1k-1. This augments JJ to an independent (r+1)(r+1)-set.

If JJ contains at least two absorbed exceptions supported on vv, those exceptions have the same sole neighbor vv in PvP_v. Hence ∣Lv∣≥k|L_v|\ge k. Leave that larger list to last in a greedy transversal, again obtaining a contradiction.

Therefore failure requires exactly one absorbed exception from JJ for each v∈Zv\in Z. Their positive-deficit owners are distinct: two exceptions with the same owner lie in the clique HuH_u. Their original support edges give a matching covering ZZ, proving (13).

This argument alone does not exclude a matching covering ZZ. The exact deletion theorem handles that case using a minimal Hall-deficient family of support lists, whose one- and two-element lists have at most two boundary incidences.

7. Structural closure of full rows and a proper overloaded core

For r≥12r\ge12, every original private clique PuP_u in the selected system has order at least four. Indeed, if ∣Pu∣≤3|P_u|\le3, then tu≥r−3t_u\ge r-3. Since T≤r−1T\le r-1, at most three anchors have positive deficit. Write their number as a≤3a\le3. There are r−a>r/2r-a>r/2 full rows, so Section 6 gives an induced residual of order ar+1ar+1 and independence number at most aa. For a=1a=1 the local cap is a contradiction; for a=2,3a=2,3, the two-row and three-row span bounds give a contradiction.

Assume now that every original ∣Pu∣≥4|P_u|\ge4. The full-row absorption can be iterated as a statement about cliques, without claiming that a reduced anchor set remains maximum independent. Keep the original independent set QQ and its original private cliques throughout the proof. Grow pairwise disjoint cliques CuC_u, initially Cu=PuC_u=P_u, by assigning exceptional vertices to their supporting anchors. A row is called full when ∣Cu∣=r|C_u|=r.

Suppose CuC_u is full and an unassigned exception xx has support {u,v}\{u,v\}. Its only neighbor in CuC_u is uu. If some original private vertex y∈Pvy\in P_v were nonadjacent to xx, the original two-support augmentation in Section 2 would make yy adjacent to every vertex of Pu∖{u}P_u\setminus\{u\}. Since ∣Pu∣−1≥3|P_u|-1\ge3, this contradicts the outside-neighbor bound for the full clique CuC_u. Thus xx is complete to the original PvP_v.

All exceptions assigned to a fixed row vv by this rule are pairwise adjacent. To verify this, let x,yx,y be two such exceptions, with other supporting anchors u,wu,w whose growing rows are full. If u=wu=w, the original private set Pu∖{u}P_u\setminus\{u\} is nonempty and anticomplete to x,yx,y. If x,yx,y were nonadjacent, one of its vertices would augment Q∖{u,v}Q\setminus\{u,v\} with x,yx,y. If u≠wu\ne w, use the original private lists

Pu∖(N(x)∪N(y)),Pw∖(N(x)∪N(y)).P_u\setminus(N(x)\cup N(y)),\qquad P_w\setminus(N(x)\cup N(y)).

Each has at least ∣Pu∣−2≥2|P_u|-2\ge2 or ∣Pw∣−2≥2|P_w|-2\ge2 vertices, respectively. Edges between the disjoint full cliques Cu,CwC_u,C_w form a matching, so nonadjacent representatives exist. With x,yx,y and Q∖{u,v,w}Q\setminus\{u,v,w\} they would again augment the original independent set. This proves the assertion.

Consequently Cv=PvC_v=P_v together with the exceptions assigned to vv is always a clique. Assign xx to vv, and mark vv full if its row reaches order rr. An unassigned exception cannot have both endpoints full: the preceding argument would make it complete to the original PvP_v, giving it more than one neighbor in the full CvC_v. Thus the process cannot overfill any row, and terminates with no unassigned exception incident to a full row.

This closure gives a purely structural restriction on the support multigraph. For a support component A⊆QA\subseteq Q, write bAb_A for its number of exceptional edges and put

δA=bA−∑u∈Atu.\delta_A=b_A-\sum_{u\in A}t_u.

If δA>0\delta_A>0, that component is not a pseudoforest. To prove this, run the closure within the component. Each assignment removes one unassigned edge and consumes one unit of the receiving row's capacity, so edges minus remaining capacity stays equal to δA\delta_A. At termination all remaining edges lie on nonfull rows, each with positive remaining capacity. Let UU be the anchors of all nonfull rows, including isolated ones. The number of remaining edges is therefore at least

∣U∣+δA≥∣U∣+1.|U|+\delta_A\ge |U|+1.

This is impossible if the original component is a pseudoforest, since every induced subgraph of a pseudoforest has at most as many edges as vertices. The argument also covers U=∅U=\varnothing, when positive excess cannot remain.

The support multigraph has total excess b−T=1b-T=1, so it has a component with δA>0\delta_A>0. Being connected and not a pseudoforest, that component has at least ∣A∣+1|A|+1 exceptional edges. Since b≤rb\le r, necessarily

a:=∣A∣≤r−1.(14)a:=|A|\le r-1. \tag{14}

Take the original private rows anchored in AA and all exceptional vertices supported there. This induced graph is anticomplete to Q∖AQ\setminus A, so its independence number is at most aa. Its order is

∑u∈A(r−tu)+bA=ar+δA>ar.(15)\sum_{u\in A}(r-t_u)+b_A=ar+\delta_A>ar. \tag{15}

Thus the reduction yields a proper overloaded core, not merely the original independence-rr graph. Restricting further gives a core on ar+1ar+1 vertices if desired.

The established span theorems rule out (15) when a≤3a\le3 (for r≥12r\ge12) or when r≥64a2r\ge64a^2. They do not exclude its remaining possible range 4≤a≤r−14\le a\le r-1. Structural closure has not supplied a bound reducing aa to the known small-independence range.

8. Excess deficit and a stronger full-row count

Proved, author-recorded. Let aa now denote the number of positive deficits, put D=∑idi=b−1D=\sum_i d_i=b-1, and set

E=D−a=∑di>0(di−1).E=D-a=\sum_{d_i>0}(d_i-1).

For r≥7r\ge7, every selected two-supported system satisfies

r−1≤(E+32).(16)r-1\le\binom{E+3}{2}. \tag{16}

The statement applies both to the original system and to the core after full-row deletion.

First, E=0E=0 is impossible. All positive deficits would equal one; Section 2 assigns every exception to a positive row in which it misses at most two vertices. Since b=D+1b=D+1, one row receives two exceptions. Their union with its private Kr−1K_{r-1} misses at most five edges, contrary to the local cap when r≥7r\ge7.

For E≥1E\ge1, give each positive row the threshold

hi=E−di+2.h_i=E-d_i+2.

Every exception has a supporting positive row ii in which it misses at most hih_i private vertices. This follows from completeness when its other supporting row is full. Otherwise, for positive endpoints i,ji,j, put x=di−1x=d_i-1, y=dj−1y=d_j-1. We have x+y≤Ex+y\le E, and

hihj−((di−1)(dj−1)+1)=(E+1)2−(E+1)(x+y)−1≥E>0.h_ih_j-\bigl((d_i-1)(d_j-1)+1\bigr) =(E+1)^2-(E+1)(x+y)-1\ge E>0.

Thus missing more than both thresholds contradicts (5).

Assign an owner using these thresholds. Some row of deficit dd owns at least d+1d+1 exceptions. Its entire private clique together with these exceptions is an (r+1)(r+1)-set with at most

(d+1)(E−d+2)+(d+12)=(d+1)(E+2−d2)≤(E+32)(d+1)(E-d+2)+\binom{d+1}{2} =(d+1)\left(E+2-\frac d2\right) \le\binom{E+3}{2}

missing edges. The last expression is maximized at d=E+1d=E+1 on 1≤d≤E+11\le d\le E+1. The local cap proves (16).

If the original system has zz full rows, then a=r−za=r-z and E=D−a≤z−1E=D-a\le z-1. Hence the exact deletion theorem can start with at least

z≥λ(r):=⌈8r−7−32⌉(17)\boxed{z\ge\lambda(r):= \left\lceil\frac{\sqrt{8r-7}-3}{2}\right\rceil} \tag{17}

full rows. This improves the earlier 2r/3+O(1)\sqrt{2r/3}+O(1) bound to 2r+O(1)\sqrt{2r}+O(1). In the final all-deficient core, E=D−a≤r−a−1E=D-a\le r-a-1 gives the same bound r−a≥λ(r)r-a\ge\lambda(r). This is still not a uniform exclusion of the core.

9. All row-plus-exception caps

Let F=H[B]‾F=\overline{H[B]}, where HH is the selected color graph (or its two-supported core). For every row ii and every S⊆BS\subseteq B,

∑x∈S(dH(x,Pi)+di−1)−e(F[S])≤di(r−1).(18)\boxed{ \sum_{x\in S}\bigl(d_H(x,P_i)+d_i-1\bigr)-e(F[S]) \le d_i(r-1).} \tag{18}

For s=∣S∣≥dis=|S|\ge d_i, the set Pi∪SP_i\cup S has order r−di+sr-d_i+s between rr and 2r2r, since ∣B∣≤r|B|\le r. On this interval,

Dr(r−di+s)=(r2)+(s−di+12).D_r(r-d_i+s)=\binom r2+\binom{s-d_i+1}{2}.

Subtract (r−di2)\binom{r-d_i}{2} and use e(H[S])=(s2)−e(F[S])e(H[S])=\binom s2-e(F[S]) to obtain (18). For s<dis<d_i, each summand is at most r−1r-1, so the inequality is immediate.

No assignment theorem has been deduced from (18). In particular, maximizing a functional w(S)−e(F[S])w(S)-e(F[S]) is not the usual capacitated orientation deficiency problem, which involves incident edges instead.

10. Independent exceptional sets and their common nonneighbors

If X⊆BX\subseteq B is independent and UU is the union of its supports, then

α ⁣(H ⁣[⋃i∈U(Pi∖NH(X))])≤∣U∣−∣X∣.(19)\alpha\!\left(H\!\left[ \bigcup_{i\in U}(P_i\setminus N_H(X))\right]\right) \le |U|-|X|. \tag{19}

Indeed, a larger independent set there could be combined with XX and all anchors outside UU to exceed the full independence bound. Consequently, when the support edges of XX form a tree, the displayed common-nonneighbor union is a clique; when ∣X∣=∣U∣|X|=|U|, it is empty. These are exact geometric constraints not captured by (18) alone.

11. Capacitated Hall constraints on exceptional cliques

Here assume the core is all-deficient, so di≥1d_i\ge1. Put q=r−1q=r-1 and

wi(x)=dH(x,Pi)+di−1=q−∣Pi∖NH(x)∣≥0.w_i(x)=d_H(x,P_i)+d_i-1 =q-|P_i\setminus N_H(x)|\ge0.

For an exception xx supported on {u,v}\{u,v\},

wu(x)+wv(x)≥q.(20)w_u(x)+w_v(x)\ge q. \tag{20}

Its nonneighbors in Pu∪PvP_u\cup P_v form a clique, by the two-support augmentation argument. This clique has order at most r−1r-1: if it contained a KrK_r, each of the outside anchors u,vu,v would have at most one neighbor there, while every chosen vertex is adjacent to its own anchor, an impossibility for r≥3r\ge3. This proves (20).

For a nonempty anchor set UU, let

BU={x∈B:Kx⊆U},DU=∑i∈Udi.B_U=\{x\in B:K_x\subseteq U\},\qquad D_U=\sum_{i\in U}d_i.

Apply (18) with S=BUS=B_U and sum over i∈Ui\in U. Nonnegativity permits retaining just the two support contributions from (20), giving

∣U∣ e(F[BU])≥(r−1)(∣BU∣−DU).(21)\boxed{|U|\,e(F[B_U])\ge(r-1)(|B_U|-D_U).} \tag{21}

In particular e(F[B])≥⌈(r−1)/a⌉e(F[B])\ge\lceil(r-1)/a\rceil, where aa is the number of core anchors.

If C⊆BC\subseteq B is a target clique, apply the same argument to CU={x∈C:Kx⊆U}C_U=\{x\in C:K_x\subseteq U\}. Now e(F[CU])=0e(F[C_U])=0, so

∣CU∣≤DU(U⊆[a]).|C_U|\le D_U\qquad(U\subseteq[a]).

These are the capacitated Hall conditions: each exceptional target clique can be assigned to supporting rows with capacities did_i. In particular BB itself is not a target clique, since ∣B∣=D+1|B|=D+1.

No argument extends these assignments to all exceptions. The missing exception-edge terms in (21) are a real remaining term, not an established contradiction. The nonnegativity used above also explains why this summation is stated only after all full rows have been removed.