Wiki
Wiki

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

Updated

Maximum set surplus


Proved necessary inequality, not a resolution of E617. Let r≥12r\ge12. Suppose a graph QQ satisfies the hereditary density caps e(Q[W])≤Dr(∣W∣)e(Q[W])\le D_r(|W|) from Reductions, is KrK_r-free, and satisfies

∣Q∣=ar+ε,α(Q)=a,δ(Q)≥r,ε∈{0,1}.|Q|=ar+\varepsilon,\qquad \alpha(Q)=a,\qquad \delta(Q)\ge r, \qquad \varepsilon\in\{0,1\}.

Then every maximum independent set II satisfies

7∑q∈I(dQ(q)−r)≥2a+12ε.(1)\boxed{7\sum_{q\in I}(d_Q(q)-r)\ge2a+12\varepsilon.} \tag{1}

No bound on aa, exception count, or support size is assumed. Neither criticality, proper-induced minimality, nor an edge budget is used.

In fact, for a specified maximum set II, the proof only needs degree at least rr at its unit-row anchors. Degrees at its other anchors and at vertices outside II need not be at least rr: their individual surpluses may be negative, but the exact identities remain valid, and only σU,σM≥0\sigma_U,\sigma_M\ge0 are used at the end. The minimum-degree hypothesis above is a convenient sufficient condition for the assertion about every maximum set.

The proof extends the unit-row routing argument in Unit-deficit propagation. Its new point is to count higher-support exceptions separately, without routing them. Only caps at orders r+1,r+2,2r−1r+1,r+2,2r-1 are needed.

Private rows and exact accounting

Write I={q1,…,qa}I=\{q_1,\ldots,q_a\} and si=dQ(qi)−r≥0s_i=d_Q(q_i)-r\ge0. Let PiP_i contain qiq_i and all outside vertices whose only neighbor in II is qiq_i. Each PiP_i is a clique: a nonadjacent pair in it, together with the other anchors, would exceed aa. Put

∣Pi∣=r−di,di≥1,E=∑i(di−1),σ=∑isi.|P_i|=r-d_i,\quad d_i\ge1,\quad E=\sum_i(d_i-1),\quad \sigma=\sum_i s_i.

The remaining vertices form an exception set BB, each with at least two anchor neighbors. For x∈Bx\in B write kx=∣N(x)∩I∣k_x=|N(x)\cap I|, and put

b=∣B∣=a+E+ε,v=∣{x∈B:kx≥3}∣,u=∑x∈B(kx−2).b=|B|=a+E+\varepsilon,\quad v=|\{x\in B:k_x\ge3\}|,\quad u=\sum_{x\in B}(k_x-2).

Thus v≤uv\le u. Counting edges incident with II in two ways gives

σ=E+u+2ε.(2)\boxed{\sigma=E+u+2\varepsilon.} \tag{2}

Indeed the private rows contribute ar−a−∑idiar-a-\sum_i d_i incidences, and BB contributes 2b+u2b+u.

Call a row unit if di=1d_i=1, and let UU be the set of its indices. Then a−∣U∣≤Ea-|U|\le E. A unit anchor has exactly 2+si2+s_i incident exceptions, regardless of their support sizes. An exception with exactly two anchor neighbors will be called two-supported below.

Local routing only uses two-supported exceptions

If a two-supported exception xx has supports i,ji,j, then Pi∖N(x)P_i\setminus N(x) and Pj∖N(x)P_j\setminus N(x) are completely joined: a nonadjacent pair from these sets, together with xx and the unused anchors, would be an independent (a+1)(a+1)-set.

Suppose both rows are unit, and let their nonneighbor counts be mi,mjm_i,m_j. Both are positive, since otherwise xx creates a KrK_r. The cap on Pi∪Pj∪{x}P_i\cup P_j\cup\{x\} gives

(mi−1)(mj−1)≤1.(3)(m_i-1)(m_j-1)\le1. \tag{3}

For an explicit count, its 2r−12r-1 vertices have at most (r−1)2−mimj+mi+mj(r-1)^2-m_im_j+m_i+m_j nonedges, and the cap requires at least (r−1)2(r-1)^2.

If mi≤2m_i\le2, then for every z∉Pi∪{x}z\notin P_i\cup\{x\} the cap on Pi∪{x,z}P_i\cup\{x,z\} gives

d(z,Pi)+1xz∈E(Q)≤mi+1≤3.(4)d(z,P_i)+\mathbf1_{xz\in E(Q)}\le m_i+1\le3. \tag{4}

Any other two-supported exception yy at ii therefore misses at least r−4r-4 vertices of PiP_i. If it missed a private vertex in its other supporting row, that vertex would be complete to all these nonneighbors, contrary to (4). Consequently yy is complete to its other row, which cannot be unit because QQ is KrK_r-free. In particular there is no parallel unit--unit exception.

If instead mi≥3m_i\ge3, (3) gives mj=1m_j=1. Let zz be the unique nonneighbor of xx in PjP_j, and put Zi=Pi∪{x,z}Z_i=P_i\cup\{x,z\}. Complete joining of the nonneighbor sets gives d(z,Pi)≥mid(z,P_i)\ge m_i; the cap on ZiZ_i gives d(z,Pi)≤mi+1d(z,P_i)\le m_i+1. Thus

e(Q[Zi])=Dr(r+1)−λ,λ∈{0,1}.e(Q[Z_i])=D_r(r+1)-\lambda,\qquad \lambda\in\{0,1\}.

Since Dr(r+2)−Dr(r+1)=2D_r(r+2)-D_r(r+1)=2, every vertex outside ZiZ_i has at most three neighbors in ZiZ_i. A second two-supported exception at ii cannot have other endpoint jj, by (4) at the active endpoint jj. Its other row therefore lies outside ZiZ_i, and the same augmentation argument makes it complete to that row. Again the other row is nonunit.

It follows that the support graph of two-supported exceptions between unit rows is a matching. Let it have tt edges, with endpoint set M⊆UM\subseteq U of order 2t2t.

At each endpoint designate its matching exception. Every other two-supported exception there is complete to its other, nonunit row. Route it to that row. These routed exceptions are distinct across sources: their only unit endpoint is their source.

For a nonunit row PjP_j, let CjC_j be the routed group. Then

Pj∪Cj is a clique,∣Cj∣≤dj−1.(5)P_j\cup C_j\text{ is a clique},\qquad |C_j|\le d_j-1. \tag{5}

To verify pairwise adjacency in CjC_j, every routed exception lies outside every source block: designated exceptions are unit--unit, routed exceptions are unit--nonunit, and private holes are not exceptions. Hence each routed exception has at most three neighbors in every source row. Two nonadjacent group members from one source have a common private nonneighbor there and augment the unused anchors. If their sources are distinct i,ki,k, their common-nonneighbor lists in those rows have at least r−7r-7 vertices. Augmentation forces the two lists to be completely joined. Choose a listed vertex in PkP_k outside the source block at ii, omitting at most its one private hole. It has at most three neighbors in PiP_i, contradicting r−7>3r-7>3. This proves (5). The total routed capacity is therefore EE.

Higher-support incidence accounting

Let KUK_U and KMK_M count incidences of exceptions with support at least three at unit anchors and matching endpoints respectively. Write σU=∑i∈Usi\sigma_U=\sum_{i\in U}s_i and σM=∑i∈Msi\sigma_M=\sum_{i\in M}s_i. Then

KM≤KU≤u+2v.K_M\le K_U\le u+2v.

The number of routed exceptions from matching endpoints is exactly 2t+σM−KM2t+\sigma_M-K_M: subtract the designated incidence at each endpoint from its 2+si2+s_i incidences and exclude the higher supports. By (5),

t≤E+u+2v−σM2.(6)t\le\frac{E+u+2v-\sigma_M}{2}. \tag{6}

The b−vb-v two-supported exceptions have 2∣U∣+σU−KU2|U|+\sigma_U-K_U incidences at unit anchors. Only the tt matching exceptions are counted twice, so

b−v≥2∣U∣+σU−KU−t.b-v\ge2|U|+\sigma_U-K_U-t.

Using ∣U∣≥a−E|U|\ge a-E and the identities above gives

a≤3E+ε+u+v+t−σU≤72(E+u)+ε−σU−12σM.a\le3E+\varepsilon+u+v+t-\sigma_U \le\tfrac72(E+u)+\varepsilon-\sigma_U-\tfrac12\sigma_M.

Substitute (2) to obtain the slightly stronger statement

7σ≥2a+12ε+2σU+σM,7\sigma\ge2a+12\varepsilon+2\sigma_U+\sigma_M,

which proves (1).

Consequences and limitation

For the least-budget minimal-ratio core, every maximum independent set has surplus at least ⌈(2a+12)/7⌉\lceil(2a+12)/7\rceil. In a balanced capped KrK_r-free graph of order arar and minimum degree at least rr, an all-degree-rr maximum independent set is impossible.

If the unproved light-set bridge gave a maximum independent set of degree sum at most a(r−1)+r+1a(r-1)+r+1, then σ≤r−a+1\sigma\le r-a+1, and (1) would yield

9a≤7r−5.9a\le7r-5.

This conclusion would no longer require transferring that set to a two-supported anchor system. The light-set bridge is still unproved, and the displayed linear bound would itself leave an unbounded range.

The local routing hypotheses, higher-support separation, and incidence algebra were checked by the author and are author-recorded. No computation is used.