Wiki
Wiki

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

Updated

Incidence minimal families


Unlike deleting edges to preserve a large transversal number, deleting an incidence shrinks an edge and can only increase the transversal number. This suggests minimizing incidences subject to P7P_7. The resulting witness conditions are proved below; they do not yield a general three-quarter upper bound.

Simultaneous minimality

Suppose a finite rank-at-most-kk P7P_7 family has transversal number at least tt. Among such finite families choose one with the fewest edges, and then with the fewest incidences. Its transversal number is exactly tt: deleting any edge leaves a family with transversal number at most t−1t-1, and adjoining one point of that edge covers the whole family.

It is therefore transversal-critical. Moreover, deleting any incidence from an edge of size at least two must destroy P7P_7. Otherwise rank is unchanged or lowered and the transversal number cannot decrease, contradicting the minimal choice. If a shrinking makes two edges equal, removing the duplicate also contradicts minimality.

This yields simultaneous edge criticality and incidence minimality. The latter alone is insufficient for a small ground set, as explained below.

A critical-cover trace constraint

There is an exact compatibility condition between edge criticality and its minimum covers. Let H\mathcal H be a finite transversal-critical family of nonempty edges with τ(H)=t≥2\tau(\mathcal H)=t\ge2. For any edge EE, choose a (t−1)(t-1)-point transversal TET_E of H∖{E}\mathcal H\setminus\{E\}. Necessarily TE∩E=∅T_E\cap E=\varnothing, since otherwise it would cover H\mathcal H.

For every u∈TEu\in T_E and every v∈Ev\in E, there is an edge FF such that

F∩TE={u},v∉F.(C)F\cap T_E=\{u\},\qquad v\notin F. \tag{C}

To prove this, TE∪{v}T_E\cup\{v\} is a minimum transversal of H\mathcal H. Removing uu leaves a set of size t−1t-1 and hence misses some edge FF. This FF misses vv and TE∖{u}T_E\setminus\{u\}. It is not EE, because v∈Ev\in E, so it meets TET_E and must do so exactly at uu.

Equivalently, each nonempty singleton-trace class satisfies

E∩⋂F∈HF∩TE={u}F=∅.E\cap\bigcap_{\substack{F\in\mathcal H\\F\cap T_E=\{u\}}}F =\varnothing.

This uses only edge criticality, not P7P_7 or incidence minimality. The edge FF may depend on both uu and vv. Selecting one such edge for each v∈Ev\in E gives a witness using at most ∣E∣|E| rows for each uu; no bound independent of ∣E∣|E|, or useful aggregate rank estimate for these classes, was proved. Combining (C) with all essential-incidence witnesses to deduce k≥4t/3−o(k)k\ge4t/3-o(k) remains unproved. In particular, (C) does not supply an extra local-test slot for a seven-row mandatory-point witness.

Exact transversal numbers on the critical-cover trace cube

The same critical cover gives an exact normalization for all its subsets. Fix EE and T=TET=T_E, with ∣T∣=t−1|T|=t-1 and T∩E=∅T\cap E=\varnothing. For S⊆TS\subseteq T, define

H[S]={F∈H:F∩T⊆S}.\mathcal H[S]=\{F\in\mathcal H:F\cap T\subseteq S\}.

Then

τ(H[S])=∣S∣+1.(C1)\boxed{\tau(\mathcal H[S])=|S|+1.} \tag{C1}

For the upper bound, SS covers all members except EE, so add any point of EE. For the lower bound, a cover of size at most ∣S∣|S|, together with T∖ST\setminus S, would cover H\mathcal H with at most t−1t-1 points.

For each x∈Ex\in E, there is also the exact residual formula

τ({F∈H[S]:x∉F})=∣S∣.(C2)\boxed{\tau(\{F\in\mathcal H[S]:x\notin F\})=|S|.} \tag{C2}

The upper bound uses SS, because EE is the only edge missed by TT and has been removed. Adding xx to a cover of the residual covers H[S]\mathcal H[S], so (C1) gives the opposite bound. For S=∅S=\varnothing, the residual is empty and has transversal number zero. All these families inherit P7P_7 when H\mathcal H has it.

These identities constrain every trace class simultaneously, but do not proportionately lower the rank as ∣S∣|S| decreases. In particular, (C1) retains the original edge EE even at S=∅S=\varnothing. No rank-sensitive extraction or three-quarter bound follows from these formulas alone.

The critical-cover projection can destroy locality

Tracing a critical family onto U=E∪TEU=E\cup T_E preserves its transversal number exactly. Shrinking edges cannot decrease it, while TET_E together with one point of EE covers all traces and has size tt. All traces are nonempty. However, even simultaneous edge criticality and incidence minimality do not make this projection preserve P7P_7, or even P3P_3.

For a finite example, take C9C_9 with vertices 0,…,80,\ldots,8 in cyclic order. Its nine independent four-subsets are the points of a new hypergraph, and its rows are

Ei={I:i∈I, ∣I∣=4, I independent in C9}.E_i=\{I:i\in I,\ |I|=4,\ I\text{ independent in }C_9\}.

There are nine such independent sets because their cyclic gaps are 2,2,2,32,2,2,3; by cyclic symmetry each row has size four. The family has transversal number three: two independent sets cannot cover an odd cycle, while, for example, {0,2,4,6}\{0,2,4,6\}, {1,3,5,7}\{1,3,5,7\}, and {8,1,3,5}\{8,1,3,5\} cover it. Deleting any row leaves a path on eight vertices, whose two four-point color classes are allowable ground points. Thus the family is transversal-critical and has P8P_8.

Every incidence i∈Ii\in I is essential for P7P_7. The independent four-set II is the larger color class of a seven-vertex induced path, determined by its three gaps of length two. Any cover of this path by two independent sets gives its unique bipartition. The four-point class can be covered only by the ground point II itself. Deleting the incidence of row ii with point II prevents that point from covering the class, and no other independent four-set contains it. Those seven rows consequently lose every two-cover.

Take E=E0E=E_0 and

TE={{1,3,5,7},{2,4,6,8}}.T_E=\{\{1,3,5,7\},\{2,4,6,8\}\}.

On U=E0∪TEU=E_0\cup T_E, the traces of E0,E1,E8E_0,E_1,E_8 are pairwise disjoint. The last two are respectively the two singleton points of TET_E: no independent set contains 00 together with 11 or 88. Thus the trace family fails P3P_3 despite both criticality conditions and τ/k=3/4\tau/k=3/4 in the original family. This fixed-rank example is not an asymptotic counterexample and does not exclude a projection theorem requiring a fixed excess above three quarters.

Mandatory-point witnesses

Let v∈Ev\in E, where ∣E∣≥2|E|\ge2, be an essential incidence. There is a subfamily C\mathcal C of m≤7m\le7 original rows containing EE such that replacing EE by E∖{v}E\setminus\{v\} makes its transversal number greater than two. Every two-cover of C\mathcal C is consequently of the form {v,w}\{v,w\} with w∉Ew\notin E.

The point vv cannot lie in every witness row. Otherwise, after the deletion, {v,x}\{v,x\} for any x∈E∖{v}x\in E\setminus\{v\} would still cover them. Choose a partner ww in a two-cover of the original rows. The rows missing ww all contain vv, include EE, and have common intersection exactly {v}\{v\}. Any other common point xx would give the cover {x,w}\{x,w\} avoiding vv. At least one row contains ww and misses vv, so there are at most six rows in this singleton-intersection certificate.

Short witnesses force smaller global covers

Suppose m≤6m\le6, and set

W=⋂F∈Cv∉FF.W=\bigcap_{\substack{F\in\mathcal C\\v\notin F}}F.

This is the intersection of a nonempty family. It is nonempty because it contains every possible partner ww, and W∩E=∅W\cap E=\varnothing: a point in that intersection would still supplement vv after the incidence deletion.

Every at-most-(7−m)(7-m) edges avoiding vv have a common point in WW. Indeed, adjoining them to C\mathcal C and applying P7P_7 forces a cover {v,w}\{v,w\} with ww in all these edges and in WW.

Partition WW into 7−m7-m parts of size at most ⌈∣W∣/(7−m)⌉\lceil |W|/(7-m)\rceil. If none of these parts covers the family of edges avoiding vv, choose an avoiding edge for each part. Their intersection with WW is empty, a contradiction. Adding vv to a cover of the residual family gives

τ(H)≤1+⌈∣W∣7−m⌉≤1+⌈k7−m⌉.(1)\tau(\mathcal H)\le1+\left\lceil\frac{|W|}{7-m}\right\rceil \le1+\left\lceil\frac{k}{7-m}\right\rceil. \tag{1}

Thus τ>1+⌈k/2⌉\tau>1+\lceil k/2\rceil excludes all witnesses with at most five rows. A six-row witness has ∣W∣≥τ(H)−1|W|\ge\tau(\mathcal H)-1.

A restriction on six-row witnesses

For k≥4k\ge4, if τ(H)>⌈3k/4⌉\tau(\mathcal H)>\lceil3k/4\rceil, a six-row witness cannot have its mandatory point vv in exactly two rows.

To prove this, denote those two rows by E,AE,A. If some w∈Ww\in W lay in AA, the only row missing ww would be EE, contradicting the preceding singleton-intersection certificate because ∣E∣≥2|E|\ge2. Hence W∩(E∪A)=∅W\cap(E\cup A)=\varnothing, and the same certificate gives E∩A={v}E\cap A=\{v\}.

Choose a witness row BB missing vv. The triple E,A,BE,A,B has empty common intersection, and its pair-intersection sizes are

1,∣E∩B∣≤k−∣W∣,∣A∩B∣≤k−∣W∣.1,\qquad |E\cap B|\le k-|W|,\qquad |A\cap B|\le k-|W|.

Writing t=τ(H)t=\tau(\mathcal H) and using t≥⌈3k/4⌉+1t\ge\lceil3k/4\rceil+1, all three sizes are at most ⌊k/4⌋\lfloor k/4\rfloor. The new rank-aware good-triple certificate (G0a) in Certificates and obstructions then gives τ≤⌈3k/4⌉\tau\le\lceil3k/4\rceil, a contradiction.

A one-sided trace classification for three containing rows

Let H\mathcal H be a rank-at-most-kk P7P_7 family with transversal number tt, and suppose

C={E,A,B,F1,F2,F3}\mathcal C=\{E,A,B,F_1,F_2,F_3\}

is a six-row witness for the essential incidence v∈Ev\in E, with ∣E∣≥2|E|\ge2. Assume that vv belongs to exactly E,A,BE,A,B, that

E∩A=E∩B={v},(2)E\cap A=E\cap B=\{v\}, \tag{2}

and that t≥⌈3k/4⌉+2t\ge\lceil3k/4\rceil+2. Put

E0=E∖{v},P=(A∩B)∖{v},W=F1∩F2∩F3.E_0=E\setminus\{v\},\qquad P=(A\cap B)\setminus\{v\},\qquad W=F_1\cap F_2\cap F_3.

Then one of the following two configurations holds:

  • Every FiF_i avoids PP, and the three traces Fi∩E0F_i\cap E_0 are pairwise intersecting with empty total intersection.
  • Every FiF_i avoids EE, and the three traces Fi∩PF_i\cap P are pairwise intersecting with empty total intersection.

To prove this, write w=∣W∣≥t−1w=|W|\ge t-1 and q=k−wq=k-w. The earlier witness argument gives W∩E0=∅W\cap E_0=\varnothing. Also W∩P=∅W\cap P=\varnothing: a point in this intersection, together with any point of the nonempty E0E_0, would cover C\mathcal C without vv. By (2), every non-vv two-cover of E,A,BE,A,B is a pair from E0×PE_0\times P.

Suppose some pair i,ji,j satisfies

E0∩Fi∩Fj=P∩Fi∩Fj=∅.(3)E_0\cap F_i\cap F_j=P\cap F_i\cap F_j=\varnothing. \tag{3}

Every non-vv cover of E,A,B,Fi,FjE,A,B,F_i,F_j then has both endpoints in

U=(E0∪P)∩(Fi∪Fj).U=(E_0\cup P)\cap(F_i\cup F_j).

Indeed, neither endpoint can meet both Fi,FjF_i,F_j, so one endpoint must meet each. Since WW avoids E0∪PE_0\cup P, we have ∣U∣≤2q|U|\le2q. Moreover, (3) gives U∩Wij=∅U\cap W_{ij}=\varnothing, where Wij=Fi∩FjW_{ij}=F_i\cap F_j.

Choose an edge DD avoiding vv, UU, and enough of WijW_{ij} to leave ∣D∩Wij∣≤t−2|D\cap W_{ij}|\le t-2. The required avoiding set has size

1+∣U∣+max⁡{0,∣Wij∣−(t−2)}≤t−1,1+|U|+\max\{0,|W_{ij}|-(t-2)\}\le t-1,

because the assumed lower bound on tt gives

∣U∣≤2(k−t+1)≤t−2,∣U∣+∣Wij∣≤2q+k≤3k−2t+2≤2t−4.|U|\le2(k-t+1)\le t-2, \qquad |U|+|W_{ij}|\le2q+k\le3k-2t+2\le2t-4.

Thus such DD exists by the definition of tt.

Every two-cover of the at-most-six rows E,A,B,Fi,Fj,DE,A,B,F_i,F_j,D must contain vv, and its partner belongs to Wij∩DW_{ij}\cap D. Appending any edge avoiding vv and using P7P_7 shows that Wij∩DW_{ij}\cap D covers all edges avoiding vv. Consequently

τ(H)≤1+∣Wij∩D∣≤t−1,\tau(\mathcal H)\le1+|W_{ij}\cap D|\le t-1,

a contradiction. Therefore every pair Fi,FjF_i,F_j has a common point in E0E_0 or in PP.

A point in E0∩Fi∩FjE_0\cap F_i\cap F_j forbids every point in P∩FℓP\cap F_\ell for the remaining index ℓ\ell: these two points would otherwise cover all six rows without vv. The same assertion holds with E0,PE_0,P interchanged. In particular, common points on opposite sides cannot occur for two different pairs of indices, since the two pairs together cover all three indices. If one pair has a common point in E0E_0, the other two pairs must therefore also have common points in E0E_0, and the three forbidden-complement assertions imply P∩Fi=∅P\cap F_i=\varnothing for every ii. Otherwise all three pairs have common points in PP and all E0∩FiE_0\cap F_i vanish. The total intersections of these traces are empty because WW avoids both E0E_0 and PP. This proves the classification.

Hypothesis (2) holds automatically if WW meets both AA and BB. For example, a point of W∩AW\cap A, together with any point of (E∩B)∖{v}(E\cap B)\setminus\{v\}, would be a forbidden non-vv cover; hence E∩B={v}E\cap B=\{v\}. The other equality is symmetric. The two configurations above have not themselves been excluded.

There is also a necessary restriction without (2). In any mandatory witness with exactly three vv-containing rows, their total intersection is {v}\{v\}: another common point could replace vv in a cover with a partner from WW. The arbitrary-common-intersection certificate (G0c) therefore implies that, when t>1+⌈3k/4⌉t>1+\lceil3k/4\rceil, some exclusive pair intersection of these three rows has size at least ⌈(t−1)/3⌉\lceil(t-1)/3\rceil.

The remaining seven-row gap

No theorem was obtained saying that a seven-row essential-incidence witness forces τ≤3k/4+O(1)\tau\le3k/4+O(1). Even the case where vv belongs to exactly two witness rows remains untreated. Denote those rows by E,AE,A and the other five by F1,…,F5F_1,\ldots,F_5. The mandatory-point argument gives

E∩A={v},W=⋂i=15Fi≠∅,W∩(E∪A)=∅.E\cap A=\{v\},\qquad W=\bigcap_{i=1}^5F_i\ne\varnothing,\qquad W\cap(E\cup A)=\varnothing.

For x∈E∖{v}x\in E\setminus\{v\} and y∈A∖{v}y\in A\setminus\{v\}, define their missing-index sets

Mx={i:x∉Fi},Ny={i:y∉Fi}.M_x=\{i:x\notin F_i\},\qquad N_y=\{i:y\notin F_i\}.

The absence of non-vv two-covers is precisely Mx∩Ny≠∅M_x\cap N_y\ne\varnothing for all such pairs. On deleting FiF_i, the new non-vv candidate pairs are exactly those with Mx∩Ny={i}M_x\cap N_y=\{i\}. The possible partners of vv become Wi=⋂j≠iFjW_i=\bigcap_{j\ne i}F_j.

This is an exact local formulation, not a small-endpoint-set bound or a counterexample construction. In contrast to a six-row witness, it leaves no room to append an arbitrary ambient edge and deduce ∣W∣≥t−1|W|\ge t-1. No argument combining this seven-row formulation with the full ambient P7P_7 condition and large tt was completed. The rank-feasible local patterns considered during this approach have small transversal number and do not constitute counterexamples to the three-quarter conjecture.

Incidence minimality alone does not bound the ground set

The family in Indispensable local cores has rank 16m+116m+1, transversal number 12m12m, and (2mm)\binom{2m}{m} added points vSv_S. Each has three incidences protected by a seven-row witness.

In that witness every core point has degree four, vSv_S has degree three, no pair of core points covers all rows, and every other exterior point has degree at most one. After deleting any one of the three vSv_S incidences, a pair using vSv_S covers at most six rows. All other pairs already failed. Therefore each incidence is essential to P7P_7.

Further incidence shrinkings cannot restore a lost two-cover. Hence any incidence-minimal P7P_7 descendant retaining these original rows still contains all these exponentially many vertices. This excludes a polynomial ground-set theorem based on incidence minimality alone, even at a ratio tending to three quarters.

It does not exclude a stronger theorem under simultaneous edge criticality and incidence minimality, nor one assuming a fixed excess above three quarters. The exceptional rows of the cited construction have a two-point transversal, so preliminary edge minimization or low-transversal deletion may change the example substantially.