Wiki
Wiki

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

Updated

Fano cap covering


Let C\mathcal C be the seven complements of lines of the Fano plane on [7][7]. For nonnegative demands q1,…,q7q_1,\ldots,q_7, let

ϕ(q)=min⁡{∑C∈CwC:wC≥0,∑C∋iwC≥qi (i∈[7])}.\phi(q)=\min\left\{\sum_{C\in\mathcal C}w_C: w_C\ge0,\quad \sum_{C\ni i}w_C\ge q_i\ (i\in[7])\right\}.

Then

ϕ(q)=max⁡{max⁡iqi, 12max⁡line L∑i∈Lqi, 14∑iqi}.(1)\boxed{\phi(q)= \max\left\{\max_iq_i,\ \frac12\max_{\text{line }L}\sum_{i\in L}q_i,\ \frac14\sum_iq_i\right\}.} \tag{1}

Excess incidences in a covering solution can be deleted. Thus (1) is also the minimum total mass of points realizing the exact row demands, when each point's spectrum must be a subset of a Fano cap.

This lemma assists the block-profile constructions. It does not turn arbitrary edges with the given profiles into members of an unsymmetrized family.

Dual polytope

By finite linear-programming duality, it suffices to show that

D={z≥0:∑i∈Czi≤1 for every cap C}D=\{z\ge0:\sum_{i\in C}z_i\le1\ \text{for every cap }C\}

is the convex hull of zero, the seven coordinate unit vectors, the seven half-line vectors 121L\frac12\mathbf1_L, and 141\frac14\mathbf1.

All the listed vectors are feasible. For the converse, put u=min⁡iziu=\min_i z_i. Then u≤1/4u\le1/4. If u=1/4u=1/4, all coordinates equal 1/41/4: every coordinate belongs to a cap whose sum is at most one. Otherwise

z=4u(141)+(1−4u)z′,z′=z−u11−4u∈D,z=4u\left(\frac14\mathbf1\right)+(1-4u)z', \qquad z'=\frac{z-u\mathbf1}{1-4u}\in D,

and z′z' has a zero coordinate. It remains to analyze this face.

The six-coordinate face

Delete the zero-coordinate Fano point. Identify the other six points with the edges of K4K_4, so that the four Fano lines not containing the deleted point become the four stars of K4K_4. The remaining cap inequalities are exactly

∑e∈Tze≤1(T a triangle),∑e∈Qze≤1(Q a four-cycle).(2)\sum_{e\in T}z_e\le1 \quad(T\text{ a triangle}),\qquad \sum_{e\in Q}z_e\le1 \quad(Q\text{ a four-cycle}). \tag{2}

We show that this polytope is the convex hull of zero, the six edge unit vectors, and the four half-star vectors. Consider

B=max⁡{∑vβv:βv≥0,βu+βv≤2zuv}.B=\max\left\{\sum_v\beta_v: \beta_v\ge0,\quad \beta_u+\beta_v\le2z_{uv}\right\}.

Its dual is the minimum-cost fractional edge-cover problem on K4K_4, with edge costs 2ze2z_e.

For nonnegative costs this problem has an integral optimum. Here is a short justification specific to K4K_4. Restrict edge-cover variables to [0,1][0,1] and take an optimum extreme point. Every tight vertex incident with a fractional edge has at least two fractional incident edges: it cannot also have an incident edge of weight one. Independence of the fractional columns on the tight vertex constraints, together with the degree count, forces every fractional component to be a cycle, all its vertices tight. An even cycle admits an alternating perturbation, so such a component must be an odd cycle. On four vertices the only possibility is a triangle, but the fourth vertex could not then be covered without meeting a tight triangle vertex. Thus no fractional edges occur.

An inclusion-minimal integral edge cover of K4K_4 is either a perfect matching or a three-edge star. Consequently

B=2min⁡{∑e∈Mze, ∑e∈Sze:M a perfect matching, S a star}.B=2\min\left\{\sum_{e\in M}z_e,\ \sum_{e\in S}z_e: M\text{ a perfect matching},\ S\text{ a star}\right\}.

Choose an optimal β\beta and set αuv=zuv−(βu+βv)/2≥0\alpha_{uv}=z_{uv}-(\beta_u+\beta_v)/2\ge0. Then

z=∑eαeee+∑vβv(121star⁡(v)).z=\sum_e\alpha_e\mathbf e_e+\sum_v\beta_v\left(\frac12\mathbf1_{\operatorname{star}(v)}\right).

The coefficient sum is

∑eαe+∑vβv=∑eze−B2=max⁡{∑e∉Mze, ∑e∉Sze}≤1,\sum_e\alpha_e+\sum_v\beta_v =\sum_ez_e-\frac B2 =\max\left\{\sum_{e\notin M}z_e,\ \sum_{e\notin S}z_e\right\}\le1,

because the complements of perfect matchings and stars are four-cycles and triangles. This proves the asserted convex-hull description on the face, and hence on DD. Evaluating its listed vertices gives (1).

The initial bounded enumeration of active constraints suggested this formula, but the proof above does not depend on that computation.