Wiki
Wiki

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

Updated

Large block profiles


This page proves two restricted results. The first applies to arbitrary capacities when the heavy supports have a transversal of size at most two. The second proves

τ(H)≤34k+15m\tau(\mathcal H)\le \frac34k+15m

for full mm-block kk-uniform families with the local seven-edge property, provided every block has size at least kk and k≥70k\ge70. The block sizes need not be equal, and there is no upper bound on their total size. These results do not resolve the arbitrary-capacity profile question or either unrestricted assertion of Problem 644.

Definitions

Let S⊆∏j=1m[0,cj]S\subseteq\prod_{j=1}^m[0,c_j] be a finite nonempty set of profiles, each of coordinate sum one. Define T(S;c)T(S;c) to be the infimum of ∑jdj\sum_jd_j over vectors 0≤dj≤cj0\le d_j\le c_j for which no A∈SA\in S satisfies A≤c−dA\le c-d coordinatewise.

A weighted realization of an ordered tuple of profiles assigns nonnegative mass to point spectra, separately in each block. A spectrum is the subset of row indices containing the point. Its total mass in block jj is at most cjc_j, and its row demands equal the chosen profiles. A realization is bad if no two appearing spectra have union equal to the whole row set. Repeated profiles are allowed.

We may first cover, rather than realize exactly, the prescribed row demands, and then delete excess incidences. Deleting incidences replaces spectra by subsets, so it cannot create a two-point cover.

The heavy support of AA is

H(A)={j:Aj>2cj/3}.H(A)=\{j:A_j>2c_j/3\}.

A profile with empty heavy support is called light on this page.

We will use subadditivity of the deletion threshold. If S=S1∪S2S=S_1\cup S_2 with both subfamilies nonempty, then

T(S;c)≤T(S1;c)+T(S2;c).(1)T(S;c)\le T(S_1;c)+T(S_2;c). \tag{1}

Indeed, combine deletion vectors for the two subfamilies by adding them coordinatewise and truncating at cc. The combined deletion removes both subfamilies and has no greater total mass than the sum of the two original deletion masses. Taking infima proves (1).

A non-Fano seven-row construction

Distinguish four rows X1,…,X4X_1,\ldots,X_4 and three rows Y1,Y2,Y3Y_1,Y_2,Y_3. Allow spectra consisting of any three XX rows together with one YY row, and also the spectrum consisting of all three YY rows. Allow subsets of these spectra as well.

No two allowed spectra cover all seven rows. Two spectra of the first kind cover at most two YY rows. A spectrum of the first kind together with the all-YY spectrum misses an XX row. Two all-YY spectra miss every XX row.

Proper-triple covering lemma. Suppose the four XX demands in a block are a1,…,a4≥0a_1,\ldots,a_4\ge0, and all three YY demands are b≥0b\ge0. Put

z=max⁡{max⁡iai, a1+a2+a3+a43}.(2)z=\max\left\{\max_i a_i,\ \frac{a_1+a_2+a_3+a_4}{3}\right\}. \tag{2}

The demands have a covering realization using the allowed spectra and total mass

max⁡{z, b+2z/3}.(3)\max\{z,\ b+2z/3\}. \tag{3}

Proof. Each z−aiz-a_i is nonnegative, and

∑i=14(z−ai)≥z.\sum_{i=1}^4(z-a_i)\ge z.

Choose wi≥0w_i\ge0 with ∑iwi=z\sum_iw_i=z and wi≤z−aiw_i\le z-a_i. Give the triple of XX rows omitting XiX_i total weight wiw_i, divided equally among its three possible accompanying YY singletons. Each XiX_i receives weight z−wi≥aiz-w_i\ge a_i, and each YY row receives z/3z/3. Add weight max⁡{b−z/3,0}\max\{b-z/3,0\} on the all-YY spectrum. The resulting mass is (3). The case z=0z=0 is included. □\square

In particular, four copies of a profile AA and three copies of BB are bad-realizable if

43A≤c,B+89A≤c.(4)\frac43A\le c, \qquad B+\frac89A\le c. \tag{4}

The sufficient conditions A≤3c/4A\le3c/4 and A+B≤cA+B\le c imply (4). The selection argument below needs the unequal-demand form (2), not just this special case.

Two coordinates meeting all heavy supports

Theorem. Suppose T(S;c)>3/4T(S;c)>3/4. If SS has a light profile, or if its heavy supports have a transversal of size at most two, then at most seven profiles from SS have a bad weighted realization.

The light-profile case

Take a light profile BB. Deleting 3B/43B/4 costs 3/43/4, so there is an A∈SA\in S with

A+3B/4≤c.A+3B/4\le c.

Place three copies of BB on a Fano line and four copies of AA on its complementary cap. In block jj, give this cap weight max⁡{Aj−3Bj/4,0}\max\{A_j-3B_j/4,0\} and each of the other six caps weight Bj/4B_j/4. These weights cover the prescribed demands, and their total is

max⁡{Aj+3Bj/4, 3Bj/2}≤cj.\max\{A_j+3B_j/4,\ 3B_j/2\}\le c_j.

Two Fano caps cannot cover all seven rows. Trimming therefore gives a bad realization. This is the same light-profile construction as in Full profile families.

The two-coordinate case

Assume that there is no light profile. If one coordinate gg is heavy for every profile, put h=min⁡A∈SAgh=\min_{A\in S}A_g. Deleting slightly more than cg−hc_g-h in that coordinate removes all profiles. But

cg−h<cg/3<1/2,c_g-h<c_g/3<1/2,

where cg<3/2c_g<3/2 follows from 1≥h>2cg/31\ge h>2c_g/3. This contradicts T(S;c)>3/4T(S;c)>3/4.

It remains to consider two distinct coordinates g1,g2g_1,g_2 meeting all the heavy supports. Partition SS into nonempty subfamilies S1,S2S_1,S_2 so that gig_i is heavy for every profile in SiS_i. Write

ti=T(Si;c),hi=min⁡A∈SiAgi,ei=cgi−hi.t_i=T(S_i;c),\qquad h_i=\min_{A\in S_i}A_{g_i},\qquad e_i=c_{g_i}-h_i.

Deleting slightly more than eie_i in coordinate gig_i removes SiS_i. Consequently

ti≤ei<cgi/3<1/2,hi>2ei≥2ti,cgi/3>ti.(5)t_i\le e_i<c_{g_i}/3<1/2, \qquad h_i>2e_i\ge2t_i, \qquad c_{g_i}/3>t_i. \tag{5}

Subadditivity and the hypothesis give t1+t2>3/4t_1+t_2>3/4. Relabel so that t1≥t2t_1\ge t_2. Then

t1>3/8,t2>1/4,t1+2t2>1.(6)t_1>3/8,\qquad t_2>1/4,\qquad t_1+2t_2>1. \tag{6}

Choose Q∈S1Q\in S_1. Since t1>1/3t_1>1/3, deletion of Q/3Q/3 leaves an A∈S1A\in S_1 with

A+Q/3≤c.(7)A+Q/3\le c. \tag{7}

For the four profiles Q,A,A,AQ,A,A,A, define

zj=max⁡{Qj, Aj+Qj/3}.(8)z_j=\max\{Q_j,\ A_j+Q_j/3\}. \tag{8}

This is exactly the quantity (2) for their four demands in block jj, and (7) gives z≤cz\le c. Every profile in S1S_1 has mass outside g1g_1 less than 1−2t11-2t_1, by (5). Since zj≤Qj+Ajz_j\le Q_j+A_j,

∑j≠g1zj<2(1−2t1).(9)\sum_{j\ne g_1}z_j<2(1-2t_1). \tag{9}

Delete 2zj/32z_j/3 in each coordinate j≠g1j\ne g_1, and delete nothing at g1g_1. This deletion is feasible, and its total mass is less than

43(1−2t1)<34−t1<t2.(10)\frac43(1-2t_1)<\frac34-t_1<t_2. \tag{10}

For the middle inequality, subtract its right-hand side from its left-hand side to obtain 7/12−5t1/3<07/12-5t_1/3<0, using t1>3/8t_1>3/8. Hence some B∈S2B\in S_2 satisfies

Bj+2zj/3≤cj(j≠g1).B_j+2z_j/3\le c_j\qquad(j\ne g_1).

The remaining coordinate also satisfies this inequality: because g1≠g2g_1\ne g_2, equations (5) and (6) give

Bg1<1−2t2<t1<cg1/3≤cg1−2zg1/3.(11)B_{g_1}<1-2t_2<t_1<c_{g_1}/3 \le c_{g_1}-2z_{g_1}/3. \tag{11}

Thus B+2z/3≤cB+2z/3\le c everywhere. Apply the proper-triple covering lemma in each block, with four XX profiles Q,A,A,AQ,A,A,A and three YY profiles B,B,BB,B,B. Its required mass is at most cjc_j by (3). Trim the excess incidences to finish the proof. □\square

No uniqueness assumption on heavy coordinates was used. A profile in SiS_i may also be heavy at other coordinates. In particular, this theorem applies to arbitrary capacities whenever there are at most two distinct profiles: choose one heavy coordinate of each nonlight profile.

Unequal capacities close to or above one

Continuous theorem. Let 0≤ϵ≤1/100\le\epsilon\le1/10, and suppose

cj≥1−ϵ(1≤j≤m),T(S;c)>34+mϵ.(12)c_j\ge1-\epsilon\quad(1\le j\le m), \qquad T(S;c)>\frac34+m\epsilon. \tag{12}

Then seven profiles from SS, with repetitions allowed, have a bad weighted realization.

Proof. A light profile has already been handled. Otherwise every profile has exactly one heavy coordinate: two such coordinates would have total demand greater than 4(1−ϵ)/3>14(1-\epsilon)/3>1. Call this coordinate its type.

For each represented type ii, choose a profile AiA^i minimizing its demand in coordinate ii. Put

hi=Aii,ei=ci−hi,δi=1−hi,D=1+2ϵ3.h_i=A^i_i,\qquad e_i=c_i-h_i,\qquad\delta_i=1-h_i, \qquad D=\frac{1+2\epsilon}{3}.

Every off-type coordinate of AiA^i is at most δi\delta_i. Moreover,

0≤ei<ci/3,ei<1/2,0≤δi<D,δi≤ei+ϵ,δi<1−2ei.(13)0\le e_i<c_i/3,\qquad e_i<1/2, \qquad 0\le\delta_i<D,\quad \delta_i\le e_i+\epsilon,\quad \delta_i<1-2e_i. \tag{13}

For example, hi>2ci/3h_i>2c_i/3 gives hi>2eih_i>2e_i and hence ei<1/2e_i<1/2, while ci≥1−ϵc_i\ge1-\epsilon gives δi≤ei+ϵ\delta_i\le e_i+\epsilon.

Deleting slightly more than eie_i in every represented type coordinate removes every profile. Therefore

∑i representedei≥T(S;c)>34+mϵ.(14)\sum_{i\text{ represented}}e_i \ge T(S;c)>\frac34+m\epsilon. \tag{14}

One or two represented types are covered by the preceding theorem. We henceforth assume that there are at least three.

We use the Fano lines

123,145,167,246,257,347,356.(15)123,\quad145,\quad167,\quad246,\quad257,\quad347,\quad356. \tag{15}

By the Fano-cap covering formula, seven demands q1,…,q7q_1,\ldots,q_7 in a block of capacity cc have a cap-spectrum covering precisely when

qi≤c,∑i∈Lqi≤2c(L a Fano line),∑iqi≤4c.(16)q_i\le c,\qquad \sum_{i\in L}q_i\le2c\quad(L\text{ a Fano line}), \qquad \sum_iq_i\le4c. \tag{16}

Individual demand bounds are automatic. Two estimates used repeatedly below are

3D≤2(1−ϵ),1+2D≤2(1−ϵ),(17)3D\le2(1-\epsilon),\qquad 1+2D\le2(1-\epsilon), \tag{17}

valid for ϵ≤1/10\epsilon\le1/10. Thus a line-coordinate check involving no own-type row, or only one own-type row, follows respectively from the bounds 3D3D and 1+2D1+2D.

Exactly three represented types

Order them so that e1≥e2≥e3e_1\ge e_2\ge e_3. Since m≥3m\ge3, equation (14) gives

e1+e2≥23(e1+e2+e3)>12+2ϵ.e_1+e_2\ge\frac23(e_1+e_2+e_3) >\frac12+2\epsilon.

It follows that

δ1<1−2e1<2e2−4ϵ,e2>2ϵ.(18)\delta_1<1-2e_1<2e_2-4\epsilon, \qquad e_2>2\epsilon. \tag{18}

Put type 1 on rows 2,3,42,3,4, type 2 on rows 5,6,75,6,7, and type 3 on row 11. Neither three-row type class is a Fano line. A line containing two type-1 rows has demand in coordinate 1 at most

2h1+max⁡{δ2,δ3}≤2h1+2e1=2c1,2h_1+\max\{\delta_2,\delta_3\}\le2h_1+2e_1=2c_1,

because δ2,δ3≤e1+ϵ≤2e1\delta_2,\delta_3\le e_1+\epsilon\le2e_1. A line containing two type-2 rows has demand in coordinate 2 at most

2h2+max⁡{δ1,δ3}≤2h2+2e2=2c2,2h_2+\max\{\delta_1,\delta_3\}\le2h_2+2e_2=2c_2,

using (18) and δ3≤e3+ϵ≤e2+ϵ≤2e2\delta_3\le e_3+\epsilon\le e_2+\epsilon\le2e_2. All other line inequalities follow from (17).

Write QiQ_i for the total demand in coordinate ii. For the two repeated types,

4c1−Q1≥h1+4e1−3δ2−δ3≥h1−4ϵ>0,4c2−Q2≥1−δ2+4e2−3δ1−δ3>2e2−4ϵ>0.(19)\begin{aligned} 4c_1-Q_1 &\ge h_1+4e_1-3\delta_2-\delta_3 \ge h_1-4\epsilon>0,\\ 4c_2-Q_2 &\ge 1-\delta_2+4e_2-3\delta_1-\delta_3 >2e_2-4\epsilon>0. \end{aligned} \tag{19}

The first strict inequality follows from h1>2(1−ϵ)/3>4ϵh_1>2(1-\epsilon)/3>4\epsilon. For the second line, use 3δ1<1+2ϵ3\delta_1<1+2\epsilon and δ2+δ3≤2e2+2ϵ\delta_2+\delta_3\le2e_2+2\epsilon.

Coordinate 3 has total demand less than

1+6D=3+4ϵ≤4(1−ϵ)≤4c3.1+6D=3+4\epsilon\le4(1-\epsilon)\le4c_3.

Every other coordinate has total demand less than

7D≤4(1−ϵ)≤4cj.7D\le4(1-\epsilon)\le4c_j.

Both displayed comparisons hold for ϵ≤1/10\epsilon\le1/10. This proves (16) in the three-type case.

At least four represented types

Order all represented types so that e1≥e2≥⋯e_1\ge e_2\ge\cdots. Then e2≥ϵe_2\ge\epsilon. Otherwise (13) and (14) would give the contradiction

T(S;c)≤∑iei<12+(m−1)ϵ<34+mϵ.T(S;c)\le\sum_i e_i <\frac12+(m-1)\epsilon<\frac34+m\epsilon.

Use type 1 on rows 1,2,41,2,4, type 2 on rows 3,53,5, type 3 on row 66, and type 4 on row 77. A line containing two type-1 rows has its third row of type 2 or 3, and hence its coordinate-1 demand is at most 2h1+2e1=2c12h_1+2e_1=2c_1. Indeed, δ2,δ3≤e1+ϵ≤2e1\delta_2,\delta_3\le e_1+\epsilon\le2e_1. The line through the two type-2 rows is 356356; its coordinate-2 demand is at most 2h2+δ3≤2h2+2e2=2c22h_2+\delta_3\le2h_2+2e_2=2c_2. Other line inequalities follow from (17).

The first two total demands satisfy

Q1≤3h1+2δ2+δ3+δ4≤3c1+e1+4ϵ≤4c1,Q2≤2h2+3δ1+δ3+δ4≤2c2+1+4ϵ≤4c2.(20)\begin{aligned} Q_1 &\le3h_1+2\delta_2+\delta_3+\delta_4 \le3c_1+e_1+4\epsilon\le4c_1,\\ Q_2 &\le2h_2+3\delta_1+\delta_3+\delta_4 \le2c_2+1+4\epsilon\le4c_2. \end{aligned} \tag{20}

For the final comparison in the first line, use e1<c1/3e_1<c_1/3 and 4ϵ≤2c1/34\epsilon\le2c_1/3. For the second line, use 3δ1<3D=1+2ϵ3\delta_1<3D=1+2\epsilon and 1+4ϵ≤2c21+4\epsilon\le2c_2. These inequalities hold since ci≥1−ϵc_i\ge1-\epsilon and ϵ≤1/10\epsilon\le1/10.

Each of coordinates 3 and 4 has total demand at most 1+6D≤4(1−ϵ)1+6D\le4(1-\epsilon). Every other coordinate, including any unselected represented type, has total demand at most 7D≤4(1−ϵ)7D\le4(1-\epsilon). Thus (16) holds in every block.

In both cases the Fano-cap formula supplies the covering realizations. Trimming excess incidences gives exact demands without a two-point cover. This proves the continuous theorem. □\square

Sparsifying a covering certificate

Lemma. Any finite weighted covering certificate on seven rows can be replaced by one using at most seven nonempty spectra, preserving every row's covered demand and not increasing total mass. Only spectra from the original certificate are needed.

Proof. Delete empty spectra and zero weights. If more than seven positive-weight spectra remain, their incidence columns in R7\mathbb R^7 have a nonzero linear dependence α\alpha. Because the columns are nonzero and nonnegative, α\alpha has both positive and negative entries. Change its sign if necessary so that ∑iαi≤0\sum_i\alpha_i\le0. Add a positive multiple of α\alpha to the weight vector, increasing the multiple until the first weight reaches zero. No weight becomes negative, row coverage is unchanged, and total mass does not increase. Delete a zero weight and repeat. □\square

This operation is performed before trimming excess incidences. Trimming may split a class into several smaller spectra, but the seven-class bound is needed only for rounding the untrimmed certificate.

Finite families on blocks of size at least the uniformity

Theorem. Let H\mathcal H be a nonempty full mm-block kk-uniform family, with every block of size at least kk. Suppose k≥70k\ge70 and every subfamily of at most seven edges has a two-point transversal. Then

τ(H)≤34k+15m.(21)\boxed{\tau(\mathcal H)\le\frac34k+15m.} \tag{21}

Proof. Suppose the inequality fails. Reserve seven points in each block, and let H′\mathcal H' consist of the edges avoiding all reserved points. A transversal of H′\mathcal H', together with the reserved points, is a transversal of H\mathcal H. Thus

τ(H′)≥τ(H)−7m>0.\tau(\mathcal H')\ge\tau(\mathcal H)-7m>0.

In particular, H′\mathcal H' is nonempty.

If the original block sizes are njn_j, normalize the profiles of H′\mathcal H' by kk and set

cj=nj−7k,ϵ=7k.c_j=\frac{n_j-7}{k},\qquad\epsilon=\frac7k.

Then cj≥1−ϵc_j\ge1-\epsilon and ϵ≤1/10\epsilon\le1/10. Any real deletion removing all normalized profiles can be rounded upward coordinatewise to an integer deletion in the unreserved blocks, increasing its size by less than mm. The full-profile transversal criterion from Block-profile convergence therefore gives

kT(S;c)≥τ(H′)−m≥τ(H)−8m>34k+7m.(22)kT(S;c)\ge\tau(\mathcal H')-m \ge\tau(\mathcal H)-8m >\frac34k+7m. \tag{22}

Consequently T(S;c)>3/4+mϵT(S;c)>3/4+m\epsilon, and the continuous theorem applies.

In each block, sparsify the resulting untrimmed covering certificate to at most seven spectra. Multiply its weights by kk and round each positive weight upward to an integer. Before rounding the total mass is at most nj−7n_j-7; the seven reserved points absorb the rounding increase. Every row has at least its prescribed integer block degree. Delete excess incidences independently in each row and block to recover the exact profile.

Full block invariance ensures that all seven resulting rows are edges of H\mathcal H. Their spectra are still subsets of allowed bad spectra, so no two points meet all rows. Removing repeated edges leaves a bad subfamily of at most seven distinct edges, a contradiction. □\square

The same bound holds under the exactly-seven convention. A family violating (21) has more than seven edges, so any smaller bad subfamily can be extended to seven distinct edges.

Scope

Taking ϵ=0\epsilon=0 in the continuous theorem handles arbitrary capacities cj≥1c_j\ge1, not just equal capacities. The finite theorem extends Equal blocks to unequal blocks of size at least the uniformity. It gives the conjectured asymptotic upper bound when the number of such blocks is o(k)o(k).

For arbitrary capacities, the two-heavy-coordinate theorem excludes every profile family whose heavy supports have a transversal of size at most two, as well as every family with a light profile. No theorem here handles all remaining heavy-support families. In particular, a heavy-support transversal number of at least three, with small block capacities, remains an unresolved case of the continuous-profile question.

All constructions require complete freedom to realize the selected profiles inside each block. There is no proved reduction of an arbitrary family to this setting preserving the local property with an asymptotically negligible loss.