Wiki
Wiki

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

Updated

Eight real points contain an affinely dissociated five-set


Write daff(X)d_{\rm aff}(X) for the largest size of a subset with no nonzero balanced ternary relation. This parameter concerns distinct equal-cardinality subset sums. It is not ordinary dissociation; the distinction is essential below.

The eight-point theorem

Theorem. Every eight distinct real numbers contain five whose restricted pair sums are distinct. Equivalently, they contain an affinely dissociated five-set.

Form the integer matrix whose rows encode all equalities xi+xj=xk+xlx_i+x_j=x_k+x_l on four distinct elements. Its real kernel contains the independent vectors 1\mathbf1 and (xi)(x_i), so its rank is at most six. Reduction modulo two cannot increase rank. There is therefore a nonconstant two-coloring such that every additive quadruple meets each color evenly.

Suppose there is no desired five-set. A four-set with distinct restricted pair sums in either color could be extended by any point of the other color: a new additive quadruple would meet that color exactly once. Thus neither color has such a four-set. Every five distinct reals do have one: for x1<⋯<x5x_1<\cdots<x_5, at most one of the quadruples (x1,x2,x3,x4)(x_1,x_2,x_3,x_4) and (x1,x2,x3,x5)(x_1,x_2,x_3,x_5) can satisfy its sole possible pair-sum equality. Hence both color classes have exactly four elements, each a parallelogram:

T={t1<t2<t3<t4},t1+t4=t2+t3,T=\{t_1<t_2<t_3<t_4\},\quad t_1+t_4=t_2+t_3,

and similarly U={u1<u2<u3<u4}U=\{u_1<u_2<u_3<u_4\}.

Assume t4−t1≥u4−u1t_4-t_1\ge u_4-u_1. Seek a five-set consisting of t1,t4t_1,t_4 and three points of UU. The coloring excludes all 3+13+1 additive quadruples. A 2+22+2 quadruple is possible only at a pair of UU satisfying

ui+uj=t1+t4oruj−ui=t4−t1.(1)u_i+u_j=t_1+t_4\quad\hbox{or}\quad u_j-u_i=t_4-t_1. \tag{1}

If the parallelogram centers differ, the sum condition gives at most one pair and never the inner pair {u2,u3}\{u_2,u_3\}. The difference condition gives at most the outer pair {u1,u4}\{u_1,u_4\}, and only when diameters agree. If both offending pairs occur, they intersect. Deleting a common endpoint leaves the required three points, a contradiction.

If the centers agree, translate them to zero and write the eight points as {±a,±b,±c,±d}\{\pm a,\pm b,\pm c,\pm d\} with 0<a<b<c<d0<a<b<c<d. The set {−d,d,a,b,c}\{-d,d,a,b,c\} has distinct restricted pair sums unless d=b+c−ad=b+c-a. In this exceptional case use {−d,d,−a,b,c}\{-d,d,-a,b,c\}. A quadruple containing both −d,d-d,d cannot have an equality. The other two quadruples would require d=a+c−bd=a+c-b or d=a+b+cd=a+b+c, both impossible under the exceptional identity. This proves the theorem.

On five distinct points a nonzero balanced ternary relation must have support four (support two would equate two points), proving the stated equivalence.

A binary split for affine dimension

If ∣X∣≥2|X|\ge2 and daff(X)=dd_{\rm aff}(X)=d, there is a partition into two nonempty sets, each of affine dimension at most d−1d-1. Use all balanced ternary relations of support at most d+1d+1 as rows of an integer matrix. Again the real kernel contains 1\mathbf1 and the distinct weight vector, so its binary kernel has dimension at least two. Choose a nonconstant binary kernel vector to color the points.

If one color contained an affinely dissociated dd-set, adjoining a point of the other would produce a balanced relation of support at most d+1d+1 using that point. The coloring contradicts the relation. For the color-one class, use the complementary coloring, which also annihilates the even-support rows.

The eight-point theorem and induction on this split give

daff(X)=d≥4⟹∣X∣≤7 2d−4.(2)\boxed{d_{\rm aff}(X)=d\ge4 \quad\Longrightarrow\quad |X|\le7\,2^{d-4}.} \tag{2}

This improves the old affine count 2d−12^{d-1} by a factor 7/87/8. It makes no unrestricted ordinary-dissociation claim.

Consequences for unit classes

If a positive integer set lies in a unit class modulo qq and its ordinary dimension dd satisfies d≤qd\le q, then daff=dd_{\rm aff}=d. Indeed, on a set of size at most d+1d+1, a nonzero relation has both signs by positivity, so its coefficient sum has magnitude at most d−1<qd-1<q. Congruence forces that sum to be zero. A larger affine-independent set would therefore give an ordinary-independent (d+1)(d+1)-set.

For even q≥4q\ge4, the proof of the one-more-dimension theorem propagates the same factor through d=q+1d=q+1. In its nontrivial cases the two remaining classes have dimension at most d−1d-1, giving 2⋅7 2d−5=7 2d−42\cdot7\,2^{d-5}=7\,2^{d-4}. The two-point exception instead gives 7 2d−5+3≤7 2d−47\,2^{d-5}+3\le7\,2^{d-4}; an empty fiber is handled directly modulo 2q2q. Thus

q≥4 even,4≤d≤q+1⟹∣O∣≤7 2d−4(3)\boxed{q\ge4\text{ even},\quad4\le d\le q+1 \quad\Longrightarrow\quad |O|\le7\,2^{d-4}} \tag{3}

for positive OO in one unit class modulo qq.

In particular an arbitrary positive odd set of ordinary dimension at most five has at most 2828 elements, by its two modulo-four classes. The known smaller odd bounds and valuation-layer augmentation then give

d(P)≤5⟹∣P∣≤28+8+4+2+1=43d(P)\le5\Longrightarrow |P|\le28+8+4+2+1=43

for positive sets, and ∣A∣≤87|A|\le87 for arbitrary real sets of dimension at most five. This is still short of the requested positive bound 3131, or real bound 6363.

Two further congruence endpoints

For even q≥6q\ge6, the weaker count ∣O∣≤2d−1|O|\le2^{d-1} also holds when d=q+2d=q+2. Split modulo 2q2q as O=A⊔BO=A\sqcup B. If both fibers have dimension at most d−1d-1, (3) applies to each. Otherwise choose a dissociated dd-set D={x1<⋯<xd=M}⊂AD=\{x_1<\cdots<x_d=M\}\subset A, and put S=∑DS=\sum D. Every b∈Bb\in B has a ternary representation on DD whose coefficient sum is d−1d-1 or 3−d3-d. Positivity leaves only

b=S−xiorb=2xj+xi−S(i≠j).(4)b=S-x_i\quad\hbox{or}\quad b=2x_j+x_i-S\quad(i\ne j). \tag{4}

There are at most dd values of the first kind. A positive value of the second kind requires j=dj=d, or i=di=d and j=d−1j=d-1: if neither index is dd, the sum of the omitted positive terms contains M>xjM>x_j; if i=di=d and j<d−1j<d-1, it contains xd−1>xjx_{d-1}>x_j. There are at most dd such values. Thus ∣B∣≤2d|B|\le2d.

Since d−1<2qd-1<2q, the full fiber has affine dimension dd and ∣A∣≤7 2d−4|A|\le7\,2^{d-4} by (2). Therefore

∣O∣≤7 2d−4+2d≤2d−1(d≥8).(5)|O|\le7\,2^{d-4}+2d\le2^{d-1}\qquad(d\ge8). \tag{5}

The endpoint q=4,d=6q=4,d=6 is outside this argument except when the opposite fiber has at most four points. The subsequent polynomial grid proof settles that endpoint in full, with the stronger bound 3131.

There is a further endpoint: for even q≥10q\ge10, the count holds through d=q+3d=q+3. If a full dd-set lies in one modulo-2q2q fiber, the possible coefficient sums for the opposite fiber are d−2d-2 and 4−d4-d. For d−2d-2, the representation has either two zeros and otherwise plus signs, or one minus sign and otherwise plus signs, giving at most (d2)+d\binom d2+d values. For 4−d4-d, positivity allows one positive coefficient and two zeros, or two positive coefficients with full support. For each omitted pair there is at most one positive dominant coordinate, giving at most (d2)\binom d2 values of the former type. The positive pairs of the latter type have sum greater than S/2S/2; any two such pairs intersect, so there are at most d−1d-1 of them. Consequently the opposite fiber has at most d2+d−1d^2+d-1 values. For d≥13d\ge13,

7 2d−4+d2+d−1≤2d−1.7\,2^{d-4}+d^2+d-1\le2^{d-1}.

If neither fiber is full-dimensional, apply the already established q+2q+2 endpoint to both. These refinements do not handle general odd sets or establish the logarithmic target.

A proportional congruence range

Corollary. For every even q≥56q\ge56, a positive integer set in one unit class modulo qq, of dimension d≤⌊4q/3⌋d\le\lfloor4q/3\rfloor, has at most 2d−12^{d-1} elements.

Proceed by induction on dd at fixed qq, starting with the proved range d≤q+1d\le q+1. Split modulo 2q2q. If both parts have dimension at most d−1d-1, induction applies. Otherwise take a maximum dd-set DD in one fiber AA, and write d=q+kd=q+k. The other fiber's ternary representations on DD have coefficient sum q+1q+1 or 1−q1-q, since d<2qd<2q. With

Cd,j=[xj](1+x+x2)d,C_{d,j}=[x^j](1+x+x^2)^d,

the numbers of such coefficient vectors are respectively Cd,k−1C_{d,k-1} and Cd,k+1C_{d,k+1}, by measuring deviations from all plus or all minus signs. Thus

∣B∣≤Cd,k−1+Cd,k+1.|B|\le C_{d,k-1}+C_{d,k+1}.

The condition d≤4q/3d\le4q/3 gives k≤d/4k\le d/4. Evaluation at x=1/4x=1/4, using nonnegative coefficients, yields

Cd,k−1+Cd,k+1≤174(21216)d≤2d−4(d≥58).C_{d,k-1}+C_{d,k+1} \le\frac{17}{4}\left(\frac{21\sqrt2}{16}\right)^d \le2^{d-4}\qquad(d\ge58).

For the last inequality, its square at d=58d=58 is the exact integer inequality 4624⋅44158<512584624\cdot441^{58}<512^{58}, and the ratio decreases with dd. Here d≥q+2≥58d\ge q+2\ge58.

Since d−1<2qd-1<2q, the full fiber has affine dimension dd and at most 7 2d−47\,2^{d-4} points. The remaining 2d−42^{d-4} slack absorbs BB, completing the induction. This supplies a range of unit congruence classes, not the unrestricted odd-set bound.