Wiki
Wiki

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

Updated

Dimension increase in odd reflected configurations


These are special augmentation results, not a proof of the target. They differ from the refuted assertion in the reflection obstruction: here the center itself, not twice the center, is included, and all points are odd.

Let cc be a positive odd integer and let XX be a dissociated set of distinct positive even integers smaller than cc. Write

Q={c}∪{c−x,c+x:x∈X}.Q=\{c\}\cup\{c-x,c+x:x\in X\}.

A proved orientation for at most seven differences

Theorem. If ∣X∣≤7|X|\le7, there are signs σx\sigma_x such that

{c}∪{c+σxx:x∈X}\{c\}\cup\{c+\sigma_xx:x\in X\}

is dissociated. In particular d(Q)≥∣X∣+1d(Q)\ge |X|+1.

A balanced relation among any oriented selection gives a ternary relation on XX, so is impossible. For the all-upper selection, every point is in [c,2c)[c,2c). Odd parity requires even relation support. An unbalanced relation supported on at most six points has at least four terms against at most two, or at least three against one; the interval bounds rule it out. This proves the theorem for ∣X∣≤6|X|\le6.

Let ∣X∣=7|X|=7. If the all-upper selection is dependent, its relation must use all eight points and have five terms against three. The pivot cc lies on the five-term side: if it lay on the three-term side, that side would be less than 5c5c, whereas the other side would be at least 5c5c. We therefore have a partition

X=A⊔B,∣A∣=3,∣B∣=4,∑A−∑B=2c.X=A\sqcup B,\qquad |A|=3,\quad |B|=4,\qquad \sum A-\sum B=2c.

Put S=∑BS=\sum B. Since the three members of AA are less than cc, we have S<cS<c, and every a∈Aa\in A exceeds SS.

Flip only b=min⁡Bb=\min B: set σb=−1\sigma_b=-1 and all other signs to 11. Suppose the resulting selection has a relation. Write it as

sc+∑x∈Xrxx=0,s=α0+∑xrxσx,sc+\sum_{x\in X}r_xx=0, \qquad s=\alpha_0+\sum_xr_x\sigma_x,

where rx,α0∈{−1,0,1}r_x,\alpha_0\in\{-1,0,1\}. Oddness makes ss even; s=0s=0 is excluded by dissociation of XX. Also ∑X=2c+2S<4c\sum X=2c+2S<4c. Reversing the relation if needed gives

∑xrxx=2c,∑xrxσx∈{−3,−2,−1}.(1)\sum_xr_xx=2c, \qquad \sum_xr_x\sigma_x\in\{-3,-2,-1\}. \tag{1}

At least two members of AA have coefficient +1+1, since one of them together with all of BB has sum less than c+S<2cc+S<2c. If all three do, every BB-coefficient must be −1-1. The signed coefficient sum in (1) is then 11, impossible. If two do and the third aa has coefficient −1-1, the total is at most 2c+2S−2a<2c2c+2S-2a<2c. Hence precisely two have coefficient +1+1 and the third aa is omitted. The equation becomes

a=S+∑u∈Bruu>S.(2)a=S+\sum_{u\in B}r_u u>S. \tag{2}

Let u,vu,v be the numbers of positive and negative coefficients on BB. There is at least one positive coefficient. If rb≤0r_b\le0, counting the other three coefficients gives ∑xrxσx≥1\sum_xr_x\sigma_x\ge1, contrary to (1). If rb=1r_b=1, that sum equals u−vu-v. To be negative, with u+v≤4u+v\le4, requires (u,v)=(1,2)(u,v)=(1,2) or (1,3)(1,3). Equation (2) then gives, respectively, a=2b+qa=2b+q for the omitted q∈B∖{b}q\in B\setminus\{b\}, or a=2ba=2b. Both are less than SS, since bb is the smallest of four distinct positive numbers. This contradicts a>Sa>S and proves the theorem.

Thus a centrally symmetric odd set containing its odd center has dimension at least d(X)+1d(X)+1 whenever its positive difference set has dimension at most seven. The unrestricted version is unproved.

Increasing dimension by two is false in general

Take X={2,4,8,16}X=\{2,4,8,16\} and c=17c=17. The set XX is dissociated, but

Q={1,9,13,15,17,19,21,25,33},d(Q)=5.Q=\{1,9,13,15,17,19,21,25,33\},\qquad d(Q)=5.

Thus the tempting stronger bound d(Q)≥d(X)+2d(Q)\ge d(X)+2 is false.

The five-set {1,9,13,15,17}\{1,9,13,15,17\} is dissociated: its ten pair sums are distinct, and its smallest triple sum exceeds 1717. Parity excludes every relation except a two-term equality, a pair-sum equality, or a triple equaling a singleton.

Every six-subset of QQ containing two complete reflected pairs is dependent, since both pairs sum to 3434. Otherwise it consists of cc, one complete pair {c−x,c+x}\{c-x,c+x\}, and one point from each other pair. If x∈{2,4,8}x\in\{2,4,8\}, then 2x∈X2x\in X, and the selected point c±2xc\pm2x, together with the complete pair and cc, gives a balanced four-term relation.

It remains to consider the complete pair {1,33}\{1,33\}. The other three choices are a∈{15,19}a\in\{15,19\}, b∈{13,21}b\in\{13,21\} and z∈{9,25}z\in\{9,25\}. If a=15a=15, use 1+15+17=331+15+17=33; if a=19,b=13a=19,b=13, use 1+13+19=331+13+19=33. In the remaining cases use

1+17+19+21=25+33,1+9+19+21=17+33.1+17+19+21=25+33,\qquad 1+9+19+21=17+33.

This covers every six-subset and proves the claimed dimension.

The obstruction persists in every dimension k≥4k\ge4. Set

Xk={2,4,…,2k},c=2k+1,Qk={c}∪(c+Xk)∪(c−Xk).X_k=\{2,4,\ldots,2^k\},\quad c=2^k+1,\quad Q_k=\{c\}\cup(c+X_k)\cup(c-X_k).

Then d(Qk)=k+1d(Q_k)=k+1. The all-upper selection with cc is dissociated: in a relation its coefficient sum at cc is even, whereas ∑Xk=2k+1−2<2c\sum X_k=2^{k+1}-2<2c, so it must be balanced and then trivial. To exclude (k+2)(k+2)-subsets, two complete pairs give an immediate relation; otherwise the set has the center, one complete pair and one choice from every other pair. Any nonlargest pair is blocked by its doubled offset. For the largest pair {1,2c−1}\{1,2c-1\}, choices at offsets 2,4,82,4,8 suffice. Successively according to those choices, use

1+(c−2)+c=2c−1,1+(c+2)+(c−4)=2c−1,1+(c+2)+(c+4)+(c−8)=c+(2c−1),1+c+(c+2)+(c+4)=(c+8)+(2c−1).\begin{split} 1+(c-2)+c&=2c-1,\\ 1+(c+2)+(c-4)&=2c-1,\\ 1+(c+2)+(c+4)+(c-8)&=c+(2c-1),\\ 1+c+(c+2)+(c+4)&=(c+8)+(2c-1). \end{split}

Thus a general reflection theorem could guarantee at most one additional dimension.

A two-dimension increase in an affine special case

Call a set affinely dissociated if its only balanced ternary relation is trivial. If a dissociated integer set XX has all elements of the same 22-adic valuation, then

{0}∪X∪(−X)contains an affinely dissociated (∣X∣+2)-set.(3)\{0\}\cup X\cup(-X) \quad\hbox{contains an affinely dissociated }(|X|+2)\hbox{-set}. \tag{3}

Scale so the entries are odd, fix a pivot xjx_j, and, for i≠ji\ne j, choose σi∈{±1}\sigma_i\in\{\pm1\} with σi≡xi(mod4)\sigma_i\equiv x_i\pmod4. Use

Y={0,−xj,xj}∪{σixi:i≠j}.Y=\{0,-x_j,x_j\}\cup\{\sigma_i x_i:i\ne j\}.

A balanced ternary relation not assigning opposite nonzero coefficients to −xj,xj-x_j,x_j reduces to a ternary relation on XX. If those two coefficients are equal, cancellation in XX would instead require a coefficient of absolute value two at 00; this is also impossible. The only remaining possibility would give

2xj=∑i≠jεixi,∣∑i≠jεiσi∣≤1.2x_j=\sum_{i\ne j}\varepsilon_i x_i, \qquad \left|\sum_{i\ne j}\varepsilon_i\sigma_i\right|\le1.

But the latter sum is 2xj≡2(mod4)2x_j\equiv2\pmod4, a contradiction. This proves (3).

For the odd reflected configuration above, (3) gives an actual dissociated (∣X∣+2)(|X|+2)-subset if, additionally,

2c>∑X+max⁡X.2c>\sum X+\max X.

Indeed, every relation on the translated selection has an even coefficient sum at cc. Its offset part has absolute value at most ∑X+max⁡X\sum X+\max X, so the displayed inequality forces balance.

Statement (3) for mixed valuation layers remains unproved. The counterexample above concerns actual, not affine, dissociation and does not refute it. It does refute a two-dimension increase for arbitrary c>max⁡Xc>\max X.

There is also a valuation-free small affine case: (3) holds for every positive dissociated XX with ∣X∣≤5|X|\le5, using Y={0,−M,M}∪(X∖{M})Y=\{0,-M,M\}\cup(X\setminus\{M\}) and M=max⁡XM=\max X. The only possible affine relation would give 2M=∑x≠Mεxx2M=\sum_{x\ne M}\varepsilon_xx with ∣∑εx∣≤1|\sum\varepsilon_x|\le1. Among at most four other coordinates this permits at most two positive coefficients, whose weighted sum is strictly less than 2M2M. This is impossible.

Neither an unrestricted largest-pivot rule nor a largest-valuation rule can extend this proof. The dissociated set {1,2,8,15}\{1,2,8,15\} has its maximum-valuation pivot 88 blocked by 16=1+15=−1+2+1516=1+15=-1+2+15: avoiding a signed coefficient sum of zero in the first representation forces the signs on 1,151,15 to agree, and the second then has coefficient sum ±1\pm1.

The largest pivot is blocked in X={8,20,36,40,42,45}X=\{8,20,36,40,42,45\} by

90=40+42+8=36+42+20−8=36+40+42−8−20.90=40+42+8=36+42+20-8=36+40+42-8-20.

The first representation forces the signs on 40,42,840,42,8 to agree; the second then forces those on 36,2036,20 to agree; the last has signed coefficient sum ±1\pm1. This XX is dissociated: successive valuation arguments eliminate 4545 and 4242, leaving four times the dissociated set {2,5,9,10}\{2,5,9,10\}, whose subset sums are 0,2,5,7,9,10,11,12,14,15,16,17,19,21,24,260,2,5,7,9,10,11,12,14,15,16,17,19,21,24,26. These examples leave free choice of the pivot open.