Wiki
Wiki

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

Updated

The rational-rank defect at most one case


Statement

The following argument is proved. It extends the full-rational-rank case in the proof sketch, but does not resolve the general problem.

Let PP be a finite nonempty set in a finite-dimensional rational vector space. Assume that 0∉P0\notin P and that no two elements of PP are opposites. Write r=dim⁡Qspan⁡QPr=\dim_{\mathbb Q}\operatorname{span}_{\mathbb Q}P and let d=d(P)d=d(P) be the maximum size of a dissociated subset. If d≤r+1d\le r+1, then

∣P∣≤2d−1.(1)|P|\le 2^d-1. \tag{1}

If, additionally, the additive group L=⟨P⟩ZL=\langle P\rangle_{\mathbb Z} admits a homomorphism

χ:L⟶F2,χ(p)=1(p∈P),(2)\chi:L\longrightarrow\mathbb F_2, \qquad \chi(p)=1\quad(p\in P), \tag{2}

then

∣P∣≤2d−1.(3)|P|\le 2^{d-1}. \tag{3}

Positive sets satisfy the no-opposites hypothesis. Positive odd integer sets satisfy (2), using reduction modulo two. The intrinsic lift below makes these claims applicable beyond the original rational rank.

Maximum-volume coordinates and the integral columns

Choose a rational basis B={b1,…,br}⊆PB=\{b_1,\ldots,b_r\}\subseteq P whose determinant has maximum absolute value among all bases chosen from PP. Use BB as coordinate basis. Every coordinate of every p∈Pp\in P lies in [−1,1][-1,1]: replacing bib_i by pp multiplies the basis determinant by the ii-th coordinate of pp. Likewise, the absolute determinant of any rr coordinate columns from PP is at most one.

Let II be the elements with integral coordinates. Their coordinates belong to {−1,0,1}r\{-1,0,1\}^r. Reduction modulo two is injective on II. Indeed, two nonzero proportional ternary vectors are equal or opposite. Thus two distinct columns x,y∈Ix,y\in I are rationally independent. If their reductions modulo two were equal, extend x,yx,y to a rational basis using columns of BB. The resulting matrix is integral and has nonzero even determinant, contrary to the determinant bound. No nonzero ternary vector reduces to zero modulo two. Hence

∣I∣≤2r−1.(4)|I|\le 2^r-1. \tag{4}

Under (2), an integral-coordinate vector satisfies χ(∑izibi)=∑izi(mod2)\chi(\sum_i z_i b_i)=\sum_i z_i\pmod2. Every column in II therefore reduces to an odd-weight binary vector, giving

∣I∣≤2r−1.(5)|I|\le 2^{r-1}. \tag{5}

If d=rd=r, every p∈P∖Bp\in P\setminus B has ternary coordinates: the dependence of B∪{p}B\cup\{p\} must use pp, since BB is rationally independent. Thus P=IP=I, and (4)--(5) prove both claims in this case. It remains to consider d=r+1d=r+1.

A single signed nonintegral residue class

Put N=P∖IN=P\setminus I. If NN is empty, (4)--(5) already give the required bounds. Suppose that N≠∅N\ne\varnothing.

For distinct x,y∈Nx,y\in N, the set B∪{x,y}B\cup\{x,y\} has r+2>dr+2>d elements and is not dissociated. Any ternary relation must use both xx and yy: a relation using just one of them would make that column integral, and a relation using neither contradicts the independence of BB. Consequently

x−y∈{−1,0,1}rorx+y∈{−1,0,1}r.(6)x-y\in\{-1,0,1\}^r \quad\hbox{or}\quad x+y\in\{-1,0,1\}^r. \tag{6}

In particular, the residues of all columns of NN in Qr/Zr\mathbb Q^r/\mathbb Z^r belong to a single pair {g,−g}\{g,-g\}, where g≠0g\ne0.

Individual changes of sign preserve dissociation, preserve (2), and do not identify two of the columns because the original set has no opposites. Thus we may orient the columns of NN as convenient. The determinant bounds and (6) remain valid after these sign changes.

The case 2g≠02g\ne0

Orient every nonintegral column to have residue gg. For two such columns, their sum is not integral. Thus (6) says that all their differences are ternary. In each coordinate, all values have the same fractional part and any two differ by at most one. There are at most two values in each coordinate, so

∣N∣≤2r.(7)|N|\le2^r. \tag{7}

Combining (4) and (7) gives ∣P∣≤2r+1−1=2d−1|P|\le2^{r+1}-1=2^d-1.

Under (2), fix x0∈Nx_0\in N. For every x∈Nx\in N, the integral coordinate vector x−x0x-x_0 has

∑i(xi−(x0)i)≡χ(x)−χ(x0)=0(mod2).(8)\sum_i(x_i-(x_0)_i)\equiv\chi(x)-\chi(x_0)=0\pmod2. \tag{8}

Embed the available coordinate choices in a product of rr pairs of consecutive values. If a coordinate is constant, add an artificial second value differing by one. Exactly half of the 2r2^r vertices of this product satisfy (8). Therefore ∣N∣≤2r−1|N|\le2^{r-1}. Together with (5), this proves ∣P∣≤2r=2d−1|P|\le2^r=2^{d-1}.

The case 2g=02g=0

There are s≥1s\ge1 coordinates where the residue is 1/21/2, and z=r−sz=r-s coordinates where the residue is zero. The first coordinates have values ±1/2\pm1/2; the latter have values in {−1,0,1}\{-1,0,1\}. Orient each column so that its first half-integral coordinate is +1/2+1/2.

Fix a support SS among the zz integral coordinates. If two sign patterns on SS are neither equal nor opposite, they agree at a nonzero coordinate and disagree at another. Their column sum then has a coordinate ±2\pm2, and their column difference also has a coordinate ±2\pm2. This contradicts (6). Thus at most two opposite sign patterns occur for a nonempty support, and one occurs for the empty support. Each fixed integral pattern has at most 2s−12^{s-1} half-integral patterns. Consequently

∣N∣≤2s−1(1+2(2z−1))=2r−2s−1<2r.(9)|N|\le2^{s-1}\bigl(1+2(2^z-1)\bigr) =2^r-2^{s-1}<2^r. \tag{9}

Together with (4), this proves (1); in this subcase it gives the slightly stronger bound ∣P∣≤2r+1−2|P|\le2^{r+1}-2.

Now assume (2). Since 2x2x is an integral-coordinate vector and χ(2x)=0\chi(2x)=0, its coordinate sum is even. Modulo two, this sum is ss, so ss is even and hence s≥2s\ge2. The total coordinate sum

t(x)=∑i=1rxit(x)=\sum_{i=1}^r x_i

is therefore an integer. For any two columns, their difference is integral, and (2) gives t(x)−t(y)≡0(mod2)t(x)-t(y)\equiv0\pmod2. Thus all these total sums have the same parity. For a fixed integral-coordinate pattern, exactly half of the 2s−12^{s-1} possible half-integral patterns have the required total parity: flipping any one of the remaining s−1s-1 half-integral signs changes the total by one. It follows that

∣N∣≤2s−2(1+2(2z−1))=2r−1−2s−2<2r−1.(10)|N|\le2^{s-2}\bigl(1+2(2^z-1)\bigr) =2^{r-1}-2^{s-2}<2^{r-1}. \tag{10}

Combining this with (5) proves (3).

Intrinsic ternary-relation rank

The rank hypothesis can be strengthened without changing any ternary relations. For a finite positive real set PP, take the vector space QP\mathbb Q^P with coordinate vectors epe_p. Let

T=span⁡Q{(εp)p∈P∈{−1,0,1}P:∑p∈Pεpp=0},V=QP/T,ρ=dim⁡QV.T=\operatorname{span}_{\mathbb Q} \left\{(\varepsilon_p)_{p\in P}\in\{-1,0,1\}^P: \sum_{p\in P}\varepsilon_p p=0\right\}, \qquad V=\mathbb Q^P/T, \qquad \rho=\dim_{\mathbb Q}V.

Call ρ\rho the intrinsic ternary-relation rank in this note. The lifted configuration is P~={ep+T:p∈P}⊆V\widetilde P=\{e_p+T:p\in P\}\subseteq V.

The original evaluation functional descends to a rational-linear map f:V→Rf:V\to\mathbb R, with f(ep+T)=p>0f(e_p+T)=p>0. In particular, the lifted columns are distinct, nonzero, and have no antipodal pair. A ternary vector vanishes in VV if and only if it gives a zero relation on the original set. One direction holds because every original ternary relation was included in the definition of TT; the reverse holds because TT is contained in the kernel of the original evaluation. Hence

d(P~)=d(P).(11)d(\widetilde P)=d(P). \tag{11}

For a maximum dissociated D⊆PD\subseteq P, every p∈P∖Dp\in P\setminus D has a ternary representation on DD. These relations show that the images of DD span VV. Thus

dim⁡Qspan⁡QP≤ρ≤d(P).(12)\dim_{\mathbb Q}\operatorname{span}_{\mathbb Q}P \le\rho\le d(P). \tag{12}

Applying the proved vector-space theorem to P~\widetilde P establishes the target whenever

ρ≥d(P)−1.(13)\rho\ge d(P)-1. \tag{13}

If the original elements are odd integers, the parity homomorphism also survives this lift. Indeed, an integral vector (zp)∈T∩ZP(z_p)\in T\cap\mathbb Z^P has ∑pzpp=0\sum_p z_p p=0, and reduction modulo two gives ∑pzp≡0(mod2)\sum_p z_p\equiv0\pmod2. Therefore the map sending every lifted generator to one descends to ⟨P~⟩Z→F2\langle\widetilde P\rangle_{\mathbb Z}\to\mathbb F_2. Under (13), the odd-slice bound follows:

∣P∣≤2d(P)−1.|P|\le2^{d(P)-1}.

Remaining limitation

For rank defect d−r=t≥2d-r=t\ge2, the residue configuration in Qr/Zr\mathbb Q^r/\mathbb Z^r has dissociation dimension at most tt: a dissociated set of t+1t+1 residues, together with BB, would give d+1d+1 dissociated columns. However, a dependence among t+1t+1 nonintegral columns need not imply any pair relation of the form (6). The coordinate counting above depends essentially on that pair relation. No general rank-defect induction, and no proof for intrinsic defect at least two, is established here.