Wiki
Wiki

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

Updated


Statement

Let kk be a positive prime power and put m=4km=4k. There is a finite set Xm⊂RdmX_m\subset\mathbb R^{d_m}, where

dm=(m2)−1,d_m=\binom m2-1,

with the following properties:

  1. ∣Xm∣=12(mm/2)|X_m|=\frac12\binom m{m/2};
  2. after a common rescaling, diam⁡(Xm)=1\operatorname{diam}(X_m)=1; and
  3. every subset of XmX_m having diameter strictly less than 11 contains at most 2(m−1m/4−1)2\binom{m-1}{m/4-1} points.

Consequently, if f(d)f(d) denotes the Borsuk partition function, then

f(dm)≥12(mm/2)2(m−1m/4−1)=(mm/2)(mm/4).f(d_m)\ge \frac{\frac12\binom m{m/2}} {2\binom{m-1}{m/4-1}} =\frac{\binom m{m/2}}{\binom m{m/4}}.

Construction and proof

Let V=[m]V=[m] and let W=(V2)W=\binom V2 be the edge set of the complete graph on VV. For every unordered bipartition P={A,B}P=\{A,B\} with ∣A∣=∣B∣=m/2=2k|A|=|B|=m/2=2k, define

SP=S(A,B):={{a,b}∈W:a∈A, b∈B}.S_P=S(A,B):=\{\{a,b\}\in W:a\in A,\ b\in B\}.

Thus SPS_P is the edge set of the complete bipartite graph across the cut. The unordered cut has exactly two ordered descriptions, (A,B)(A,B) and (B,A)(B,A), and its crossing-edge set determines those two sides: for each vertex, its non-neighbors in SPS_P are exactly the other vertices on its side. Hence the family K\mathcal K of all such SPS_P has

∣K∣=12(mm/2).|\mathcal K|=\frac12\binom m{m/2}.

Every member has size ∣SP∣=∣A∣∣B∣=m2/4|S_P|=|A||B|=m^2/4. Represent SPS_P by its incidence vector xP∈{0,1}Wx_P\in\{0,1\}^{W}. All these vectors lie in the affine hyperplane

∑e∈Wxe=m2/4,\sum_{e\in W}x_e=m^2/4,

whose dimension is ∣W∣−1=(m2)−1=dm|W|-1=\binom m2-1=d_m. An affine isometry identifies this hyperplane with Rdm\mathbb R^{d_m}.

Consider two cuts P={A,B}P=\{A,B\} and Q={C,D}Q=\{C,D\}. After choosing the labels, write r=∣A∩C∣r=|A\cap C|. The four cells determined by the two bipartitions have sizes

r,2k−r,2k−r,r.r,\quad 2k-r,\quad 2k-r,\quad r.

An edge crosses both cuts precisely when its endpoints lie in the two opposite cells of one of the two matching pairs. Hence

∣SP∩SQ∣=r2+(2k−r)2=2(r−k)2+2k2.|S_P\cap S_Q| =r^2+(2k-r)^2 =2(r-k)^2+2k^2.

The minimum possible intersection is therefore 2k2=m2/82k^2=m^2/8, attained exactly when r=kr=k; such distinct cuts exist by choosing two 2k2k-sets with intersection kk. Since all incidence vectors have the same weight,

∥xP−xQ∥22=∣SP△SQ∣=m22−2∣SP∩SQ∣.\|x_P-x_Q\|_2^2 =|S_P\mathbin\triangle S_Q| =\frac{m^2}{2}-2|S_P\cap S_Q|.

Thus the minimum intersection corresponds to the maximum squared distance m2/4m^2/4, and this value is attained. The diameter of the incidence configuration is m/2m/2; multiplying every vector by 2/m2/m makes the diameter one without changing which pairs attain it.

It remains to bound a subset L⊆K\mathcal L\subseteq\mathcal K of diameter strictly smaller than the full configuration. Choose one side APA_P of each cut P∈LP\in\mathcal L. If distinct chosen sides AP,AQA_P,A_Q had ∣AP∩AQ∣=k=m/4|A_P\cap A_Q|=k=m/4, then the displayed calculation would put xP,xQx_P,x_Q at the diameter, a contradiction. The chosen sides therefore form a family of m/2m/2-subsets with the intersection size m/4m/4 forbidden. Applying the exact Frankl–Wilson interface gives

∣L∣≤2(m−1m/4−1).|\mathcal L|\le 2\binom{m-1}{m/4-1}.

If a partition of XmX_m has qq parts of smaller diameter, counting points in those parts gives

12(mm/2)≤q 2(m−1m/4−1).\frac12\binom m{m/2} \le q\,2\binom{m-1}{m/4-1}.

Finally,

(m−1m/4−1)=14(mm/4),\binom{m-1}{m/4-1}=\frac14\binom m{m/4},

which proves the asserted ratio.

Source wording

This is a complete expansion of Section 2 on physical PDF p. 2 (journal p. 61). The printed sentence says that two sets with minimum intersection “realize the minimal distance.” The incidence-vector identity above shows that they realize the maximal distance; the next printed sentence also uses the minimum-intersection condition. The proof here follows the displayed construction and records that one-word source defect explicitly.