ON THE NUMBER OF HALVING PLANES

EXTENDED ABSTRACT

IMRE BARANY’, ZOLTAN F~~REDI* AND LAsz~6 Loviszt

* Mathematical Institute of the Hungarian Academy of Sciences,
1364 Budapest, P. 0. B. 127, Hungary
and
t Department of Computer Science,
EGtvijs University, 1088 Budapest, Mtizeum krt. 6-8, Hungary, and
Princeton University, Princeton, NJ 08544, USA

Research supported partly by the Hungarian National Science Foundation grant No.
1812

Abstract. Let S C R3 be an n-set in general position. A plane con-

taining three of the points is. called a halving plane if it dissects S into
two parts of equal cardinelity. It is proved that the number of halving
planes is at most O(na-ggs).
As a main tool, for every set Y of n points in the plane a set N of size
O(n’) is constructed such that the points of N are distributed almost
evenly in the triangles determined by Y.

1. HALVING PLANES Clearly, h(n) 5 (J). Moreover, h(n) 2 i(;), as any two
points are contained in a halving plane. A slightly bet-
A point-set S c Rd is in general position if no d + 1 ter lower bound is known [Ed] that h(n) = R(n*logn).
points of it lie in a hyperplane. The plane determined by The aim of this paper is to improve the trivial upper
the non-collinear points a, b, c is denoted by P(a, b, c). bound:
In general, the afflne subspace spanned by the set A is
denoted by aff(A). As usual, conv(A) stands for the THEOREM 1.1. h(n) = O(n2.ssa).
convex hull of A.
Assume that S is an n-element point-set in the three- The proof, which is postponed to Section 7, is similar
dimensional Euclidean space in general position. A to the proof of the a-dimensional case given in [L], but
plane P(u, b, c), where a, b, c E S, is called a halving the crucial step requires new tools (Theorem 2.1). (We
plane if it dissects S into two equal parts, that is, on will prove h(n) 5 O(n3-“) with a = l/343: with more
both sides of P there are exactly $(n - 3) points of S. effort, one could prove the result with a = l/64.)
Denote the number of halving planes by h(S), and set 2. COVERING TRIANGLES BY CROSSINGS
h(n) =max{h(S) : S c R3, ISI = n, A point-set S in Rd is said to be in totally geneml
position if
S is in general position}.
dim(f$=, aff Ai) 2 max{-1, Cdim(Ai) - (S - l)d}

Permission to copy without fee all or part of this material is granted pro-
vided that the copies are not made or distributed for direct commercial
advantage, the ACM copyright notice and the title of the publication and
its date appear, and notice is given that copying is by permission of the
Association for Computing Machinery. To copy otherwise, or to
republish, requires a fee and/or specific permission.
0 1989 ACM 0-89791-318-3/89/0006/0140 $1.50
holds for all subsets Ai c S. Prom now on we always
suppose, if it is not otherwise stated, that the (finite)
point-sets are in totally general position. A set F couers
t triangles from the set Y c R2 if at least t open tri-
angles (~1, ~2, ~3) ( w h ere yi E Y) contain a point of F.
Obviously, no set can cover more than (1:‘) triangles.

140
THEOREM2.1. For every n element set Y c R2 there
exists a set F with JF( < no.ggs which covers all but at
most O(n2.gga) triangles from Y.
Two lines determined by four distinct points of Y in-
tersect in a crossing. Define C(Y) as the set of crossings.
We have IC(Y)I = $(g)(“T”) = O(n*). Let N(R) de-
note the number of crossings in the interior of the region
R, and N(abc) = N(conv(a, b, c)).
It is perhaps instructive to show at this step that the
average number of crossings in a triangle with vertices
from Y is s2(n4). Our first observation is that every set
of nine points, E c Y, contains a triangle such that at
least one of the crossings defined by four of the remain-
ing 6 points lies inside the triangle. Indeed, a theorem
of Tverberg [T] (cf. also Reay [R]) states that there is
a partition {ar,b~,cl} U {u2,b2,c2} U {aa,ba,ca} = E
such that the intersection of the three triangular re-
gions conv(qbici) (1 5 i 5 3) is non-empty. Then
n; conv(e;b;c;) is a convex polygon. Assume that the
line aaba contains an edge of this polygon. The pro-
longation of this edge in any direction will leave one of
the triangles conv(arbrq) or conv(aabacz) first; assume
it leaves conv(azb2cz) first, at a point p. Then p is a
crossing, defined by four of the points a2bzcZasba, and
it is contained in the triangle conv(aibrcr).
So every nine-tuple from Y contains an (ordered)
seven-tuple abctyvv such that such that the lines my
and uv intersect at an interior point of the triangle xyz.
Let s denote the number of such 7-tuples in Y. As
every seven-tuple is contained in (“i’) nine-tuples we
have that s > (:)/(“G’) = (;)/36, and so
(2.2) tyxye N(abc) = fi 2 (n 4 “> /I266 I planar convex regions.
,, 3 THEOREM 3.3. There is a set F C R2 of size at most
Unfortunately, this computation is not enough to
guarantee that most triangles contain Q(n”) crossings.
For this we need a colored version of Tverberg’s theo-
rem:
LEMMA 2.3. There is a positive integer t such that the
following holds. Assume that A, B, C c R2 are disjoint
sets with at least t elements each, such that their union
is in general position. Then there exist three disjoint
triples aibici, ai E A, bi E B, ci E C (1 5 i 5 3) such
that fIi conv(aibic;) # 0.
The smallest value oft for which we managed to prove
this lemma is 4, and we do not have a counterexample
even for t = 3. For brevity’s sake we give the proof for
t = 7.
The other tool of the proof is the following lemma,
which strengthens the averaging in (2.2). This lemma
will imply that the number of triangles with vertices
from Y containing “few” crossings is “small”.
141
LEMMA 2.4. Let t satisfy the previous Lemma. Then
there exist positive constants c’ and c” such that the
following holds. Assume that 1 _< k _< c’n11t2, and 7-t
is a set of triples from Y with 1x1 > (1)/k. Then the
average number of crossings in the members of 31 is at
least cnn4/kt3 - l.
3. COROLLARIES AND AN ALGORITHM
In this section t is a value that satisfies Lemma 2.3,
k > 1 and Y C It2 is an n-element set in general po-
sition. A straighforward application of the Lemma 2.4
yields the following covering theorem, where c is again
an absolute constant.
THEOREM 3.1. There is a set F c R2 of size at most
&t3 - 1 such that the number of triangles with vertices
from Y containing no point of F is at most (I;)/k.
It is interesting to compare 3.1 to a result from [BF],
which states that there is a point contained in at least
$ (1) triangles from Y. (For higher dimensions, see [B]).
The covering set F in Theorem 3.1 is obtained by a
random process. We have a deterministic, polynomial
time algorithm to construct a suitable F as well, but
now F will have larger size:
THEOREM 3.2. There is an algorithm, polynomial in n,
which supplies a set F with IFI _<exp(c’kgooo) such that
the number of triangles from Y containing no point of
F is at most (3/k.
Here c’ is another absolute constant. The following
corollary of 3.1 concerns the difference between the be-
havior of a continuous and a discrete measure of the
&t3 - 3 such that any convexregion R with IRnYl 2
n/k contains a point of F.
This follows from Theorem 3.1, as if IR fl YI _> n/k,
and RnF= 8, then F avoids at least (ni’) triangles.
4. THE PROOF OF LEMMA 2.3
LEMMA 4.1. Let El, Ez, Ea C R2 be finite nonempty
subsets and p any point. Then p is not contained in any
triangle conv(ere2ez) with e; E Ei if and only if there
exist a k E {1,2,3} and two closed halfplanes H, H’
such that p 4 H’ U H”, Ei c H’ n H” if i # k and
Ek C H’ U H”.
PROOF: By Theorem 2.3 of [B], p is contained in a tri-
angle conv(ereZez) if it is contained in the convex hull
of two of the sets El, E2, Es. So we may suppose that
p 4 conv Ei for i = 1,2, say. Write Cr and C2 for the
smallest cone containing E1 and E2 and having apex p.
It is easy to see that if Cr UC, contains a line, then p is
contained in a triangle conv(ere2eg). Then the smallest
cone containing Cl U C2 and having apex p is of the
form Hi II Hz, where Hi and Hs are two-hyperplanes.
It follows readily that HI and Hz satisfy the require-
ments. 1
Let U = AU B UC, jcTl = 3t 2 21, i 2 0. The i-th
con2rez hull, convi(r/), is the intersection of all of the
(open) halfplanes containing at least ]U] - i elements
of U. Then convi(U) is a convex polygonal region for
0 < i 2 t - 1. Let p be a point from intconvl-l(U),
such that U U {p} is in general position. Then for all
open halfplanes H we have that
(4-2) p E H implies IH n UI > t.
Every copy of K(t,t, t) contains three multicolored tri-
angles with a common interior point, and so, as we have
seen in the argument leading to (2.2), it contains an (or-
dered) seven-tuple abczyvv such that such that the lines
xy and uv intersect at an interior point of the triangle
xyz. If, as before, s denotes the number of such 7-tuples
in Y, then by (5.2), s 2 (cin”‘//~~“,l(~~~) 2 c2n7/kt3.
Then, as in (2.2), one has
Average N(abc) 2 L!- > c3n4/kt3 - ‘. m
ia,b,cJE~ WI -

6. THE PROOFS OF THEOREMS 3.1 AND 2.1

We claim that p is contained in at least three vertex-
disjoint multicolored triangles of U. As Theorem 2.1 is a trivial corollary of 3.1 (with k =
Suppose, to the contrary, that one can find only s c7w3) we turn to the proof of Theorem 3.1. Define the
(S = 0,l or 2) triangles aibici (i = 1,. . . ,s) such that triangle system X(i) as the set of triangles {a, b, c} c Y
p E conv(sibici). Let U’ = U \ {ai,bi,ci : i < s}, A’ = satisfying
AflU’ and so on. We have IU’l = 3t-3s. Apply Lemma
4.1 to A’, B’, C’ and p. We obtain two halfplanes H’,
H” such that (say) A’UB’ C H’nH#, C’ C H’UH”, and dtn4 5 N(abc) < c”n4
t3 - 1

p $I!H’UH”. The complementary halfplanes H’ and H”

both contain at most 2s points from {ai, bi,ci : i 5 s}.
One of them, say H’ contains only at most one half of for i = O,l,.... If A > (2n)1/(t3 - ‘), then the
the points of C’ from U’. So altogether H’ contains at bound on IFI is larger than 2n, and it is easy to see
most 2s + $t - s) p oints of U. This contradicts (4.2) ([KM] or [B&F]) that 2n points are always sufficient to
asl>7>3s.I cover all triangles from Y. So we may suppose that
k < (27~)l/(~~ - ‘) 5 crn11t2. Then Lemma 2.4 implies
5. THE PROOF OF LEMMA 2.4 that
A hypergraph H is a pair H = (V,C), where V is a fi-
nite set, the set of vertices, and & is a family of subsets of
V, the set of edges. If all the edges have r elements, then
(6.1) I~(O)l+l’H(l)l+...+I’H(i)l
5 0; %.
H is called r-graph, or r-uniform hypergraph. The com-
plete r-partite hypergraph K(ti, t2,. . . ,t,) has a parti- Now we are going to give a random construction for
tion of its vertex set V = VllJ. . UV,, such that ]K] = ti, the covering set F. A crossing from C(Y) is put into
and t = {E : IE n Vi1 = 1) for all 1 5 i 5 r. ErdBs F with probability p, where p = akt3 - lnm4 andcris
[E65] proved the following theorem in an implicit form. an absolute constant to be fixed later. The expected
(More explicit formulations see in ErdGs and Simonovits number of points in F is
[ES] or in Frank1 and Rod1 [FR]).
LEMMA 5.1. For any positive integers-r and tl 5 . . . 5
t, there exist positive constants c’ and c” such that the
following holds. If H is an arbitrary r-graph with n We estimate the expected number of triangles from X(i)
vertices e 2 c’n*-’ edges where E = l/(tl . . .&-I), then containing no point of F:
H contains at least
$1 . . . t, (6.2)
C
II Xi =: E(# {a, b, c) E N(i) : F n conv(abc) = 0) =
nrtl...t,-tl-...t,
c (1 - py(“bc) 5 ]7+)](1 - p)minN(abC) 5
obcER(i)
copiesofK(t1,...,t,).
IX(i)1 exp(-pmin N(abc)) 5
Now consider the triangle system ‘?f, and consider it
as a J-regular hypergraph with vertex set Y. Lemma 5.1 IX(i)1 exp(-c&it3 - l/4).
implies that there is a constant cl such that the number
Then (6.1) and (6.2) imply that
of copies of K(t, t, t) in R is at least
(6.3)
t3
(5.2) cln3’/k

142
Then the expectation of the random variable One can divide A2 into two equal parts by a halfline
hi starting from the intersection point of 12and la. Sim-
PI + cxi ilarly, a halfline h2 parallel to 1 divides Bz into two equal
akt3 - l/S n 1+4/cxP parts, and finally, a halfline hs starting from the point
4k 111sdivides Cs. Then consider the triangle T formed
03
by hr, hs and the continuation of ha. The sides of T
is less then 2. So there is a choice of F such that IFI 5 divide the plane into 7 regions. Let A’ the part of A2
( 1/4)akt3 - ’ and contained in the region with 2 sides.. The definition of
B’ and C’ are similar. Then every point p E T satisfies
the requirements in the Lemma. 1
# empty triangles 2 i ’ +$o”‘.
0 For the proof of Theorem 3.2 we only need from the
above argument that for arbitrary sets A, B and C there
Chosing cr = 4/c”, one obtains Theorem 3.1. I
is a point p which is contained in at least
7. THE PROOF OF THE MAIN THEOREM 1.1
We have to prove the upper bound. Suppose that (8.2) I4l~llCt/1728
S c R3 is an n-set and P(&), and P(abd) are halving
triangles abc with a E A, b E B, c E C; and moreover,
planes, {a, b,c,d} c S. Let H, and Hd be halfspaces
it is easy to find such a point p algorithmically.
with boundary planes P(abc) and P(abd), resp., such
Suppose we have an algorithm supplying a cover F,
that {a, b, c, d} c H, II Hd. Then there is a point z E S
which avoids at most (!-J)/k triangles of Y. Let IFI =
outside of H, U Hd such that abz is again a halving
f(k) or briefly f. From any point x E F one can start
triangle. This can be seen by rotating (around the line
halflines hr(x), hs(z), . . . , h,(z) such that the cone de-
ab) P(abc) into P(abd).
fined by hi(z) and hi+r(z) contains about n/m points
Now take a plane & in general position with respect
of Y. Let Rl,Ra,. . . , RIM be the cell-decomposition of
to S, and consider Y, the image of S on Q under or-
the plane defined by the halflines {hi(Z) : x E F, 1 5
thogonal projection. Denote the system of images of the
i 5 m). Then M < (fm)2. Call a three-tuple of the
halving triangles in S by 31. Let x denote the sum of
regions R,, Rb, R, uncovered if all triangles conv(tyz)
the characteristic functions of the open triangles in 31.
with x E R,, y E Rb and z E R, avoid F. They are
By the above observation, x changes by at most 1 when
covered if all triangles contain a point from F, and am-
one crosses a segment uw with u,v E Y. This means
biguous if both of the above constraints fail.
that x is at most (y). To put this differently, every line
The number of triangles from Y which are covered by
orthogonal to Q and in general position with respect to
at most 2 regions Ri, Rj is at most 0(1/m)(z). It is
S intersects at most (;) halving triangles of S.
easy to see that the number of triangles in ambiguous
Let F be a point set according to Theorem 2.1. Then
triples is at most (9f/m)(“,).
The above Lemma yields a point p(a, b,c) for each
WI I IFI ; -t(# empty triangles ) 5 O(n3-‘lt3). I uncovered triple R,RbR,. Then, the set F U {~(a, b, c) :
0 1 5 a, b, c 5 M} avoids less than
8. SKETCH OF THE ALGORITHM
LEMMA 8.1. -Assume that A, B and C are sets in (8.3) (;) ((;-~)~+~)
general position in the plane. Then there are subsets
A’ c A, B’ c B and C’ c C with IA’1 > IA1/12,
IB’I 1 IBl/12 and IC’l > ICI/12 and a point p such triangles by (8.2). If we choose m = 20fk, then (8.3)
that p is contained in all triangles abc whenever a E A’, gives 3455(:)/3456k. This leads to the recursion
b E B’ and c E C’.
PROOF: First, for any direction 1, one can find two lines (8.4) .f(gk) 5 f+ (7) 5 (fm)’ = O(f(k)12k6).
Ir and 12 parallel to 1 which divide R2 into three regions
Ro, RI and Rz (where Ro and Ra are halfplanes with
It is easy to check that (8.4) implies f(k) 2 exp(ckgooO).
boundaries 11 and Is, resp., and RI is a strip), such
that each fi contains one third of the points of some 9. PROBLEMS
color class. Say, e.g., Al =: A rl Ro, (AlI 2 IAl/ and
Define hd(n) as the maximum number of halving hy-
B1 =: B rl RI, lBl[ 2 IBl/3, finally Cr =: C II RX,
perplanes in d dimensions. It is known that
ICrl > ICl/3. By the Ham-Sandwich theorem, there
exists a line 1s that divides both Al and Ci into almost
equal parts. Denote by Hs the halfplane with boundary
h(j(n) = i-2(7&-1(n)) = R(ndlog n)
1s and containing the larger part of B1. Then let Bz =:
B1 nH3, we have (8212 lBrl/2 1 IBl/S and A:! =: A1 \
Ha, lAzl L 14/S, finally CZ=: G \ H3, .IGI 2 ICl/6.

143
 To extend our arguments to the (d + 1)-dimensional
case, only a d-dimensional version of the Lemma 2.3 is
missing. This generalization could be that there exists
a positive integer t = t(d) such that if we have d + 1
disjoint sets of points Ao, . . . ,Ad in R2, each of size t,
then there exists distinct points eij (0 ,< i, j 5 d) such
that u;j E Ai and
COnV{aOo, . . . , ado} n . . . n COnV{aOd, . . . , add}
This would imply that for an appropriate c = c(d) > 0
one has
hd(n) = o(n d-c ).
It would also be interesting to find the higher dimen-
sional analogues of Lemma 4.1 and of the algorithm in
Theorem 3.2.
What is c(f), the maximum ratio of the number of
covered triangles by f points? The only known value,
as it was mentioned, is ~(1) = $. We conjecture that
E(J7-i) = 0(1/&i).
Of course, it would be also interesting to find the best
values for the constants in our lemmas, like in Lemma
2.3.
10. ACKNOWLEDGEMENTS
The authors are grateful to J. Beck, J. Path, J.
Tordcsik, W. Steiger and E. Szemeredi for stimulating
discussions.
REFERENCES
[B] I. BARANY, A generalization of Caratheodory’s theo-
rem, Discrete Math. 40 (1982), 141-152.
[B&F] I. BARANY AND Z. F~~REDI, Empty simplices in
Euclidean space, Canad. Math, Bull. 30 (1987), 436-
445.
[BF] E. BOROS and Z. F~REDI, The number of triangles
covering the center of an n-set, Geometn’ae Dedicata,
# 0. 17 (1984), 69-77.
[Ed] H. EDELSBRUNNERAlgorithms in Combinatorial Ge-
ometry, Springer, (1987), Section 3.5.
[E65] P. ERDBS, On extremal problems of graphs and gen-
eralized graphs, Israel J. Math. 2 (1964), 183-190.
[ELSS] P. ERD~S, L. LovAsz, A. SIMMONS AND E. G.
STRAUS, Dissection graphs of planar point sets, in A
Survey of Combinatorial Theory (J. N. Shrivastava et
al., eds.), North-Holland, Amsterdam, (1973), 139-149.
[ES] P. ERD~S AND M. SIMONOVITS, Supersaturated
graphs and hypergraphs, Combinatorics 3 (1983), 181-
192.
[FR] P. FRANKL, AND V. R~DL, Hypergraphs do not
jump, Combinatorics 4 (1984), 149-159.
[L] L. LOVASZ, On the number of halving lines, Annales
Univ. R. Eiitviis 14, 107-108.
[KM] M. KATCHALSKI AND A. MEIR, On empty triangles
determined by points in the plane, Acta Math. Hungar.
51 (1988), 323-328.
[R] J. REAY, An extension of Radon’s theorem, Illinois J.
Math. 12, (1968), 184-189.
[T] H. TVERBERG, A generalization of Radon’s theorem,
J. London Math. Sot. 41 (1966), 123-128.
144