COMBINATORICA 2(4) (1982) 333--340

THE UNIQUENESS OF THE NEAR HEXAGON

ON 729 POINTS

A. E, BROUWER
Received 6 Augustus 1981

We prove that any regular near hexagon with 729 vertices and lines of size 3 is derived
from the ternary Golay code, thus settling the last case in doubt a m o n g the regular near hexagons
with lines of size 3.

Introduction

A near hexagon is a partial linear space (X, ~ ) such that

(a) For any point pEX and line { ~ t h e r e is a unique point on g nearestp.
(b) Every point is on at least one line.
(c) The distance between any two points is at most three.
(The distances are measured in the point graph: d(p, q)=:1 iffp and q are collinear.)
A regular near hexagon with parameters (s, t, t2) is a near hexagon such that each
line contains 1 + s points, each point is in 1 + t lines, and a point at distance 2 from
a fixed point x0 is in 1 + t2 lines containing a neighbour of x0.
Shult and Yanushka [4] showed that there are exactly eleven possibilities for
the parameters of a regular near hexagon with s = 2 . For ten parameter sets the cor-
responding near hexagons have been classified completely (by Cameron, Shult and
Yanushka, Tits, Cohen and Brouwer). Here we settle the last case by showing that
there is a unique regular near hexagon with parameters (s, t, t~)=(2, 11, 1). As Shult
and Yanushka indicate, an example is given by the 729 vectors of the extended ternary
Golay code, where lines are triples {x,y,z} with x + y + z = O and dH(x,y)=
=dn(x, z)=dn(y, z)= 12 (where dn is the Hamming distance in F~Z). One finds that
distance 0, I, 2, 3 in the point graph corresponds to vectors at Hamming distance
0, 12, 6, 9, respectively.
Here we find another way to describe this near-hexagon in terms of the exten-
ded ternary Golay code, and prove that any regular near hexagon with (s, t, t~)=
= (2, 11, 1) can be obtained in this way.

AMS subject classification (1980): 05 B 30, 51 E 10, 94 B 25

1"
334 A.E. BROUWER

1. P a r a m e t e r s

Let -Z~ be a regular near hexagon with parameters (s, t, t2)=(2, 11, 1). Let
ki = [Fi(xo)l = + {xld(x, xo)=i} for some fixed x o E ~ . Then

]Co = 1

/"1 = 24 (= s(t+l))

I¢o = 264 (= k~. s. t./(t2+ 1))

k3=440 (=l%.s.(t-t2)/(t+l))
so that v = Zk~= 729.
Diagram of the distance regular point graph"

1 2 12

(pkoj).ik =
1°°!/
o~ is an association scheme with intersection numbers (p~j), where

01 0
00 1 / '
(p~j) =
24 1
0 22 2 12 '
2 00

(p~j) =
°°1°/ / /
000

0 22 2 12
264 22 131 78 ' (P~J)=
0 0 20 12

0 20 12
220 130 174 "
0 220 130 174 440 220 290 253

Using some properties of ~ that will be proved in the next section we give here also
the parameters w.r.t, a quad. If Qo is a fixed quad (cf. Shult and Yanushka [4]) then
there are 9, 180, 540 points at distance 0, 1,2 from Q0, respectively. Diagram :

/. 5 18

(A point in Qo is in 12 lines, 2 in Qo and 10 with one point in Qo and two points in

FI(Qo). A point in F~(Qo) is in 12 lines, 1 meeting Qo, 2 contained in FI(Qo), 9 with
two points in F2(Qo). A point in F2(Qo) is in 12 lines, 6 contained in F2(Qo) and 6
having one point in FI(Qo).)
 NEAR HEXAGONS ON 729 POINTS 335

2. Quads and cubes

(For definition, existence and properties of quads, see [4].) In ~ the quads
have 9 points and 6 lines and look like the picture shown, i.e., they are G Q (2, 1)'s.
Any two intersecting lines determine a quad. We shall show that any three concurrent
lines determine a 3 × 3 × 3 cube (called H (27) by Shult and Yanushka). •

A point x at distance one from a quad Q has a unique neighbour in the quad,
say 7ze(x) or just re(x). In any near hexagon the following is true:
Lemma. Let Q be a quad and x, y two adjacent points at distance one from Q, where
~(x) ¢ ~ ( y ) (i.e., the line xy does not intersect Q). Then
(i) any pohTt on xy has distance one to Q;
(ii) ~(xy) is a line in Q.
ProoL First of all ~ ( x ) ~ ( y ) ( ~ denotes adjacency, i.e., collinearity), since
d(x, n ( y ) ) = 2 and a shortest path from x to n(y) can be found through re(x). If
zExy and d(z, Q) = 2 then the collection O of points in Q closest to z forms an oval
in O (i.e., a set of points in Q meeting each line of Q in exactly one point), but ~z(x),
rr(y)EO, a contradiction. This proves (i), and shows that 7r(xy) is a clique in Q, so is
contained in a line f. Let p e g . Let q be the point on xy closest t o p . Since there are
two points on xy at distance 2 from p and xy does not meet Q it follows that d(p, q) =
= l , i . e . pETz(xy). I
In our case there are 6 ovals in any quad - - the quad together with (lines and
ovals) as lines is an atfine plane A G (2, 3) - - and the ovals fall into two parallel clas-
ses. Fix a quad Q and call the two parallel c l a s s e s / / a n d \\. If x and y are two adja-
cent points in F~(Q) then either the line xy is contained within F2(Q), and the ovals
determined by x and y are disjoint, hence parallel, or the third point z of xy is in
FI(Q), and the ovals determined by x and y meet in rr(z), hence are not parallel.
Fix a point xEF~(Q) and consider the quads Q' containing x. These are of
seven possible types indicated by a diagram : an open dot is a point of Q, a solid dot
is in F~(Q) a n d / / a n d \\ label points belonging to ovals in these two parallel classes.
We may assume that x itself is labelled.

(Note that patterns with one or three solid dots (not on a line) are impossible. The
former since it cannot be completed with / / a n d \\ labels. The latter because if p, q, r
are three points in Q'NFI(Q) and y is a common neighbour o f p and q in Q' then the
336 A.E. BROUWER

oval determined by y in Q contains n(p) and n(q) and therefore is {n(p), n(q), n(r)};
now y must be adjacent ro r, a contradiction.)
In order to exclude possibility 1, let us write down some equations for ni,
the number of quads Q" of type i.
(a) Count lines yz in FI(Q) with x ~ y : X is in 12 lines, 6 contained in F2(Q) and
6'w-ith a unique point in FI(Q). Thus we have 6 choices for y, and given y we have 2
choices for the line yz. This shows that

2nl+n2 = 12.
(b) Count pairs of lines passing through x and contained in F~ (Q):

na+n4 = 15.

(c) Count pairs of lines passing through x and meeting F~(Q):

no+hi+n8 = 15.
(d) Count pairs at distance 2 from x in F~(Q): One one hand there are 2n0+
+3n1+2n2+2n4+n5 such points (for: such a point determines together with x a
unique quad Q'); on the other hand, for points zCF~(Q) with d(x, z ) = 2 consider
y=r~(z). I f d(x,y)=2 then there are p ~ = 2 points z with z~(z)=y. I f d(x,y)=3
then there are p ~ l - 2 = 10 points z with r~(z)=y. Altogether we find 3 . 2 + 6 - 10=
= 66 points z. Thus:
2no+3nl+2n2+2n4+n5 = 66.
(e) Count points at distance 2 from x in Q:

r~0 = 3.

(f) Count pairs o f lines passing through x:

no+nl+n2+na+n4+ns+n6 = 66.
This gives us six equations with seven unknowns. Putting n l = ~ we find

1 1
no = 3, nl = ~, n2 = 12--2~, na = 9+~-ce, na = 6--~-cq

n5 = 24+2c¢, n6 = 12--cc

N o w put ~ = c~(x), nt = ni(x) and vary the point x. Clearly, each quad Q' with 2 points
in Fx(Q) is of type 4 for one point x, of type 5 for four points x and of type 6 for two
points x. Consequently, averaging over x, we have ~ = 4 ~ 4 and /ie=2~a, so that
~ = 0 . But i f n 1 is zero on the average then it is always zero, and we have for each x:

n = (3, 0, 12, 9, 6, 24, 12).

In particular case (1) does not occur, so that we proved:

 NEAR HEXAGONS ON 729 POINTS 337

Lemma. Let Q, Q" be two points such that Q' (] Fi(Q) contains two intersecting lines.
Then Q"=Fi(Q). |
Now we are in a position to prove the announced e:,istence of cubes:
Proposition. Let g~, C2, {3 be three concurrent lines. Then there is a unique minimal
set C containing these three lines and such that iJ it contains two points at distance 2
then also the quad determined by them. We have [C 1=27 and the graph induced on C is
isomorphic to the graph on the Hamming scheme F3a where adjacency =- Hamming
distance one. Such sets C are called cubes.
Proof. Let Qij=(gi, fi) be the quad defined by g'i and g'j ( i , j = l , 2, 3; i ~ j ) .
Let C=QI~.UQlaU{yld(y, Q12)=d(y, Q13)=1, d(y, .(1)=2}. Then Q~acC, and
if yEC",,(Qi2UQiaUQ23) and zE{i with d(y, z ) = 2 then by the lemma above each
point of the quad Q(y, z) has distance one to Q23. Using the fact that if Q is a quad
and xEFI(Q) there are exactly two lines passing through the point x and contained
entirely within Fx(Q) one easily verifies all claims. |

3. Vector space structure

Assuming the validity of this proposition Shult and Yanushka proved the
existence of a regular abelian group of automorphisms of ~ . In fact they first proved
(tediously) :
Let d(x, y ) = 3 . Then there are precisely 4 cubes containing both x and y.
These cubes meet, except for x and y, in a unique third point z, where z is the unique
point at distance 3 from both x and y in each cube.
Now fixing a vertex and calling it 0 we may define an involution x ~ - x by
fixing 0, interchanging points S 0 on a line through _0, interchanging points ~ 0
on an oval through 0, and interchanging points y, z if d(_0, y) =d(0, z) =d(y, z)--3,
and z is in each of the 4 cubes containing 0 and y.
The product of two such involutions (with different fixpoints) is an automorph-
ism without fixpoints and is uniquely determined by the image of some vertex 0.
These automorphisms are our translations, and we have the group Z~.
Identify the vertices with group elements. Now we have given ~ the structure
of a vector space over F3, and lines and ovals have the property that the sum of their
elements is zero. Lines, quads and cubes are one-, two- and three-dimensional sub-
spaces (but there are also other subspaces of these dimensions). [Our subspaces are
affine subspaces - - they need not contain _0.]

4. Buekenhout diagram

At this point I did not know how to continue: the extended ternary Golay code
is a 6-dimensional subspace of a 12-dimensional vector space over Fa. We found the
subspace, but where is the big space? One would like to identify the 12 coordinate
positions, and an obvious choice seems to be the set of twelve lines through 0. But
in Shutt and Yanushka's construction these lines do not correspond to coordinate
positions. A coordinate position gives us a coclique of size 243, but it is not clear how
338 A . E . I]ROUWER

to find such cocliques. Arnold Neumaier suggested first to identify the objects belong-
ing to the Buekenhout diagram

:2 C C f~,f

of the extended Golay code, and indeed, these are easy to find. But having found these
one may define a Golay code where the 12 coordinate positions are in fact the 12 lines
through 13. Thus we find a newlconstruction of the near hexagon ~ and prove that
any near hexagon with these parameters can be obtained in this way.
So, let us look at the diagram. The first three species are points, lines and
quads; the next type is formed by the cubes. Objects F of the fifth type have the pro-
perty that any two cubes with a ]quad in common determine a unique such F. So these
must be hypercubes, generated I by four concurrent lines. (We shall call our objects
/-spaces i f = 0 , 1, 2, 3, 4, 5); note that not each subspace of dimension i is an
/-space.)
We must show that the 4-space Fgenerated by four lines ( o f ~ ) through a point
x0 cannot contain other lines (of .~). Two cubes with a quad in common cannot meet
outside this quad since the intersection of subspaces is a subspace. Thus we find that
the four cubes determined by the triples from the four given lines contain together
65 points (1, 8, 24, 32 points at distance 0, 1, 2, 3 from x0, respectively) so that F
contains 16 more points. If we set x 0 = 0 then these points are sum of four vectors
along the four given lines, but each sum of these three vectors has distance 3 to 0 so
our new points have distance at least 2 to 0 (in fact 8 are at distance 2 and 8 at di-
stance 3), and all points at distance 1 from 0 in F lie on the four given lines.
The last type of object is determined by two 4-spaces with a 3-space in common,
hence must be a 5-space generated by five lines through a point xo. The stroke A~
o----o

denotes an A G (2, 3) in this case, so we want to prove (i) that this partial linear space
has 9 points - - and indeed, given a cube containing x0 there are 9 4-spaces F contain-
ing this cube, determined by the 9 lines on x0 not in the cube, and (ii) that this partial
linear space has lines of size 3, i.e., that 5-spaces generated by five lines through x0
contain in fact six lines through x0.
Suppose that the 5-space G on _0 contains only 5 lines (of -Jd) through 0, i.e.
only 10 points at distance one from _0. The point set of aed is partitioned by G and
its two cosets G1 and G2. Since kl =24 there is a coset, say G1, containing at least 7
points at distance one from 13. G1 is union of three 4-spaces, at least one of which
contains 3 points at distance one from 0. Now we find a contradiction from
Lemma. L e t F be a 4-space. Then a point x ~. F has at most 2 neighbours in F.
Proof. Suppose x has three neighbours in F. I f C is a cube contained in F then x has
precisely one neighbour in C (not more since a cube is a Cameron set, i.e., is closed
under formation of quads - - not less since F is union of C and two translates of C).
Coordinatize F with Z~. Then the 3 points in F at distance 1 from x have no coordi-
nate in common so are 1111, 2222 and 0000, say. Then x is at distance 2 from 1000,
1011 in the quad 10"*, hence also from 1022. Looking at the line *022 we have two
points, 1022, 2022 at distance 2 from x, so that 0022 must be at distance 1 from x,
a contradiction. II
 NEAR HEXAGONS ON 729 P O I N T S 339

Consequently, G contains at least 6 lines through _0, and the linear space of
4-spaces and 5-spaces on a fixed 3-space has lines of size at least 3. But there are only
two linear spaces on 9 points with lines of size at least 3, namely A G (2, 3) and the
space with a single line of size 9. The latter, however, is impossible - - if G contains all
lines through 0 then G contains all of ~ and cannot be a 5-space.
This proves that we indeed have the diagram given above.

5. Uniqueness
C C C AF
It is well-known that there is a unique object with diagram ~-o=o=-o=o
where A f denotes an A G (2, 3), namely the Witt design S (5, 6, 12). We use only the
simple property that the complement of a block of this design again is a block. Choose
six lines :; ( 0 N i ~ 5 ) through 0 not forming a block in the local design S (5, 6, 12).
Then the remaining six lines :; ( 6 ~ i N 11) do not form a block either. Choose a point
ei.~0 on each of the lines :i. Then both B={el]0=<i=<5} and B'={ei[6<-_i<=ll}
form a basis for the vector space ,g'. Associate a vector of length 12 with each point
xC--gY, writing x,--(% . . . . . ~u) if X=~Y'i=l
5
~iei--Si=6'xle,,
__ 11
The image of ~ under
this map is a 6-dimensional subspace of Z~2. Let us prove that it is the extended G o l a y
code. To this end it suffices to show that the miniinum distance is 6 (cf. Delsarte and
Goethals [2], Pless [3]).
Suppose v = ( % . . . . . an) is a nonzero vector of weight at most 5. Let WL(V)=
= ~ {ilcq~0 and 0 N i = 5 } and we(v)= :~ { i l c ~ 0 and 6_-<i_ <- 11} be the left and
right weights of v. Then v can be written as sum of WL multiples of vectors in B and
also as sum of wR multiples of vectors in B'. Therefore there is a representation of _0
as linear combination with nonzero cooefficients of W=WL+WR vectors in BUB', or
also a representation of one vector in BUB" as linear combination of at most four
others. But we know that the 4-sp'~ce generated by these others does not contain other
points of B U B', a contradiction.
This proves that the image of J:" is the extended Golay code. But conversely,
if ~ is the extended Golay code then let Dc:g be the set of 24 vectors with exactly
one nonzero coordinate in the first 6 positions or exactly one nonzero coordinate in
the last 6 positions. We find ~ back again by calling two vectors adjacent if their
difference belongs to D. (Now the lines of ,~f are the triangles in the point graph.)
Note that we obtain only one near hexagon Y in this way (up to isomorphism) :
the automorphism group of 9" contains M~2 acting on the positions (cf. [1], page 85),
and under this group the 6-sets fall into two orbits: those corresponding to a block
o r S (5, 6, 12), i.e., those forming the support o f a codeword, and the remaining ones.
Clearly, when talking about 'the first six positions' we mean a 6-set not forming a
block. (Note: here automorphisms are monomial transformations, i.e., matrices
with one nonzero entry in each row and column. That there are only two orbits, is
seen as follows" M~2 is sharply 5-transitive and the setwise stabilizer of a 5-set has
order 120; it fixes a 6th point and acts as P G L (2,5) on the remaining 6 points.) Thus
there is a unique regular near hexagon with parameters (s, t, t.,)=(2, 11, 1).
340 A. E. BROUWER" NEAR HEXAGONS ON 729 POINTS

References

[I] P. J. CAMERON and J. H. VAN LINT, Graph Theory, Coding Theory and Block Designs,
LMS Lecture Notes 19, Cambridge Univ. Press, Cambridge, 1975.
[2] PH. DELSARTE and J. M. GOETHALS, Unrestricted codes with the Golay parameters are unique,
Diser. Math. 12 (1975) 211--224.
[3] V. PLESS, On the uniqueness of the Golay codes, J. Comb. Theory 5 (1968) 215--228.
[4] E. E. SHULT and A. YANUSHKA, Near n-gons and line systems, Geometriae Dedicata 9 (1980)
1- - 7 2 .

A. E. Brouwer
]l/[athematisch Centrum
Kruislaan 413, 1098 SJ Amsterdam
Tile Netherlands
 COMBINATORICA2(4) (1982) 341--345

ON THE VECTOR SPACE OF 0 - C O N F I G U R A T I O N S

M. D E Z A a n d P. F R A N K L
Received 3 February 1982

Let e be a rational-valued set-function on the n-element set X i.e. e(B)EQ for every
B~=X. We say that ct defines a 0-configuration with respect to .N'~2 x if for every A C,~ we have
e(B)=0. The 0-configurations form a vector space of dimension 2 " - r , ~ l (Theorem 1).
A~BC_X
Let O~_t<k~_n and let ~ = { A ~ X : IAl~_t}. We show that in this case the O-configurations sat-
isfying ~(B)=O for IBl > k form a vector space of dimension r<i~_k
Z ( ni ) , we exhibit a basis
for this space (Theorem 4). Also a result of lZrankl, Wilson [3] is strengthened (Theorem 6).

I. Introduction and statement of the results

Let X = {xl . . . . . x,} be a finite set o f n elements. F o r an element A o f 2 x =

= {F: Fc=X} we define the m o n o m i a l p ( A ) = 17 x , p ( 0 ) = l .
x~A
Set I / = V(2X)={ ~ ~(A)p(A): c~(A) is rational}, i.e. Vis the set o f a l l s q u a r e -
a~_x
free p o l y n o m i a l s in the variables xl .... , x,. O f course, V is a vector space o f d i m e n -
sion 2" over Q, the field o f rationals.
F o r a f a m i l y o f subsets, d _c_2 x we define
V(d)={a==~xCt(A)p(A): ~(A) is r a t i o n a l , ~ ( A ) = 0 unless A ~ d } .
F o r an ~ ¢ ~ 2 x a n d f = ~ ~(A)p(A)C V we say t h a t f i s ~¢-orthogonal o r a
ac=x
O-configuration with respect to d if

(1) ~' c~(B) = 0 holds for all A E ~ ¢.

A~B~X

We prove:
Theorem 1. The set V * ( d ) o f all ~-orthogonal elements o f V(2 x) is a vector space o f
dimension 2 " - l d l , moreover V ( ~C) f~ V * ( ~ ) = {0}.

AMS subject classification (1980): 05 C 65; 05 C 35, 15 A 03.

342 M. DEZA, P. F R A N K L

In [1] the following m o r e general definition was considered:

Definition 2. For d , ~ c=2 x we set

V*(J, ~) = {fCV*(s~) J~V(~)} = V*(d)ff)V(~)

Tile special case ~._: t !

negative integers) was considered by Graver, Jurkat [5], and G r a h a m , Li, Li [4]. The
elements of V*(,~¢, ,~) are called in this special case 0-designs.

, ° , - a r k . No e that - k we have

v'f( x I , ,~],
and for d c = ~ ' c = 2 x, ~ ' c _ ~ c _ 2 x
V*(.ff, :~) _D V * ( d ' , B ' ) holds.

(xi, - xi~)(xi~-- xi:)...(xi .... - xi~ ÷~)xi.,,. ~...xl . . . . . • (xi~, ..., xl ...... are distinct elements

d,..

Here we prove
{.

~,oo,o°.. ~ ~,~;~ o j , ~ ; ° c ~
( x i l - 1),--(xi~+l-- 1) xi,+....xi~
~*/[
K-
where
Xt' _~k
t+2~l<-k and l--xiz<...<iz<=n. Thus

Remark. In [2] the case t = 2 , k = n was considered. There the terminology

0-measure (or isometry) was used and a entirely different generator system was
exhibited.
The following theorem of R a y - - C h a u d h u r i , Wilson [6] can be formulated in
terms o f 0-designs.

Theorem 5. [6]. Suppose ~ c ( X ) is such that ] B A B ' , takesatmosttvaluesfor

f-- N

non-trivial O-design,
and consequently
 0-CONFIGURATIONS 343

Singhi [7] pointed out that one can localize Theorem 5, i.e. in any 0-design
f= ~ c~(B)p(B)onecanfinda B ~ [ X / stick that e ( B ) ¢ 0 and
kr~)
I{IBNB'I: B¢

We prove a similar strengthening of a theorem of Frankl--Wilson [3].

Theorem 6.Suppose .f= Z c~(A)p(A)~V*[[ X ) ( 2k)l" Then there exists an

A_c_X ~_ _ t_ ' _
Ac=X such that c~(A)~-O and
I{IANA'I: A ~ A" c= X, ~{1") ~0}l--> t + l .

2. The proof of the results

Proof of Theorem 1. The fact that V*(d) is a vector space is evident. The solutions
of (1) for a fixed AC2x form a subspace of dimension 2 " - 1 in V - - as the solution
of any non-trivial homogenous linear equation. Now obviously V*(z~¢)= ['1 V*({A}).
Thus V* ( d ) is the intersection of Id] subspaces of dimension 2 " - 1 , yielding
(2) dim V*(~/) => 2"-1,~'[.
As dim V ( d ) = I d 1, dim V * ( o ~ ) = 2 " - I d l will follow from (2) if we establish
V(d)(] V*(d)--- {0}. To prove this let f - ~ ~(A)p(A) be an arbitary non-zero
element of V(~4), Choose an A ~ d such that ~(A)~0 but c~(A')=0 for every
A' D A. Then checking for A the condition (1) we conclude .f~ V*(d). |

ProofofTheorem4. In thecase n~t obviously V*[[ X]kt<=t),/ X k ) / = 0 , thus the

statement is true. We apply induction on n, simultaneously for all k, t. Let
f=a~=xCZ(A)p(A)belongto V ' [ ( X t ) , ( SIk. )_ Then we can write f=fo+.f~ where

fo = x , , ¢ / t = ( cX
x-{x })
c~(A)p(A), f~ = ~- cac(~x
=
x - {~
..
i) ~(AU{x,})p(A)x,.
Let us set f2=f~/x,. Then

(fo+f2)CV*[(X-<={7"}), (X-_<{k"})}, and

t<= k-1))"
As f=(fo+fz)+(x,-1)f2 the induction hypothesis yields the decomposition
of f into a linear combinations of polynomials in V, each of them of the form
(xi~- 1)...(xi, ÷~- 1)xi,+ :...x~, l<=k.
For]0+fz there are no problems, however for (x,-1)./~ the monotonicity is
violated for every term g in the decomposition off2 having l>t+ 1.
344 M. DEZA, P. FRANKL

For such g we can write

{ x,--1 x,--1 }
(x,,-1)g = ~ g x.-g + ( x ~ - l ) ... ( x ~ , - U ( x . - 1 ) ,
J=,+l /7 x;~ H xi,,
t+l~v~_j t+l~v~j

which procures a decomposition with the desired property.

Now we calculate the dimension of the space W=V*{( X)<=t' ( L=) ]_

=V*[(XIu{t _->~(+l)j' t h u s T h e o r e m l yields dimW=t+x~_,_~k

~ [ n' p) r ° v i n ~ t h e l
second part of the theorem.
Since in new generator systems the number of polynomials is just dim W,
they are linearly independent, i.e. they form a basis. II

Remark. The proof also shows that these polynomials form a system of generators
for I,I,_-<t)' _ as a Z-module, i.e. i f J h a s integer coefficients then it can be
obtained as an integer linear combination of the generators.
Proof of Theorem6. For Be_2 x we define matrices M~, O<_i<--t. For that let
A1 . . . . . A(,.) be a fixed ordering of the elements of ( f ) a n d B1,...,B~ of those

of~.i.e. ]~'l=m. N o w f o r l<-r<=m, l=<sN{7 } the element

mi(r,s)={O1 if A~= B~}

if A . ~ B , "

Let M be the m by + t - 1 + "" + matrix which we obtain by put-

ting side by side M, . . . . . M0. Let us denote by u~ the i'th row vector of M. In this case
seq iv tto in case

row vectors of M are not independent and consequently the rank of M is less than m.
Suppose now that ~3~2 x is such that for every BCN, ]B(~B'I takes at most
t values different from IB]. In view of the above observations it is sufficient to show
that in this case the rank of M is at least m.
To do this let vi.j denote the i'th column vector of M i, l<=i<-m, O~-j<-t,
and let W be the vector space spanned by these vectors over Q.
Let us calculate the matrices Nj = M j M f for 0<-j~ t. Nj is an m by m matrix
with column vectors w{, 1 = z = m , w~EW. The (r, s)-element of Nj is IN, B~I
l<=.
Without loss of generality we assume ]B,I>=IBsl for l<-r<s<-m. Fix r,
l<-r<=m and let ll . . . . . lp be the different values of IB,(3B~I for r<s<=m. Thus,
 0-CONFIGURATIONS 345

by our assumption p<=t, there exist rational constanst cj such that (x-ll).,.
...(x-/p) =l~_j~_,
~ ci (j) . L e t u s s e t u,=j~=oCjW~ and let N be the m by m matrix
formed by the column vectors urEW, 1 <=r<-m.
By definition N is an upper-triangular matrix with non-zero diagonal
p
(the t'th diagonal entry is ] - / ( I B , ] - / 0 ~ 0 , while the (r,s)-entry for r < s is
i=l
P
l I (]B, AB~I--l~)=0), thus N h a s full rank m. As the columns of N are from W, we
i=I
deduce dim W = r a n k M~=m. I
Open problem. Find a basis for V * ( d , ~ ) in the general case, in particular determine
dim V*(s~, N).
In the particular case ~ = [ x t }, ~ = / 2 / ~ ] a basiscan beobtained from the
basis in Theorem 4 by adding all the monomials of degree less than t.
It is not hard to see that in the case ~ c ~ we have dim V* ( d , ~ ) = ]~1 - I~¢1.

References
[1] M. DEZA, Probl6me de l'existence de ( H 1; H ~, b)-hypergraphs, Cahiers du C. E. R. O. Bruxelles,
17, 2 - - 3 ~ 4 , (1975) 185--190.
[2] M. DEZA, Isometries of hypergraphs, in Proc. Symposium on Graph Theory (Calcutta, 1976),
ISl Lecture Notes 4, (ed. A. R. Rao), Mc Millan India, 1979, 174--189.
[3] P. FRANKL and R. M. WILSON, Intersection theorems with geometric consequences, Com-
binatorica 1 (1981), 357--368.
[4] R. L. GRAHAM, S-Y. R. LI and W.-W. W. LI, On the structure of t-designs, S I A M J. Alg.
Disc. Methods, 1 (1980), 8--14.
[5] J. B. GRAVER and W. B. JURrAT, The module Structure of Integral Designs, Journal o f Com-
binatorial Theory A 15 (1973), 75--90.
[6"] D. K. RAY-CHAUDHURI and R. M. WILSON, On t-designs, Osaka J. Math. 12 (1975), 737--744.
[7] N. M. SINGHI, personal communication, 198l.

M. Deza, P. Frankl
C.N.R.S.,
54 Bd. Raspail
Paris 75006, France
 COMBINATORICA2(4) (1982) 347--359

ON THE CONNECTIVITY OF R A N D O M m-ORIENTABLE

GRAPHS AND DIGRAPHS

T. I. F E N N E R and A. M. F R I E Z E
Received 18 August 1981

We consider graphs and digraphs obtained by randomly generating a prescribed number

of arcs incident at each vertex.
We analyse their ahnost certain connectivity and apply these results to the expected value
of random minimum length spanning trees and arborescences.
We also examine the relationship between our results and certain results of Erd6s and
R6nyi.

1. Introduction

This paper is concerned with the connectedness of certain random graphs and
digraphs. The most common approach to studying random graphs is either to consi-
der a graph Gn,N with n vertices and N edges chosen at random (see Erd6s and R6nyi
[1], [2], [3]), or to fix p, 0 < p < 1, and include each possible edge independently with
probabilityp.
However, in his paper on the expected value of a random assignment problem,
Walkup [6] uses results on a random bipartite graph that is generated in the following
manner: for each vertex v, generate independently at random m distinct edges con-
taining v.
In this paper we consider general graphs generated in an analogous manner.
Our motivation for studying these graphs is that various interesting properties of
G,.N, like connectivity and the existence of hamiltonian cycles, seem to occur when N
is large enough to force a lower bound on the degrees of each vertex with probability
tending to 1. In this paper we show, for example, that choosing 2 random edges con-
taining v, for each vertex v, is sufficient to ensure connectedness (mor :over nonsepa-
rability) with probability tending to 1. In a future paper we intend to discussthe exist-
ence of hamiltonian cycles in such graphs.*
The structure of the paper is as follows: in Section 2 we consider the vertex
and edge connectivities of a random graph G~n) with n vertices where, for each vertex

AMS subject classification (1980): 05 C 40; 60 C 05.

* We have shown there exists m0 such that for m~_mo Gm is hamiltonian with probabili-
ty~ 1, [7]. We are currently trying to reduce m0 (from 23).
348 T . I . FENNER, A. M. FRIEZE

v independently, m edges are generated containing v. We show that, with probability

tending to 1 as n ~ , G~ ) has both vertex and edge connectivities equal to m for
m_->2~ and that G~"~ is not connected.
In Section 3 we apply these results to establish that the expected length of a
minimum spanning tree in the complete n vertex graph, with edge lengths indepen-
dently drawn from the uniform distribution on [0, 1], never exceeds 2(1 + l o g n/n).
In Section4 we obtain a characterisation of those graphs which can be
instances of GC,''), ; this enables us to relate our results to those of Erdd~s and R6nyi [3].
Lastly, in Section 5 we outline similar results for a corresponding class of ran-
dom digraphs.

2. Connectivity

We begin with the definition of the random digraph DG[2 }, where 1 ~m<_ -
n - 1. The vertex set of DG~ ) is V, = {1, ..., n}. The arcs of DG~p) are obtained by
independently taking each vE V, and then randomly choosing m distinct arcs (v, w),
where w( V , - { v } , so that, for each v~ V,, each of the -In 2 1/- possible sets of arcs
]
has the same probability of being chosen.
G~ ) is the random graph obtained from DG~"} by ignoring the orientation of
the arcs. (Strictly, G~"1 is a multigraph as some edges may occur twice.)
For a graph G, the vertex connectivity C,(G) is the minimum number of ver-
tices the deletion of which disconnects G. The edge connectivity C~(G) is defined simi-
larly in terms of edges.
Theorem2.1. (a) For rn>-2, !irn Prob(C~(G[2})=m)=l.
(bJ For m>=2, lira Prob(C~(G~">)=m)=l.
(c) lirn Prob (GI "} is connected)=O.

Proof. (a) Consider the deletion of p vertices from the random graph G~ ), where
0_-<p_-<m - 1 . If G~ ) can be disconnected by deleting p vertices then there exists a
partition (P, S, T) of V,, where IP]=p, [ S [ = k and I T [ = n - p - k , with m - p + l <=
k<=n-m - I, such that G O') has no edge joining a vertex in S to a vertex in T.
For an arbitrary such partition, the probability of this is

[(p-t-k-1)/(nml)f{(n-k--1)/Cn~l))"-P-k ~ fk+plk"In-kl "-p-'>''

Thus,
(2.1) A (m, n, V) = Prob (Co (G~)) ~ p)

fk + pik'V-kl
k = m"~
--p+l k!p!(n-p-k)t t---~-} I,----~1 "
Now,
n! 12n n ~+~ c~(m, n, p)n"
k!p!(n-p-k)! 12n--1 p!ePkk+~r(n-- p--k)"-P-k+ ~ ]/2--n~ k k (n - p - k)"- p- k ,
 CONNEC"rlvYrY IN RANDOM GRAPHS 349

where (m, n, p) = ( 1 2 n / ( 1 2 n - 1)) (n/( (m - p + 1) (n - m - 1)) )½/(p ! e p 1/2"~).

Thus
L'~(n- p) A
A ( m , n, p) <=~(m, n, p) k z~+~p ((k + p ) / k ) k ( ( n - k ) / ( n - p - k ) ) " - P - ~ u k

where uk = ( k + p ) ( " - a)k ( n - k)C" - ~) c. - p - ~) n " - " ( " - P).

N o w ((k +p)/k) k ((n - k)/(n - p - k))" - ~'-k <_e v × e", and (k +p)k (n - k)" - p- k decreases
monotonically with increasing k for k<-½(n-p).
Therefore,

(2.2) A (m, n, p) <- o~(m, n, p)e2P(Um_p+ l + k ( n - 2m + p - 2)u,,_r+ ~)

<- o~(m, n, p)e2P(an ~-"(m-p) + bn 3-mC"-p+ l))

where a =- ( r e + l ) (m-1)(m-p+l) and b = ½ ( m + 2 ) (m-a)(m-p+2). It follows that

lirn A ( m , n , p ) = O for m=~2 and O<-p<m. Thus lim Prob(Co(am (n)
)->m)=l.
We complete the p r o o f o f (a) by showing that

B(m, n ) = Prob (G(,") has a vertex o f degree m)-+ 1 as n ~ ~o.

This will imply that lim P r o b (C,,(Gm)<m)= 1 a n d hence the result.

Let E ('° be the event that vertex k has indegree zero in the associated digraph
DG~"). Thus the degree of k in G ~ ) is m if and only i~
"~'~(")
z~i occurs,
N o w consider SC= V. with [S[=s. Then

q(m,n,s)=Prob(,~sE (',)=
([o--)/(,
ms 1 m ,) l,', m //~. m =

--
t=l
(1-s/(n- 0)"(1 + ( m / ( n - s - m)))'.
Thus, keeping s fixed, we see that lirn q (m, n, s) -- e - m~. If we let t be a fixed positive in-

teger then,forn>=t,g(m,n)>-C(m,n, t)=Prob = ~ ( - 1 ) ~-1 q(m, n, s).

Therefore lijn C(m, n, t ) = l - ( 1 - e - m ) ', so for n>-no(t ), B ( m , n ) > = l - ( 1 - e - m ) t-
- 1/t. However, since t is arbitrary, we see in fact that lirn B(m, n) = 1, which c o m -
pletes the p r o o f o f (a).
(b) N o w Ce(G~'))~C,(G~ ")) and if G ~ ) has a vertex o f degree m then Ce(G(~"))<=m.
T h e result thus follows f r o m (a).
(c) It was shown by K a t z [4] that P r o b ( G x(') is c o n n e c t e d ) = 2 ( n - l ) ! .
11--2

2*
350 T. I. F E N N E R , A. M. F R I E Z E

3. Expected length of a minimum spanning tree

In this section we obtain some results concerning the following question : given
a complete graph CG. with n vertices and with edge lengths drawn independently and
uniformly from [0, 1], what is the expected value E,, of the minimum length of a
spanning tree of CG,?
Using an approach derived from that used by Walkup [6] for the assignment
problem, we shall show that
(3.1) E, < 2(1+1og n/n).
For the moment we consider n as fixed. For each distinct pair i,j with 1 -<i,
j<-n, let Yi~' be a random variable with distribution function
(3.2) F(2)=1-(1-2)*, 0=<2=< 1,
the Yij all being independent. For e = {i,j}, if we let Xe=min (Yij, Yji) then Xe
is unilorm in [0, 1].
Walkup observed that F(2)_->H(2)=2/2 where H is the distribution function
for a random variable uniform in [0, 2]. For each i, let Yi. tk) and Utk), respectively,
denote the U h smallest of the n - 1 random variables Yii and the U h smallest of n - 1
uniform random variables in [0, 1], then
(3.3) E(Y~ck)) <-- 2 E ( U ~ ) = 2k/n.
(In order to overcome the technical difficulty of ties in the definition of the k th small-
est of n random variables we delete the tie set, which has probability zero, from the
underlying probability space.)
The edge lengths in CG,, are obtained by sampling the Y~j and then computing
X~=min(Yii, Yji), as above, for each e={i,j}.
Given {Yii} and m-<_n-1, let Gm be the random graph obtained as follows:
for each vertex i = 1, 2 .... , n, select the m smallest of the Yo', 1 <-j<-n, and then
take the m corresponding edges {i,j}. We note that the graph G,, has the same dis-
tribution as the graph G},~') of Section 2. The length of edge e = { i , j } in G,, is Y~j
or Yji or rain (Yij, Yj~) depending on whether one or other or both of Yij and Yji are
selected.
For a given Y we obtain a spanning tree T of CG. as follows:
l. Construct G~.
2. I f G2 is connected then
(a) Construct Gt and delete the longest edge from any cycle.
(b) Connect up the remahffng forest using edges o f Gz-G1.
3. I f G.~ is not connected then choose an arbitrary spanning tree in CGn.
Let F, be the expected length of T and, for l_<-k<=[n]2J, let qk =
Prob (G~ has k components and G2 is connected). Let rc,=Prob (G~ is not con-
nected) = A (2, n, 0).
Now
knl2 J
e. r. <
 CONNECTIVITY IN RANDOM GRAPHS 351

The summation corresponds to the case when G2 is connected because if G1 has

k components then T uses n - k edges of G1, whose expected lengths do not exceed
2In, by (3.3), and k - 1 edges of G2-G1, whose expected lengths do not exceed 4/n.
The term ( n - 1 ) n , corresponds to the case where G,, is not connected, in which case
the length of T does not exceed n - 1 .
We note that n, = 0 for n =<5, and from (2.2) it is straightforward to show that
(3.4) re,<= 1/n(n--l) for n =>2.
Now qk<=pk=Prob(G~ has k components) and thus
knl21
E, < • (2(n-k)/n+4(k-1)/n)pk+(n-1)n,=
k=l
kn/2Z [ kn/22 "1
= 2-4/n+2 ~y kpk/n+(n--1)n, < 21,1+ k__~
k=~ i= kpk/nJ, by (3.4).

Now Kruskal [5] shows that

Ln/SJ 2
X kPk ~ nI/((n--j)ljnO < log n
k=l j=l

(3.1) now follows immediately.

It seems reasonable to conjecture that E~ is monotonically increasing in n,
in which case (3.1) would imply E,<=2 for all n.

4. m-orientable graphs

Every instance of the graph G,(,") has n vertices and mn edges where an edge
can be repeated at most twice and each vertex has degree at least m. It is not true,
however, that these properties characterise G}"). In this section, therefore, we obtain
such a characterisation.
We first consider graphs without repeated edges. Given a graph G=(V, E),
an orientation f2(G) of G is a digraph (V, f2(E)) in which
(4.1a) (v, w)EI2(E) ~ {v, w}~E,
(4.1b) {v, w}~E:=~ I{f2(E)('l {(v, w), (w, v)}[ = 1,
i.e. each edge of E is oriented by f2.
Now let c~: V ~ Z +, the set of non-negative integers. The graph G is c~-
orientable if there is some orientation f2(G) for which the outdegree of v is at least
c~(v) for all vC V. (Our aim is to find a way of determining whether a graph G is a
G(m"), i.e. whether it can be oriented to yield a DG~,",).)
For a set S ~ V, we let c~(S)= ~Yc~(v). We now obtain a characterisation
of a-orientability. ~ s
Theorem 4.1.* A graph G=(V, E) is c~-orientabte if and only if, for all SC= V,
(4.1) I{eCE: eNS # O}I >=c~(S).

* This result was also obtained by Frank and GyfirfS.s[8].

352 T. I. F E N N E R , A. M. F R I E Z E

Proof. The necessity of 4.1 is evident, for if f2(G) is an a-orientation then at least
e ( S ) arcs must have their tails in S.
To prove sufficiency, we consider the following maximum flow problem (see
Figure 1):

V E

/
/

s
.f// v (v.wl
~t

Fig. I

Let D be the digraph (X,A) where X = { s , t } U V U E , A=A1UA~UA~,

= {(s, v):
A~ = {(v, e): vEeEE},
A3 = {(e, t): eEE}.
The flow capacities c: A-~Z + are defined by

i(v) if a = (s, v)EA1,

c(a) = if aEAz,
if aEA3.
We observe first that G is e-orientable if and only if the maximum value of an
s-t flow in D is e(V).
Condition (4.1) ensures that a cut separating s and t which contains no arcs of
As has capacity at least e(V). Then, on applying the Max-Flow-Min-Cut Theorem
of Ford and Fulkerson, we obtain the result. II
In order to deal with multiple edges, we now define a graph as G=(V, E, I~)
where V, E are as before and p: E ~ { 1 , 2} gives the edge multiplicities. An edge
e = {v, w} with /~(e)=2 should be oriented in both directions to correspond with
the way double edges arise in obtaining G(~n) from DG~n). We therefore change the
definition of an orientation, replacing (4.1b) by
(4.1b') {v, wIEE =, [12(E)71 {(v, w), (w, v)}[ = tt(e).
The definition of e-orientable remains unchanged. It now follows easily that, if
eEE and #(e)=2, G = ( V , E , p ) is e-orientable if and only if G'=(V,E-{e},u')
is e'-orientable, where e'(v)=e(v) for vC~e and e ' ( v ) = m a x (0, e ( v ) - I ) for vEe,
and #' is the restriction of p to E - {e}. On combining this remark with Theorem 4.1
we obtain
 CONNECTIVITY IN RANDOM GRAPHS 353

Corollary 4.1. A graph G=(V, E, p) is ~-orientable if and only i[, for all SC= V,
(4.2) I{eEE: eOS ~ 0}l+ • (#(e)--l) ~c~(S). 1
eC=S

If m is a positive integer, we use the term m-orientable to denote c,.-orientable

where c,.(v)=m for all vE V. Hence we see that the graphs G~.") generated as in
Section 2 are precisely the m-orientable graphs for which ~ I~(e)=mn.
eEE
Erd6s and R6nyi [3] have shown that, for a fixed integer m=>0 and real c,
if N=N(n)=½nlogn+½mnloglogn+cn and Go,N is a random graph with
n vertices and N edges then
(4.3) lifn Prob (C~(G..N) = m) = 1 -- exp (-- e-ZC/m !).

This is derived by showing that, with probability tending to 1, Cv(G.,r~)>=m

and with probability tending to 1 - e x p (-e-2C/m!) G.,N has a vertex of degree m.
We next show that
(4.4) lira Prob (G,,.u is m-orientable) = 1,

which combined with a modification of Theorem 2.1 for the case with no repeated
edges, leads us to conjecture that the m-connectivity of G,,~ could be deduced from
its almost certain m-orientahility, provided m ~ 2 .
Now if G,,N is not m-orientable, by Theorem 4.1 there exists a set S of s
vertices which meet k edges, where k <-ms-1. Thus Prob (G,,u is not m-orientable)<-
An, where

~.r,¢

the range of summation. We show in an appendix, that lifn A,=O, which yields
the desired result (4.4).

5. Digraphs
Our results on the connectivity of G}2) extend easily to digraphs. The random
digraph D~") has vertex set V,. The arcs of D[2 ) are obtained by first independently
taking each v~c V,, then randomly choosing m distinct arcs (v, iv) where wE V , - {v}
and finally randomly choosing another m distinct arcs (w', v) where w'E II,-{v}.
One can easily show that, if .4(m, n, p) is the probability that there exists a set
of p vertices whose deletion from D[2 ) leaves the remaining digraph not strongly
connected, and if/~(m, n) is the probability that" 1_,,,
'~(") contains a vertex with outdegree
m, then
(5.1a) A(m, n, p) -<_ 2×[R.H.S. of (2.1)],
(5.1b) /~(m, n) = B(m, n).
 354 T. L FENNER, A. M. FRIEZE

We thus have
Theorem 5.1. (a) For m~2, ,!!m Prob (SCo(D¢,",))= m ) = 1,

(b) For m>=2, lira P r o b ( S C e t"D m

n ~ (n)]~= m ) = 1

where, for any digraph D, SC~(D) is the minimum number of vertices that must be
deleted in order that the remainhTg digraph be not strongly connected. SCe(D) is de-
fined similarly in terms of arcs (directed edges). II
It is at present not known whether or not the probability that D[ ") is strongly
connected tends to 1.

Minimum length arborescences

The results on the expected length o f a minimum spanning tree can be partially
generalis:d to the expected length of a minimum spanning arborescence rooted at
vertex 1. (An arborescence rooted at vertex 1 is a digraph in which vertex 1 has
indegree 0, every other vertex has indegree 1, and the graph obtained by ignoring arc
orientations is a tree.)
Assuming that we are given a complete digraph CD, with n vertices and with
arc lengths drawn independently and uniformly from [0, 1], we shall show that the
expected length E, of a minimum spanning arborescence rooted at vertex 1 is less
than 3.
For each distinct pair i, i with 1 <_-i, j_<-n, let Y~j and Z~j be independent
random variables with distribution function (3.2). Let X~j=min (Yii, Zij) be the
length of the arc (i, j ) in CD,. Let Dr, be the random digraph obtained as follows:
for each vertex i = 1, 2 .... , n, select the m smallest of the Y~j, 1 ~_j<-n, and then
take corresponding arcs (i,j) and similarly select the m smallest of the Zj~ and take
the corresponding arcs (j, i). We see that the digraph D,, has the same distribution as
the digraph D~ ').
For given Y and Z, we construct a spanning arborescence R as follows: if
D2 is strongly connected then it contains at least one spanning arborescence rooted
at vertex I so we choose the shortest of these, and if D2 is not strongly connected
then we choose an arbitrary spanning arborescence in CD,.
Let F, be the expected length of R, then we see that

~, ~_ F. <=3(n-1)/n+(n-1)t~,

where r2,=Prob(D2 is not strongly connected)=A(2, n,O). The term 3 ( n - 1 ) / n ,

which corresponds to the case when Dz is strongly connected, is n - 1 times the ex-
pected length of an arc of D~.
The result now follows as 3/n>(n-1)~, since ~,<-2rr,~_2/n(n-l) by
(5.1) and (3.4).
As in Section 4 we pursue a characterisation of the digraphs D~,"). This time,
given ~,/~: V ~ Z + and a digraph D=(V, A), we determine whether there is a
 CONNECTIVITY IN R A N D O M G R A P H S 355

function ~o: A ~ V such that

(5.2a) ~o((v, w))E {v, w} for (v, w)EA,

(5.2b) ]{(v, w)EA: q~((v, w)) = v}l _->~(v) for vEV,

]{(w,v)EA: q)((w,v))=v}l _->fl(v) for vEV.

The digraphs D~"~ without repeated arcs satisfy (5.2) with et=fl=e m, where
c,,(v)=m for vE V.

Fig. 2

By considering the maximum flow problem (see figure 2):

Vertices: {s, t } U V U V ' U A , where V" = {v': vEV}
Ares: A1U A~ U Aa
A1 = {(s, v): vEV}U{(s, v'): v'EV}, Capacity: c(s, ~) = ~(v)
c(s, v3 = 11(~)
A2 = U {(v, a), (w', a)}, Capacity: ~o(_---~(v)+~(v)+0
(v, w) £ A

A3 = {(a, t): aEA}, Capacity: 1, we obtain

Theorem 5.2. Given ~ and fl, there exists q~ satisfying (5.2) i f and only i f for all
S, Tc= V we have
(5.3) I(v,w)EA: yES or wET}I ->~(S)+fl(T). 1
Repeated arcs are dealt with as in section 4, i.e. by considering the reduced
problem obtained by ignoring repetitions and modifying c~and fl appropriately. This
leads to an extra term ~'(#(v, w) - 1 ) on the left side of (5.3), where the summation is
over all (v,w)EA with yES and wET.
 IiV

I
0
+ o tl

I
¢}

~ + N} I
o~ {'}
~m

II ?
F e +

~T~ ~ ~
r~ ~x 0
IIA ml
IIA ~ , + + IIA z
c,l B1
v I
~ J ~
ttA ~ .>
.~ r~ I
]Iv ~ ~ 0"~1
I o ml
[ I IIV .~, IIA
I I
N
I + + ~ I IIA
N-"
%
+ - % + + I
I
x ,,

t~
I1 "
I
M
IIV
"" 1~/

%
0
O~
[IV
 CONNECrlVrrY IN RANDOM GRAPHS 357

N o w let

where
A" = ,2 ,~,1
nlZ n s
<= ,wiZ ~ ms R, _il
2
PQ

p =
N- ms + 1 =

for some ~ > 0 when n is large, and where

Q• t=O (2}__tJ [-~j_<-{x,~)

~_ exp (-- s log n - s m log log n -- 2cs) = n -~ (log n ) - " s e-2%
Also

and so
R; il
2 < (s(n-½s-½))
= ( m s -- 1) !
"~-i
~_ (2en)m~-i

, m (2ae)m~_ i = O
A,<_- = ( s - l ) ! logn
and so A~-~0.

N o w let A"= .Z v~ . Now

n
s=--ff+l

v~+___.L= __n-- s U V
0s S
where

U=H
t=O
m-1

{nTs}-(N-m(s+ l)+t+ l)
[ N-m(s+l)+t+2]

~ctIsl°gn] m
k n--S ]
358 T . I . FENNER, A. M, FRIEZE

for some ~ > 0 when n is large and

n--s--2
V= 17
I [n2 s}-t-(N-ms)
~ -----7-dn- - - ~ ~ "
t=O
) ÷'÷1

for some ~>0 when n is large.

Thus
"Ds+a ( S ,~m-1 n
(log n)"_~ ~fl (log n)" =~ 1 =2
and this is =~1, for large n.
Thus

(AI) zl': ~= ½ nv=~.~

From its definition a=a(n)satisfies

n - a - - l } < N-m(a+l)+ 1 or
2
(A2)
(n2a)-(n-a-1)<(N-ma+l)-m
from which we deduce r-x(n2aj_(N_ma+ 1 ) < n - a .

From (A2) we can also deduce that

(n-a)Z-3(n-a)+2 -<2(N-m(a+l)+l)
from which we obtain n-a<=21/2-N
assuming n large.
It then follows that

,N--ma+[
( n 2 a ) I1 ~ [ ( [ 2 1 ) ] -~ ( 2y2-N)N as n--a~--½(2n a) for large n
 CONNECTIVITY IN RANDOM GRAPHS 359

and so

( )
Itfollowseasilyfrom(A1),since[(2)]ismuchlargerth
vo ma -- "

Thus An-~0 as was to be shown.

Note added in proof. A. M. Frieze has recently improved the result of Section
3 by showing that En-+ ~ (3).

References

[1] P. ERD6S and A. RANvl, On Random Graphs I, Publ. Math. Debreeen, 6 (1959), 290--297.
[2] P. ERD6S and A. R~NYI, On the Evolution of Random Graphs, Magyar Tud. Akad. Mat.
Kut. Int. K6zL 5 (1960), 17--61.
[3] P. ERD6S and A. R~NvI, On the Strength of Connectedness of a Random Graph, Acta. Math.
Acad. Sci. Hungar., 12 (1961), 261--267.
[4]. L. KATZ, Probability of Indecomposability of a Random Mapping Function, Annals of
Mathematical Statistics, 25 (1955), 512--517.
[5] M. KRUSKAL, The Expected Number of Components under a Random Mapping Function,
American Mathematical Monthly, 61 (1954), 392--397.
[6] D. W. WALKUP, On the Expected Value of a Random Assignment Problem, S I A M Journal
on Computing, 8 (1979), 440---442.
[7] T. I. FENNER and A. M. FRIEZE, On the existence of Hamiltonian cycles in a class of random
graphs, to appear in Discr. Math.
[8] A. FRANK and A. GY/mFLS, How to orient the edges of a graph?, 1976 Colloq. Math. Soch. J.
Bolyai 18 (1978), 353--364.

T. I. Fenner A . M . Frieze
Dept. of Computer Science Dept. of Computer Science
Birkbeck College and Statistics
University of London Queen Mary College
London WC1E 7HX London E1 4NS
England England
 COMBINATORICA 2(4) (1982) 361--371

DISJOINT PATHS IN A RECTILINEAR GRID

Andr~is F R A N K *
Dedicated to Tibor Gallai on his seventieth birthday

Received 16 November 1981

We give a good characterization and a good algorithm for a special case of the integral
multicommodity flow problem when the graph is defined by a rectangle on a rectilinear grid.
The problem was raised by engineers motivated by some basic questions of constructing printed
circuit boards.

1. Introduction

One of the central topics in graph theory concerns the existence of disjoint
paths under various constraints. Basic results are due to Gallai [3, 4], Tutte [13],
Menger [10], K6nig [6], Mader [9], Seymour [12], Lov~sz [8]. In the present paper we
discuss the following problem in this field.
Given an undirected graph, find k edge-disjoint paths between k pairs of ver-
tices specified in advance. This problem, often called the disjoint paths problem,
is a specialization of the integral multicommodity flow problem which belongs to the
hard class of NP-complete problems even if k = 2 [5].
For the disjoint paths problem there is a good characterization, due to Seymour
[12], for k = 2 , but the question is open for any fixed k=>3 [5]. However, the follow-
ing special case, related to wiring problems of printed circuit boards, can be well-
characterized. The proof also provides a good algorithm.
In a rectilinear grid (or plane lattice) we are given a closed rectangle T (boun-
ded by lattice lines) and k pairs of distinct lattice points of the boundary. T defines
a finite subgraph G of the plane grid in the natural way (which has n. m vertices when
m horizontal and n vertical lines intersect T). The purpose of this paper is to give an
answer to the disjoint paths problem in this special case.
The problem was formulated by I. Abos, an electrical engineer who (with
his co-workers) worked out a general computer program for designing printed cir-
cuits boards [1].

* Research partly supported by Sonderforschungsbereich 21 (DFG), Institute ffir Operat-

ions Research, Universit~it Bonn, West Germany.
AMS subject classification (1980): 90 B 10, 68 C 25; 68 E 10.
362 A. F R A N K

It should be noted that the engineering literature is quite rich in works concern-
ing such designing questions and a large number of models, approaches, and algo-
rithms are offered. However, the great part of these procedures uses heuristics and
a typical conclusion o f such a paper is that "the computer program was able to cons-
truct 90--95 per cent of the wirings". Of course, such programs may work quite well
in practice, but mathematically well-founded models may help to reach further impro-
vements.
In the mathematical context our problem is strongly related to that of Oka-
mura and Seymour [11]. The precise relationship will be discussed in Section 4.
In the first part of the paper we consider the even more special problem when
one member of each pair is on the upper horizontal line while the other one is on the
lower line. We refer to this case as bipartite. The non-bipartite case, when the termi-
nals are allowed to be arbitrarily positioned on the boundary of T, will be presented
in Section 3.
By a column (row) of T w e mean a region in Tbetween two consecutive vertical
(horizontal) lattice lines. The path congestion (or congestion) of a column e is the num-
ber of terminal pairs separated by c. We cal! a lattice point exposed if it is not a ter-
minal. An obvious necessary condition for a solution is the column criterion: the
congestion o f any column is at most m, the number of horizontal lines.

2. The bipartite case

Unfortunately the column criterion is not sufficient in general, as each of the

next examples shows.

1 2 1 2 1 23 1 23 & 5 67

21 2 1 I
I
231 2563 71 /~

Fig. 1

Surprisingly, a small restriction makes the column criterion sufficient. Namely,

we have the following result.

Theorem 1. I f at least one corner point of T is not a terminal vertex, the column criterion
is necessary and sufficient for the existence of edge-disjoint paths between the corres-
ponding terminals.
Proof. We prove the suffiency. Suppose that the four corners of T are A = ( I , m),
B=(n, m), C=(m, 1), D = ( I , 1). For a terminal pair i let U(i) and L(i) denote the
x-coordinates of the upper and lower member of i, respectively. We shall adopt the
notational convention that the x-coordinate of a point P will be denoted by the same
term P.
It may be assumed that m is equal to the maximal congestion A. We proceed by
induction on m. The case m = 1 being trivial, let m > 1. Assume that point A (the left
upper corner of T) is exposed.
 D I S J O I N T PATHS IN A G R I D 363

We can assume that L(i)~ U(i), that is no trivial path exists. For otherwise,
join them by a vertical path, then remove this lattice line from the rectangle. In the
reduced problem the column criterion holds again and A is also exposed.
The algorithm will consist of m phases. During each phase, both the number
of horizontal lines and the maximal congestion will be reduced by one. We describe
and analyse only the first phase. This will result in a sequence U'(i) (i= 1, 2, ..., k)
of distinct points which tells path i where to go to on the upper line. In particular,
U" (i) = U(i) means that path i will start vertically down.
In order to have a valid algorithm and proof, the sequence U'(i) (i= 1. . . . . k)
must be determined in such a way that the induction hypothesis should hold for the
terminal pairs U'(i), L(i) (i= 1, 2 . . . . . k) on the rectangle T" consisting of one less
horizontal lattice line (i.e. the corners of T' are A ' = ( 1 , m - 1 ) , B'-=(n,m-1),
C=(n, 1), D = ( 1 , 1)). In other words, the new maximal congestion should be m - 1
and one corner point of T' must be exposed again.
Our procedure has the feature that the upper corner points are alternately
exposed, that is, at the end o f the first phase the right upper corner of T" will be expo-
sed, after the second phase the left upper corner will, and so forth.
A path i is said to be a left-path (right-path) if U(i)>L(i) (U(i)<L(i)). Let
X, Y be two points on segment AB with X < Y. The basic step of our procedure is:
"left-pushing" on XY. Assume that X is exposed. An elementary left-pushing on X Y
tells a single left-path how to go on the segment XY. Decide whether there exists a left
path i for which X < U(i)~= Y and if so, select one, say j, for which L ( j ) is as small
as possible. (If no such path exists, the elementary left-pushing is called vanishi~g.)
Set U ' ( j ) = X if L(j)<=X and U ' ( j ) = m a x ( Z : X~-Z<=L(j), (Z, m) is exposed)
if L (j) > X.
Note that U ' ( j ) is well-defined since X was exposed. Furthermore j becomes
trivial for the second phase if Z = L ( j ) and becomes a right path if Z-<L(j). In
this latter case the congestion of columns between Z and L ( j ) has been increased by
one. Moreover, by the minimality of j, each point Z + 1, Z + 2 .... , L ( j ) on the upper
line is a terminal of a right path.
A left-pushing on X Y tells several left-paths how to go on the segment XY.
First, apply an elementary left-pushing on XY. Assume that p a t h j has been moved
to the left. Then the point X ' = U ( j ) on the upper line has become exposed. Apply
now an elementary left-pushing on X ' Y and repeat this procedure until the actual
left-pushing is vanishing.
An elementary right-pushing is defined analogously.
The first phase of the algorithm consists of two parts. First, apply a left-push-
ing on the whole AB. In the second part consider each maximal subsegment X Y o f A B
which has not been covered by any path in the first part and put Y ' = m a x (Z: X ~
Z < Y, (Z, m) is exposed). (By definition of a left-pushing, X has been exposed and
thus Y' does make sense.) For each XY" apply a right-pushing on XY'.
The description of the first phase of the algorithm is now complete. We set
U" ( j ) = U(j ) for each path j having not appeared in the left or right-pushing of the
first phase.
To prove the correctness of.the algorithm, first observe that U'(i)~ U'(j)
when i ¢ j and the subpaths which have been defined on AB are edge-disjoint. Furt-
hermore, one can see that, after deleting the trivial paths which have arisen, the right
364 A. FRANK

upper corner of the rectangle is exposed. Let us prove now that the maximum con-
gestion in the reduced problem is one less.
The congestion of any column was changed by 0, + 1 or - 1. First, consider
those columns whose congestion increased by one. This might occur only in Part
1, if, while accomplishing an elementary left-pushing on XY, we had L ( j ) > X and
L ( j ) > Z . As mentioned, the congestion of each column between Z and L ( j ) has
increased by 1.
Claim 1. The original congestion of any column between Z and L ( j ) is at most A - 2 .
Proof. Since the points ( L ( j ) , m ) and ( L ( j ) , 1) are terminals o f a right-path
and a left-path, respectively, we have c ( L ( j ) - 1, L ( j ) ) = c(L(j), L ( j ) + 1 ) - 2 ~ A - 2
where c(S, S + I ) stands for the congestion of the column between S and S + 1.
We know that all points Z + 1.... , L ( j ) on the upper line are terminals of
right paths. Therefore c(Z, Z + 1 ) ~ c ( Z + 1, Z + 2 ) N ... <=c(L(j)- 1, L ( j ) ) ~ A - 2 ,
as required. II
The congestion of any column, which separates the terminals of a left-path, is
reduced by 1 in the first part. So we have to deal with the maximal segments XYwhich
have not been covered in the first part. Recall the definition of Y'.
Claim 2. The (unchanged) congestion of each column between Y" and Y is less than A.
Proof. For Y ' < Z < Y each point (Z, m) is a terminal of a right path and thus
c(Y', Y ' + 1)<-c(Y'+ 1, Y ' + 2 ) < - . . . <=c(Y- 1, Y). One can see that the point (II, m)
was originally exposed. We have three possibilities for Y.
(i) Y= B, that is Y is the right upper corner of T. Then

c ( Y - 1 , Y) <=I < m = A,

(ii) (Y, 1) is a terminal of a left path in which case

c ( Y - 1 , Y) < c(Y, Y + I ) <- A,
(iii) (Y, 1) is not a terminal of a left path and Y<n. ByClaim 1
c(g, Y + I ) <= a - 2 .
Since c ( Y - 1 , Y)<=c(Y, Y + I ) + I the claim follows. II
Claim 3. The new congestion of any column between X and Y" is at most A - 1.
Proof. Assume that the column (Z, Z + 1) violates the claim (X<=Z< Y'). Then
(Z, Z + 1) originally separates A terminal pairs of right paths. If anyone of these has
an upper terminal U ( i ) > X then one path has crossed the column (Z, Z + 1) in the
second part of the algorithm and therefore the new congestion of (Z, Z + 1) is less
than A, contradicting the assumption.
In the other case U ( i ) < X for all the A (right) paths i separated by (Z, Z + 1).
Then X can not be A and (X, m) is a terminal of a left path while (X, 1) not. Hence
c ( X - 1, X ) > c ( Z , Z + 1 ) - A , a contradiction. II
 D I S J O I N T PATHS I N A G R I D 365

The p r o o f of Theorem I is now complete. II

Remark. Observe that a non-boundary vertical lattice line can be deleted if no ter-
minal point is on it since the column criterion continues to hold. So we can assume
that m-< k, n _<-2k. Therefore the complexity of the algorithm depends only on k and
not on m and n.
Theorem 1 does not answer the case when none of the four corners is exposed.
In the first three examples in Figure 1 the reader can easily convince himself or herself,
that no solution exists. But what simple reasons may prove the non-existence of the
solution in the fourth example? Theorem 2 provides an answer. In order to incorporate
Theorem 1, it will be convenient to consider the points (0, m), (n + 1, m) as fictitious
exposedpoints.

Theorem 2. (a) I f there is no (proper) exposed po&t on the upper l#w, the problem
has a solution i f and only if each path is trivial.
(b) Assume that there is at least one proper exposed point.
bl. I f m < A , there is no solution.
b2. l f m > A, there always exists a solution.
b3. I f m = A , there exists a solution if and only if there are two exposed points
(fictitious or not) on the upper line such that the congestion of any column between them
is less than A.

(Observe that Theorem 1 is included. I f m = A and A is exposed, the "column"

between 0 and A is of zero congestion.)
ProoL Part (a) is trivial. We have seen bl. To see b2, let (X, m) be an exposed point.
Apply a right-pushing on AX. Then we get a new problem on a rectangle of m - 1
horizontal lines. But now the left upper corner became exposed and m - 1 _->A so
Theorem 1 applies.
b3. Necessity. Assume that the exposed points J~o, X~ . . . . , X t + 1 (t_>l) on
the upper line are separated by saturated columns (columns of congestion A) which
are (Z1, Z I + 1)..... (Zt+ 1, Z,+~+ 1). Each of the sets {1, 2 . . . . . Z1}, {Z~+I .... , Z~},
{Zt+ a + 1. . . . , n} on the upper line has the property that the number of terminals in it
and the number of edges leaving it have different parity. Therefore no solution uses
every edge leaving such a set. But the edge which is not used cannot be (Zi, Z i + 1)
since each edge of a saturated column must be in a path. Hence these edges are vertical
edges of the uppermost row ( m - 1, m). This row contains n vertical edges. On the
other hand, it must contain at least t + 2 edges not in any path and n - t edges which
are in a path. This is impossible.
Suffiency. Suppose X and Y are two exposed points not separated by saturated
columns. I f Y is proper, first apply a left-pushing on YB then apply right-pushings on
the maximal non-covered segments of YB. I f X is proper, do the same on A X by
changing the terms "left" and "right". We get therefore a problem in a rectangle of
m - 1 horizontal lattice lines in which the maximal congestion is A - 1. This latter
follows from the p r o o f o f Theorem I applied to the segments A X and YB and from
the hypothesis for the segment XY. Moreover, the left (right) corner became exposed
i f X ( Y ) was proper. Apply Theorem 1. II
Return to the examples in Figure 1. The first and third do not satisfy condition
(a). The second and fourth example do not satisfy b3. (See Figure 2.)

3*
366 A. F R A N K

f exp°~ed P°ints~ ' , ~ exposed points

4- 1 2 + + I
I ; I I I 6#3 tl +
2 1 2 5 # 7 1 4

s a t u r a t e d columns soturflted columns

Fig. 2

3. The non-bipartite case

Henceforth we assume the terminals are positioned arbitrarily on the boundary

of T. Furthermore, the corners are allowed to be assigned to two terminals. The two
examples in Figure 3 may indicate some difficulties of findir~g a good characterization
lbr this problem. One of them has a solution but the other has not.
2 3 4 5 2 3 4 5

1
1

5 I 4 3 2 5 1 4 3 2

Fig. 3

We need the notion of odd sets. To this end, join each terminal pair by a new
edge. A subset X of vertices o f G is called odd if, in the extended graph, the number
of edges (new or old) leaving X is odd. (In the sequel we do not need the new edges any
longer.)
A simple parity argument shows that no solution uses every edge leaving X.
Call a cut of G saturated if its congestion is equal to the number of edges in it.
Remark that columns and rows define cuts of special kind.
Let {rl, r~ . . . . , r,} be the set of saturated rows (t~O) and let c be any column.
These define t + 1 disjoint sets Ti on the left-hand side of e. See Fig. 4.

[ D~%U" .......
.,~ Y / / / S > ~ ' J ,
,[ ~>>/};¢>/A~///////////~.
t~

F]
gd
EJ
r1

r2 (t=2)

Fig. 4
 D I S J O I N T PATHS IN A G R I D 367

The number of odd sets T~ is called the parity congestion of c.

The revised congestion or r-congestion of a column c is the sum of the parity
congestion and the path congestion of c. (Note that the r-congestion would be the
same if, in the definition of T;, we had said "right-hand" instead of "left-hand".)
Revised cohtmn criterion. The r-congestion of any column is at most m.
The revised row criterion is defined in an analogous manner.

Theorem 3. We are given a rectangle in a rectilinear grid and k pairs of terminals on

its botmdao~. There exist k edge-disjoint paths between the corresponding terminals if
and only if the revised row and cohlmn criteria hold.
Remark. This theorem involves Theorem 2 but its proof provides a much less efficient
(though polynomial-bounded) algorithm than that described in Section 2.

Pro@ Necessity (of the revised column criterion). If T i is odd, at least one edge e leav-
ing T; will not be used in any solution. Edge e is in column c, for othervise e would be
in a saturated row which is impossible. Therefore the number m of edges in c should
be at least the number of odd sets T i plus the congestion of c, i.e. at least the r-congest-
ion of e.
Su~cieney. The theorem is trivial for m = 1, or n = 1 so suppose that m, n > 1.
First we prove the theorem when there are no odd sets at all. In this case the congest-
ion and r-congestion are the same. Instead of rectangles, we prove the assertion for
n-rectangles. By a near-rectangle or n-rectangle we mean a region R bounded by the
segments AY, YX, XB, BC, CD, DA where A = ( 1 , m - 1 ) , Y=(v,m-1), X=
= ( v , m ) , B=(n,m), C=(n, 1), D = ( t , 1). Here n>v>-I and m > 2 . See Fig. 5.

Fig. 5

The row (column) criterion is now slightly modified: the congestion of any
row (column) r should be at most the number of edges in r.
The convex corners A, X, B, C, D may be assigned to either two or no termi-
nals. The corner Y is assigned to no terminal and any other point of the boundary
is assigned to exactly one terminal. We say that an n-rectangle is satisfactory if these
assumptions and the row and column criteria hold.

Lemma. In satisfactory n-rectangles there exist edge-disjoint points between the cor-
responding terminals.
Proof. By induction on the area of R (which is integer). We are using a 'cutting off'
operation at X which replaces R by R' defined by X ' = ( v + 1, m) Y' = ( v + 1, m - 1),
368 A. F R A N K

A ' = A , B'=B, C ' = C , D ' = D if v < n - 1 and b y A ' = A , B ' = ( m - l , n ) , C ' = C ,

D ' = D if v~-n - 1. Moreover, a cutting offoperation may change some terminals and
introduce some new ones. Our general strategy is that we determine a cutting off
in such a way that R' should be satisfactory and a solution in R' (that we have by the
induction hypothesis) should imply a solution in R.
For a terminal pair i let i(1) and i(2) denote the two members of i.
For two points V, W ~ X of the boundary we say that W < ' V if X W V is
clockwise on the boundary.
Case 1. X is not exposed. Then two terminals are in X, say i ( 1 ) = j ( 1 ) = X . Suppose
that i(2)<-'j(2). Set i ( 1 ) : = X ' and j ( 1 ) : = Y if v < n - 1 and i(1):=B', j ( 1 ) : =
: = ( n - l , m - 1 ) if v=n.
Claim. R" is satisfactory.
We prove the row and column criteria. The only row which can violate the
row criterion is (m - 1, m). In this case j (2) is on the segment X'B and so is i(2).
Furthermore this row was saturated in R. But this is impossible.
The only column which can violate the column criterion is c=(v, v + l ) .
Then both i(2) a n d j (2) are on the left hand side of c and c was saturated in R. This
is impossible if v = 1. If v > I, the congestion of the column ( v - 1, v) is at least that
of(v, v + 1), that is the column ( v - 1, v) was oversaturated in R, a contradiction. |
From a solution in R' we obtain one in R by extending the paths i and j with
the edges (X, X') and (Y, X), respectively.

Case 2. Xis exposed.

Subcase 2.1. Neither the column (v, v + 1) nor the row ( m - 1, m) is saturated.
Cut off X and let i be a new terminal pair with i ( 1 ) = X ' , i(2)= Y. Obviously
R" is satisfactory and omitting the path i from the solution in R' we obtain a solution
in R.
Subcase 2.2. The row ( m - 1, m) is saturated.
Then there is no terminal p a i r j with both j (1) and j (2) on the segment XB.
For notational convenience suppose t h a t j (2) is not on (X, B) for anyj. Let v < n - 1.
Choose the terminal pair i in such a way that i(1) is on XB and i f j (1) is on XB then
i(2)->'j (2). Cut off X and replace the terminal pair i by i' and i" where i'(1)=i(1),
i ' ( 2 ) = X ' , i"(1)=Y, i"(2)=i(2).
If v = n - 1 then i ( 1 ) = j ( l ) = B for some i a n d j . Suppose that i(2) _<-~/(2).
Cut o f f X a n d set i(1):=(n, m - 1), j (1)= Y.

Claim. R' is satisfactory.

Proof. The only danger may arise when X is less than the X-coordinate of i(2) and
there is a column c separating X and i(2) such that c was saturated in R. But the choice
of i implies that j (2) is on the right-hand side of c whenever j (1) is on XB. Thus the
congestion o f c may be at most m - 2 . II
From a solution in R" we obtain one in R by gluing the paths i', YXY', i".
 DISJOINT PATHS IN A GRID 369

Subcase 2. The column c = ( v , v + 1) is saturated but the row ( m - I, m) is not.

Suppose that j (1) is on the right-hand side of e and j (2) is on the other for all
j~ ~ , where .@denotes the set of terminal pairs separated by c.
Choose i ~ in such a way that j (1)<='i(1) for any j ~ .
Cut off X and replace i by i' and i" where i ' ( 1 ) = i ( 1 ) , i ' ( 2 ) = Y, i " ( 1 ) = X ' ,
i " ( 2 ) = X ' . (The case v = n - 1 is left to the reader.)
Claim. R" is satisfactory.
I f Uis a member of a terminal pair, denote the other by t(U).
Proof. The only danger may arise when i(1) is on CD and there is a row r---
= ( # , kt+ 1) above i(1) and i(2) which was saturated in R. By a simple induction we
can prove that each column is saturated on the left-hand side of c and t(k, 1) is on
the right-hand side of e for k = 1, 2 . . . . , v. Consequently the column (l, 2) is satura-
ted, there are two terminals in D and no terminal in C and no terminal pair j with
j (1), j (2) on CD. Using this, it can be shown that each row above r is saturated contra-
dicting the assumption that ( m - 1, m) is not saturated. |
From a solution in R" we obtain one in R by gluing the paths i', Y X Y ' , i'.
The p r o o f of the lemma is complete. |
Turning to the general case, observe that a set is odd if and only if it contains
an odd number of odd vertices. Furthermore, in our case only the vertices on the
boundary may be odd. What we are going to show is that there exists a pairing of the
odd vertices such that the column and row criteria will hold if these pairs are consi-
dered as new terminal pairs to be connected. In this case the Lemma can be applied.
Suppose that the saturated rows and columns are rl, r~ . . . . . rt (t~O) and
Cl, c~ . . . . , cs (s>=O), respectively. These rows and columns divide the graph into
( r + 1). ( s + 1) disjoint parts P~. It can be seen that each set Pi is even. Condiser one
Pi in which the odd points are X1, X2 . . . . , Xk (k is even) in this orde ron the bound-
ary of T. Pair them in the natural way, that is, the pairs are )(1 X2, Xa X4, ..., XR-1Xk.
Claim. The row and column criteria hold i f these pairs are new terminal pairs to be
connected.
Proof. We prove the column criterion. Let e be any column. If c is saturated then it
does not separate any new pair and thus c will satisfy the column criterion.
Let c be not saturated. Assume that c separates at most two new pairs. In this
case c will violate the column criterion only if its congestion is one less than m and c
separates exactly two new pairs. But this is impossible because of parity reasons.
I f c separates more than two new pairs then t > 0 and s = 0 .
Observe that in one P~ there may arise at most one new pair which is separated
by c and one does arise if and only if the corresponding T~ (see Figure 6) is odd. This
means that the number of new pairs separated by c is just equal to the parity conges-
tion of c. Now the revised column criterion implies the claim. |
The p r o o f of Theorem 3 is complete. |
Turn back to the examples in Figure 3. The first example has no solution since
column c violates the revised column criterion. The second example has a solution.
See Figure 7 (next page).
370 A. FRANK

Remark. The proof of Theorem 3 involves an algorithm. In the first part we have to
pair the odd vertices which is easy. The second part consists of the applications of
about m • n cutting off operations. So the complexity is proportional to m • n compar-
ing with the algorithm in Section 2 whose complexity depends only on the number of
paths to be constructed.
Consequenee. I f no saturated columns and rows exist, the problem alwayx has a sohltion.

x2 x3 X/, X 5 X5 X7 X8 x9

x 1 ~ ~_ ~ . XlO

Ti~;X3 :~-X2
~;"$5~5f.5f,f.'_.'.,f,f.5~: //////////'//////~J//////////~
. . . . . xt,
'1X1
~///////f/////Y/////////////~
×2 x3
I l
Xl I.//J//J.//////U//////////////////////////////////////////////Z/S/
xt.

Fig. 6

3 /, 5 2 3 4 5

soturoted " ~ ~ ~ --
rOWS

I 4 3 2 5 1 4 3 2

t
c
4+~¢6
Fig. 7

4. Planar graphs

A related result of Okamura and Seymour [11] concerns the disjoint path
problem for arbitrary planar graphs with a fixed embedding. Again, the terminals
are specified on the boundary. The cut criterion requires that the congestion of any
cut should not exceed the cardinality of the cut.
Theorem 4. (Cf. [I 2.]) I f there are no odd sets, the cut criterion is necessary and suffici-
ent for the existence of edge disjoint paths between the corresponding terminals. |
The Lemma, used in the proof of Theorem 3, can easily be derived from this
theorem. Actually, the proof of the Lemma is an adaption of Seymour and Okamura's
proof but it seemed to be worthwhile to work out the details here since the algorithm
is significantly simpler tbr this special case.
It would be nice to find a common generalization of Theorem 3 and 4.
 DISJOINT PATHS IN A GRID 371

References
[1] I. ABOS, On some problems of computer aided layout design, Proc. Sixth Colloquium on
Microwave Communication, Budapest, 1978. Vol. 1.
[2] I. ABos, L. BODROG, A. FRANK, A. SOMOS, P. SCSAURSZKI, Programs and Algorithms for
Printed Circuit Board Design and Documentation, Annual o f the Research b~stitute for
Telecommunication, 1975. Vol. II., 247--258.
[3] T. GALLAI, Maximum-Minimum S/itze fiber Graphen, Acta Math. Acad. Sci. Hungar. 9
(1958) 395--434.
[4] T. GALLAI, Ober extreme Punkt- und Kantenmengen, Ann. Univ. Sci., Budapest, EOtv6s
Sect. Math. 2 (1959) 133--138.
[5] M. R. GAREYand D. S. JOHNSON, Computers and Intractability, W. H. Freeman and Company
San Francisco, 1979.
[6] D. KONIG, Grfiphok 6s mfitrixok, Mat. Fiz. Lapok 38 (1931) 116--119.
[7] S. LINS, A minimax Theorem on Circuits in Projective Graphs, J. Combinatorial Theory ( B )
30 (1981) 253--262.
[8] L. Lov/,sz, Generalized factors of graphs, in "Combinatorial Theory and its Applications,
Proc. Conf. Balatonftired, 1969, Bolyai" North-Holland, Amsterdam, 1970, 773--781.
[9] W. MADER, f2ber die Maxinaalzahl kantendisjunkter A-Wege, Arch. Math. (Basel) 30
(1978) 325--336.
[10] K. MENGER, Zur allgemeinen Kurventheorie, Fund. Math. 10 (1927) 96--115.
[11] H. OKAMURAand P. SEYMOUR, Multicommodity Flows in Planar Graphs, J. Combinatorial
Theory, Ser. B 31, (1981) 75--81.
[12] P. SEYMOUR, On Odd Cuts and Plane Multicommodity Flows, Proc. London Math. Soc.
(3) 42 (1981) 178--192.
[13] W. T. TUTTE, The factors of graphs, Canadian J. Math. 4 (1952) 314--328.

A. F r a n k
Research Inst. f o r Telecommunication
G6bor /~. u. 6 5 - - 6 7
Budapest, Hungary H--1026
 CO~BINATORICA2(4) (1982) 373--376

ON 3-SKEIN ISOMORPHISMS OF GRAPHS

R. L. H E M M I N G E R * , H. A. J U N G and A. K. K E L M A N S
Received 15 September 1981

It is shown that a 3-skein isomorphism between 3-connected graphs with at least 5 vertices
is induced by an isomorphism. These graphs have no loops but may be infinite and have multiple
edges.

I. Introduction

A graph S consisting o f n openly disjoint paths joining two vertices is called an

n-skein. Note that S is a 2-skein if and only if S is a circuit. A bijection between the
edge sets E(G) and E(G*) o f graphs G and G* is called an n-skein isomorphism if it
induces a bijection between the sets of n-skeins contained in G and G*, respectively.
Whitney proved in [6] that circuit isomorphism of 3-connected finite graphs
are induced by isomorphisms. In [7] he observed that one need only assume the 3-
connectedness of one of the graphs to obtain the same conclusion. Various generali-
sations were given in [1], [2], [3], [4] and [5]. The main result of this paper is the
following.
Theorem. Each 3-skein isomorphism from a 3-connected graph with at least 5 vertices
onto a graph without isolated vertices is induced by an isomorphism.
Actually it will be shown that each 3-skein isomorphism from a 3-connected
graph with at least 5 vertices is a circuit isomorphism. The theorem then follows by
applying Whitney's theorem (in a suitably generalized form). On the other hand it
is easily seen that each circuit isomorphism is a 3-skein isomorphism. This fact is a
special case of Theorem 1 in [3].
In [1] the Theorem was proved for 4-connected graphs, and in [3] for 3-connec-
ted graphs containing at least one 4-skein. These results were special cases of results
about n-skeins.
The theorem in the present form was proved by Kelmans and independently
by Hemminger and Jung.

* The paper was written while this author was visiting Universit/it f/Jr Bildungswissen-
schaften, Klagenfurt, Austria.
AMS subject classification (1980): 05 C 40; 05 C 38.
374 R. L. H E M M I N G E R , H. A. J U N G , A. K. KELMANS

2. Proof of the Theorem

In all that follows let G, G* and a be as in the Theorem; i.e. G is 3-connected

and has at least 5 vertices, G* has no isolated vertices and a is a 3-skein isomorphism
from G onto G*. For a subgraph H of G with E ( H ) ¢ 0 , let H ' denote the unique
subgraph of G* without isolated vertices that has E ( H ' ) = a ( E ( H ) ) . A nontrivial
path P in H is considered a subgraph of H and is called a constituent path of H if
the inner vertices of P, and only these, have valency 2 in H.
Observe that for a path P in G the image P ' need not to be a path; even if P"
is a path, incident edges need not be mapped by a onto incident edges.
We will achieve the proof in stages. In the first stage we consider homeomorphs
(some authors say subdivisions) of/£4, the complete graph on 4 vertices. Note that
each permutation of the edges of K4 is a 3-skein automorphism o f / ( 4 , hence the
reqmrement that G have at least 5 vertices.

Lemma 1. I f H is a homeomorph of 1(4 in G, then H" is a homeomorph of K4 in G*.

Moreover, constituent paths of H are mapped onto constituent paths of H'.

ProoL Let H be a homeomorph of K~, and let P be a constituent path of H. Further

let S denote the unique 3-skein in H such that E ( S ) = E ( H ) - E ( P ) . In H ' there
exists a path Q between distinct vertices of S ' such that E ( S ' ) ~ E ( Q ) = O . Then
H ' = S ' U Q since S ' U Q contains at least four 3-skeins while each proper sub-
graph of H contains at most one 3-skein.Hence E ( Q ) = E ( P ' ) and, by construction,
Q is a constituent path of S ' U Q. Since H and consequently H ' contains six 3-skeins
S'CJQ must be a homeomorph o f / ( 4 . I

For a graph H we let n(H)=IV(H)I.

Lemma 2. l f Pa, 1)2 and P3 are constituent paths o f a 3-skein S in G and n(P;)=>3
for each i, then each P[ is a constituent path of S'.
ProoL Since G is 3-connected there exists, for i = 1, 2 and 3, a path Ji, openly
disjoint from S, joining an inner vertex of Pi to an inner vertex of some Pj, j~-i;
since some pair of J1, J2, Jz is incident with all three of P~, P2 and P~, we can assume
that J1 and Jz join /'1 and P~, respectively, to /'3. Now the graph Hi=SUJ~ is a
homeomorph of K4 and so, by Lemma 1, H i is also. For i = 1 and 2 moreover,
H i contains P~_~ as a constituent path. Hence P~_i is a subgraph of some constituent
path P~_~ of S'. But P? (i= 1, 2) is composed of either one or two constituent paths
of H i, since S" is a 3-skein in H ' ; on the other hand P~ is composed of two consti-
tuent paths of H~, and hence, by Lemma 1, also P[ is the union of two constituent
paths of H~.' It follows that PI=PI' * and P2-P2,
' - * and hence that P~, ' P2' and P3'
are the constituent paths of S'. 1[
Lemma 3. I f C is a circuit in G with 3<-n(C)<n(G), then C" is a circuit.
ProoL Let y~ V ( G ) - V(C). Since G is 3-connected, there are paths Q1, Qz and Qa
in G, openly disjoint from each other and C, which join y to distinct vertices on C,
say Xl, x2 and x3 respectively.
We consider cases and subcases.
 3-SKEIkN ISOMORPHISMS OF GRAH'PS 375

(I) n(C)=>4. Then there exists a path P=Q~UQ2UQa such that P joins
nonadjacent vertices on C. By L e m m a 2, P ' is a constituent path of the 3-skein
C'I,.)p', and hence C ' is a circuit.
(H) n ( C ) - - 3 . We consider two possibilities here.
(a) n ( Q i ) ~ 3 for some i, say n(Q3)->_3. We further divide this subcase into
two.
(a.l) n(COQ3)<n(G). In this case n(CUQ3UQ~)<n(G) for i = 1 or 2,
say for i = 1. Now the path P = Q~ CI Q3 is a constituent path ofthe 3-skein S=C U P
and P is a subgraph of two different circuits C1 and C2 in S. Moreover 4 ~ n ( C ~ ) <
n(G). Hence, by Case I, C~ and C~ are circuits in the 3-skein S ' . Since C~NC~=P',
P" is a constituent path of S'=C'UP" and so C ' is a circuit.
(a.2) n(COQ3)=n(G). Since G is 3-connected, some inner vertexx4 of Q3
is adjacent to x~ or x2, say to x~. Let R:~ be the subpath of Q3 between the vertices x3
and x4. Remembering that Qa and Q2 are edges, let e~, e2, e.~ and e4 denote respect-
ively the edges x~x3, x2x3, x,x~ and x~x~. And let S be the 3-skein with constituent
paths Q~Ue~, Q.2Ue.,,and Q3. Thus, by Lemma 2, the 3-skein S ' has constituent
paths Q[Ue;., Q~Ue~ and Q~. Moreover, by applying L e m m a 1 to both SUe3
and SUe4, we have that S'Ue~Ue'4 is as in the figure where {Q~, e~}={e',f'}
and Q£ is on the left. The only question is whether e ~ = e ' or./'.

Fig. 1

But SI=CUe4URa is a 3-skein in G so we must have e ~ = e ' ; or S~ will

not be a 3-skein. Hence C ' is a circuit.
(b) n(QO=n(Q2)=n(Q3)=2. Since n(G)=>5, there is a vertex z~y,
xl, x~, x~. By the previous cases we can assume that z is also adjacent to xl, x2 and
x~. Then, by L e m m a 2, the 3-skein S consisting of all edges xly and xiz has con-
stituent p a t h s / 1 , Pz a n d / ' 3 which are mapped onto constituent paths of S ' . More-
over, by Lemma 1, each pair of the inner vertices of the P[ are joined by the image
of one of the edges xlx2, xtx3 and xzx3. Hence C" is a circuit. II
Proof of the theorem. Let C be a circuit in G. We wish to show that C ' is a circuit.
Hence, by L e m m a 3, we can assume that either n(C)=n(G) or n ( C ) = 2 .
If n(C)=n(G), then we can find an edge e joining nonadjacent vertices on C
(since G is 3-connected and n(G)=>5). But then, by Lemma 3, the two circuits C~
and C2 in CUe that contain e are mapped onto circuits in the 3-skein C'Oe" and
since they have only e' in common, C ' is a circuit.
I f n ( C ) = 2 , say V ( C ) = {y, z}, then there exists a path P in G such that
n ( P ) ~ 3 and CUP is a 3-skein. But then, by Lemma 3, POe for each edge e of
C is a circuit. It follows that C ' is a circuit.
We conclude that a maps circuits in G onto circuits in G*. To complete the
p r o o f that a is a circuit isomorphism let C* be a circuit in G*. Pick e~E(G) such
376 R. L. HEMMINGER, H. A. JUNG, A. K. KELMANS: 3-SKEIN ISOMORPHISMS OF GRAPHS

that e'CE(C*) and let C1 be a circuit in G containing e. By Lemma 3, C~ is a circuit,

which of course also contains e'. Hence, assuming C~#C* (for otherwise we are
already done), there exists a 3-skein S* in G* such that C* is a circuit in S*. N o w
S* is the image of some 3-skein S in G and the three circuits in S are mapped onto
distinct circuits of S*, one of which must therefore be C*.
We have shown that ~ is a circuit isomorphism. Hence, by Whitney's theorem
on circuit isomorphisms and its extension to infinite graphs with multiple edges
[3, 4], the claim now follows. II
We note that Theorem 2 in [3], for n = 2 , explicitly says that a circuit iso-
morphism from a 3-connected graph onto a graph without isolated vertices is in-
duced by an isomorphism.

References

[1] R. HALIN and H. A. JUNG, N o t e o n i s o m o r p h i s m s o f graphs, Y. London Math. Soc. 42 (1967),

254--256.
[2] R. L. HEMMINOER, I s o m o r p h i s m - i n d u c e d line i s o m o r p h i s m s o n p s e u d o g r a p h s , Czechoslovak
Math. J. (96) (1971), 672--679.
[3] R. L. HEMMINGER and H. A. JUNO, On n-skein Isomorphisms of Graphs, J. Combinatorial
Theory (B), 32 (1982), 103--111.
14] H. A. JuNo, Zu einem Isomorphiesatz von H. Whitney ffir Graphen, Math. Ann. 164 (1966),
270--271.
[5] J. H. SANDERSand D. SANDERS, Circuit preserving edge maps, J. Combinatorial Theory (B)
22 (1977), 91--96.
[6] H. WmTNEY, Congruent graphs and the c o n n e c t i v i t y o f graphs, Amer. J. Math. 54 (1932),
150--168.
[7] H. WHITNEY, 2-isomorphic graphs, Amer. J. Math. 55 (1933), 245--254.

H. A. Jung R. L. Hemminger
Fachbereich Mathematik Department o f Mathematics
Teehnische Universit8t Berlin Vanderbilt University
Berlin, West Germany Nashville, Tennesse 37235 U.S.A.

A. K. Kelmans
Profsoyuznaya Str. 130
K. 3, kv, 33
117321 Moscow
U.S.S.R.
 COMBINATORICA2(4) (1982) 377--383

NEARLY BIPARTITE GRAPHS WITH LARGE

CHROMATIC NUMBER

Vojt~ch Rt3DL
Received I I July 1981

P. Erd6s and A. Hajnal asked the following question. Does there exist a constant e>0
with the following property : If every subgraph H of a graph G can be made bipartite by the omis-
sion of at most e IHI edges where IH I denotes the number of vertices of H then Z (H) <=3.
The aim of this note is to give a negative answer to this question and consider the ana-
logous problem for hypergraphs. The first was done also by L. Lov~isz who used a different
construction.

O. Introduction

For a positive e and a positive integer k we shall say that the graph G has the
(e, k)-vertex property ((e, k)-edge property resp.) if
(i) Z (G) =>k;
(ii) if H is a subgraph of G then H can be made bipartite by the omission of at
most e In] vertices (e IH[ edges resp.). (Here IHI denotes the number of vertices of H.)
Using this terminology, a question of P. Erd6s and A. Hajnal can be formula-
ted now as follows:
(*) Given e and k, do there exist graphs with the (e, k)-edge property?
Note that the (e, k)-vertex property is clearly weaker than the (e, k)-edge pro-
perty. The Kneser graphs K ( n , k) ([4], cf. the next section) have the (e, k)-vertex
property for sufficiently large n. Further examples are the Borsuk graphs defined by
Erd6s and Hajnal [5] as follows. The vertex set is the k-dimensional unit sphere. Two
points are adjacent iff their distance is greater than 2 - 6 (where 6 = 6 ( e ) is a suffi-
ciently small positive real). It is easy to see that any sufficiently dense finite subset of
the sphere induces an appropriate f i n i t e subgraph. Still other examples were con-
structed by P. Erd6s, A. Hajnal and E. Szemer6di [6]. They have shown that the graphs,
the vertices of which are the r-tuples ( r > r ( e ) ) of elements of a sufficiently large
linearly ordered set and two of them, x l < x 2 < . . . . x r and y l < y 2 < . . . < y r are
adjacent if either x i + a = y i or Yi+a=xi for i = 1, 2 . . . . , r - 1, have the (e, k)-vertex
property.

AMS subject classification (1980): 05 C 15; 05 C 65.

378 v. RODL

Our note is divided into two parts. In the first part we give a construction
which assigns to every graph G which has vertex transitive automorphism group and
- ~ U ] - ' k -vertex property a graph G* having the (e, k)-edge property. We also
give another example of graphs having the (e, k)-edge property. We shall use the
graphs previously considered by A. Schrijver [I1] in another context. A different
example of graphs with the (e, k)-edge-property was constructed by L. Lovfisz.
In the second part we show that the problem analogous to (*) for 3-graphs can
be answered positively in a very strong sense. We show that for any sequence {k,}
of positive integers such that ,!ira k , , = ~ there exists a 3-uniform hypergraph
F=(V, E) with
(i) x(F) = 1%
(ii) Every subgraph H of F with lHl<--k, can be made bipartite by omitting at
most n 3-tuples.
Acknowledgement. My thanks are due to A. Hajnal who told me about the result
of L. Lovfisz.

1. Graphs which have the (e, k)-edge-property

Construction 1.1. Let G=(V, E ) b e a graph with 7~(G)=k and let < be an or-
dering of its vertices. Set {r~, v2.... , v,,} where v~<vj if i<j. We construct the
graphs G~, G2 .... , G,,,=G* successively as follows:
(I) GI=G.
(II) Suppose G~=(Vt, Ei), l<=i<=m-1 has been constructed and I~ partitioned
as Vi= U U~ such that U}={vj} for any j>i.
j~_m
Denote by W~ the set of those vertices of Gi which are elements of (J U}
and are adjacent to vi+~. We define Gi+1 as follows:
(~) if ]W~l~k-1 set Gi=Gi+l;
by .oiots that
\-- ]

each (k-1)-subset of W~ is adjacent to one of them. Moreover if {v~+~, vv}~E~

tbr i ' > i + 1 , put {u}+~,vv}~E/+l for all l<./<=t. Set now
U~.+I
j ~ O ji for j <- i, r[i+l1
vi+ ~
{ld~+1, ...~
/,/~+1}
and
Ui+X={vj}
J
for j > i + l .

Denote by Fi+ x the set of all edges incident to vertices u~+1.... , u~+1 and put

E,+~ = E, • [ v - {v,+,}p u F,+~.

Claim 1.2. x(G*)=k.

Proof. Clearly x(G*)<=k as ~9: V*~ V defined by q~(U~')=v s for any)', l<=j<=m
is a homomorphism. For the reverse inequality it suffices to prove that ~((Gi+,)<k
implies g(G;)<k for any 1 <=i<=m- 1. Suppose therefore that there is an r-coloring
 LARGE CHROMATIC GRAPHS 379

c o f G i + l ( r < k ) . Now, no (k-1)-subset of W~gets all the r colors, because one of

the vertices u~+~, u~+~, ..., u~+1 (say u~+1) must be colored by a color different from
the colors of points of W/since it is connected to a ( k - D - s u b s e t of W~ which has all
those colors. Define now a new graph

, /Gi+l+all edges connecting vertices of 14/,- t o b/~+1 if G i 7~ G i + l

G~+I
=\Gi+l if Gi = Gi+l.
Clearly c is also a coloring of G;+a and moreover Gi is a subgraph of G[+~. Thus
z(Gi) <= =z(o,+O=,
t . < , -=~:.
< I
Lemma 1.3. Let G=(V, E) be a graph with vertex transitive automorphism group
which can be made bipartite by deletion of a t most s [G[ vertices. Then Jbr any nonnega-
tire real valued function f: V ~ R + there exists a set V" of vertices such that
(i) the subgraph of G hMuced on a set V - V" is bipartite.
(ii) ~ f(v) <=e ~ f(v).
vCV" v~V

Proof. Let G and f be given. Suppose that G can be made bipartite by deleting the
vertices vl, vz ..... cp; p<=elG[. It follows by vertex symmetry of G that

IAut G1
Z Z f ( 9 ( v 3 ) = p Z~vf(V)
l~=i~--p ~ O E A u t G IGI
and thus there exists q)0EAut G such that

Z f(v)
v~ V
Z f(q)o(V,)) <= p - - -- e Z f(v).
Set

Let Z be a set with 2 n + k elements. Consider the Kneser graph K(n, k) the
vertices of which are the n-element subsets of Z ; two such sets are adjacent iff they
are disjoint. It was conjectured by Kneser and proved by L. LovAsz [8] and B~ir~ny
[ 1] that
z(K(,,, I,)) -- k + Z
A few years ago L. Lovfisz told me that the following lemma holds.
Lemma 1.4. For any e > 0 and positive integer k there exists no such that for all posi-
tire integers n>=no the Kneser graph K = K ( n , k) can be made bipartite by the omis-
sion of e Igl vertices.
Proof. Split the (2n+k)-element set Z into two parts 3(. Y of cardinalities IX[=
and ,~ /="--;~/'' , resp. Consider the set M of all n-tuples inter-
/ - - / L - - . I

n k
secting either X or Y in more than - f + T elements. Clearly, the subgraph of Kinduced
on ~ is bipartite.
380 v. RODL

It can be shown that

]K] <e if n >: no(k, ~) > t---e) t-2 1 e1/3. |

Let us take any ordering of the vertices of K=K(n, k) and construct the graph
K*=K*(n,k) as in 1.1.
Then the following holds
Theorem 1.5. For any ~>0 and any positive integer k the graphs K*=K*(n, k),
n>n0 k, ~ have the following properties:
(i) z(K*)=k+2.
(ii) Any subgraph H of K* can be made bipartite by deleting at most ~]gl edges.
g
Proof. Let e and k be given. Set e ' - k + 1 and n>no(k, ~'). (i) follows immediately
by Claim 1.2 and the result of Lovi~sz [8]. We prove (ii). Let H be a subgraph of
K*=K*(n,k). Set K = ( V , E ) , V={vt,v2 .... ,v,,}. Let K*=(V,,,,E,,) where
Vm= 1.J Up where the notation is taken from Construction 1.1. Consider the func-
j_~rn
tion
1
f: V~R + definedby f(vfl=~- ~ d(u), l~=j<=m.
u E V(H) Cl Oj
Then as K is clearly vertex symmetric, we obtain by Lemma 1.3 that K can be made
bipartite omitting a subset V ' c V such that
g

v
~vZ f(v)<=k-~ l v~E V f ( v ) = ~
"
IE(H)[ NelV(H) 1.
Clearly omission of all edges incident to vertices uE Us. such that vs.EV' makes H
bipartite. Moreover there are at most
Z X d(u) = Z f(v) <- elV(O)l
V iEV" u E v j n v ( H ) vEV"
such edges. 1
The graphs having the (e, k)-edge property are clearly examples of graphs with
large chromatic number not containing short cycles of odd length. I would like to
know whether this can be generalized to avoid all short cycles? For instance, do there
exist graphs having the (e, k)-vertex (or edge) property and girth greater than four?
Another way of constructing graphs for which (*) holds was indicated in [11]
by A. Schrijver who described a class of color critical subgraphs of the Kneser graphs
but did not examine how they are related to Conjecture (*). This will be done here.
We shall call an n-subset X' of X = {1, 2 .... , 2 n + k } stable (cf. [6]) if it con-
tains no pair of neighbours in the cyclic ordering of {1, 2 .... , 2 n + k } . In [11] A.
Schrijver introduced the reduced Kneser graph K" (n, k) = K" the vertices of which are
the stable n-subsets of {1,2 .... ,2n +k}, two of them being adjacent iff they are dis-
joint. It is also proved in [11] that Z(K (n, k ) ) = k + 2 .

Note added in proof." The author has answered this question in the affirmative.
 LARGE CHROMATIC GRAPHS 381

Now we shall show that the reduced Kneser graphs provide examples of
graphs satisfying (*). Let k be fixed and n large. It can be seen quite easily that K"
has c~n~(1 + O(1)) vertices and c2nk(1 +O(1)) edges, where cl and c~ are constants
depending on k only.
If e is an edge and vl, v2 its endpoints then denote by v(e) the set of all
iE {1, 2, ..., 2n+k} such that both v~ N {i, i + 1} and v2N {i, i + 1} are nonempty
(addition is mod2n+k). Clearly Iv(e)]>=2n-k.
Let H be a subgraph of K'(n, k) with h edges. There exists an iE {1, 2 ..... 2n + k}
such that

As {eEE(H); i-(v(e)} is clearly the edge set of a bipartite graph we obtain the fol-
lowing result:
Theorem 1.6. Let H be a subgraph of K'(n, k) the reduced Kneser graph. Set h =
h
IE(H)I and m= [E(K'(n, k))[. Then H can be made bipartite by deleting c ml/£
edges (where c is a constant depending only on k). |
We close this section by proving a stronger version of Conjecture (*).
Theorem 1.7. For every e > 0 and k there exists a uniquely k-colorable graph H
having the (e, k)-edge-property.
Proof. Set H = K * ( n , k - 1 ) × K k (where × denotes the (weak) direct product,
cf. [8, p. 538]). Let n be so large that K*(n, k - l ) has the k ( k - - l) ' k -edge-pro-
perty. The unique colorability of H follows from [7] where it was proved that the
direct product of a (k+l)-chromatic graph with Kk is uniquely (k+l)-colorable.
Let H" be a subgraph of H. Consider the projection n: H'~K*(n, k - l )
and assign to every edge e of K*(n, k - l ) the number of edges e'EE(H') such that
7z(e')=e. Denote this number by c~(e). Clearly O~c~(e)<-k(k-1) for every
eEE(K*(n, k - l ) ) . Consider the subgraph F of K*(n, k - l ) formed by the edges
e with e(e)>0. F can be made bipartite by omitting k ( k e- l ) IF[ edges. Since
c~(e)<=k(k- 1) for every eEE(F) it follows that H" can be made bipartite by omitting
~[FI--e]H'[ edges. |

2. Results concerning hypergraphs

In the previous section we have shown that for every ~>0 and positive in-
teger k there exists a graph having the (e, k)-edge property. Theorem 1.6 actually
states a stronger result. It follows from this Theorem 1.6 that every subgraph H of
K' (n, k) having h edges can be made bipartite by omitting ch1-1/k edges. Other graphs
with this property were constructed by L. Lovfisz. The question, whether chl-Ilk
can be replaced by a slowlier growing function, remains open. Note that using similar
4"
382 v. R~DL

examples this cannot be considerably improved and candidates for graphs which
could improve Theorem 1.6 must be very asymmetrical. This follows from results
proved in [10] where among others, the following has been noticed: Let G=(V, E)
be a k-chromatic (k_>5) graph with vertex (edge) transitive automorphism group.
Set n = 1VI, m = IEI. Then G cannot be made bipartite by omission of less than I/n
in
vertices ( ( - - ~ edges, resp.).
In contrast, the analogous question for hypergraphs has a positive answer in
a very strong sense.

Theorem 2.1. For erery sequence {k,} of positive integers such that .!in~k, = ,,o there
exists a 3-uniform hypergraph F = ( V, E) such that
(i) z(F) = s0;
(ii) every subhypergraph H o f F with IHl<=k, vertices can be made bipartite by
the omission of at most n triples.
Remark. One can prove in a similar way that the above Theorem is valid for r-uni-
form hypergraphs for any r ~ 3 .
In the proof of Theorem 2.1 we shall use the following lemma which we state
here without proof because it can be done by standard methods due to P. Erd6s
(see e.g. [2]).
Lemma 2.2. For any positive integers t, p and k there exists a system of graphs
GJ(t,k,p)=GJ=(V, Ej), 1 <'<=J=t such that
(a) (V, E) does not contain any cycles of length <=k, where E= U Ej.
l~j~t
(b) None o f the G i contains an independent set of more than IVI/p vertices. |

Proof of Theorem 2.1. We construct the sets of vertices and edges of the hyper-
graph F as unions of pairwise disjoint sets V~, V2. . . . and E~, E2, ..., resp.
Put I V~]--k~ and denote by Gt the empty graph (the graph with no edges)
with the vertex set V1. Set E~=O.
Suppose now that V~, V~.... , Vii and E~, E2, ..., E~ have been constructed.
Set h=lv~l+lv~l+...+lv, I and ui=IE~I+IEzI+...+]E~I.
Take a system of graphs G{+~=GJ(ti, k,,, i+ 1) with the properties defined
in Lemma 2.2. Put Vi+~= V(GI+a) . . . . . V(G[%0 and consider a one-to-one mapp-
ing
1 ti
~0: ~.J Vj ~ {Gi+ 1 .... , GI+I}.
l~j~i
Set now

Let
V= ~Vi and E= ~Ei.
i=l i=l

We show that F=(V, E) has the required properties. Suppose first that z ( F ) ~ k
for some finite k. Then there exist t > t = k and nonempty monochromatic sets
 L A R G E CHROMATIC GRAPHS 383

MicVi, M v c V v , ]Mi[=>s-~- which are colored by the same color. Hence, by

property (b) of the graphs G1 and the construction of Ei there is a monochromatic
triple eEE~ in M~ U M v , a contradiction, proving (i).
Let now W be a finite subset of V. Set rI=I[WI3AE~] where [W] 3 denotes the
set of triples from W. If r~+l-<_k,, for every i = 1 , 2 .... then [W]ef-q,E*+~ is a forest
for any i, where E*+x- [.j E(G{+z). This implies that the hypergraph induced on
l~j~t l
W is bipartite and there is nothing to prove. Take therefore the largest i (if exists)
such that ri+l>k~,. Then ]W] >k~, but the subhypergraph of E induced on W can be
made bipartite by omitting the (at most ui) edges of U Ejfq[W] 8. II
j=l

References

[1] I. B/,RANY, A short proof of Kneser's conjecture, Journal of Comb. Theory, A, 25 (1978),
325--326.
[2] B. BOLLOBAS,Extremal graph theory, Academic Press, London 1978.
[3] P. ERD6S, Graph theory and probability, Canadian Jr. of Math. 11 (1959), 34--38.
[4] P. ERD6S, On the combinatorial problems which I would most like to see solved, Combina-
torica 1 (1981), 25---42.
[5] P. ERD6S and A. HmNAL, Kromatikus grfifokr61 (On chromatic graphs, in Hungarian),
Matematikai Lapok 18 (I967), 1--4.
[6] P. ERD6S, A. HAJNAL and E. SZEMER~Dr, On almost bipartite large chromatic graphs, to
appear in the volume dedicated to the 60th birthday of A. Kotzig.
[7] D. GREENWELL and L. LovLsz, Applications of product colouring, Acta Math. Acad. Sci.
Hung. 25 (1974), 335--340.
[8] L. LovA.sz, Kneser's conjecture, chromatic number, and homotopy, J. Combinatorial Theory,
A, 25 (1978), 319--324.
[9] L. Lov~,sz, Combinatorial Problems and Exercises, Akad6miai Kiad6 - - North-Holland,
Budapest--Amsterdam, I979.
[I0] V. R6DL and Zs. TUZA, On color critical graphs (manuscript).
[11] A. SCHRIJVER, Vertex-critical subgraphs of Kneser graphs, reprint, Amsterdam (1978).

Vojt~ch R6dl
FJFI ~VUT,
Dept. of Mathematics,
11000 PRAHA, 1 ~ S S R
 COMBINATORICA2(4) (1982) 385--393

A N ANALYSIS OF THE GREEDY ALGORITHM FOR

THE SUBMODULAR SET COVERING PROBLEM

Laurence A. W O L S E Y
Received 19 October 1981

We consider the problem: rain ~{~fi: z(S)=z(N), S==N) where z is a nondecreasing

j ~ - -

submodular set function on a finite set N. When z is integer-valued and z(0) =0, it is shown that
the value of a greedy heuristic solution never exceeds the optimal value by more than a factor
d 1
H(maxz({j})), where H(d)= Z ' - - .
i=1 /
This generalises earlier results of Dobson and others on the applications of the greedy
algorithm to the integer covering problem: rain {f~v: Ay~b, yE{O, 1}} where a~j, b~_~0 are
integer, and also includes the problem of finding a minimum weight basis in a matroid.

1. Introduction

Several authors have very recently studied the behaviour o f the greedy heuris-
tic for various versions o f the integer covering p r o b l e m

(C) min fjj: Ay>=b, yiC{O, 1} j = l , . . . , n

J
where alj >-- O, bi,fj>O for all i and j.
On the other h a n d the optimality o f the greedy algorithm for finding a mini-
m u m weight basis in a matroid is by now a classic result. Here we consider a genera-
lisation o f both problems, namely the s u b m o d u l a r set covering p r o b l e m :
(Q) z= min{
ScN j
~s f j : z ( S ) = z(N)}

where z: N(N)--,R is a nondecreasing, s u b m o d u l a r set function on N = {1, ..., n}.

A function is submodular if z(A)+z(B)>=z(AtSB)+z(Af-lB).
To see that the integer covering problem (C) is a special case o f (Q), it suffi-
m
ces to take z ( S ) = ,--~irain {j~s alj, b,}, while we obtain the minimum weight spann-
ing set o f a matroid by taking z to be the rank function o f the matroid. A n o t h e r

AMS subject classification (1980): 68 C 05, 68 C 25; 90 C 10, 05 B 35

386 L.A. WOLSEY

case of (Q) of practical interest is the set covering problem with capacity restrictions:

min f;yj: aijxi j >=l i = 1.... m, xij <=djyj j = 1..... n,

j=l i=1

xij >- O, yj6{O, 1} with a~jE{O, 1}.

Note that z' defined by z ' ( S ) = m i n {z0, z(S)} is submodular and nondec-
reasing whenever z is, so the apparently more general constraint z(S)>-zo also fits
the model.
The main result of the paper is to show that if a greedy heuristic is applied
to problem (Q), the value Z ~ of a greedy heuristic solution always satisfies Z a <-
<- (1 + 1Oge?)Z where ? is one of several possible problem parameters. In the special
case that z is integer-valued, the analysis gives Z~/Z<=H(maxz({j})-z(O)). This

leads to an error factor of H n~ax aii for problem (C) with integer data, which
i
is a result of Dobson [3], generalising earlier results of Johnson [5], Lov~sz [6] and
Chvfital [1] for the set covering problem. If z is the rank function of a matroid,
ma.x (z({j}) - z(0)) = 1, H(I) = 1, and greedy is optimal, see Rado [9] and many others.
J
The problem ({2) and its analysis is also closely related to the problem
sm~__x{z(S): z~ f]--<f0} which has been studied extensively in [7, 8, 10].
= j~S
The outline of the paper is as follows. In the following section we give an
integer programming reformulation of (Q), describe the greedy heuristic and prove
the main result. In Section 3 we indicate how a similar analysis can be carried out for
a continuous version of (Q), the problem: Z R = m i n {w(y): w(y)=w(h), y>=0}
where w is a concave submodular nondecreasing function on R~.. In addition we de-
duce a highly negative result for the family of set covering problems with unit costs
and duplicate rows: rain yj: aijyj>=l i = 1 .... ,m, yjE{O, I}./E where
j J=
aijE {0, 1} for all i and./'. Among all "'black box" algorithms looking only at values
m m

of the subroutine z ( S ) = j--~irain {i~s a~j, 1/=, ~=~ZmaXimsa~j (i.e., z ( S ) is the number
of rows covered by the set S of columns), there is no approximation algorithm making
a polynomial number of calls of the subroutine that guarantees less than ~ times the
optimal value for all problem instances for any fixed value of z.

2. Problem reformulation and the greedy heuristic

First we present a reformulation of (Q) as a linear integer program. For this

it is useful to view two alternative properties of submodular functions.
Let ~;(S) = z ( S U { j } ) - z ( S ) .
Proposition 1 [7]. A set function z: ~ ( N ) ~ R is submodular and nondecreasing i f
and only i f either
a) oj(S)>= o j ( T ) ~ O VSC= Tc= N or
b) z(T)<= z ( S ) + ~ 9j(S) VS, TC= N. II
j cr \ s
 THE SUBMODULAR COVERING PROBLEM 387

Consider now the linear integer program:

Zi=min ~fiyi s.t.
(Q,) i~
Z ~j(S)Yj >=z ( N ) - z ( S ) VS c= N yjE{O, 1} jCN.
jEN

The following result shows us that problems (Q) and (Qt) are equivalent, and
that Z=Zz.
Proposition 2. TC=N is feasible in (Q) if and only if its characteristic vector yr is
feasible h~ (Q,).
Proof. Suppose T is feasible in (Q). Then ~ Qj(S)yy= ~ o~(S)>=z(T)-
j( N jE T ~ S
-z(S)=z(N)-z(S)VS=N. The first equality holds as Qj(S)=0 if jES, the
inequality follows from Proposition 1, and z(T)=z(N) as T is feasible in (Q).
Conversely if y r is feasible in (QZ), and we consider the constraint indexed
by T, 0 = ~ ~j(T)y~>=z(N)-z(T), and hence z(T)=z(N). I
jE/v

Now we present the greedy algorithm.

A Greedy Heuristic for (Q)

Set t = l . S°=0. Stop if z(O)=z(N).
Iteration t. Let 0 ' = rain ~ fJ "~
~ N,,~,-~ [ oj(s'- 1) 1.

Let arg min{.oj(~7_l)}=j,.

Let O, = Oj,(St-1).
Set S t = S t - I U {Jr}, and a, = z(St)--z(St-1),
Stop if z(S')=z(N), and set T=t.
Otherwise set t=t+ I.
We say that S r is a greedy heuristic solution with value Z o = c~STfj. Evi-
J
dently Z ° provides an upper bound for Z. Note also that because of submodularity
0<01---02-< ... <--Or.
Theorem 1. I f the greedy algorithm is applied to (Q)

i)
DAs°)
Za/Z <- 1 +log~ max,_-:-7~,,: oj(SO > 0
J,, t~j(S )
}
ii) Z°/Z <- 1 + loge 0r/01,

iii) Z ° / Z ="<I + loge {z (N)--

z ( N ) z (Sz~(~r)l'
- )}"
388 L . A . WOLSEY

I f z is integer-valued,

iv) Z G / Z ~ H(max z({j})- z(O)),

where H ( d ) = i=1
~ 1i ./or d a positive integer.

Before proving the theorem we need one preliminary result.

Proposition 3. Let O<ux <

=u2 -<
= •. <
= u. . .. and. .x. l ~ x 2 ~ > x . > O.
= /f S=
n--1 n--1
,~--1= u i ( x l - x i + O + u " x " = u l x l + , = l Z ( u , + l - u i ) x i + l , then

S <= (maxu, x,)[I +loge min ( x , , u__~_l)].

k Xn

I f (xi}7=x are integer, S<=(maxu, x i ) H ( x O .

I f {u/}7=1 are integer, S-<_(maxu, x~)H(u,).
-n--I - X
Z 1 -i+1 +l]_<_(max
n--1
"= x~+lJl i ~ , where the second inequality
1 1

uses the fact that 1 !~logexYx>=l •

-
If ,tx~"
m=l are integer, 1 - x~+l ~_
X Xi
1 1
--+ +...-t - - l
i f x i > x i + l = > l , s o t h a t S<=kmaxu, x l ) , ~ . ~ , l ,
- ' ' x.l
+ 1 +1 1<=
xl xi+l xi+l+l ~ ti=l ~, Xl ) J
(m x u,x,) 1 +
•
+ 1 + l[<_(maxu,x,)n(x,)"
. . . ,
Taking x~<_-(maxu~x~)/u~, und using an identical argument completes the
proof. |
f Qj (S °) ~ 0 7` z(N) -- z(O)
Let kl - max~--z--7-~: oi(S') > Oj; k2 = - - " kz =
J., [ o j ( S ) 01 ' z(N)--z(S r-l)

Proof of Theorem 1. To analyse the heuristic it is necessary to obtain lower bounds

on Z. For this we consider the following linear programming relaxation of (Qt):
Z L = mJn ~a f j Y j : ~ o j ( S t ) Y j >= z ( N ) - z ( S t) t : 0 ..... T - - 1
(QL) ~u jcu
yj>=O, jEN.
Our aim will be to find appropriate dual feasible solutions for (QL) whose
value will provide a lower bound on Z L and hence on Z.
i) and ii). Let 0*=(01, 0~-01, ..., 0 r - 0 r - O . For a given j, there exists r~_T
such that oj(sr-1)>0 and Qj(S')=0. Apply Proposition 3 with 0<01=<...=<0"
 THE SUBMODULAR COVERING PROBLEM 389

and oj(S°)=>..._->oj(S'-~)>O. We obtain that

o~oAs °) + (o ~- oDo~ (sD +... + (o" - o" -Ooj (s'-1)

<={max
'=~,..., O,oj(Xt_,)}[l+log~nfin { oj(S,_I)
Qj(S °, ~0"}] ~ f j [ l + l o g ~ m i n {k~,k2}],

where O'oj(S'-l)<-fi is a consequence of the greedy heuristic.

Hence (l+log~min {kl, k2})-10 * is dual feasible for (QZ), and therefore
(1 + loge min {kl, k~})-~ [01 (z ( N ) - z (S°)) +
+ (0 2- O0 ( z ( N ) - z ($2)) +... + (0 T - 0 T-l) ( z ( N ) - z(Sr-1))] ~ Z L <=Z.
But
T T
Za = Z O ' ( z ( S t ) - z ( S ' - * ) ) = O X ( z ( N ) - z ( S ° ) ) + Z ( O ' - O ' - ~ ) ( z ( U ) - z ( S t - 1 ) )
t=l t=2

and hence
Z ~ ~ Z(1 +log~ min {k~, k~}).
iii) Define u'~R T by u[=O t if i=t, u~=0 otherwise.

~'(QAs0), ..., 0 A s ~ - o ) = o%(s--) <=f~,

and h.nce zd is dual feasible for t = l , ..., T. It follows that

max u ' ( z ( N ) - z ( S °), .... z ( N ) - z ( S r - 1 ) ) = max O ' ( z ( N ) - - z ( S ' - O ) <=Z L <=Z.

t=l ..,, T t = l , ..., T

Now applying Proposition 3 with 0 < 01 ~... <- Or, and z ( N ) - z ( S °) >=z(N)-
-z(SO>=... > = z ( N ) - z ( S T-l) gives
T--1
Z~ = ~ Ot(z(S ' ) - z(St-O) + OT(z(N) -- z(Sr-1))

z(N)-z(SO) ]
~_ max{O'(z(N)-z(St-O)} 1 + log e z ~ l ) j ~ Z(1 +loge k3).

iv). Ifz is integer-valued, o~(S t) is integer for allj and t, and from Proposition 3, we
obtain
010j (S °) +... + (0" - 6*-1)Qj(S" -1) <_f j H (max Oj (S°)).

The rest of the proof follows that of i) and ii) above. |

Corollary. For the problem o f finding a minimum weight set that is a spanning set in
each o f p matroids, there exists a greedy heuristic for which Z o / Z ~ H ( p ) .
P
Proof, Let r~ be the rank function ofmatroid i. Take z = ~ r~ and apply the greedy
i=1
heuristic to the resulting problem (Q). As z ( S ) = z ( N ) only if r~(S)=r~(N) for all
i, the result follows from Theorem 1. |
390 L.A. WOLSEY

It is perhaps of interest to note that when z ( S ) is the rank function o f a matroid

i.e. p = 1 above, the proof of Theorem 1 shows not only that the greedy algorithm is
optimal but also the polyhedron

{Y: Z [ r ( S U { j } ) - r ( S ) l y j > z ( N ) - z ( S ) V S C = N , 3i=0, ./E N}

jEN

has integer vertices, and hence is the blocker of the bases of the matroid.

3. Further results and extensions

First we consider a continuous version of the earlier model, namely the pro-
blem :

(R) Z R = rnin yj: w(y) = w(h), y.i >=O, j = 1..... n ,

j=

where w: R"+~R is nondecreasing, submodular ( w ( x ) + w ( y ) > = w ( x V y ) + w ( x A y ) ) ,

piecewise linear and concave. Again there is no gain in generality with the constraint
set {y: w(y)~w0, O~y<=h} as both ~ ( y ) = w ( y A h ) and w'O,)=min (w(y), ,%)
are submodular whenever w is submodular.
Using the properties of w, it is easily shown, see [9], that:

w(y) = w(x)+{~: ~

Paralleling the earlier development, we now describe the continuous greedy heuristic
for (R):
Let 0 j ( x ) = lim w ( x + e e i ) - w ( x ) where ej is the unit vector in direction/.
elO + S

A Continuous Greedy Algorithm for (R)

Set t = l . y ° = ( 0 , . . . , 0 ) . Stop if w(0)=w(h).

Iteration t. Let 0 t = m i n {fj/0~.(y'-l)}
jEN
Let arg min {fj./~o~(y'-l)} =.it.
Let ~ot=~o~,(y~-l).
Let a,=e,Q,, where e , = m a x {e: w ( y t - l + e % ) - w ( y ' - l ) = e O , } .
Set yt=yt-l~-atej,, so that w ( f l ) = w ( y t - X ) + a t .
Stop if w(yt)=w(h), and set T = t .
Otherwise set t = t + i.

We call y r a continuous greedy solution with value Z cG = Z" f i yr. A lower bound
j=l
 T I I E S U B M O D U L A R COVERING PROBLEM 391

on ZR is now obtained from the linear program:

n

Z~ = min Z f / J ' j
j=l

(R z) ~_~ O j ( y t - a ) y i >= w ( h ) - w ( y ' - a ) , t = 1. . . . . T

j=l

3,j -> 0, j= 1, ..., n.

Defining all other terms in identical fashion, we obtain:

Theorem 2. I f the continuous greedy algorithm is applied to (R) and terminates

with a .feasible solution yCG,
f[ Qj(yO) ] Or w(h)-w(0) I,
ZC°/zR <--_l + l o g ~ r n i n ~ / m a x ~ : Oj(Y') > Ol
tt J,, oj(y ) ' 01' w(h)-w(Yr-~)l"

Proof. The p r o o f is identical to that of Theorem 1, once we have shown that (R L)

is indeed a relaxation of (R). This will follow as in Proposition 2 if we can show that

w(y) <_-w(x)+ Z ~j(x)yjVx, y6R. +.

j=I

es - w ( x Vt~--> 1, and settin~

ts = (y~- x~)/~

w(y) <=w(x) + ~ x ~(ys-x~) w(x+ees)- w(x) Ve ~ 0 +.

Hence w(y) ~ w(x)+ g ~ (Ys-Xs)O~(x) ~ w(x)+j~_xOj(x)Y j as ~oj(x)-> 0

{~ y ,1
; s . =
and x,y~0. I

The linear programming covering problem: rain {fy: A y ~ b , y=>0} treated

in [4] is one special case of problem (R). The results here and in [4] suggest that initial
problem scaling is of importance for the worstcase results. Dobson [3] has taken
this further, and shown how rescaling in the course of the greedy algorithm (i.e.,
changing the submodular function) can significantly improve certain worst case per-
formance.
To conclude the paper, we now consider a somewhat different question. Let
_q. cg denote the families of all problems of the form (Q), (C) respectively. Given an
algorithm tbr ~, or some subclass of.@ such as the family of integer covering problems
~, that works only by looking at function values z(S), can we say anything about the
performance guarantees we can obtain from looking at a given number of function
values?
392 L.A.. WOLSEY

For a class N of problems, we say that an approximation algorithm H has

performance measure ccn if Zn<=enZ for all problems D~N, where Z n is any solu-
tion value from the algorithm H, and Z is the optimal value.
We say that algorithm H is an O (nq) black box algorithm for N if there exists
a constant M such that for every problem DEN with [Nl=n, algorithm H always
terminates before looking at more than Mn q function values.
It is immediate that the greedy algorithm is an O(n ~) black box algorithm for
.~, and if .~P~.~ is the subclass of problems with integer data and z(N)<=p, we
know from Theorem 1 that the greedy algorithm has a performance measure of
{ l + l o g , p}. However, for the whole class .~ our results do not given an obvious
performance measure.
Let ~* be the subclass of .~ consisting of set covering problems with unit
n

costs and duplicate rows, i.e., min~'yj: .~ aiyy>=l, i = 1 . . . . ,m, yS{O, 1}

U=I d=l
j~N} with aij~ {0, 1} for all i and./'. Perhaps not surprisingly in view of the above re-
mark, we have:

Theorem 3. There is no polynomial (O(nq) for an), q) black box algorithm for c~*
having performance measure less than ~ for any finite value of z.
Proof. Fix v and q. We consider two families of functions v.,r,~ and ur,.
..
Given the set N = { I . . . . . n}, we take a set RC=N with ]Rn=r as a special
set. We define v.r,, and u."'" as follows:
v'.,'(S) = IS] if IS[<zr, and R~S
v~"(S) = zr otherwise,
and
u~"(S) = ]SI if IS} < ~r
u';~(S) = zr otherwise.

We observe immediately that the optimal value o f problem (Q) equals r

when g =" v," ' , and equals zr when z - u , r, .
What is more, we claim that any black box algorithm requires at least
n zr = 0 (r)t' calls of the function z to distinguish between v, u, .
r r
first that any set S that gives information (i.e., v~'~(S)¢u~'~(S)) has ISl<~r,.~
and hence contains less than ( 7 ) r-tuples. However the total number of r-tuples is

( : / ' andhence at least [n/fr)r r sets must be examined so as to be sure that ff~e spe-
cial set R has not been missed (if it exists).
Now we invoke a result from [2] that any nondecreasing set function satis-
fying the following condition:
'If z(SU{j})=z(S) for some S~N and jEN-S, then
Z(TU{j}) = Z(T) for all T D S '
 THE SUBMODULAR COVERING PROBLEM 393

is ' o r d e r equivalent' (z(A)>=z(B) if a n d only if z'(A)>=z'(B)) to a location func-

m
tion z'(S)= Zmaxaij., where aij~{O, 1}.
t=1 jES
B o t h v,r,~ a n d u,,"~ satisfy the a b o v e c o n d i t i o n , so let v,""~ a n d u,,""" be two
o r d e r equivalent location functions. I t follows that any black b o x a l g o r i t h m H requi-
res at least O ( r f ) function calls to distinguish between v~,r ' ' a n d u~,~'', a n d hence to
g u a r a n t e e that ~ n < z .
Letting r = q + 1, we have shown t h a t for a n y q a n d z, t h e r e is n o O(n q)
b l a c k b o x a l g o r i t h m with p e r f o r m a n c e m e a s u r e less than z. II
Acknowledgement. I a m m o s t grateful to G . de G h e l l i n c k for his f o r m u l a t i o n o f
P r o p o s i t i o n 3 p e r m i t t i n g a f u r t h e r s h o r t e n i n g of the p r o o f of T h e o r e m 1.

References

[1] V. CHVATAL, A Greedy Heuristics for the Set-Covering Problem, Math. of Oper. Res.
4 (3), (1979), 233--235.
[2] G. CORNU~JOLS, G. L. NEMHAUSERand L. A. WOLSEY, A Canonical Representation of
Simple Plant Location Problems and Its Applications, S l A M J. Alg. Disc. Math. I,
(1980), 261--272.
[3] G. DOBSON, Worst Case Analysis of Greedy Heuristics for Integer Programming with
Nonnegative Data, Technical Report SOL 80--25, Stanford, October 1980.
[4] M. L. FISHER and L. A. WOLSEY,On the Greedy Heuristic for Covering and Packing Pro-
blems, CORE, DP 8124, Louvain-la-Neuve, May 1981.
[5] D. S. JOrL~SON, Approximation Algorithms for Combinatorial problems, J. Comput.
System Sci. 9, (1974), 256--298.
[6] L. LovAsz. On the Ratio of Optimal Integral and Fractional Covers, Discrete Math.,
13, (1975), 383--390.
[7] G. L. NEMHAUSER,L. A. WOLSEYand M. L. FISHER, An Analysis of Approximations for
Maximizing Submodular Set Functions - - I. Math. Prog., 14 (1978), 265--294.
[8] G. L. NEMHAUSERand L. A. WOLSEY, Maximising Submodular Set Functions : Formulations
and Analysis of Algorithms, CORE, DP 7832, Louvain La-Neuve, August 1978, to appear
in Ann. of. Disc. Afath.
[9] R. RAt)O, Note on Independence Functions, Proc. London Math. Soc., 7, (1957), 300--320.
[10] L. A. WOLSEY, Maximising Real-Valued Submodular Functions: Primal and Dual Heuris-
tics for Location Problems, Math. of Oper. Res. 7, (1982), 410---425,

L. A . W o l s e y
Center for Operations Research and
Econometrics
Universitd Catholique de Louvaht
1348 Louvain-la-Neuve
Belgium
 COMBINATORICA2(4) (1982) 395--397

ON A GEOMETRIC PROPERTY OF PERFECT G R A P H S

E. S. Z A R E M B A and S. P E R Z
Received 19 October 1981

Let G be a graph, VP(G) its vertex packing polytope and let A(G) be obtained by reflect-
ing VP(G) in all Cartersian coordinates. Denoting by A*(G) the set obtained similarly from the
fractional vertex packing polytope, we prove that the segment connecting any two non-anti-
podal vertices of A(G) is contained in the surface of A (G) and that G is perfect if and only if
A*(G) has a similar property.

1. Introduction

A graph G is called v-perfect if and only if for every one of its induced sub-
graphs Go, the chromatic number of Go equals the maximum number of vertices
forming a clique in Go; a graph G is c~-perfect if and only if for every one of its indu-
ced subgraphs Go, the minimum number of cliques covering Go is equal to the size of
a largest independent (stable) set in Go. These concepts were introduced by C. Berge
in the early nineteen-sixties to formulate two conjectures. One of them stating that a
graph is v-perfect if and only if it is a-perfect was proved in 1971 by Lovfisz [4, 5]
and became known as the Perfect G r a p h Theorem, whereas the second, called the
Strong Perfect Graph Conjecture, is still unsettled.
A further result [6, p. 86] (essentially used in this paper) that is strongly related
to perfect (i.e., ~, or, equivalently a-perfect graphs) is due to Chv~tal and Fulkerson
[1, 2, 3]. It can be restated as follows. Let G be a finite undirected graph on n vertices;
let A be the m X n clique-vertex incidence matrix of G, and B be the r X n clique-vertex
incidence matrix of G, the complement of G. Then the following conditions are equi-
valent (below er and em denote the vectors in R r and R m, respectively, with all their
components equal to one)
(i) G is perfect,
(ii) P(A)= {x( R": x>-O, Ax<-e,,} has only integer vertices,
(iii) Q ( B ) = { y ( R " : y>=O, By<=e,} has only integer vertices.
In this paper, a new condition characterizing perfect graphs is stated and
proved. Namely, let G be a graph, VP(G) its vertex packing polytope and let A (G)

AMS subject classification 1980:05 C 99; 52 A 25.

396 L. S. Z A R E M B A , S. P E R Z

be obtained by reflecting VP(G) in all Cartesian coordinates. Let, moreover, A*(G)

be obtained similarly from the fractional vertex packing polytope. It is proved that
the segment connecting any two non-antipodal vertices of A (G) lies in the surface o f
A(G) and that G is perfect if and only if A*(G) has similar property.

2. Basic notation

Let G = ( V , F) be a graph with n vertices. By A we shall denote the m×n

clique matrix of G, and by B the r × n clique matrix of G. We call
(1) Q(B)={yER": y>=O, By<=e~}
the fractional vertex packing polytope of G; above e~ is the vector with all his r com-
ponents equal to one.
By g w e shall denote the family of all maximal cliques in G and by g t h e family
of all maximal stable sets in G. Besides, for every W c V, let x w mean the incidence
vector of W, i.e. x W = l if yEW and xW=O if v~W.
Therefore
(2) VP(G) = c o n v {xWER": WE~}
is the vertex packing potytope of G. Reflecting this set in all coordinate planes we
obtain the set
(3) A (G) = c o n v {xER": Ix[EVP(G)},
where [x I= (txl ..... Ix, I). Let A * (G) be obtained similarly from the fractional vertex
packing polytope of G, i.e.,
(4) A*(G) = {xER": txIEQ(B)}.

3. Main results

Theorem 1. For any undirected graph G, the segment connecting any two non-antipodal
vertices of A (G) lies in the surface of A (G).
Proof. Let x 1 and x 2 be any pair of non-antipodal extreme points of A(G). Assume
that, for a certain number ~, 0<)~<1, the point Y~=~xl+(l-2)x~EintA(G),
the interior of A(G). Then also ~EintA*(G), which means that, for some ~>0,

(5) BIX[ < ( 1 - ~ ) e , .

Let VI= {rE V: x~#O} and Vz= {rE V: x~#O}. If V~= II2, then, for some
index vE V, we have x~=x~E {+ 1, - 1} and consequently l.%]= 1, which obviously
contradicts (5).
In the opposite case, we get nodes vie I/1 and vie V such that, for some clique
EEl, {vl, v~}cE. Then x~.12l=2[x~l+(l-2)[x~]=l, which again contradicts
(5). I
Now we can state and prove our main result.
 GEOMETRIC PROPERTY OF PERFECT GRAPHS 397

Theorem 2. Let G be a graph on n vertices. Then the following conditions are equivalent:
(i) G is perfect,
(ii) A(G)=A*(G),
(iii) the segment connecting any two non-antipodal vertices of A*(G) lies in the sur-
face of A*(G).
Proof. We shall prove the equivalences (i)¢*(ii) and (ii)~,(iii). To prove the
first one, observe that condition (ii) is equivalent to the equality
(6) VP(G) = Q (B),
which, by virtue of the well-known characterization of perfect graphs due to V. Chvfi-
tal and D. Fulkerson [6, p. 86], is equivalent to condition (i).
Assume now that condition (ii) does not hold. Then there exists a vertex
xEA*(G) with Ixil<l for some iC v. Putting y=(yo), v~V,
(7) y~ ~- xo, v = i, and Yl = --Xi, c ~: i,
and z=½(x+y) we infer that the point z does not lie in the surface of A*(G).
In this way we have proved (iii)::,(ii). The reverse implication follows easily from
Theorem 1. |

References

[1] V. CHVATAL,On certain polytopes assotiated with graphs, Centre de Recherche Mathematique,
Univ. de Montreal, Que., CRM~238, 1972.
[2] D. R. FULKERSON,Antiblocking polyhedra, J. Combinatorial Theory (B) 12 (1972) 50--71.
[3] D. R. FULKERSON,Blocking and antiblocking pairs of polyhedra, Mathematical Programming
I (1971) 168--194.
[4] L. Lov~.sz, Normal hypergraphs and the perfect graph conjecture, Discrete Mathematics 2
(1972) 353--267.
[5] L. LovAsz, A characterization of perfect graphs, J. Combinatorial Theory (B) 13 (1972)
95--98.
[6l M. PADBERG, Almost integral polyhedra related to certain combinatorial optimzation
problems, Linear Algebra and its Applications, Vol. 15 (1976) 63--88.

L. S. Zaremba S. Perz
ul. Szekspira 4m 130, Central Office of
01--913 Warszawa, Poland Interurban Telecommunications
Warszawa, Poland

$*