C OM BIN A TORIC A

Bolyai Society Springer-Verlag

Combinatorica 29 (1) (2009) 4964

DOI: 10.1007/s00493-009-2264-1

COLORING NUMBER AND ON-LINE RAMSEY THEORY

FOR GRAPHS AND HYPERGRAPHS
H. A. KIERSTEAD, GORAN KONJEVOD
Received December 31, 2005

Let c, s, t be positive integers. The (c, s, t)-Ramsey game is played by Builder and Painter.
Play begins with an s-uniform hypergraph G0 = (V, E0 ), where E0 = and V is determined
by Builder. On the ith round Builder constructs a new edge ei (distinct from previous
edges) and sets Gi = (V, Ei ), where Ei = Ei1 {ei }. Painter responds by coloring ei with
one of c colors. Builder wins if Painter eventually creates a monochromatic copy of Kst ,
the complete s-uniform hypergraph on t vertices; otherwise Painter wins when she has
colored all possible edges.
We extend the denition of coloring number to hypergraphs so that (G) col(G) for
any hypergraph G and then show that Builder can win (c, s, t)-Ramsey game while building
a hypergraph with coloring number at most col(Kst ). An important step in the proof is
the analysis of an auxiliary survival game played by Presenter and Chooser. The (p, s, t)survival game begins with an s-uniform hypergraph H0 = (V, ) with an arbitrary nite
number of vertices and no edges. Let Hi1 = (Vi1 , Ei1 ) be the hypergraph constructed
in the rst i 1 rounds. On the i-th round Presenter plays by presenting a p-subset
Pi Vi1 and Chooser responds by choosing an s-subset Xi Pi . The vertices in Pi Xi
are discarded and the edge Xi added to Ei1 to form Ei . Presenter wins the survival game
if Hi contains a copy of Kst for some i. We show that for positive integers p, s, t with s p,
Presenter has a winning strategy.

1. Introduction


For a positive integer n and a set S, let [n] denote the set {1, . . . , n} and Sn
denote {X
S : |X| = n}. An s-uniform hypergraph is a structure H = (V, E),
where E Vs . Elements of V are called vertices and elements of E are
Mathematics Subject Classication (2000): 05D10, 05C55, 05C65, 03C13, 03D99
c
02099683/109/$6.00
2009
J
anos Bolyai Mathematical Society and Springer-Verlag

50

H. A. KIERSTEAD, GORAN KONJEVOD

called edges or s-edges. At times we will refer to s-uniform hypergraphs as

s-graphs. Then ordinary graphs are 2-graphs. Let Kst denote the complete s


graph on t vertices dened by Kst = [t], [t]
s . For s-graphs G and H we write
G c H, if every c-coloring of the s-edges of G results in a monochromatic
copy of the target H.
For positive integers s, c, t the Ramsey number Ramsc (t) is the least integer
n such that Ksn c Kst . Estimating Ramsey numbers, even for graphs, is
a notoriously dicult problem. This has led researchers to consider other
versions of the problem. Suppose that f is an increasing graph parameter.
For positive integers s, c, t, dene the f -Ramsey number, f -Ramsc (t), to be
the least integer n for which there exists an s-graph G such that f (G) = n
and G c Kst . Trivially, we have
f (Kst ) f - Ramsc (t) f (Ksn ), where n = Ramsc (t).
Erd
os, Faudree, Rousseau and Schelp [2] studied the size Ramsey number, obtained when f (G) = size(G), the number of edges of G. Erd
os et al.
showed, with a proof attributed to Chvat
al,
that
the
size
Ramsey
number
for

graphs is exactly the trivial upper bound n2 , n = Ram2c (t). In other words,
allowing more vertices will not reduce the number of edges needed to force
a monochromatic clique. The coloring number of a graph G = (V, E) is the
least integer d such that its vertices can be ordered as v1 vn so that


 vi : i < j {i, j} E  < d for all j [n].
Let (G) be the chromatic number of G. Clearly, (G) col(G), since its
vertices can be colored with col(G) colors using First-Fit on the enumeranski [4] considered the chromatic
tion v1 , . . . , vn . Recently, Kurek and Ruci
and coloring Ramsey numbers obtained when f is and col, respectively.
They observed that the trivial upper bound for the chromatic Ramsey number is again tight, and thus the trivial upper bound is also tight for the
coloring Ramsey number.
From a Ramsey theoretic perspective, these results are disappointing.
Kurek and Ruci
nski suggested a more promising line of inquiry might be
to study on-line versions of f -Ramsey numbers. This seems encouraging
because many Ramsey theoretic constructions are inherently on-line.
In the on-line setting we consider a process in which edges (not vertices)
are generated one at a time and then immediately and irrevocably colored by
an on-line algorithm. This is best understood as a game played between two
players, Builder and Painter. For positive integers c, s, t the (c, s, t)-Ramsey
game (with target Kst ) is played as follows. Play begins with an empty sgraph G0 = (V, E0 ) on an arbitrarily large, but nite, vertex set V determined

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

51

by Builder. (So E0 = .) The game is played in rounds. At the beginning of

the ith round Builder will have constructed an s-graph Gi1 = (V, Ei1 ) with
|Ei1 | = i 1 and Painter will have constructed a coloring fi1 : Ei1 [c].
On the ith round Builder constructs a new edge ei (distinct from previous
edges) and sets Gi = (V, Ei ), where Ei = Ei1 {ei }. Painter responds by
coloring ei to obtain a coloring fi : Ei [c] with fi1 fi . Builder wins if
Painter eventually creates a monochromatic copy of Kst ; otherwise Painter


wins when she has colored all |Vs | edges.
For positive integers s, c, t and an increasing graph parameter f , dene
the on-line f -Ramsey number, f -oRamsc (t), to be the least integer n such
that Builder can force a win in the (c, s, t)-Ramsey game while constructing
a graph G with f (G) = n. Trivially,
f (Kst ) f - oRamsc (t) f - Ramsc (t) f (Ksn ), where n = Ramsc (t).
Kurek and Ruci
nski conjectured:
size-oRam2c (t)
(t) = 0.
2
t size-Ramc (t)

Conjecture 1. For all positive integers c and t, lim

In some sense this appears to be a trick question, since the heart of the
problem still seems to be to improve known estimates of Ram2c (t). A standard construction for bounding Ram2c (t), and hence size-Ram2c (t), provides
a much stronger bound for size-oRam2c (t): The game is played on a set of
vertices V0 of cardinality cctc . During the ith stage of the game Builder
picks a vertex xi Vi1 and constructs all possible edges between xi and
Vi1 \{xi }. Painter must color at least 1c |Vi1 | of these edges with the same
color i . Now Builder will only play on the set Vi of vertices linked to xi by
an edge colored i . At the end of this process the set {xi : i [ct c + 1]}
will contain a monochromatic copy of K2t . This shows that Ram2c (t) cct ,

ct 
size-Ram2c (t) c2 and
size-oRam2c (t)

ctc


ci cct .

i=0

Thus if cct is a good bound on Ram2c (t) then we have a positive answer to
the question of Kurek and Ruci
nski. Of course, there may exist a dierent
strategy for Builder that requires even fewer edges.
Another approach is to change the question. Let us consider the on-line
coloring Ramsey number, col-oRam2c (t). The construction above shows that
col-oRam2c (t) ct: Order the vertices so that for all i the vertices in Vi \Vi+1
precede the vertices in Vi+1 . Since each vertex in Vi+1 is adjacent to at most

52

H. A. KIERSTEAD, GORAN KONJEVOD

one vertex in each Vj \Vj+1 for j < i, the coloring number is at most ct. Thus
by using a well known lower bound on the Ramsey number, we have
col-oRam2c (t)
ct
t = 0.
2
t col-Ram (t)
2
c
lim

The main result of this paper is that much more is true. First we prove the
following theorem that shows that the trivial lower bound on col-oRam2c (t)
is tight even though the trivial upper bound is tight for -Ram2c (t).
Theorem 2. For all positive integers c, t, col-oRam2c (t) = (K2t ) = col(K2t ) =
t.
Next we extend the denition of coloring number to hypergraphs in a
natural way so that (G) col(G) for all hypergraphs G. Finally we prove
our main result:
Theorem 3. For all positive integers c, s, t, col-oRamsc (t) = (Kst ) = col(Kst ).
Our techniques were rst used in [3], where it is shown that -oRam22 (t) =
t for every positive integer t and col-oRam2c (3) = 3 for every positive integer c.
As in [3] our main tool is the analysis of an auxiliary game called survival,
which seems to be interesting in its own right. The novelty of the current
paper is that our previous analysis of survival for graphs is extended to
hypergraphs. This is needed, even in the case of graphs (Theorem 2) to
extend the results of [3] to arbitrary c and t. This paper is organized as
follows. In Section 2 we introduce the survival game and prove our key
technical result about it. In Section 3 we use our result on survival to prove
Theorem 2. In Section 4 we introduce our generalization of coloring number
to hypergraphs and show how to modify the proof of Theorem 2 to prove
Theorem 3.
1.1. Notation and terminology
Let tn denote the sequence t1 , . . . , tn . So t0 is the empty sequence . It will
also occasionally be convenient to let t0 = . Also we may abuse notation
and write tn for {t1 , . . . , tn }. Let G = (V, E) be a graph and X, Y V . Then
E (X, Y ) denotes the set of edges with at least one vertex in X and at least
one vertex in Y . Let K(X, Y ) denote the bipartite complete graph with
bipartition {X, Y }. Notice that thisnotation
makes sense evenwhen G is a


: eX = = eY . If X = {x}
hypergraph. In this case K(X, Y ) = e XY
s
we may write E(x, Y ) and K(x, Y ) instead of E({x}, Y ) and K({x}, Y ). If

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

53

V has been ordered as v1 vn and vi vj is an edge with 1 i < j n then

we say that vi is a back-neighbor of vj and vj is a forward-neighbor of vi . The
back-degree of a vertex is the number of back-neighbors that it has. Thus a
graph has coloring number t if there exists an ordering of its vertex set such
that every vertex has back-degree at most t1. A forward-edge of a vertex v
is an edge to a forward-neighbor of v.
2. The Survival Game
Let p, s, t be positive integers with s p. The (p, s, t)-survival game is
played by two players, Presenter and Chooser. Play begins with the s-graph
H0 = (S0 , E0 ), where S0 is an arbitrarily large, but nite, set of vertices
determined at the beginning of the game by Presenter and E0 = . The
game is played in rounds. At the beginning of the ith round the players
will have constructed an s-graph Hi1 = (Si1 , Ei1 ). During the ith round
they construct Hi = (Si , Ei ) as follows. Presenter plays by presenting a psubset Pi Si1 . Chooser responds by choosing an s-set Xi Pi . The remaining vertices in Pi Xi are discarded, leaving Si = Si1 (Pi Xi ) and
Ei = (Ei1 {Xi }){Xj Ei1 : Xj
Si }. The vertices in Si are called surviving vertices. Presenter wins if Hi contains a copy of Kst for some i. Otherwise
Chooser wins when eventually |Si | < t as then Presenter cannot make a play.
Theorem 4. For all positive integers p, s, t with s p, Presenter has a
winning strategy in the (p, s, t)-survival game.
Our approach to proving the theorem will be to show recursively that
Presenter can meet a certain series of goals. We view the starting position,
a graph with a huge number of vertices, but no edges, as having great potential, but being very far from the nal goal of a complete subgraph. As
Presenter progress through these goals he will be sacricing potential for
graphs that in a certain usable sense are closer and closer to being complete. In order to express these goals we will rst develop a new logical
system consisting of models (partitioned s-graphs), basic formulas and rules
for satisfaction.
We assume that the initial set S0 of vertices on which the game is played
is ordered by the relation <. By a partitioned s-graph we mean a structure
H = (U, W, E), where (U W, E) is an s-uniform hypergraph and U and W
are disjoint subsets of S0 . The set U W inherits the order of S0 . We call U
the universal set and W the witness set. The partition of vertices into two
sets does not imply any kind of bipartite property. In particular, edges may
be contained in either subset of vertices. The order of a partitioned s-graph,

54

H. A. KIERSTEAD, GORAN KONJEVOD

which measures its potential, is the cardinality of its universal set. Given a
partitioned s-graph (U, W, E), we dene a relation  on (U W ) by v  v 
i v < v  and if v  W then {u U : v < u < v  } = . Also  v  means that
if v  W then {u U : u < v  } = . We view v  v  as indicating that v  is
bigger than v, but not by too much. Given a partitioned s-graph (U, W, E)
and a sequence of vertices vn U W , we say that vn is strongly increasing
i  v1 and vi  vi+1 for all i [n 1].
The concept of a basic formula is dened recursively as follows. The
language for basic formulas consists of an s-ary predicate P , intended to
indicate strongly increasing s-edges, and exactly s variables s = 1 , . . . , s .
The variable i is only allowed to appear in the ith position of P . There are
no logical connectives, but universal and existential quantiers may appear.
1. P (s ) is the only basic formula with free variables s .
2. = h+1 is a basic formula with free variables h i is a basic formula
with free variables h+1 .
3. = h+1 is a basic formula with free variables h i is a basic formula
with free variables h+1 .
Finally we recursively dene satisfaction for basic formulas. In doing so
we give a special (and denitely nonstandard) interpretation of the meaning of universal and existential quantiers. Roughly, it says that universal
quantication only applies to suciently large vertices in the universal set U ,
while existential quantication only applies to vertices in a specic range of
the witness set W . For notational convenience, let v0 = = v0 . Also u > is
dened to be true.
Denition 5. Let H be a partitioned s-graph as above, be a basic formula
with free variables h , 0 h s, and vh U W be a sequence of vertices.
Then H satises (
vh ) i one of the following holds:
1. = P (s ), vh E, and vh is increasing (so h = s).
vh , u) for all u U with u > vh .
2. = h+1 and H satises (
vh , w) for some w W with w  vh .
3. = h+1 and H satises (
In particular, H satises a basic formula with no free variables, if
H satises (
v0 ), i.e., (). Roughly, the next lemma states that a basic
formula satised by a partitioned s-graph remains true if we restrict the
universal set, while expanding the witness set. This is needed for the main
step in the proof of Lemma 7. The intricate denitions above and hypotheses
below are all needed to prove a statement that can be applied.

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

55

Lemma 6. Suppose H = (U, W, E) and H  = (U  , W  , E  ) are partitioned sgraphs, is a basic formula with free variables h and vh (UW )(U W  )
is a sequence of vertices. Suppose the following conditions are all satised:
1. If ys E then ys E  for all ys (U W ) (U  W  ).
2. U  {u U  : u vh } U .
3. W {w W : w vh } W  .
If H satises (
vh ) then H  satises (
vh ).
Proof. We argue by induction on the denition of a basic formula. For
vs ). Then by
the base step, take = P (s ) and suppose that H satises (
vs ).
Denition 5.1 and Hypothesis 1, H  satises (
Let us now consider the induction step. First suppose that H satises (
vh ), where = h+1 and consider u U  with u > vh . Then, by
Hypothesis 2, u U . So H satises (
vh , u) by Denition 5.2. By the inducvh , u). Since u was arbitrary, we conclude by
tion hypothesis H  satises (
vh ).
Denition 5.2 that H  satises (
Now suppose that H satises (
vh ), where = h+1 . By Denition 5.3,
vh , w). By Hypoththere exists a w W with wvh such that H satises (
vh , w). By
esis 3, w W  . So, by the induction hypothesis, H  satises (


vh ).
Hypothesis 2, H satises wvh . Thus by Denition 5.3, H satises (
The rank r() of a basic formula is dened recursively as follows.
r(P (s )) = 0;
r(h+1 ) = 1 + 2r();
r(h+1 ) = 2r().
A sentence is a formula with no free variables. A basic sentence has the
form
= Q1 1 . . . Qs s P (s )
where each Qi {, }. Note that there are exactly 2s basic sentences and
that each of them has a unique rank between 0 and 2s 1. Let r denote the
basic sentence with rank r. The type of is the maximum  such that Qi =
for all i []. So has type 0 if Q1 = . Suppose =

+1 has type .
Then we set + =

+1 . Notice that +
r = r+1 . Presenters ith goal will
be to force an s-graph with suciently large potential that satises i .
Let be a basic sentence and consider the set A of partitioned s-graphs
G = (U, W, E) such that G has order at least n and G satises . Recall
that Hi = (Si , Ei ) is the s-graph constructed after i rounds of the survival
game. We say that Presenter can construct a partitioned s-graph in A, if he

56

H. A. KIERSTEAD, GORAN KONJEVOD

can force the play so that eventually Hi contains some subgraph (V, E) such
that V can be partitioned as {U, W } so that G = (U, W, E) A. Let f : N N
be a function. The sentence is f -satisable if, starting from f (n) vertices,
Presenter can construct a partitioned s-graph of order n that satises .
Our plan is to show by induction on rank that for every basic sentence r ,
there exists a function fr such that r is fr -satisable. This will prove the
theorem, since, if H = (U, W, E) satises 2s 1 = 1 . . . s P (s ) and has
order t then U induces Kst . The next lemma provides the induction step.
Lemma 7. Let =

+1 be a basic sentence of type  < s. Suppose
that for some function f , the sentence is f -satisable. Then the sentence
+ =

+1 is F -satisable, where F is dened recursively by
F (0) = s
F (j + 1) = f (F (j)), if j 0.
Proof. Fix a positive integer n. We shall construct a partitioned s-graph
H  = (U  , W  , E  ) of order n that satises + from a set U0 of F (n) vertices
ordered by <.
We claim that, starting from U0 , Presenter can construct a sequence
(Hi : i [n]) with Hi = (Ui , Wi , Ei ) such that for all i [n]:
1. |Ui | = F (n i),
2. Ui Wi Ui1 , and
3. Hi satises .
By the denition of F we have |U0 | = F (n) = f (F (n 1)). Since is f satisable, Presenter can construct H1 in U0 such that H1 has order F (n1)
and satises . Arguing by induction on i, if |Ui1 | = F (ni+1) = f (F (ni)),
then Presenter can construct Hi in Ui such that Hi has order F (n i) and
satises .

= = u
).
Let u

Un be an increasing sequence of length  (if  = 0 then u
Such a sequence exists since |Un | F (0) = s  + 1. Since Ui+1 Ui for all
i [n 1], u

Ui for all i [n].
By construction, Hi satises for each i. Hence by Denition 5.2, Hi
satises
+1 (
u
,
+1 ). Thus by Denition 5.3, for each i there exists a
u
, wi ) and u
 wi .
wi Wi such that Hi satises (




Dene H = (U , W , E ) by
U  = {w1 , . . . , wn },

(W1 {w1 }) (Wn {wn }), and
W = u
E  = E[U  W  ].

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

57

By the construction of Hi , Wi Ui1 ; so Wi Wi1 = and thus wi1 = wi .

Moreover, wi1 < wi , since wi Ui1 and u
wi1 in Hi1 implies that there
are no elements of Ui1 between u
 wi1 . It follows that H  has order n.

is a strongly increasing
We claim that H  satises + . First note that u

W  . Thus by Denition 5 it suces to show
sequence in H  and that u
u
, u) for all u U  .
that H  satises (
Consider u U  . Then u = wi for some i [n]. Recall that Hi satises (
u
, wi ). Furthermore, the partitioned s-graphs Hi and H  together with

, u) satisfy the conditions of Lemma 6: if wi < wj
the sequence u

, wi (= u
then as above wj Ui ; if w Wi and wi < w then w Wi {wi } W  .
u
, u) as well.
Hence by the lemma, H  satises (
We are nally ready to prove Theorem 4.
Proof of Theorem 4. Dene the function F p :

n + p, r = 0
p
F (n, r) =
p
p
F (F (n 1, r), r 1), r > 1.
We show by induction on r = r() that for all positive integers n and basic
sentences , Presenter can, starting from a set S0 of F p (n, r()) vertices,
construct partitioned s-graphs of order n that satisfy .
For the base step r = 0 we have 0 = sP (s ). Presenter should present
the rst p vertices vp of S0 on his rst move. Chooser must return an s-edge
s are discarded. Setting U = S0
vp and W = w
s ,
w
s and the vertices in vp w

we have a partitioned s-graph H = (U, W, {ws }), which satises s P (s ) and

has order n.
Now consider the induction step r > 0. By the induction hypothesis,
for every k Presenter can construct a partitioned s-graph from F p (k, r 1)
vertices that has order k and satises r1 =

+1 . Take k = F p (n
1, r). Then by Lemma 7 and the denition of F p , Presenter can construct a
partitioned s-graph of order n that satises r =

+1 .
So, starting from F p (t, 2s 1) vertices, Presenter can construct a partitioned s-graph H = (U, W, E) of order t that satises 2s 1 = 1 . . . s P (s ).
As remarked above, H[U ] = Kst .
For future reference, let h(p, s, t) be the number of vertices and l(p, s, t)
be the number of rounds required for Presenter to win the (p, s, t)-survival
game. From Theorem 4, it follows that h(p, s, t) F p (t, 2s 1).
The function F p grows very quickly. It is in fact very closely related to the
Ackermann function [1], which was the rst example of a number-theoretic
function that is not primitive recursive. Many versions of the Ackermann

58

H. A. KIERSTEAD, GORAN KONJEVOD

function have been dened. For comparison, we now give the complete definition of one version from Peter [5]:
(0, n) = n + 1
(m + 1, 0) = (m, 1)
(m + 1, n + 1) = (m, (m + 1, n)).
The recursive part of the denition of F p is the same as that of , except
that the order of arguments is reversed. Thus F p grows faster than any
primitive recursive function. However, for any xed r, F p (n, r) is primitive
recursive as a function of n. For example, if p and s are xed, Theorem 4
gives a primitive recursive bound on the number of vertices needed for the
Presenter to win the (p, s, t)-survival game.
3. Builders Winning Strategy
In this section we prove Theorem 2. We must show that there exists a function R(c, t) such that Builder can win the (c, 2, t)-Ramsey game while constructing a graph with R(c, t) vertices, whose coloring number is t. In the
standard proof of Ramseys Theorem for graphs, outlined in the introduction, one begins by nding a vertex v1 which is linked to a huge set V1 of
other vertices by edges of the same color. This is then iterated ct times, always working inside the set Vi constructed at the previous step, to construct
a sequence of vertices such that all vertices have monochromatic forwardedges in Vtc . Then the vertices are partitioned according to the common color
of their forward-edges. One part of this partition must contain a monochromatic K2t . However, this produces a graph with coloring number ct.
Instead of constructing single vertices with large monochromatic sets of
incident edges, Builder will construct a large set S such that for every (t1)subset X S there exists some vertex w such that E(w, X) = K(w, X) and
E(w, X) is monochromatic. We call such a vertex w a witness for X; if
all the edges of K(w, X) are colored with , then we say that w is a witness for X. Builder will construct a huge independent set S such that
every (t 1)-subset X S has a witness w V S. Moreover, Builder will
be able to order the vertices so that the vertices of S precede V S and the
back-degree of any vertex is at most t 1. Next, by Ramseys Theorem for
(t1)-graphs, S has a large subset V1 such that there exists a color so that
every (t 1)-subset X S has a -witness w V S. This is then iterated
ct times, always working inside the set Vi constructed at the previous step.
It will then follow that the constructed graph has coloring number t and
contains a monochromatic K2t .

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

59

Here are the details. Fix c and t.

Denition 8. For a function r : N N, we say that Builder has an rstrategy, if starting from r(n) vertices he can play so that eventually G := Gi
has the form G = (V, E), where V can be partitioned as {S, W, T } and ordered
by so that
1. |S| = Ramt1
c (n);
2. for every (t 1)-subset X S there exists a witness w W ;
3. all of S precedes all of W T in and every vertex has back-degree at
most t 1;
4. S is independent, and so the back-degree of every vertex in S is 0.
Next we use Presenters strategy in the survival game to provide Builder
with an r-strategy. Recall that h(p, s, n) and l(p, s, n) are the numbers of
vertices and rounds, respectively, required for Presenter to win the (p, s, n)survival game.
Lemma 9. Let p = c(t1)+2 and s = t1. Builder has an r-strategy, where
r : N N is dened by






t1
r(n) = h p, s, Ramt1
c (n) + l p, s, Ramc (n) .
Proof. Fix n and set N := Ramt1
c (n), h := h(p, s, N ) and l = l(p, s, N ).
Builder begins by selecting a set V of r(n) vertices and partitioning it into
two sets S0 and W = {wi : i [l]} so that |S0 | = h and |W | = l. He also sets
T0 := . The set W is ordered so that
wl wl1 w1 .
Builder will simulate a play of the (p, s, N )-survival game using Presenters winning strategy as follows. The vertices of S0 correspond to the vertices
that Presenter uses in a winning play of the survival game. The vertices of
W correspond to the plays that Presenter makes in this game. Suppose that
on the rst move of his winning strategy, Presenter would play a p-subset
P1 S0 . Then on Builders rst round of play, he simulates Presenters move
by constructing the edges in K(w1 , P1 ). After Painter colors these edges,
there exists an s-subset X1 P1 such that K(w1 , X1 ) is monochromatic, i.e.,
w1 is a witness for X1 . Builder interprets this as a response of X1 by Chooser
in the survival game. He removes the set D1 := P1 X1 of discarded vertices
from S0 to form S1 := S0 D1 and adds D1 to T0 to form T1 := T0 D1 .
The vertices of D1 are placed in the ordering so that w1 D1 , where this
means that w1 comes before every vertex of D1 in .

60

H. A. KIERSTEAD, GORAN KONJEVOD

The game continues in this manner. On his ith round of play, Builder
considers the move Pi called for by Presenters winning strategy and plays
the edges in K(wi , Pi ). After Painter colors them, there exists an s-subset
Xi such that wi is a witness for Xi . Builder interprets this as a response Xi
by Chooser in the simulated game. He sets Di := Pi Xi , Si := Si1 Di and
Ti := Ti1 Di . The vertices of Di are placed between wi and wi1 in so
that wi Di wi1 .
After l rounds the simulated game ends with a win for Presenter. Set S :=
Sl and T := Tl , and place S before wl in the order . By the denition of a win
in the (p, s, N )-survival game and our method of simulation, Conditions 1
and 2 of Denition 8 hold. Clearly S is independent. Finally, V has been
ordered by so that
S wl Dl wl1 Dl1 w1 D1 .
Thus every vertex in S precedes every vertex in W , every vertex in W has
back-degree s and every vertex in T has back-degree 1.
Now we iterate this construction.
Lemma 10. Dene R : N N recursively by
R (0) = 1
R (j + 1) = r(R (j)).


If Builder starts with a set V of R (n) vertices then he can construct a

graph G , a sequence (i : i [n 1]) of colors and a sequence (Vi : i [n]) of
subsets of V such that for every i [n 1] and every (t 1)-subset X Vi :
1. Vi Vi+1 ;
2. there exists an i -witness w Vi+1 for X; and
3. col(G ) = t.
Proof. We argue by induction on n. The base step is trivial so consider the
induction step n = j + 1. By denition R (j + 1) = r(R (j)). By Lemma 9
Builder has an r-strategy. Thus he can construct a graph G, a set S V and
an ordering on V satisfying Conditions 14 of Denition 8. Let Vj+1 := V .

V 

By Condition 1, |S| = Ramt1
c (R (j)). Dene a function : t1 [c] by
setting (X) to be the least color [c] such that X has an -witness.
The color exists by Condition 2. By the denition of Ramsey
number,

V
; let
there exists Vj V such that |Vj | = R (j) and is constant on t1
j be this constant value. We may choose so that Vj precedes S Vj .
By Conditions 3 and 4 all vertices in Vj have back-degree 0 and all other

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

61

vertices have back-degree at most t1. From now on Builder will only play
on the vertices of Vj . By the induction hypothesis, he can do this so that
for the resulting graph there exists a sequence (i : i [j 1]) of colors and
a sequence (Vi : i [j 1]) of subsets of V such that for every i [j 1] and
every (t 1)-subset X Vi Conditions 13 of Lemma 10 hold.
We are now ready to prove our main theorem.
Proof of Theorem 2. Let R(c, t) := R (c(t1)+2), where R is the function
dened in the statement of Lemma 10, and set n = c(t1)+2. We must show
that starting from R(c, t) vertices, Builder has a winning strategy in the
(c, 2, t)-Ramsey game. By Lemma 10, Builder can construct a graph G with
coloring number t for which there exists a sequence (i : i [n 1]) of colors
and a sequence (Vi : i [n]) of subsets of V such that for every i [n1], every
(t1)-subset X Vi has an i -witness wi Vi+1 . By the pigeonhole principle
there exists a color and a subsequence (ih : h [t]) such that = ih . Let
v1 Vi1 . Now suppose recursively that we have constructed (vh : h [j]) such
that each v Vih and X = {vh : h [j]} is complete. Choose vj+1 Vij+1 so
that vj+1 is a -witness for X. Then {vh : h [t]} is a K2t , all of whose edges
have been colored . This completes the proof.
According to Theorem 2, Builder has a winning strategy in the (c, 2, t)Ramsey game when starting from R(c, t) vertices.
Recall now the denition of r(n) as r(n) = h(p, s, Ramt1
c (n)) +
(n)),
where
h
and
l
stand
for
the
number
of
vertices
and numl(p, s, Ramt1
c
ber of rounds, respectively, used in the winning strategy for the (p, s, t)survival game. Regardless of the strategy used, in every round Presenter
uses p vertices, some of which may survive to be used again. Therefore,
p l(p, s, t) h(p, s, t). On the other hand, in each round p s vertices are
thrown away, and so h(p, s, t) (p s)l(p, s, t). Therefore, we have






t1
(n)

r(n)

h
p,
s,
Ram
(n)
1+
h p, s, Ramt1
c
c

1
.
ps

Writing the denition of R(c, t) in terms of R , r and h, we nd that

R(c, t) is computed by c(t 1) + 2 recursive applications of h (recall that
R (0) = 1).
R(c, t) = R (c(t 1) + 2)
R (j + 1) = r(R (j))




1

h c(t 1) + 2, t 1, Ramt1
1+
c (R (j) .
(c 1)(t 1) + 2

62

H. A. KIERSTEAD, GORAN KONJEVOD

Thus our bound R(c, t) on the number of vertices needed to win the (c, 2, t)Ramsey game is not primitive recursive even when one of the parameters
(c or t) is xed.
4. Hypergraphs
In this section we give a natural generalization of Theorem 2 to hypergraphs.
We rst review hypergraph coloring and extend the notion of coloring number to hypergraphs in an appropriate way. A proper c-coloring of a hypergraph H is a function f : V [c] such that every edge contains vertices of
at least two dierent colors. The chromatic number, denoted (H),

t of H is
t
the least c such that H has a proper c-coloring. Then (Ks ) = s1 .
Let v1 vn be an ordering of the vertices of an s-graph H = (V, E).
We call the largest vertex of an s-edge X the root of X. An edge X is a
back-edge of a vertex v if v is the root of X. Two back-edges of v are almost
disjoint if v is their only common vertex. We dene the coloring number
col(H) of H to be the least d such that for some ordering
of V every vertex
t
. Notice that
has fewer than d almost disjoint back-edges. So col(Kst ) = s1
when s = 2 this is the same denition as the coloring number for graphs.
Also, if the vertices of H are colored by First-Fit in an order that achieves
the coloring number d of H then at most d colors will be used. To see this
note that if First-Fit is forced to use a color k on a vertex v then for each
color [k 1], there is a back-edge X of v such that all s 1 vertices in
X{v} have been colored with . It follows that v has at least (k1) almost
disjoint back-edges. Since we started with an optimal ordering, the coloring
number of H is at least k.
Now we are ready to sketch the proof of Theorem 3. We must show that
there exists a function R(c, s, t) such that Builder can win the (c, s, t)-Ramsey
game while constructing a graph with R(c, s, t) vertices, whose coloring number is t. Fix c, s and t. The argument is essentially the same as in the proof
of Theorem 2. We rst must prove Lemmas 9 and 10 for s-graphs. Note that
the denition of a witness still makes sense in the case of hypergraphs. The
only signicant dierence from the proof of Lemma 9 is in the details of how
Builder uses a simulation of the survival game. This time he simulates the
s1
(p, t1, Ramt1
c (n))-survival game, where p = Ramc (t1). When Presenter
would play a p-set Pi on round i, Builder constructs the s-edges of K(wi , Pi ).
By the choice of p, after Painter colors the edges of K(wi , Pi ) there exists
a (t 1)-subset Xi Pi such that K(wi , Xi ) is monochromatic. Builder interprets this as Choosers response. One other change involves checking the
maximum number of almost disjoint back-edges of a vertex. For vertices of

COLORING NUMBER AND ON-LINE RAMSEY THEORY FOR GRAPHS

63

 t1
t 
W this is at most s1
< s1 . For vertices v Di this is at most one since
every back-edge of v contains wi . The proofs of Lemma 10 and Theorem 3
proceed as before.
5. Conclusion and Open Questions
We have shown a striking contrast between the on-line and o-line Ramsey
coloring number for graphs. We have extended our evaluation of the online
coloring Ramsey number for graphs to hypergraphs in a natural way. However the magnitude of the o-line coloring Ramsey number of hypergraphs
remains unknown. It would be nice to clear this up.
The KurekRuci
nski conjecture is still open and of great interest. One
approach might be to prove the conjecture for specic classes of graphs. Our
results may give hope that the on-line size Ramsey number is signicantly
less than the size Ramsey number, but our techniques are not directly applicable because they require such an incredibly large vertex set. This leads
to the question of whether Presenter really needs so many vertices to win
the survival game.
If one examines the survival game more closely, it is relatively easy to
argue that a winning game (on s-uniform hypergraphs) for Presenter must
pass through stages i = 0, 1, . . . , 2s 1 where the graph built by stage i
satises the positive sentence of rank i. However, our strategy for the game
does ignore many edges that might be useful later in the game and it is by
no means clear that the strategy is optimal.
Further, the application of the survival game to the Ramsey game is
where the numbers really blow up, and it would be interesting to know if
there is a better strategy for Builder. For example, is it possible to give a
primitive recursive upper bound on the number of vertices needed to win
the (c, 2, t)-Ramsey game, even for a xed small c?
References
[1] W. Ackermann: Zum Hilbertschen Aufbau der reellen Zahlen, Math. Annalen 99
(1928), 118131.
s, R. J. Faudree, C. Rousseau and R. H. Schelp: The size Ramsey
[2] P. Erdo
number, Period. Math. Hungar. 9 (1978), 145161.
[3] J. A. Grytczuk, M. Ha
luszczak and H. A. Kierstead: On-line Ramsey theory,
Electronic J. of Combinatorics 11 (2004), #R57.
ski: Two variants of the size Ramsey number, Discuss. Math.
[4] A. Kurek and A. Rucin
Graph Theory 25(12) (2005), 141149.
[5] R. P
eter: Recursive Functions, Academic Press, New York, 1967.

64

KIERSTEAD, KONJEVOD: COLORING NUMBER AND RAMSEY THEORY. . .

H. A. Kierstead

Goran Konjevod

Department of Mathematics and Statistics

Arizona State University
Tempe, AZ 85287
USA
kierstead@asu.edu

Department of Computer Science

Arizona State University
Tempe, AZ 85287
USA
goran@asu.edu