Chip-firing games on graphs

A. Björner
Department of Mathematics, Royal Institute of Technology,
S-10044 Stockholm, Sweden
L. Lovász
Department of Computer Science, Eötvös Loránd University,
Budapest, Hungary H-1088, and
Department of Computer Science, Princeton University,
Princeton, NJ 08544
P. W. Shor
AT&T Bell Laboratories, Murray Hill, NJ 07974
 Abstract

We analyze the following (solitaire) game: each node of a graph contains a

pile of chips, and a move consists of selecting a node with at least as many
chips on it as its degree, and let it send one chip to each of its neighbors.
The game terminates if there is no such node. We show that the finiteness
of the game and the terminating configuration are independent of the moves
made. If the number of chips is less than the number of edges, the game
is always finite. If the number of chips is at least the number of edges, the
game can be infinite for an appropriately chosen initial configuration. If the
number of chips is more than twice the number of edges minus the number
of nodes, then the game is always infinite.
The independence of the finiteness and the terminating position follows
from simple but powerful “exchange properties” of the sequences of legal
moves, and from some general results on “antimatroids with repetition”, i.e.,
languages having these exchange properties. We relate the number of steps
in a finite game to the least positive eigenvalue of the Laplace operator of
the graph.
0 Introduction

Studying a certain “balancing game”, J. Spencer (1986) introduced the fol-

lowing (solitaire) game. We start with N chips in a pile. We move bN/2c of
them to the right and bN/2c of them to the left by one unit (if the number
of chips is odd, one chip stays in its original position). In the next step, we
do the same with each of the piles etc. Spencer proved that for the first N
steps, the procedure is well approximated by the “Galton process”, i.e., the
procedure in which a fractional number of chips is allowed and the pile of N
is split into two piles of size N/2.
Inspired by these results, R. Anderson, L. Lovász, P. Shor, J. Spencer,
É. Tardos and S. Winograd (1987) examined this procedure in greater de-
tail. In some respects it turned out to be more natural to study the refined
procedure in which an elementary step was to select a pile with more that
one chip and move one chip of the pile to the left and one to the right.
Among other results, it was shown that this procedure terminates in exactly
k(k − 1)(2k − 1)/6 steps, where k = b N 2+2 c. Moreover, independently of the
selections made during the procedure, the terminating configuration consists
of N consecutive single chip piles, if N is odd, and N + 1 consecutive single
chip piles with the middle pile missing, if N is even.
It is natural to consider this procedure on an arbitrary finite graph G.
We start with a pile of some chips on each node, N chips altogether. A step
consists of selecting a node v which has at least as many chips as its degree,
and move one chip from v to each of its neighbors. We call this step firing the
node v. The game terminates if each node has fewer chips than its degree.
The original game corresponds to the case when the graph is a sufficiently
long path and all chips are piled up on the middle node.
In general, the finiteness of this procedure will depend on both the graph
and the distribution of the chips. However, we shall show that it does not
depend on the choices made during the game: the graph and the original
position of the chips determines both the number of steps and (if this num-
ber is finite) the terminating configuration. It seems that there are various
ways to put this fact in a more general context. First, the positions of the
game can be viewed as a so-called “Church-Rosser system”. Second, the
feasible sequences of moves form a language that has strong “exchange prop-

1
erties”. These properties are extensions of exchange properties of greedoids
and, more specifically, of antimatroids, to languages that allow repetition.
We shall pursue this second approach because it will yield a more thorough
understanding of the structure of feasible games. Among others, it follows
that the positions that can be reached from a given beginning position form a
locally free semimodular lattice. (For more on connections between greedoids
and Church-Rosser systems, see also Faigle, Goecke and Schrader (1987)). A
third way of putting this game in a more general context is to view it as a
Petri net. (For an introduction to Petri nets, see Reisig (1985)). Although
we have not been able to use this connection, it is possible that it could also
contribute to understanding the game.
Our main concern will be to study the finiteness of the procedure. It is
clear that if the number of chips, N , is more than 2 · |E(G)| − |V (G)| then
the procedure cannot terminate: with so many chips, one node will have at
least as many as its degree. If N ≤ 2 · |E(G)| − |V (G)| then clearly we can
have a terminating position. We shall prove that if the number of chips is
less than |E(G)| then the game always terminates; while if |E(G)| ≤ N ≤
2 · |E(G)| − |V (G)| then the game may or may not terminate, depending on
the original configuration of the chips (both possibilities occurring for every
graph and every number of chips in this range).
Our proof gives a rather poor (exponential) bound on the number of steps
in which the procedure terminates (if it terminates at all). G. Tardos (1987)
proved that if the procedure terminates than it terminates in O(|V (G)|4 )
steps. In section 3, we relate the chip-firing game to the eigenvalues of the
Laplace matrix of the graph. This approach yields a bound on the length
of this procedure in terms of the smallest positive eigenvalue of the Laplace
matrix. This bound is, in terms of the number of nodes of the graph, weaker
then the bound given by Tardos, but if the least positive eigenvalue of the
Laplace matrix is not too small (the graph has good expanding properties),
then it may be better.

2
1 The chip firing game and greedoids with
repetition
Let us fix a connected graph G (without loops or multiple edges) on node
set 1, . . . , n, and start by putting ai chips on node i, i = 1, . . . , n. So a ∈ Z+n
P
and i ai = N . Recall that firing the node i means that we decrease ai by
the degree deg(i) of node i, and increase aj by 1 for each neighbor j of i.
Formally, we can define the vector wi by

 deg(i),
 if j = i,
(wi )j = −1, if ij ∈ E(G),


0, otherwise.
Then firing i means subtracting wi from a; this step is legal if a − wi ≥ 0.
A position in the game is any distribution of the chips on the graph,
P
i.e., any vector b ∈ Z+n such that i bi = N . A legal game is any sequence
of positions, starting with a, such that each position is obtained from the
previous by legal step.
Let us record the order in which the nodes are fired in a game; this gives
us a word formed from the nodes as letters. Let L denote the set of records of
all legal games. We shall show that L has special exchange properties which
imply that if it is finite then it is an “antimatroid with repetition” (see below
for definition). This will yield the following theorem:

1.1 Theorem. Given a connected graph and an initial distribution of

chips, either every legal game can be continued indefinitely, or every legal
game terminates after the same number of moves with the same final position.
The number of times a given node is fired is the same in every legal game.

Let E be a finite set and L, a language over E, i.e., a set of finite strings
formed from the elements of E. A subword of a word α is obtained by deleting
letters from α arbitrarily (so a subword need not consist of consecutive letters
of α). We denote by |α| the length of the word α, and by [α], the “score” of
the word α, i.e., the vector in Z+n defined by
[α]i = k, if i occurs k times in α.
For two vectors u, v ∈ Rn , we shall denote by u ∨ v their (coordinate-wise)
maximum and by u ∧ v, their coordinate-wise minimum. We shall denote by
|u|1 the l1 -norm (sum of absolute values of entries) of the vector u.

3
 We say that the language L is left-hereditary, or briefly hereditary, if
whenever a string belongs to it, every beginning section of this string also
belongs to it. We say that L is locally free if the following holds:

(LF) Let α ∈ L and x 6= y, two letters in E such that αx ∈ L, and

αy ∈ L. Then αxy ∈ L.

We say that L is permutable if:

(PM) Whenever α, β ∈ L, [α] = [β] and αx ∈ L for some x ∈ E, we also

have βx ∈ L.

It will be easy to verify that the records of legal games have these proper-
ties. On the other hand, just these properties of a language have quite strong
implications. The key to these is the following “strong” exchange property:

(SE) If α, β ∈ L then α contains a subword α0 such that βα0 ∈ L and

[βα0 ] = [α] ∨ [β].

This last property is a very strong version of the greedoid exchange prop-
erty:

(GE) If α, β ∈ L and |β| < |α| then there exists a letter x in α such that
βx ∈ L.

If L is simple, i.e., no word in L contains the same letter more than once,
then (GE) defines greedoids (among all left-hereditary languages). Since this
case does not concern us in this paper, we refer the reader to Korte and
Lovász (1983) for other definitions of greedoids and some basic examples and
properties. In the case of simple languages, (SE) defines a special class of
greedoids, called antimatroids. Antimatroids were introduced by Edelman
(1980) and Jamison (1982) as closure operations abstracting the combina-
torial properties of convexity and (equivalently) as locally free semimodular
lattices; as special left-hereditary languages they were characterized by Korte
and Lovász (1984). The characterization as left-hereditary simple languages
with the strong exchange property was given by Björner (1985). So the
languages considered in this paper may be viewed as “antimatroids with rep-
etition”. Björner (1985) and Björner and Ziegler (1987) extended the notion
of greedoids to languages with repeated letters. We refer to the latter for a

4
discussion of other related exchange properties.

1.2 Lemma. Every locally free permutable left-hereditary language has

the strong exchange property. Conversely, every language with the strong
exchange property is locally free and permutable.

Proof. We use induction on |[α] ∨ [β]|1 . Let α0 consist of those letters

x in α which are preceded by at least [β]x occurrences of x. We show that
βα0 ∈ L; the fact that [βα0 ] = [α] ∨ [β] is obvious.
Let α00 be the longest prefix of α0 such that βα00 ∈ L. Assume, by way of
contradiction, that α00 6= α0 , and let x be the letter in α0 following α00 . The
fact that x occurs in α0 at all implies that

[α]x = [β]x + [α0 ]x > [β]x + [α00 ]x = [βα00 ]x ,

and hence we can write α as α1 xα2 where [α1 ]x = [βα00 ]x .

Now [α1 ] ≤ [βα00 ], and hence [α1 ] ∨ [βα00 ] = [βα00 ]. Moreover, |[βα00 ]|1 <
|[βα0 ]|1 = |[α] ∨ [β]|1 , and so we can apply the induction hypothesis and
find a subword γ of βα00 such that α1 γ ∈ L and [α1 γ] = [βα00 ]. Clearly x
does not occur in γ. Since both α1 x and α1 γ are in L, we can apply (LF)
repeatedly and obtain that α1 γx ∈ L. Now (PM) implies that βα00 x ∈ L,
which contradicts the choice of α00 .
The converse, which is more straightforward, is left to the reader.

Let L be any left-hereditary language. We say that a word α ∈ L is basic,

if it is not a proper beginning section of any word in the language. It is clear
that if a left-hereditary language has the strong exchange property (or just
the greedoid exchange property) and has a basic word, then every basic word
has the same length, and no word is longer. We call the common length of
basic words the rank of the language. We say that the rank is infinite if there
are no basic words. Note that in this latter case, every word can be extended
indefinitely.
The strong exchange property also implies that if α and β are basic words
then [α] = [β]. We can generalize this observation as follows. We say that
two words α and β in a left-hereditary language L are equivalent (in notation
α ∼ β), if for every string γ, αγ ∈ L iff βγ ∈ L. The equivalence classes of

5
this equivalence relation are called f lats. We say that a flat f is a subflat of
flat g, if every word in f can be extended to a word in g. Note that it would
be enough to require this for a single word in f : if α ∈ f and αγ ∈ g, and
β is equivalent to α, then βγ is in L by the definition of equivalence, and
βγ is trivially equivalent to αγ. For a study of flats in general greedoids, see
Björner (1985), Crapo (1984), and Björner, Korte and Lovász (1985).
We show now that the flats of a left-hereditary language with the strong
exchange property have a particularly simple structure:

1.3 Lemma. Let L be a locally free permutable left-hereditary language

of finite rank and α, β ∈ L. Then the flat defined by α is a subflat of the flat
defined by β iff [α] ≤ [β]. In particular, α ∼ β iff [α] = [β].

Proof. Assume that [α] ≤ [β]. Then by the strong exchange property,
there exists a word γ such that αγ ∈ L and [αγ] = [β]. By permutability,
αγ belongs to the same flat as β, and so α belongs to a subflat of this.
Conversely, assume that α defines a subflat of the flat defined by β, and
let αγ be an extension of α in this flat. Let αγδ be an extension to a basic
word. Then βδ is also in the language, and is clearly also basic. But then,
as remarked, [αγδ] = [βδ] and hence [α] ≤ [β].

So the flats of such a language can be identified with the scores of words.
These scores are partially ordered (by the coordinate-wise ordering) and form
a lattice (since by the strong exchange property, the coordinate-wise maxi-
mum of two scores is again a score). It follows easily from the local freeness
of the language that this lattice is locally free (i.e., the interval from any lat-
tice element x to the join of all elements covering x is a boolean algebra; see
Edelman (1980) and Crapo (1984)). This property implies that the lattice is
semimodular (i.e., if x covers x ∧ y then x ∨ y covers y). The lattice rank of
a score vector equals its l1 -norm.
Now we apply these general results to the language of legal games on a
given graph, starting from a given position.

1.4 Lemma. The records of legal games form a locally free permutable
left-hereditary language.

6
 Proof. It is trivial that this language is left-hereditary. To check property
(LF), note that if node y can be fired after the game α, i.e., it has enough
chips on it, then this remains true after the firing of node x, (this can only
increase the number of chips on y), i.e., αxy is a legal game. To show that
the language is permutable, note that if the games α and β have the same
score then they lead to the same position, and hence any legal continuation
of one is also a legal continuation of the other.

The positions in the game have a very simple interpretation in terms of

the language.

1.5 Lemma. Suppose the language of legal games is of finite rank. Then
two legal games lead to the same position if and only if they have the same
score. (Hence, the positions in the game can be identified with the flats of
the language.)

Proof. Let α and β lead to the same position. Then they have the same
continuations, and hence α ∼ β. By Lemma 1.3, this implies that [α] = [β].
Conversely, if [α] = [β] then, as already used, they lead to the same position.

Now the proof of Theorem 1.1 is immediate: terminating legal games

correspond to basic words, and so by the above, they have the same length
and same score, and lead to the same position.

1.6 Remark. The argument of Lemma 1.4 carries over to the following
more general “game”: let Ax ≤ b (b ≥ 0) define a non-empty polyhedron P
in Rd . Let E be any finite set of vectors in Rd with the property that for each
row vector a of A, at most one of the inner products ax (x ∈ E) is positive.
Consider the language of those sequences x1 x2 . . . xn , xi ∈ E, which satisfy
x1 + . . . + xj ∈ P for all 1 ≤ j ≤ n. Then this language is left-hereditary,
locally free and permutable.

7
2 The finiteness of the game
In this section we study the question of which chip firing games are infinite
and which are finite. Let G be a connected graph with n nodes and m edges,
and let us have N chips. We start with a simple lemma.

2.1 Lemma. If a chip-firing game is infinite then every node is fired

infinitely often.

Proof. There is a node v which is fired infinitely often. Let u be any

neighbor of v. Then every time v is fired, it sends a chip to u. Since u cannot
compile more than N chips, it must be fired infinitely often itself. Since the
graph is connected, this means that every node is fired infinitely often.

Let us contrast this with the following fact proved by G. Tardos (1987):

2.2 Lemma. If the chip-firing game terminates then there is a node

which is not fired at all.
Proof. For sake of completeness, we give Tardos’s simple proof. We
want to show that once every node has been fired, the game cannot get
stuck. Consider the node v that has been idle for the longest time. Then all
of its neighbors have been fired since the last firing of v, and since it received
a chip from each of its neighbors, v must have at least deg(v) chips. So v
can be fired.

The main result in this section relates the finiteness of the game to the
number of chips.

2.3 Theorem.
(a) If N > 2m − n then the game is infinite.

(b) If m ≤ N ≤ 2m − n then there exists an initial configuration guaran-

teeing finite termination and also one guaranteeing infinite game.

8
 (c) If N < m then the game is finite.

Proof. It is obvious that if N > 2m − n then the game cannot terminate:

there is always a node v with at least deg(v) chips on it. It is also obvious
that if N ≤ 2m − n then we can place at most deg(v) − 1 chips on each node
v, and so there are configurations with no legal move.
Next we show that if the number of chips is at least m then there is
an initial configuration that leads to an infinite game. Clearly it suffices to
show this for N = m. Consider any acyclic orientation of G, and let deg+ (v)
denote the out-degree of node v. Let us place deg+ (v) chips on each node v;
this is clearly possible since there are m chips altogether. We claim that this
game is infinite.
Observe first that there must exist a node that can be fired in the first
step. In fact, the orientation is acyclic, which implies that there is a source,
i.e., a node v with deg(v) = deg+ (v). Now fire this node, and observe that
the resulting distribution of chips can also be obtained from an orientation:
if we reverse the edges incident with v, we decrease the outdegree of v by
deg+ (v), and increase the outdegree of each of its neighbors by 1. Since
reversing the edges incident with a source does not create any directed cycle,
we can find a source in the resulting digraph which can again be fired, etc.
The proof of (c) is motivated by the previous construction. Consider any
distribution of N < m chips on the nodes; let f (v) denote the number of
chips on node v. Also consider an acyclic orientation of the graph G and the
quantity X
T = max{0, f (v) − deg+ (v)}.
v∈V (G)

We say that a node u is deficient if f (u) < deg+ (u); by our hypothesis that
N < m, there must exist a deficient node. We are going to show that we
can modify the orientation during the game so that T never increases and
if the set of deficient nodes changes then T must actually decrease. If the
game is infinite then every node gets fired infinitely often, and hence the set
of deficient nodes must change infinitely often (since a deficient node cannot
be fired). Since T cannot decrease infinitely often, this implies that the game
is finite.
Consider the node v that is first fired; we have f (v) ≥ deg(v). Fire v
and reverse the orientation of all edges leaving v. We do not create any

9
cycle. Moreover, we do not increase T since the term in T corresponding to
v decreases by deg(v) − deg+ (v) while each of the deg(v) − deg+ (v) terms
corresponding to the nodes u for which uv ∈ E(G) increases by at most 1.
Also note that if such a node u was deficient then T actually decreases; if
none of these was deficient then the set of deficient nodes did not change. As
remarked, this proves the theorem.

2.4 Remark. Suppose the chip-firing game is played on a directed graph

with n nodes and m arcs, so that when vertex v is fired then deg+ (v) chips
are moved from it along the outgoing arcs to the neighbors dominated by v.
Suppose furthermore that the graph has no sinks, so that every node has an
arc v leaving it. The same analysis as in Lemma 1.4 shows that the language
of legal games is left-hereditary, locally free and permutable (see also Remark
1.6). So Theorem 1.1 is valid also for this directed version, and one can ask
for a characterization of those digraphs and initial chip configurations that
guarantee finite termination. Clearly, if N > m − n then the game is infinite
(there must be a node v with at least deg+ (v) chips on it). For N ≤ m − n
the situation is more complicated, and we leave open the problem of whether
there is some general result for the directed games of which parts (b) and (c)
of Theorem 2.3 are special cases.

3 Chips and the eigenvalues of the Laplace

matrix
The Laplace matrix L = LG = (lij ) of the graph G with n nodes is the n × n
matrix whose rows and columns are indexed by the nodes and

 deg(i),
 if i = j;
lij =  −1, if ij ∈ E(G);

0, otherwise.

This matrix is symmetric and clearly 0 is an eigenvalue of it (with corre-

sponding eigenvector (1, . . . , 1)T ). Moreover, L is positive semidefinite, which

10
is easily seen by expressing the corresponding quadratic form as a sum of
squares: X
xT Lx = (xi − xj )2 .
ij∈E(G)

If G is connected (which we assume throughout this paper), then all the

other eigenvalues of L are positive, since the above quadratic form is clearly
non-zero unless all entries of x are the same. The least of these non-zero eigen-
values, which we shall denote by λ1 , is closely related to several “expanding”
properties of G (see Alon 1986). We shall need the following (probably folk-
lore) lower bound on it (cf. also Alon 1986). Let d denote the diameter of
G.

3.1 Lemma. λ1 ≥ 1/(nd).

Proof. Let w be an eigenvector belonging √ to λ1 with unit length. Then

there is a node, say node 1, such that |w1 | ≥ 1/ n. We may assume without
loss of generality that w1 > 0. Since w is orthogonal to the eigenvector
P
belonging to the eigenvector 0, we have j wj = 0, and hence there exists a
k such that wk < 0. Let, say, nodes 1, 2, . . . , k form a minimal path from 1
to k (so that k − 1 ≤ d in this labelling).
Now we have

λ1 = wT Lw
X k−1
X
= (wi − wj )2 ≥ (wi − wi+1 )2
ij∈E(G) i=1
k−1
X
1
≥ ( (wi − wi+1 ))2
k − 1 i=1
1 1 1
= (w1 − wk )2 ≥ ≥ .
k−1 (k − 1)n nd

11
 Let us return to the chip firing game. We prove the following bound on
the length of the game:

3.2 Theorem. The number of steps in any terminating chip-firing game

with N chips is at most 2nN/λ1 .

Proof. Assume that we started with ai chips on node i and after a total
of s steps, we have bi chips on node i. Let xi denote the number of times
node i was fired. By the lemma of Tardos mentioned in the previous section,
if the game is finite then there is a node, say node n, such that xn = 0.
Straightforward counting yields the formula

Lx = a − b.

We can express L in the following form:

n−1
X
L= λi · vi viT ,
i=1

where λ1 ≤ . . . ≤ λn−1 are the non-zero eigenvalues of L and v1 , . . . , vn−1 are

corresponding orthogonal eigenvectors of unit length. Consider the general-
ized inverse L0 of L, defined by
n−1
X
0 1
L = vi viT .
i=1 λi

Then, setting vn = √1 1, we get by straightforward computation

n

n−1
X
0 1
LL= vi viT = I − vn vnT = I − J,
i=1 n

where I is the identity matrix and J is the all-1 matrix (since {v1 , . . . , vn } is
an orthonormal basis). Hence
1 T
eTn L0 L = eTn − 1 ,
n
and
s = 1T x = (neTn − neTn L0 L)x = −neTn L0 Lx

12
 n−1
X n T
= −neTn L0 (a − b) = − (e vi )(viT (a − b)).
i=1 λi n

So we can estimate s as follows (using the Cauchy-Schwartz inequality):

n n−1
X
s ≤ |eT vi | · |viT (a − b)|
λ1 i=1 n
v v
u
n−1 u
nu X un−1
X
≤ t (eTn vi )2 t (viT (a − b))2
λ1 i=1 i=1
n 2nN
≤ |a − b| ≤ .
λ1 λ1

Using Lemma 3.1, we get that s < 2n2 dN < 2n3 N . This is slightly worse
than the result of Tardos mentioned in the Introduction, since N may be as
large as Ω(n2 ). On the other hand, for “expanding” graphs λ1 is bounded
from below by a constant, so in this case our bound is tighter.

13
 References

N. Alon (1986), Eigenvalues and expanders, Combinatorica 6, 83-96.

R. J. Anderson, L. Lovász, P. W. Shor, J. Spencer, É. Tardos, and S. Wino-
grad (1987), Disks, Balls, and Walls: Analysis of a combinatorial game,
to appear in Amer. Math. Monthly).
A. Björner (1985), On matroids, groups, and exchange languages, in Matroid
Theory and its Applications (ed. L. Lovász and A. Recski), Coll. Math.
Soc. J. Bolyai 40; North-Holland, Amsterdam; 25-60.
A. Björner, B. Korte and L. Lovász (1985), Homotopy properties of gree-
doids, Advances in Appl. Math. 6, 447-494.
A. Björner and G. M. Ziegler (1987), Introduction to greedoids, to appear
in: Combinatorial Geometries, Advanced Theory (N. White, ed.), Cam-
bridge Univ. Press.
H. Crapo (1984), Selectors. A theory of formal languages, semimodular
lattices branchings, and shelling processes, Advances in Math. 54, 233-
277.
P. Edelman (1980), Meet-distributive lattices and the anti-exchange closure,
Alg. Univ. 10, 290-299.
U. Faigle, O. Goecke and R. Schrader (1987), Church-Rosser decomposi-
tion in combinatorial structures, preprint, Inst. f. Operations Research,
Univ. Bonn.
R. E. Jamison (1982), A perspective on abstract convexity: classifying align-
ments by varieties, in Convexity and Related Combinatorial Geometry
(ed. D. C. Kay and M. Breehm), 113-150.
B. Korte and L. Lovász (1983), Structural properties of greedoids, Combi-
natorica 3, 359-374.
B. Korte and L. Lovász (1984), Shelling structures, convexity, and a happy
end, in Graph Theory and Combinatorics, Proc. Cambridge Combina-
torial Conference in Honour of P. Erdős (ed. B. Bollobàs), Academic
Press, New York/London, 219-232.

14
W. Reisig (1985), Petri Nets: An Introduction, Springer-Verlag, New York.

J. Spencer (1986), Balancing vectors in the max norm, Combinatorica 6,

55-66.

G. Tardos (1987), Polynomial bound for a chip firing game on graphs

(preprint).

15