MSE 334 Computation of Equilibria Lecture 4 - 03/28/2009

Lemke-Howson Algorithm
Prof. Amin Saberi Scribe: Yu Wu

Consider a two person bimatrix game where the payoff matrices are Am×n and Bm×n . A pair of strategies
(x, y) is a Nash equilibrium for game (A, B) if and only if

∀ 1 ≤ i ≤ m, xi > 0 ⇒ (Ay)i = max(Ay)k

k

∀ m + 1 ≤ j ≤ m + n, yj > 0 ⇒ (xT B)j = max(xT B)k

k

Let M = {1, 2, · · · , m} and N = {m + 1, m + 2, · · · , m + n}. Define the support of x by S(x) = {i | xi > 0}.
Define the support of y similarly.
Definition 1. A bimatrix game (A, B) is non-degenerate if and only if for every strategy x of the row player,
|S(x)| is at least the number of pure best responses to x, and for every strategy y of the column player, |S(y)|
is bigger than or equal to the number of pure best responses to y.
An equivalent definition is: for any y 0 that is a best response to x, |S(x)| ≥ |S(y 0 )|, and for any x0 that
is a best response to y, |S(y)| ≥ |S(x0 )|.
Also note that we can slightly perturb the payoff matrices to make the game non-degenerate. Therefore
WLLOG (with little loss of generality!), we can assume that game (A, B) is non-degenerate.
The following proposition is directly implied by the definition:
Proposition 2. If (x, y) is a Nash equilibrium of a non-degenerate bimatrix game, then |S(x)| = |S(y)|.
Now consider the following Polytopes:
X
P = {(u, x) | xi ≥ 0, xi = 1, xT B ≤ u · 1}
X
Q = {(v, y) | yj ≥ 0, yj = 1, Ay ≤ v · 1}
By the above proposition it is easy to see that every Nash equilibrium can be described as a pair of corner
points of P and Q. For simplicity of notation, consider the following transformations”

P = {x | xi ≥ 0, xT B ≤ 1}
and
Q = {y | yj ≥ 0, Ay ≤ 1}.
There is a one to one correspondence between the corners of P and P , except the zero corner P of P
P . In
fact, for each corner (u, x) of P , x/u is a corner of P ; and for each nonzero corner x of P , (1/ xi , x/ xi )
is a corner of P . The same correspondence exists for Q and Q.
The corner points of P and Q are of our interest because they correspond to special set of strategies of
the players. x is a corner point of P implies some inequalities among {x | xi ≥ 0, xT B ≤ 1} bind. If xi = 0,
then row i is not used in the mixed strategy x; if (xT B)j = 1, then column j is a best response to row
player’s strategy x. Next we give an explicit connection of the corner points of P , Q and Nash equilibria.
Define graph G1 , G2 as follows: The vertices of G1 , G2 are the corner points of P , Q respectively. There
is an edge between x1 and x2 in G1 if and only if x1 and x2 are adjacent corner points of P . Define the
edges of G2 similarly. Then label each vertex x of G1 with the indices of the tight constraints in P , i.e.

L(x) = {i | xi = 0} ∪ {j | (xT B)j = 1}

1
Label G2 similarly. By the non-degeneracy of the game, |L(x)| ≤ m and |L(y)| ≤ n. We have the following
theorem.
Theorem 3. A pair (x, y) is a Nash equilibrium if and only if (x, y) is completely labeled: L(x) ∪ L(y) =
M ∪ N = {1, 2, · · · , m + n}.
Proof Suppose L(x) ∪ L(y) = {1, 2, · · · , m + n}. For each i ∈ M that is in the label set of x, row i is
not used in x, for each j ∈ N that is in the label set of x, column j for the other player is a best response
to x. These conclusions are symmetric for the label set of y. Let M1 = {i|xi = 0}, N2 = {j|(xT B)j = 1};
N1 = {j|yj = 0}, M2 = {i|(Ay)i = 1}. Since |L(x)| ≤ m and |L(y)| ≤ n, then L(x) ∪ L(y) = M ∪ N implies
(M1 , M2 ) is a partition of M and (N1 , N2 ) is a partition of N . Therefore x consists of strategies only in M2 ,
and is a best response to y, y consists of strategies only in N2 and is a best response to x.
On the other hand, if (x, y) is a pair of Nash equilibrium, then M \S(x)) ⊂ L because those rows
are not used in x, and S(y) ∈ L because those columns are best responses to x. Note the game is non-
degenerate, so |S(x)| = |S(y)|, then L(x) = (M \S(x)) ∪ S(y). Similarly, L(y) = (N \S(y)) ∪ S(x). Hence
L(x) ∪ L(y) = M ∪ N .

Finally, we use this connection of Nash equilibrium and graphs G1 , G2 to give a combinatorial (albeit
exponential-time) algorithm of finding a Nash equilibrium in a bimatrix game. The algorithm is by Lemke
and Howson. The basic idea is to pivot alternatingly in P and Q until we find a pair that is completely
labeled.
Let G = G1 × G2 , i.e., vertices of G are defined as v = (v1 , v2 ) where v1 ∈ V (G1 ) and v2 ∈ V (G2 ). There
is an edge between v = (v1 , v2 ) and v 0 = (v10 , v20 ) in G if and only if (v1 , v10 ) ∈ E(G1 ) or (v2 , v20 ) ∈ E(G2 ).
Then for each vertex v = (v1 , v2 ) ∈ V (G), define its label by L(v) = L(v1 ) ∪ L(v2 ). For each k ∈ M ∪ N ,
define the set of ”k-almost” completely labeled vertices by

Uk = {v ∈ V (G)|L(v) ⊇ M ∪ N \{k}}
We have the following key results of Uk :
Theorem 4. For any k ∈ M ∪ N ,

1. (0, 0) and all Nash equilibrium points belong to Uk . Furthermore, their degree in the graph induced by
Uk is exactly one.
2. The degree of every other vertex in the graph induced by Uk is two.

Proof First, note that the label set of (0, 0) and any Nash equilibrium is exactly M ∪ N , so (0, 0) and all
Nash equilibrium points are in Uk for any k. Furthermore, let v = (v1 , v2 ) be (0, 0) or any Nash equilibrium
point. Without loss of generality, suppose k ∈ L(v1 ), where v1 is a corner point of the polytope P . Among
all edges in G1 that v1 is incident to, there is only one direction leading to a vertex v10 without label k (i.e.
loosing the binding constraint corresponding to label k). It is easy to see that (v10 , v2 ) ∈ Uk , therefore there
is only one neighbor of v in Uk .
For part (2), let v = (v1 , v2 ) be any other point in Uk . Then there must be a duplicated label in L(v1 )
and L(v2 ), denoted by l. Similarly to (2), there is exactly one direction of v1 ’s edges in P to drop the label
l, and the new vertex v10 has all labels v1 has except l, so (v10 , v2 ) ∈ Uk . It is symmetric for v2 . Hence there
are two neighbors of v in Uk .

In other words, in a non-degenerate bimatrix game (A, B) the set of k-almost completely labeled vertices in
G and their induced edges consist of disjoint paths and cycles. The endpoints of the paths are the artificial
equilibrium (0,0) and the equilibria of the game.
Corollary 5. A non-degenerate bimatrix game has an odd number of Nash equilibria.

2
Algorithm (Lemke-Howson)
Input: A Non-degenerate bimatrix game (A, B).
Output: One Nash equilibrium of the game.

1. Choose k ∈ M ∪ N .
2. Start with (x, y) = (0, 0) ∈ G. Drop label k from (x, y) (from x ∈ P if k ∈ M , from y ∈ Q if k ∈ N ).
3. Let (x, y) be the current vertex. Let l be the label that is picked up by dropping label k. If l = k,
terminate and (x, y) is a Nash equilibrium of the game. If l 6= k, drop l in the other polytope and
repeat this step.

The Lemke-Howson algorithm starts from the artificial equilibrium (0, 0) and follows the path in Uk .
Since the number of vertices of G is exponential in n and m, so the algorithm may take an exponential time
to find a Nash equilibrium .

Reference

[1]. C. E. Lemke and J. T. Howson, Jr., Equilibrium points of bimatrix games. Journal of the Society for
Industrial and Applied Mathematics 12.
[2]. B. V. Stengel, Computing Equilibria for Two-Person Games. Chapter 45, Handbook of Game Theory
with Economic Applications, Vol. 3.
[3]. B. V. Stengel, Equilibrium Computation for Two-Person Games in Strategic and Extensive Form.
Chapter 3, Algorithmic Game Theory.