Quantum

Game Theory
Steven E. Landsburg

An n-person game consists of Nash equilibrium either is or is not a good predic-

1. n sets Si (i = 1, . . . , n ). tor of actual behavior depending on exactly what
2. n real-valued functions sort of behavior is being modeled.
Turning now to normative criteria (criteria
Pi : S1 × · · · × Sn → R (i = 1, . . . , n) .
intended to judge the desirability of outcomes),
The set Si is called Player i’s strategy space. The the least controversial is the criterion of Pareto-
function Pi is called Player i’s payoff function. Optimality. Given two n -tuples of strategies
This formulation is general enough to model s = (s1 , . . . , sn ) and s = (s1 , . . . , sn ) , we say that s
pretty much any real-world strategic interaction: we (weakly) Pareto dominates s if Pi (s) ≥ Pi (s ) for all
take Si to be the set of actions available to i; it is easy to verify that Pareto dominance is a
the character called Player i, we imagine partial order, and we say that s is Pareto optimal
that each player must choose some ac- if it is maximal for this order.
tion, we imagine that those actions have In general, we have no guarantees of existence
some joint consequence and that Pi or uniqueness for Nash equilibria or for Pareto op-
measures Player i’s assessment of the tima. When they do exist, Nash equilibria and Pareto
value of that consequence. Given such a optima might or might not coincide. The single
model, one can ask questions that tend to most famous example in the history of game
fall into two broad categories: First, what do we
theory is the Prisoner’s Dilemma, one form of
think the players will do (assuming, ordinarily, that
which can be represented by the following matrix:
they are sublimely selfish and sublimely rational)?
Second, what do we think the players ought to do Player Two
(according to some standard of fairness or justice
or morality)? Questions of the first sort call for so- C D
lution concepts; questions of the second sort call
for normative criteria.
The most thoroughly studied solution concept (1)
is Nash equilibrium, an outcome that results when
C (3, 3) (0, 5)
players maximize their own payoffs, taking other Player One
players’ behavior as given. More precisely, an
n-tuple of strategies (s1 , . . . , sn ) is a Nash equilib- D (5, 0) (1, 1)
rium if for every i and for every s ∈ Si ,
Pi (s1 , . . . , si−1 , si , si+1 . . . sn )
≥ Pi (s1 , . . . , si−1 , s, si+1 , . . . , sn ). Here the rows are indexed by Player 1’s strategy set
{C, D}, the columns are indexed by Player 2’s strat-
egy set {C, D} , and the (i, j) entry is
(P1 (i, j), P2 (i, j)) . The associated story, which the
Steven E. Landsburg is adjunct associate professor of reader can see enacted every Tuesday night on
economics at the University of Rochester. His email N.Y.P.D. Blue, goes as follows: Two prisoners have
address is [email protected]. jointly committed a crime. They are separated and

394 NOTICES OF THE AMS VOLUME 51, NUMBER 4

invited to testify against each other. Each receives and Even with probability 1 − p ”. It is then natural
the same menu of options: You’re facing a ten- to define new payoff functions
year sentence for this crime. You can either coop-
erate with your buddy (strategy C) by refusing to Pi+ (p, q) = pqPi (Odd, Odd)
testify against him or defect from your criminal + p(1 − q)Pi (Odd, Even)
partnership (strategy D) by testifying. If you both
+ (1 − p)qPi (Even, Odd)
cooperate (i.e., if you both stay silent), we’ll have
to convict you of a lesser crime, which will take + (1 − p)(1 − q)Pi (Even, Even).
three years off each of your sentences. But if you More generally, given any game G with finite
defect while your buddy cooperates, we’ll take five strategy sets Si , we define a new game G+ as fol-
years off your sentence (while he serves a full lows: Let Si+ be the set of all probability distribu-
term). And if you both defect, we’ll take one year tions on Si , and define
off each of your sentences.
As the reader can (and should) verify, the unique Pi+ : S1+ × · · · × Sn
+
→R
Nash equilibrium (D, D) is also the unique out-
come that is not Pareto optimal. Rational selfish by
prisoners always choose the one strategy pair that
(3) Pi+ (s1+ , . . . , sn+ )
both can agree is undesirable—in the sense that 
they would both prefer (C, C) .1 = Pi (s1 , . . . , sn )ds1+ (s1 ) . . . dsn+ (sn ).
The simplest game without a Nash equilibrium S1 ×···×Sn
is “odds and evens”, represented by the game ma-
(The restriction to games with finite strategy sets
trix
Player Two is so that we do not have to worry about conver-
gence issues in (3).) One proves via standard fixed
O E point theorems that the game G+ has at least one
Nash equilibrium, the key point being that each Si+,
unlike the original Si , can be identified with a con-
(2) vex subset of a Euclidean space on which the Pi+
O (1, −1) (−1, 1) are continuous.
Player One Thus in the case of Odds and Evens, G+ is a
better model of reality than G is. I want to argue
E (−1, 1) (1, −1) that the same thing is true more generally: If G
is any game with finite strategy spaces intended
to model some real-world interaction, then G+ is
always a better model of that same interaction.
Suppose there is a Nash equilibrium in which
Here’s why: In the real world, players must com-
Player 1 plays Even. Then Player 2 plays Odd, so
municate their strategies either to each other or
Player 1 plays Odd—contradiction; and similarly
to a referee or to an interrogating detective, who
with Odd and Even reversed. Thus there is no Nash
then computes the payoffs. And as a practical
equilibrium, and without an alternative solution
matter, it is quite impossible for a referee or any-
concept we are unable to predict anything other
one else to prohibit the use of mixed strategies.
than paralysis on the part of both players. But of
Player 1 announces “I defect!” How can the referee
course anyone who has ever played this game
know whether Player 1 arrived at this strategy
knows what actually happens: players randomize
through a legal deterministic process or an illegal
their strategies, and each wins half the time.
random one?
One might be tempted to conclude that Nash Because there is no way to prohibit mixed
equilibrium is the wrong solution concept for this strategies in practice, we might as well allow
game. A better conclusion is that the mathemati- them in the model. More generally, whenever
cal structure (2) is a poor model for the real-world the real world imposes limits on referees’ ability
game of odds and evens. A better model would to observe and/or calculate, we should improve
allow for mixed (i.e. randomized) strategies. So we the model by adjusting the strategy spaces and
replace the strategy space Si = {Odd, Even} with payoff functions accordingly.
the unit interval Si+ = [0, 1] , using p ∈ [0, 1] to Quantum game theory begins with the obser-
model the strategy “play Odd with probability p vation that the technology of the (near?) future is
1 In fact, things are even worse than that. Each player has likely to dictate that much communication will
D as a dominant strategy, which means that D is the op- occur through quantum channels. For example,
timal play regardless of whether the other player chooses players might communicate their strategies to a ref-
C or D . This is a far more powerful reason to anticipate eree via email composed on quantum computers.
an outcome of (D, D) than the mere fact that (D, D) is a Such communication creates the prospect of new
Nash equilibrium. strategies whose use the referee cannot detect and

APRIL 2004 NOTICES OF THE AMS 395

 therefore cannot prohibit: Instead of cooperating on his first turn and the matrix U −1 on his second.
or defecting (or randomizing between the two), a Here’s why that’s clever: If Player Two fails to flip,
player might send a message that is some quantum then the net result of the three operations is
superposition of the messages “I cooperate” and
“I defect”. To read the message, the referee must (4) U −1 ◦ I ◦ U = I;
destroy the superposition, along with any evidence
that the superposition ever existed, which makes whereas if Player Two flips, the net result is
superpositions effectively impossible to prohibit.  √ √ 
− 2+i 2
0
What cannot be prohibited must be (5) U −1 ◦ F ◦ U =  2 √ √ .
2−i 2
allowed; therefore, if we want to model accurately 0 2
the behavior of games in which players have access
to “quantum moves”, we should expand our This is great for Player One, because both (4)
strategy spaces accordingly. and (5) map the state represented by H into itself
One might guess that a quantum move is just (remember that any scalar multiple of H is equiv-
one more way to implement a mixed strategy, so alent to H). Thus whether or not Player Two flips
that there is nothing new here for game theory. The the coin, Player One is guaranteed a win.
physicist David Meyer [M] was the first to publish Meyer’s example shows that quantum moves
a counterexample to that guess. In Meyer’s exam- can be more powerful than mere mixed strategies,
ple a single coin is passed back and forth between at least in a context where quantum moves are
two blindfolded players. The coin starts out heads available to only one player. But of course it is
up (call this state H ). Player One has the option ei-
more natural to ask what happens if both players
ther to flip the coin or to return it unflipped. Then
are given access to a full set of quantum moves.
Player Two (still blindfolded so he doesn’t know
The first example of a full-fledged quantum
Player One’s first move) has the same option: Flip
game is due to Jens Eisert, Martin Wilkens, and
or don’t flip. And finally, Player One gets another
Maciej Lewenstein [EWL]. Let G be a two-by-two
turn. If the coin ends up in its initial state H, Player
game, that is, a game with two players, each of
One wins. If it ends up in the opposite state T (tails
whom has a two-point strategy space, say {C, D}.
up), Player Two wins.
Here Player One has four strategies (“flip, flip”, (The reader will have no difficulty extending this
“flip, don’t flip”, etc.). Player Two has two strate- construction to n-by-m games.) Each player is given
gies (“flip” and “don’t flip”). In any mixed strategy a coin that he returns to the referee either in its
Nash equilibrium, Player Two flips with probabil- original state (to indicate a play of C) or flipped (to
ity .5, Player One flips an even number of times with indicate a play of D). A player with access to quan-
probability .5, and each wins half the games. tum moves can act on his coin with any unitary
Now suppose we treat the coin as a quantum matrix and therefore return it in a state
object. Its state is an equivalence class of nonzero αH + βT
vectors in the complex vector space spanned by H
(“heads”) and T (“tails”); two vectors are equivalent where α and β are arbitrary complex numbers,
if one is a scalar multiple of the other. A physical not both zero. (Here and in everything that follows
operation on the coin corresponds to a unitary I will freely abuse notation by writing αH + βT
operation on the state space; in particular, we can both for an element of the vector space C2 and for
set things up so the operation “not flip” is repre- the state represented by that vector.) When the
sented by the identity transformation and the referee observes the coin, it will appear to be un-
operation “flip” is represented (with respect to flipped or flipped with probabilities proportional
the basis {H, T} ) by the unitary matrix to |α|2 and |β|2 . As long as the coins can be treated
  as independent quantum entities, then indeed all
0 1 we have is a fancy way to implement a mixed strat-
F= .
−i 0 egy—in other words, nothing new for game theory.
Meyer’s example was more interesting because
both players acted on a single coin. Eisert, Wilkens,
Now suppose that Player One (but not Player Two)
and Lewenstein (referred to henceforth as EWL)
has access to a full array of quantum moves; that
make their example more interesting by assuming
is, instead of simply flipping or not flipping, he can
the players’ coins are entangled so that there
apply any unitary matrix he chooses. In particular,
is a single state space for the pair of coins.
if he is clever, Player One will choose the matrix
 √  Explicitly, let C2 be the two-dimensonal complex
−1−i 2 vector space spanned by symbols H and T; then the
−
U =  √22 2 
−1+i state space for an entangled pair is
2 2
(6) S = (C2 ⊗ C2 − {0})/ ∼

396 NOTICES OF THE AMS VOLUME 51, NUMBER 4

where ∼ is the equivalence relation that identifies SU2 /{±1} → U2 /S1 ,
a vector with all its nonzero scalar multiples. As
before, I will write, for example, H ⊗ H both for so we can just as well define Player One’s strategy
a vector in the space C2 ⊗ C2 and for the state it space to be SU2 /{±1} . Moreover, the group SU2 can
represents. A physical operation on the first coin be identified with the group S3 of unit quaternions
is represented by a two-by-two unitary matrix U via the map
acting on the state space S as U ⊗ 1 . A physical SU2 → S3
operation on the second coin is represented by a  
A B
two-by-two unitary matrix V acting on the state → A + Bj .
space as 1 ⊗ V T . C D
Now EWL conjure the following scenario: A pair The same analysis applies to Player Two. Thus we
of coins starts in the state2 define the strategy spaces
Q Q
(H ⊗ H) + (T ⊗ T). (8) S1 = S2 = S3 /{±1} = RP3 .
As before, each player is handed one of the Using language loosely, I will often identify a
coins and invited to indicate a play of C by strategy with either of the two quaternions that
applying the identity matrix (“leaving the coin represent it.
untouched”) or to indicate a play of D by applying Next we motivate definitions for the payoff
the matrix   functions. Start with the game G:
0 1
F=
−i 0 Player Two

(“flipping the coin”). As long as the players restrict

C D
themselves to the two-point strategy space {C, D},
the pair of coins lands in one of the four states
(9) C (X1 , Y1 ) (X2 , Y2 )
(7a) CC = (H ⊗ H) + (T ⊗ T) ,
Player One
(7b) CD = (H ⊗ T) − i(T ⊗ H) ,
D (X3 , Y3 ) (X4 , Y4 )
(7c) DC = (H ⊗ T) + i(T ⊗ H) ,

(7d) DD = (H ⊗ H) − (T ⊗ T) . Suppose Player One plays the quaternion p

and Player Two plays the quaternion q. Write the
The referee now performs an observation to product as
determine which of the states (7a–d) the coins
occupy and makes appropriate payoffs. (10) pq = π1 (pq) + π2 (pq)i + π3 (pq)j + π4 (pq)k
If players cannot be trusted to restrict them-
where the πα are real numbers unique up to a
selves to the two strategies C and D, then the
sign, because p and q are defined up to a sign. Using
mathematical modeler should replace the game G
the notation of (7a–d), a chase through the iso-
with a new game GQ that expands the strategy
morphisms reveals that the coin is transformed
spaces accordingly. Player One’s strategy space
from the initial state CC to a final state
should consist of the operations that can be
effected on the state space S (defined in (6)) via the π1 (pq)CC + π2 (pq)CD + π3 (pq)DC + π4 (pq)DD.
action of unitary matrices on the first variable. Let
U2 be the group of two-by-two unitary matrices. The When the referee observes the coins’ joint state, he
matrices that fix S are the scalar matrices, which observes each of the four outcomes with proba-
form a subgroup S1 ⊂ U2 . Therefore, we define bilities
Player One’s strategy space to be the group U2 /S1 . Prob (CC) = π1 (pq)2 Prob (CD) = π2 (pq)2
Let SU2 ⊂ U2 be the subgroup of matrices with Prob (DC) = π3 (pq) 2
Prob (DD) = π4 (pq)2 .
determinant one; then inclusion induces an iso-
morphism Thus we should define the payoff functions by
2 Everything to follow depends heavily on the assumption
that the coins start in the “maximally entangled” state Q

4

(H ⊗ H) + (T ⊗ T) . If they were to start in a state of the form (11) P1 (p, q) = πi (pq)2 Xi ,

s ⊗ t , then the construction that follows would only re- α=1
construct the classical mixed strategy game G+ . [EWL]
4
Q
studies families of quantum games parameterized by the P2 (p, q) = πi (pq)2 Yi .
choice of initial state. α=1

APRIL 2004 NOTICES OF THE AMS 397

 Equations (8) and (11) define the quantum game Pi+ (s1+ , . . . , sn+ )
GQ associated to the game G of equation (9). The 
quantum game GQ is not at all the same as the = Pi (s1 , . . . , sn )ds1+ (s1 ) . . . dsn+ (sn ).
S1 ×···×Sn
mixed strategy game G+ . Using mixed strategies,
the players can jointly effect some but not all (We identify the original strategy space Si with a
probability distributions over the four possible subset of Si+ by identifying the point s with the prob-
outcomes; there is, for example, no pair of ability measure concentrated on s .) Now if we equip
mixed strategies that can effect the probability RP3 with its Borel σ -algebra, then GQ acquires, quite
distribution naturally, the structure of a measurable game. Thus
Prob (CC) = 1/2 Prob (CD) = 0 equation (12) becomes a meaningful definition.
The payoff functions in G! should be called Pi!
Prob (DC) = 0 Prob (DD) = 1/2 . Q+
or Pi , but I will just call them Pi .
By contrast, in the game GQ any probability dis- In the game G! there is always at least one Nash
tribution at all is realizable; in fact, more is true: equilibrium, namely (µ, µ) where µ is the uniform
taking Player One’s strategy as given, Player Two probability distribution on RP3 . There are usually
can choose a strategy that effects any desired prob- more interesting equilibria as well. For example, let
ability distribution. (Proof: Let Player One choose us return to the Prisoner’s Dilemma (1). It is easy
strategy p, and let r be an arbitrary unit quaternion; to verify that the following pair of mixed strategies
then Player Two can play p−1 r .) constitutes a Nash equilibrium:
Thus Nash equilibria must be a great rarity in
quantum games; in fact, Nash equilibria exist only µ: Player 1 plays the quaternions 1 and k , each
when there exists α ∈ {1, 2, 3, 4} that maximizes with probability 1/2 .
both Xα and Yα ; in other words, Nash equilibria (13)
exist only when there is no conflict between the ν : Player 2 plays the quaternions i and j , each
players. with probability 1/2 .
But nobody would ever actually play the game
GQ anyway. Just as there is nothing to stop our Indeed, for any unit quaternion q = a + bi + cj + dk ,
players from adopting quantum strategies, there we have
is also nothing to stop them from adopting mixed 5 5
quantum strategies. So we really want to study the P2 (µ, q) = 2a2 + b2 + c 2 + 2d 2 ,
2 2
game
so that q is an optimal response to µ if and only if
(12) G! = (GQ )+ . a = d = 0 . Thus q = i , q = j are optimal responses;
So far, we have defined the game G+ only when whence so is the strategy ν . Similarly with the play-
the game G has finite strategy spaces, which is cer- ers reversed.
tainly not the case for GQ . So to turn (12) into a In the Nash equilibrium (13), each player’s pay-
definition, we must first give a more general defi- off is 5/2, so (13) Pareto dominates the unique
nition of G+ . classical equilibrium (D, D) (where the payoffs are
both 1 ). Nevertheless, (13) is still Pareto subopti-
Definition. A measurable game consists of mal, being dominated by (C, C) .
More generally, we would like to classify the
1. n measure spaces (Si , Oi ) (i = 1, . . . , n ). (That
Nash equilibria in G! where G is an arbitrary two-
is, Si is a set and Oi is a σ -algebra on Si .)
by-two game. The results that follow are from the
2. n bounded measurable functions
forthcoming article [L].
Pi : S1 × · · · × Sn → R (i = 1, . . . , n). Given a strategy µ, we define the optimal response
sets
Definition. Let G be a measurable game. Then a
mixed strategy for Player i consists of a probability
O1 (µ) = {p ∈ RP3 |P1 (p, µ) is maximized},
measure on the space (Si , Oi ) .
O2 (µ) = {q ∈ RP3 |P2 (µ, q) is maximized}.
Definition. Let G be a measurable game. Then we
Thus (µ, ν) is a Nash equilibrium if and only if ν
define a game G+ as follows: Let Si+ be the set of all
is supported on O2 (µ) and µ is supported on O1 (ν) .
probability measures on Si and define
This leads us to ask: Which subsets of RP3 can occur
as Oi (µ) ? The answer is:
Pi+ : S1+ × · · · × Sn
+
→R
Theorem. For any µ , each Oi (µ) is a projective
by hyperplane in RP3 .

398 NOTICES OF THE AMS VOLUME 51, NUMBER 4

Proof. We have That cuts the set of (equivalence classes of)
 potential Nash equilibria way down to size. But
(14) P1 (p, µ) = P1 (p, q)dµ(q), we can cut it down much further. Call a pair of
weighted frames (µ, ν) realizable if it is a Nash
which, for fixed µ, is a quadratic form in the coef- equilibrium for some quantum game. A reason-
ficients πα (p) and hence maximized (over S3 ) on able first guess is that every pair (µ, ν) is realizable.
the intersection of S3 with the linear subspace of But the truth is quite the opposite: the main
R4 corresponding to the maximum eigenvalue of theorem of [L] establishes highly restrictive (and
that form. O1 (µ) is the image in RP3 of that linear easily checkable) necessary conditions for realiz-
subspace. Similarly, of course, for O2 . ability. Modulo some minor technical provisos,
The theorem (or more precisely, the proof of the the theorem implies that for each µ there are
theorem) has some immediate corollaries of con- only a small finite number of ν such that (µ, ν) is
realizable. In many cases, that small finite number
siderable interest.
is zero; in the remaining cases, the relevant strate-
Corollary. Let G be the arbitrary two-by-two game gies ν are easy to describe. Given a particular game,
of expression (9). Then in any mixed strategy quan- this makes the search for Nash equilibria a tractable
tum Nash equilibrium, Player One earns a payoff problem.
of at least (X1 + X2 + X3 + X4 )/4 , and Player Two One can view quantum game theory as an exercise
earns a payoff of at least (Y1 + Y2 + Y3 + Y4 )/4 . in pure mathematics: Given a game G, we create a
new game G! and we study its properties. But game
Proof. The quadratic form (14) has trace theory has historically been interesting primarily
X1 + X2 + X3 + X4 and hence a maximum eigen- for its applications. As with quantum computing,
value of at least one fourth that amount. Similarly, the applications of quantum game theory lie in the
of course, for Player Two. future.3 The immediate task is to prove theorems
that we expect will be useful a generation from now.
Corollary. Suppose the game (9) is zero-sum, mean-
References
ing that Xα + Yα = 0 for all α. Then in any mixed
[EWL] J. EISERT, M. WILKENS, and M. LEWENSTEIN, Quantum
strategy quantum equilibrium, Player One earns games and quantum strategies, Phys. Rev. Lett. 83
exactly (X1 + X2 + X3 + X4 )/4 and Player Two earns (1999), 3077.
exactly (Y1 + Y2 + Y3 + Y4 )/4 . [L] S. LANDSBURG, Nash equilibria in quantum games,
preprint available at http://www.landsburg.
To describe Nash equilibria in general, we need com/qgt.pdf.
to describe probability measures on RP3 in general. [M] D. MEYER, Quantum strategies, Phys. Rev. Lett. 8 (1999),
Of course, there are a huge number of such mea- 1052.
sures, but fortunately they fall quite naturally into
large equivalence classes. In particular, we say that
two mixed quantum strategies µ and µ are equiv-
alent if for all mixed quantum strategies ν and for
all α ∈ {1, 2, 3, 4} , we have

πα (pq)2 dµ(p)dν(q)
RP3 ×RP3

= 3 3
πα (pq)2 dµ (p)dν(q)
RP ×RP
where the πα are the coordinate functions on the
quaternions as in (10). That is, two strategies are
equivalent if one can be substituted for the other
without changing either player’s payoffs in any
game.
It turns out that equivalence classes of strate-
gies are quite large. I proved in [L] that every mixed
quantum strategy is equivalent to a strategy sup-
ported on (at most) four mutually orthogonal points
(that is, four points that are the images of mutu-
ally orthogonal points in S3 ). Thus an equivalence 3 It has been suggested that quantum games might have
class of measures can be identified with a “weighted immediate applications in the theory of evolution,
frame” in R4 , that is, a set of mutually orthogonal assuming that genetic mutations are driven by quantum
vectors, each assigned a positive weight in such a events. To my knowledge there is no evidence to support
way that the weights add up to one. this admittedly intriguing speculation.

APRIL 2004 NOTICES OF THE AMS 399