Quantum Games
Jens Eisert and Martin Wilkens
arXiv:quant-ph/0004076v1 19 Apr 2000
Institut für Physik, Universität Potsdam, 14469 Potsdam, Germany
(February 1, 2008)
In these lecture notes we investigate the implications of the identification of strategies with quantum operations in game theory beyond the
results presented in [J. Eisert, M. Wilkens, and M. Lewenstein, Phys. Rev.
Lett. 83, 3077 (1999)]. After introducing a general framework, we study
quantum games with a classical analogue in order to flesh out the peculiarities of game theoretical settings in the quantum domain. Special
emphasis is given to a detailed investigation of different sets of quantum
strategies.
PACS-numbers: 03.67.-a, 03.65.Bz, 02.50.Le
I. INTRODUCTION
Game theory is the theory of decision making, which provides powerful tools
for investigating situations in which several parties make decisions according to
their personal interest [1–4]. It gives an account of how the parties would decide
in a situation which involves contest, rivalry, or struggle. Such have been found to
be relevant to social sciences, biology, or economics. Of particular interest to the
theory are games of incomplete information in which the parties may choose their
plan of action with complete knowledge of the situation on rational grounds, but
without knowing what decision the other parties have actually taken.
One important two player game is the so-called Prisoners’ Dilemma [5]. In this
game two players – in the following referred to as Alice and Bob – can independently decide whether they intend to ”cooperate” or ”defect”. Being well aware
of the consequences of their decisions the players obtain a certain pay-off according to their respective choices. This pay-off provides a quantitative characterisation of their personal preferences. Both players are assumed to want to maximise
their individual pay-off, yet they must pick their choice without knowing the other
player’s decision. Fig. 1 indicates the pay-off of Alice and Bob. The players face a
dilemma since rational reasoning in such a situation dictates the players to defect,
although they would both benefit from mutual cooperation. As Alice is better off
with defection regardless of Bob’s choice, she will defect. The game being symmetric, the same argument applies to Bob.
Bob sB
Alice
sA
C
D
C
(3,3)
(0,5)
D
(5,0)
(1,1)
FIG. 1. The pay-off matrix in the Prisoners’ Dilemma game. The first entry refers to Alice’s pay-off, the second to Bob’s. If both players cooperate, they both get 3 units pay-off. If
Bob defects and Alice happens to cooperate, he obtains 5 units, while Alice is in the unfortunate situation in which she does not receive any pay-off at all. Bob faces the same situation
if he chooses to cooperate while Alice prefers to defect. If both defect, they get only 1 unit
pay-off.
1
Formally, Alice has two basic choices, meaning that she can select from two possible strategies sA = C (cooperation) and sA = D (defection). Bob may also take
sB = C or sB = D. The game is defined by these possible strategies on the one
hand, and on the other hand by a specification of how to evaluate the pay-off once
the combination of chosen strategies (sA , sB ) is known, i.e., the utility functions
mapping (sA , sB ) on a number [2]. The expected pay-off quantifies the preference
of the players.
In these lecture notes the idea of identifying strategic moves with quantum operations as introduced in Refs. [6,7] is further developed. This approach appears
to be fruitful in at least two ways [6–9]. On the one hand several recently proposed
applications of quantum information theory can already be conceived as competitive situations where several parties with more or less opposed motives interact.
These parties may, for example, apply quantum operations on a bi-partite quantum
system [10]. In the same context, quantum cloning has been formulated as a game
between two players [11]. Similarly, eavesdropping in quantum cryptography [12]
can be regarded as a game between the eavesdropper and the sender, and there
are similarities of the extended form of quantum versions of games and quantum
algorithms [13,14]. On the other hand a generalisation of the theory of decisions
into the domain of quantum probabilities seems interesting, as the roots of game
theory are partly in probability theory. In this context it is of interest to investigate
what solutions are attainable if superpositions of strategies are allowed for [6,7].
Game theory does not explicitly concern itself with how the information is transmitted once a decision is taken. Yet, it should be clear that the practical implementation of any (classical) game inevitably makes use of the the exchange of voting
papers, faxes, emails, ballots, and the like. In the Prisoners’ Dilemma, e.g., the
two parties have to communicate with an advocate by talking to her or by writing a short letter on which the decision is indicated. Bearing in mind that a game
is also about the transfer of information, it becomes legitimate to ask what happens if these carriers of information are taken to be quantum systems, quantum
information being a fundamental notion of information.
By classical means a two player binary choice game may be played as follows:
An arbiter takes two coins and forwards one coin each to the players. The players
then receive their coin with head up and may keep it as it is (”cooperate”) or turn it
upside down so that tails is up (”defection”). Both players then return the coins to
the arbiter who calculates the players’ final pay-off corresponding to the combination of strategies he obtains from the players. Here, the coins serve as the physical
carrier of information in the game. In a quantum version of such a game quantum
systems would be used as such carriers of information. For a binary choice two
player game an implementation making use of minimal resources involves two
qubits as physical carriers.
II. QUANTUM GAMES AND QUANTUM STRATEGIES
Any quantum system which can be manipulated by two parties or more and
where the utility of the moves can be reasonably quantified, may be conceived as
a quantum game.
A two-player quantum game Γ = (H, ρ, SA , SB , PA , PB ) is completely specified
by the underlying Hilbert space H of the physical system, the initial state ρ ∈
S(H), where S(H) is the associated state space, the sets SA and SB of permissible
quantum operations of the two players, and the utility functionals PA and PB , which
specify the utility for each player. A quantum strategy sA ∈ SA , sB ∈ SB is a
quantum operation, that is, a completely positive trace-preserving map mapping
the state space on itself [15]. The quantum game’s definition also includes certain
implicit rules, such as the order of the implementation of the respective quantum
strategies. Rules also exclude certain actions, as the alteration of the pay-off during
the game.
2
The quantum games proposed in Refs. [6], [7], and [9] can be cast into this form.
Also, the quantum cloning device as described in [11] can be said to be a quantum
game in this sense. A quantum game is called zero-sum game, if the expected payoffs sum up to zero for all pairs of strategies, that is, if PA (sA , sB ) = −PB (sA , sB )
for all sA ∈ SA , sB ∈ SB . Otherwise, it is called a non-zero sum game.
It is natural to call two quantum strategies of Alice sA and s′A equivalent, if
PA (sA , sB ) = PA (s′A , sB ) and PB (sA , sB ) = PA (s′A , sB ) for all possible sB . That
is, if sA and s′A yield the same expected pay-off for both players for all allowed
strategies of Bob. In the same way strategies sB and s′B of Bob will be identified.
ρ
σ
(sA , sB )
PA , PB
FIG. 2. The general setup of a quantum game.
A solution concept provides advice to the players with respect to the action they
should take. The following solution concepts will be used in the remainder of this
lecture. These definitions are fully analogous to the corresponding definitions in
standard game theory [2].
A quantum strategy sA is called dominant strategy of Alice if
PA (sA , s′B ) ≥ PA (s′A , s′B )
(1)
for all s′A ∈ SA , s′B ∈ SB . Analogously we can define a dominant strategy for Bob.
A pair (sA , sB ) is said to be an equilibrium in dominant strategies if sA and sB are
the players’ respective dominant strategies. A combination of strategies (sA , sB ) is
called a Nash equilibrium if
PA (sA , sB )≥ PA (s′A , sB ),
PB (sA , sB )≥ PB (sA , s′B )
(2a)
(2b)
for all s′A ∈ SA , s′B ∈ SB . A pair of strategies (sA , sB ) is called Pareto optimal, if it
is not possible to increase one player’s pay-off without lessening the pay-off of the
other player.
A solution in dominant strategies is the strongest solution concept for a non-zero
sum game. In the Prisoner’s Dilemma defection is the dominant strategy, as it is
favourable regardless what strategy the other party picks. Typically, however, the
optimal strategy depends on the strategy chosen by the other party. A Nash equilibrium implies that neither player has a motivation to unilaterally alter his or her
strategy from this equilibrium solution, as this action will lessen his or her pay-off.
Given that the other player will stick to the strategy corresponding to the equilibrium, the best result is achieved by also playing the equilibrium solution. The
concept of Nash equilibria is of paramount importance to studies of non-zero-sum
games. It is, however, only an acceptable solution concept if the Nash equilibrium is unique. For games with multiple equilibria the application of a hierarchy
of natural refinement concepts may finally eliminate all but one of the Nash equilibria. Note that a Nash equilibrium is not necessarily efficient. In the Prisoners’
Dilemma for example there is a unique equilibrium, but it is not Pareto optimal,
meaning that there is another outcome which would make both players better off.
III. TWO-QUBIT QUANTUM GAMES
In the subsequent investigation we turn to specific games where the classical version of the game is faithfully entailed in the quantum game. In a quantum version
3
of a binary choice game two qubits are prepared by a arbiter in a particular initial
state, the qubits are sent to the two players who have physical instruments at hand
to manipulate their qubits in an appropriate manner. Finally, the qubits are sent
back to the arbiter who performs a measurement to evaluate the pay-off. For such
a bi-partite quantum game the system of interest is a quantum system with underlying Hilbert space H = HA ⊗ HB , HA = HB = C2 , and associated state space
S(H). Quantum strategies sA and sB of Alice and Bob are local quantum operations acting in HA and HB respectively [16]. That is, Alice and Bob are restricted to
implement their respective quantum strategy sA and sB on their qubit only. In this
step they may choose any quantum strategy that is included in the set of strategies
S. They are both well aware of the set S, but they do not know which particular
quantum strategy the other party would actually implement. As the application of
both quantum strategies amounts to a map sA ⊗sB : S(H) → S(H), after execution
of the moves the system is in the state
σ = (sA ⊗ sB )(ρ).
(3)
Particularly important will be unitary operations sA and sB . They are associated
with unitary operators UA and UB , written as sA ∼ UA and sB ∼ UB . In this case
the final state σ is given by
σ = (UA ⊗ UB )ρ(UA ⊗ UB )† .
(4)
From now on both the sets of strategies of Alice and Bob and the pay-off functionals
are taken to be identical, that is,
SA = SB = S and PA = PB = P,
(5)
such that both parties face the same situation.
The quantum game Γ = (C2 ⊗ C2 , ρ, S, S, P, P ) can be played in the following
way: The initial state ρ is taken to be a maximally entangled state in the respective
state space. In order to be consistent with Ref. [6] let ρ = |ψihψ| with
√
|ψi = (|00i + i|11i)/ 2,
(6)
where the first entry refers to HA and the second to HB . The two qubits are forwarded to the arbiter who performs a projective selective measurement on the final
state σ with Kraus operators πCC , πCD , πDC , and πDD , where
√
πCC = |ψCC ihψCC |, |ψCC i = (|00i + i|11i)/ 2,
(7a)
√
πCD = |ψCD ihψCD |, |ψCD i = (|01i − i|10i)/ 2,
(7b)
√
(7c)
πDC = |ψDC ihψDC |, |ψDC i = (|10i − i|01i)/ 2,
√
(7d)
πDD = |ψDD ihψDD |, |ψDD i = (|11i + i|00i)/ 2.
According to the outcome of the measurement, a pay-off of ACC , ACD , ADC , or
ADD is given to Alice, Bob receives BCC , BCD , BDC , or BDD . The utility functionals, also referred to as expected pay-off of Alice and Bob, read
PA (sA , sB ) = ACC tr[πCC σ] + ACD tr[πCD σ] + ADC tr[πDC σ] + ADD tr[πDD σ], (8a)
PB (sA , sB ) = BCC tr[πCC σ] + BCD tr[πCD σ] + BDC tr[πDC σ] + BDD tr[πDD σ]. (8b)
It is important to note that the Kraus operators are chosen in such a way that the
classical game is fully entailed in the quantum game: The classical strategies C and
D are associated with particular unitary operations,
0 1
1 0
.
(9b)
, D∼
C∼
−1 0
0 1
4
C does not change the state at all, D implements a ”spin-flip”. If both parties
stick to these classical strategies, Eq. (8a) and Eq. (8b) guarantee that the expected
pay-off is exactly the pay-off of the corresponding classical game defined by the
numbers ACC , ACD , ADC , ADD , BCC , BCD , BDC , and BDD . E.g., if Alice plays C
and Bob chooses D, the state σ after implementation of the strategies is given by
σ = (C ⊗ D)(ρ) = |ψCD ihψCD |,
(10b)
such that Alice obtains ACD units and Bob BCD units pay-off (see Fig. 1). In this
way the peculiarities of strategic moves in the quantum domain can be adequately
studied. The players may make use of additional degrees of freedom which are
not available with randomisation of the classical strategies, but they can also stick
to mere classical strategies. This scheme can be applied to any two player binary
choice game and is to a high extent canonical.
A. Prisoners’ Dilemma
We now investigate the solution concepts for the quantum analogue of the Prisoners’ Dilemma (see Fig. 1) [17],
ACC = BCC = 3, ADD = BDD = 1,
ACD = BDC = 0, ADC = BCD = 5.
(11a)
(11b)
In all of the following sets of allowed strategies S the classical options (to defect
and to cooperate) are included. Several interesting sets of strategies and concomitant solution concepts will at this point be studied. The first three subsections
involve local unitary operations only, while in the last subsection other quantum
operations are considered as well.
1. One-parameter set of strategies.
The first set of strategies S (CL) involves quantum operations sA and sB which
are local rotations with one parameter. The matrix representation of the corresponding unitary operators is taken to be
cos(θ/2) sin(θ/2)
(12)
U (θ) =
− sin(θ/2) cos(θ/2)
with θ ∈ [0, π]. Hence, in this simple case, selecting strategies sA and sB amounts
to choosing two angles θA and θB . The classical strategies of defection and cooperation are also included in the set of possible strategies, C ∼ U (0), D ∼ U (π). An
analysis of the expected pay-offs PA and PB ,
PA (θA , θB ) = 3| cos(θA /2) cos(θB /2)|2 + 5| cos(θB /2) sin(θA /2)|2
+ | sin(θA /2) sin(θB /2)|2 ,
PB (θA , θB ) = 3| cos(θA /2) cos(θB /2)|2 + 5| sin(θB /2) cos(θA /2)|2
+ | sin(θA /2) sin(θB /2)|2 ,
(13a)
(13b)
shows that this game is the classical Prisoners’ Dilemma game [6]. The pay-off
functions are actually identical to the analogous functions in the ordinary Prisoners’ Dilemma with mixed (randomised) strategies, where cooperation is chosen
with the classical probability p = cos2 (θ/2). The inequalities
PA (D, sB ) ≥ PA (sA , sB ),
PB (sA , D) ≥ PB (sA , sB )
5
(14a)
(14b)
hold for all sA , sB ∈ S (CL) , therefore, (D, D) is an equilibrium in dominant strategies and thus the unique Nash equilibrium. As explained in the introduction this
equilibrium is far from being efficient, because PA (D, D) = PB (D, D) = 1 instead
of the Pareto optimal pay-off which would be 3.
2. Two-parameter set of strategies
A more general set of strategies is the following two-parameter set S (T P ) . The
matrix representation of operators corresponding to quantum strategies from this
set is given by
iφ
e cos(θ/2)
sin(θ/2)
U (θ, φ) =
(15)
− sin(θ/2) e−iφ cos(θ/2)
with θ ∈ [0, π] and φ ∈ [0, π/2]. Selecting a strategy sA , sB then means choosing
appropriate angles θA , φA and θB , φB . The classical pure strategies can be realised
as
C ∼ U (0, 0) and D ∼ U (π, 0).
(16)
This case has also been considered in Ref. [6]. The expected pay-off for Alice, e.g.,
explicitly reads
2
PA (θA , φA , θB , θB ) = 3 |cos(φA + φB ) cos(θA /2) cos(θB /2)|
(17)
2
+ 5 |sin(φA ) cos(θA /2) sin(θB /2) − cos(φB ) cos(θB /2) sin(θA /2)|
2
+ |sin(φA + φB ) cos(θA /2) cos(θB /2) + sin(θA /2) sin(θB /2)| .
It turns out that the previous Nash equilibrium (D, D) of S (CL) is no longer an
equilibrium solution, as both players can benefit from deviating from D. However,
concomitant with the disappearance of this solution another Nash equilibrium has
emerged, given by (Q, Q). The strategy Q is associated with a matrix
i 0
.
(18)
Q ∼ U (0, π/2) =
0 −i
This Nash equilibrium is unique [6] and serves as the only acceptable solution of
the game. The astonishing fact is that PA (Q, Q) = PB (Q, Q) = 3 (instead of 1) so
that the Pareto optimum is realised. No player could gain without lessening the
other player’s expected pay-off. In this sense one can say that the dilemma of the
original game has fully disappeared. In the classical game only mutual cooperation is Pareto optimal, but this pair of strategies does not correspond to a Nash
equilibrium.
3. General unitary operations
One can generalise the previous setting to the case where Alice and Bob can implement operations sA and sB taken from S (GU ) , where S (GU ) is the set of general
local unitary operations. Here, it could be suspected that the solution becomes
more efficient the larger the sets of allowed operations are. But this is not the case.
The previous Pareto optimal unique Nash equilibrium (Q, Q) ceases to be an equilibrium solution if the set is enlarged: For any strategy sB ∈ S (GU ) there exists an
optimal answer sA ∈ S (GU ) resulting in
(sA ⊗ sB )(ρ) = |ψDC ihψDC |,
6
(19)
with ρ given in Eq. (6). That is, for any strategy of Bob sB there is a strategy sA of
Alice such that
PA (sA , sB ) = 5 and PB (sA , sB ) = 0 :
(20)
Take
sA ∼
a b
c d
, sB ∼
−ib a
−d −ic
,
(21)
where a, b, c, d are appropriate complex numbers. Given that Bob plays the strategy sB associated with a particular Nash equilibrium (sA , sB ), Alice can always
apply the optimal answer sA to achieve the maximal possible pay-off. However,
the resulting pair of quantum strategies can not be an equilibrium since again, the
game being symmetric, Bob can improve his pay-off by changing his strategy to
his optimal answer s′B . Hence, there is no pair (sA , sB ) of pure strategies with the
property that the players can only lose from unilaterally deviating from this pair
of strategies.
Yet, there remain to be Nash equilibria in mixed strategies which are much
more efficient than the classical outcome of the equilibrium in dominant strategies PA (D, D) = PB (D, D) = 1. In a mixed strategy of Alice, say, she selects a
particular quantum strategy sA (which is also called pure strategy) from the set of
strategies SA with a certain classical probability. That is, mixed strategies of Alice
and Bob are associated with maps of the form
X (i) (j) (i)
(j)
(i)
(j)
pA pB (UA ⊗ UB )ρ(UA ⊗ UB )† ,
(22)
ρ 7−→ σ =
i,j
(i)
(i)
pA , pB ∈ [0, 1], i, j = 1, 2, ..., N , with
X (i) X (j)
pB = 1.
pA =
(i)
(23)
j
i
(j)
(i)
UA and UB are local unitary operators corresponding to pure strategies sA and
(j)
sB .
The map given by Eq. (27) acts in HA and HA as a doubly stochastic map, that is,
as a completely positive unital map [18]. As a result, the final reduced states trB [σ]
and trA [σ] must be more mixed than the reduced initial states trB [ρ] and trA [ρ] in
the sense of majorisation theory [19]. As the initial state ρ is a maximally entangled
state, all accessible states after application of a mixed strategy of Alice and Bob are
locally identical to the maximally mixed state 1/dim(HA ) = 1/dim(HB ), which is
a multiple of 1.
The following construction, e.g., yields an equilibrium in mixed quantum strate(2)
(1)
gies: Allow Alice to choose from two strategies sA and sA with probabilities
(1)
(2)
(1)
(2)
pA = 1/2 and pA = 1/2, while Bob may take sB or sB , with
−i 0
1 0
(2)
(1)
,
(24a)
, sA ∼
sA ∼
0 i
0 1
0 −i
0 1
(2)
(1)
.
(24b)
, sB ∼
sB ∼
−i 0
−1 0
(2)
(1)
His probabilities are also given by pB = 1/2 and pB = 1/2. The quantum strategies of Eq. (24a) and Eq. (24b) are mutually optimal answers and have the property
that
(i)
(i)
(i)
(i)
PA (sA , sB ) = 0, PB (sA , sB ) = 5,
(i) (3−i)
PA (sA , sB )
= 5,
(i) (3−i)
PB (sA , sB )
7
= 0,
(25a)
(25b)
for i = 1, 2. Due to the particular constraints of Eq. (25a) and Eq. (25b) there exists
no other mixed strategy for Bob yielding a better pay-off than the above mixed
strategy, given that Alice sticks to the equilibrium strategy. This can be seen as
follows. Let Alice use this particular mixed quantum strategy as above and let Bob
use any mixed quantum strategy
(1)
(N )
sB , ..., sB
(26)
(N )
(1)
together with pA , ..., pA . The final state σ after application of the strategies is
given by the convex combination
X X (i) (j) (i)
(j)
(27)
pA pB (sA ⊗ sB )(ρ),
σ=
i=1,2
j
This convex combination can only lead to a smaller expected pay-off for Bob than
(k)
the optimal pure strategy sB in Eq. (26), k ∈ {1, ..., N }. Such optimal pure strate(2)
(1)
gies are given by sB and sB as in Eq. (24b) leading to an expected pay-off for
Bob of PB (sA , sB ) = 2.5; there are no pure strategies which achieve a larger ex(2)
(1)
pected pay-off. While both pure strategies sB and sB do not correspond to an
(1)
(2)
(1)
equilibrium, the mixed strategy where sB and sB are chosen with pB = 1/2
(2)
and pB = 1/2 actually does. Nash equilibria consist of pairs of mutually optimal answers, and only for this choice of Bob the original mixed quantum strategy
of Alice is her optimal choice, as the same argument applies also to her, the game
being symmetric.
This Nash equilibrium is however not the only one. There exist also other fourtuples of matrices than the ones presented in Eq. (24a) and Eq. (24b) that satisfy Eq.
(25a) and Eq. (25b). Such matrices can be made out by appropriately rotating the
matrices of Eq. (24a) and Eq. (24b). In the light of the fact that there is more than
one equilibrium it is not obvious which Nash equilibrium the players will realise.
It is at first not even evident whether a Nash equilibrium will be played at all. But
the game theoretical concept of the focal point effect [20,2] helps to resolve this issue.
To explore the general structure of any Nash equilibrium in mixed strategies we
continue as follows: let
(N )
(1)
UA , ..., UA
(28)
(N )
(1)
together with pA , ..., pA specify the mixed strategy pertinent to a Nash equilib(1)
(N ) (1)
(N )
rium of Alice. Then there exists a mixed strategy UB , ..., UB , pB , ..., pB of Bob
which rewards Bob with the best achievable pay-off, given that Alice plays this
mixed strategy. Yet, the pair of mixed strategies associated with
(1)
(N )
(1)
QUA Q† , ..., QUA Q† ,
(1)
(N )
(1)
(N )
QUB Q† , ..., QUB Q†
(29)
(N )
with pA , ..., pA , pB , ..., pB is another Nash equilibrium. This equilibrium leads
to the same expected pay-off for both players, and is fully symmetric to the previ(N )
(1)
ous one. Doubly applying Q as QQUA Q† Q† , ..., QQUA Q† Q† results again into a
situation with equivalent strategies as the original ones. For a given Nash equilibrium as above the one specified by Eq. (29) will be called dual equilibrium.
However, there is a single Nash equilibrium (R, R) which is the only one which
gives an expected pay-off of PA (R, R) = PB (R, R) = 2.25 and which is identical to
its dual equilibrium: it is the simple map
ρ 7−→ σ = 1/dim(H).
(N )
(1)
(30)
(1)
(N )
Indeed, there exist probabilities pA , ..., pA and unitary operators UA , ..., UA
P (i) (i)
(i)
such that i pA (UA ⊗1)ρ(UA ⊗1)† = 1/dim(H) [19]. If Alice has already selected
8
sA = R, the application of sB = R will not change the state of the quantum system
any more.
Assume that Eq. (28) and Eq. (29) are associated with equivalent quantum strategies. This means that they have to produce the same expected pay-off for all quantum strategies sB of Bob. If Alice and Bob apply sA ⊗ sB they get an expected
pay-off according to Eq. (8a) and Eq. (8b); if Alice after implementation of sA manipulates the quantum system by applying the local unitary operator Q ⊗ 1, they
obtain
PA′ (sA , sB ) = ADD tr[πCC σ] + ADC tr[πCD σ] + ACD tr[πDC σ] + ACC tr[πDD σ], (31a)
PB′ (sA , sB ) = BDD tr[πCC σ] + BDC tr[πCD σ] + BCD tr[πDC σ] + BCC tr[πDD σ]. (31b)
The only sA with the property that PA′ (sA , sB ) = PA (sA , sB ) and PB′ (sA , sB ) =
PB (sA , sB ) for all sB is the map given by Eq. (30).
In principle, any Nash equilibrium may become a self-fulfilling prophecy if the
particular Nash equilibrium is expected by both players. It has been pointed out
that in a game with more than one equilibrium, anything that attracts the players’
attention towards one of the equilibria may make them expect and therefore realise
it [20]. The corresponding focal equilibrium [2] is the one which is conspicuously
distinguished from the other Nash equilibria. In this particular situation there is
indeed one Nash equilibrium different from all the others: it is the one which is
equivalent to its dual equilibrium, the map which simply maps the initial state
on the maximally mixed state. For all other expected pay-offs both players are
ambivalent between (at least) two symmetric equilibria. The expected pay-off the
players will receive in this focal equilibrium –
PA (R, R) = PB (R, R) = 2.25
(32)
– is not fully Pareto optimal, but it is again much more efficient than the classically
achievable outcome of 1 [21].
4. Completely positive trace-preserving maps corresponding to local operations
In this scenario both Alice and Bob may perform any operation that is allowed
by quantum mechanics. That is, the set of strategies S (CP ) is made up of (sA , sB ),
where both sA and sB correspond to a completely positive trace-preserving map
XX
(Ai ⊗ Bj )ρ(Ai ⊗ Bj )†
(33)
(sA ⊗ sB )(ρ) =
i
j
corresponding to a local operation, associated with Kraus operators Ai and Bj with
P
P
i, j = 1, 2, ... . The trace-preserving property requires i A†i Ai = 1 and i Bi† Bi =
1. This case has already been mentioned in Ref. [6]. The quantum strategies sA
and sB do no longer inevitably act as unital maps in the respective Hilbert spaces
as before. In other words, the reduced states of Alice and Bob after application of
the quantum strategy are not necessarily identical to the maximally mixed state
1/dim(HA ).
As already pointed out in Ref. [6], the pair of strategies (Q, Q) of the twoparameter set of strategies S (T P ) is again no equilibrium solution. It is straightforward to prove that the Nash equilibria of the type of Eq. (24a) and Eq. (24b) of
mixed strategies with general local unitary operations are, however, still present,
and each of these equilibria yields an expected pay-off of 2.5.
In addition, as strategies do no longer have to be locally unital maps, it is not
surprising that new Nash equilibria emerge: Alice and Bob may, e.g., perform a
measurement associated with Kraus operators
A1 = |0ih0| A2 = |1ih1|, B1 = D|0ih0| B2 = D|1ih1|.
9
(34)
This operation yields a final state σ = (sA ⊗sB )(ρ) = (|01ih01|+|10ih10|)/2. Clearly
neither Alice nor Bob can gain from unilaterally deviating from their strategy.
One can nevertheless argue as in the previous case. Again, all Nash equilibria
occur at least in pairs. First, there are again the dual equilibria from S (GU ) . Second, there are Nash equilibria (sA , sB ), sA 6= sB , with the property that (sB , sA )
is also a Nash equilibrium yielding the same expected pay-off. The only Nash
equilibrium invariant under application of Q and exchange of the strategies of the
players is again (R, R) defined in the previous subsection, which yields a pay-off
PA (R, R) = PB (R, R) = 2.25. This is the solution of the game is the most general
case. While both players could in principle do better (as the solution lacks Pareto
optimality), the efficiency of this focal equilibrium is much higher than the equilibrium in dominant strategies of the classical game. Hence, also in this most general
case both players gain from using quantum strategies.
This study shows that the efficiency of the equilibrium the players can reach in
this game depends on the actions the players may take. One feature, however is
present in each of the considered sets: both players can increase their expected
pay-offs drastically by resorting to quantum strategies.
B. Chicken
In the previous classical game – the Prisoners’ Dilemma – an unambiguous solution can be specified consisting of a unique Nash equilibrium. However, this
solution was not efficient, thus giving rise to the dilemma. The situation of the
players in the Chicken game [2,3],
ACC = BCC = 6, ACD = BDC = 8,
ADC = BCD = 2, ADD = BDD = 0,
(35a)
(35b)
can be described by the matrix of Fig. 3.
Bob sB
Alice
sA
C
D
C
(6,6)
(2,8)
D
(8,2)
(0,0)
FIG. 3. The pay-off matrix of the so-called Chicken game.
This game has two Nash equilibria (C, D) and (D, C): it is not clear how to anticipate what the players’ decision would be. In addition to the two Nash equilibria
in pure strategies there is an equilibrium in mixed strategies, yielding an expected
pay-off 4 [2].
In order to investigate the new features of the game if superpositions of classical
strategies are allowed for, three set of strategies are briefly discussed:
1. One-parameter set of strategies
Again, we consider the set of strategies S (CL) of one-dimensional rotations. The
strategies sA and sB are associated with local unitary operators
cos(θ/2) sin(θ/2)
(36)
U (θ) =
− sin(θ/2) cos(θ/2)
10
with θ ∈ [0, π],
C ∼ U (0) =
1 0
0 1
, D ∼ U (π) =
0 1
−1 0
.
(37)
Then as before, the quantum game yields the same expected pay-off as the classical
game in randomised strategies. This means that still two Nash equilibria in pure
strategies are present.
2. Two-parameter set of strategies
The players can actually take advantage of an additional degree of freedom
which is not accessible in the classical game. If they may apply unitary operations
from S (T P ) of the type
iφ
e cos(θ/2)
sin(θ/2)
U (θ, φ) =
(38)
− sin(θ/2) e−iφ cos(θ/2)
with θ ∈ [0, π] and φ ∈ [0, π/2] the situation is quite different than with S (CL) .
(C, D) and (C, D) with C ∼ U (0, 0) and D ∼ U (π, 0) are no longer equilibrium
solutions. E.g., given that sA = D the pair of strategies (D, Q) with Q ∼ U (0, π/2)
yields a better expected pay-off for Bob than (D, C), that is to say PB (D, Q) = 8,
PB (D, C) = 2. In fact (Q, Q) is now the unique Nash equilibrium with PA (Q, Q) =
PB (Q, Q) = 6, which follows from an investigation of the actual expected payoffs of Alice and Bob analogous to Eq. (17). This solution is not only the unique
acceptable solution of the game, but it is also an equilibrium that is Pareto optimal.
This contrasts very much with the situation in the classical game, where the two
equilibria were not that efficient.
3. Completely positive trace-preserving maps corresponding to local operations
As in the considerations concerning the Prisoner’s Dilemma game, more than
one Nash equilibrium is present, if both players can take quantum strategies from
the set S (CP ) , and all Nash equilibria emerge at least in pairs as above. The focal
equilibrium is given by (R, R), resulting in a pay-off of PA (R, R) = PB (R, R) = 4,
which is the same as the mixed strategy of the classical game.
IV. SUMMARY AND CONCLUSION
In these lecture notes the idea of implementing quantum operations as strategic
moves in a game is explored [6]. In detail, we investigated games which could
be conceived as a generalisation into the quantum domain of a two player binary choice game. As a toy model for more complex scenarios we studied quantum games where the efficiency of the equilibria attainable when using quantum strategies could be contrasted with the efficiency of solutions in the corresponding classical game. We investigated a hierarchy of quantum strategies
S (CL) ⊂ S (T P ) ⊂ S (GU ) ⊂ S (CP ) . Again [6,7], we found superior performance
of quantum strategies as compared to classical strategies.
The nature of a game is determined by the rules of the game. In particular, the
appropriate solution concept depends on the available strategic moves. Obviously,
a player cannot make a meaningful choice without knowing the options at his or
her disposal. So it comes to no surprise that also the actual achievable pay-off in
such a game depends on the set of allowed strategies. Roughly speaking, one can
11
say that the possibility of utilising strategies which are not feasible in the analogous
classical game implicates a significant advantage. In the models studied in detail
two kinds of ”dilemmas” were ”resolved”: (i) On the one hand there are quantum games with an efficient unambiguous solution, while in the classical analogue
only an inefficient equilibrium can be identified. By taking advantage of appropriate quantum strategies much more efficient equilibria could be reached. In certain
sets of strategies even a maximally efficient solution – the Pareto optimum – was attainable. (ii) On the other hand, there exist quantum games with a unique solution
with a classical equivalent which offers two Nash equilibria of the same quality.
This paper deals with simple set-ups in which information is exchanged
quantum-mechanically. The emphasis was to examine how situations where
strategies are identified with quantum operations applied on quantum mechanical carriers of information are different from the classical equivalent. It is the hope
that these investigations may enable us to better understand competitive structures
in a game theoretical sense in applications of quantum information theory.
V. ACKNOWLEDGEMENTS
We would like to thank Maciej Lewenstein, Onay Urfalıoglu, Joel Sobel, Tom
Cover, Charles H. Bennett, Martin B. Plenio, and Uta Simon for helpful suggestions. We also acknowledge fruitful discussions with the participants of the A2
Consortial Meeting. This work was supported by the European Union and the
DFG.
[1] J. von Neumann and O. Morgenstern, The Theory of Games and Economic Behaviour
(Princeton University Press, Princeton, 1947).
[2] R.B. Myerson, Game Theory: An Analysis of Conflict (MIT Press, Cambridge, 1991).
[3] W. Poundstone, Prisoners’ Dilemma. John von Neumann, Game Theory, and the Puzzle of the
Bomb (Doubleday, New York, 1992).
[4] M.D. Davis, Game Theory. A Nontechnical Introduction (Dover, New York, 1970).
[5] A.W. Tucker, unpublished (1950).
[6] J. Eisert, M. Wilkens, and M. Lewenstein, Phys. Rev. Lett. 83, 3077 (1999).
[7] D. Meyer, Phys. Rev. Lett. 82, 1052 (1999).
[8] P. Ball, Nature, Science Update 991021-3 (1999); G.P. Collins, Scientific American, January issue (2000).
[9] L. Goldenberg, L. Vaidman, and S. Wiesner, Phys. Rev. Lett. 82, 3356 (1999).
[10] M.B. Plenio and V. Vedral, Contemp. Phys. 39, 431 (1998).
[11] R.F. Werner, Phys. Rev. A 58, 1827 (1998).
[12] C.H. Bennett and G. Brassard, in Proceedings of the IEEE International Conference on Computers, Systems, and Signal Processing, Bangalore, India (IEEE, New York, 1984); C.H. Bennett, F. Bessette, G. Brassard, L. Salvail, and J. Smolin, J. Crypto. 5, 3 (1992).
[13] P.W. Shor, SIAM J. Comput. 26, 1484 (1997).
[14] A. Ekert, Phys. Rev. Lett. 67, 661 (1991).
[15] E.g., Alice may allow a coupling of the original quantum system to an auxiliary quantum system and let the two unitarily interact. After performing a projective measurement on the composite system she could eventually consider the original system again
by taking a partial trace with respect to the auxiliary part.
[16] The quantum strategies sA ⊗ 1 and 1 ⊗ sB are in the following identified with sA and
sB , respectively.
[17] From a game theoretical viewpoint, any positive numbers satisfying the symmetry conditions ACC = BCC , ADD = BDD , ACD = BDC , ADC = BCD and the inequalities
12
[18]
[19]
[20]
[21]
ADC > ACC > ADD > ACD and ACC ≥ (ACD + ADC )/2 define a (strong) Prisoners’
Dilemma.
P.M. Alberti and A. Uhlmann, Stochasticity and Partial Order: Doubly Stochastic Maps and
Unitary Mixing (VEB Deutscher Verlag der Wissenschaften, Berlin, 1982).
A.W. Marshall and I. Olkin, Inequalities: Theory of Majorisation and its Applications (Academic Press, New York, 1979);
T.C. Schelling, The Strategy of Conflict (Harvard University Press, Cambridge (MA),
1960).
In the classical game both players could also play C and D with probabilities p = 1/2
yielding the same expected pay-off of 2.25 for both players, but this pair of mixed
strategies would be no equilibrium solution, as any player could benefit from simply
choosing the dominant strategy D.
13