Team decision problems with classical and quantum signals.

We study team decision problems where communication is not possible, but coordination among team members can be realized via signals in a shared environment. We consider a variety of decision problems that differ in what team members know about one another's actions and knowledge. For each type of decision problem, we investigate how different assumptions on the available signals affect team performance. Specifically, we consider the cases of perfectly correlated, i.i.d., and exchangeable classical signals, as well as the case of quantum signals. We find that, whereas in perfect-recall trees (Kuhn 1950 Proc. Natl Acad. Sci. USA 36, 570-576; Kuhn 1953 In Contributions to the theory of games, vol. II (eds H Kuhn, A Tucker), pp. 193-216) no type of signal improves performance, in imperfect-recall trees quantum signals may bring an improvement. Isbell (Isbell 1957 In Contributions to the theory of games, vol. III (eds M Drescher, A Tucker, P Wolfe), pp. 79-96) proved that, in non-Kuhn trees, classical i.i.d. signals may improve performance. We show that further improvement may be possible by use of classical exchangeable or quantum signals. We include an example of the effect of quantum signals in the context of high-frequency trading.

Introduction

A team is a group of agents unified by common goals. Characteristic of team problems is that members of a team have access to different information depending on their local environments. Communication of this information among team members may or may not be possible, depending on economic and physical constraints. An example of the latter arises in high-frequency trading (see Pagnotta & Philippon [1] for a survey), where messaging across widely dispersed members of a team would be too slow to be useful. In this paper, we study scenarios where direct communication is indeed unavailable, but team members can use a shared global environment to achieve highly effective coordination. We undertake a systematic examination of how the informational properties of the environment interact with the informational structure of a decision problem to bring about changes in performance.

In the absence of communication, team problems become formally equivalent to one-agent decision problems with memory limitations. This equivalence was noted by Marschak & Radner [2] in their pioneering work on teams. We shall refer to these scenarios as team decision problems. Figure 1 is a simple example.

Figure 1.

A team decision problem.

In the scenario accompanying this tree, there is a task that requires the completion of two different steps in sequence. If the two steps are completed in the correct order, the pay-off is 4. If the second step is undertaken before the completion of the first, the pay-off is 0. If the first step is mistakenly repeated, the pay-off is 1. There are two people assigned to the task and each has to act without knowing if the other has already completed the first step. The tree of figure 1 captures this scenario. In particular, the two square nodes belong to the two team members, but the nodes are enclosed in an information set I to capture the idea that they do not know whether they are acting first or second.[1]

As a one-agent problem, this same scenario has been extensively discussed as the absent-minded driver's problem [4]. While the focus in that literature was on conceptual aspects of this scenario, our interest is more ‘engineering-like’. Specifically, we shall investigate how well a team can perform tasks such as the above one, as a function of the assumptions made on its shared environment.

A concrete example of how environmental information can affect performance in a decision problem of the type in figure 1 was offered by Isbell [5]. He showed that if the players have access to i.i.d. signals (pay-off-irrelevant chance moves), then they can do better relative to no signals. There are other possibilities. Members of a team might have access to exchangeable (not necessarily i.i.d.) signals. Will this make a difference—in particular, will it allow still better performance?

Another possibility is that the physical make-up of the environment matters. In other areas of information theory, it is well established that access to quantum rather than classical information resources has profound consequences for various tasks. One main distinction between classical and quantum signals is that they arise at different physical scales. Classical signals are encoded in the macroscopic state of some physical system—for example, in an electrical current or light beam (or even in smoke signals …). By contrast, quantum signals are encoded in the microscopic state of a system—for example, in the spin of an electron or of a photon. Most importantly, quantum signals can exhibit patterns of behaviour that are impossible with any choice of classical signals. In particular, quantum signals may be not only correlated but even entangled, where this term refers to exotic correlations that cannot arise in the classical case.[2]

While quantum signals permeate any physical environment, their controlled use as information resources has only recently become possible and implementable. One case in which quantum techniques have already entered the practical arena is quantum cryptography [7,8], where the security of communication is protected by the very laws of Nature; by contrast, analogous classical schemes do not offer similar guarantees. In computer science, there are important quantum algorithms that have been proved superior to classical algorithms [9-11]. Could it be that in the area of decision-making, as in cryptography and computing, the availability of quantum resources might lead to improved performance over what is possible in a classical environment? We will identify conditions under which this is indeed the case. This may not be of theoretical interest alone. We will come back later to the example of high-frequency trading, where access to quantum resources could have practical significance.

Team problems with classical signals have been studied by Lehrer et al. [12]. Their focus is on signals which are informative about the underlying (‘physical’) state, while our interest is in signals as coordinating devices when there is a fixed information structure concerning the underlying state. La Mura [13] provides an example of a team problem where quantum signals yield an improvement over classical signals. (We make use of this example later.) Brunner & Linden [14] go beyond quantum information resources to so-called no-signalling information resources (see §6 for more on quantum information and no signalling) and show that this can allow even further improvement over the classical regime in a game. Kargin [15] provides a necessary condition for quantum signals to yield no improvement in a specific family of team problems.

Results

Kuhn [16,17] introduced into decision theory the fundamental distinction between perfect and imperfect recall. Isbell [5] extended this classification further to include other trees with limited recall, which do not belong to the family of Kuhn trees. Those include decision problems in which, as in the example of figure 1, an information set may contain nodes which are met in sequence. We will call these Isbell trees. This classification is equally important in team decision problems, where it refers to the availability or unavailability of information about what other team members do or know. This three-way taxonomy of decision problems—perfect-recall Kuhn, imperfect-recall Kuhn, Isbell—is the one we will use.

We now add a framework for talking about the different kinds of signals to which members of a team might have access. In the simplest case, there is one signal per information set. But this is restrictive and does not fit well with cases where the different nodes in a given information set could be widely separated from one another in space or time. In such cases, it may be more appropriate to think of distinct but perfectly correlated signals operating at different nodes within the same information set. In fact, other assumptions on signals are possible that still preserve the indistinguishability of nodes in an information set. In particular, the signals could be i.i.d., but, more generally, we can require them to be exchangeable.[3] Figure 2 depicts an information set in some tree. Figure 3a is the simple case where one coin is tossed at this information set and the choices can be pegged on the outcome of the toss. Figure 3b depicts two coins, one per node, where the coins are exchangeable (which includes the case that they are i.i.d.).

Figure 2.

An information set.

Figure 3.

(a) One signal at an information set. (b) Two exchangeable signals at an information set.

Within information sets, there are some clear considerations of indistinguishability. We also need to consider what are the appropriate conditions to impose on signals across information sets. We will want to know how these conditions, too, affect the potential performance in a task. To uncover these capabilities, it becomes important to specify the physical embodiment of the signals that are available. In particular, what correlations across signals are possible fundamentally depends on whether the signal carrier obeys classical or quantum physical laws.

Table 1 shows, for each type of decision problem we consider, the effect on team performance of different assumptions about the type of signals available. We denote in green the baseline performance which can be achieved in all types of problem with no signals. Along a given row, higher performance is indicated by moving from green to orange to yellow to red (as in a heat map). Our results can be summarized as follows. For perfect-recall Kuhn trees, no type of signal brings any improvement over the baseline. For imperfect-recall Kuhn trees, no classical signal type (perfectly correlated, i.i.d. or exchangeable) can improve over the baseline, but quantum signals may do so. For Isbell trees, classical i.i.d. signals improve over the baseline, exchangeable signs improve further and quantum signals still further.

Table 1.

Summary of results.

We see from the table that, in situations where communication among members of a team would be helpful but is unavailable, signals can act as substitutes, at least in part. Another implication of our results is that decision-making is not invariant to the physical embodiment of the decision environment. In particular, we see that access to quantum signals may yield improvements over any choice of classical signals. We will make the last point concretely via an example later in the paper. We do not explore the case of access to super-quantum signals (cf. [14]) in this paper, although it would certainly be of interest to extend table 1 to this case.

Signal structures

Figure 4 depicts a team decision problem which begins with a chance move. This is represented by a circular node belonging to Nature, where the numbers in parentheses give the probabilities of Nature's move. One team member, Ann, moves at information set I1, and the other member, Bob, moves at I2. When Ann moves, she knows that Nature chose left. But, when Bob moves, he does not know whether Nature chose right, or Nature chose left and then Ann chose In. Thus, Ann may have information—that Nature chose left and she chose In—that Bob does not get. In terms of our three-way taxonomy, the team problem is a Kuhn tree with imperfect recall. The reader should refer to definitions A.1 (Kuhn tree) and A.4 (perfect recall) in appendix A to check this last statement.

Figure 4.

A team decision problem with a chance move.

The team's expected pay-offs are: from the pair of strategies (In, Left), from (In, Right), from (Out, Left) and from (Out, Right). The team's highest expected pay-off is therefore .

Figure 5 adds two signals, in the form of coin tosses, to figure 4. There is one coin toss at information set I1, with outcomes Heads1 and Tails1, and another coin toss at I2, with outcomes Heads2 and Tails2. Ann can make her choice at I1 contingent on the outcome of the coin toss at I1. Likewise, Bob can make his choice at I2 contingent on the outcome of the coin toss at I2.

Figure 5.

The problem with added signals.

It will be convenient to describe the probability structure of the coin tosses in the following way. Each possible path through a tree crosses certain information sets of the team members in a certain order. A signal structure associates to each sequence of information sets that arises in this fashion, a probability measure on the product space of the associated signals. In figure 5, the sequences of information sets that can arise are I1, I2 and I1I2. Figure 6 gives the general form of the associated probability measures. Here, α to θ are numbers between 0 and 1 satisfying α+β=1, γ+δ=1 and ϵ+ζ+η+θ=1.

Figure 6.

An associated signal structure.

Consider the following strategies for the team. At her information set, Ann chooses Out if her coin comes up Heads1, and In if her coin comes up Tails1. At his information set, Bob chooses Left if his coin comes up Heads2, and Right if his coin comes up Tails2. The team's expected pay-off is then: If α=0, δ=1 (so that γ=0) and η=1 (so that θ=0), then the team gets an expected pay-off of 2, which is greater than the best possible () without signals.

On a closer look, it is apparent that the information structure of figure 6 when α=0 and δ = η = 1 is conceptually unsatisfactory. The second coin comes up Heads (H2) almost surely if it is tossed after the first coin has been tossed, but comes up Tails (T2) almost surely if it is tossed when the first coin has not been tossed. If we allow this amount of information ‘flow’ between the two coins, it is not surprising that we can engineer an improvement over the case of no coins. Our interest is in situations where such direct communication is impossible. What is needed, then, is a condition that rules out such information flow or communication, without going so far as to rule out all correlation across signals. The next section presents a suitable condition.

Indistinguishability

We want a condition that applies across information sets and that is in line with the basic requirement that nodes within an information set are indistinguishable. Suppose, for a moment, that the outcomes of a signal at one information set could reveal, probabilistically at least, which other signals are activated elsewhere in the tree. Then, simply by observation of the signal, a team member at a given information set might learn something—probabilistically, say—about which node in the information set was reached. (This was the case in the example of figures 5 and 6. If Bob observes H2, then he knows almost surely he is at the middle node of I2. If he observes T2, he knows almost surely he is at the right-hand node of I2.) Here is the formal condition to rule out this possibility:

Indistinguishability condition. Consider two sequences of information sets and the two associated signal probability measures. The marginals of these two measures—with respect to common subsequences—must agree.

Let us see how this condition works in the signal structure of figure 6. Looking at the two sequences I1 and I1I2, with common subsequence I1, we see that the condition is that the probabilities of H1 must be equal: α=ϵ+ζ. Likewise, looking at the two sequences I2 and I1I2, with common subsequence I2, we see that the condition is that the probabilities of H2 must be equal: γ=ϵ+η. Of course, the first (resp. second) condition implies that the probabilities of T1 (resp. T2) are also equal.

Indistinguishability rules out the choice of parameters α=0 and δ=η=1 we had before in figure 6. (This choice contradicts γ=ϵ+η.) We can go further. Indistinguishability reduces the five free parameters in figure 6 to three, which we will take to be ϵ, ζ and η (we will still write θ=1−ϵ−ζ−η). The expected pay-off under the previous strategy can then be written as Since this is a convex combination of expected pay-offs to the team in the tree without signals, we see that no improvement in the team's (maximum) expected pay-off is now possible under signals. The same is easily seen to be true for any other strategy for the tree of figure 5.

Classicality

The example of the previous section might suggest that, provided the indistinguishability condition is satisfied, the addition of signals to a (Kuhn) decision problem can never result in an increase in the team's maximum expected pay-off. Table 1 tells us that this is false once quantum signals are allowed. We will see an example of this phenomenon in the next section. Before that, we will establish the correct baseline for signals to yield no improvement in Kuhn trees.

Fix a Kuhn tree, let I1,I2,… be the information sets for the DM and Ω,Ω,… be associated finite signal sets which we add.[4] Write Ω=Ω×Ω×⋯ .

Classicality condition. There is a probability measure μ on Ω such that, for each subsequence II⋯I of information sets that arises in the tree, the associated probability measure is given by:

Note that this condition is well defined since, in a Kuhn tree, each path through the tree crosses a given information set at most once. Classicality says that there is a joint probability space (Ω,μ) from which a given signal structure, of the kind we explored in the previous section, can be derived. It is immediate from the properties of marginals that:

Proposition 5.1

Classicality implies indistinguishability.

Next, let M,M,… be the sets of moves at the information sets I1,I2,…, respectively, and write M=M×M×⋯ .

Proposition 5.2

Fix a Kuhn tree. The highest expected pay-off a team can achieve with signals satisfying classicality is the same as that without signals.

Proof.

A strategy profile for the team in the underlying tree is an element m∈M. A strategy profile for the team in the extended tree with signals is a tuple of maps . Write f=f×f×⋯ . Also, write π(m) for the expected pay-off to the team in the underlying tree, when it chooses strategy profile m and we average over Nature. Then, the expected pay-off to the team in the extended tree, when it chooses strategy profile f (and we again average over Nature), is . That is, in the tree with signals, the expected pay-off to any particular strategy profile is a convex combination of expected pay-offs to strategy profiles in the underlying tree.

This argument applies when there is one signal per information set. Since, in a Kuhn tree, every path from the root to a terminal node passes through a given information set at most once, it immediately extends to the case of multiple signals per information set, whether these signals are perfectly correlated, i.i.d. or exchangeable. ▪

Throughout, we assume independence between Nature (in the underlying tree) and signals. Independence seems like the right assumption for our purpose. We do not want signals to give the team information it never had. When independence is violated, proposition 5.1 may fail. Appendix B provides an example of such a case.

In decision theory (and game theory), one normally takes for granted the existence of a joint probability space that yields whatever signals one has in mind. This makes sense in the classical physical world where every physical mechanism can be associated with an appropriate joint probability space. But it may fail in the quantum realm. It turns out that existence or non-existence of this joint space actually defines the classical-quantum divide [19,20]. This is the reason behind the naming of our classicality condition.

Quantum improvement

We now show:

Proposition 6.1

There is a Kuhn tree (with imperfect recall) in which the team can achieve a higher expected pay-off with quantum signals than with any classical signals.

The decision problem of figure 7 will suffice to establish this claim. It represents a situation in which two team members are imperfectly informed of Nature's initial move and must coordinate their actions. One team member operates at either information set I1 or I2, according to Nature's initial move. The other team member operates at either information set I3 or I4, again according to Nature's initial move. We suppose that information sets I1 and I2 are at a common physical location; likewise, information sets I3 and I4 are at a (different) common physical location. (Location plays no role in classical decision and game theory, but it will matter below when we bring in quantum information resources.) Assume that the pay-offs satisfy 0

Figure 7.

A decision problem allowing quantum improvement.

We first show that the team's expected pay-off with classical signals is at most 0. To see this, start without signals. Observe that the only way for the team to get the +m pay-off (with positive probability) is if it chooses Up2 at information I2 and Up4 at information set I4. But then, to avoid the −M pay-off on the right-hand side of the tree, it must choose Down3 at I3. Then, to avoid the upper −M pay-off on the left-hand side, it must choose Up1 at I1. Then, to avoid the lower −M pay-off on the left-hand side, it must choose Down4 at I4, not Up4 as we supposed. It follows that the +m pay-off cannot arise unless at least one −M pay-off also arises. Moreover, it will arise with the same probability. Since M>m, we have shown that the team's expected pay-off is at most 0. Now use proposition 5.2 to conclude that the team's highest possible expected pay-off in any extended tree with classical signals is also 0.

Next, consider the signal structure of figure 8. Here, and is the inverse of the Golden Ratio. One can check that our indistinguishability condition is satisfied (use the fact that Φ2+Φ=1). Now consider the following strategy profile for the team in the extended tree: (i) at I1, choose Up1 after H1 and Down1 after T1; (ii) at I2, choose Up2 after H2 and Down2 after T2; (iii) at I3, choose Up3 after H3 and Down3 after T3; (iv) at I4, choose Up4 after H4 and Down4 after T4. The team's expected pay-off from this strategy profile is . An improvement in the team's highest expected pay-off is achieved.

Figure 8.

A quantum signal structure.

By proposition 5.2, we know that the signal structure of figure 8 cannot be realized classically. (This can also be verified directly; see appendix C.) It can, however, be realized quantum mechanically [21]. The physical mechanism involves the creation of what is called an entangled pair of particles. The basic set-up is that two particles—two photons, for example—are prepared in a special state and sent off on different trajectories. Each particle then enters a detector, placed some distance from the source on that particle's trajectory. Detectors have various settings, and the setting chosen determines which property of a particle is measured. For example, a detector might be set to measure the so-called spin of a photon along a particular direction. The outcome of each measurement is binary and can take one of two values, conventionally labelled spin +1 or spin −1.

Such a quantum system can be used to generate the signal structure of figure 8. The spin of one particle is measured at information set I1 or I2. It is measured along one direction at I1 and along a different direction at I2. In either case, the measurement has two possible outcomes. We call them Heads1 or Tails1, and Heads2 or Tails2, respectively. The spin of the second particle is measured at information set I3 or I4. It is measured along one direction at I3 and along a different direction at I4. Again, in either case, the measurement has two possible outcomes. We call them Heads3 or Tails3, and Heads4 or Tails4, respectively. This gives us the form of the signal structure of figure 8. The specific probabilities come from the preparation of a particular entangled quantum state [21,22].

The choice of a tree with imperfect recall (figure 7) and of a signal structure with indistinguishability (figure 8) was quite deliberate. If a tree has perfect recall, then no signal—even quantum—can bring any improvement. This is not surprising since there is nothing for team members to learn about one another (see appendix A for a formal argument). As for indistinguishability, this is a necessary feature of any signal structure that is built using quantum information resources. This follows from an important property of quantum mechanics called ‘no signalling’ [23].

Isbell trees

We now examine a class of non-Kuhn trees first studied by Isbell [5]. We have already seen an example of an Isbell tree in figure 1. This will be the first example where it matters which formulation of classical signals we choose.

We first dispatch the case of perfectly correlated signals. We argue that any strategy profile using perfectly correlated signals cannot do better than a strategy that does not use any signals. Indeed, observe that a strategy profile based on perfectly correlated signals cannot prescribe different moves at two nodes in the same information set. But then, for each realization of the signal, the resulting common move can be replicated as part of a (deterministic) strategy profile that simply prescribes this common move at both nodes.

It is no longer true that classical signals have no effect in Isbell trees, once we move from perfectly correlated to i.i.d. signals. We review Isbell's [5] argument. We go back to the tree of figure 1 and extend it by adding two coins at the information set I. The extended tree is depicted in figure 9. The coin which is tossed at the root of the tree comes up Heads1 or Tails1, and the coin tossed at the subsequent node comes up Heads2 or Tails2. The team has two information sets in the extended tree—one where team members see a coin land Heads, and one where they see a coin land Tails. The three nodes in the first information set are shaded with the right-side-up triangles, and the three nodes in the second information set with upside-down triangles.

Figure 9.

A non-Kuhn tree with added signals.

In this tree, with either no signals or perfectly correlated signals, the team's best expected pay-off is 1 (from choosing In). Now let the signals be i.i.d., as in figure 10a, and suppose team members adopt the strategy of choosing In at the first information set in figure 9 (right-side-up triangles) and Out at the second information set (upside-down triangles). Then, the team's expected pay-off is . This effect of i.i.d. signals was first noted by Isbell [5]. Under exchangeability, the team can do even better. For the signal distribution of figure 10b, its expected pay-off is .

Figure 10.

(a,b) Two associated signal structures.

Cabello & Calsamiglia [24] also studied the game of figure 1 and showed that the availability of quantum signals there allows a team to achieve an expected pay-off of 2. We have just seen that one does not need to resort to quantum signals in this tree to obtain this improvement. However, one can easily build other Isbell trees where quantum signals can improve still further on classical signals. For example, we could simply glue together the tree of figure 7 (where quantum signals improve on classical signals) and the tree of figure 1 (where i.i.d. or exchangeable classical signals improve on no signals). This would yield an Isbell tree where quantum signals improve on all classical signals.

An economic application

A natural scenario in which team problems arise but direct communication is impossible is high-frequency financial trading. As a concrete example of communication limitations in this setting, consider two markets located in New York and Shanghai, respectively. Typically, a new trade is accepted every 0.5 ms by the stock exchange servers. Even at the speed of light, communication between the two locations takes approximately 40 ms. This makes classical arbitrage impossible, since any information about prices on one exchange is already out of date once it reaches the other exchange [25].

Now consider a team problem involving two markets (1 and 2) and two traders (Ann and Bob) engaged in local high-frequency strategies. We assume that the traders are located at a significant distance from each other and from the two markets. The distances are such that communication prior to their trading decisions is too slow. We will show how access to quantum signals can enable the two traders to improve their joint performance relative to any classical signals. The mechanism is based on a well-studied quantum set-up going back to Bell [26] and discussed as a team decision problem in La Mura [13].

There are three assets X, Y and Z, and, at each point in time, each trader needs to sell one of the three assets (chosen with equal probability) against the other two. When Ann and Bob want to sell the same asset, they do better trading on separate markets, in which case they get pay-offs of 0 (a normalization), rather than on the same market, where they would directly compete against each other and get pay-offs of −M. When Ann and Bob want to sell different assets, they do better trading on the same market since each increases the demand for the asset the other wants to sell. This yields both a pay-off of +m, as compared with 0 if they trade on separate markets. We assume m

If M is sufficiently large compared with m, any good pair of strategies for the traders must preclude their selling the same asset on the same market. In fact, the following is optimal for the traders. Ann goes to market 2 only when she needs to sell asset Z, while Bob goes to market 1 only when he needs to sell Z. (By symmetry, we can replace Z by X or Y , and market 2 with market 1.) To calculate the resulting expected pay-off to the traders, note that there are nine equally likely cases according to whether Ann wants to sell asset X, Y or Z, and similarly for Bob. In four of these cases, the above strategy profile secures a pay-off of +m, and otherwise 0. So, the expected pay-off is . This is under the assumption of no signals, but, since the scenario corresponds to a Kuhn tree (with imperfect recall), we know from proposition 5.2 that the addition of classical signals cannot improve the baseline pay-off.

Now bring in quantum information resources. We give the traders access to an entangled quantum system on which they can make certain measurements and thereby condition their choices. Specifically, we assume that the system consists of two particles, one per trader. The particle pair is prepared in a so-called Bell state [26], which gives rise to the signal structure of figure 11.

Figure 11.

(a,b) Quantum signals for the trading example.

Each trader chooses one of three possible local measurements on the system, which, for convenience, we also label X, Y or Z. The table in figure 11a gives the probabilities of the joint outcomes (each outcome can be Up or Down) when the traders make the same choice of measurement, and the table in figure 11b gives the probabilities when they make different choices of measurement. Consider the following strategy for Ann. If she wants to sell X, then she performs measurement X and, if she observes Up, she executes the trade on market 1, while if she observes Down, she executes the trade on market 2. Similarly, if Ann observes Y or Z, she performs the corresponding measurement and acts accordingly. Bob adopts the same strategy. The expected pay-off is calculated as follows. These strategies always avoid selling the same asset on the same market. Moreover, in each of the six cases where the traders want to sell different assets, they manage, with probability , to trade on the same market. This leads to an expected pay-off of , which is greater than the baseline pay-off of .

Table 2.

Joint probability space for a signal structure.