1 Introduction
As artificial agents proliferate, it is increasingly important to analyze, predict and control their collective behavior (Parkes and Wellman, 2015; Rahwan et al., 2019). Unfortunately, despite almost a century of intense research since von Neumann (1928), game theory provides little guidance outside a few special cases such as twoplayer zerosum, auctions, and potential games (von Neumann and Morgenstern, 1944; Vickrey, 1961; Monderer and Shapley, 1996; Nisan et al., 2007). Nash equilibria provide a general solution concept, but are intractable in almost all cases for many different reasons (Hart and MasColell, 2003; Daskalakis et al., 2009; Babichenko, 2016). These and other negative results (Palaiopanos et al., 2017) suggest that understanding and controlling societies of artificial agents is near hopeless. Nevertheless, human societies – of billions of agents – manage to organize themselves reasonably well and mostly progress with time, suggesting game theory is missing some fundamental organizing principles.
In this paper, we investigate how markets structure the behavior of agents. Market mechanisms have been studied extensively (Nisan et al., 2007). However, prior work has restricted to concrete examples, such as auctions and prediction markets, and strong assumptions, such as convexity. Our approach is more abstract and more directly suited to modern machine learning where the building blocks are neural nets. Markets, for us, encompass discriminators and generators trading errors in GANs (Goodfellow et al., 2014) and agents trading wins and losses in StarCraft (Vinyals et al., 2019).
1.1 Overview
The paper introduces a class of games where optimization and aggregation make sense. The phrase requires unpacking. “Optimization” means gradientbased methods. Gradient descent (and friends) are the workhorse of modern machine learning. Even when gradients are not available, gradient estimates
underpin many reinforcement learning and evolutionary algorithms. “Aggregation” means weighted sums. Sums and averages are the workhorses for analyzing ensembles and populations across many fields. “Makes sense” means we can draw conclusions about the gradientbased dynamics of the collective by summing over properties of its members.
As motivation, we present some pathologies that arise in even the simplest smooth games. Examples in section 2 show that coupling strongly concave profit functions to form a game can lead to uncontrolled behavior, such as spiraling to infinity and excessive sensitivity to learning rates. Hence, one of our goals is to understand how to ‘glue together agents’ such that their collective behavior is predictable.
Section 3 introduces a class of games where simultaneous gradient ascent behaves well and is amenable to analysis. In a smooth market (SMgame), each player’s profit is composed of a personal objective and pairwise zerosum interactions with other players. Zerosum interactions are analogous to monetary exchange (my expenditure is your revenue), doubleentry bookkeeping (credits balance debits), and conservation of energy (actions cause equal and opposite reactions). SMgames explicitly account for externalities. Remarkably, building this simple bookkeeping mechanism into games has strong implications for the dynamics of gradientbased learners. SMgames generalize adversarial games (Cai et al., 2016) and codify a common design pattern in machine learning, see section 3.1.
Section 4 studies SMgames from two points of view. Firstly, from that of a rational, profitmaximizing agent that makes decisions based on firstorder profit forecasts. Secondly, from that of the game as a whole. SMgames are not potential games, so the game does not optimize any single function. A collective of profitmaximizing agents is not rational because they do not optimize a shared objective (Drexler, 2019). We therefore introduce the notion of legibility, which quantifies how the dynamics of the collective relate to that of individual agents.
Finally, section 5 applies legibility to prove some basic theorems on the dynamics of SMgames under gradientascent. We show that (i) Nash equilibria are stable; (ii) that if profits are strictly concave then gradient ascent converges to a Nash equilibrium for all learning rates; and (iii) the dynamics are bounded under reasonable assumptions.
The results are important for two reasons. Firstly, we identify a class of games whose dynamics are, at least in some respects, amenable to analysis and control. The kinds of pathologies described in section 2 cannot arise in SMgames. Secondly, we identify the specific quantities, forecasts, that are useful to track at the level of individual firms and can be meaningfully aggregated to draw conclusions about their global dynamics. It follows that forecasts should be a useful lever for mechanism design.
1.2 Related work
A wide variety of machine learning markets and agentbased economies have been proposed and studied: Selfridge (1958); Barto et al. (1983); Minsky (1986); Wellman and Wurman (1998); Baum (1999); Kwee et al. (2001); Kearns et al. (2001); Kakade et al. (2003, 2005); Lay and Barbu (2010); Sutton et al. (2011); Abernethy and Frongillo (2011); Storkey (2011); Storkey et al. (2012); Hu and Storkey (2014); Balduzzi (2014). The goal of this paper is different. Rather than propose another market mechanism, we abstract an existing design pattern and elucidate some of its consequences for interacting agents.
Our approach draws on work studying convergence in generative adversarial networks (Mescheder et al., 2017; Mescheder, 2018; Balduzzi et al., 2018; Gemp and Mahadevan, 2018; Gidel et al., 2019), related minimax problems (Bailey and Piliouras, 2018; Abernethy et al., 2019), and monotone games (Nemirovski et al., 2010; Gemp and Mahadevan, 2017; Tatarenko and Kamgarpour, 2019).
1.3 Caveat
We consider dynamics in continuous time in this paper. Discrete dynamics, require a more delicate analysis, e.g. Bailey et al. (2019). In particular, we do not claim that optimizing GANs and SMgames is easy in discrete time. Rather, our analyis shows that it is relatively easy in continuous time, and therefore possible in discrete time, with some additional effort. The contrast is with smooth games in general, where gradientbased methods have essentially no hope of finding local Nash equilibria even in continuous time.
1.4 Notation
Vectors are columnvectors. The notations and refer to a positivedefinite matrix and vector with all entries positive respectively. Rather than losses, we work with profits. Proofs are in the appendix. We use economic terminology (firms, profits, forecasts, and sentiment) even though the examples of SMgames, such as GANs and adversarial training, are taken from mainstream machine learning. We hope the economic terminology provides an invigorating change of perspective. The underlying mathematics is no more than first and secondorder derivatives.
2 Smooth games
Smooth games model interacting agents with differentiable objectives. They are the kind of games that are played by neural nets. In practice, the differentiability assumption can be relaxed by replacing gradients with gradient estimates.
Definition 1.
A smooth game (Letcher et al., 2019) consists in players , equipped with twice continuously differentiable profit functions . The parameters are with where . Player controls the parameters .
If players update their actions via simultaneous gradient ascent, then a smooth game yields a dynamical system specified by the differential equation for
(1) 
where is a vector. The Jacobian of a game is the matrix of secondderivatives . The setup can be recast in terms of minimizing losses by substituting for all .
Smooth games are too general to be tractable since they encompass all dynamical systems.
Lemma 1.
Every continuous dynamical system on , for any , arises as simultaneous gradient ascent on the profit functions of a smooth game.
The next two sections illustrate some problems that arise in simple smooth games.
Definition 2.
We recall some solution concepts from dynamical systems and game theory:

A stable fixed point^{1}^{1}1Berard et al. (2019) use a different notion of stable fixed point that requires
has positive eigenvalues.
satisfies and for all vectors . 
A local Nash equilibrium has neighborhoods of for all , such that all .

A classical Nash equilibrium satisfies for all and all players .
Example 1 below shows that stable fixed points and local Nash equilibria do not necessarily coincide. The notion of classical Nash equilibrium is illsuited to nonconcave settings.
Intuitively, a fixed point is stable if all trajectories sufficiently nearby flow into it. A joint strategy is a local Nash if each player is harmed if it makes a small unilateral deviation. Local Nash differs from the classic definition in two ways. It is weaker, because it only allows small
unilateral deviations. This is necessary since players are neural networks and profits are not usually concave. It is also stronger, because unilateral deviations decrease (rather than
not increase) profits.2.1 Problems with potential games
A game is a potential game if for some function , see Balduzzi et al. (2018) for details.
Example 1 (potential game).
Fix a small . Consider the twoplayer games with profit functions
(2) 
The game has a unique local Nash equilibrium at with .
The game is chosen to be as nice as possible: and are strongly concave functions of and respectively. The game is a potential game since for . Nevertheless, the game exhibits three related problems.
Firstly, the Nash equilibrium is unstable. Players at the Nash equilibrium can increase their profits via the joint update , so . The existence of a Nash equilibrium where players can improve their payoffs by coordinated action suggests the incentives are not welldesigned.
Secondly, the dynamics can diverge to infinity. Starting at and applying simultaneously gradient ascent causes the norm of vector to increase without limit as – and at an accelerating rate – due to a positive feedback loop between the players’ parameters and profits. Finally, players impose externalities on each other. The decisions of the first player affect the profits of the second, and vice versa. Obviously players must interact for a game to be interesting. However, positive feedback loops arise because the interactions are not properly accounted for.
In short, simultaneous gradient ascent does not converge to the Nash – and can diverge to infinity. It is open to debate whether the fault lies with gradients, the concept of Nash, or the game structure. In this paper, we take gradients and Nash equilibria as given and seek to design better games.
2.2 Problems with learning rates
Gradientbased optimizers rarely follow the actual gradient. For example RMSProp and Adam use adaptive, parameterdependent learning rates. This is not a problem when optimizing a function. Suppose
is optimized with reweighted gradient where is a vector of learning rates. Even though is not necessarily the gradient of any function, it behaves like because they have positive inner product when :(3) 
Parameterdependent learning rates thus behave well in potential games where the dynamics derive from an implicit potential function . Severe problems can arise in general games.
Example 2 (“half a game”).
Consider the following game, where the player is indifferent to :
(4) 
The dynamics are clear by inspection: the player converges to , and then the player does the same. It is hard to imagine that anything could go wrong. In contrast, behavior in the next example should be worse because convergence is slowed down by cycling around the Nash:
Example 3 (minimal SMgame).
A simple SMgame, see definition 3, is
(5) 
Figure 1 shows the dynamics of the games, in discrete time, with small learning rates and small gradient noise. In the top panel, both players have the same learning rate. Both games converge. Example 2 converges faster – as expected – without cycling around the Nash.
In the bottom panels, the learning rate of the second player is decreased by a factor of eight. The SMgame’s dynamics do not change significantly. In contrast, the dynamics of example 2 become unstable: although player 1 is attracted to the Nash, it is extremely sensitive to noise and does not stay there for long. One goal of the paper is to explain why SMgames are more robust, in general, to differences in relative learning rates.
2.3 Stop_gradient and learning rates
Tools for automatic differentiation (AD) such as TensorFlow and PyTorch include
stop_gradient operators that stop gradients from being computed. For example, let . The use of stop_gradient means is not strictly speaking a function and so we use to refer to its gradient under automatic differentiation. Then(6) 
which is the simultaneous gradient from example 2. Any smooth vector field is the gradient of a function augmented with stop_gradient operators, see appendix D. Stop_gradient is often used in complex neural architectures (for example when one neural network is fed into another leading to multiplicative interactions), and is thought to be mostly harmless. Section 2.2 shows that stop_gradients can interact in unexpected ways with parameterdependent learning rates.
2.4 Summary
It is natural to expect individually wellbehaved agents to also behave well collectively. Unfortunately, this basic requirement fails in even the simplest examples.
Maximizing a strongly concave function is wellbehaved: there is a unique, finite global maximum. However, example 1 shows that coupling concave functions can cause simultaneous gradient ascent to diverge to infinity. The dynamics of the game differs in kind from the dynamics of the players in isolation. Example 2 shows that reducing the learning rate of a wellbehaved (strongly concave) player in a simple game destabilizes the dynamics. How collectives behave is sensitive not only to profits, but also to relative learning rates. Offtheshelf optimizers such as Adam (Kingma and Ba, 2015) modify learning rates under the hood, which may destabilize some games.
3 Smooth Markets (SMgames)
Let us restrict to more structured games. Take an accountant’s view of the world, where the only thing we track is the flow of money. Interactions are pairwise. Money is neither created nor destroyed, so interactions are zerosum. If we model the interactions between players by differentiable functions that depend on their respective strategies then we have an SMgame. All interactions are explicitly tracked. There are no externalities off the books. Positive interactions, , are revenue, negative are costs, and the difference is profit. The model prescribes that all firms are profit maximizers. More formally:
Definition 3 (SMgame).
A smooth market is a smooth game where interactions between players are pairwise zerosum. The profits have the form
(7) 
where for all .
The functions can act as regularizers. Alternatively, they can be interpreted as natural resources or dummy players that react too slowly to model as players. Dummy players provide firms with easy (nonadversarial) sources of revenue.
Humans, unlike firms, are not profitmaximizers; humans typically buy goods because they value them more than the money they spend on them. Appendix C briefly discusses extending the model.
3.1 Examples of SMgames
SMgames codify a common design pattern:

Optimizing a function. A neartrivial case is where there is a single player with profit .

Zerosum polymatrix games are SMgames where and for some matrices
. Weights are constrained to probability simplices. The games have nice properties including: Nash equilibria are computed via a linear program and correlated equilibria marginalize onto Nash equilibria
(Cai et al., 2016). 
Intrinsic curiosity modules use games to drive exploration. One module is rewarded for predicting the environment and an adversary is rewarded for choosing actions whose outcomes are not predicted by the first module (Pathak et al., 2017). The modules share some weights, so the setup is nearly, but not exactly, an SMgame.

Tasksuites where a population of agents are trained on a population of tasks, form a bipartite graph. If the tasks are parametrized and adversarially rewarded based on their difficulty for agents, then the setup is an SMgame.
Monetary exchanges in SMgames are quite general. The error signals traded between generators and discriminators and the wins and losses traded between agents in StarCraft are two very different special cases.
4 From Micro to Macro
How to analyze the behavior of the market as a whole? Adam Smith claimed that profitmaximizing leads firms to promote the interests of society, as if by an invisible hand (Smith, 1776). More formally, we can ask: Is there a measure that firms collectively increase or decrease? It is easy to see that firms do not collectively maximize aggregate profit (AP) or aggregate revenue (AR):
(9) 
Maximizing aggregate profit would require firms to ignore interactions with other firms. Maximizing aggregate revenue would require firms to ignore costs. In short, SMgames are not potential games; there is no function that they optimize in general. However, it turns out the dynamics of SMgames aggregates the dynamics of individual firms, in a sense made precise in section 4.3.
4.1 Rationality: Seeing like a Firm
Give an objective function to an agent. The agent is rational, relative to the objective, if it chooses actions because it forecasts they will lead to better outcomes as measured by the objective. In SMgames, agents are firms, the objective is profit, and forecasts are computed using gradients.
Firms aim to increase their profit. Applying the firstorder Taylor approximation obtains
(10) 
where refers to higherorder terms. Firm ’s forecast of how profits will change if it modifies production by is . The Taylor expansion implies that for small updates . Forecasts encode how individual firms expect profits to change ceteris paribus^{2}^{2}2All else being equal – i.e. without taking into account updates by other firms..
4.2 Profit changes do not add up
How does profit maximizing by individual firms look from the point of view of the market as a whole? Summing over all firms obtains
(11) 
where is the aggregate forecast. Unfortunately, the lefthand side of Eq. (11) is incoherent. It sums the changes in profit that would be experienced by firms updating their production in isolation. However, firms change their production simultaneously. Updates are not ceteris paribus and so profit is not a meaningful macroeconomic concept. The following minimal example illustrates the problem:
Example 4.
Suppose and . Fix and let . The sum of the changes in profit expected by the firms, reasoning in isolation, is
(12) 
whereas the actual change in aggregate profit is zero because for any .
Tracking aggregate profits is therefore not useful. The next section shows forecasts are better behaved.
4.3 Legibility: Seeing like an Economy
Give a target function to every agent in a collective. The collective is legible, relative to the targets, if it increases or decreases the aggregate target according to whether its members forecast, on aggregate, they will increase or decrease their targets. We show that SMgames are legible. The targets are profit forecasts (note: not profits).
Let us consider how forecasts change. Define the sentiment as the directional derivative of the forecast . The firstorder Taylor expansion of the forecast shows that the sentiment is a forecast about the profit forecast:
(13) 
The perspective of firms can be summarized as:

Choose an update direction that is forecast to increase profit.

The firm is then in one of two main regimes:

If sentiment is positive then forecasts increase as the firm modifies its production – forecasts become more optimistic. The firm experiences increasing returnstoscale.

If sentiment is negative then forecasts decrease as the firm modifies its production – forecasts become more pessimistic. The firm experiences diminishing returnstoscale.

Our main result is that sentiment is additive, which means that forecasts are legible:
Proposition 2 (forecasts are legible in SMgames).
Sentiment is additive
(14) 
Thus, the aggregrate profit forecast increases or decreases according to whether individual forecasts are expected to increase or decrease in aggregate.
Section 5.1 works through an example that is not legible.
5 Dynamics of Smooth Markets
Finally, we study the dynamics of gradientbased learners in SMgames. Suppose firms use gradient ascent. Firm ’s updates are, infinitesimally, in the direction so that . Since updates are gradients, we can simplify our notation. Define firm ’s forecast as and its sentiment, ceteris paribus, as .
We allow firms to choose their learning rates; firms with higher learning rates are more responsive. Define the weighted dynamics and weighted forecast as
(15) 
In this setting, proposition 2 implies that
Proposition 3 (legibility under gradient dynamics).
Fix dynamics . Sentiment decomposes additively:
(16) 
Thus, we can read off the aggregate dynamics from the dynamics of forecasts of individual firms.
5.1 Example of a failure of legibility
The pairwise zerosum structure is crucial to legibility. It is instructive to take a closer look at example 1, where the forecasts are not legible.
Suppose and . Then and the firms’ sentiments are and which are always nonpositive. However, the aggregate sentiment is
(17) 
which for small is dominated by , and so can be either positive or negative.
When we have and . Each firm expects their forecasts to decrease, and yet the opposite happens due to a positive feedback loop that ultimately causes the dynamics to diverge to infinity.
5.2 Stability, Convergence and Boundedness
We provide three fundamental results on the dynamics of smooth markets. Firstly, we show that stability, from dynamical systems, and local Nash equilibrium, from game theory, coincide in SMgames:
Theorem 4 (stability).
A fixed point in an SMgame is a local Nash equilibrium iff it is stable. Thus, every local Nash equilibrium is contained in an open set that forms its basin of attraction.
Secondly, we consider convergence. Lyapunov functions are tools for studying convergence. Given dynamical system with fixed point , recall that is a Lyapunov function if: (i) ; (ii) for all ; and (iii) for all . If a dynamical system has a Lyapunov function then the dynamics converge to the fixed point. Aggregate forecasts share properties (i) and (ii) with Lyapunov functions.

Shared global minima: iff for all , which occurs iff is a stationary point, for all .

Positivity: for all points that are not fixed points, for all .
We can therefore use forecasts to study convergence and divergence across all learning rates:
Theorem 5.
In continuous time, for all positive learning rates ,

Convergence: If is a stable fixed point (), then there is an open neighborhood where for all , so the dynamics converge to from anywhere in .

Divergence: If is an unstable fixed point (), there is an open neighborhood such that for all , so the dynamics within are repelled by .
The theorem explains why SMgames are robust to relative differences in learning rates – in contrast to the sensitivity exhibited by the game in example 2. If a fixed point is stable, then for any dynamics , there is a corresponding aggregate forecast that can be used to show convergence. The aggregate forecasts provide a family of Lyapunovlike functions.
Finally, we consider the setting where firms experience diminishing returnstoscale for sufficiently large production vectors. The assumption is realistic for firms in a finite economy since revenues must eventually saturate whilst costs continue to increase with production.
Theorem 6 (boundedness).
Suppose all firms have negative sentiment for sufficiently large values of . Then the dynamics are bounded for all .
The theorem implies that the kind of positive feedback loops that caused example 1 to diverge to infinity, cannot occur in SMgames.
5.3 Legibility and the landscape
One of our themes is that legibility allows to read off the dynamics of games. We make the claim visually explicit in this section. Let us start with a concrete game.
Example 5.
Consider the SMgame with profits
(18) 
Figure 3AB plots the dynamics of the SMgame in example 5, under two different learning rates for player 1. There is an unstable fixed point at the origin and an ovoidal cycle. Dynamics converge to the cycle from both inside and outside the ovoid. Changing player 1’s learning rate, panel B, squashes the ovoid. Panels CD provide a cartoon map of the dynamics. There are two regions, the interior and exterior of the ovoid and the boundary formed by the ovoid itself.
In general, the phase space of any SMgame is carved into regions where sentiment is positive and negative, with boundaries where sentiment is zero. The dynamics can be visualized as operating on a landscape where height at each point corresponds to the value of the aggregate forecast . The dynamics does not always ascend or always descend the landscape. Rather, sentiment determines whether the dynamics ascends, descends, or remains on a levelset. Since sentiment is additive, , the decision to ascend or descend comes down to a weighted sum of the sentiments of the firms.^{3}^{3}3Note: the dynamics do not necessarily follow the gradient of . Rather, they move in directions with positive or negative inner product with according to sentiment. Changing learning rates changes the emphasis given to different firms’ opinions, and thus changes the shapes of the boundaries between regions in a relatively straightforward manner.
SMgames can thus express richer dynamics than potential games (cycles will not occur when performing gradient ascent on a fixed objective), which still admit a relatively simple visual description in terms of a landscape and decisions about which direction to go (upwards or downwards). Computing the landscape for general SMgames, as for neural nets, is intractable.
6 Discussion
Machine learning has got a lot of mileage out of treating differentiable modules like plugandplay lego blocks. This works when the modules optimize a single loss and the gradients chain together seamlessly. Unfortunately, agents with differing objectives are far from plugandplay. Interacting agents form games, and games are intractable in general. Worse, positive feedback loops can cause individually wellbehaved agents to collectively spiral out of control.
It is therefore necessary to find organizing principles – constraints – on how agents interact that ensure their collective behavior is amenable to analysis and control. The pairwise zerosum condition that underpins SMgames is one such organizing principle, which happens to admit an economic interpretation. Our main result is that SMgames are legible: changes in aggregate forecasts are the sum of how individual firms expect their forecasts to change. It follows that we can translate properties of the individual firms into guarantees on collective convergence, stability and boundedness in SMgames, see theorems 46.
Legibility is a localtoglobal principle, whereby we can draw qualitative conclusions about the behavior of collectives based on the nature of their individual members. Identifying and exploiting games that embed localtoglobal principles will become increasingly important as artificial agents become more common.
References
 Abernethy and Frongillo (2011) Abernethy, J. and Frongillo, R. (2011). A Collaborative Mechanism for Crowdsourcing Prediction Problems. In NeurIPS.
 Abernethy et al. (2019) Abernethy, J., Lai, K. A., and Wibisono, A. (2019). Lastiterate convergence rates for minmax optimization. In arXiv:1906.02027.
 Babichenko (2016) Babichenko, Y. (2016). Query Complexity of Approximate Nash Equilibria. Journal ACM, 63(4).
 Bailey et al. (2019) Bailey, J. P., Gidel, G., and Piliouras, G. (2019). Finite Regret and Cycles with Fixed StepSize via Alternating Gradient DescentAscent. In arXiv:1907.04392.
 Bailey and Piliouras (2018) Bailey, J. P. and Piliouras, G. (2018). Multiplicative Weights Update in ZeroSum Games. In ACM EC.
 Balduzzi (2014) Balduzzi, D. (2014). Cortical prediction markets. In AAMAS.
 Balduzzi et al. (2018) Balduzzi, D., Racanière, S., Martens, J., Foerster, J., Tuyls, K., and Graepel, T. (2018). The mechanics of player differentiable games. In ICML.
 Barto et al. (1983) Barto, A. G., Sutton, R. S., and Anderson, C. W. (1983). Neuronlike Adapative Elements That Can Solve Difficult Learning Control Problems. IEEE Trans. Systems, Man, Cyb, 13(5):834–846.
 Baum (1999) Baum, E. B. (1999). Toward a Model of Intelligence as an Economy of Agents. Machine Learning, 35(155185).
 Berard et al. (2019) Berard, H., Gidel, G., Almahairi, A., Vincent, P., and LacosteJulien, S. (2019). A Closer Look at the Optimization Landscapes of Generative Adversarial Networks. In arXiv:1906.04848.
 Cai et al. (2016) Cai, Y., Candogan, O., Daskalakis, C., and Papadimitriou, C. (2016). Zerosum Polymatrix Games: A Generalization of Minmax. Mathematics of Operations Research, 41(2):648–655.
 Daskalakis et al. (2009) Daskalakis, C., Goldberg, P. W., and Papadimitriou, C. (2009). The Complexity of Computing a Nash Equilibrium. SIAM J. Computing, 39(1):195–259.
 Drexler (2019) Drexler, K. E. (2019). Reframing Superintelligence: Comprehensive AI Services as General Intelligence. Future of Humanity Institute, University of Oxford, Technical Report #20191.
 Gemp and Mahadevan (2017) Gemp, I. and Mahadevan, S. (2017). Online Monotone Games. In arXiv:1710.07328.
 Gemp and Mahadevan (2018) Gemp, I. and Mahadevan, S. (2018). Global Convergence to the Equilibrium of GANs using Variational Inequalities. In arXiv:1808.01531.
 Gidel et al. (2019) Gidel, G., Hemmat, R. A., Pezeshki, M., Lepriol, R., Huang, G., LacosteJulien, S., and Mitliagkas, I. (2019). Negative Momentum for Improved Game Dynamics. In AISTATS.
 Goodfellow et al. (2014) Goodfellow, I. J., PougetAbadie, J., Mirza, M., Xu, B., WardeFarley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative Adversarial Nets. In NeurIPS.
 Hart and MasColell (2003) Hart, S. and MasColell, A. (2003). Uncoupled Dynamics Do Not Lead to Nash Equilibrium. American Economic Review, 93(5):1830–1836.
 Hu and Storkey (2014) Hu, J. and Storkey, A. (2014). Multiperiod Trading Prediction Markets with Connections to Machine Learning. In ICML.
 Kakade et al. (2003) Kakade, S., Kearns, M., and Ortiz, L. (2003). Graphical economics. In COLT.
 Kakade et al. (2005) Kakade, S., Kearns, M., Ortiz, L., Pemantle, R., and Suri, S. (2005). Economic properties of social networks. In NeurIPS.
 Kearns et al. (2001) Kearns, M., Littman, M., and Singh, S. (2001). Graphical models for game theory. In UAI.
 Kingma and Ba (2015) Kingma, D. P. and Ba, J. L. (2015). Adam: A method for stochastic optimization. In ICLR.
 Kurakin et al. (2017) Kurakin, A., Goodfellow, I., and Bengio, S. (2017). Adversarial Machine Learning at Scale. In ICLR.
 Kwee et al. (2001) Kwee, I., Hutter, M., and Schmidhuber, J. (2001). Marketbased reinforcement learning in partially observable worlds. In ICANN.

Lay and Barbu (2010)
Lay, N. and Barbu, A. (2010).
Supervised aggregation of classifiers using artificial prediction markets.
In ICML.  Letcher et al. (2019) Letcher, A., Balduzzi, D., Racanière, S., Martens, J., Foerster, J., Tuyls, K., and Graepel, T. (2019). Differentiable Game Mechanics. JMLR, 20:1–40.

Madry et al. (2018)
Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. (2018).
Towards Deep Learning Models Resistant to Adversarial Attacks.
In ICLR.  Mescheder (2018) Mescheder, L. (2018). On the convergence properties of GAN training. In ArXiv:1801:04406.
 Mescheder et al. (2017) Mescheder, L., Nowozin, S., and Geiger, A. (2017). The Numerics of GANs. In NeurIPS.
 Minsky (1986) Minsky, M. (1986). The society of mind. Simon and Schuster, New York NY.
 Monderer and Shapley (1996) Monderer, D. and Shapley, L. S. (1996). Potential Games. Games and Economic Behavior, 14:124–143.
 Nemirovski et al. (2010) Nemirovski, A., Onn, S., and Rothblum, U. G. (2010). Accuracy certificates for computational problems with convex structure. Mathematics of Operations Research, 35(1).
 Nisan et al. (2007) Nisan, N., Roughgarden, T., Tardos, É., and Vazirani, V., editors (2007). Algorithmic Game Theory. Cambridge University Press, Cambridge.
 Palaiopanos et al. (2017) Palaiopanos, G., Panageas, I., and Piliouras, G. (2017). Multiplicative Weights Update with Constant StepSize in Congestion Games: Convergence, Limit Cycles and Chaos. In NeurIPS.

Parkes and Wellman (2015)
Parkes, D. C. and Wellman, M. P. (2015).
Economic reasoning and artificial intelligence.
Science, 349(6245):267–272.  Pathak et al. (2017) Pathak, D., Agrawal, P., Efros, A. A., and Darrell, T. (2017). Curiositydriven Exploration by Selfsupervised Prediction. In ICML.
 Rahwan et al. (2019) Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.F., Breazeal, C., Crandall, J. W., Christakis, N. A., Couzin, I. D., Jackson, M. O., Jennings, N. R., Kamar, E., Kloumann, I. M., Larochelle, H., Lazer, D., Mcelreath, R., Mislove, A., Parkes, D. C., Pentland, A. S., Roberts, M. E., Shariff, A., Tenenbaum, J. B., and Wellman, M. (2019). Machine behaviour. Nature, 568:477–486.
 Scott (1999) Scott, J. (1999). Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press.
 Selfridge (1958) Selfridge, O. G. (1958). Pandemonium: a paradigm for learning. In Mechanisation of Thought Processes: Proc Symposium Held at the National Physics Laboratory.
 Silver et al. (2016) Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T. P., Leach, M., Kavukcuoglu, K., Graepel, T., and Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489.
 Smith (1776) Smith, A. (1776). The Wealth of Nations. W. Strahan and T. Cadell, London.
 Storkey (2011) Storkey, A. (2011). Machine Learning Markets. In AISTATS.
 Storkey et al. (2012) Storkey, A., Millin, J., and Geras, K. (2012). Isoelastic Agents and Wealth Udates in Machine Learning Markets. In ICML.
 Sutton et al. (2011) Sutton, R., Modayil, J., Delp, M., Degris, T., Pilarski, P. M., White, A., and Precup, D. (2011). Horde: A Scalable Realtime Architecture for Learning Knowledge from Unsupervised Motor Interaction. In AAMAS.
 Tatarenko and Kamgarpour (2019) Tatarenko, T. and Kamgarpour, M. (2019). Learning Generalized Nash Equilibria in a Class of Convex Games. IEEE Transactions on Automatic Control, 64(4):1426–1439.
 Vickrey (1961) Vickrey, W. (1961). Counterspeculation, Auctions and Competitive Sealed Tenders. J Finance, 16:8–37.

Vinyals et al. (2019)
Vinyals, O., Babuschkin, I., Chung, J., Mathieu, M., Jaderberg, M., Czarnecki,
W. M., Dudzik, A., Huang, A., Georgiev, P., Powell, R., Ewalds, T., Horgan,
D., Kroiss, M., Danihelka, I., Agapiou, J., Oh, J., Dalibard, V., Choi, D.,
Sifre, L., Sulsky, Y., Vezhnevets, S., Molloy, J., Cai, T., Budden, D.,
Paine, T., Gulcehre, C., Wang, Z., Pfaff, T., Pohlen, T., Wu, Y., Yogatama,
D., Cohen, J., McKinney, K., Smith, O., Schaul, T., Lillicrap, T., Apps, C.,
Kavukcuoglu, K., Hassabis, D., and Silver, D. (2019).
AlphaStar: Mastering the RealTime Strategy Game StarCraft II.
https://deepmind.com/blog/alphastarmastering
realtimestrategygamestarcraftii/.  von Neumann (1928) von Neumann, J. (1928). Zur Theorie der Gesellschaftsspiele. Mathematische Annalen, 100(1):295–320.
 von Neumann and Morgenstern (1944) von Neumann, J. and Morgenstern, O. (1944). Theory of Games and Economic Behavior. Princeton University Press, Princeton NJ.
 Wellman and Wurman (1998) Wellman, M. P. and Wurman, P. R. (1998). Marketaware agents for a multiagent world. Robotics and Autonomous Systems, 24:115–125.
 Wu et al. (2019) Wu, Y., Donahue, J., Balduzzi, D., Simonyan, K., and Lillicrap, T. (2019). LOGAN: Latent Optimisation for Generative Adversarial Networks. In arXiv:1912.00953.

Zhu et al. (2017)
Zhu, J.Y., Park, T., Isola, P., and Efros, A. A. (2017).
Unpaired ImagetoImage Translation using CycleConsistent Adversarial Networks.
In CVPR.
A Mechanics of Smooth Markets
This section provides a physicsinspired perspective on smooth markets. Consider a dynamical system with particles moving according to the differential equations:
(19)  
(20)  
(21) 
The kinetic energy of a particle is mass times velocity squared, , or in our case
(22) 
where we interpret the learning rate squared of particle as its mass and as its velocity. The total energy of the system is the sum over the kinetic energies of the particles:
(23) 
For example, in a Hamiltonian game we have that energy is conserved:
(24) 
since , see Balduzzi et al. (2018); Letcher et al. (2019) for details.
Energy is measured in joules (). The rate of change of energy with respect to time is power, measured in joules per second or watts (). Conservation of energy means that a (closed) Hamiltonian system, in aggregate, generates no power. The existence of an invariant function makes Hamiltonian systems easy to reason about in many ways.
Smooth markets are more general than Hamiltonian games in that total energy is not necessarily conserved. Nevertheless, they are much more constrained than general dynamical systems. Legibility, proposition 3, says that the total power (total rate of energy generation) in smooth markets is the sum of the power (rate of energy generation) of the individual particles:
(25) 
Example where legibility fails. Once again, it is instructive to look at a concrete example where legibility fails. Recall the potential game in example 1 with profits
(26) 
and sentiments
(27) 
Physically, the negative sentiments and mean that that each “particle” in the system, considered in isolation, is always dissipating energy. Nevertheless as shown in section 5.1 the system as a whole has
(28) 
which is positive for some values of . Thus, the system as a whole can generate energy through interaction effects between the (dissipative) particles.
B Proofs
Proof of lemma 1.
Lemma 1.
Every continuous dynamical system on , for any , arises as simultaneous gradient ascent on the profit functions of a smooth game.
Proof.
Specifically, we mean that every dynamical system of the form arises as simultaneous gradient ascent on the profits of a smooth game.
Given continuous vector field on , we need to construct a smooth game with dynamics given by . To that end, consider a player game where player controls coordinate . Set the profit of player to
(29) 
and observe that by the fundamental theorem of calculus. ∎
Proof of proposition 2.
Before proving proposition 2, we first prove a lemma.
Lemma 7 (generalized Helmholtz decomposition).
The Jacobian decomposes into where and are symmetric and antisymmetric, respectively, for all .
Proof.
Follows immediately. See Letcher et al. (2019) for details and explanation. ∎
Proposition 2.
Sentiment is additive: .
Proof.
For any collection of updates , we need to show that
(30) 
Direct computation obtains because is antisymmetric and is blockdiagonal. ∎
Proof of proposition 3.
First we prove a lemma.
Lemma 8.
.
Proof.
Observe by direct computation that
(31) 
It is then easy to see that . Thus,
(32) 
where since is symmetric. By antisymmetry of , we have that for all . The expression thus simplifies to
(33) 
by the blockdiagonal structure of . ∎
Proposition 3 (legibility under gradient dynamics).
Fix dynamics . Sentiment decomposes additively:
(34) 
Proof.
Applying the chain rule obtains that
(35) 
where the second equality follows by construction of the dynamical system as . Lemma 8 shows that
(36)  
(37)  
(38)  
(39) 
Finally, since by construction, we have
(40)  
(41) 
for all as required. ∎
Proof of theorem 4.
Theorem 4.
A fixed point in an SMgame is a local Nash equilibrium iff it is stable.
Proof.
Suppose that is a fixed point of the game, that is suppose .
Recall from lemma 7 that the Jacobian of decomposes uniquely into two components where is symmetric and is antisymmetric. It follows that since is antisymmetric. Thus, is a stable fixed point iff is negative definite.
In an SMgame, the antisymmetric component is arbitrary and the symmetric component is block diagonal – where blocks correspond to players’ parameters. That is, for because the interactions between players and are pairwise zerosum – and are therefore necessarily confined to the antisymmetric component of the Jacobian. Since is blockdiagonal, it follows that is negative definite iff the submatrices along the diagonal are negative definite for all players .
Finally, is negative definite iff profit is strictly concave in the parameters controlled by player at . The result follows. ∎
Proof of theorem 5.
Theorem 5.
In continuous time, for all positive learning rates ,

If is a stable fixed point (), then there is an open neighborhood where for all , so the dynamics converge to from anywhere in .

If is an unstable fixed point (), there is an open neighborhood such that for all , so the dynamics within are repelled by .
Proof.
We prove the first part. The second follows by a symmetric argument. First, strict concavity implies is negative definite for all . Second, since is blockdiagonal, with zeros in all blocks for pairs of players , it follows that is also negative definite. Observe that
(42) 
for all since is negative definite. Thus, simultaneous gradient ascent on the profits acts to infinitesimally reduce the function .
Since reduces , it will converge to a stationary point satisfying . Observe that iff since and the symmetric component of the Jacobian is negative definite. Finally, observe that all stationary points of , and hence , are stable fixed points of because is negative definite, which implies that the fixed point is a Nash equilibrium. ∎
Proof of theorem 6.
Theorem 6.
Suppose all firms have negative sentiment, , for sufficiently large values of . Then the dynamics are bounded for any learning rates .
Proof.
Fix and also fix such that for all satisfying . Let and suppose is sufficiently large such that . We show that
(43) 
for the dynamical system defined by . Since we are operating in continuous time, all that is required is to show that implies that for all sufficiently small .
Recall that . It follows immediately that for all in a sufficiently small ball centered at . In other words, the dynamics reduce and the result follows. ∎
C Near SMgames: Experiential Value and the Exchange of Goods
Definition 3 proposes a model of monetary exchange in smooth markets. It ignores some major aspects of actual markets. For example, SMgames do not model inventories, investment, borrowing or interest rates. Moreover, in practice money is typically exchanged in return for goods or services – which are ignored by the model.
In this section, we sketch one way to extend SMgames to model the exchange of both money and goods  although still without accounting for inventories, which would more significantly complicate the model. The proposed extension is extremely simplistic. It is provided to indicate how the model’s expressive power can be increased, and complications that results.
Suppose