# Research

## Linear programming formulation

We specify conditions leading to a desired subgame-perfect equilibrium solution. The main idea is to have a solution that corresponds to the right play. That is, subgame-perfect equilibrium moves – optimal rational moves – are the same as honest moves.

Each decision at each internal node of the decision tree gives a linear constraint on the variable set $$T$$. These are so-called incentive compatibility constraints. There are also budget-balance constraints for each final state, namely that $$t^s_S+t^s_B\leq D_S+D_B+p$$. In simple terms, it means that the mechanism can not add any money into the contract, but it can burn money.

Feasible linear programs normally have more than $$1$$ solution. To pin down the solution set to one, we need to come up with the objective function. The first candidate is to minimize the sum of deposits. The friction coming from putting deposits has to be acknowledged. The obvious cost of making high deposits for any (especially long) time period is opportunity cost. Depositing involves an implicit cost for the participants in the smart contract: This can be the opportunity cost of not using the deposit while the contract is executed, borrowing costs of the agent, risk of loss of the deposit if the consensus protocol fails, etc. In general, mechanisms that use punishment through huge negative transfers in case of “misbehavior” of agents are impractical. Even though the punishments are only executed off the equilibrium path, huge deposits have to be made to make the threat of punishment credible. This would make such mechanisms very costly to implement as a smart contract. Therefore, minimizing the sum of deposits can be an objective function. Opportunity cost in smart contracts and subgame-perfect equilibria is a topic of Mamageishvili and Schlegel (2020). With instant exchange procedures, however, the opportunity cost is not of first-order importance.

An alternative objective function could be to maximize the money burnt (transferred to Boson protocol) on the off-equilibrium path. The transferred money could be used to analyze what went wrong in this particular game. This incentive for network nodes can be useful in the early stages of protocol functioning.

In the first move of the seller, H corresponds to sending a high-quality item, while L corresponds to sending a low-quality item. In the second move, the buyer plays either redeem (R) or no redeem (refund or expire - RfE). In the third move, the buyer complains (C) or does not complain (NC). Finally, the seller acknowledges a fault (CF) or does not acknowledge the fault (NC).

Suppose the seller is deciding the root vertex S. Then the example incentive compatibility constraint at the vertex $$S$$ is the following:

$t*{S_8}^S + u^S*{H, R, NC,NC} \geq t*{S_1}^S + u^S*{L,RfE,C,CF}$

where $$u^X_{p}$$ denotes the utility of player $$X\in \{S,B\}$$ when the mechanism takes path $$p$$ from the root to (any) leaf.

We consider quasi-linear utilities. That is, utility of the path for a player $$X\in \{S,B\}$$ in the decision tree is the sum of utilities of each move. Formally, $$u_p^{X}=u_{e_1}+\cdots+u_{e_n}$$, where $$p$$ is a path consisting of edges $$p=\{e_1,\cdots, e_n\}$$.

There are $$15$$ internal nodes of the decision tree. Each node generates one incentive compatibility constraint of the linear program. Each final state adds $$1$$ constraint. There are $$2$$ additional individual rationality constraints, $$1$$ for each player. In the subgame-perfect equilibrium solution, the expected payoff of the player should be non-negative. In other words, any player may opt-out for not participating in the contract. In total, there are $$15+8+2=25$$ constraints and $$8$$ (transfer) variables. Optimizing target function given these constraints determines the values of transfers, and consequently, smart contract.

## Discussion

The subgame-perfect implementation can be used with machine-to-machine commerce since it avoids the behavioral play of humans. Machines are programmed by rational people, therefore there is no difference between these two. Machines are, on the other hand, better than irrational–behavioral players, or even rational players. Examples include but are not limited to: they do not press the wrong button by accident, they do not make decisions based on a bad mood (e.g, to punish other machines), they do not form opinions about other people’s skills depending on how they look (this phenomenon is extensively studied in the experimental economic literature), etc. Programs simply follow the rules rational people define for them. The core exchange mechanism could be useful with machine-to-machine scenarios as long as machines can do proper verification of the service or item that the seller provides. Examples include digital goods (files, codes, passwords) - the example of a similar solution is filebounty.

With physical goods commerce, the mechanism can still be used for the machine-to-machine approach. The only difference is that some steps (those that machines can not perform) will be performed by humans. Integrating machine and human is not the focus of the protocol. Another advantage of machine-to-machine use of the mechanism is that moves do not need to be intuitive and have a logical explanation. All the relevant information is final transfers and that the subgame-perfect equilibrium corresponds to the correct play. To design transfers in any game, we apply the same linear programming approach as in the section on Linear programming formulation. We need to design how the subgame-perfect equilibrium solution looks, make appropriate assumptions on the utilities for both players coming from each move, and solve the linear program to find transfers in the final states leading to this subgame-perfect equilibrium solution.

The theoretical importance and applicability of subgame-perfect equilibria solutions were found much earlier. Moore and Repullo (1988) developed a general mechanism. The paper proposes subgame perfect implementation of most of the allocation problems given some mild conditions are satisfied. In other words, the authors prove that almost all social choice functions can be implemented as a subgame-perfect equilibrium solution of some game. However, the game forms can be complicated, they have at most $$3$$ stages.

A more recent experiment by Aghion et al. (2018) suggests that participants fail to play these games correctly and that often subgame-perfect equilibrium solution is not implemented. The authors argue that the problem is not in the inability of the participants to use backward induction for calculating the right solution, but in their beliefs about their counterparts’ abilities. $$30\%$$ of the buyers who receive a high-quality good lie, while $$10\%$$ of the buyers lie all the time. In the context of the Boson exchange mechanism, it would be the seller forming pessimistic beliefs about the buyer.

This mechanism can be easily generalized to general values of $$h$$ and $$l$$, where $$h$$ denotes the value of the high-quality item/service and $$l$$ denotes the value of the low-quality item/service. A similar generic example was developed by Gans (2019), with the difference that the seller has different costs of sending/providing different quality item/service.

Two additional recent experimental papers, Fehr et al. (2020) and Chen et al. (2020) test theoretical predictions of how players should play if they were rational and believed that other players were also rational. The first one discusses the role of risk-aversion in sequential games. Players act irrationally if the mechanism introduces too high fines for misbehavior. The second one claims to have a simpler (2-stage) and more efficient (in the sense that players deviate from subgame-perfect play very rarely) mechanism.

These experimental papers explain why games with subgame-perfect equilibrium solutions are not implemented in real life, even though they are efficient and well understood in the theoretical economic literature. We could argue that one additional reason for this is the absence of low-cost intermediaries, which can be replaced by smart contracts and decentralized systems. In the Boson core exchange mechanism, on the other hand, common knowledge is guaranteed by informing both parties joining the smart contract by default. Smart contracts can be seen as a way of generating common knowledge. Informally, common knowledge of a proposition $$P$$ means: every player knows $$P$$, every player knows that every other player knows $$P$$, every player knows that every player knows that every player knows $$P$$, and so on and so forth, infinitely. In our setting, $$P = \text{\{all players are rational\}}$$.

## Future development

Experimenting with the exchange mechanisms with different parameters in different environments, assessing their performance, and finding the best contracts depending on the application is at the center of the core mechanism development. Creating a suitable environment for common knowledge in the game is the ultimate goal to achieve. Such an environment would allow the efficient deployment of not only exchange mechanisms but also more sophisticated economic mechanisms, which on its own would save intermediary fees and a negative impact powerful intermediaries or centralized authorities have on the general ecosystem. Natural live experiments using Boson’s core exchange mechanism should contribute to our knowledge in experimental economics as well. We have so far analyzed a game where only the seller makes a hidden move. In some applications, the buyer may also make a move, before they start reporting to the smart contract. This adds additional complexity to the analysis since the order of moves is not observed by the mechanism, therefore, the analysis of both possible sequences is required. We also need to include mechanism/smart contract fee calculation in the design and analysis.

## References

Aghion, P., Fehr, E., Holden, R., and Wilkening, T. (2018). The role of bounded rationality and imperfect information in subgame perfect implementation – an empirical investigation. Journal of European Economic Association, 16(1):232–274.

Blumrosen, L. and Dobzinski, S. (2016). (almost) efficient mechanisms for bilateral trading. CoRR, abs/1604.04876.

Chen, Y.-C., Holden, R., Kunimoto, T., Sun, Y., and Wilkening, T. (2020). Getting dynamic implementation to work. Working Paper.

Fehr, E., Powell, M., and Wilkening, T. (2020). Behavioral constraints on the design of subgame-perfect implementation mechanisms. Forthcoming at American Economic Review.

Gans, J. S. (2019). The fine print in smart contracts. Working paper.

Janin, S., Qin, K., Mamageishvili, A., and Gervais, A. (2020). Filebounty: Fair data exchange. In IEEE European Symposium on Security and Privacy Workshops, Euro S&P Workshops 2020, Genoa, Italy, September 7-11, 2020, pages 357–366. IEEE.

Mamageishvili, A. and Schlegel, J. C. (2020). Optimal smart contracts with costly verification. In IEEE International Conference on Blockchain and Cryptocurrency, ICBC 2020,Toronto, ON, Canada, May 2-6, 2020, pages 1–8. IEEE.

Moore, J. and Repullo, R. (1988). Subgame perfect implementation. Econometrica, 56(5):1191–1220.