Abstract
This paper studies competitions with rank-based reward among a large number of teams. Within each sizable team, we consider a mean-field contribution game in which each team member contributes to the jump intensity of a common Poisson project process; across all teams, a mean field competition game is formulated on the rank of the completion time, namely the jump time of Poisson project process, and the reward to each team is paid based on its ranking. On the layer of teamwise competition game, three optimization problems are introduced when the team size is determined by: (i) the team manager; (ii) the central planner; (iii) the team members’ voting as partnership. We propose a relative performance criteria for each team member to share the team’s reward and formulate some special cases of mean field games of mean field games, which are new to the literature. In all problems with homogeneous parameters, the equilibrium control of each worker and the equilibrium or optimal team size can be computed in an explicit manner, allowing us to analytically examine the impacts of some model parameters and discuss their economic implications. Two numerical examples are also presented to illustrate the parameter dependence and comparison between different team size decision making.
Similar content being viewed by others
Notes
Using the information generated by \(\rho (t)\) would lead to the closed loop equilibrium in the two-layer mean-field games that will be investigate later. As is well known in the literature of mean-field game, such closed loop equilibrium would also give an open loop equilibrium, which means that each team member does not need to observe \(\rho (t)\) at the equilibrium.
One can also think of the team sizes as determining the change of measure from \({\mathbb {P}}\) to some \({\mathbb {Q}}\) under which \(Z^i\) is an exponential random variable with rate \(z_i\).
To simplify the presentation, let us assume that \(\beta \) is a constant. The main results can be easily extend to the cases when \(\beta \) is a function of the team size z. In particular, Theorem 1 remains valid.
Here the intra-team division effect only applies to the regular team members’ share of the reward, namely, \((1-\theta )K(1+p)(1-\rho (\tau ))^p\). In the public good allocation scheme (\(\varepsilon =0\)), the manager and the each member’s reward have the same order of magnitude; in the budget allocation scheme (\(\varepsilon =1\)), the manager receives a chunk of the fixed pie while each member shares a negligible piece of the remaining pie.
One could also consider other criteria for the central planner, such as minimizing a given quantile of the completion time distribution or maximizing the total welfare of a team.
Here we refer to \(V_{z\alpha _z, z}(0)\) as “equilibrium reward” to separate it from size-related costs in the definition of \(V^c\). It should be understood that \(V_{z\alpha _z, z}(0)\) includes the cost of effort.
Here and in the sequel, the average is taken in the space \((J, {\mathscr {J}}, \nu )\).
References
Aliprantis, C.D., Border, K.C.: Infinite Dimensional Analysis: A Hitchhiker’s Guide, 3rd edn. Springer, Berlin (2006)
Battaglini, M., Nunnari, S., Palfrey, T.R.: Dynamic free riding with irreversible investments. Am. Econ. Rev. 104(9), 2858–71 (2014)
Bayraktar, E., Zhang, Y.: A rank-based mean field game in the strong formulation. Electron. Commun. Probab. 21(72), 1–12 (2016)
Bayraktar, E., Zhang, Y.: Terminal ranking games (2019). Forthcoming in Mathematics of Operations Research
Bayraktar, E., Cvitanić, J., Zhang, Y.: Large tournament games. Ann. Appl. Probab. 29(6), 3695–3744 (2019)
Bensoussan, A., Huang, T., Laurière, M.: Mean field control and mean field game models with several populations. Minimax Theory Appl. 3(2), 173–209 (2018)
Bonatti, A., Hörner, J.: Collaborating. Am. Econ. Rev. 101(2), 632–63 (2011)
Campbell, A., Ederer, F., Spinnewijn, J.: Delay and deadlines: freeriding and information revelation in partnerships. Am. Econ. J. 6(2), 163–204 (2014)
Carmona, R., Wang, P.: Finite-state contract theory with a principal and a field of agents. Forthcoming in Management Science (2020)
Compte, O., Jehiel, P.: Gradualism in bargaining and contribution games. Rev. Econ. Stud. 71(4), 975–1000 (2004)
Cvitanić, J., Georgiadis, G.: Achieving efficiency in dynamic contribution games. Am. Econ. J. 8(4), 309–42 (2016)
Élie, R., Mastrolia, T., Possamaï, D.: A tale of a principal and many many agents. Math. Oper. Res. 44(2), 440–467 (2019)
Fujii, M.: Probabilistic approach to mean field games and mean field type control problems with multiple populations. Forthcoming in Minimax Theory and its Applications (2020)
Georgiadis, G.: Projects and team dynamics. Rev. Econ. Stud. 82(1), 187–218 (2015)
Huang, M., Nguyen, S.L.: Linear-quadratic mean field teams with a major agent. In: 2016 IEEE 55th Conference on Decision and Control (CDC), pp. 6958–6963 (2016)
Legros, P., Matthews, S.A.: Efficient and nearly-efficient partnerships. Rev. Econ. Stud. 60(3), 599–611 (1993)
Lockwood, B., Thomas, J.P.: Gradualism and irreversibility. Rev. Econ. Stud. 69(2), 339–356 (2002)
Nutz, M., Zhang, Y.: A mean field competition. Math. Oper. Res. 44(4), 1245–1263 (2019)
Sanjari, S., Yüksel, S.: Optimal solutions to infinite-player stochastic teams and mean-field teams. IEEE Trans. Autom. Control 66(3), 1071–1086 (2021)
Sun, Y.: The exact law of large numbers via Fubini extension and characterization of insurable risks. J. Econ. Theory 126(1), 31–69 (2006)
Yildirim, H.: Getting the ball rolling: voluntary contributions to a large-scale public project. J. Public Econ. Theory 8(4), 503–528 (2006)
Acknowledgements
Yuchong Zhang is supported by NSERC Discovery Grant RGPIN-2020-06290.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Proofs
Proofs
1.1 Proof of Theorem 1
Proof
(i) The function \(V_{\lambda , z}\) in (6) is well-defined by the non-negativity and integrability of \(\lambda \) (see the definition of \({\mathscr {A}}\)). It is straightforward to verify that \(V_{\lambda , z}\) satisfies (5) and equivalently, (3) with \({\bar{\alpha }}=\alpha _z\), and that \(\alpha _z\in {\mathscr {A}}\). Standard verification argument shows that \(V_{\lambda , z}\) is the value function (in response to \((\lambda , z,\alpha _z)\)), and that \(\alpha _z\) is an optimal control.
(ii) Let \({{\hat{\alpha }}} \in {\mathscr {A}}\) be any equilibrium control and \({{\hat{V}}}\) be the corresponding equilibrium value function within the team. Since the best response problem within a team is time-consistent, the restriction of \({{\hat{\alpha }}}\) on [r, 1] is optimal for \({{\hat{V}}}(r)\) (in response to \((\lambda , z,{{\hat{\alpha }}})\)) for any \(r<1\). By the optimality of \({{\hat{\alpha }}}\), we have \({{\hat{V}}}(r)\le G_z(r)\). On the other hand, taking the admissible control \(\alpha =\epsilon G_z\in {\mathscr {A}}\), we obtain
Letting \(\epsilon \rightarrow 0+\) yields \({{\hat{V}}}(r)\ge \beta G_z(1)\). Because \(G_z(1-)=G_z(1)=0\), we must have \({{\hat{V}}}(1-)=0\).
Claim that \({{\hat{V}}}\) is absolutely continuous. Once this is proved, dynamic programming yields that \({{\hat{V}}}\) must a.e. satisfy (3) with \({\bar{\alpha }}={{\hat{\alpha }}}\), and that \({{\hat{\alpha }}}\) coincides with \(\alpha _z\), which further implies that \({{\hat{V}}}\) satisfies (5). It is easy to check that (5) has at most one absolutely continuous solution, namely, (6). By its uniqueness, we must have \({{\hat{V}}}=V_{\lambda , z}\).
The rest is devoted to the proof of absolutely continuity of \({{\hat{V}}}\) by a control-theoretical argument adapted from [18]. Fix an arbitrary \(r_0<1\). Since \(\lambda \) is assumed to be locally piecewise Lipschitz and strictly positive on [0, 1), it is uniformly bounded away from zero on \([0,r_0]\). This implies that \(\rho \) will reach \(r_0\) in finite time. Let \(0\le r<r+h\le r_0\), we wish to bound \({{\hat{V}}}(r)-{{\hat{V}}}(r+h)\) by a constant times h. There are two subtle differences from the proof in [18]: First, due to the bonus payment, monotonicity of the value function is unclear; thus, a lower bound for \({{\hat{V}}}(r)-{{\hat{V}}}(r+h)\) is no longer trivial. Second, in our model a single member has negligible impact on the team’s completion time. Hence \(\tau ^{r}\) and \(\tau ^{r+h}\) (where the superscript indicates the dependence on \(\rho (0)\)) are different regardless of how we choose a single member’s control.
Denote by \(\rho ^r\) the state process starting at \(\rho (0)=r\), and let \(t_h\) be the first time \(\rho ^r\) hits \(r+h\), which is finite. By the memoryless property of exponential random variables and the flow property of \(\rho ^r\), we have that
In other words, the distribution of \(\tau ^r-t_h\) conditioned on the event \(\tau ^r\ge t_h\) is the same as the distribution of \(\tau ^{\tau +h}\). Using this, we deduce that
Similarly, with
we can show that
Taking limit as \(\epsilon \rightarrow 0+\) and combining the two chains of inequalities, we obtain
It remains to note that
We conclude that \({{\hat{V}}}\) is Lipschitz continuous on \([0, r_0]\) for any \(r_0<1\) and thus, absolutely continuous on [0, 1). \(\square \)
1.2 Proof of Theorem 2
Proof
We only show (ii) and (iii). Suppose \(K(1+p)\theta >\kappa _0\) so that zero is not an equilibrium team size. For each \({{\bar{z}}} >0\), define
By (11), \(z^*\) is an equilibrium team size if and only if \(z^*\in \mathop {{\mathrm{arg~max}}}\limits _{z> 0}F_{z^*}(z)\) and \(K\theta -\kappa _0-\kappa (z^*)\ge 0\). Since \(F_{{{\bar{z}}}}\) is continuous and \(\lim _{z\rightarrow \infty }F_{{{\bar{z}}}}(z)=-\infty \), the maximum of \(F_{{{\bar{z}}}}\) is attained either at \(z=0\) or at some interior point where the first derivative
vanishes. Any positive equilibrium team size \(z^*\) must satisfy \(F_{z^*}'(z^*)=0\), giving the unique candidate \(z_m^*\) in (15). It remains to check that \(F_{z_m^*}(z)\) attains global maximum at \(z=z_m^*\) and that \(K\theta -\kappa _0-\kappa (z_m^*)\ge 0\).
Let us rewrite the function
where
\(F_{z_m^*}(z)\) attains global maximum at \(z=z_m^*\) if and only if f(x) attains global maximum on \({\mathbb {R}}_+\) at \(x=1\). We have
It is easy to see that for \(x>0\), \({{\,\mathrm{sgn}\,}}(f'(x))={{\,\mathrm{sgn}\,}}(h(x))\), where
Notice that \(h(1)=0\) and \(h'(1)=p+\varepsilon -1-(1+p)(\varepsilon +\delta )/2.\) Consider two cases:
(i) \(\varepsilon +\delta \ge 2\). In this case, h is strictly decreasing, which implies \(f'\) is positive when \(0<x<1\) and negative when \(x>1\). Consequently, the global maximum of f is attained at \(x=1\) as desired.
(ii) \(\varepsilon +\delta < 2\). In this case, h is strictly concave, which implies that it can cross the x-axis at most twice. As \(h(0)=-p<0\), \(x=1\) is a global maximum of f if and only if \(h'(1)<0\) and \(f(1)\ge f(0)\), i.e.,
Note that \(\delta \ge (2-\varepsilon )p/(1+p)\) is equivalent to \(\varepsilon +\delta \ge 2-\delta /p\). Combining the two cases, we see that f(x) attains global maximum at \(x=1\) if and only if \(\delta \ge (2-\varepsilon )p/(1+p)\). We also have that
if and only if
which implies \(\delta \ge (2-\varepsilon )p/(1+p)\). The rest of the theorem statement follows from direct computation using (9), (15), (13) and (14). \(\square \)
1.3 Proof of Theorem 4
Proof
Part 1: Let us first examine the candidate equilibrium team size \(z^*=0\).
Case I: \(\epsilon =0\). For \(\kappa _0=0\), the conclusion I(i) of Theorem 4 is easy to verify. Now assume \(\kappa _0>0\) and \(\delta \in (0,1]\cup [2,\infty )\). Let us define
We get \(J'(z)=\kappa _0z^{-2}-k(\delta -1)z^{\delta -2}\) and \(J''(z)=-2\kappa _0z^{-3}-k(\delta -1)(\delta -2)z^{\delta -3}<0\) as \(\delta \le 1\) or \(\delta \ge 2\). Therefore, the unique interior critical point \({\hat{z}}:=\left( \frac{\kappa _0}{k(\delta -1)}\right) ^{\frac{1}{\delta }}\) is the global maximum point. We have that
Then \(z^*=0\) is the equilibrium team size if and only if \({{\hat{z}}} J({{\hat{z}}})\le 0\), which is equivalent to
Case II: \(\epsilon =1\). It is clear that \(z^*=0\) is an equilibrium team size if \(\frac{(1+\beta )}{2} K(1+p)\le \kappa _0\). Now suppose that \(\frac{(1+\beta )}{2} K(1+p)>\kappa _0\) and let us define
We get
It then follows that \(\lim _{z\rightarrow 0+}J(z)=+\infty \) and hence \(z^*=0\) is not an equilibrium team size in view of its definition.
Part 2: Next, consider the candidate equilibrium team size \(z^*>0\). We can compute from (19) that
Again as team sizes are required to be positive, we only need to consider interior maxima of \(H(z;z^*)\). Therefore, \(z^*\) is the equilibrium team size with partnership implies that \(z^*\) satisfies that \(H'(z^*;z^*)=0\), which gives that \(z^*\) solves the algebraic equation
Case I: \(\varepsilon =0\). Suppose that \(\delta \ge 3\). If \(\kappa _0=0\), it is clear that the algebraic equation (22) admits a unique positive root \(z^*_p=\left( \frac{A}{k(\delta -1)}\right) ^{\frac{1}{\delta -1}}\), where we denote \(A:=\frac{pK(1+\beta )}{(p+1)}>0\). If \(\kappa _0>0\), let us denote \(\gamma (x):=Ax+k(1-\delta )x^{\delta }+\kappa _0\). We have that \(\lim _{x\rightarrow 0} \gamma (x)=\kappa _0>0\) and \(\lim _{x\rightarrow \infty } \gamma (x)=-\infty \). Therefore, the equation \(\gamma (x)=0\) admits at least one positive root. Moreover, we also know that \(\gamma '(x)=A+k\delta (1-\delta )x^{\delta -1}\) and therefore \(\gamma (x)\) is strictly increasing for \(x\le x^*\) and strictly decreasing for \(x>x^*\), where
It then follows that the curve \(y=\gamma (x)\) only hits x-axis once, which implies that \(\gamma (x)=0\) admits a unique positive root \(z_p^*\).
Case II: \(\varepsilon =1\). The algebraic equation (22) can be simplified as
It is clear that if \(\delta \ge 2\) and \(\frac{K(1+\beta )}{2(p+1)}<\kappa _0\), we can obtain the unique positive solution given in (20).
It then suffices to verify that \(H(z;z^*_p)\) attains its global maximum at the unique point \(z=z^*_p\) in two cases.
Case I: \(\varepsilon =0\). Let us assume that \(\delta \ge 3\) and \(p\ge 1/3\). We first have
and
It is straightforward to verify that the sign of \(H'(z;z_p^*)\) coincides with the sign of \(h(z;z_p^*)\), which is defined by
Note that \(z_p^*\) solves the equation \(Ax+k(1-\delta )x^{\delta }+\kappa _0=0\), we get that
After changing variable \(x=\frac{z}{z_p^*}\), we can consider the function
with \(B:=pK(1+p)(1+\beta )\) and \(C:=\kappa _0(z_p^*)^{-1}\). First, we have \(h(1)=B-(p+1)^2A=0\) by recalling that \(A=\frac{pK(1+\beta )}{(p+1)}\). Moreover, we have that
As \(A,C>0\) and \(\delta \ge 3\), it follows that \((A+C)\delta \ge 3A\). As \(4A(p+1)>0\), we can then deduce that \(h'(1)<3B-3A(p+1)^2=0\) and hence \(z=z_p^*\) is a local maximum of the function \(H(z;z_p^*)\).
We then claim that the equation \(h(x)=0\), \(x>0\), admits a unique solution at \(x=1\). As we already know that \(h(1)=0\) and \(h'(1)<0\), we will show that for any other \({\bar{x}}>0\) such that \(h({\bar{x}})=0\), we always have \(h'({\bar{x}})<0\) and therefore \({\bar{x}}=1\) must be the unique solution as h(x) is a continuous function. Let us then assume that \({\bar{x}}\ne 1\) that also satisfies
and we have
To show \(h'({\bar{x}})<0\) for \({\bar{x}}\ne 1\), it is equivalent to show that \(g({\bar{x}})<0\). First, for the second term of \(g({\bar{x}})\), the condition \(\delta \ge 3\) clearly implies that \((1+\delta ){\bar{x}}^2+(\delta -3)p>0\) for any \(x>0\) and therefore the second term is always negative for \(x>0\). For the first term of \(g({\bar{x}})\), recall the condition that \(p\ge 1/3\) and hence \(\sqrt{3p}\ge 1\). It is then clear to see that either if \({\bar{x}}\le 1\) or \({\bar{x}}\ge \sqrt{3p}\), we have \(({\bar{x}}^2-3p)(1-{\bar{x}}^{\delta })\le 0\) and it follows that the first term is nonpositive. Otherwise, for \(1<{\bar{x}}<\sqrt{3p}\), we can also write
This verifies the claim that \(g({\bar{x}})<0\) for any \({\bar{x}}\ge 0\). We can then conclude that the claim holds and \(x=1\) is the unique solution such that \(h(x)=0\). It follows that \(H(z;z_p^*)\) admits a unique critical point \(z=z_p^*\) and hence the local maximum is also the global maximum.
Case II: \(\varepsilon =1\). We have that
as well as
It is clear that the sign of the function \(H'(z;z_p^*)\) coincides with the sign of the function
After changing variable, we can consider the function
where \(A:=\frac{-K(1+p)(1+\beta )}{2}\) and \(B:=(z_p^*)^{\delta }k(1-\delta )=\frac{K(1+\beta )}{2(p+1)}-\kappa _0\).
Note that \(h(1)=0\). Moreover, we have that
By the assumption that \(2(1+\delta )>(p+1)^2\), we get \(2(1+\delta )>(p+1)\) as well and hence \(\frac{\frac{2(1+\delta )}{(p+1)^2}-1}{4(1+\delta )-2(p+1)}\le \frac{1}{2(p+1)^2}\). We obtain that \(h'(1){<}\left[ \frac{K(1{+}\beta )}{2(p{+}1)}{-}\kappa _0\right] \left[ 4(1+\delta ){-}2(p{+}1)\right] <0\) as we assume \(\frac{K(1+\beta )}{2(p+1)}<\kappa _0\). Again, it follows that \(z_p^*\) is a local maximum of the function \(H(z;z_p^*)\).
We then claim that the equation \(h(x)=0\), \(x>0\), admits a unique solution. We again show that for any point \({\bar{x}}\) such that \(h({\bar{x}})=0\), we always have \(h'({\bar{x}})<0\). Let us assume that \({\bar{x}}\) satisfies
and check that
As \(\delta \ge 2\), the quadratic function \(\delta {\bar{x}}^{2}+(2\delta -2){\bar{x}}+(\delta -2)>0\) for any \(x>0\). Thanks to \(B<0\), we have \(h'({\bar{x}})<0\) for any \({\bar{x}}>0\) if \(h({\bar{x}})=0\). This leads to the fact that \({\bar{x}}=1\) is the unique solution to the equation \(h(x)=1\) as h(x) is a continuous function. It then yields that the function \(H(z;z_p^*)\) admits a unique critical point. Therefore, \(z_p^*\) is the global maximum of \(H(z;z_p^*)\), which completes the proof. \(\square \)
1.4 Proof of Proposition 1
Proof
Recall the function H defined in (19). \(z_p^*>0\) is an equilibrium only if \(H'(z;z_p^*)|_{z=z_p^*}=0\), which is equivalent to
where
As \(\delta >1\), the functions \(z\mapsto K\alpha (\varepsilon ,p)(1+\beta )z^{1-\varepsilon }\) and \(z\mapsto -\kappa _0+k(\delta -1)z^\delta \) have at most one intersection point \(z_p^*>0\) for \(z>0\).
Note that \(\alpha (\varepsilon ,p)>0 \Leftrightarrow 2p-2p\varepsilon -\varepsilon >0\), and if this holds, it can be easily seen that \(z_p^*\) is increasing w.r.t. \(\beta \). Similarly we can show the monotonicity of \(z_p^*\) w.r.t. \(\beta \) when \(\alpha (\varepsilon ,p)=0\) and \(\alpha (\varepsilon ,p)<0\). \(\square \)
1.5 Proof of Proposition 2
Proof
Consider
We have that
where the second equality follows from (24). Therefore,
By Proposition 1, when \(2p-2p\varepsilon -\varepsilon \le 0\), \(\frac{d}{d\beta }z_p^*(\beta )\le 0\) and thus \(\frac{d}{d\beta }V^p(\beta ,z_*^p(\beta ))\ge 0\). For the rest of the proof, we assume \(2p-2p\varepsilon -\varepsilon >0\).
By (24) we have that
Since \(2p-2p\varepsilon -\varepsilon>0\Leftrightarrow \alpha (\varepsilon ,p)> 0\), we deduce that
It follows that
where the second equality follows from (24). This completes the proof. \(\square \)
Rights and permissions
About this article
Cite this article
Yu, X., Zhang, Y. & Zhou, Z. Teamwise Mean Field Competitions. Appl Math Optim 84 (Suppl 1), 903–942 (2021). https://doi.org/10.1007/s00245-021-09789-1
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00245-021-09789-1
Keywords
- Teamwork formulation
- Rank-based reward
- Mean field game of mean field games
- Optimal team size
- Equilibrium team size