Abstract
We investigate the resolution of second-order, potential, and monotone mean field games with the generalized conditional gradient algorithm, an extension of the Frank-Wolfe algorithm. We show that the method is equivalent to the fictitious play method. We establish rates of convergence for the optimality gap, the exploitability, and the distances of the variables to the unique solution of the mean field game, for various choices of stepsizes. In particular, we show that linear convergence can be achieved when the stepsizes are computed by linesearch.
Similar content being viewed by others
References
Achdou, Y., Laurière, M.: Mean field games and applications: numerical aspects. Mean Field Games: Cetraro, Italy 2019, 249–307 (2020)
Benamou, J.-D., Carlier, G.: Augmented Lagrangian methods for transport optimization, mean field games and degenerate elliptic equations. J. Optim. Theory Appl. 167(1), 1–26 (2015)
Benamou, J.-D., Carlier, G., Di Marino, S., Nenna, L.: An entropy minimization approach to second-order variational mean-field games. Math. Models Methods Appl. Sci. 29(08), 1553–1583 (2019)
Benamou, J.-D., Carlier, G., Santambrogio, F.: Variational mean field games. In: Active Particles, Vol. 1, pp. 141–171. Springer, New York (2017)
Bonnans, J.F., Hadikhanloo, S., Pfeiffer, L.: Schauder estimates for a class of potential mean field games of controls. Appl. Math. Optim. 83, 1431–1464 (2021)
Bonnans, J.F., Lavigne, P., Pfeiffer, L.: Discrete potential mean field games: duality and numerical resolution. Math. Program. (2023)
Bredies, K., Lorenz, D.A., Maass, P.: A generalized conditional gradient method and its connection to an iterative shrinkage method. Comput. Optim. Appl. 42(2), 173–193 (2009)
Briani, A., Cardaliaguet, P.: Stable solutions in potential mean field game systems. Nonlinear Differ. Equ. Appl. 25(1), 1 (2018)
Briceño-Arias, L., Kalise, D., Kobeissi, Z., Laurière, M., González, A.M., Silva, F.: On the implementation of a primal-dual algorithm for second order time-dependent mean field games with local couplings. In: ESAIM: Proceedings and Surveys, Vol. 65, pp. 330–348 (2019)
Briceño-Arias, L., Kalise, D., Silva, F.: Proximal methods for stationary mean field games with local couplings. SIAM J. Control Optim. 56(2), 801–836 (2018)
Cacace, S., Camilli, F., Goffi, A.: A policy iteration method for mean field games. ESAIM 27, 85 (2021)
Camilli, F., Tang, Q.: Rates of convergence for the policy iteration method for mean field games systems. J. Math. Anal. Appl. 512(1), 126–138 (2022)
Cardaliaguet, P., Graber, P.J., Porretta, A., Tonon, D.: Second order mean field games with degenerate diffusion and local coupling. Nonlinear Differ. Equ. Appl. 22(5), 1287–1317 (2015)
Cardaliaguet, P., Hadikhanloo, S.: Learning in mean field games: the fictitious play. ESAIM 23(2), 569–591 (2017)
Cardaliaguet, P., Lehalle, C.-A.: Mean field game of controls and an application to trade crowding. Math. Financ. Econ. 12(3), 335–363 (2018)
Elie, R., Pérolat, J., Laurière, M., Geist, M., Pietquin, O.: Approximate fictitious play for mean field games. arXiv preprint arXiv:1907.02633 (2019)
Fleming, W.H., Soner, H.M.: Controlled Markov Processes and Viscosity Solutions, vol. 25. Springer, New York (2006)
Geist, M., Pérolat, J., Laurière, M., Elie, R., Perrin, S., Bachem, O., Munos, R., Pietquin, O.: Concave utility reinforcement learning: the mean-field game viewpoint. arXiv preprint arXiv:2106.03787 (2021)
Gilbarg, D., Trudinger, N.S.: Elliptic Partial Differential Equations of Second Order. Springer, New York (2015)
Graber, P.J., Mouzouni, C.: Variational mean field games for market competition. In: PDE Models for Multi-agent Phenomena, pp. 93–114. Springer, New York (2018)
Graber, P.J., Mullenix, A., Pfeiffer, L.: Weak solutions for potential mean field games of controls. Nonlinear Differ. Equ. Appl. 28(5), 1–34 (2021)
Hadikhanloo, S.: Learning in anonymous nonatomic games with applications to first-order mean field games. arXiv preprint arXiv:1704.00378 (2017)
Hadikhanloo, S., Silva, F.J.: Finite mean field games: fictitious play and convergence to a first order continuous mean field game. J. Math. Pures Appl. 132, 369–397 (2019)
Huang, M., Malhamé, R.P., Caines, P.E.: Large population stochastic dynamic games: closed-loop McKean-Vlasov systems and the Nash certainty equivalence principle. Commun. Inf. Syst. 6(3), 221–252 (2006)
Jaggi, M.: Revisiting Frank-Wolfe: projection-free sparse convex optimization. In: International Conference on Machine Learning, pp. 427–435. PMLR (2013)
Kunisch, K., Walter, D.: On fast convergence rates for generalized conditional gradient methods with backtracking stepsize. Numerical Algebra, Control and Optimization (2022)
Lacoste-Julien, S.: Convergence rate of frank-wolfe for non-convex objectives. arXiv preprint arXiv:1607.00345, (2016)
Ladyzhenskaia, O.A., Solonnikov, V.A., Ural’tseva, N.N.: Linear and Quasi-linear Equations of Parabolic Type, vol. 23. American Mathematical Soiety, Providence (1988)
Lasry, J.-M., Lions, P.-L.: Mean field games. Jpn. J. Math. 2(1), 229–260 (2007)
Lions, J.-L.: Optimal Control of Systems Governed by Partial Differential Equations, vol. 170. Springer, New York (1971)
Nocedal, J., Wright, S.J.: Numerical Optimization. Springer, New York (1999)
Perrin, S., Pérolat, J., Laurière, M., Geist, M., Elie, R., Pietquin, O.: Fictitious play for mean field games: Continuous time analysis and applications. arXiv preprint arXiv:2007.03458 (2020)
Santambrogio, F., Shim, W.: A cucker-smale inspired deterministic mean field game with velocity interactions. SIAM J. Control Optim. 59(6), 4155–4187 (2021)
Sorin, S.: Continuous time learning algorithms in optimization and game theory. In: Dynamic Games and Applications, pp. 1–22 (2022)
Wang, W., Han, J., Yang, Z., Wang, Z.: Global convergence of policy gradient for linear-quadratic mean-field control/game in continuous time. In: International Conference on Machine Learning, pp. 10772–10782. PMLR (2021)
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Financial or Non-financial interests
This work was supported by a public grant as part of the Investissement d’avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH. The authors have no relevant financial or non-financial interests to disclose.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was supported by a public grant as part of the Investissement d’avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH.
A Appendix: Regularity of the auxiliary mappings
A Appendix: Regularity of the auxiliary mappings
This appendix contains the proofs of the technical lemmas of Subsection 5.1.
1.1 A.1 Parabolic estimates
In this section we provide estimates for the following parabolic equation:
for different assumptions on \(b :Q \rightarrow \mathbb {R}^d\), \(c :Q \rightarrow \mathbb {R}\), \(h :\rightarrow \mathbb {R}\), and \(u_0 :\mathbb {T}^d \rightarrow \mathbb {R}\). The proofs of the following results can be found in the Appendix of [5]; they largely rely on [28]. We recall that q is a fixed parameter and \(q>d+2\).
In the next theorem, we consider the Sobolev space \(W^{2-2/p,p}(\mathbb {T}^d)\) with a fractional order of derivation, see [28, section II.2] for a definition.
Theorem 29
For all \(R>0\), there exists \(C>0\) such that for all \(u_0 \in W^{2-2/q,q}(\mathbb {T}^d)\), for all \(b \in L^q(Q;\mathbb {R}^d)\), for all \(c \in L^q(Q)\), and for all \(h \in L^q(Q)\) satisfying
equation (31) has a unique solution u in \(W^{2,1,q}(Q)\). Moreover, \(\Vert u \Vert _{W^{2,1,q}(Q)} \le C\).
Theorem 30
There exists \(C>0\) such that for all \(u_0 \in W^{2-2/q,q}(\mathbb {T}^d)\) and for all \(h \in L^q(Q)\), the unique solution u to (31) (with \(b = 0\) and \(c=0\)) satisfies the following estimate:
Theorem 31
For all \(\beta \in (0,1)\), for all \(R>0\), there exist \(\alpha \in (0,1)\) and \(C>0\) such that for all \(u_0 \in \mathcal {C}^{2+ \beta }(\mathbb {T}^d)\), \(b \in \mathcal {C}^{\beta ,\beta /2}(Q;\mathbb {R}^d)\), \(c \in \mathcal {C}^{\beta ,\beta /2}(Q)\) and \(h \in \mathcal {C}^{\beta ,\beta /2}(Q)\) satisfying \(\Vert u_0 \Vert _{\mathcal {C}^{2+ \beta }(\mathbb {T}^d)} + \Vert b \Vert _{\mathcal {C}^{\beta ,\beta /2}(Q;\mathbb {R}^d)} + \Vert c \Vert _{\mathcal {C}^{\beta ,\beta /2}(Q)} + \Vert h \Vert _{\mathcal {C}^{\beta ,\beta /2}(Q)} \le R\), the solution to (31) lies in \(\mathcal {C}^{2+\alpha ,1+\alpha /2}(Q)\) and satisfies \(\Vert u \Vert _{\mathcal {C}^{2+\alpha ,1+\alpha /2}(Q)} \le C\).
1.2 A.2 Fokker-Planck equation
Proof of Lemma 13
Let us write the Fokker-Planck equation in the form of equation (31): \(\partial _t m - \Delta m + (\nabla \cdot v) m + \langle v, \nabla m \rangle = 0\). The first part of lemma follows from Theorem 29. The nonnegativity of \(\varvec{M}[v]\) is proved in [5, Lemma 3]. \(\square \)
Proof of Lemma 14
Set \(w= v_2-v_1\) and \(\mu = m_2-m_1\). Then \(\mu \) is the solution to
Set \(V=W^{1,2}(\mathbb {T}^d)\) and consider the Gelfand triple \((V,L^2(\mathbb {T}^d),V^*)\), where \(V^*\) denotes the dual of V. Then \(\mu \) is solution of a parabolic equation of the form
where \(B(t) \in L(V, V^{*})\) and \(f(t) \in V^{*}\). For any \(m \in V\), we have
where the second equality is obtained by integration by parts. Using Cauchy-Schwarz inequality and \(\Vert v_1 \Vert _{L^\infty (Q;\mathbb {R}^d)} \le R\), we obtain the following inequality:
where the constant C is independent of t (but depends on R). A direct application of Young’s inequality yields the existence of C (depending on R) such that
Thus B(t) is semi-coercive, uniformly in time. With similar techniques, one can show that \(\langle B(t)m,m' \rangle _V \le C \Vert m \Vert _V \Vert m' \Vert _V\), for a.e.@ \(t \in (0,T)\) and for all m and \(m'\) in V. We can apply [30, Chapter 3, Theorems 1.1 and 1.2], from which we derive
Finally, since \(\Vert m_2 \Vert _{L^\infty (Q)} \le R\), we have \(\Vert w m_2 \Vert _{L^2(Q;\mathbb {R}^d)}^2 \le C \int _Q |w|^2 m_2 \, \textrm{d}x \, \textrm{d}t. \) Combining the two last obtained inequalities, we obtain the announced result. \(\square \)
1.3 A.3 HJB equation
Lemma 32
The Hamiltonian H is differentiable with respect to p and \(H_p\) is differentiable with respect to x and p. Moreover, H, \(H_p\), \(H_{px}\), and \(H_{pp}\) are locally Hölder-continuous.
Proof
See [5, Lemma 1]. \(\square \)
The analysis of the HJB equation relies on its connection with the value function of an optimal control problem, that was introduced in (11). This connection allows first to show a uniform bound for \(\varvec{u}[\gamma ,P]\).
Lemma 33
Let \(R> 0\) and let \((\gamma ,P) \in \Xi _R\). There exists a constant \(C(R) > 0\) such that \(\Vert \varvec{u}[\gamma ,P] \Vert _{L^\infty (Q)} \le C(R)\) and such that u is C(R)-Lipschitz continuous with respect to x. Moreover, for any \((x,t) \in Q\),
In the above relation, \(\mathbb {L}_{\mathbb {F}}^{2,C(R)}(t,T)\) denotes the set of stochastic processes \(\nu \in \mathbb {L}_{\mathbb {F}}^{2}(t,T)\) such that \(\mathbb {E} \big [ \int _t^T |\nu _s|^2 \textrm{d}s \big ] \le C(R)\).
Proof
We first derive a lower bound of L. By assumption (H4), L(x, t, 0) and \(L_v(x,t,0)\) are bounded. It follows then from the strong convexity assumption (Assumption (H1)) that there exists a constant \(C> 0\) such that
Then, for any \((x,s) \in Q\) and for any \(\nu \in \mathbb {R}^d\), we have the following estimates:
Now we show that \(\varvec{u}[\gamma ,P]\) is bounded in \(L^{\infty }(Q)\). For any \((x,t) \in Q\), using the above bound for the running cost L, the bound \(\Vert \gamma \Vert _{L^{\infty }(Q)} \le R\), together with Assumption (H4) on the terminal cost g, we obtain that \(\varvec{u}[\gamma ,P](x,t) \ge - C(R)\). In addition, using Assumption (H3) and the fact that that \(\Vert \gamma \Vert _{L^{\infty }(Q)} \le R\), we deduce that
from which we conclude that \(\Vert \varvec{u}[\gamma ,P] \Vert _{L^\infty (Q)} \le C(R)\).
Finally we show equation (32). Let \(t\in [0,T]\), let \(\varepsilon \in (0,1)\) and let \({\tilde{\nu }} \in \mathbb {L}_{\mathbb {F}}^2(t,T)\) be an \(\varepsilon \)-optimal process. Since g is bounded (Assumption (H4)) and since \((\gamma ,P) \in \Xi _R\), we deduce from the above inequality that
where the constant C does not depend on t and \(\varepsilon \). Thus any \(\varepsilon \)-optimal process lies in \(\mathbb {L}_{\mathbb {F}}^{2,C}(t,T)\), which concludes the proof. \(\square \)
Proof of Lemma 15
Let \((\gamma _1,P_1)\) and \((\gamma _2,P_2)\) be in \(\Xi _R\). Let \(u_1= \varvec{u}[\gamma _1,P_1]\) and \(u_2= \varvec{u}[\gamma _2,P_2]\). By Lemma 33, there exists \(C>0\) such that
for any \((x,t) \in Q\). We denote \((X^\nu _{s})_{s \in [t,T]}\) the solution to the stochastic differential equation \(\textrm{d}X_s = \nu _s \textrm{d}s + \sqrt{2} \textrm{d}B_s\) with \(X^\nu _t = x\), for any \(\nu \in \mathbb {L}_{\mathbb {F}}^{2}(t,T)\). Then
For any \((x,s) \in Q\) and \(\nu \in \mathbb {R}^d\), the Cauchy-Schwarz inequality yields
Using again Cauchy-Schwarz inequality and \(\Vert a\Vert _{L^\infty (Q;\mathbb {R}^{k \times d})} \le C\), we conclude that
as was to be proved. \(\square \)
We prove Proposition 16 with a density argument. In a nutshell: we prove in Proposition 34 below that the result of Proposition 16 holds true when \(\gamma \) and P are Hölder continuous. Then we pass to the limit, using Lemma 15.
Proposition 34
Let \(R>0\) and let \(\beta \in (0,1)\). For any \((\gamma ,P) \in \Xi _R\cap \mathcal {C}^\beta (Q) \times \mathcal {C}^\beta (0,T;\mathbb {R}^k)\), the viscosity solution to the Hamilton-Jacobi-Bellman equation (9) is a classical solution. Moreover, there exists \(\alpha \in (0,1)\) such that \(\varvec{u}[\gamma ,P]\) lies in \(\mathcal {C}^{2+\alpha ,1+\alpha /2}(Q)\) and there exists a constant C(R), depending only on R, such that \(\Vert \varvec{u}[\gamma ,P] \Vert _{W^{2,1,q}(Q)} \le C\).
The proof of Proposition 34 is given at page 32 and relies on a fixed point approach which requires some preparatory work. We introduce the map \(\mathcal {T} :W^{2,1,q}(Q) \times [0,1] \rightarrow W^{2,1,q}(Q)\) which associates to any \(u \in W^{2,1,q}(Q)\) and \(\tau \in [0,1]\) the classical solution \({\tilde{u}} = \mathcal {T}[u,\tau ]\) to the linear parabolic equation
For any \((u,\tau ) \in W^{2,1,q}(Q) \times [0,1]\), we have \(\tau ( \gamma - \varvec{H}[\nabla u + A^\star P]) \in L^\infty (Q)\), by Lemma 32 and Lemma 1. It follows then from Theorem 29 that \(\mathcal {T}[u,\tau ]\) lies in \(W^{2,1,q}(Q)\), proving that \(\mathcal {T}\) is well-defined.
Lemma 35
The mapping \(\mathcal {T}\) is continuous and compact. In addition, for all \(K >0\), there exists \(\alpha \in (0,1)\) and \(C>0\) depending on K, \(\gamma \), and P such that \(\Vert u\Vert _{W^{2,1,q}(Q)} \le K\) implies \(\Vert \mathcal {T}[u,\tau ]\Vert _{\mathcal {C}^{2+\alpha ,1+ \alpha /2}(Q)} \le C\).
Proof
Step 1: Continuity of \(\mathcal {T}\). Let \((u_k,\tau _k) \in W^{2,1,q}(Q) \times [0,1]\) be a sequence converging to \((u,\tau ) \in W^{2,1,q}(Q) \times [0,1]\). Then \(\nabla u_k \rightarrow \nabla u\) in \(L^\infty (Q;\mathbb {R}^d)\) by Lemma 1. Then \(\tau _k (\gamma - \varvec{H}[\nabla u_k + A^\star P]) \rightarrow \tau (\gamma - \varvec{H}[\nabla u + A^\star P])\) in \(L^\infty (Q;\mathbb {R}^d)\) by continuity of the Hamiltonian (see Lemma 32). Finally \(\mathcal {T}\) is continuous, by Theorem 30.
Step 2: Compactness of \(\mathcal {T}\). Let \(K>0\) and let \((u,\tau ) \in W^{2,1,q}(Q) \times [0,1]\) be such that \(\Vert u\Vert _{W^{2,1,q}(Q)} \le K\). Combining Lemma 1 and Lemma 32 there exist \(\alpha \in (0,1)\) and \(C>0\) such that \( \Vert \gamma - \varvec{H}[\nabla u + A^\star P]\Vert _{C^\alpha (Q)} \le C\). Then applying Theorem 31, there exist \(\alpha \in (0,1)\) and \(C>0\) such that \(\Vert \mathcal {T}[u,\tau ]\Vert _{\mathcal {C}^{2+\alpha ,1+ \alpha /2}(Q)} \le C\). By the Arzela-Ascoli Theorem the centered ball of \(\mathcal {C}^{2+\alpha ,1+ \alpha /2}(Q)\) of radius \(C>0\) is a relatively compact subset of \(W^{2,1,q}(Q)\). As a consequence \(\mathcal {T}[u,\tau ]\) is a compact mapping and the conclusion follows. \(\square \)
Theorem 36
(Leray-Schauder) Let X be a Banach space and let \(T : X \times [0, 1] \rightarrow X\) be a continuous and compact mapping. Assume that \(T (x,0) = 0\) for all \(x\in X\) and assume there exists \(C>0\) such that \(\Vert x\Vert _X < C\) for all \((x,\tau ) \in X\times [0,1]\) such that \(T (x,\tau ) = x\). Then, there exists \(x \in X\) such that \(T(x,1) = x\).
Proof
See [19, Theorem 11.6]. \(\square \)
Proof of Proposition 34
We prove that under the assumptions of the proposition, the HJB equation has a classical solution in \(\mathcal {C}^{2+\alpha ,1+\alpha /2}(Q)\) (for some \(\alpha \in (0,1)\)), which is then necessarily the unique viscosity solution \(\varvec{u}[\gamma ,P]\). To this purpose, we prove the existence of a solution to the fixed point equation \(u= \mathcal {T}[u,1]\). We have \(\mathcal {T}[u,0] = 0\) for all \(u \in W^{2,1,q}(Q)\). Now let \((u,\tau ) \in W^{2,1,q}(Q) \times [0,1]\) be such that \(\mathcal {T}[u,\tau ] = u\). From Lemma 35, the mapping \(\mathcal {T}\) is continuous and compact, in addition u is a classical solution and thus the viscosity solution to the Hamilton-Jacobi-Bellman equation
and can be interpreted as the value function associated to the following stochastic control problem
where \((X^\tau _{s})_{s \in [t,T]}\) is the solution to \(\textrm{d}X_s = \tau \nu _s \textrm{d}s + \sqrt{2} \textrm{d}B_s\), \(X_0 = Y\). Following [5, Proposition 1, Step 2], there exists a constant \(C>0\), depending only on R, such that \(\Vert u\Vert _{L^\infty (Q)} + \Vert \nabla u\Vert _{L^\infty (Q;\mathbb {R}^d)} \le C\). Then using Lemma 32 and recalling that \((\gamma ,P) \in \Xi _R\), we deduce that \(\Vert \varvec{H}[\nabla u + A^\star P] - \gamma \Vert _{L^\infty (Q)} \le C\). It follows that u is the solution to a parabolic PDE with bounded coefficients and thus \(\Vert u \Vert _{W^{2,1,q}(Q)} \le C\), by Theorem 29. Again, C only depends on R. Finally, by the Leray-Schauder theorem (Theorem 36), there exists a solution to \(u= \mathcal {T}[u,1]\), which is necessarily \(\varvec{u}[\gamma ,P]\). \(\square \)
Proof of Proposition 16
Take \((\gamma ,P) \in \Xi _R\) and fix \(\beta \in (0,1)\). Let \((\gamma _n,P_n)\) be a sequence in \({\Xi }_{R+1} \cap \mathcal {C}^{\beta }(Q) \times \mathcal {C}^{\beta }(0,T;\mathbb {R}^k)\) such that \(\Vert \gamma _n - \gamma \Vert _{L^\infty (Q)} \longrightarrow 0\) and such that \(\Vert P_n - P \Vert _{L^2(0,T;\mathbb {R}^k)} \longrightarrow 0\). We do not detail the construction of such a sequence, this can be done by convolution. Define \(u^n= \varvec{u}[\gamma ^n,P^n]\) and \(u=\varvec{u}[\gamma ,P]\). By Lemma 15, \(u_n \rightarrow u\) for the \(L^\infty \)-norm. Moreover, by Proposition 34,
Thus, the three sequences \((\partial _t u^n)_{n \in \mathbb {N}}\), \((\Delta u^n)_{n \in \mathbb {N}}\), and \((\nabla u^n)_{n \in \mathbb {N}}\) are bounded in \(L^q(Q)\). By the Banach-Alaoglu theorem, the three sequences have at least one accumulation point for the weak topology of \(L^q(Q)\). These three accumulation points are necessarily (by definition of weak derivatives) equal to \(\partial _t u\), \(\Delta u\), and \(\nabla u\), respectively. Since the \(L^q\)-norm is weakly lower semi-continuous, we deduce that \(\Vert u \Vert _{W^{2,1,q}(Q)} \le C(R)\), where C(R) is as in (34). This concludes the proof. \(\square \)
1.4 A.4 The other mappings
Proof of Lemma 17
Let \((\gamma ,P) \in \Xi _R\). Let \(u= \varvec{u}[\gamma ,P]\). We already know from Proposition 16 that \(\Vert u \Vert _{W^{2,1,q}(Q)} \le C(R)\). Then Lemma 1 implies that u and \(\nabla u\) are continuous and that \(\Vert u \Vert _{L^\infty (Q)} \le C(R)\) and \(\Vert \nabla u \Vert _{L^\infty (Q;\mathbb {R}^d)} \le C(R)\). Let \(v=\varvec{v}[\gamma ,P]= - \varvec{H}_p[ \nabla u + A^\star P]\). We have
Using the regularity of u, the regularity properties of the Hamiltonian given in Lemma 20, and the regularity assumptions on a (Assumption (H4)), we deduce that \(\Vert v \Vert _{L^\infty (Q;\mathbb {R}^d)} \le C(R)\) and that \(\Vert D_x v \Vert _{L^q(Q;\mathbb {R}^{d \times d})} \le C(R)\). Moreover, v is continuous.
Next, let \(m= \varvec{m}[\gamma ,P]= \varvec{M}[v]\). A direct application of Lemma 13 yields that \(\Vert m \Vert _{W^{2,1,q}(Q)} \le C\). Finally, let \(w= \varvec{w}[\gamma ,P]= mv\). Using again Lemma 1, we obtain that m is continuous and that \(\Vert m \Vert _{L^\infty (Q)} \le C(R)\) and \(\Vert \nabla m \Vert _{L^\infty (Q;\mathbb {R}^d)} \le C(R)\). Then \(w \in \Theta {}\), with a norm bounded by some constant C(R). The lemma is proved. \(\square \)
Proof of Lemma 18
The two statements concerning \(\varvec{\gamma }\) are directly deduced from Assumptions (H3) and (H5). Let \(w \in \Theta {}\). Recalling the definition of the operator A (page 7), it is easy to see with Assumption (H4) that \(Aw \in \mathcal {C}(0,T;\mathbb {R}^d)\). Assumptions (H3) and (H5) ensure then that \(\varvec{P}[w]= \varvec{\phi } \big [ A[w] \big ]\) lies in \(\mathcal {C}(0,T;\mathbb {R}^k)\) and that \(\Vert \varvec{P}[w] \Vert _{L^\infty (0,T;\mathbb {R}^k)} \le C\). Let us next consider \(w_1\) and \(w_2\) in \(\Theta {}\). We have
by Assumption (H4). Using next the Lipschitz-continuity of \(\phi \) (Assumption (H5)), we obtain that \(\Vert \varvec{\phi }[Aw_2] - \varvec{\phi }[Aw_1] \Vert _{L^2(0,T;\mathbb {R}^k)} \le C \Vert w_2-w_1 \Vert _{L^2(Q;\mathbb {R}^d)}\), as was to be proved. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Lavigne, P., Pfeiffer, L. Generalized Conditional Gradient and Learning in Potential Mean Field Games. Appl Math Optim 88, 89 (2023). https://doi.org/10.1007/s00245-023-10056-8
Accepted:
Published:
DOI: https://doi.org/10.1007/s00245-023-10056-8
Keywords
- Mean field games
- Generalized conditional gradient
- Fictitious play
- Mean field optimal control
- Learning
- Exploitability