Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
MEAN FIELD GAMES WITH COMMON NOISE
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF MATHEMATICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Saran Ahuja
July 2015
http://creativecommons.org/licenses/by-nc/3.0/us/
This dissertation is online at: http://purl.stanford.edu/vt075xr1988
© 2015 by Saran Ahuja. All Rights Reserved.
Re-distributed by Stanford University under license with the author.
This work is licensed under a Creative Commons Attribution-Noncommercial 3.0 United States License.
ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
George Papanicolaou, Primary Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Georg Menz
I certify that I have read this dissertation and that, in my opinion, it is fully adequatein scope and quality as a dissertation for the degree of Doctor of Philosophy.
Leonid Ryzhik
Approved for the Stanford University Committee on Graduate Studies.
Patricia J. Gumport, Vice Provost for Graduate Education
This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file inUniversity Archives.
iii
Abstract
Mean field games (MFG) are a limit of stochastic differential games with a large
number of identical players. They were proposed and first studied by Lasry and Lions
and independently by Caines, Huang, and Malhame in 2006. They have attracted a
lot of interest in the past decades due to their application in many fields. By assuming
independence among each agent, taking the limit as N → ∞ reduces a problem to a
fully-coupled system of forward-backward partial differential equations (PDE). The
backward one is a Hamilton-Jacobi-Bellman (HJB) equation for the value function of
each player while the forward one is the Fokker-Planck (FP) equation for the evolution
of the players distribution. This limiting system is more tractable and one can use
its solution to approximate the Nash equilibrium strategy of N -player games.
In this thesis, we consider the MFG model in the presence of common noise,
relaxing the usual independence assumption of individual random noise. The presence
of common noise clearly adds an extra layer of complexity to the problem as the
distribution of players now evolves stochastically. Our first task is proving existence
and uniqueness of a Nash equilibrium strategy for this game, showing wellposedness
of MFG with common noise. We use a probabilistic approach, namely the Stochastic
Maximum Principle (SMP), instead of a PDE approach. This approach gives us a
forward-backward stochastic differential equation (FBSDE) of McKean-Vlasov type
instead of coupled HJB-FP equations. This was first done by Carmona and Delarue
in the case of no common noise and we extend their results to MFG with common
noise. We are able to extend their results under a linear-convexity framework and a
weak monotonicity assumption on the cost functions. In addition to wellposedness
results, we also prove the Markov property of McKean-Vlasov FBSDE by proving the
iv
existence of a decoupling function.
In the second part of this thesis, we consider ε-MFG models when the common
noise is small. For simplicity, we assume a quadratic running cost function while
keeping a general terminal cost function satisfying the same assumptions as in the
first part. Our goal is to give an approximation of Nash equilibrium of this game
using the solution from the original MFG with no common noise, which could be
described through a finite-dimensional system of PDEs. We characterize the first
order approximation terms as the solution to a linear FBSDE of mean-field type.
We then show that the solution to this FBSDE is a centered Gaussian process with
respect to the common noise. By assuming regularity of the decoupling function of
the 0-MFG problem, we can find an explicit solution showing that they are in the
form of a stochastic integral with respect to the common noise with the integrands
adapted to the information from the 0-MFG only. We then are able to compute the
covariance function explicitly.
v
Acknowledgments
I would like to express my deepest gratitude to Professor George Papanicolaou, my
principal advisor, for his continuous supports and thoughtful guidances towards my
research. His insights and advices have been extremely helpful and his enthusiasm
for mathematics has been inspiring. I am very grateful to have learned and grown so
much as a researcher from working closely with him in the past five years. I would
like to thank the other members of my committee; Professor Lenya Ryzhik, Professor
Georg Menz, Professor Lexing Ying, and Professor Kay Giesecke for taking time out
from their busy schedules to serve in my committee.
I would also like to acknowledge Dr.Tzu-Wei Yang for his tremendous help and
encouragement. He is always available to talk whenever I seek advice and always
gives helpful feedbacks. I also thank David Ren for all the discussions we had and for
carefully reading this thesis and providing useful comments on my work.
I am very thankful for all the friends and colleagues I have met during my time at
Stanford. I am particularly thankful for being a part of the strong Thai community
here which always makes me feel like home. My special thanks goes to Eve, my
wonderful girlfriend and my best friend, whose love and support has been invaluable.
I cannot express how fortunate I am to have her by my side throughout this eventful
journey. Lastly and most importantly, I am very grateful for my family in Thailand;
my mom and dad, my brother and sister, who always believe in me and continuously
send their love, supports, and encouragement from afar.
vi
Contents
Abstract iv
Acknowledgments vi
1 Introduction 1
1.1 Model setup, notations, and main assumptions . . . . . . . . . . . . . 4
1.1.1 N -player stochastic differential game . . . . . . . . . . . . . . 5
1.1.2 Definition of mean field games . . . . . . . . . . . . . . . . . . 6
1.1.3 Derivative with respect to a probability measure . . . . . . . . 9
1.1.4 Main assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Summary of the main results . . . . . . . . . . . . . . . . . . . . . . . 19
1.2.1 Two approaches to mean field games . . . . . . . . . . . . . . 20
1.2.2 Mean field games with common noise . . . . . . . . . . . . . . 27
1.2.3 Asymptotic analysis of mean field games . . . . . . . . . . . . 33
1.2.4 Linear quadratic mean field games with common noise . . . . 39
1.3 Summary of contributions . . . . . . . . . . . . . . . . . . . . . . . . 42
2 Two approaches to mean field games 44
2.1 Dynamic Programming Principle (DPP) . . . . . . . . . . . . . . . . 44
2.2 Stochastic Maximum Principle (SMP) . . . . . . . . . . . . . . . . . 52
2.3 Comparison between the two approaches . . . . . . . . . . . . . . . . 60
3 Mean field games with common noise 64
3.1 Wellposedness of MFG with common noise . . . . . . . . . . . . . . . 64
vii
3.1.1 FBSDE with monotone functionals . . . . . . . . . . . . . . . 65
3.1.2 A priori estimate . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.1.3 Wellposedness result . . . . . . . . . . . . . . . . . . . . . . . 68
3.2 Markov property and a decoupling function . . . . . . . . . . . . . . 74
3.3 Master equation and connection to DPP approach . . . . . . . . . . . 82
3.4 Proof of lemmas, propositions, and theorems . . . . . . . . . . . . . . 86
3.4.1 Proof of Theorem 3.1.1 . . . . . . . . . . . . . . . . . . . . . . 86
3.4.2 Proof of Theorem 3.1.2 . . . . . . . . . . . . . . . . . . . . . . 91
3.4.3 Proof of Theorem 3.2.8 . . . . . . . . . . . . . . . . . . . . . . 95
4 Asymptotic analysis of mean field games 97
4.1 Linear variational FBSDE . . . . . . . . . . . . . . . . . . . . . . . . 98
4.1.1 Wellposedness result . . . . . . . . . . . . . . . . . . . . . . . 100
4.1.2 Convergence result . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2 Approximate Nash equilibrium . . . . . . . . . . . . . . . . . . . . . . 102
4.3 Gaussian property and decoupling function . . . . . . . . . . . . . . . 104
4.3.1 Centered Gaussian process . . . . . . . . . . . . . . . . . . . . 104
4.3.2 Decoupling function . . . . . . . . . . . . . . . . . . . . . . . 105
4.4 Explicit solution and covariance function . . . . . . . . . . . . . . . . 108
4.4.1 Sensitivity functional . . . . . . . . . . . . . . . . . . . . . . . 110
4.4.2 Explicit solution . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4.3 Covariance function . . . . . . . . . . . . . . . . . . . . . . . . 114
4.5 Proof of lemmas, propositions, and theorems . . . . . . . . . . . . . . 116
4.5.1 Proof of Theorem 4.1.2 . . . . . . . . . . . . . . . . . . . . . . 116
4.5.2 Proof of Theorem 4.2.5 . . . . . . . . . . . . . . . . . . . . . . 118
4.5.3 Proof of Theorem 4.3.1 . . . . . . . . . . . . . . . . . . . . . . 119
4.5.4 Proof of Theorem 4.3.2 . . . . . . . . . . . . . . . . . . . . . . 120
5 Linear-quadratic MFG with common noise 123
5.1 DPP approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.1.1 Master Equation . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.1.2 Optimal controlled process and stochastic value function . . . 127
viii
5.2 SMP approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3 Asymptotic analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bibliography 131
ix
List of Tables
2.1 Summary of the types of equations from applying DPP and SMP ap-
proaches to MFG problems . . . . . . . . . . . . . . . . . . . . . . . . 62
x
List of Figures
xi
Chapter 1
Introduction
Mean field games, or MFG for short, were recently proposed by Lasry and Lions in
a series of papers [13, 42, 43, 44] and independently by Caines, Huang, and Mal-
hame [37], who called them Nash Certainty Equivalence. An MFG is a limit of a
stochastic differential game with a large number of players, symmetric cost, and weak
interaction. Specifically, each player executes a stochastic control problem whose cost
and/or dynamics depend not only on their own state and control but also on other
players’ states. However, this interaction is weak in a sense that a player feels the
effect of other players only through their empirical distribution. Searching for a Nash
equilibrium strategy for an N -player games is known to be intractable when N is
large as the dimensionality grows with the number of players. However, by assuming
independence of the random noise in the players’ state processes, symmetry of the
cost functions, and a mean-field interaction, taking the limit as N → ∞ reduces
a problem to solving a fully-coupled system of forward-backward partial differential
equations (FBPDEs). The backward one is a Hamilton-Jacobi-Bellman (HJB) equa-
tion for the value function for each player while the forward one is the Fokker-Planck
(FP) equation for the evolution of the player’s probability distribution. This limit
system is more tractable and one can use its solution to approximate Nash equilibrium
strategies of the N -player games. Lasry and Lions studied this model extensively and
rigorously analyzed the limit in some cases [42, 43, 44].
Alternatively, one can apply the Stochastic Maximum Principle (SMP) to the
1
CHAPTER 1. INTRODUCTION 2
control problem as opposed to the Dynamic Programming Principle (DPP). Instead
of the HJB equation, we have a forward backward stochastic differential equations
(FBSDE). In the MFG setting, this FBSDE is of McKean-Vlasov type as it also
involves the law of the process. This probabilistic analysis of MFG was first done by
Carmona and Delarue in [17] and is the framework that we follows in this thesis.
In the past decade, active research has been done in this area and tremendous
progress has been made in several directions. See [32] for a brief survey and [9]
for a more extensive reference. These include a finite state model [29, 30], a model
with major/minor players [8, 23, 36, 50, 51], and a study of convergence from finite
player games to MFG [7, 27, 30, 40]. One important assumption in most of the prior
work is independence of the random factors in each player’s state processes. From
this assumption and the Law of Large Numbers, the distribution of players’ state
evolves deterministically in the limit. It is this property that plays a key role in
reducing the dimension of this complex problem and making it tractable. However,
many models in applications do not satisfy this assumption. For instance, in financial
applications, any reasonable model which attempts to understand the interactions of
a large number of market participants will have to assume that they are exposed to
some type of overall market randomness. This common random factor is applied to
all the players, and, as a result, the independence assumption does not hold. See
[2, 21, 33, 45] for examples of mean field game models with common noise in finance
and economics.
The presence of common noise clearly adds an extra layer of complexity to the
problem as the empirical distribution of players in the limit now evolves stochastically.
One approach would be to abandon the forward backward coupling and simply add the
players’ distribution as an argument in the value function. Using the Dynamic Pro-
gramming Principle, this generalized value function will satisfy an infinite-dimensional
HJB equation referred to as the master equation. This equation requires new theo-
ries and tools as it is a second order infinite-dimensional HJB involving derivatives
with respect to probability measures. See [10, 18, 31] for some discussion on the
master equation. Alternatively, we could revisit the forward-backward equations in
the presence of common noise. Following the PDE approach of Lasry and Lions, the
CHAPTER 1. INTRODUCTION 3
common noise then turns HJB-FP equations to stochastic HJB-FP equations which
are of forward backward stochastic partial differential equation (FBSPDE for short)
type. This FBSPDE is clearly more complicated than the FBPDE counterpart in
the original mean field game. On the other hands, as we shall see, the probabilistic
approach introduced by Carmona and Delarue in [17] can be extended more naturally
to accommodate the common noise. The law of the state process which occurs in the
McKean-Vlasov FBSDE from the SMP will simply be replaced by its conditional law
given the common Brownian motion.
In the first part of this thesis, we are interested in a general MFG model with
common noise. By using SMP, we turn the problem to analyzing a McKean-Vlasov
FBSDE. Our main result is establishing existence and uniqueness of a strong solution
to this FBSDE under a linear-convexity framework and a weak monotonicity condition
on the cost functions. We also prove a Markov property of the solution by showing
the existence of a deterministic function that decouples the FBSDE. The latter gives
a clear connection to the DPP approach. The material in this part is mostly from
our work in [3].
Recently, in addition to [3], there has been progress in this direction concerning
mostly well-posedness results. Carmona et. al. [20] give a general existence and
uniqueness result of a weak solution. In [10, 18], the master equation was discussed
from the perspective of both HJB and probabilistic approaches. Despite these results,
a general common noise model is still difficult and impractical to solve numerically
or explicitly as it does not enjoy the dimension reduction property as in the case of
MFG without common noise. In the original MFG model of Lasry and Lions, we deal
with a system of PDEs, hence numerical schemes can be developed to approximate
the solution. See [1] for analysis and applications. Intuitively, when there is no
common noise, the law m0,αt is expected to be deterministic, so we only need to find
the optimal strategy along the equilibrium distribution. A common noise model,
on the other hand, is much more complex as the flow of the players’ distribution is
stochastic. This means that we need to specify the optimal action for all possible
trajectories of the players’ distribution which is infinite-dimensional. We would like
to point out that while a common noise model will most likely be intractable, under
CHAPTER 1. INTRODUCTION 4
special circumstances, it might be explicitly solvable through a certain transformation
that turns the problem to the original MFG. See [21, 33] for some examples of this
sort and [41] for a more general treatment of what is called a translation-invariant
MFG with common noise.
In the second part of this thesis, we consider MFG problems when the common
noise is small as shall be indicated by the parameter ε. In this set up, it is reasonable
to seek an approximate solution using only finite-dimensional information from the
ε = 0 problem, i.e. the original MFG with no common noise. We prove that the
first order approximation can be characterized as the solution to a linear FBSDE of
mean-field type. We then show that this FBSDE is uniquely solvable and its solution
is a mean-zero Gaussian process with respect to a common Brownian motion. Our as-
sumptions are similar to those in the first part with additional regularity assumptions.
Furthermore, by assuming that the decoupling function of McKean-Vlasov FBSDE is
sufficiently regular, the linear FBSDE described above can be solved explicitly. The
solution is Wt-path dependent in the form of a stochastic integral of dWt with the
integrand adapted to the information from 0-MFG only. We then are able to compute
the covariance function explicitly.
This chapter is organized as follows; In Section 1.1.1-1.1.2, we introduce a model
for an N -player stochastic differential game and formulate its limit, an MFG problem
with common noise. We then discuss a notion of differentiating with respect to a
probability measure in Section 1.1.3, and list all the main assumptions that will be
used throughout the thesis in Section 1.1.4. Lastly, we summarize all the results in
Section 1.2 and summarize our contributions in Section 1.3.
1.1 Model setup, notations, and main assumptions
In this section, we describe a stochastic differential game model with N players, then
formulate an MFG model by formally taking the limit as N → ∞.
CHAPTER 1. INTRODUCTION 5
1.1.1 N-player stochastic differential game
Let T be a fixed terminal time, W i = (W it )0≤t≤T , i = 1, 2, . . . , N, W = (Wt)0≤t≤T
are one dimensional independent Brownian motions. Consider a stochastic dynamic
game with N players, where each player i ∈ 1, 2, . . . , N controls a state process
(X it)0≤t≤T in R given by
dX it = αi
tdt+ σdW it + εdWt, X i
0 = ξi0
by selecting a control αi = (αit)0≤t≤T in H2([0, T ];R), the set of progressively measur-
able processes β = (βt)0≤t≤T such that
E
[∫ T
0
β2t dt
]
<∞
where ξi0 is an initial state of player i. We assume that (ξi0)1≤i≤N are independent
identically distributed, independent of all Brownian motions, and satisfy E[(ξi0)2] <∞
for all 1 ≤ i ≤ N . We will refer to W i as an individual noise or idiosyncratic noise
and W as a common noise.
Given the other players’ strategies, player i selects a control αi ∈ H2([0, T ];R) to
minimize his/her expected total cost given by
J i(αi|(αj)j 6=i) , E
[∫ T
0
f(t, X it , mt, α
it)dt+ g(X i
T , mT )
]
where (αj)j 6=i denotes a strategy profile of other players excluding i, f : [0, T ]× R×
P(R)× R, g : R× P(R) → R are the running and terminal cost functions which are
identical for all players, P(R) denotes the space of Borel probability measure on R,
and mt denote the empirical distribution of (X it)1≤i≤N , i.e.
mt =1
N
N∑
i=1
δXit(dx)
Note that the strategies of the other players have an effect on the cost of player i
CHAPTER 1. INTRODUCTION 6
through mt. This interaction is the main feature that makes this set up a game. We
are seeking a type of equilibrium solution widely used in game theory settings called
a Nash equilibrium.
Definition 1.1.1. A set of strategies (αi)1≤i≤N is a Nash Equilibrium if for every
player i, αi is an optimal control given the other players’ strategies are (αj)j 6=i. In
other words,
J i(αi|(αj)j 6=i) = minα∈H2([0,T ];R)
J i(α|(αj)j 6=i), ∀i ∈ 1, 2, . . . , N
Since the cost function of each player is identical, if the control problem for each
player has a unique solution, then clearly the Nash equilibrium strategy must be
symmetric. However, solving this problem is impractical when N is large due to
high dimensionality, so we formally take the limit as N → ∞ and consider the limit
problem, called a mean field game, instead. Solving a mean field game problem yields
a control that can be used to approximate the exact Nash equilibrium for an N -player
game. See [9, 17] for discussions and results on an approximate Nash equilibrium for
N -player games.
1.1.2 Definition of mean field games
We now formulate an MFG problem in the presence of common noise by formally
taking a limit as N → ∞. By considering the limit problem and assuming that each
player adopts the same strategy, one can represent the distribution of players mt by
the law of a single representative player. However, since the common noise is applied
to all the players, this law is a conditional law given a common noise path. In other
words, we formulate the MFG with common noise as a stochastic control problem for
a single (representative) agent with an equilibrium condition involving a conditional
law of the state process.
Fix a terminal time T > 0. Let (Wt)0≤t≤T , (Wt)0≤t≤T be two independent Brown-
ian motions defined on a complete filtered probability space (Ω,F ,F = Ft0≤t≤T ,P),
augmented by all the P-null sets. (Wt)0≤t≤T is called individual or idiosyncratic noise
CHAPTER 1. INTRODUCTION 7
and (Wt)0≤t≤T is called common noise. Let P2(R) denote the space of Borel proba-
bility measures on R with finite second moment, i.e. all Borel probability measure µ
such that∫
R
x2dµ(x) <∞
It is a complete separable metric space equipped with a Wasserstein metric defined
as
W2(m1, m2) =
(
infγ∈Γ(m1,m2)
∫
R2
|x− y|2γ(dx, dy)
)12
(1.1)
where Γ(m1, m2) denotes the collection of all probability measures on R2 with marginals
m1 andm2. Let F st denote the filtration generated by Wr−Ws, s ≤ r ≤ t and Ft = F0
t .
Suppose G is a sub σ-algebra of F and G = Gt0≤t≤T is a sub filtration of F, then
let L2G denote the set of G-measurable real-valued square integrable random vari-
ables, L2G(P2(R)) denote the set of G-measurable random probability measures µ on
R with finite second moment, and H2G([0, T ];R) denote the set of all Gt-progressively-
measurable process β = (βt)0≤t≤T such that
E
[∫ T
0
β2t dt
]
<∞
We define similarly the space H2G([s, t];R) for any 0 ≤ s < t ≤ T and we will often
omit the subscript and writeH2([0, T ];R) forH2F([0, T ];R). We also letM([0, T ];P2(R))
denote the space of continuous Ft-adapted stochastic flow of probability measures
µ = (µt)0≤t≤T with values in P2(R) such that
E
[
sup0≤t≤T
∫
R
y2dµt(y)
]
<∞
and define similarly M([s, t],R).
Let ξ0 ∈ L2F0
be an initial state. For any control α ∈ H2([0, T ];R), we denote by
Xα = (Xαt )0≤t≤T the corresponding state process, i.e.
Xαt = ξ0 +
∫ t
0
αtdt+ σWt + εWt, (1.2)
CHAPTER 1. INTRODUCTION 8
and mαt is the law of Xα
t conditional on Ft, i.e.
mαt , L(Xα
t |Ft), ∀t ∈ [0, T ] (1.3)
where L(·|Ft) denotes the conditional law given Ft. It is easy to check that when
α ∈ H2([0, T ];R) and E[ξ20 ] < ∞, mα = (mαt )0≤t≤T ∈ M([0, T ];P2(R)), i.e. it has a
continuous trajectory with finite second moment for all t ∈ [0, T ] a.s. We first define
a control problem for a single player given a stochastic flow of probability measures.
Let f : [0, T ] × R × P2(R) × R → R, g : R × P2(R) → R be measurable functions
denoting a running and terminal cost function.
Definition 1.1.2. Given a stochastic flow of probability measure m = (mt)0≤t≤T ∈
M([0, T ];P2(R)), a single player stochastic control problem with state process
dXαt = αtdt+ σdWt + εdWt, X0 = ξ0
and cost
J ε(α|m) = E
[∫ T
0
f(t, Xαt , mt, αt)dt+ g(Xα
T , mT )
]
is called an individual control problem given m.
Remark 1.1.3. Note that because m = (mt)0≤t≤T is stochastic, the individual control
problem given m is a control problem with random cost.
We are now ready to state the definition of an MFG with common noise.
Definition 1.1.4. a control α = (αt)0≤t≤T ∈ H2([0, T ];R) is called a solution to MFG
with common noise if an optimal control for an individual player given (mαt )0≤t≤T is
α. In other words, α satisfies
J ε(α|mα) ≤ J ε(α|mα), ∀α ∈ H2([0, T ];R)
We will often refer to the MFG problem described above as ε-MFG to emphasize
the existence and magnitude of the common noise and call α a solution to ε-MFG
CHAPTER 1. INTRODUCTION 9
with initial ξ0. Note that 0-MFG is simply the original MFG with no common noise
arising from an N -player game with independent Brownian motions.
We would like to emphasize that in the control problem above, mα is exogenous
and is not affected by a player’s control. Thus, MFG is a standard control problem
with an additional equilibrium condition. A type of problem where a player’s control
can affect the law is referred to as Mean Field Type Control Problem. In that setting,
we are searching for a control α such that
J ε(α|mα) ≤ J ε(α|mα), ∀α ∈ H2([0, T ];R)
Notice the difference in the RHS of the inequality. See [9, 16, 19] for analysis on this
different model and some discussion of the differences between the two problems.
The MFG problem described above can also be viewed as a fixed point problem
as follows. Given a strategy α ∈ H2([0, T ];R), then mα is determined as defined in
(1.3). We then solve an individual control problem given mα (see Definition 1.1.2).
This step yields a new optimal control α. The following diagram summarizes the
process
α = (αt)0≤t≤T → mα = (mαt )0≤t≤T → α = (αt)0≤t≤T (1.4)
By the definition of ε-MFG, α is an ε-MFG solution if and only if it is a fixed point
of this map.
1.1.3 Derivative with respect to a probability measure
From the set up of MFG problems, we see that the distribution of a player evolves
stochastically and, as a result, some notion of optimization, hence differentiation, over
a probability measure is unavoidable. In this section, we discuss a notion of derivative
for a function with a probability measure as its argument.
A notion of derivative of a function on the space of probability measures was first
defined using a geometric approach. See [4, 57] for extensive treatments on the subject
in this direction. In this work, however, it is more convenient to use an alternative
approach which is more probabilistic in nature. This method was introduced by Lions
CHAPTER 1. INTRODUCTION 10
in his lecture at the College de France[13]. Since then, many works on MFG have
employed this notion of derivative. While we will only discuss the results that are
necessary for our work here, we refer interested readers to [16] or to [24] for more
details on this framework.
The idea is based on lifting up a function on a space of probability measures to a
function on a space of random variables. When the space of probability measures we
are working on is P2(R), this method is extremely useful because it allows us to work
on a Hilbert space of square integrables random variable instead of a metric space
P2(R). Consequently, we are able to use a notion of Frechet derivative in Hilbert
space to help define a derivative.
Given a functional V : P2(R) → R, we define the lifting function V : L2(Ω, F , P) →
R by
V (X) = V (L(X))
where (Ω, F , P) is an arbitrary probability space with Ω being a Polish space and
P being atomless. This is sufficient to ensure that given any m ∈ P2(R), there
exist a random variable X such that L(X) = m, where L(X) denotes the law of X .
Suppose there exists a random variable X0 ∈ L2(Ω, F , P) such that L(X0) = m0 and
V is Frechet differentiable at X0, denote its derivative by V ′(X0)(·), then by Reiz
representation theorem, there exist an element DV (X0) in L2(Ω, F , P) such that
V ′(X0)(Y ) = E[DV (X0) · Y ]
It then can be shown that if V is differentiable at X0, then it is differentiable at
any X ∈ L2(Ω, F , P) with L(X) = m0 and the law of DV (X) is independent of the
choice of X . See Theorem 6.2 in [13]. Furthermore, it can be shown that there exist
a function h ∈ L2m0
(R;R) uniquely defined m0-a.s. such that
DV (X) = h(X)
See Theorem 6.5 in [13]. As a result, it is natural to call this function h the derivative
of V with respect to m and we shall denote it by ∂mV . Thus, the following relation
CHAPTER 1. INTRODUCTION 11
holds, for any X, Y ∈ L2(Ω, F , P)
V (L(X + Y )) = V (L(X)) + E[∂mV (L(X))(X)Y ] + o(‖Y ‖2)
From the definition, ∂mV (m0)(·) is only uniquely-defined m0-a.s.. Thus, to consider
jointly the map (m, z) → ∂mV (m)(z), we need to specify a canonical version to be
used. When DV is Lipschitz, then it can be shown (see Lemma 3.2 in [16]) that
there exist a version of ∂mV such that z → ∂mV (m)(z) is Lipshitz with the same
constant for all m ∈ P2(R). We shall take this version whenever possible. The higher
order derivative can be defined similarly by using the Frechet derivative on the lifting
function to get the second derivative of the form ∂mmV (m0)(·, ·) ∈ L2m0
(R2;R).
Example 1.1.5. Consider a general class of functionals V : P2(R) → R of the form
V (m) = F
(∫
φ(z)dm(z)
)
where φ, F : R → R are continuously differentiable. Then the lifting function V :
L2(Ω, F , P) → R is
V (X) = F (E [φ(X)])
Thus,
DV (X)(Y ) = F ′(E [φ(X)])(E [φ′(X)Y ]
so that
∂mV (m)(z) = F ′
(∫
φ(u)dm(u)
)
φ′(z)
In the context of MFG with common noise, we will be dealing with a stochastic
flow of probability measures m = (mt)0≤t≤T ∈ M([0, T ];P2(R)) which arises from the
law conditional on Ft of a state process
Xt = ξ0 +
∫ t
0
αsds+ σWt + εWt, ∀t ∈ [0, T ]
To take the derivative using this notion at mt = L(Xt|Ft), we need to find a random
variable to represent it. An obvious choice is simply the state process X = (Xt)0≤t≤T
CHAPTER 1. INTRODUCTION 12
itself. To do this in an explicit manner, we first need to separate the path space
for individual noise and common noise. From now and throughout the rest of the
thesis, we assume that (Ω,F ,P) is in the form (Ω0 × Ω,F0 ⊗ F ,P0 ⊗ P) where the
individual noise Wt and common noise Wt are supported in the space (Ω0,F0,P0)
and (Ω, F , P) respectively. We will also assume that (Ω, F , P) is the canonical sample
space of the Brownian motion (Wt)0≤t≤T , so that F0 is trivial. To avoid confusion
between a lifting space and the original space, we let (Ω0, F0, P0) denote a copy of
(Ω0,F0,P0) and write
(Ω, F , P) = (Ω0 × Ω, F0 ⊗ F , P0 ⊗ P)
to denote a copy of (Ω,F ,P) sharing a common noise space. For any X ∈ L2(Ω,F ,P),
we denote by X , a copy of X in L2(Ω, F , P). Also, we let E0[·] denote the expectation
with respect to P0 only, or equivalently, the expectation with respect to P conditional
on FT .
In addition to introducing a “copy” space sharing a common noise, it will also
be convenient to have a random variable independent of all Brownian motion to
represent an arbitrary law m0 ∈ P2(R). So we further assume that (Ω,F ,P), and
hence (Ω, F , P) as well, is sufficiently rich that for any m0 ∈ P2(R), there exist an
F0-measurable random variable ξ such that L(ξ) = m0. We end this section with the
following lemma which will be useful in the later chapters. The result is standard
in the optimal transport theory and we refer the readers to [55] (see, in particular,
Theorem 3.2.9).
Lemma 1.1.6. Let (Ω, F , P) be a probability space with Ω being a Polish space and
P being an atomless measure. Given m0 ∈ P2(R) and ξ0 ∈ L2(Ω, F , P) with law m0,
then for any m ∈ P2(R), there exists ξ ∈ L2(Ω, F , P) with law m such that
E[(ξ0 − ξ)2] = W22 (m0, m)
CHAPTER 1. INTRODUCTION 13
1.1.4 Main assumptions
In this section, we list all the main assumptions on the cost functions that will be
used in the thesis.
The first set of assumptions, assumption A, is essential for ensuring that given any
stochastic flow of probability measure m = (mt)0≤t≤T ∈ M([0, T ],R), the stochastic
control for an individual player given m (see Definition 1.1.2) is uniquely solvable.
Assumption A consists of standard Lipschitz, linear and quadratic growth, and con-
vexity conditions. For notational convenience, we will use the same constant K for
all the conditions below.
(A1). (Lipschitz) ∂xf, ∂αf, ∂xg exist and are K-Lipschitz continuous in (x, α) uni-
formly in (t,m); for any t ∈ [0, T ], x, x′, α, α′ ∈ R, m ∈ P2(R).
|∂xg(x,m)− ∂xg(x′, m)| ≤ K|x− x′|
|∂xf(t, x,m, α)− ∂xf(t, x′, m, α′)| ≤ K (|x− x′|+ |α− α′|)
|∂αf(t, x,m, α)− ∂αf(t, x′, m, α′)| ≤ K (|x− x′|+ |α− α′|)
(1.5)
(A2). (Growth) ∂xf, ∂αf, ∂xg satisfy a linear growth condition; for any t ∈ [0, T ], x, α ∈
R, m ∈ P2(R),
|∂xg(x,m)| ≤ K
(
1 + |x|+
(∫
R
y2dm(y)
)12
)
|∂xf(t, x,m, α)| ≤ K
(
1 + |x|+ |α|+
(∫
R
y2dm(y)
)12
)
|∂αf(t, x,m, α)| ≤ K
(
1 + |x|+ |α|+
(∫
R
y2dm(y)
)12
)
(1.6)
In addition, f, g satisfy a quadratic growth condition in m.
|g(0, m)| ≤ K
(
1 +
∫
R
y2dm(y)
)
|f(t, 0, m, 0)| ≤ K
(
1 +
∫
R
y2dm(y)
) (1.7)
CHAPTER 1. INTRODUCTION 14
(A3). (Convexity) g is convex in x and f is convex jointly in (x, α) with strict con-
vexity in α. That is, for any x, x′ ∈ R, m ∈ P2(R),
(∂xg(x,m)− ∂xg(x′, m))(x− x′) ≥ 0 (1.8)
and there exist a constant cf > 0 such that for any t ∈ [0, T ], x, x′, α, α′ ∈ R, m ∈
P2(R),
f(t, x′, α′, m) ≥ f(t, x, α,m)+∂xf(t, x, α,m)(x′−x)+∂αf(t, x, α,m)(α′−α)+cf |α′−α|2
(1.9)
The Lipschitz and linear growth conditions (A1),(A2) are standard in SDE theory
to ensure the existence of a strong solution. The convexity assumption (A3) is essen-
tial to our setup in various ways. First, it ensures that the Hamiltonian is strictly
convex, so that there is a unique minimizer in a feedback form. In addition, it satisfies
sufficient conditions for the SMP so that solving an optimal control problem can be
translated to solving the corresponding FBSDE. See Section 6.4.2 in [53] for instance.
Lastly, it gives a monotonicity property for the FBSDE corresponding to an individual
player control problem so that it is uniquely solvable. We refer to [34, 52] for general
solvability properties of standard FBSDEs related to convex control problems.
The second set of assumptions, assumption B, is specific to MFG and are condi-
tions on the m-argument in the cost functions. These assumptions are essential in
showing the wellposed-ness of ε-MFG.
(B1). (Lipschitz in m) ∂xg, ∂xf is Lipschitz continuous in m uniformly in (t, x), i.e.
there exist a constant K such that
|∂xg(x,m)− ∂xg(x,m′)| ≤ KW2(m,m
′)
|∂xf(t, x,m, α)− ∂xf(t, x,m′, α)| ≤ KW2(m,m
′)(1.10)
for all t ∈ [0, T ], x, α ∈ R, m,m′ ∈ P2(R), where W2(m,m′) is the second order
Wasserstein metric defined by (1.1). This is equivalent to the following; for any
CHAPTER 1. INTRODUCTION 15
t ∈ [0, T ], x, α ∈ R, ξ, ξ′ ∈ L2(Ω, F , P;R) where (Ω, F , P) is arbitrary,
|∂xg(x,L(ξ))− ∂xg(x,L(ξ′))| ≤ K‖ξ − ξ′‖2
|∂xf(t, x,L(ξ), α)− ∂xf(t, x,L(ξ′), α)| ≤ K‖ξ − ξ′‖2
(1.11)
where ‖ · ‖2 denote the L2-norm
(B2). (Separable in α,m) f is of the form
f(t, x,m, α) = f 0(t, x, α) + f 1(t, x,m) (1.12)
where f 0 is assumed to be convex in (x, α) strictly in α, f 1 is assumed to be convex
in x.
(B3). (Weak monotonicity) For all t ∈ [0, T ], m,m′ ∈ P2(R) and γ ∈ P2(R2) with
marginals m,m′ respectively,
∫
R2
[(∂xg(x,m)− ∂xg(y,m′))(x− y)] γ(dx, dy) ≥ 0
∫
R2
[(∂xf(t, x,m, α)− ∂xf(t, y,m′, α))(x− y)]γ(dx, dy) ≥ 0
(1.13)
Equivalently, for any x ∈ R, ξ, ξ′ ∈ L2(Ω, F , P;R) where (Ω, F , P) is arbitrary,
E[∂xg(ξ,L(ξ))− ∂xg(ξ′,L(ξ′))(ξ − ξ′)] ≥ 0
E[∂xf(t, ξ,L(ξ), α)− ∂xf(t, ξ′,L(ξ′), α)(ξ − ξ′)] ≥ 0
(1.14)
Lipschitz (in m) condition is similar to that used in other works for 0-MFG model.
Separability condition (B2) is not necessary in establishing existence for 0-MFG, but
is needed for the uniqueness result. See Proposition 3.7 and 3.8 in [17] for instance.
In our case, we rely on the monotonicity property of the McKean-Vlasov FBSDE
and this separability condition is necessary for establishing this property. Our last
assumption (B3) is new and can be viewed as a stronger version of the weak mean-
reverting assumption used in [17] to show the existence result of 0-MFG and a weaker
version of Lasry and Lions’ monotonicity assumption used in [13, 17, 31] to show the
CHAPTER 1. INTRODUCTION 16
uniqueness result. We discuss this point in more detail below.
The next set of assumptions, assumption C, concerns the second order derivative
of the cost functions. This is necessary for our asymptotic analysis in Chapter 4.
(C1). ∂xf, ∂xg are differentiable in (x,m) with Lipschitz continuous and bounded
derivative. Denote their bound and Lipschitz constant by the same K. Specifically for
∂mxf, ∂mxg, they satisfy, for all t ∈ [0, T ], x, α ∈ R, m,m′ ∈ P2(R), and ξ, ξ′ ∈ L2
F
with law m,m′,
E[∂mxf(t, x,m, α)(ξ)2]
12 ≤ K
E[∂mxg(x,m)(ξ)2]12 ≤ K
E[(∂mxf(t, x,m, α)(ξ)− ∂mxf(t, x,m′, α)(ξ′))2]
12 ≤ K‖ξ − ξ′‖2
E[(∂mxg(x,m)(ξ)− ∂mxg(x,m′)(ξ′))2]
12 ≤ K‖ξ − ξ′‖2
(1.15)
Remark 1.1.7. The function ∂mxf, ∂mxg involve the derivative with respect to a
probability measure. We follow the framework introduced by Lasry and Lions in [13]
which is based on a Frechet derivative of a lifting function defined on a space of
random variables. See Section 1.1.3 for more detail.
Our last assumption, assumption D, concerns the second derivative with respect
to m. This assumption was used in [24] to prove differentiability in m-argument of
the decoupling function and we will use this result in Chapter 4 to find the decoupling
function of the linear variational FBSDE.
(D1). For all m ∈ P2(R), the map (x, z) 7→ ∂mxf(t, x,m, α)(z), ∂mxg(x,m)(z), are
continuously differentiable and satisfy; for all t ∈ [0, T ], x, x′, α ∈ R,m,m′ ∈ P2(R),
and ξ, ξ′ ∈ L2F with law m,m′,
E
[
(∂z∂mxf(t, x,m, α)(ξ)− ∂z∂mxf(t, x′, m′, α)(ξ′))
2]
12≤ K (|x− x′|+ ‖ξ − ξ′‖2)
E
[
(∂z∂mxg(x,m)(ξ)− ∂z∂mxg(x′, m′)(ξ′))
2]
12≤ K (|x− x′|+ ‖ξ − ξ′‖2)
(1.16)
CHAPTER 1. INTRODUCTION 17
Comparison to assumptions for 0-MFG
Assumptions A and (B1) are similar to those used in [17] where they establishes
existence of 0-MFG solution under standard Lipschitz and linear-convexity model.
In addition, they also assume the following weak mean reverting assumption which
states that there exist a constant C > 0 such that for all t ∈ [0, T ], x ∈ R
x∂xf(t, 0, δx, 0) ≥ −C(1 + |x|)
x∂xg(0, δx) ≥ −C(1 + |x|)(1.17)
where δx denote the Dirac measure at x. Our weak monotonicity assumption (B3)
can be viewed as a stronger version of this assumption. This can be seen easily by
plugging in deterministic ξ = x, ξ′ = 0 in (1.14).
From the uniqueness result, the main assumption in the literature [13, 17, 31] is
the separability assumption and the Lasry and Lions’ monotonicity assumption which
states that∫
(h(x,m1)− h(x,m2))d(m1 −m2)(x) ≥ 0
for anym1, m2 ∈ P2(R). This condition can be expressed in terms of random variables
as follows; For any ξ, ξ′ ∈ L2 (over an arbitrary probability space).
E [h(ξ′,L(ξ′)) + h(ξ,L(ξ))− h(ξ,L(ξ′))− h(ξ′,L(ξ))] ≥ 0 (1.18)
Our weak monotonicity assumption (B3) is, as the name suggests, a weaker version
of (1.18) when the cost functions are convex. This is shown in the proposition below
Proposition 1.1.8. Let h : R × P2(R) → R be a convex continuously differentiable
function satisfying the monotonicity condition (1.18) stated above, then h satisfies the
weak monotonicity condition (1.14).
Proof. Suppose (1.18) holds. Let ξ, ξ′ ∈ L2, then by convexity in x of h, we get
h(ξ′,L(ξ′))− h(ξ,L(ξ′)) ≤ ∂xh(ξ′,L(ξ′))(ξ′ − ξ)
CHAPTER 1. INTRODUCTION 18
and
h(ξ,L(ξ))− h(ξ′,L(ξ)) ≤ −∂xh(ξ,L(ξ))(ξ′ − ξ)
Summing up, taking expectation, and using (1.18) yields (1.14).
The converse of the proposition above does not hold. Example (1.20) below gives
a cost function that is convex in x, satisfies (1.14), but does not satisfy (1.18). Thus,
our result gives a more general uniqueness theorem for 0-MFG.
Examples of cost functions
We consider a rather general class of f, g of the form
f(t, x,m, α) = F0(t, x, α) + F1
(
t, x,
∫
φf(z)dm(z)
)
+ F2
(∫
ψf (z)dm(z)
)
g(x,m) = G1
(
x,
∫
φg(z)dm(z)
)
+G2
(∫
ψg(z)dm(z)
) (1.19)
where F0, F1 : [0, T ] × R2 → R, G1 : R2 → R, φf , ψf , φg, ψg, F2, G2 : R → R are
all smooth with bounded second derivatives. Then the assumptions A,B,C,D are
satisfied if
∂xxF0, ∂xxF1, ∂xxG0 ≥ 0, ∂ααF0 ≥ c > 0, ∂xxF0∂ααF0 − (∂xαF0)2 ≥ 0
∂xxF1 + ∂xyF1‖φ′f‖∞ ≥ 0, ∂xxG1 + ∂xyG1‖φ
′g‖∞ ≥ 0
The first line is the convexity assumption in x, α strictly in α. The second line is
sufficient to guarantee that the weak monotonicity condition holds. This form of f, g
includes the following examples
f(t, x,m, α) = Aα2 +B
(
x−
∫
zdm(z)
)2
,
g(x,m) = C
(
x−
∫
zdm(z)
)2(1.20)
CHAPTER 1. INTRODUCTION 19
or
f(t, x,m, α) = Aα2 +B
∫
(x− z)2dm(z),
g(x,m) = C
∫
(x− z)2dm(z),
where A,B,C > 0, which occur frequently in applications (see [21, 33] for instance).
It also includes the cost functions in the general linear-quadratic mean field games
(LQMFG) discussed in [11] where f, g takes the form
f(t, x,m, α) =1
2
(
qx2 + α2 + q(x− sm)2)
g(x,m) =1
2
(
qTx2 + (x− sT m)2qT
)
(1.21)
where
m =
∫
R
zdm(z)
and q, q, s, qT , qT , sT are constant satisfying
q + q − qs ≥ 0, qT + qT − qT sT ≥ 0
We will solve this LQMFG explicitly in Chapter 5.
1.2 Summary of the main results
We now summarize all the main results of this thesis. We begin by discussing two ap-
proaches to MFG, namely the Dynamic Programming Principle (DPP) and Stochastic
Maximum Principle (SMP), which are two different approaches to tackle a stochastic
optimal control problem. We then summarize our results on the wellposed-ness of
a general MFG with common noise and an asymptotic analysis when the common
noise is small. Lastly, we work out explicitly the linear-quadratic mean field games
(LQMFG) model.
CHAPTER 1. INTRODUCTION 20
1.2.1 Two approaches to mean field games
There are two general approaches to stochastic optimal control problems. We discuss
here briefly how to apply both of them to an MFG problem.
Dynamic Programming Principle
We begin by assuming that there exists an ε-MFG solution αε = (αεt )0≤t≤T ∈ H2([0, T ];R)
which is given in the feedback form
αεt = αε(t, Xε
t , mεt )
where Xεt = X αε
t , mεt = L(Xε
t |Ft), αε : [0, T ]× R× P2(R) is deterministic Lipschitz
function. Using this function αε, we know that mε = (mεt )0≤t≤T is a solution to the
stochastic Fokker-Planck equation
dmε(t, x) =
(
−∂x(αε(t, x,mε)mε) +
σ2 + ε2
2∂xxm
ε
)
dt− ε∂xmε dWs (1.22)
Given the stochastic flow of probability measures mε, we have a standard control
problem parametrized by mε, so we define the value function Vε : [0, T ]×R×P2(R)
by
Vε(t, x,m) = inf(αs)t≤s≤T
E
[∫ T
t
f(s,X t,x,αs , mε
s, αs)ds+ g(X t,x,αT , mε
T )∣
∣
∣X
t,x,αt = x,mε
t = m
]
(1.23)
Suppose that Vε is sufficiently regular, then we can use standard Principal of Opti-
mality and Ito’s formula to show that Vε satisfies
∂tVε(t, x,m) + H(t, x,m, ∂xV
ε(t, x,m)) +σ2 + ε2
2∂xxV
ε(t, x,m)
+ E0[
∂mVε(t, x,m)(X)(α(t, X,m))
]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(1.24)
CHAPTER 1. INTRODUCTION 21
with the terminal condition
Vε(T, x,m) = g(x,m)
where
H(t, x,m, y) , infα∈R
(αy + f(t, x,m, α)) (1.25)
is the Hamiltonian, X is a lifting random variable with law m, and ζ is a N (0, 1)-
random variable independent of X .
Remark 1.2.1. The Ito’s lemma used to derive (1.26) is non-standard as it involves
the derivative with respect to m. We refer to Proposition 6.5 in [18] for a full state-
ment and a proof of this Ito’s formula.
It will be useful to note here that under assumption A, there exist a unique
minimizer α : [0, T ]×R×P2(R)×R to (1.25). Thus, by a verification theorem for a
standard control problem, it can be shown that if Vε is a classical solution to (1.26),
then the optimal control is given by
αε(t, x,m) = α(t, x,m, ∂xVε(t, x,m))
But by definition of ε-MFG, this must be the same as αε. That is, we must have
αε(t, x,m) = αε(t, x,m) = α(t, x,m, ∂xVε(t, x,m))
Plugging this back in (1.24) yields
∂tVε(t, x,m) + H(t, x,m, ∂xV
ε(t, x,m)) +σ2 + ε2
2∂xxV
ε(t, x,m)
+ E0[
∂mVε(t, x,m)(X)(α(t, X,m, ∂xV
ε(t, X,m)))]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(1.26)
The equation (1.26) contains all the information about the ε-MFG, in a sense that
CHAPTER 1. INTRODUCTION 22
having a classical solution to (1.26) yields an ε-MFG solution in feedback form by
α(t, x,m) = α(t, x,m, ∂xVε(t, x,m))
See Proposition 4.1 in [18] for this verification-type result. For this reason, it is called
the master equation for MFG. However, solving this equation even numerically is
not feasible since it is a second-order non-linear infinite-dimensional HJB equation
involving derivatives with respect to a probability measure. We refer to [10, 18, 24]
for more discussion on the master equation.
Next, we proceed in a similar way as done by Lasry, Lions [13] for 0-MFG by
considering the value function along the trajectory of the equilibrium distribution
mε. We define a random function uε : [0, T ]× R× Ω by
uε(t, x) = Vε(t, x,mεt )
When ε > 0, uε is a stochastic value function which is Wt-path dependent. By Ito-
Kunita lemma, it can be shown that uε(t, x) satisfies the backward stochastic partial
differential equation (BSPDE)
duε(t, x) =
(
−H(t, x,mεt , ∂xu
ε(t, x))−σ2
2∂xxu
ε(t, x)−ε2
2
(
∂xxuε(t, x)− 2∂xv
ε(t, x))
)
dt
− εvε(t, x)dWt
where
vε(t, x) , E0[
∂mVε(t, x,mε
t )(X)1]
with terminal condition
uε(T, x) = g(x,mεT )
More details on the derivation of this BSPDE is provided in Section 2.1. Similarly,
solving this BSPDE yields an optimal feedback control;
αε(t, x) = α(t, x,mεt , ∂xu
ε(t, x))
CHAPTER 1. INTRODUCTION 23
Plugging this back in (1.22) yields a system of forward backward stochastic partial
differential equation (FBSPDE),
duε(t, x) =
(
−H(t, x,mε, ∂xuε)−
σ2
2∂xxu
ε −ε2
2∂xxu
ε − ε2∂xvε
)
dt− εvεdWt
dmε(t, x) =
(
−∂x(α(t, x,mε, ∂xu
ε)mε) +σ2 + ε2
2∂xxm
ε
)
dt− ε∂xmε dWt
(1.27)
with boundary conditions
mε(0, x) = m0(x) = L(ξ0), uε(T, x) = g(x,mεT )
Similar to the master equation above, there is a verification-type theorem (see Section
4.2 in [10]) which states that a sufficiently regular solution (uε, vε, mε) of (1.27) yields
ε-MFG solution in a feedback form by
αεt = α(t, Xε
t , mεt , ∂xu
ε(t, Xεt ))
Remark 1.2.2. As in the standard BSDE or BSPDE, vε is not specified and is part
of the solution to ensure adaptivity of uε. See Ch.1 in [49] for a brief introduction to
BSDE.
While this FBSPDE seems simpler than the master equation due to an absence
of the terms involving the derivative with respect to m, it is still difficult to solve
and to the best of our knowledge, there is no wellposedness result for FBSPDE on
unbounded domains. See [58] for a recent work on the case of bounded domains.
One can also view FBSPDE (1.27) simply as a different way to represent the
value function. As opposed to Vε which gives the minimum expected cost given
(t, x,m), uε describes the same quantity as a function of (t, x, ω), where ω is a common
Brownian motion path. Both of these equations are infinite-dimensional due to their
last arguments and are difficult to solve. While we have a verification-type theorem
for both equations, there is no existence theory for either of them and to solve these
equations numerically is difficult.
However, when ε = 0, we can reduce (1.27) to a system of fully-coupled finite
CHAPTER 1. INTRODUCTION 24
dimensional PDE. This can be seen by plugging ε = 0 in (1.27) which results in
∂tu0 = −H(t, x,m0, ∂xu
0)−σ2
2∂xxu
0, u0(T, x) = g(x,m0T )
∂tm0 = −∂x(α(t, x,m
0, ∂xu0)m0) +
σ2
2∂xxm
0, m0(0, x) = m0(x) = L(ξ0)
(1.28)
This is precisely the system of FBPDE that was introduced and studied by Lasry and
Lions [13, 42, 43, 44].
Stochastic Maximum Principle
SMP is an alternative approach to a stochastic control problem which studies opti-
mality conditions satisfied by an optimal control. It gives sufficient and necessary
conditions for the existence of an optimal control in terms of a backward stochastic
differential equation (BSDE) of an adjoint process. We summarize here how the SMP
can be applied to ε-MFG model. For a general control problem, we refer to [53, 61].
To state the SMP, we first introduce an adjoint process corresponding to an indi-
vidual control problem given m ∈ M([0, T ];P2(R)). Given a control α ∈ H2([0, T ];R)
and its corresponding controlled process X (see (1.2)), then the adjoint equation is
defined by
dYt = −∂xH(t, αt, Xt, mt, Yt)dt+ ZtdWt + ZtdWt, YT = ∂xg(XT ) (1.29)
where H : [0, T ]× Rd+3 → R is a generalized Hamiltonian defined as
H(t, a, x, y,m) , αy + f(t, x,m, a) (1.30)
Equation (1.29) is a backward stochastic differential equation (BSDE). We would
like to emphasize that the solution to (1.29) is a pair (Yt, Zt, Zt)0≤t≤T . The processes
(Zt, Zt)0≤t≤T arise naturally from the martingale representation theorem and are nec-
essarily part of a solution to ensure that an adapted solution (Yt)0≤t≤T can be found.
We refer the readers to Ch.1 in [49] for a brief introduction to BSDE.
We can now state a necessary and sufficient condition for an optimal control in
CHAPTER 1. INTRODUCTION 25
terms of the adjoint process. In short, the necessary condition (see Theorem 2.2.2)
states that if α is an optimal control, then the corresponding adjoint process is solvable
and the following maximum condition (or minimum in this case) holds
H(t, αt, Xt, mt, Yt) = mina∈R
H(t, a, Xt, mt, Yt), 0 ≤ t ≤ T, a.s. (1.31)
where (Yt)0≤t≤T is a solution to (1.29). The sufficient condition (see Theorem 2.2.4)
requires extra convexity assumptions on H, g, which hold under A, and the state-
ment is similar. It says that, given an optimal control α and its corresponding state
process X , if the adjoint BSDE is solvable and its solution (Yt)0≤t≤T satisfies the max-
imum condition (1.31), then α is an optimal control. In fact, under strict convexity
assumption (see (A3)), we have an estimate
J (β) ≥ J (α) + C
∫ T
0
|βt − αt|2dt, ∀β ∈ H2([0, T ];R)
This estimate implies that an optimal control, if exists, must be unique.
Under a strict convexity assumption on f , the generalized Hamiltonian has a
unique minimizer α : [0, T ]× R× P2(R)× R → R. It is then easy to check that
∂xH(t, α(t, x,m, y), x,m, y) = ∂xH(t, x,m, y)
where H is the Hamiltonian defined in (1.25). Consequently, we can plug this min-
imizer function into both the forward controlled process and the backward adjoint
process; as a result, we have a system of forward backward stochastic differential
equation (FBSDE)
dXt = α(t, Xt, mt, Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt, mt, Yt)dt+ ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT , mT )
(1.32)
We can now state the SMP for an individual control problem givenm ∈ M([0, T ];P2(R))
(recall Definition 1.1.2) in term of FBSDE (1.32) as follows;
CHAPTER 1. INTRODUCTION 26
Theorem 2.2.7. Assume A holds, then the individual control problem given (mt)0≤t≤T ∈
M([0, T ];P2(R)) has an optimal control if and only if FBSDE (2.17) is solvable. In
that case, the optimal control is given by
αt = α(t, Xt, mt, Yt)
for all t ∈ [0, T ], where (Xt, Yt, Zt, Zt)0≤t≤T is a solution to FBSDE (2.17).
Recall the definition of ε-MFG solution which says that given the stochastic flow
of probability measure mα corresponding to a control α, the optimal control of an
individual control problem given mα is again α. This definition is equivalent to the
following consistency condition
mαt = L(Xα
t |Ft)
Plugging this to (1.32), we have our main result for this section which establishes the
SMP for ε-MFG.
Theorem 2.2.8. Assume that A holds, then ε-MFG is solvable if and only if the
FBSDEdXt = α(t, Xt,L(Xt|Ft), Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt,L(Xt|Ft), Yt)dt+ ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT ,L(XT |FT ))
(1.33)
is solvable. In that case, ε-MFG solution is given by
αt = α(t, Xt,L(Xt|Ft), Yt), ∀t ∈ [0, T ]
Equation (1.33) is a McKean-Vlasov FBSDE. It was first introduced in [17] from
the 0-MFG problem. When there is no common noise, the conditional law L(Xt|Ft)
is simply the law L(Xt). In [17], Carmona and Delarue show that the McKean-Vlasov
FBSDE corresponding to a 0-MFG is solvable under assumptions similar to A, (B1),
plus what they call a weak mean reverting assumptions (see (1.17)). The proof is
based on Schauder fixed point theorem similar to Lasry and Lions’s existence proof
CHAPTER 1. INTRODUCTION 27
for the system (1.28). In the next section (Section 1.2.2), we discuss the existence
and uniqueness of the solution to both (1.32) and (1.33). The latter requires extra
weak monotonicity conditions which can be viewed as a stronger version of the weak
mean-reverting condition (1.17).
Connection to the DPP approach
The connection between Vε or (uε, vε, mε) from the DPP approach and a solution
(Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T of McKean-Vlasov FBSDE (1.33) from the SMP approach can
be stated simply through the relations
mεt = L(Xε
t |Ft)
Y εt = ∂xV
ε(t, Xεt , m
εt) = ∂xu
ε(t, Xεt )
Zεt = σ∂xxV
ε(t, Xεt , m
εt )
Zεt = ε
(
∂xxVε(t, Xε
t , mεt ) + E
0[
∂xmVε(t, Xε
t , mεt)(Xt)
])
(1.34)
To prove this relation, extra regularity for Vε(t, x,m) (C1,3,3) is needed and the proof
is based on applying Ito’s formula to the process (∂xVε(t, Xεt ,L(X
εt |Ft)))0≤t≤T . For
the full statement, we refer to Theorem 2.3.1.
1.2.2 Mean field games with common noise
In Section 1.2.1, we have shown that under assumption A, the ε-MFG problem is
equivalent to solving the McKean-Vlasov FBSDE (1.33). In this section, we summa-
rize our results on a wellposed-ness theory of this FBSDE, hence for a general MFG
with common noise. We will discuss the existence and uniqueness of a solution to
ε-MFG, a Markov property, and a connection to the HJB approach. Our main result
for this section is the following
Theorem 3.1.6 (Wellposedness of McKean-Vlasov FBSDE). Assume that A and B
CHAPTER 1. INTRODUCTION 28
hold, then there exist a unique solution (Xt, Yt, Zt, Zt)0≤t≤T to FBSDE (1.33) satisfy-
ing
E
[
sups≤t≤T
[X2t + Y 2
t ] +
∫ T
s
[Z2t + Z2
t ]dt
]
≤ C
(
E[ξ2] + (∂xg(0, δ0))2 +
∫ T
s
(α(t, 0, δ0, 0))2 + (∂xH(t, 0, δ0, 0))
2dt+ σ2 + ε2)
(1.35)
where δx denote the Dirac measure at x. Moreover, two solutions (X it , Y
it , Z
it , Z
it)s≤t≤T , i =
1, 2 to FBSDE (2.2.8) with initial ξi satisfies the estimate
E
[
sups≤t≤T
1A∆X2t + sup
s≤t≤T1A∆Y
2t +
∫ T
s
[1A∆Z2t + 1A∆Z
2t ]dt
]
≤ CK,TE[1A∆ξ2]
(1.36)
where ∆Xt = X1t − X2
t , ∆Yt,∆Zt,∆Zt,∆ξ are defined similarly, CK,T is a constant
depends only on K, T , A is an Fs-measurable set, and 1A denotes the indicator func-
tion.
In the existence proof of both Lasry and Lions PDE approach and Carmona
and Delarue probabilistic approach, they applied Schauder fixed point theorem to
establish the existence of a fixed point. However, this strategy cannot be extended
to the common noise case as the flow of probability measures given by the law of
the optimal state process is no longer deterministic. Instead, we have to deal with
a random flow of probability measures from a conditional law. Working with this
larger space, we cannot establish compactness which is necessary to apply Schauder
fixed point theorem in the same way.
As a result, we will use an alternative approach, namely the Banach fixed point
theorem. In the same way as in the proof of wellposedness of FBSDE, the Banach
fixed point theorem can be used to establish the existence of a solution when the
time duration T is sufficiently small. This method can usually be applied in that
case because the solution estimate depends on T , thus we can get a contraction map
when T is sufficiently small. See [5, 49] for a proof of existence and uniqueness of
a solution to FBSDE for a small time duration. However, the small time restriction
is not a desirable assumption for obvious reasons, so we wish to extend the solution
CHAPTER 1. INTRODUCTION 29
to an arbitrary time duration. To do so, we need an extra condition to be able to
control the Lipschitz constant of a new terminal condition as we move backwards in
time. This is precisely the weak monotonicity condition (B3).
In our subsequent analysis, we will see a similar type of FBSDE which requires a
similar technique involving a monotonicity property, so we state and prove the result
in a slightly more general framework. That is, we will consider the following FBSDE
over [s, T ]
dXt = b(t, Xt, Yt)dt+ σdWt + εdWt
dYt = F (t, Xt, Yt)dt+ ZtdWt + ZtdWt
Xs = ξ, YT = G(XT )
(1.37)
where ξ ∈ L2Fs
and b, F,G are measurable maps
b : [0, T ]×L2F × L2
F × Ω → R
F : [0, T ]× L2F ×L2
F × Ω → R
G : L2F × Ω → R
(1.38)
Notice that the map b, F,G are “functional” in a sense that their inputs are random
variables. In addition to standard measurability, Lipschitz, and linear growth condi-
tion, we need the following monotonicity condition which reads; For any t ∈ [0, T ],
X,X ′, Y, Y ′ ∈ L2FT
, and A ∈ Gt,
E [1A(F (t, X, Y )− F (t, X ′, Y ′))(X −X ′) + 1A(b(t, X, Y )− b(t, X ′, Y ′))(Y − Y ′)] ≤
−β1E[
1A(b(t, X, Y )− b(t, X ′, Y ′))2]
− β2E[
1A(Y − Y ′)2]
E [1A(G(X)−G(X ′))(X −X ′)] ≥ 0
(1.39)
for some constant β1, β2 ≥ 0 with β1 + β2 > 0. For the full statement of all the
assumptions, we refer to assumption H in Section 3.1.1. Then we have the following
result
Theorem 3.1.2. Let ξ ∈ L2Fs
and b, F,G be functionals satisfying assumption H, then
there exist a unique adapted solution (Xt, Yt, Zt, Zt)s≤t≤T to FBSDE (1.37) satisfying
CHAPTER 1. INTRODUCTION 30
the estimate
E
[
1A sups≤t≤T
X2t + 1A sup
s≤t≤TY 2t + 1A
∫ T
s
[Z2t + Z2
t ]dt
]
≤ CK,T
(
E
[
1Aξ2 + 1AG(0)
2 + 1A
∫ T
s
(
b(t, 0, 0)2 + F (t, 0, 0)2)
dt
]
+ σ2 + ε2)
(1.40)
for some constant CK,T depends only on K, T and for all A ∈ Gt.
Using assumption A,B, we can show that (see Theorem 3.1.6) b, F,G defined by
b(t, X, Y ) = α(t, X,L(X|Ft), Y ),
F (t, X, Y ) = −∂xH(t, X,L(X|Ft), Y ),
G(X) = ∂xg(X,L(X|FT ))
Gt0≤t≤T = Ft0≤t≤T
(1.41)
satisfies all the assumption in Theorem 3.1.2. As a result, we can apply Theorem
3.1.2 to the McKean-Vlasov FBSDE (1.33) which, combining with Theorem 2.2.8,
yields the existence and uniqueness result for an ε-MFG solution.
Markov property and a decoupling function
In this section, we discuss a Markov property of McKean-Vlasov FBSDE (1.33). A
classical FBSDE is said to have a Markov property if there exist a deterministic
function θ : [0, T ]× R → R such that
Yt = θ(t, Xt)
This function is often called a decoupling function of an FBSDE. By plugging it into
the FBSDE, it can be seen that θ is a solution to a quasilinear PDE. Solving this PDE,
one can decouple the system and reduce the problem to solving just the forward SDE.
This method of solving FBSDE is called Four-step scheme [46]. Markov property of
a classical FBSDE holds when the coefficients are deterministic.
CHAPTER 1. INTRODUCTION 31
Even though all the coefficients of McKean-Vlasov FBSDE (1.33) are determinis-
tic, it is not obvious that this Markovian property holds particularly in the case of
common noise. For a fixed m ∈ M([0, T ];P2(R)), we are in fact dealing with FBSDE
with non-deterministic coefficients. Specifically, we have a path dependent functions
∂xH(t, x,mt(ω), y), g(x,mT (ω))
However, as f, g are still deterministic functions of m, it is reasonable to expect a
Markov property if we include the current distribution of players as an additional
input, or in FBSDE context, the conditional distribution of the state process. Our
main result for this section is the following theorem
Theorem 3.2.1. Let (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T denote the solution to McKean-Vlasov
FBSDE (1.33), then there exist a deterministic function Uε : [0, T ]×R×P2(R) such
that
Y εt = Uε(t, Xε
t ,L(Xεt |Ft)) (1.42)
Moreover, Uε satisfies the estimates
1. |Uε(t, x,m)− Uε(t, x′, m′)| ≤ CK,T (|x− x′|+W2(m,m′))
2. (Uε(t, x,m)− Uε(t, x′, m)) (x− x′) ≥ 0
for all t ∈ [0, T ], x, x′ ∈ R, m,m′ ∈ P2(R) where CK,T depends only on K, T .
We explain briefly here how Uε(s, x,m) is defined; Fix (s, x,m) ∈ [0, T ] × R ×
P2(R). First, we solve McKean-Vlasov FBSDE over [s, T ] using initial Xs = ξ with
L(ξ) = m to get a solution (Xs,mt , Y
s,mt , Z
s,mt , Z
s,mt )s≤t≤T , then we define a stochastic
flow of probability measure (ms,mt )s≤t≤T ∈ M([s, T ];P2(R)) by ms,m
t = L(Xs,mt |F s
t )
where F st = σ(Wr − Ws; s ≤ r ≤ t). Intuitively, ms,m
t represents the distribution of
players at time t under ε-MFG solution strategy given that the distribution at time
s is m. Then we consider an individual control problem over [s, T ] given (ms,mt )s≤t≤T
with the initial position Xs = x by solving the corresponding FBSDE (1.32) to get a
solution (Xs,x,mt , Y
s,x,mt , Z
s,x,mt , Z
s,x,mt )s≤t≤T . We then define Uε(s, x,m) to be Y s,x,m
s
which is F ss -measurable, hence deterministic. See Section 3.2 for more details.
CHAPTER 1. INTRODUCTION 32
Having defined Uε, we are left to show that it is indeed a decoupling function, i.e.
(1.42) holds. The proof is based on a discretizing argument and a priori estimate of
Uε. See Theorem 3.2.12 for the full statement and its proof.
Connection to DPP approach
In the previous section, we discussed the relation between Vε, uε and the solution
(Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T of FBSDE (1.33). Having introduced the deterministic decou-
pling function Uε, we can restate the connection between the two approaches through
this function. Comparing (1.34) and (1.42), we have
Uε(t, x,m) = ∂xVε(t, x,m) (1.43)
Consequently, we can deduce other relations which are summarized below
mεt = L(Xε
t |Ft)
Y εt = Uε(t, Xε
t , mεt ) = ∂xV
ε(t, Xεt , m
εt ) = ∂xu
ε(t, Xεt )
Uε(t, x,mεt ) = ∂xV
ε(t, x,mεt ) = ∂xu
ε(t, x)
Uε(t, x,m) = ∂xVε(t, x,m)
E0[
∂mUε(t, x,mε
t )(X)1]
= E0[
∂xmVε(t, x,mε
t )(X)1]
= ∂xvε(t, x)
(1.44)
Using the relation (1.43) and the master equation (1.26) for Vε, the equation for Uε
can be derived easily and is given by
∂tUε(t, x,m) + ∂xH(t, x,m,Uε(t, x,m)) + ∂yH(t, x,m,Uε(t, x,m))∂xU
ε(t, x,m)
+σ2 + ε2
2∂xxU
ε(t, x,m)− E0[
∂mUε(t, x,m)(X)α(t, X,m,Uε(t, X,m))
]
+σ2
2∂mmU
ε(t, x,m)(X)[ζ, ζ ] +ε2
2∂mmU
ε(t, x,m)(X)[1, 1]
+ ε2E0[
∂xmUε(t, x,m)(X)1
]
= 0
(1.45)
Similar to Vε and its master equation, we have a verification theorem which says
that a classical solution to (1.45) is a decoupling function to McKean-Vlasov FBSDE
CHAPTER 1. INTRODUCTION 33
(1.33). See Theorem 3.3.2 for the full statement.
1.2.3 Asymptotic analysis of mean field games
Recently, there has been progress in studying the MFG with common noise concerned
mostly with the general well-posedness theory. See [3, 20, 10, 18]. Despite this
progress, a general common noise model is still difficult and impractical to solve
numerically or analytically as it does not enjoy the dimension reduction property
as in the case of MFG without common noise. When the common noise is small,
however, it is reasonable to seek an approximate solution using only finite-dimensional
information from the ε = 0 problem.
Let (αεt , X
εt )0≤t≤T be an ε-MFG solution, and its corresponding state process, we
are interested in the following ε-expansion
αεt = α0
t + εδαt + o(ε), Xεt = X0
t + εδXt + o(ε)
Through the SMP, this is equivalent to studying the limit as ε→ 0 of
Xεt −X0
t
ε,
Y εt − Y 0
t
ε(1.46)
where (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T denote the solution to McKean-Vlasov FBSDE (1.33).
We now state additional regularity assumption; From now and throughout the rest
of the section, we assume, in addition to A,B, that C holds.
(C1) ∂xxf, ∂xmf, ∂xxg, ∂xmg exist and are continuous and bounded. Denote their
bounds by the same constant K.
To reduce notations, particularly those from the Hamiltonian H and its deriva-
tives, we will assume in this section that
f(t, x,m, α) =α2
2
CHAPTER 1. INTRODUCTION 34
while keeping a general terminal cost function g. The same result still holds for a
general running cost f satisfying A,B,C. Note that with this running cost, it follows
that
α(t, x,m, y) = −y, H(t, x,m, y) = −y2
2
so that the McKean-Vlasov FBSDE (1.33) reads
dXt = −Ytdt+ σdWt + εdWt
dYt = ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT ,L(XT |FT ))
(1.47)
By formally taking the limit (1.46) through (1.47), we get the following linear varia-
tional mean-field FBSDE
dUt = −Vtdt+ dWt
dVt = QtdWt + QtdWt
U0 = 0, VT = ∂xxg(X0T , m
0T )UT + E[∂xmg(X
0T , m
0T )(X
0T )UT ]
(1.48)
where
m0t = L(X0
t |Ft) = L(X0t )
and X0 and U are identical copies ofX0 and U in (Ω, F , P) and E0[·] is the expectation
with respect to ω0 only. We can write the terminal function explicitly as
E0[∂xmg(X
0T , m
0T )(X
0T )UT ] =
∫
Ω0
∂xmg(X0T (ω
0), m0T )(X
0T (ω
0))UT (ω0, ω)dP0(ω0)
where we suppress the ω in X0T , X
0T as they do not depend on it. We can see that the
term E0[∂xmg(X0T , m
0T )(X
0T )UT ] is a mean-field term that couples UT (ω, ·) for different
ω ∈ Ω0 together. Intuitively, each different path ω ∈ Ω0 represents different players
and the mean-field term represents the effect on a single player from all other players.
First, by applying Theorem 3.1.2, we can show that this FBSDE is uniquely solvable
(see Theorem 4.1.1). We are now ready to state our first main theorem which gives
a convergence result of the limit formally taken above.
CHAPTER 1. INTRODUCTION 35
Theorem 4.1.2. Assume A,B,C hold, for all ε > 0, let (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T de-
note the solution to McKean-Vlasov FBSDE (1.33) corresponding to ε-MFG and
(Ut, Vt, Qt, Qt)0≤t≤T denote the solution to (1.48), then there exist a constant CK,T
dependent only on K, T such that
E sup0≤t≤T
[
(
Xεt −X0
t
ε− Ut
)2
+
(
Y εt − Y 0
t
ε− Vt
)2]
≤ CK,Tε2 (1.49)
Approximate Nash equilibrium
Using the result above, we can construct a first order approximation for ε-MFG by
taking
βεt , α0
t − εVt (1.50)
for all t ∈ [0, T ]. Being a game, an appropriate notion must be used to check that
β is a good approximate strategy. This notion is what is called δ-approximate Nash
equilibrium. An admissible strategy α = (αt)0≤t≤T ∈ H2([0, T ];R) is called a δ-Nash
equilibrium for ε-MFG if
J ε(α|mα) ≤ J ε(β|mα) + δ
for all β = (βt)0≤t≤T where J ε(·) denote the cost function and (mαt )0≤t≤T denote
the law conditional on the common noise of a player adopting strategy α. From this
definition, a 0-approximate Nash equilibrium is an ε-MFG solution.
This notion of approximate Nash equilibrium has applications in many areas in-
cluding infinite horizon stochastic games or algorithmic game theory. In those cases,
an exact Nash equilibrium either does not exist or is computationally expensive. In
MFG, recall that the motivation for considering this model is its application for find-
ing a good approximate strategy for an N -player game when N is large. This notion
is used mainly in the study of this approximation. See [17, 22, 37, 39]. In these
works, however, we are only concerned with the model at the continuum limit. That
is, we are more interested in the approximate solution for ε-MFG using information
available from the 0-MFG solution. Our main result for this section is the following;
CHAPTER 1. INTRODUCTION 36
Theorem 4.2.5. Assume A,B,C hold. For ε > 0, let αε = (αεt )0≤t≤T denote the
solution to ε-MFG and (Ut, Vt, Qt, Qt)0≤t≤T denote the solution to the linear variation
FBSDE (4.5). Define an approximate strategy βε = (βεt )0≤t≤T by
βεt , α0
t − εVt (1.51)
Then βε is an ε2-Nash equilibrium for ε-MFG.
Gaussian properties of (Ut, Vt) and decoupling function
In this section, we would like to discuss some properties of (Ut, Vt)0≤t≤. The first
property is that (Ut, Vt)0≤t≤ is a Gaussian process. This follows from the fact that
the FBSDE (1.48) is linear with respect to common noise. Furthermore, using the
weak monotonicity condition and the fact that the initial U0 = 0, it can be shown
that their means are zero (see Theorem 4.3.1).
To compute the covariance function, however, the standard approach for a linear
SDE involving taking Ito’s lemma is no longer applicable due to the fully-coupled for-
ward backward structure, particularly due to the presence of the terms (Qt, Qt)0≤t≤T .
As a result, we need to resort to the decoupling function of (1.48). Recall that we
have the decoupling function Uε for McKean-Vlasov FBSDE (1.33) which satisfies
the relation
Y εt = Uε(t, Xε
t ,L(Xεt |Ft))
The following theorem is crucial in obtaining the decoupling function for FBSDE
(1.48)
Theorem 4.3.2. Let Uε denote the decoupling function of FBSDE (1.33) as defined
in (1.42), then the following holds;
limε→0
Uε(t, x,m)− U0(t, x,m)
ε= 0 (1.52)
uniformly in (t, x,m) ∈ [0, T ]× R× P2(R).
CHAPTER 1. INTRODUCTION 37
Remark 1.2.3. The theorem above implies that to approximate the ε-MFG solu-
tion at the first order, we simply need to use 0-MFG solution along the trajectory
(t, Xεt ,L(X
εt |Ft)). However, we would like to emphasize that we do not usually know
U0(t, x,m) for all (t, x,m) since that would require us to solve the master equation
(2.8) or (3.48) which is an infinite-dimensional problem and clearly non-trivial to do
so.
While there is no result on the regularity of Uε and all the discussions so far has
been formal [10, 18, 13], there has been a recent progress for the original no common
noise model. In a recent paper by Chassagneux et al.[24], they showed the existence
and uniqueness of a classical solution to the master equation (3.48) when ε = 0. As
a by-product, we have that U0 is continuously differentiable in (x,m). To apply this
result, we need assumption D which provides additional regularity assumption on the
cost functions involving second derivative with respect to m-argument. As a result,
we have the decoupling function for linear variational FBSDE (4.5).
Proposition 4.3.3. Assume that A,B,C,D holds. Let (Ut, Vt, Qt, Qt)0≤t≤T denote
the unique solution to FBSDE (4.5), then
Vt = ∂xU0(t, X0
t , m0t )Ut + E
0[∂mU0(t, X0
t , m0t )(X
0t )Ut] (1.53)
Explicit solution and covariance function
From (1.48) and (1.53) above, we see that the functions U0(t, x,m0t ), ∂xU
0(t, x,m0t )
and E0[∂mU0(t, X0t , m
0t )(X
0t )Ut] are essential terms in our asymptotic analysis. The
first two terms are simply ∂xu0(t, x), ∂xxu
0(t, x) where (u0, m0) is the solution to Lasry
and Lions’ FBPDE (2.10). We let
η(t, x, ω) , E0[∂mU
0(t, x,m0t )(X
0t )Ut]
To proceed and further analyze the linear mean-field FBSDE (4.5), particularly
the random function η(t, x), an extra regularity assumption on the decoupling func-
tion of 0-MFG U0 is required to derive the equations for η(t, x) and compute the
CHAPTER 1. INTRODUCTION 38
covariance functions. While there are some results that give sufficient conditions for
the differentiability with respect to (t, x) when m = m0t , since it is directly related
to u0(t, x) from Lasry and Lions’s FBSDE, the only work that provides similar result
for the derivative with respect to m-argument for U0 is [24]. However, their result
is still not sufficient for our application below. So going forwards, we will proceed
formally assuming that U0 is sufficiently regular with bounded derivatives. Our goal
is to derive the equation for η(t, x) and to compute the covariance function of the
process (Ut, Vt)0≤t≤T .
By using Ito’s lemma, we can show that η(t, x) satisfies the SPDE
dη(t, x) =
[
η(t, x)∂xU0(t, x,m0
t ) + ∂xη(t, x)U0(t, x,m0
t )−σ2
2∂xxη(t, x)
]
dt− w(t, x)dWt
(1.54)
with initial condition η(0, x) = 0 and w : [0, T ]× R → R is given by
w(t, x) , E0[
∂mU0(t, x,m0
t )(X0t )]
To fully describe the system, we are left to find w(t, x), which requires us to compute
∂mU0(t, x,m0t )(·). Again, by using Ito’s lemma and the master equation for U0 (3.48),
it can be shown that the resulting derivative satisfies an infinite-dimensional Riccati-
type equation. More precisely, w(t, x) is given by
w(t, x) = E0[
∂mU0(t, x,m0
t )(X)]
=
∫
R
h(t, x, z)m0(t, z)dz
for all (t, x) ∈ [0, T ]× R, where h(t, x, z) satisfies
∂th(t, x, z) = h(t, x, z)∂xxu0(t, x) + ∂xh(t, x, z)∂xu
0 −σ2
2∂xxh(t, x, z)
+
∫
R
h(t, x, u)h(t, u, z)m0(t, u)du+ ∂z(h(t, x, z)∂zu0(t, z))−
σ2
2∂zzh(t, x, z)
(1.55)
where the equation is to be interpreted weakly in the z-argument. We provide more
details of all these calculation in Section 4.3.2.
We have described the decoupling function for linear variational FBSDE (1.48)
CHAPTER 1. INTRODUCTION 39
explicitly in term of η(t, x) and w(t, x) defined above. In addition, we know that
η(t, x) satisfies a linear SPDE and w(t, x) can be expressed in term of h(t, x, z), a
solution to a deterministic partial integro-differential equation (1.55). Being a linear
SPDE, we have an explicit solution of (1.54) by mean of Duhamel’s principle. That
is,
η(t, x) =
∫ t
0
ψs(t, x)dWs
where ψs is a solution to the PDE
∂tφ(t, x) = ∂xφ(t, x)U0(t, x,m0
t ) + φ(t, x)∂xU0(t, x,m0
t )−σ2
2∂xxφ(t, x), (t, x) ∈ [s, T ]× R
φ(s, x) = w(s, x), x ∈ R
(1.56)
Using this and the relation (1.53) above, we can derive an explicit solution to FBSDE
(1.48) as follows; First, we find w by solving the PDE (1.55), then solve of η by
solving for ψ satisfying (1.56). Next, we plug in η in the forward SDE of Ut and we
can easily solve for Ut by using integrating factor. We then use the relation (1.53) to
get Vt explicitly.
Similarly, by using Ito’s lemma, we can fully describe the covariance function for
Ut, Vt in terms of ∂xxu0, a solution from 0-MFG system (2.10), and ψs, a solution to
PDE (4.22). We omit the detail here and refer the readers to Section 4.4.2 for all the
equations.
1.2.4 Linear quadratic mean field games with common noise
We consider a linear quadratic model with controlled process
dXt = αtdt + σdWt + εdWt, X0 = ξ0
CHAPTER 1. INTRODUCTION 40
and cost functions
f(t, a, x,m) =1
2
(
qx2 + α2 + q(x− sm)2)
g(x,m) =1
2
(
qTx2 + (x− sT m)2qT
)
,
Under this LQ model, we can solve for the generalized value function Vε by assuming
that it takes a quadratic form. As a result, we get that
Vε(t, x,m) =1
2p(t)x2 + q(t)xm+
1
2r(t)m2 + s(t)
where p(t), q(t), r(t), s(t) : [0, T ] → R satisfies
p′(t) = p2(t)− q − q, p(T ) = qT + qT
q′(t) = 2p(t)q(t) + q2(t) + qs, q(T ) = −sT qT
r′(t) = 2(p(t) + q(t))r(t) + q(t)2 − qs2, r(T ) = s2T qT
s′(t) = −1
2(σ2 + ε2)p(t)−
1
2ε2r(t)− ε2q(t), s(T ) = 0
(1.57)
Consequently, all other functions including the solutions to FBSPDE, McKean-
Vlasov FBSDE, and linear variational FBSDE can be found explicitly as they are
directly related to Vε. The resulting functions are summarized below, while more
detail on the calculation are provided in Chapter 5. Let
a(t) =σ2
2
∫ T
t
p(s)ds, b(t) =
∫ T
t
(
1
2(p(s) + r(s)) + q(s)
)
ds (1.58)
CHAPTER 1. INTRODUCTION 41
then from the DPP approach, we have
Vε(t, x,m) =1
2p(t)x2 + q(t)xm+
1
2r(t)m2 + a(t) + ε2b(t)
uε(t, x) =1
2p(t)x2 + q(t)x
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)
+1
2r(t)
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)2
+ a(t) + ε2b(t)
vε(t, x) = q(t)x+ r(t)
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)
From the SMP approach, we have
Uε(t, x,m) = p(t)x+ q(t)m
Y εt = p(t)Xε
t + q(t)E[Xεt |Ft]
Zεt = σp(t)
Zεt = −εq(t)(p(t) + q(t))E[Xε
t |Ft]
where (Xεt )0≤t≤T solves
dXεt = (−p(t)Xε
t − q(t)E[Xεt |Ft])dt+ σdWt + εWt, Xε
0 = ξ0
For the asymptotic analysis, we have
Vt = (p(t) + q(t))Ut
where (Ut)0≤t≤T solves
dUt = −(p(t) + q(t))Utdt+ dWt, U0 = 0
Consequently, (Ut, Vt)0≤t≤T are Gaussian processes with mean zero and the following
CHAPTER 1. INTRODUCTION 42
covariance functions
dE[U2t ]
dt= −2(p(t) + q(t))E[U2
t ] + 1, E[U20 ] = 0
E[UtVt] = (p(t) + q(t))E[U2t ]
E[V 2t ] = (p(t) + q(t))2E[U2
t ]
Lastly, η, w, h are given by
η(t, x) = q(t)Ut = q(t)
∫ t
0
e−∫ t
s(p(r)+q(r))drdWs
w(t, x) = h(t, x, z) = ∂mU0(t, x,m0
t )(z) = q(t)
Note that the functions from the asymptotic analysis of LQMFG are trivial. This
result is mainly due to the fact that the second derivative of f, g in LQ model are
constant which results in the second derivative of the value function and related
functions being constant.
1.3 Summary of contributions
Motivated by applications in finance and economics, we are interested in an MFG
model with common noise which extends the original MFG model proposed by Lasry
and Lions. We operate under a linear-convexity framework and what we call the weak
monotonicity assumption (see (B3)) on the cost functions. Our contributions can be
summarized as follows;
• Wellposed theory. We show the existence and uniqueness of a strong solu-
tion to an MFG model with common noise using the SMP approach. That is,
we prove that the McKean-Vlasov FBSDE corresponding to ε-MFG is uniquely
solvable (Theorem 3.1.6) and its solution yields an ε-MFG solution (Theorem
2.2.8). Our main assumption, in addition to standard assumptions, is the weak
monotonicity condition which is stronger than the weak-mean reverting assump-
tion used to show the existence of a 0-MFG solution and is weaker than Lasry
CHAPTER 1. INTRODUCTION 43
and Lions’ monotonicity assumption used to show the uniqueness (Proposition
1.1.8). As a by-product, we give a more general uniqueness result for 0-MFG.
• Markov property. We show that the ε-MFG solution is in the feedback form.
This is done by proving the existence of a deterministic decoupling function of
the corresponding McKean-Vlasov FBSDE. Our main result is Theorem 3.2.1.
• Asymptotic analysis. We obtain the first order expansion of the ε-MFG
solution which is characterized as the solution to a linear FBSDE of mean-field
type (Theorem 4.1.2). We then show that the first order approximate solution is
ε2-approximate Nash equilibrium (Theorem 4.2.5). Furthermore, we show that
the solution is a Gaussian process with mean zero (Theorem 4.3.1) and find
the decoupling function for this FBSDE (Proposition 4.3.3) in terms of U0, the
decoupling function for 0-MFG. We then formally derive, assuming regularity
on U0, the explicit solution and compute the covariance functions (Section 4.4).
All the results in this chapter are specialized to the case of quadratic running
cost for simplicity, although the same method can be applied to general running
cost functions satisfying convexity and weak monotonicity assumptions.
Chapter 2
Two approaches to mean field
games
There are two main approaches to a stochastic optimal control problem, namely
Dynamic Programming Principle (DPP) and Stochastic Maximum Principle (SMP).
The DPP is a PDE approach which involves the study of an HJB equation for a
value function. The SMP is a probabilistic approach which involves the backward
stochastic differential equation (BSDE) of an adjoint process.
In this section, we describe how to apply these two approaches to an MFG model
in the presence of common noise. We begin with the DPP approach in Section 2.1
and the SMP approach in Section 2.2.4. We then discuss the relation between the
two approaches in Section 2.3.
2.1 Dynamic Programming Principle (DPP)
The DPP approach to a continuous-time stochastic control problem is the study of
its value function which describes the optimal cost that can be obtained as a function
of the current time and state. By exploiting the Markovian structure of the problem,
we can derive formally the PDE satisfied by the value function. For a stochastic
control problem where the controlled process is a diffusion process, this method gives
a second-order non-linear PDE called Hamilton-Jacobi-Bellman (HJB) equation. We
44
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 45
can show that a classical solution of this equation gives an optimal cost and, as a by-
product, the optimal policy for each state. This type of result is called a verification
theorem. We refer to Ch.3 of [28] for more details on the DPP approach to a stochastic
control problem.
Applying this approach to an MFG problem is non-trivial since the cost depends
on the distribution (mt)0≤t≤T which is governed by other players’ decisions. How-
ever, if each player has the same control problem, it is reasonable to expect that, at
equilibrium, the optimal control of each player is identical and is given in a feedback
form, i.e. as a function of his/her own state Xt and the distribution of the players mt.
Let us suppose that there exist a unique solution αε ∈ H2([0, T ];R) to the ε-MFG
problem which is of the form
αεt = αε(t, Xε
t ,L(Xεt |Ft))
whereXεt := X αε
t and αε : [0, T ]×R×P2(R) → R is a deterministic Lipschitz function.
The feedback form and Lipschitz property of an optimal control does indeed hold
under assumption A, a fact that will be proven in Section 2.2.4 through the SMP
approach.
Next, let mεt := mαε
t denote the corresponding conditional law, that is,
mεt = L(Xε
t |Ft)
then we have that mε = (mεt )0≤t≤T is a weak solution of the following stochastic
Fokker-Planck equation
dmε(t, x) =
(
−∂x(αε(t, x,mε)mε) +
σ2 + ε2
2∂xxm
ε
)
dt− ε∂xmε dWs (2.1)
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 46
Now we can define the generalized value function Vε : [0, T ]× R× P2(R) by
Vε(t, x,m) = inf(αs)t≤s≤T∈H2([t,T ];R)
E
[∫ T
t
f(s,X t,x,αs , mε
s, αs)ds+ g(X t,x,αT , mε
T )∣
∣
∣X
t,x,αt = x,mε
t = m
]
= E
[∫ T
t
f(s,Xεs , m
εs, α
ε(s,Xεs , m
εs))ds+ g(Xε
T , mεT )∣
∣
∣Xε
t = x,mεt = m
]
(2.2)
where
dX t,x,αs = αsds+ σdWs + εdWs, X
t,x,αt = x, s ∈ [t, T ]
Given that the αε is a solution to ε-MFG, the value function above represents the
minimum expected cost from t to T given the state of the game at time t. By the
Principal of Optimality, we expect to have
Vε(t, x,m) = infα∈H2([t,t+h];R)
E
[∫ t+h
t
f(s,X t,x,αs , mt,m
s , αs)ds+ Vε(t + h,Xt,x,αt+h , m
t,mt+h)
]
(2.3)
where
dmt,ms =
(
−∂x(αε(s, x,mt,m
s )mt,ms ) +
σ2 + ε2
2∂xxm
t,ms
)
dt−ε∂xmt,ms dWs, m
t,mt = m
or equivalently,
mt,ms = L(X t,m
s |Fs)
dX t,ms = αε(s, X t,m
s , mt,ms )ds+ σdWs + εdWs, X
t,mt = ξ
Here ξ is independent of Ft with law m, (Ws)t≤s≤T is a Brownian motion independent
of all other Brownian motion, (X t,ms )t≤s≤T is the dynamics of a typical or representa-
tive player under the ε-MFG solution, and (mt,ms )t≤s≤T represents the corresponding
distribution of players with distribution m at time t.
Suppose that Vε is sufficiently regular, then we can use Ito’s formula in (2.3) to
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 47
show that Vε satisfies
∂tVε(t, x,m) + H(t, x,m, ∂xV
ε(t, x,m)) +σ2 + ε2
2∂xxV
ε(t, x,m)
+ E0[
∂mVε(t, x,m)(X)(α(t, X,m))
]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(2.4)
with the terminal condition
Vε(T, x,m) = g(x,m)
where
H(t, x,m, y) , infα∈R
(αy + f(t, x,m, α)) (2.5)
is the Hamiltonian, X is a lifting random variable, i.e. L(X) = m, and ζ is a N (0, 1)-
random variable independent of X .
Remark 2.1.1. The derivative with respect to the m-argument is based on the prob-
abilistic framework proposed by Lasry and Lions in [13]. See Section 1.1.3 for more
detail. Also, an extended version of Ito’s lemma is necessary to incorporate this notion
of derivative. This result is shown in [18], see specifically Proposition 6.5.
When f is strictly convex in α (see (1.9)), then there is a unique minimizer to
(2.5) which we will denote by
α(t, x,m, y) = argminα∈R
(αy + f(t, x,m, α)) (2.6)
From the equation (2.4) above, the optimal control in feedback form is given by
αε(t, x,m) = α(t, x,m, ∂xVε(t, x,m)) (2.7)
By the definition of ε-MFG, we are searching for an optimal control such that given
a control (αεt )0≤t≤T and its corresponding stochastic flow of distribution (mαε
t )0≤t≤T ,
the optimal control for an individual problem given (mαε
t )0≤t≤T is again (αεt )0≤t≤T .
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 48
Equivalently, we would like to find αε such that the following holds
αε(t, x,m) = αε(t, x,m) = α(t, x,m, ∂xVε(t, x,m))
Plugging in this relation back in (2.4), it follows that Vε(t, x,m) must satisfy
∂tVε(t, x,m) + H(t, x,m, ∂xV
ε(t, x,m)) +σ2 + ε2
2∂xxV
ε(t, x,m)
+ E0[
∂mVε(t, x,m)(X)(α(t, X,m, ∂xV
ε(t, X,m)))]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(2.8)
with terminal condition
Vε(T, x,m) = g(x,m)
Equation (2.8) is often referred to as the master equation. It is an infinite-
dimensional non-linear second order HJB equation involving derivative with respect
to a probability measure. It was first introduced by Lasry and Lions in a heuristic
fashion and was discussed more extensively in [10, 18, 24]. In [10], Bensoussan de-
scribed the master equation through the HJB approach, while in [18], Carmona et
al. viewed the master equation through the decoupling function of McKean-Vlasov
FBSDE associated with the ε-MFG. This is related to Section 2.3 where we discuss
the relation between DPP and SMP. Recently, Delrarue et al. in [24] showed the ex-
istence of a classical solution to the master equation. The result was obtained under
a no common noise model and with monotonicity-type assumptions.
We have derived this equation by assuming the existence of an ε-MFG solution,
its Markovian property, and the regularity of the value function. We now state the
verification theorem which says that if there exist a classical solution to this master
equation satisfying an integrability condition, then the ε-MFG solution is indeed given
in the feedback form as
αεt = α
(
t, Xεt ,L(X
εt |Ft), ∂xV
ε(t, Xεt , m
εt ))
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 49
Let C1,2,2([0, T )× R× P2(R)) denote the space of real-valued function h on [0, T )×
R× P2(R) whose partial derivatives ∂th, ∂xxh, ∂mmh, ∂xmh exist and are continuous,
C([0, T ]×R×P2(R)) denote the space of continuous real-valued function on [0, T ]×
R×P2(R), and L2m(R) denote the space of measurable function h such that
∫
R
h2(x)dm(x) <∞
Theorem 2.1.2. Suppose Vε is a function in C1,2,2([0, T )×R×P2(R))∩C([0, T ]×
R×P2(R)) satisfying the master equation (2.8) with terminal condition Vε(T, ·) = g.
Suppose further that Vε satisfies
|∂xVε(t, x,m)|+ ‖∂mV
ε(t, x,m)(·)‖L2m(R) ≤ C
(
1 + |x|+
∫
R
y2dm(y)
)
for any (t, x,m) ∈ [0, T ]× R× P2(R) and there exist a unique solution to the SDE
dXt = α(t, Xt,L(Xt|Ft), ∂xVε(t, Xt,L(Xt|Ft)))dt+ σdWt + εdWt, X0 = ξ0
then
(αεt )0≤t≤T = (α(t, Xε
t , mεt , ∂xV
ε(t, Xεt , m
εt )))0≤t≤T
is a solution to ε-MFG with initial ξ0.
For the proof, we refer to Proposition 4.1 in [18]. We would like to note that
while this master equation contains all the information about the ε-MFG in the
same way that the classical HJB equation contains all the information about the
control problem, unlike the classical HJB equation, solving the master equation even
numerically is not feasible as we need to discretize the P2(R) space.
Next, we proceed in a similar way as done by Lasry and Lions in [13] for the
case of no common noise. That is, we consider the value function along the path of
the optimal distribution. Recall that we assume that there exists a unique solution
(αεt )0≤t≤T to the ε-MFG and let (mε
t )0≤t≤T ∈ M([0, T ];P2(R)) be the corresponding
conditional law given by the stochastic Fokker-Plank equation (2.1). We consider the
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 50
value function along (mεt )0≤t≤T by letting
uε(t, x) , Vε(t, x,mεt )
Using Ito-Kunita formula and the master equation (2.8), it follows that
duε(t, x) =
(
∂tVε(t, x,mε
t )− E0[
∂mVε(t, x,mε
t )(X)(α(t, X,mεt , ∂xV
ε(t, X,mεt )))]
+σ2
2∂mmV
ε(t, x,mεt )(X)[ζ, ζ ] +
ε2
2∂mmV
ε(t, x,mεt )(X)[1, 1]
)
dt
− εE0[
∂mVε(t, x,mε
t )(X)1]
dWt
=
(
− H(t, x,mεt , ∂xV
ε(t, x,mεt ))−
σ2
2∂xxV
ε(t, x,mεt )−
ε2
2∂xxV
ε(t, x,mεt )
− ε2E0[
∂xmVε(t, x,mε
t )(X)1]
)
dt− εE0[
∂mVε(t, x,mε
t )(X)1]
dWt
=
(
− H(t, x,mεt , ∂xu
ε(t, x))−σ2
2∂xxu
ε(t, x)−ε2
2∂xxu
ε(t, x)
− ε2E0[
∂xmVε(t, x,mε
t )(X)1]
)
dt− εE0[
∂mVε(t, x,mε
t )(X)1]
dWt
=
(
−H(t, x,mεt , ∂xu
ε(t, x))−σ2
2∂xxu
ε(t, x)−ε2
2
(
∂xxuε(t, x)− 2∂xv
ε(t, x))
)
dt
− εvε(t, x)dWt
where
vε(t, x) , E0[
∂mVε(t, x,mε
t )(X)1]
Combining with (2.1) and (2.7), we have arrived at a system of forward backward
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 51
stochastic partial differential equations (FBSPDE)
duε(t, x) =
(
−H(t, x,mε, ∂xuε)−
σ2
2∂xxu
ε −ε2
2∂xxu
ε − ε2∂xvε
)
dt− εvεdWt
dmε(t, x) =
(
−∂x(α(t, x,mε, ∂xu
ε)mε) +σ2 + ε2
2∂xxm
ε
)
dt− ε∂xmε dWt
(2.9)
with boundary conditions
mε(0, x) = m0(x) = L(ξ0), uε(T, x) = g(x,mεT )
Recall that H denotes the Hamiltonian (see (2.11)) and α denotes the minimizer of
H (see (2.6)).
Similarly to Theorem 2.1.2, we have a verification theorem for (2.9) which states
that if we have a sufficiently regular solution (uε, mε, vε) to the FBSPDE (2.9) above,
then the ε-MFG solution is given in a feedback form as
αεt = α(t, Xε
t , mεt , ∂xu
ε(t, Xεt )),
We refer the reader to Section 4.2 in [10] for this result. The triple (uε, mε, vε) then
gives, respectively, the value function, distribution of the optimal state process, and
the sensitivity of the value function with respect to a spatial shift of the distribution
process. Consequently, despite the fact that (2.9) above is derived from a solution of
the master equation, it actually contains the same information as the master equation.
The function Vε represents the value function at time t as a function of (x,m) while
uε represents the value function at time t as a function of (x, ω) where ω is a common
Brownian motion path.
Remark 2.1.3. As mentioned earlier, the FBSPDE above is simply a different rep-
resentation of a value function and is not necessarily easier to deal with than the
master equation. Nevertheless, by considering (uε, mε), we can avoid working with a
derivative with respect to a probability measure. Also, it might be useful in the case
of a small ε since we can attempt to analyze the approximation of (uε, mε) around
(u0, m0).
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 52
When ε = 0, the system of SPDEs above reduces to a system of PDEs
du0(t, x) =
(
−H(t, x,m0, ∂xu0)−
σ2
2∂xxu
0
)
dt
dm0(t, x) =
(
−∂x(α(t, x,m0, ∂xu
0)m0) +σ2
2∂xxm
0
)
dt
or equivalently,
∂tu0 = −H(t, x,m0, ∂xu
0)−σ2
2∂xxu
0, u0(T, x) = g(x,m0T )
∂tm0 = −∂x(α(t, x,m
0, ∂xu0)m0) +
σ2
2∂xxm
0, m0(0, x) = m0(x) = L(ξ0)
(2.10)
Recall that H denotes the Hamiltonian (see (2.11)) and α denotes the minimizer of
H (see (2.6)).
The system of PDEs above is precisely the 0-MFG model originally derived and
studied by Larsy and Lions in a series of papers [42, 43, 44, 13]. The system (2.10)
says that when there is no common noise, we can turn the infinite-dimensional HJB
equation into a system of finite-dimensional fully-coupled forward backward PDEs.
Solving this system gives an optimal control and the distribution of the optimal state
process.
However, this dimension reduction cannot be obtained when there is common noise
since we need to specify an optimal control for each trajectory of the common Brow-
nian motion. As we shall see in Chapter 4, some finite dimensional approximation
can be done when the common noise is small.
2.2 Stochastic Maximum Principle (SMP)
Stochastic Maximum Principle (SMP) or Pontryagin Maximimum Principle is an
approach to a control problem which studies optimality conditions satisfied by an
optimal control. It gives sufficient and necessary conditions for the existence of an
optimal control in terms of a backward stochastic differential equation (BSDE) of an
adjoint process. In this section, we review general SMP results and apply them to
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 53
the MFG problem. For more details about SMP in general, see [53, 61] for instance.
Consider a general continuous time stochastic control problem with state process
in R given by the following dynamic
dXs = b(t, Xs, αs)dt+ σ(t, Xs)dBs, X0 = ξ0
where (Bt)0≤t≤T denotes a d-dimensional Brownian motion over a filtered probability
space (Ω,F , Ft0≤t≤T ,P), (αt)0≤t≤T denotes a control, b, σ are given functions taking
values in R,Rd respectively. Our goal is to select α = (αt)0≤t≤T ∈ H2([0, T ];R) to
minimize the expect cost
J (α) = E
[∫ T
0
r(t, Xt, αt)dt+ h(XT )
]
where r, h are running and terminal cost functions taking values in R. We assume
that b, σ, r, h are C1 in x with bounded derivative, r, h are also continuous in (t, x) for
all α and satisfy a quadratic growth condition. Define the generalized Hamiltonian
H : [0, T ]× Rd+3 → R by
H(t, a, x, y, z) , b(t, x, a)y + σ(t, x)T z + r(t, x, a) (2.11)
Given a control and the corresponding state process, we consider a pair of Ft-adapted
processes (Yt, Zt)0≤t≤T which satisfies
Yt = ∂xh(XT ) +
∫ T
t
∂xH(t, αt, Xt, Yt, Zt)dt−
∫ T
t
ZtdBt (2.12)
The equation above is called an adjoint equation and (Yt)0≤t≤T , if exist, is called
the adjoint process. Equation (2.12) is a backward stochastic differential equation
(BSDE) and can be written in a differential form as
dYt = −∂xH(t, αt, Xt, Yt, Zt)dt+ ZtdBt, YT = ∂xh(XT ) (2.13)
Remark 2.2.1. The process (Zt)0≤t≤T is part of a solution to ensure Ft-adaptivity
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 54
of Yt. For a quick introduction to BSDE, we refer to Ch.1 of [49].
The following theorem gives necessary conditions for an optimal control
Theorem 2.2.2. Suppose (αt)0≤t≤T is an optimal control with (Xt)0≤t≤T being a
corresponding state process, then the adjoint equation
dYt = −∂xH(t, αt, Xt, Yt, Zt)dt+ ZtdBt, YT = ∂xh(XT )
has an Ft-adapted solution (Yt, Zt)0≤t≤T and the following maximum condition holds
αt = argmaxa∈R
H(t, a, Xt, Yt, Zt), 0 ≤ t ≤ T, a.s.
Remark 2.2.3. There are more general statements involving the second-order varia-
tional process and a more general Hamiltonian when the diffusion coefficient depends
on a control and/or when the control space is non-convex. Proof for Theorem 2.2.2 is
based on the Taylor expansion of the cost functional J (·) around an optimal control
which is perturbed by a “spike” or “needle” variation. See Ch.3 in [61] for details.
When the cost functions are convex, then we have sufficient conditions which gives
an optimal control in term of a solution to the adjoint process.
Theorem 2.2.4. Let α = (αt)0≤t≤T ∈ H2([0, T ];R) be a control and X = (Xt)0≤t≤T
be the corresponding state process. Suppose that there exist a solution (Yt, Zt)0≤t≤T to
the corresponding adjoint equation such that
H(t, αt, Xt, Yt, Zt) = mina∈R
H(t, a, Xt, Yt, Zt), 0 ≤ t ≤ T, a.s.
and both h and (a, x) → H(t, a, x, Yt, Zt) are convex for all t ∈ [0, T ], then αt is an
optimal control, i.e.
J (α) = infα∈H2([0,T ];R)
J (α)
If, in addition, H(t, a, x, Yt, Zt) is strongly convex in a uniformly in t ∈ [0, T ] and
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 55
x ∈ R, i.e. there exist a constant C such that for all t ∈ [0, T ] and x, a, a′ ∈ R,
H(t, a, x, Yt, Zt) ≥ H(t, a′, x, Yt, Zt) + ∂aH(t, a′, x, Yt, Zt)(a− a′) + C|a− a′|2 (2.14)
then for any control β = (βt)0≤t≤T , the following estimate holds;
J (β) ≥ J (α) + C
∫ T
0
|βt − αt|2dt (2.15)
Proof. Let β ∈ H2([0, T ];R) be an arbitrary admissible control and Xt be the cor-
responding state process. Let ∆rt = r(t, Xt, αt) − r(t, Xt, βt) and define similarly
∆Xt,∆h,∆bt,∆σt, then using convexity of h and Ito’s lemma, it follows that
J (α)− J (β) = E
[∫ T
0
∆rtdt+∆h
]
≤ E
[∫ T
0
∆rtdt+ ∂xh(XT )∆XT
]
= E
[∫ T
0
∆rtdt+ YT∆XT
]
= E
[∫ T
0
(
∆rtdt+ Yt∆bt + Zt∆σt − ∂xH(t, αt, Xt, Yt, Zt)∆Xt
)
dt
]
= E
[∫ T
0
(
H(t, αt, Xt, Yt, Zt)−H(t, βt, Xt, Yt, Zt)− ∂xH(t, αt, Xt, Yt, Zt)∆Xt
)
dt
]
≤ 0
where the last inequality is from the fact that ∂aH(t, αt, Xt, Yt, Zt) = 0 and (a, x) →
H(t, a, x, Yt, Zt) is convex. Thus,
J (α) = minβ∈H2([0,T ];R)
J (β)
If H(t, a, x, Yt, Zt) is strongly convex in a (see (2.14)), then the last inequality can be
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 56
strengthened,
J (α)− J (β) = E
[∫ T
0
(
H(t, αt, Xt, Yt, Zt)−H(t, βt, Xt, Yt, Zt)− ∂xH(t, αt, Xt, Yt, Zt)∆Xt
)
dt
]
≤ −CE
[∫ T
0
|βt − αt|2dt
]
and we get the estimate (2.15) as desired.
Remark 2.2.5. The estimate (2.15) implies that an optimal control, if it exists, is
unique.
Remark 2.2.6. Note that for our particular ε-MFG model as defined in Section 1.1,
we have b(t, x, a) = a, σ(t, x, a) = (σ, ε), and Bt = (Wt, Wt).
When applying SMP to solve a control problem, if possible, one usually begins by
solving for a minimizer of the Hamiltonian as a function of (t, x, y, z), i.e.
α(t, x, y, z) = arg infa∈R
H(t, a, x, y, z)
The minimizer is unique when the Hamiltonian is (strictly) convex. We plug this
minimizer function back to the SDEs which results in a coupled system of SDE where
one is a forward SDE of the state process and the other is a backward SDE of the
adjoint process,
dXt = b(t, Xt, α(t, Xt, Yt, Zt))dt+ σ(t, Xt)dBt
dYt = −∂xH(t, α(t, Xt, Yt, Zt), Xt, Yt, Zt)dt+ ZtdBt
X0 = ξ0, YT = ∂xh(XT )
(2.16)
This type of system is called a forward backward stochastic differential equation
(FBSDE). FBSDE was first introduced by Bismut in [12] from a stochastic control
problem similar to what is presented here. Since then, the wellposed-ness theory of
FBSDE has expanded significantly. See [49] for a general reference on FBSDE and
[47] for a more recent development on a wellposed-ness theory.
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 57
If we can solve the FBSDE (2.16) above to get a solution (Xt, Yt, Zt)0≤t≤T and
show that
αt , α(t, Xt, Yt, Zt) ∈ H2([0, T ];R),
it then follows from Theorem 2.2.4 that αt is an optimal control. In fact, from both
necessary (Theorem 2.2.2) and sufficient (Theorem 2.2.4) conditions, when operating
under convexity assumption as we are, solving for an optimal control is equivalent to
solving the corresponding FBSDE.
Next, we would like to apply the SMP to our ε-MFG problem. We begin by consid-
ering the individual control problem given a stochastic continuous flow of probability
measure as defined in (1.1.2). Given (mt)0≤t≤T ∈ M([0, T ];P2(R)), then we have a
standard control problem with running and terminal cost functions given by
r(t, x, α, ω) = f(t, x,mt(ω), α), h(x, ω) = g(x,mT (ω))
Note that with (mt)0≤t≤T being random, r, h is now a random function. A represen-
tative player attempts to control his/her state process
dXαt = αtdt+ σdWt + εdWt
to minimize the cost
J (α) = E
[∫ T
0
(f(t, Xαt , mt, αt)dt+ g(Xα
T , mT )
]
With m as an argument in the cost functions, the generalized Hamiltonian now reads
H(t, a, x,m, y, z) = ay + f(t, x,m, a) + σz
Consequently,
∂xH(t, a, x,m, y, z) = ∂xf(t, x,m, a)
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 58
and the minimizer is given by
α(t, x, y, z) = argmina∈R
(ay + f(t, x,m, a) + σz)
Note that both ∂xH and α is independent of z, so we can reduce our notation and
exclude the z-argument in both H and α. In addition, it is more convenient to work
with the Hamiltonian (as opposed to the generalized Hamiltonian)
H(t, x,m, y) = mina∈R
(ay + f(t, x,m, a)) = H(t, α(t, x, y,m), x,m, y)
Using optimality of α, it is easy to check that
∂xH(t, x,m, y) = ∂xH(t, α(t, x, y,m), x,m, y)
Thus, the corresponding FBSDE for an individual control problem given m is given
by
dXt = α(t, Xt, Yt, mt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt, mt, Yt)dt+ ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT , mT )
(2.17)
By applying the necessary and sufficient conditions above, we have the following
theorem which turns the control problem of each individual player to an FBSDE with
random coefficient
Theorem 2.2.7. Assume A holds, then the individual control problem given (mt)0≤t≤T ∈
M([0, T ];P2(R)) has an optimal control if and only if the FBSDE (2.17) is solvable.
In that case, the optimal control is given by
αt = α(t, Xt, mt, Yt)
for all t ∈ [0, T ], where (Xt, Yt, Zt, Zt)0≤t≤T is a solution to FBSDE (2.17).
In fact, as we shall see in Chapter 3 (see particularly Theorem 3.1.3), under the
same set of assumptions, the FBSDE (2.17) is uniquely solvable.
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 59
Next, we proceed from a single player control problem to a mean field game. From
the definition of ε-MFG (see Definition 1.1.4), it simply requires that the following
equilibrium or consistency condition holds
mt = L(Xt|Ft)
where Xt is an optimal controlled process for an individual control problem given
m = (mt)0≤t≤T . Adding this condition into FBSDE (2.17) yields the SMP for ε-
MFG.
Theorem 2.2.8. Assume that A holds, then ε-MFG is solvable if and only if the
FBSDEdXt = α(t, Xt,L(Xt|Ft), Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt,L(Xt|Ft), Yt)dt+ ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT ,L(XT |FT ))
(2.18)
is solvable. In that case, ε-MFG solution is given by
αt = α(t, Xt,L(Xt|Ft), Yt), ∀t ∈ [0, T ]
Proof. Suppose (αt)0≤t≤T is an ε-MFG solution, then let (mt)0≤t≤T be the conditional
law of the corresponding state process, i.e.
mt = L(Xt|Ft)
where (Xt)0≤t≤T follows the dynamic
dXt = αtdt + σdWt + εdWt, X0 = ξ0
By definition of an ε-MFG solution, α must be an optimal control for an individual
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 60
control problem given m as well. Thus by Theorem 2.2.2, the FBSDE
dXt = α(t, Xt, mt, Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt, mt, Yt)dt+ ZtdWt + ZtdWt
X0 = ξ0, YT = ∂xg(XT , mT )
(2.19)
is solvable and αt = α(t, Xt, mt, Yt). Thus, the forward solution (Xt)0≤t≤T to (2.19)
has the same law as (Xt)0≤t≤T , so that mt = L(Xt|Ft) = L(Xt|Ft). Plug this back to
(2.19) and we get (2.18) as desired.
Now, suppose (2.18) is solvable and denote a solution by (Xt, Yt, Zt, Zt)0≤t≤T . Let
αt = α(t, Xt,L(Xt|Ft), Yt), then the result follows directly from the definition of ε-
MFG and Theorem 2.2.4 using the fact that f, g are convex under assumption A.
The FBSDE (2.18) above is a mean-field type or more specifically McKean-Vlasov
type FBSDE as it involves the law of the process. The McKean-Vlasov forward SDE
was first introduced by McKean as a limit of interacting particles system following
diffusion processes. We refer to [25, 56] for classical references on a general McKean-
Vlasov theory. The FBSDE version of McKean-Vlasov equation was first introduced
by Carmona and Delarue in [17] as a result of SMP applied to MFG with no common
noise or 0-MFG. In that case, the conditional law L(Xt|Ft) is simply L(Xt). To the
best of our knowledge, McKean-Vlasov FBSDE with conditional law has not been
considered. We will show existence and uniqueness result for this FBSDE in Chapter
3 under an additional monotonicity type condition.
2.3 Comparison between the two approaches
We now describe the connection between the two approaches in the MFG context.
In a general stochastic control problem, it is known that the adjoint process is a
gradient of a value function along the optimal path provided that the value function
is sufficiently regular. See Theorem 6.4.2 in [53] for instance. The following theorem
gives a similar result in the MFG context with the generalize value function Vε.
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 61
Theorem 2.3.1. Suppose that Vε ∈ C1,3,3([0, T )×R×P2(R))∩C([0, T ]×R×P2(R))
is a classical solution to the master equation (2.8). Suppose that ∂xxVε, ∂xmVε, and
∂xVε(t, 0, δ0) are bounded then
(Xt, Yt, Zt, Zt)0≤t≤T =(
Xεt , ∂xV
ε(t, Xεt , m
εt), σ∂xxV
ε(t, Xεt , m
εt), ε
(
∂xxVε(t, Xε
t , mεt ) + E
0[
∂xmVε(t, Xε
t , mεt)(Xt)
]))
(2.20)
is a solution to FBSDE (2.18) satisfying
E
[
sup0≤t≤T
(Xεt )
2 + sup0≤t≤T
(Y εt )
2 +
∫ T
0
(Zεt )
2 +
∫ T
0
(Zεt )
2
]
<∞ (2.21)
where (Xεt )0≤t≤T is a solution to the SDE
dXt = α(t, Xt,L(Xt|Ft), ∂xVε(t, Xt,L(Xt|Ft)))dt+ σdWt + εdWt, X0 = ξ0
mεt = L(Xε
t |Ft), and (Xt)0≤t≤T is a lifting random variable of (Xεt )0≤t≤T sharing a
common noise space. In particular,
αεt , α(t, Xt,L(Xt|Ft), ∂xV
ε(t, Xt,L(Xt|Ft))) ∈ H2([0, T ];R)
is a solution to ε-MFG.
Proof. The bounded second derivative condition of Vε implies that the SDE
dXt = α(t, Xt,L(Xt|Ft), ∂xVε(t, Xt,L(Xt|Ft)))dt+ σdWt + εdWt, X0 = ξ0
admits a unique solution. Combining this with the fact that ∂xVε(t, 0, δ0) is bounded,
we have that ∂xVε satisfies the linear growth condition. By standard SDE estimate, we
have that (Xεt , Y
εt , Z
εt , Zt)0≤t≤T satisfying (2.21). Then we can apply generalized Ito’s
lemma to ∂xVε(t, Xε
t ,L(Xεt |Ft)) (see Proposition 6.5 in [18]) to verify directly that
(Xεt , Y
εt , Z
εt , Zt)0≤t≤T satisfies FBSDE (2.18). The last remark follows from Theorem
2.2.8
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 62
Remark 2.3.2. Observe that extra regularity on a value function is required to estab-
lish the connection between DPP and SMP. For a classical stochastic control problem,
there is a more general result involving a viscosity solution and a sub/super gradient
that does not require as much regularity. We refer to [61] for more details on the
relation between the two approaches in a classical stochastic control set up.
We now summarize the discussion of the two approaches to MFG in Table 2.1 by
describing the types of equations that results from applying both approaches to MFG
model in both cases: with and without common noise.
Approachε = 0
(no common noise)
ε > 0
(with common noise)
Dynamic Programming Principle FBPDEInfinite-dimensional HJB
or FBSPDE
Stochastic Maximum Principle McKean-Vlasov FBSDEMcKean-Vlasov FBSDE
with conditional law
Table 2.1: Summary of the types of equations from applying DPP and SMP ap-
proaches to MFG problems
Most of the prior works has been on the upper-left corner of the table, i.e. the
study of FBPDE (2.10) and some of its variations. However, extending those results
to the ε > 0 case is non-trivial because we now either have to work with an infinite-
dimensional non-linear PDE or a system of FBSPDE. In either cases, they are much
more complicated than a system of finite-dimensional PDEs and are much less studied.
For the SMP, the main difference between a model with and without common noise
is in the mean-field terms appearing in the FBSDE. Changing from a deterministic
unconditional law to a law conditional on a common brownian motion filtration, the
FBSDE for a given m is changed from having deterministic coefficients to the one
with random coefficients from a stochastic flow of probability measure. These two
kinds of FBSDE are very different as the former can be solved via PDE techniques
CHAPTER 2. TWO APPROACHES TO MEAN FIELD GAMES 63
while the tools to tackle the latter are much more limited. In addition, the spaces
of deterministic flows and stochastic flows of probability measures are considerably
different. As a result, the proof based on compactness and the Schauder fixed point
theorem previously used in the 0-MFG by Carmona and Delarue in [17] cannot be
extended to the ε-MFG case. Despite this fact, there have been active researches on
FBSDE with random coefficients in the past decades. See [5, 35, 47, 48, 52, 59, 60],
and these results are helpful in dealing with Mckean-Vlasov FBSDE with conditional
law.
Chapter 3
Mean field games with common
noise
In this chapter, we discuss a general MFG with common noise. In Section 3.1, we show
the existence and uniqueness of ε-MFG solutions for a linear-convexity model with a
weak monotonicity condition using SMP approach discussed in the previous chapter.
In Section 3.2, we discuss the Markov property of ε-MFG by showing existence of a
deterministic decoupling function. We then discuss the connection to HJB through
this decoupling function in Section 3.3.
3.1 Wellposedness of MFG with common noise
Recall that by applying the SMP to ε-MFG, we turn the problem to solving the
following McKean-Vlasov FBSDE (see Theorem 2.2.8)
dXt = α(t, Xt,L(Xt|Ft), Yt)dt+ σdWt + εdWt
dYt = ∂xH(t, Xt,L(Xt|Ft), Yt)dt+ ZtdWt + ZtdWt
X0 = ξ, YT = ∂xg(XT ,L(XT |FT ))
(3.1)
Our goal is to show that the system above is uniquely solvable under assumption
A and B outlined in Section 1.1.4. We will in fact prove a slightly more general
result, namely the existence and uniqueness of an FBSDE with monotone functional,
64
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 65
since this type of FBSDE will appear frequently in our subsequent analysis.
3.1.1 FBSDE with monotone functionals
We are interested in the following FBSDE over [s, T ]
dXt = b(t, Xt, Yt)dt+ σdWt + εdWt
dYt = F (t, Xt, Yt)dt+ ZtdWt + ZtdWt
Xs = ξ, YT = G(XT )
(3.2)
where ξ ∈ L2Fs
and b, F,G are measurable maps
b : [0, T ]×L2F × L2
F × Ω → R
F : [0, T ]× L2F ×L2
F × Ω → R
G : L2F × Ω → R
(3.3)
Notice that the maps b, F,G are “functionals”. An example of b, F,G to keep in mind
is precisely that from McKean-Vlasov FBSDE (3.1) of ε-MFG. That is,
b(t, X, Y ) = α(t, X,L(X|Ft), Y )
F (t, X, Y ) = −∂xH(t, X,L(X|Ft), Y )
G(X) = ∂xg(X,L(X|FT ))
(3.4)
In this example, they are functionals only with respect to X . We now list a set of
assumptions (denoted by H) on b, F,G. Being a functional, we first need to ensure
that the stochastic integral is well-defined, so the following measurability assumption
is necessary
(H1). (b(t, Xt, Yt))0≤t≤T , (F (t, Xt, Yt))0≤t≤T are Ft-progressively measurable for any
Ft-progressively measurable (Xt)0≤t≤T ,(Yt)0≤t≤T .
The second assumption is a standard SDE assumption, namely Lipschitz and
linear growth condition. In a functional form, it is stated under expectation, or
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 66
for our application to ε-MFG, under a conditional expectation. Fix a sub-filtration
G = Gt0≤t≤T of F = Ft0≤t≤T , then the assumption reads
(H2). There exist a constant K such that for any t ∈ [0, T ], X,X ′, Y, Y ′ ∈ L2FT
,
A ∈ Gt.
E[
1A(b(t, X, Y )− b(t, X ′, Y ′))2]
≤ KE[
1A((X −X ′)2 + (Y − Y ′)2)]
E[
1A(F (t, X, Y )− F (t, X ′, Y ′))2]
≤ KE[
1A((X −X ′)2 + (Y − Y ′)2)]
E[
1A(G(X)−G(X ′))2]
≤ KE[
1A(X −X ′)2]
(3.5)
and
E
[∫ T
0
b(t, 0, 0)2 + F (t, 0, 0)2dt
]
<∞ (3.6)
In addition to the standard assumptions above, the following monotonicity con-
dition on b, F,G is needed.
(H3). For any t ∈ [0, T ], X,X ′, Y, Y ′ ∈ L2FT
, and A ∈ Gs,
E [1A(F (t, X, Y )− F (t, X ′, Y ′))(X −X ′) + 1A(b(t, X, Y )− b(t, X ′, Y ′))(Y − Y ′)] ≤
−β1E[
1A(b(t, X, Y )− b(t, X ′, Y ′))2]
− β2E[
1A(Y − Y ′)2]
E [1A(G(X)−G(X ′))(X −X ′)] ≥ 0
(3.7)
for some constant β1, β2 ≥ 0 with β1 + β2 > 0.
This monotonicity condition is motivated by results in standard FBSDE theory
that have successfully dealt with the case of random coefficients. See [52, 49] for in-
stance. This type of equation, with these three assumptions on the drift and terminal
functionals, covers most, if not all, of the FBSDEs that we will encounter in this
thesis including those in Chapter 4 where we analyze asymptotic behavior when the
common noise is small. For instance, given m = (mt)0≤t≤T ∈ M([0, T ];P2(R)), then
b(t, X, Y ) = α(t, X,mt, Y )
F (t, X, Y ) = −∂xH(t, X,mt, Y )
G(X) = ∂xg(X,mT )
(3.8)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 67
would correspond to FBSDE (2.17) arising from an individual control problem given
m with f, g satisfying assumption A. As we shall see in Theorem 3.1.3, the convexity
on f, g gives a monotonicity condition on b, F,G defined above. Another case and
perhaps a more important one is (3.4) as it corresponds to the McKean-Vlasov FBSDE
(2.2.8) arising from the ε-MFG as discussed above.
3.1.2 A priori estimate
We first present a priori estimates necessary for our proof of existence. Working with
FBSDE, this type of estimate can usually be obtained only when the time duration
is sufficiently small. In this case, we can achieve this estimate for arbitrary time
s ∈ [0, T ] by using the monotonicity condition (H3)
Theorem 3.1.1. Let ξ1, ξ2 ∈ L2Fs
and, for i = 1, 2, let (X it , Y
it , Z
it , Z
it)s≤t≤T denote
a solution to FBSDE (3.2) with initial condition ξi and coefficients (σi, εi, bi, Fi, Gi).
That is, they satisfy
dX it = bi(t, X
it , Y
it )dt+ σidWt + εidWt
dY it = Fi(t, X
it , Y
it ) + Z i
tdWt + Z itdWt
X is = ξi, Y i
T = Gi(XiT )
(3.9)
Assume that (bi, Fi, Gi) satisfies assumptions H. Then there exists a constant CK,T
that depends only on K, T such that the following estimate holds for any A ∈ Gs:
E
[
sups≤t≤T
1A(∆Xt)2 + sup
s≤t≤T1A(∆Yt)
2 +
∫ T
s
[1A(∆Zt)2 + 1A(∆Zt)
2]dt
]
≤ CK,T
(
E
[
1A(∆ξ)2 + 1AG(X
1T )
2 + 1A
∫ T
s
(
F (θ1t )2 + b(θ1t )
2)
dt
]
+ (∆σ)2 + (∆ε)2)
(3.10)
where∆Xt = X1t −X
2t , b = b1−b2, θ1t = (t, X1
t , Y1t ) and∆Yt,∆Zt,∆Zt,∆ξ,∆σ,∆ε,F , G
are defined similarly.
Proof. See Section 3.4.1.
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 68
3.1.3 Wellposedness result
Using estimate (3.10), we are ready to prove our main theorem which shows that
FBSDE (3.2) is wellposed. We then apply this result to establish unique solvability
of two FBSDEs: FBSDE (2.17) arising from an individual control problem given m,
and McKean-Vlasov FBSDE (3.1) from ε-MFG.
Theorem 3.1.2. Let ξ ∈ L2Fs
and b, F,G be functionals satisfying assumption H, then
there exists a unique adapted solution (Xt, Yt, Zt, Zt)s≤t≤T to FBSDE (3.2) satisfying
the estimate
E
[
1A sups≤t≤T
X2t + 1A sup
s≤t≤TY 2t + 1A
∫ T
s
[Z2t + Z2
t ]dt
]
≤ CK,T
(
E
[
1Aξ2 + 1AG(0)
2 + 1A
∫ T
s
(
b(t, 0, 0)2 + F (t, 0, 0)2)
dt
]
+ σ2 + ε2)
(3.11)
for some constant CK,T depends only on K, T and for all A ∈ Gs.
Proof. The proof can be summarized as follows; we begin by showing existence of a
solution over a small time duration by means of Banach fixed point theorem. This
small time depends on the Lipschitz constants of the functionals. We then define
a new terminal functional and repeat the process backward in time. The estimate
(3.10) above allows us to control the Lipschitz constant of the terminal functional
as we move backward to construct a solution. See Section 3.4.2 for details of the
proof.
We now apply Theorem 3.1.2 to FBSDE (2.17) (over [s, T ]) by setting
b(t, X, Y ) = α(t, X,mt, Y ),
F (t, X, Y ) = −∂xH(t, X,mt, Y ),
G(X) = ∂xg(X,mT )
Gts≤t≤T = Fts≤t≤T
(3.12)
for a given (mt)s≤t≤T ∈ M([s, T ];P2(R)), where H is the Hamiltonian, α is its min-
imizer function, and g is the terminal cost function. See (2.5) and (2.6) for their
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 69
definitions. This FBSDE corresponds to an individual stochastic control given m.
See Theorem 2.2.7.
Theorem 3.1.3. Assume thatA holds, there exist a unique solution (Xt, Yt, Zt, Zt)s≤t≤T
to the FBSDE (2.17) over [s, T ] satisfying
E
[
sups≤t≤T
[X2t + Y 2
t ] +
∫ T
s
[Z2t + Z2
t ]dt
]
≤ CK,TE
[
ξ2 + (∂xg(0, mT ))2 +
∫ T
s
(α(t, 0, mt, 0))2 + (∂xH(t, 0, mt, 0))
2dt+ σ2 + ε2]
(3.13)
Moreover, two solutions (X it , Y
it , Z
it , Z
it)s≤t≤T , i = 1, 2 to FBSDE (2.17) with initial ξi
satisfies the estimate
E
[
sups≤t≤T
1A∆X2t + sup
s≤t≤T1A∆Y
2t +
∫ T
s
[1A∆Z2t + 1A∆Z
2t ]dt
]
≤ CK,TE[1A∆ξ2]
(3.14)
where ∆Xt = X1t − X2
t , ∆Yt,∆Zt,∆Zt,∆ξ are defined similarly, CK,T is a constant
depends only on K, T , and A ∈ Fs.
Proof. We simply need to check that under assumption A on f, g, the correspond-
ing b, F,G as defined in (3.12) satisfy assumptions H with filtration Gt0≤t≤T =
Ft0≤t≤T . Then we can apply Theorem 3.1.2 to get the desired result.
First, by measurability of f, g, hence for α, ∂xH , and the assumption that (mt)0≤t≤T
is Ft-progressively measurable, (H1) follows. Next, by definition, α satisfies
∂αf(t, x,m, α(t, x,m, y)) + y = 0 (3.15)
for any x, y ∈ R, m ∈ P2(R). Consequently, by the Lipschitz property of ∂αf in
(x, α,m), it follows that α is Lipschitz in x, y. In addition, using the optimality
condition on α, we have
∂xH(t, x,m, y) = ∂xf(t, x,m, α(t, x,m, y)) (3.16)
Combining with Lipschitz assumption (A1) on ∂xf, ∂αf, ∂xg, (H2) follows easily. We
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 70
now check the monotonicity condition (H3). By strict convexity assumption on f (see
(1.9) in (A3)), we have
f(t, x,m, α)+∂xf(t, x,m, α)(x′−x)+∂α(t, x,m, α)(α
′−α)+cf |α′−α|2 ≤ f(t,m, x′, α′)
where
α = α(t, x,m, y), α′ = α(t, x′, m, y′)
Interchanging (x, α) and (x′, α′), we also have
f(t, x′, m, α′)+∂xf(t, x′, m, α′)(x−x′)+∂αf(t, x
′, m, α′)(α−α′)+cf |α′−α|2 ≤ f(t, x,m, α)
Summing both equations and using (3.15), it follows that
−(∂xf(t, x,m, α)−∂xf(t, x′, m, α′))(x−x′)+(y−y′)(α− α′)+cf (α− α
′)2 ≤ 0 (3.17)
for any x, x′, y, y′ ∈ R. The condition (H3) then follows by (3.16),(3.17) with β1 =
cf , β2 = 0. The monotonicity of G is obvious from the convexity of g. The estimate
(3.14) follows immediately from estimate (3.10) in Theorem 3.1.2
In addition to estimate (3.14), we are also interested in an estimate between two
solutions from two different stochastic flows of probability measure.
Proposition 3.1.4. In addition to A, we assume that ∂xf, ∂αf, ∂xg are Lipschitz
in m with the same Lipschitz constant K. Let ξ1, ξ2 ∈ L2Fs, (m1
t )s≤t≤T , (m2t )s≤t≤T ∈
M([s, T ];P2(R)), and (Xit , Y
it , Z
it , Z
it)s≤t≤T denote the solution to FBSDE (2.17) given
mi and initial ξi, then the following estimate holds
E
[
sups≤t≤T
1A∆X2t + sup
s≤t≤T1A∆Y
2t +
∫ T
s
[1A∆Z2t + 1A∆Z
2t ]dt
]
≤ CK,TE[1A∆ξ2 + 1A
∫ T
s
(∆mt)2dt]
(3.18)
where∆Xt = X1t −X
2t , ∆Yt,∆Zt,∆Zt,∆ξ are defined similarly, and∆mt = W2(m
1t , m
2t ).
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 71
Proof. Let
bi(t, X, Y ) = α(t, X,mit, Y ),
F i(t, X, Y ) = −∂xH(t, X,mit, Y ),
Gi(X) = ∂xg(X,miT )
Gt0≤t≤T = Ft0≤t≤T
(3.19)
for i = 1, 2. Then from the Lipschitz in m of ∂xf, ∂αf, ∂xg, it follows that α, ∂xH, ∂xg
are Lipschitz in m and the result follows by applying estimate (3.10) in Theorem 3.1.2
with (ξi, bi, F i, Gi), i = 1, 2 above.
Remark 3.1.5. Additional assumption in Proposition 3.1.4 holds under (B2),(B1).
Going back to McKean-Vlasov FBSDE (2.2.8), we now apply Theorem 3.1.2 by
setting
b(t, X, Y ) = α(t, X,L(X|Ft), Y ),
F (t, X, Y ) = −∂xH(t, X,L(X|Ft), Y ),
G(X) = ∂xg(X,L(X|FT ))
Gts≤t≤T = Fts≤t≤T
(3.20)
Notice that the filtration used here is not as general as that in the previous theorem.
The result is the following;
Theorem 3.1.6. [Wellposedness of McKean-Vlasov FBSDE] Assume that A,B hold,
then there exist a unique solution (Xt, Yt, Zt, Zt)s≤t≤T to FBSDE (2.2.8) satisfying
E
[
sups≤t≤T
[X2t + Y 2
t ] +
∫ T
s
[Z2t + Z2
t ]dt
]
≤ C
(
E[ξ2] + (∂xg(0, δ0))2 +
∫ T
s
(α(t, 0, δ0, 0))2 + (∂xH(t, 0, δ0, 0))
2dt+ σ2 + ε2)
(3.21)
where δa denote the dirac measure at a. Moreover, given two solutions (X it , Y
it , Z
it , Z
it)s≤t≤T , i =
1, 2 to FBSDE (2.2.8) with initial ξi, the following estimate holds:
E
[
sups≤t≤T
1A∆X2t + sup
s≤t≤T1A∆Y
2t +
∫ T
s
[1A∆Z2t + 1A∆Z
2t ]dt
]
≤ CK,TE[1Aξ2] (3.22)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 72
where ∆Xt = X1t − X2
t , ∆Yt,∆Zt,∆Zt,∆ξ are defined similarly, CK,T is a constant
depends only on K, T , and A ∈ Fs.
Proof. Similar to Theorem 3.1.3, we simply need to check that under assumption A,
B on f, g, the corresponding b, F,G as defined in (3.20) satisfy assumption H with
filtration Gt0≤t≤T = Ft0≤t≤T .
First, (H1) follows from the measurability of f, g and the fact that (Xt)0≤t≤T 7→
(L(Xt|Ft))0≤t≤T is measurable. Next, by separable assumption (B2), α is independent
of m, that is, α := α(t, x, y) satisfies
∂αf0(t, x, α(t, x, y)) + y = 0 (3.23)
Consequently, by Lipschitz property of ∂αf0 in (x, α), it follows that α is Lipschitz in
x, y. In addition, using the optimality condition on α, we have
∂xH(t, x,m, y) = ∂xf0(t, x, α(t, x, y)) + ∂xf
1(t, x,m) (3.24)
Combining with Lipschitz assumption (A1),(B1) on ∂xf, ∂αf, ∂xg, (H2) follows easily.
We now check the monotonicity condition (H3). By strict convexity assumption on
f 0, (see (1.9) in (A3)), we follow the same argument as in Theorem 3.1.3 to get
−(∂xf0(t, x, α)− ∂xf
0(t, x′, α′))(x− x′) + (y − y′)(α− α′) + cf(α− α′)2 ≤ 0 (3.25)
for any x, x′, y, y′ ∈ R. where
α = α(t, x, y), α′ = α(t, x′, y′)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 73
Now, by weak monotonicity condition (B3) on f, g, we have
E
[
1A(∂xf1(t, X,L(X|Ft))− ∂xf
1(t, X ′,L(X ′|Ft))(X −X ′)]
= E
[
1AE
[
(∂xf1(t, X,L(X|Ft))− ∂xf
1(t, X ′,L(X ′|Ft))(X −X ′)|Ft
]]
≥ 0
E
[
1A(∂xg(X,L(X|Ft))− ∂xg(X′,L(X ′|Ft))(X −X ′)
]
= E
[
1AE
[
(∂xg(X,L(X|Ft))− ∂xg(X′,L(X ′|Ft))(X −X ′)|Ft
]]
≥ 0
(3.26)
for any X ∈ L2F , A ∈ Ft. The condition (H3) then follows by (3.24),(3.25),(3.26) with
β1 = cf , β2 = 0.
Combining with Theorem 2.2.8, we have the wellposedness result for ε-MFG with
common noise.
Corollary 3.1.7 (Wellposedness of ε-MFG). Under assumption A and B, there exist
a unique ε-MFG solution for any initial ξ0 ∈ L2F0.
Remarks on the functional framework
The existence and uniqueness proof of (3.2) is based on the standard techniques in
FBSDE involving Lipschitz and monotonicity condition. The essential difference is the
m-argument in the coefficients. However, working under the second order Wasserstein
metric, we have the inequality
W2(m,m′) ≤ E[(ξ − ξ′)2]
12
where ξ, ξ′ are square integrable random variables over an arbitrary probability space
with law m,m′. As a result, we have an estimate of the form
E[ϕ(ξ,L(ξ))− ϕ(ξ′,L(ξ′))] ≤ KE[(ξ − ξ′)2]12
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 74
whenever ϕ isK-Lipschitz in (x,m)-arguments. Using this estimate, we can deal with
the m-argument in the same way as we do with the x-argument. Under this general
framework, these two arguments are then grouped into the same argument in the
functional and their Lipschitz property are treated together. By brining the mono-
tonicity condition, which is widely used in the standard FBSDE, to this functional
framework, we arrive at the weak monotonicity condition.
However, it is worth noting that while this set of assumptions is rather general,
it does not include many interesting coefficient functions. A main example includes
local coupling function on m where the output depends on a local information of a
probability measure such that the density function. An example of such functions is
g(x,m) = ϕm(x)
where ϕm denotes the density function of m. See [14, 15, 54] for some works on the
MFG with local coupling.
3.2 Markov property and a decoupling function
Generally, to solve an FBSDE, we need to find what is called a decoupling field, a pos-
sibly random function describing the relation of the backward process Yt as a function
of the forward process Xt. When the coefficients are deterministic, this function is
deterministic and is a solution to a quasilinear PDE. In that case, the FBSDE is said
to be Markovian and we call the function a decoupling function. Using this function,
we can decouple an FBSDE and reduce the problem to solving a standard forward
SDE. This method of solving FBSDE is called Four-steps scheme and was first pro-
posed by Ma et al. in [46]. See [26] for more detail on a decoupling function of a
classical FBSDE in a deterministic case and [47] for a decoupling field in a general
case.
While we should expect a similar result in ε-MFG problem when the running
and terminal cost functions are deterministic, it is not obvious a priori that this
Markovian property holds particularly in the case of common noise. For a fixed
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 75
m ∈ M([0, T ];P2(R)), we are dealing with FBSDE with non-deterministic coefficients
from the fact that m is stochastic. Specifically, we have a path dependent functions
∂xH(t, x,mt(ω), y), g(x,mT (ω))
However, as f, g are still deterministic functions of m, it is reasonable to expect a
Markov property if we include an additional input, the current distribution of players,
or in FBSDE context, the conditional law ofXt. Our main result for this section is the
following theorem which shows the existence of a deterministic decoupling function
for ε-MFG thereby proving the Markov property of MFG with common noise.
Theorem 3.2.1. There exist a deterministic function Uε : [0, T ] × R × P2(R) such
that
Y εt = Uε(t, Xε
t ,L(Xεt |Ft)) (3.27)
Moreover, Uε satisfies the estimates
1. |Uε(t, x,m)− Uε(t, x′, m′)| ≤ CK,T (|x− x′|+W2(m,m′))
2. (Uε(t, x,m)− Uε(t, x′, m)) (x− x′) ≥ 0
for all t ∈ [0, T ], x, x′ ∈ R, m,m′ ∈ P2(R) where CK,T depends only on K, T .
We will state and prove a slightly more general result in Theorem 3.2.12 towards
the end of this section. For the estimates, see Lemma 3.2.11.
Remark 3.2.2. We will focus here only on Yt as it directly relates to the ε-MFG
solution in our model, but the existence of a decoupling function indeed holds for
Zt, Zt as well.
Remark 3.2.3. Similar to a classical FBSDE, it is natural to ask if Uε is a solution
to a certain PDE, and this question brings us to the connection between the SMP and
HJB approach in Section 3.3
As we shall see in Chapter 4, the decoupling function Uε and its PDE are essential
in our asymptotical analysis. The main idea for proving Theorem 3.2.1 is to define
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 76
Uε(t, x,m) through a solution of a certain FBSDE, then show that this solution is
well-defined and deterministic, and the resulting function satisfies the relation
Y εt = Uε(t, Xε
t ,L(Xεt |Ft)) (3.28)
where (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T is the solution to McKean-Vlasov FBSDE (2.2.8).
From now and throughout the rest of this section, we will operate under assump-
tion A,B. Under these assumptions, α := α(t, x, y) is independent of m and
∂xH(t, x,m, y) = ∂xf0(t, x, α(t, x, y)) + ∂xf
1(t, x,m)
We begin with the definition of the FBSDE corresponding to a subgame over [s, T ].
This FBSDE will be used for defining Uε.
Definition 3.2.4. Let s ∈ [0, T ], ξ ∈ L2Fs
and Gts≤t≤T be a sub-filtration of Fts≤t≤T .
We define FBSDE with data (s, ξ, Gts≤t≤T ) or simply FBSDE (s, ξ, Gts≤t≤T ) to
be the following FBSDE
dXt = α(t, Xt, Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt,L(Xt|Gt), Yt)dt+ ZtdWt + ZtdWt
Xs = ξ, YT = ∂xg(XT ,L(XT |GT ))
(3.29)
(Xt, Yt, Zt, Zt)s≤t≤T is called a solution to FBSDE (s, ξ, Gts≤t≤T ) if they are Ft-
adapted and satisfy FBSDE (3.29).
From this definition, the McKean-Vlasov FBSDE corresponding to the ε-MFG
with initial ξ ∈ L2F0
is simply FBSDE (0, ξ, Ft0≤t≤T ). Under assumption A,B,
existence and uniqueness of the solution to FBSDE above is guaranteed by Theorem
3.1.6.
Recall that F st is a σ-algebra generated by the common brownian motion start-
ing at time s, i.e. F st = σ
(
Wr − Ws; s ≤ r ≤ t)
. For ξ ∈ L2Fs, we denote by
(Xs,ξt , Y
s,ξt , Z
s,ξt , Z
s,ξt )s≤t≤T the unique solution to FBSDE (s, ξ, F s
t s≤t≤T ), i.e. they
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 77
satisfies
dXs,ξt = α(t, Xs,ξ
t , Ys,ξt )dt+ σdWt + εdWt
dYs,ξt = −∂xH(t, Xs,ξ,L(Xs,ξ|F s
t ), Ys,ξt ) + Z
s,ξt dWt + Z
s,ξt dWt
Xs,ξs = ξ, Y
s,ξT = ∂xg(X
s,ξT ,L(Xs,ξ
T |F sT ))
(3.30)
The FBSDE above corresponds to the ε-MFG over [s, T ] with initial distribution at
time s being L(ξ). We define, for 0 ≤ s ≤ t ≤ T , the following map
Θs,t :P2(R) → L2Fs
t(P2(R))
m→ L(Xs,ξt |F s
t )(3.31)
where ξ ∈ L2Fs
with L(ξ) = m. Intuitively, Θs,t(m) gives the unique distribution of
players’ state at time t corresponding to the ε-MFG solution over [s, T ] with initial
law at time s being m. We will sometimes use the following notation
ms,mt , Θs,t(m) (3.32)
First, we check that this map is well-defined, that is, L(Xs,ξt |F s
t ) is independent of the
choice of ξ ∈ L2Fs
provided that L(ξ) = m. This is equivalent to a conditional weak
uniqueness for FBSDE (3.30). This property is implied from a pathwise uniqueness
of FBSDE (3.30). See Proposition 3.2.6 for a more general statement.
Corollary 3.2.5. Suppose that ξ1, ξ2 ∈ L2Fs
such that L(ξ1) = L(ξ2) = m ∈ P2(R),
then L(Xs,ξ1t |F s
t ) = L(Xs,ξ2t |F s
t ) for all t ∈ [s, T ] where Xs,ξ1t , X
s,ξ2t are as defined
above.
For a slightly more general result, we consider the FBSDE (s, ξ, F rt s≤t≤T ) where
0 ≤ r ≤ s ≤ T , i.e.
dXt = α(t, Xt, Yt)dt+ σdWt + εdWt
dYt = −∂xH(t, Xt,L(Xt|Frt ), Yt)dt+ ZtdWt + ZtdWt
Xs = ξ, YT = ∂xg(XT ,L(XT |FrT ))
(3.33)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 78
Observe that there is a subtle difference between (3.33) and (3.30) in the filtration
used in the equations. Now, we state a more general weak uniqueness result which
implies Corollary 3.2.5 above by setting r = s.
Proposition 3.2.6. Suppose that ξ1, ξ2 ∈ L2Fs
such that L(ξ1|F rs ) = L(ξ2|F r
s ) ∈
L2Fs(P2(R)), then L(Xs,ξ1
t |F rt ) = L(Xs,ξ2
t |F rt ) for all t ∈ [s, T ] where Xs,ξ1
t , Xs,ξ2t are
as defined above.
Proof. Fix a path of the common Brownian motion ω ∈ Ω, then follow the same
argument as in Theorem 5.1 in [6] which shows that pathwise uniqueness implies
weak uniqueness for an FBSDE.
Next, we show the following estimate
Proposition 3.2.7. For 0 ≤ s ≤ t ≤ T , m,m′ ∈ P2(R), there exists a constant CK,T
that depends only on K, T such that
E[
W2(Θs,t(m),Θs,t(m′))
]
≤ CK,TW2(m,m′) (3.34)
Proof. Let ξ, ξ′ be arbitrary elements of L2Fs
with law m,m′. Let (Xt, Yt, Zt, Zt)0≤t≤T
and (X ′t, Y
′t , Z
′t, Z
′t)0≤t≤T denote the solutions of FBSDE (2.18) with initial Xs =
ξ,X ′s = ξ′, then by the estimate (3.22), it follows that
E[W2(Θs,t(m),Θs,t(m′))] ≤ E[(Xt −X ′
t)2] ≤ CK,TE[(ξ − ξ′)2]
for a constant CK,T depends only on K, T . Since ξ, ξ′ are arbitrary, we conclude that
E[W2(Θs,t(m),Θs,t(m′))] ≤ CW2(m,m
′)
With Proposition 3.2.6 and 3.2.7, we can now proceed to show the following
property of Θ. The result below is essentially the Markov property of the stochastic
flow of probability measure corresponding to ε-MFG.
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 79
Theorem 3.2.8. For any m ∈ P2(R) and 0 ≤ s ≤ t ≤ u ≤ T
Θt,u(Θs,t(m)) = Θs,u(m) (3.35)
where Θ is as defined in (3.31).
Proof. The proof is based on the localization argument and a prior global Lipschitz
estimate (3.34). Given an arbitrary random variable ξ ∈ L2Fs
with lawm ∈ P2(R), the
values Θs,t(m) and Θs,u(m) are known immediately from its definition (see (3.31)).
However, since Θs,t(m) is random, Θt,u(Θs,t(m)) cannot be computed easily from its
definition. So we construct an approximaton of Θs,t(m) and use the estimate (3.34)
to bound the difference. We refer to Section 3.4.3 for the details of the proof.
Remark 3.2.9. Using notation (3.32), this is equivalent to ms,mu = m
t,ms,mt
u for any
m ∈ P2(R) and 0 ≤ s ≤ t ≤ u ≤ T .
Next, we let ξ ∈ L2Fs
and define (Xs,ξ,mt , Y
s,ξ,mt , Z
s,ξ,mt , Z
s,ξ,mt )s≤t≤T to be the Ft-
adapted solution to the following FBSDE
dXs,ξ,mt = α(t, Xs,ξ,m
t , Ys,ξ,mt )dt+ σdWt + εdWt
dYs,ξ,mt = −∂xH(t, Xs,ξ,m
t , ms,mt , Y
s,ξ,mt )dt+ Z
s,ξ,mt dWt + Z
s,ξ,mt dWt
Xs,ξ,ms = ξ, Y
s,ξ,mT = ∂xg(X
s,ξ,mT , m
s,mT )
(3.36)
Remark 3.2.10. The FBSDE (3.36) is a classical FBSDE with random coefficients
and not a McKean-Vlasov FBSDE since the stochastic law (ms,mt )0≤t≤T in the system
is given exogenously.
The existence and uniqueness of the FBSDE above is guaranteed under assumption
A by setting.
b(t, X, Y ) = α(t, X, Y )
F (t, X, Y ) = −∂xH(t, X,ms,mt , Y )
G(X) = ∂xg(X,ms,mT )
(3.37)
See Theorem 3.1.3. Note that the initial ξ here is arbitrary and does not necessarily
have law m. When ξ = x is a constant, (3.36) corresponds to the FBSDE from an
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 80
individual control problem over [s, T ] given (ms,mt )s≤t≤T with initial state Xs = x.
In that case, Y s,x,ms is deterministic since it is F s
s -measurable. This fact allows us to
define the following map
Uε :[0, T ]× R× P2(R) → R
(s, x,m) 7→ Y s,x,ms
(3.38)
To complete the proof of the Markov property of FBSDE (2.2.8), we are left to show
(3.27). We first state a necessary estimate for Uε.
Lemma 3.2.11. Let Uε : [0, T ]×R×P2(R) → R be as defined above, then it satisfies
1. |Uε(t, x,m)− Uε(t, x′, m′)| ≤ CK,T (|x− x′|+W2(m,m′))
2. (Uε(t, x,m)− Uε(t, x′, m)) (x− x′) ≥ 0
for all t ∈ [0, T ], x, x′ ∈ R, m,m′ ∈ P2(R), where CK,T depends only on K, T .
Proof. The first estimate directly follows from estimate (3.14) and (3.34). Next,
Let (Xt, Yt, Zt, Zt)s≤t≤T and (X ′t, Y
′t , Z
′t, Z
′t)s≤t≤T denote the solutions to the FBSDE
corresponding to the definition of Uε(s, x,m) and Uε(s, x′, m) respectively. Let ∆Xt =
Xt−X ′t and define similarly ∆Yt,∆∂xf,∆∂xg. Applying Ito’s lemma to ∆Xt∆Yt and
using convexity assumption (A3), particularly inequality (3.17), gives
E[∆Ys∆Xs] = E[∆∂xg∆XT ] + E
[∫ T
s
∆∂xft∆Xt −∆α∆Ytdt
]
≥ 0
By definition of Uε and the fact that it is deterministic, we deduce that
Uε(s, x,m)− Uε(s, x′, m)(x− x′) ≥ 0
Now we are ready to state and prove the existence of a deterministic decoupling
function thereby establishing the Markov result. Using Theorem 3.2.8 above, we can
show (3.27) using a similar argument as was done for a classical FBSDE (see Corollary
1.5 in [26] for instance).
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 81
Theorem 3.2.12. Let s ∈ [0, T ], m ∈ P2(R), ξ ∈ L2Fs, consider (ms,m
t )s≤t≤T and
(Xs,ξ,mt , Y
s,ξ,mt , Z
s,ξ,mt , Z
s,ξ,mt )s≤t≤T as defined above, then it follows that for all t ∈
[s, T ]
Ys,ξ,mt = Uε(t, Xs,ξ,m
t , ms,mt ) (3.39)
Remark 3.2.13. (3.27) in Theorem 3.2.1 follows from (3.39) by setting s = 0 and
L(ξ) = m.
Proof. We will use a similar argument as in the proof of Theorem 3.2.8 which is based
on the localization argument and global Lipschitz property. Note that R× P2(R) is
separable, hence there exist a countable disjoint set Ann∈N such that⋃∞
n=1An =
P2(R) and diag(An) < δ. Let (xn, mn) ∈ R × P2(R) be a fixed element of An, then
let
Bn = ω ∈ Ω; (Xs,ξ,mt , m
s,mt ) ∈ An (3.40)
Then by Lemma 3.2.11, we have
∑
n∈N
|Uε(t, Xs,ξ,mt , m
s,mt )− Uε(t, xn, mn)|
21Bn
≤ C1δ2 (3.41)
On the other hands, using (3.35), it follows that (Xs,ξ,mr , Y s,ξ,m
r , Zs,ξ,mr , Zs,ξ,m
r )t≤r≤T
satisfies the FBSDE
dXr = α(t, Xr, Yr)dr + σdWr + εdWr
dYr = −∂xH(r,Xr,Θt,r(ms,m
t ), Yr)dt+ ZrdWr + ZrdWr
Xt = Xs,ξ,mt , YT = gx(XT ,Θ
t,T (ms,mt ))
(3.42)
Thus, we get by (3.18), (3.34), and (3.40) that
∑
n∈N
E
[
(Y s,ξ,mt − Y
t,xn,mn
t )21Bn
]
≤ C2δ2 (3.43)
Combining (3.41) and (3.43), it follows that
E
[
(
Uε(t, Xs,ξ,mt , m
s,mt )− Y
s,ξ,mt
)2]
≤ C3δ2
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 82
Since δ is arbitrary, (3.39) holds as desired.
3.3 Master equation and connection to DPP ap-
proach
In this section, we state the connection between the decoupling function and the
generalized value function. Consequently, we obtain the master equation for Uε and
provide a verification-type theorem. We then discuss the connection to the solution
to FBSPDE of ε-MFG and FBPDE of 0-MFG.
One of the consequences of this Markov property is the fact that the ε-MFG
solution is in the feedback form. That is,
αεt = α(t, Xε
t , Yεt ) = α(t, Xε
t ,Uε(t, Xε
t ,L(Xεt |Ft))) (3.44)
With this fact, we can proceed as we did in Section 2.3 and define the value function
Vε for ε-MFG as a minimum expected cost given the current state and the current
distribution of players. We have shown that it satisfies the master equation
∂tVε(t, x,m) + H(t, x,m, ∂xV
ε(t, x,m)) +σ2 + ε2
2∂xxV
ε(t, x,m)
+ E0[
∂mVε(t, x,m)(X)(α(t, X,m, ∂xV
ε(t, X,m)))]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(3.45)
with terminal condition
Vε(T, x,m) = g(x,m)
where X is a lifting random variable, i.e. L(X) = m, and ζ is a N (0, 1)-random
variable independent of X . Furthermore, in Theorem 2.3.1, we have established the
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 83
connection between the SMP and DPP approach through the relation
Y εt = ∂xV
ε(t, Xεt ,L(X
εt |Ft)) (3.46)
Now we are ready to restate this relation through a decoupling function Uε by com-
bining (3.46) and (3.44).
Corollary 3.3.1. Assume that A,B hold and suppose that Vε ∈ C1,3,3([0, T )× R ×
P2(R)) ∩ C([0, T ] × R × P2(R)) is a classical solution to the master equation (2.8),
then the following holds.
Uε(t, x,m) = ∂xVε(t, x,m) (3.47)
Using (3.47) and the master equation for Vε, we can now get the PDE for Uε. We
state the result through a verification theorem for Uε.
Theorem 3.3.2. Suppose that Uε : [0, T ]×R×P2(R) → R is a classical solution to
∂tUε(t, x,m) + ∂xH(t, x,m,Uε(t, x,m)) + ∂yH(t, x,m,Uε(t, x,m))∂xU
ε(t, x,m)
+σ2 + ε2
2∂xxU
ε(t, x,m)− E0[
∂mUε(t, x,m)(X)α(t, X,m,Uε(t, X,m))
]
+σ2
2∂mmU
ε(t, x,m)(X)[ζ, ζ ] +ε2
2∂mmU
ε(t, x,m)(X)[1, 1]
+ ε2E0[
∂xmUε(t, x,m)(X)1
]
= 0
(3.48)
with terminal condition
Uε(T, x,m) = ∂xg(x,m)
where X is a lifting random variable, i.e. L(X) = m, and ζ is a N (0, 1)-random
variable independent of X. Suppose further that ∂xUε, ∂mUε, and Uε(t, 0, δ0) are
bounded then
(Xt, Yt, Zt, Zt)0≤t≤T =(
Xεt ,U
ε(t, Xεt , m
εt ), σ∂xU
ε(t, Xεt , m
εt ), ε
(
∂xUε(t, Xε
t , mεt ) + E
0[
∂mUε(t, Xε
t , mεt )(X
εt )]))
(3.49)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 84
is a solution to FBSDE (2.18) satisfying
E
[
sup0≤t≤T
(Xεt )
2 + sup0≤t≤T
(Y εt )
2 +
∫ T
0
(Zεt )
2 +
∫ T
0
(Zεt )
2
]
<∞ (3.50)
where (Xεt )0≤t≤T is a solution to the SDE
dXt = α(
t, Xt,L(Xt|Ft),Uε(t, Xt,L(Xt|Ft))
)
dt+ σdWt + εdWt, X0 = ξ0
mεt = L(Xε
t |Ft), and (Xεt )0≤t≤T is a lifting random variable of (Xε
t )0≤t≤T sharing a
common noise space. As a result,
αεt , α(t, Xε
t ,L(Xεt |Ft), Y
εt ) = α(t, Xε
t ,L(Xεt |Ft),U
ε(t, Xεt ,L(X
εt |Ft)))
is a solution to ε-MFG.
Proof. The proof is identical to Theorem 2.3.1.
Similarly to Vε, equation (3.48) is a second order quasilinear PDE involving the
derivative with respect to a probability measure. We will refer to this PDE as the
master equation for Uε. The PDE (3.48) will be useful in our subsequent analysis
particularly in solving the linear variational FBSDE introduced in the next section.
From a similar result in the general FBSDE setting [26], the converse, which says
that Uε defined by (3.38) is a classical solution to (3.48), is expected to hold when
the coefficients are sufficiently regular. This is still open problem for a general MFG
with common noise and was recently shown for the no common noise case in [24].
Having established the relation between Uε and Vε, we can obtain in a similar
fashion the relation between the decoupling function Uε and the solution (uε, mε, vε)
of the FBSPDE
duε(t, x) =
(
−H(t, x,mε, ∂xuε)−
σ2
2∂xxu
ε −ε2
2∂xxu
ε − ε2∂xvε
)
dt− εvεdWt
dmε(t, x) =
(
−∂x(α(t, x,mε, ∂xu
ε)mε) +σ2 + ε2
2∂xxm
ε
)
dt− ε∂xmε dWt
(3.51)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 85
corresponding to ε-MFG as described in Section 2.1. Recall that the first equation
denotes the backward stochastic HJB equation for the value function of each players
given the flow of distribution (mεt )0≤t≤T . Relating to Vε, this simply means
uε(t, x) = Vε(t, x,m0t ),
The second equation is the forward Fokker-Planck equation describing the distribution
of players’ state given all the players adopt the strategy
αε(t, x,mεt ) = α(t, x,mε
t , ∂xVε(t, x,m0
t )) = α(t, x,mεt , ∂xu
ε(t, x))
From this and Ito-Kunita formula, we also have
vε(t, x) = E0[
∂mVε(t, x,mε
t )(X)1]
Combining with (3.47), it follows that
∂xuε(t, x) = Uε(t, x,mε
t ), ∂xvε(t, x) = E
0[
∂mUε(t, x,mε
t )(X)1]
(3.52)
We are particularly interested in the case ε = 0. In that case, we have the original
MFG model proposed by Lasry and Lions in [13] through the forward backward PDE
∂tu0 = −H(t, x,m0, ∂xu
0)−σ2
2∂xxu
0, u0(T, x) = g(x,m0T )
∂tm0 = −∂x(α(t, x,m
0, ∂xu0)m0) +
σ2
2∂xxm
0, m0(0, x) = m0(x) = L(ξ0)
(3.53)
where m0t (·) = m0(t, ·). The relation (3.52) then reads
U0(t, x,m0t ) = ∂xu
0(t, x) (3.54)
We would like to emphasize the relation (3.54) as the functions U0(t, x,m0t ), ∂xU
0(t, x,m0t ),
∂mU0(t, x,m0t )(·) are the main terms that will appear in the asymptotic analysis in
the next chapter. Relation (3.54) tells us that the first two terms can be found from
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 86
the system of PDEs describing 0-MFG. The last term, which represents the sensitiv-
ity of the solution around the optimal path (m0t )0≤t≤T , is new and we shall derive its
PDE in the next section.
We end this section with the summary of the relations between all the equations
related to ε-MFG. From the DPP approach, we have a generalized value function Vε
and a solution (uε, vε, mε) from FBSPDE (2.9), which reduces to FBPDE of Lasry and
Lion when ε = 0. From the SMP approach, we have a solution (Xεt , Y
εt , Z
εt , Zt)0≤t≤T
of McKean-Vlasov FBSDE (2.18) and its decoupling function Uε defined by (3.38).
We list all the relations below; for any (t, x,m) ∈ [0, T ]× R×P2(R),
mεt = L(Xε
t |Ft)
Y εt = Uε(t, Xε
t , mεt ) = ∂xV
ε(t, Xεt , m
εt ) = ∂xu
ε(t, Xεt )
Uε(t, x,mεt ) = ∂xV
ε(t, x,mεt ) = ∂xu
ε(t, x)
Uε(t, x,m) = ∂xVε(t, x,m)
E0[
∂mUε(t, x,mε
t )(X)1]
= E0[
∂xmVε(t, x,mε
t )(X)1]
= ∂xvε(t, x)
(3.55)
It is important to note that some of these relations only hold under a verification-
type argument. That is, we need to assume that there exists a classical solution Vε to
the master equation or a classical solution (uε, vε, mε) to the FBSPDE with certain
regularity assumption in order to justify these relations rigorously. However, these
existence results are not known in the case of common noise.
3.4 Proof of lemmas, propositions, and theorems
3.4.1 Proof of Theorem 3.1.1
Proof. Let s ∈ [0, T ], A ∈ Gs and for i = 1, 2, let (X it , Y
it , Z
it , Z
it)s≤t≤T denote a
solution to FBSDE (3.2) with data (ξi, bi, Fi, Gi). Let
∆Xt = X1t −X2
t
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 87
and define similarly ∆Yt,∆Zt,∆Zt,∆ξ,∆σ,∆ε. For the coefficients, we let
∆bt = b2(t, X1t , Y
1t )− b2(t, X
2t , Y
2t ), b(·) = b1(·)− b2(·)
and defined similarly ∆Ft, F and ∆G, G. Lastly, we write
θ1(t) = (t, X1t , Y
1t )
By applying Ito’s lemma on ∆Xt∆Yt and using the monotonicity condition (H3), we
have
E[1A∆Xs∆Ys] = E[1A∆XT (∆G+ G(X1T ))]− E
[
1A
∫ T
s
(∆Ft∆Xt +∆Yt∆bt)dt
]
− E
[
1A
∫ T
s
(F (θ1t )∆Xt + b(θ1t )∆Yt +∆σ∆Zt +∆ε∆Zt)dt
]
≥ E[1A∆XT G(X1T )] + β1E
[
1A
∫ T
s
(∆bt)2dt
]
+ β2E
[∫ T
s
1A(∆Yt)2)dt
]
− E
[
1A
∫ T
s
(F (θ1t )∆Xt + b(θ1t )∆Yt +∆σ∆Zt +∆ε∆Zt)dt
]
(3.56)
Thus, we have the estimate
β1E
[
1A
∫ T
s
(∆bt)2dt
]
+ β2E
[∫ T
s
1A(∆Yt)2dt
]
≤ E
[
1A
∫ T
s
(|F (θ1t )∆Xt|+ |b(θ1t )∆Yt|+ |∆σ∆Zt|+ |∆ε∆Zt|)dt
]
+ E[
1A(|∆ξ∆Ys|+ |∆XT G(X1T )|)
]
(3.57)
We will deal with the case β1 > 0 or β2 > 0 separately. For notational convenience,
we will use the same notation CK or CK,T for all the constants that depend only on
K or on K and T while they might be different from line to line.
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 88
Case I: β2 > 0. By Ito’s lemma on ∆X2t , Young’s inequality, and Lipschitz prop-
erty of b2, we have
E[1A∆X2t ] ≤ E[1A∆ξ
2] + E
[
1A
∫ t
s
2|∆Xu(∆bu + b(θ1u))|+∆σ2 +∆ε2du
]
≤ E[1A∆ξ2] + CKE
[
1A
∫ t
s
(∆Xu)2 + (∆Yu)
2 + b(θ1u)2 +∆σ2 +∆ε2du
]
(3.58)
By Gronwall lemma, it follows that
E[1A∆X2t ] ≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2) + 1A
∫ t
s
(∆Yu)2 + b(θ1u)
2du
]
(3.59)
Next, we apply Ito’s lemma on ∆X2t again, then use Doob’s martingale inequality
and (3.59), we have
E[1A sups≤t≤T
∆X2t ] ≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2) + 1A
∫ T
s
(∆Xt)2 + (∆Yt)
2 + b(θ1t )2dt
]
≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2) + 1A
∫ T
s
(∆Yt)2 + b(θ1t )
2dt
]
(3.60)
Similarly, applying Ito’s lemma on ∆Y 2t , using Lipschitz property, we arrive at an
estimate
E
[
1A∆Y2t + 1A
∫ T
t
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A((∆XT )2 + G(X1
T )2) + 1A
∫ T
t
(∆Xu)2 + (∆Yu)
2 + F (θ1u)2du]
(3.61)
To get a bound on the sup norm, we first use Burkholder-Davis-Gundy (see Theorem
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 89
3.28 in [38]) and Young’s inequality to bound the martingale term.
E[1A sups≤t≤T
∫ T
t
∆Yu∆ZudWu] ≤ CE
[
1A
(∫ T
s
(∆Yt∆Zt)2dt
)
12
]
≤ CE
[
1A
(
sups≤t≤T
(∆Yt)2
∫ T
s
(∆Zt)2dt
)
12
]
≤ δE
[
1A sups≤t≤T
(∆Yt)2
]
+ CδE
[
1A
∫ T
s
(∆Zt)2dt
]
(3.62)
Using this fact, we can go back to Ito’s lemma on ∆Y 2t , take supremum, and use (3.62),
(3.61), (3.60) and select sufficiently small δ > 0 (independent of all the constants) to
get an estimate
E
[
1A sups≤t≤T
∆Y 2t + 1A
∫ T
s
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A
(
∆ξ2 +∆σ2 +∆ε2 + G(X1T )
2)
+ 1A
∫ T
s
(∆Yt)2 + b(θ1t )
2 + F (θ1t )2dt]
(3.63)
Combining with (3.60), and using (3.57), we then have an estimate
E
[
1A sups≤t≤T
∆X2t + sup
s≤t≤T∆Y 2
t + 1A
∫ T
s
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A
(
∆ξ2 +∆σ2 +∆ε2 + G(X1T )
2)
+ 1A
∫ T
s
(∆Yt)2 + b(θ1t )
2 + F (θ1t )2dt]
≤ CK,T
(
E
[
1A
(
∆ξ2 +∆σ2 +∆ε2 + G(X1T )
2)
+ 1A
∫ T
s
b(θ1t )2 + F (θ1t )
2dt
]
+ E
[
1A
∫ T
s
(|F (θ1t )∆Xt|+ |b(θ1t )∆Yt|+ |∆σ∆Zt|+ |∆ε∆Zt|)dt
]
+ E[
1A|∆ξ∆Ys|+ 1A|∆XT G(X1T )|]
)
(3.64)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 90
Using Young’s inequality, we have
E
[
1A sups≤t≤T
∆X2t + sup
s≤t≤T∆Y 2
t + 1A
∫ T
s
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A
(
∆ξ2 +∆σ2 +∆ε2 + G(X1T )
2)
+ 1A
∫ T
s
b(θ1t )2 + F (θ1t )
2dt]
(3.65)
as desired.
Case II: β1 > 0 This case is fairly similar to the first case except that we keep ∆bt
on the RHS instead of ∆Yt so we can apply (3.57) in the end. Similar to (3.60), we
get the bound
E[1A sups≤t≤T
∆X2t ] ≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2) + 1A
∫ T
s
(∆bt)2 + b(θ1t )
2dt
]
(3.66)
For the bound on ∆Y 2t , we proceed in the same way as we did in (3.61), but we apply
Gronwall lemma to get an estimate of the form
E
[
1A∆Y2t + 1A
∫ T
t
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A((∆XT )2 + G(X1
T )2) + 1A
∫ T
t
(∆Xu)2 + F (θ1u)
2du]
(3.67)
Combining with (3.66), we get
E
[
1A∆Y2t + 1A
∫ T
t
(∆Zu)2 + (∆Zu)
2du
]
≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2 + G(X1
T )2) + 1A
∫ T
t
(∆bu)2 + b(θ1u)
2 + F (θ1u)2du
]
(3.68)
The next step is to get a bound on the sup-norm of the backward process. Similar
to (3.62) and (3.63), we apply Ito’s lemma on ∆Y 2t , take the sup-norm, use BDG
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 91
inequality on the martingale term, and use (3.68) to arrive at the estimate
E
[
1A sups≤t≤T
∆Y 2t + 1A
∫ T
s
(∆Zt)2 + (∆Zt)
2dt
]
≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2 + G(X1
T )2) + 1A
∫ T
s
(∆bt)2 + b(θ1t )
2 + F (θ1t )2dt
]
(3.69)
Then we combine with (3.66), use (3.57), and apply Young’s inequality in the same
way at the first case to get
E
[
1A sups≤t≤T
∆X2t + 1A sup
s≤t≤T∆Y 2
t + 1A
∫ T
s
(∆Zt)2 + (∆Zt)
2dt
]
≤ CK,TE
[
1A(∆ξ2 +∆σ2 +∆ε2 + G(X1
T )2) + 1A
∫ T
s
b(θ1t )2 + F (θ1t )
2dt
]
(3.70)
as desired.
3.4.2 Proof of Theorem 3.1.2
First, we show that FBSDE (3.2) is solvable over small time interval and that the
length of the interval depends only on the Lipschitz constant K of the functionals
b, F,G. Let τ ∈ [0, T ) and ξ ∈ L2Fτ, then we let A = X ∈ H2([τ, T ];R);Xτ = ξ be
a metric space with its metric induced by the sup norm. We define the map
Φτ,ξ : A → A
(Xt)τ≤t≤T → (Xt)τ≤t≤T
as follows; given X = (Xt)τ≤t≤T ∈ H2([τ, T ];R) with initial Xτ = ξ, we solve the
following BSDE
dYt = F (t, Xt, Yt)dt+ ZtdWt + ZtdWt, YT = G(XT ) (3.71)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 92
We then define Φτ,ξ(X) to be X = (Xt)τ≤t≤T where X is given by
dXt = b(t, Xt, Yt)dt+ σdWt + εdWt, Xτ = ξ
Then clearly the fixed point of this map gives a solution to FBSDE (3.2) over [t, T ]
with initial condition Xτ = ξ. First, we note that the BSDE in the definition above
is uniquely solvable under assumption (H1),(H2) by standard SDE argument. Even
though we are dealing with functionals, the result still applies in the same way since
its proof only relies on probabilistic estimates.
Now we show that when T − τ is sufficiently small, the map Φτ,ξ is indeed a con-
traction map and we have a fixed point as desired. Suppose X1, X2 ∈ H2([τ, T ];R)
and for i = 1, 2, let (Y it , Z
it , Z
it)τ≤t≤T denote the solution to the BSDE (3.71) corre-
sponding to X i. We let
∆Xt = X1t −X2
t
and define similarly ∆Yt,∆Zt,∆Zt,∆ξ,∆σ,∆ε. For the coefficients, we let
∆bt = b(t, X1t , Y
1t )− b(t, X2
t , Y2t )
and defined similarly ∆Ft,∆G. Lastly, we write
∆Φ = Φτ,ξ(X1)− Φτ,ξ(X2)
By applying Ito’s lemma on ∆Y 2t , using Lipschitz property of F,G and Gronwall
lemma, we arrive at an estimate
E[
1A∆Y2t
]
≤ CK,TE
[
(∆XT )2 +
∫ T
t
(∆Xu)2du
]
≤ CK,TE
[
supτ≤t≤T
(∆Xt)2
] (3.72)
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 93
Using this estimate, we have
E
[
supτ≤t≤T
(∆Xt)2
]
= E
[
supτ≤t≤T
(∫ t
τ
∆budu
)2]
≤ (T − τ)E
[∫ T
τ
(∆bt)2dt
]
≤ 2K(T − τ)E
[∫ T
τ
(∆Xt)2 + (∆Yt)
2dt
]
≤ CK,T (T − τ)E
[
supτ≤t≤T
(∆Xt)2
]
for some constant CK,T depends only on K, T . Let γ =C−1
K,T
2, then for any τ ∈ [0, T )
such that T − τ ≤ γ, we have a contraction map and hence a fixed point for Φτ,ξ as
desired.
Next, we attempt to extend the solution over to [0, T ]. Let Γ denote the set of all
t ∈ [0, T ] such that the following holds: For any ξ ∈ L2Ft, the FBSDE (3.2) over [t, T ]
with initial ξ is solvable. Let τm = inf Γ, thus τm ≤ T − γ. Let δ > 0 be arbitrary
but sufficiently small so that τm + δ < T − γ2, and let τ ∈ Γ ∩ [τm, τm + δ]. Note that
T − τ ≥ γ2. Since τ ∈ Γ, for any ξ ∈ L2
Fτ, the FBSDE (3.2) is solvable over [τ , T ] and
we denote its solution by (Xξt , Y
ξt , Z
ξt , Z
ξt )τ≤t≤T .
Now we define G : L2Fτ
→ L2Fτ
by G(ξ) = Yξτ . Our goal is to show that G satisfies
(3.5) and (3.7) with G replaced by G and the constant denoted by K depends only
on K, T .
Let ξ1, ξ2 ∈ L2Fτ
and, for i = 1, 2, let (X it , Y
it , Z
it , Z
it)τ≤t≤T denote the solution to
FBSDE (3.2) corresponding to the initial X iτ = ξi. Let ∆Xt = X1
t − X2t and define
similarly ∆Yt,∆Zt,∆Zt,∆G,∆Ft,∆bt. Applying Ito’s lemma on ∆Xt∆Yt and using
(3.7) yield
E[1A∆Xs∆Ys] = E[1A∆XT∆G]− E
[
1A
∫ T
s
(∆Ft∆Xt +∆Yt∆bt)dt
]
≥ β1E
[
1A
∫ T
s
(∆bt)2dt
]
+ β2E
[∫ T
s
1A(∆Yt)2)dt
]
≥ 0
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 94
That is,
E[1A(G(ξ1)− G(ξ2))(ξ1 − ξ2)] ≥ 0 (3.73)
In addition, from (3.10) in Theorem 3.1.1, we have
E[(G(ξ1)− G(ξ2))2] ≤ KE[(ξ1 − ξ2)
2] (3.74)
where K depends only on K, T and is independent of τ . WLOG, we assume that
K ≥ K, otherwise we will take K to be the maximum of K,K instead.
Having shown (3.73) and (3.74), we can use the same argument as before to show
that there exist γ > 0 sufficiently small depending only on K such that FBSDE (3.2)
is solvable over [τ − γ, τ ] for any initial ξ ∈ L2Fτ−γ
and terminal functional G. By
definition of G, we can patch up a solution from [τ − γ, τ ] and [τ , T ] to get a solution
over [τ−γ, T ] for any arbitrary initial stateXτ−γ = ξ ∈ L2Fτ−γ
and terminal functional
G. As a result, we have shown that τ − γ ∈ Γ.
Note that γ is independent of δ. Thus, if τm > 0, then we can select δ sufficiently
small so that τ − γ < τm yielding a contradiction. Therefore, τm = 0 and, in that
case, we can select δ so that τ − γ = τm = 0. That is, 0 ∈ Γ so FBSDE (3.2) is
solvable over [0, T ].
Next, we show that the solution is in fact unique. For i = 1, 2, let (X it , Y
it , Z
it , Z
it)0≤t≤T
denote solutions to FBSDE (3.2) with the same initial ξ ∈ L2F0. Let ∆Xt = X1
t −
X2t and we define similarly ∆Yt,∆Zt,∆Zt,∆G,∆bt,∆Ft. Applying Ito’s lemma on
∆Xt∆Yt and using the fact that ∆X0 = 0 (due to the same initial ξ) and (H3), it
then follows that
β1E
[∫ T
s
(∆bt)2dt
]
+ β2E
[∫ T
s
(∆Yt)2)dt
]
≤ 0
In either cases (β1 > 0 or β2 > 0), it follows easily that ∆Xt = 0 for all t ∈ [0, T ] a.s.
and thus, so does ∆Yt,∆Zt,∆Zt.
The estimate (3.11) follows from Theorem 3.1.1 by taking the difference to the
zero solution (when all coefficients b, F,G are zero).
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 95
3.4.3 Proof of Theorem 3.2.8
Proof. Let η ∈ L2Fs
with L(η) = m and (Xs,ηt , Y
s,ηt , Z
s,ηt , Z
s,ηt )s≤t≤T denote the solution
to FBSDE (3.36) corresponding to the definition of Θs,u, so
Θs,t(m) = L(Xs,ηt |F s
t ), Θs,u(m) = L(Xs,ηu |F s
u)
Since P2(R) is separable, for any δ > 0, there exist a sequence of disjoint Borel
measurable subsets Ann∈N of P2(R) such that diam(An) < δ and ∪n∈NAn = P2(R).
Let mn be a representative element of An so that W2(m,mn) < δ for all m ∈ An. Let
Bn = ω ∈ Ω;L(Xs,ηt |F s
t )(ω) ∈ An. Consider
ξ ,∑
n∈N
1Bnξn
where ξn ∈ L2Ft
has law mn and independent of Ft, thus independent of Bn. That is,
L(ξn|F st ) = L(ξn) = mn
Then it follows by construction that
W2(L(ξ|Fst ),L(X
s,ηt |F s
t )) < δ
Using this type of discretization and Lemma 1.1.6, we can redivide An further and
proceed sequentially to construct a sequence ξNN∈N of the form
ξN ,∑
n∈N
1Bn,Nξn,N
such that ξNN∈N is Cauchy in L2Ft, ξn,N is independent of F s
t and
W2(L(ξN |F s
t ),L(Xs,ηt |F s
t )) <1
N
CHAPTER 3. MEAN FIELD GAMES WITH COMMON NOISE 96
Let ξ = limN→∞ ξN in L2Ft, then we have
L(ξ|F st ) = L(Xs,η
t |F st ) = Θs,t(m) (3.75)
Now consider FBSDE (t, ξ, F srt≤r≤T ) and denote its solution by (X t,ξ
r , Y t,ξr , Zt,ξ
r , Zt,ξr ).
By (3.75) and Theorem 3.2.6, it follows that
L(X t,ξu |F s
u) = L(Xs,ηu |F s
u) = Θs,u(m) (3.76)
Let
XNr ,
∑
n∈N
1Bn,NXn,N
r
where (Xn,Nr )t≤r≤u is a solution to FBSDE (t, ξn,N , F t
rt≤r≤T ). It is easy to check
that (XNr )t≤r≤u is a solution to FBSDE (t, ξN , F s
rt≤r≤T ) with initial ξN .
Note that Xn,Nr is F t
r-measurable which is independent of Ft, hence independent
of Bn. Thus, we have
L(XNu |F s
u) =∑
n∈N
1Bn,NL(Xn,N
u |F tu) =
∑
n∈N
1BnΘt,u(L(ξn,N))
Taking limit in L2Fu
as N → ∞ both sides, it follows from the fact that E[(ξ−ξN)2] →
0 and from estimate (3.10) that
L(X t,ξu |F s
u) = Θt,u(L(ξ|F st ))
Combine with (3.75) and (3.76), we get (3.35) as desired.
Chapter 4
Asymptotic analysis of mean field
games
In this chapter, we analyze ε-MFG model when ε is small. When the common noise
is small, it is reasonable to seek an approximate solution using the finite-dimensional
information from the 0-MFG. This chapter is organized as follows. In Section 4.1, we
consider the linear variational FBSDE of mean-field type describing the ε-first order
approximation term. We show that it is uniquely solvable and prove the convergence
result. In Section 4.2, we show that the approximate solution using the first order ap-
proximation is ε2-Nash equilibrium. In Section 4.3, we discuss the Gaussian property
of the solution and discuss the decoupling function. Lastly, in Section 4.4, we find
the explicit solution to this FBSDE and compute the covariance functions assuming
regularity of the decoupling function from 0-MFG.
To reduce notations, we will assume throughout this chapter that
f(t, x,m, α) =α2
2
while keeping a general cost function g. We would like to note that all the results in
this chapter still hold for a running cost function f satisfying A,B,C by doing the
same analysis on f as will be done on g. Working with this running cost, we now
97
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 98
have
α(t, x,m, y) = −y, ∂xH(t, x,m, y) = 0 (4.1)
4.1 Linear variational FBSDE
In the previous chapters, we have shown that, under linear-convexity framework,
finding a solution of ε-MFG is equivalent to solving the corresponding McKean-Vlasov
FBSDE (2.2.8). We also show that this FBSDE is in fact uniquely solvable under
assumptions A, B. Let us denote the solution to ε-MFG by (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T .
Under 4.1, the McKean-Vlasov FBSDE reads
dXεt = −Y ε
t dt+ σdWt + εdWt
dY εt = Zε
t dWt + Zεt dWt
Xε0 = ξ, Y ε
T = ∂xg(XεT ,L(X
εT |FT ))
(4.2)
Solving this FBSDE yields the ε-MFG solution by setting
αεt = −Y ε
t
From the discussion in Section 3.3, we see that solving the 0-MFG problem for
(X0t , Y
0t , Z
0t , Z
0t )0≤t≤T requires us to find U0(t, x,m0
t ), which by (3.54) is reduced to
solving a system of PDEs. However, when adding common noise, we need to seek Uε,
a solution to the master equation (3.48). Instead of solving this infinite dimensinonal
equation, our goal here is to consider the approximation (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T around
(X0t , Y
0t , Z
0t , Z
0t )0≤t≤T when the common noise is small. Equivalently, we would like
to consider the limit as ε→ 0 of
Xεt −X0
t
ε,
Y εt − Y 0
t
ε(4.3)
As a result, we need additional regularity assumption on g. Thus, from now and
throughout the rest of this chapter, we will also assume C which is the following.
(C1) ∂xxg and ∂xmg exist and are continuous and bounded. Denote their bounds
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 99
by the same constant K.
We derive formally the FBSDE for the limit of (4.3) as ε → 0. First, we write
∆Xεt =
Xεt−X0
t
εand denote similarly ∆Y ε
t ,∆Zεt ,∆Z
εt , then they satisfy
d∆Xεt = −∆Y ε
t dt+ dWt
d∆Y εt = ∆ZtdWt +∆ZtdWt
∆Xε0 = 0, ∆Y ε
T =∂xg(X
εT ,L(X
εT |FT ))− ∂xg(X
0T ,L(X
0T |FT ))
ε
(4.4)
Formally taking ε→ 0, we get the following linear variational FBSDE
dUt = −Vtdt+ dWt
dVt = QtdWt + QtdWt
U0 = 0, VT = ∂xxg(X0T , m
0T )UT + E
0[∂xmg(X0T , m
0T )(X
0T )UT ]
(4.5)
where
m0t = L(X0
t |Ft) = L(X0t )
and X0 and U are identical copies ofX0 and U in (Ω, F , P) and E0[·] is the expectation
with respect to P0 only. We can write the terminal function explicitly as
E0[∂xmg(X
0T , m
0T )(X
0T )UT ] =
∫
Ω0
∂xmg(X0T (ω
0), m0T )(X
0T (ω
0))UT (ω0, ω)dP0(ω0)
where we suppress the ω in X0T , X
0T as they do not depend on it. We can see that the
term E0[∂xmg(X0T , m
0T )(X
0T )UT ] is a mean-field term that couples UT (ω
0, ·);ω0 ∈ Ω0
together.
We would like to note here that the description of each term as a function of path
is in the nature of SMP approach which describes the optimal control in an open-loop
form. Intuitively, each different path of Wt being coupled is simply a result of the
interaction between each players.
This FBSDE is called a linear variational FBSDE. Mathematically, it is simply a
linearization of the McKean-Vlasov FBSDE around the 0-MFG solution (X0t , Y
0t , Z
0t , Z
0t )0≤t≤T .
Notice that (4.5) is not in the standard form of McKean-Vlasov FBSDE as it involves
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 100
not just the law of Ut, but the joint law of (Ut, X0t ).
4.1.1 Wellposedness result
We first show that the linear variational FBSDE (4.5) is uniquely solvable. Derived
from FBSDE with monotone functional, it still possess the same monotonicity prop-
erty. As a result, the wellposedness theorem of FBSDE with monotone functional
(Theorem 3.1.2) can be applied in the same way. We will state the result for linear
FBSDE over [s, T ].
Theorem 4.1.1. Let (X0t , Y
0t , Z
0t , Z
0t )s≤t≤T denote the solution to McKean-Vlasov
FBSDE (2.18) corresponding to 0-MFG. Assume A,B,C hold. There exists a unique
adapted solution (Ut, Vt, Qt, Qt)s≤t≤T to FBSDE
dUt = −Vtdt+ dWt
dVt = QtdWt + QtdWt
Us = 0, VT = ∂xxg(X0T , m
0T )UT + E
0[∂xmg(X0T , m
0T )(X
0T )UT ]
(4.6)
satisfying
E
[
sups≤t≤T
[U2t + V 2
t ] +
∫ T
s
[Q2t + Q2
t ]dt
]
≤ CK,T (4.7)
Proof. We will apply Theorem 3.1.2. Define G : L2FT
→ L2FT
by
G(ξ) = ∂xxg(X0T , m
0T )ξ + E
0[
∂xmg(X0T , m
0T )(X
0T )ξ]
where X0T , ξ are identitcal copies of X
0T , ξ in the space (Ω, F , P), E0 is the expectation
with respect to Ω0 only (or equivalent conditional on FT ). By assumption (C1), there
exist a constant C depends on K such that
E[
1A(G(ξ)−G(ξ′))2]
≤ 2KE[1A(ξ − ξ)2]
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 101
for any ξ, ξ′ ∈ L2FT
, A ∈ FT . From the monotonicity condition (B3),
1
δ2E
[
∂xg(X0T + δξ,L(X0
T + δξ|FT ))− ∂xg(X0T ,L(X
0T |FT )))ξ
∣
∣
∣FT
]
≥ 0
Since ∂xxg and ∂xm are bounded by assumption (C1), we can take δ → 0 which yields,
for any ξ ∈ L2FT
,
E[G(ξ)ξ|FT ] ≥ 0
Since G is linear, we have
E[(G(ξ)−G(ξ′))(ξ − ξ′)|FT ] ≥ 0
or equivalently,
E[1A(G(ξ)−G(ξ′))(ξ − ξ′)] ≥ 0
or any ξ, ξ′ ∈ L2FT
, A ∈ FT . Having shown the Lipschitz and monotonicity property
of G, we can apply Theorem 3.1.2 as desired.
4.1.2 Convergence result
We are now ready to state our first main result of this chapter which justifies the
formal limit taken above. As a result, we characterize the ε-first order expansion
terms of ε-MFG solution as the solution to linear variational mean-field type FBSDE
(4.5).
Theorem 4.1.2. Assume A,B,C hold, for all ε > 0, let (Xεt , Y
εt , Z
εt , Z
εt )s≤t≤T denote
the solution to McKean-Vlasov FBSDE (4.2) corresponding to ε-MFG and (Ut, Vt, Qt, Qt)0≤t≤T
denote the solution to (4.6), then there exist a constant CK,T dependent only on K, T
such that
E sups≤t≤T
[
(
Xεt −X0
t
ε− Ut
)2
+
(
Y εt − Y 0
t
ε− Vt
)2]
≤ CK,Tε2 (4.8)
Proof. The proof is based on the estimates (3.11) and (3.10) with the bounded second
derivative of the cost function g. See Section 4.5.1 for the proof.
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 102
4.2 Approximate Nash equilibrium
In the previous section, we have shown that
αεt − α0
t
ε=
−Y εt + Y 0
t
ε→ −Vt as ε → 0
where the limit is in H2([0, T ];R). Using this result, we can construct the first order
approximate strategy by
βεt , α0
t − εVt (4.9)
for all t ∈ [0, T ]. We would like to see if (βεt )0≤t≤T serves as a good approximation.
Being a game, an appropriate notion of approximation is required. In this case,
it is reasonable to assume that each player adopts this approximate solution, and
analyze the gap between the expected cost under this set of strategies and that the
optimal cost each player could have obtained given that others adopt this approximate
strategy. For an exact Nash equilibrium, this gap is precisely zero by definition as
every player is optimal given other players’ strategy. This notion of approximate
optimality is called δ-Nash equilibrium. In the finite player game, it is defined as
follows.
Definition 4.2.1. Under the same notations as defined in Section 1.1, for the N-
player game, a set of admissible strategies (αit)0≤t≤T,1≤i≤N is called a δ-Nash equilib-
rium if for each i = 1, 2, . . . , N ,
J i(
αi|(αj)j 6=i
)
≤ J i(
β|(αj)j 6=i
)
+ δ
for all β = (βt)0≤t≤T ∈ H2([0, T ];R) where J i(·) denote the cost function of player i.
To go from a finite-player symmetric game to a continuum limit version, we for-
mally take N → ∞, assume that each player adopts the same strategy, and use a
single player as a representative player. As a result, we can define an approximate
Nash equilibrium similarly for MFG as follows;
Definition 4.2.2. Under the same notations as defined in Section 1.1, an admissible
strategy α = (αt)0≤t≤T ∈ H2([0, T ];R) is called a δ-Nash equilibrium for ε-MFG
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 103
problem if
J ε(α|mα) ≤ J ε(β|mα) + δ
for all β = (βt)0≤t≤T where J ε(·) denotes the cost function and mαt denotes the
conditional law of Xαt with (Xα
t )0≤t≤T being the state process corresponding to α.
Remark 4.2.3. By definition, an ε-MFG solution is a 0-Nash equilibrium for ε-MFG
problem.
Remark 4.2.4. It is conventionally called ε-Nash equilibrium. We use parameter δ
here to avoid confusion with parameter ε denoting the level of common noise.
The notion of approximate Nash equilibrium is important in the theory of stochas-
tic games with infinite horizon. In many problems, there is no exact Nash equilibrium
while there exists a δ-Nash equilibrium for any δ > 0. It is also a widely used notion in
an area called algorithmic game theory. In algorithmic game theory, we are interested
in finding polynomial time algorithms that yield an approximate Nash equilibrium
solution for problems where finding exact Nash equilibrium is computationally ex-
pensive.
In MFG, this notion is used mainly in the study of the relation between an MFG
and a symmetric N -player stochastic differential game. Recall that the motivation
for considering an MFG model is in its application for finding a good approximate
strategy for an N -player game when N is large. In [17], Carmona and Delarue showed,
under a linear-convexity MFG model without common noise, the 0-MFG strategy is
εN -Nash equilibrium for the corresponding N -player game where εN ∼ O(N−1/(d+4))
where d is the dimension of the underlying Euclidean space. See also [22, 37, 39] for
other similar results. The converse, which asks whether the Nash equilibrium from
N -player game converges to a corresponding MFG solution, is also of interest and is
more challenging. For interested readers, we refer to [27] and reference therein for
results in this direction all of which are for MFG models without common noise.
In this work, we are only concerned with the model at the continuum limit. We are
particularly interested in an approximate solution for ε-MFG using the information
from 0-MFG solution. Our main result for this section is the following theorem
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 104
Theorem 4.2.5. Assume A,B,C hold. For ε > 0, let αε = (αεt )0≤t≤T denote the
solution to the ε-MFG and (Ut, Vt, Qt, Qt)0≤t≤T denote the solution to the linear vari-
ation FBSDE (4.5). Define a first order approximate strategy βε = (βεt )0≤t≤T by
βεt , α0
t − εVt (4.10)
Then βε is an ε2-Nash equilibrium for ε-MFG as defined in 4.2.2.
Proof. See Section 4.5.2
4.3 Gaussian property and decoupling function
Having characterized the first order approximation of ε-MFG solution as the solution
of a linear variational FBSDE of mean-field type, we now proceed to analyze some
properties of the solution (Ut, Vt)0≤t≤T .
4.3.1 Centered Gaussian process
While the FBSDE (4.5) describing them seems complicated as it involves both the
individual noise and common noise, this is simply in the nature of SMP approach
as it describes the control in the open-loop form (a function of path) instead of the
closed-loop feedback form (a function of state). However, if we only analyze the effect
from the common noise, or equivalently if we fix a path ω ∈ Ω0 of (Wt)0≤t≤T , hence of
(X0t )0≤t≤T , and look at the distribution of (Ut(ω, ·), Vt(ω, ·))0≤t≤T with respect to the
common noise ω ∈ Ω only, then (Ut(ω, ·), Vt(ω, ·))0≤t≤T are simply centered Gaussian
processes.
Theorem 4.3.1. Let (Ut, Vt, Qt, Qt)0≤t≤T denote the solution to FBSDE (4.5). Then
for all ω ∈ Ω0, (Ut(ω, ·), Vt(ω, ·))0≤t≤T are Gaussian processes in (Ω, Ft0≤t≤T , P)
with mean zero.
Proof. The Gaussian property is from the linearity of FBSDE, while the mean zero
is proved by taking the conditional expectation and show that the resulting system
has a unique solution which is zero. See Section 4.5.3 for the proof.
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 105
Having shown that (Ut, Vt)0≤t≤T are Gaussian processes with mean zero, we would
like to compute its covariance functions. In a classical linear SDE, we can achieve
this by using Ito’s lemma on the SDE to derive a deterministic equation describing
the covariance function. However, when dealing with FBSDE, even though it is a
linear system, to describe the covariance structure of (4.5) will require the decoupling
function.
4.3.2 Decoupling function
Despite being a linear FBSDE, the mean field term E0[
∂xmg(X0T , m
0T )(X
0T )UT
]
in the
terminal condition makes solving this FBSDE explicitly a nontrivial task. However,
we proceed in the similar way as solving a classical FBSDE by attempting to find a
decoupling function describing the relation between Vt and Ut. To do so, recall that
we have a decoupling function Uε which satisfies the relation
Y εt = Uε(t, Xε
t ,L(Xεt |Ft))
Therefore, we have
Vt = limε→0
Y εt − Y 0
t
ε
= limε→0
Uε(t, Xεt ,L(X
εt |Ft))− U0(t, X0
t ,L(X0t ))
ε
= limε→0
Uε(t, Xεt ,L(X
εt |Ft))− U0(t, Xε
t ,L(Xεt |Ft)) + U0(t, Xε
t ,L(Xεt |Ft))− U0(t, X0
t ,L(X0t ))
ε
= limε→0
Uε(t, Xεt ,L(X
εt |Ft))− U0(t, Xε
t ,L(Xεt |Ft))
ε+ lim
ε→0
U0(t, Xεt ,L(X
εt |Ft))− U0(t, X0
t ,L(X0t ))
ε
where the limit is in L2F . The following proposition shows that the first part is in fact
zero.
Theorem 4.3.2. Let Uε denote the decoupling function of FBSDE (2.18) as defined
in (3.27), then the following holds;
limε→0
Uε(t, x,m)− U0(t, x,m)
ε= 0 (4.11)
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 106
uniformly in (t, x,m) ∈ [0, T ]× R× P2(R).
Proof. See Section 4.5.4
The interpretation of (4.11) is the following; to approximate the ε-MFG solu-
tion at the first order, we simply need to use 0-MFG solution along the trajectory
(t, Xεt ,L(X
εt |Ft)), i.e.
αεt = −Y ε
t = −Uε(t, Xεt ,L(X
εt |Ft)) ≈ −U0(t, Xε
t ,L(Xεt |Ft))
However, we would like to emphasize that we do not usually know U0(t, x,m) for all
(t, x,m) since that would require us to solve the master equation (2.8) or (3.48) which
is infinite-dimensional problem and is non-trivial to do so. On the other hand, we can
solve for U0(t, x,m) along the trajectory (m0t )0≤t≤T where m0
t = L(X0t ) corresponds
to the 0-MFG solution, since U0(t, x,m0t ) is simply the gradient of the solution of
FBPDE (2.10) of Lasry and Lions. So unless we know the function U0(t, x,m), this
process means that to get our optimal control at time t, we need to resolve 0-MFG
problem over [t, T ] with initial mεt = L(Xε
t |Ft). This method is very computationally
expensive, so we will instead approximate U0 at the current state (t, Xεt ,L(X
εt |Ft))
around (t, X0t , m
0t ).
In fact, it is not necessary to approximate at X0t if the current state Xε
t is observ-
able. In other words, making use of (4.11), we can get a slightly simpler approximation
of αεt as follows. Assuming that ∂mU0 exists and is bounded,
αεt = −Y ε
t = −Uε(t, Xεt ,L(X
εt |Ft))
= −U0(t, Xεt ,L(X
εt |Ft)) + o(ε)
= −U0(t, Xεt , m
0t ) + εE0[∂mU
0(t, X0t , m
0t )(X
0t )Ut] + o(ε)
(4.12)
From both (4.12) and (4.13) below, we see that in any case, the crucial term in our
approximation is E0[∂mU0(t, X0t , m
0t )(X
0t )Ut].
The master equation for MFG with common noise is rather difficult and most of
the discussions so far has been formal. There is a recent paper by Chassagneux et
al.[24] which shows the existence and uniqueness of a classical solution to the master
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 107
equation (3.48) when ε = 0, MFG without common noise, which implies that U0 is
continuously differentiable in (x,m). To apply this result, we need assumption D
which provides additional regularity on the cost functions involving second derivative
with respect to the m-argument. Combining with the discussion above, we have the
decoupling function for linear variational FBSDE (4.5) as follow.
Proposition 4.3.3. Assume that A,B,C,D holds. Let (Ut, Vt, Qt, Qt)0≤t≤T denote
the unique solution to FBSDE (4.5), then
Vt = ∂xU0(t, X0
t , m0t )Ut + E
0[∂mU0(t, X0
t , m0t )(X
0t )Ut] (4.13)
Proof. Under A,B,C,D, it follows from [24] (see specifically Theorem 5.3) that U0 is
continuously differentiable. From Theorem 4.1.2 and 4.3.2, we have that
Vt = limε→0
U0(t, Xεt ,L(X
εt |Ft))− U0(t, X0
t ,L(X0t ))
ε
Note that from Lemma 3.2.11, ∂xU0, ∂mU0 are bounded. Using Theorem 4.1.2, it
follows that
limε→0
U0(t, Xεt ,L(X
εt |Ft))− U0(t, X0
t ,L(X0t ))
ε= ∂xU
0(t, X0t , m
0t )Ut+E
0[∂mU0(t, X0
t , m0t )(X
0t )Ut]
where all the limits above are in L2F .
Plugging (4.13) back in and we have decoupled the FBSDE (4.5) and reduced it
to solving the following forward SDE
dUt = −[
∂xU0(t, X0
t , m0t )Ut + E
0[∂mU0(t, X0
t , m0t )(X
0t )Ut]
]
dt+ dWt, U0 = 0
(4.14)
Recall that (Ut)0≤t≤T is a copy of (Ut)0≤t≤T sharing the common noise space. We
would like to point out that this SDE is non-standard due to the mean field term
in the form of the conditional law of Ut. Even if we have explicitly the functional
∂mU0(t, x,m0t )(·), solving this linear mean-field SDE is non-trivial since it involves
the joint law of (Ut, X0t ).
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 108
Next, we proceed by letting γt, βs,t, ηt ∈ L2Ft, η : [0, T ]×R× Ω → R be defined as
γt , ∂xU0(t, X0
t , m0t )
βs,t , e−∫ t
sγrdr
η(t, x, ω) , E0[∂mU
0(t, x,m0t )(X
0t )Ut]
ηt , η(t, X0t , ω)
then from (4.13) and (4.14);
Ut = −
∫ t
0
βs,tηsds−
∫ t
0
βs,tdWs
Vt = γtUt + ηt
(4.15)
is a solution to FBSDE (4.5). The terms βs,t, γt can be obtained from a 0-MFG
solution straightforwardly, by solving the system of FBPDE (2.10) and using the
relation
U0(t, x,m0t ) = ∂xV
0(t, x,m0t ) = ∂xu
0(t, x)
We are now left to analyze a more non-trivial term, the random function η(t, x, ω)
which involves analyzing the derivative with respect to the m-argument of U0.
4.4 Explicit solution and covariance function
At this point, we have proved all the results using only assumptions A,B,C,D. To
proceed and further analyze the linear mean-field FBSDE (4.5), particularly the term
η(t, x), extra regularity assumptions on the decoupling function of 0-MFG U0 is re-
quired to derive the equations for η(t, x) and compute the covariance functions. While
there are some results that give sufficient conditions for the differentiability with re-
spect to (t, x) when m = m0t since it is related to u0(t, x) from Lasry and Lions’s
FBSDE, the only work that provides similar result for the derivative with respect to
m-argument for U0 is [24]. As mentioned above, they showed existence and unique-
ness of a classical solution to the master equation U0 when ε = 0, as a by product,
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 109
showing that U0 is continuously differentiable in m. However, their result is still not
sufficient for our application below. In this section, we will proceed formally assum-
ing that U0 is sufficiently regular with bounded derivatives. Our goal is to derive the
equation for η(t, x) and compute the covariance function of the process (Ut, Vt)0≤t≤T .
SPDE for η(t, x, ω)
Notice that η involves the law of Ut, and Ut depends on η in its dynamic (4.14). To
derive an equation for η, we need to resort to the master equation. Having shown
(4.8), we write
η(t, x, ω) = E0[∂mU
0(t, x,m0t )(X
0t )Ut] = lim
ε→0
U0(t, x,mεt (ω))− U0(t, x,m0
t )
ε
where mεt = L(Xε
t |Ft), m0t = L(X0
t ). Now we apply Ito’s lemma on each term and
use the master equation (3.48);
dU0(t, x,mεt ) =
[
∂tU0(t, x,mε
t )− E0[
∂mU0(t, x,mε
t )(Xεt )(U
ε(t, Xεt , m
εt ))]
+σ2
2∂mmU
0(t, x,mεt )(X
εt )[ζ, ζ ]
]
dt+ εE0[∂mU0(t, x,mε
t )]dWt
=
[
U0(t, x,mεt )∂xU
0(t, x,mεt )−
σ2
2∂xxU
0(t, x,mεt )
]
dt
+ εE0[
∂mU0(t, x,mε
t )(Xεt )]
dWt
and
dU0(t, x,m0t ) =
[
U0(t, x,m0t )∂xU
0(t, x,m0t )−
σ2
2∂xxU
0(t, x,m0t )
]
dt
Thus, taking the difference and ε→ 0, we arrive at the SPDE for η;
dη(t, x) =
[
η(t, x)∂xU0(t, x,m0
t ) + ∂xη(t, x)U0(t, x,m0
t )−σ2
2∂xxη(t, x)
]
dt
− E0[
∂mU0(t, x,m0
t )(X0t )]
dWt
(4.16)
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 110
Due to the fact that mε(0, x) = m0(0, x) = m0(x), we have the initial condition
η(0, x) = 0
Notice that the diffusion term in the dynamic of η is deterministic due to the absence
of common noise. We define w : [0, T ]× R → R by
w(t, x) , E0[
∂mU0(t, x,m0
t )(X0t )]
(4.17)
To fully describe the dynamic for η, we need to solve for a deterministic w(t, x). Note
that
w(t, x) = E0[
∂mU0(t, x,m0
t )(X0t ) · 1
]
= limε→0
U0(t, x,L(X0t + ε))− U0(t, x,L(X0
t ))
ε
That is, w(t, ·) measures the sensitivity of the solution of 0-MFG PDE system at time
t with respect to the spatial shift of the equilibrium distribution m0t . However, to
do so, we will need to compute ∂mU0(t, x,m0t )(·). We will refer to this term as the
sensitivity functional of 0-MFG and we shall derive its equation in the next section.
4.4.1 Sensitivity functional
Our goal in this section is to study ∂mU0(t, x,m0t ) or more specifically to compute
w(t, x) = E0[
∂mU0(t, x,m0t )(X
0t ) · 1
]
. This function represents the sensitivity of 0-
MFG solution with respect to a perturbation on m0t . As it turns out, the Gauteax
derivative of U0(t, x, ·) is more suitable for this calculation. See [10] for instance. In
this framework, we are working with the density function, so we denote the density
function of m0t by the same notation m0(t, x). The existence of a density function for
m0t is guaranteed for t > 0 given that σ > 0 (recall that σ is the diffusion term of the
individual noise). We assume that U0 is Gauteax differentiable, that is , there exist
a unique ∂mU0(t, x,m) ∈ L2(R;R) such that
limε→0
U0(t, x,m+ εψ)− U0(t, x,m)
ε=
∫
R
∂mU0(t, x,m)(z)ψ(z)dz
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 111
for all test function ψ ∈ C∞(R;R) with compact support and∫
Rψ(x)dx = 0. We
use the notation ∂m to avoid confusion with the notion of the derivative of the lifting
function. We are specifically interested in this functional along m = (m0t )0≤t≤T , so
we let h(t, x, z) = ∂mU0(t, x,m0t )(z). Also, we let 〈∂mU0(t, x,m0
t )(·), ψ〉 or simply
〈∂mU0, ψ〉 to be∫
R∂mU0(t, x,m)(z)ψ(z)dz. Using this notion, the master equation
for U0 reads
∂tU0(t, x,m)− U0(t, x,m)∂xU
0(t, x,m) +σ2
2∂xxU
0(t, x,m)
+
⟨
∂mU0(t, x,m), ∂x(U
0(t, x,m)m) +σ2
2∂xxm
⟩
= 0
with terminal condition
U0(T, x,m) = ∂xg(x,m)
See [10] for the master equation for V0(t, x,m) and recall that U0 = ∂xV0. Using this
fact, we have
d
dtU0(t, x,m0
t + εψ)
= ∂tU0(t, x,m0
t + εψ) +
⟨
∂mU0,−∂z(−U0(t, z,m0
t )m0t ) +
σ2
2∂zzm
0t
⟩
= U0(t, x,m0t + εψ)∂xU
0(t, x,m0t + εψ)−
σ2
2∂xxU
0(t, x,m0t + εψ)
+⟨
∂mU0, ∂z(U
0(t, z,m0t )m
0t ) +
σ2
2∂zzm
0t − ∂z(U
0(t, z,m0t + εψ)(m0
t + εψ))
−σ2
2∂zz(m
0t + εψ)
⟩
(4.18)
andd
dtU0(t, x,m0
t ) = U0(t, x,m0t )∂xU
0(t, x,m0t )−
σ2
2∂xxU
0(t, x,m0t )
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 112
where we use the z-variable to denote the variable in the inner product. Therefore,
we have
d
dt
⟨
∂mU0, ψ⟩
=⟨
∂mU0, ψ⟩
∂xU0 +
⟨
∂xmU0, ψ⟩
U0 −σ2
2
⟨
∂xxmU0, ψ⟩
+
⟨
∂mU0,−∂z
(⟨
∂mU0(t, z,m0
t )(k), ψ(k)⟩
m0(t, z) + U0(t, z,m0t )ψ)
−σ2
2∂zzψ
⟩
(4.19)
If we write h(t, x, z) = ∂mU0(t, x,m0t )(z) as mentioned above and recall that U0(t, x,m0
t ) =
∂xu0(t, x) where u0(t, x) is the solution to FBPDE (2.10) of Lasry and Lions, then
the equation above reads
∂t
⟨
h(t, x, z), ψ(z)⟩
=⟨
h(t, x, z), ψ(z)⟩
∂xxu0(t, x) +
⟨
∂xh(t, x, z), ψ(z)⟩
∂xu0 −
σ2
2
⟨
∂xxh(t, x, z), ψ(z)⟩
+
⟨
∂zh(t, x, z),⟨
h(t, z, k), ψ(k)⟩
m0(t, z) + ∂zu0(t, z)ψ(z)−
σ2
2ψzz
⟩
(4.20)
with terminal condition
⟨
h(T, x, z), ψ(z)⟩
=⟨
∂mxg(x,m0T )(z), ψ(z)
⟩
Note that the domain for this functional is only those test functions ψ such that∫
Rψ(x)dx = 0. We can then write an equation for h(t, x, z) by formally plugging in
ψ = δ′z which yields
∂tzh(t, x, z)
= ∂zh(t, x, z)∂xxu0(t, x) + ∂xzh(t, x, z)∂xu
0 −σ2
2∂xxzh(t, x, z)
+
∫
R
∂zh(t, x, u)∂zh(t, u, z)m0(t, u)du+ ∂z
(
∂zh(t, x, z)∂zu0(t, z)
)
−σ2
2∂zzzh(t, x, z),
h(T, x, z) = ∂mxg(x,m0T )(z)
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 113
Let h : [0, T ]× R× R be a weak-derivative of h with respect to z, then h(t, x, z) is a
weak solution to
∂th(t, x, z)
= h(t, x, z)∂xxu0(t, x) + ∂xh(t, x, z)∂xu
0 −σ2
2∂xxh(t, x, z)
+
∫
R
h(t, x, u)h(t, u, z)m0(t, u)du+ ∂z(h(t, x, z)∂zu0(t, z))−
σ2
2∂zzh(t, x, z)
h(T, x, z) = ∂z∂mxg(x,m0T )(z)
(4.21)
To find w(t, x), letmεt denote the right shift by ε ofm
0t , that is, m
ε(t, x−ε) = m0(t, x).
Recall the definition of w(t, x);
w(t, x) = limε→0
U0(t, x,mεt )− U0(t, x,m0
t )
ε
Assuming ∂mU0 is continuous in m and ∂xxm0 is bounded, then we have
U0(t, x,mεt )− U0(t, x,m0
t )
ε=
⟨
∂mU0(t, x,mλ∗
t ),mε
t −m0t
ε
⟩
for some λ∗ ∈ [0, 1] where mλt = m0
t + λ(mεt −m0
t ). Thus, we have
U0(t, x,mεt )− U0(t, x,m0
t )
ε=⟨
∂mU0(t, x,m0
t ,−∂zm0t
⟩
= −
∫
R
h(t, x, z)∂zm0(t, z)dz
=
∫
R
h(t, x, z)m0(t, z)dz
where h is a solution to (4.21)
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 114
4.4.2 Explicit solution
Having shown that the random function η satisfies a linear SPDE, we can solve (4.16)
by means of Duhamel’s Principle. Let P sf denote a solution to the problem;
∂tφ(t, x) = ∂xφ(t, x)U0(t, x,m0
t ) + φ(t, x)∂xU0(t, x,m0
t )−σ2
2∂xxφ(t, x), (t, x) ∈ [s, T ]× R
φ(s, x) = f(x), x ∈ R
(4.22)
Then, from the initial condition η(0, x) = 0, it follows that
η(t, x) = (P 0η(0, ·))(t, x) +
∫ t
0
(P sw(s, ·))(t, x)dWs =
∫ t
0
(P sw(s, ·))(t, x)dWs (4.23)
We let
ψ(s, t, x) , (P sw(s, ·))(t, x), 0 ≤ s ≤ t ≤ T, x ∈ R
Combining with (4.15), we have an explicit solution to FBSDE (4.5)
Ut = −
∫ t
0
∫ s
0
βs,tψ(r, s,X0s )dWrds−
∫ t
0
βs,tdWs
Vt = ∂xU0(t, X0
t , m0t )Ut +
∫ t
0
ψ(s, t, X0t )dWs
Equivalently, by relation (3.54), we can write these in terms of the solution (u0, m0)
of the system of PDE (2.10) from 0-MFG and the function ψ.
Ut = −
∫ t
0
∫ s
0
e−∫ t
s∂xxu0(k,X0
k)dkψ(r, s,X0
s )dWrds−
∫ t
0
e−∫ t
s∂xxu0(k,X0
k)dkdWs
Vt = ∂xxu0(t, X0
t )Ut +
∫ t
0
ψ(s, t, X0t )dWs
(4.24)
4.4.3 Covariance function
Having shown that the solution to (4.5) is a centered Gaussian process with respect to
the common noise, we now compute its covariance function making use of the explicit
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 115
solution we just derived. Recall our notations defined earlier;
γt , ∂xU0(t, X0
t , m0t )
βs,t , e−∫ t
sγrdr
η(t, x, ω) , E0[∂mU
0(t, x,m0t )(X
0t )Ut]
ηt , η(t, X0t , ω)
We define the following covariance functions
ϕUt , E
[
U2t
]
, ϕVt , E
[
V 2t
]
, ϕUVt , E [UtVt]
ϕUηt , E [Utηt] , ϕ
V ηt , E [Vtηt] , ϕ
ηt , E
[
η2t]
where E[·] is the expectation with respect to P only, i.e .with respect to the common
Brownian motion only. Observe that the full covariance functions can be derived
easily in terms of these functions. Using Ito’s lemma and the decoupling relation
(4.13), it is easy to check that
d
dtϕUt = −γtϕ
Ut − ϕ
Uηt + 1, ϕU
0 = 0
ϕUVt = γtϕ
Ut + ϕ
V ηt
ϕVt = γ2tϕ
Ut + 2γtϕ
Uηt + ϕ
ηt
ϕV ηt = γtϕ
Uηt + ϕ
ηt
(4.25)
From this system, we see that every term can be expressed in terms of ϕUηt and ϕη
t .
Using the explicit form of Ut, ηt in (4.15) and (4.23);
Ut = −
∫ t
0
βs,tηsds−
∫ t
0
βs,tdWs, ηt =
∫ t
0
ψ(s, t, X0t )dWs
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 116
we have
ϕUηt = −
∫ t
0
βs,tE[ηsηt]ds− E
[
ηt
∫ t
0
βs,tdWs
]
= −
∫ t
0
βs,t
∫ s
0
ψ(r, t, X0t )ψ(r, s,X
0s )drds−
∫ t
0
βs,tψ(s, t, X0t )ds
= −
∫ t
0
e−∫ t
s∂xxu0(k,X0
k)dk
∫ s
0
ψ(r, t, X0t )ψ(r, s,X
0s )drds−
∫ t
0
e−∫ t
s∂xxu0(k,X0
k)dkψ(s, t, X0
t )ds
and
ϕηt =
∫ t
0
(ψ(s, t, X0t ))
2ds
Thus, the covariance structure of (Ut(ω, ·), Vt(ω, ·))0≤t≤T can be fully described in
terms of ∂xxu0, where u0 is a solution from 0-MFG system (2.10), and ψ, a solution
to PDE (4.22).
4.5 Proof of lemmas, propositions, and theorems
4.5.1 Proof of Theorem 4.1.2
Proof. Let ∆Xεt =
Xεt −X0
t
εand δX,ε
t = ∆Xεt −Ut and define similarly ∆Y ε
t , ∆Zεt , ∆Z
εt ,
δY,εt , δZ,εt , δZ,ε
t , then (δX,εt , δ
Y,εt , δ
Z,εt , δ
Z,εt )0≤t≤T satisfies
dδX,εt = −δY,εt dt, dδ
Y,εt = dδ
Z,εt dWt + δ
Z,εt dWt, δ
X,ε0 = 0
δY,εT =
∂xg(XεT ,L(X
εT |FT ))− ∂xg(X
0T , m
0T )
ε− ∂xxg(X
0T , m
0T )UT + E
0[
∂xmg(X0T , m
0T )(X
0T )UT
]
Let
Xλ,εt := X0
t + λ(Xεt −X0
t ), 0 ≤ λ ≤ 1
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 117
Note that
∂xg(XεT ,L(X
εT |FT ))− ∂xg(X
0T ,L(X
0T ))
ε− ∂xxg(X
0T ,L(X
0T ))UT + E
0[
∂xmg(X0T , m
0T )(X
0T )UT
]
=
∫ 1
0
(
∂xxg(Xλ,εT ,L(Xλ,ε
T |FT ))∆Xεt + E
0[
∂xmg(Xλ,εT ,L(Xλ,ε
T |FT ))(Xλ,εT )∆Xε
t
])
dλ
− ∂xxg(X0T , m
0T )UT + E
0[
∂xmg(X0T , m
0T )(X
0T )UT
]
=
[∫ 1
0
∂xxg(Xλ,εT ,L(Xλ,ε
T |FT ))dλ
]
δX,εT +
∫ 1
0
E0[
∂xmg(Xλ,εT ,L(Xλ,ε
T |FT ))(Xλ,εT )δX,ε
T
]
dλ
+
[∫ 1
0
∂xxg(Xλ,εT ,L(Xλ,ε
T |Ft))dλ− ∂xxg(X0T ,L(X
0T ))
]
UT
+
∫ 1
0
E0[(
∂xmg(Xλ,εT ,L(Xλ,ε
T |FT ))(Xλ,εT )− ∂xmg(X
0T ,L(X
0T ))(X
0T ))
UT
]
dλ
= Iε1(δX,εT ) + Iε2
where Iε1 : L2FT
→ L2FT
is a linear functional defined by
Iε1(ξ) =
[∫ 1
0
∂xxg(Xλ,εT ,L(Xλ,ε
T |FT ))dλ
]
ξ+
∫ 1
0
E0[
∂xmg(Xλ,εT ,L(Xλ,ε
T |FT ))(Xλ,εT )ξ
]
dλ
and Iε2 ∈ L2FT
is given by
Iε2 =
[∫ 1
0
∂xxg(Xλ,εT ,L(Xλ,ε
T |Ft))dλ− ∂xxg(X0T ,L(X
0T ))
]
UT
+
∫ 1
0
E0[(
∂xmg(Xλ,εT ,L(Xλ,ε
T |FT ))(Xλ,εT )− ∂xmg(X
0T ,L(X
0T ))(X
0T ))
UT
]
dλ
Because ∂xxg, ∂xmg are bounded and UT , UT are bounded in L2, it follows that Iε2 is
bounded in L2. Thus, we can apply estimate (3.11) by setting
G(X) := Iε1(X) + Iε2
to deduce that
E[ sup0≤t≤T
(δX,εt )2] ≤ CK,T
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 118
where CK,T depends only on K, T . In particular, it is independent of ε. As a result,
we get
E[ sup0≤t≤T0≤λ≤1
(Xλ,εt −X0
t )2] ≤ CK,Tε
2
Thus, there exist a constant CK,T depending only on K, T such that
E[(Iε2)2] ≤ CK,Tε
2
Thus, by estimate (3.11) again, we get (4.8) as desired.
4.5.2 Proof of Theorem 4.2.5
Proof. Let αε = (αεt )0≤t≤T denote the ε-MFG solution, βε = (βε
t )0≤t≤T be the approx-
imate strategy defined by
βεt = α0
t − εVt
where V = (Vt)0≤t≤T is the backward process of the linear variational FBSDE (4.5).
For notational convenience, we will write J ε(α|β) to denote J ε(α|mβ) for any α, β ∈
H2([0, T ];R). Recall the definition of J ε(α|mβ) and mβ in Section 1.1.2.
For any control α, β(1), β(2) ∈ H2([0, T ];R), let Xα, Xβ(1), Xβ(2)
denote the corre-
sponding state processes, then we have
E[(Xβ(1)
T −Xβ(2)
T )2] ≤ CT
∫ T
0
|β(1)t − β
(2)t |2dt
Thus, combining with Lipschitz assumption on g, it follows that
|J ε(α|β(1))− J ε(α|β(2))| ≤ E
[
g(XαT ,L(X
β(1)
T |FT )− g(XαT ,L(X
β(2)
T |FT ))]
≤ K(E[(Xβ(1)
T −Xβ(2)
T )2])12
≤ CK,T
(∫ T
0
|β(1)t − β
(2)t |2dt
)
12
(4.26)
Also, since αε is the ε-MFG solution, we can use the estimate (2.15) in Theorem 2.2.4
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 119
to get
J ε(αε|αε) + C
∫ T
0
|αεt − αt|
2dt ≤ J ε(α|αε) (4.27)
for any α ∈ H2([0, T ];R). Lastly, from the definition of ε-MFG strategy, we have
J ε(αε|αε) ≤ J ε(α|αε) (4.28)
for any α ∈ H2([0, T ];R). Combining (4.26),(4.27), and (4.28) yields
J ε(βε|βε)− J ε(α|βε) ≤ J ε(βε|βε)−J ε(αε|αε) + J ε(α|αε)− J ε(α|βε)
= J ε(βε|βε)− J ε(βε|αε) + J ε(βε|αε)−J ε(αε|αε)
+ J ε(α|αε)− J ε(α|βε)
≤ C
[
(∫ T
0
|αεt − βε
t |2dt
)
12
+
∫ T
0
|αεt − βε
t |2dt
]
Using estimate (4.8) in Theorem 4.1.2,
(∫ T
0
|αεt − βε
t |2dt
)
12
= |ε|
∥
∥
∥
∥
αε − α0
ε− Vt
∥
∥
∥
∥
H2([0,T ];R)
≤ CK,Tε2
and we have
J ε(βε|βε)− J ε(α|βε) ≤ CK,Tε2
for any α ∈ H2([0, T ];R) and some constant CK,T as desired.
4.5.3 Proof of Theorem 4.3.1
Proof. Since FBSDE (4.5) is linear, one can easily check that for any ω ∈ Ω0 and
ω1, ω2 ∈ Ω
(Ut(ω, ω1) + Ut(ω, ω2), Vt(ω, ω1) + Vt(ω, ω2))0≤t≤T
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 120
satisfy the FBSDE for a given path (ω, ω1 + ω2). Thus, by pathwise uniqueness of
this FBSDE, one get that
Ut(ω, ω1 + ω2) = Ut(ω, ω1) + Ut(ω, ω2)
Vt(ω, ω1 + ω2) = Vt(ω, ω1) + Vt(ω, ω2)
Thus, we can conclude that they are both Gaussian processes. Next, we find their
mean and covariance functions for each fixed ω ∈ Ω0. Let first denote the mean by
µU , µV , µQ : [0, T ]× Ω0 → R. By direct calculation, i.e. taking expectation E[·] with
respect to P, they satisfies
dµUt = −µV
t dt
dµVt = µ
Qt dWt
µU0 = 0, µV
T = ∂xxg(X0T ,L(X
0T ))µ
UT + E
0[
∂xmg(X0T , m
0T )(X
0T )µ
UT
]
Applying Ito’s lemma on µUt µ
Vt , it follows that
µUTµ
VT − µU
0 µV0 =
∫
µVt µ
Qt dWt −
∫
(µVt )
2dt
Thus,
E
[∫
(µVt )
2dt
]
= −∂xxg(X0T ,L(X
0T ))(µ
UT )
2 − E0[
∂xmg(X0T , m
0T )(X
0T )µ
UT
]
µUT ≤ 0
Then it follows easily that µUt = µV
t = 0 for all t ∈ [0, T ] a.s.
4.5.4 Proof of Theorem 4.3.2
Proof. Fix (s, x,m) ∈ [0, T ]×R×P2(R). Let (Xεt , Y
εt , Z
εt , Z
εt )0≤t≤T denote the solution
to FBSDE (3.30) corresponding to ε-MFG over [s, T ] with the initial distributionms =
m and let (Xεt , Y
εt , Z
εt ,
˜Zεt )s≤t≤T denote the solution to FBSDE (3.36) corresponding
to an individual problem given (ms,mt )s≤t≤T and the initial state Xs = x. Recall the
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 121
definition of ms,m,Uε, we have
ms,mt = L(Xε
t |Fst ), Y ε
s = Uε(s, x,m)
By Theorem 4.1.2, we know that
E sups≤t≤T
[
(
Xεt −X0
t
ε− Ut
)2
+
(
Y εt − Y 0
t
ε− Vt
)2]
≤ Cε2 (4.29)
where (Ut, Vt, Qt, Qt)0≤t≤T is the solution to FBSDE (4.6) which reads
dUt = −Vtdt+ dWt
dVt = QtdWt + QtdWt
Us = 0, VT = ∂xxg(X0T , m
0T )UT + E
0[∂xmg(X0T , m
0T )(X
0T )UT ]
(4.30)
By the same argument as in Theorem 4.1.2, we also get
E sups≤t≤T
[
(
Xεt − X0
t
ε− Ut
)2
+
(
Y εt − Y 0
t
ε− Vt
)2]
≤ Cε2 (4.31)
where C is a constant depends only onK, T and not on ε, s, x,m, and (Ut, Vt, Qt,˜Qt)s≤t≤T
is the linear variational process for FBSDE (3.36) satisfying
dUt = −Vtdt+ dWt
dVt = QtdWt +˜QtdWt
Us = 0, VT = ∂xxg(X0T , m
0T )UT + E
0[∂xmg(X0T , m
0T )(X
0T )UT ],
(4.32)
Notice a slight difference between the FBSDE (4.32) and (4.30) in the mean-field
term. The mean-field term in (4.32) is given exogenously by (Ut, Vt)0≤t≤T , or more
precisely its copy (Ut, Vt)0≤t≤T . Thus, (4.32) is an ordinary FBSDE with random
coefficients. On the other hands, the mean-field term in (4.30) is part of the solution,
so we are dealing with mean-field type FBSDE in this case.
CHAPTER 4. ASYMPTOTIC ANALYSIS OF MEAN FIELD GAMES 122
From Theorem 4.3.1, we have
E[Ut] = E[Vt] = 0, for s ≤ t ≤ T (4.33)
where E[·] denote the expectation with respect to the common Brownian motion
(Wt)0≤t≤T , i.e. with respect to P, only. We claim that the same holds for (Ut, Vt)0≤t≤T .
We take expectation in FBSDE (4.32) to get that (At, Bt)0≤t≤T , (E[Ut], E[Vt])0≤t≤T
satisfiesdAt = −Btdt, dBt = CtdWt
As = 0, BT = ∂xxg(X0T , m
0T )AT
(4.34)
Note that zero is a solution to this FBSDE and from convexity assumption on g, this
FBSDE is in fact monotone, so we can apply Theorem 3.1.2 to conclude that it has
a unique solution. Therefore, we have
E[Ut] = E[Vt] = 0, for s ≤ t ≤ T (4.35)
Combining with (4.31) and the fact that Uε(s, x,m),U0(s, x,m) are deterministic, we
get, uniformly in (s, x,m),
limε→0
Uε(s, x,m)− U0(s, x,m)
ε= 0
Chapter 5
Linear-quadratic MFG with
common noise
In this chapter, we consider a linear-quadratic MFG (LQMFG for short) with common
noise. Similar to a classical stochastic control problem, LQMFG model is explicitly
solvable and we can illustrate all the concepts discussed in the previous chapters
through this model. We assume a linear state process is given by
dXt = αtdt+ σdWt + εWt, X0 = ξ0
and the running and terminal cost functions are of the form
f(t, x,m, α) =1
2
(
qx2 + α2 + q(x− sm)2)
g(x,m) =1
2
(
qTx2 + (x− sT m)2qT
)
where q, q, s, qT , qT , sT are constant and
m =
∫
R
ydm(y)
123
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 124
With these cost functions, the generalized Hamiltonian becomes
H(t, a, x, y,m) =1
2
(
qx2 + α2 + q(x− sm)2)
+ αy
As a result, the minimizer is given by
α(t, x, y,m) = −y
so the Hamiltonian is
H(t, x, y,m) = −y2
2+
1
2(q + q)x2 − qsxm+
1
2qs2m2
5.1 DPP approach
As discussed in Chapter 2, there are two ways to formulate an MFG problem through
DPP approach. The first one is to study the master equation of the generalize value
function denoted by Vε. The second method is to solve the FBSPDE of (uε, mε) where
the backward SPDE of uε describes a stochastic HJB of the value function along the
optimal distribution and the forward one gives the stochastic Fokker-Planck equation
of mε. Note that we have the relation
uε(t, x, ω) = Vε(t, x,mεt (ω)) (5.1)
So we will begin by solving the master equation to find Vε, then use the relation
above to get uε.
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 125
5.1.1 Master Equation
Using (2.8), the master equation for LQMFG model reads
∂tVε(t, x,m) + H(t, x, ∂xV (t, x,m), m) +
σ2 + ε2
2∂xxV
ε(t, x,m)
− E0[
∂mVε(t, x,m)(X)(∂xV
ε(t, X,m))]
+σ2
2∂mmV
ε(t, x,m)(X)[ζ, ζ ]
+ε2
2∂mmV
ε(t, x,m)(X)[1, 1] + ε2E0[
∂xmVε(t, x,m)(X)1
]
= 0
(5.2)
with terminal condition
Vε(T, x,m) = qTx+ (x− sT m)qT = (qT + qT )x− sT qT m
Solving this equation yields the optimal control in the feedback form
αε(t, x,m) = −∂xVε(t, x,m) (5.3)
Motivated by classical LQ problems, we seek a solution of the form
Vε(t, x,m) =1
2p(t)x2 + q(t)xm+
1
2r(t)m2 + s(t)
where p(t), q(t), r(t), s(t) : [0, T ] → R are deterministic functions. Using this form,
we have
∂tVε(t, x,m) =
1
2p′(t)x2 + q′(t)xm+
1
2r′(t)m2 + s′(t)
∂xVε(t, x,m) = p(t)x+ q(t)m
∂xxVε(t, x,m) = p(t)
E0[∂mV
ε(t, x,m)(X)ξ] = (q(t)x+ r(t)m)ξ
∂mmVε(t, x,m)(X)[ξ, ξ′] = r(t)ξξ′
E0[∂xmV
ε(t, x,m)(X)ξ] = q(t)ξ
(5.4)
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 126
for any (t, x,m) ∈ [0, T ]× R×P2(R), X is a lifting random variable with L(ξ) = m,
ξ, ξ′ are arbitrary random variables in the same lifting space and ξ, ξ′ denote the mean
of ξ, ξ′ respectively. Therefore, we have
E0[∂mV
ε(t, x,m)(X)∂xVε(t, X,m)] = (q(t)x+ r(t)m)(p(t) + q(t))m
= q(t)(p(t) + q(t))xm+ r(t)(p(t) + q(t))m2
∂mmVε(t, x,m)(X)[1, 1] = r(t)
∂mmVε(t, x,m)(X)[ζ, ζ ] = 0
E0[∂xmV
ε(t, x,m)(X)1] = q(t)
(5.5)
Plugging (5.4),(5.5) into (5.2) yields
1
2p′(t)x2 + q′(t)xm+
1
2r′(t)m2 + s′(t)−
1
2p2(t)x2 −
1
2q2(t)m2 − p(t)q(t)xm
+1
2(q + q)x2 − qsxm+
1
2qs2m2 +
1
2(σ2 + ε2)p(t)− q(t)(p(t) + q(t))xm
− r(t)(p(t) + q(t))m2 +ε2
2r(t) + ε2q(t) = 0
Grouping the coefficients of x2, xm, m2, 1 and using the terminal conditions, we get
the system of ODEs
p′(t) = p2(t)− q − q, p(T ) = qT + qT
q′(t) = 2p(t)q(t) + q2(t) + qs, q(T ) = −sT qT
r′(t) = 2(p(t) + q(t))r(t) + q(t)2 − qs2, r(T ) = s2T qT
s′(t) = −1
2(σ2 + ε2)p(t)−
1
2ε2r(t)− ε2q(t), s(T ) = 0
(5.6)
The system above is decoupled since we can solve in successive order from the top
equation of p(t) to the bottom equation of s(t). From the ODE of s(t), we know that
s(t) is of the form
s(t) = a(t) + ε2b(t)
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 127
where
a(t) =σ2
2
∫ T
t
p(s)ds, b(t) =
∫ T
t
(
1
2(p(s) + r(s)) + q(s)
)
ds (5.7)
So we have the generalized value function
Vε(t, x,m) =1
2p(t)x2 + q(t)xm+
1
2r(t)m2 + a(t) + ε2b(t)
where p(t), q(t), r(t) are given by the ODEs (5.6) and a(t), b(t) are given by (5.7).
Going back to (5.3), we have the optimal control in the feedback form
αε(t, x,m) = −p(t)x− q(t)m
5.1.2 Optimal controlled process and stochastic value func-
tion
Now we would like to write out explicitly the solution (uε, mε, vε) of the FBSPDE
(2.9). First, note that the optimal controlled process is given by
dXεt = (−p(t)Xε
t − q(t)mεt )dt+ σdWt + εdWt, Xε
0 = ξ0
where
mεt , E[Xε
t |Ft]
Thus, we have
dmεt = −(p(t) + q(t))mε
tdt + εdWt, mε0 = E[ξ0] (5.8)
Solving this SDE yields
mεt = β0,tm
ε0 + ε
∫ t
0
βs,tdWs, βs,t = exp
(
−
∫ t
s
[p(r) + q(r)]dr
)
(5.9)
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 128
Using relation (5.1), we have the stochastic value function
uε(t, x) =1
2p(t)x2 + q(t)x
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)
+1
2r(t)
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)2
+ a(t) + ε2b(t)
We also have the diffusion term for duε(t, x) which is given by
vε(t, x) = E0[∂mV
ε(t, x,mεt )(X)1] = q(t)x+r(t)mε
t = q(t)x+r(t)
(
β0,tmε0 + ε
∫ t
0
βs,tdWs
)
5.2 SMP approach
Recall that the Hamiltonian and the optimal feedback control are given by
H(t, x, y,m) = −y2
2+
1
2(q + q)x2 − qsxm+
1
2qs2m2, α(t, x, y) = −y
Therefore, the corresponding McKean-Vlasov FBSDE is
dXt = −Ytdt+ σdWt + εdWt
dYt =(
(q + q)Xt − qsE[Xt|Ft])
dt+ ZtdWt + ZtdWt
X0 = ξ, YT = (qT + qT )XT − sT qTE[XT |FT ]
(5.10)
We can take the ansatz
Yt = P (t)Xt +Q(t)E[Xt|Ft], P (T ) = qT + qT , Q(T ) = −sT qT
to solve this system. However, having found the generalized value function, we can
use Theorem (2.3.1) to see that the decoupling function is simply
Uε(t, x,m) = ∂xVε(t, x,m) = p(t)x+ q(t)m
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 129
Thus, we have that
Y εt = p(t)Xε
t + q(t)E[Xεt |Ft], Zε
t = σp(t), Zεt = −εq(t)(p(t) + q(t))E[Xε
t |Ft]
solve FBSDE (5.11) where (Xεt )0≤t≤T satisfies the mean-field SDE
dXεt = (−p(t)Xε
t − q(t)E[Xεt |Ft])dt+ σdWt + εWt, Xε
0 = ξ0
The SDE above can be solved explicitly by plugging in (mεt )0≤t≤T from (5.9).
5.3 Asymptotic analysis
The linear variational process reads
dUt = −Vtdt+ dWt
dVt =(
(q + q)Ut − qsE[Ut|Ft])
dt+QtdWt + QtdWt
U0 = 0, YT = (qT + qT )UT − sT qTE[UT |FT ]
(5.11)
Note that
∂xU0(t, x,m) = p(t), E
0[∂mU(t, x,m)(X)ξ] = q(t)E0[ξ], h(t, x, z) = q(t)
We can use (4.13) to get
Vt = p(t)Ut + q(t)E[Ut|Ft] (5.12)
where (Ut)0≤t≤T solves
dUt =(
−p(t)Ut − q(t)E[Ut|Ft])
dt+ dWt, U0 = 0
We can deduce from the SDE above that (Ut)0≤t≤T is independent of ω ∈ Ω0 and is
Ft-adpated, thus
Ut = E[Ut|Ft]
CHAPTER 5. LINEAR-QUADRATIC MFG WITH COMMON NOISE 130
In other words, Ut is independent of the current state X0t . The main reason for this
fact is the fact that ∂xxu0(t, x) is constant in linear-quadratic case. Thus, the SDE
for (Ut)0≤t≤T is simply
dUt = −(p(t) + q(t))Utdt+ dWt, U0 = 0 (5.13)
From (5.12), we also have all the related terms from Section 4.4.2
η(t, x) = q(t)Ut, w(t, x) = h(t, x, z) = q(t)
Lastly, we describe the distribution of (Ut, Vt)0≤t≤T . The fact that (Ut, Vt)0≤t≤T are
centered Gaussian processes follows directly from (5.12) and (5.13). Now we compute
the covariance function
ϕUt , E
[
U2t
]
, ϕVt , E
[
V 2t
]
, ϕUVt , E [UtVt]
directly using the (5.12) and (5.13) again to get
dϕUt
dt= −2(p(t) + q(t))ϕU
t + 1, ϕU0 = 0
ϕUVt = (p(t) + q(t))ϕU
t
ϕVt = (p(t) + q(t))2ϕU
t
Notice that all the terms in the asymptotic analysis is trivial; (Ut)0≤t≤T is independent
of (X0t )0≤t≤T and η, w, ∂mU0 are independent of x. The main reason for this is due
to the fact that ∂xxf, ∂xxg are constant, hence so does ∂xxu0 and other second order
terms.
Bibliography
[1] Yves Achdou and Italo Capuzzo-Dolcetta. Mean field games: Numerical meth-
ods. SIAM Journal on Numerical Analysis, 48(3):1136–1162, 2010.
[2] Yves Achdou, Jean-Michel Lasry, Pierre-Louis Lions, and Benjamin Moll. Het-
erogeneous agent models in continuous time. Technical report, mimeo, 2013.
[3] Saran Ahuja. Wellposedness of mean field games with common noise under a
weak monotonicity condition. arXiv preprint arXiv:1406.7028, 2014.
[4] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savarae. Gradient flows. Springer,
2005.
[5] Fabio Antonelli. Backward-forward stochastic differential equations. The Annals
of Applied Probability, pages 777–793, 1993.
[6] Fabio Antonelli and Jin Ma. Weak solutions of forward–backward sde. Stochastic
Analysis and Applications, 21(3):493–514, 2003.
[7] Martino Bardi and Fabio S Priuli. Linear-quadratic n-person and mean-
field games with ergodic cost. SIAM Journal on Control and Optimization,
52(5):3022–3052, 2014.
[8] Alain Bensoussan, Michael Chau, and Phillip Yam. Mean field games with a
dominating player. arXiv preprint arXiv:1404.4148, 2014.
[9] Alain Bensoussan, Jens Frehse, and Phillip Yam. Mean Field Games and Mean
Field Type Control Theory. Springer, 2013.
131
BIBLIOGRAPHY 132
[10] Alain Bensoussan, Jens Frehse, and Sheung Chi Phillip Yam. The master equa-
tion in mean field theory. Journal de Mathematiques Pures et Appliquees, 2014.
[11] Alain Bensoussan, Joseph Sung, Phillip Yam, and Siu Pang Yung. Linear-
quadratic mean field games. arXiv preprint arXiv:1404.5741, 2014.
[12] Jean-Michel Bismut. Theorie probabiliste du controle des diffusions, volume 167.
American Mathematical Soc., 1976.
[13] Pierre Cardaliaguet. Notes on mean field games. from P.-L. Lions’ lectures at
College de France, 2010.
[14] Pierre Cardaliaguet, P Jameson Graber, Alessio Porretta, and Daniela Tonon.
Second order mean field games with degenerate diffusion and local coupling.
Nonlinear Differential Equations and Applications NoDEA, pages 1–31, 2014.
[15] Pierre Cardaliaguet, J-M Lasry, P-L Lions, and Alessio Porretta. Long time
average of mean field games with a nonlocal coupling. SIAM Journal on Control
and Optimization, 51(5):3558–3591, 2013.
[16] Rene Carmona and Francois Delarue. Forward-backward stochastic differ-
ential equations and controlled mckean vlasov dynamics. arXiv preprint
arXiv:1303.5835, 2013.
[17] Rene Carmona and Francois Delarue. Probabilistic analysis of mean-field games.
SIAM Journal on Control and Optimization, 51(4):2705–2734, 2013.
[18] Rene Carmona and Francois Delarue. The master equation for large popula-
tion equilibriums. In Stochastic Analysis and Applications 2014, pages 77–128.
Springer, 2014.
[19] Rene Carmona, Francois Delarue, and Aime Lachapelle. Control of mckean–
vlasov dynamics versus mean field games. Mathematics and Financial Eco-
nomics, 7(2):131–166, 2013.
BIBLIOGRAPHY 133
[20] Rene Carmona, Francois Delarue, and Daniel Lacker. Mean field games with
common noise. arXiv preprint arXiv:1407.6181, 2014.
[21] Rene Carmona, Jean-Pierre Fouque, and Li-Hsien Sun. Mean field games and
systemic risk. Available at SSRN 2307814, 2013.
[22] Rene Carmona and Daniel Lacker. A probabilistic weak formulation of mean
field games and applications. arXiv preprint arXiv:1307.1152, 2013.
[23] Rene Carmona and Xiuneng Zhu. A probabilistic approach to mean field games
with major and minor players. arXiv preprint arXiv:1409.7141, 2014.
[24] Jean-Francois Chassagneux, Dan Crisan, and Francois Delarue. Classical so-
lutions to the master equation for large population equilibria. arXiv preprint
arXiv:1411.3009, 2014.
[25] Donald Dawson and Jean Vaillancourt. Stochastic mckean-vlasov equations.
Nonlinear Differential Equations and Applications NoDEA, 2(2):199–229, 1995.
[26] Francois Delarue. On the existence and uniqueness of solutions to fbsdes in a
non-degenerate case. Stochastic processes and their applications, 99(2):209–286,
2002.
[27] Markus Fischer. On the connection between symmetric n-player games and mean
field games. arXiv preprint arXiv:1405.1345, 2014.
[28] Wendell H Fleming and Halil Mete Soner. Controlled Markov processes and
viscosity solutions, volume 25. Springer Science & Business Media, 2006.
[29] Diogo A Gomes, Joana Mohr, and Rafael Rigao Souza. Discrete time, finite
state space mean field games. Journal de mathematiques pures et appliquees,
93(3):308–328, 2010.
[30] Diogo A Gomes, Joana Mohr, and Rafael Rigao Souza. Continuous time finite
state mean field games. Applied Mathematics & Optimization, pages 1–45, 2012.
BIBLIOGRAPHY 134
[31] Diogo A Gomes and Vardan K Voskanyan. Extended mean field
games-formulation, existence, uniqueness and examples. arXiv preprint
arXiv:1305.2600, 2013.
[32] DiogoA. Gomes and Joao Saude. Mean field games models—a brief survey.
Dynamic Games and Applications, pages 1–45, 2013.
[33] Olivier Gueant, Jean-Michel Lasry, and Pierre-Louis Lions. Mean field games
and applications. In Paris-Princeton Lectures on Mathematical Finance 2010,
pages 205–266. Springer, 2011.
[34] Ying Hu and Shige Peng. Maximum principle for semilinear stochastic evolution
control systems. Stochastics and Stochastic Reports, 33(3-4):159–180, 1990.
[35] Ying Hu and Shige Peng. Solution of forward-backward stochastic differential
equations. Probability Theory and Related Fields, 103(2):273–283, 1995.
[36] Minyi Huang. Large-population lqg games involving a major player: the nash
certainty equivalence principle. SIAM Journal on Control and Optimization,
48(5):3318–3353, 2010.
[37] Minyi Huang, Peter E Caines, and Roland P Malhame. Large-population cost-
coupled lqg problems with nonuniform agents: Individual-mass behavior and
decentralized ε-nash equilibria. Automatic Control, IEEE Transactions on,
52(9):1560–1571, 2007.
[38] Ioannis Karatzas. Brownian motion and stochastic calculus, volume 113. Springer
Science & Business Media, 1991.
[39] Vassili N Kolokoltsov, Jiajie Li, and Wei Yang. Mean field games and nonlinear
markov processes. arXiv preprint arXiv:1112.3744, 2011.
[40] Daniel Lacker. A general characterization of the mean field limit for stochastic
differential games. arXiv preprint arXiv:1408.2708, 2014.
BIBLIOGRAPHY 135
[41] Daniel Lacker and Kevin Webster. Translation invariant mean field games with
common noise. arXiv preprint arXiv:1409.7345, 2014.
[42] Jean-Michel Lasry and Pierre-Louis Lions. Jeux a champ moyen. i–le cas sta-
tionnaire. Comptes Rendus Mathematique, 343(9):619–625, 2006.
[43] Jean-Michel Lasry and Pierre-Louis Lions. Jeux a champ moyen. ii–horizon fini
et controle optimal. Comptes Rendus Mathematique, 343(10):679–684, 2006.
[44] Jean-Michel Lasry and Pierre-Louis Lions. Mean field games. Japanese Journal
of Mathematics, 2(1):229–260, 2007.
[45] Jean-Michel Lasry, Pierre-Louis Lions, Olivier Gueant, et al. Application of mean
field games to growth theory. 2008.
[46] Jin Ma, Philip Protter, and Jiongmin Yong. Solving forward-backward stochastic
differential equations explicitly—a four step scheme. Probability Theory and
Related Fields, 98(3):339–359, 1994.
[47] Jin Ma, Zhen Wu, Detao Zhang, and Jianfeng Zhang. On wellposedness of
forward-backward sdes—a unified approach. arXiv preprint arXiv:1110.4658,
2011.
[48] Jin Ma, Hong Yin, and Jianfeng Zhang. On non-markovian forward–backward
sdes and backward stochastic pdes. Stochastic Processes and Their Applications,
122(12):3980–4004, 2012.
[49] Jin Ma and Jiongmin Yong. Forward-backward stochastic differential equations
and their applications. Number 1702. Springer, 1999.
[50] Son Luu Nguyen and Minyi Huang. Linear-quadratic-gaussian mixed games with
continuum-parametrized minor players. SIAM Journal on Control and Optimiza-
tion, 50(5):2907–2937, 2012.
[51] Mojtaba Nourian and Peter E Caines. ǫ-nash mean field game theory for nonlin-
ear stochastic dynamical systems with major and minor agents. SIAM Journal
on Control and Optimization, 51(4):3302–3331, 2013.
BIBLIOGRAPHY 136
[52] Shige Peng and Zhen Wu. Fully coupled forward-backward stochastic differential
equations and applications to optimal control. SIAM Journal on Control and
Optimization, 37(3):825–843, 1999.
[53] Huyen Pham. Continuous-time stochastic control and optimization with financial
applications, volume 61. Springer, 2009.
[54] Alessio Porretta. Weak solutions to fokker–planck equations and mean field
games. Archive for Rational Mechanics and Analysis, 216(1):1–62, 2015.
[55] Svetlozar T Rachev and Ludger Ruschendorf. Mass Transportation Problems:
Volume I: Theory, volume 1. Springer Science & Business Media, 1998.
[56] Alain-Sol Sznitman. Topics in propagation of chaos. In Ecole d’ete de probabilites
de Saint-Flour XIX—1989, pages 165–251. Springer, 1991.
[57] Cedric Villani. Optimal transport: old and new, volume 338. Springer, 2009.
[58] Hong Yin. Solvability of forward–backward stochastic partial differential equa-
tions. Stochastic Processes and their Applications, 124(8):2583–2604, 2014.
[59] Jiongmin Yong. Finding adapted solutions of forward–backward stochastic differ-
ential equations: method of continuation. Probability Theory and Related Fields,
107(4):537–572, 1997.
[60] Jiongmin Yong. Linear forward-backward stochastic differential equations with
random coefficients. Probability theory and related fields, 135(1):53–83, 2006.
[61] Jiongmin Yong and Xun Yu Zhou. Stochastic controls: Hamiltonian systems and
HJB equations, volume 43. Springer, 1999.