MFG: Francois Delarue Lectures

Ramesh Kadambi

February 17, 2025

1
May, 09, 2021

Contents
1 Motivation 4

2 Asymptotic Formulation 4
2.1 Paradigm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 The Problem Formulation 4

3.1 The Equilibria Finite Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.3 The Controls and Feasibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

4 Interaction Model 5

5 Asymptotic Formulation under the limit N → ∞ 6

6 Propagation of Chaos 6
6.1 Distance in Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
6.2 Taking N to ∞. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

7 The distribution µt = P (X t )⊗k 7

8 Back to the Optimization Problem 8

8.1 Representative Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
8.2 Intrinsic Formulation for µt (x) t ∈ [0, T ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

9 Tractable Formulation The MFG System 9

9.1 The PDE Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
9.2 Solving the HJB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
9.3 The fixed point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
9.4 Uniqueness of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
9.5 Existence of Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
9.6 Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
9.7 Relative Compactness of the Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
9.8 If f ̸= 0 ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

10 Other Approaches 15
10.1 The fixed point problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
10.2 Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

11 McKean Vlasov FBSDE for Optimal Feedback 16

12 Examples - Linear Quadratic Systems 16

13 Back to the N player Game 16

13.1 Implementing the limit Optimal Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

14 MASTER EQUATION AND CONVERGENCE PROBLEM 17

14.1 Equilibrium Value for the MFG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
14.2 The ROAD to the PDE in U . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

15 Convergence Problem 18
15.1 Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
15.2 The Master Equation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
15.3 Finding the NASH Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
15.4 The Classical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
15.5 The ASNATZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
15.6 From Master Equation to Nash . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
May, 09, 2021

15.7 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

16 Schrodinger Approach to MFG [1] 21

May, 09, 2021

1 Motivation
Mean Field Games are extension of finite player games to Infinitely many players. the main characteristics of the Mean Field
models are,

1. The number of players/particles is large.

2. In the mean field the interaction between particles is weak.
3. Individual states are not observable however the overall state or a function of the overall state is observable. Players see
the others collectively.

4. Individuals act to minimize/maximize their cost function.

2 Asymptotic Formulation
From the finite to infinite reduces the complexity via the notions of propagation of chaos and LLN. We do this in two steps and
are interested in how equilibria are formed and the nature of these equilibria.

2.1 Paradigm
1. The main ideas used are mean-field/symmetry justified by propagation of chaos and LLN.

2. We reduce the asymptotic analysis to one typical player with interaction with a theoretical distribution of the population?
3. Decrease the complexity of the asymptotic formulation.

2.2 Program
1. Are there asymptotic equilibria? Are they unique? If they exist what is the shapre of this equilibrium?
2. We want to then go backward to finite games. We want to use Asymptotic equilibria as quasi-equilibria in finite game.
3. Prove the convergence of equilibria in finite-player games.

3 The Problem Formulation

Here we would like to let the number of players go to infinity. This facilitates complexity reduction. Going to infinity creates an
averaging effect that reduces complexity. The Averaging is connected to LLN.

3.1 The Equilibria Finite Games

Equilibria in finite games are states of the players that they are happy with. This is described in the lectures as compromises.
The players do not see benefit in changing the positions when in equilibrium.

3.2 The Model

We model N players. The states of the players are described by the dynamics below,

dXti = αti dt + dWti

The state Xti ∈ Rd , Wti are N -independent Brownian motions. The cost function for each player.for a finite horizon,

(  )
Z T 
i 1 i2
J =E g(XTi , <>) + i
f (Xi , <>) + |αt | dt
0 2

Above the Wti are independent Brownian Motions. On the same token, X0i are independent. The next issue is the interaction
of the agents. That is where the place holder <> comes into play. The functions g : Rd × <>7→ R and similarly f : Rd × <>7→ R
May, 09, 2021

take in the interaction term specified by the empirical distribution of the agents. The empirical distribution is essentially given
by,
N
1 X
µN
t = δ j
N j=1 Xt

The collective state of the population is described by this measure. The cost function now becomes,
( Z T  )
i i N i N 1 i2
J = E g(XT , µt ) + f (Xi , µt ) + |αt | dt
0 2

Now we can define the maps g : Rd × P(Rd ) 7→ Rd and f : Rd × P(Rd ) 7→ R.

R
Remark 1 (Example of a Cost Function). An example for the cost functions we use a convolution f (x, µ) = Rd
F (x − y)dµ(y)
where x ∈ Rd and µ ∈ P(Rd ).

1 1
PN 1
PN
If now we consider µ = µN
P
X = N j δxj , we can write f (xi , N j=1 δxj ) = N j=1 F (xi − xj ). This is an interesting
function that indeed depends on the distribution of the agent states. If F is the distance between the states, then it would capture
the average distance of all the points from xi . This is compared to the potential energy in physics.

3.3 The Controls and Feasibility

The controls in the problem are αti . A feasible control has the following properties,
Definition 1 (Feasible Control). A process αti is said to be feasible if,
1. It is adapted.
2. it is progressively measurable, i.e. αti = [0, t] × Ω 7→ Rd is measurable w.r.t Ft the filtration.
RT
3. E[ 0 |αti |2 dt < ∞.
Now that we know what is admissible control, we are ready to define the Nash-Equilibrium.
Definition 2 (Nash Equilibrium). Given the cost function or performance criteria Ji (αti , αt−i . The N -tuple {α∗,1 , · · · , α∗,N } is
a Nash-Equilibrium given function J i if and only if,

J i (α∗,1 , · · · , α∗,i , · · · , α∗,N ) ≤ J(α∗,1 , · · · , β i , · · · , α∗,N )

In the above definition the equilibrium is only based on the equilibrium values of the remaining agents. It is fixed and solved as
an open loop control problem. There is a closed loop control solution where the optimal is a function of the current state of the
system. In the case of the open loop set up the optimal is the function just of the noises. In the case of the closed loop system,
the control or strategy is given as,

αti = ui (t, Xt1 , · · · , XtN ) (3.1)

The performance criteria in that case would be,

J i (u∗,1 , · · · , u∗,i , · · · , u∗,N ) ≤ J i (u∗,1 , · · · , ui , · · · , u∗,N ).

From the optimal control problem for example in the case of a robot, the open loop control is just a minimization problem that
reveals a function u(t) that drives the robot from point a to b given the performance criteria. In the case of a feed back control,
u(t) = f (Xd − Xt ) and it will not require a path planning where you specify the tracking trajectory.

4 Interaction Model
The interaction can be modeled as a function of the mutual states and the empirical distribution of the states of the agents, or
it can be modeled as the empirical distribution of their actions. In the case of price formation and price setting one is probably
setting ones price based on the mean of all the prices that are set. The latter problem where the interaction is modeled via the
strategies/actions of the agents is called MFG of control type.
May, 09, 2021

5 Asymptotic Formulation under the limit N → ∞

Our performance criteria is givens as,
( Z T  )
i i N i N 1 i2
J = E g(XT , µt ) + f (Xi , µt ) + |αt | dt .
0 2
As an be seen, the agents all us the same functions g, f which introduces a strong symmetry in the problem. We expect this
symmetry to present itself as N → ∞. This symmetry translation cannot really be proved. However we make a leap of faith
and chose our control as follows.
αti = α(t, Xti , µN
t ) (5.1)
The final choice (5.1) is different from the choice (3.1) in that we are not now dependent on the state of the other agents. We
only are looking at the agents as a whole. We have also make another simplification using our symmetry. The choice is now
independent of the number of agents as well. The dependence on the number of agents is to the extent that the empirical
distribution is dependent on it. This is a difficult thing to prove. In reality the choice is dependent on N as below,
αti = αN (t, Xti , µN
t ).

We conjecture that αN ≈ α which implies that αt∗,i = α(t, Xt∗,i , µ∗,N

t ). So the particle system becomes,
dXt∗,i ≈ α(t, Xt∗,i , µ∗,N
t )dt + dBti . (5.2)

Where µ∗,N
N
= N1 j=1 δX ∗,j . The question we would like to answer now is what happens to this new particle system that has
P
t t
the modifications using symmetry as N → ∞. Enter propagation of chaos.

6 Propagation of Chaos
Now that we have the asymptotic formulation with a fixed α as in (5.2), our question is for this large system of particles as
follows.
1. Existence.
2. Well-Posedness.
Apparently the Brownian motions make easier and are helpful. The conditions for this are similar to that for first order ODE.
That is,
N
d×N 1 X
[0, T ] × R ∋ (t, (x1 , x2 , · · · , xN )) 7→ α(t, xi , δx ) (6.1)
N j=1 j

where xi ∈ Rd , is Lipschitz. Note that here the empirical measure is fixed. It also needs to be bounded in time. The issue is that
α is defined on the space of probability measures and not in Euclidean space. The smoothness and Lipschitz conditions have to
be specified in he measure space. If you see the definition of (6.1), you see that we replaced the {x−i } by the empirical measure.
So essentially the functions α takes the states and maps it into a measure space.

6.1 Distance in Measure Spaces

We first define the space of measures that we would like to consider. Given p ≥ 1 we define the space of feasible measures as,
 Z 
Pp (Rd ) = µ ∈ P(Rd ) : |x|p dµ(x) < ∞
Rd

We now given µ, ν ∈ Pp (Rd ) define a measure Π on Rd × Rd such that µ, ν are marginal densities of Π. We basically have,
µ = Π ◦ e−1
1 ; e1 : (x1 , x2 ) ∈ Rd × Rd 7→ x1 (6.2)
ν =Π◦ e−1
2 ;
d d
e2 : (x1 , x2 ) ∈ R × R 7→ x2 (6.3)
The distance function ∀µ, ν ∈ P(Rd ) now is given as,
Z  p1
p
Wp (µ, ν) = inf |x1 − x2 | dΠ(x1 , x2 )
Π Rd ×Rd
May, 09, 2021

Remark 2 (Bound for distance). Consider two random variables X1 and X2 , if Px1 and Px2 are the probability densities
of these randomR variables. We have the joint density given by PX1 ,X2 . Clearly, PX1 , PX2 ∈ mathcalP (R). We now have
E[|x1 − x2 |p ] = R×R |x1 − x2 |p dPX1 ,X2 (x1 , x2 ). Now to find a bound to our Wp (µ, ν), we define Πmax = PX1 ,X2 , clearly PX1 ,X2
satisfies the conditions (6.2), (6.3). We also have,
Z  p1
1
Π p p
Wp (PX1 , PX2 ) ≤ E [|x1 − x2 | ] ≤ p |x1 − x2 | dPX1 ,X2 (x1 , x2 ) . (6.4)
R×R

We see that (6.4) is true by definition. This is a major advantage. If we have random variables then it is fairly straight forward
to construct a distance function that is bounded.

PN PN
Remark 3 (Empirical Measures Distance). Consider X, Y ∈ RN , we then have Wp ( N1 i=1 δxi , N1 i=1 δyi ) where PX =
1
PN 1
Pn
N i=1 δxi and PY = N i=1 δyi . This is fairly straight forward to compute. The distance function is given by,

( n
) p1
1X
Wp (P1 , P2 ) ≤ E[|x1 − x2 |p ] = |xi − yi |p .
n i=1

Remark 4 (Lipschitz requirement for α). The standard assumption regarding the functions α(t, xi , µN t ) is that it is Lipschitz
continuous in x and µN
t ∈ Pp (Rd
) where p ∈ {1, 2}, and Pp (Rd
) is equipped with distance functions Wp.

6.2 Taking N to ∞.
When N → ∞ our particle system can be written independent of the optimization problem as,

dXti = α(t, xi , µN i
t )dt + dBt (6.5)

We observe that as N → ∞ the individual particles Xi will be impacted less by the noises Bt−i . We have ansatz, particles are
uncorrelated as N → ∞. This also impacts our early assumption that the initial distribution of Xi were independent. As N → ∞
we claim that they are indentically distributed as well. The initial state of Xi is that of N -iid particles. Given our symmetry
they are exchangeable and iid for all time. The level of uncorrelation is strong enough to satisfy independence of particles. Now
we can use LLN. We therefore have,

µN
t ≈ Px .
N →∞

Proposition 6.1 (The same distribution.). Let us denote the particles as {XtN,1 , ·, XtN,N }t∈[0,T ] , the solution to (6.5) as N → ∞
has the probability law,
L
P (XtN,1 , · · · , XtN,N ) ⇒ P (X t )⊗k , t ∈ [0, T ]
N →∞

Now we have (X)t∈[0,T ] satisfies the equation,

dX t = α(t, X t , P (X t )⊗k )dt + dBt (6.6)

The above equation is called the McKean-Vlasov stochastic differential equation. The solution here is interacting with its own
distribution and we have introduced a new Brownian motion that is independent of the label. This again boils down to the
indistinguishability of the particles.

7 The distribution µt = P (X t )⊗k

In order to obtain an expression for P (X t )⊗k , we first assume that (6.6) is well posed. We then let µt = P (X t )⊗k . The question
is can we get an equation for µt . We use a PDE approach here. We take a smooth test function h : Rd 7→ R with compact
May, 09, 2021

support. We want to do the following,

 
1 2
dE[h(X t )] = E ∇h(X t ).(α(t, X t , µt )dt + dBt ) + ∆ h(X)dt
2
1 2
= E[∇h(X t ).α(t, X t , µt ) + ∆ h(X t )]dt
Z 2 
1 2
= (∇h(X t ).α(t, X t , µt ) + ∆ h(X t ))dµt (x) dt
d 2
 R Z
= h(X t ).(α(t, X, µt )µt (x))|Rd − h(X t ).∇(α(t, X, µt )µt (x))dµt (x)
Rd
+∆h(X t )µ(t)|Rd − h(X t ).∇µt (X)|Rd
Z 
+ h(X t ).∇µt (x)dx dt
Rd
Z  Z 
1 T
d h(X t )µt (x)dx = h(X t ).[−∇(α(t, X t , µt )µt (x)) + ∇ ∇µt (x)]dx dt
Rd d 2
Z  ZR 
∂µt (x) 1 T
h(X t ) dx dt = h(X t ).[−∇(α(t, X t , µt )µt (x)) + ∇ ∇µt (x)]dx dt
Rd ∂t Rd 2

Since this is true for any arbitrary distribution function, we have the following for the density function.
1
∂t µt (x) = −∇.[α(t, X t , µt (x))µt (x)] + ∇T ∇µt (x)
2
1
We solve this as an initial value problem with m0 (x) = P (X 0 ). These result are from the notes Sznitman [2].

8 Back to the Optimization Problem

What we have done so far is to go from a particle system with finite particles and moved to infinite particles where interactions
are absent and we have replaced the original dynamics with a new dynamics where individual particles interact with their
distribution.

Remark 5 (N → ∞ Dynamics). We just recall the N → ∞ dynamics and the Fokker-Plank equations.
∗ ∗
dX t = α(t, X t , µt (x))dt + dBt
∗ 1
∂t µt (x) = −∇.(α(t, X t , µt (x))µt (x)) + ∇T ∇µt (x)
2
∗
µ0 (x) = P (X 0 )

Our main observations can be summerized as,

1. As N → ∞, the players at equillibrium become independent.

2. The empirical distribution converges to a some theoretical distribution.

However, we are looking for an intrinsic characterization of the equilibrium. In order to do so we go back t our optimization
problem and try to simplify our original model.

Remark 6 (Our original model). Minimize :

( Z T  )
i i N i N 1 i2
J = E g(XT , µt ) + f (Xi , µt ) + |αt | dt (8.1)
0 2

subject to:

dXti = αti dt + dWti

May, 09, 2021

8.1 Representative Agent

We now replace the empirical distribution in (8.1) by our new theoretical distribution (flow of the theoretical distribution). The
theoretical distribution µt (x) describes the equilibrium state of the system or population. We also replace the individual agents
with our equilibrium agent as N → ∞. To find this distribution we make one more simplification, we just consider a system
with just one particle. It really does not matter which since they are indistinguishable. This particle evolves in an environment
where the other particles are described only by a distribution µt (x). The particle tries to optimize its own cost function. Each
player therefore tries to optimize with respect to the state of the population. When the particle or agent acts in this manner,
we characterize the equilibrium of the game as, the best response of the player is the state of the population itself.

Model 1 (The Representative Agent Problem). Minimize:

( Z T  )
i i i 1 i2
J = E g(XT , µt (x)) + f (Xi , µt (x)) + |αt | dt
0 2

Subject to:

dXt = α(t, Xt , µt (x))dt + dBt

1
∂t µt (x) = −∇.[α(t, X t , µt (x))µt (x)] + ∇T ∇µt (x)
2

8.2 Intrinsic Formulation for µt (x) t ∈ [0, T ]

Definition 3 (MFG Equilibrium). The MFG equilibrium is the path of {µt }t∈[0,T ] with values in P(Rd ). Such that:
1. The optimization problem:

dXt = αt dt = dBt X0 ≈ m0

with cost
(  )
Z T 
1 2
J(α, µ) = E g(XT , mT ) + f (Xt , mt ) + |αt | dt
0 2

The above problem has one solution {Xt∗ }t∈[0,T ] .

2. For all t ∈ [0, T ] and P (Xt∗ ) = µt (x).
Essentially what we are saying is that if we freeze the population the best response is the theoretical distribution. This distribution
should fit the distribution of the population. The law of the best response is the environment itself. a In this model, the distribution
of the best response under the environment {µt (x)}t∈[0,T ] is the {µt (x)}t∈[0,T ] .
a Is the conclusion therefore in equilibrium the state is always in an optimal distribution. Since if it were not the distribution will change and move

towards it.

9 Tractable Formulation The MFG System

An MFG equilibrium is a fixed point. We now would like to formulate the problem of the MFG so as to facilitate the tractability
of the fixed point. This is just optimization of the performance criteria.

9.1 The PDE Approach

When we freeze the environment the control problem can be reduced to a stochastic optimal control problem. This is not a game
anymore. We can just solve the optimal control problem by minimizing the performance criteria subject to the dynamics of the
representative agent. The idea is to freeze the environment and let the initial condition vary. This is a one agent problem, the
tagged player. The initial condition is varied for the tagged player.
May, 09, 2021

Model 2 (The tagged player). The value function u(t, x) is given as,
( Z T  )
1 2
u(t, x) = inf E g(XT , mT ) + f (Xs , ms ) + |αs | ds
αs t 2
s∈[t,T ]

subject to:

dXt = αs ds + dBs , t ≤ s ≤ T, Xt = x

where mt is the equilibrium distribution. Note we have used m here as later we need to distinguish between m and the µ we
have used all along.

The above problem has a Hamilton Jacobi Bellman equation.

 
1 T 1 2
∂t u(t, x) + ∇ ∇u(t, x) + inf β∂x u(t, x) + |β| + f (x, mt ) = 0
2 β∈Rd 2
u(T, x) = g(x, mT )

βδx u(t, x) + 12 |β|2 we obtain the term − 21 |δx u(t, x)|2 . This would make the above system non-linear.

The minimization of inf
β∈Rd
If f, g are bounded and smooth, there are classical solutions to this problem. The solution is obtained through linearization via
the Cole-Hopf transformation. The optimal trajectory on solving the system is obtained to be,

α∗ = β ∗ = −∂x u(t, x).

This leads to the following equations.

dXt∗ = −∂x u(t, Xt∗ )dt + dBt .

The well posedness is as usual the lipschitz and boundedness of ∂x u(t, Xt∗ ). We still have to identify our distribution Xt∗ . This
is done with the Fokker Planck equation, after appropriately replacing the α term.
1
∂t µt (x) = −∇.[−∂x u(t, <>)µt (x)] + ∇T ∇µt (x)
2
Not that u(t, <>) depends on mt our theoretical distribution. We resolve this by setting µt = mt . Also taking into note that
f does not depend on all the Xi , ∇T ∇ the Hessian reduces to just the Laplacian1 . We now have the following,

Definition 4 (MFG System).

1
−∂t mt + ∇.[−∂x u(t, .)mt ] − ∇T ∇mt = 0
2
1 2 1 2
∂t u(t, x) + ∇ u(t, x) − |∂x u(t, x)| + f (., mt ) = 02 (9.1)
2 2
u(T, .) = g(., mT )

The understanding here is that as in a typical optimal control problem, you have a backward and forward pair. The
distribution equations (mt ) are forward equations. The state equations are backward equations. The distribution represents an
equilibrium distribution of the population at t. The function u(t, x) is the cost of the representative agent or the tagged player.

9.2 Solving the HJB

The HJB equation solves for the cost function and the theoretical distribution. The HJB equation is solvable under the following
conditions,
1. f, g functions are bounded, smooth in x with bounded derivatives of sufficiently high order.
1 We are using the notation from H M Schey.
2 The sign of ∂t is positive, I think the lectures had it as negative.
May, 09, 2021

The classical solution to the HJB when u, ∂x u are bounded is obtained using the transformation v(t, x) = eu(t,x) this linearizes
the system 3 .This may not work in all situations. This is called the Cole-Hopf transformation.

The problem we have is that f : Rd × P(R) 7→ R, g : Rd × P(R) 7→ R need to be bounded uniformly in the measure argu-
ment. This is because the Cole-Hopf transformation is an exponential transformation. Depending on the requirements you need
more and more conditions on these functions.

Once the HJB equation are solved, we can design a mapping as below for the measure and measure flow. The environment (the
entire path) µ = µt , t ∈ [0, T ] : [0, T ] 7→ P2 (Rd . We require the measure to be continuous for W2 -topology. This is a weak
topology on the space of measures using the Wasserstein-distance. The interesting point is that this continuity comes for free
in our setting 4 . We are also looking for time continuity of the cost function. This requires f to be continuous in the measure
arguments with respect to W2 . We call our classical solution that meet these criteria uµ .

9.3 The fixed point

Once the optimal control problem is solved. We now go to the dynamics and try to find the fixed point. Note that the actions
are fixed point of the best response function. We have the following,
dXt = −∂x uµ (t, Xt )dt + dBt
P (X0 ) = µ0

Proposition 1. Prove that P (Xt ) is automatically H2 (Holder Continuous) in W2 for a constant C that is independent of
µ.
1
Proof. Consider W2 (P (Xt ), P (Xs )), 0 ≤ s ≤ t. We know that W2 (P (Xt ), P (Xs )) ≤ E[|Xt − Xs |2 ] 2 . We now substitute for
Xt , Xs . We have,
"Z
t 2 # 12
1
2
E[|Xt − Xs | ] = E
2 −∂x u(r, Xr )dr + Bt − Bs
s

now we have the following, since ∂x u(t, Xt ) is bounded we have the following,
"Z
t 2 # 21
W2 (P (Xt ), P (Xs )) ≤ E −∂x u(r, Xr )dr + Bt − Bs
s
"Z
t 2 Z t
#
2
=E −∂x u(r, Xr )dr − 2(Bt − Bs ) ∂x u(t, Xr )dr + (Bt − Bs )
s s
√ √
≤α+β+γ t−s≤C t−s

Despite all this, it is not easy to solve this fixed point problem. To show that this a fixed point we need to show that the
mapping Φ : P(Rd ) 7→ P(Rd ) that creates the probability law of Xt is a contraction mapping in a suitable topology, W2 topology.
This is difficult to prove unless time is small. 5

Proposition 2. Φ is not a contraction mapping except if T is small.

Proof. The main issue for the proof is that the times directions are conflicting in the backward and forward equations.
Consider a one dimensional system.

ẋt = b(xt , yt ), x0 = a Forward ODE

ẏt = −f (xt , yt ) yT = g(xT ) Backward ODE

3 Allpapers by Olivier Gueant use this in the initial papers of MFG.

4I am not sure what exactly this means.
5 Note the difference between bounded and small. Small typically means relative to some other parameters of the problem.
May, 09, 2021

this is a two point boundary value problem in one dimension. Even when things are linear there are examples where a
solution does not exist. However, if b, f, g are Lipschitz there exists by contraction mapping arguments using Cauchy-
Lipschitz theory a solution to the problem if T is small. The meaning of small is interpreted relative to the Lipschitz
constants of b, f, g.

A simple counter example can be found in [3]. If there is no contraction argument, then you do them separately. Existence
and uniqueness as separate entities. Especially because we do not have a direct contraction argument. In conclusion we need to
separate uniqueness and existence.

9.4 Uniqueness of Solutions

This is non-typical. Lack of uniqueness causes difficulties in numerically solving the problem. This happens when the problem
has multiple equillibria. The question then becomes which one to select. In economics and markets, it may be that the equillibria
determines the current regime. The main condition is the Lasry-Lions monotonicity that guarantees uniqueness. There are other
criteria bu this is the most used. The question then becomes why or how does monotonicity help? Consider the system below,

ẋt = −yt , x0 = a, Forward Equations

ẏ = 0, yT = g(xT ), Backward Equations
Assuming a form for yt = U (t, xt ) we have he following,
ẏt = ∂t U (t, xt ) + ∂x U (t, xt )ẋt
using he fact that ẋt = −yt and that ẏt = 0, we have.
∂t U − U ∂x U = 0 (9.2)
(9.2) is the infamous inviscid Burgers equation. In the above system it is well posed if g is monotonically increasing. Try and
solve this for g.




 0 xT < −l

g(xT ) = xT −l ≤ xT ≤ l



 0 xT > l

you will see how the solution behaves. In the case of monotonically increasing g the solutions are indeed unique. When it is
not monotonically decreasing, the solution is unique depending on the time T . If T is small relative to Lipschitz constant.

The question is, what is the analogue for MFG?

Definition 5 (Lasry-Lions monotonicity). We say that function f satisfies Lasry-Lions monotonicity condition if, ∀m, m′ ∈
P(Rd ) and f we have,
Z
[f (x, m) − f (x, m′ )] d(m, m′ )(x) ≥ 0, ∀m ̸= m′ .
Rd

This is a fairly intuitive definition. All it is saying that there is a unique probability measure that maximizes expected value of
the given function. Stated simply this means,
′
E m [f ] ≥ E m [f ]. ∀m ̸= m′ .
R
Example 1 (MFG monotonicity). Consider h(x, µ) = Rd L(z, ρ∗µ(z))ρ(x−z)dz, we are trying to minimize this integral. Given
a µ that describes the empirical distribution, the minimum happens only when z is not in a dense neighborhood of x. This is
interesting, the question to ask is how does a choice of this distance function impact the MFG solution? This
works for some problems such as seat selection when you enter a bus, or train. You tend to find a spot where there is no others.
This phenomenon is expected to provide uniqueness.
May, 09, 2021

Proposition 3 (Monotonicity and Uniqueness of Solution). Monotonicity forces uniqueness.

Proof. If f = 0, and there are two equilibrium {µt }, {µ′ }. We have,

dXt = −∂x uµ (t, Xt ) + dBt

′
dXt′ = −∂x uµ (t, Xt′ ) + dBt

where µ = P (Xt ) and µ′ = P (Xt′ ), we must have the following

Z
′
E |∂x uµ (t, Xt ) − ∂x uµ (t, Xt′ |2 dt > 0
R

let us consider the cost functions.

uµ (t, Xt ) ≥ uµ (t, Xt′ )

′ ′
uµ (t, Xt′ ) ≥ uµ (t, Xt )

I am not sure the above condition is the right way to look at it. Since we solve a minimization problem, the right way to
do it would be,
′ ′ ′ ′
E µ [uµ (t, Xt )] ≤ E µ [uµ (t, Xt )], E µ [uµ (t, Xt′ )] ≤ E µ [uµ (t, Xt′ )]

Now take expectations of the performance criteria and since we are at a minimum we have,a
Z " Z T  # Z " Z T  #
1 2 1
g(x, µ) + f (x, t, µ) + αs ds dµ ≤ g(x, µ) + f (x, t, µ) + (αs ) ds dµ′
2
R d 0 2 R d 0 2

similarly for (Xt′ , µ′ ),

Z " Z T   # Z " Z T  #
′ ′ 1 ′ 2 ′ ′ ′ 1 ′ 2
g(x, µ ) + f (x, t, µ ) + (αs ) ds dµ ≤ g(x, µ ) + f (x, t, µ ) + (αs ) ds dµ
Rd 0 2 Rd 0 2

Noting that αs is not a function of µ adding the two will give us the following result,
Z T Z Z T Z
′
f (x, t, µ)d(µ − µ )ds − f (x, t, µ′ )d(µ − µ′ )ds
0 Rd 0 Rd
Z
+ (g(x, µ) − g(x, µ′ )) d(µ − µ′ ) ≤ 0
Rd

By our assumption f = 0 and this leaves us with,

Z
(g(x, µ) − g(x, µ′ )) d(µ − µ′ ) ≤ 0
Rd

The claim is that, this is a contradiction to our assumption that g is non-decreasing.

a Notice that the measures embed the Xt , Xt′ µ′X ′ , µXt and the X’s really get buried when taking expectations. The variable of integration
t
has nothing to do with it. The function is evaluated at the variable of integration against the measure.

We still do not know the structure of the utility function that actually gives you uniqueness. The issue is that contraction
mapping arguments cannot be made. That means we have to do roundabout conditions that facilitate uniqueness and existence.

9.5 Existence of Solution

We use Schauder Fixed Point theorem for proving fixed point. This depends on convexity and closedness of the space.
Theorem 1 (Schauder Fixed Point Theorem). Let T : C 7→ C, be a map from the closed convex subset C of a normed vector
space E. We assume that T ( C) is relatively compact in E. T has a fixed point.
May, 09, 2021

We make the following choices, Φ is the T . This can be found in [4] chapter 4. We chose
 Z 
d
E = ν : finite signed measures on R : |x|d|ν|(x) < ∞ .
Rd

We define on this space as.

Z
d
∥ν∥ = |ν(R )| + sup l(x)dν(x) .
l1−lipschitz
l(0)=0

Clearly C = P1 (Rd ).
Theorem 2 (Kantorovich-Rubinstein Duality). This can be found in Villani [5]. Given probability measures µ and µ′ on Rd
such that,
Z Z
|x|dµ(x) < ∞, |x|dµ′ (x) < ∞,
Rd Rd

then, ∥µ.µ′ ∥ = W1 (µ, µ′ ).

We are missing time, we will do the same as we did before, we set f = 0. Take µ ∈ P1 (Rd ). Looking to solve the HJB
equation we have,
1 1
∂t uµ + ∇2 uµ − |∂x uµ |2 = 0
2 2
uµ (T, .) = g(, µ)

These differential equations should map to the P (XT ), and dXtµ = −∂x uµ (t, Xtµ )+dBt . Our operator T in the Schauder theorem
is the PDE operator. We also have the t C = P1 (Rd ) is closed convex under W1 . What we need to prove is that our PDE
operator Lµ = ∂t uµ + 21 ∇2 uµ − 12 |∂x uµ |2 is continuous.

9.6 Continuity
′
Given µ, µ′ ∈ P1 (Rd ), there is not time. So these are simple measures. We look at the distance between P (XTµ ) and P (XTµ ).
We consider the Wasserstein distance.
′ ′
W1 (P (XTµ , P (XTµ )) ≤ E[|XTµ − XTµ |]
′
The above means that we compare ∂xµ and ∂xµ .
Z T
′ ′ ′
E[|XTµ − XTµ |] ≤ E |∂x u(t, Xtµ ) − ∂xµ (t, X µ )|dt
0

Applying ITO’s lemma we have,

1
d[∂x uµ (t, Xt )] = ∂xt uµ dt − ∂xx uµ ∂x uµ dt + ∂xx uµ dBt + ∂xxx uµ dt
2
1 1
= ∂x [∂t u + ∂xx u − |∂x u | ]dt + ∂xx uµ dBt 6
µ µ µ 2
(9.3)
2 2
1 µ
= ∂xx u dBt
2

We use the fact that given f = 0, we have from the HJB systems ∂t uµ + 12 ∂xx uµ − 12 |∂x uµ |2 = 0. We now have established
that ∂x u is a martingale. Now we can do something interesting,
Z T Z T
′ ′ ′
E |∂x u(t, Xtµ ) − ∂xµ (t, X µ )|dt ≤ E |∂x g(XTµ , µ) − ∂x g(XTµ , µ′ )|dt
0 0
6 Note u]2
that ∂x [∂x = 2∂x u∂xx u and this gives us the result here. This is an interesting result. So the delta of an option is a Martingale? Not
really, we need it to satisfy HJB. That is interesting.
May, 09, 2021

′
We have not much of an advantage since we still have to compare XTµ and XTµ . Instead we compare.
′ ′ ′ ′ ′
d(XTµ − dXTµ = −(∂x uµ (t, XTµ ) − ∂x uµ (t, XTµ ))dt ≤ ∥∂x uµ − ∂x uµ ∥∞ dt + C|Xtµ − Xtµ |dt
′
Where C ≈ ∥∂xx W∥∞ . We now have to show that if W1 (µn , µ) → 0 ⇒ ∥∂x uµn − ∂x uµ ∥∞ → 0. The latter relation (∥∂x uµn −
′
∂x uµ ∥∞ ) is the regularity condition of the HJB equation with respect to parameter. The remainder is in he references provided.

9.7 Relative Compactness of the Range

Looking at Φ : P(Rd ) 7→ P(Rd ), the result of the map is P (XTµ ). From the SDE of Xt we have the following,
Z T
XT = X0 − ∂x uµ (t, Xtµ )dt + BT
0

we have X0 = µ0 and given that |∂x uµ (t, Xtµ )| ≤ C, we can see that.
Z
E[|XT |2 ] ≤ C, if E[|X02 |] = x2 dµ0 (x) < ∞.
Rd

The above result implies that, Rd |x|2 dP (XTµ )(x) ≤ C. We now use the result from [5] that states that µ ∈ P(Rd ) : Rd |x|2 dµ(x) ≤ C
R  R

is relatively compact in P(Rd ) equipped with W1 . We now have all the pieces for applying the Shauder fixed point theorem.

9.8 If f ̸= 0 ?

Here we cannot just work with the terminal value of the population. We need to do as follows. Define M1 = E = ν : finite signed measur
We now define E = C 0 ([0, T ], M1 ) (set of continuous mappings from [0, T ] 7→ M1 ). We define C = C 0 ([0, T ], P1 (Rd )). We define
Φ : C 7→ C. We do the same thing as before, the continuity Φ is the continuity of the HJB differential operator with respect to a
parameter. For relative compactness is similar as before when f = 0. We prove that sup Rd |x|2 dP (Xtµ )(x) ≤ C. This essen-
R
0≤t≤T
√
tially needs that the variance be bounded for the entire process Xtµ . We then show that W1 (P1 (Xtµ ), P1 (Xsµ )) ≤ C t − s, 0 ≤
s ≤ t ≤ T . This proves ideas of equicntinuity and uniform continuity. For more details chapter 4 of [4].

10 Other Approaches
So far we looked at PDE approach to the solution of mean-field-games. There are also probabilistic approach to mean-field-games.
In the probabilistic approach,
1. Reformulate the MFG system as a McKean-Vlasov FBSDE.
2. Unlike in the PDE approach here we follow the cost along the evolution of the optimal trajectory.
With the given MFG system, optimal trajectory in the environment µ = {µt }t ,

dXt = −∂x uµ (t, Xt )dt + dBt

We define Yt = uµ (t, Xt ) as the cost of the remaining path of the system at time t. We expand Yt using Ito’s lemma.
1
dYt = ∂t uµ (t, Xt )dt + ∂x uµ (t, Xt )[−∂x uµ (t, Xt )dt + dBt ] + ∂xx uµ (t, Xt )dt
  2
1
= ∂t uµ (t, Xt ) + ∂xx uµ (t, Xt ) − |∂x uµ (t, Xt )|2 dt + ∂x uµ (t, Xt )dBt
2

So we use the PDE (9.1) for replacing the drift term by − f (Xt , µt ) + 12 |∂x uµ (t, Xt )|2 to obtain,


 
1
dYt = − f (Xt , µt ) + |∂x uµ (t, Xt )|2 dt + ∂x uµ (t, Xt )dBt
2

Now we define Zt = ∂x uµ (t, Xt ) to get a Backward SDE.

May, 09, 2021

 
1
dYt = − f (Xt , µt ) + |Zt |2 dt + Zt dBt
2
YT = g(XT , µT )

The theory for these BSDE is by Peng and Pardoux in 1990s. We note that there are two unknowns in this BSDE, the value
function Yt = uµ (t, Xt ) and the gradient of the value function, Zt = ∂x uµ (t, Xt ), which is unusual. The options pricing system is
a BSDE. It should be noted that Yt is adapted process, this implies that Yt is σ(X0 , Bs |s≤t ) measurable. The Brownian motion
makes the process non-anticipative.

10.1 The fixed point problem

Here we identify µt = P (Xt ), we end up with a McKean-Vlasov type equations.

dXt = −Zt dt + dBt

 
1
dYt = − f (Xt , µt ) + |Zt |2 dt + Zt dBt
2
YT = g(XT , µT )

10.2 Solvability
We assume that f, g have bounded derivatives uniformly in the measure argument. f, g are Lipschitz continuous in the measure
argument, w.r.t W1 . These are the same type of constraints as the PDE. This is again Chapter 4, [4]. The equations here are
for the cost.

11 McKean Vlasov FBSDE for Optimal Feedback

This approach uses the Pontryagin Minimum principle. This is based on the necessary condition for the optimal trajectory. The
Potryagin Minimum Principle is very robust. Under the randomization of the environment this is more robust as coming with
the HJB equations is harder. We again return to the HJB equation. We here follow the optimal feedback instead of the cost,
The optimal feedback,

Yt = ∂x u(t, Xtµ )

we apply Ito’s lemma, but we have a non-zero f unlike the last time we did in 9.6, (9.3). We have,
1 1
dYt = ∂x [∂t uµ + ∂xx uµ − |∂x uµ |2 ] + ∂xx uµ dBt
2 2
= −∂x f (Xtµ , µt )dt + ∂xx uµ (t, Xt )dBt dXt = −∂x uµ (t, Xtµ )dt + dBt

Simplifying the above we get,

dXt = −Yt dt + dBt

dYt = −∂x f (Xtµ , µt )dt + ∂x Zt dBt
YT = g(XT , P (XT ))

12 Examples - Linear Quadratic Systems

13 Back to the N player Game
Our goal is to say how the solution of N → ∞ game is meaningful relative to the finite player game. There are two approaches,
1. Solve the finite game and let N → ∞ and prove that the equilibria converge to a solution of the MFG.
2. Start from the solution of the MFG. Use the equilibrium feedback strategy of the MFG in the original game. See what
happens.
May, 09, 2021

13.1 Implementing the limit Optimal Feedback

1. Do not need Uniqueness.

Achdou and Lavriere are the references for this, Chapter 6 [6].
Refrences: Gangbo, Meszarios, Zhang, Mou, Daniel Lacber

14 MASTER EQUATION AND CONVERGENCE PROBLEM

References Cardaliaguet D, Lasry Lions. Chassagneux Crisan D). Lions 2008 Lectures at College De France. 2013 Seminar.
Gangbo Swiech (2015) Bensoussan co-authors. Gangbo Mesgaros Mou Zhang (2020) - Weaker solution but uniqueness of equi-
libria.

There is a requirement of uniqueness of Equilibrium is subtle and important. Mean field games in finite state space and
potential games that facilitate solutions to Master Equations.

14.1 Equilibrium Value for the MFG

We want to define a equilibrium value for the MFG. This is an equilibrium cost for a given player for any possible initial
conditions. Here we ask a simple question, Can we associate an equilibrium wiht (t, x, µ). Our setting still remains the same,

1. Dynamics : dXt = αs ds + dBs , Xt = x.

2. Control acts during the time t ≤ s ≤ T .
3. Performance Criteria:
" #
Z T
1 2
J(α, m) = E g(XT , mT ) + [f (Xs , ms ) + |αs | ]ds
t 2

4. HJB:
1 1
∂t usm + ∇2 um m 2
s − |∂x us | + f (x, ms ) = 0
2 2
um (T, x) = g(x, mT ) (14.1)

5. Marginal laws m of the optimal trajectory flow according to the Fokker-Planck equation,
1
∂s ms − ∇2 ms − div(∂x um
s ms ) = 0.
2

The above system is solvable if f, g are smooth in x, bounded with bounded derivatives and Lipschitz in measure. In order to
avoid shocks in the system (not sure what this means) we require the Lasry-Lions monotonicity.
Z
[f (x, m) − f (x, m′ )] d(m − m′ )(x) > 0,
Rd
Z
[g(x, m) − g(x, m′ )] d(m − m′ )(x) > 0.
Rd

The monotonicity condition enforce uniqueness to solutions of the FB MFG system.

Goal 1. Construct an equilibrium for some time initial t. Given an initial population µt . We would like to construct µs
t≤s≤T
given by the MFG system.

We know the optimal control that is solved by the tagged player. The best cost function under µ is uµ (t, x) if Xt = x. Since
u is determined uniquely by µ, we denote u as uµ (t, x) as the value of the game. Here implicitly t is the initial time for both x
and µ.
May, 09, 2021

Definition 6 (Value of the Game). We call the value uµ (t, x) the value of the game. Mathematically,

U : [0, T ] × Rd × P2 (Rd ) 7→ R
(t, x, µ) 7→ uµ (t, x)

Goal 2. Show that U solves a PDE. This PDE is the master equation.
We rewrite our dynamics with a new notation to denote the time the population distribution was initialized,
dXs = −∂x ut,µ (s, Xs )ds + dBs
Here we know that ut,µ solves the HJB (14.1) on [t, T ]. We now take an incremental step h and restrict s ∈ [t + h, T ] where
0 < h <= T − h. Now we change the environment from (µs )t≤s≤T ⇒ (µs )t+h≤s≤T . What we are essentially doing is moving the
environment along time using the uniqueness property. Instead of keeping the environment fixed at a chosen time. Our notation
essentially is µt = µ and µt+h = µt,µ
t+h . This has an impact on our value function notation. We claim that,
t,µ
ut,µ (t + h, .) = ut+h,µt+h (t + h, .)
t,µ
The above is notationally ugly. ut+h,µt+h (t + h, .) is the value function at t + h using the environment µt+h at t + h. It is obtained
t,µ
as a result of the flow from t, µt → t + h, µt+h . This latter term ut+h,µt+h (t + h, .) is essentially our newly defined function U .
t,µ
We basically have U (t + h, x, µt+h ). Summarizing we have,

ut,µ (t + h, x) = U (t + h, x, µt,µ
t+h )

.
We now go back to the optimal trajectory and rewrite the dynamics in terms of U .
dXs = −∂x U (s, Xs , µt,µ
s )ds + dBs

If we now say that we initialize Xt ∼ µ then P (Xt ) = µ, Since we are at equilibrium at each step,
dXs = −∂x U (s, Xs , P (Xs ))ds + dBs
In the above equation µs = P (Xs ). We have another interesting observation to make. The action function that we had in the
original McKean-Vlasof equation and termed αi has an excellent candidate. This is the ∂x U .

14.2 The ROAD to the PDE in U

As before we look at the Dynamic Programming Principle of Bellman to arrive at the PDE for U . For this we need to expand
U . Chapter 5 [4] describes how this is done.
BUCHDAHN Li PENG REINER AOP 2017
This becomes the master equation problem. This requires derivatives in measure space and some other mathematics that I
do not know yet. I will have to revisit this.

15 Convergence Problem
Going from the infinite to the finite problem is one approach The other is to solve the N -player game and see if as N → ∞ it
converges. The question is, in the limit do we get a solution to the MFG? This is called the convergence problem. This is a
difficult problem.

15.1 Approaches
1. The Master equation comes to the rescue. The rate of convergence is the benefit from this approach.
2. Compactness on the equilibrium controls of the N-player game. This is too demanding. There is no proof at this time
using Compactness.
3. Relaxed Controls: This is a compactification of standard controls. The controls are replaced by measures. This is a
relaxation process. Daniel Laker This is less demanding in terms of assumptions. there is no rate of convergance.
All the above methods are complementary.
May, 09, 2021

15.2 The Master Equation Method

We start from our original problem, We have stochastic dynamics and we have replaced the interaction with empirical distribution.

dXti = αti dt + dBti

"  #
Z T 
i i 1 i2
J (α · · · ) = E g(XTi , µN
T ) + XTi , µN
T ) + |αt | dt
0 2

15.3 Finding the NASH Equilibrium

We assume Markov strategies. Here we assume the action process will be as given below,
αti = αi (t, Xti , · · · , XtN )
You then look for a PDE for the game. We then freeze the responses of all the players but i. We then compute the best response
of player i. This is done by calculating the Value function and the optimal feedback as before using the HJB equations. The
PDE for player i. We denote the value function of player i as u(t, x1 , · · · , xN ).
 
1 X 1
∂t ui (t, x) + ∇2 ui (t, x) + αj (t, x).∂xj ui (t, x) + inf β.∂xi ui (t, x) + |β|2 + f (xi , µN
x )=0
2 2
j̸=i

ui (T, x) = g(xi , µN
x )

We now argue that the equilibrium is when the best response is the optimal response. So we go back to our optimal response
αi = −∂xi ui (t, x). Substituting this back to our Nash system under the assumption that every ones optimal action is given by
the partial derivative of ui w.r.t xi . We now have,

1 X 1
∂t ui (t, x) + ∇2 ui (t, x) − ∂xj uj (t, x).∂xj ui (t, x) + |∂xi ui (t, x)|2 + f (xi , µN
x )=0
2 2
j̸=i

ui (T, x) = g(xi , µN
x )

This is a PDE system. The curse of dimensionality damns it. This is a non-degenerate PDE system, it becomes degenerate
in the limit. There are a series of paper by Bensoussan, Fiehse in the 90s that look at these problems. Carmonas book gives all
these references Chapter 6 [6].

15.4 The Classical Solution

The classical solution implies uniqueness and a lot of symmetry. Changing the index of the PDE still solves the same problem.
The solution retains the symmetry of the model.
ui (t, x) = ui (t, xi , x−i )
We see that the solution cost function is symmetric in the argument x−i . It also seems to me that every player has the same
optimal cost. Somehow we have reduced a heterogeneous system to a homogeneous system. How does information asymmetry
impact the MFG? Is the MFG approach really saying it is really not relevant this information Asymmetry. in our equation
above, ui : [0, T ] × Rd × Rd⊗N −1 7→ R. We can use this symmetry and say the following,
ui (t, x) = uN (t, xi , x−i ) (15.1)
In (15.1) we should note that the cost is dependent on the number of players, N . Interestingly, in the case of asymmetric
information the cost for players will be different. From here we use the symmetry and replace the individuals with their
distribution.
1 X
ui (t, x) = uN (t, xi , x−i ) = U (t, xi , δx−i )
N −1
We note that we can redefine xi in a manner as to include the it h-particle in the distribution calculation.
1 X 1 X
U (t, xi , δx−i ) = U (t, xi , δ xi )
N −1 N
We note that N1 i δxi ∈ P(Rd ), where the measures are all of he uniform type. The question is then the extension of this to
P
space of all probability measures. The master equation approach makes an attempt at this.
May, 09, 2021

15.5 The ASNATZ

This is the same as we did in Lecture #1. The ansatz is the existence of a U : [0, T ]×Rd ×Rd 7→ R and that ui (t, x) ≈ U (t, xi , µN
x ).
We take this as exact equality and we use the U in our Nash-System. Now the previous lecture on derivatives on measure space
dU
come in. Note that the measure is a function of xj here. So dx j
= ∂xj U + ∂µ U ∂x µ. We have the following

∂t ui (t, x) ≈ ∂t U (t, xi , µN
x )

X N
X
− ∂xj uj (t, x).∂xj ui (t, x) ≈ − ∂xj U (t, xj , µN N
x ).∂xj U (t, xi , µx )
j̸=i j=1
X 1
 
1

N N N
=− ∂xj U (t, xj , µx ) + ∂µ U (t, xj , µx ) . ∂µ U (t, xi , µx )
N N
j̸=i
   
1 1
− ∂xi U (t, xi , µNx ) + ∂ µ U (t, xi , µN
x ) . ∂ xi U (t, x i , µN
x ) + ∂µ U (t, xi , µN
x )
N N
 
1 X 1
= −|∂xi U (t, xi , µNx )|2
− ∂ µ U (t, xi , µN
x )(x j ).∂ xj U (t, xj , µN
x ) + O
N j N
h P i
N
Note that ∂xi N1 1 δxi (xi ) = N1 . We now write the middle term N1 j ∂µ U (t, xj , µN N
P
x )(xj ).∂xj U (t, xj , µx ) as an integral. This
was shown in the master equation lectures. We then obtain,
Z
1 X
∂µ U (t, xi , µN
x )(x j ).∂xj U (t, xj , µN
x ) = ∂xj U (t, xj , µN N N
x ).∂µ U (t, xi , µx )(xj )(ν)dµx (ν)
N j R d

Substituting back we have,

 
1 X 1
−|∂xi U (t, xi , µN 2
x )| − ∂µ U (t, xi , µN
x )(x j ).∂ xj U (t, xj , µN
x ) + O =
N j N
Z  
1
−|∂xi U (t, xi , µN
x )|2
− ∂xj U (t, xj , µN
x ).∂µ U (t, xi , µN
x )(x j )(ν)dµ N
x (ν) + O
R d N

We next take a look at 12 ∇2 ui (t, x)

N N   
1 2 i 1X n 
N
o 1 X
N 1 N
∇ u (t, x) ≈ T r ∂xj ∂xj U (t, xi , µx ) = T r ∂xj ∂xi U (t, xi , µx ) + ∂µ U (t, xi , µx )(xj )
2 2 j=1 2 j=1 N
 
1 1 X 1 X 1 X  
= ∂xi xi U (t, xi , µN
x )+ ∂xi U (t, xi , µN
x )+ ∂µ U (t, xi , µN
x )+ T r ∂xi µ U (t, xi , µN
x )(xj )

2 N N2 N j
j̸=i j̸=i
Z  
1 2 N
h
N
i
N 1
= ∇ U (t, xi , µx ) + T r ∂xi ∂µ U (t, xi , µx )(xj ) dµx (xj ) + O
2 Rd N

We again note that this is a term from the master equation. Essentially when we consider the Nash system and apply DPP
we do recover the terms from the Master Equation. We can piece the other two remaining pieces f, g.

15.6 From Master Equation to Nash

Here we go from the Master Equation to the Nash. We assume Lasry-Lions monotonicity, smooth coefficients. We start from
the classical solution U to the master equation. Let u(t, x) = U (t, xµ , µN
x ). The interpretation is that U is a projection to R
d⊗N
.

Proposition 4 (Close Solution to Nash). The claim is that the projection (ui )1≤i≤N is almost a solution to the Nash
system.
May, 09, 2021

Our Nash system is as below using u.

1 X 1
∂t ui (t, x) + ∇2 ui (t, x) − ∂xj uj (t, x).∂xj ui (t, x) + |∂xi u(t, x)|2 + f (xi , µN i
x ) + r (t, x) = 0
2 2
j̸=i

ui (T, x) = g(xi , µN
x )

1


The term ri (x, x) is of the order O N .

15.7 Comparison

16 Schrodinger Approach to MFG [1]

May, 09, 2021

References
[1] I. Swiecicki, T. Gobron, and D. Ullmo, “Schrödinger approach to mean field games,” Physical review letters, vol. 116, no. 12,
p. 128701, 2016.

[2] A.-S. Sznitman, “Topics in propagation of chaos,” Lecture notes in mathematics, pp. 165–251, 1991.
[3] J. Ma and J. Yong, Forward-backward stochastic differential equations and their applications. No. 1702, Springer Science &
Business Media, 1999.
[4] R. Carmona and F. Delarue, “Probabilistic theory of mean field games: vol. i, mean field fbsdes, control, and games,”
Stochastic Analysis and Applications. Springer Verlag, 2018.

[5] C. Villani et al., Optimal transport: old and new, vol. 338. Springer, 2009.
[6] R. Carmona and F. Delarue, “Probabilistic theory of mean field games: vol. ii, mean field games with common noise and
master equations,” Stochastic Analysis and Applications. Springer Verlag, 2017.