Price Formation and Mean-Field-Games: A Review [14] [16] [3] [5] [4] [7] [12]

[11]

Ramesh Kadambi

March 1, 2025

1
May, 09, 2021

Contents
1 Introduction - Mean Field Games Formulation [2] 4
1.1 Game Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.1 Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Differential Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 The Mean Field Game 5

2.1 The HJB formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Optimal Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Bellman’s Optimal Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Dynamic Programming Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 The HJB Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.6 Stochastic Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6.1 Ito’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.7 Stochastic Optimal Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.8 The HJB Equation Stochastic Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.8.1 The Stochastic HJB equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.9 Differential Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.9.1 Stochastic Differential Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Mean Field Game 10

3.1 The Mean Field m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Linear Quadratic Cost Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Cole Hopf Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Numerical Methods for MFG 14

4.1 MFG PDE System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

5 Price Formation Lasry-Lyons[14] 16

5.1 Invariance Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5.2 Research Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.4 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.5 Stationary Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.6 Some comments on the PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

6 Asymptotics for a symmetric equation in price Formation [13] 20

6.1 The Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
6.2 The Modeling Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.3 Existence Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.3.1 Motivation in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.3.2 Existence in a Bounded Interval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6.4 Asymptotic Decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6.5 Fokker Planck Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

7 Asymptotics for a free boundary model in price Formation [9] 24

8 The Long Term Behavior 24

9 Global existence and uniqueness of solutions to a model of price formation [8] 25

10 On a price formation free boundary model by Lasry and Lions [4] 26

10.1 The Lasry Lions Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
10.2 Connection to the heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
10.3 Asymptotic Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
May, 09, 2021

11 On A Price Formation Free Boundary Model by Lasry Lions: Neumann Problem [5] 29
11.1 Analysis of the Neumann problem - transformation to the heat equation . . . . . . . . . . . . . . . . . . . . . . . 29
11.2 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
May, 09, 2021

1 Introduction - Mean Field Games Formulation [2]

The mean field game is a formulation where agent positions are modeled as diffusion processes. The drift of the position reflects
the agents actions. The agent actions are dependent on the distribution of the population states. Each agent acts to optimize
his payoff. MFG borrows ideas from different fields such as Game Theory, Optimal Control and Differential Games.

1.1 Game Theory

Definition 1 (Game). A game in strategic form is an ordered triple G = (N, (Si )i∈N, (ui )i∈N) where,
1. N = {1, 2, 3, · · · , N } is a finite set of players.
Q
2. Si is the set of strategies of player i, for every player i ∈ N. The set of all vectors of strategies is denoted as S = i Si =
S1 × S2 × · · · × SN ; and

3. ui : S → R is a utility function of an individual agent for every agent i ∈ N a .

a Here we just use a utility function interchangeably with the pay off function. In reality one could consider a utility function and a pay off function

both separately, and agents maximize their utilities.

Definition 2 (Pure Strategy). A pure strategy is an action taken by an agent with certainty given the information of the agent.

Definition 3 (Mixed Strategy). Given a game G = (N, (Si )i∈N, (ui )i∈N) in strategic form with finite strategy set for each player
i. A mixed strategy of an agent i is the probability distribution over a setPof strategies Si . So the optimal strategy is a distribution
Pi : Si → [0, 1], over the set of all probabilities Σi = {Pi : Si → [0, 1] : si Pi (si ) = 1} for each agent i ∈ N.

The strategic form game in definition (1) is extended to a mixed strategy game by maximizing expected utility of the payoff
as opposed to just the payoff.

1.1.1 Nash Equilibrium

Definition 4. A Nash equilibrium is defined as a strategy specification s∗ = {s∗1 , · · · , s∗N }, by all the agents such that for each
agent,

u(s∗ ) ≥ u(si , s∗−i ).

in the case of pure strategy and,

EP∗ [u(s∗ )] ≥ EPi [u(si , s∗−i )].

in the case of mixed strategies.

The Nash equilibrium is the fixed point of the best response function. The best response function is a vector function B : S → S ,
where S is the strategy product spaces as defined in definition (1). The best response function B = [B1 (s1 , s−1 ) B2 (s2 , s−2 ) · · · BN (sN , s−
where Bi (si , s−i ) is the best response function of the individual agent i given as,

Bi (sk , s−k ) = max Bi (sj , s−k ).

sj ∈Si

Given the structure of distributions to check if a mixed strategy is a Nash equilibrium or not, it is sufficient to check only the
deviations in pure strategies.

1.2 Optimal Control

Optimal control extends the static optimization problem to a dynamic situation. The state variables are now functions of time
that are defined by a differential equation. An optimal control problem is stated as below,

J(u∗ ) =minJ(u)
u∈ U
subject to: ẋ = f (x, u, t).

Where u∗ represents the optimal value of u and J(u∗ ) is the minimum cost based on u∗ .
May, 09, 2021

1.3 Differential Game

As seen in section (1.1), an N -player game has N decision makers, each optimizing their pay off Ji (u∗i , u−i ) = min Ji (ui , u−i )
u∈ Ui
for all i ∈ N, where u∗i represents the optimal action of the ith agent ui . The cost Ji (u∗i , u−i ) is the minimum cost based on
(u∗i , u−i ). A differential game extends an optimal control problem to multiple decision makers. A differential game is stated as
below,
Ji (u∗i , u−i ) = min Ji (u∗i , u−i )
ui ∈ Ui

subject to: ẋ = f (x, ui , u−i , t),

as before u∗i represents the optimal value of ui and Ji (u∗i , u−i ) denotes the minimum cost based on u∗i and u−i .

2 The Mean Field Game

A mean field game combines the ideas from sections (1.1,1.2,1.3). A mean field game can be stated as,
J(u∗ , m) =minJ(u, m)
u∈ U
subject to: ẋ = f (x, u, m, t)
In the case of the mean field game, the agent takes into account the mean field m and not individual actions. The mean field
appears as one takes the limit as the number of agents goes to infinity. The control u∗ is the optimal value against the mean field
m. The cost J(u∗ , m) denotes the minimum cost for the given optimum action u∗ and the mean field m. The question what
is the mean field? Could it be any observable quantity that reflects the action of the agents? The mean field
game converts a differential game to an optimal control problem in the mean field. The figure above illustrates the

1
Figure 2: Relation of MFG to other fields

relation of MFG to the other fields.

2.1 The HJB formulation

In the case of the Mean-Field-Game the number of agents is large. The agents therefore maximize against the collective behavior
of the other players. In this scenario the collective behavior can be represented by a mean-field against whose density the agents
maximize their objective function. In such a case the Mean-Field-Game system reduces to a HJB system of the value function
v(x, t) and a Fokker-Planck-Kolmogorov equation for the mean-field density m(x, t),
∂v
− − H(x, m, ∇v, t) = κ∆2 v(x, t)
∂t
1 This figure is taken from [2].
May, 09, 2021

∂m
(x, t) + ∇ . [f (x, u, m, t)m(x, t)] = κ∆2 m(x, t)
∂t
with boundary conditions m(x, 0) = m0 (x) and v(x, T ) = g(x) and a non-negative parameter κ. The function H is called the
Hamiltonian, which is optimized varying the control variable u. ∇ is the gradient operation, and ∆2x is the Laplacian.

2.2 Optimal Control Theory

In standard optimal control theory there is a single actor that is trying to optimize an objective function. The objective function
is dependent on a control function and the state of the system. The deterministic optimal control problem is setup as below,

Definition 5 (Optimal Control Problem). Let f (x, u, τ ) : Rn × Rm × [t, T ] → Rn be a function referring to instantaneous change
of x w.r.t. τ . The state of the system at time τ ∈ [t, T ] is denoted by mapping x(τ ) : [t, T ] → Rn . The evolution of the system is
given as,

ẋ = f (x(τ ), u(τ ), τ )
x(t) = x

where x ∈ Rn is fixed initial state, t ≥ 0 is the initial time, and T is the final time or terminal time. The control
u(τ ) : [t, T ] → Rm is a measurable function with values from a compact subset of U of Rm . It belongs to a set of admissible
controls U = {u(τ ) : [t, T ] → U }.

The cost function or the running cost function is given by r(x, u, τ ) : Rn × Rm × [t, T ] → R and g(x) : Rn → R the
terminal cost. The performance index is given by a functional form,
Z T
J(u) = r(x(τ ), u(τ ), τ )dτ + g(x, T )
t

Given an initial state x(t) = x, deterministic control problem is finding an optimal control u∗ (.) that minimizes the cost functional,
J,
"Z #
T
J(u∗ ) = min J(u) = min r(x(τ ), u(τ ), τ )dτ + g(x(T ), T ) ,
u(.)∈ U u(.)∈ U t

subject to: ẋ = f (x(τ ), u(τ ), τ )

x(t) = x.

Definition 6 (The Value Function). The value function is he cost of implementing an optimal control. It is given as,
"Z #
T
v(x, t) = min J(u) = min r(x(τ ), u(τ ), τ )dτ + g(x(T ), T ) ,
u(.)∈ U u(.)∈ U t

v(x, T ) = g(x(T ), T )

furthermore the optimal control satisfies the relation,

Z T
∗
J(u ) = r(x(τ ), u(τ ), τ )dτ + g(x(T ), T ) = v(x, t).
t

2.3 Bellman’s Optimal Policy

Problems of optimization in time have an interesting property. As time goes on the choice of optimality will only depend on the
future. The current state should typically reflect any decisions taken in the past and the effects of such a decision. Any future
decision should account for this impact. This is what Bellman says,
An optimal policy has property that, whatever the initial state and initial decision are, the remaining decision
must constitute an optimal policy with regard to the outcome resulting from the initial decision.
May, 09, 2021

2.4 Dynamic Programming Principle

Application of of the Bellman principle we get for the value function v(x, t) at some initial time t given some later time s and
state xs = x(s),
Z s 
v(x, t) = min r(x(τ ), u(τ ), τ )dτ + v(xs , s) .
u(.)∈ U t

Theorem 1 (DPP). Let v(x, t) be the value function at initial time t,

"Z #
T
v(x, t) = min r(x(τ ), u(τ ), τ )dτ + g(x(T ), T ) ,
u(.)∈ U t

and v(x, T ) = g(x(T ), T ) be the value function at terminal time T . The dynamic programming principle states that for an time
s that satisfies s ∈ (t, T ], the value function in the state x at initial time t can be calculated backwards from v(xs , s) as,
Z s 
v(x, t) = min r(x(τ ), u(τ ), τ )dτ + v(xs , s)
u(.)∈ U t

where xs = x(s).

2.5 The HJB Equation

One approach to solve the Optimal Control problem is to find a PDE for the value function. In many cases the value function
may not be differentiable. In this case one subdivides the interval [t, T ] into sub-intervals of length δ and apply DPP on each
sub-interval leading to a PDE that the value function v(x, t) needs to satisfy.
"Z #
t+δ
v(x, t) = min r(x(τ ), u(τ ), τ )dτ + v(x(t + δ), t + δ) (2.1)
u(.)∈ U t

Now as always, Taylor series comes to the rescue.

∂v
v(x(t + δ), t + δ) = v(x, t) + δ + [x(t + δ) − x(t)] . ∇v(x(t), t) + o(δ),
∂t
where ∇ is the gradient operator in the spacial co-ordinates. We next note from the differential equation for x,
Z t+δ
x(t + δ) − x(t) = f (x(τ ), u(τ ), τ )dτ.
t

Now the value function can be written as,

Z t+δ
∂v
v(x(t + δ), t + δ) = v(x(t), t) + δ + f (x(τ ), u(τ ), τ )dτ . ∇v(x(t), t) + o(δ)
∂t t

substituting into (2.1) we have,

"Z #
t+δ Z t+δ
∂v
v(x, t) = min r(x(τ ), u(τ ), τ ) dτ + v(x(t), t) + δ + f (x(τ ), u(τ ), τ )dτ . ∇v(x(t), t) + o(δ)
u(.)∈ U t ∂t t

Noting that v(x(t), t), δ ∂v

∂t and o(δ) are all independent of u(t). By canceling v(x, t) and dividing by δ we obtain,
"Z #
t+δ Z t+δ
∂v 1 ∂v
(x(t), t) + min r(x(τ ), u(τ ), τ ) dτ + δ + f (x(τ ), u(τ ), τ )dτ . ∇v(x(t), t) + o(δ) = 0
∂t δ u(.)∈ U t ∂t t

by taking limit as δ → 0 and noting the following,

R t+δ
t
ϕ(x(τ ), u(τ ), τ )dτ
lim → ϕ(x(t), u(t).t)
δ→0 δ
May, 09, 2021

o(δ)
lim →0
δ→0 δ
we obtain the HJB equation,
∂v
(x, t) + min [r(x, u, t) + f (x, u, t) . ∇v(x, t)] = 0
∂t u(.)∈ U

subject to: v(x, T ) = g(x(T ), T ).

We now define a function called the Potryagin State Function[17] (also referred to as Hamitonian),
H(x, p, t) = min [r(x, u, t) + f (x, u, t) . p]
u(.)∈ U

where p = ∇v(x, t).

2.6 Stochastic Optimal Control

If the dynamics of the state variables is given by stochastic differential equation, then we need to apply Ito’s lemma to obtain
our HJB equations. Ito’s lemma adds another term to our HJB equation that results from the volatility term.

2.6.1 Ito’s Lemma

Given X(t) ∈ Rn , f (t) ∈ Rn and σ(t) ∈ Rn×l and W (t) ∈ Rl2 . The dynamics of X given as below,
dX(t) = f (t)dt + σ(t) dW (t)
X(0) = x0
For a twice differentiable function ϕ with respect to x. We have,
 
∂ϕ 1
dϕ(X(t), t) = + f . ∇ϕ + T r[σ ⊤ H(ϕ)σ] dt + (∇ϕ)⊤ σdW (t)
∂t 2

2.7 Stochastic Optimal Control Problem

The stochastic optimal problem only has dynamics defined by a stochastic differential equation instead of deterministic dynamics.
This adds a volatility term to our HJB equation. The stochastic optimal control is defined as below,
Definition 7 (Stochastic Optimal Control). The objective is to minimize the functional,
"Z #
T
J(u∗ ) = minJ(u) = min E r(X(τ ), u(τ ), τ )dτ + g(X(T ), T ) ,
u∈ U u∈ U t

subject to: dX(t) = f (t)dt + σ(t) dW (t),

X(0) = x0 .

2.8 The HJB Equation Stochastic Optimal Control

We follow the same steps as before with the deterministic dynamics. However we have one additional term due to volatility. The
Bellman principle is just applied to the expected value. We therefore have the following for Bellman principle.
Theorem 2 (Stochastic Bellman Principle). Let v(x, t) be the value function at initial time t,
"Z #
T
v(x, t) = minJ(u) = min E r(X(τ ), u(τ ), τ )dτ + g(X(T ), T )
u∈ U u∈ U t

and v(x, T ) = g(x(T ), T ). The dynamics programming principle states that, at any time t < s ≤ T , the value function in state
X at initial state time t can be calculated backwards from v(Xs , s) as below,
Z t 
v(x, t) = min E r(X(τ ), u(τ ), τ )dτ + v(Xs , s) .
u∈ U s
2 We assume these are independent Brownian Motions.
May, 09, 2021

2.8.1 The Stochastic HJB equation

The stochastic HJB follow fairly easily from Ito’s lemma. We have the following,
(Z
t+δ Z t+δ  
∂v
v(x, t) = min E r(X(τ ), u(τ ), τ ) + v(x, t) + + f . ∇v(X(τ ), τ ) dτ
u∈ U t t ∂t
Z t+δ Z t+δ )
1  ⊤ 
+ T r σ H(v(X(τ ), τ ))σ dτ + [∇v(X(τ ), τ )σdW (τ )]
t 2 t

Dividing across by δ and taking limits one obtains,

∂v 1
+ min[r(X(τ ), u(τ ), τ ) + f . ∇v(X(τ ), τ )] + T r[σ ⊤ Hσ] = 0, 3
∂t u∈ U 2
subject to: v(X(T ), T ) = g(X(T ), T ).

As before we define the Pontryagin (Hamiltonian) state function as,

H(x, p, t) = min[r(X(t), u(t), t) + f . p(t)]

u∈ U
where p(t) = ∇v(X(t), t)

2.9 Differential Game

In a differential game we have a set of N players whose cardinality N = |N|. The state of the system is now defined by
x(t) : [0, T ] 7→ Rn . The control of player i at time t is denoted by ui (t) : [0, T ] 7→ Rm . The dynamics is given by a system of
ODEs,

dx(t) = fi (x, ui , u−i , t)dt

x(0) = 0

where all the terms have their standard meanings from game-theory. The function f (x, ui , u−i , t) : Rn ×Rm ×Rm · · ·×Rm ×[0, T ] 7→
RN is continuous with respect to all its arguments. Each agent has a cost functional,
Z T
Ji (ui , u−i ) = ri (x, ui , u−i , t)dt + gi (x(T ))
0

where ri (x, ui , u−i , t) : Rn × Rm × · · · × Rm × [0, T ] 7→ R refers to the running cost and gi (x) : Rn 7→ R refers to the terminal cost.
As before solving the minimization problem,
"Z #
T
v(x, t) = min Ji (ui , u−i ) = min ri (x, ui , u−i , t)dt + gi (x, T ) .
ui ∈ Ui ui ∈ Ui 0

This leads to a system of HJB equations,

∂vi
+ Hi (x, pi , t) = 0
∂t
subject to: vi (x(T), T) = gi (x(T).T).

Where,

Hi (x, pi , t) = min [ri (x, ui , u−i , t) + fi (x, ui , u−i , t) . pi (t)]

ui ∈ Ui

where pi (t) = ∇vi (x, t)

3 Here we assume that σ is a constant vector or at least not a function of u.
May, 09, 2021

2.9.1 Stochastic Differential Game

As before if the dynamics are given by stochastic differential equations then there is an additional volatility term that comes
into the picture.

Definition 8 (Stochastic Differential Game). A stochastic differential game with agents in N with |N| = N we have,
"Z #
T
Ji (u∗i , u−i ) = min Ji (ui , u−i ) = min E ri (x, ui , u−i , t)dt + gi (x, T )
ui ∈ Ui ui ∈ Ui 0

subject to: dXi (t) = f (x, ui , u−i , t)dt + σi (x, ui , u−i , t)dWi (t).

Definition 9 (Nash Equilibrium). The Nash equilibrium of the solution of the above differential game refers to the control u∗
that satisfies,

Ji (u∗i , u∗−i ) ≤ Ji (ui , u∗−i ),

for all i ∈ N.

The Nash equilibrium, above is obtained by simultaneously solving Hamilton Jacobi system as below,

∂vi
(x, t) + Hi (x, pi , qi , t) = 0
∂t  
1
where Hi (x, pi , qi , t) = min ri (x, ui , u−i , t) + fi (x, ui , u−i , t) . pi (t) + T r[σi⊤ (x, ui , u−i , t)]qi (t)σi (x, ui , u−i , t) ,
ui ∈ Ui 2
and pi (t) = ∇vi (x, t), qi (t) = H[vi (t)].

3 Mean Field Game

In the MFG the agents are large in number and the agents are replaced by an dynamic evolving mean field. Our starting point
is a non-cooperative differential game with N -players. The state of the game is represented by X(t) : [0, T ] 7→ Rn . The control
of player i at time t is denoted by ui (t), while the control profile u−i contains control of players other than player i. The running
cost function is ri (x, ui , u−i , t) : Rn × Rm × cdotsRm × [0, T ] 7→ R, and gi (x) : Rn 7→ R is a terminal cost function at state X(T ),
and fi (x, ui , u−i , t) : Rn × Rm × · · · Rm × [0, T ] 7→ R is a drift function.
The goal of each player is to find a control ui ∈ Ui that minimizes the cost function Ji (ui , u−i ) subject to the dynamics
dX(t).

Definition 10 (MFG problem).

"Z #
T
min Ji (ui , u−i ) = E ri (x, ui , u−i , t)dt + gi (X(T ))
ui ∈ Ui 0

subject to: dX(t) = fi (x, ui , u−i , t)dt + σi dWi (t)

X(0) = x0

where Wi (t) denotes a standard Wiener process and accounts for the uncertainties in the state, and σi the diffusion constant.

Definition 11 (The Value function). The value function vi (x, t) be defined mathematically as,

vi (x, t) = min Ji (ui , u−i ), ∀t ∈ [0, T ].

ui ∈ Ui

Then the optimal control u∗i satisfies,

"Z #
T
Ji (u∗i , u−i ) =E ri (x, u∗i , u−i , t) + gi (X(T )) = vi (x, t), ∀t ∈ [0, T ]
0
May, 09, 2021

The value function as usual satisfies the HJB system,

∂vi 1
− (x, t) − Hi (x, pi , t) = σi2 ∇2 vi (x, t)
∂t 2
where Hi (x, p, t) = min [ri (x, ui , u−i , t) + fi (x, ui , u−i , t) . p(t)] , and pi = ∇vi .
ui ∈ Ui

When the number of players becomes large we can replace the other players actions by a mean field distribution of the state of
the players. The system is reduced to describing a representative player with state x ∈ X and control u ∈ U. The mean field
game can now be expressed as a pair of HJB and FPK equations.

∂v 1
− (x, t) − H(x, m, p, t) = σi2 ∇2 v(x, t)
∂t 2
∂m 1 2 2
(x, t) + ∇.(f (x, u, m)m(x, t)) = σ ∇ m(x, t),
∂t 2
where H(x, m, p, t) = min [r(x, u, m, t) + f (x, u, m, t) . p(t)] , the Pontryagin state function.
u∈ U

3.1 The Mean Field m

Definition 12. The mean field m(x, t) is probability distribution of the players state x at time t. Mathematically,
N
1 X
m(x, t) = lim δxi (x).
N →∞ N
i=1

where,

 1 if x = xi
δxi (x) = .
 0 if x ̸= xi

The interesting part of this mean-field game is the notion of what the mean field can represent. For the limit order book, it can
perhaps represent the mean wait time. For a given or chosen random quantity, the mean wait time can determine the shape of
the book.
May, 09, 2021

Theorem 3 (The Mean Field Game). Consider a non-cooperative game among large number of indistinguishable players. If
every player faces the optimization problem,
"Z #
T
minJ(u, m) = E r(x, u, m) + g(X(T ), m(x, T )) ,
u∈ U 0

subject to: dX(t) = f (x, u, m, t)dt + σdW (t), (3.1)

X(0) = x0

the equivalent non-cooperative mean field game is represented by the pair of HJB and Fokker Planck Equations,

∂v σ2 2
− (x, t) − H(x, m, p, t) = ∇ v(x, t)
∂t  2
∂m ∂H 1 2 2
(x, t) + ∇ . (x, m, p, t)m(x, t) = σ ∇ m(x, t), (3.2)
∂t ∂p 2
subject to: v(x, T ) = g(x, m(x, T )), m(x, 0) = m0 (x),
where, H(x, m, p, t) = min[r(x, u, m, t) + f (x, u, m, t) . p(t)],
u∈ U
p(t) = ∇v(x, t)

In addition the optimal control u∗ (x, t) is the solution to the equation,

∂H
f (x, u∗ , m, t) = (x, m, p, t)
∂p

such that the mean field m when u = u∗ (3.1) coincides with the solution m∗ of the FPK equation in (3.2).

Proof. The HJB equations are pretty straight forward to prove as we have done in the case of optimal control problems. It is a
straight forward application of Ito’s Lemma, taking expectations and limits.
Now consider a distribution function ϕ which is twice differentiable of the process X(t) with dynamics governed by (3.1).

σ2 2
 
d[ϕ(x)] = f (x, u, m, t).∇ϕ(x) + ∇ ϕ(x) dt + ∇ϕ(x)σdW (t)4
2

dividing by dt and taking expectations,

σ2 2
   
dϕ(x)
E = E f (x, u, m, t) . ∇ϕ(x) + ∇ ϕ(x)
dt 2

since E[∇ϕ(x)σdW (t)] = 0.

Rewriting in terms of the mean field m(x, t) and using integration by parts,

σ2
Z Z Z
∂ϕ
(x)m(x, t)dx = [f (x, u, m, t)] . ∇ϕ(x)]m(x, t)dx + ∇2 ϕ(x)m(x, t)dx
X ∂t X 2 X
σ2
Z Z Z
∂
m(x, t)ϕ(x)dx = − ∇.(f (x, u, m, t)m(x, t))ϕ(x)dx + ∇2 m(x, t)ϕ(x)dx
∂t X X 2 X
∂m 1
+ ∇ . (f (x, u, m, t)m(x, t)) = σ 2 ∇2 m(x, t) (3.3)
∂t 2
∂H
using the fact that f (x, u, m, t) = ∂p we can rewrite the equation (3.3) as,
 
∂m ∂H 1
(x, t) + ∇ . (x, m, p, t)m(x, t) = σ 2 ∇2 m(x, t)
∂t ∂p 2

Most papers in literature start from (3.1) and proceed.

4 Note here that the last multiplication is a matrix style multiplication. So we do not need a ∇.ϕ(x) dot product.
May, 09, 2021

3.2 Linear Quadratic Cost Function

The simplest of MFG are the linear quadratic cost function problems. The liner quadratic cost function is as below,
Z T 
1
J(x, u) = g(XT , mT ) + r(Xs , ms ) + ∥αs ∥2 ds
0 2
subject to: dXs = αs du + σs dWs

To obtain the optimal control we solve the following a HJB system with Hamiltonian given by,
∥αt ∥2
H(x, m, p, t) = min + αt . p(t)
αt ∈ U 2
This is is easy to solve and we have,
∂H
(x, m, p, t) = 0
∂αt
αt + p(t) = 0
αt = −p(t) = −∇v(x, t)
⇒ αt = −∇v(x, t) (3.4)

Substituting (3.4) into our Hamiltonian we get,

∥∇v(x, t)∥2 ∥∇v(x, t)∥2

H(x, m, p, t) = − ∇v(x, t).∇v(x, t) = −
2 2
We now have a HJB formulation that looks as,
Z T
∂v ∥∇v(x, t)∥2 σ2 2
− (x, t) − + r(Xs , ms )ds = ∇ v(x, t)
∂t 2 0 2
∂m 1 2 2
− ∇ . (∇v(x, t) m(x, t)) = σ ∇ m(x, t)
∂t 2
In one dimension we have the following if r(Xs , ms ) = 0,

1 σ2 2
∂t v(x, t) − (∂x v(x, t))2 + ∂ v(x, t) = 0
2 2 xx
1 2
∂t m(x, t) − ∂x ((∂x v(x, t)) m(x, t)) − ∂xx m(x, t) = 0
2

3.2.1 Cole Hopf Transformation

Consider the following PDE,

∂t u − a∇2 u + b∥∆u∥2 = 0
Given: u(0, x) = g(x)

where x ∈ Rn , a, b are constants. ∇ is the gradient operator, and ∇2 , ∆ is the Laplacian. ∥.∥ is the Euclidean norm in Rn . Let
w = ϕ(u),

∂t w = ϕ′ (u)∂t u (3.5)
′ ′′
∆w = ϕ ∆u + ϕ ∥∇u∥2

substituting for ∂ut into (3.5) we have,

∂t w = ϕ′ [a∆u − b∥∇u∥2 ] (3.6)

again substituting for ∆u into the (3.6)

∂t w = [a∆w − aϕ′′ ∥∇u∥2 ] − bϕ′ ∥∇u∥2

May, 09, 2021

∂t w = a∆w − [aϕ′′ (u) + bϕ′ (u)]∥∇u∥2

Now constrain ϕ(u) such that,

aϕ′′ (u) + bϕ′ (u) = 0

ϕ′′ (u) b
=−
ϕ′ (u) a
b
⇒ ϕ′ (u) = ke− a u

By letting ϕ′ (0) = 1 we have k = 1 and the following,

b a
w = e− a u ⇒ u(x, t) = − ln w(x, t)
b
After the transformation we are left with,

∂t w = a∆w

we can use separation of variables to solve this problem, Let

w(x, t) = f (t)g(x)

we therefor have,

gft = af gxx
ft gxx
=a = −k
f g
solving for the time equation we have,

f (t) = αe−kt

solving for the space equation we have,

agx x + kgx = 0
r
2 k k
⇒ D + = 0 ⇒ D = ±i
a a
The solution,
h √k √k i
w(x, t) = αe−kt c1 ei a x + c2 e−i a x

we no substitute for u to obtain,

a  h √k √ k i
u(x, t) = − ln αe−kt c1 ei a x + c2 e−i a x
b

4 Numerical Methods for MFG

The notation and problem setup here is a bit different, we will use the notation from the YouTube lectures5 .

dXti = b(t, Xti , α(t, Xti ), mN i

t )dt + σdWt , X0i ∼ m0
1
∂t m(x, t) − σ 2 ∇2 m(t, x) + ∇ . (b(x, t, α(x, t))m(x, t)) = 0
2

The cost functions, the two common cases,

5 Open Doctoral Lectures: Numerical methods for Mean Field Games.
May, 09, 2021

1. the non-local (typically - regularized using a Gaussian Kernel)

Z T
J(x, t) = f (u, Xu , αu , mu )du
t

2. The local (if the population distribution has a density, still denoted by m),
Z T
J(x, t) = f (u, Xu , αu , m(u, Xu ))dt
t

The only complication here is that one needs to calculate the gradient of the value function ∇u(x, t). We proceed as we always
do in solving these problems. We discretize the system and solve them. The issue is the initialization

4.1 MFG PDE System

The starting point of the MFG problem is the following system of equations.

0 = −∂t u(x, t) − ν∇2 u(x, t) + H(x, m(t, .), ∇u(x, t))

0 = ∂t m(x, t) − ν∇2 m(x, t) − ∇ , (m(., t)∂p H(., m(t, ∇u(t, .))))(X)
with: u(T, x) = g(x, m(T, ∗)), m(0, x) = m0 (x)
May, 09, 2021

5 Price Formation Lasry-Lyons[14]

Lasry-Lyons makes an explicit periodicity assumption. How important is it? Lasry-Lyons models the population density (shape
of the book). They postulate an external randomness in price preferences. The densities are described fb , fs , two non-negative
functions of (x, t). Where x is a price and t is time. The price p(t) is the price obtained by dynamic equilibrium. There is
friction measured by a positive parameter a such that 2a can be interpreted as the bid-ask spread. The mean field equations are
described to be,

Figure 3: The bid-ask spread and Moving Boundary.

1
∂t fb − σ 2 ∂xx fb = λδ(x − p(t) + a), if x < p(t), t > 0
2
fb ≥ 0, fb (x, t) = 0 if x ≥ p(t), t ≥ 0 (5.1)

on the sell side we have,

1
∂t fs − σ 2 ∂xx fs = −λδ(x − p(t) + a), if x > p(t), t > 0
2
fv ≥ 0, fv (x, t) = 0, if x ≤ p(t), t ≥ 0.

In additions we have the following for λ,

1 1
λ = − σ 2 ∂x fb (p(t), t) = σ 2 ∂x fs (p(t), t) (5.2)
2 2
The constant λ measures the number of transactions at time t, (i..e., the flux of buyers which must be equal to the flux of
vendors). The parameter σ > 0 measures the randomness. And δ denotes the usual delta functions
R δ0 or a smoothed version
of it, that is a smooth non-negative function with compact support in (−a, +a) and such that δ = 1. We have the following
initial condition,

fb |t=0 = fb0 , fs |t=0 = fs0 .

In addition we assume,

fb0 > 0 if x < p0 , fb0 (x) = 0, if x ≥ p0

fv0 > 0 if x > p0 , fs0 (x) = 0, if x ≤ p0

for some p0 ∈ R.

5.1 Invariance Property

The Bak paper [1] and Slanina [18] make an assumption regarding the total sum of buyers and sellers,
Z Z
fb (x, t)dx = k = − fs (x, t)dx
R R
May, 09, 2021

We can now differentiate with respect to t and we have the following,

Z Z p(t) Z p(t)
d ∂fb
fb (x, t)dx = fb (x, t)dx = (x, t)dx
dt R −∞ −∞ ∂t
p(t) p(t)
σ2
Z Z
= ∂xx fb + λδ(x − p(t) + a)dx
−∞ 2 −∞
p(t)
σ2
= ∂x fb + λ(t)
2 −∞
σ2
= ∂x fb (p(t), t) + λ(t) = 0.
2
Similarly,

d ∞
Z Z Z ∞
d ∂fs
fs (x, t)dx = fs (x, t)dx = (x, t)dx
dt R dt p(t) p(t) ∂t
Z ∞ 2 Z ∞
σ
= ∂xx fs (x, t)dx − λδ(x − p(t) + a)dx
p(t) 2 p(t)
∞
σ2
= ∂x fs − λ(t)
2 p(t)
σ2
=− ∂x fs (p(t), t) − λ(t) = 0.
2

5.2 Research Topics

1. Derivation of the model from Nash Points and utility maximization.
2. Extension to more general dynamics.

3. To situations with several possible goods or where transactions may involve more than one unit quantity of good.

5.3 Main Results

Reduction of the system (5.1)-(5.2) which is not really necessary for our argument but allows to simplify the presentation,
consider

f (x, t) = fb (x, t) if x ≤ p(t) = −fs (x, t) if x > p(t)

essentially this is he same as doing f = fb − fs . The system now reduces to,


2 2
 ft − σ2 fxx = − σ2 fx (p(t), t) {δ(x − p(t) + a) − δ(x − p(t) − a)} on R × (0, ∞)



f (x, t) > 0, if x < p(t), t ≥ 0; f (x, t) < 0 if x > p(t), t ≥ 0, (5.3)


 f|

= f , on R, p(0) = p .
t=0 0 0

5.4 Assumptions
C
1. We assume that f0 is a smooth function on R with fast decay at infinity (fast decay to infinity means that 1+|x|2 for some
positive C).
2. We assume f0 (x) > 0 if x < p0 and f0 (x) < 0 if x > p0 . Essentially there exists a price p0 such that the function f0 is split
into a positive and a negative section.
3. Essentially it is required that p0 to be an equilibrium price at t = 0, i.e. f ′ (p0+ ) = f ′ (p0− ) or (fb0 )′ (p0 ) = −(fs0 )′ (p0 ).

Though this is not strictly necessary for our analysis but if we do not make this assumption, f 0 cannot be smooth
at p0 and quite a few technicalities that we wish to avoid in this survey have to be incorporated.
May, 09, 2021

Theorem 4. Under the above conditions, there exists a smooth solution (f, p) of (5.3) such that f has fast decay for all t ≥ 0.
Both from a theoretical viewpoint and from a numerical approximation viewpoint, we also investigate the time-implicit dis-
cretization of (5.3) which takes the following form,

 λ2 f − fxx = g − fx (p)(δp−a − δp+a )



f (x) > 0 if x < p (5.4)



 f (x) < 0 if x > p

2
where λ > 0 (in a discretization, λ2 = σ 2 ∆t where ∆t is he time step...)

This is straight forward to arrive at, ft = f (t + ∆t) − f (t) ∆t. We then denote f (t) = g, we obtain (5.4). Here, g is given a
smooth function (with fast decay) such that we have for some p.

g > 0, if x < p0 , g < 0 if x > p0 . (5.5)

1 −λ|x| σ 2 d2
The function 2λ e is the Green’s function of the operator λ2 − 2 dx2 and we can recast (5.4) as,

1 df n o
f =G− (p) e−λ|x−p+x| − e−λ|x−p−a| (5.6)
2λ dx
1 −λ|x|
where G = 2λ e ∗ g. Hence, f (p) = G(p). Hence f (p) = G(p). Therefore, p will be determined provided we show that G
df
has a unique zero. Furthermore, in order to determine dx (p), we observe that, if (5.5) is differentiated and we chose x = p, we
deduce from (5.6),
df dG df
(p) = (p) + e−λa (p)
dx dx dx

5.5 Stationary Problems

The stationary form of (5.3) on a fixed interval (0, A), and A > 2a, we consider the following,

 − σ2 f = − σ2 f (δ
2 xx 2 x p−a − δp+a )
 f (0) = f (A) = 0, f > 0 if x < p, f < 0 if x > p, a < p < A − a
x x

Rp
where f models the total number of agents and the total number of goods, or N1 = 0
f dx is the number of buyers and
RA
N2 = p (−f )dx the number of sellers. Observing that we have,

fx = 0 if x < p − a or if x > p + a
fx = fx (p) if p − a < x < p + a

The solution to the equation is pretty straight forward, straight up integration gives.

f (x) = Θ(x − p + a)1{x>p−a} − Θ(x − p − a)1{x>p+a} + C1 x + C2

Solving for constants based on the specification of the derivatives leads to, the following function as solution.



 Θa if x < p − a

f (x) = −Θ(x − p) if x ∈ [−a, a] (5.7)


 −Θa

if x > p + a

Now we can compute N1 and N2 as,

Z p−a Z p
N1 = Θadx + (−Θ)(x − p)dx
0 p−a
2
(p − a)2
 
p
= Θa(p − a) − Θ − + Θp2 − Θp(p − a)
2 2
May, 09, 2021

Θa2
= Θap −
2
2
Similarly, we can find that N2 = Θa(A − p) − Θ a2 . From the paper we get the following conditions,
N1 −N2 |N1 −N2 |
1. N1 +N2 = 2p−A a 6
A−a and a solution exists if N1 +N2 < 1− A−a . A restriction which corresponds to the restriction p ∈ (a, A−a)...
N1
We also note that p is an increasing function of the ratio N 2
. This is natural property from an economic view point, as
buyers grow the price goes up.

5.6 Some comments on the PDE

What does this PDE really mean? What is the notion of equilibrium? Are we determining the equilibrium price or the equilibrium
shape? What is the significance of λ(t)?
I do not have answers to all the questions. However, the one question that I have an answer for is λ(t). This is the rate of
trades occurring. If we discretize the problem, we have this,
1 2
fb (t) − fb (t − ∆t) = σ fxx ∆t − λ(t)∆tδ(x − p(t) + a)
2
1
fb (t) = fb (t − ∆t) + σ 2 fxx ∆t − λ(t)∆tδ(x − p(t) + a)
2
At the boundary,
1
fbn (t) = ft−∆t
n
+ σ 2 fxx ∆t − λ(t)∆t
2
Essentially the quantity depletes at some rate of trading. Note that if the price p(t) has dynamics (Here looks exogenous), or a
notion of equilibrium price dynamics then the buy and sell curves move and price formation occurs.
6 This seems like a statement on the liquidity and perhaps toxic order flow and market price not existing. Essentially such no-trade theorems are

interesting in the study of toxic-order flow.

May, 09, 2021

6 Asymptotics for a symmetric equation in price Formation [13]

The paper studies a particular case of the Lasry-Lions mean-field BVP and the existence and asymptotics for large time of the
solutions to a one dimensional evolution equation with non-standard right hand side. The right hand side is a derivative of the
solution computed at a given point. The paper proves the existence through a fixed point argument. The problem is then solved
on a bounded interval and is shown to decay to a stationary state.

6.1 The Problem Setup

The setup is that of idealized players of tow groups, buyers and sellers. There is one good. The buyers and sellers are distributed
according to densities fb and fs on (x, t) ∈ R × R+ . The dynamic equilibrium price is represented by p(t). There is friction
measured by a positive parameter a. There is a parameter σ that measures the randomness. The problem can be described as
below,

2 2
∂fb
 − σ2 ∂∂xf2b = −λ(t)δx (p(t) − a) if x ≤ p(t), t > 0
 ∂x2


(6.1)


 f (t)(x, t) > 0 if x < p(t), f (x, t) = 0, f (x, t) if x ≥ p(t)

b b b

together with the sell side given by,


2 2
∂fs
 − σ2 ∂∂xf2s = λ(t)δx (p(t) + a) if x > p(t), t > 0
 ∂x2


(6.2)


 f (t)(x, t) > 0 if x > p(t), f (x, t) = 0, f (x, t) if x ≤ p(t)

s b b

where,

σ 2 ∂fb σ 2 ∂fs
λ(t) = − (p(t), t) = (p(t), t) (6.3)
2 ∂x 2 ∂x
The constant λ(t) represents the number of transactions at time t 7 , so (6.3) means that the flux of buyers which must be equal
to flux of vendors. The initial conditions,

fb (x, 0) = fb0 (x) and fs (x, 0) = fs0 (x)

are such that, for some p0 in R,

fb0 (x) > 0 if x < p0 , fb0 (x) = 0 if x ≥ p0

fs0 (x) > 0 if x > p0 , fs0 (x) = 0 if x ≤ p0

The equation satisfy the property of conservation of mass. Indeed, both

Z p(t) Z ∞
fs (x)dx and fb (x)dx
−∞ p(t)

remain constant for all time t ≥ 0.

Equations (6.1),(6.2) and (6.3) describe a mean field model for the dynamical formation of the price of some good. The
important question answered here is the following, will the good reach a stable price or will the price oscillate in time? The
answer to this question can be found through the study of the large time behavior of the system. However, the asymptotics
when t → ∞ of such a model is not known.
This paper establishes the asymptotic behavior of (6.1)-(6.3), under certain assumptions on initial data.
1. The initial data are symmetric with respect to a general given point p0 , i.e. fb0 (p0 − x) = fs0 (p0 + x) for all x > 0. This
essentially says that the amount of buy vs sell is the same at a given price level. In this case the price p(t) resulting from
the dynamical equilibria of the system is a constant function p(t) = p0 for all t > 0 8 .
∂fb
7I need to think about this a bit. On the buy side at p(t) we can actually have ∂t
= −λ(t) and sell side the opposite is true ( ∂fs
∂t
) = λ(t)) if fs > 0.
8I do not understand this aspect of the dynamic equillibrium.
May, 09, 2021

The above problem reduces to the single equation,

∂f σ2 ∂ 2
− = λ(t)[δx=−a − δx=a ] (6.4)
∂t 2 ∂x2
f (x, 0) = fI (x)
where
σ 2 ∂f
λ(t) := − (0, t)
2 ∂x
and
f :=fb − fs , and fI = fb0 − fs0 .

The paper does the following,

1. A direct proof of existence,
2. The solution to the symmetric problem (6.7) decays exponentially to a stationary state when the problem is considered in
a bounded interval.
Benefits from symmetry:
1. The symmetric solution eliminates the free boundary p(t) and makes the problem accessible while giving an idea of the
general picture.
2. The linearized problem (6.1)-(6.3) has surprisingly no free boundary and maintains similarity to (6.7). This facilitates
understanding the asymptotics of the general problem through the linear problem.

6.2 The Modeling Perspective

A convection term is introduce to keep the prices reasonable 9 . This implies a convective term with confining effects.

σ2
ft − fxx − (xf )x = λ(t)[δx=−a − δx=a ]
2

σ2
f (x, 0) = fI (x), λ(t) = − fx (0, t) (6.5)
2
The problem (6.5) has a non-trivial stationary solution de to the presence of confinement. The asymptotics cannot be proven,
but existence of solutions can be shown. The systems (6.7) and (6.7) are related by some suitable self-similar re-scaling as
indicated in [6].

6.3 Existence Proof

We assume without loss of generality, that σ 2 /2 = 1.

6.3.1 Motivation in R
We seek a function f (x, t), x ∈ R, t > 0, f ∈ L∞ (0, T, L1x (R)), ∀T > 0, odd in x-axis, that solves

ft − fxx = λ(t)[δx=−a − δx=a ]

f (x, 0) = fI (x) (6.6)

where the initial condition satisfies fI (x) = −fI (−x) such that fI (x) ≥ 0 for x ≤ 0. The condition λ(t) = − ∂f
∂x (0, t) implies
conservation of mass for the equation. Since the mass is conserved, it should be conserved for all time,
Z 0 Z 0
f (x, t)dx = fI (x)dx ∀t > 0.
−∞ −∞
9I am not sure how this works out? Is this a kind of damping?
May, 09, 2021

Theorem 5 (Proof Existence). Let fI ∈ L1 (R) ∩ C (R) be an odd function in R, positive for x < 0, fI (0) = 0, such that fI ∈ C 0,1
for x = 0. Problem (6.6) has a unique odd, positive for x < 0, solution f (x, t) ∈ L∞ (0, T, C 0,1 (R) ∩ L1 (R)) for any T > 0. In
particular,

f (x, t) ∈ C (0, T, C ∞ (R −a, a)).

Moreover, λ(t) > 0 for all t ∈ [0, T ].

Proof. The proof is purely based on Duhamel’s principle and a fixed point argument.

6.3.2 Existence in a Bounded Interval

The problem is solved in a bounded interval (−π/2, π/2). The boundary conditions that are chosen are Neumann conditions.
The problem is reformulated into a reflection problem with zero-Dirichlet boundary conditions on the interval [0, π]. It is problem
(6.7) with Neumann conditons.

∂f σ2 ∂ 2
− = λ(t)[δx=−a − δx=a ] (6.7)
∂t 2 ∂x2
fx (−π/2, t) = fx (π/2, t) = 0, f (x, 0) = fI (x) (6.8)

where,
∂f
λ(t) := − (0, t)
∂x

The problem is recast or reformulated around the point x = π/2, as boundary value value problem with zero-Dirichlet boundary
conditions on the interval [0, π]: we seek an even function f with respect to x = π/2, positive in (0, π) with f (0) = f (π) = 0,
that solves

ft − fxx = λ(t)[δx=a + δx=π−a ], xin(0, π), t > 0

f (0, t) = f (π, t) = 0, f (x, 0) = fI (x), (6.9)
λ(t) = fx (0, t)

The initial datum fI is odd with respect to the point x = 0 and satisfies compatibility conditions at the boundary with fI (x) ≥ 0
for x ∈ [−π/2, 0]. The Neumann boundary condition in (6.8) reduces to fx (π/2, t) = 0 for (6.9). The paper states the following
theorem,

Theorem 6 (Existence Neumann Condition). Let Ω = (0, π), 0 < a < π/2, and let fI ∈ C(Ω) be a symmetric function in Ω
with respect to x = π/2, positive, such that the Fourier series for fIx (x) converges at x = 0 and fI (0) = fI (π) = 0. Then problem
(6.9) has a unique solution f (x, t) ∈ L∞ (0, T, C 0,1 (Ω)) for all T > 0 that is positive and symmetric with respect to x = π/2. In
particular,

f (x, t) ∈ C (0, T, C ∞ (Ω{a, π − a}))

Moreover, λ(t) > 0 for all t ∈ [0, T ].

The proof is similar to the one before, we use Frourier series and a fixed point argument for λ(t).

6.4 Asymptotic Decay

The exponential decay of solutions of (6.9) towards the steady state given. For simplicity in the computations, we translate the
problem from the bounded interval (0, π) to Ω := (−π/2, π/2).
May, 09, 2021

Theorem 7 (Asymptotic Decay). Consider the following problem,

ft − fxx = λ(t)[δx=−π/2+a + δx=π/2−a ], x ∈ (−π/2, π/2), t > 0

f (−π/2, t) = f (π/2, t) = 0, f (x, 0) = fI (x) (6.10)
λ(t) = −fx (π/2, t)

Under the hypothesis of theorem 6 for fI , after translation, the unique solution of f of (6.10) decays exponentially to the unique
stationary state f∞ , given by

 β(x − π/2) if π/2 − a ≤ x ≤ π/2,
f∞ :=
 β if 0 ≤ x ≤ π/2 − a
R0 R0
and extended evenly to the negative axis. Here β is the only constant that preserves mass, i.e. π/3
f∞ (x)dx = −π/2
f (x)dx for
all t > 0. Moreover, the following limit exists,

lim λ(t) = β.
t→∞

The conservation of mass is key. If there are sources and sinks, then the dynamics is such that they have to preserve the mass.
The proof revolves around Fourier expansions and convergence, need to look into it in a bit more detail.

6.5 Fokker Planck Version

The Fokker-Planck version of the problem involves an extra convection term. The paper seeks solutions f (x, t), x ∈ R, f ∈
R0
L∞ (0, T, L1 (R)), ∀T > 0, f odd in x, and with finite first moment −∞ |x|f (x, t)dx < ∞, of he problem

ft − fxx − (xf )x = λ(t)[δx=−a − δx=a ]

f (x, 0) = fI (x), λ(t) = −fx (0, t) (6.11)

The above equation is a FPK equation with a harmonic confining potential.

Theorem 8 (FPK Variant Existence). Let fI ∈ L1 (R) be an odd function, positive for x < 0, fI (0) = 0, such that fI ∈ C 0,1 at
x = 0, and with finite first moment,
Z 0
|x|fI (x)dx < ∞
−∞

The problem (6.11) has unique odd, positive for x < 0, solution f (x, t) ∈ L∞ (0, T, C 0,1 (R) ∩ L1 (R)) for all T > 0, and with finite
first moment. In particular,

f (x, t) ∈ C (0, T, C ∞ (R −a, a)).

Moreover, λ(t) > 0 for all t ∈ [0, T ].

Proof. Look at [13]

May, 09, 2021

7 Asymptotics for a free boundary model in price Formation [9]

This paper and [13] show that the free boundary disappears in the linearized problem. The paper then looks at it as a pertur-
bation through semi-group theory. In carefully crafted and selected spaces that give better regularity properties near the free
boundary, the paper shows global existence for solutions with initial data in a small neighborhood of any equilibrium point, the
system decays exponentially towards a stationary state. The observation is that the family of equilibria of the equation is stable
and follows from center manifold theory.

The details of the problem are setup as before, (6.1), we reproduce here,


2
∂fb σ 2 ∂ fb


 ∂x2 − 2 ∂x2 = −λ(t)δx (p(t) − a) if x ≤ p(t), t > 0

(7.1)


 f (t)(x, t) > 0 if x < p(t), f (x, t) = 0, f (x, t) if x ≥ p(t)

b b b

together with the sell side given by,


2 2
∂fs
 − σ2 ∂∂xf2s = λ(t)δx (p(t) + a) if x > p(t), t > 0
 ∂x2





 f (t)(x, t) > 0 if x > p(t), f (x, t) = 0, f (x, t) if x ≤ p(t)

s b b

where,
σ 2 ∂fb σ 2 ∂fs
λ(t) = − (p(t), t) = (p(t), t) (7.2)
2 ∂x 2 ∂x
10
The constant λ(t) represents the number of transactions at time t , so (7.2) means that the flux of buyers which must be equal
to flux of vendors. The initial conditions,

fb (x, 0) = fb0 (x) and fs (x, 0) = fs0 (x)

are such that, for some p0 in R,

fb0 (x) > 0 if x < p0 , fb0 (x) = 0 if x ≥ p0

fs0 (x) > 0 if x > p0 , fs0 (x) = 0 if x ≤ p0

The equation satisfy the property of conservation of mass. Indeed, both

Z p(t) Z ∞
fs (x)dx and fb (x)dx
−∞ p(t)

remain constant for all time t ≥ 0. fb and fs are the buy and sell densities. The price is supposed to form based on equilibrium
11
.

8 The Long Term Behavior

One of the points of interest is the long term behavior of the price. Does the price reach a stable value or will it oscillate? as
time t → ∞. The problem (7.1)-(7.2) is investigated under the assumptions,
1. the domain is a bounded interval [−A, B], A, B > 0, for a < min{A/2, B/2},
2. under Neumann boundary conditions.
The objective is to show that if initial conditions are in some neighborhood of a general equilibrium point in some suitable
function space, then there is a unique solution to (7.1)-(7.2) that decays exponentially fast in time to a unique stationary state.
In addition, it can be shown that the problem presents a two dimensional family of equilibria, and that family is stable.
10 I need to think about this a bit. On the buy side at p(t) we can actually have ∂fb
∂t
= −λ(t) and sell side the opposite is true ( ∂fs
∂t
) = λ(t)) if fs > 0.
11 How is the equilibrium formed? Purely by one diffusion process hitting the other. The price is basically calculated as part of the solution?
May, 09, 2021

9 Global existence and uniqueness of solutions to a model of price formation [8]

This is one of the first papers that analyzed the Lasry Lions model. The PDE and the setup,

 ft − fxx = [δp(t)+a−δp(t)−a ]fx (p(t), t) in (−1, 1) × [0, ∞)



fx (1, t) = fx (−1, t) = 0



 f (x, 0) = f (x)
I

where p(t) = {x : f (x, t) = 0} presumed. for a.e. t, to be a singleton and

 
1
a := a(p(t)) = min a, |p(t) ± 1| with a < 1
2
May, 09, 2021

10 On a price formation free boundary model by Lasry and Lions [4]

This is the first in the series of two papers that deals with the Lasry-Lions price formation model. The paper looks at the global
existence and asymptotic behavior of a price formation. The paper transforms the problem into a heat equation with specially
prepared initial condition. The key points is that the free boundary present in the original problem becomes the zero level set of
the new solution. The paper then produces results on the global existence, regularity and asymptotic results of free-boundary.

10.1 The Lasry Lions Setup

The setup is as follows,

1. Large group of buyers and sellers trading a certain good at a certain price p(t).
2. The transaction cost is fixed at a.
3. The model is given by a non-linear parabolic boundary evolution equation that describes the dynamical behavior of the
densities of buyers and sellers which in turn define the price.

4. The domain of the set up is the whole real line, therefore in principle take arbitrarily large values.
5. The model is given by the equation,

ft − fx x = λ(t)(δ(x − p(t) + a) − δ(x − p(t) − a)), x ∈ R, t ∈ R+ (10.1)

λ(t) = −fx (p(t), t), f (p(t), t) = 0, (10.2)
f (x, 0) = fI , p(0) = p0 , for some p0 ∈ R, (10.3)

with compatibility conditions at time t = 0:

(A1) fI (p0 ) = 0 and fI (x) > 0 for x < p0 and fI (x) < 0 for x > p0 .

For the following we assume that fI is in L1 (R) and bounded. This model has been studied in number of papers prior to this
[8] [10] and [15]. The current paper presents the first global existence result of a smooth solution on the whole real line. In the
following we shall denote f = f + − f − the decomposition of a function into its positive and negative part.

10.2 Connection to the heat equation

It will be shown that the system (10.1) (10.2)(10.3) and solutions of the heat equation supplemented with specially prepared
initial data. This results in an global-existence result that is fairly elegant.

Theorem 9. (The Diffusion Equation) Let f = f (x, t) be a solution to (10.1) - (10.3) on the time interval [0, T ] withe T > 0.
Then there exists a linear transformation from f to a function F = F (x, t), being a solution of he heat equation, such that the
graph of the zero level set of F is p(t). By reversing the transformation, each solution of the heat equation such that the zero
level seta of the solution is a smooth graph for 0 ≤ t ≤ T can be transformed into a solution of the FBP with the same p(t).
a What is the zero level set? It is the set of point where the solution of the PDE has a value of zero. Note that f is the difference function of Buyers

an Sellers densities. The claim here is that the equilibrium price is where this function has the value zero.

Proof. The construction is based on the observation that the second derivative of −f − at free boundary p(t) is precisely the
negative value of he weighted delta mass centered at p(t) − a, as it appears in equation (10.1). Analogously, the second derivative
of f + is the negative of he weighted delta mass of equation (10.1) centered at p(t) − a. Essentially, this is the same as saying
− +
fxx = λ(t)(δp−a − δp+a ). Essentially, −fxx = −δ(p(t) + a) and fxx = δ(p(t) − a).
Let f = fI (x) be a given initial condition satisfying the assumption (A1). Let f = f (x, t) be the solution of (10.1)-(10.3) in
the time interval [0, T ] (such a solution exists due to [15] theorem 2.6). Now we define,

 P∞ f + (x + na, t), x < p(t)
0
F (x, t) =
 − ∞ f − (x − na, t), x > p(t)
P
0
May, 09, 2021

and the initial condition,


 P∞ f + (x + na, 0), x < p0 ,
0 I
FI (x) = (10.4)
 − P∞ f − (x − na, 0) x > p0
0 I

for an arbitrary function fI satisfying (A1).

Here the transaction cost is fixed. One of the properties of f is that it is a fast decreasing function. The function f being
bounded the sums defined above converge in D′ (R × [0, ∞]). It is easy to check that F satisfies, in the sense of distributions the
heat equation with the initial condition FI (x) := F (x, t = 0), given by (10.4). Clearly, the free boundary p = p(t) is now the
zero level set of F 12 . We next consider a given F = F (x, t), solution of the heat equation in [0, T ] with the initial value (10.4).
A solution can now be constructed for the FBP (10.1) - (10.3). Assume that the initial condition fI is as given in (10.4), then
we construct f (x, t) in the following way,

 F + (x, t) − F + (x + a), x < p(t)
f (x, t) =
 F − (x − a) − F − (x, t), x > p(t)

again by construction, the zero level set of F becomes the free boundary of (10.1) - (10.3).

Theorem 10 (Global Existence). There exists a unique smooth solution f = f (x, t) of (10.1) - (10.3) for t ∈ [0, ∞). Furthermore,
p ∈ C([0, ∞)).

Proof. Let FI be the transformed initial condition corresponding to fI and let F be the solution of the heat equation within
initial condition FI . Denoting (abusing notation) let p = p(t) the zero level set of F . First we note that oscillations of p(t)
yielding a ’fat’ free-boundary cannot occur as they contradict the x-analyticity of solutions of the heat equation. Furthermore,
due to [15] Lemma 2.9, we know that fx (p(t), t) < 0 for all t > 0 (by Hopf Lemma) and the min-max principle implies that
p = p(t) is the graph of a function.Hence we only need to exclude existence of t∗ such that |p(t)| becomes unbounded as t → t∗ .
We write,
Z ∞ Z x−p0 Z ∞
−
F (x, t) = G(t, z)FI (x − z)dz = G(t, z)FI (x − z)dz − G(t, z)FI+ (x − z)dz (10.5)
−∞ −inf ty x−p0

2
1
where G(t, x) = √4πt e−x /4t is the 1-d heat kernel. Since f is bounded by its construction, F grows at most linearly at |x| = ∞.
Thus second term on the RHS in (10.5) tends to zero as x → ∞. For the first term we have,

Z p0 Z ∞ Z p0 +a
G(t, z)FI− (x − z)dz = G(t, z)F (x + z)dz ≥ C |FI (x + z)dz
−∞ p0 −x p0

Due to (10.4) we have,

Z p0 +a ∞ Z
X x+p0 −(n−1)a
FI (x + z)dz = fI− (y)dy ≥ const > 0.
p0 0 x+p0 −na

For large enough x, this term dominates in (10.5) an thus F (., t) becomes negative. By the same argument we show that for
large negative x, F (., t) becomes positive and thus there must exist a unique x with −∞ < x < ∞ such that F (x, t) = 0. From
these arguments we conclude that p(t) is defined and continuous for all t.

Remark 1. A similar analysis produces solutions of the Neumann problem in the interal [−L, L] and certain examples of
non-existence. In this case the associated solution of the heat equation satisfies the unusual Neuman type boundary condition
Fx (±L, t) = Fx (±L ∓ a, t).
12 This is by construction. At the point x = p(t)
May, 09, 2021

10.3 Asymptotic Behavior

R∞ 2
The assumption is that the initial condition fI = 0 and a = 1. For the following we define that erf (u) = √1
4π u
e−x /4
dx.
R0 R∞
Theorem 11 (Asymptotic Behavior). Let f = f (x, t) be a solution of (10.1)-(10.3). If M + = −∞
f + (z)dx ̸= 0
f − (z)dz =:
M − , i.e. the total mass of f is zero the,
R∞
z|f (z)|dz
 
−∞ 1
p(t) = +O √
M+ + M− t

Proof. From (10.4) and (10.5), (essentially the infinite sum and the solution ideas) we obtain at x = p(t),
Z ∞ ∞ Z 0 ∞
1 X − |p(t)−z+n|2 − 1 X − |p(t)−z+n|2 +
0=− √ e 4t f (z)dz + √ e 4t f (z)dz
0 4πt n=0 −∞ 4πt n=0
√ P∞ 2 √ √
The point he makes is that 1/ 4πt n=0 e−|p(t)−z+n| /(4t) = erf ((p(t)−z−1)/ t) The final conclusion is as follows, p(t) = q(t) t
and if M + ̸= M −
R∞  
0
z|f (z)|dz 1
p(t) = + −
+O √ .
M +M t
May, 09, 2021

11 On A Price Formation Free Boundary Model by Lasry Lions: Neumann

Problem [5]
This paper mostly deals with existence and uniqueness for the price formation free boundary model with homogeneous Neumann
Boundary Conditions 13 . This looks at the model (5.3). The problem is restated as follows,




 ft − fxx = λ(t)(δ(x − p(t) + a) − δ(x − p(t) − a)), x ∈ (−L, L), t > 0


 λ(t) = −f (p(t), t), f (p(t), t) = 0

x
(11.1)


 f (x, 0) = fI , p(0) = p0 , for some po ∈ (−L + a, L − a)



 f (±L, t) = 0, t > 0
x

The range of a is 0 < a < L, compatibility conditions at time t = 0 : fI (p0 ) = 0 and fI (x) > 0 for x < p0 and fI < 0 for x > p0 .
The paper assumes the following,
1. fI is in L2 (−L, L) and L2 norm is denoted as ∥.∥.
This paper is an extension of the work in [4] [8] [15] . In this paper Global (non-)existence results are presented for (11.1) on
the bounded interval (−L, L).

11.1 Analysis of the Neumann problem - transformation to the heat equation

The positive and negative part of a function f defined for x ∈ (−L, L) by, f + := max(f, 0), f − := max(−f, 0) 14
. The functions
f + , f − are extended outside the interval (−L, L) using this notation.
Proposition 1 (Conditions for Existence). Let f be a solution of the modified price formation problem (11.1) with the initial
and boundary conditions a .

 
 0, p(t) > −L + a  0, p(t) < L − a
fx (−L, t) = fx (L, t) = (11.2)
 f (−L + a, t) p(t) ≤ −L + a,  f (L − a, t) p(t) ≥ L − a,
x x

Then the function,


 P∞ f + (x + na, t), x < p(t)
0
F (x, t) = (11.3)
 − P∞ f − (x − na, t), x > p(t)
0

is a solution to the following BVP for the heat equation,

Ft = Fxx , for all x ∈ (−L, L), t > 0,

Fx (±L, t) = Fx (±L ∓ a, t), t > 0 (11.4)
F (x, t = 0) = FI (x) ∀ x ∈ (−L, L)

where FI is constructed from fI according to (11.3). Conversely if F is a solution of (11.4) then

f (x, t) = F (x, t) − F + (x + a, t) + F − (x − a, t) (11.5)

satisfies (11.1) and (11.2).

a The derivative conditions, the Neumann problem.

The proof for the proposition is just hand waved in terms of how it is constructed. The idea is that F was constructed using the
fact that f + and f − has jump discontinuities at x = p(t) − a (x = p(t) + a) and x = p(t) of equal magnitude but opposite signs.
13 Specifies the derivative of the function.
14 This is terrible notation. Here f + (x) = max(f (x), 0).
May, 09, 2021

11.2 Solution
This is a standard diffusion equation. So we proceed with separation of variables,
F = ψ(t)ϕ(x)
ψt ϕxx
= = −z 2
ψ ϕ
2
ψ(t) = αe−z t

ϕ(z) = αeizx + βA−izx = B cos zx + A sin zx

Applying the Neumann boundary conditions we obtain,
Fx (L, t) = Fx (L − a, t) ⇒ Az(cos zl − cos z(L − a)) + Bz(sin zl − sin z(L − a)) = 0
Fx (−L, t) = Fx (−L + a, t) ⇒ Az(cos −zl − cos z(a − L)) + Bz(sin (−zL) − sin z(a − L))
Solving across the two sets of equations by elimination we get,
G(z) := cos zL − cos z(L − a) = 0
H(z) := sin zL − sin z(L − a) = 0
2πl π(2l−1)
The solutions for the above are G(z) = 0 iff z = { 2πl 2πl
a , 2L−a) }, and H(z) = 0 ⇒ z ∈ { a , 2L−a }. The solution (11.4) is given
by,
X 2
[Al sin(c1,l x) + Bl cos(c1,l x)+] e−c1,l t + Cl sin(c2,l x)ec2,l t + Dl cos(c3,l x)ec3,l t + A0 x + B0
 
F (x, t) = (11.6)
l

Where the constants are determined by the initial condition.

Theorem 12 (Global Existence). The BVP (11.1) (1-3), (11.2) has a unique global solution f = f (x, t) for t > 0. Furthermore
the free boundary p = p(t) is a smooth graph p(t) ∈ (−L, L) for all t > 0.
The proof is a result of the following according to the paper,
1. The construction.
2. The min-max principle.
3. The Hopf principle.
R L+a RL
4. Since the condition 0 < −L F (x, t)dx and 0 > L−a F (x, t)dx are conserved in time (by both the equation and the BC)
it concludes that p(t) ∈ (−L, L) for all t > 0.
5. The solution to (11.1)(1-3), (11.2), is a solution to (11.1) on a time interval [0, T ] iff p(t) ∈ (−L + a, L − a) for t ∈ [0, T ].
R p(t) R∞
Then the total mass of buyers MB = −∞ f (x, t)dx and vendors/sellers MS = p(t) f (x, t)dx are time conserved quantities
15
.
Lemma 1 (Solution Convergence). The solution F converges exponentially fast to F∞ = A∞ x + B∞ in L2 (−L, L).
Proof. Our constructed solutions satisfying G(z) = 0 and H(z) = 0 form the set of functions {e2iπlx/a , eiπ(2l−1x/(2L−a)) }. These
form a Riesz bais in L2 (−L, L). It is claimed that G(z) is a sine-type function of type L with separated zeros except for z = 0
being a zero of order 2. The conclusion is that {ei2πlx/a , ei2πlx/(2L−a),x } is a Riesz basis in L2 (−L, L). The consideration of odd
and even functions separately it is conclude that, {cos(2πlx/a), sin(2πlx/a), sin(2πlx/(2L − a)), cos((2l − 1)πx/(2L − a), 1, x)}
form a Riesz basis of eigen-functionsP of the heat equation (11.4). From the Reisz base property we deduce that there exists
∞
c1 , c2 mahtbbR+ such that c1 ∥FI ∥ ≤ l=1 (A2l + Bl2 + Cl2 + Dl2 ) + A20 + B02 ≤ c2 ∥F ∥2 , for F̃ = F − (A0 x + B0 ) we deduce
∞
X ∞
X
c1 ∥F ∥2 ≤ Al + Bl2 + Cl2 + Dl2 e−2γl t ≤ e−γ1 t Al + Bl2 + Cl2 + Dl2 ≤ c2 e−γ1 ∥F ∥2
 2   2 

l=1 l=1

where γl = min{4π 2 l2 /a2 , 4π 2 l2 /(2L − a)2 , ((2l − 1)2 π 2 )/(2L − a)2 }, l = 1, 2, · · · a .

a We assume that the eigen-functions form a basis. However that need not be true. One needs to establish the fact that the eigen-functions are

indeed a basis. Especially since the coefficients are dependent on boundary conditions.
15 If this is not conserved or if a is a function p(t) then M
B and MS may not be conserved. Or if MS and MB are functions of p(t) then a increases
and MS and MB will go to zero under toxic order flow. Essentially here p(t) is exogenous determined by flow of market orders.
May, 09, 2021

Lemma 2. (The Maximum Principle) The solution (11.6) satisfies |Fx (x, t) ≤ sup{x∈(−L,L)} a |(FI )x | and |F (x, t)| ≤ c, ∀x ∈
(−L, L), t > 0.
Proof. The function V = Fx satisfies the heat equation with V (−L, t) = V (−L + a, t), V (L, t) = V (L − a, t) and initial condition
V (x, t = 0) = (FI )x (x) b . If V assumes its maximum on a cylinder [−L, L] × [0, T ] at either boundary x = ±L, then it must
also assume a maximum (with the same value) in the interior (due to BC on V). This contradicts the maximum principle, thus
V must assume its maximum at t = 0. The same arguments can be made for the minimum.
Rx R −L+a R −L+a R x
Since F (x, t) = F (−L, t) + −L Fx (y, t)dy we deduce aF = − −L F (x, t)dx + −L −L
F (x, t)dx = I1 + I2 . We know Fx
d
R −L+a R −L+a R −L+a
is bounded, there fore |I2 | ≤ K. In addition dt −L
F (x, t)dx = 0 and therefore −L
F (x, t)dx = −L F (x, t)dx = 0 and
R L+a R −L+a
therefore −L F (x, t)dx = −L FI (x)dx. Thus F (−L, t) as well as F = F (x, t) are bounded uniformly on (−L, L) × R+ .
a Note that L is the boundary and the PDE need only be valid in (−L, L).
b Essentially the derivative satisfies the heat equation with the boundary conditions that are derivatives of the boundary conditions of the function.

Theorem 13 (BVP Global Solution). The BVP (11.1) has a global solution conserving the total mass of buyers and vendors iff
the zero level set p of the solution of (11.4) satisfies p(t) ∈ (−L + a, L − a) for all t > 0. The the free boundary p(t) converges to
p∞ ∈ (−L + a, L − a).
Proof. We know that F converges exponentially fast to F{ ∞} = A∞ x+B∞ . This follows from Lemma (1). There exists a smooth
graph p = p(t) such that F (p(t), t) = 0a . We next assume that p(t) ∈ (−L + a, L − a)b for t > 0, choose any sequence tn → ∞ and
conclude that there is a sub-sequence tnk such that p(tnk ) → p∞ ∈ [−L + a, L − a]. Let ϕ be a test function in D(p∞ , L) c . If k is
RL
sufficiently large we conclude that f (x, tnk ) < 0d for x ∈ suppϕ. Therefore −L f (x, tnk )ϕ(x)dx < 0. Since f (., t) → f∞ there is a
subsequence tnkl converges to f∞ (x) pointwise a.e. in (−L, L). The function |f (x, tnk )| ≤ K on [−L, L] for all k. Now, Lebesgue
RL RL
DCT results in −L f (x, tnk )ϕ(x)dx → −L f∞ ϕ(x) ≤ 0. Since f∞ = A∞ x + B∞ + (A∞ (x − a) + B∞ )− − (A∞ (x + a) + B∞ )+e
we conclude that f∞ ≤ 0, for x > p∞ and f∞ ≥ 0 for x < ′∞ f . From Lebesque’s dominated convergence theorem, we conclude
RL RL L R p∞
− p(tn ) f (x, t)dx → − p∞ f∞ f∞ (x)dx ≥ 0. On the other side we similarly have −L f∞ dx ≥ 0g . Next we show that F∞ has
kl
− +
a unique zero in (−L, L). If F∞ = A∞ x + B∞ ≥ 0 on (−L, L) then F∞ = 0, F∞ = F∞ h . This implies that f∞ is a constant in
(−L, L), which is a contradiction. The same argument holds for F∞ < 0. Therefore the function f∞ is given by f∞ (x) = ±α
for x ∈ (−L, p∞ − a) and x ∈ (p∞ + a, L) respectively and f∞ = −α/a(x − p∞ )i for x ∈ [p∞ − a, p∞ + a], with α ∈ R+ . We
conclude that p∞ is unique and that p(t) → p∞ as t → ∞, since every sequence has a sub-sequence which converges to the same
limitj .
a This says that there is a price at which there are no buyers or sellers.
b This assumes that the equilibrium price is inside the best bid/ask for all time t > 0.
c What is the space D(p , L)? Space of distributions?
∞
d Basically this say that there is a set where ϕ(x) > 0 around p
∞ such that there are buyers or sellers.
e From definition of f in terms of F (11.5).
f Essentially there is a function of buyers and sellers around this equilibrium price.
g In this proof the p(t) is determined from F and it is shown that it solves f from (11.1).
h If F
∞ is strictly positive then, this is obvious.
i Like our solution in the case of Lasry Lyons setup (5.7).
j Interestingly you have to wait infinitely long for a price to form.

Theorem 14 (Conservation Laws and Non-Existence). Let fI be such that MB /MV ∈ / [a/(4L−3a), (4L−3a)/a], where MB , MV
denotes the initial mass of buyers and vendors. Then (11.1) does not have a global-in-time solution, which conserves both buyers
and vendors masses.
R p∞
Proof. The result follows since α, p∞ cannot be adjusted so that p∞ ∈ [−L + a, L − a] where MB = −L f∞ dx, MV =
RL
− p∞ f∞ dx.

The choice of L, which corresponds to the maximally attainable price, is more or less arbitrary but a bad (too small) choice of
L might impede global existence. In this case clearly looses its ’practical’ significance. The global existence results for (11.1)
(with a free boundary which remains in (−L + a, L − a)) for initial data which are small perturbations of stationary solutions
are straightforward, without using the analytical machinery of [8].
May, 09, 2021

References
[1] P. Bak, M. Paczuski, and M. Shubik. Price variations in a stock market with many agents. Physica A: Statistical Mechanics
and its Applications, 246(3-4):430–453, Dec 1997.

[2] Reginald A Banez, Lixin Li, Chungang Yang, and Zhu Han. Mean field game and its applications in wireless networks.
Springer, 2021.
[3] Martin Burger, Luis Caffarelli, Peter A. i, and Marie-Therese Wolfram. On a boltzmann-type price formation model.
Proceedings: Mathematical, Physical and Engineering Sciences, 469(2157):1–20, 2013.
[4] Luis A. Caffarelli, Peter A. Markowich, and Jan-Frederik Pietschmann. On a price formation free boundary model by lasry
and lions, 2011.
[5] Luis A. Caffarelli, Peter A. Markowich, and Marie-Therese Wolfram. On a price formation free boundary model by lasry
and lions: The neumann problem, 2011.
[6] José A Carrillo and Giuseppe Toscani. Asymptotic l 1-decay of solutions of the porous medium equation to self-similarity.
Indiana University Mathematics Journal, pages 113–142, 2000.
[7] Patrick Chan and Ronnie Sircar. Bertrand and cournot mean field games. Applied Mathematics & Optimization, 71(3):533–
569, 2015.
[8] Lincoln Chayes, María del Mar González, Maria Pia Gualdani, and Inwon Kim. Global existence and uniqueness of solutions
to a model of price formation. SIAM Journal on Mathematical Analysis, 41(5):2107–2135, 2009.

[9] María del Mar González and Maria Pia Gualdani. Asymptotics for a free boundary model in price formation. Nonlinear
Analysis: Theory, Methods & Applications, 74(10):3269–3294, 2011.
[10] María del Mar González and Maria Pia Gualdani. Asymptotics for a free boundary model in price formation. Nonlinear
Analysis: Theory, Methods & Applications, 74(10):3269–3294, 2011.

[11] David Evangelista, Yuri Saporito, and Yuri Thamsten. Price formation in financial markets: a game-theoretic perspective.
arXiv preprint arXiv:2202.11416, 2022.
[12] Thierry Foucault, Ohad Kadan, and Eugene Kandel. Limit order book as a market for liquidity. The Review of Financial
Studies, 18(4):1171–1217, 2005.

[13] María del Mar González and Maria Pia Gualdani. Asymptotics for a symmetric equation in price formation. Applied
Mathematics and Optimization, 59(2):233–246, 2009.
[14] Jean-Michel Lasry and Pierre-Louis Lions. Mean field games. Japanese journal of mathematics, 2(1):229–260, 2007.
[15] PA Markowich, N Matevosyan, J-F Pietschmann, and M-T Wolfram. On a parabolic free boundary equation modeling price
formation. Mathematical Models and Methods in Applied Sciences, 19(10):1929–1957, 2009.

[16] Peter A. Markowich, Josef Teichmann, and Marie-Therese Wolfram. Parabolic free boundary price formation models under
market size fluctuations, 2016.
[17] D. G. Schultz and James L. Melsa. State Functions and Linear Control Systems. 1967.
[18] F. Slanina. Essentials of Econophysics Modelling. OUP Oxford, 2013.