Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Mean Field Games: numerical methods
Y. Achdou and I. Capuzzo-Dolcetta
November 19, 2008
Introduction
• Considera very large numberN of rational agents(or players) with
a limited global information on the game. Nash equilibria.
• In symmetric situations, Lasry and Lions have made a theory allowing
for passing to the limit whenN → ∞.
They have introduced, justified and analyzed models of mean field
games. Analogy with statistical physics and mechanics.
• New systems of PDEs with uniqueness, stability...
• For the justification, P-L. Lions has proposed a new mathematical
framework including
– differential calculus for functions of probability measures
– viscosity solutions for PDEs on probability measures.
1
Applications
• Finance
– Self-consistent models for volatility
– Price formation
• Economics
– Economical equilibrium with anticipation.
– Economic models for the behaviour of customers in front of a
technological breakthrough; ex: heat insulation of housesand
appartments.
2
Content
• Mean field games: some aspects of the theory of Lasry and Lions:
– The stationary case (infinite horizon)
– The non stationary case (finite horizon)
• Monotone numerical schemes
• Theoretical results on the numerical schemes
– Proof of existence in the non stationary case
– Uniqueness
– Proof of existence in the stationary case
• Numerical experiments.
• Open questions
3
Mean field games: some aspects of the theory of Lasry and Lions
ConsiderN players whose dynamics are
dXit =
√2νidW i
t − αidt, Xi0 = xi ∈ R
d, for 1 ≤ i ≤ N,
• (W 1t , . . . ,W
Nt ) independent Brownian motions inRd,
• νi > 0,
• The control of the playeri, i.e.αi is a bounded processassumed to beadapted to(W i
t ).
For simplicity, we assume thatXit ∈ T (the unit torus ofRd), and all the
functions used below will be periodic with period1 in every direction.
4
Cost function of the player i
J i(α1, . . . , αN ) = lim infT→∞
1
TE
(∫ T
0
(Li(Xi
t , αit) + F i(X1
t , . . . , XNt ))dt
),
given the initial conditionsX = (x1, . . . , xN ).
• F i is Lipschitz inTN ,
• Li is Lipschitz inxi uniformly in αi bounded,
•lim
|αi|→∞infxi
Li(xi, αi)
|αi| = +∞.
5
Nash equilibrium, given the initial conditions (x1, . . . , xN )
(α1, . . . , αN ) is a Nash point if
αi = Argminαi
J i(α1, . . . , αi−1, αi, αi+1, . . . , αN ), ∀i = 1, . . . , N.
Introduce the Hamiltonians
Hi(x, p) = supα∈Rd
(p · α− Li(x, α)
), i = 1, . . . , N, x ∈ T, p ∈ R
d.
Assume thatHi is C1.
6
Existence of a Nash equilibrium (Lasry-Lions)If there existsθ ∈ (0, 1) such that for|p| large,
infx∈T
„∂Hi
∂x· p +
θ
d(Hi)2 > 0
«> 0, ∀i, 1 ≤ i ≤ N,
then
1. ∃λ1, . . . , λN ∈ R, v1, . . . , vN ∈ C2(T), m1, . . . ,mN ∈ W 1,p(T) ∀p < ∞,
such that, for alli, 1 ≤ i ≤ N ,
8>>>>>>>><>>>>>>>>:
− νi∆vi + H
i(x,∇vi) + λi =
Z
TN−1
Fi(x1
, . . . , xi−1
, x, xi+1
, . . . , xN )
Y
j 6=i
mj(xj)dx
j,
− νi∆mi − div
„mi
∂Hi
∂p(x,∇vi)
«= 0,
Z
T
vidx = 0,
Z
T
midx = 1 and mi > 0 in T.
7
2. For any solutionλ1, . . . , λn, v1, . . . , vN ,m1, . . . ,mN of this system,
•αi(·) =
∂Hi
∂p(·,∇vi(·))
defines a feedback which is a Nash point for all(x1, . . . , xN ) ∈ TN .
•λi = J i(α1, . . . , αN )
= limT→∞
1
TE
(∫ T
0
(Li(Xi
t , αit(X
it)) + F i(X1
t , . . . , XNt ))dt
),
whereXit is the process corresponding to the feedbackαi.
8
Sketch of the proof:1) The existence for the system of PDEs follows from
• A priori estimate onλi: by the maximum principle,
|λi| ≤ ‖F i‖∞ + supx
|Hi(x, 0)|.
• A priori estimate on|∇vi|: using the technical hypothesis andBernstein method (i.e. write the equation satisfied by|∇vi|2 and use amaximum principle).
2) The existence of a Nash equilibria follows from a verification argumentusing Ito’s formula and the ergodicity formula
0 = limT→∞
0BBBB@
E
„1
T
Z T
0
Fi(X1
t , . . . , Xi−1t , X
it , X
i+1t , . . . , X
Nt )dt
«
−E
0@ 1
T
Z T
0
dt
Z
TN−1
Fi(x1
, . . . , xi−1
, Xit , x
i+1, . . . , x
N )Y
j 6=i
mj(xj)dx
j
1A
1CCCCA
.
9
Particular casesIf Li(x, α) = µi|α|2/2, then the change the unknown
φi = e−
vi
2νiµi
(∫
T
e−
vi
νiµi
)−1/2
and mi = φ2i
leads to the generalized Hartree equation: findφi > 0 s.t.
−2(νi)2µi∆φi+φi
∫
TN−1
F i(x1, . . . , xi−1, x, xi+1, . . . , xN )∏
j 6=i
φ2j(x
j)dxj = λiφi.
In particular, if
F i = 1/2∑
j 6=i
V (xi − xj) + V0(xi), νi = 1/2, µi = 1,
then one gets the Hartree equation: findφi > 0 s.t.
−1
2∆φi + φi
∑
j 6=i
V ∗ φ2j + V0φi = λiφi.
10
In a totally symmetric situation, one can find symmetrical solutions
Theorem: Symmetric caseIf the players are identical, i.e.νi = ν,
Hi = H and if
F i(x1, . . . , xN ) = F j(x1, . . . , xi−1, xj , xi+1, . . . , xj−1, xi, xj+1, . . . , xN ),
then there exists a solution of the previous system of PDE s.t.
λ1 = · · · = λN , v1 = · · · = vN , m1 = · · · = mN .
Uniqueness of the Nash equilibrium is false in general, evenin totallysymmetric situations.
Even symmetrical Nash equilibria may not be unique.
11
The limit N → ∞
We consider situations where there is a very large number of players and
each player chooses his optimal strategy in view of a limitedglobal
information on the game.
Assumptions
• The players are identical, thusνi = ν andHi = H.
• F i depends onxi andon the empirical density of the other players,namely
1
N − 1
∑
j 6=i
δxj .
12
• More precisely, assume that there exists an operatorV from the spaceof probability measures onT into a bounded set of Lipschitzfunctions onT such that
F i(x1, . . . , xN ) = V
1
N − 1
∑
j 6=i
δxj
(xi).
i.e.
Ji(α1
, . . . , αN ) = lim inf
T→∞
1
TE
0@
Z T
0
0@L(Xi
t , αit) + V
24 1
N − 1
X
j 6=i
δX
jt
35 (Xi
t)
1A dt
1A .
•
V [mn] converges uniformly onT to V [m] if mn weakly converges tom.
Typical examples forV include nonlocal smoothing operators.
13
TheoremUnder the above conditions, any Nash equilibriaλN1 , . . . , λN
N ,
vN1 , . . . , vN
N , mN1 , . . . ,mN
N satisfies:
• (λNi )i,N is bounded inR, (vN
i )i,N is relatively compact inC2(T), (mNi )i,N is
relatively compact inW 1,p(T), for any1 ≤ p < ∞.
•
supi,j
“|λN
i − λNj | + ‖vN
i − vNj ‖∞ + ‖mN
i − mNj ‖∞
”→ 0.
• Any converging subsequence(λN′
1 , vN′
1 , mN′
1 ) in R × C2(T) × W 1,p(T)
converges to(λ, v, m) such that8>>>>><>>>>>:
− ν∆v + H(x,∇v) + λ = V [m],
− ν∆m − div
„m
∂H
∂p(x,∇v)
«= 0,
Z
T
vdx = 0,
Z
T
mdx = 1, and m > 0 in T.
(∗)
14
Uniqueness for the mean field problem
By contrast with the preceding systems of PDEs, the system (*) is well
posed, under some conditions:
Theorem If V is strictly monotone, i.e.∫
T
(V [m1] − V [m2])(m1 −m2)dx ≤ 0 ⇒ m1 = m2,
then the solution of the mean field system (*) is unique.
15
Proof
Consider two solutions of (*):(λ1, v1,m1) and(λ2, v2,m2):
• multiply the first equation bym1 −m2
∫
T
(ν∇(v1 − v2) · ∇(m1 −m2) + (H(x,∇v1) −H(x,∇v2))(m1 −m2)
)dx
=
∫
T
(V [m1] − V [m2])(m1 −m2)dx
• multiply the second equation byv1 − v2
0 =
∫
T
ν∇(v1 − v2) · ∇(m1 −m2)dx
+
∫
T
(m1
∂H
∂p(x,∇v1) −m2
∂H
∂p(x,∇v2)
)· ∇(v1 − v2)dx.
16
• subtract:
0 =
8>>>>><>>>>>:
Z
T
m1
„H(x,∇v1) − H(x,∇v2) −
∂H
∂p(x,∇v1) · ∇(v1 − v2)
«dx
+
Z
T
m2
„H(x,∇v2) − H(x,∇v1) −
∂H
∂p(x,∇v2) · ∇(v2 − v1)
«dx
+
Z
T
(V [m1] − V [m2])(m1 − m2)dx
SinceH is convex andV is monotone, the 3 terms vanish.
The strict monotonicity ofV implies thatm1 = m2.
The identitiesv1 = v2 andλ1 = λ2 come from the uniqueness for the
ergodic problem:
−ν∆v +H(x,∇v) + λ = f.
17
Finite horizon Nash equilibrium with N playersCost:
E
∫ T
0
Li(Xi
s, αis) + V
1
N − 1
∑
j 6=i
δXjs
ds+ V0
1
N − 1
∑
j 6=i
δXj
T
Lasry and Lions show (formally only in most cases) that the value functions
atT − t: vNi (x1, . . . , xN , t) satisfy
∂vNi
∂t− ν
∑
j
∆xjvNi
+H(xi,∇xivNi ) +
∑
j 6=i
∂H
∂p(xj ,∇xjvN
i ) · ∇xjvNi
= V
1
N − 1
∑
j 6=i
δxj
(xi).
18
and
vi|t=0 = V0
1
N − 1
∑
j 6=i
δxj
.
The densitymN (x1, . . . , xN , t) of players onTN atT − t, obeys the
Fokker-Planck equation onTN × (0, T ):
∂mN
∂t+ ν
∑
i
∆ximN +∑
i
divxi
(mN ∂H
∂p(xi,∇xivN
i )
)= 0,
mN (x1, . . . , xN , T ) =N∏
i=1
m0(xi),
assuming that theN players initial conditions are random, independent,
with the same probability distributionm0.
19
Formal passage to the limit whenN → ∞
• vNi is expected to depend very little onxj , j 6= i.
• Since the players are independent,
mN (x1
, . . . , xN
, t) ≈Y
i
m(xi, t).
• Consider
vNi (xi, t) =
∫
TN−1
vNi (x1, . . . , xN , t)
∏
j 6=i
m(xj , t)dxj .
This should be asymptotically independent ofi:
vNi (x, t) ≈ v(x, t).
20
Multiplying the pde forvNi by
∏j 6=im(xj , t) and integrating onTN−1,
L.L. obtain asymptotically
∂v
∂t− ν∆v +H(x,∇v) = V [m],
∂m
∂t+ ν∆m+ div
(m∂H
∂p(x,∇v)
)= 0,
∫
T
mdx = 1, and m > 0 in T,
v(t = 0) = V0[m(t = 0)], m(t = T ) = m◦,
(∗∗)
since, from the law of large numbers,
Z
TN−1
V
24 1
N − 1
X
j 6=i
δxj
35 (xi)
Y
j 6=i
m(xj, t)dx
j → V [m](xi).
21
Existence for (**)
TheoremIf
• same kind of assumptions onV andV0 as in the stationary case (V and
V0 are nonlocal smoothing operators).
• H is smooth onT × Rd and
∣∣∣∣∂H
∂x(x, p)
∣∣∣∣ ≤ C(1 + |p|), ∀x ∈ T, ∀p ∈ Rd,
then (**) has at least a smooth solution.
22
Uniqueness for (**)
TheoremIf the operatorsV andV0 are strictly monotone, i.e.
∫
T
(V [m] − V [m])(m− m) ≤ 0 ⇒ V [m] = V [m],
∫
T
(V0[m] − V0[m])(m− m) ≤ 0 ⇒ V0[m] = V0[m],
then (**) has a unique solution.
23
Numerical Methods: taked = 2 and approximate numerically8>>>>>>>>><>>>>>>>>>:
∂u
∂t− ν∆u + H(x,∇u) = V [m],
∂m
∂t+ ν∆m + div
„m
∂H
∂p(x,∇u)
«= 0,
Z
T
mdx = 1, m > 0 in T,
u(t = 0) = V0[m(t = 0)], m(t = T ) = m◦,
(∗∗)
Let Th be a uniform grid on the torus with mesh steph, andxij be a genericpoint inTh.
LetNT be a positive integer and∆t = T/NT , tn = n∆t.
The values ofu andm at (xi,j , tn) are respectively approximated byUni,j
andMni,j .
24
Goal: use a fully implicit scheme, prove existence and possiblyuniqueness
Notation:
• The discrete Laplace operator:
(∆hW )i,j = − 1
h2(4Wi,j −Wi+1,j −Wi−1,j −Wi,j+1 −Wi,j−1).
• Right-sided finite difference formulas for∂1w(xi,j) and∂2w(xi,j):
(D+1 W )i,j =
Wi+1,j −Wi,j
h, and (D+
2 W )i,j =Wi,j+1 −Wi,j
h.
• The set of 4 finite difference formulas atxi,j :
[DhW ]i,j =((D+
1 W )i,j , (D+1 W )i−1,j, (D
+2 W )i,j , (D
+2 W )i,j−1
).
25
The discrete version of
∂u
∂t− ν∆u+H(x,∇u) = V [m]
is chosen as
Un+1i,j − Un
i,j
∆t− ν(∆hU
n+1)i,j + g(xi,j , [DhUn+1]i,j) =
(Vh[Mn+1]
)i,j,
where
•g(xi,j , [DhU
n+1]i,j)
=g`xi,j , (D
+
1 Un+1)i,j , (D
+
1 Un+1)i−1,j , (D
+
2 Un+1)i,j , (D
+
2 Un+1)i,j−1
´,
• for instance,(Vh[M ])i,j = V [mh](xi,j),
callingmh the piecewise constant function onT taking the valueMi,j
in the square|x− xi,j |∞ ≤ h/2.
26
Classical assumptions on the discrete Hamiltoniang
(q1, q2, q3, q4) → g (x, q1, q2, q3, q4) .
• Monotonicity: g is nonincreasing with respect toq1 andq3 and
nondecreasing with respect to toq2 andq4.
• Consistency:
g (x, q1, q1, q3, q3) = H(x, q), ∀x ∈ T, ∀q = (q1, q3) ∈ R2.
• Differentiability: g is of classC1, and there exists a constantC such
that∣∣∣ ∂g∂x (x, (q1, q2, q3, q4))
∣∣∣ ≤ C(1 + |q1| + |q2| + |q3| + |q4|) for all
x ∈ T andq1, q2, q3, q4.
• Convexity: the function(q1, q2, q3, q4) → g (x, q1, q2, q3, q4) is convex.
27
The discrete version of
∂m
∂t+ ν∆m+ div
(m∂H
∂p(x,∇v)
)= 0.
It is chosen so each time step corresponds to a linear system with a matrix
• whose diagonal coefficients are negative,
• whose off-diagonal coefficients are nonnegative,
in order to hopefully use somediscrete maximum principle.
• The argument for uniqueness should hold in the discrete case,
sothe discrete Hamiltoniang should be used for the second pde as well.
28
Principle
Discretize
−∫
T
div
(m∂H
∂p(x,∇u)
)w =
∫
T
m∂H
∂p(x,∇u) · ∇w
by
h2∑
i,j
mi,j∇qg(xi,j , [DhU ]i,j) · [DhW ]i,j
29
This yields the following scheme:
0 =Mn+1
i,j − Mni,j
∆t+ ν(∆hM
n)i,j
+1
h
8>>>>>>>><>>>>>>>>:
Mni,j
∂g
∂q1
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
−Mni−1,j
∂g
∂q1
`xi−1,j , (D
+
1 Un)i−1,j , (D
+
1 Un)i−2,j , (D
+
2 Un)i−1,j , (D
+
2 Un)i−1,j−1
´
+Mni+1,j
∂g
∂q2
`xi+1,j , (D
+
1 Un)i+1,j , (D
+
1 Un)i,j , (D
+
2 Un)i+1,j , (D
+
2 Un)i+1,j−1
´
−Mni,j
∂g
∂q2
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
+1
h
8>>>>>>>><>>>>>>>>:
Mni,j
∂g
∂q3
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
−Mni,j−1
∂g
∂q3
`xi,j−1, (D
+
1 Un)i,j−1, (D
+
1 Un)i−1,j−1, (D
+
2 Un)i,j−1, (D
+
2 Un)i,j−2
´
+Mni,j+1
∂g
∂q4
`xi,j+1, (D
+
1 Un)i,j+1, (D
+
1 Un)i,j+1, (D
+
2 Un)i,j+1, (D
+
2 Un)i,j
´
−Mni,j
∂g
∂q4
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
30
Classical discrete Hamiltoniansg can be chosen.For example, assume that the Hamiltonian is of the form
H(x,∇u) = ψ(x, |∇u|).
A possible choice is the Godunov scheme
g(x, q1, q2, q3, q4) =
ψ(x,√
min(q1, 0)2 + max(q2, 0)2 + min(q3, 0)2 + max(q4, 0)2).
If ψ is convex and nondecreasing with respect to its second argument, theng is a convex function of(q1, q2, q3, q4); it is nonincreasing with respect toq1 andq3 and nondecreasing with respect toq2 andq4.
Finally, it can be proven that the global scheme is consistent if H andg aresmooth enough.
31
Kushner-Dupuis scheme
More generally, ifH is given by
H(x, p) = supα∈R2
(p · α− L(x, α)) , x ∈ T, p ∈ R2,
then one may chooseg as
g(x, q1, q2, q3, q4) = supα∈R2
(−α−
1 q1 + α+1 q2 − α−
2 q3 + α+2 q4 − L(x, α)
).
The numerical Hamiltoniang is clearly convex, nonincreasing with respect
to q1 andq3 and nondecreasing with respect toq2 andq4.
32
Existence for the discrete problem
Theorem Assume thatMNT ≥ 0 and that∑
i,j MNT
i,j = 1. Under theassumptions above onV , V0 andg, the discrete problem has a solution.
Strategy of proof
K =
(Mi,j)0≤i,j<N : h2
∑
i,j
Mi,j = 1,Mi,j ≥ 0
.
Apply Brouwer fixed point theorem to a mapping
KNT → KNT ,
(Mn)n → (Un)n → (Mn)n.
33
Proof: a fixed point method in the compact and convex setKNT ,
Step 1: a mapping(Mn)n → (Un)n.Given(U0
i,j) and(MNT
i,j ), define the mapΦ:
Φ :(Mn)0≤n<NT∈ KNT → (Un)1≤n≤NT
, s.t.
Un+1i,j − Un
i,j
∆t− ν(∆hU
n+1)i,j + g(xi,j , [DhUn+1]i,j) =
(Vh[Mn+1]
)i,j.
• Existence is classical: (Leray-Schauder fixed point theorem at each time
step, making use of the monotonicity ofg, the uniform boundedness
assumption onV and ofH(·, 0)).• Uniqueness stems from the monotonicity ofg.
34
Step 2: estimates• There exists a constantC independent of(Mn)n s.t.
‖Un‖∞ ≤ C(1 + T ).
• From the assumption on the continuity ofV and well known results on
continuous dependence on the data for monotone schemes, themapΦ is
continuous.
•∣∣∣ ∂g∂x (x, (q1, q2, q3, q4)
∣∣∣ ≤ C(1 + |q1| + |q2| + |q3| + |q4|) implies
‖DhUn‖∞ ≤ LT, ∀n,
whereL does not depend on(Mn)0≤n<NT∈ KNT andh.
35
Step 3: A map(Mn)n → (Mn)n
For (Un)n = Φ((Mn)n) andµ > 0 large enough,
consider thelinear backward parabolic problem:
find (Mn)0≤n≤NTsuch that
MNT = MNT
and
36
−µMni,j =
fMn+1
i,j − fMni,j
∆t+ ν(∆h
fMn)i,j − µfMni,j
+1
h
8>>>>>>>><>>>>>>>>:
fMni,j
∂g
∂q1
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
−fMni−1,j
∂g
∂q1
`xi−1,j , (D
+
1 Un)i−1,j , (D
+
1 Un)i−2,j , (D
+
2 Un)i−1,j , (D
+
2 Un)i−1,j−1
´
+fMni+1,j
∂g
∂q2
`xi+1,j , (D
+
1 Un)i+1,j , (D
+
1 Un)i,j , (D
+
2 Un)i+1,j , (D
+
2 Un)i+1,j−1
´
−fMni,j
∂g
∂q2
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
+1
h
8>>>>>>>><>>>>>>>>:
fMni,j
∂g
∂q3
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
−fMni,j−1
∂g
∂q3
`xi,j−1, (D
+
1 Un)i,j−1, (D
+
1 Un)i−1,j−1, (D
+
2 Un)i,j−1, (D
+
2 Un)i,j−2
´
+fMni,j+1
∂g
∂q4
`xi,j+1, (D
+
1 Un)i,j+1, (D
+
1 Un)i,j+1, (D
+
2 Un)i,j+1, (D
+
2 Un)i,j
´
−fMni,j
∂g
∂q4
`xi,j , (D
+
1 Un)i,j , (D
+
1 Un)i−1,j , (D
+
2 Un)i,j , (D
+
2 Un)i,j−1
´
37
This is adiscrete backward linear parabolic problem.
From the previous estimates on(Un)n, one can findµ large enough andindependent of(Mn)n such that the iteration matrix is the opposite of aM-matrix, thusthere is a discrete maximum principle.
Therefore, there exists a unique solution(Mn)n. Moreover,
Mn ≥ 0 ⇒ Mn ≥ 0, ∀n,h2∑
i,j Mn = 1 ⇒ h2
∑i,j M
n = 1, ∀n.
ThusMn ∈ K for all n. Define the map
χ : KNT 7→ KNT ,
(Mn)0≤n<NT→ (Mn)0≤n<NT
38
Step 4: existence of a fixed point ofχ
From the boundedness and continuity of the mappingΦ, and from the fact
thatg is C1, we obtain thatχ : KNT 7→ KNT is continuous.
From Brouwer fixed point theorem,χ has a fixed point, which yields a
solution of (**).
39
Discrete weak maximum principle-1The discrete version of
∂m
∂t+ ν∆m+ div
(m∂H
∂p(x,∇v)
)= 0
can be written
−Mn − ∆tAnMn = −Mn+1.
Introduce the semi-norm|||W |||:
|||W |||2 =∑
i,j
((D+
1 W )2i,j + (D+2 W )2i,j
).
From the estimates onUn, there existsγ ≥ 0 (independent ofh and∆t) s.t.
(AnW,W )2 ≥ ν
2|||W |||2 − γ‖W‖2
2, ∀W, ∀n.
This is a discrete Garding inequality.
40
Discrete weak maximum principle-2
One can also prove that
−(AnW,W−)2 ≥ ν
2|||W−|||2 − γ‖W−‖2
2, ∀W.
Taking(Mn)− as a test-function in the discrete equation yields
(1 − 2γ∆t)∥∥(Mn)−
∥∥2
2+ ν∆t|||(Mn)−|||2 ≤
∥∥(Mn+1)−∥∥2
2,
so there is a discrete weak maximum principle if2γ∆t < 1.
41
Uniqueness
Theorem We make the same assumptions as above onV , V0,H andg, and
we assume furthermore that the operatorsVh andV0,h are strictly
monotone, i.e.(Vh[M ] − Vh[M ],M − M
)2≤ 0 ⇒ Vh[M ] = Vh[M ],
(V0,h[M ] − V0,h[M ],M − M
)2≤ 0 ⇒ V0,h[M ] = V0,h[M ].
There exists a nonnegative constantγ such that if2γ∆t < 1, then the
discrete problem has a unique solution.
Proof The choice of the scheme makes it possible to mimic the proof used
in the continuous case: uses the convexity assumption ong and the discrete
weak maximum principle.
42
A finite difference scheme for the stationary problem8>>>>>>>>>>>>>>>>><>>>>>>>>>>>>>>>>>:
−ν(∆hU)i,j + g(xi,j , [DhU ]i,j) + λ = (Vh[M ])i,j ,0BBBBBBBBBBB@
−ν(∆hM)i,j
−1
h
0B@
Mi,j∂g
∂q1
(xi,j , [DhU ]i,j) − Mi−1,j∂g
∂q1
(xi−1,j , [DhU ]i−1,j)
+Mi+1,j∂g
∂q2
(xi+1,j , [DhU ]i+1,j) − Mi,j∂g
∂q2
(xi,j , [DhU ]i,j)
1CA
−1
h
0B@
Mi,j∂g
∂q3
(xi,j , [DhU ]i,j) − Mi,j−1
∂g
∂q3
(xi,j−1, [DhU ]i,j−1)
+Mi,j+1
∂g
∂q4
(xi,j+1, [DhU ]i,j+1) − Mi,j∂g
∂q4
(xi,j , [DhU ]i,j)
1CA
1CCCCCCCCCCCA
= 0,
Mi,j ≥ 0,
and
h2∑
i,j
Mi,j = 1, and∑
i,j
Ui,j = 0.
43
Existence for the discrete problem: strategy of proof
• Use Brouwer fixed point theorem in the set of discrete probabilitymeasures for a mappingM → U →M .
• The mapM → U consists of solving
−ν(∆hU)i,j + g(xi,j , [DhU ]i,j) + λ = (Vh[M ])i,j ,∑i,j Ui,j = 0
• Ergodic approximation: pass to the limit asρ→ 0:
−ν(∆hU(ρ))i,j + g(xi,j , [DhU
(ρ)]i,j) + ρU (ρ))i,j = (Vh[M ])i,j ,
We need estimates onU (ρ) − U(ρ)0,0 uniform inρ andh.
44
DifficultyThe proof of existence for the continuous problem used the estimate
‖∇u‖∞ ≤ C, which was proven using the Bernstein method and the
assumption: there existsθ ∈ (0, 1) such that for|p| large,
infx∈T
(∂H
∂x· p+
θ
2H2 > 0
)> 0.
In the discrete case, this argument seems difficult to reproduce, and we were
not able so far to obtain an estimate on‖Dhu‖∞ independent ofh, under
the same assumptions.
45
Assumptions on the Hamiltonian
H(x, p) = maxα∈A
(p · α− L(x, α)) ,
where
• A is a compact subset ofR2,
• L is aC0 function onT ×A,
• Kushner-Dupuis scheme:
g(x, q1, q2, q3, q4) = supα∈A
(−α−
1 q1 + α+1 q2 − α−
2 q3 + α+2 q4 − L(x, α)
).
46
Proposition (using Kuo-Trudinger(1992) and Camilli-Marchi(2008))
Consider a grid functionV and make the assumptions:
• L is a continuous function,A is compact,
• g aC1 function onT × R4 such that∇qg is bounded.
• ‖V ‖∞ is bounded uniformly w.r.th,
For any real numberρ > 0, there exists a unique grid functionUρ such that
ρUρi,j − ν(∆hU
ρ)i,j + g(xi,j , [DhUρ]i,j) = Vi,j ,
and there exist two constantsδ, δ ∈ (0, 1) andC, C > 0, uniform inh andρ
s.t.
|Uρ(ξ) − Uρ(ξ′)| ≤ C|ξ − ξ′|δ, ∀ξ, ξ′ ∈ Th.
47
Proposition (using Krylov(2007) and Camilli-Marchi(2008))
Same assumptions as before, and furthermore
• L is uniformly Lipschitz w.r.t.x,
• ‖DhV ‖∞ is bounded uniformly w.r.th,
For any real numberρ > 0, there exists a unique grid functionUρ such that
ρUρi,j − ν(∆hU
ρ)i,j + g(xi,j , [DhUρ]i,j) = Vi,j ,
and there exist a constantC, C > 0, uniform inh andρ s.t.
|Uρ(ξ) − Uρ(ξ′)| ≤ C|ξ − ξ′|, ∀ξ, ξ′ ∈ Th.
48
Proposition Under the first set of assumptions, there exists a unique grid
functionU and a real numberλ such that
−ν(∆hU)i,j + g(xi,j , [DhU ]i,j) + λ = (Vh[M ])i,j ,∑i,j Ui,j = 0,
and there exist two constantsδ, δ ∈ (0, 1) andC, C > 0, uniform inh s.t.
|U(ξ) − U(ξ′)| ≤ C|ξ − ξ′|δ, ∀ξ, ξ′ ∈ Th.
Under the second set of assumptions,
|U(ξ) − U(ξ′)| ≤ C|ξ − ξ′|, ∀ξ, ξ′ ∈ Th.
49
Existence and uniqueness for the stationary problem
Theorem Under the above assumptions onV andg, the discrete stationary
problem has at least a solution and we have either a uniform Holder or a
Lipschitz estimate onU , depending on the assumptions.
Uniqueness: Ok if(Vh[M ] − Vh[M ],M − M
)2≤ 0 ⇒M = M.
50
The case whenV is a local operatorAssume thatV is a local operator, i.e.V [m](x) = F (m(x), x), whereF is
aC1 function onR × T.
The above mentioned existence theorem in its weakest form contains the
case whenVh is a continuous map from discrete probability measures to
bounded grid functions. In particular,Vh can be a local operator.
Existence and Holder estimates.
51
Numerical method: Long time approximation of the stationaryproblem
∂u
∂t− ν∆u+H(x,∇u) = V [m],
∂m
∂t− ν∆m− div
(m∂H
∂p(x,∇u)
)= 0,
u(0, x) = u0(x), m(0, x) = m0(x),
with∫
Tm0 = 1 and m0 ≥ 0, and we expect that
limt→∞
(u(t, x) − λt) = u(x), limt→∞
m(t, x) = m(x),
Same thing at the discrete level.We use a semi-implicit linearized scheme. It requires the numerical solutionof a linearized problem. Linearizing must be done carefullyand is notalways possible. In such cases, an explicit method can be used.
52
ν = 1,
H(x, p) = sin(2πx2) + sin(2πx1) + cos(4πx1) + |p|2,F (x,m) = m2
2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.23.89e-16 -0.2 -0.4 -0.6 -0.8 -1 -1.2 -1.4 -1.6 -1.8 -2 -2.2 -2.4 -2.6 -2.8
sin(2πx2) + sin(2πx1) + cos(4πx1).
53
0.98
1
1.02
1.04
1.06
1.08
1.1
1.12
1.14
1.16
1.18
200 400 600 800 1000 1200 1400 1600 1800 2000
"mean_8_200"
Convergence1t∫
Tu(x, t)dx→ λ ast→ ∞
54
0.05 0.045 0.04 0.035 0.03 0.025 0.02 0.015 0.01 0.0051.73e-18 -0.005 -0.01 -0.015 -0.02 -0.025 -0.03 -0.035
1.08 1.07 1.06 1.05 1.04 1.03 1.02 1.01 1 0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 0.91 0.9
left: u, rightm.
55
0
2e-05
4e-05
6e-05
8e-05
0.0001
0.00012
0.00014
0.00016
0.00018
0.0002
0.01 0.015 0.02 0.025 0.03 0.035 0.04
"erroru"0.005*x
5e-05
0.0001
0.00015
0.0002
0.00025
0.0003
0.01 0.015 0.02 0.025 0.03 0.035 0.04
"errorm"0.005*x
Convergence ash→ 0
56
ν = 0.01,
H(x, p) = sin(2πx2) + sin(2πx1) + cos(4πx1) + |p|2,F (x,m) = m2
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2
2.1
2.2
200 400 600 800 1000 1200 1400 1600 1800 2000
"mean_8_200"
Convergence1t∫
Tu(x, t)dx→ λ ast→ ∞
57
0.15 0.14 0.13 0.12 0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 -0.01
1.8 1.7 1.6 1.5 1.4 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
left: u, rightm.
Note that the supports of∇u and ofm tend to be disjoint asν → 0.
58
Deterministic limit
Theorem (Lasry-Lions)
If
• H(x, p) ≥ H(x, 0) = 0,
• V [m] = F (m) + f0(x) whereF ′ > 0,
then
limν→0
(λν ,mν) = (λ,m),
where
m(x) =(F−1(λ− f0(x))
)+and
∫
T
mdx = 1.
59
ν = 0.001,
H(x, p) = |p|2,
V [m](x) = 4 cos(4πx) + m(x)
"u.gp" "m.gp"
left: u, rightm.The supports of∇u and ofm tend to be disjoint.
m(x) ≈ (λ− 4 cos(4πx))+
60
A nonlocal operator V
0.3 0.25 0.2 0.15 0.1 0.05-1.39e-17 -0.05 -0.1 -0.15 -0.2
7 6.5 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5
ν = 0.1,
H(x, p) = sin(2πx2) + sin(2πx1) + cos(4πx1) + |p|3/2,
F (x, m) = 200(1 − ∆)−1(1 − ∆)−1m
left: u, right m.
61
0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05-6.94e-17 -0.05 -0.1 -0.15 -0.2 -0.25 -0.3
13 12 11 10 9 8 7 6 5 4 3 2 1
ν = 0.1,
H(x, p) = sin(2πx2) + sin(2πx1) + cos(4πx1) + (0.6 + 0.59 cos(2πx))|p|3/2,
F (x, m) = 200(1 − ∆)−1(1 − ∆)−1m
left: u, right m.
62
Perspectives
• Obtain an estimate on‖DhU‖∞ uniform inh with more general
structure assumptions in the stationary case (important for stability and
convergence).
• Prove that the long-time strategy converges to the solutionof the
stationary problem.
• When convergence is OK, prove error estimates?
• Tackle practical situations.
63