Wuchen Li Mean eld games summer school, June 26,...
Transcript of Wuchen Li Mean eld games summer school, June 26,...
![Page 1: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/1.jpg)
Mean field games via probability manifold I
Wuchen Li
Mean field games summer school, June 26, 2018
![Page 2: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/2.jpg)
Introduction
Lecture I
I Static Population games;
I Wasserstein Gradient flow on graphs;
I Entropy dissipation.
Lecture II
I Differential Population games;
I Wasserstein Hamiltonian flow on graphs;
I Schrodinger equation on graphs;
I Schrodinger bridge problems on graphs.
2
![Page 3: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/3.jpg)
Games
Game contains: Players; Strategies; Payoffs.
I Players: 2;
I Strategies: S1 = S2 = {Rock, Paper, Scissors};I Payoffs: F1, F2 : S1 × S2 → {+1, 0,−1}.
3
![Page 4: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/4.jpg)
Stag hunt
I Players: Infinity;
I Strategy set: S = {C,D}; Players form (ρC , ρD) with ρC + ρD = 1;
I Payoffs: F (ρ) = (FC(ρ), FD(ρ))T = Aρ, where A =
(3 02 2
),
meaning a deer worthing 6, a rabbit worthing 2.
4
![Page 5: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/5.jpg)
Population game
Population games model the strategic interactions in large populations ofsmall, anonymous agents.
I Strategy setS = {1, · · · , n} ;
I Players (Simplex)
P = {(ρi)ni=1 ∈ Rn :
n∑i=1
ρi = 1 , ρi ≥ 0} ;
I Payoff function to strategy i: Fi : P → R. E.g.
F (ρ) = (Fi(ρ))ni=1 = Aρ , where A ∈ Rn×n .
ApplicationsSocial Network, Biology species, Virus, Trading, Cancer, Congestion andmany more (See Sandholm’s textbook). We plan to design new dynamicsto model for the evolution of a game, and study their asymptoticproperties.
5
![Page 6: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/6.jpg)
Nash Equilibrium
Nash Equilibrium (NE): Players have no unilateral incentive to deviatefrom their current strategies.ρ∗ = (ρ∗i )
ni=1 is a Nash equilibrium (NE) if
ρ∗i > 0 implies that Fi(ρ∗) ≥ Fj(ρ∗) for all j ∈ S.
E.g. if S = {C,D}, FC(ρ) = 3ρC , FD(ρ) = 2, it is simple to check
(1, 0), ( 23 ,
13 ), (0, 1) are three NEs.
A particular type of game, named Potential games, are widely considered:There exists a potential F : P → R, such that
∂
∂ρiF(ρ) = Fi(ρ) .
If F (ρ) = Aρ, consider F(ρ) = 12ρTAρ, where A is a symmetric matrix.
In potential games, NE is the critical points of
maxρF(ρ) : ρ ∈ P .
6
![Page 7: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/7.jpg)
Evolutionary dynamics
In literature, people have designed many dynamics, named mean orevolutionary dynamics, to model games. Typical examples are BNN(Brown-von Neumann-Nash 1950), Best response dynamics(Gilboa-Matsui 1991), Logit (Fudenberg-Levine 1998), Smith dynamics(Smith 1983) etc.
A famous dynamics is Replicator dynamics (Taylor and Jonker 1978)
dρidt
= ρi(Fi(ρ)− F (ρ)), where F (ρ) =∑j∈S
ρjFj(ρ).
In potential games, the Replicator dynamics is a gradient flow inprobability set P w.r.t a Fisher-Rao (Shahshahani) metric (Akin (1980)).
7
![Page 8: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/8.jpg)
Goal
Design the other dynamics for evolutionary games with
I Evolution only using local information in strategies;
I Gradient flow in potential games;
I Ability to include white noise perturbations.
Related materials:Dynamical system; Optimal transport; Riemannian Geometry; Partialdifferential equations; Topology; Graph theory; Information theory.
8
![Page 9: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/9.jpg)
Best Reply dynamics
Choose S = Td. Consider the model (P. Degond, J. G. Liu, C. Ringhofer,2014.)
dXt = ∇XtF (Xt, ρ)dt+√
2βdWt , Xt ∈ Td.
where Wt is the standard Brownian motion (or noise level) and ρ(t, x) isthe density function of Xt. In this case, the mean field equation refers tothe evolution of density:
∂ρ
∂t+∇x ·
(ρ∇xF (x, ρ)
)= β∆xρ .
I Individual players change their pure strategies according to thedirection that maximizes their own payoff functions most rapidly.And the Brownian motion represents uncertainties.
I In potential games, this PDE is the gradient flow w.r.t the optimaltransport metric.
9
![Page 10: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/10.jpg)
Density manifold
Optimal transport has a variational formulation (Benamou-Brenier 2000):
infv
∫ 1
0
Ev(t,Xt)2 dt ,
where E is the expectation operator and the infimum runs over all vectorfields vt, such that
Xt = v(t,Xt) , X0 ∼ ρ0 , X1 ∼ ρ1 .
Under this metric, the probability set has a Riemannian geometrystructure1.
1John D. Lafferty: the density manifold and configuration space quantization, 1988.10
![Page 11: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/11.jpg)
Brownian motion and Entropy dissipation
The gradient flow of the entropy
H(ρ) =
∫Tdρ(x)log ρ(x)dx ,
w.r.t. optimal transport metric is:
∂ρ
∂t= ∇ · (ρ∇log ρ) = ∆ρ .
Entropy dissipation:
d
dtH(ρ) =
∫Td
log ρ∇ · (ρ∇log ρ)dx = −∫Td
(∇ log ρ)2ρdx .
11
![Page 12: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/12.jpg)
Evolutionary games via optimal transport
Question:Can we derive a Best-Reply dynamics on a discrete strategy set?
Answer:Yes, we need a discrete dynamical optimal transport metric. Using thismetric, we derive the gradient flow as the evolutionary dynamics forpotential games.
Recent Developments:Mielke, Maas, Chow, Zhou, Li, Huang, Erbar, Fathi, Gangbo, Mou andmany more.
12
![Page 13: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/13.jpg)
Basic setting
Graph with finite vertices
G = (S,E, ω), S = {1, · · · , n}, E is the edge set, ω is the weight ;
Probability set
P = {(ρi)ni=1 |n∑i=1
ρi = 1, ρi ≥ 0} ;
Noise potential:
F(ρ) =1
2
n∑i=1
n∑j=1
Aijρiρj − β
n∑i=1
ρi log ρi ,
Interaction Potential energy Boltzmann-Shannon entropy
where A is a given symmetric matrix and β > 0 is a given constant.
13
![Page 14: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/14.jpg)
Definition I
We plan to find the discrete analog of density manifold (Maas, Mielke,Chow et.al).
First, it is natural to define a vector field on a graph
v = (vij)(i,j)∈E , satisfying vij = −vji.
Given a potential Φ = (Φi)ni=1, a gradient vector field refers
∇GΦij =√ωij(Φi − Φj).
14
![Page 15: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/15.jpg)
Definition II
We next define an inner product of two vector fields v1, v2:
(v1, v2)ρ :=1
2
∑(i,j)∈E
v1ijv
2ijθij(ρ);
and a divergence of a vector field v at ρ ∈ P:
divG(ρv) :=(−∑
j∈N(i)
√ωijvijθij(ρ)
)ni=1
.
Here θ represents the probability weight on the edge. E.g. θij is given bya upwind scheme:
θij(ρ) =
ρi/di if ∂
∂ρiF(ρ) > ∂
∂ρjF(ρ), j ∈ N(i);
ρj/dj if ∂∂ρiF(ρ) < ∂
∂ρjF(ρ), j ∈ N(i);
ρi/di+ρj/dj2 if ∂
∂ρiF(ρ) = ∂
∂ρjF(ρ), j ∈ N(i).
where di =∑j∈N(i) ωij∑(i,j)∈E ωij
is the volume form on the node. θij has the
other choices.15
![Page 16: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/16.jpg)
Optimal transport on a graph
DefinitionFor any ρ0, ρ1 ∈ P, define the Wasserstein metric W : P × P → R by
W (ρ0, ρ1)2
=: infvt{∫ 1
0
(vt, vt)ρtdt :dρtdt
+ divG(ρtvt) = 0, ρ(0) = ρ0, ρ(1) = ρ1}.
16
![Page 17: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/17.jpg)
Hodge decomposition on graphs
Consider a Hodge decomposition of vector field v = (vij)(i,j)∈E on agraph G
vij = ∇GΦij + uij
Gradient Divergence free
where Φ ∈ Rn and u ∈ R|E|, with
∇GΦij =√ωij(Φi − Φj),
and uij = −uji satisfying divG(ρu) = 0, i.e.∑j∈N(i)
uijθij = 0.
In this case,(v, v)ρ = (∇GΦ,∇GΦ)ρ + (u, u)ρ.
17
![Page 18: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/18.jpg)
Effect of Hodge decomposition
LemmaThe discrete Wasserstein metric is equivalent to
W (ρ0, ρ1)2 = inf∇GΦ
∫ 1
0
(∇GΦ,∇GΦ)ρ dt ,
where the infimum is taken among all discrete potential vector fields∇GΦ, such that
dρ
dt+ divG(ρ∇GΦ) = 0 , ρ(0) = ρ0, ρ(1) = ρ1 .
(P,W ) has a Riemannian geometry structure.
18
![Page 19: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/19.jpg)
Probability manifold
Denotedρ
dt= −divG(ρ∇GΦ) = L(ρ)Φ .
With this discrete Wasserstein metric, P forms a Riemannian manifold
infρ(t){∫ 1
0
ρTL(ρ)−1ρdt : ρ(0) = ρ0, ρ(1) = ρ1} .
Here L(ρ) ∈ R|V |×|V | is the weighted Laplacian matrix
L(ρ) = −divG(ρ∇G) = −DTΘ(ρ)D ,
where D ∈ R|E|×|V | is a discrete gradient matrix, DT ∈ R|V |×|E| is adiscrete divergence matrix, and Θ(ρ) ∈ R|E|×|E| is a diagonal weightmatrix
Θ(ρ)(i,j)∈E,(k,l)∈E =
{θij if (i, j) = (k, l) ∈ E ;
0 otherwise .
19
![Page 20: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/20.jpg)
Gradient flow in Riemannian manifold
The gradient flow in abstract form
dρ
dt= gradWF(ρ) ,
where the gradient is defined by
I Tangency:
gradWF(ρ) ∈ TρP ={
(σi)ni=1 :
n∑i=1
σi = 0}.
I Duality:
gradWF(ρ)TL(ρ)−1σ = dF(ρ) · σ, for any σ ∈ TρP ,
where dF(ρ) = ( ∂∂ρiF(ρ))ni=1.
20
![Page 21: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/21.jpg)
Derivation
TheoremGiven a potential game with a strategy graph G = (S,E, ω), a payoffmatrix A. Then the gradient (ascent) flow of the free energy F(ρ) on Pwith respect to W is
dρ
dt= L(ρ)dρF(ρ).
I.e.
dρidt
=∑
j∈N(i)
ωijρj/dj [Fi(ρ)− Fj(ρ) + β logρjρi
]+
−∑
j∈N(i)
ωijρi/di[Fj(ρ)− Fi(ρ) + β logρiρj
]+ .(1)
21
![Page 22: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/22.jpg)
Asymptotical behavior
TheoremFor any initial condition ρ0 ∈ P+(S), (1) has a unique solutionρ(t) : [0,∞)→ P+(S).
(i) The free energy F(ρ) is a Lyapunov function of (1);
(ii) If limt→∞ ρ(t) exists, call it ρ∞, then ρ∞ is one of the possibleGibbs measures, i.e.
ρ∞i =1
KeFi(ρ
∞)
β , K =
n∑i=1
eFi(ρ
∞)
β for all i ∈ S.
22
![Page 23: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/23.jpg)
Entropy dissipation on population games
What is the rate of convergence to a Gibbs measure?
Motivation
I Entropy dissipation: Carrillo, McCann and Villani’s work2 fornonlinear Fokker-Planck equations on Td;
I Gradient flows: dynamical systems viewpoint!
2Carrillo, McCann and Villani, “Kinetic equilibration rates for granular media andrelated equations: entropy dissipation and mass transportation estimates”, 2003.
23
![Page 24: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/24.jpg)
Entropy dissipation
Theorem (Entropy dissipation)If the Gibbs measure ρ∞ is a strict maximizer of F(ρ), then there existsa constant C > 0, such that
F(ρ∞)−F(ρ(t)) ≤ e−Ct(F(ρ∞)−F(ρ0)) .
24
![Page 25: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/25.jpg)
Idea of Proof
The speed of convergence comes from comparing the ratio between thefirst and second derivative of F(ρ(t)) along with the ODE. If one canfind a constant C > 0, such that
d2
dt2F(ρ(t)) ≥ −C d
dtF(ρ(t)) ,
holds for all t ≥ 0. Then by integrating the above formula in [t,+∞],one obtains
d
dt[F(ρ∞)−F(ρ(t))] ≥ −C[F(ρ∞)−F(ρ(t))] .
Proceed with the Gronwall’s inequality, the result is proved.
25
![Page 26: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/26.jpg)
Proof
In our case, the first derivative of energy along the gradient flow is
d
dtF(ρ(t)) = F (ρ)T ρ = F (ρ)TL(ρ)F (ρ) = ρTL−1(ρ)ρ ,
while the second derivative forms
d2
dt2F(ρ(t)) =2 ρT dρρF(ρ)ρ+ ρTL−1(ρ)L(ρ)L−1(ρ)ρ .
Compare ddtF(ρ(t)) with d2
dt2F(ρ(t)) to find
C := infρ∈B(ρ0)
2ρT dρρF(ρ)ρ
ρTL−1(ρ)ρ+ρTL−1(ρ)L(ρ)L−1(ρ)ρ
ρTL−1(ρ)ρ.
Quadratic Cubic
26
![Page 27: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/27.jpg)
Hessian operator at Gibbs measure
Let
λF (ρ) = minΦ∈Rn
∑(i,j)∈E
∑(k,l)∈E
hij,kl(Φi − Φj)θij(Φk − Φl)θkl
s.t. ∑(i,j)∈E
(Φi − Φj)2θij = 1.
Here
hij,kl = (∂2
∂ρi∂ρk+
∂2
∂ρj∂ρl− ∂2
∂ρi∂ρl− ∂2
∂ρj∂ρk)F(ρ) .
This rate connects with Yano formula3, which is related to Riccicurvature in geometry.
3Kentaro Yano, “On Harmonic and Killing Vector Fields”, 38-45, Annals ofMathematics, 1958.
27
![Page 28: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/28.jpg)
Stag Hunt
I Strategy set {C,D};I Players ρ = (ρC , ρD)T ;
I Payoff F (ρ) = Aρ with A =
(3 02 2
).
28
![Page 29: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/29.jpg)
Stag Hunt
We draw the vector field of the Fokker-Planck equation. Different noiselevels lead to different NEs.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(c) β = 5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(d) β = 0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(e) β = 0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(f) β = 0
29
![Page 30: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/30.jpg)
Rock-Scissors-Paper
I Strategy set {r, s, p}; Players ρ = (ρr, ρs, ρp)T ;
I Payoff F (ρ) = Aρ with payoff matrix A =
0 −1 11 0 −1−1 1 0
.
0
0.2
0.4
0.6
0.8
11
0.8
0.6
0.4
0.2
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0
(g) β = 0
0
0.2
0.4
0.6
0.8
11
0.8
0.6
0.4
0.2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
(h) β = 0.1
30
![Page 31: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/31.jpg)
Bad Rock-Scissors-Paper
I Payoff F (ρ) = Aρ with payoff matrix A =
0 −2 11 0 −2−2 1 0
.
We demonstrate a Hopf Bifurcation. If β is large, there is a uniqueequilibrium around ( 1
3 ,13 ,
13 ). If β is small, a limit cycle exists.
0
0.2
0.4
0.6
0.8
11
0.8
0.6
0.4
0.2
0
0.1
0.2
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0
(i) β = 0.5
0
0.2
0.4
0.6
0.8
11
0.8
0.6
0.4
0.2
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
(j) β = 0.1
0
0.2
0.4
0.6
0.8
11
0.8
0.6
0.4
0.2
0.1
0.2
0.5
0
0.3
0.4
1
0.9
0.8
0.7
0.6
0
(k) β = 0
31
![Page 32: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/32.jpg)
Optimal transport+Dynamical system
Boltzman-Shannon entropy
H(ρ) =
n∑i=1
ρi log ρi ⇒ HessRnH(ρ) = diag(1
ρi).
Thus our asymptotically dissipation rate forms
λH(ρ) = min{n∑i=1
1
ρi(divG(ρ∇GΦ)|i)2 :
∑(i,j)∈E
(Φi − Φj)2θij = 1} > 0.
32
![Page 33: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/33.jpg)
Linear Entropy+ Yano formula
Consider
H(ρ) =
∫Mρ(x) log ρ(x)dx,
whose Gibbs measure is a uniform measure. Then
(HessP(M)H · ∇Φ,∇Φ)ρ∗
=
∫M
[Ric(∇Φ,∇Φ) + tr(D2ΦTD2Φ)]ρ∗(x)dx
=
∫M
[∇ · (ρ∗∇Φ)]21
ρ∗(x)dx.
The first equality is well known derived through Bochner’s formula,while the second equality is new. It shows Yano’s formula.
33
![Page 34: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/34.jpg)
Reference
Cedric VillaniOptimal transport: Old and new, 2008.
B. FriedenScience from Fisher Information: A Unification, 2004.
Shui-Nee Chow, Wuchen Li and Haomin ZhouEntropy dissipation on finite graphs, DCDS, series A, 2018.
Shui-Nee Chow, Wuchen Li, Jun Lu and Haomin ZhouPopulation games and discrete optimal transport, 2017.
Shui-Nee Chow, Wuchen Li and Haomin ZhouA discrete Schrodinger equation via optimal transport, 2017.
Wilfrid Gangbo, Wuchen Li and Chenchen Mou.Geodesic of Minimal Length in the Set of Probability Measures onGraph, 2018.
Wuchen Li.Geometry of probability simplex via optimal transport, 2018.
34
![Page 35: Wuchen Li Mean eld games summer school, June 26, 2018helper.ipam.ucla.edu/publications/gss2018/gss2018_15289.pdf · Evolutionary dynamics In literature, people have designed many](https://reader035.fdocuments.in/reader035/viewer/2022071018/5fd217b8239f685636728c3a/html5/thumbnails/35.jpg)
Thanks.
35