A canonical martingale coupling
A canonical martingale coupling
Workshop on Optimal Transportation and AppplicationsNicolas JUILLET
Université de Strasbourg
Pisa, November 2012
A canonical martingale coupling
Outline
1 The martingale transport plans
2 Tools for the martingale transport problem
3 Results
Nicolas JUILLET A canonical martingale coupling
A canonical martingale coupling The martingale transport plans
Definition: martingale transport plan
A probability measurePonR×Ris termed a martingale transport plan if P=Law(X,Y)where(X,Y)is a two-times martingale process.
Equivalently if(Px)x∈Ris a disintegration (allias conditional laws, allias Markov kernel) ofP, it has to satisfy
Barycenter(Px) = Z
ydPx(y) =x
for(projx#P)-almost everyx.
A canonical martingale coupling The martingale transport plans
Some examples
P=Law(x,Y)wherex=E(Y).
P=∑2i=1∑3j=1ai,jδ(xi,yj)wherex∈ {−1,1}andy∈ {−2,0,2}and (ai,j) =
1/4 1/4 0 1/12 1/12 1/3
+t
1/12 −1/6 1/12
−1/12 1/6 −1/12
for somet∈[0,1].
P=Law(X,X+I)where the incrementIis independent fromX. (for instanceX andIare Gaussian)
P=P1+2P2 whereP1,P2are martingale transport plans.
P=Law(E(Y|
F
),Y)for someF
⊆σ(Y).Nicolas JUILLET A canonical martingale coupling
A canonical martingale coupling The martingale transport plans
The general problem
The problem Minimize
P7→
Z
c(x,y)dP(x,y) among the martingale transport plans fromµtoν.
For different cost functionscwe would like to know:
How do the minimizers look like?
What are their properties?
Is there a unique minimizer?
A canonical martingale coupling The martingale transport plans
Model theorem
Model theorem in the classical setting
Forµandνin
P
2in the convex order andPa transport plan fromµtoν. The following statements are equivalent:The planPis optimal for the transport problem withc(x,y) = (y−x)2, The planPis concentrated on a monotone setΓ,
The planPis the quantile coupling.
We have proved a theorem similar to this one in the martingale setting.
Nicolas JUILLET A canonical martingale coupling
A canonical martingale coupling Tools for the martingale transport problem
The convex order
Definition: the convex order We write
µCν
and say thatµis smaller thanνin the convex order if and only if there exists a martingale transport planPwith
projx#P=µ and projy#P=ν.
According to a (non constructive) theorem of Strassen, it is equivalent to assume Z
ϕdµ≤ Z
ϕdν
for every convex functionϕ:R→R.
A canonical martingale coupling Tools for the martingale transport problem
The extended order and the shadows
Proposition - Definition
We writeµEνand say thatµis smaller thanνin the extended order if Fµν:={θ:µCθandθ≤ν}
is not empty.
The partially ordered set(Fµν,C)has a minimum. We call it the shadow ofµin νand denote it bySν(µ).
µ
ν γ1
γ2
Sν(γ1) =ν1
Sν−ν1(γ2)
γ1
µ
γ2
ν Sν(γ1) =ν1 Sν−ν1(γ2)
Figure:Shadow ofµinνand associativity of the shadow projection.
Nicolas JUILLET A canonical martingale coupling
A canonical martingale coupling Tools for the martingale transport problem
The variational lemma
This lemma is a kind ofc-cyclical monotonicity lemma for the martingale setting.
Variational Lemma
LetPbe optimal, there existsΓ⊆R×Rsuch that for any finitely supported measure αwithα(Γ) =1, the minimum ofα07→Rc(x,y)dα0(x,y)over
Competitor(α) =
α0:α0
has the same marginals asα
∀x∈R, Z
ydαx= Z
ydα0x
is obtained inα.
A canonical martingale coupling Results
Martingale theorem
Theorem
Forµandνin
P
3in the convex order andPa martingale transport plan fromµtoν. The following statements are equivalent:The planPis optimal for the martingale transport problem with cost c(x,y) = (y−x)3,
The planPis concentrated on a martingale-monotone setΓ(see the figure), The planPis the left- curtain coupling (i.e., transportsµ]−∞,x]to its shadow)
x x0
y− y0 y+
Figure:This configuration of three points(x,y),(x0,y−)and(x0,y+)is forbidden on martingale-monotone setsΓ.
Nicolas JUILLET A canonical martingale coupling
A canonical martingale coupling Results
One example
x x0
T2(x0) T1(x0) T1(x) =T2(x)
Figure:Optimal transport plan between Gaussian measures.
A canonical martingale coupling Results
Corollary
Corollary forµcontinuous
Ifµis continuous (=no atom), there areT1,T2:R→Rsuch that the optimalPis concentrated ongraph(T1)∪graph(T2).
The variational lemma is of general use, especially whenµis continuous.
Examples
c(x,y) =−|y−x| c(x,y) =|y−x| c(x,y) = (y−x)n
Nicolas JUILLET A canonical martingale coupling