Multivariate Risk Measures
Ivar Ekeland and Walter Schachermayer
Zurich, May 5, 2011
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 1 / 16
Monovariate risk measures
A monovariate risk measure(1-rm) is a functionρ :L∞(Ω,F,P)!∞ such that:
ρ(0) =0
X Y =)ρ(X) ρ(Y)
ρ(X +m) =ρ(X) m for m2R It is
convex ifρ is a convex function
coherent if it is convex and positively homogeneous:
ρ(X+Y) ρ(X) +ρ(Y) ρ(λX) = λρ(X)
Law-invariant 1-r.m.
We shall write X Y to mean thatX andY have the same law. ρ is law-invariant if X Y implies that ρ(X) =ρ(Y)
De…nition
A 1-rm is strongly coherent if it is convex, law-invariant and:
ρ(X) +ρ(Y) =supn
ρ X+Ye j Y Yeo There are two fundamental examples of s.c. 1-r.m.
1 Let F 2L1 be a probability density, so thatF 0 and E[F] =1.
De…ne
ρF :=supn Eh
FXei
j Xe Xo
2 Set:
ρ∞(X):=esssup X Both ρF and ρ∞ are strongly coherent.
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 3 / 16
Kusuoka’s theorem
Kusuoka (2001) has shown that all strongly coherent risk measures can be built from ρF and ρ∞
Theorem
ρis a strongly coherent 1-r.m. if and only if there is a probability density F and a number s with 0 s 1 such that:
8X 2L∞, ρ(X) =sρ∞(X) + (1 s)ρF(X)
Multivariate risk measures
A d-dimensional risk measure(d-r.m.) is a function ρ:L∞ Ω,F,P;Rd !R such that:
ρ(0) =0
X Y =)ρ(X) ρ(Y)
ρ(X +me) =ρ(X) m for m2R ande = (1, ...,1) It is
convex ifρ is a convex function
coherent if it is convex and positively homogeneous:
ρ(X+Y) ρ(X) +ρ(Y) ρ(λX) = λρ(X)
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 5 / 16
Law-invariant d-r.m.
ρis law-invariant if X Y implies that ρ(X) =ρ(Y) De…nition
(Galichon, Henry) A d-rm is strongly coherentif it is convex, law-invariant and:
ρ(X) +ρ(Y) =supn
ρ X+Ye j Y Yeo
Examples
There are two fundamental examples of s.c. 1-r.m.
1 Let F = (F1, ...,Fd)2L1 Rd satisfy Fi 0,1 i d, and
∑iE[Fi] =1. De…ne ρF :=supn
Eh FXei
jXe Xo=sup (
∑
iEh FiXei
i jXe X )
2 Denote bySd the unit simplex in Rd and let ξ 2Sd, so thatξi 0 and ∑ξi =1. De…ne
ρξ(X):=esssup Xξ =esssup
∑
Xiξi3 Let µbe a probability on Sd. Set:
ρµ(X):=
Z
Sdρξ(X)dµ(ξ) ρF andρµ are strongly coherent d-r.m.
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 7 / 16
Extension of Kusuoka’s theorem
Schachermayer and IE have shown that all strongly coherent risk measured-r.m. can be built from ρF andρ∞
Theorem
ρis a strongly coherent d-r.m. if and only if there is some F 2L1+ Rd with ∑iE[Fi] =1, a probabilityµon Sd and a number s with0 s 1 such that:
8X 2L∞, ρ(X) =sρµ(X) + (1 s)ρF (X)
This result builds on an earlier result by Galichon, Henry and IE, which we will explain in due course.
Duality
If ρ:L∞ Rd !R is a convex and law-invariant, it is lower
semi-continuous wrt σ L∞,L1 (Jouini, Schachermayer, Touzi, 2006). It follows that ρ is coherent (homogeneous) if and only if there exists a closed, convex, law-invariant subsetC of L1 Rd such that.
ρ(X) = sup
F2C hF,Xi
We shall set ρ=ρC to remember this fact. Denote by T the set of measure-preserving maps τ fromΩ into itself. Set:
K :=f(F,F τ) jF 2C, τ2 T g Since C is law-invariant,K C C. Conversely:
Lemma
ρis strongly coherent i¤ C C is the closed convex hull of K
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 9 / 16
Proof
Suppose ρ=ρC andC C is the closed convex hull ofK. Then:
ρC(X) +ρC(Y) = supf hF,Xi hG,Xi j (F,G)2C Cg
= supf hF,Xi hF τ,Yi j F 2C, τ2 T g
= sup hF,X +Y τ 1i j F 2C, τ2 T sup
n
hF,X+Yei j Ye Yo=ρC X+Ye
and the reverse inequality is always true
Geometry of strongly coherent sets
Theorem
Let C L1 Rd be strongly coherent. Then there exists a number t, with 0 t 1, and two closed convex law-invariant subsets Cr and Cs of L1 Rd such that:
1 C = (1 t)Cr+tCs
2 Cr is σ L1,L∞ -compact
3 Cs (L∞) isσ (L∞) ,L∞ -compact, and its extreme points are purely singular
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 11 / 16
The regular case
This is the case when Cs =?. In other words, C =Cr is weakly compact in L1 Rd . For instance, C Lp Rd , so thatρ= ρC is continuous in theLp Rd norm. This case was dealt with in an earlier paper by Galichon, Henry and IE.
SinceC is weakly compact, it is the closed convex hull of its set of strongly exposed points. Let F 2C be such a point. We claim that ρ=ρF.
Proof
By de…nition, there is some X 2L∞ Rd such that supH2C
hH,Xi= hF,Xiand if Hn 2 C is a maximizing sequence, then Hn !F strongly inLi Rd . Sinceρ is strongly coherent, we have:
ρ(X) +ρ(Y) =supf hG,Xi hG τ,Yi j G 2C, τ2 T g Let (Gn,τn) be a maximizing sequence. Since we have= instead of , we must have ρ(X) =sup hGn,Xiandρ(Y) =sup hGn τ,Yiso Gn !F and:
ρ(Y) =sup
τ hF τ,Yi= sup
τ hF,Y τ 1i= sup
Y Ye
hF,Yei so ρ =ρF as desired
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 13 / 16
The singular case 1
Let G = (Ai) be a …nite partition ofΩ. On eachAi we are given a point xi 2Sd and an element βi 2(L∞(R)) . We then build an element βG 2 L∞ Rd by the formula
hβG,Xi=
∑
hβi,(xi X)iDenote by δi the Dirac mass atxi. We associate withβG the Borel measure µG onSd de…ned by:
µG =
∑
δiβi(Ai)Lemma
Suppose βG is law-invariant. Then, for X 2L∞ Rd we have:
Z
Proof
Z
Sd
sup (x X)dµ(x) =
Z
Sd
sup x X τ 1 dµ(x)
=
∑
sup xi X τ 1 βi(Ai)hβG τ,Xi
and the reverse inequality follows from the fact that βG is purely singular.
Taking points (ξi)and disjoint sets(Bi) with P(Bi)>0 such that
esssup xi X = xi ξi andBi fjX ξij<εg
we can …nd a set N with 0<P(N)<infP(Bi) such that βG1N = βG. We then …nd a measure-preserving transformation which maps N\Ai into Bi and the result follows
Ivar Ekeland and Walter Schachermayer () Multivariate Risk Measures Zurich, May 5, 2011 15 / 16
The singular case: simple elements
Let β2 L∞ Rd 0 be a purely singular element. We de…ne a …nitely additive measure jβj by:
jβj[A] =he1A,βi Let G = (Ai) be a …nite partition ofΩ. Set:
βG =
∑
ξi(jβj[Gi])ξi = hei1Gi,βi he1Gi,βi
We can approximate βby the βG which have a simpler form. We associate with βG the Borel measure:
µG =