Initial enlargement of filtrations and entropy of Poisson compensators

Initial enlargement of filtrations and entropy of
Poisson compensators
Stefan Ankirchner and Jakub Zwierz
“Enlargement of Filtrations and Applications to Finance and Insurance”
Jena, May 31 - June 4, 2010
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
1
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
Hypothesis (H’)
F = (Ft ) a filtration
G ⊃ F an enlargement
Hypothesis (H’):
every F-martingale is a G-semimartingale.
Question: Does the enlargement preserve integrability?
For example, when do we have for any r , p ≥ 1:
Every Hp (F)-martingale is a S r (G)-semimartingale.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
2
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
S p norms for semimartingales
Recall the definition of the k · kS p norm:
Definition
X a special F-semimartingale,
canonical F-decomposition: X = M + A
Let 1 ≤ p < ∞. Then
Z
1
2
kX kS p (F) := [M, M]∞
+
0
∞
|dAs | p .
L
S p (F) is the set of all F-semimartingales with kX kS p (F) < ∞.
Hp (F) is the set of all F-martingales with kX kS p (F) < ∞.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
3
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
0
Hypothesis (Hp,r
)
Suppose (H’) is satisfied. If M is an F-martingale and ψ is
G-predictable and bounded, then
(ψ · M) is defined wrt G.
0 ):
Hypothesis (Hp,r
there exists a constant Cp,r such that for every Fmartingale M and every bounded G-predictable ψ
Z
kψ · MkS r (G) ≤ Cp,r
E
∞
p ! p1
2
ψ 2 dhM, Mi
0
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
4
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
Initial enlargements by finite partitions
let F be a filtration s.th. every F-martingale is continuous,
let A1 , A2 , . . . be a countable measurable partition,
let G = (Gt ) be the initial enlargement
\
Gt =
(Fs ∨ σ(A1 , A2 , . . .)).
s>t
Theorem (Yor 1985, LNM 1118)
Let p, γ > 0 and r ≥ 1 such that
0 ) is satisfied iff
(Hp,r
X
P(An ) log
n≥1
Stefan Ankirchner and Jakub Zwierz
1
r
=
1
P(An )
1
p
+
1
2γ .
Then Hypothesis
γ
< ∞.
Initial enlargement & Poisson compensators
5
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
Initial enlargements by finite partitions
Special case: γ = 1. Then
X
n≥1
P(An ) log
1
P(An )
γ
is the absolute entropy of the partition A1 , A2 , . . .
Corollary
If the partition’s absolute entropy is finite, then
H2 (F) → S 1 (G), X 7→ X , is continuous.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
6
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
Arbitrary initial enlargements
let F be a filtration s.th. every F-martingale is continuous,
let G = (Gt ) be an arbitrary initial enlargement by a random
variable G
\
Gt =
(Fs ∨ σ(G )).
s>t
Theorem (A., Dereich, Imkeller, Sem. de Prob. XL, 2007)
1
Let p, γ > 0 and r ≥ 1 such that 1r = p1 + 2γ
. Then Hypothesis
0
(Hp,r ) is satisfied iff the generalized mutual information satisfies
I p (σ(G ), F∞ ) < ∞.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
7
Beyond Hypothesis (H’)
Continuous embedding of continuous martingales
Information drifts
Definition
A G-predictable process α is called G-information drift of an
F-martingale M if
Z t
Mt −
αs dhM, Mis
0
is a G-local martingale.
Theorem (A., Dereich, Imkeller, ’07)
If I (σ(G ), F∞ ) < ∞, then every F-martingale has a square
integrable G-information drift.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
8
Poisson random measures and filtrations
Information drift in terms of compensators
Problem description
So far: Every F-martingale was supposed to be continuous.
Questions:
When is integrability of semimartingales with jumps preserved
under enlargements?
0 ) for jump semimartingales?
When do we have (Hp,r
Does finite entropy guarantee integrable information drifts?
Does finite entropy imply that every jump semimartingale M
in H 2 (F) belongs to S 1 (G)?
To simplify the analysis we make the following restrictions:
initial enlargements
pure jump martingales driven by a comp. Poisson random
measure
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
9
Poisson random measures and filtrations
Information drift in terms of compensators
Poisson random measures
jump space: a st. Borel space (E , E)
jump measure: ν (assumed to be σ-finite)
Let µ be a homogeneous Poisson random measure on
(R+ × E , λ ⊗ ν). In particular
a) if λ ⊗ ν(A) < ∞, then µ(ω; A) is Poisson distributed with
intensity λ ⊗ ν(A),
b) if A ∩ B = ∅, then µ(ω; A) and µ(ω; B) are independent.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
10
Poisson random measures and filtrations
Information drift in terms of compensators
Poisson compensators depend on the filtration
Lemma and Definition
F = the filtration generated by µ
H ⊃ F an enlargement
P(H) = the predictable σ-field on Ω × R+ associated with H
There exists a unique predictable random measure π H , called
compensator of µ relative to H, s.th.
!
!
Z Z
Z Z
∞
∞
φ(s, z)µ(dz, ds)
E
0
E
φ(s, z)π H (dz, ds)
=E
0
E
for every nonnegative P(H) ⊗ E-measurable function φ.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
11
Poisson random measures and filtrations
Information drift in terms of compensators
Poisson compensators depend on the filtration
Remark.
Compensators exist for all enlargements of F!
π F = λ ⊗ ν.
Let µH := µ − π H . If ψ is square integrable, then
Z tZ
H
(ψ ∗ µ )t =
ψ(s, z)µH (ds, dz)
0
E
is a square-int. H-martingale.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
12
Poisson random measures and filtrations
Information drift in terms of compensators
Information drifts of pure jump martingales
F = (Ft ) a filtration
G ⊃ F an enlargement
Definition
A G-predictable process α is called G-information drift of an
F-martingale M if
Z t
Mt −
αs ds
0
is a G-local martingale.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
13
Poisson random measures and filtrations
Information drift in terms of compensators
Linking information drifts with compensators
Rt R
let Mt = 0 E ψ(s, z)µF (ds, dz) be an F-martingale
G ⊃ F an enlargement
α = G-information drift of M
Notice that
t
Z
Mt −
αs ds
(ψ ∗ µF )t −
=
0
t
Z
αs ds
0
(ψ ∗ µG )t + (ψ ∗ (π G − π F ))t −
|
{z
G − predictable
=
t
Z
0
αs ds
}
Z
=⇒
αs
=
ψ(s, z)(π G − π F )(ds, dz)
G
Z
dπ
ψ(s, z)
−
1
π F (ds, dz)
dπ F
E
E
=
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
14
Poisson random measures and filtrations
Information drift in terms of compensators
Linking information drifts with compensators
The information drift, provided it exists, satisfies
G
Z
dπ
αs =
ψ(s, z)
− 1 π F (ds, dz)
dπ F
E
Question: Is the Lebesgue integral defined P-a.s.?
a necessary condition: π G × P π F × P on P(G) ⊗ E
a sufficient condition: ψ ∈ L2 (π F × P) ∩ L1 (π F × P)
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
15
Entropy and mutual information
Poisson compensators of initial enlargements
Mutual information
Definition
let A, B be two sub-σ-algebras of F
P(·|B) = regular conditional probability of P with respect to B
The mutual information between A and B is defined as
I (AkB) = EHA (P(·|B)kP),
where HA (P(·|B)kP) is the relative entropy of P(·|B) wrt P on A.
This means: if P(·|B) P on A, then
Z Z
P(dω 0 |B)(ω) I (AkB) =
log
P(dω 0 |B)P(dω).
A
P(dω 0 )
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
16
Entropy and mutual information
Poisson compensators of initial enlargements
Mutual information between random variables
Lemma
Let G , H be two random variables, and denote by PG and PH their
distributions. Moreover let PG ,H be the joint distribution of G and
H. Setting A = σ(G ) and B = σ(H), we have:
I (AkB) = H(PG ,H kPG ⊗ PH ).
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
17
Entropy and mutual information
Poisson compensators of initial enlargements
Properties of mutual information
Properties:
1)
2)
3)
4)
I (A, B) = I (B, A),
A, B independent =⇒ I (A, B) = 0,
if C ⊂ B, then I (A, C)) ≤ I (A, B),
if X and Y are Gaussian, then
1
I (σ(X ), σ(Y )) = − log 1 − corr(X , Y )2 .
2
5) Let A be generated by a countable partition
{A1 , A2 , . . .}. Then I (A, A) = absolute entropy of
A, i.e.
X
I (A, A) = −
P(Ai ) log P(Ai ).
i≥1
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
18
Entropy and mutual information
Poisson compensators of initial enlargements
Initial enlargement of Poisson filtrations
Back to Poisson random measures:
F = (Ft ) = filtration generated by µ(dt, dz)
W
F∞ = t Ft
G = (Gt ) = initial enlargement by a random variable G :
\
Gt =
Fs ∨ σ(G )
s>t
Question: What does I (σ(G )kF∞ ) tell us about π G (dt, dz), or
the information density
G
dπ
−1 ?
dπ F
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
19
Entropy and mutual information
Poisson compensators of initial enlargements
Mutual information and entropy of compensators
Theorem
(i) If T ∈ R+ and ν(E ) < ∞, then
HPT (G)×E (π G × Pkπ F × P) = I (σ(G )kFT ).
(ii) Let Tn ∈ R+ such that Tn ↑ ∞, and E1 , E2 , . . . an increasing
sequence of sets in E with En ↑ E and ν(En ) < ∞. Then we
have
sup HPTn (G)×(E∩En ) (π G × Pkπ F × P) = I (σ(G )kF∞ ).
n
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
20
Entropy and mutual information
Poisson compensators of initial enlargements
Convex conjugate of the information density
The G-information drift of an F-martingale M = (ψ ∗µF ):
G
Z
dπ
αs =
− 1 π F (ds, dz)
ψ(s, z)
F
dπ
E
If
)kF
∞ ) is finite, then we have convex integrability of
I (σ(G
dπ G
− 1 wrt
dπ F
f (x) = (x + 1) log(x + 1) − x.
Namely,
Z G
Z
G
dπ
dπ G
dπ
F
F
E
f
− 1 π (ds, dz) = E
log
π (ds, dz)
dπ F
dπ F
dπ F
= I (σ(G )kF∞ )
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
21
Entropy and mutual information
Poisson compensators of initial enlargements
Convex conjugates
Recall the definition of the convex conjugate or Legendre
transform of a convex function f : R → R ∪ {+∞}:
f ∗ : R → R ∪ {+∞},
f ∗ (y ) := sup(xy − f (x)).
x∈R
Young’s Inequality:
xy ≤ f (x) + f ∗ (y ).
The convex conjugate of
f (x) = (x + 1) log(x + 1) − x,
x > −1,
is given by
f ∗ (y ) = e y − y − 1.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
22
Entropy and mutual information
Poisson compensators of initial enlargements
Convex conjugates
Young’s inquality implies
xy ≤ e y − y − 1 + (x + 1)log (x + 1) − x.
Proposition
Let M = (ψ ∗ µF ) be an F-loc. martingale, and suppose
Z
f ∗ (|ψ|) d(π F × P) < ∞.
If I (σ(G )kF∞ ) is finite, then M possesses an integrable
G-information drift. It is given by
G
Z
dπ
αs =
ψ(s, z)
− 1 π F (ds, dz)
F
dπ
E
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
23
Finite entropy implies integrable information drifts
Theorem
Let G = (Gt ), where Gt =
If I (σ(G )kF∞ ) < ∞, then
T
s>t
Fs ∨ σ(G ).
(i) every X ∈ H2 (F) possesses an integrable G-information drift,
(ii) H2 (F) ⊂ S 1 (G),
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
24
Example I
Let ψ ∈ L1 (π F ) be a deterministic positive function and define
Z ∞Z
G=
ψ(s, z)µ(ds, dz).
0
E
ThenR G Ris Poisson distributed with parameter
∞
λ = 0 E ψ(s, z)π F (ds, dz) ∈ R+ . The mutual information
between G and F∞ is given by
I (σ(G )kF∞ ) = −
∞
X
i=0
e −λ
λi λi
log e −λ
,
i!
i!
which is easily shown to be finite. Consequently every square
integrable F-martingale has a G-information drift and belongs to
S 1 (G),
T where the filtration G = (Gt ) is defined by
Gt = s>t Fs ∨ σ(G ).
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
25
Example II
Let τ be the first jump time of a standard Poisson process
N = (Nt )t≥0 . Notice that I (σ(τ ), F∞ ) = ∞, since the distribution
of τ is absolutely continuous wrt the Lebesgue measure.
Add some noise:
Let X be normally distributed, and independent of F∞ . Let
τ̂ = τ + X ,
and consider the enlargement
Gt =
\
Fs ∨ σ(τ̂ ).
s>t
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
26
Example II cont’d
Since X is normally distributed, the mutual information satisfies
1 Var(τ ) + Var(X ) I (σ(τ̂ )kF∞ ) ≤ log
< ∞.
2
Var(X )
Notice that the variance Var(τ ) is defined since τ is exponentially
distributed.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
27
References
S.A. and Jakub Zwierz: Initial enlargement of filtrations and
entropy of Poisson compensators. Journal of Theoretical
Probability 2010. In Press.
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
28
Thanks for your attention !!
Stefan Ankirchner and Jakub Zwierz
Initial enlargement & Poisson compensators
29