Transcript Document

On some integral estimates for solutions
of SDEs driven by symmetric stable processes
Vladimir P. KURENOK
Department of Electrical and Systems Engineering,
Washington University in St. Louis,
One Brookings Drive, St. Louis, MO 63130-4899, USA
e-mail: [email protected]
Workshop on Stochastic Analysis - Jena 2015
Introduction
We assume Z to be a one-dimensional symmetric stable process of
index  ∈(0, 2] and X to be of the form
𝑋𝑑 = π‘₯0 +
𝑑
𝑏
0
π‘‹π‘ βˆ’ 𝑑𝑍𝑠 +
𝑑
π‘Ž
0
𝑋𝑠 𝑑𝑠 , π‘₯0 ∈ ℝ.
(A)
For a measurable function 𝑓 ∢ ℝ β†’ 0, ∞ , define
π‘š
1
𝑝
𝑓 𝑝,π‘š =
( 𝑓 𝑦 𝑑𝑦) 𝑝
βˆ’π‘š
to be its 𝐿𝑝 -norm on the interval [-m, m], m ∈ β„• , p ο‚³ 1 , and let
πœπ‘š (X) = inf 𝑑 β‰₯ 0 ∢ 𝑋𝑑 β‰₯ π‘š .
The estimates of the form
π‘‘βˆ§πœπ‘š (𝑋)
𝑬
𝑒 βˆ’πœ™π‘  πœ“π‘  𝑓 𝑋𝑠 𝑑𝑠 ≀ 𝑁 𝑓
𝑝,π‘š
(1)
0
where (πœ™π‘‘ ) and (πœ“π‘‘ ) are nonnegative processes are called (local) Krylov's estimates.
οƒ˜ N. V. Krylov (Controlled Diffusion Processes, Springer, New York, 1980) proved
them first for diffusion processes, that is when Z is a Brownian motion
process W (Ξ± = 2).
οƒ˜ Krylov's estimates for diffusion processes with jumps (W β‰  0):
- S. Anulova and H. Pragarauskas: Liet. Math. Rinkinys, XVII (1977), 5-26;
- J. P. Lepeltier and B. Marchal: Annales IHP, Vol. 12, No. 1 (1976), 43-103;
- A.V. Melnikov: Stochastics and Stoch. Rep., 10 (1983), 81-102.
οƒ˜ Krylov's estimates for purely discontinuous processes:
- H. Pragarauskas, In: "Probab. Theory and Math. Statist.", B.Grigelionis et al.
(eds.), 579-588, VSP, Utrecht/TEV, Vilnius, 1999 (1 < Ξ± < 2 and a = 0);
- V. P. Kurenok: Transactions AMS, Vol. 360, No. 2, 925-938, 2008 (1 < Ξ± < 2);
- Xicheng Zhang: AIHP, Vol. 49, no.4, 1057-1079, 2013 (b = 1 and
multidimensional drift a).
οƒ˜ Our goal here is to discuss some versions of Krylov's
estimates for equation (A) in π‘³πŸ -norm and their use to prove
the existence of weak solutions for corresponding SDEs.
We recall:
If Z is a symmetric stable process with index 𝛼 ∈ (0, 2 , then its characteristic
function is given by
𝑬 π‘–πœ‰π‘π‘‘ = exp βˆ’π‘‘ πœ‰
𝛼
, 𝑑 > 0, πœ‰ ∈ ℝ.
For α ∈ 0,2 , Z is a purely discontinuous Markov process that can be
described by its infinitisimal generator given by
ℒ𝑔 π‘₯ =
𝑔 π‘₯+𝑧 βˆ’π‘” π‘₯ βˆ’1
where 𝑐𝛼 is a constant.
∞
𝑧 <1
𝑔′
π‘₯ 𝑧
𝑐𝛼
𝑧
1+𝛼
𝑑𝑧
For 𝑔 ∈ 𝐿1 ℝ , define
𝑒 𝑖𝑧π‘₯ 𝑔 𝓏 𝑑𝓏
𝐹𝑔(π‘₯) ≔
βˆ’βˆž
to be the Fourier transform of 𝑔. For any function 𝑔 ∈ 𝐢 2 ℝ such that ℒ𝑔 ∈ 𝐿1 , it
holds then
𝐹 ℒ𝑔 π‘₯ = βˆ’ π‘₯ 𝛼 𝐹𝑔 π‘₯
and
𝐹𝑔′ π‘₯ = 𝑖π‘₯𝐹𝑔 π‘₯ .
PART I: Estimates in the case of 1<Ξ±<2
We assume that:
οƒ˜ 𝐾 > 0 is a constant;
οƒ˜ 𝑓
is a nonnegative, measurable function such that 𝑓 ∈ 𝐢0 ∞ ℝ where
𝐢0 ∞ ℝ denotes the class of all infinitely differentiable real valued functions
with compact support defined on ℝ;
οƒ˜ 𝑍 is a symmetric stable process of index Ξ± ∈ 1,2 adapted to a filtration 𝔽;
οƒ˜ D denotes the class of all 𝔽 -predictable processes (𝛾𝑑 ) such that 𝛾𝑑 ≀ 𝐾.
For a process 𝑋 𝛾 (controlled process) of the form
𝑑𝑋𝑑
𝛾
= 𝑑𝑍𝑑 + 𝛾𝑑 𝑑𝑑
and any πœ† > 0, define the value function 𝑣 π‘₯ , π‘₯ ∈ ℝ, by
∞
𝑒 βˆ’πœ†π‘‘ 𝑓 π‘₯ + 𝑋𝑑
𝑣 π‘₯ = sup 𝑬
π›Ύβˆˆπ·
0
𝛾
𝑑𝑑.
οƒ˜ By standard arguments (N.V. Krylov and H. Pragarauskas: "Traditional
derivation of Bellman equation for general controlled stochastic processes", Lit.
Math. Rink., 21, 146-152, 1982), the function v will satisfy the following Bellman
equation (  is deterministic!)
𝑠𝑒𝑝 ℒ𝑣 π‘₯ βˆ’ πœ†π‘£ π‘₯ + 𝛾𝑣 β€² π‘₯ + 𝑓 π‘₯
=0
𝛾≀𝐾
which holds a.e. in ℝ.
οƒ˜ The above equation is then equivalent to
ℒ𝑣 βˆ’ πœ†π‘£ + 𝐾 𝑣 β€² + 𝑓 = 0.
(2)
Lemma 1.
𝑣 π‘₯ ≀𝑁 𝑓
For all π‘₯ ∈ ℝ, it holds
∞
2
1
2
(3)
𝑓 2 𝑦 𝑑𝑦
≔𝑁
βˆ’βˆž
where the constant N depends on K and Ξ± only.
οƒ˜ The proof uses Fourier transform technique combined with the Parseval's
identity. As the result, it holds for all 𝑦 ∈ ℝ and πœ† β‰₯ πœ‡
∞
𝑣 2 (𝑦) ≀ 𝑁
𝑓 2 (π‘₯)𝑑π‘₯
βˆ’βˆž
where
1
𝑁= 2
πœ‹
and µ > 0 is such that
for all π‘₯ ∈ ℝ.
∞
π‘₯
𝛼
+πœ†
βˆ’2 𝑑π‘₯
π‘₯
𝛼
+πœ‡
2
<∞
βˆ’βˆž
β‰₯ 4𝐾 2 π‘₯
2
Theorem 1.
π‘Ž π‘₯
Suppose X is a solution of the equation (A) with 𝛼 ∈ 1,2 and
≀𝐾𝑏 π‘₯
𝛼
(4)
for all π‘₯ ∈ ℝ.
Then, for any π‘₯ ∈ ℝ, πœ† β‰₯ πœ‡ , and any measurable function 𝑓 ∢ ℝ β†’ 0, ∞ , it holds
∞
𝐸
0
where πœ“π‘‘ =
𝑑
0
𝑒 βˆ’πœ†πœ“π‘’ 𝑏 𝑋𝑒
𝑏 𝑋𝑠
𝛼 𝑑𝑠
𝛼𝑓
π‘₯ + 𝑋𝑒 𝑑𝑒 ≀ 𝑁 𝑓
2
(5)
and the constant N depends on K and Ξ± only.
οƒ˜ Accordingly, a local version holds as well:
π‘‘βˆ§πœπ‘š 𝑋
𝐸
𝑏 𝛼 𝑓 𝑋𝑒 𝑑𝑒 ≀
0
𝑁 𝑓
where N depends then on K, Ξ±, m, and t.
οƒ˜ The proof uses Ito's formula and Lemma 1.
2,π‘š
(6)
Application of the estimates in case 1 < Ξ± < 2:
Theorem 2. Assume that Z is a symmetric stable process of index 1 < Ξ± < 2
and there exist positive constants Ξ΄1 and Ξ΄2 such that
𝑏 π‘₯
β‰₯ 𝛿1 ,
π‘Ž π‘₯ + |𝑏 π‘₯ |≀ 𝛿2
for all π‘₯ ∈ ℝ.
Then, for any π‘₯0 ∈ ℝ, there exists a solution of the equation (A).
οƒ˜ In particular, if π‘Ž π‘₯
≀ 𝛿2 for all π‘₯ ∈ ℝ, then the equation (A) with b = 1 has a
solution for any π‘₯0 ∈ ℝ.
οƒ˜ To compare: N. I. Portenko (Random Oper. and Stoch. Equ., Vol. 2, No. 3, 1994)
proved the existence of solutions for the equation (A) with b = 1 if there is 𝑝 >
π›Όβˆ’1
βˆ’1
such that π‘Ž ∈ 𝐿𝑝 ℝ .
οƒ˜ Sufficient conditions in the Theorem 2 are different from those of Portenko and the
proof method used by Portenko is a different one from using the Krylov's estimates.
PART II:
Integral estimates in the case of 0 < Ξ± < 2
We assume:
οƒ˜ Z is a symmetric stable process of index 0 < Ξ± < 2;
οƒ˜ 𝑓 is a nonnegative, measurable function such that 𝑓 ∈ 𝐢0 ∞ ℝ ;
οƒ˜ 𝛾 is a fixed constant.
For  > 0, we consider the linear fractional ODE:
ℒ𝑣 π‘₯ βˆ’ πœ†π‘£ π‘₯ + 𝛾𝑣 β€² π‘₯ + 𝑓 π‘₯ = 0.
Lemma 2.
(7)
Let 𝛾 β‰  0. Then, for all π‘₯ ∈ ℝ and 0 < Ξ± < 2, it holds
𝑣 π‘₯ ≀𝑁 𝑓
2,
where the constant 𝑁 depends on Ξ± and 𝛾.
If 𝛾 = 0, then the estimate holds only for all 1/2 < Ξ± < 2.
οƒ˜ The proof uses again Fourier transform technique combined with the Parseval
identity. As the result, one has for all π‘₯ ∈ ℝ and > 0:
∞
𝑣 2 (π‘₯)
where
𝑁
≀ 2
4πœ‹
∞
𝑁≔
βˆ’βˆž
𝑓 2 𝑧 𝑑𝑧
βˆ’βˆž
π‘‘πœ‰
πœ‰ 𝛼 + πœ† 2 + 𝛾2πœ‰2
One sees that N < ∞ for 𝛾 β‰  0 and 0 < Ξ± < 2 while N < ∞ in the case of
𝛾 = 0 and 1/2 < Ξ± < 2.
οƒ˜ Addition of the drift term 𝛾𝑑 (𝛾 β‰  0) to the process Z allows to expand the
analytical estimate of the sup-norm of the function 𝑣 π‘₯ through the 𝐿2 -norm of
the function 𝑓 π‘₯ from the range 1/2 < Ξ± < 2 to the entire range 0 < Ξ± < 2 of the
index. In this sense the addition of the drift plays a "regularizing effect" on the
estimates.
Assume now that X is of the form
𝑑𝑋𝑑 = 𝑏 π‘‹π‘‘βˆ’ 𝑑𝑍𝑑 + 𝛾 𝑏
𝛼
𝑋𝑑 𝑑𝑑,
(B)
where 𝑑 β‰₯ 0 and 𝑋0 = π‘₯0 ∈ ℝ .
One has then the following (local) Krylov's estimate:
Theorem 3. Let 𝛾 β‰  0 and X be of the form (B). Then, for all 0 < Ξ± < 2, 𝑑 β‰₯ 0,
m ∈ β„•, and any measurable function f: ℝ β†’ 0, ∞ , one has
𝐄
π‘‘βˆ§πœπ‘š 𝑋
0
𝑏 𝛼 (𝑋𝑠 )𝑓 𝑋𝑠 𝑑𝑠 ≀ 𝑁 𝑓
2,π‘š
where the constant N depends on α, , m, and t.
If 𝛾 = 0 , then the estimate is true for 1/2 < Ξ± < 2 only.
(8)
Application of the estimates in case 0 < Ξ± < 2:
Theorem 4. Let 0 < Ξ± < 2 and assume that there exists a constant
𝐾 > 0 such that 𝑏 π‘₯
≀ 𝐾 for all π‘₯ ∈ ℝ and 𝑏
βˆ’2𝛼
∈ πΏπ‘™π‘œπ‘ ℝ . Then,
for any π‘₯0 ∈ ℝ, there exists a solution of the equation (B) for any 𝛾 β‰ 
0 and 𝛾 = 0.
οƒ˜ πΏπ‘™π‘œπ‘ ℝ denotes the class of locally integrable functions on ℝ.
Proof idea:
For any fixed 𝜸 β‰  𝟎 :
οƒ˜ One chooses a sequence of Lipshitz continuous and uniformly bounded
functions 𝑏𝑛 , 𝑛 β‰₯ 1, converging to b so that, for any fixed n and the given
process Z, there is a solution of the equation
𝑑
𝑋𝑑 = π‘₯0 +
0
𝑑
𝑏𝑛 π‘‹π‘ βˆ’ 𝑑𝑍𝑠 + 𝛾
0
𝑏𝑛
𝛼
𝑋𝑠 𝑑𝑠,
𝑑 β‰₯ 0.
οƒ˜ One proves then that the sequence of processes 𝑋 𝑛 , 𝑍 , 𝑛 = 1,2, … is tight and
converges in weak sense so that there is a limiting process 𝑋, 𝑍 with 𝑍 being a
symmetric stable process of the same index Ξ±.
οƒ˜ One shows then - with help of Krylov's estimates – that the process 𝑋, 𝑍
satisfies the equation (B).
For 𝜸 = 𝟎:
οƒ˜ One chooses a sequence of real numbers 𝛾𝑛 β‰  0, 𝑛 = 1, 2, … such that lim 𝛾𝑛 = 0.
π‘›β†’βˆž
For any fixed 𝑛 = 1, 2, … . , there is a solution 𝑋 𝑛 , 𝑍 𝑛 of the equation
𝑑𝑋𝑑 = 𝑏 π‘‹π‘‘βˆ’ 𝑑𝑍𝑑 + 𝛾𝑛 𝑏
𝛼
𝑋𝑑 𝑑𝑑,
𝑋 = π‘₯0 ∈ ℝ (9)
where 𝑍 𝑛 is a symmetric stable process of index 0 < Ξ± < 2.
οƒ˜ One proves as before that the sequence of processes 𝑋 𝑛 , 𝑍 𝑛 , 𝑛 = 1,2, … is tight
and converges in weak sense so that there is a limiting process 𝑋, 𝑍 with 𝑍
being a symmetric stable process of the same index Ξ±.
οƒ˜ One shows then that the limiting process 𝑋, 𝑍 satisfies the equation
𝑑𝑋𝑑 = 𝑏 π‘‹π‘‘βˆ’ 𝑑𝑍𝑑 ,
(𝐢)
where 𝑑 β‰₯ 0 and 𝑋0 = π‘₯0 ∈ ℝ .
οƒ˜ In the last step, one uses the Skorokhod's lemma for the convergence of stable
integrals instead of Krylov's estimates since the coefficient b in (9) does not
depend on n.
Discussion of the existence result in case 𝜸 = 𝟎 :
οƒ˜ 𝛼 = 2 (H.J. Engelbert and W. Schmidt): equation (C) has a non-trivial
solution for any π‘₯0 ∈ ℝ if and only if 𝑏 βˆ’2 ∈ πΏπ‘™π‘œπ‘ ;
οƒ˜ 𝛼 ∈ 1,2 (P.A. Zanzotto, 2002): the same result as for 𝛼 = 2 where the
condition 𝑏 βˆ’2 ∈ πΏπ‘™π‘œπ‘ was replaced by the condition 𝑏
βˆ’π›Ό ∈
πΏπ‘™π‘œπ‘ ;
οƒ˜ 0 < Ξ± ≀ 1: there are only sufficient conditions, for example:
1) There exists a number 𝛿 > 1 such that
𝑏
βˆ’π›Ώ ∈
πΏπ‘™π‘œπ‘ ;
2) There exists a number π‘ˆ > 0 with 𝑙 π΅π‘ˆ < ∞ such that
π΅π‘ˆ = 𝑦 ∈ ℝ: 𝑏 𝑦
>π‘ˆ
where l is the Lebesgue measure onℝ (P.A. Zanzotto, 2002).
οƒ˜ The existence conditions of Theorem 4 are different from those of Zanzotto.
Moreover, consider the function
𝑏 π‘₯ =
βˆ’2, π‘₯ < 1
π‘₯, π‘₯ ∈ βˆ’1,1
2, π‘₯ > 1
in the case of 0 < Ξ± < 1/2 .
οƒ˜ It does not satisfy the existence conditions of Zanzotto but does satisfy the
assumptions of Theorem 4 so that there is a solution of the equation (C).
οƒ˜ Besides, Theorem 4 also improves the local integrability condition on the
coefficient b found by Zanzotto for the case of 0 < Ξ± < 1/2 in general.
οƒ˜ Contrary to the condition required by Zanzotto, Theorem 4 assumes only
𝑏
βˆ’π›Ώ ∈
πΏπ‘™π‘œπ‘ where 𝛿 = 2𝛼 < 1 for 0 < Ξ± < 1/2 .
Summary
οƒ˜ One starts with an integro-differential equation, for example:
ℒ𝑣 βˆ’ πœ†π‘£ + 𝐾 𝑣 β€² + 𝑓 = 0
where f is a given function.
οƒ˜ One proves that
𝑠𝑒𝑝 𝑣 π‘₯ ≀ 𝑁 𝑓
οƒ˜ One derives a Krylov type estimate
𝑑
𝑬 𝑒 βˆ’πœ‘π‘  πœ“π‘  𝑓 𝑠, 𝑋𝑠
0
𝑝
𝑑𝑠 ≀ 𝑁 𝑓
𝑝
where process X is related to the integro-differential equation.
οƒ˜ The estimates can be used to prove the existence of solutions of related SDEs
with measurable coefficients.