Analytical properties of scale functions for spectrally negative Lévy processes

Transcription

1 Analytical properties of scale functions for spectrally negative Lévy processes Andreas E. Kyprianou 1 Department of Mathematical Sciences, University of Bath 1 Joint work with Terence Chan and Mladen Savov 1/ 11

2 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. 2/ 11

3 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. Use P x to denote the law of X when X 0 = x > 0 with associated expectation operator E x. 2/ 11

4 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. Use P x to denote the law of X when X 0 = x > 0 with associated expectation operator E x. Let us define the Laplace exponent ψ on [0, ) by E 0(e βxt ) = e ψ(β)t. 2/ 11

5 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. Use P x to denote the law of X when X 0 = x > 0 with associated expectation operator E x. Let us define the Laplace exponent ψ on [0, ) by E 0(e βxt ) = e ψ(β)t. Then for q 0 we may define the q-scale function W (q) : R [0, ) by W (q) (x) = 0 for x < 0 and on (0, ) it is the unique right continuous function such that for β > Φ(q) Z 0 e βx W (q) (x)dx = 1 ψ(β) q. 2/ 11

6 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. Use P x to denote the law of X when X 0 = x > 0 with associated expectation operator E x. Let us define the Laplace exponent ψ on [0, ) by E 0(e βxt ) = e ψ(β)t. Then for q 0 we may define the q-scale function W (q) : R [0, ) by W (q) (x) = 0 for x < 0 and on (0, ) it is the unique right continuous function such that for β > Φ(q) Z Notation! W := W (q). 0 e βx W (q) (x)dx = 1 ψ(β) q. 2/ 11

7 Spectrally negative Lévy processes and scale functions Suppose that X = {X t : t 0} is a general spectrally negative Lévy process. Use P x to denote the law of X when X 0 = x > 0 with associated expectation operator E x. Let us define the Laplace exponent ψ on [0, ) by E 0(e βxt ) = e ψ(β)t. Then for q 0 we may define the q-scale function W (q) : R [0, ) by W (q) (x) = 0 for x < 0 and on (0, ) it is the unique right continuous function such that for β > Φ(q) Z Notation! W := W (q). 0 e βx W (q) (x)dx = 1 ψ(β) q. Note, formally we need to prove that scale functions exist - they do! 2/ 11

8 Why are scale functions important? They appear in virtually all fluctuation identities for spectrally negative Lévy processes. 3/ 11

9 Why are scale functions important? They appear in virtually all fluctuation identities for spectrally negative Lévy processes. For example, resolvent in a strip: for any a > 0, x, y [0, a], q 0 Z 0 where e qt P x (X t dy, t < τ a + τ 0 )dt j W (q) (x)w (q) (a y) = W (q) (a) ff W (q) (x y) dy. τ + a = inf{t > 0 : X t > a} and τ 0 = inf{t > 0 : X t < 0}. 3/ 11

10 Why are scale functions important? They appear in virtually all fluctuation identities for spectrally negative Lévy processes. For example, resolvent in a strip: for any a > 0, x, y [0, a], q 0 Z 0 where e qt P x (X t dy, t < τ a + τ 0 )dt j W (q) (x)w (q) (a y) = W (q) (a) ff W (q) (x y) dy. τ + a = inf{t > 0 : X t > a} and τ 0 = inf{t > 0 : X t < 0}. Or another example is: for a > 0, x [0, a], E x (e qτ 0 1 {τ 0 <τ + a } ) = Z (q) (x) W (q) (x) Z (q) (a) W (q) (a) where Z x Z (q) (x) = 1 + q W (q) (y)dy. 0 3/ 11

11 A basis of fundamental solutions to the generator equation 4/ 11

12 A basis of fundamental solutions to the generator equation It is not difficult to check that for t 0 and a (0, ] e q(t τ 0 τ + a ) W (q) (X t τ 0 τ + a are martingales. ) and e q(t τ0 τ a + ) Z (q) (X t τ 0 τ a + ) 4/ 11

13 A basis of fundamental solutions to the generator equation It is not difficult to check that for t 0 and a (0, ] e q(t τ 0 τ + a ) W (q) (X t τ 0 τ + a are martingales. ) and e q(t τ0 τ a + ) Z (q) (X t τ 0 τ a + ) Suppose that Γ is the infinitessimal generator of X. Then another way of expressing these martingale properties is by writing (in a loose sense) on (0, a) (Γ q)w (q) (x) = 0 and (Γ q)z (q) (x) = 0 4/ 11

14 A basis of fundamental solutions to the generator equation It is not difficult to check that for t 0 and a (0, ] e q(t τ 0 τ + a ) W (q) (X t τ 0 τ + a are martingales. ) and e q(t τ0 τ a + ) Z (q) (X t τ 0 τ a + ) Suppose that Γ is the infinitessimal generator of X. Then another way of expressing these martingale properties is by writing (in a loose sense) on (0, a) (Γ q)w (q) (x) = 0 and (Γ q)z (q) (x) = 0 As many known fluctuation identites turn out to be linear combinations of W (q) and Z (q), this suggest that potentially a small theory would be possible in which these functions play the role of a fundamental basis of solutions to the equation (Γ q)u(x) = 0 on (0, a). 4/ 11

15 A basis of fundamental solutions to the generator equation It is not difficult to check that for t 0 and a (0, ] e q(t τ 0 τ + a ) W (q) (X t τ 0 τ + a are martingales. ) and e q(t τ0 τ a + ) Z (q) (X t τ 0 τ a + ) Suppose that Γ is the infinitessimal generator of X. Then another way of expressing these martingale properties is by writing (in a loose sense) on (0, a) (Γ q)w (q) (x) = 0 and (Γ q)z (q) (x) = 0 As many known fluctuation identites turn out to be linear combinations of W (q) and Z (q), this suggest that potentially a small theory would be possible in which these functions play the role of a fundamental basis of solutions to the equation (Γ q)u(x) = 0 on (0, a). Before even pursuing that objective, one needs to ask whether (Γ q)w (q) (x) makes sense mathematically. 4/ 11

16 How smooth is W (q)? Where to start looking? 5/ 11

17 How smooth is W (q)? Where to start looking? Enough to answer the question for q = 0 and E(X 1) = ψ (0+) 0 as all other cases can be reduced to this case via the relation W (q) (x) = e Φ(q)x W Φ(q) (x) where W Φ(q) plays the role of W after an exponential change of measure under which X drifts to +. 5/ 11

18 How smooth is W (q)? Where to start looking? Enough to answer the question for q = 0 and E(X 1) = ψ (0+) 0 as all other cases can be reduced to this case via the relation W (q) (x) = e Φ(q)x W Φ(q) (x) where W Φ(q) plays the role of W after an exponential change of measure under which X drifts to +. Two key theories that will help analyse the issue of smoothness: 5/ 11

19 How smooth is W (q)? Where to start looking? Enough to answer the question for q = 0 and E(X 1) = ψ (0+) 0 as all other cases can be reduced to this case via the relation W (q) (x) = e Φ(q)x W Φ(q) (x) where W Φ(q) plays the role of W after an exponential change of measure under which X drifts to +. Two key theories that will help analyse the issue of smoothness: Excursion theory 5/ 11

20 How smooth is W (q)? Where to start looking? Enough to answer the question for q = 0 and E(X 1) = ψ (0+) 0 as all other cases can be reduced to this case via the relation W (q) (x) = e Φ(q)x W Φ(q) (x) where W Φ(q) plays the role of W after an exponential change of measure under which X drifts to +. Two key theories that will help analyse the issue of smoothness: Excursion theory Renewal theory 5/ 11

21 The connection with excursion theory 6/ 11

22 The connection with excursion theory It is known that j Z a ff W (x) = W (a) exp ν(ɛ > t)dt x where ν( ) is the intensity measure associated with the Poisson point process of excursions and ɛ is the canonical excursion and ɛ is its maximum value. 6/ 11

23 The connection with excursion theory It is known that j Z a ff W (x) = W (a) exp ν(ɛ > t)dt x where ν( ) is the intensity measure associated with the Poisson point process of excursions and ɛ is the canonical excursion and ɛ is its maximum value. Hence W (x+) W (x ) = ν(ɛ = x) (which immediately implies that for unbounded variation processes W C 1 (0, )). 6/ 11

24 The connection with excursion theory 7/ 11

25 The connection with excursion theory For processes with bounded variation paths (then necessarily X t = δt S t where δ > 0 and S is a driftless subordinator): ν(ɛ > t) = 1 δ Π(, t) + 1 Z «W (t + z) Π(dz) 1 δ W (t) [ t,0] showing W C 1 (0, ) Π has no atoms. 7/ 11

26 The connection with excursion theory For processes with bounded variation paths (then necessarily X t = δt S t where δ > 0 and S is a driftless subordinator): ν(ɛ > t) = 1 δ Π(, t) + 1 Z «W (t + z) Π(dz) 1 δ W (t) [ t,0] showing W C 1 (0, ) Π has no atoms. In the presence of a Gaussian component (σ 0): «ν(ɛ passes x continuously) = σ2 W (x) 2 2 W (x) W (x) showing that W C 2 (0, ). 7/ 11

27 The connection with renewal theory 8/ 11

28 The connection with renewal theory An important observation which helps us understand what kind of answer we might expect comes from the Wiener-Hopf factorization: ψ(β) = βφ(β) where φ is the Laplace exponent of the descending ladder height process and hence an integration by parts gives Z e βx W (dx) = 1 φ(β). [0, ) 8/ 11

29 The connection with renewal theory An important observation which helps us understand what kind of answer we might expect comes from the Wiener-Hopf factorization: ψ(β) = βφ(β) where φ is the Laplace exponent of the descending ladder height process and hence an integration by parts gives Z e βx W (dx) = 1 φ(β). [0, ) This means that W can be seen as the renewal function of a (killed) subordinator, say H, which has Laplace exponent φ, from which it can be calculated that its jump measure is given by Π(, x)dx, its drift is given σ 2 /2 and its killing rate E(X 1) > 0. Here, renewal function means W (dx) = Z 0 P(H t dx)dt. 8/ 11

30 The connection with renewal theory 9/ 11

31 The connection with renewal theory Then appealing to the formula 1 = P(H τ + x = x) + P(H τ + x > x) using Kesten s classical result for the probability of continuous crossing for a subordinator and Kesten-Horowitz-Bertoin formula for overshoots of subordinators one derrives that 1 = σ2 2 W (x) + Z x where Π(x) = R Π(, y)dy. x 0 W (x y){π(y) + E(X 1)}dy 9/ 11

32 The connection with renewal theory Then appealing to the formula 1 = P(H τ + x = x) + P(H τ + x > x) using Kesten s classical result for the probability of continuous crossing for a subordinator and Kesten-Horowitz-Bertoin formula for overshoots of subordinators one derrives that 1 = σ2 2 W (x) + Z x 0 W (x y){π(y) + E(X 1)}dy where Π(x) = R x Π(, y)dy. This can be seen as a renewal equation of the form f = 1 + f g for appropriate f and g. But only when σ 0, otherwise it takes the form f = f g. 9/ 11

33 The connection with renewal theory Then appealing to the formula 1 = P(H τ + x = x) + P(H τ + x > x) using Kesten s classical result for the probability of continuous crossing for a subordinator and Kesten-Horowitz-Bertoin formula for overshoots of subordinators one derrives that 1 = σ2 2 W (x) + Z x 0 W (x y){π(y) + E(X 1)}dy where Π(x) = R Π(, y)dy. x This can be seen as a renewal equation of the form f = 1 + f g for appropriate f and g. But only when σ 0, otherwise it takes the form f = f g. When X has bounded variation paths (X t = δt S t) it a direct inverse of the Laplace transform for W also gives us δw (x) = 1 + Z x 0 W (x y)π(, y)dy which is again a renewal function of the form f = 1 + f g. 9/ 11

34 Some more results 10/ 11

35 Some more results Suppose that σ 0 and Z inf{β 0 : x β Π(dx) < } [0, 2). x <1 Then for k = 0, 1, 2, W C k+3 (0, ) if and only if Π C k (0, ). 10/ 11

36 Some more results Suppose that σ 0 and Z inf{β 0 : x <1 x β Π(dx) < } [0, 2). Then for k = 0, 1, 2, W C k+3 (0, ) if and only if Π C k (0, ). Suppose that X has paths of bounded variation and Π has a derivative π such that π(x) Cx 1 α in the neighbourhood of the origin for some α < 1 and C > 0. Then for k = 0, 1, 2, W C k+3 (0, ) if and only if Π C k (0, ). 10/ 11

37 Doney s conjecture for scale functions For k = 0, 1, 2, 11/ 11

38 Doney s conjecture for scale functions For k = 0, 1, 2, If σ 0 (X has a Gaussian component) then W C k+3 (0, ) Π C k (0, ) 11/ 11

39 Doney s conjecture for scale functions For k = 0, 1, 2, If σ 0 (X has a Gaussian component) then W C k+3 (0, ) Π C k (0, ) If σ = 0 and R x Π(dx) = then ( 1,0) W C k+2 (0, ) Π C k (0, ) 11/ 11

40 Doney s conjecture for scale functions For k = 0, 1, 2, If σ 0 (X has a Gaussian component) then W C k+3 (0, ) Π C k (0, ) If σ = 0 and R x Π(dx) = then ( 1,0) W C k+2 (0, ) Π C k (0, ) If σ = 0 and R x Π(dx) < then ( 1,0) W C k+1 (0, ) Π C k (0, ) 11/ 11