Intertwining of the Wright-Fisher diffusion Tobiáš Hudec e-mail: [email protected] 7 January 20, 2017 1 0 2 n Institute of Information Theory and Automation of the CAS a Pod Vodárenskou věží 4 J CZ-182 08 9 1 Prague 8 Czech Republic ] R P . Abstract h at Itisknownthatthetimeuntilabirthanddeathprocessreachesacertainlevelisdistributed m asasumofindependentexponentialrandomvariables. Diaconis,MicloandSwartgaveaproba- bilistic proof of this fact by coupling the birth and death process with a pure birth process such [ thatthe twoprocessesreachthe givenlevelatthe sametime. Theircouplingis ofaspecialtype 1 called intertwining of Markov processes. We apply this technique to couple the Wright-Fisher v diffusion with reflectionat1/2 anda pure birthprocess. We show that in our coupling the time 8 1 ofabsorptionofthediffusionisa.s.equaltothetimeofexplosionofthepurebirthprocess. The 4 couplingalsoallowsus to interpretthe diffusionas being initially reluctantto getabsorbed,but 5 later getting more and more compelled to get absorbed. 0 . 1 Keywords: intertwining of Markov processes, Wright-Fisher diffusion, pure birth pro- 0 cess, time of absorption, coupling. 7 1 Classification: 60J60, 60J35, 60J27. : v i X 1 Introduction and the main result r a 1.1 Introduction It is known that the time until a birth and death process X started at the origin reaches a certain t levelisdistributedasasumofindependentexponentialvariableswhoseparameters arethenegatives of the non-zero eigenvalues of the generator of the process stopped at the given level (see Karlin [6]). Diaconis and Miclo [2] and Swart [11] gave a probabilistic proof of this fact by finding a pure birth process Y which reaches the given level at the same time as X . The technique that Diaconis, Miclo t t and Swart employ is called intertwining of Markov processes. This technique was developed by Rogers and Pitman [10], Diaconis and Fill [1], and Fill [4]. It allows them to add structure to the process X such that it is initially reluctant to be absorbed, but after each exponential time (which t corresponds to jump times of Y ) it changes its behavior to be more and more compelled to be t absorbed. Since one-dimensional diffusions can beobtained as limits of birth and death processes, it isinteresting toinvestigate whether thistechnique canbeextended tothecasethatX isadiffusion. t 1 Since the general case is too difficult, in this paper we consider the case that X is the Wright-Fisher t diffusion with reflection at 1/2, which has state-space [1/2,1] and is absorbed at 1. The generator andthesemigroup ofthisdiffusionhavetheniceproperty thatthey mappolynomials topolynomials of the same order, which simplifies our proofs. We need that the diffusion is reflected at 1/2 for technical reasons; without it, one of our proofs would not work (see Remark 8). We construct an explosive pure birth process Y such that X is absorbed at the same time as Y explodes. t t t The idea of Diaconis, Miclo and Swart can be summarized as follows. For a given transition semigroup P of a birth and death process X on {0,...,n} absorbed at n, Swart finds a transition t t semigroup Q of a pure birth process Y on {0,...,n} and a probability kernel K which satisfies t t P K = KQ (t ≥ 0). (1) t t The algebraic relation (1) is called intertwining, which gives the name to the intertwining of Markov processes. Swart builds on an earlier work of Diaconis and Miclo, who found an intertwining of the form KP = Q K. However, we focus on Swart’s construction, because our work on the Wright- t t Fisher diffusion is closer in spirit to his. He uses a result proved by Fill [4] which says that if X t and Y are Markov processes with finite state-spaces related by (1), then the two processes can be t 1 coupled (i.e. defined on the same probability space) such that P(Y = y|X ,0 ≤ u≤ t)= K(X ,y) a.s. (t ≥ 0). (2) t u t Using (2) and the fact that his kernel satisfies 1, ifx = n, K(x,n) =1 := (3) [x=n] (0, otherwise, Swart proves that X and Y can be coupled such that the times of absorption of X and Y are a.s. t t t t the same. In this paper, we derive analogue results for the case that X is the Wright-Fisher diffusion with t reflection at zero. We find an explosive pure-birth process Y on N¯:=N∪{∞} := {0,1,...,∞} and t a probability kernel K from [0,1] to N¯ satisfying the intertwining relation (1) and K(x,∞) = 1 . [x=1] We couple the two processes such that they satisfy (2) which allows us to conclude that the time of absorption of X is a.s. equal to the time of explosion of Y . But since the time of explosion of the t t pure birth process is the sum of independent exponential variables whose intensities are the birth 2 rates of Y , this gives us a new proof of the distribution of the time to absorption of X . t t 1.2 Intertwining of the Wright-Fisher diffusion Define ∂ D(G) = f ∈ C2[0,1]; f(0)= 0 ∂x (cid:26) (cid:27) and ∂2 Gf(x) = 1−x2 f(x), f ∈D(G),x ∈ [0,1]. (4) ∂x2 1Under certain conditions, Fill show(cid:0)ed that(cid:1)this coupling can be extended to countably infinite state-spaces. Fill built on earlier work of Diaconis and Fill [1], where an analogous result is provedfor processes with discrete time. 2 Justaswiththebirthanddeathprocesses,thedistributionofthetimeofabsorptionoftheWright-Fisherdiffusion haslong been known (see e.g. Kent [7]), butour proof is new. 2 In the appendix we show that G is closable and its closure generates a Feller semigroup, which we denote P . We also show that the associated Markov process, which we call the Wright-Fisher t diffusion with reflection at zero, has continuous sample paths. Note that the generator of the Wright-Fisher diffusion is usually defined as 1 ∂2 x(1−x) f(x), f ∈ C2[0,1], x ∈ [0,1] (5) 2 ∂x2 (see e.g. Liggett [8, Example 3.48]). However, if X˜ is generated by (5), then X = 2X˜ −1 is t t 2t generated by (4). (cid:12) (cid:12) (cid:12) (cid:12) Define H as the generator of an explosive pure birth process on N¯ which jumps fro(cid:12)m y to y+(cid:12) 1 with the rate λ = (2y+1)(2y+2), y ∈ N. y That is, define H as an operator from RN¯ to RN¯ by Hf(y) = λ (f(y+1)−f(y)), y ∈ N, y Hf(∞) = 0, where f is in RN¯. It is shown in the Appendix that the restriction of H to a suitable domain is the generator of a Feller semigroup on C N¯ , which we denote by Q . Define a probability kernel from t [0,1] to N¯ by (cid:0) (cid:1) 1−x2 x2y, if0≤ y < ∞, K(x,y)= (6) (1(cid:0)[x=1], (cid:1) ify = ∞. It can be shown that K maps C N¯ into C[0,1]. We claim that there is an intertwining relation: Theorem 1. We have (cid:0) (cid:1) P K = KQ , (t ≥ 0). (7) t t Using (7), we are able to couple the two processes in the spirit of [2, 4, 11]. We define the state-space S of the coupled process as the one-point compactification of [0,1]×N, where we denote the point at infinity by (1,∞). Using this notation, we can think of S as a subset of [0,1]×N¯, but keep in mind that the topology of S is not the one induced by [0,1]×N¯. Note that if the coupled process is to satisfy an analogue of (2), then it would be natural to construct it on the space (x,y) ∈ [0,1]×N¯; K(x,y) > 0 , as was done for Markov process(cid:8)es with discrete state-spaces [(cid:9)1, 4]. However, this space is not compact, soifwewanttousethetheoryofFellersemigroupswemustcompactifyit(eitherimplicitly or explicitly). It turns out that S is the right compactification. An analogous result to the following theorem was proved by Fill [4, Theorem 2] for processes with discrete state-spaces and by Diaconis and Fill [1, Theorem 2.33] for processes with discrete time and space. Theorem 2. There exists a Feller process (X ,Y ) on S such that t t E(f(Y )|X ,Y ,0 ≤ u≤ s)= (Q f)(Y ) a.s. (8) s+t u u t s for all f ∈ C N¯ and s,t ≥ 0. Hence, Y on its own is a pure birth process on N¯ with birth rates λ . t y If the initial distribution satisfies (cid:0) (cid:1) π(X,Y)(A×{y}) = K(x,y)πX(dy), (9) 0 ˆ 0 A 3 where πX is an arbitrary probability measure on [0,1], then X on its own is the Wright-Fisher 0 t diffusion with reflection at zero with initial distribution πX and we have 0 P(Y = y|X ,0 ≤ s≤ t) = K(X ,y) a.s. (10) t s t for all y ∈ N¯ and t ≥ 0. Note that if both X and Y start from zero, then (9) is satisfied. Using Theorem 2 we can prove t t that the time of absorption of the diffusion is a.s. equal to the time of explosion of the pure birth process. Indeed, from (10) we have P(X ∈ A,Y ∈ B)= E(1 (X )K(X ,B)). t t A t t But since K(x,·) is concentrated on N for x < 1 and on {∞} for x = 1, we have P(X < 1,Y = ∞) = P(X = 1,Y < ∞)= 0. t t t t 1.3 Discussion In addition to proving the a.s. equality of the time of absorption and the time of explosion of the two processes, Theorem 2 shows more about the underlying structure. By inspection of the formula for the coupled generator (17) below we see that conditionally on Y = y ∈ N we can interpret X t t as the Wright-Fisher diffusion with reflection at zero and with additional drift, which at the point X = x equals t 4y −4(y+1)x. x It can be shown that the scale function u(x) and the speed measure m(dx) of this diffusion satisfy 1 u′(x) = , x4y(1−x2)2 m(dx) = x4y 1−x2 dx. (cid:0) (cid:1) Hence, by Mandl [9, pp. 24–25], both boundaries are entrance for y > 0 and 0 is a regular boundary while 1 is an entrance boundary for y = 0. In particular, the coupled process lives on the set (x,y) ∈ [0,1]×N¯; K(x,y) > 0 as one might expect. Moreover, we can see that there is an equilibrium point (cid:8) (cid:9) y x = y y+1 r such that the drift is positive when x < x and negative when x > x . We can interpret this as, y y conditionally on Y = y, X is pushed toward the equilibrium point x . Obviously, x is monotonous t t y y iny andgoesfrom0to1asy goes from0to∞. Thus, we canthink ofX asbeinginitially reluctant t to be absorbed, but later getting more and more compelled to get absorbed. Inour paperwe construct Markov processes fromgenerators using the Hille-Yosida theorem. We could also construct them as solutions to Martingale problems or stochastic differential equations. However, we chose theHille-Yosida theorem forits simplicity. Theorem 1extends results of Diaconis andMiclo [2] and Swart [11], whoproved similar theorems for birth and death processes. Theorem 2 extends results of Fill [4], who proved similar result for Markov processes with continuous time and discrete state-space, and of Diaconis and Fill [1], who proved it for the case of discrete time and space. It remains an open problem whether results like Theorems 1 and 2 hold for other diffusions 4 than the modified version of the Wright-Fisher diffusion we consider in our paper. It seems that our proof of Theorem 1 does not exploit any peculiarity of the Wright-Fisher diffusion and we believe it could be extended to other types of diffusions as well. On the other hand, our proof of Theorem 2 depends strongly on the fact that the generator of the Wright-Fisher diffusion maps polynomials to polynomials of the same order, and it seems that entirely different proof techniques would be required for other diffusions. 2 Proofs 2.1 Intertwining To prove Theorem 1 we need to show that there is an intertwining between semigroups P and Q . t t The following theorem says that we can show this by proving that there is an intertwining between the generators. An analogous result for Markov processes with discrete state-spaces was proved by Fill[4,Lemma3]. Althoughweusethefollowingtheoremonly whenP andQ areFellersemigroups t t and K is a probability kernel, we are able to prove it more generally. Theorem 3. Let L ,L be Banach spaces. Let P and Q be strongly continuous contraction semi- 1 2 t t groups defined on L ,L and let G and H be their generators. Let K : L → L be a continuous 1 2 2 1 linear operator. Then the following are equivalent: 1. For all t ≥ 0, P K =KQ (11) t t on L , 2 2. K maps D(H) into D(G) and GK = KH (12) on D(H), 3. There exists a core D of H (i.e. D is a dense subspace of D(H) such that the closure of the restriction of H to D is H) such that K maps D into D(G) and (12) holds on D. Proof. To prove (1) ⇒ (2), fix f ∈ D(H). Then 1 (Q f −f) converges to Hf, so by the continuity t t of K, 1(KQ f −Kf) converges to KHf. By (11), 1(P Kf −Kf) is also convergent, so Kf is in t t t t D(G) and GKf = KHf. In order to prove (2)⇒ (1), fix f ∈ D(H) and define u(t) = KQ f. Since Q f is in D(H) by [3, t t Proposition 1.1.5], u(t) ∈ D(G) for all t ≥ 0. By the continuity of K d d u(t) = K Q f = KHQ f = GKQ f = Gu(t). t t t dt dt Since d u(t) = KQ Hf, t dt Gu(t) = du(t) is continuous as a function of t. By Proposition 1.3.4 in Ethier and Kurtz [3], dt u(t) = P u(0) = P Kf which proves that (11) holds on D(H). Since all operators involved in (11) t t are continuous, the assertion now follows from the density of D(H) in L . 2 The implication (2) ⇒ (3) is trivial by taking D = D(H). To prove the converse, let f be in D(H). Then there exist f ∈ D such that f → f and Hf → Hf. Since K is continuous, n n n Kf → Kf and GKf = KHf → KHf, where we have used (12) for f . Since G is a closed n n n n operator, Kf is in D(G) and GKf = KHf. 5 Theorem 3 shows that it suffices to prove (12) on a core of H. In the Appendix it is shown that D = f ∈ C N¯ ; ∃y ∈ Ns.t.∀y > y f(y)= f(∞) (13) H 0 0 is a core of H. The following(cid:8)theorem(cid:0) v(cid:1)erifies condition 3 of Theorem 3. (cid:9) Theorem 4. K maps D into D(G) and H GK = KH on D . H Proof. Fix f in D and let y ∈ N be such that f(y) =f(∞) for all y ≥ y . Then for x ∈ [0,1), 0 0 y0−1 Kf(x) = 1−x2 x2yf(y)+x2y0f (y ) 0 y=0 (cid:0) (cid:1) X y0 = f(0)+ x2y(f(y)−f(y−1)). (14) y=1 X As x approaches 1, Kf(x) approaches f(y ). Now 0 Kf(1) = f(∞) = f(y ). 0 Hence (14) holds also for x = 1, and therefore Kf is in C∞[0,1] ⊆ C2(G). Moreover y0 ∂ Kf(x)= 2yx2y−1(f(y)−f(y−1)), ∂x y=1 X hence ∂ Kf(0)= 0. ∂x We have shown that Kf is in D(G). From (14) we have that for x ∈[0,1], y0 GKf(x) = 1−x2 2y(2y −1)x2y−2(f(y+1)−f(y)) y=1 (cid:0) (cid:1)X y0−1 = 1−x2 λ x2y(f(y+1)−f(y)). y y=0 (cid:0) (cid:1) X Now for y < y 0 Hf(y) =λ (f(y+1)−f(y)), y and for y ≥y , 0 Hf(y)= 0, hence for x ∈ [−1,1], y0−1 KHf(x)= 1−x2 λ x2y(f(y+1)−f(y)). y y=0 (cid:0) (cid:1) X 6 Proof of Theorem 1. We use Theorem 3. In the present context, L =C[0,1] and L = C N¯ . Thus 1 2 weneed to showthat K maps C N¯ to C[0,1]. This isequivalent tosaying thatthe measures K(x,·) (cid:0) (cid:1) are continuous in x with respect to the weak convergence. But this is easy to prove, since K(x,·) is (cid:0) (cid:1) geometric distribution with success parameter 1−x2 if x < 1, and it is the degenerate distribution δ if x = 1. Theorem 4 verifies condition 3 of Theorem 3. Theorem 3 shows that this is equivalent ∞ to condition 1, and this is what we had to prove. 2.2 Coupling Inorder to find a coupling of Theorem 2, recall that S is the one-point compactification of [0,1]×N, where (1,∞) denotes the point at infinity. It is easy to see that f : S → R is continuous if and only if f(·,y) is continuous for all y ∈ N and f(x,y) → f(1,∞) as y → ∞, uniformly in x. It is also easy to see that {f ∈ C(S);f(x,y) = f(1,∞)forallx ∈[0,1]andy > y forsomey ∈ N} 0 0 is dense in C(S). Since even polynomials are dense in C[0,1] by the Stone-Weierstrass theorem, it follows that D(G)={f ∈ C(S);∃y ∈ Ns.t.f(x,y)= f(1,∞)forallx ∈ [0,1]andy > y , 0 0 (15) f(·,y)isanevenpolynomialforally ≤ y } 0 is dense in C(S). We now define an operator G with domain D(G) which we later prove generates a Feller process satisfying Theorem 2. For the motivation of this definition, see section 2.3. For f ∈ D(G) and (x,y) ∈ (0,1)×N define (Gf (·,y)K(·,y))(x)−f(x,y)(GK(·,y))(x) Gf(x,y) = (Hf(x,·))(y)+ . (16) K(x,y) Here (Gf (·,y)K(·,y)) denotes the application of the operator G to the product of f(·,y) and K(·,y), and (GK(·,y)) is the application of the operator G to K(·,y). In both cases, y is held fixed, so f(·,y) and K(·,y) are viewed as functions of x only. (Hf(x,·)) is interpreted similarly, but here x is held fixed. Note that f(·,y)K(·,y) and K(·,y) are even polynomials, hence they are in D(G). Moreover, f(x,·) is in D of (13), which is a core of H as shown in the Appendix, hence f(x,·) is H in D(H). Finally, K(x,y) > 0 since x is in (0,1) and y < ∞. Therefore, all the expressions in (16) are well defined. After plugging in the definitions of H and G, we can get an explicit formula for G. Lemma 5. Let f be in D(G) and (x,y) ∈ (0,1)×N. Let Gf be defined by (16). Then ∂2 y ∂ Gf(x,y)= λ (f(x,y+1)−f(x,y))+ 1−x2 f(x,y)+4 −(y+1)x f(x,y). (17) y ∂x2 x ∂x h i (cid:0) (cid:1) Proof. Observe that for (x,y) ∈ (0,1)×N, ∂2 (Gf (·,y)K(·,y))(x) = 1−x2 f(x,y) K(x,y) ∂x2 (cid:18) (cid:19) (cid:0) (cid:1) ∂ ∂ +2 1−x2 f(x,y) K(x,y) ∂x ∂x (cid:0) (cid:1) ∂2 + 1−x2 f(x,y) K(x,y). ∂x2 (cid:0) (cid:1) 7 Hence (Gf (·,y)K(·,y))(x)−f(x,y)(GK(·,y))(x) K(x,k) ∂2 ∂ K(x,y) ∂ = 1−x2 f(x,y)+2 1−x2 ∂x f(x,y). ∂x2 K(x,y) ∂x (cid:0) (cid:1) (cid:0) (cid:1) Noting that ∂ K(x,y) 2yx2y−1−(2y+2)x2y+1 2 1−x2 ∂x = 2 K(x,y) y2k (cid:0) (cid:1) 4y = −4(y+1)x, x we get (Gf (·,y)K(·,y))(x)−f(x,y)(GK(·,y))(x) K(x,k) ∂2 y ∂ = 1−x2 f(x,y)+4 −(y+1)x f(x,y). (18) ∂x2 x ∂x h i (cid:0) (cid:1) Plugging (18) and the definition of H into (16), we get (17). Formula (17) is well defined even for x = 1. Moreover, since f(x,y) is an even polynomial in x, it follows that y ∂ lim f(x,y) x↓0 x∂x exists, hence we can define Gf on [0,1]×N by taking the limit. Observe that for y > y (where y 0 0 is as in (15)), Gf(x,y)= 0. Therefore, if we define Gf(1,∞) = 0, then Gf is in D(G) ⊆ C(S) and we can view G :D(G)→ C(S) as a linear operator. Theorem 6. Operator G is closable and its closure generates a Feller semigroup. In order to prove Theorem 6, we use the following corollary to the Hille-Yosida theorem. Proposition 7. Let E be a compact metric space, D(G) a subspace of C(E), and G: D(G) → C(E) a linear operator. Suppose that 1 is in D(G) and G1 = 0, G satisfies the positive maximum principle, and there exist a sequence (L ) of finite-dimensional subspaces of D(G) such that L is n n∈N n∈N n dense in C(E) and G :L → L . Then G is closable and its closure generates a Feller semigroup. n n S Proof. Lemma 4.2.1 in Ethier and Kurtz [3] shows that G is dissipative, and Proposition 1.3.5 in [3] thenproves that G isclosable and its closure generates astrongly continuous contraction semigroup. Finally, the fact that G1 = 0 proves that G is conservative. Proof of Theorem 6. Let us first prove that G satisfies the positive maximum principle. Let f ∈ D and (x ,y ) ∈ S be such that 0 0 sup f(x,y)= f (x ,y )≥ 0. 0 0 (x,y)∈S Firstify = ∞,thenGf(z ) = 0bydefinition. Second,letusassumethat(x ,y ) ∈[0,1]×N. Then 0 0 0 0 we have f(x ,y +1)−f(x ,y ) ≤ 0. If x ∈ (0,1), then ∂ f(x ,y ) = 0 and ∂2 f(x ,y ) ≤ 0, 0 0 0 0 0 ∂x 0 0 ∂x2 0 0 hence Gf (x ,y )≤ 0. If x = 1, then 0 0 0 y ∂ 0 4 −(y +1)x f(x ,y )≤ 0 0 0 0 0 x ∂x (cid:20) 0 (cid:21) 8 and ∂2 1−x2 f(x ,y ) = 0, 0 ∂x2 0 0 so Gf(x ,y ) ≤ 0. And if x = 0, t(cid:0)hen the(cid:1) second-order term of the polynomial f(·,y ) must be 0 0 0 0 non-positive, for otherwise (0,y ) could not be a point of maximum. Hence 0 y ∂ 0 lim f(x,y )≤ 0 0 x↓0 x ∂x and ∂2 1−x2 f(x ,y ) ≤ 0, 0 ∂x2 0 0 so we again get that Gf(x ,y ) ≤(cid:0) 0. W(cid:1)e have shown that G satisfies the positive maximum 0 0 principle. Define L = {f ∈ C(S); f(·,y)isanevenpolynomialof degreeatmost2nforally ≤ n, n f(·,y) =f(1,∞)forally > n}. It is easy to see that G : L → L and L = D(G) is dense in C[0,1]. Finally, G1 = 0, so by n n n∈N n Proposition 7, G is closable and its closure generates a Feller semigroup. S Remark 8. The proof of Theorem 6 is the only place where our argument fails for the Wright-Fisher diffusion without reflection at zero. Indeed, we could take P to be the semigroup of the diffusion t on the whole interval [−1,1], that is, P would be generated by t ∂2 Gf(x)= 1−x2 f(x), f ∈ C2[−1,1],x ∈ [−1,1]. (19) ∂x2 We could now extend kernel K(cid:0) to be(cid:1)from [−1,1] to N¯ using the same formula (6). Our proof of Theorem 1 would still work. We could define G by (16) where G would now be defined by (19). But now we could not take D(G) to be functions such that f(·,y) are even polynomials, because they are not dense in C[−1,1]. But if we allowed all polynomials, then for y > 0 we could not extend (Gf)(·,y) to a continuous function on [−1,1] because of the term y ∂ f(x,y) in (17). The deeper x∂x reason for this problem is that K(0,y) = 0 for y > 0, so if the process (X ,Y ) satisfies (10), then t t after Y departs from zero, X is no longer allowed to cross zero, so the behavior of the diffusion on t t [−1,0] and [0,1] are independent. To overcome this problem, we could define S = [−1,1]×{0}∪ −1,0− ×{1,2,...}∪ 0+,1 ×{1,2...}∪{−1,1}×{∞}, where we think of 0− and 0+(cid:2) as two(cid:3) different point(cid:2)s. No(cid:3)w we could take D(G) to be functions such that f(·,0) is a polynomial, f(·,y) is an even polynomial with possibly different coefficients on [−1,0−] and on [0+,1] and from some y , f(x,y) equals either f(−1,∞) or f(1,∞) depending on 0 whether x is in [−1,0−] or [0+,1]. This set is dense in C(S). Then, however, (Gf)(·,0) could be discontinuous at x = 0 because of the term λ (f(x,y+1)−f(x,y)) in (17). To get around this y problem, we decided to work with the Wright-Fisher diffusion with reflection at zero. In order to prove Theorem 2, we need the following theorem due to Rogers and Pittman [10]. Theorem 9. Let (S,SSS) and (S,S) be measurable spaces and let φ : S → S be a measurable transformation. Let Λ be a probability kernel from S to S and define a probability kernel from S to S by Φf = f ◦φ. 9 Let X be a continuous-time Markov process with state space (S,SSS), transition semigroup P and t t initial distribution π = π Λ, for some distribution π on S. Suppose further: 0 0 0 1. ΛΦ = I, the identity kernel on S, 2. for each t ≥ 0 the probability kernel P := ΛP Φ from S to S satisfies t t ΛP = P Λ. (20) t t Then P is a transition semigroup on S, φ◦ X is Markov with transition semigroup P and the t t t initial distribution π and 0 P(X ∈ A|φ◦X ,0 ≤ s≤ t) = Λ(φ◦X ,A) t s t a.s. for all t ≥ 0 and A∈SSS. Proof. Rogers and Pittman [10, Theorem 2] proved this for the case that π = δ for some y ∈ S. 0 y The general case follows by integration with respect to π . 0 Proof of Theorem 2. It is intuitively clear from the form of G that Y on its own is generated by H, t but here we give a short formal proof. Note that for f ∈ C N¯ and s,t ≥ 0, E(f(Y )|X ,Y ,0 ≤ u≤ s) =(P(cid:0) Ψ(cid:1)f)(X ,Y ), s+t u u t s s where P is the semigroup generated by G and Ψ is a kernel given by Ψf = f◦ψ where ψ(x,y) =y. t Hence, in order to prove (8) we need to show that P Ψf = ΨQ f (21) t t for all f ∈ C N¯ . By Theorem 3 it suffices to prove that there exists a core D of H such that Ψ H maps D into D(G) and H (cid:0) (cid:1) GΨ= ΨH (22) on D . It is shown in Lemma 12 that H D = f ∈ C N¯ ; ∃y s.t.f(y)= f(∞)forally > y H 0 0 is a core of H and it is easy to(cid:8)see th(cid:0)at(cid:1)for f ∈ D , Ψf is in D(G). Moreov(cid:9)er, for f ∈ D we have H H (Gf(y)K(·,y))(x) = f(y)(GK(·,y))(x), so from (16) it is easy to see that (22) holds. In order to prove the claims about X , we will use Theorem 9. In the present setting, S = [0,1] t and φ(x,y) = x. Define a probability kernel from [0,1] to S by Λ(x,A×{y}) = δ (A)K(x,y) x Λ(x,(1,∞)) = K(x,∞) where x is in [0,1], A is in B[0,1] and y ∈ N. In other words, Λf(x) = K(x,y)f(x,y) (23) 0≤y≤∞ X for f ∈ C(S) and x ∈ [0,1]. Observe that πXΛ = π(X,Y). Also observe that for f ∈ C[0,1] and 0 0 x ∈ [0,1] we have ΛΦf(x)= K(x,y)f(x) = f(x), 0≤y≤∞ X 10