Introduction
The large-time behaviour of semilinear equations with absorption is governed by the competition between spatial dispersal and nonlinear dissipation. Consider
∂ t v = L v − v 1 + β , β > 0 , \partial_t v = \mathcal{L}v - v^{1+\beta}, \qquad\beta> 0, ∂ t v = L v − v 1 + β , β > 0 ,
where L \mathcal{L} L generates a positive linear semigroup preserving the underlying reference measure. For nonnegative solutions with finite initial mass, absorption makes the total mass nonincreasing. The main analytic question is whether this mass tends to zero or converges to a strictly positive limit. We address this question through decay estimates for the linear semigroup, obtaining quantitative lower bounds for the remaining mass. The applications include stable, elliptic and Bessel-type generators, as well as time-dependent diffusion rates; a hyperbolic diffusion gives a positive limiting mass for every β > 0 \beta> 0 β > 0 .
The connection with branching populations can be expressed through their Laplace functionals. For exponential lifetimes and a homogeneous Poisson initial field, these functionals are determined by the spatial integrals of solutions to semilinear equations with absorption. This representation relates the long-time behaviour of the remaining mass to persistence and local extinction of the population.
Critical spatial branching systems exhibit a competition between spatial dispersal and genealogical clustering. In the classical ( d , α , β ) (d,\alpha,\beta) ( d , α , β ) model, particles move as isotropic α \alpha α -stable processes on R d \mathbb{R}^d R d , where d ∈ N d \in\mathbb{N} d ∈ N and 0 < α ≤ 2 0 < \alpha\le2 0 < α ≤ 2 , and reproduce at death with offspring generating function f ( s ) = s + c ( 1 − s ) 1 + β f(s) = s + c(1-s)^{1+\beta} f ( s ) = s + c ( 1 − s ) 1 + β , 0 ≤ s ≤ 1 0 \le s \le1 0 ≤ s ≤ 1 , where 0 < β ≤ 1 0 < \beta\le1 0 < β ≤ 1 and 0 < c < ( 1 + β ) − 1 0 < c < (1+\beta)^{-1} 0 < c < ( 1 + β ) − 1 . Particles evolve independently, with motion, lifetime, and offspring number independent of one another; children are born at the parent’s death position. For exponential lifetimes and an initial Poisson field of newborn ancestors with Lebesgue intensity, the critical relation between persistence and local extinction is d = α / β d = \alpha/\beta d = α / β ; see Gorostiza and Wakolbinger [1 ] .
This probabilistic question has a direct formulation in terms of a nonlinear diffusion equation. To describe it, let Z t Z_t Z t be the counting measure of descendants alive at time t t t from one newborn ancestor at x x x , and let E x \mathbb{E}_x E x denote expectation for this family. For φ ∈ C c + ( R d ) \varphi\in C_c^+(\mathbb{R}^d) φ ∈ C c + ( R d ) , the space of nonnegative continuous functions with compact support, set
a φ ( x ) : = 1 − e − φ ( x ) , v φ ( t , x ) : = E x [ 1 − e − ⟨ Z t , φ ⟩ ] , ⟨ Z t , φ ⟩ : = ∫ φ d Z t . a_\varphi(x) := 1-e^{-\varphi(x)}, \qquad v_\varphi(t,x) := \mathbb{E}_x\left[1-e^{-\langle Z_t,\varphi\rangle}\right], \qquad\langle Z_t,\varphi\rangle:= \int\varphi\,\mathrm{d}Z_t. a φ ( x ) := 1 − e − φ ( x ) , v φ ( t , x ) := E x [ 1 − e − ⟨ Z t , φ ⟩ ] , ⟨ Z t , φ ⟩ := ∫ φ d Z t .
The function v φ ( t , x ) v_\varphi(t,x) v φ ( t , x ) is denoted by V t φ ( x ) V_t^\varphi(x) V t φ ( x ) in Section 4. Write P t P_t P t for the transition semigroup of the motion, normalized so that its generator is L α = − ( − Δ ) α / 2 L_\alpha=-( -\Delta)^{\alpha/2} L α = − ( − Δ ) α /2 , with Fourier symbol − ∣ ξ ∣ α -|\xi|^\alpha − ∣ ξ ∣ α . When lifetimes are exponential with rate θ e > 0 \theta_e>0 θ e > 0 , the evolution is governed by the fractional diffusion equation with absorption
∂ t v φ = − ( − Δ ) α / 2 v φ − c θ e v φ 1 + β , v φ ( 0 , ⋅ ) = a φ . (1) \partial_t v_\varphi=-( -\Delta)^{\alpha/2}v_\varphi-c\theta_e v_\varphi^{1+\beta}, \qquad v_\varphi(0,\cdot)=a_\varphi. \tag*{(1)} ∂ t v φ = − ( − Δ ) α /2 v φ − c θ e v φ 1 + β , v φ ( 0 , ⋅ ) = a φ . (1)
For the initial data considered here, this equation is understood in its mild form,
v φ ( t , ⋅ ) = P t a φ − c θ e ∫ 0 t P t − s [ v φ ( s , ⋅ ) 1 + β ] d s . (2) v_\varphi(t,\cdot)=P_ta_\varphi-c\theta_e\int_0^t P_{t-s}\left[v_\varphi(s,\cdot)^{1+\beta}\right]\,\mathrm{d}s. \tag*{(2)} v φ ( t , ⋅ ) = P t a φ − c θ e ∫ 0 t P t − s [ v φ ( s , ⋅ ) 1 + β ] d s . (2)
These identities are the formulation in terms of v φ v_\varphi v φ of the equations in Gorostiza and Wakolbinger [1 Eqs. (2.2)–(2.4)] , with the offspring coefficient c c c used throughout the present paper.
The relevance of the absorption term becomes explicit after spatial integration. With ∥ u ∥ 1 : = ∫ R d ∣ u ( x ) ∣ d x \|u\|_1:=\int_{\mathbb{R}^d}|u(x)|\,\mathrm{d}x ∥ u ∥ 1 := ∫ R d ∣ u ( x ) ∣ d x , define G φ ( t ) : = ∥ v φ ( t , ⋅ ) ∥ 1 G_\varphi(t):=\|v_\varphi(t,\cdot)\|_1 G φ ( t ) := ∥ v φ ( t , ⋅ ) ∥ 1 and b ( φ ) : = ∥ a φ ∥ 1 b(\varphi):=\|a_\varphi\|_1 b ( φ ) := ∥ a φ ∥ 1 . Criticality gives 0 ≤ v φ ( t , ⋅ ) ≤ P t φ 0\leq v_\varphi(t,\cdot)\leq P_t\varphi 0 ≤ v φ ( t , ⋅ ) ≤ P t φ , so invariance of Lebesgue measure implies G φ ( t ) ≤ ∥ φ ∥ 1 < ∞ G_\varphi(t)\leq\|\varphi\|_1<\infty G φ ( t ) ≤ ∥ φ ∥ 1 < ∞ . Integrating (1.2) then gives the mass balance
G φ ( t ) = b ( φ ) − c θ e ∫ 0 t ∥ v φ ( s , ⋅ ) 1 + β ∥ 1 d s . (3) G_\varphi(t)=b(\varphi)-c\theta_e\int_0^t\|v_\varphi(s,\cdot)^{1+\beta}\|_1\,\mathrm{d}s. \tag*{(3)} G φ ( t ) = b ( φ ) − c θ e ∫ 0 t ∥ v φ ( s , ⋅ ) 1 + β ∥ 1 d s . (3)
If X t X_t X t is the population obtained from independent families whose newborn ancestors form a Poisson field of Lebesgue intensity, the Poisson formula yields E e − ⟨ X t , φ ⟩ = e − G φ ( t ) \mathbb{E}e^{-\langle X_t,\varphi\rangle}=e^{-G_\varphi(t)} E e − ⟨ X t , φ ⟩ = e − G φ ( t ) . Since G φ G_\varphi G φ is nonnegative and nonincreasing, the limiting Laplace functional is determined, for all such test functions, by the remaining mass
ℓ φ : = lim t → ∞ G φ ( t ) = b ( φ ) − c θ e ∫ 0 ∞ ∥ v φ ( s , ⋅ ) 1 + β ∥ 1 d s ≥ 0. \ell_\varphi:=\lim_{t\to\infty}G_\varphi(t)=b(\varphi)-c\theta_e\int_0^\infty\|v_\varphi(s,\cdot)^{1+\beta}\|_1\,\mathrm{d}s\geq0. ℓ φ := t → ∞ lim G φ ( t ) = b ( φ ) − c θ e ∫ 0 ∞ ∥ v φ ( s , ⋅ ) 1 + β ∥ 1 d s ≥ 0.
Local extinction corresponds to ℓ φ = 0 \ell_\varphi=0 ℓ φ = 0 for every φ \varphi φ , whereas a nontrivial limiting population leaves ℓ φ > 0 \ell_\varphi>0 ℓ φ > 0 for some φ \varphi φ . The analytic issue is therefore how much mass is absorbed over time. The mere finiteness of the absorption integral does not distinguish these alternatives: it already follows from (1.3) in both regimes.
For general lifetimes, the corresponding equation retains the lifetime law explicitly. Let Γ \Gamma Γ be its distribution function, with Γ ( 0 ) = 0 \Gamma(0)=0 Γ ( 0 ) = 0 , write Γ ‾ = 1 − Γ \overline{\Gamma}=1-\Gamma Γ = 1 − Γ , and denote the associated probability measure by Γ ( d s ) \Gamma(\mathrm{d}s) Γ ( d s ) . Conditioning on the first death gives the nonlinear renewal equation
v φ ( t , ⋅ ) = Γ ‾ ( t ) P t a φ + ∫ [ 0 , t ] P s [ v φ ( t − s , ⋅ ) − c v φ ( t − s , ⋅ ) ] 1 + β Γ ( d s ) . (4) v_\varphi(t,\cdot)=\overline{\Gamma}(t)P_ta_\varphi+\int_{[0,t]}P_s\left[v_\varphi(t-s,\cdot)-cv_\varphi(t-s,\cdot)\right]^{1+\beta}\Gamma(\mathrm{d}s). \tag*{(4)} v φ ( t , ⋅ ) = Γ ( t ) P t a φ + ∫ [ 0 , t ] P s [ v φ ( t − s , ⋅ ) − c v φ ( t − s , ⋅ ) ] 1 + β Γ ( d s ) . (4)
This is Vatutin and Wakolbinger [2, Eq. (10)] in the present notation. For Γ ( t ) = 1 − e − θ e t \Gamma(t)=1-e^{-\theta_e t} Γ ( t ) = 1 − e − θ e t it reduces to (1.2); hence the fractional absorption equation is the exponential-lifetime case of this renewal formulation.
Vatutin and Wakolbinger [2 ] considered the same branching mechanism and stable motion with a general lifetime law. For a nonarithmetic law of finite mean they proved persistence for d > α / β d>\alpha/\beta d > α / β and local extinction for d ≤ α / β d\leq\alpha/\beta d ≤ α / β . For a nonarithmetic law satisfying Γ ‾ ( t ) ∼ c 1 t − γ \overline{\Gamma}(t)\sim c_1t^{-\gamma} Γ ( t ) ∼ c 1 t − γ , where c 1 > 0 c_1>0 c 1 > 0 and 0 < γ ≤ 1 0<\gamma\leq1 0 < γ ≤ 1 , the threshold becomes d = α γ / β d=\alpha\gamma/\beta d = α γ / β : above it the population converges to the Poisson field of full Lebesgue intensity, below it the population becomes locally extinct, and at equality a nontrivial subsequential limit exists. Their result at equality does not identify the limiting law.
Multitype extensions retaining symmetric stable motions include the Markovian models with exponential lifetimes studied in Gorostiza et al. [3 ] , López-Mimbela and Wakolbinger [4 ] , the extinction results for multitype Bellman–Harris systems of Kevei and López-Mimbela [5 ] , and the finite-mean and mixed finite–infinite mean persistence results of Kolkovska et al. [6 ] , Ramírez-González and Machado [7 ] . These works retain stable spatial motions and address the additional type structure.
The present paper retains this critical reproduction mechanism and replaces the stable motion by a conservative Hunt process on E = R d E=\mathbb{R}^d E = R d or E = [ 0 , ∞ ) d E=[0,\infty)^d E = [ 0 , ∞ ) d , with transition semigroup P t P_t P t and Lebesgue measure λ \lambda λ . The definitions of v φ v_\varphi v φ , G φ G_\varphi G φ , and b ( φ ) b(\varphi) b ( φ ) now refer to families on E E E , test functions φ ∈ C + ( E ) \varphi\in C_+ (E) φ ∈ C + ( E ) , and the norm in L 1 ( E , λ ) L^1(E,\lambda) L 1 ( E , λ ) . The same first-death decomposition gives (1.4) with this semigroup. If P t P_t P t preserves λ \lambda λ , spatial integration and resolution of the scalar renewal equation yield
G φ ( t ) = b ( φ ) − c ∫ [ 0 , t ] ∥ v φ ( t − s , ⋅ ) ∥ 1 1 + β U ( d s ) , (5) G_\varphi(t)=b(\varphi)-c\int_{[0,t]}\left\|v_\varphi(t-s,\cdot)\right\|_1^{1+\beta}U(\mathrm{d}s), \tag*{(5)} G φ ( t ) = b ( φ ) − c ∫ [ 0 , t ] ∥ v φ ( t − s , ⋅ ) ∥ 1 1 + β U ( d s ) , (5)
where U : = ∑ n ≥ 1 Γ ∗ n U:=\sum_{n\geq1}\Gamma^{*n} U := ∑ n ≥ 1 Γ ∗ n is the renewal measure; Γ ∗ n \Gamma^{*n} Γ ∗ n denotes the law of the sum of n n n independent lifetimes. This identity is proved in Lemma 4.4. For exponential lifetimes, U ( d s ) = θ e d s U(\mathrm{d}s)=\theta_e\,\mathrm{d}s U ( d s ) = θ e d s , recovering (1.3). In general, the Laplace functionals are therefore determined by the renewal convolution of the L 1 L^1 L 1 norm of v φ 1 + β v_\varphi^{1+\beta} v φ 1 + β . The assumptions on P t P_t P t are stated in (H1)–(H3). They require invariance of Lebesgue measure, a uniform estimate P t 1 K ≤ C K ( 1 + t ) − q P_t1_K \leq C_K(1+t)^{-q} P t 1 K ≤ C K ( 1 + t ) − q for each compact K ⊂ E K \subset E K ⊂ E , with an exponent q > 0 q>0 q > 0 and a constant C K < ∞ C_K<\infty C K < ∞ , and a localization property on regions of volume O ( t γ ) O(t^\gamma) O ( t γ ) . These are the properties of the stable semigroup that enter the renewal argument of Vatutin and Wakolbinger [2 ] . They control the nonlinear term in (5) through dispersal and localization estimates; their distinct roles and logical independence are established in § 2 .
Under these assumptions the stable exponent d / α d/\alpha d / α is replaced by q q q . For spread-out lifetimes of finite mean, the population persists when q β > 1 q\beta>1 q β > 1 and becomes locally extinct when q β ≤ 1 q\beta\leq1 q β ≤ 1 . In this regime, the spread-out assumption is stronger than the nonarithmetic hypothesis in Vatutin and Wakolbinger [2 ] ; it is used below through Stone’s decomposition of the renewal measure. For a nonarithmetic lifetime law with Γ ‾ ( t ) ∼ c 1 t − γ \overline{\Gamma}(t)\sim c_1t^{-\gamma} Γ ( t ) ∼ c 1 t − γ , the threshold is q β = γ q\beta=\gamma q β = γ . The renewal estimates needed at the endpoint γ = 1 \gamma=1 γ = 1 are included.
The semigroup hypotheses hold for several standard and nonhomogeneous Markov motions. Besides full-dimensional symmetric stable processes, they include anisotropic stable products with different indices and all their nonsingular linear images; in particular, the Brownian case allows every positive definite covariance matrix. They also hold for symmetric uniformly elliptic divergence-form diffusions. A further class is obtained by a power transformation of squared-Bessel processes, parametrized by ρ > 0 \rho>0 ρ > 0 as in § 3.3 . The resulting one-dimensional kernel is symmetric with respect to Lebesgue measure and decays locally as t − 1 / ρ t^{-1/\rho} t − 1/ ρ , so the independent d d d -fold product has exponent q = d / ρ q=d/\rho q = d / ρ . When ρ > 1 \rho>1 ρ > 1 , the symmetric extension through zero gives a conservative Bessel-type diffusion on the whole line, and nonsingular linear images yield examples on R d \mathbb{R}^d R d with dependent coordinates.
For each listed example we also determine recurrence and transience of the individual motion. The estimate (H2) gives transience whenever q > 1 q>1 q > 1 . For the symmetric classes in Table 3.1 , recurrence holds when q ≤ 1 q\leq1 q ≤ 1 , while the Brownian motion with nonzero drift considered after the table shows that this conclusion is not implied by (H1)–(H3). This comparison is separate from the branching threshold: finite-mean persistence implies transience of the individual motion, while for heavy-tailed lifetimes a Poisson limit may occur in a recurrent regime.
Table 3.1. Parameter ranges, dispersal exponents, and recurrence classifications for the Markov motions considered in § 3 . Throughout, d ∈ N d\in\mathbb{N} d ∈ N .
The paper is organized as follows. Section 2 introduces the model, states the hypotheses on the motion, and gives the extinction–survival results. Section 3 first verifies the motion hypotheses and then treats recurrence and transience. The estimates for one family and the renewal equations are proved in Section 4; the proofs of the extinction–survival results are given in Section 5. Finally, Section 6 applies the semigroup decay estimates to persistence of the total mass of semilinear equations with absorption, including equations with time-dependent diffusion rates.
Model and extinction–survival results
Fix d ∈ N d \in\mathbb{N} d ∈ N . Let E = R d E = \mathbb{R}^{d} E = R d or E = [ 0 , ∞ ) d E = [0,\infty)^{d} E = [ 0 , ∞ ) d , with Borel σ \sigma σ -field and Lebesgue measure λ \lambda λ . Let W = ( W t ) t ≥ 0 W = (W_t)_{t\geq0} W = ( W t ) t ≥ 0 be a conservative Hunt process with Markov family ( P x ) x ∈ E (\mathbb{P}_x)_{x\in E} ( P x ) x ∈ E and semigroup P t h ( x ) = E x [ h ( W t ) ] P_t h(x) = \mathbb{E}_x[h(W_t)] P t h ( x ) = E x [ h ( W t )] , h ≥ 0 h \geq0 h ≥ 0 . Each particle moves as a copy of W W W from its birth position, has lifetime distribution Γ \Gamma Γ with Γ ( 0 ) = 0 \Gamma(0) = 0 Γ ( 0 ) = 0 , and at death is replaced at its death position by ζ ∈ N 0 \zeta\in\mathbb{N}_0 ζ ∈ N 0 children. For each particle, its motion, lifetime, and offspring number are mutually independent; the corresponding marks of distinct particles are independent. Write Γ ‾ ( t ) = 1 − Γ ( t ) \overline{\Gamma}(t) = 1 - \Gamma(t) Γ ( t ) = 1 − Γ ( t ) and Γ ( d t ) \Gamma(dt) Γ ( d t ) for the corresponding probability measure. The offspring generating function is f ( s ) = E [ s ζ ] = s + c ( 1 − s ) 1 + β f(s) = \mathbb{E}[s^\zeta] = s + c(1-s)^{1+\beta} f ( s ) = E [ s ζ ] = s + c ( 1 − s ) 1 + β , where 0 < β ≤ 1 0 < \beta\leq1 0 < β ≤ 1 and 0 < c < ( 1 + β ) − 1 0 < c < (1+\beta)^{-1} 0 < c < ( 1 + β ) − 1 . This is exactly the nondegenerate critical mechanism of Vatutin and Wakolbinger [2 Eq. (1)] . Since f ( s ) > s f(s) > s f ( s ) > s for 0 ≤ s < 1 0 \leq s < 1 0 ≤ s < 1 , the associated Galton–Watson process becomes extinct after a finite generation almost surely. We use the standard spatial Bellman–Harris model of Vatutin and Wakolbinger [2 Section 2] : particles reproduce only at death, all children are born at the parent’s death position, and distinct families evolve independently.
Let Z t x Z_t^x Z t x be the counting measure of descendants alive at time t t t from one newborn ancestor at x x x . If Π \Pi Π is a Poisson random measure of intensity λ \lambda λ , attach independent one-family marks and set X t = ∑ x ∈ Π Z t x X_t = \sum_{x\in\Pi} Z_t^x X t = ∑ x ∈ Π Z t x . Let M ( E ) \mathcal{M}(E) M ( E ) be the space of nonnegative Radon measures on E E E , equipped with the vague topology. Every convergence statement for X t X_t X t below means convergence in distribution on this space. Throughout, local extinction means X t ⇒ 0 X_t \Rightarrow0 X t ⇒ 0 . In particular, P { X t ( K ) > 0 } → 0 \mathbb{P}\{X_t(K) > 0\} \to0 P { X t ( K ) > 0 } → 0 for every compact K K K : choose ψ ∈ C + c ( E ) \psi\in C_+^c(E) ψ ∈ C + c ( E ) with ψ ≥ 1 K \psi\geq1_K ψ ≥ 1 K and use { X t ( K ) > 0 } ⊂ { ⟨ X t , ψ ⟩ ≥ 1 } \{X_t(K) > 0\} \subset\{\langle X_t,\psi\rangle\geq1\} { X t ( K ) > 0 } ⊂ {⟨ X t , ψ ⟩ ≥ 1 } . Recall that Γ \Gamma Γ is spread-out if some Γ ∗ n \Gamma^{*n} Γ ∗ n has a nonzero absolutely continuous component, and nonarithmetic if it is not supported by h N 0 h\mathbb{N}_0 h N 0 for any h > 0 h > 0 h > 0 .
Hypotheses. For some q > 0 q > 0 q > 0 , the semigroup satisfies
(H1) Invariant reference measure: ∫ E P t h d λ = ∫ E h d λ \int_E P_t h\,d\lambda= \int_E h\,d\lambda ∫ E P t h d λ = ∫ E h d λ for every t ≥ 0 t \geq0 t ≥ 0 and every nonnegative Borel h h h , with either side allowed to be infinite.
(H2) Uniform local dispersal: for each compact K ⊂ E K \subset E K ⊂ E , there is 0 < C K < ∞ 0 < C_K < \infty 0 < C K < ∞ such that sup x P t 1 K ( x ) ≤ C K ( 1 + t ) − q \sup_x P_t1_K(x) \leq C_K(1+t)^{-q} sup x P t 1 K ( x ) ≤ C K ( 1 + t ) − q , t ≥ 0 t \geq0 t ≥ 0 .
(H3) Dispersal-scale localization: there are Borel sets D t ( N ) ⊂ E D_t(N) \subset E D t ( N ) ⊂ E , t , N ≥ 1 t,N \geq1 t , N ≥ 1 , and constants C N < ∞ C_N < \infty C N < ∞ such that λ ( D t ( N ) ) ≤ C N t q \lambda(D_t(N)) \leq C_N t^q λ ( D t ( N )) ≤ C N t q and, for every compact K ⊂ E K \subset E K ⊂ E , lim N → ∞ sup t ≥ 1 ∫ E ∖ D t ( N ) P t 1 K d λ = 0 \lim_{N\to\infty}\sup_{t\geq1}\int_{E\setminus D_t(N)}P_t1_K\,d\lambda= 0 lim N → ∞ sup t ≥ 1 ∫ E ∖ D t ( N ) P t 1 K d λ = 0 .
Distinct roles and logical independence of (H1)–(H3). The three semigroup hypotheses have distinct roles in the argument. Condition (H1) preserves Lebesgue intensity under the motion and permits the spatially integrated equations to be reduced to scalar renewal equations. Condition (H2) gives the uniform local decay that determines the dispersal exponent q q q , whereas (H3) ensures that, uniformly in time, the initial positions that may contribute to a fixed compact set can be confined to sets of volume of order t q t^q t q .
We next record that none of these assumptions is a consequence of the other two. Since the statement of the hypotheses requires the existence of a single exponent q > 0 q>0 q > 0 for which both (H2) and (H3) hold, the exponent must be quantified carefully. For a conservative Hunt semigroup P P P , define
Q 2 ( P ) : = { q > 0 : P satisfies (H2) with exponent q } , \mathcal{Q}_2(P):=\{q>0:P\text{ satisfies (H2) with exponent }q\}, Q 2 ( P ) := { q > 0 : P satisfies (H2) with exponent q } ,
Q 3 ( P ) : = { q > 0 : P satisfies (H3) with exponent q } . \mathcal{Q}_3(P):=\{q>0:P\text{ satisfies (H3) with exponent }q\}. Q 3 ( P ) := { q > 0 : P satisfies (H3) with exponent q } .
These sets have opposite monotonicity properties:
q 0 ∈ Q 2 ( P ) ⟹ ( 0 , q 0 ] ⊆ Q 2 ( P ) , q 0 ∈ Q 3 ( P ) ⟹ [ q 0 , ∞ ) ⊆ Q 3 ( P ) . q_0\in\mathcal{Q}_2(P)\ \Longrightarrow\ (0,q_0]\subseteq\mathcal{Q}_2(P),\qquad
q_0\in\mathcal{Q}_3(P)\ \Longrightarrow\ [q_0,\infty)\subseteq\mathcal{Q}_3(P). q 0 ∈ Q 2 ( P ) ⟹ ( 0 , q 0 ] ⊆ Q 2 ( P ) , q 0 ∈ Q 3 ( P ) ⟹ [ q 0 , ∞ ) ⊆ Q 3 ( P ) .
Indeed, the first implication follows from ( 1 + t ) − q 0 ≤ ( 1 + t ) − q (1+t)^{-q_0}\leq(1+t)^{-q} ( 1 + t ) − q 0 ≤ ( 1 + t ) − q for 0 < q ≤ q 0 0<q\leq q_0 0 < q ≤ q 0 , and the second follows by using the same localization sets and observing that t q 0 ≤ t q t^{q_0}\leq t^q t q 0 ≤ t q for t ≥ 1 t\geq1 t ≥ 1 and q ≥ q 0 q\geq q_0 q ≥ q 0 . Thus it would not be sufficient merely to exhibit one exponent for which one condition holds and the other fails. The examples below prove the stronger assertion that changing the exponent cannot restore the missing hypothesis.
(i) Conditions (H1) and (H3) do not imply (H2). On E = R d E=\mathbb{R}^d E = R d , consider the constant process W t x = x W_t^x=x W t x = x . It is deterministic, has continuous paths and infinite lifetime, and satisfies the strong Markov property. Hence it is a conservative Hunt process. Its semigroup is P t h = h P_th=h P t h = h , and therefore (H1) holds. For an arbitrary q > 0 q>0 q > 0 , let
D t ( N ) = B ( 0 , N t q / d ) , t , N ≥ 1. D_t(N)=B(0,Nt^{q/d}),\qquad t,N\geq1. D t ( N ) = B ( 0 , N t q / d ) , t , N ≥ 1.
Then
λ ( D t ( N ) ) = λ ( B ( 0 , 1 ) ) N d t q . \lambda(D_t(N))=\lambda(B(0,1))N^dt^q. λ ( D t ( N )) = λ ( B ( 0 , 1 )) N d t q .
If K ⊂ B ( 0 , R ) K\subset B(0,R) K ⊂ B ( 0 , R ) is compact, then, for every N ≥ R N\geq R N ≥ R and t ≥ 1 t\geq1 t ≥ 1 ,
∫ R d ∖ D t ( N ) P t 1 K d λ = ∫ R d ∖ D t ( N ) 1 K d λ = 0. \int_{\mathbb{R}^d\setminus D_t(N)}P_t1_K\,\mathrm{d}\lambda
=
\int_{\mathbb{R}^d\setminus D_t(N)}1_K\,\mathrm{d}\lambda
=0. ∫ R d ∖ D t ( N ) P t 1 K d λ = ∫ R d ∖ D t ( N ) 1 K d λ = 0.
Consequently, (H3) holds for every q > 0 q>0 q > 0 . On the other hand, for every nonempty compact set K K K ,
sup x ∈ R d P t 1 K ( x ) = 1 , t ≥ 0. \sup_{x\in\mathbb{R}^d}P_t1_K(x)=1,\qquad t\geq0. x ∈ R d sup P t 1 K ( x ) = 1 , t ≥ 0.
so (H2) fails for every q > 0 q>0 q > 0 . Therefore
Q 3 ( P ) = ( 0 , ∞ ) , Q 2 ( P ) = ∅ . Q_3(P)=(0,\infty),\qquad Q_2(P)=\varnothing. Q 3 ( P ) = ( 0 , ∞ ) , Q 2 ( P ) = ∅ .
(ii) Conditions (H1) and (H2) do not imply (H3). On E = R 2 E=\mathbb{R}^2 E = R 2 , let
A = ( 1 0 0 − 1 ) A=\begin{pmatrix}1&0\\0&-1\end{pmatrix} A = ( 1 0 0 − 1 )
and consider the linear diffusion
d W t = A W t d t + d B t . \mathrm{d}W_t=AW_t\,\mathrm{d}t+\mathrm{d}B_t. d W t = A W t d t + d B t .
where B B B is a standard Brownian motion in R 2 \mathbb{R}^2 R 2 . The coefficients are globally Lipschitz and have at most linear growth, so the equation has a unique nonexplosive strong solution with continuous paths and a Feller transition semigroup. Hence W W W is a conservative continuous Feller process, and therefore a Hunt diffusion. Starting from x x x , it has the representation
W t x = e t A x + Z t , Z t : = ∫ 0 t e ( t − s ) A d B s , W_t^x=e^{tA}x+Z_t,\qquad Z_t:=\int_0^t e^{(t-s)A}\,\mathrm{d}B_s, W t x = e t A x + Z t , Z t := ∫ 0 t e ( t − s ) A d B s ,
where Z t Z_t Z t is centered Gaussian with covariance matrix
Q t = ( ( e 2 t − 1 ) / 2 0 0 ( 1 − e − 2 t ) / 2 ) , det Q t = sinh 2 t . Q_t=\begin{pmatrix}(e^{2t}-1)/2&0\\0&(1-e^{-2t})/2\end{pmatrix},\qquad\det Q_t=\sinh^2t. Q t = ( ( e 2 t − 1 ) /2 0 0 ( 1 − e − 2 t ) /2 ) , det Q t = sinh 2 t .
Since det ( e t A ) = e t tr A = 1 \det(e^{tA})=e^{t\operatorname{tr}A}=1 det ( e t A ) = e t tr A = 1 , for every nonnegative Borel function h h h , Tonelli’s theorem and the change of variables y = e t A x + Z t y=e^{tA}x+Z_t y = e t A x + Z t give
∫ R 2 P t h ( x ) d x = E [ ∫ R 2 h ( e t A x + Z t ) d x ] = ∫ R 2 h ( y ) d y . \int_{\mathbb{R}^2}P_th(x)\,\mathrm{d}x
=\mathbb{E}\left[\int_{\mathbb{R}^2}h(e^{tA}x+Z_t)\,\mathrm{d}x\right]
=\int_{\mathbb{R}^2}h(y)\,\mathrm{d}y. ∫ R 2 P t h ( x ) d x = E [ ∫ R 2 h ( e t A x + Z t ) d x ] = ∫ R 2 h ( y ) d y .
Thus (H1) holds. The transition density is
p t ( x , y ) = 1 2 π ( det Q t ) 1 / 2 exp [ − 1 2 ( y − e t A x ) T Q t − 1 ( y − e t A x ) ] . p_t(x,y)=\frac{1}{2\pi(\det Q_t)^{1/2}}\exp\left[-\frac{1}{2}(y-e^{tA}x)^{\mathsf T}Q_t^{-1}(y-e^{tA}x)\right]. p t ( x , y ) = 2 π ( det Q t ) 1/2 1 exp [ − 2 1 ( y − e t A x ) T Q t − 1 ( y − e t A x ) ] .
and hence
sup x , y ∈ R 2 p t ( x , y ) = 1 2 π sinh t , t > 0. \sup_{x,y\in\mathbb{R}^2}p_t(x,y)=\frac{1}{2\pi\sinh t},\qquad t>0. x , y ∈ R 2 sup p t ( x , y ) = 2 π sinh t 1 , t > 0.
It follows that, for every compact K ⊂ R 2 K\subset\mathbb{R}^2 K ⊂ R 2 and t ≥ 1 t\geq1 t ≥ 1 ,
sup x ∈ R 2 P t 1 K ( x ) ≤ λ ( K ) 2 π sinh t . \sup_{x\in\mathbb{R}^2}P_t1_K(x)\leq\frac{\lambda(K)}{2\pi\sinh t}. x ∈ R 2 sup P t 1 K ( x ) ≤ 2 π sinh t λ ( K ) .
Since ( sinh t ) − 1 (\sinh t)^{-1} ( sinh t ) − 1 decreases faster than every negative power of t t t , and since P t 1 K ≤ 1 P_t\mathbf{1}_K\le1 P t 1 K ≤ 1 for 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , condition (H2) holds for every q > 0 q>0 q > 0 .
We claim that (H3) fails for every q > 0 q>0 q > 0 . Fix q > 0 q>0 q > 0 and suppose, to the contrary, that there are sets D t ( N ) D_t(N) D t ( N ) satisfying (H3) with this exponent. Choose a compact set K K K with 0 < λ ( K ) < ∞ 0<\lambda(K)<\infty 0 < λ ( K ) < ∞ . For each fixed N N N ,
∫ D t ( N ) P t 1 K ( x ) d x ≤ λ ( D t ( N ) ) sup x ∈ R 2 P t 1 K ( x ) ≤ C N λ ( K ) t q 2 π sinh t ⟶ 0 , t → ∞ . \int_{D_t(N)} P_t\mathbf{1}_K(x)\,\mathrm{d}x
\le\lambda(D_t(N))\sup_{x\in\mathbb{R}^2}P_t\mathbf{1}_K(x)
\le\frac{C_N\lambda(K)t^q}{2\pi\sinh t}\longrightarrow0,\qquad t\to\infty. ∫ D t ( N ) P t 1 K ( x ) d x ≤ λ ( D t ( N )) x ∈ R 2 sup P t 1 K ( x ) ≤ 2 π sinh t C N λ ( K ) t q ⟶ 0 , t → ∞.
By (H1),
∫ R 2 P t 1 K d λ = λ ( K ) , \int_{\mathbb{R}^2}P_t\mathbf{1}_K\,\mathrm{d}\lambda=\lambda(K), ∫ R 2 P t 1 K d λ = λ ( K ) ,
and therefore, for every fixed N N N ,
∫ R 2 ∖ D t ( N ) P t 1 K d λ ⟶ λ ( K ) , t → ∞ . \int_{\mathbb{R}^2\setminus D_t(N)}P_t\mathbf{1}_K\,\mathrm{d}\lambda\longrightarrow\lambda(K),\qquad t\to\infty. ∫ R 2 ∖ D t ( N ) P t 1 K d λ ⟶ λ ( K ) , t → ∞.
Consequently, for every fixed N N N ,
lim t → ∞ ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K ) . \lim_{t\to\infty}\int_{\mathbb{R}^2\setminus D_t(N)}P_t\mathbf{1}_K\,\mathrm{d}\lambda=\lambda(K). t → ∞ lim ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K ) .
Moreover, since
0 ≤ ∫ R 2 ∖ D t ( N ) P t 1 K d λ ≤ ∫ R 2 P t 1 K d λ = λ ( K ) , 0\le\int_{\mathbb{R}^2\setminus D_t(N)}P_t\mathbf{1}_K\,\mathrm{d}\lambda\le\int_{\mathbb{R}^2}P_t\mathbf{1}_K\,\mathrm{d}\lambda=\lambda(K), 0 ≤ ∫ R 2 ∖ D t ( N ) P t 1 K d λ ≤ ∫ R 2 P t 1 K d λ = λ ( K ) ,
we obtain
sup t ≥ 1 ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K ) \sup_{t\ge1}\int_{\mathbb{R}^2\setminus D_t(N)}P_t\mathbf{1}_K\,\mathrm{d}\lambda=\lambda(K) t ≥ 1 sup ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K )
for every N N N . Therefore,
lim N → ∞ sup t ≥ 1 ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K ) > 0 , \lim_{N\to\infty}\sup_{t\ge1}\int_{\mathbb{R}^2\setminus D_t(N)}P_t\mathbf{1}_K\,\mathrm{d}\lambda=\lambda(K)>0, N → ∞ lim t ≥ 1 sup ∫ R 2 ∖ D t ( N ) P t 1 K d λ = λ ( K ) > 0 ,
which contradicts (H3). Hence
Q 2 ( P ) = ( 0 , ∞ ) , Q 3 ( P ) = ∅ . Q_2(P)=(0,\infty),\qquad Q_3(P)=\varnothing. Q 2 ( P ) = ( 0 , ∞ ) , Q 3 ( P ) = ∅ .
(iii) Conditions (H2) and (H3) do not imply (H1). Let B B B be a standard Brownian motion on R \mathbb{R} R , fix 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 , and set
T ( x ) = x + ε sin x , m : = 1 − ε , M : = 1 + ε . T(x)=x+\varepsilon\sin x,\qquad m:=1-\varepsilon,\qquad M:=1+\varepsilon. T ( x ) = x + ε sin x , m := 1 − ε , M := 1 + ε .
Then T T T is an increasing C 1 C^1 C 1 -diffeomorphism of R \mathbb{R} R and
m ≤ T ′ ( x ) ≤ M , x ∈ R . m\le T'(x)\le M,\qquad x\in\mathbb{R}. m ≤ T ′ ( x ) ≤ M , x ∈ R .
Define the homeomorphic image of Brownian motion by
W ~ t x = T ( T − 1 ( x ) + B t ) , \widetilde{W}_t^x=T\left(T^{-1}(x)+B_t\right), W t x = T ( T − 1 ( x ) + B t ) ,
whose semigroup is
P ~ t f ( x ) = P t B ( f ∘ T ) ( T − 1 x ) , \widetilde{P}_t f(x)=P_t^B(f\circ T)(T^{-1}x), P t f ( x ) = P t B ( f ∘ T ) ( T − 1 x ) ,
where ( P t B ) t ≥ 0 (P_t^B)_{t\ge0} ( P t B ) t ≥ 0 denotes the standard Brownian semigroup. Since T T T is a homeomorphism, the map U f = f ∘ T Uf=f\circ T U f = f ∘ T is an isometric isomorphism of C 0 ( R ) C_0(\mathbb{R}) C 0 ( R ) , and P ~ t = U − 1 P t B U \widetilde{P}_t=U^{-1}P_t^B U P t = U − 1 P t B U is a conservative Feller semigroup. Moreover, W ~ \widetilde{W} W has continuous paths. Hence it is a conservative continuous Feller process, and therefore a Hunt diffusion. If K ⊂ R K\subset\mathbb{R} K ⊂ R is compact and K ∗ : = T − 1 ( K ) K^*:=T^{-1}(K) K ∗ := T − 1 ( K ) , then, for t ≥ 1 t\ge1 t ≥ 1 ,
sup x ∈ R P ~ t 1 K ( x ) = sup z ∈ R P t B 1 K ∗ ( z ) ≤ λ ( K ∗ ) 2 π t ≤ λ ( K ) m 2 π t . \sup_{x\in\mathbb{R}}\widetilde{P}_t1_K(x)=\sup_{z\in\mathbb{R}}P_t^B1_{K^*}(z)\le\frac{\lambda(K^*)}{\sqrt{2\pi t}}\le\frac{\lambda(K)}{m\sqrt{2\pi t}}. x ∈ R sup P t 1 K ( x ) = z ∈ R sup P t B 1 K ∗ ( z ) ≤ 2 π t λ ( K ∗ ) ≤ m 2 π t λ ( K ) .
Together with the bound P ~ t 1 K ≤ 1 \widetilde{P}_t1_K\le1 P t 1 K ≤ 1 on 0 ≤ t ≤ 1 0\le t\le1 0 ≤ t ≤ 1 , this proves (H2) with q = 1 / 2 q=1/2 q = 1/2 .
For t , N ≥ 1 t,N\ge1 t , N ≥ 1 , set
D t ( N ) = T ( [ − N t , N t ] ) . D_t(N)=T([-N\sqrt{t},N\sqrt{t}]). D t ( N ) = T ([ − N t , N t ]) .
Then
λ ( D t ( N ) ) ≤ 2 M N t . \lambda(D_t(N))\le2MN\sqrt{t}. λ ( D t ( N )) ≤ 2 M N t .
Let K ∗ ⊂ [ − R , R ] K^*\subset[-R,R] K ∗ ⊂ [ − R , R ] . By the change of variables x = T ( z ) x=T(z) x = T ( z ) , the bound T ′ ≤ M T'\le M T ′ ≤ M , Tonelli's theorem, and symmetry of the Brownian transition density, for N > R N>R N > R and t ≥ 1 t\ge1 t ≥ 1 we obtain
∫ R ∖ D t ( N ) P ~ t 1 K ( x ) d x ≤ M ∫ R ∖ [ − N t , N t ] P t B 1 K ∗ ( z ) d z = M ∫ K ∗ P { y + B t ∉ [ − N t , N t ] } d y ≤ M λ ( K ∗ ) P { ∣ B 1 ∣ > N − R } . \begin{aligned}
\int_{\mathbb{R}\setminus D_t(N)}\widetilde{P}_t1_K(x)\,\mathrm{d}x
\le M\int_{\mathbb{R}\setminus[-N\sqrt{t},N\sqrt{t}]}P_t^B1_{K^*}(z)\,\mathrm{d}z \\
&= M\int_{K^*}\mathbb{P}\{y+B_t\notin[-N\sqrt{t},N\sqrt{t}]\}\,\mathrm{d}y \\
&\le M\lambda(K^*)\mathbb{P}\{|B_1|>N-R\}.
\end{aligned} ∫ R ∖ D t ( N ) P t 1 K ( x ) d x ≤ M ∫ R ∖ [ − N t , N t ] P t B 1 K ∗ ( z ) d z = M ∫ K ∗ P { y + B t ∈ / [ − N t , N t ]} d y ≤ M λ ( K ∗ ) P { ∣ B 1 ∣ > N − R } .
The last expression tends to zero as N → ∞ N \to\infty N → ∞ , uniformly in t ≥ 1 t \ge1 t ≥ 1 . Thus (H3) also holds with the same exponent q = 1 / 2 q = 1/2 q = 1/2 .
It remains to show that Lebesgue measure is not invariant. For every nonnegative compactly supported Borel function f f f , symmetry of the Brownian semigroup and the change of variables x = T ( z ) x = T(z) x = T ( z ) give
∫ R P ~ t f ( x ) d x = ∫ R P t B ( f ∘ T ) ( z ) T ′ ( z ) d z = ∫ R ( f ∘ T ) ( z ) P t B T ′ ( z ) d z = ∫ R f ( y ) P t B T ′ ( T − 1 y ) T ′ ( T − 1 y ) d y . \int_{\mathbb{R}} \widetilde{P}_t f(x)\,\mathrm{d}x = \int_{\mathbb{R}} P_t^B(f \circ T)(z)T'(z)\,\mathrm{d}z
= \int_{\mathbb{R}} (f \circ T)(z)P_t^B T'(z)\,\mathrm{d}z
= \int_{\mathbb{R}} f(y)\frac{P_t^B T'(T^{-1}y)}{T'(T^{-1}y)}\,\mathrm{d}y. ∫ R P t f ( x ) d x = ∫ R P t B ( f ∘ T ) ( z ) T ′ ( z ) d z = ∫ R ( f ∘ T ) ( z ) P t B T ′ ( z ) d z = ∫ R f ( y ) T ′ ( T − 1 y ) P t B T ′ ( T − 1 y ) d y .
Therefore, invariance of Lebesgue measure would imply P t B T ′ = T ′ P_t^B T' = T' P t B T ′ = T ′ almost everywhere. However,
T ′ ( z ) = 1 + ε cos z , P t B T ′ ( z ) = 1 + ε e − t / 2 cos z , T'(z) = 1 + \varepsilon\cos z,\qquad P_t^B T'(z) = 1 + \varepsilon e^{-t/2}\cos z, T ′ ( z ) = 1 + ε cos z , P t B T ′ ( z ) = 1 + ε e − t /2 cos z ,
which are not equal for any t > 0 t > 0 t > 0 . Hence (H1) fails.
The three examples prove, within the class of conservative Hunt processes, that none of (H1)–(H3) follows from the other two, even when the exponent q q q is allowed to vary.
Theorem 2.1. Assume (H1)–(H3). Suppose that Γ \Gamma Γ is spread-out and has finite mean μ = E σ = ∫ 0 ∞ Γ ‾ ( t ) d t \mu= \mathbb{E}\sigma= \int_0^\infty\overline{\Gamma}(t)\,\mathrm{d}t μ = E σ = ∫ 0 ∞ Γ ( t ) d t . Then:
(a) if q β > 1 q\beta> 1 q β > 1 , then X t ⇒ X ∞ X_t \Rightarrow X_\infty X t ⇒ X ∞ for a nontrivial random Radon measure satisfying E ⟨ X ∞ , φ ⟩ = ∫ E φ d λ \mathbb{E}\langle X_\infty,\varphi\rangle= \int_E \varphi\,\mathrm{d}\lambda E ⟨ X ∞ , φ ⟩ = ∫ E φ d λ for all φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) ;
(b) if q β ≤ 1 q\beta\le1 q β ≤ 1 , then X t ⇒ 0 X_t \Rightarrow0 X t ⇒ 0 .
Theorem 2.2. Assume (H1)–(H3). Suppose that Γ \Gamma Γ is nonarithmetic and
Γ ‾ ( t ) ∼ c 1 t − γ , c 1 > 0 , 0 < γ ≤ 1. (6) \overline{\Gamma}(t) \sim c_1 t^{-\gamma},\qquad c_1 > 0,\qquad0 < \gamma\le1. \tag*{(6)} Γ ( t ) ∼ c 1 t − γ , c 1 > 0 , 0 < γ ≤ 1. (6) Then:
(a) if q β > γ q\beta> \gamma q β > γ , then X t ⇒ Π λ X_t \Rightarrow\Pi_\lambda X t ⇒ Π λ , a Poisson random measure with intensity λ \lambda λ ;
(b) if q β < γ q\beta< \gamma q β < γ , then X t ⇒ 0 X_t \Rightarrow0 X t ⇒ 0 ;
(c) if q β = γ q\beta= \gamma q β = γ , there are t n → ∞ t_n \to\infty t n → ∞ and a random Radon measure X ∞ X_\infty X ∞ such that X t n ⇒ X ∞ X_{t_n} \Rightarrow X_\infty X t n ⇒ X ∞ and P { X ∞ ≠ 0 } > 0 \mathbb{P}\{X_\infty\ne0\} > 0 P { X ∞ = 0 } > 0 .
Markov process examples
We first verify (H1)–(H3) for four classes of conservative Hunt processes and identify the corresponding exponent q q q . Recurrence and transience are treated separately in § 3.5 . The estimate (H2) gives a general sufficient condition for transience when q > 1 q>1 q > 1 ; recurrence is then established separately for each symmetric class. Throughout this section, G L d ( R ) GL_d(\mathbb{R}) G L d ( R ) denotes the group of invertible d × d d\times d d × d real matrices.
For F = R d F=\mathbb{R}^d F = R d or [ 0 , ∞ ) d [0,\infty)^d [ 0 , ∞ ) d , C 0 ( F ) C_0(F) C 0 ( F ) is the space of continuous functions vanishing at infinity, equipped with the supremum norm ∥ h ∥ ∞ = sup x ∈ F ∣ h ( x ) ∣ \lVert h\rVert_\infty=\sup_{x\in F}|h(x)| ∥ h ∥ ∞ = sup x ∈ F ∣ h ( x ) ∣ . We write C c ( F ) C_c(F) C c ( F ) for the continuous functions with compact support and C c + ( F ) C_c^+(F) C c + ( F ) for its nonnegative cone. For an open set O O O , C c k ( O ) C_c^k(O) C c k ( O ) consists of k k k times continuously differentiable functions with compact support contained in O O O . In C c 2 ( [ 0 , ∞ ) ) C_c^2([0,\infty)) C c 2 ([ 0 , ∞ )) , the derivatives extend continuously to zero from the right.
Two recurring facts.. If a conservative semigroup admits a transition density satisfying p t ( x , y ) = p t ( y , x ) p_t(x,y)=p_t(y,x) p t ( x , y ) = p t ( y , x ) , then Lebesgue measure is invariant, the semigroup is reversible, and its L 2 L^2 L 2 -extension is self-adjoint. We shall also use the localization identity
∫ E ∖ D P t 1 K ( x ) λ ( d x ) = ∫ K P y { W t ∉ D } λ ( d y ) . \int_{E\setminus D} P_t1_K(x)\lambda(dx)=\int_K \mathbb{P}_y\{W_t\notin D\}\lambda(dy). ∫ E ∖ D P t 1 K ( x ) λ ( d x ) = ∫ K P y { W t ∈ / D } λ ( d y ) .
valid for Borel D ⊂ E D\subset E D ⊂ E and compact K ⊂ E K\subset E K ⊂ E . Both statements follow directly from Tonelli’s theorem, symmetry, and conservativity. Moreover, once an estimate of the form sup x P t 1 K ( x ) ≤ C K t − q \sup_x P_t1_K(x)\le C_Kt^{-q} sup x P t 1 K ( x ) ≤ C K t − q has been proved for t ≥ 1 t\ge1 t ≥ 1 , the bound P t 1 K ≤ 1 P_t1_K\le1 P t 1 K ≤ 1 extends it to C K ′ ( 1 + t ) − q C'_K(1+t)^{-q} C K ′ ( 1 + t ) − q for all t ≥ 0 t\ge0 t ≥ 0 . We use this extension without further comment.
Lemma 3.1. Let W W W be a conservative Hunt process on R d \mathbb{R}^d R d satisfying (H1)–(H3) with exponent q q q and localization sets D t ( N ) D_t(N) D t ( N ) . For A ∈ G L d ( R ) A\in GL_d(\mathbb{R}) A ∈ G L d ( R ) , the process W t A = A W t W_t^A=AW_t W t A = A W t also satisfies (H1)–(H3) with exponent q q q and localization sets A D t ( N ) AD_t(N) A D t ( N ) . If the semigroup of W W W is symmetric with respect to Lebesgue measure, then so is the semigroup of W A W^A W A . If U U U and U A U^A U A denote the corresponding potential operators, then
U A 1 K ( A x ) = U 1 A − 1 K ( x ) (7) U^A1_K(Ax)=U1_{A^{-1}K}(x) \tag*{(7)} U A 1 K ( A x ) = U 1 A − 1 K ( x ) (7) for every Borel set K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d .
Proof. Since A A A is a homeomorphism, W A W^A W A is again a conservative Hunt process, with semigroup
P t A h ( x ) = P t ( h ∘ A ) ( A − 1 x ) . P_t^Ah(x)=P_t(h\circ A)(A^{-1}x). P t A h ( x ) = P t ( h ∘ A ) ( A − 1 x ) .
A change of variables shows that (H1) is preserved. For every compact K ⊂ R d K \subset\mathbb{R}^{d} K ⊂ R d ,
sup x P t A 1 K ( x ) = sup z P t 1 A − 1 K ( z ) , \sup_{x} P_{t}^{A}\mathbf{1}_{K}(x) = \sup_{z} P_{t}\mathbf{1}_{A^{-1}K}(z), x sup P t A 1 K ( x ) = z sup P t 1 A − 1 K ( z ) ,
∫ ( A D t ( N ) ) c P t A 1 K d λ = ∣ det A ∣ ∫ D t ( N ) c P t 1 A − 1 K d λ , \int_{(AD_{t}(N))^{c}} P_{t}^{A}\mathbf{1}_{K}\,\mathrm{d}\lambda= |\det A| \int_{D_{t}(N)^{c}} P_{t}\mathbf{1}_{A^{-1}K}\,\mathrm{d}\lambda, ∫ ( A D t ( N ) ) c P t A 1 K d λ = ∣ det A ∣ ∫ D t ( N ) c P t 1 A − 1 K d λ ,
and λ ( A D t ( N ) ) = ∣ det A ∣ λ ( D t ( N ) ) \lambda(AD_{t}(N)) = |\det A|\lambda(D_{t}(N)) λ ( A D t ( N )) = ∣ det A ∣ λ ( D t ( N )) . Hence (H2) and (H3) hold with the same exponent. The same change of variables in the bilinear symmetry identity proves symmetry of the transformed semigroup, and integration of the semigroup identity over time gives (3.1).
Symmetric stable motions and anisotropic stable products
We first record two classes of stable motions satisfying (H1)–(H3). Although Vatutin and Wakolbinger [2 ] refers simply to a symmetric α \alpha α -stable motion, the estimates used there are those for the spherically symmetric stable process with generator − ( − Δ ) α / 2 -(-\Delta)^{\alpha/2} − ( − Δ ) α /2 , corresponding to the isotropic case. The first class below removes this isotropy restriction and consists of arbitrary full-dimensional symmetric stable Lévy processes, while the second is obtained from products of one-dimensional stable motions by an invertible linear transformation.
Proposition 3.2. (a) Let 0 < α ≤ 2 0 < \alpha\le2 0 < α ≤ 2 . A full-dimensional symmetric α \alpha α -stable Lévy process on R d \mathbb{R}^{d} R d , that is, one whose law at time one is not supported on a proper linear subspace of R d \mathbb{R}^{d} R d , satisfies (H1)–(H3) with q = d / α q = d/\alpha q = d / α and D t ( N ) = B ( 0 , N t 1 / α ) D_{t}(N) = B(0,Nt^{1/\alpha}) D t ( N ) = B ( 0 , N t 1/ α ) .
(b) Let S ( i ) S^{(i)} S ( i ) , i = 1 , … , d i = 1,\ldots,d i = 1 , … , d , be independent nondegenerate symmetric α i \alpha_{i} α i -stable Lévy processes on R \mathbb{R} R , with 0 < α i ≤ 2 0 < \alpha_{i} \le2 0 < α i ≤ 2 , and let A ∈ G L d ( R ) A \in GL_{d}(\mathbb{R}) A ∈ G L d ( R ) . Then S t A = A ( S t ( 1 ) , … , S t ( d ) ) S_{t}^{A} = A(S_{t}^{(1)},\ldots,S_{t}^{(d)}) S t A = A ( S t ( 1 ) , … , S t ( d ) ) satisfies (H1)–(H3) with q = ∑ i = 1 d α i − 1 q = \sum_{i=1}^{d}\alpha_{i}^{-1} q = ∑ i = 1 d α i − 1 and
D t ( N ) = A ( ∏ i = 1 d [ − N t 1 / α i , N t 1 / α i ] ) . D_{t}(N) = A\left(\prod_{i=1}^{d}[-Nt^{1/\alpha_{i}},Nt^{1/\alpha_{i}}]\right). D t ( N ) = A ( i = 1 ∏ d [ − N t 1/ α i , N t 1/ α i ] ) . In particular, this class contains every centered Brownian motion on R d \mathbb{R}^{d} R d with positive definite covariance matrix.
Proof. Let W W W be the process in part (a) and write
E e i ξ ⋅ W t = e − t Ψ ( ξ ) , ξ ∈ R d . \mathbb{E}e^{i\xi\cdot W_{t}} = e^{-t\Psi(\xi)}, \qquad\xi\in\mathbb{R}^{d}. E e i ξ ⋅ W t = e − t Ψ ( ξ ) , ξ ∈ R d .
Suppose first that 0 < α < 2 0 < \alpha< 2 0 < α < 2 . By Sato [8 Theorem 14.13] , there exists a unique finite symmetric measure σ \sigma σ on S d − 1 \mathbb{S}^{d-1} S d − 1 such that
Ψ ( ξ ) = ∫ S d − 1 ∣ ⟨ ξ , θ ⟩ ∣ α σ ( d θ ) . \Psi(\xi) = \int_{\mathbb{S}^{d-1}} |\langle\xi,\theta\rangle|^{\alpha}\,\sigma(\mathrm{d}\theta). Ψ ( ξ ) = ∫ S d − 1 ∣ ⟨ ξ , θ ⟩ ∣ α σ ( d θ ) .
The finiteness of σ \sigma σ implies, by dominated convergence, that Ψ \Psi Ψ is continuous, and its representation gives Ψ ( r ξ ) = ∣ r ∣ α Ψ ( ξ ) \Psi(r\xi)=|r|^\alpha\Psi(\xi) Ψ ( r ξ ) = ∣ r ∣ α Ψ ( ξ ) . If Ψ ( ξ 0 ) = 0 \Psi(\xi_0)=0 Ψ ( ξ 0 ) = 0 for some ξ 0 ≠ 0 \xi_0\ne0 ξ 0 = 0 , this homogeneity would give E e i r ⟨ ξ 0 , W 1 ⟩ = 1 \mathbb{E}e^{ir\langle\xi_0,W_1\rangle}=1 E e i r ⟨ ξ 0 , W 1 ⟩ = 1 for every r ∈ R r\in\mathbb{R} r ∈ R . Then ⟨ ξ 0 , W 1 ⟩ = 0 \langle\xi_0,W_1\rangle=0 ⟨ ξ 0 , W 1 ⟩ = 0 almost surely, contradicting full dimensionality. Hence Ψ \Psi Ψ is strictly positive on S d − 1 S^{d-1} S d − 1 , and compactness and homogeneity yield
a ∣ ξ ∣ α ≤ Ψ ( ξ ) ≤ b ∣ ξ ∣ α , ξ ∈ R d . a|\xi|^\alpha\leq\Psi(\xi) \leq b|\xi|^\alpha,\qquad\xi\in\mathbb{R}^d. a ∣ ξ ∣ α ≤ Ψ ( ξ ) ≤ b ∣ ξ ∣ α , ξ ∈ R d .
for some 0 < a ≤ b < ∞ 0<a\leq b<\infty 0 < a ≤ b < ∞ . For α = 2 \alpha=2 α = 2 , W 1 W_1 W 1 is centered Gaussian Sato [8, Theorem 14.1], with a positive definite covariance matrix Σ \Sigma Σ by full dimensionality. The same homogeneity and bounds follow from
Ψ ( ξ ) = 1 2 ⟨ Σ ξ , ξ ⟩ , a = 1 2 λ min ( Σ ) , b = 1 2 λ max ( Σ ) . \Psi(\xi)=\frac{1}{2}\langle\Sigma\xi,\xi\rangle,\qquad a=\frac{1}{2}\lambda_{\min}(\Sigma),\qquad b=\frac{1}{2}\lambda_{\max}(\Sigma). Ψ ( ξ ) = 2 1 ⟨ Σ ξ , ξ ⟩ , a = 2 1 λ m i n ( Σ ) , b = 2 1 λ m a x ( Σ ) .
Since Ψ ( ξ ) ≥ a ∣ ξ ∣ α \Psi(\xi)\geq a|\xi|^\alpha Ψ ( ξ ) ≥ a ∣ ξ ∣ α , for every t > 0 t>0 t > 0 ,
∫ R d e − t Ψ ( ξ ) d ξ ≤ ∫ R d e − a t ∣ ξ ∣ α d ξ < ∞ . \int_{\mathbb{R}^d}e^{-t\Psi(\xi)}\,d\xi\leq\int_{\mathbb{R}^d}e^{-at|\xi|^\alpha}\,d\xi<\infty. ∫ R d e − t Ψ ( ξ ) d ξ ≤ ∫ R d e − a t ∣ ξ ∣ α d ξ < ∞.
Thus the characteristic function of W t W_t W t is integrable. By Sato [8, Proposition 2.5(xii)], the law of W t W_t W t has a bounded continuous density
p t ( z ) = ( 2 π ) − d ∫ R d e − i ξ ⋅ z e − t Ψ ( ξ ) d ξ . p_t(z)=(2\pi)^{-d}\int_{\mathbb{R}^d}e^{-i\xi\cdot z}e^{-t\Psi(\xi)}\,d\xi. p t ( z ) = ( 2 π ) − d ∫ R d e − i ξ ⋅ z e − t Ψ ( ξ ) d ξ .
Changing variables η = t 1 / α ξ \eta=t^{1/\alpha}\xi η = t 1/ α ξ in this absolutely convergent integral and using homogeneity gives
p t ( z ) = t − d / α p 1 ( t − 1 / α z ) . p_t(z)=t^{-d/\alpha}p_1(t^{-1/\alpha}z). p t ( z ) = t − d / α p 1 ( t − 1/ α z ) .
The symmetry of Ψ \Psi Ψ also gives p t ( z ) = p t ( − z ) p_t(z)=p_t(-z) p t ( z ) = p t ( − z ) . Every Lévy process is a Hunt process Applebaum [9, Theorem 3.6.11], and its semigroup is conservative. The convolution form of the semigroup and translation invariance of Lebesgue measure give (H1). Moreover, for every compact K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d and t ≥ 1 t\geq1 t ≥ 1 ,
sup x ∈ R d P t 1 K ( x ) ≤ λ ( K ) ∥ p t ∥ ∞ = λ ( K ) ∥ p 1 ∥ ∞ t − d / α . \sup_{x\in\mathbb{R}^d}P_t1_K(x)\leq\lambda(K)\|p_t\|_\infty=\lambda(K)\|p_1\|_\infty t^{-d/\alpha}. x ∈ R d sup P t 1 K ( x ) ≤ λ ( K ) ∥ p t ∥ ∞ = λ ( K ) ∥ p 1 ∥ ∞ t − d / α .
Thus (H2) holds with q = d / α q=d/\alpha q = d / α .
For (H3), take K ⊂ B ( 0 , R ) K\subset B(0,R) K ⊂ B ( 0 , R ) and D t ( N ) = B ( 0 , N t 1 / α ) D_t(N)=B(0,Nt^{1/\alpha}) D t ( N ) = B ( 0 , N t 1/ α ) . By the localization identity at the beginning of the section and W t = t 1 / α W 1 W_t=t^{1/\alpha}W_1 W t = t 1/ α W 1 , for t ≥ 1 t\geq1 t ≥ 1 and N > R N>R N > R ,
∫ D t ( N ) c P t 1 K ( x ) d x = ∫ K P { y + t 1 / α W 1 ∉ D t ( N ) } d y ≤ λ ( K ) P { ∣ W 1 ∣ > N − R } . \int_{D_t(N)^c}P_t1_K(x)\,dx
=\int_K\mathbb{P}\{y+t^{1/\alpha}W_1\notin D_t(N)\}\,dy
\leq\lambda(K)\mathbb{P}\{|W_1|>N-R\}. ∫ D t ( N ) c P t 1 K ( x ) d x = ∫ K P { y + t 1/ α W 1 ∈ / D t ( N )} d y ≤ λ ( K ) P { ∣ W 1 ∣ > N − R } .
Indeed, the event in the first line implies ∣ W 1 ∣ > N − ∣ y ∣ t − 1 / α ≥ N − R |W_1| > N - |y|t^{-1/\alpha} \ge N - R ∣ W 1 ∣ > N − ∣ y ∣ t − 1/ α ≥ N − R . The bound tends to zero as N → ∞ N \to\infty N → ∞ , uniformly in t ≥ 1 t \ge1 t ≥ 1 , and λ ( D t ( N ) ) = λ ( B ( 0 , N ) ) t d / α \lambda(D_t(N)) = \lambda(B(0,N))t^{d/\alpha} λ ( D t ( N )) = λ ( B ( 0 , N )) t d / α . This proves (H3).
For part (b), set S t = ( S t ( 1 ) , … , S t ( d ) ) S_t = (S_t^{(1)},\ldots,S_t^{(d)}) S t = ( S t ( 1 ) , … , S t ( d ) ) and q = ∑ i = 1 d α i − 1 q = \sum_{i=1}^{d}\alpha_i^{-1} q = ∑ i = 1 d α i − 1 . The product S S S is again a conservative Lévy process, hence Hunt by the result used in part (a). Applying that part in dimension one and using independence gives the symmetric density
p t ( z ) = t − q ∏ i = 1 d p 1 ( i ) ( t − 1 / α i z i ) , ∥ p 1 ( i ) ∥ ∞ < ∞ . p_t(z) = t^{-q}\prod_{i=1}^{d}p_1^{(i)}(t^{-1/\alpha_i}z_i), \qquad\|p_1^{(i)}\|_{\infty}<\infty. p t ( z ) = t − q i = 1 ∏ d p 1 ( i ) ( t − 1/ α i z i ) , ∥ p 1 ( i ) ∥ ∞ < ∞.
Its convolution semigroup satisfies (H1), and
sup x P t 1 K ( x ) ≤ λ ( K ) t − q ∏ i = 1 d ∥ p 1 ( i ) ∥ ∞ , t ≥ 1 , \sup_x P_t1_K(x) \le\lambda(K)t^{-q}\prod_{i=1}^{d}\|p_1^{(i)}\|_{\infty}, \qquad t\ge1, x sup P t 1 K ( x ) ≤ λ ( K ) t − q i = 1 ∏ d ∥ p 1 ( i ) ∥ ∞ , t ≥ 1 ,
proves (H2).
Set
Q t ( N ) = ∏ i = 1 d [ − N t 1 / α i , N t 1 / α i ] = Δ t [ − N , N ] d , Q_t(N) = \prod_{i=1}^{d}[-Nt^{1/\alpha_i},Nt^{1/\alpha_i}] = \Delta_t[-N,N]^d, Q t ( N ) = i = 1 ∏ d [ − N t 1/ α i , N t 1/ α i ] = Δ t [ − N , N ] d ,
Δ t = diag ( t 1 / α 1 , … , t 1 / α d ) . \Delta_t = \operatorname{diag}(t^{1/\alpha_1},\ldots,t^{1/\alpha_d}). Δ t = diag ( t 1/ α 1 , … , t 1/ α d ) .
Then λ ( Q t ( N ) ) = ( 2 N ) d t q \lambda(Q_t(N))=(2N)^dt^q λ ( Q t ( N )) = ( 2 N ) d t q . For K ⊂ [ − R , R ] d K\subset[-R,R]^d K ⊂ [ − R , R ] d , the localization identity and S t = d Δ t S 1 S_t \overset{d}{=} \Delta_tS_1 S t = d Δ t S 1 give
∫ Q t ( N ) c P t 1 K ( x ) d x = ∫ K P { Δ t − 1 y + S 1 ∉ [ − N , N ] d } d y ≤ λ ( K ) P { ∥ S 1 ∥ ∞ > N − R } , t ≥ 1 , N > R . \begin{aligned}
\int_{Q_t(N)^c}P_t1_K(x)\,dx = \int_K \mathbb{P}\{\Delta_t^{-1}y+S_1\notin[-N,N]^d\}\,dy \\
&\le\lambda(K)\mathbb{P}\{\|S_1\|_{\infty}>N-R\}, \qquad t\ge1,\ N>R.
\end{aligned} ∫ Q t ( N ) c P t 1 K ( x ) d x = ∫ K P { Δ t − 1 y + S 1 ∈ / [ − N , N ] d } d y ≤ λ ( K ) P { ∥ S 1 ∥ ∞ > N − R } , t ≥ 1 , N > R .
Here ∥ Δ t − 1 y ∥ ∞ ≤ R \|\Delta_t^{-1}y\|_{\infty}\le R ∥ Δ t − 1 y ∥ ∞ ≤ R , so leaving the box forces ∥ S 1 ∥ ∞ > N − R \|S_1\|_{\infty}>N-R ∥ S 1 ∥ ∞ > N − R . The right-hand side tends to zero uniformly in t t t , proving (H3) together with the volume bound.
The assertion for S A = A S S^A=AS S A = A S now follows from Lemma 3.1, with localization sets A Q t ( N ) AQ_t(N) A Q t ( N ) . Taking α i = 2 \alpha_i=2 α i = 2 and S ( i ) S^{(i)} S ( i ) to be standard Brownian motions, and writing a positive definite covariance matrix as Σ = A A T \Sigma=AA^{\mathsf{T}} Σ = A A T , gives the final claim. □ \square □
Uniformly elliptic symmetric diffusions
The next class consists of symmetric diffusions whose transition densities satisfy two-sided Gaussian bounds. In this case the natural dispersal scale is t 1 / 2 t^{1/2} t 1/2 , and the corresponding exponent in (H1)–(H3) is d / 2 d/2 d /2 .
Proposition 3.3. Let W W W be a conservative Hunt diffusion on R d \mathbb{R}^{d} R d whose transition density with respect to Lebesgue measure is symmetric, p t ( x , y ) = p t ( y , x ) p_t(x,y)=p_t(y,x) p t ( x , y ) = p t ( y , x ) . Suppose that there are constants a 1 , a 2 , A 1 , A 2 > 0 a_1,a_2,A_1,A_2>0 a 1 , a 2 , A 1 , A 2 > 0 such that, for t ≥ 1 t\ge1 t ≥ 1 ,
a 1 t − d / 2 exp ( − a 2 ∣ x − y ∣ 2 t ) ≤ p t ( x , y ) ≤ A 1 t − d / 2 exp ( − A 2 ∣ x − y ∣ 2 t ) . (8) a_1t^{-d/2}\exp\left(-a_2\frac{|x-y|^2}{t}\right)\le p_t(x,y)\le A_1t^{-d/2}\exp\left(-A_2\frac{|x-y|^2}{t}\right). \tag*{(8)} a 1 t − d /2 exp ( − a 2 t ∣ x − y ∣ 2 ) ≤ p t ( x , y ) ≤ A 1 t − d /2 exp ( − A 2 t ∣ x − y ∣ 2 ) . (8) Then (H1)–(H3) hold with q = d / 2 q=d/2 q = d /2 and D t ( N ) = B ( 0 , N t ) D_t(N)=B(0,N\sqrt{t}) D t ( N ) = B ( 0 , N t ) . In particular, this applies to the conservative symmetric Hunt diffusion associated with any bounded measurable symmetric uniformly elliptic divergence-form matrix field on R d \mathbb{R}^{d} R d .
Proof. The symmetric-kernel observation at the beginning of the section gives (H1). If K ⊂ R d K\subset\mathbb{R}^{d} K ⊂ R d is compact, the upper bound in (8) yields, for t ≥ 1 t\ge1 t ≥ 1 ,
sup x ∈ R d P t 1 K ( x ) ≤ A 1 λ ( K ) t − d / 2 . \sup_{x\in\mathbb{R}^{d}}P_t\mathbf{1}_K(x)\le A_1\lambda(K)t^{-d/2}. x ∈ R d sup P t 1 K ( x ) ≤ A 1 λ ( K ) t − d /2 .
Hence (H2) holds with q = d / 2 q=d/2 q = d /2 .
For (H3), choose R > 0 R>0 R > 0 such that K ⊂ B ( 0 , R ) K\subset B(0,R) K ⊂ B ( 0 , R ) and set D t ( N ) = B ( 0 , N t ) D_t(N)=B(0,N\sqrt{t}) D t ( N ) = B ( 0 , N t ) . For t ≥ 1 t\ge1 t ≥ 1 and N > R N>R N > R ,
∫ D t ( N ) c P t 1 K ( x ) d x ≤ A 1 ∫ K ∫ ∣ x ∣ > N t t − d / 2 exp ( − A 2 ∣ x − y ∣ 2 t ) d x d y ≤ A 1 λ ( K ) ∫ ∣ z ∣ > N − R e − A 2 ∣ z ∣ 2 d z . \begin{aligned}
\int_{D_t(N)^c}P_t\mathbf{1}_K(x)\,dx\le A_1\int_K\int_{|x|>N\sqrt{t}}t^{-d/2}\exp\left(-A_2\frac{|x-y|^2}{t}\right)\,dx\,dy \\
&\le A_1\lambda(K)\int_{|z|>N-R}e^{-A_2|z|^2}\,dz.
\end{aligned} ∫ D t ( N ) c P t 1 K ( x ) d x ≤ A 1 ∫ K ∫ ∣ x ∣ > N t t − d /2 exp ( − A 2 t ∣ x − y ∣ 2 ) d x d y ≤ A 1 λ ( K ) ∫ ∣ z ∣ > N − R e − A 2 ∣ z ∣ 2 d z .
Under the substitution z = ( x − y ) / t z=(x-y)/\sqrt{t} z = ( x − y ) / t , the conditions ∣ x ∣ > N t |x|>N\sqrt{t} ∣ x ∣ > N t , ∣ y ∣ ≤ R |y|\le R ∣ y ∣ ≤ R , and t ≥ 1 t\ge1 t ≥ 1 imply ∣ z ∣ > N − R |z|>N-R ∣ z ∣ > N − R . The Gaussian tail tends to zero uniformly in t t t , and λ ( D t ( N ) ) = λ ( B ( 0 , N ) ) t d / 2 \lambda(D_t(N))=\lambda(B(0,N))t^{d/2} λ ( D t ( N )) = λ ( B ( 0 , N )) t d /2 , proving (H3).
For the concrete class in the last assertion, let A : R d → R d × d A:\mathbb{R}^{d}\to\mathbb{R}^{d\times d} A : R d → R d × d , A ( x ) = ( a i j ( x ) ) 1 ≤ i , j ≤ d A(x)=(a_{ij}(x))_{1\le i,j\le d} A ( x ) = ( a ij ( x ) ) 1 ≤ i , j ≤ d , be a measurable matrix field such that A ( x ) = A ( x ) T A(x)=A(x)^{\mathsf{T}} A ( x ) = A ( x ) T for almost every x ∈ R d x\in\mathbb{R}^{d} x ∈ R d . Assume that A A A is uniformly elliptic: there exist constants 0 < λ 0 ≤ Λ 0 < ∞ 0<\lambda_0\le\Lambda_0<\infty 0 < λ 0 ≤ Λ 0 < ∞ such that
λ 0 ∣ ξ ∣ 2 ≤ ξ T A ( x ) ξ ≤ Λ 0 ∣ ξ ∣ 2 , ξ ∈ R d , \lambda_0|\xi|^2\le\xi^{\mathsf{T}}A(x)\xi\le\Lambda_0|\xi|^2,\qquad\xi\in\mathbb{R}^{d}, λ 0 ∣ ξ ∣ 2 ≤ ξ T A ( x ) ξ ≤ Λ 0 ∣ ξ ∣ 2 , ξ ∈ R d ,
for almost every x ∈ R d x\in\mathbb{R}^{d} x ∈ R d .
Let L 2 ( R d ) = L 2 ( R d , d x ) L^2(\mathbb{R}^{d})=L^2(\mathbb{R}^{d},dx) L 2 ( R d ) = L 2 ( R d , d x ) be the Hilbert space of square-integrable functions, identified up to equality almost everywhere, and let
H 1 ( R d ) = { u ∈ L 2 ( R d ) : ∂ i u ∈ L 2 ( R d ) , 1 ≤ i ≤ d } . H^1(\mathbb{R}^{d})=\left\{u\in L^2(\mathbb{R}^{d}):\partial_i u\in L^2(\mathbb{R}^{d}),\ 1\le i\le d\right\}. H 1 ( R d ) = { u ∈ L 2 ( R d ) : ∂ i u ∈ L 2 ( R d ) , 1 ≤ i ≤ d } .
Here ∂ i u \partial_i u ∂ i u denotes the weak derivative of u u u : a function g i ∈ L 2 ( R d ) g_i \in L^2(\mathbb{R}^d) g i ∈ L 2 ( R d ) is equal to ∂ i u \partial_i u ∂ i u if
∫ R d u ( x ) ∂ i φ ( x ) d x = − ∫ R d g i ( x ) φ ( x ) d x , φ ∈ C c ∞ ( R d ) . \int_{\mathbb{R}^d} u(x)\,\partial_i\varphi(x)\,dx=-\int_{\mathbb{R}^d}g_i(x)\varphi(x)\,dx,\qquad\varphi\in C^\infty_c(\mathbb{R}^d). ∫ R d u ( x ) ∂ i φ ( x ) d x = − ∫ R d g i ( x ) φ ( x ) d x , φ ∈ C c ∞ ( R d ) .
For u ∈ H 1 ( R d ) u\in H^1(\mathbb{R}^d) u ∈ H 1 ( R d ) , write ∇ u = ( ∂ 1 u , … , ∂ d u ) ∈ L 2 ( R d ; R d ) \nabla u=(\partial_1u,\ldots,\partial_du)\in L^2(\mathbb{R}^d;\mathbb{R}^d) ∇ u = ( ∂ 1 u , … , ∂ d u ) ∈ L 2 ( R d ; R d ) . On L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) , define the bilinear form
F = H 1 ( R d ) , E ( u , v ) = 1 2 ∫ R d ∇ u ( x ) ⊤ A ( x ) ∇ v ( x ) d x , u , v ∈ F . \mathcal{F}=H^1(\mathbb{R}^d),\qquad\mathcal{E}(u,v)=\frac{1}{2}\int_{\mathbb{R}^d}\nabla u(x)^\top A(x)\nabla v(x)\,dx,\qquad u,v\in\mathcal{F}. F = H 1 ( R d ) , E ( u , v ) = 2 1 ∫ R d ∇ u ( x ) ⊤ A ( x ) ∇ v ( x ) d x , u , v ∈ F .
The ellipticity bounds imply
∣ E ( u , v ) ∣ ≤ Λ 0 2 ∥ ∇ u ∥ 2 ∥ ∇ v ∥ 2 , λ 0 2 ∥ ∇ u ∥ 2 2 ≤ E ( u , u ) ≤ Λ 0 2 ∥ ∇ u ∥ 2 2 . \lvert\mathcal{E}(u,v)\rvert\leq\frac{\Lambda_0}{2}\lVert\nabla u\rVert_2\lVert\nabla v\rVert_2,\qquad\frac{\lambda_0}{2}\lVert\nabla u\rVert_2^2\leq\mathcal{E}(u,u)\leq\frac{\Lambda_0}{2}\lVert\nabla u\rVert_2^2. ∣ E ( u , v )∣ ≤ 2 Λ 0 ∥ ∇ u ∥ 2 ∥ ∇ v ∥ 2 , 2 λ 0 ∥ ∇ u ∥ 2 2 ≤ E ( u , u ) ≤ 2 Λ 0 ∥ ∇ u ∥ 2 2 .
Thus E \mathcal{E} E is well defined on F \mathcal{F} F , and the norm induced by E 1 ( u , u ) : = E ( u , u ) + ∥ u ∥ 2 2 \mathcal{E}_1(u,u):=\mathcal{E}(u,u)+\lVert u\rVert_2^2 E 1 ( u , u ) := E ( u , u ) + ∥ u ∥ 2 2 is equivalent to the usual H 1 ( R d ) H^1(\mathbb{R}^d) H 1 ( R d ) -norm. By Chen and Fukushima [10 Example 6.7.15] , ( E , F ) (\mathcal{E},\mathcal{F}) ( E , F ) is a regular local Dirichlet form on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) , and its associated Hunt process is a conservative diffusion with a jointly continuous transition density.
The L 2 L^2 L 2 -generator of this form is the weak realization of the divergence-form operator
L = 1 2 ∇ ⋅ ( A ∇ ) , \mathcal{L}=\frac{1}{2}\nabla\cdot(A\nabla), L = 2 1 ∇ ⋅ ( A ∇ ) ,
in the sense that E ( u , v ) = − ( L u , v ) L 2 \mathcal{E}(u,v)=-(\mathcal{L}u,v)_{L^2} E ( u , v ) = − ( L u , v ) L 2 for u u u in the domain of L \mathcal{L} L and v ∈ H 1 ( R d ) v\in H^1(\mathbb{R}^d) v ∈ H 1 ( R d ) . Since the form is symmetric with respect to Lebesgue measure, its semigroup is self-adjoint on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) . Consequently, p t ( x , y ) = p t ( y , x ) p_t(x,y)=p_t(y,x) p t ( x , y ) = p t ( y , x ) for almost every ( x , y ) (x,y) ( x , y ) , and joint continuity extends this identity to every t > 0 t>0 t > 0 and x , y ∈ R d x,y\in\mathbb{R}^d x , y ∈ R d .
For t > s t>s t > s , the kernel p t − s ( x , y ) p_{t-s}(x,y) p t − s ( x , y ) , as a function of ( x , t ) (x,t) ( x , t ) with pole ( y , s ) (y,s) ( y , s ) , is the weak fundamental solution of ∂ t u = L u \partial_tu=\mathcal{L}u ∂ t u = L u . Aronson’s Gaussian estimates, proved for smooth coefficients in Aronson [11 Theorem 7, p. 661] and extended to bounded measurable coefficients in Aronson [11 Theorem 10(ii), p. 679] , therefore apply. Since the coefficients are time independent and there are no lower-order terms, the constants are independent of the time horizon Aronson [12 §5, p. 895] . This gives (8) , and the first part of the proof yields (H1)–(H3) with q = d / 2 q=d/2 q = d /2 . □
Products of generalized squared-Bessel diffusions
Fix ρ , κ > 0 \rho,\kappa>0 ρ , κ > 0 and set a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) and δ = 2 / ρ \delta=2/\rho δ = 2/ ρ . For each x ∈ [ 0 , ∞ ) x\in[0,\infty) x ∈ [ 0 , ∞ ) , let Z Z Z be the unique nonnegative strong solution of
d Z t = δ d t + 2 Z t d B t , Z 0 = a ρ , κ x ρ , (9) dZ_t=\delta\,dt+2\sqrt{Z_t}\,dB_t,\qquad Z_0=a_{\rho,\kappa}x^\rho, \tag*{(9)} d Z t = δ d t + 2 Z t d B t , Z 0 = a ρ , κ x ρ , (9)
where B B B is a standard Brownian motion. Thus Z Z Z is the squared-Bessel process of dimension δ \delta δ started from a ρ , κ x ρ a_{\rho,\kappa}x^{\rho} a ρ , κ x ρ . Define
Y t = ( Z t / a ρ , κ ) 1 / ρ . Y_t = \left(Z_t/a_{\rho,\kappa}\right)^{1/\rho}. Y t = ( Z t / a ρ , κ ) 1/ ρ .
so that Y 0 = x Y_0=x Y 0 = x . The existence and uniqueness of the nonnegative solution of (3.3), together with the Feller property and transition kernel of the squared-Bessel family, are given in Revuz and Yor [13 Chapter XI, §1, pp. 439–442] .
Proposition 3.4. The process Y Y Y is a conservative Feller diffusion on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . The domain of its strong generator contains C c 2 ( ( 0 , ∞ ) ) C_c^2((0,\infty)) C c 2 (( 0 , ∞ )) , and on this class the generator is
L ρ , κ ϕ ( x ) = κ d d x ( x 2 − ρ ϕ ′ ( x ) ) . (10) L_{\rho,\kappa}\phi(x)=\kappa\frac{\mathrm{d}}{\mathrm{d}x}\left(x^{2-\rho}\phi'(x)\right). \tag*{(10)} L ρ , κ ϕ ( x ) = κ d x d ( x 2 − ρ ϕ ′ ( x ) ) . (10) For every t > 0 t>0 t > 0 , its transition kernel has a jointly continuous, strictly positive, symmetric density with respect to Lebesgue measure. The independent product
W = ( Y ( 1 ) , … , Y ( d ) ) W=\left(Y^{(1)},\ldots,Y^{(d)}\right) W = ( Y ( 1 ) , … , Y ( d ) ) is a conservative Hunt process on [ 0 , ∞ ) d [0,\infty)^d [ 0 , ∞ ) d satisfying (H1)–(H3) with q = d / ρ q=d/\rho q = d / ρ and
D t ( N ) = [ 0 , N t 1 / ρ ] d . D_t(\mathbb{N})=[0,Nt^{1/\rho}]^d. D t ( N ) = [ 0 , N t 1/ ρ ] d . Its semigroup is reversible with respect to Lebesgue measure and self-adjoint on L 2 ( [ 0 , ∞ ) d , λ ) L^2([0,\infty)^d,\lambda) L 2 ([ 0 , ∞ ) d , λ ) . On the linear span of functions of the form
f ( x ) = ∏ j = 1 d f j ( x j ) , f j ∈ C c 2 ( ( 0 , ∞ ) ) , f(x)=\prod_{j=1}^{d}f_j(x_j),\qquad f_j\in C_c^2((0,\infty)), f ( x ) = j = 1 ∏ d f j ( x j ) , f j ∈ C c 2 (( 0 , ∞ )) , its generator acts as
L [ d ] f ( x ) = κ ∑ j = 1 d ∂ x j ( x j 2 − ρ ∂ x j f ( x ) ) . L^{[d]}f(x)=\kappa\sum_{j=1}^{d}\partial_{x_j}\left(x_j^{2-\rho}\partial_{x_j}f(x)\right). L [ d ] f ( x ) = κ j = 1 ∑ d ∂ x j ( x j 2 − ρ ∂ x j f ( x ) ) .
Proof. Construction and generator. Let
T ( z ) = ( z / a ρ , κ ) 1 / ρ , T − 1 ( x ) = a ρ , κ x ρ . T(z)=(z/a_{\rho,\kappa})^{1/\rho},\qquad T^{-1}(x)=a_{\rho,\kappa}x^\rho. T ( z ) = ( z / a ρ , κ ) 1/ ρ , T − 1 ( x ) = a ρ , κ x ρ .
If ( Q t ) t ≥ 0 (Q_t)_{t\geq0} ( Q t ) t ≥ 0 denotes the squared-Bessel semigroup, then, for every f ∈ C 0 ( [ 0 , ∞ ) ) f\in C_0([0,\infty)) f ∈ C 0 ([ 0 , ∞ )) ,
P t f ( x ) = Q t ( f ∘ T ) ( T − 1 ( x ) ) . P_tf(x)=Q_t(f\circ T)(T^{-1}(x)). P t f ( x ) = Q t ( f ∘ T ) ( T − 1 ( x )) .
Since T T T is a homeomorphism of [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , the map f ↦ f ∘ T f\mapsto f\circ T f ↦ f ∘ T is an isometric isomorphism of C 0 ( [ 0 , ∞ ) ) C_0([0,\infty)) C 0 ([ 0 , ∞ )) . It follows that ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 is a Feller semigroup. Conservativity follows from that of the squared-Bessel process. Moreover, Y t = T ( Z t ) Y_t=T(Z_t) Y t = T ( Z t ) has continuous paths because Z Z Z does. Thus Y Y Y is a conservative continuous Feller process, hence a Hunt process.
Fix ϕ ∈ C c 2 ( ( 0 , ∞ ) ) \phi\in C_c^2((0,\infty)) ϕ ∈ C c 2 (( 0 , ∞ )) and set h = ϕ ∘ T h=\phi\circ T h = ϕ ∘ T . Since ϕ \phi ϕ has compact support in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , the same is true of h h h , so h h h extends by zero to a function in C c 2 ( [ 0 , ∞ ) ) C_c^2([0,\infty)) C c 2 ([ 0 , ∞ )) . The squared-Bessel generator is
A h ( z ) = 2 z h ′ ′ ( z ) + δ h ′ ( z ) . \mathcal{A}h(z)=2zh''(z)+\delta h'(z). A h ( z ) = 2 z h ′′ ( z ) + δ h ′ ( z ) .
Writing x = T ( z ) x=T(z) x = T ( z ) and using a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) and δ = 2 / ρ \delta=2/\rho δ = 2/ ρ , direct differentiation gives
A h ( z ) = κ x 2 − ρ ϕ ′ ′ ( x ) + κ ( 2 − ρ ) x 1 − ρ ϕ ′ ( x ) = L ρ , κ ϕ ( x ) . (11) \mathcal{A}h(z)=\kappa x^{2-\rho}\phi''(x)+\kappa(2-\rho)x^{1-\rho}\phi'(x)=L_{\rho,\kappa}\phi(x). \tag*{(11)} A h ( z ) = κ x 2 − ρ ϕ ′′ ( x ) + κ ( 2 − ρ ) x 1 − ρ ϕ ′ ( x ) = L ρ , κ ϕ ( x ) . (11)
In particular, L ρ , κ ϕ L_{\rho,\kappa}\phi L ρ , κ ϕ extends by zero to an element of C 0 ( [ 0 , ∞ ) ) C_0([0,\infty)) C 0 ([ 0 , ∞ )) . Applying Itô's formula to h ( Z t ) h(Z_t) h ( Z t ) gives
h ( Z t ) − h ( Z 0 ) = ∫ 0 t A h ( Z s ) d s + 2 ∫ 0 t Z s h ′ ( Z s ) d B s . h(Z_t)-h(Z_0)=\int_0^t\mathcal{A}h(Z_s)\,\mathrm{d}s+2\int_0^t\sqrt{Z_s}\,h'(Z_s)\,\mathrm{d}B_s. h ( Z t ) − h ( Z 0 ) = ∫ 0 t A h ( Z s ) d s + 2 ∫ 0 t Z s h ′ ( Z s ) d B s .
Since h ′ h' h ′ has compact support in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , z ↦ z h ′ ( z ) z\mapsto\sqrt{z}h'(z) z ↦ z h ′ ( z ) is bounded, so the stochastic integral is a martingale. Moreover, A h = ( L ρ , κ ϕ ) ∘ T \mathcal{A}h=(L_{\rho,\kappa}\phi)\circ T A h = ( L ρ , κ ϕ ) ∘ T is bounded. Hence Fubini's theorem and Y s = T ( Z s ) Y_s=T(Z_s) Y s = T ( Z s ) give
P t ϕ ( x ) − ϕ ( x ) = ∫ 0 t P s ( L ρ , κ ϕ ) ( x ) d s , x ∈ [ 0 , ∞ ) . P_t\phi(x)-\phi(x)=\int_0^tP_s(L_{\rho,\kappa}\phi)(x)\,\mathrm{d}s,\qquad x\in[0,\infty). P t ϕ ( x ) − ϕ ( x ) = ∫ 0 t P s ( L ρ , κ ϕ ) ( x ) d s , x ∈ [ 0 , ∞ ) .
Using this identity and the strong continuity of the Feller semigroup,
∥ P t ϕ − ϕ t − L ρ , κ ϕ ∥ ∞ ≤ 1 t ∫ 0 t ∥ P s ( L ρ , κ ϕ ) − L ρ , κ ϕ ∥ ∞ d s ⟶ 0. \left\|\frac{P_t\phi-\phi}{t}-L_{\rho,\kappa}\phi\right\|_\infty\leq\frac{1}{t}\int_0^t\left\|P_s(L_{\rho,\kappa}\phi)-L_{\rho,\kappa}\phi\right\|_\infty\,\mathrm{d}s\longrightarrow0. t P t ϕ − ϕ − L ρ , κ ϕ ∞ ≤ t 1 ∫ 0 t ∥ P s ( L ρ , κ ϕ ) − L ρ , κ ϕ ∥ ∞ d s ⟶ 0.
Thus C c 2 ( ( 0 , ∞ ) ) C_c^2((0,\infty)) C c 2 (( 0 , ∞ )) is contained in the strong generator domain and the generator is given there by (3.4).
Transition density and invariance. Put
ν = δ 2 − 1 = 1 ρ − 1 > − 1. \nu=\frac{\delta}{2}-1=\frac{1}{\rho}-1>-1. ν = 2 δ − 1 = ρ 1 − 1 > − 1.
For z , w > 0 z,w>0 z , w > 0 , the squared-Bessel transition density is
q t ( z , w ) = 1 2 t ( w z ) ν / 2 I ν ( z w t ) exp ( − z + w 2 t ) , q_t(z,w)=\frac{1}{2t}\left(\frac{w}{z}\right)^{\nu/2}I_\nu\left(\frac{\sqrt{zw}}{t}\right)\exp\left(-\frac{z+w}{2t}\right), q t ( z , w ) = 2 t 1 ( z w ) ν /2 I ν ( t z w ) exp ( − 2 t z + w ) ,
where I ν I_\nu I ν is the modified Bessel function of the first kind. If the squared-Bessel process starts from zero, its transition density is
q t ( 0 , w ) = w ν ( 2 t ) ν + 1 Γ ( ν + 1 ) exp ( − w 2 t ) , w > 0. q_t(0,w)=\frac{w^\nu}{(2t)^{\nu+1}\Gamma(\nu+1)}\exp\left(-\frac{w}{2t}\right),\qquad w>0. q t ( 0 , w ) = ( 2 t ) ν + 1 Γ ( ν + 1 ) w ν exp ( − 2 t w ) , w > 0.
Here Γ ( ⋅ ) \Gamma(\cdot) Γ ( ⋅ ) denotes Euler’s gamma function; see Revuz and Yor [13, Corollary (1.4), pp. 441–442]. For x , y > 0 x,y>0 x , y > 0 , the change of variables w = T − 1 ( y ) = a ρ , κ y ρ w=T^{-1}(y)=a_{\rho,\kappa}y^\rho w = T − 1 ( y ) = a ρ , κ y ρ , with d w = a ρ , κ ρ y ρ − 1 d y \mathrm{d}w=a_{\rho,\kappa}\rho y^{\rho-1}\,\mathrm{d}y d w = a ρ , κ ρ y ρ − 1 d y , gives
p t ( ρ ) ( x , y ) = q t ( a ρ , κ x ρ , a ρ , κ y ρ ) a ρ , κ ρ y ρ − 1 . p_t^{(\rho)}(x,y)=q_t(a_{\rho,\kappa}x^\rho,a_{\rho,\kappa}y^\rho)a_{\rho,\kappa}\rho y^{\rho-1}. p t ( ρ ) ( x , y ) = q t ( a ρ , κ x ρ , a ρ , κ y ρ ) a ρ , κ ρ y ρ − 1 .
Since a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) and ρ ν = 1 − ρ \rho\nu=1-\rho ρ ν = 1 − ρ , this gives
p t ( ρ ) ( x , y ) = ( x y ) ( ρ − 1 ) / 2 κ ρ t exp ( − x ρ + y ρ κ ρ 2 t ) I ν ( 2 ( x y ) ρ / 2 κ ρ 2 t ) , x , y > 0. (12) p_t^{(\rho)}(x,y)=\frac{(xy)^{(\rho-1)/2}}{\kappa\rho t}\exp\left(-\frac{x^\rho+y^\rho}{\kappa\rho^2t}\right)I_\nu\left(\frac{2(xy)^{\rho/2}}{\kappa\rho^2t}\right),\qquad x,y>0. \tag*{(12)} p t ( ρ ) ( x , y ) = κ ρt ( x y ) ( ρ − 1 ) /2 exp ( − κ ρ 2 t x ρ + y ρ ) I ν ( κ ρ 2 t 2 ( x y ) ρ /2 ) , x , y > 0. (12)
When x = 0 x=0 x = 0 , applying the same change of variables to q t ( 0 , w ) q_t(0,w) q t ( 0 , w ) and using ν + 1 = 1 / ρ \nu+1=1/\rho ν + 1 = 1/ ρ yields
p t ( ρ ) ( 0 , y ) = ρ Γ ( 1 / ρ ) ( κ ρ 2 t ) 1 / ρ exp ( − y ρ κ ρ 2 t ) , y > 0. (13) p_t^{(\rho)}(0,y)=\frac{\rho}{\Gamma(1/\rho)(\kappa\rho^2t)^{1/\rho}}\exp\left(-\frac{y^\rho}{\kappa\rho^2t}\right),\qquad y>0. \tag*{(13)} p t ( ρ ) ( 0 , y ) = Γ ( 1/ ρ ) ( κ ρ 2 t ) 1/ ρ ρ exp ( − κ ρ 2 t y ρ ) , y > 0. (13)
To obtain a continuous expression on the closed quadrant, define
I ~ ν ( r ) : = r − ν I ν ( r ) , r > 0. \widetilde I_\nu(r):=r^{-\nu}I_\nu(r),\qquad r>0. I ν ( r ) := r − ν I ν ( r ) , r > 0.
By Equation (10.30.1) of Olver et al. [14 ] , this function extends continuously to [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with
I ~ ν ( 0 ) : = lim r ↓ 0 I ~ ν ( r ) = 2 − ν Γ ( ν + 1 ) . \widetilde I_\nu(0):=\lim_{r\downarrow0}\widetilde I_\nu(r)=\frac{2^{-\nu}}{\Gamma(\nu+1)}. I ν ( 0 ) := r ↓ 0 lim I ν ( r ) = Γ ( ν + 1 ) 2 − ν .
Since I ν ( r ) = r ν I ~ ν ( r ) I_\nu(r)=r^\nu\widetilde I_\nu(r) I ν ( r ) = r ν I ν ( r ) and ρ ν = 1 − ρ \rho\nu=1-\rho ρ ν = 1 − ρ , (3.6) becomes, for x , y > 0 x,y>0 x , y > 0 ,
p t ( ρ ) ( x , y ) = [ 2 / ( κ ρ 2 t ) ] ν κ ρ t exp ( − x ρ + y ρ κ ρ 2 t ) I ~ ν ( 2 ( x y ) ρ / 2 κ ρ 2 t ) . p_t^{(\rho)}(x,y)=\frac{[2/(\kappa\rho^2t)]^\nu}{\kappa\rho t}\exp\left(-\frac{x^\rho+y^\rho}{\kappa\rho^2t}\right)\widetilde I_\nu\left(\frac{2(xy)^{\rho/2}}{\kappa\rho^2t}\right). p t ( ρ ) ( x , y ) = κ ρt [ 2/ ( κ ρ 2 t ) ] ν exp ( − κ ρ 2 t x ρ + y ρ ) I ν ( κ ρ 2 t 2 ( x y ) ρ /2 ) .
The right-hand side extends continuously to x , y ≥ 0 x,y\geq0 x , y ≥ 0 , and at x = 0 x=0 x = 0 it agrees with (3.7). It therefore defines a jointly continuous, strictly positive, symmetric version of the transition density on [ 0 , ∞ ) 2 [0,\infty)^2 [ 0 , ∞ ) 2 .
Since Y Y Y is conservative and p t ( ρ ) p_t^{(\rho)} p t ( ρ ) is symmetric, the symmetric-kernel observation gives invariance of Lebesgue measure, reversibility, and self-adjointness on L 2 ( [ 0 , ∞ ) , λ ) L^2([0,\infty),\lambda) L 2 ([ 0 , ∞ ) , λ ) .
Scaling and local dispersal. Direct substitution in the continuous transition density gives
p t ( ρ ) ( x , y ) = t − 1 / ρ p 1 ( ρ ) ( t − 1 / ρ x , t − 1 / ρ y ) , x , y ≥ 0. (14) p_t^{(\rho)}(x,y)=t^{-1/\rho}p_1^{(\rho)}(t^{-1/\rho}x,t^{-1/\rho}y),\qquad x,y\geq0. \tag*{(14)} p t ( ρ ) ( x , y ) = t − 1/ ρ p 1 ( ρ ) ( t − 1/ ρ x , t − 1/ ρ y ) , x , y ≥ 0. (14)
We claim that, for every R < ∞ R < \infty R < ∞ ,
sup x ≥ 0 , 0 ≤ y ≤ R p 1 ( ρ ) ( x , y ) < ∞ . (15) \sup_{\substack{x \ge0,\,0 \le y \le R}} p_1^{(\rho)}(x,y) < \infty. \tag*{(15)} x ≥ 0 , 0 ≤ y ≤ R sup p 1 ( ρ ) ( x , y ) < ∞. (15)
Set
r = 2 ( x y ) ρ / 2 κ ρ 2 . r = \frac{2(xy)^{\rho/2}}{\kappa\rho^2}. r = κ ρ 2 2 ( x y ) ρ /2 .
If 0 ≤ r ≤ 1 0 \le r \le1 0 ≤ r ≤ 1 , the continuous representation of the density in terms of I ~ ν \widetilde{I}_{\nu} I ν gives
p 1 ( ρ ) ( x , y ) = [ 2 / ( κ ρ 2 ) ] ν κ ρ exp ( − x ρ + y ρ κ ρ 2 ) I ~ ν ( r ) . p_1^{(\rho)}(x,y) = \frac{[2/(\kappa\rho^2)]^\nu}{\kappa\rho}\exp\left(-\frac{x^\rho+y^\rho}{\kappa\rho^2}\right)\widetilde{I}_{\nu}(r). p 1 ( ρ ) ( x , y ) = κ ρ [ 2/ ( κ ρ 2 ) ] ν exp ( − κ ρ 2 x ρ + y ρ ) I ν ( r ) .
Since I ~ ν \widetilde{I}_{\nu} I ν is continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] , it is bounded there; hence p 1 ( ρ ) p_1^{(\rho)} p 1 ( ρ ) is uniformly bounded on the region 0 ≤ r ≤ 1 0 \le r \le1 0 ≤ r ≤ 1 .
For fixed ν \nu ν , Equation (10.40.1) of Olver et al. [14 ] gives
I ν ( r ) = e r 2 π r ( 1 + O ( r − 1 ) ) , r → ∞ . I_{\nu}(r) = \frac{e^r}{\sqrt{2\pi r}}(1+O(r^{-1})), \qquad r \to\infty. I ν ( r ) = 2 π r e r ( 1 + O ( r − 1 )) , r → ∞.
Hence, after increasing the constant on a bounded interval if necessary,
I ν ( r ) ≤ C r − 1 / 2 e r , r ≥ 1. I_{\nu}(r) \le Cr^{-1/2}e^r,\qquad r \ge1. I ν ( r ) ≤ C r − 1/2 e r , r ≥ 1.
Substitution in (3.6) gives
p 1 ( ρ ) ( x , y ) ≤ C ( x y ) ( ρ − 2 ) / 4 exp [ − ( x ρ / 2 − y ρ / 2 ) 2 κ ρ 2 ] . (16) p_1^{(\rho)}(x,y) \le C(xy)^{(\rho-2)/4}\exp\left[-\frac{(x^{\rho/2}-y^{\rho/2})^2}{\kappa\rho^2}\right]. \tag*{(16)} p 1 ( ρ ) ( x , y ) ≤ C ( x y ) ( ρ − 2 ) /4 exp [ − κ ρ 2 ( x ρ /2 − y ρ /2 ) 2 ] . (16)
Moreover, r ≥ 1 r \ge1 r ≥ 1 implies
x y ≥ ( κ ρ 2 2 ) 2 / ρ . xy \ge\left(\frac{\kappa\rho^2}{2}\right)^{2/\rho}. x y ≥ ( 2 κ ρ 2 ) 2/ ρ .
If 0 < ρ ≤ 2 0 < \rho\le2 0 < ρ ≤ 2 , the exponent ( ρ − 2 ) / 4 (\rho-2)/4 ( ρ − 2 ) /4 is nonpositive, so (3.10) is uniformly bounded. If ρ > 2 \rho> 2 ρ > 2 , then, for 0 ≤ y ≤ R 0 \le y \le R 0 ≤ y ≤ R ,
( x y ) ( ρ − 2 ) / 4 ≤ C R x ( ρ − 2 ) / 4 . (xy)^{(\rho-2)/4} \le C_Rx^{(\rho-2)/4}. ( x y ) ( ρ − 2 ) /4 ≤ C R x ( ρ − 2 ) /4 .
For all sufficiently large x x x , uniformly in 0 ≤ y ≤ R 0 \le y \le R 0 ≤ y ≤ R ,
( x ρ / 2 − y ρ / 2 ) 2 ≥ 1 4 x ρ . (x^{\rho/2}-y^{\rho/2})^2 \ge\frac{1}{4}x^\rho. ( x ρ /2 − y ρ /2 ) 2 ≥ 4 1 x ρ .
Hence, on this range,
p 1 ( ρ ) ( x , y ) ≤ C R x ( ρ − 2 ) / 4 e − c x ρ , p_1^{(\rho)}(x,y) \le C_Rx^{(\rho-2)/4}e^{-cx^\rho}, p 1 ( ρ ) ( x , y ) ≤ C R x ( ρ − 2 ) /4 e − c x ρ ,
which is uniformly bounded. The remaining region with bounded x x x and r ≥ 1 r \ge1 r ≥ 1 is compact and stays away from the axes. This proves (15) .
Products. For the independent d d d -fold product, the transition density is
p t [ d ] ( x , y ) = ∏ j = 1 d p t ( ρ ) ( x j , y j ) . p_t^{[d]}(x,y)=\prod_{j=1}^{d}p_t^{(\rho)}(x_j,y_j). p t [ d ] ( x , y ) = j = 1 ∏ d p t ( ρ ) ( x j , y j ) .
For a product function
f ( x ) = ∏ j = 1 d f j ( x j ) , f j ∈ C 0 ( [ 0 , ∞ ) ) , f(x)=\prod_{j=1}^{d}f_j(x_j),\qquad f_j\in C_0([0,\infty)), f ( x ) = j = 1 ∏ d f j ( x j ) , f j ∈ C 0 ([ 0 , ∞ )) ,
the product semigroup satisfies
P t [ d ] f ( x ) = ∏ j = 1 d P t f j ( x j ) . P_t^{[d]}f(x)=\prod_{j=1}^{d}P_tf_j(x_j). P t [ d ] f ( x ) = j = 1 ∏ d P t f j ( x j ) .
Finite linear combinations of such product functions are dense in C 0 ( [ 0 , ∞ ) d ) C_0([0,\infty)^d) C 0 ([ 0 , ∞ ) d ) . The strong continuity of the one-dimensional semigroup, together with contractivity and density, therefore implies that ( P t [ d ] ) t ≥ 0 (P_t^{[d]})_{t\ge0} ( P t [ d ] ) t ≥ 0 is a Feller semigroup. Since each coordinate semigroup is conservative, so is the product semigroup. Moreover, the coordinatewise realization has continuous paths. Thus W W W is a conservative continuous Feller process, hence a Hunt process.
Let
f ( x ) = ∏ j = 1 d f j ( x j ) , f j ∈ C c 2 ( ( 0 , ∞ ) ) . f(x)=\prod_{j=1}^{d}f_j(x_j),\qquad f_j\in C_c^2((0,\infty)). f ( x ) = j = 1 ∏ d f j ( x j ) , f j ∈ C c 2 (( 0 , ∞ )) .
The identity
∏ j = 1 d P t f j − ∏ j = 1 d f j = ∑ j = 1 d ( P t f j − f j ) ∏ i < j P t f i ∏ i > j f i \prod_{j=1}^{d}P_tf_j-\prod_{j=1}^{d}f_j
=\sum_{j=1}^{d}(P_tf_j-f_j)\prod_{i<j}P_tf_i\prod_{i>j}f_i j = 1 ∏ d P t f j − j = 1 ∏ d f j = j = 1 ∑ d ( P t f j − f j ) i < j ∏ P t f i i > j ∏ f i
shows, after division by t t t and passage to the uniform limit as t ↓ 0 t\downarrow0 t ↓ 0 , that f f f belongs to the strong generator domain and
L [ d ] f ( x ) = κ ∑ j = 1 d ∂ ∂ x j ( x j 2 − ρ ∂ f ∂ x j ( x ) ) . L^{[d]}f(x)=\kappa\sum_{j=1}^{d}\frac{\partial}{\partial x_j}\left(x_j^{2-\rho}\frac{\partial f}{\partial x_j}(x)\right). L [ d ] f ( x ) = κ j = 1 ∑ d ∂ x j ∂ ( x j 2 − ρ ∂ x j ∂ f ( x ) ) .
The same conclusion holds for finite linear combinations of such product functions.
The product process is conservative and its transition density is symmetric. Hence the symmetric-kernel observation gives (H1), reversibility, and self-adjointness on L 2 ( [ 0 , ∞ ) d , λ ) L^2([0,\infty)^d,\lambda) L 2 ([ 0 , ∞ ) d , λ ) .
To verify (H2), let K ⊂ [ 0 , R ] d K \subset[0,R]^d K ⊂ [ 0 , R ] d be compact and set
M R : = sup u ≥ 0 , 0 ≤ v ≤ R p 1 ( ρ ) ( u , v ) < ∞ , M_R := \sup_{u\geq0,\,0\leq v\leq R} p_1^{(\rho)}(u,v) < \infty, M R := u ≥ 0 , 0 ≤ v ≤ R sup p 1 ( ρ ) ( u , v ) < ∞ ,
as established in (15) . By (14) , for t ≥ 1 t\geq1 t ≥ 1 ,
p t ( ρ ) ( x j , y j ) ≤ t − 1 / ρ M R , x j ≥ 0 , 0 ≤ y j ≤ R . p_t^{(\rho)}(x_j,y_j) \leq t^{-1/\rho}M_R,\qquad x_j\geq0,\quad0\leq y_j\leq R. p t ( ρ ) ( x j , y j ) ≤ t − 1/ ρ M R , x j ≥ 0 , 0 ≤ y j ≤ R .
Therefore,
sup x ∈ [ 0 , ∞ ) d P t [ d ] 1 K ( x ) ≤ λ ( K ) M R d t − d / ρ , t ≥ 1. \sup_{x\in[0,\infty)^d}P_t^{[d]}1_K(x)\leq\lambda(K)M_R^d t^{-d/\rho},\qquad t\geq1. x ∈ [ 0 , ∞ ) d sup P t [ d ] 1 K ( x ) ≤ λ ( K ) M R d t − d / ρ , t ≥ 1.
Hence (H2) holds with q = d / ρ q=d/\rho q = d / ρ .
Localization. Let Y Y Y start from z ≥ 0 z\geq0 z ≥ 0 , so that Z s = a ρ , κ Y s ρ Z_s=a_{\rho,\kappa}Y_s^\rho Z s = a ρ , κ Y s ρ and Z 0 = a ρ , κ z ρ Z_0=a_{\rho,\kappa}z^\rho Z 0 = a ρ , κ z ρ . For m > Z 0 m>Z_0 m > Z 0 , set τ m = inf { s ≥ 0 : Z s ≥ m } \tau_m=\inf\{s\geq0:Z_s\geq m\} τ m = inf { s ≥ 0 : Z s ≥ m } . The stochastic integral in (9) , stopped at 1 ∧ τ m 1\wedge\tau_m 1 ∧ τ m , has bounded integrand and hence mean zero. Thus
E [ Z 1 ∧ τ m ] = a ρ , κ z ρ + δ E [ 1 ∧ τ m ] ≤ a ρ , κ z ρ + δ . \mathbb{E}[Z_{1\wedge\tau_m}]=a_{\rho,\kappa}z^\rho+\delta\mathbb{E}[1\wedge\tau_m]\leq a_{\rho,\kappa}z^\rho+\delta. E [ Z 1 ∧ τ m ] = a ρ , κ z ρ + δ E [ 1 ∧ τ m ] ≤ a ρ , κ z ρ + δ .
Conservativity and path continuity imply τ m ↑ ∞ \tau_m\uparrow\infty τ m ↑ ∞ almost surely. Fatou’s lemma therefore yields
E z [ a ρ , κ Y 1 ρ ] = E [ Z 1 ] ≤ a ρ , κ z ρ + δ . \mathbb{E}_z[a_{\rho,\kappa}Y_1^\rho]=\mathbb{E}[Z_1]\leq a_{\rho,\kappa}z^\rho+\delta. E z [ a ρ , κ Y 1 ρ ] = E [ Z 1 ] ≤ a ρ , κ z ρ + δ .
Markov’s inequality then yields
P z { Y 1 > N } ≤ a ρ , κ z ρ + δ a ρ , κ N ρ , z ≥ 0 , N > 0. (17) \mathbb{P}_z\{Y_1>N\}\leq\frac{a_{\rho,\kappa}z^\rho+\delta}{a_{\rho,\kappa}N^\rho},\qquad z\geq0,\quad N>0. \tag*{(17)} P z { Y 1 > N } ≤ a ρ , κ N ρ a ρ , κ z ρ + δ , z ≥ 0 , N > 0. (17)
We now verify (H3). Let
D t ( N ) = [ 0 , N t 1 / ρ ] d D_t(N)=[0,Nt^{1/\rho}]^d D t ( N ) = [ 0 , N t 1/ ρ ] d
and let K ⊂ [ 0 , R ] d K\subset[0,R]^d K ⊂ [ 0 , R ] d be compact. The localization identity, the product form of (14) , and the change of variables u = t − 1 / ρ x u=t^{-1/\rho}x u = t − 1/ ρ x give
∫ [ 0 , ∞ ) d ∖ D t ( N ) P t [ d ] 1 K ( x ) λ ( d x ) = ∫ K P t − 1 / ρ y { W 1 ∉ [ 0 , N ] d } λ ( d y ) . (18) \int_{[0,\infty)^d\setminus D_t(N)}P_t^{[d]}1_K(x)\lambda(dx)
=
\int_K\mathbb{P}_{t^{-1/\rho}y}\{W_1\notin[0,N]^d\}\lambda(dy). \tag*{(18)} ∫ [ 0 , ∞ ) d ∖ D t ( N ) P t [ d ] 1 K ( x ) λ ( d x ) = ∫ K P t − 1/ ρ y { W 1 ∈ / [ 0 , N ] d } λ ( d y ) . (18)
Let t ≥ 1 t\geq1 t ≥ 1 and y ∈ K y\in K y ∈ K . Then every coordinate of t − 1 / ρ y t^{-1/\rho}y t − 1/ ρ y belongs to [ 0 , R ] [0,R] [ 0 , R ] . Hence, by the union bound and (17) ,
P t − 1 / ρ y { W 1 ∉ [ 0 , N ] d } ≤ ∑ j = 1 d P t − 1 / ρ y j { Y 1 > N } ≤ d a ρ , κ R ρ + δ a ρ , κ N ρ . \mathbb{P}_{t^{-1/\rho}y}\{W_1\notin[0,N]^d\}
\leq\sum_{j=1}^{d}\mathbb{P}_{t^{-1/\rho}y_j}\{Y_1>N\}
\leq d\frac{a_{\rho,\kappa}R^\rho+\delta}{a_{\rho,\kappa}N^\rho}. P t − 1/ ρ y { W 1 ∈ / [ 0 , N ] d } ≤ j = 1 ∑ d P t − 1/ ρ y j { Y 1 > N } ≤ d a ρ , κ N ρ a ρ , κ R ρ + δ .
Substituting this estimate into (18) yields
sup t ≥ 1 ∫ [ 0 , ∞ ) d ∖ D t ( N ) P t [ d ] 1 K ( x ) λ ( d x ) ≤ λ ( K ) d a ρ , κ R ρ + δ a ρ , κ N ρ ⟶ 0 \sup_{t\ge1}\int_{[0,\infty)^d\setminus D_t(N)}P_t^{[d]}1_K(x)\,\lambda(\mathrm{d}x)\le\lambda(K)d\frac{a_{\rho,\kappa}R^\rho+\delta}{a_{\rho,\kappa}N^\rho}\longrightarrow0 t ≥ 1 sup ∫ [ 0 , ∞ ) d ∖ D t ( N ) P t [ d ] 1 K ( x ) λ ( d x ) ≤ λ ( K ) d a ρ , κ N ρ a ρ , κ R ρ + δ ⟶ 0
as N → ∞ N\to\infty N → ∞ . Finally, λ ( D t ( N ) ) = N d t d / ρ \lambda(D_t(N))=N^d t^{d/\rho} λ ( D t ( N )) = N d t d / ρ , so the volume bound in (H3) holds with C N = N d C_N=N^d C N = N d . Together with the preceding localization estimate, this proves (H3).
Symmetric bilateral Bessel-type motions on R d \mathbb{R}^d R d
Fix ρ > 1 \rho>1 ρ > 1 and κ > 0 \kappa>0 κ > 0 , retain a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) , and set
δ = 2 ρ ∈ ( 0 , 2 ) , ν = δ 2 − 1 = 1 ρ − 1 ∈ ( − 1 , 0 ) . \delta=\frac{2}{\rho}\in(0,2),\qquad\nu=\frac{\delta}{2}-1=\frac{1}{\rho}-1\in(-1,0). δ = ρ 2 ∈ ( 0 , 2 ) , ν = 2 δ − 1 = ρ 1 − 1 ∈ ( − 1 , 0 ) .
Let
R ∼ SSBES ( ν , 1 / 2 , 1 , 1 ) R\sim\operatorname{SSBES}(\nu,1/2,1,1) R ∼ SSBES ( ν , 1/2 , 1 , 1 )
be the symmetric bilateral Bessel diffusion of Alili and Aylwin [15 Section 2] . By Alili and Aylwin [15 Lemma 1] , ∣ R ∣ |R| ∣ R ∣ is a Bessel process of index ν \nu ν ; equivalently, R 2 R^2 R 2 is a squared-Bessel process of dimension δ \delta δ . Moreover, the excursion construction in Alili and Aylwin [15 Remark 6] shows that, after the first visit to zero, the excursions of ∣ R ∣ |R| ∣ R ∣ are assigned independent signs, with probabilities 1 / 2 1/2 1/2 and 1 / 2 1/2 1/2 . The restriction ρ > 1 \rho>1 ρ > 1 is precisely the range ν ∈ ( − 1 , 0 ) \nu\in(-1,0) ν ∈ ( − 1 , 0 ) considered in this bilateral construction.
Define the homeomorphism
T ρ , κ ( r ) = a ρ , κ − 1 / ρ sgn ( r ) ∣ r ∣ 2 / ρ , T ρ , κ − 1 ( x ) = a ρ , κ 1 / 2 sgn ( x ) ∣ x ∣ ρ / 2 . (19) T_{\rho,\kappa}(r)=a_{\rho,\kappa}^{-1/\rho}\operatorname{sgn}(r)|r|^{2/\rho},\qquad T_{\rho,\kappa}^{-1}(x)=a_{\rho,\kappa}^{1/2}\operatorname{sgn}(x)|x|^{\rho/2}. \tag*{(19)} T ρ , κ ( r ) = a ρ , κ − 1/ ρ sgn ( r ) ∣ r ∣ 2/ ρ , T ρ , κ − 1 ( x ) = a ρ , κ 1/2 sgn ( x ) ∣ x ∣ ρ /2 . (19)
For x ∈ R x\in\mathbb{R} x ∈ R , start R R R from T ρ , κ − 1 ( x ) T_{\rho,\kappa}^{-1}(x) T ρ , κ − 1 ( x ) and set
Y ~ t = T ρ , κ ( R t ) . \widetilde{Y}_t=T_{\rho,\kappa}(R_t). Y t = T ρ , κ ( R t ) .
Then Y ~ 0 = x \widetilde{Y}_0=x Y 0 = x .
Proposition 3.5. The process Y ~ \widetilde{Y} Y is a conservative Hunt diffusion on R \mathbb{R} R . For every t > 0 t>0 t > 0 , its transition kernel has a strictly positive density with respect to Lebesgue measure, and its semigroup is reversible with respect to Lebesgue measure. On C c 2 ( R ∖ { 0 } ) C^2_c(\mathbb{R}\setminus\{0\}) C c 2 ( R ∖ { 0 }) , its extended infinitesimal generator is
L ~ ρ , κ f ( x ) = κ d d x ( ∣ x ∣ 2 − ρ f ′ ( x ) ) . (20) \widetilde{L}_{\rho,\kappa}f(x)=\kappa\frac{\mathrm{d}}{\mathrm{d}x}\left(|x|^{2-\rho}f'(x)\right). \tag*{(20)} L ρ , κ f ( x ) = κ d x d ( ∣ x ∣ 2 − ρ f ′ ( x ) ) . (20)
The independent product
W ~ = ( Y ~ ( 1 ) , … , Y ~ ( d ) ) \widetilde{W}=\left(\widetilde{Y}^{(1)},\ldots,\widetilde{Y}^{(d)}\right) W = ( Y ( 1 ) , … , Y ( d ) )
is a conservative Hunt process on R d \mathbb{R}^{d} R d satisfying (H1)–(H3) with
q = d ρ , D t ( N ) = [ − N t 1 / ρ , N t 1 / ρ ] d . q=\frac{d}{\rho},\qquad D_t(N)=[-Nt^{1/\rho},Nt^{1/\rho}]^d. q = ρ d , D t ( N ) = [ − N t 1/ ρ , N t 1/ ρ ] d .
Every nonsingular linear image of W ~ \widetilde{W} W also satisfies (H1)–(H3) with the same exponent.
Proof. One-dimensional construction and transition density. By definition, R \mathrm{R} R is a conservative continuous strong Markov diffusion. Since T ρ , κ T_{\rho,\kappa} T ρ , κ is a homeomorphism of R \mathbb{R} R , Y ~ = T ρ , κ ( R ) \widetilde{Y}=T_{\rho,\kappa}(\mathrm{R}) Y = T ρ , κ ( R ) is again a conservative continuous strong Markov process, hence a Hunt diffusion.
Let q t R ( r , s ) q_t^{\mathrm{R}}(r,s) q t R ( r , s ) denote the transition density of R \mathrm{R} R . For p = 1 / 2 p=1/2 p = 1/2 and equal scale parameters, Theorem 2 of Alili and Aylwin [15 ] gives, for r , s ≠ 0 r,s\ne0 r , s = 0 ,
q t R ( r , s ) = ∣ s ∣ 2 t ( ∣ s ∣ ∣ r ∣ ) ν exp ( − r 2 + s 2 2 t ) { I ν ( ∣ r s ∣ t ) + I − ν ( ∣ r s ∣ t ) , r s > 0 , I ν ( ∣ r s ∣ t ) − I − ν ( ∣ r s ∣ t ) , r s < 0. q_t^{\mathrm{R}}(r,s)=\frac{|s|}{2t}\left(\frac{|s|}{|r|}\right)^{\nu}\exp\left(-\frac{r^2+s^2}{2t}\right)
\begin{cases}
I_{\nu}\left(\frac{|rs|}{t}\right)+I_{-\nu}\left(\frac{|rs|}{t}\right), & rs>0,\\
I_{\nu}\left(\frac{|rs|}{t}\right)-I_{-\nu}\left(\frac{|rs|}{t}\right), & rs<0.
\end{cases} q t R ( r , s ) = 2 t ∣ s ∣ ( ∣ r ∣ ∣ s ∣ ) ν exp ( − 2 t r 2 + s 2 ) ⎩ ⎨ ⎧ I ν ( t ∣ r s ∣ ) + I − ν ( t ∣ r s ∣ ) , I ν ( t ∣ r s ∣ ) − I − ν ( t ∣ r s ∣ ) , r s > 0 , r s < 0.
Here I α I_{\alpha} I α denotes the modified Bessel function of the first kind. For x , y ≠ 0 x,y\ne0 x , y = 0 , changing variables by s = T ρ , κ − 1 ( y ) s=T_{\rho,\kappa}^{-1}(y) s = T ρ , κ − 1 ( y ) gives the Jacobian
d s = ρ 2 a ρ , κ 1 / 2 ∣ y ∣ ρ / 2 − 1 d y ds=\frac{\rho}{2}a_{\rho,\kappa}^{1/2}|y|^{\rho/2-1}\,dy d s = 2 ρ a ρ , κ 1/2 ∣ y ∣ ρ /2 − 1 d y
and hence
p ~ t ( ρ ) ( x , y ) = q t R ( T ρ , κ − 1 ( x ) , T ρ , κ − 1 ( y ) ) ρ 2 a ρ , κ 1 / 2 ∣ y ∣ ρ / 2 − 1 . \widetilde{p}_t^{(\rho)}(x,y)=q_t^{\mathrm{R}}\left(T_{\rho,\kappa}^{-1}(x),T_{\rho,\kappa}^{-1}(y)\right)\frac{\rho}{2}a_{\rho,\kappa}^{1/2}|y|^{\rho/2-1}. p t ( ρ ) ( x , y ) = q t R ( T ρ , κ − 1 ( x ) , T ρ , κ − 1 ( y ) ) 2 ρ a ρ , κ 1/2 ∣ y ∣ ρ /2 − 1 .
Since ρ ν = 1 − ρ \rho\nu=1-\rho ρ ν = 1 − ρ and a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) , this simplifies to
p ~ t ( ρ ) ( x , y ) = ∣ x y ∣ ( ρ − 1 ) / 2 2 κ ρ t exp ( − ∣ x ∣ ρ + ∣ y ∣ ρ κ ρ 2 t ) { I ν ( u ) + I − ν ( u ) , x y > 0 , I ν ( u ) − I − ν ( u ) , x y < 0 , (21) \widetilde{p}_t^{(\rho)}(x,y)=\frac{|xy|^{(\rho-1)/2}}{2\kappa\rho t}\exp\left(-\frac{|x|^\rho+|y|^\rho}{\kappa\rho^2t}\right)
\begin{cases}
I_{\nu}(u)+I_{-\nu}(u), & xy>0,\\
I_{\nu}(u)-I_{-\nu}(u), & xy<0,
\end{cases} \tag*{(21)} p t ( ρ ) ( x , y ) = 2 κ ρt ∣ x y ∣ ( ρ − 1 ) /2 exp ( − κ ρ 2 t ∣ x ∣ ρ + ∣ y ∣ ρ ) { I ν ( u ) + I − ν ( u ) , I ν ( u ) − I − ν ( u ) , x y > 0 , x y < 0 , (21)
where
u = 2 ∣ x y ∣ ρ / 2 κ ρ 2 t . u=\frac{2|xy|^{\rho/2}}{\kappa\rho^2t}. u = κ ρ 2 t 2∣ x y ∣ ρ /2 .
The density in (21) is strictly positive. Indeed, for u > 0 u>0 u > 0 and α > − 1 \alpha>-1 α > − 1 , the power series
I α ( u ) = ( u 2 ) α ∑ k = 0 ∞ ( u 2 / 4 ) k k ! Γ ( α + k + 1 ) I_{\alpha}(u)=\left(\frac{u}{2}\right)^{\alpha}\sum_{k=0}^{\infty}\frac{(u^2/4)^k}{k!\Gamma(\alpha+k+1)} I α ( u ) = ( 2 u ) α k = 0 ∑ ∞ k ! Γ ( α + k + 1 ) ( u 2 /4 ) k
has strictly positive terms; see Olver et al. [14 Eq. (10.25.2)] . Since ν \nu ν , − ν > − 1 -\nu> -1 − ν > − 1 , it follows that I ν ( u ) > 0 I_{\nu}(u) > 0 I ν ( u ) > 0 and I − ν ( u ) > 0 I_{-\nu}(u) > 0 I − ν ( u ) > 0 . Hence the same-sign part of (21) is strictly positive.
For the opposite-sign part, Equation (14) of Alili and Aylwin [15 ] gives
I ν ( u ) − I − ν ( u ) = − 2 π sin ( π ν ) K ν ( u ) , I_{\nu}(u) - I_{-\nu}(u) = -\frac{2}{\pi}\sin(\pi\nu)K_{\nu}(u), I ν ( u ) − I − ν ( u ) = − π 2 sin ( π ν ) K ν ( u ) ,
where K ν K_{\nu} K ν is the modified Bessel function of the second kind.
Since u > 0 u > 0 u > 0 , the condition ∣ ph u ∣ < π / 2 |\operatorname{ph} u| < \pi/2 ∣ ph u ∣ < π /2 in Olver et al. [14 Eq. (10.32.9)] is automatically satisfied. The integral representation
K ν ( u ) = ∫ 0 ∞ e − u cosh s cosh ( ν s ) d s K_{\nu}(u) = \int_{0}^{\infty} e^{-u\cosh s}\cosh(\nu s)\,\mathrm{d}s K ν ( u ) = ∫ 0 ∞ e − u c o s h s cosh ( ν s ) d s
therefore shows that K ν ( u ) > 0 K_{\nu}(u) > 0 K ν ( u ) > 0 . Since ν ∈ ( − 1 , 0 ) \nu\in(-1,0) ν ∈ ( − 1 , 0 ) , sin ( π ν ) < 0 \sin(\pi\nu) < 0 sin ( π ν ) < 0 , and hence I ν ( u ) − I − ν ( u ) > 0 I_{\nu}(u) - I_{-\nu}(u) > 0 I ν ( u ) − I − ν ( u ) > 0 .
To identify the transition law from zero, recall the symmetric excursion signs in the construction: under P 0 \mathbb{P}_{0} P 0 , the law of R R R is invariant under reflection. Since R 2 R^{2} R 2 is a squared-Bessel process of dimension δ \delta δ , also started from zero, the transformation gives
∣ Y ~ t ∣ = a ρ , κ − 1 / ρ ∣ R t ∣ 2 / ρ = d Y t , |\widetilde{Y}_{t}| = a_{\rho,\kappa}^{-1/\rho}|R_{t}|^{2/\rho} \mathrel{\overset{\mathrm{d}}{=}} Y_{t}, ∣ Y t ∣ = a ρ , κ − 1/ ρ ∣ R t ∣ 2/ ρ = d Y t ,
where Y Y Y is the half-line process of Proposition 3.4 started from zero. Thus, for every Borel set A ⊂ ( 0 , ∞ ) A \subset(0,\infty) A ⊂ ( 0 , ∞ ) ,
P 0 { Y ~ t ∈ A } = P 0 { Y ~ t ∈ − A } = 1 2 P 0 { Y t ∈ A } . \mathbb{P}_{0}\{\widetilde{Y}_{t} \in A\} = \mathbb{P}_{0}\{\widetilde{Y}_{t} \in-A\} = \frac{1}{2}\mathbb{P}_{0}\{Y_{t} \in A\}. P 0 { Y t ∈ A } = P 0 { Y t ∈ − A } = 2 1 P 0 { Y t ∈ A } .
Therefore,
p ~ t ( ρ ) ( 0 , y ) = 1 2 p t ( ρ ) ( 0 , ∣ y ∣ ) = ρ 2 Γ ( 1 / ρ ) ( κ ρ 2 t ) 1 / ρ exp ( − ∣ y ∣ ρ κ ρ 2 t ) , y ≠ 0. (22) \widetilde{p}_{t}^{(\rho)}(0,y) = \frac{1}{2}p_{t}^{(\rho)}(0,|y|) = \frac{\rho}{2\Gamma(1/\rho)(\kappa\rho^{2}t)^{1/\rho}}\exp\left(-\frac{|y|^{\rho}}{\kappa\rho^{2}t}\right), \qquad y \ne0. \tag*{(22)} p t ( ρ ) ( 0 , y ) = 2 1 p t ( ρ ) ( 0 , ∣ y ∣ ) = 2Γ ( 1/ ρ ) ( κ ρ 2 t ) 1/ ρ ρ exp ( − κ ρ 2 t ∣ y ∣ ρ ) , y = 0. (22)
Since p t ( ρ ) ( 0 , ⋅ ) p_{t}^{(\rho)}(0,\cdot) p t ( ρ ) ( 0 , ⋅ ) integrates to one on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , (22) accounts for the entire transition law from zero; in particular, P 0 { Y ~ t = 0 } = 0 \mathbb{P}_{0}\{\widetilde{Y}_{t}=0\}=0 P 0 { Y t = 0 } = 0 . Together with (21) , this proves that the transition law from every starting point has a strictly positive density with respect to Lebesgue measure.
Reversibility and extended generator. For p = 1 / 2 p=1/2 p = 1/2 and σ + = σ − = 1 \sigma_{+}=\sigma_{-}=1 σ + = σ − = 1 , the speed measure in Alili and Aylwin [15 Eq. (2)] is
m R ( d r ) = ∣ r ∣ 2 ν + 1 d r = ∣ r ∣ δ − 1 d r . m_{\mathrm{R}}(\mathrm{d}r) = |r|^{2\nu+1}\,\mathrm{d}r = |r|^{\delta-1}\,\mathrm{d}r. m R ( d r ) = ∣ r ∣ 2 ν + 1 d r = ∣ r ∣ δ − 1 d r .
Using the Jacobian above and δ = 2 / ρ \delta= 2/\rho δ = 2/ ρ , for x > 0 x > 0 x > 0 we obtain
r δ − 1 d r = ρ 2 a ρ , κ δ / 2 x ρ δ / 2 − 1 d x = ρ 2 a ρ , κ 1 / ρ d x , r = T ρ , κ − 1 ( x ) . (23) r^{\delta-1}\,\mathrm{d}r = \frac{\rho}{2}a_{\rho,\kappa}^{\delta/2}x^{\rho\delta/2-1}\,\mathrm{d}x = \frac{\rho}{2}a_{\rho,\kappa}^{1/\rho}\,\mathrm{d}x,\qquad r=T_{\rho,\kappa}^{-1}(x). \tag*{(23)} r δ − 1 d r = 2 ρ a ρ , κ δ /2 x ρ δ /2 − 1 d x = 2 ρ a ρ , κ 1/ ρ d x , r = T ρ , κ − 1 ( x ) . (23)
The same calculation holds on the negative half-line. Therefore the image of m R m_{\mathrm{R}} m R under T ρ , κ T_{\rho,\kappa} T ρ , κ is
m Y ~ ( d x ) = ρ 2 a ρ , κ 1 / ρ d x . m_{\widetilde{\mathrm{Y}}}(\mathrm{d}x)=\frac{\rho}{2}a_{\rho,\kappa}^{1/\rho}\,\mathrm{d}x. m Y ( d x ) = 2 ρ a ρ , κ 1/ ρ d x .
The resolvent of R has a symmetric density with respect to m R m_{\mathrm{R}} m R ; see Alili and Aylwin [15 Section 2, p. 100] . Since this symmetry holds for every Laplace parameter, uniqueness of Laplace transforms implies that the semigroup of R is symmetric with respect to m R m_{\mathrm{R}} m R . Conjugation by T ρ , κ T_{\rho,\kappa} T ρ , κ then makes the semigroup of Y ~ \widetilde{\mathrm{Y}} Y symmetric with respect to m Y ~ m_{\widetilde{\mathrm{Y}}} m Y . Since m Y ~ m_{\widetilde{\mathrm{Y}}} m Y is a positive constant multiple of Lebesgue measure, the semigroup of Y ~ \widetilde{\mathrm{Y}} Y is reversible with respect to Lebesgue measure.
Now let f ∈ C c 2 ( R ∖ { 0 } ) f\in C^2_c(\mathbb{R}\setminus\{0\}) f ∈ C c 2 ( R ∖ { 0 }) , viewed as extended by zero at 0, and set
g = f ∘ T ρ , κ . g=f\circ T_{\rho,\kappa}. g = f ∘ T ρ , κ .
Since f f f has compact support in R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } , it vanishes in a neighborhood of zero, and the same is true of g g g . Hence the expression in Alili and Aylwin [15 Eq. (4)] is bounded and continuous, vanishes near zero, and satisfies the interface condition Alili and Aylwin [15 Eq. (5)] . The scale–speed characterization of the extended infinitesimal generator then implies that, under every P r \mathbb{P}_r P r ,
M t g : = g ( R t ) − g ( R 0 ) − ∫ 0 t A R g ( R s ) d s M_t^g:=g(R_t)-g(R_0)-\int_0^t\mathcal{A}_{\mathrm{R}}g(R_s)\,\mathrm{d}s M t g := g ( R t ) − g ( R 0 ) − ∫ 0 t A R g ( R s ) d s
is a local martingale; see also Revuz and Yor [13 Theorem VII.3.12, p. 308] . Since both g g g and A R g \mathcal{A}_{\mathrm{R}}g A R g are bounded, M g M^g M g is a martingale on every bounded time interval.
For r ≠ 0 r\ne0 r = 0 , Equation (4) of Alili and Aylwin [15 ] reduces, since σ + = σ − = 1 \sigma_+=\sigma_-=1 σ + = σ − = 1 , to
A R g ( r ) = 1 2 g ′ ′ ( r ) + δ − 1 2 r g ′ ( r ) . \mathcal{A}_{\mathrm{R}}g(r)=\frac{1}{2}g^{\prime\prime}(r)+\frac{\delta-1}{2r}g^\prime(r). A R g ( r ) = 2 1 g ′′ ( r ) + 2 r δ − 1 g ′ ( r ) .
Writing x = T ρ , κ ( r ) x=T_{\rho,\kappa}(r) x = T ρ , κ ( r ) , we have
T ρ , κ ′ ( r ) = 2 ρ a ρ , κ − 1 / ρ ∣ r ∣ 2 / ρ − 1 , T ρ , κ ′ ′ ( r ) = 2 ρ ( 2 ρ − 1 ) a ρ , κ − 1 / ρ sgn ( r ) ∣ r ∣ 2 / ρ − 2 . T_{\rho,\kappa}^\prime(r)=\frac{2}{\rho}a_{\rho,\kappa}^{-1/\rho}|r|^{2/\rho-1},\qquad T_{\rho,\kappa}^{\prime\prime}(r)=\frac{2}{\rho}\left(\frac{2}{\rho}-1\right)a_{\rho,\kappa}^{-1/\rho}\operatorname{sgn}(r)|r|^{2/\rho-2}. T ρ , κ ′ ( r ) = ρ 2 a ρ , κ − 1/ ρ ∣ r ∣ 2/ ρ − 1 , T ρ , κ ′′ ( r ) = ρ 2 ( ρ 2 − 1 ) a ρ , κ − 1/ ρ sgn ( r ) ∣ r ∣ 2/ ρ − 2 .
The chain rule, together with δ = 2 / ρ \delta=2/\rho δ = 2/ ρ and a ρ , κ = 2 / ( κ ρ 2 ) a_{\rho,\kappa}=2/(\kappa\rho^2) a ρ , κ = 2/ ( κ ρ 2 ) , gives
A R ( f ∘ T ρ , κ ) ( r ) = κ ∣ x ∣ 2 − ρ f ′ ′ ( x ) + κ ( 2 − ρ ) sgn ( x ) ∣ x ∣ 1 − ρ f ′ ( x ) = L ~ ρ , κ f ( x ) . \mathcal{A}_{\mathrm{R}}(f\circ T_{\rho,\kappa})(r)=\kappa|x|^{2-\rho}f^{\prime\prime}(x)+\kappa(2-\rho)\operatorname{sgn}(x)|x|^{1-\rho}f^\prime(x)=\widetilde{L}_{\rho,\kappa}f(x). A R ( f ∘ T ρ , κ ) ( r ) = κ ∣ x ∣ 2 − ρ f ′′ ( x ) + κ ( 2 − ρ ) sgn ( x ) ∣ x ∣ 1 − ρ f ′ ( x ) = L ρ , κ f ( x ) .
Because f ( Y ~ t ) = g ( R t ) f(\widetilde{Y}_t)=g(R_t) f ( Y t ) = g ( R t ) , the process
f ( Y ~ t ) − f ( Y ~ 0 ) − ∫ 0 t L ~ ρ , κ f ( Y ~ s ) d s f(\widetilde{Y}_t)-f(\widetilde{Y}_0)-\int_0^t \widetilde{L}_{\rho,\kappa}f(\widetilde{Y}_s)\,\mathrm{d}s f ( Y t ) − f ( Y 0 ) − ∫ 0 t L ρ , κ f ( Y s ) d s
is a martingale. Thus (3.14) is the extended infinitesimal generator of Y ~ \widetilde{Y} Y on C c 2 ( R ∖ { 0 } ) C^2_c(\mathbb{R}\setminus\{0\}) C c 2 ( R ∖ { 0 }) .
Products and verification of (H1)–(H3). For a general initial point x x x , the squared-Bessel process ∣ R ∣ 2 |R|^2 ∣ R ∣ 2 identified in the construction starts from ∣ T ρ , κ ( x ) ∣ 2 = a ρ , κ ∣ x ∣ ρ |T_{\rho,\kappa}(x)|^2=a_{\rho,\kappa}|x|^\rho ∣ T ρ , κ ( x ) ∣ 2 = a ρ , κ ∣ x ∣ ρ . Thus the same transformation as in Proposition 3.4 gives the path-law identity
( ∣ Y ~ t ∣ ) t ≥ 0 under P x = d ( Y t ) t ≥ 0 under P ∣ x ∣ . (24) (|\widetilde{Y}_t|)_{t\geq0}\text{ under }\mathbb{P}_x \ \mathrel{\stackrel{\mathrm{d}}{=}}\ (Y_t)_{t\geq0}\text{ under }\mathbb{P}_{|x|}. \tag*{(24)} ( ∣ Y t ∣ ) t ≥ 0 under P x = d ( Y t ) t ≥ 0 under P ∣ x ∣ . (24)
Let
W ~ = ( Y ~ ( 1 ) , … , Y ~ ( d ) ) \widetilde{W}=(\widetilde{Y}^{(1)},\ldots,\widetilde{Y}^{(d)}) W = ( Y ( 1 ) , … , Y ( d ) )
be the independent d d d -fold product, and let
W + = ( Y ( 1 ) , … , Y ( d ) ) W^+=(Y^{(1)},\ldots,Y^{(d)}) W + = ( Y ( 1 ) , … , Y ( d ) )
denote the corresponding independent half-line product from Proposition 3.4. Throughout this paragraph, absolute values of vectors are taken coordinatewise:
∣ W ~ t ∣ : = ( ∣ Y ~ t ( 1 ) ∣ , … , ∣ Y ~ t ( d ) ∣ ) , ∣ x ∣ : = ( ∣ x 1 ∣ , … , ∣ x d ∣ ) . |\widetilde{W}_t|:=\bigl(|\widetilde{Y}^{(1)}_t|,\ldots,|\widetilde{Y}^{(d)}_t|\bigr),\qquad|x|:=\bigl(|x_1|,\ldots,|x_d|\bigr). ∣ W t ∣ := ( ∣ Y t ( 1 ) ∣ , … , ∣ Y t ( d ) ∣ ) , ∣ x ∣ := ( ∣ x 1 ∣ , … , ∣ x d ∣ ) .
Since measurable functions of independent random variables remain independent, (3.18) yields
( ∣ W ~ t ∣ ) t ≥ 0 under P x = d ( W t + ) t ≥ 0 under P ∣ x ∣ . (25) (|\widetilde{W}_t|)_{t\geq0}\text{ under }\mathbb{P}_x \ \mathrel{\stackrel{\mathrm{d}}{=}}\ (W^+_t)_{t\geq0}\text{ under }\mathbb{P}_{|x|}. \tag*{(25)} ( ∣ W t ∣ ) t ≥ 0 under P x = d ( W t + ) t ≥ 0 under P ∣ x ∣ . (25)
The process W ~ \widetilde{W} W is a conservative continuous Hunt process. Its transition density is
p ~ t [ d ] ( x , y ) = ∏ j = 1 d p ~ t ( ρ ) ( x j , y j ) . \widetilde{p}^{[d]}_t(x,y)=\prod_{j=1}^{d}\widetilde{p}^{(\rho)}_t(x_j,y_j). p t [ d ] ( x , y ) = j = 1 ∏ d p t ( ρ ) ( x j , y j ) .
The product process is conservative and its kernel is symmetric. Hence the symmetric-kernel observation gives (H1) and reversibility.
To verify (H2), let K ⊂ [ − R , R ] d K\subset[-R,R]^d K ⊂ [ − R , R ] d be compact. Then
{ W ~ t ∈ K } ⊂ { ∣ W ~ t ∣ ∈ [ 0 , R ] d } . \{\widetilde{W}_t\in K\}\subset\{|\widetilde{W}_t|\in[0,R]^d\}. { W t ∈ K } ⊂ { ∣ W t ∣ ∈ [ 0 , R ] d } .
Therefore, by (3.19),
sup x ∈ R d P ~ t [ d ] 1 K ( x ) ≤ sup z ∈ [ 0 , ∞ ) d P z { W t + ∈ [ 0 , R ] d } ≤ C R ( 1 + t ) − d / ρ . \sup_{x\in\mathbb{R}^d}\widetilde{\mathbb{P}}^{[d]}_t1_K(x)\leq\sup_{z\in[0,\infty)^d}\mathbb{P}_z\{W^+_t\in[0,R]^d\}\leq C_R(1+t)^{-d/\rho}. x ∈ R d sup P t [ d ] 1 K ( x ) ≤ z ∈ [ 0 , ∞ ) d sup P z { W t + ∈ [ 0 , R ] d } ≤ C R ( 1 + t ) − d / ρ .
Here x ↦ ∣ x ∣ x \mapsto|x| x ↦ ∣ x ∣ maps R d \mathbb{R}^d R d onto [ 0 , ∞ ) d [0,\infty)^d [ 0 , ∞ ) d . Thus (H2) holds with q = d / ρ q=d/\rho q = d / ρ .
To verify (H3), set
D t ( N ) = [ − N t 1 / ρ , N t 1 / ρ ] d D_t(N)=[-Nt^{1/\rho},Nt^{1/\rho}]^d D t ( N ) = [ − N t 1/ ρ , N t 1/ ρ ] d
and let K ⊂ [ − R , R ] d K\subset[-R,R]^d K ⊂ [ − R , R ] d be compact. The localization identity and (25) give
∫ D t ( N ) c P ~ t [ d ] 1 K d λ = ∫ K P ∣ y ∣ { W t + ∉ [ 0 , N t 1 / ρ ] d } d y . \int_{D_t(N)^c}\widetilde{P}_t^{[d]}1_K\,\mathrm{d}\lambda
=\int_K\mathbb{P}_{|y|}\{W_t^+\notin[0,Nt^{1/\rho}]^d\}\,\mathrm{d}y. ∫ D t ( N ) c P t [ d ] 1 K d λ = ∫ K P ∣ y ∣ { W t + ∈ / [ 0 , N t 1/ ρ ] d } d y .
Decomposing K K K into orthants and using the change of variables z = ∣ y ∣ z=|y| z = ∣ y ∣ , this is bounded by
2 d ∫ [ 0 , R ] d P z { W t + ∉ [ 0 , N t 1 / ρ ] d } d z . 2^d\int_{[0,R]^d}\mathbb{P}_z\{W_t^+\notin[0,Nt^{1/\rho}]^d\}\,\mathrm{d}z. 2 d ∫ [ 0 , R ] d P z { W t + ∈ / [ 0 , N t 1/ ρ ] d } d z .
By the localization identity for the half-line product, the last expression equals
2 d ∫ [ 0 , ∞ ) d ∖ [ 0 , N t 1 / ρ ] d P t + 1 [ 0 , R ] d ( x ) d x , 2^d\int_{[0,\infty)^d\setminus[0,Nt^{1/\rho}]^d}P_t^+1_{[0,R]^d}(x)\,\mathrm{d}x, 2 d ∫ [ 0 , ∞ ) d ∖ [ 0 , N t 1/ ρ ] d P t + 1 [ 0 , R ] d ( x ) d x ,
which tends to zero uniformly in t ≥ 1 t\geq1 t ≥ 1 as N → ∞ N\to\infty N → ∞ by (H3) for W + W^+ W + . Since λ ( D t ( N ) ) = ( 2 N ) d / ρ \lambda(D_t(N))=(2N)^{d/\rho} λ ( D t ( N )) = ( 2 N ) d / ρ , (H3) follows. The assertion for nonsingular linear images follows from Lemma 3.1 . □
Recurrence and transience
For a nonnegative Borel function h h h , write
U h ( x ) : = ∫ 0 ∞ P t h ( x ) d t . Uh(x):=\int_0^\infty P_th(x)\,\mathrm{d}t. U h ( x ) := ∫ 0 ∞ P t h ( x ) d t .
We use the following potential-theoretic notion of transience: the process W W W is transient if there exists a nonnegative Borel function g ∈ L 1 ( E , λ ) g\in L^1(E,\lambda) g ∈ L 1 ( E , λ ) such that g > 0 g>0 g > 0 λ \lambda λ -almost everywhere and U g < ∞ Ug<\infty U g < ∞ λ \lambda λ -almost everywhere. For the symmetric semigroups considered below, recurrence is understood in the sense of Chen and Fukushima [10 Definition 2.1.1] . We use repeatedly the following consequence of Chen and Fukushima [10 Proposition 2.1.3(ii)] : if a compact set K K K of positive Lebesgue measure satisfies U 1 K = ∞ U1_K=\infty U 1 K = ∞ λ \lambda λ -almost everywhere, then the semigroup is recurrent.
Proposition 3.6. Let W W W be a conservative Hunt process on E = R d E=\mathbb{R}^d E = R d or E = [ 0 , ∞ ) d E=[0,\infty)^d E = [ 0 , ∞ ) d . If (H2) holds with q > 1 q>1 q > 1 , then W W W is transient.
Proof. For every compact K ⊂ E K \subset E K ⊂ E ,
sup x ∈ E U 1 K ( x ) ≤ C K ∫ 0 ∞ ( 1 + t ) − q d t = C K q − 1 < ∞ . \sup_{x \in E} U1_K(x) \le C_K \int_0^\infty(1+t)^{-q}\,\mathrm{d}t = \frac{C_K}{q-1} < \infty. x ∈ E sup U 1 K ( x ) ≤ C K ∫ 0 ∞ ( 1 + t ) − q d t = q − 1 C K < ∞.
Let K 0 = ∅ K_0=\varnothing K 0 = ∅ and choose an increasing compact exhaustion ( K n ) n ≥ 1 (K_n)_{n\ge1} ( K n ) n ≥ 1 of E E E . Set
M n : = 1 + λ ( K n ) + C K n q − 1 , a n : = 2 − n M n , g : = ∑ n ≥ 1 a n 1 K n ∖ K n − 1 . M_n := 1+\lambda(K_n)+\frac{C_{K_n}}{q-1}, \qquad a_n := \frac{2^{-n}}{M_n}, \qquad g := \sum_{n\ge1} a_n 1_{K_n\setminus K_{n-1}}. M n := 1 + λ ( K n ) + q − 1 C K n , a n := M n 2 − n , g := n ≥ 1 ∑ a n 1 K n ∖ K n − 1 .
Since ( K n ) (K_n) ( K n ) exhausts E E E , the function g g g is strictly positive. Moreover, a n ≤ 2 − n a_n\le2^{-n} a n ≤ 2 − n , so g g g is bounded, and
∫ E g d λ ≤ ∑ n ≥ 1 a n λ ( K n ) ≤ ∑ n ≥ 1 2 − n < ∞ . \int_E g\,\mathrm{d}\lambda\le\sum_{n\ge1} a_n\lambda(K_n) \le\sum_{n\ge1}2^{-n}<\infty. ∫ E g d λ ≤ n ≥ 1 ∑ a n λ ( K n ) ≤ n ≥ 1 ∑ 2 − n < ∞.
Finally, Tonelli’s theorem gives, for every x ∈ E x\in E x ∈ E ,
U g ( x ) ≤ ∑ n ≥ 1 a n U 1 K n ( x ) ≤ ∑ n ≥ 1 a n C K n q − 1 ≤ ∑ n ≥ 1 2 − n < ∞ . Ug(x)\le\sum_{n\ge1}a_nU1_{K_n}(x)\le\sum_{n\ge1}a_n\frac{C_{K_n}}{q-1}\le\sum_{n\ge1}2^{-n}<\infty. U g ( x ) ≤ n ≥ 1 ∑ a n U 1 K n ( x ) ≤ n ≥ 1 ∑ a n q − 1 C K n ≤ n ≥ 1 ∑ 2 − n < ∞.
Thus W W W is transient. □
Stable motions
Proposition 3.7. The process in Proposition 3.2 (a) is recurrent if and only if d ≤ α d\le\alpha d ≤ α . The anisotropic product in Proposition 3.2 (b), as well as every invertible linear image of it, is recurrent if and only if
q = ∑ i = 1 d α i − 1 ≤ 1. q=\sum_{i=1}^{d}\alpha_i^{-1}\le1. q = i = 1 ∑ d α i − 1 ≤ 1. In the complementary cases the corresponding processes are transient. In particular, every centered Brownian motion on R d \mathbb{R}^d R d with positive definite covariance matrix is recurrent for d = 1 , 2 d=1,2 d = 1 , 2 and transient for d ≥ 3 d\ge3 d ≥ 3 .
Proof. For either untransformed process, the scaling established in Proposition 3.2 reads p t ( z ) = t − q p 1 ( Δ t − 1 z ) p_t(z)=t^{-q}p_1(\Delta_t^{-1}z) p t ( z ) = t − q p 1 ( Δ t − 1 z ) , where q = d / α q=d/\alpha q = d / α and Δ t = t 1 / α I \Delta_t=t^{1/\alpha}I Δ t = t 1/ α I in part (a), while q = ∑ i α i − 1 q=\sum_i\alpha_i^{-1} q = ∑ i α i − 1 and Δ t = diag ( t 1 / α 1 , … , t 1 / α d ) \Delta_t=\operatorname{diag}(t^{1/\alpha_1},\ldots,t^{1/\alpha_d}) Δ t = diag ( t 1/ α 1 , … , t 1/ α d ) in part (b). In both cases p 1 p_1 p 1 is continuous and p 1 ( 0 ) > 0 p_1(0)>0 p 1 ( 0 ) > 0 , by the Fourier representation and, for the product, its coordinatewise application. For a compact K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d of positive measure and fixed x x x , the arguments Δ t − 1 ( y − x ) \Delta_t^{-1}(y-x) Δ t − 1 ( y − x ) tend to zero uniformly on K K K . Integrating the scaled density over K K K therefore gives
P t 1 K ( x ) ∼ p 1 ( 0 ) λ ( K ) t − q . (26) P_t1_K(x)\sim p_1(0)\lambda(K)t^{-q}. \tag*{(26)} P t 1 K ( x ) ∼ p 1 ( 0 ) λ ( K ) t − q . (26)
If q ≤ 1 q\le1 q ≤ 1 , integration in time yields U 1 K ( x ) = ∞ U1_K(x)=\infty U 1 K ( x ) = ∞ for every x x x , so the recurrence criterion above applies. For q > 1 q>1 q > 1 , use Proposition 3.6 . The classification is preserved by invertible linear images through Lemma 3.1 and Equation (7) ; setting all α i = 2 \alpha_i=2 α i = 2 gives the Brownian case. □
Uniformly elliptic symmetric diffusions
Proposition 3.8. Every diffusion in Proposition 3.3 is recurrent for d ≤ 2 d \le2 d ≤ 2 and transient for d ≥ 3 d \ge3 d ≥ 3 .
Proof. The semigroup is symmetric by Proposition 3.3 . Fix R > 0 R > 0 R > 0 , set K = B ( 0 , R ) K = B(0,R) K = B ( 0 , R ) , and put T x = 1 + ( ∣ x ∣ + R ) 2 T_x = 1 + (|x| + R)^2 T x = 1 + ( ∣ x ∣ + R ) 2 . For t ≥ T x t \ge T_x t ≥ T x and y ∈ K y \in K y ∈ K , we have ∣ x − y ∣ 2 / t ≤ 1 |x-y|^2/t \le1 ∣ x − y ∣ 2 / t ≤ 1 , so (8) yields
P t 1 K ( x ) ≥ a 1 e − a 2 λ ( K ) t − d / 2 , t ≥ T x . (27) P_t1_K(x) \ge a_1e^{-a_2}\lambda(K)t^{-d/2}, \qquad t \ge T_x. \tag*{(27)} P t 1 K ( x ) ≥ a 1 e − a 2 λ ( K ) t − d /2 , t ≥ T x . (27)
For d ≤ 2 d \le2 d ≤ 2 , the right-hand side has infinite time integral; hence U 1 K ( x ) = ∞ U1_K(x) = \infty U 1 K ( x ) = ∞ for every x x x , and the recurrence criterion applies. For d ≥ 3 d \ge3 d ≥ 3 , (H2) holds with q = d / 2 > 1 q = d/2 > 1 q = d /2 > 1 , and Proposition 3.6 gives transience.
Products of generalized squared-Bessel diffusions
Proposition 3.9. The product process in Proposition 3.4 is recurrent on [ 0 , ∞ ) d [0,\infty)^d [ 0 , ∞ ) d if and only if d ≤ ρ d \le\rho d ≤ ρ , and it is transient if d > ρ d > \rho d > ρ .
Proof. Use the symmetric product density p t [ d ] p_t^{[d]} p t [ d ] from Proposition 3.4 and set q = d / ρ q = d/\rho q = d / ρ . By (13) ,
b ρ , κ : = p 1 ( ρ ) ( 0 , 0 ) = ρ Γ ( 1 / ρ ) ( κ ρ 2 ) 1 / ρ > 0. b_{\rho,\kappa} := p_1^{(\rho)}(0,0) = \frac{\rho}{\Gamma(1/\rho)(\kappa\rho^2)^{1/\rho}} > 0. b ρ , κ := p 1 ( ρ ) ( 0 , 0 ) = Γ ( 1/ ρ ) ( κ ρ 2 ) 1/ ρ ρ > 0.
For a compact K ⊂ [ 0 , ∞ ) d K \subset[0,\infty)^d K ⊂ [ 0 , ∞ ) d of positive measure and fixed x ∈ [ 0 , ∞ ) d x \in[0,\infty)^d x ∈ [ 0 , ∞ ) d , scaling gives
t q P t 1 K ( x ) = ∫ K p 1 [ d ] ( t − 1 / ρ x , t − 1 / ρ y ) d y . (28) t^qP_t1_K(x) = \int_K p_1^{[d]}(t^{-1/\rho}x,t^{-1/\rho}y)\,\mathrm{d}y. \tag*{(28)} t q P t 1 K ( x ) = ∫ K p 1 [ d ] ( t − 1/ ρ x , t − 1/ ρ y ) d y . (28)
Continuity of p 1 [ d ] p_1^{[d]} p 1 [ d ] at ( 0 , 0 ) (0,0) ( 0 , 0 ) implies uniform convergence of the integrand on K K K to b ρ , κ b_{\rho,\kappa} b ρ , κ . Thus
P t 1 K ( x ) ∼ b ρ , κ d λ ( K ) t − d / ρ , t → ∞ . (29) P_t1_K(x) \sim b_{\rho,\kappa}^{d}\lambda(K)t^{-d/\rho}, \qquad t \to\infty. \tag*{(29)} P t 1 K ( x ) ∼ b ρ , κ d λ ( K ) t − d / ρ , t → ∞. (29)
For d ≤ ρ d \le\rho d ≤ ρ , time integration gives U 1 K ( x ) = ∞ U1_K(x) = \infty U 1 K ( x ) = ∞ for every x x x , and hence recurrence. For d > ρ d > \rho d > ρ , (H2) and Proposition 3.6 give transience.
Symmetric bilateral Bessel-type motions
Proposition 3.10. For ρ > 1 \rho> 1 ρ > 1 , the bilateral d d d -fold product in Proposition 3.5 , and every nonsingular linear image of it, is recurrent if and only if d ≤ ρ d \le\rho d ≤ ρ . It is transient if d > ρ d > \rho d > ρ .
Proof. Let W ~ \widetilde W W and W + W^+ W + be the bilateral and half-line products, with semigroups P ~ t [ d ] \widetilde P_t^{[d]} P t [ d ] and P t + P_t^+ P t + and potential operators U ~ \widetilde U U and U + U^+ U + , respectively. For R > 0 R > 0 R > 0 , set C R = [ − R , R ] d C_R=[-R,R]^d C R = [ − R , R ] d and K R = [ 0 , R ] d K_R=[0,R]^d K R = [ 0 , R ] d . The coordinatewise absolute-value law (25) gives P ~ t 1 C R ( x ) = P t + 1 K R ( ∣ x ∣ ) \widetilde{P}_t\mathbf{1}_{C_R}(x)=P_t^+\mathbf{1}_{K_R}(|x|) P t 1 C R ( x ) = P t + 1 K R ( ∣ x ∣ ) ; integrating in time,
U ~ 1 C R ( x ) = U + 1 K R ( ∣ x ∣ ) . (30) \widetilde{U}\mathbf{1}_{C_R}(x)=U^+\mathbf{1}_{K_R}(|x|). \tag*{(30)} U 1 C R ( x ) = U + 1 K R ( ∣ x ∣ ) . (30)
When d ≤ ρ d\leq\rho d ≤ ρ , the proof of Proposition 3.9 shows that the right-hand side is infinite for every x x x . Symmetry of the bilateral semigroup and the recurrence criterion therefore give recurrence. The same argument applies to every nonsingular linear image by Lemma 3.1 and Equation (7) : such a map preserves compactness and rescales Lebesgue measure by its nonzero constant Jacobian. When d > ρ d>\rho d > ρ , the bilateral product and its linear images satisfy (H2) with q = d / ρ > 1 q=d/\rho>1 q = d / ρ > 1 by Proposition 3.5 , so Proposition 3.6 gives transience. □
Remark 3.11. For every motion listed in Table 3.1 , recurrence changes to transience at q = 1 q=1 q = 1 . This is not the Bellman–Harris threshold. Since 0 < β ≤ 1 0<\beta\leq1 0 < β ≤ 1 , persistence with finite-mean lifetimes, q β > 1 q\beta>1 q β > 1 , can occur only when the individual motion is transient. For a regularly varying lifetime tail, the inequality q β > γ q\beta>\gamma q β > γ may hold with q ≤ 1 q\leq1 q ≤ 1 , so the Poisson limit may occur while the individual motion is recurrent.
The admissible parameter ranges, dispersal exponents, and recurrence classifications obtained above are summarized in Table 3.1 .
Remark 3.12. The following example shows that the recurrence–transience threshold of the underlying motion need not coincide with the spatial threshold governing extinction and persistence of the branching system. For all the symmetric examples in Table 3.1 , the recurrence–transience transition occurs at q = 1 q=1 q = 1 , but this coincidence does not persist for general Hunt processes satisfying (H1)–(H3).
Let B B B be a standard Brownian motion on R d \mathbb{R}^d R d , let b ∈ R d ∖ { 0 } b\in\mathbb{R}^d\setminus\{0\} b ∈ R d ∖ { 0 } , and set
W t x = x + b t + B t . W_t^x=x+bt+B_t. W t x = x + b t + B t .
This is a conservative Hunt Lévy process. Translation invariance of its semigroup gives (H1). If K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d is compact, then
sup x ∈ R d P t 1 K ( x ) ≤ ( 2 π t ) − d / 2 λ ( K ) , t ≥ 1 , \sup_{x\in\mathbb{R}^d}P_t1_K(x)\leq(2\pi t)^{-d/2}\lambda(K),\qquad t\geq1, x ∈ R d sup P t 1 K ( x ) ≤ ( 2 π t ) − d /2 λ ( K ) , t ≥ 1 ,
so (H2) holds with q = d / 2 q=d/2 q = d /2 .
To verify (H3), choose R > 0 R>0 R > 0 such that K ⊂ B ( 0 , R ) K\subset B(0,R) K ⊂ B ( 0 , R ) and set D t ( N ) = B ( − b t , N t ) D_t(N)=B(-bt,N\sqrt{t}) D t ( N ) = B ( − b t , N t ) . Tonelli’s theorem and the Gaussian transition density give
∫ D t ( N ) c P t 1 K ( x ) d x = ∫ K P { y − b t − B t ∉ D t ( N ) } d y = ∫ K P { ∣ y − B t ∣ > N t } d y ≤ λ ( K ) P { ∣ B 1 ∣ > N − R } , t ≥ 1 , N > R . \begin{aligned}
\int_{D_t(N)^c}P_t1_K(x)\,\mathrm{d}x=\int_K\mathbb{P}\{y-bt-B_t\notin D_t(N)\}\,\mathrm{d}y \\
&=\int_K\mathbb{P}\{|y-B_t|>N\sqrt{t}\}\,\mathrm{d}y \\
&\leq\lambda(K)\mathbb{P}\{|B_1|>N-R\},\qquad t\geq1,\quad N>R.
\end{aligned} ∫ D t ( N ) c P t 1 K ( x ) d x = ∫ K P { y − b t − B t ∈ / D t ( N )} d y = ∫ K P { ∣ y − B t ∣ > N t } d y ≤ λ ( K ) P { ∣ B 1 ∣ > N − R } , t ≥ 1 , N > R .
The inequality uses ∣ y ∣ ≤ R |y|\leq R ∣ y ∣ ≤ R , the triangle inequality, and B t = d t B 1 B_t\mathrel{\overset{\mathrm{d}}{=}}\sqrt{t}B_1 B t = d t B 1 . The last bound tends to zero uniformly in t ≥ 1 t\geq1 t ≥ 1 , while λ ( D t ( N ) ) = λ ( B ( 0 , N ) ) t d / 2 \lambda(D_t(N))=\lambda(B(0,N))t^{d/2} λ ( D t ( N )) = λ ( B ( 0 , N )) t d /2 , proving (H3) with q = d / 2 q=d/2 q = d /2 .
Nevertheless, the motion is transient. Take g ( y ) = e − ∣ y ∣ g(y)=e^{-|y|} g ( y ) = e − ∣ y ∣ , which is bounded, integrable, and strictly positive. For fixed x x x , choose T x ≥ 1 T_x\geq1 T x ≥ 1 such that ∣ x + b t ∣ ≥ ∣ b ∣ t / 2 |x+bt|\geq|b|t/2 ∣ x + b t ∣ ≥ ∣ b ∣ t /2 for t ≥ T x t\geq T_x t ≥ T x . On { ∣ B t ∣ ≤ ∣ b ∣ t / 4 } \{|B_t|\leq|b|t/4\} { ∣ B t ∣ ≤ ∣ b ∣ t /4 } , the triangle inequality gives ∣ x + b t + B t ∣ ≥ ∣ b ∣ t / 4 |x+bt+B_t|\geq|b|t/4 ∣ x + b t + B t ∣ ≥ ∣ b ∣ t /4 . A Gaussian tail estimate therefore yields constants C , c > 0 C,c>0 C , c > 0 , independent of x x x , such that
P t g ( x ) ≤ e − ∣ b ∣ t / 4 + P { ∣ B t ∣ > ∣ b ∣ t / 4 } ≤ e − ∣ b ∣ t / 4 + C e − c t , t ≥ T x . P_tg(x)\leq e^{-|b|t/4}+\mathbb{P}\{|B_t|>|b|t/4\}\leq e^{-|b|t/4}+Ce^{-ct},\qquad t\geq T_x. P t g ( x ) ≤ e − ∣ b ∣ t /4 + P { ∣ B t ∣ > ∣ b ∣ t /4 } ≤ e − ∣ b ∣ t /4 + C e − c t , t ≥ T x .
Since P t g ≤ 1 P_tg\leq1 P t g ≤ 1 on [ 0 , T x ] [0,T_x] [ 0 , T x ] , these estimates give U g ( x ) < ∞ Ug(x)<\infty U g ( x ) < ∞ for every x x x , proving transience in the potential-theoretic sense above. In particular, for d = 1 , 2 d=1,2 d = 1 , 2 the motion is transient although q = d / 2 ≤ 1 q=d/2\leq1 q = d /2 ≤ 1 . Thus the recurrence–transience threshold q = 1 q=1 q = 1 for the symmetric examples in Table 3.1 does not follow from (H1)–(H3) alone.
Auxiliary estimates
The proofs in this section and in § 5 follow the scheme of Vatutin and Wakolbinger [2 Sections 3–7] . We retain the steps that require modification in the present setting: the first-moment identity is stated for a general Markov semigroup, (H1) replaces spatial homogeneity after integration over the initial position, (H2) replaces the stable transition-density estimate, and (H3) replaces localization in the stable-scale sets C ( t , N ) C(t,N) C ( t , N ) . Thus the exponent d / α d/\alpha d / α in their argument is replaced by q q q .
For a family issued from one ancestor, write N ( t ) = Z t ( E ) N(t)=Z_t(E) N ( t ) = Z t ( E ) . Call an age θ ≥ 0 \theta\ge0 θ ≥ 0 admissible if Γ ‾ ( θ ) > 0 \overline{\Gamma}(\theta)>0 Γ ( θ ) > 0 . If the initial particle has age θ \theta θ , its residual lifetime has distribution
Γ θ ( [ 0 , r ] ) = Γ ( θ + r ) − Γ ( θ ) Γ ‾ ( θ ) , Γ ‾ θ ( r ) = Γ ‾ ( θ + r ) Γ ‾ ( θ ) . (31) \Gamma_\theta([0,r])=\frac{\Gamma(\theta+r)-\Gamma(\theta)}{\overline{\Gamma}(\theta)}, \qquad\overline{\Gamma}_\theta(r)=\frac{\overline{\Gamma}(\theta+r)}{\overline{\Gamma}(\theta)}. \tag*{(31)} Γ θ ([ 0 , r ]) = Γ ( θ ) Γ ( θ + r ) − Γ ( θ ) , Γ θ ( r ) = Γ ( θ ) Γ ( θ + r ) . (31)
We denote the corresponding one-family law and expectation by P x θ \mathbb{P}^{\theta}_x P x θ and E x θ \mathbb{E}^{\theta}_x E x θ . Let ( F u ) u ≥ 0 (\mathcal{F}_u)_{u\ge0} ( F u ) u ≥ 0 denote the natural filtration generated by the branching system, so that F u \mathcal{F}_u F u contains the complete history of the system up to time u u u . If the particles alive at u u u have positions and ages ( y 1 , θ 1 ) , … , ( y N ( u ) , θ N ( u ) ) (y_1,\theta_1),\ldots,(y_{N(u)},\theta_{N(u)}) ( y 1 , θ 1 ) , … , ( y N ( u ) , θ N ( u ) ) , then, conditionally on F u \mathcal{F}_u F u , their future descendant families are independent and the family issued from particle j j j has law P y j θ j \mathbb{P}^{\theta_j}_{y_j} P y j θ j .
First moments and tightness
Lemma 4.1. For x ∈ E x\in E x ∈ E , every admissible age θ \theta θ , every t ≥ 0 t\ge0 t ≥ 0 , and every measurable function h : E → [ 0 , ∞ ] h:E\to[0,\infty] h : E → [ 0 , ∞ ] ,
E x θ ⟨ Z t , h ⟩ = P t h ( x ) . (32) \mathbb{E}^{\theta}_x\langle Z_t,h\rangle=P_t h(x). \tag*{(32)} E x θ ⟨ Z t , h ⟩ = P t h ( x ) . (32) Consequently, for every compact set K ⊂ E K\subset E K ⊂ E ,
P x θ { Z t ( K ) > 0 } ≤ P t 1 K ( x ) . (33) \mathbb{P}^{\theta}_x\{Z_t(K)>0\}\le P_t1_K(x). \tag*{(33)} P x θ { Z t ( K ) > 0 } ≤ P t 1 K ( x ) . (33)
Proof. Assume first that h h h is bounded. For m ≥ 0 m\ge0 m ≥ 0 , let Z [ m ] Z^{[m]} Z [ m ] denote the branching family obtained by truncating the genealogical tree after generation m m m . Thus Z t [ m ] Z_t^{[m]} Z t [ m ] is the counting measure of all particles alive at time t t t whose genealogical distance from the initial ancestor is at most m m m .
Let L 0 L_0 L 0 have the residual-lifetime law Γ θ \Gamma_\theta Γ θ , and let L 1 , L 2 , … L_1,L_2,\ldots L 1 , L 2 , … be independent with law Γ \Gamma Γ . Set
S 0 = 0 , S n = L 0 + ⋯ + L n − 1 , n ≥ 1. S_0=0,\qquad S_n=L_0+\cdots+L_{n-1},\qquad n\ge1. S 0 = 0 , S n = L 0 + ⋯ + L n − 1 , n ≥ 1.
The variables S n S_n S n represent the successive birth times along a single genealogical line: a particle of generation j j j on that line is alive at time t t t precisely when
S j ≤ t < S j + 1 . S_j\le t<S_{j+1}. S j ≤ t < S j + 1 .
We claim that
E x θ ⟨ Z t [ m ] , h ⟩ = P t h ( x ) ∑ j = 0 m P { S j ≤ t < S j + 1 } . (34) \mathbb{E}^{\theta}_{x}\left\langle Z_t^{[m]},h\right\rangle=P_t h(x)\sum_{j=0}^{m}\mathbb{P}\{S_j\le t<S_{j+1}\}. \tag*{(34)} E x θ ⟨ Z t [ m ] , h ⟩ = P t h ( x ) j = 0 ∑ m P { S j ≤ t < S j + 1 } . (34)
For m = 0 m=0 m = 0 , Z [ 0 ] Z^{[0]} Z [ 0 ] consists only of the initial ancestor. Hence, by independence of its motion and residual lifetime,
E x θ ⟨ Z t [ 0 ] , h ⟩ = Γ ‾ θ ( t ) P t h ( x ) = P t h ( x ) P { S 0 ≤ t < S 1 } . \mathbb{E}^{\theta}_{x}\left\langle Z_t^{[0]},h\right\rangle=\overline{\Gamma}_{\theta}(t)P_t h(x)=P_t h(x)\mathbb{P}\{S_0\le t<S_1\}. E x θ ⟨ Z t [ 0 ] , h ⟩ = Γ θ ( t ) P t h ( x ) = P t h ( x ) P { S 0 ≤ t < S 1 } .
Suppose now that (4.4) holds for a newborn family truncated after generation m m m . Consider a family truncated after generation m + 1 m+1 m + 1 . If the initial ancestor is still alive at time t t t , its expected contribution is Γ ‾ θ ( t ) P t h ( x ) \overline{\Gamma}_{\theta}(t)P_t h(x) Γ θ ( t ) P t h ( x ) . Otherwise, condition on its residual lifetime L 0 = u ≤ t L_0=u\le t L 0 = u ≤ t , its death position W u W_u W u , and its number ζ \zeta ζ of children. Each child then starts, from W u W_u W u , an independent newborn family truncated after generation m m m and observed for the remaining time t − u t-u t − u . Since E ζ = 1 \mathbb{E}\zeta=1 E ζ = 1 , Tonelli’s theorem and the induction hypothesis give
E x θ ⟨ Z t [ m + 1 ] , h ⟩ = Γ ‾ θ ( t ) P t h ( x ) + ∫ [ 0 , t ] P u [ E ⋅ θ ⟨ Z t − u [ m ] , h ⟩ ] ( x ) Γ θ ( d u ) = P t h ( x ) [ Γ ‾ θ ( t ) + ∫ [ 0 , t ] ∑ j = 0 m P { L 1 + ⋯ + L j ≤ t − u < L 1 + ⋯ + L j + 1 } Γ θ ( d u ) ] = P t h ( x ) ∑ j = 0 m + 1 P { S j ≤ t < S j + 1 } . \begin{aligned}
\mathbb{E}^{\theta}_{x}\left\langle Z_t^{[m+1]},h\right\rangle
&=\overline{\Gamma}_{\theta}(t)P_t h(x)+\int_{[0,t]}P_u\left[\mathbb{E}^{\theta}_{\cdot}\left\langle Z_{t-u}^{[m]},h\right\rangle\right](x)\Gamma_{\theta}(\mathrm{d}u)\\
&=P_t h(x)\left[\overline{\Gamma}_{\theta}(t)+\int_{[0,t]}\sum_{j=0}^{m}\mathbb{P}\{L_1+\cdots+L_j\le t-u<L_1+\cdots+L_{j+1}\}\Gamma_{\theta}(\mathrm{d}u)\right]\\
&=P_t h(x)\sum_{j=0}^{m+1}\mathbb{P}\{S_j\le t<S_{j+1}\}.
\end{aligned} E x θ ⟨ Z t [ m + 1 ] , h ⟩ = Γ θ ( t ) P t h ( x ) + ∫ [ 0 , t ] P u [ E ⋅ θ ⟨ Z t − u [ m ] , h ⟩ ] ( x ) Γ θ ( d u ) = P t h ( x ) [ Γ θ ( t ) + ∫ [ 0 , t ] j = 0 ∑ m P { L 1 + ⋯ + L j ≤ t − u < L 1 + ⋯ + L j + 1 } Γ θ ( d u ) ] = P t h ( x ) j = 0 ∑ m + 1 P { S j ≤ t < S j + 1 } .
Under the natural coupling,
⟨ Z t [ m ] , h ⟩ ↑ ⟨ Z t , h ⟩ . \left\langle Z_t^{[m]},h\right\rangle\uparrow\left\langle Z_t,h\right\rangle. ⟨ Z t [ m ] , h ⟩ ↑ ⟨ Z t , h ⟩ .
Moreover, since Γ ( 0 ) = 0 \Gamma(0)=0 Γ ( 0 ) = 0 , there exists η > 0 \eta>0 η > 0 such that P { L 1 > η } > 0 \mathbb{P}\{L_1>\eta\}>0 P { L 1 > η } > 0 . The events { L n > η } \{L_n>\eta\} { L n > η } , n ≥ 1 n\ge1 n ≥ 1 , are independent and have the same positive probability, so the second Borel–Cantelli lemma implies that they occur infinitely often almost surely. Consequently,
S n = L 0 + ⋯ + L n − 1 ⟶ ∞ a.s. S_n=L_0+\cdots+L_{n-1}\longrightarrow\infty\qquad\text{a.s.} S n = L 0 + ⋯ + L n − 1 ⟶ ∞ a.s.
Hence, for every fixed t ≥ 0 t \ge0 t ≥ 0 , the disjoint events
{ S j ≤ t < S j + 1 } , j ≥ 0 , \{S_j \le t < S_{j+1}\}, \qquad j \ge0, { S j ≤ t < S j + 1 } , j ≥ 0 ,
exhaust the sample space up to a null set, and therefore
∑ j = 0 ∞ P { S j ≤ t < S j + 1 } = 1. \sum_{j=0}^{\infty} \mathbb{P}\{S_j \le t < S_{j+1}\} = 1. j = 0 ∑ ∞ P { S j ≤ t < S j + 1 } = 1.
For bounded h h h , (4.4) also gives
E x θ ⟨ Z t [ m ] , h ⟩ ≤ P t h ( x ) ≤ ∥ h ∥ ∞ . \mathbb{E}^{\theta}_{x}\langle Z^{[m]}_t,h\rangle\le P_t h(x) \le\|h\|_{\infty}. E x θ ⟨ Z t [ m ] , h ⟩ ≤ P t h ( x ) ≤ ∥ h ∥ ∞ .
Thus, by monotone convergence,
E x θ ⟨ Z t , h ⟩ = lim m → ∞ E x θ ⟨ Z t [ m ] , h ⟩ = P t h ( x ) . \mathbb{E}^{\theta}_{x}\langle Z_t,h\rangle= \lim_{m\to\infty}\mathbb{E}^{\theta}_{x}\langle Z^{[m]}_t,h\rangle= P_t h(x). E x θ ⟨ Z t , h ⟩ = m → ∞ lim E x θ ⟨ Z t [ m ] , h ⟩ = P t h ( x ) .
Now let h : E → [ 0 , ∞ ] h:E\to[0,\infty] h : E → [ 0 , ∞ ] be arbitrary and measurable, and set h n = h ∧ n h_n=h\wedge n h n = h ∧ n . Applying the preceding result to h n h_n h n gives
E x θ ⟨ Z t , h n ⟩ = P t h n ( x ) . \mathbb{E}^{\theta}_{x}\langle Z_t,h_n\rangle=P_t h_n(x). E x θ ⟨ Z t , h n ⟩ = P t h n ( x ) .
Since h n ↑ h h_n \uparrow h h n ↑ h , monotone convergence on both sides yields
E x θ ⟨ Z t , h ⟩ = P t h ( x ) , \mathbb{E}^{\theta}_{x}\langle Z_t,h\rangle=P_t h(x), E x θ ⟨ Z t , h ⟩ = P t h ( x ) ,
with either side allowed to be infinite. Finally, 1 { Z t ( K ) > 0 } ≤ Z t ( K ) \mathbf{1}_{\{Z_t(K)>0\}}\le Z_t(K) 1 { Z t ( K ) > 0 } ≤ Z t ( K ) gives (4.3). □ \square □
Corollary 4.2. For every measurable h : E → [ 0 , ∞ ] h:E\to[0,\infty] h : E → [ 0 , ∞ ] and every t ≥ 0 t\ge0 t ≥ 0 ,
E ⟨ X t , h ⟩ = ∫ E h d λ . (35) \mathbb{E}\langle X_t,h\rangle=\int_E h\,d\lambda. \tag*{(35)} E ⟨ X t , h ⟩ = ∫ E h d λ . (35) Consequently, X t X_t X t is almost surely a Radon counting measure for every t ≥ 0 t\ge0 t ≥ 0 , and the family { L ( X t ) : t ≥ 0 } \{\mathcal{L}(X_t):t\ge0\} { L ( X t ) : t ≥ 0 } is tight in M ( E ) \mathcal{M}(E) M ( E ) endowed with the vague topology.
Proof. Campbell’s formula, Lemma 4.1, and (H1) give
E ⟨ X t , h ⟩ = ∫ E E x ⟨ Z t , h ⟩ λ ( d x ) = ∫ E P t h ( x ) λ ( d x ) = ∫ E h d λ . \mathbb{E}\langle X_t,h\rangle=\int_E\mathbb{E}_x\langle Z_t,h\rangle\lambda(dx)=\int_E P_t h(x)\lambda(dx)=\int_E h\,d\lambda. E ⟨ X t , h ⟩ = ∫ E E x ⟨ Z t , h ⟩ λ ( d x ) = ∫ E P t h ( x ) λ ( d x ) = ∫ E h d λ . Thus E X t ( K ) = λ ( K ) < ∞ \mathbb{E}X_t(K)=\lambda(K)<\infty E X t ( K ) = λ ( K ) < ∞ for every compact K ⊂ E K\subset E K ⊂ E . A countable compact exhaustion of E E E therefore shows that X t X_t X t is almost surely a Radon counting measure.
For every compact K ⊂ E K\subset E K ⊂ E , choose ψ ∈ C c + ( E ) \psi\in C^+_c(E) ψ ∈ C c + ( E ) with ψ ≥ 1 K \psi\ge\mathbf{1}_K ψ ≥ 1 K . Then
sup t ≥ 0 P { X t ( K ) > a } ≤ 1 a ∫ E ψ d λ ⟶ 0 , a → ∞ . \sup_{t\ge0}\mathbb{P}\{X_t(K)>a\}\le\frac{1}{a}\int_E\psi\,d\lambda\longrightarrow0,\qquad a\to\infty. t ≥ 0 sup P { X t ( K ) > a } ≤ a 1 ∫ E ψ d λ ⟶ 0 , a → ∞. The tightness criterion for random Radon measures [16 Lemma 4.5] now gives the result. □ \square □
Renewal estimates
Let
R = ∑ n = 0 ∞ Γ ∗ n , U = R − δ 0 = ∑ n = 1 ∞ Γ ∗ n , U ( t ) = U ( [ 0 , t ] ) . R=\sum_{n=0}^{\infty}\Gamma^{*n},\qquad U=R-\delta_0=\sum_{n=1}^{\infty}\Gamma^{*n},\qquad U(t)=U([0,t]). R = n = 0 ∑ ∞ Γ ∗ n , U = R − δ 0 = n = 1 ∑ ∞ Γ ∗ n , U ( t ) = U ([ 0 , t ]) .
and write
( U ∗ r ) ( t ) : = ∫ [ 0 , t ] r ( t − s ) U ( d s ) . (U*r)(t):=\int_{[0,t]}r(t-s)U(\mathrm{d}s). ( U ∗ r ) ( t ) := ∫ [ 0 , t ] r ( t − s ) U ( d s ) .
Suppose first that Γ \Gamma Γ is spread-out and has finite mean μ \mu μ . For every bounded r ∈ L 1 ( [ 0 , ∞ ) ) r\in L^1([0,\infty)) r ∈ L 1 ([ 0 , ∞ )) with r ( t ) → 0 r(t)\to0 r ( t ) → 0 , the key renewal theorem for spread-out distributions gives
( U ∗ r ) ( t ) ⟶ 1 μ ∫ 0 ∞ r ( s ) d s . (36) (U*r)(t)\longrightarrow\frac{1}{\mu}\int_0^\infty r(s)\,\mathrm{d}s. \tag*{(36)} ( U ∗ r ) ( t ) ⟶ μ 1 ∫ 0 ∞ r ( s ) d s . (36)
Indeed, after extending r r r by zero to ( − ∞ , 0 ) (-\infty,0) ( − ∞ , 0 ) , Sgibnev [17 Section 5, Eq. (5.2), p. 2428] gives
( R ∗ r ) ( t ) ⟶ 1 μ ∫ 0 ∞ r ( s ) d s . (R*r)(t)\longrightarrow\frac{1}{\mu}\int_0^\infty r(s)\,\mathrm{d}s. ( R ∗ r ) ( t ) ⟶ μ 1 ∫ 0 ∞ r ( s ) d s .
Since R = δ 0 + U R=\delta_0+U R = δ 0 + U and r ( t ) → 0 r(t)\to0 r ( t ) → 0 , (4.6) follows.
Assume now (2.1). The renewal asymptotics for regularly varying tails in Bingham et al. [18 Section 8.6.2, Eqs. (8.6.1)–(8.6.4), pp. 360–361] when 0 < γ < 1 0<\gamma<1 0 < γ < 1 , and in Bingham et al. [18 Section 8.6.3, p. 365] when γ = 1 \gamma=1 γ = 1 , imply that for every a ∈ ( 0 , 1 ) a\in(0,1) a ∈ ( 0 , 1 ) and ε > 0 \varepsilon>0 ε > 0 there are C a , ε > 0 C_{a,\varepsilon}>0 C a , ε > 0 and t a , ε < ∞ t_{a,\varepsilon}<\infty t a , ε < ∞ such that
U ( a t ) ≥ C a , ε t γ − ε , t ≥ t a , ε . (37) U(at)\ge C_{a,\varepsilon}t^{\gamma-\varepsilon},\qquad t\ge t_{a,\varepsilon}. \tag*{(37)} U ( a t ) ≥ C a , ε t γ − ε , t ≥ t a , ε . (37)
We shall also use
sup t ≥ 0 ∫ [ 0 , t ] ( 1 + t − s ) − γ U ( d s ) < ∞ . (38) \sup_{t\ge0}\int_{[0,t]}(1+t-s)^{-\gamma}U(\mathrm{d}s)<\infty. \tag*{(38)} t ≥ 0 sup ∫ [ 0 , t ] ( 1 + t − s ) − γ U ( d s ) < ∞. (38)
Indeed, let S 0 = 0 S_0=0 S 0 = 0 and S n = σ 1 + ⋯ + σ n S_n=\sigma_1+\cdots+\sigma_n S n = σ 1 + ⋯ + σ n . As in the proof of Lemma 4.1, S n → ∞ S_n\to\infty S n → ∞ almost surely, and therefore
∫ [ 0 , t ] Γ ‾ ( t − s ) R ( d s ) = ∑ n ≥ 0 P { S n ≤ t < S n + 1 } = 1. \int_{[0,t]}\overline{\Gamma}(t-s)R(\mathrm{d}s)=\sum_{n\ge0}\mathbb{P}\{S_n\le t<S_{n+1}\}=1. ∫ [ 0 , t ] Γ ( t − s ) R ( d s ) = n ≥ 0 ∑ P { S n ≤ t < S n + 1 } = 1.
Moreover, (2.1) implies that there exists c γ > 0 c_\gamma>0 c γ > 0 such that
Γ ‾ ( r ) ≥ c γ ( 1 + r ) − γ , r ≥ 0. \overline{\Gamma}(r)\ge c_\gamma(1+r)^{-\gamma},\qquad r\ge0. Γ ( r ) ≥ c γ ( 1 + r ) − γ , r ≥ 0.
Since U ≤ R U \le R U ≤ R as positive measures, for every t ≥ 0 t \ge0 t ≥ 0 ,
∫ [ 0 , t ] ( 1 + t − s ) − γ U ( d s ) ≤ ∫ [ 0 , t ] ( 1 + t − s ) − γ R ( d s ) ≤ c γ − 1 ∫ [ 0 , t ] Γ ‾ ( t − s ) R ( d s ) = c γ − 1 . \int_{[0,t]} (1+t-s)^{-\gamma}U(\mathrm{d}s) \le\int_{[0,t]} (1+t-s)^{-\gamma}R(\mathrm{d}s) \le c_{\gamma}^{-1}\int_{[0,t]} \overline{\Gamma}(t-s)R(\mathrm{d}s)=c_{\gamma}^{-1}. ∫ [ 0 , t ] ( 1 + t − s ) − γ U ( d s ) ≤ ∫ [ 0 , t ] ( 1 + t − s ) − γ R ( d s ) ≤ c γ − 1 ∫ [ 0 , t ] Γ ( t − s ) R ( d s ) = c γ − 1 .
This proves (38) . Finally, (6) implies
E σ = ∫ 0 ∞ Γ ‾ ( s ) d s = ∞ , \mathbb{E}\sigma=\int_{0}^{\infty}\overline{\Gamma}(s)\,\mathrm{d}s=\infty, E σ = ∫ 0 ∞ Γ ( s ) d s = ∞ ,
since 0 < γ ≤ 1 0<\gamma\le1 0 < γ ≤ 1 . If, in addition, Γ \Gamma Γ is nonarithmetic, Blackwell’s renewal theorem in the infinite-mean case gives, for every fixed ℓ > 0 \ell>0 ℓ > 0 ,
U ( t + ℓ ) − U ( t ) ⟶ 0. (39) U(t+\ell)-U(t)\longrightarrow0. \tag*{(39)} U ( t + ℓ ) − U ( t ) ⟶ 0. (39)
See Thorisson [19 Eq. (1.1), p. 88] . For every ϑ > γ \vartheta>\gamma ϑ > γ ,
∫ [ 0 , t ] ( 1 + t − s ) − ϑ U ( d s ) ⟶ 0. (40) \int_{[0,t]}(1+t-s)^{-\vartheta}U(\mathrm{d}s)\longrightarrow0. \tag*{(40)} ∫ [ 0 , t ] ( 1 + t − s ) − ϑ U ( d s ) ⟶ 0. (40)
Indeed, fix T > 0 T>0 T > 0 . For t > T t>T t > T , (38) gives
∫ [ 0 , t − T ] ( 1 + t − s ) − ϑ U ( d s ) ≤ C ( 1 + T ) − ( ϑ − γ ) . \int_{[0,t-T]}(1+t-s)^{-\vartheta}U(\mathrm{d}s)\le C(1+T)^{-(\vartheta-\gamma)}. ∫ [ 0 , t − T ] ( 1 + t − s ) − ϑ U ( d s ) ≤ C ( 1 + T ) − ( ϑ − γ ) .
On the other hand,
∫ ( t − T , t ] ( 1 + t − s ) − ϑ U ( d s ) ≤ U ( t ) − U ( t − T ) ⟶ 0 \int_{(t-T,t]}(1+t-s)^{-\vartheta}U(\mathrm{d}s)\le U(t)-U(t-T)\longrightarrow0 ∫ ( t − T , t ] ( 1 + t − s ) − ϑ U ( d s ) ≤ U ( t ) − U ( t − T ) ⟶ 0
by (39) . Hence
lim sup t → ∞ ∫ [ 0 , t ] ( 1 + t − s ) − ϑ U ( d s ) ≤ C ( 1 + T ) − ( ϑ − γ ) . \limsup_{t\to\infty}\int_{[0,t]}(1+t-s)^{-\vartheta}U(\mathrm{d}s)\le C(1+T)^{-(\vartheta-\gamma)}. t → ∞ lim sup ∫ [ 0 , t ] ( 1 + t − s ) − ϑ U ( d s ) ≤ C ( 1 + T ) − ( ϑ − γ ) .
Letting T → ∞ T\to\infty T → ∞ proves (40) .
Hitting estimates
Since N ( t ) = Z t ( E ) N(t)=Z_t(E) N ( t ) = Z t ( E ) depends only on the Bellman–Harris genealogy, its law does not depend on the spatial motion or the starting position. For 0 ≤ s ≤ 1 0\le s\le1 0 ≤ s ≤ 1 , set
Q ( t ; s ) : = 1 − E [ s N ( t ) ] . Q(t;s):=1-\mathbb{E}[s^{N(t)}]. Q ( t ; s ) := 1 − E [ s N ( t ) ] .
Lemma 4.3. Let K ⊂ E K \subset E K ⊂ E be compact and 0 ≤ u < t 0 \le u < t 0 ≤ u < t . If C K ( 1 + t − u ) − q ≤ 1 C_K(1+t-u)^{-q} \le1 C K ( 1 + t − u ) − q ≤ 1 , then
P x { Z t ( K ) > 0 } ≤ ( C K ( 1 + t − u ) − q c U ( u ) ) 1 / ( 1 + β ) \mathbb{P}_x\{Z_t(K)>0\}\le\left(\frac{C_K(1+t-u)^{-q}}{cU(u)}\right)^{1/(1+\beta)} P x { Z t ( K ) > 0 } ≤ ( c U ( u ) C K ( 1 + t − u ) − q ) 1/ ( 1 + β ) with the right-hand side interpreted as + ∞ +\infty + ∞ when U ( u ) = 0 U(u)=0 U ( u ) = 0 . Moreover,
lim N → ∞ sup t ≥ 1 ∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) = 0. (41) \lim_{N\to\infty}\sup_{t\ge1}\int_{E\setminus D_t(N)}\mathbb{P}_x\{Z_t(K)>0\}\lambda(\mathrm{d}x)=0. \tag*{(41)} N → ∞ lim t ≥ 1 sup ∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) = 0. (41)
Proof. Set p = C K ( 1 + t − u ) − q p=C_K(1+t-u)^{-q} p = C K ( 1 + t − u ) − q . Conditionally on F u \mathcal{F}_u F u , if the particles alive at time u u u have positions and ages ( y j , θ j ) 1 ≤ j ≤ N ( u ) (y_j,\theta_j)_{1\le j\le N(u)} ( y j , θ j ) 1 ≤ j ≤ N ( u ) , then (4.3) and (H2) give
p j : = P y j θ j { Z t − u ( K ) > 0 } ≤ p . p_j:=\mathbb{P}_{y_j}^{\theta_j}\{Z_{t-u}(K)>0\}\le p. p j := P y j θ j { Z t − u ( K ) > 0 } ≤ p .
Conditionally on F u \mathcal{F}_u F u , the future descendant families are independent. Hence
P x { Z t ( K ) = 0 ∣ F u } = ∏ j = 1 N ( u ) ( 1 − p j ) ≥ ( 1 − p ) N ( u ) . \mathbb{P}_x\{Z_t(K)=0\mid\mathcal{F}_u\}=\prod_{j=1}^{N(u)}(1-p_j)\ge(1-p)^{N(u)}. P x { Z t ( K ) = 0 ∣ F u } = j = 1 ∏ N ( u ) ( 1 − p j ) ≥ ( 1 − p ) N ( u ) .
Taking complements and then expectations gives
P x { Z t ( K ) > 0 } ≤ 1 − E [ ( 1 − p ) N ( u ) ] = Q ( u ; 1 − p ) . \mathbb{P}_x\{Z_t(K)>0\}\le1-\mathbb{E}[(1-p)^{N(u)}]=Q(u;1-p). P x { Z t ( K ) > 0 } ≤ 1 − E [( 1 − p ) N ( u ) ] = Q ( u ; 1 − p ) .
If p < 1 p<1 p < 1 , Vatutin and Wakolbinger [2 Lemma 2, Eq. (7)] , applied with s = 1 − p s=1-p s = 1 − p , gives
Q ( u ; 1 − p ) ≤ ( p c U ( u ) ) 1 / ( 1 + β ) . Q(u;1-p)\le\left(\frac{p}{cU(u)}\right)^{1/(1+\beta)}. Q ( u ; 1 − p ) ≤ ( c U ( u ) p ) 1/ ( 1 + β ) .
If p = 1 p=1 p = 1 , the same estimate follows by letting s ↓ 0 s\downarrow0 s ↓ 0 in Vatutin and Wakolbinger [2, Lemma 2, Eq. (7)]. This proves the first assertion. The preceding argument is the analogue of Vatutin and Wakolbinger [2 Lemma 3] `, with (H2) replacing the stable transition estimate.
For the second assertion, the one-family bound (4.3) yields
∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) ≤ ∫ E ∖ D t ( N ) P t 1 K ( x ) λ ( d x ) . \int_{E\setminus D_t(N)}\mathbb{P}_x\{Z_t(K)>0\}\lambda(\mathrm{d}x)\le\int_{E\setminus D_t(N)}P_t1_K(x)\lambda(\mathrm{d}x). ∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) ≤ ∫ E ∖ D t ( N ) P t 1 K ( x ) λ ( d x ) .
Taking the supremum over t ≥ 1 t\ge1 t ≥ 1 and then letting N → ∞ N\to\infty N → ∞ , (H3) gives (4.11).
The Laplace renewal equation
For φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) , set
V t φ ( x ) : = E x [ 1 − e − ⟨ Z t , φ ⟩ ] , G φ ( t ) : = ∫ E V t φ ( x ) λ ( d x ) , V_t^\varphi(x) := \mathbb{E}_x\left[1-e^{-\langle Z_t,\varphi\rangle}\right], \qquad G_\varphi(t):=\int_E V_t^\varphi(x)\lambda(\mathrm{d}x), V t φ ( x ) := E x [ 1 − e − ⟨ Z t , φ ⟩ ] , G φ ( t ) := ∫ E V t φ ( x ) λ ( d x ) ,
b ( φ ) : = ∫ E ( 1 − e − φ ( x ) ) λ ( d x ) , g φ ( t ) : = c ∫ E V t φ ( x ) 1 + β λ ( d x ) . b(\varphi):=\int_E\left(1-e^{-\varphi(x)}\right)\lambda(\mathrm{d}x), \qquad g_\varphi(t):=c\int_E V_t^\varphi(x)^{1+\beta}\lambda(\mathrm{d}x). b ( φ ) := ∫ E ( 1 − e − φ ( x ) ) λ ( d x ) , g φ ( t ) := c ∫ E V t φ ( x ) 1 + β λ ( d x ) .
Lemma 4.4. For every φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) ,
G φ ( t ) = b ( φ ) − ( U ∗ g φ ) ( t ) . (42) G_\varphi(t)=b(\varphi)-(U*g_\varphi)(t). \tag*{(42)} G φ ( t ) = b ( φ ) − ( U ∗ g φ ) ( t ) . (42)
Moreover,
0 ≤ G φ ( t ) ≤ b ( φ ) , V t φ ( x ) ≤ P t φ ( x ) , 0\leq G_\varphi(t)\leq b(\varphi), \qquad V_t^\varphi(x)\leq P_t\varphi(x), 0 ≤ G φ ( t ) ≤ b ( φ ) , V t φ ( x ) ≤ P t φ ( x ) ,
g φ ( t ) ≤ c C φ β ( 1 + t ) − q β G φ ( t ) ≤ c C φ β b ( φ ) ( 1 + t ) − q β , for some 0 < C φ < ∞ . (43) g_\varphi(t)\leq cC_\varphi^\beta(1+t)^{-q\beta}G_\varphi(t)\leq cC_\varphi^\beta b(\varphi)(1+t)^{-q\beta}, \qquad\text{for some }0<C_\varphi<\infty. \tag*{(43)} g φ ( t ) ≤ c C φ β ( 1 + t ) − q β G φ ( t ) ≤ c C φ β b ( φ ) ( 1 + t ) − q β , for some 0 < C φ < ∞. (43)
and
E e − ⟨ X t , φ ⟩ = e − G φ ( t ) . (44) \mathbb{E}e^{-\langle X_t,\varphi\rangle}=e^{-G_\varphi(t)}. \tag*{(44)} E e − ⟨ X t , φ ⟩ = e − G φ ( t ) . (44)
Proof. The inequality 1 − e − z ≤ z 1-e^{-z}\leq z 1 − e − z ≤ z and Lemma 4.1 give
0 ≤ V t φ ( x ) ≤ P t φ ( x ) . 0\leq V_t^\varphi(x)\leq P_t\varphi(x). 0 ≤ V t φ ( x ) ≤ P t φ ( x ) .
Therefore, by (H1),
G φ ( t ) ≤ ∫ E P t φ d λ = ∫ E φ d λ < ∞ , G_\varphi(t)\leq\int_E P_t\varphi\,\mathrm{d}\lambda=\int_E\varphi\,\mathrm{d}\lambda<\infty, G φ ( t ) ≤ ∫ E P t φ d λ = ∫ E φ d λ < ∞ ,
since φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) . Moreover, 0 ≤ V t φ ≤ 1 0\leq V_t^\varphi\leq1 0 ≤ V t φ ≤ 1 , and hence
g φ ( t ) = c ∫ E ( V t φ ) 1 + β d λ ≤ c G φ ( t ) < ∞ . g_\varphi(t)=c\int_E(V_t^\varphi)^{1+\beta}\,\mathrm{d}\lambda\leq cG_\varphi(t)<\infty. g φ ( t ) = c ∫ E ( V t φ ) 1 + β d λ ≤ c G φ ( t ) < ∞.
Conditioning on the first death, exactly as in Vatutin and Wakolbinger [2 Eqs. (9)–(13), p. 626] , and using 1 − f ( 1 − z ) = z − c z 1 + β 1-f(1-z)=z-cz^{1+\beta} 1 − f ( 1 − z ) = z − c z 1 + β , then integrating over the initial position and using (H1), gives
G φ = Γ ‾ b ( φ ) + Γ ‾ ∗ G φ − Γ ‾ ∗ g φ . G_\varphi=\overline{\Gamma}b(\varphi)+\overline{\Gamma}*G_\varphi-\overline{\Gamma}*g_\varphi. G φ = Γ b ( φ ) + Γ ∗ G φ − Γ ∗ g φ .
Solving this scalar renewal equation gives (42) , as in Vatutin and Wakolbinger [2 Eqs. (13)–(15), p. 626] . Since g φ ≥ 0 g_\varphi\geq0 g φ ≥ 0 , (4.12) yields 0 ≤ G φ ≤ b ( φ ) 0\leq G_\varphi\leq b(\varphi) 0 ≤ G φ ≤ b ( φ ) .
If K = supp φ K=\operatorname{supp}\varphi K = supp φ , then (H2) gives sup x V t φ ( x ) ≤ C φ ( 1 + t ) − q \sup_x V_t^\varphi(x)\leq C_\varphi(1+t)^{-q} sup x V t φ ( x ) ≤ C φ ( 1 + t ) − q , for some constant 0 < C φ < ∞ 0<C_\varphi<\infty 0 < C φ < ∞ and therefore
g φ ( t ) ≤ c ( sup x V t φ ( x ) ) β G φ ( t ) ≤ c C φ β ( 1 + t ) − q β G φ ( t ) ≤ c C φ β b ( φ ) ( 1 + t ) − q β , g_\varphi(t)\leq c\left(\sup_x V_t^\varphi(x)\right)^\beta G_\varphi(t)\leq cC_\varphi^\beta(1+t)^{-q\beta}G_\varphi(t)\leq cC_\varphi^\beta b(\varphi)(1+t)^{-q\beta}, g φ ( t ) ≤ c ( x sup V t φ ( x ) ) β G φ ( t ) ≤ c C φ β ( 1 + t ) − q β G φ ( t ) ≤ c C φ β b ( φ ) ( 1 + t ) − q β ,
which is (4.13). Finally, (4.14) is the Poisson exponential formula applied to the independently marked initial field.
Proofs of the extinction–survival results
By Corollary 4.2, the laws of X t X_t X t are tight in M ( E ) \mathcal{M}(E) M ( E ) . We use the Laplace-functional convergence criterion Kallenberg [16, Theorem 4.2] below.
Regularly varying lifetimes
Proof of Theorem 2.2. Fix φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) and a compact K ⊃ supp φ K \supset\operatorname{supp}\varphi K ⊃ supp φ . The argument follows Vatutin and Wakolbinger [2, Sections 4–5, pp. 626–630].
Case q β < γ q\beta< \gamma q β < γ . Choose ε ∈ ( 0 , γ − q β ) \varepsilon\in(0,\gamma-q\beta) ε ∈ ( 0 , γ − q β ) and a ∈ ( 0 , 1 ) a \in(0,1) a ∈ ( 0 , 1 ) . Applying the first assertion of Lemma 4.3 with u = a t u=at u = a t , and using (4.7) together with λ ( D t ( N ) ) ≤ C N t q \lambda(D_t(N)) \le C_N t^q λ ( D t ( N )) ≤ C N t q , gives, for every fixed N ≥ 1 N \ge1 N ≥ 1 and all sufficiently large t t t ,
∫ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) ≤ C K , N , a , ε t ( q β − γ + ε ) / ( 1 + β ) ⟶ 0. \int_{D_t(N)} \mathbb{P}_x\{Z_t(K)>0\}\lambda(\mathrm{d}x) \le C_{K,N,a,\varepsilon}t^{(q\beta-\gamma+\varepsilon)/(1+\beta)} \longrightarrow0. ∫ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) ≤ C K , N , a , ε t ( q β − γ + ε ) / ( 1 + β ) ⟶ 0.
The second assertion of Lemma 4.3 gives
lim sup N → ∞ sup t ≥ 1 ∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) = 0. \limsup_{N\to\infty}\sup_{t\ge1}\int_{E\setminus D_t(N)} \mathbb{P}_x\{Z_t(K)>0\}\lambda(\mathrm{d}x)=0. N → ∞ lim sup t ≥ 1 sup ∫ E ∖ D t ( N ) P x { Z t ( K ) > 0 } λ ( d x ) = 0.
Hence, letting first t → ∞ t\to\infty t → ∞ and then N → ∞ N\to\infty N → ∞ ,
∫ E P x { Z t ( K ) > 0 } λ ( d x ) ⟶ 0. \int_E \mathbb{P}_x\{Z_t(K)>0\}\lambda(\mathrm{d}x)\longrightarrow0. ∫ E P x { Z t ( K ) > 0 } λ ( d x ) ⟶ 0.
Since
1 − e − ⟨ Z t , φ ⟩ ≤ 1 { Z t ( K ) > 0 } , 1-e^{-\langle Z_t,\varphi\rangle}\le\mathbf{1}_{\{Z_t(K)>0\}}, 1 − e − ⟨ Z t , φ ⟩ ≤ 1 { Z t ( K ) > 0 } ,
we obtain G φ ( t ) → 0 G_\varphi(t)\to0 G φ ( t ) → 0 , and (4.14) yields X t ⟹ 0 X_t\Longrightarrow0 X t ⟹ 0 .
Case q β > γ q\beta>\gamma q β > γ . By (4.13),
g φ ( r ) ≤ c C φ β b ( φ ) ( 1 + r ) − q β , r ≥ 0. g_\varphi(r)\le cC_\varphi^\beta b(\varphi)(1+r)^{-q\beta},\qquad r\ge0. g φ ( r ) ≤ c C φ β b ( φ ) ( 1 + r ) − q β , r ≥ 0.
Hence
( U ∗ g φ ) ( t ) = ∫ [ 0 , t ] g φ ( t − s ) U ( d s ) ≤ c C φ β b ( φ ) ∫ [ 0 , t ] ( 1 + t − s ) − q β U ( d s ) . \begin{aligned}
(U*g_\varphi)(t)=\int_{[0,t]}g_\varphi(t-s)\,U(\mathrm{d}s) \\
&\le cC_\varphi^\beta b(\varphi)\int_{[0,t]}(1+t-s)^{-q\beta}\,U(\mathrm{d}s).
\end{aligned} ( U ∗ g φ ) ( t ) = ∫ [ 0 , t ] g φ ( t − s ) U ( d s ) ≤ c C φ β b ( φ ) ∫ [ 0 , t ] ( 1 + t − s ) − q β U ( d s ) .
Since q β > γ q\beta>\gamma q β > γ , (4.10), applied with ϑ = q β \vartheta=q\beta ϑ = q β , gives
( U ∗ g φ ) ( t ) ⟶ 0. (U*g_\varphi)(t)\longrightarrow0. ( U ∗ g φ ) ( t ) ⟶ 0.
Therefore (4.12) yields
G φ ( t ) ⟶ b ( φ ) . G_\varphi(t)\longrightarrow b(\varphi). G φ ( t ) ⟶ b ( φ ) .
Since b ( φ ) b(\varphi) b ( φ ) is the Laplace exponent of the Poisson random measure Π λ \Pi_\lambda Π λ , (44) and Kallenberg [16 Theorem 4.2] give X t → d Π λ X_t \xrightarrow{\mathrm{d}} \Pi_\lambda X t d Π λ .
Case q β = γ q\beta=\gamma q β = γ . Fix a nonzero φ \varphi φ and suppose, for contradiction, that G φ ( t ) → 0 G_\varphi(t)\to0 G φ ( t ) → 0 . For an arbitrary η > 0 \eta>0 η > 0 , choose T η < ∞ T_\eta<\infty T η < ∞ such that
G φ ( s ) ≤ η , s ≥ T η . G_\varphi(s)\le\eta,\qquad s\ge T_\eta. G φ ( s ) ≤ η , s ≥ T η .
Since q β = γ q\beta=\gamma q β = γ , (43) gives
g φ ( s ) ≤ c C φ β η ( 1 + s ) − γ , s ≥ T η . g_\varphi(s)\le cC_\varphi^\beta\eta(1+s)^{-\gamma},\qquad s\ge T_\eta. g φ ( s ) ≤ c C φ β η ( 1 + s ) − γ , s ≥ T η .
Set
M γ : = sup r ≥ 0 ∫ [ 0 , r ] ( 1 + r − u ) − γ U ( d u ) < ∞ M_\gamma:=\sup_{r\ge0}\int_{[0,r]}(1+r-u)^{-\gamma}U(\mathrm{d}u)<\infty M γ := r ≥ 0 sup ∫ [ 0 , r ] ( 1 + r − u ) − γ U ( d u ) < ∞
by (38) . For t > T η t>T_\eta t > T η , splitting the renewal convolution at t − T η t-T_\eta t − T η and using again (43) gives
( U ∗ g φ ) ( t ) = ∫ [ 0 , t − T η ] g φ ( t − u ) U ( d u ) + ∫ ( t − T η , t ] g φ ( t − u ) U ( d u ) ≤ c C φ β η ∫ [ 0 , t − T η ] ( 1 + t − u ) − γ U ( d u ) + c C φ β b ( φ ) [ U ( t ) − U ( t − T η ) ] ≤ c C φ β M γ η + c C φ β b ( φ ) [ U ( t ) − U ( t − T η ) ] . \begin{aligned}
(U*g_\varphi)(t)
&=\int_{[0,t-T_\eta]}g_\varphi(t-u)U(\mathrm{d}u)+\int_{(t-T_\eta,t]}g_\varphi(t-u)U(\mathrm{d}u)\\
&\le cC_\varphi^\beta\eta\int_{[0,t-T_\eta]}(1+t-u)^{-\gamma}U(\mathrm{d}u)\\
&\quad+cC_\varphi^\beta b(\varphi)\left[U(t)-U(t-T_\eta)\right]\\
&\le cC_\varphi^\beta M_\gamma\eta+cC_\varphi^\beta b(\varphi)\left[U(t)-U(t-T_\eta)\right].
\end{aligned} ( U ∗ g φ ) ( t ) = ∫ [ 0 , t − T η ] g φ ( t − u ) U ( d u ) + ∫ ( t − T η , t ] g φ ( t − u ) U ( d u ) ≤ c C φ β η ∫ [ 0 , t − T η ] ( 1 + t − u ) − γ U ( d u ) + c C φ β b ( φ ) [ U ( t ) − U ( t − T η ) ] ≤ c C φ β M γ η + c C φ β b ( φ ) [ U ( t ) − U ( t − T η ) ] .
By (39) ,
U ( t ) − U ( t − T η ) ⟶ 0 , t → ∞ , U(t)-U(t-T_\eta)\longrightarrow0,\qquad t\to\infty, U ( t ) − U ( t − T η ) ⟶ 0 , t → ∞ ,
and therefore
lim sup t → ∞ ( U ∗ g φ ) ( t ) ≤ c C φ β M γ η . \limsup_{t\to\infty}(U*g_\varphi)(t)\le cC_\varphi^\beta M_\gamma\eta. t → ∞ lim sup ( U ∗ g φ ) ( t ) ≤ c C φ β M γ η .
Since η > 0 \eta>0 η > 0 is arbitrary and ( U ∗ g φ ) ( t ) ≥ 0 (U*g_\varphi)(t)\ge0 ( U ∗ g φ ) ( t ) ≥ 0 ,
( U ∗ g φ ) ( t ) ⟶ 0. (U*g_\varphi)(t)\longrightarrow0. ( U ∗ g φ ) ( t ) ⟶ 0.
Equation (42) would then imply
G φ ( t ) ⟶ b ( φ ) > 0 , G_\varphi(t)\longrightarrow b(\varphi)>0, G φ ( t ) ⟶ b ( φ ) > 0 ,
contradicting the assumption G φ ( t ) → 0 G_\varphi(t)\to0 G φ ( t ) → 0 . Since
lim sup t → ∞ G φ ( t ) > 0 , \limsup_{t\to\infty}G_\varphi(t)>0, t → ∞ lim sup G φ ( t ) > 0 ,
choose
0 < a φ < lim sup t → ∞ G φ ( t ) 0 < a_{\varphi} < \limsup_{t \to\infty} G_{\varphi}(t) 0 < a φ < t → ∞ lim sup G φ ( t )
and a sequence t n → ∞ t_n \to\infty t n → ∞ such that
G φ ( t n ) ≥ a φ , n ≥ 1. G_{\varphi}(t_n) \ge a_{\varphi}, \qquad n \ge1. G φ ( t n ) ≥ a φ , n ≥ 1.
Since
a φ ≤ G φ ( t n ) ≤ b ( φ ) , n ≥ 1 , a_{\varphi} \le G_{\varphi}(t_n) \le b(\varphi), \qquad n \ge1, a φ ≤ G φ ( t n ) ≤ b ( φ ) , n ≥ 1 ,
there exist a subsequence ( t n , 1 ) n ≥ 1 (t_{n,1})_{n \ge1} ( t n , 1 ) n ≥ 1 of ( t n ) n ≥ 1 (t_n)_{n \ge1} ( t n ) n ≥ 1 and some ℓ φ ∈ [ a φ , b ( φ ) ] \ell_{\varphi} \in[a_{\varphi}, b(\varphi)] ℓ φ ∈ [ a φ , b ( φ )] such that
G φ ( t n , 1 ) ⟶ ℓ φ . G_{\varphi}(t_{n,1}) \longrightarrow\ell_{\varphi}. G φ ( t n , 1 ) ⟶ ℓ φ .
In particular, ℓ φ > 0 \ell_{\varphi} > 0 ℓ φ > 0 .
Since { L ( X t ) : t ≥ 0 } \{L(X_t) : t \ge0\} { L ( X t ) : t ≥ 0 } is tight in M ( E ) \mathcal{M}(E) M ( E ) , Prokhorov’s theorem yields a subsequence ( t n , 2 ) n ≥ 1 (t_{n,2})_{n \ge1} ( t n , 2 ) n ≥ 1 of ( t n , 1 ) n ≥ 1 (t_{n,1})_{n \ge1} ( t n , 1 ) n ≥ 1 and a random Radon measure X ∞ X_{\infty} X ∞ such that
X t n , 2 ⟹ X ∞ . X_{t_{n,2}} \Longrightarrow X_{\infty}. X t n , 2 ⟹ X ∞ .
Because ( t n , 2 ) n ≥ 1 (t_{n,2})_{n \ge1} ( t n , 2 ) n ≥ 1 is a subsequence of ( t n , 1 ) n ≥ 1 (t_{n,1})_{n \ge1} ( t n , 1 ) n ≥ 1 ,
G φ ( t n , 2 ) ⟶ ℓ φ . G_{\varphi}(t_{n,2}) \longrightarrow\ell_{\varphi}. G φ ( t n , 2 ) ⟶ ℓ φ .
Now define
F φ ( ν ) : = e − ⟨ ν , φ ⟩ , ν ∈ M ( E ) . F_{\varphi}(\nu) := e^{-\langle\nu,\varphi\rangle}, \qquad\nu\in\mathcal{M}(E). F φ ( ν ) := e − ⟨ ν , φ ⟩ , ν ∈ M ( E ) .
Since φ ∈ C + ( E ) \varphi\in C^{+}(E) φ ∈ C + ( E ) , the map ν ↦ ⟨ ν , φ ⟩ \nu\mapsto\langle\nu,\varphi\rangle ν ↦ ⟨ ν , φ ⟩ is continuous for the vague topology. Hence F φ ∈ C b ( M ( E ) ) F_{\varphi} \in C_b(\mathcal{M}(E)) F φ ∈ C b ( M ( E )) . Therefore
E e − ⟨ X ∞ , φ ⟩ = lim n → ∞ E e − ⟨ X t n , 2 , φ ⟩ = lim n → ∞ e − G φ ( t n , 2 ) = e − ℓ φ < 1 , \begin{aligned}
\mathbb{E}e^{-\langle X_{\infty},\varphi\rangle}
= \lim_{n \to\infty} \mathbb{E}e^{-\langle X_{t_{n,2}},\varphi\rangle} \\
&= \lim_{n \to\infty} e^{-G_{\varphi}(t_{n,2})}
= e^{-\ell_{\varphi}} < 1,
\end{aligned} E e − ⟨ X ∞ , φ ⟩ = n → ∞ lim E e − ⟨ X t n , 2 , φ ⟩ = n → ∞ lim e − G φ ( t n , 2 ) = e − ℓ φ < 1 ,
where the second equality follows from (44) . Consequently,
P { X ∞ ≠ 0 } > 0. \mathbb{P}\{X_{\infty} \ne0\} > 0. P { X ∞ = 0 } > 0.
Finite-mean lifetimes
Proof of Theorem 2.1. Fix φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) . The renewal argument follows Vatutin and Wakolbinger [2 Theorems 4–5, pp. 630–632] .
Case q β > 1 q\beta> 1 q β > 1 . By (43) ,
0 ≤ g φ ( t ) ≤ c C β b ( φ ) ( 1 + t ) − q β , t ≥ 0. 0 \leq g_\varphi(t) \leq cC^\beta b(\varphi)(1+t)^{-q\beta}, \qquad t \geq0. 0 ≤ g φ ( t ) ≤ c C β b ( φ ) ( 1 + t ) − q β , t ≥ 0.
Since q β > 1 q\beta> 1 q β > 1 , the function g φ g_\varphi g φ is bounded and integrable, and g φ ( t ) → 0 g_\varphi(t) \to0 g φ ( t ) → 0 . Hence (36) gives
( U ∗ g φ ) ( t ) ⟶ 1 μ ∫ 0 ∞ g φ ( s ) d s . (U * g_\varphi)(t) \longrightarrow\frac{1}{\mu}\int_0^\infty g_\varphi(s)\,\mathrm{d}s. ( U ∗ g φ ) ( t ) ⟶ μ 1 ∫ 0 ∞ g φ ( s ) d s .
Together with (42) , this yields
G φ ( t ) ⟶ G ( φ ) : = b ( φ ) − 1 μ ∫ 0 ∞ g φ ( s ) d s . G_\varphi(t) \longrightarrow G(\varphi) := b(\varphi)-\frac{1}{\mu}\int_0^\infty g_\varphi(s)\,\mathrm{d}s. G φ ( t ) ⟶ G ( φ ) := b ( φ ) − μ 1 ∫ 0 ∞ g φ ( s ) d s .
Therefore, by (44) and Kallenberg [16 Theorem 4.2] ,
X t ⇒ X ∞ X_t \Rightarrow X_\infty X t ⇒ X ∞
for some random Radon measure X ∞ X_\infty X ∞ whose Laplace functional is
E e − ⟨ X ∞ , φ ⟩ = e − G ( φ ) . \mathbb{E}e^{-\langle X_\infty,\varphi\rangle}=e^{-G(\varphi)}. E e − ⟨ X ∞ , φ ⟩ = e − G ( φ ) .
It remains to identify its intensity. Let 0 < ε ≤ 1 0 < \varepsilon\leq1 0 < ε ≤ 1 . Since 1 − e − ε z ≤ ε z 1-e^{-\varepsilon z}\leq\varepsilon z 1 − e − ε z ≤ ε z , Lemma 4.1 gives
V t ε φ ( x ) ≤ ε P t φ ( x ) . V_t^{\varepsilon\varphi}(x)\leq\varepsilon P_t\varphi(x). V t εφ ( x ) ≤ ε P t φ ( x ) .
Consequently,
G ε φ ( t ) ≤ ε ∫ E P t φ d λ = ε ∫ E φ d λ G_{\varepsilon\varphi}(t)\leq\varepsilon\int_E P_t\varphi\,\mathrm{d}\lambda=\varepsilon\int_E\varphi\,\mathrm{d}\lambda G εφ ( t ) ≤ ε ∫ E P t φ d λ = ε ∫ E φ d λ
by (H1). Moreover, the same constant C φ C_\varphi C φ as in (43) gives
sup x V t ε φ ( x ) ≤ ε C φ ( 1 + t ) − q , \sup_x V_t^{\varepsilon\varphi}(x)\leq\varepsilon C_\varphi(1+t)^{-q}, x sup V t εφ ( x ) ≤ ε C φ ( 1 + t ) − q ,
and therefore
g ε φ ( t ) ≤ c ( sup x V t ε φ ( x ) ) β G ε φ ( t ) ≤ c C φ β ε 1 + β ( ∫ E φ d λ ) ( 1 + t ) − q β . \begin{aligned}
g_{\varepsilon\varphi}(t)\leq c\left(\sup_x V_t^{\varepsilon\varphi}(x)\right)^\beta G_{\varepsilon\varphi}(t) \\
&\leq cC_\varphi^{\beta}\varepsilon^{1+\beta}\left(\int_E\varphi\,\mathrm{d}\lambda\right)(1+t)^{-q\beta}.
\end{aligned} g εφ ( t ) ≤ c ( x sup V t εφ ( x ) ) β G εφ ( t ) ≤ c C φ β ε 1 + β ( ∫ E φ d λ ) ( 1 + t ) − q β .
Since q β > 1 q\beta>1 q β > 1 ,
1 ε ∫ 0 ∞ g ε φ ( t ) d t ⟶ 0 , ε ↓ 0. \frac{1}{\varepsilon}\int_0^\infty g_{\varepsilon\varphi}(t)\,\mathrm{d}t\longrightarrow0,\qquad\varepsilon\downarrow0. ε 1 ∫ 0 ∞ g εφ ( t ) d t ⟶ 0 , ε ↓ 0.
Also, by dominated convergence,
b ( ε φ ) ε = ∫ E 1 − e − ε φ ( x ) ε λ ( d x ) ⟶ ∫ E φ d λ . \frac{b(\varepsilon\varphi)}{\varepsilon}=\int_E \frac{1-e^{-\varepsilon\varphi(x)}}{\varepsilon}\lambda(\mathrm{d}x)\longrightarrow\int_E\varphi\,\mathrm{d}\lambda. ε b ( εφ ) = ∫ E ε 1 − e − εφ ( x ) λ ( d x ) ⟶ ∫ E φ d λ .
Thus
G ( ε φ ) ε ⟶ ∫ E φ d λ . \frac{G(\varepsilon\varphi)}{\varepsilon}\longrightarrow\int_E\varphi\,\mathrm{d}\lambda. ε G ( εφ ) ⟶ ∫ E φ d λ .
Now set Y = ⟨ X ∞ , φ ⟩ Y=\langle X_\infty,\varphi\rangle Y = ⟨ X ∞ , φ ⟩ . Since
1 − e − ε Y ε ↑ Y as ε ↓ 0 , \frac{1-e^{-\varepsilon Y}}{\varepsilon}\uparrow Y\quad\text{as }\varepsilon\downarrow0, ε 1 − e − ε Y ↑ Y as ε ↓ 0 ,
monotone convergence and the Laplace functional of X ∞ X_\infty X ∞ give
E Y = lim ε ↓ 0 1 − E e − ε Y ε = lim ε ↓ 0 1 − e − G ( ε φ ) ε = ∫ E φ d λ . \begin{aligned}
\mathbb{E}Y&=\lim_{\varepsilon\downarrow0}\frac{1-\mathbb{E}e^{-\varepsilon Y}}{\varepsilon}\\
&=\lim_{\varepsilon\downarrow0}\frac{1-e^{-G(\varepsilon\varphi)}}{\varepsilon}=\int_E\varphi\,\mathrm{d}\lambda.
\end{aligned} E Y = ε ↓ 0 lim ε 1 − E e − ε Y = ε ↓ 0 lim ε 1 − e − G ( εφ ) = ∫ E φ d λ .
Hence X ∞ X_\infty X ∞ has intensity λ \lambda λ and, for every nonzero φ ∈ C c + ( E ) \varphi\in C_c^+(E) φ ∈ C c + ( E ) ,
E ⟨ X ∞ , φ ⟩ > 0 , \mathbb{E}\langle X_\infty,\varphi\rangle>0, E ⟨ X ∞ , φ ⟩ > 0 ,
so X ∞ X_\infty X ∞ is nontrivial.
Case q β ≤ 1 q\beta\le1 q β ≤ 1 . For T > 0 T>0 T > 0 , set
r T ( t ) : = g φ ( t ) 1 [ 0 , T ] ( t ) . r_T(t):=g_\varphi(t)1_{[0,T]}(t). r T ( t ) := g φ ( t ) 1 [ 0 , T ] ( t ) .
Since 0 ≤ r T ≤ g φ 0\le r_T\le g_\varphi 0 ≤ r T ≤ g φ , (42) gives
0 ≤ ( U ∗ r T ) ( t ) ≤ ( U ∗ g φ ) ( t ) = b ( φ ) − G φ ( t ) ≤ b ( φ ) . 0\le(U*r_T)(t)\le(U*g_\varphi)(t)=b(\varphi)-G_\varphi(t)\le b(\varphi). 0 ≤ ( U ∗ r T ) ( t ) ≤ ( U ∗ g φ ) ( t ) = b ( φ ) − G φ ( t ) ≤ b ( φ ) .
The function r T r_T r T is bounded, integrable, and vanishes at infinity, so (36) yields
1 μ ∫ 0 T g φ ( s ) d s = lim t → ∞ ( U ∗ r T ) ( t ) ≤ b ( φ ) . \frac{1}{\mu}\int_0^T g_\varphi(s)\,\mathrm{d}s=\lim_{t\to\infty}(U*r_T)(t)\le b(\varphi). μ 1 ∫ 0 T g φ ( s ) d s = t → ∞ lim ( U ∗ r T ) ( t ) ≤ b ( φ ) .
Therefore
∫ 0 T g φ ( s ) d s ≤ μ b ( φ ) , T > 0 , \int_0^T g_\varphi(s)\,\mathrm{d}s\le\mu b(\varphi),\qquad T>0, ∫ 0 T g φ ( s ) d s ≤ μ b ( φ ) , T > 0 ,
and letting T → ∞ T\to\infty T → ∞ gives
g φ ∈ L 1 ( 0 , ∞ ) . g_\varphi\in L^1(0,\infty). g φ ∈ L 1 ( 0 , ∞ ) .
Since (43) also gives g φ ( t ) → 0 g_\varphi(t) \to0 g φ ( t ) → 0 , another application of (36) gives
( U ∗ g φ ) ( t ) → 1 μ ∫ 0 ∞ g φ ( s ) d s . (U * g_\varphi)(t) \to\frac{1}{\mu}\int_0^\infty g_\varphi(s)\,\mathrm{d}s. ( U ∗ g φ ) ( t ) → μ 1 ∫ 0 ∞ g φ ( s ) d s .
Hence, by (42) , G φ ( t ) G_\varphi(t) G φ ( t ) has a finite limit. We next show that this limit is zero. Since g φ ≥ 0 g_\varphi\ge0 g φ ≥ 0 is integrable and q β ≤ 1 q\beta\le1 q β ≤ 1 ,
lim inf t → ∞ t q β g φ ( t ) = 0. \liminf_{t\to\infty} t^{q\beta}g_\varphi(t)=0. t → ∞ lim inf t q β g φ ( t ) = 0.
Indeed, otherwise there would exist c > 0 c>0 c > 0 and T < ∞ T<\infty T < ∞ such that g φ ( t ) ≥ c t − q β g_\varphi(t)\ge ct^{-q\beta} g φ ( t ) ≥ c t − q β for all t > T t>T t > T , contradicting g φ ∈ L 1 ( 0 , ∞ ) g_\varphi\in L^1(0,\infty) g φ ∈ L 1 ( 0 , ∞ ) because ∫ T ∞ t − q β d t = ∞ \int_T^\infty t^{-q\beta}\,\mathrm{d}t=\infty ∫ T ∞ t − q β d t = ∞ . Thus there exists t n → ∞ t_n\to\infty t n → ∞ such that
t n q β g φ ( t n ) ⟶ 0. t_n^{q\beta}g_\varphi(t_n)\longrightarrow0. t n q β g φ ( t n ) ⟶ 0.
Fix N ≥ 1 N\ge1 N ≥ 1 . Hölder’s inequality and λ ( D t ( N ) ) ≤ C N t q \lambda(D_t(N))\le CNt^q λ ( D t ( N )) ≤ C N t q give
∫ D t ( N ) V t φ d λ ≤ λ ( D t ( N ) ) β / ( 1 + β ) ( ∫ E ( V t φ ) 1 + β d λ ) 1 / ( 1 + β ) = λ ( D t ( N ) ) β / ( 1 + β ) ( g φ ( t ) c ) 1 / ( 1 + β ) ≤ ( C β c t q β g φ ( t ) ) 1 / ( 1 + β ) . \begin{aligned}
\int_{D_t(N)} V_t^\varphi\,\mathrm{d}\lambda
&\le\lambda(D_t(N))^{\beta/(1+\beta)}
\left(\int_E (V_t^\varphi)^{1+\beta}\,\mathrm{d}\lambda\right)^{1/(1+\beta)}\\
&=\lambda(D_t(N))^{\beta/(1+\beta)}
\left(\frac{g_\varphi(t)}{c}\right)^{1/(1+\beta)}\\
&\le\left(\frac{C^\beta}{c}t^{q\beta}g_\varphi(t)\right)^{1/(1+\beta)}.
\end{aligned} ∫ D t ( N ) V t φ d λ ≤ λ ( D t ( N ) ) β / ( 1 + β ) ( ∫ E ( V t φ ) 1 + β d λ ) 1/ ( 1 + β ) = λ ( D t ( N ) ) β / ( 1 + β ) ( c g φ ( t ) ) 1/ ( 1 + β ) ≤ ( c C β t q β g φ ( t ) ) 1/ ( 1 + β ) .
Consequently,
∫ D t n ( N ) V t n φ d λ ⟶ 0 for every fixed N . \int_{D_{t_n}(N)} V_{t_n}^\varphi\,\mathrm{d}\lambda\longrightarrow0
\qquad\text{for every fixed }N. ∫ D t n ( N ) V t n φ d λ ⟶ 0 for every fixed N .
Let K = supp φ K=\operatorname{supp}\varphi K = supp φ . Since
V t φ ( x ) ≤ P t φ ( x ) ≤ ∥ φ ∥ ∞ P t 1 K ( x ) , V_t^\varphi(x)\le P_t\varphi(x)\le\|\varphi\|_\infty P_t1_K(x), V t φ ( x ) ≤ P t φ ( x ) ≤ ∥ φ ∥ ∞ P t 1 K ( x ) ,
we also have
∫ E ∖ D t ( N ) V t φ d λ ≤ ∥ φ ∥ ∞ ∫ E ∖ D t ( N ) P t 1 K d λ . \int_{E\setminus D_t(N)} V_t^\varphi\,\mathrm{d}\lambda
\le\|\varphi\|_\infty\int_{E\setminus D_t(N)}P_t1_K\,\mathrm{d}\lambda. ∫ E ∖ D t ( N ) V t φ d λ ≤ ∥ φ ∥ ∞ ∫ E ∖ D t ( N ) P t 1 K d λ .
It follows that, for every fixed N N N ,
lim sup n → ∞ G φ ( t n ) ≤ ∥ φ ∥ ∞ sup t ≥ 1 ∫ E ∖ D t ( N ) P t 1 K d λ . \limsup_{n\to\infty}G_\varphi(t_n)
\le\|\varphi\|_\infty\sup_{t\ge1}\int_{E\setminus D_t(N)}P_t1_K\,\mathrm{d}\lambda. n → ∞ lim sup G φ ( t n ) ≤ ∥ φ ∥ ∞ t ≥ 1 sup ∫ E ∖ D t ( N ) P t 1 K d λ .
Letting N → ∞ N \to\infty N → ∞ and using (H3) gives
G φ ( t n ) ⟶ 0. G_{\varphi}(t_n) \longrightarrow0. G φ ( t n ) ⟶ 0.
Since G φ ( t ) G_{\varphi}(t) G φ ( t ) already has a limit, necessarily
G φ ( t ) ⟶ 0. G_{\varphi}(t) \longrightarrow0. G φ ( t ) ⟶ 0.
Finally, (4.14) and Kallenberg [16 Theorem 4.2] yield
X t ⟹ 0. X_t \Longrightarrow0. X t ⟹ 0.
Survival of semilinear reaction–diffusion equations driven by Hunt processes
In this final section, we investigate the persistence of the total mass of nonnegative solutions to semilinear equations with absorption, similar to (1.2). The linear part of each equation generates a semigroup associated with one of the Hunt processes considered in the preceding examples. The underlying Hunt process governs the linear spatial evolution, whereas survival refers throughout to the solution of the semilinear equation.
Let D D D be a locally compact Hausdorff space, let β > 0 \beta> 0 β > 0 , and let { U ( t , s ) : 0 ≤ s ≤ t } \{\mathcal{U}(t,s): 0 \leq s \leq t\} { U ( t , s ) : 0 ≤ s ≤ t } be a strongly continuous positive contraction evolution family on C 0 ( D ) C_0(D) C 0 ( D ) . Assume that a positive Radon measure ϱ \varrho ϱ on B ( D ) \mathcal{B}(D) B ( D ) is invariant:
∫ D ( U ( t , s ) g ) ( x ) ϱ ( d x ) = ∫ D g ( x ) ϱ ( d x ) , 0 ≤ s ≤ t , (45) \int_D (\mathcal{U}(t,s)g)(x)\,\varrho(dx)=\int_D g(x)\,\varrho(dx), \qquad0 \leq s \leq t, \tag*{(45)} ∫ D ( U ( t , s ) g ) ( x ) ϱ ( d x ) = ∫ D g ( x ) ϱ ( d x ) , 0 ≤ s ≤ t , (45)
for every g ∈ C 0 ( D ) ∩ L 1 ( D , ϱ ) g \in C_0(D) \cap L^1(D,\varrho) g ∈ C 0 ( D ) ∩ L 1 ( D , ϱ ) .
For a fixed initial time s ≥ 0 s \geq0 s ≥ 0 , consider
v f ( t ) = U ( t , s ) f − ∫ s t U ( t , r ) [ v f ( r ) ] 1 + β d r , t ≥ s , (46) v_f(t)=\mathcal{U}(t,s)f-\int_s^t \mathcal{U}(t,r)[v_f(r)]^{1+\beta}\,dr, \qquad t \geq s, \tag*{(46)} v f ( t ) = U ( t , s ) f − ∫ s t U ( t , r ) [ v f ( r ) ] 1 + β d r , t ≥ s , (46)
with f ∈ C 0 + ( D ) ∩ L 1 ( D , ϱ ) f \in C_0^+(D) \cap L^1(D,\varrho) f ∈ C 0 + ( D ) ∩ L 1 ( D , ϱ ) . Powers are taken pointwise. We assume that this equation has a unique global nonnegative mild solution for every such f f f . The positive evolution family and the scalar absorption term imply that the solution map preserves order:
0 ≤ g ≤ f ⟹ v g ( t ) ≤ v f ( t ) , t ≥ s . (47) 0 \leq g \leq f \quad\Longrightarrow\quad v_g(t) \leq v_f(t), \qquad t \geq s. \tag*{(47)} 0 ≤ g ≤ f ⟹ v g ( t ) ≤ v f ( t ) , t ≥ s . (47)
Indeed, this is the case without delay of the comparison theorem of Martin and Smith [20 Corollary 5, p. 25] . To check its quasimonotonicity condition, let 0 ≤ u ≤ w 0 \le u \le w 0 ≤ u ≤ w in C 0 ( D ) C_0(D) C 0 ( D ) with ∥ w ∥ ∞ ≤ M \|w\|_\infty\le M ∥ w ∥ ∞ ≤ M . The mean value theorem gives
0 ≤ w 1 + β − u 1 + β ≤ ( 1 + β ) M β ( w − u ) . 0 \le w^{1+\beta}-u^{1+\beta} \le(1+\beta)M^\beta(w-u). 0 ≤ w 1 + β − u 1 + β ≤ ( 1 + β ) M β ( w − u ) .
Thus the reaction is locally Lipschitz, and w − u − h ( w 1 + β − u 1 + β ) w-u-h(w^{1+\beta}-u^{1+\beta}) w − u − h ( w 1 + β − u 1 + β ) belongs to the positive cone for all sufficiently small h > 0 h>0 h > 0 . This verifies Martin and Smith [20 condition (2.19)] for the reaction evaluated at the present state. The initial time is suppressed in the notation v f v_f v f .
We say that v f v_f v f survives if
inf t ≥ s ∥ v f ( t ) ∥ L 1 ( D , ϱ ) > 0. \inf_{t\ge s}\|v_f(t)\|_{L^1(D,\varrho)}>0. t ≥ s inf ∥ v f ( t ) ∥ L 1 ( D , ϱ ) > 0.
All applications below are understood within this framework. Differential equations involving generators are interpreted through their mild formulations.
The C 0 C_0 C 0 framework in the applications. Here a Feller semigroup is a positive contraction semigroup preserving C 0 ( D ) C_0(D) C 0 ( D ) and satisfying ∥ P t f − f ∥ ∞ → 0 \|P_t f-f\|_\infty\to0 ∥ P t f − f ∥ ∞ → 0 as t ↓ 0 t\downarrow0 t ↓ 0 . For a right-continuous Markov process, preservation of C 0 ( D ) C_0(D) C 0 ( D ) suffices for this strong continuity: path continuity at zero gives P t f ( x ) → f ( x ) P_t f(x)\to f(x) P t f ( x ) → f ( x ) , and the weak-to-strong continuity criterion applies; see Bonaccorsi and Priola [21 Proposition 6.16] . Thus preservation of C 0 ( D ) C_0(D) C 0 ( D ) is checked below in addition to the Hunt property.
The stable semigroups in Proposition 3.2 are convolution semigroups. If S S S is the corresponding Lévy process and f ∈ C 0 ( R d ) f\in C_0(\mathbb{R}^d) f ∈ C 0 ( R d ) , then P t f ( x ) = E f ( x + S t ) P_t f(x)=\mathbb{E}f(x+S_t) P t f ( x ) = E f ( x + S t ) is continuous and vanishes at infinity by dominated convergence. Stochastic continuity of S S S gives strong continuity by the preceding criterion. For the half-line Bessel products, both preservation of C 0 C_0 C 0 and strong continuity were proved in Proposition 3.4 .
For the uniformly elliptic realization in Proposition 3.3 , the jointly continuous kernel satisfies the Gaussian upper bound for every t > 0 t>0 t > 0 , by the Aronson estimates used in its proof. For f ∈ C c ( R d ) f\in C_c(\mathbb{R}^d) f ∈ C c ( R d ) , integration over its compact support gives continuity of P t f P_t f P t f , and
∣ P t f ( x ) ∣ ≤ A 1 t − d / 2 ∥ f ∥ L 1 ( R d ) exp ( − A 2 dist ( x , supp f ) 2 t ) ⟶ 0 ( ∣ x ∣ → ∞ ) . |P_t f(x)|\le A_1t^{-d/2}\|f\|_{L^1(\mathbb{R}^d)}\exp\left(-A_2\frac{\operatorname{dist}(x,\operatorname{supp}f)^2}{t}\right)\longrightarrow0\qquad(|x|\to\infty). ∣ P t f ( x ) ∣ ≤ A 1 t − d /2 ∥ f ∥ L 1 ( R d ) exp ( − A 2 t dist ( x , supp f ) 2 ) ⟶ 0 ( ∣ x ∣ → ∞ ) .
Contractivity and density extend preservation to C 0 ( R d ) C_0(\mathbb{R}^d) C 0 ( R d ) ; the criterion above gives strong continuity. The abstract Gaussian application below explicitly assumes the Feller property, since the large-time bounds alone are not used to infer it.
For the bilateral Bessel process of Proposition 3.5, (3.15)–(3.16) give continuity of the density in the starting point for each t > 0 t>0 t > 0 and y ≠ 0 y\ne0 y = 0 . At x = 0 x=0 x = 0 , the asymptotic I α ( u ) ∼ ( u / 2 ) α / Γ ( 1 + α ) I_{\alpha}(u)\sim(u/2)^{\alpha}/\Gamma(1+\alpha) I α ( u ) ∼ ( u /2 ) α /Γ ( 1 + α ) shows that the I ν I_{\nu} I ν contribution converges to (3.16), while the I − ν I_{-\nu} I − ν contribution is O ( ∣ x ∣ ρ − 1 ) O(|x|^{\rho-1}) O ( ∣ x ∣ ρ − 1 ) and vanishes because ρ > 1 \rho>1 ρ > 1 . Since the densities integrate to one, Scheffé’s lemma gives their L 1 L^{1} L 1 continuity in the starting point. Hence P ~ t f \widetilde{P}_{t}f P t f is continuous for f ∈ C 0 ( R ) f\in C_{0}(\mathbb{R}) f ∈ C 0 ( R ) . Put h ( r ) = max { ∣ f ( r ) ∣ , ∣ f ( − r ) ∣ } ∈ C 0 ( [ 0 , ∞ ) ) h(r)=\max\{|f(r)|,|f(-r)|\}\in C_{0}([0,\infty)) h ( r ) = max { ∣ f ( r ) ∣ , ∣ f ( − r ) ∣ } ∈ C 0 ([ 0 , ∞ )) . The absolute-value identity (3.18) gives
∣ P ~ t f ( x ) ∣ ≤ P t h ( ∣ x ∣ ) ⟶ 0 ( ∣ x ∣ → ∞ ) , |\widetilde{P}_{t}f(x)|\le P_{t}h(|x|)\longrightarrow0 \qquad(|x|\to\infty), ∣ P t f ( x ) ∣ ≤ P t h ( ∣ x ∣ ) ⟶ 0 ( ∣ x ∣ → ∞ ) ,
where P t P_{t} P t is the half-line Feller semigroup. The bilateral semigroup therefore preserves C 0 ( R ) C_{0}(\mathbb{R}) C 0 ( R ) and is strongly continuous by the same criterion. Finite independent products preserve these properties, first on products of C 0 C_{0} C 0 functions and then by density and contractivity. Nonsingular linear images preserve them by the isometric conjugation P t B f ( x ) = P t ( f ∘ B ) ( B − 1 x ) P^{B}_{t}f(x)=P_{t}(f\circ B)(B^{-1}x) P t B f ( x ) = P t ( f ∘ B ) ( B − 1 x ) . Finally, if k k k is continuous and nonnegative, then U ( t , s ) = P R s t k ( r ) d r \mathcal{U}(t,s)=P^{\mathsf{R}^{t}_{s}k(r)\,dr} U ( t , s ) = P R s t k ( r ) d r is a strongly continuous positive contraction evolution family on C 0 ( D ) C_{0}(D) C 0 ( D ) .
A sufficient condition for survival
Theorem 6.1. Let f ∈ C + 0 ( D ) ∩ L 1 ( D , ϱ ) f\in C_{+}^{0}(D)\cap L^{1}(D,\varrho) f ∈ C + 0 ( D ) ∩ L 1 ( D , ϱ ) satisfy ∥ f ∥ L 1 ( D , ϱ ) > 0 \|f\|_{L^{1}(D,\varrho)}>0 ∥ f ∥ L 1 ( D , ϱ ) > 0 . Define
J s ( f ) : = ∫ s ∞ ∥ U ( r , s ) f ∥ ∞ β d r . (48) J_{s}(f):=\int_{s}^{\infty}\|\mathcal{U}(r,s)f\|_{\infty}^{\beta}\,dr. \tag*{(48)} J s ( f ) := ∫ s ∞ ∥ U ( r , s ) f ∥ ∞ β d r . (48) For t ≥ s t\ge s t ≥ s , set A f ( t ) : = ∫ s t ∥ U ( r , s ) f ∥ ∞ β d r A_{f}(t):=\int_{s}^{t}\|\mathcal{U}(r,s)f\|_{\infty}^{\beta}\,dr A f ( t ) := ∫ s t ∥ U ( r , s ) f ∥ ∞ β d r . Then the pointwise bounds
U ( t , s ) f [ 1 + β A f ( t ) ] 1 / β ≤ v f ( t ) ≤ U ( t , s ) f \frac{\mathcal{U}(t,s)f}{[1+\beta A_{f}(t)]^{1/\beta}}\le v_{f}(t)\le\mathcal{U}(t,s)f [ 1 + β A f ( t ) ] 1/ β U ( t , s ) f ≤ v f ( t ) ≤ U ( t , s ) f hold. If J s ( f ) < ∞ J_{s}(f)<\infty J s ( f ) < ∞ , then
∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ ∥ f ∥ L 1 ( D , ϱ ) [ 1 + β J s ( f ) ] 1 / β > 0 , t ≥ s . (49) \|v_{f}(t)\|_{L^{1}(D,\varrho)}\ge\frac{\|f\|_{L^{1}(D,\varrho)}}{[1+\beta J_{s}(f)]^{1/\beta}}>0,\qquad t\ge s. \tag*{(49)} ∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ [ 1 + β J s ( f ) ] 1/ β ∥ f ∥ L 1 ( D , ϱ ) > 0 , t ≥ s . (49) In particular, v f v_{f} v f survives.
Proof. Write w ( t ) = U ( t , s ) f w(t)=\mathcal{U}(t,s)f w ( t ) = U ( t , s ) f and a ( t ) = ∥ w ( t ) ∥ ∞ β a(t)=\|w(t)\|_{\infty}^{\beta} a ( t ) = ∥ w ( t ) ∥ ∞ β . Positivity and (6.2) give 0 ≤ v f ( t ) ≤ w ( t ) 0\le v_{f}(t)\le w(t) 0 ≤ v f ( t ) ≤ w ( t ) . Define
c ( t ) = [ 1 + β A f ( t ) ] − 1 / β , z ( t ) = c ( t ) w ( t ) , h ( t ) = c ( t ) 1 + β [ a ( t ) w ( t ) − w ( t ) 1 + β ] ≥ 0. c(t)=[1+\beta A_{f}(t)]^{-1/\beta},\qquad z(t)=c(t)w(t),\qquad h(t)=c(t)^{1+\beta}[a(t)w(t)-w(t)^{1+\beta}]\ge0. c ( t ) = [ 1 + β A f ( t ) ] − 1/ β , z ( t ) = c ( t ) w ( t ) , h ( t ) = c ( t ) 1 + β [ a ( t ) w ( t ) − w ( t ) 1 + β ] ≥ 0.
Since c ( s ) = 1 c(s)=1 c ( s ) = 1 and c ′ = − a c 1 + β c'=-ac^{1+\beta} c ′ = − a c 1 + β , the evolution identity gives
z ( t ) = w ( t ) − ∫ s t U ( t , r ) [ a ( r ) c ( r ) 1 + β w ( r ) ] d r = w ( t ) − ∫ s t U ( t , r ) [ z ( r ) 1 + β + h ( r ) ] d r . \begin{aligned}
z(t)=w(t)-\int_{s}^{t}\mathcal{U}(t,r)\left[a(r)c(r)^{1+\beta}w(r)\right]\,dr \\
&=w(t)-\int_{s}^{t}\mathcal{U}(t,r)\left[z(r)^{1+\beta}+h(r)\right]\,dr.
\end{aligned} z ( t ) = w ( t ) − ∫ s t U ( t , r ) [ a ( r ) c ( r ) 1 + β w ( r ) ] d r = w ( t ) − ∫ s t U ( t , r ) [ z ( r ) 1 + β + h ( r ) ] d r .
This identity does not require differentiating the generator. To compare z z z and v f v_f v f , put M = ∥ f ∥ ∞ M = \|f\|_\infty M = ∥ f ∥ ∞ and choose λ ≥ ( 1 + β ) M β \lambda\ge(1+\beta)M^\beta λ ≥ ( 1 + β ) M β . On [ 0 , M ] [0,M] [ 0 , M ] , F λ ( u ) = λ u − u 1 + β F_\lambda(u)=\lambda u-u^{1+\beta} F λ ( u ) = λ u − u 1 + β is increasing and λ \lambda λ -Lipschitz, and 0 ≤ z , v f ≤ M 0\le z,v_f\le M 0 ≤ z , v f ≤ M . Writing both mild equations with U λ ( t , r ) = e − λ ( t − r ) U ( t , r ) U_\lambda(t,r)=e^{-\lambda(t-r)}U(t,r) U λ ( t , r ) = e − λ ( t − r ) U ( t , r ) and using h ≥ 0 h\ge0 h ≥ 0 yields
∥ ( z ( t ) − v f ( t ) ) + ∥ ∞ ≤ λ ∫ s t ∥ ( z ( r ) − v f ( r ) ) + ∥ ∞ d r . \|(z(t)-v_f(t))_+\|_\infty\le\lambda\int_s^t\|(z(r)-v_f(r))_+\|_\infty\,dr. ∥ ( z ( t ) − v f ( t ) ) + ∥ ∞ ≤ λ ∫ s t ∥ ( z ( r ) − v f ( r ) ) + ∥ ∞ d r .
Gronwall gives z ≤ v f z\le v_f z ≤ v f , proving the pointwise lower bound. Integration against ϱ \varrho ϱ and invariance give (49) .
Moreover, v f ≤ w v_f\le w v f ≤ w and contractivity imply ∫ D v f ( r , x ) 1 + β ϱ ( d x ) ≤ M β ∥ f ∥ L 1 ( D , ϱ ) \int_D v_f(r,x)^{1+\beta}\,\varrho(dx)\le M^\beta\|f\|_{L^1(D,\varrho)} ∫ D v f ( r , x ) 1 + β ϱ ( d x ) ≤ M β ∥ f ∥ L 1 ( D , ϱ ) . Thus invariance and Tonelli’s theorem applied to (46) give
∥ v f ( t ) ∥ L 1 ( D , ϱ ) = ∥ f ∥ L 1 ( D , ϱ ) − ∫ s t ∫ D v f ( r , x ) 1 + β ϱ ( d x ) d r . \|v_f(t)\|_{L^1(D,\varrho)}=\|f\|_{L^1(D,\varrho)}-\int_s^t\int_D v_f(r,x)^{1+\beta}\,\varrho(dx)\,dr. ∥ v f ( t ) ∥ L 1 ( D , ϱ ) = ∥ f ∥ L 1 ( D , ϱ ) − ∫ s t ∫ D v f ( r , x ) 1 + β ϱ ( d x ) d r .
The mass is therefore nonincreasing. Finally, log ( 1 + β J s ( f ) ) ≤ β J s ( f ) \log(1+\beta J_s(f))\le\beta J_s(f) log ( 1 + β J s ( f )) ≤ β J s ( f ) shows that (49) also implies the exponential bound ∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ e − J s ( f ) ∥ f ∥ L 1 ( D , ϱ ) \|v_f(t)\|_{L^1(D,\varrho)}\ge e^{-J_s(f)}\|f\|_{L^1(D,\varrho)} ∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ e − J s ( f ) ∥ f ∥ L 1 ( D , ϱ ) .
Reduction to decay estimates for the linear evolution
To apply Theorem 6.1, it suffices to find a compactly supported minorant of the initial datum for which the linear evolution satisfies the integrability condition. Order preservation then transfers survival to the original solution.
Proposition 6.2. Let f ∈ C + 0 ( D ) ∩ L 1 ( D , ϱ ) f\in C_+^0(D)\cap L^1(D,\varrho) f ∈ C + 0 ( D ) ∩ L 1 ( D , ϱ ) have positive integral. Suppose that there exist φ ∈ C c + ( D ) \varphi\in C_c^+(D) φ ∈ C c + ( D ) and ε 0 > 0 \varepsilon_0>0 ε 0 > 0 such that
∥ φ ∥ L 1 ( D , ϱ ) > 0 , ε 0 φ ≤ f , J s ( φ ) < ∞ . \|\varphi\|_{L^1(D,\varrho)}>0,\qquad\varepsilon_0\varphi\le f,\qquad J_s(\varphi)<\infty. ∥ φ ∥ L 1 ( D , ϱ ) > 0 , ε 0 φ ≤ f , J s ( φ ) < ∞. Then v f v_f v f survives. More precisely, for every 0 < ε ≤ ε 0 0<\varepsilon\le\varepsilon_0 0 < ε ≤ ε 0 ,
∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ ε ∥ φ ∥ L 1 ( D , ϱ ) [ 1 + β ε β J s ( φ ) ] 1 / β > 0 , t ≥ s . (50) \|v_f(t)\|_{L^1(D,\varrho)}\ge\frac{\varepsilon\|\varphi\|_{L^1(D,\varrho)}}{[1+\beta\varepsilon^\beta J_s(\varphi)]^{1/\beta}}>0,\qquad t\ge s. \tag*{(50)} ∥ v f ( t ) ∥ L 1 ( D , ϱ ) ≥ [ 1 + β ε β J s ( φ ) ] 1/ β ε ∥ φ ∥ L 1 ( D , ϱ ) > 0 , t ≥ s . (50) Proof. Linearity gives J s ( ε φ ) = ε β J s ( φ ) J_s(\varepsilon\varphi)=\varepsilon^\beta J_s(\varphi) J s ( εφ ) = ε β J s ( φ ) . Apply Theorem 6.1 to v ε φ v_{\varepsilon\varphi} v εφ and then use (47) .
The required minorant exists for every initial datum of positive integral. Indeed, choose η > 0 \eta>0 η > 0 such that ϱ ( { f > η } ) > 0 \varrho(\{f>\eta\})>0 ϱ ({ f > η }) > 0 . Inner regularity provides a compact set F ⊂ { f > η } F\subset\{f>\eta\} F ⊂ { f > η } with ϱ ( F ) > 0 \varrho(F)>0 ϱ ( F ) > 0 . Choose φ ∈ C c ( D ) \varphi\in C_c(D) φ ∈ C c ( D ) satisfying
0 ≤ φ ≤ 1 , φ = 1 on F , supp φ ⊂ { f > η } . 0\le\varphi\le1,\qquad\varphi=1\text{ on }F,\qquad\operatorname{supp}\varphi\subset\{f>\eta\}. 0 ≤ φ ≤ 1 , φ = 1 on F , supp φ ⊂ { f > η } .
Then η φ ≤ f \eta\varphi\le f η φ ≤ f and ∫ D φ d ϱ > 0 \int_D \varphi\,d\varrho> 0 ∫ D φ d ϱ > 0 .
To verify J s ( φ ) < ∞ J_s(\varphi) < \infty J s ( φ ) < ∞ , it suffices to find T > s T > s T > s and a measurable bound H s , φ H_{s,\varphi} H s , φ such that
∥ U ( t , s ) φ ∥ ∞ ≤ H s , φ ( t ) , t ≥ T , ∫ T ∞ H s , φ ( t ) β d t < ∞ . \lVert U(t,s)\varphi\rVert_\infty\le H_{s,\varphi}(t), \qquad t \ge T, \qquad\int_T^\infty H_{s,\varphi}(t)^\beta\,dt < \infty. ∥ U ( t , s ) φ ∥ ∞ ≤ H s , φ ( t ) , t ≥ T , ∫ T ∞ H s , φ ( t ) β d t < ∞.
Contractivity then yields
J s ( φ ) ≤ ( T − s ) ∥ φ ∥ ∞ β + ∫ T ∞ H s , φ ( t ) β d t < ∞ . J_s(\varphi) \le(T-s)\lVert\varphi\rVert_\infty^\beta+\int_T^\infty H_{s,\varphi}(t)^\beta\,dt < \infty. J s ( φ ) ≤ ( T − s ) ∥ φ ∥ ∞ β + ∫ T ∞ H s , φ ( t ) β d t < ∞.
For the autonomous equations below, this observation takes the following convenient form.
Corollary 6.3. Suppose that U ( t , s ) = P t − s U(t,s)=P_{t-s} U ( t , s ) = P t − s and that, for some q > 0 q>0 q > 0 , every φ ∈ C + ( D ) \varphi\in C_+^{}(D) φ ∈ C + ( D ) admits a constant C φ < ∞ C_\varphi<\infty C φ < ∞ such that
∥ P r φ ∥ ∞ ≤ C φ r − q , r ≥ 1. (51) \lVert P_r\varphi\rVert_\infty\le C_\varphi r^{-q}, \qquad r\ge1. \tag*{(51)} ∥ P r φ ∥ ∞ ≤ C φ r − q , r ≥ 1. (51) If β q > 1 \beta q>1 β q > 1 , then v f v_f v f survives for every f ∈ C + ( D ) ∩ L 1 ( D , ϱ ) f\in C_+^{}(D)\cap L^1(D,\varrho) f ∈ C + ( D ) ∩ L 1 ( D , ϱ ) of positive integral.
Proof. For every such test function,
J s ( φ ) = ∫ 0 ∞ ∥ P r φ ∥ ∞ β d r ≤ ∥ φ ∥ ∞ β + C φ β β q − 1 . J_s(\varphi)=\int_0^\infty\lVert P_r\varphi\rVert_\infty^\beta\,dr\le\lVert\varphi\rVert_\infty^\beta+\frac{C_\varphi^\beta}{\beta q-1}. J s ( φ ) = ∫ 0 ∞ ∥ P r φ ∥ ∞ β d r ≤ ∥ φ ∥ ∞ β + β q − 1 C φ β .
The conclusion follows from Proposition 6.2. □
Only estimates for compactly supported test functions are required. In particular, a global L 1 L^1 L 1 -to-L ∞ L^\infty L ∞ estimate is sufficient but unnecessary.
Equations driven by anisotropic stable motions
Let L A \mathcal{L}_A L A and P t P_t P t denote the generator and semigroup of the process in Proposition 3.2(b). Consider
∂ t v f = L A v f − v f 1 + β , v f ( s ) = f (52) \partial_t v_f=\mathcal{L}_A v_f-v_f^{1+\beta}, \qquad v_f(s)=f \tag*{(52)} ∂ t v f = L A v f − v f 1 + β , v f ( s ) = f (52)
on R d \mathbb{R}^d R d .
Proposition 6.4. If
β ∑ i = 1 d 1 α i > 1 , \beta\sum_{i=1}^d\frac{1}{\alpha_i}>1, β i = 1 ∑ d α i 1 > 1 , then every nonnegative mild solution of (6.8) with nonzero initial datum f ∈ C + ( R d ) ∩ L 1 ( R d ) f\in C_+^{}(\mathbb{R}^d)\cap L^1(\mathbb{R}^d) f ∈ C + ( R d ) ∩ L 1 ( R d ) satisfies
inf t ≥ s ∥ v f ( t ) ∥ L 1 ( R d ) > 0. \inf_{t\ge s}\lVert v_f(t)\rVert_{L^1(\mathbb{R}^d)}>0. t ≥ s inf ∥ v f ( t ) ∥ L 1 ( R d ) > 0.
Proof. Set q = ∑ i = 1 d α i − 1 q=\sum_{i=1}^{d}\alpha_i^{-1} q = ∑ i = 1 d α i − 1 . Independence, stable scaling, and the change of variables induced by A A A give
p r A ( z ) = r − q ∣ det A ∣ ∏ i = 1 d p 1 ( i ) ( r − 1 / α i ( A − 1 z ) i ) . p_r^A(z)=\frac{r^{-q}}{|\det A|}\prod_{i=1}^{d}p_1^{(i)}\left(r^{-1/\alpha_i}(A^{-1}z)_i\right). p r A ( z ) = ∣ det A ∣ r − q i = 1 ∏ d p 1 ( i ) ( r − 1/ α i ( A − 1 z ) i ) .
Consequently,
∥ p r A ∥ ∞ = C A r − q , C A : = 1 ∣ det A ∣ ∏ i = 1 d ∥ p 1 ( i ) ∥ ∞ . \|p_r^A\|_\infty=C_A r^{-q},\qquad C_A:=\frac{1}{|\det A|}\prod_{i=1}^{d}\|p_1^{(i)}\|_\infty. ∥ p r A ∥ ∞ = C A r − q , C A := ∣ det A ∣ 1 i = 1 ∏ d ∥ p 1 ( i ) ∥ ∞ .
and therefore
∥ P r A φ ∥ ∞ ≤ C A r − q ∥ φ ∥ L 1 ( R d ) . \|P_r^A\varphi\|_\infty\leq C_A r^{-q}\|\varphi\|_{L^1(\mathbb{R}^d)}. ∥ P r A φ ∥ ∞ ≤ C A r − q ∥ φ ∥ L 1 ( R d ) .
The convolution semigroup preserves Lebesgue measure. Corollary 6.3 applies with ϱ ( d x ) = d x \varrho(dx)=dx ϱ ( d x ) = d x .
The matrix A A A affects the constant in the decay estimate, but not its exponent. If α i = α \alpha_i=\alpha α i = α for every i i i , the sufficient condition becomes d > α / β d>\alpha/\beta d > α / β . For Brownian motion with positive definite covariance matrix, it becomes d > 2 / β d>2/\beta d > 2/ β .
Equations driven by symmetric diffusions with Gaussian bounds
Let L \mathcal{L} L and P t P_t P t denote the generator and semigroup of a diffusion in Proposition 3.3 , and assume that ( P t ) t ≥ 0 (P_t)_{t\geq0} ( P t ) t ≥ 0 is Feller on C 0 ( R d ) C_0(\mathbb{R}^d) C 0 ( R d ) , as verified above for the uniformly elliptic realization. Consider
∂ t v f = L v f − v f 1 + β , v f ( s ) = f (53) \partial_t v_f=\mathcal{L}v_f-v_f^{1+\beta},\qquad v_f(s)=f \tag*{(53)} ∂ t v f = L v f − v f 1 + β , v f ( s ) = f (53)
on R d \mathbb{R}^d R d .
Proposition 6.5. If β d / 2 > 1 \beta d/2>1 β d /2 > 1 , then every nonnegative mild solution of (53) with nonzero initial datum f ∈ C 0 ( R d ) ∩ L 1 ( R d ) f\in C_0(\mathbb{R}^d)\cap L^1(\mathbb{R}^d) f ∈ C 0 ( R d ) ∩ L 1 ( R d ) satisfies
inf t ≥ s ∥ v f ( t ) ∥ L 1 ( R d ) > 0. \inf_{t\geq s}\|v_f(t)\|_{L^1(\mathbb{R}^d)}>0. t ≥ s inf ∥ v f ( t ) ∥ L 1 ( R d ) > 0. In particular, this conclusion applies when L \mathcal{L} L is the conservative symmetric realization of
1 2 ∇ ⋅ ( A ∇ ) \frac{1}{2}\nabla\cdot(A\nabla) 2 1 ∇ ⋅ ( A ∇ ) under the coefficient assumptions of Proposition 3.3 .
Proof. Symmetry and conservativity imply invariance of Lebesgue measure. The Gaussian upper bound gives
∥ P r φ ∥ ∞ ≤ A 1 r − d / 2 ∥ φ ∥ L 1 ( R d ) , r ≥ 1. \|P_r\varphi\|_\infty\leq A_1r^{-d/2}\|\varphi\|_{L^1(\mathbb{R}^d)},\qquad r\geq1. ∥ P r φ ∥ ∞ ≤ A 1 r − d /2 ∥ φ ∥ L 1 ( R d ) , r ≥ 1.
Thus Corollary 6.3 applies with q = d / 2 q=d/2 q = d /2 .
Equations driven by generalized squared-Bessel product diffusions
Let D = [ 0 , ∞ ) d D=[0,\infty)^d D = [ 0 , ∞ ) d , and let L [ d ] L^{[d]} L [ d ] and P t [ d ] P_t^{[d]} P t [ d ] be the generator and semigroup in Proposition 3.4 . Consider
∂ t v f = L [ d ] v f − v f 1 + β , v f ( s ) = f . (54) \partial_t v_f=L^{[d]}v_f-v_f^{1+\beta},\qquad v_f(s)=f. \tag*{(54)} ∂ t v f = L [ d ] v f − v f 1 + β , v f ( s ) = f . (54)
Here L [ d ] L^{[d]} L [ d ] denotes the generator realization determined by the process, including its boundary behavior. On the test functions specified in Proposition 3.4 , it acts as
L [ d ] g = κ ∑ j = 1 d ∂ x j ( x j 2 − ρ ∂ x j g ) . L^{[d]}g=\kappa\sum_{j=1}^{d}\partial_{x_j}\left(x_j^{2-\rho}\partial_{x_j}g\right). L [ d ] g = κ j = 1 ∑ d ∂ x j ( x j 2 − ρ ∂ x j g ) .
Proposition 6.6. If β d / ρ > 1 \beta d/\rho>1 β d / ρ > 1 , then every nonnegative mild solution of (54) with nonzero initial datum f ∈ C 0 + ( D ) ∩ L 1 ( D ) f\in C_0^+(D)\cap L^1(D) f ∈ C 0 + ( D ) ∩ L 1 ( D ) satisfies
inf t ≥ s ∥ v f ( t ) ∥ L 1 ( D ) > 0. \inf_{t\ge s}\|v_f(t)\|_{L^1(D)}>0. t ≥ s inf ∥ v f ( t ) ∥ L 1 ( D ) > 0.
Proof. Lebesgue measure is invariant by Proposition 3.4 . Fix φ ∈ C c + ( D ) \varphi\in C_c^+(D) φ ∈ C c + ( D ) and choose R > 0 R>0 R > 0 such that supp φ ⊂ [ 0 , R ] d \operatorname{supp}\varphi\subset[0,R]^d supp φ ⊂ [ 0 , R ] d . The local kernel bound established there gives
M R : = sup u ≥ 0 , 0 ≤ z ≤ R p 1 ( ρ ) ( u , z ) < ∞ . M_R:=\sup_{u\ge0,\ 0\le z\le R}p_1^{(\rho)}(u,z)<\infty. M R := u ≥ 0 , 0 ≤ z ≤ R sup p 1 ( ρ ) ( u , z ) < ∞.
Scaling and the product representation imply, for r ≥ 1 r\ge1 r ≥ 1 ,
p r [ d ] ( x , y ) ≤ M R d r − d / ρ , x ∈ D , y ∈ [ 0 , R ] d . p_r^{[d]}(x,y)\le M_R^d r^{-d/\rho},\qquad x\in D,\quad y\in[0,R]^d. p r [ d ] ( x , y ) ≤ M R d r − d / ρ , x ∈ D , y ∈ [ 0 , R ] d .
Hence
∥ P r [ d ] φ ∥ ∞ ≤ M R d r − d / ρ ∥ φ ∥ L 1 ( D ) . \|P_r^{[d]}\varphi\|_\infty\le M_R^d r^{-d/\rho}\|\varphi\|_{L^1(D)}. ∥ P r [ d ] φ ∥ ∞ ≤ M R d r − d / ρ ∥ φ ∥ L 1 ( D ) .
Corollary 6.3 applies with q = d / ρ q=d/\rho q = d / ρ .
Equations driven by symmetric bilateral Bessel-type product diffusions
Let P ~ t [ d ] \widetilde P_t^{[d]} P t [ d ] be the semigroup of the product process in Proposition 3.5 , with ρ > 1 \rho>1 ρ > 1 , and let L ~ \widetilde L L be its generator. Consider
∂ t v f = L ~ v f − v f 1 + β , v f ( s ) = f (55) \partial_t v_f=\widetilde L v_f-v_f^{1+\beta},\qquad v_f(s)=f \tag*{(55)} ∂ t v f = L v f − v f 1 + β , v f ( s ) = f (55)
on R d \mathbb{R}^d R d . The generator realization is determined by the bilateral construction, including its behavior at the coordinate hyperplanes. Away from these hyperplanes, its differential expression on the corresponding test functions is
L ~ g = κ ∑ j = 1 d ∂ x j ( ∣ x j ∣ 2 − ρ ∂ x j g ) . \widetilde Lg=\kappa\sum_{j=1}^{d}\partial_{x_j}\left(|x_j|^{2-\rho}\partial_{x_j}g\right). L g = κ j = 1 ∑ d ∂ x j ( ∣ x j ∣ 2 − ρ ∂ x j g ) .
Proposition 6.7. If β d / ρ > 1 \beta d/\rho> 1 β d / ρ > 1 , then every nonnegative mild solution of (55) with nonzero initial datum f ∈ C 0 + ( R d ) ∩ L 1 ( R d ) f \in C_0^+(\mathbb{R}^d) \cap L^1(\mathbb{R}^d) f ∈ C 0 + ( R d ) ∩ L 1 ( R d ) satisfies
inf t ≥ s ∥ v f ( t ) ∥ L 1 ( R d ) > 0. \inf_{t\geq s}\|v_f(t)\|_{L^1(\mathbb{R}^d)} > 0. t ≥ s inf ∥ v f ( t ) ∥ L 1 ( R d ) > 0. The same conclusion holds for the semilinear equation whose linear generator corresponds to any nonsingular linear image of the bilateral product process.
Proof. Reversibility and conservativity imply invariance of Lebesgue measure. Fix φ ∈ C c + ( R d ) \varphi\in C_c^+(\mathbb{R}^d) φ ∈ C c + ( R d ) and take R > 0 R > 0 R > 0 such that supp φ ⊂ [ − R , R ] d \operatorname{supp}\varphi\subset[-R,R]^d supp φ ⊂ [ − R , R ] d . Let W + W^+ W + denote the corresponding half-line product. The absolute-value identity in Proposition 3.5 gives
P ~ r [ d ] φ ( x ) ≤ ∥ φ ∥ ∞ P x { W ~ r ∈ [ − R , R ] d } = ∥ φ ∥ ∞ P ∣ x ∣ { W r + ∈ [ 0 , R ] d } , \widetilde{P}^{[d]}_r\varphi(x) \leq\|\varphi\|_\infty\mathbb{P}_x\{\widetilde{W}_r \in[-R,R]^d\}
= \|\varphi\|_\infty\mathbb{P}_{|x|}\{W^+_r \in[0,R]^d\}, P r [ d ] φ ( x ) ≤ ∥ φ ∥ ∞ P x { W r ∈ [ − R , R ] d } = ∥ φ ∥ ∞ P ∣ x ∣ { W r + ∈ [ 0 , R ] d } ,
where absolute values are taken coordinatewise. The half-line product estimate therefore yields
∥ P ~ r [ d ] φ ∥ ∞ ≤ C R ∥ φ ∥ ∞ r − d / ρ , r ≥ 1. \|\widetilde{P}^{[d]}_r\varphi\|_\infty\leq C_R\|\varphi\|_\infty r^{-d/\rho}, \qquad r\geq1. ∥ P r [ d ] φ ∥ ∞ ≤ C R ∥ φ ∥ ∞ r − d / ρ , r ≥ 1.
Corollary 6.3 applies with q = d / ρ q=d/\rho q = d / ρ .
For a nonsingular linear image B W ~ B\widetilde{W} B W , the semigroup is
P ~ r B g ( x ) = P ~ r [ d ] ( g ∘ B ) ( B − 1 x ) . \widetilde{P}^{B}_r g(x) = \widetilde{P}^{[d]}_r(g\circ B)(B^{-1}x). P r B g ( x ) = P r [ d ] ( g ∘ B ) ( B − 1 x ) .
A change of variables shows that it preserves Lebesgue measure. Since g ∘ B g\circ B g ∘ B has compact support whenever g g g does, the same decay estimate holds with exponent d / ρ d/\rho d / ρ and a possibly different constant. The corollary applies again. □
Equations with time-dependent spatial coefficients
The four applications admit a common nonautonomous extension. Let P r P_r P r be one of the preceding semigroups, with generator L \mathcal{L} L , and consider
∂ t v f = k ( t ) L v f − v f 1 + β , v f ( s ) = f , (56) \partial_t v_f = k(t)\mathcal{L}v_f - v_f^{1+\beta}, \qquad v_f(s)=f, \tag*{(56)} ∂ t v f = k ( t ) L v f − v f 1 + β , v f ( s ) = f , (56)
where k : [ 0 , ∞ ) → [ 0 , ∞ ) k:[0,\infty)\to[0,\infty) k : [ 0 , ∞ ) → [ 0 , ∞ ) is continuous. Set
K ( t , s ) : = ∫ s t k ( r ) d r . K(t,s) := \int_s^t k(r)\,dr. K ( t , s ) := ∫ s t k ( r ) d r .
The linear evolution is U ( t , s ) = P K ( t , s ) \mathcal{U}(t,s)=P_{K(t,s)} U ( t , s ) = P K ( t , s ) and preserves the same invariant measure.
Proposition 6.8. Suppose that (6.7) holds with exponent q > 0 q > 0 q > 0 . If, for some T > s T > s T > s ,
K ( t , s ) ≥ 1 for t ≥ T , ∫ T ∞ K ( t , s ) − β q d t < ∞ , (57) K(t,s) \ge1 \quad\text{for } t \ge T,\qquad\int_T^\infty K(t,s)^{-\beta q}\,\mathrm{d}t < \infty, \tag*{(57)} K ( t , s ) ≥ 1 for t ≥ T , ∫ T ∞ K ( t , s ) − β q d t < ∞ , (57) then every nonnegative mild solution of (6.12) with initial datum of positive finite ϱ \varrho ϱ -integral survives.
Proof. For every φ ∈ C + c ( D ) \varphi\in C_+^c(D) φ ∈ C + c ( D ) , contraction and (6.7) give
J s ( φ ) = ∫ s ∞ ∥ P K ( t , s ) φ ∥ ∞ β d t ≤ ( T − s ) ∥ φ ∥ ∞ β + C φ β ∫ T ∞ K ( t , s ) − β q d t < ∞ . \begin{aligned}
J_s(\varphi)=\int_s^\infty\lVert P_{K(t,s)}\varphi\rVert_\infty^\beta\,\mathrm{d}t \\
&\le(T-s)\lVert\varphi\rVert_\infty^\beta+C_\varphi^\beta\int_T^\infty K(t,s)^{-\beta q}\,\mathrm{d}t<\infty.
\end{aligned} J s ( φ ) = ∫ s ∞ ∥ P K ( t , s ) φ ∥ ∞ β d t ≤ ( T − s ) ∥ φ ∥ ∞ β + C φ β ∫ T ∞ K ( t , s ) − β q d t < ∞.
Apply Proposition 6.2. □
The relevant decay exponents are
q = { ∑ i = 1 d α i − 1 , for anisotropic stable products , d / 2 , for symmetric diffusions with Gaussian bounds , d / ρ , for either Bessel-type product . q=\begin{cases}
\sum_{i=1}^{d}\alpha_i^{-1}, & \text{for anisotropic stable products},\\
d/2, & \text{for symmetric diffusions with Gaussian bounds},\\
d/\rho, & \text{for either Bessel-type product}.
\end{cases} q = ⎩ ⎨ ⎧ ∑ i = 1 d α i − 1 , d /2 , d / ρ , for anisotropic stable products , for symmetric diffusions with Gaussian bounds , for either Bessel-type product .
In particular, if
K ( t , s ) ≥ c ( t − s ) θ K(t,s)\ge c(t-s)^\theta K ( t , s ) ≥ c ( t − s ) θ
for all sufficiently large t t t , with c , θ > 0 c,\theta>0 c , θ > 0 , then (6.13) holds whenever θ β q > 1 \theta\beta q>1 θ β q > 1 . For k ≡ 1 k\equiv1 k ≡ 1 , this recovers the autonomous sufficient conditions β q > 1 \beta q>1 β q > 1 . These results concern survival of the semilinear solutions in the total-mass sense defined above. Failure of the sufficient conditions, including equality at the autonomous threshold, does not by itself imply extinction.
An application to the hyperbolic linear diffusion
The linear diffusion used in Section 2 to show that (H1) and (H2) do not imply (H3) gives an application beyond the joint hypotheses of the branching theorems. Its generator is
L h y p = 1 2 Δ + x 1 ∂ x 1 − x 2 ∂ x 2 . \mathcal{L}_{\mathrm{hyp}}=\frac{1}{2}\Delta+x_1\partial_{x_1}-x_2\partial_{x_2}. L hyp = 2 1 Δ + x 1 ∂ x 1 − x 2 ∂ x 2 .
Corollary 6.9. Let β > 0 \beta>0 β > 0 and let f ∈ C + c ( R 2 ) ∩ L 1 ( R 2 ) f\in C_+^c(\mathbb{R}^2)\cap L^1(\mathbb{R}^2) f ∈ C + c ( R 2 ) ∩ L 1 ( R 2 ) have positive integral. The nonnegative mild solution of
∂ t v f = L h y p v f − v f 1 + β , v f ( s ) = f , \partial_t v_f=\mathcal{L}_{\mathrm{hyp}}v_f-v_f^{1+\beta},\qquad v_f(s)=f, ∂ t v f = L hyp v f − v f 1 + β , v f ( s ) = f ,
satisfies
lim t → ∞ ∥ v f ( t ) ∥ L 1 ( R 2 ) ≥ ∥ f ∥ L 1 ( R 2 ) [ 1 + β J s ( f ) ] 1 / β > 0 , \lim_{t\to\infty}\lVert v_f(t)\rVert_{L^1(\mathbb{R}^2)} \ge\frac{\lVert f\rVert_{L^1(\mathbb{R}^2)}}{[1+\beta J_s(f)]^{1/\beta}} > 0, t → ∞ lim ∥ v f ( t ) ∥ L 1 ( R 2 ) ≥ [ 1 + β J s ( f ) ] 1/ β ∥ f ∥ L 1 ( R 2 ) > 0 ,
where J s ( f ) = ∫ 0 ∞ ∥ P r f ∥ ∞ β d r J_s(f)=\int_0^\infty\lVert P_r f\rVert_\infty^\beta\,dr J s ( f ) = ∫ 0 ∞ ∥ P r f ∥ ∞ β d r and P r P_r P r is the semigroup generated by L h y p L_{\mathrm{hyp}} L hyp .
Proof. As shown in § 2 , the linear stochastic equation d W t = A W t d t + d B t dW_t=AW_t\,dt+dB_t d W t = A W t d t + d B t , with A = diag ( 1 , − 1 ) A=\operatorname{diag}(1,-1) A = diag ( 1 , − 1 ) and B B B standard Brownian motion in R 2 \mathbb{R}^2 R 2 , has a conservative Feller semigroup preserving Lebesgue measure. Its transition density satisfies
sup x , y ∈ R 2 p r ( x , y ) = 1 2 π sinh r , r > 0. \sup_{x,y\in\mathbb{R}^2}p_r(x,y)=\frac{1}{2\pi\sinh r},\qquad r>0. x , y ∈ R 2 sup p r ( x , y ) = 2 π sinh r 1 , r > 0.
Contractivity on [ 0 , 1 ] [0,1] [ 0 , 1 ] and this estimate on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) give
J s ( f ) ≤ ∥ f ∥ ∞ β + ( ∥ f ∥ L 1 ( R 2 ) 2 π ) β ∫ 1 ∞ ( sinh r ) − β d r < ∞ . J_s(f)\le\lVert f\rVert_\infty^\beta+\left(\frac{\lVert f\rVert_{L^1(\mathbb{R}^2)}}{2\pi}\right)^\beta\int_1^\infty(\sinh r)^{-\beta}\,dr<\infty. J s ( f ) ≤ ∥ f ∥ ∞ β + ( 2 π ∥ f ∥ L 1 ( R 2 ) ) β ∫ 1 ∞ ( sinh r ) − β d r < ∞.
The tail integral is finite for every β > 0 \beta>0 β > 0 because ( sinh r ) − β = O ( e − β r ) (\sinh r)^{-\beta}=O(e^{-\beta r}) ( sinh r ) − β = O ( e − β r ) as r → ∞ r\to\infty r → ∞ . Theorem 6.1 gives the lower bound, and its mass balance gives the existence of the limit.
Thus this nonsymmetric motion preserves a positive limiting mass for every β > 0 \beta>0 β > 0 and every admissible nonzero datum, even though (H3) fails for every polynomial dispersal exponent.
Acknowledgments
The first and second authors were partially supported by SECIHTI Grant CBF2023-2024-3562. The third author acknowledges financial support from FAPESP under Grant 2025/03804-0.