Introduction
In a recent paper [3 ] , the author introduced a nonlinear kinetic Fokker–Planck equation which is Lorentz invariant and preserves particle number, energy, and momentum. The purpose of the present paper is to establish a global Cauchy theory for the spatially homogeneous version of this equation. A fundamental analytical difficulty is that both the drift vector field and the scalar diffusion coefficient depend nonlocally and nonlinearly on the solution. In particular, the diffusion coefficient is not a priori bounded away from zero. A key step in the proof of global existence is therefore to establish a uniform positive lower bound for this coefficient on every finite time interval.
The equation under consideration is
∂ t G = Φ G ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( W G G ) . (1) \partial_t G = \Phi_G \nabla_p \cdot\left(D \nabla_p G\right) + \nabla_p \cdot\left(W_G G\right). \tag*{(1)} ∂ t G = Φ G ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( W G G ) . (1)
Here, G = G ( t , p ) G = G(t,p) G = G ( t , p ) is the probability density at time t ≥ 0 t \ge0 t ≥ 0 for a system of particles with momentum p ∈ R 3 p \in\mathbb{R}^3 p ∈ R 3 . The diffusion matrix D ( p ) D(p) D ( p ) is given by
D ( p ) = I + p ⊗ p ⟨ p ⟩ , ⟨ p ⟩ = 1 + ∣ p ∣ 2 . (2) D(p) = \frac{I+p\otimes p}{\langle p\rangle}, \qquad\langle p\rangle= \sqrt{1+|p|^2}. \tag*{(2)} D ( p ) = ⟨ p ⟩ I + p ⊗ p , ⟨ p ⟩ = 1 + ∣ p ∣ 2 . (2)
The scalar diffusion coefficient Φ G ( t ) \Phi_G(t) Φ G ( t ) and the drift vector field W G ( t , p ) W_G(t,p) W G ( t , p ) are given by
Φ G ( t ) = μ 3 [ 1 − ∫ R 3 G ( t , p ) ⟨ p ⟩ β G ( t , p ) d p ] , W G ( t , p ) = μ ⟨ p ⟩ ( p − j G ( t ) β G ( t , p ) ) . (3) \Phi_G(t) = \frac{\mu}{3}\left[1-\int_{\mathbb{R}^3}\frac{G(t,p)}{\langle p\rangle\beta_G(t,p)}\,dp\right], \qquad W_G(t,p) = \frac{\mu}{\langle p\rangle}\left(p-\frac{j_G(t)}{\beta_G(t,p)}\right). \tag*{(3)} Φ G ( t ) = 3 μ [ 1 − ∫ R 3 ⟨ p ⟩ β G ( t , p ) G ( t , p ) d p ] , W G ( t , p ) = ⟨ p ⟩ μ ( p − β G ( t , p ) j G ( t ) ) . (3)
where μ > 0 \mu>0 μ > 0 is the drift constant and
β G ( t , p ) = ⟨ p ⟩ ( 1 − j G ( t ) ⋅ v ( p ) ) , j G ( t ) = ∫ R 3 v ( p ) G ( t , p ) d p , v ( p ) = p ⟨ p ⟩ . (4) \beta_G(t,p)=\langle p\rangle\left(1-j_G(t)\cdot v(p)\right), \qquad j_G(t)=\int_{\mathbb{R}^3}v(p)G(t,p)\,dp, \qquad v(p)=\frac{p}{\langle p\rangle}. \tag*{(4)} β G ( t , p ) = ⟨ p ⟩ ( 1 − j G ( t ) ⋅ v ( p ) ) , j G ( t ) = ∫ R 3 v ( p ) G ( t , p ) d p , v ( p ) = ⟨ p ⟩ p . (4)
Thus, both the diffusion coefficient and the drift vector field are integral nonlinear functionals of the solution G G G . We employ units in which the speed of light and the particle mass are equal to one.
For a probability density G G G , we define the energy E G E_G E G and the momentum P G P_G P G as
E G = ∫ R 3 ⟨ p ⟩ G ( p ) d p , P G = ∫ R 3 p G ( p ) d p . (5) E_G=\int_{\mathbb{R}^3}\langle p\rangle G(p)\,dp, \qquad P_G=\int_{\mathbb{R}^3}pG(p)\,dp. \tag*{(5)} E G = ∫ R 3 ⟨ p ⟩ G ( p ) d p , P G = ∫ R 3 pG ( p ) d p . (5)
and the energy tail by
T R ( G ) = ∫ ⟨ p ⟩ > R ⟨ p ⟩ G ( p ) d p , R > 1. (6) T_R(G)=\int_{\langle p\rangle>R}\langle p\rangle G(p)\,dp, \qquad R>1. \tag*{(6)} T R ( G ) = ∫ ⟨ p ⟩ > R ⟨ p ⟩ G ( p ) d p , R > 1. (6)
Our first main result gives global well-posedness of the Cauchy problem in the class of probability densities with finite energy and, at the same time, a quantitative estimate excluding finite time degeneration of the nonlinear diffusion coefficient.
Theorem 1 (Global well-posedness). Let G 0 ∈ L 1 ( R 3 ) G_0\in L^1(\mathbb{R}^3) G 0 ∈ L 1 ( R 3 ) , E > 1 E>1 E > 1 , P ∈ R 3 P\in\mathbb{R}^3 P ∈ R 3 satisfy
G 0 ≥ 0 a.e. , ∫ R 3 G 0 ( p ) d p = 1 , E G 0 = E , P G 0 = P . G_0\ge0\ \text{a.e.}, \qquad\int_{\mathbb{R}^3}G_0(p)\,dp=1, \qquad E_{G_0}=E, \qquad P_{G_0}=P. G 0 ≥ 0 a.e. , ∫ R 3 G 0 ( p ) d p = 1 , E G 0 = E , P G 0 = P . Then the Cauchy problem for (1) with initial data G 0 G_0 G 0 has a unique global weak solution
G ∈ C ( [ 0 , ∞ ) ; L 1 ( R 3 ) ) G\in C([0,\infty);L^1(\mathbb{R}^3)) G ∈ C ([ 0 , ∞ ) ; L 1 ( R 3 )) such that
G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , t ≥ 0. G(t,\cdot)\ge0\ \text{a.e.}, \qquad\int_{\mathbb{R}^3}G(t,p)\,dp=1, \qquad t\ge0. G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , t ≥ 0. The energy and momentum are conserved:
E G ( t ) = E , P G ( t ) = P , t ≥ 0. E_{G(t)}=E, \qquad P_{G(t)}=P, \qquad t\ge0. E G ( t ) = E , P G ( t ) = P , t ≥ 0.
Moreover,
Δ : = E 2 − ∣ P ∣ 2 − 1 > 0. \Delta:= E^2 - \lvert P\rvert^2 - 1 > 0. Δ := E 2 − ∣ P ∣ 2 − 1 > 0.
The energy tails are uniformly integrable on every finite time interval; that is, for all T > 0 T > 0 T > 0 ,
lim R → ∞ sup 0 ≤ t ≤ T T R ( G ( t ) ) = 0. (7) \lim_{R\to\infty}\sup_{0\leq t\leq T} T_R(G(t)) = 0. \tag*{(7)} R → ∞ lim 0 ≤ t ≤ T sup T R ( G ( t )) = 0. (7)
Finally, for every R > 1 R > 1 R > 1 and t ≥ 0 t \geq0 t ≥ 0 ,
Φ G ( t ) ≥ 9 μ 8 R 2 ( Δ 4 E − T R ( G ( t ) ) ) + 2 , (8) \Phi_G(t) \geq\frac{9\mu}{8R^2}\left(\frac{\Delta}{4E}-T_R(G(t))\right)_+^2, \tag*{(8)} Φ G ( t ) ≥ 8 R 2 9 μ ( 4 E Δ − T R ( G ( t )) ) + 2 , (8)
where ( x ) + = max { 0 , x } (x)_+ = \max\{0,x\} ( x ) + = max { 0 , x } . In particular, by (7) , for every T > 0 T > 0 T > 0 , there exists δ T > 0 \delta_T > 0 δ T > 0 such that
Φ G ( t ) ≥ δ T , 0 ≤ t ≤ T . \Phi_G(t) \geq\delta_T,\qquad0\leq t\leq T. Φ G ( t ) ≥ δ T , 0 ≤ t ≤ T .
The main steps in the proof of Theorem 1 are the following. First, local existence is obtained by solving a fixed point problem on the nonlinear functionals ( Φ G , j G ) (\Phi_G,j_G) ( Φ G , j G ) . The local solution can be continued globally provided that, on every finite time interval, the energy of the solution remains bounded and Φ G \Phi_G Φ G remains bounded away from zero. The mechanism preventing the breakdown of local solutions is a consequence of the relativistic structure of the equation. We first establish conservation of energy E E E and momentum P P P directly at the level of weak solutions and prove the uniform integrability property (7) of the energy tails. The strict relativistic inequality E 2 − ∣ P ∣ 2 > 1 E^2-\lvert P\rvert^2>1 E 2 − ∣ P ∣ 2 > 1 yields a quantitative lower bound for the velocity variance in terms of the energy tails. A coercivity estimate then gives (8) . Combined with (7) , this estimate shows that the diffusion coefficient remains uniformly positive on every finite time interval, thereby allowing the local solution to be continued globally.
The second main result shows that the global weak solutions immediately acquire full regularity and become strictly positive.
Theorem 2 (Regularization and strict positivity). Let G G G be the global weak solution given by Theorem 1. Then
G ∈ C ∞ ( ( 0 , ∞ ) × R 3 ) . G \in C^\infty((0,\infty)\times\mathbb{R}^3). G ∈ C ∞ (( 0 , ∞ ) × R 3 ) . Moreover,
G ( t , p ) > 0 for every t > 0 , p ∈ R 3 . G(t,p)>0\qquad\text{for every }t>0,\ p\in\mathbb{R}^3. G ( t , p ) > 0 for every t > 0 , p ∈ R 3 .
The proof of Theorem 2 involves a second difficulty: the diffusion matrix D ( p ) D(p) D ( p ) is not uniformly elliptic on the whole momentum space. It is, however, uniformly elliptic on every bounded momentum set. We first derive evolution identities for the nonlinear coefficients and use a class of admissible moment observables to prove that these coefficients are smooth for positive times. Once this regularity has been established, the equation satisfied by a given weak solution can be regarded as a linear, nonautonomous parabolic equation with smooth coefficients that is locally uniformly elliptic in momentum. Standard interior parabolic regularity then yields instantaneous smoothness of the solution. Finally, the parabolic strong maximum principle gives strict positivity for every positive time.
Our third result concerns the long time behavior of the nonlinear diffusion coefficient. For general initial data, the estimate (8) does not exclude the possibility that Φ G ( t ) \Phi_G(t) Φ G ( t ) approaches zero as t → ∞ t \to\infty t → ∞ . In the following theorem, we prove that for all even initial data with finite energy, the diffusion coefficient remains uniformly positive for all time. Moreover, if the conserved energy satisfies 1 < E < 2 1 < E < 2 1 < E < 2 , the energy tails are uniformly integrable on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) .
Theorem 3 (Uniform positivity of the diffusion coefficient for even initial data). Let G 0 ∈ L 1 ( R 3 ) G_0 \in L^1(\mathbb{R}^3) G 0 ∈ L 1 ( R 3 ) satisfy
G 0 ≥ 0 a . e . , G 0 ( p ) = G 0 ( − p ) a . e . , ∫ R 3 G 0 ( p ) d p = 1 , G_0 \ge0\ a.e., \qquad G_0(p) = G_0(-p)\ a.e., \qquad\int_{\mathbb{R}^3} G_0(p)\,dp = 1, G 0 ≥ 0 a . e . , G 0 ( p ) = G 0 ( − p ) a . e . , ∫ R 3 G 0 ( p ) d p = 1 , and assume that E G 0 = E < ∞ E_{G_0} = E < \infty E G 0 = E < ∞ . Let G G G be the global weak solution given by Theorem 1 . Then
G ( t , p ) = G ( t , − p ) , t > 0 , G(t,p) = G(t,-p), \qquad t > 0, G ( t , p ) = G ( t , − p ) , t > 0 , and there exists δ > 0 \delta> 0 δ > 0 , depending only on G 0 G_0 G 0 and μ \mu μ , such that
Φ G ( t ) ≥ δ , t ≥ 0. (9) \Phi_G(t) \ge\delta, \qquad t \ge0. \tag*{(9)} Φ G ( t ) ≥ δ , t ≥ 0. (9) If, in addition, E < 2 E < 2 E < 2 , then
lim sup R → ∞ t ≥ 0 TR ( G ( t ) ) = 0. (10) \limsup_{\substack{R \to\infty\\ t \ge0}} \operatorname{TR}(G(t)) = 0. \tag*{(10)} R → ∞ t ≥ 0 lim sup TR ( G ( t )) = 0. (10)
The key observation in the proof of Theorem 3 is that the number current j G j_G j G vanishes identically for even solutions, in which case the particle energy distribution satisfies a one-dimensional Fokker–Planck equation. When 1 < E < 2 1 < E < 2 1 < E < 2 , a PDE comparison argument with an explicit stationary supersolution yields the uniform integrability property (10). For arbitrary finite energy, we establish a comparison principle for the particle energy distributions of even solutions. This allows us to compare the diffusion coefficient with that of a radially nonincreasing solution whose conserved energy is less than two, thereby proving the uniform lower bound (9).
The validity of the lower bound
inf t ≥ 0 Φ G ( t ) > 0 \inf_{t \ge0} \Phi_G(t) > 0 t ≥ 0 inf Φ G ( t ) > 0
for general initial data remains an open problem, as is the convergence of solutions to equilibrium. The latter problem presents additional difficulties compared with the linear case studied in [1 , 2 ] , in part because the equilibrium distribution of (1) decays only polynomially. Indeed, as shown in [3 ] , the equilibrium distribution of (1) is not the standard exponentially decaying Maxwell–Jüttner distribution; rather, it belongs to the family
G e q ( p ) = Z ( λ , y ) ( ⟨ p ⟩ − y ⋅ p ) 2 λ , λ > 2 , ∣ y ∣ < 1 , G_{\mathrm{eq}}(p) = \frac{Z(\lambda,y)}{\left(\langle p\rangle- y \cdot p\right)^{2\lambda}}, \qquad\lambda> 2,\qquad|y| < 1, G eq ( p ) = ( ⟨ p ⟩ − y ⋅ p ) 2 λ Z ( λ , y ) , λ > 2 , ∣ y ∣ < 1 ,
where
Z ( λ , y ) = ( 1 − ∣ y ∣ 2 ) λ + 1 / 2 Γ ( λ ) π 3 / 2 Γ ( λ − 3 / 2 ) Z(\lambda,y)=\frac{(1-|y|^2)^{\lambda+1/2}\Gamma(\lambda)}{\pi^{3/2}\Gamma(\lambda-3/2)} Z ( λ , y ) = π 3/2 Γ ( λ − 3/2 ) ( 1 − ∣ y ∣ 2 ) λ + 1/2 Γ ( λ )
is the normalization constant. For these distributions,
j ˙ G e q = y , Φ G e q = μ 2 λ . \dot{j}_{G_{\mathrm{eq}}}=y,\qquad\Phi_{G_{\mathrm{eq}}}=\frac{\mu}{2\lambda}. j ˙ G eq = y , Φ G eq = 2 λ μ .
The parameters λ \lambda λ and y y y are related to the conserved energy and momentum of the solution. In particular, when y = 0 y=0 y = 0 , the equilibrium is isotropic and takes the form
G e q ( p ) = Γ ( λ ) π 3 / 2 Γ ( λ − 3 / 2 ) ⟨ p ⟩ − 2 λ . G_{\mathrm{eq}}(p)=\frac{\Gamma(\lambda)}{\pi^{3/2}\Gamma(\lambda-3/2)}\langle p\rangle^{-2\lambda}. G eq ( p ) = π 3/2 Γ ( λ − 3/2 ) Γ ( λ ) ⟨ p ⟩ − 2 λ .
The full asymptotic behavior of solutions to (1) will be addressed in a subsequent publication.
Preliminaries
We begin by noticing that, for ∣ p ∣ ≤ R |p|\le R ∣ p ∣ ≤ R and ξ ∈ R 3 \xi\in\mathbb{R}^3 ξ ∈ R 3 , the diffusion matrix (2) satisfies
1 1 + R 2 ∣ ξ ∣ 2 ≤ ξ T D ( p ) ξ ≤ 1 + R 2 ∣ ξ ∣ 2 . (11) \frac{1}{\sqrt{1+R^2}}|\xi|^2\le\xi^T D(p)\xi\le\sqrt{1+R^2}|\xi|^2. \tag*{(11)} 1 + R 2 1 ∣ ξ ∣ 2 ≤ ξ T D ( p ) ξ ≤ 1 + R 2 ∣ ξ ∣ 2 . (11)
Thus, the operator G → ∇ p ⋅ ( D ∇ p G ) G\to\nabla_p\cdot(D\nabla_pG) G → ∇ p ⋅ ( D ∇ p G ) in the right hand side of (1) is uniformly elliptic on every bounded momentum set.
Definition 4 (Weak probability density solution). Let T > 0 T>0 T > 0 . A function
G ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) G\in C([0,T];L^1(\mathbb{R}^3)) G ∈ C ([ 0 , T ] ; L 1 ( R 3 )) is a weak probability density solution of (1) on [ 0 , T ] [0,T] [ 0 , T ] with initial data G 0 G_0 G 0 if
G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , 0 ≤ t ≤ T , G(t,\cdot)\ge0\ \text{a.e.},\qquad\int_{\mathbb{R}^3}G(t,p)\,dp=1,\qquad0\le t\le T, G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , 0 ≤ t ≤ T , and, for every ψ ∈ C c ∞ ( [ 0 , T ) × R 3 ) \psi\in C_c^\infty([0,T)\times\mathbb{R}^3) ψ ∈ C c ∞ ([ 0 , T ) × R 3 ) ,
∫ 0 T ∫ R 3 G [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p d t + ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p = 0. (12) \int_0^T\int_{\mathbb{R}^3}G\left[\partial_t\psi+\Phi_G\nabla_p\cdot(D\nabla_p\psi)-W_G\cdot\nabla_p\psi\right]\,dp\,dt+\int_{\mathbb{R}^3}G_0(p)\psi(0,p)\,dp=0. \tag*{(12)} ∫ 0 T ∫ R 3 G [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p d t + ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p = 0. (12)
The following lemma will be used to extend local weak solutions to larger time intervals.
Lemma 5. Let G ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) G \in C([0,T];L^1(\mathbb{R}^3)) G ∈ C ([ 0 , T ] ; L 1 ( R 3 )) be a weak solution of (1) in the sense of Definition 4. Then, for every 0 ≤ t 1 < t 2 ≤ T 0 \le t_1 < t_2 \le T 0 ≤ t 1 < t 2 ≤ T and every ψ ∈ C 1 , 2 ( [ 0 , T ] × R 3 ) \psi\in C^{1,2}([0,T] \times\mathbb{R}^3) ψ ∈ C 1 , 2 ([ 0 , T ] × R 3 ) with compact support in the momentum variable uniformly on [ 0 , T ] [0,T] [ 0 , T ] , there holds
∫ R 3 G ( t 2 , p ) ψ ( t 2 , p ) d p − ∫ R 3 G ( t 1 , p ) ψ ( t 1 , p ) d p = ∫ t 1 t 2 ∫ R 3 G ( t , p ) [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p d t . (13) \int_{\mathbb{R}^3} G(t_2,p)\psi(t_2,p)\,dp-\int_{\mathbb{R}^3}G(t_1,p)\psi(t_1,p)\,dp
=
\int_{t_1}^{t_2}\int_{\mathbb{R}^3}G(t,p)\left[\partial_t\psi+\Phi_G\nabla_p\cdot(D\nabla_p\psi)-W_G\cdot\nabla_p\psi\right]\,dp\,dt. \tag*{(13)} ∫ R 3 G ( t 2 , p ) ψ ( t 2 , p ) d p − ∫ R 3 G ( t 1 , p ) ψ ( t 1 , p ) d p = ∫ t 1 t 2 ∫ R 3 G ( t , p ) [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p d t . (13) In particular, the restriction of a weak solution to any subinterval [ t 0 , T ] [t_0,T] [ t 0 , T ] is a weak solution with initial data G ( t 0 , p ) G(t_0,p) G ( t 0 , p ) almost everywhere. Moreover, if G 1 G_1 G 1 and G 2 G_2 G 2 are weak solutions on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] and [ t 0 , T ] [t_0,T] [ t 0 , T ] , respectively, and G 1 ( t 0 , p ) = G 2 ( t 0 , p ) G_1(t_0,p)=G_2(t_0,p) G 1 ( t 0 , p ) = G 2 ( t 0 , p ) almost everywhere, then their concatenation is a weak solution on [ 0 , T ] [0,T] [ 0 , T ] .
Proof. Let ψ \psi ψ be as in the statement and define
F ( t ) = ∫ R 3 G ( t , p ) ψ ( t , p ) d p F(t)=\int_{\mathbb{R}^3}G(t,p)\psi(t,p)\,dp F ( t ) = ∫ R 3 G ( t , p ) ψ ( t , p ) d p
and
B ( t ) = ∫ R 3 G ( t , p ) [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p . B(t)=\int_{\mathbb{R}^3}G(t,p)\left[\partial_t\psi+\Phi_G\nabla_p\cdot(D\nabla_p\psi)-W_G\cdot\nabla_p\psi\right]\,dp. B ( t ) = ∫ R 3 G ( t , p ) [ ∂ t ψ + Φ G ∇ p ⋅ ( D ∇ p ψ ) − W G ⋅ ∇ p ψ ] d p .
Since G ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) G\in C([0,T];L^1(\mathbb{R}^3)) G ∈ C ([ 0 , T ] ; L 1 ( R 3 )) and ∣ v ( p ) ∣ < 1 |v(p)|<1 ∣ v ( p ) ∣ < 1 , the function j G j_G j G is continuous and satisfies
∣ j G ( t ) ∣ ≤ ∫ R 3 ∣ v ( p ) ∣ G ( t , p ) d p < 1 , 0 ≤ t ≤ T . |j_G(t)|\le\int_{\mathbb{R}^3}|v(p)|G(t,p)\,dp<1,\qquad0\le t\le T. ∣ j G ( t ) ∣ ≤ ∫ R 3 ∣ v ( p ) ∣ G ( t , p ) d p < 1 , 0 ≤ t ≤ T .
By compactness, there exists γ T < 1 \gamma_T<1 γ T < 1 such that
∣ j G ( t ) ∣ ≤ γ T , 0 ≤ t ≤ T . |j_G(t)|\le\gamma_T,\qquad0\le t\le T. ∣ j G ( t ) ∣ ≤ γ T , 0 ≤ t ≤ T .
It follows directly from the definitions of Φ G \Phi_G Φ G and W G W_G W G in (3) that these coefficients are continuous in time, uniformly on compact momentum sets. Consequently, both F F F and B B B are continuous on [ 0 , T ] [0,T] [ 0 , T ] .
Let η ∈ C c ∞ ( [ 0 , T ) ) \eta\in C^\infty_c([0,T)) η ∈ C c ∞ ([ 0 , T )) . Using η ( t ) ψ ( t , p ) \eta(t)\psi(t,p) η ( t ) ψ ( t , p ) in (12), after approximation by smooth test functions, gives
∫ 0 T η ′ ( t ) F ( t ) d t + ∫ 0 T η ( t ) B ( t ) d t + η ( 0 ) I 0 = 0 , \int_0^T\eta'(t)F(t)\,dt+\int_0^T\eta(t)B(t)\,dt+\eta(0)I_0=0, ∫ 0 T η ′ ( t ) F ( t ) d t + ∫ 0 T η ( t ) B ( t ) d t + η ( 0 ) I 0 = 0 ,
where
I 0 = ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p . I_0=\int_{\mathbb{R}^3}G_0(p)\psi(0,p)\,dp. I 0 = ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p .
First take η ∈ C c ∞ ( ( 0 , T ) ) \eta\in C^\infty_c((0,T)) η ∈ C c ∞ (( 0 , T )) . Then F ′ = B F'=B F ′ = B in the sense of distributions on ( 0 , T ) (0,T) ( 0 , T ) . Since F F F and B B B are continuous on [ 0 , T ] [0,T] [ 0 , T ] , it follows that
F ( t ) = F ( 0 ) + ∫ 0 t B ( s ) d s , 0 ≤ t ≤ T . F(t)=F(0)+\int_0^tB(s)\,ds,\qquad0\le t\le T. F ( t ) = F ( 0 ) + ∫ 0 t B ( s ) d s , 0 ≤ t ≤ T .
Substituting this identity into the preceding weak identity and integrating by parts gives
η ( 0 ) ( I 0 − F ( 0 ) ) = 0. \eta(0)(I_0-F(0))=0. η ( 0 ) ( I 0 − F ( 0 )) = 0.
Choosing η \eta η with η ( 0 ) = 1 \eta(0)=1 η ( 0 ) = 1 , we obtain
∫ R 3 G ( 0 , p ) ψ ( 0 , p ) d p = ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p . \int_{\mathbb{R}^3}G(0,p)\psi(0,p)\,\mathrm{d}p=\int_{\mathbb{R}^3}G_0(p)\psi(0,p)\,\mathrm{d}p. ∫ R 3 G ( 0 , p ) ψ ( 0 , p ) d p = ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p .
In particular, taking ψ ( t , p ) = φ ( p ) \psi(t,p)=\varphi(p) ψ ( t , p ) = φ ( p ) for arbitrary φ ∈ C c ∞ ( R 3 ) \varphi\in C^\infty_c(\mathbb{R}^3) φ ∈ C c ∞ ( R 3 ) shows that
G ( 0 , p ) = G 0 ( p ) for almost every p ∈ R 3 . G(0,p)=G_0(p)\qquad\text{for almost every }p\in\mathbb{R}^3. G ( 0 , p ) = G 0 ( p ) for almost every p ∈ R 3 .
Finally, subtracting the identity
F ( t ) = F ( 0 ) + ∫ 0 t B ( s ) d s F(t)=F(0)+\int_0^t B(s)\,\mathrm{d}s F ( t ) = F ( 0 ) + ∫ 0 t B ( s ) d s
at t 1 t_1 t 1 and t 2 t_2 t 2 gives (13) . Taking t 1 = t 0 t_1=t_0 t 1 = t 0 shows that the restriction of G G G to [ t 0 , T ] [t_0,T] [ t 0 , T ] satisfies the integrated weak formulation with initial data G ( t 0 , p ) G(t_0,p) G ( t 0 , p ) . In particular, it satisfies the corresponding weak formulation after translating the initial time to t 0 t_0 t 0 .
Finally, suppose that G 1 G_1 G 1 and G 2 G_2 G 2 are weak solutions on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] and [ t 0 , T ] [t_0,T] [ t 0 , T ] with G 1 ( t 0 ) = G 2 ( t 0 ) G_1(t_0)=G_2(t_0) G 1 ( t 0 ) = G 2 ( t 0 ) . Define
G ( t , p ) = { G 1 ( t , p ) , 0 ≤ t ≤ t 0 , G 2 ( t , p ) , t 0 ≤ t ≤ T . G(t,p)=
\begin{cases}
G_1(t,p), & 0\leq t\leq t_0,\\
G_2(t,p), & t_0\leq t\leq T.
\end{cases} G ( t , p ) = { G 1 ( t , p ) , G 2 ( t , p ) , 0 ≤ t ≤ t 0 , t 0 ≤ t ≤ T .
Then G ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) G\in C([0,T];L^1(\mathbb{R}^3)) G ∈ C ([ 0 , T ] ; L 1 ( R 3 )) . Applying (13) on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] and [ t 0 , T ] [t_0,T] [ t 0 , T ] and adding the resulting identities, the boundary terms at t 0 t_0 t 0 cancel. Hence, G G G satisfies (13) on [ 0 , T ] [0,T] [ 0 , T ] . Taking t 1 = 0 t_1=0 t 1 = 0 and integrating by parts in time recovers the weak formulation (12) in Definition 4. Therefore, G G G is a weak solution on [ 0 , T ] [0,T] [ 0 , T ] .
For a probability density G G G , we define
j G = ∫ R 3 v ( p ) G ( p ) d p , β G ( p ) = ⟨ p ⟩ ( 1 − j G ⋅ v ( p ) ) , j_G=\int_{\mathbb{R}^3}v(p)G(p)\,\mathrm{d}p,\qquad\beta_G(p)=\langle p\rangle\bigl(1-j_G\cdot v(p)\bigr), j G = ∫ R 3 v ( p ) G ( p ) d p , β G ( p ) = ⟨ p ⟩ ( 1 − j G ⋅ v ( p ) ) ,
and
Φ G = μ 3 [ 1 − ∫ R 3 G ( p ) ⟨ p ⟩ β G ( p ) d p ] . \Phi_G=\frac{\mu}{3}\left[1-\int_{\mathbb{R}^3}\frac{G(p)}{\langle p\rangle\beta_G(p)}\,\mathrm{d}p\right]. Φ G = 3 μ [ 1 − ∫ R 3 ⟨ p ⟩ β G ( p ) G ( p ) d p ] .
Recall also the definition (5) of the energy-momentum ( E G , P G ) (E_G,P_G) ( E G , P G ) . If G G G has finite energy, then
since ⟨ p ⟩ ≥ 1 \langle p\rangle\geq1 ⟨ p ⟩ ≥ 1 , with equality only at p = 0 p=0 p = 0 , and G G G is an L 1 L^1 L 1 probability density. For E > 1 E>1 E > 1 , we set
γ E = 1 − E − 2 . \gamma_E=\sqrt{1-E^{-2}}. γ E = 1 − E − 2 .
Lemma 6. Let G ∈ L 1 ( R 3 ) G \in L^1(\mathbb{R}^3) G ∈ L 1 ( R 3 ) be a probability density . Then
∣ j G ∣ < 1 , 1 2 ⟨ p ⟩ ≤ β G ( p ) ≤ 2 ⟨ p ⟩ . (14) \lvert j_G\rvert< 1,\qquad\frac{1}{2\langle p\rangle} \le\beta_G(p) \le2\langle p\rangle. \tag*{(14)} ∣ j G ∣ < 1 , 2 ⟨ p ⟩ 1 ≤ β G ( p ) ≤ 2 ⟨ p ⟩ . (14) If, in addition, G G G has finite energy, then
∣ j G ∣ ≤ γ E G . (15) \lvert j_G\rvert\le\gamma E_G. \tag*{(15)} ∣ j G ∣ ≤ γ E G . (15)
Proof. Since ∣ v ( p ) ∣ < 1 \lvert v(p)\rvert< 1 ∣ v ( p )∣ < 1 for all p ∈ R 3 p \in\mathbb{R}^3 p ∈ R 3 and G G G is a probability density, we have
∣ j G ∣ ≤ ∫ R 3 ∣ v ( p ) ∣ G ( p ) d p < ∫ R 3 G ( p ) d p = 1. \lvert j_G\rvert\le\int_{\mathbb{R}^3} \lvert v(p)\rvert G(p)\,dp < \int_{\mathbb{R}^3} G(p)\,dp = 1. ∣ j G ∣ ≤ ∫ R 3 ∣ v ( p )∣ G ( p ) d p < ∫ R 3 G ( p ) d p = 1.
The upper bound on β G ( p ) \beta_G(p) β G ( p ) in (14) then follows immediately. For the lower bound on β G ( p ) \beta_G(p) β G ( p ) we write
1 ⟨ p ⟩ β G ( p ) = 1 − ∣ v ( p ) ∣ 2 1 − j G ⋅ v ( p ) ≤ 1 − ∣ v ( p ) ∣ 2 1 − ∣ v ( p ) ∣ = 1 + ∣ v ( p ) ∣ ≤ 2. \frac{1}{\langle p\rangle\beta_G(p)}=\frac{1-\lvert v(p)\rvert^2}{1-j_G\cdot v(p)}\le\frac{1-\lvert v(p)\rvert^2}{1-\lvert v(p)\rvert}=1+\lvert v(p)\rvert\le2. ⟨ p ⟩ β G ( p ) 1 = 1 − j G ⋅ v ( p ) 1 − ∣ v ( p ) ∣ 2 ≤ 1 − ∣ v ( p )∣ 1 − ∣ v ( p ) ∣ 2 = 1 + ∣ v ( p )∣ ≤ 2.
Now assume that G G G has finite energy. By Jensen’s inequality,
∣ j G ∣ 2 = ∣ ∫ R 3 v ( p ) G ( p ) d p ∣ 2 ≤ ∫ R 3 ∣ v ( p ) ∣ 2 G ( p ) d p . \lvert j_G\rvert^2=\left\lvert\int_{\mathbb{R}^3}v(p)G(p)\,dp\right\rvert^2\le\int_{\mathbb{R}^3}\lvert v(p)\rvert^2G(p)\,dp. ∣ j G ∣ 2 = ∫ R 3 v ( p ) G ( p ) d p 2 ≤ ∫ R 3 ∣ v ( p ) ∣ 2 G ( p ) d p .
Since 1 − ∣ v ( p ) ∣ 2 = ⟨ p ⟩ − 2 1-\lvert v(p)\rvert^2=\langle p\rangle^{-2} 1 − ∣ v ( p ) ∣ 2 = ⟨ p ⟩ − 2 , we obtain
1 − ∣ j G ∣ 2 ≥ ∫ R 3 ⟨ p ⟩ − 2 G ( p ) d p . (16) 1-\lvert j_G\rvert^2\ge\int_{\mathbb{R}^3}\langle p\rangle^{-2}G(p)\,dp. \tag*{(16)} 1 − ∣ j G ∣ 2 ≥ ∫ R 3 ⟨ p ⟩ − 2 G ( p ) d p . (16)
Applying Jensen’s inequality to the convex function y ↦ y − 2 y\mapsto y^{-2} y ↦ y − 2 gives
∫ R 3 ⟨ p ⟩ − 2 G ( p ) d p ≥ ( ∫ R 3 ⟨ p ⟩ G ( p ) d p ) − 2 = 1 E G 2 . \int_{\mathbb{R}^3}\langle p\rangle^{-2}G(p)\,dp\ge\left(\int_{\mathbb{R}^3}\langle p\rangle G(p)\,dp\right)^{-2}=\frac{1}{E_G^2}. ∫ R 3 ⟨ p ⟩ − 2 G ( p ) d p ≥ ( ∫ R 3 ⟨ p ⟩ G ( p ) d p ) − 2 = E G 2 1 .
Consequently, (16) gives
1 − ∣ j G ∣ 2 ≥ 1 E G 2 , 1-\lvert j_G\rvert^2\ge\frac{1}{E_G^2}, 1 − ∣ j G ∣ 2 ≥ E G 2 1 ,
which proves (15) . □
Bounds on the diffusion coefficient Φ G \Phi_G Φ G are particularly important, so we state them in a separate lemma.
Lemma 7. Let G ∈ L 1 ( R 3 ) G \in L^1(\mathbb{R}^3) G ∈ L 1 ( R 3 ) be a probability density . Then
0 < 9 μ 32 ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p ≤ Φ G ≤ μ 3 . (17) 0<\frac{9\mu}{32}\int_{\mathbb{R}^3}\lvert v(p)-j_G\rvert^2G(p)\,dp\le\Phi_G\le\frac{\mu}{3}. \tag*{(17)} 0 < 32 9 μ ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p ≤ Φ G ≤ 3 μ . (17) Moreover, if G G G has finite energy, then
Φ G ≤ μ 3 γ 2 E G 2 . (18) \Phi_G\le\frac{\mu}{3}\gamma^2E_G^2. \tag*{(18)} Φ G ≤ 3 μ γ 2 E G 2 . (18)
Proof. The upper bound in (17) is immediate, because β G ( p ) > 0 \beta_G(p)>0 β G ( p ) > 0 . For the lower bound, set j = j G j=j_G j = j G . As in the definition of Φ G \Phi_G Φ G ,
3 Φ G μ = ∫ R 3 ∣ v ∣ 2 − j ⋅ v 1 − j ⋅ v G d p . \frac{3\Phi_G}{\mu}=\int_{\mathbb{R}^3}\frac{|v|^2-j\cdot v}{1-j\cdot v}G\,dp. μ 3 Φ G = ∫ R 3 1 − j ⋅ v ∣ v ∣ 2 − j ⋅ v G d p .
Using
∣ v ∣ 2 − j ⋅ v = ∣ v − j ∣ 2 + j ⋅ ( v − j ) |v|^2-j\cdot v=|v-j|^2+j\cdot(v-j) ∣ v ∣ 2 − j ⋅ v = ∣ v − j ∣ 2 + j ⋅ ( v − j )
and
1 − j ⋅ v = 1 − ∣ j ∣ 2 − j ⋅ ( v − j ) , 1-j\cdot v=1-|j|^2-j\cdot(v-j), 1 − j ⋅ v = 1 − ∣ j ∣ 2 − j ⋅ ( v − j ) ,
we obtain
3 Φ G μ = ∫ R 3 ∣ v − j ∣ 2 + [ j ⋅ ( v − j ) ] 2 1 − ∣ j ∣ 2 1 − j ⋅ v G d p . (19) \frac{3\Phi_G}{\mu}=\int_{\mathbb{R}^3}\frac{|v-j|^2+\frac{[j\cdot(v-j)]^2}{1-|j|^2}}{1-j\cdot v}G\,dp. \tag*{(19)} μ 3 Φ G = ∫ R 3 1 − j ⋅ v ∣ v − j ∣ 2 + 1 − ∣ j ∣ 2 [ j ⋅ ( v − j ) ] 2 G d p . (19)
We claim that
∣ v − j ∣ 2 + [ j ⋅ ( v − j ) ] 2 1 − ∣ j ∣ 2 1 − j ⋅ v ≥ ∣ v − j ∣ 2 ( 1 − ∣ j ∣ ) ( 1 + ∣ j ∣ ) 2 . (20) \frac{|v-j|^2+\frac{[j\cdot(v-j)]^2}{1-|j|^2}}{1-j\cdot v}\geq\frac{|v-j|^2}{(1-|j|)(1+|j|)^2}. \tag*{(20)} 1 − j ⋅ v ∣ v − j ∣ 2 + 1 − ∣ j ∣ 2 [ j ⋅ ( v − j ) ] 2 ≥ ( 1 − ∣ j ∣ ) ( 1 + ∣ j ∣ ) 2 ∣ v − j ∣ 2 . (20)
For j = 0 j=0 j = 0 , or v = j v=j v = j , this is an equality. Otherwise, set
a = ∣ j ∣ , u = j ⋅ v a , r = ∣ v ∣ . a=|j|,\qquad u=\frac{j\cdot v}{a},\qquad r=|v|. a = ∣ j ∣ , u = a j ⋅ v , r = ∣ v ∣.
Then ∣ u ∣ ≤ r < 1 |u|\leq r<1 ∣ u ∣ ≤ r < 1 , and
∣ v − j ∣ 2 + [ j ⋅ ( v − j ) ] 2 1 − a 2 ( 1 − j ⋅ v ) ∣ v − j ∣ 2 = 1 1 − a u ( 1 + a 2 ( u − a ) 2 ( 1 − a 2 ) ( r 2 + a 2 − 2 a u ) ) . (21) \frac{|v-j|^2+\frac{[j\cdot(v-j)]^2}{1-a^2}}{(1-j\cdot v)|v-j|^2}
=\frac{1}{1-au}\left(1+\frac{a^2(u-a)^2}{(1-a^2)(r^2+a^2-2au)}\right). \tag*{(21)} ( 1 − j ⋅ v ) ∣ v − j ∣ 2 ∣ v − j ∣ 2 + 1 − a 2 [ j ⋅ ( v − j ) ] 2 = 1 − a u 1 ( 1 + ( 1 − a 2 ) ( r 2 + a 2 − 2 a u ) a 2 ( u − a ) 2 ) . (21)
Since r < 1 r<1 r < 1 , we have
r 2 + a 2 − 2 a u ≤ 1 + a 2 − 2 a u . r^2+a^2-2au\leq1+a^2-2au. r 2 + a 2 − 2 a u ≤ 1 + a 2 − 2 a u .
Using the identity
( 1 − a 2 ) ( 1 + a 2 − 2 a u ) + a 2 ( u − a ) 2 = ( 1 − a u ) 2 , (1-a^2)(1+a^2-2au)+a^2(u-a)^2=(1-au)^2, ( 1 − a 2 ) ( 1 + a 2 − 2 a u ) + a 2 ( u − a ) 2 = ( 1 − a u ) 2 ,
we deduce that the right hand side of (21) is bounded below by
1 − a u ( 1 − a 2 ) ( 1 + a 2 − 2 a u ) . \frac{1-au}{(1-a^2)(1+a^2-2au)}. ( 1 − a 2 ) ( 1 + a 2 − 2 a u ) 1 − a u .
The last expression is increasing in u u u , since
d d u ( 1 − a u ( 1 − a 2 ) ( 1 + a 2 − 2 a u ) ) = a ( 1 + a 2 − 2 a u ) 2 > 0. \frac{d}{du}\left(\frac{1-au}{(1-a^2)(1+a^2-2au)}\right)=\frac{a}{(1+a^2-2au)^2}>0. d u d ( ( 1 − a 2 ) ( 1 + a 2 − 2 a u ) 1 − a u ) = ( 1 + a 2 − 2 a u ) 2 a > 0.
Its minimum for − 1 ≤ u ≤ 1 -1 \le u \le1 − 1 ≤ u ≤ 1 is therefore attained at u = − 1 u=-1 u = − 1 , where it equals
1 ( 1 − a ) ( 1 + a ) 2 . \frac{1}{(1-a)(1+a)^2}. ( 1 − a ) ( 1 + a ) 2 1 .
This proves the pointwise inequality (20) . Integrating against G G G gives
Φ G ≥ μ 3 ( 1 − ∣ j G ∣ ) ( 1 + ∣ j G ∣ ) 2 ∫ R 3 ∣ v − j G ∣ 2 G d p . \Phi_G \ge\frac{\mu}{3(1-|j_G|)(1+|j_G|)^2}\int_{\mathbb{R}^3}|v-j_G|^2G\,\mathrm{d}p. Φ G ≥ 3 ( 1 − ∣ j G ∣ ) ( 1 + ∣ j G ∣ ) 2 μ ∫ R 3 ∣ v − j G ∣ 2 G d p .
Since the function
a ⟼ ( 1 − a ) ( 1 + a ) 2 a \longmapsto(1-a)(1+a)^2 a ⟼ ( 1 − a ) ( 1 + a ) 2
is increasing on [ 0 , 1 / 3 ] [0,1/3] [ 0 , 1/3 ] and decreasing on [ 1 / 3 , 1 ) [1/3,1) [ 1/3 , 1 ) , we obtain
Φ G ≥ 9 μ 32 ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p . \Phi_G \ge\frac{9\mu}{32}\int_{\mathbb{R}^3}|v(p)-j_G|^2G(p)\,\mathrm{d}p. Φ G ≥ 32 9 μ ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p .
The integral is strictly positive because v v v is injective and G G G is an L 1 L^1 L 1 probability density. Hence, the lower bound in (17) follows.
Next assume that G G G has finite energy. To prove the upper bound on Φ G \Phi_G Φ G in this case, we write, using the Cauchy–Schwarz inequality,
( ∫ R 3 G ⟨ p ⟩ d p ) 2 ≤ ( ∫ R 3 β G ⟨ p ⟩ G d p ) ( ∫ R 3 G ⟨ p ⟩ β G d p ) = ( 1 − ∣ j G ∣ 2 ) ∫ R 3 G ⟨ p ⟩ β G d p . \left(\int_{\mathbb{R}^3}\frac{G}{\langle p\rangle}\,\mathrm{d}p\right)^2
\le\left(\int_{\mathbb{R}^3}\frac{\beta_G}{\langle p\rangle}G\,\mathrm{d}p\right)
\left(\int_{\mathbb{R}^3}\frac{G}{\langle p\rangle\beta_G}\,\mathrm{d}p\right)
=(1-|j_G|^2)\int_{\mathbb{R}^3}\frac{G}{\langle p\rangle\beta_G}\,\mathrm{d}p. ( ∫ R 3 ⟨ p ⟩ G d p ) 2 ≤ ( ∫ R 3 ⟨ p ⟩ β G G d p ) ( ∫ R 3 ⟨ p ⟩ β G G d p ) = ( 1 − ∣ j G ∣ 2 ) ∫ R 3 ⟨ p ⟩ β G G d p .
Moreover, Jensen’s inequality implies
∫ R 3 G ⟨ p ⟩ d p ≥ 1 E G . \int_{\mathbb{R}^3}\frac{G}{\langle p\rangle}\,\mathrm{d}p \ge\frac{1}{E_G}. ∫ R 3 ⟨ p ⟩ G d p ≥ E G 1 .
Hence
∫ R 3 G ⟨ p ⟩ β G d p ≥ 1 E G 2 ( 1 − ∣ j G ∣ 2 ) ≥ 1 E G 2 . \int_{\mathbb{R}^3}\frac{G}{\langle p\rangle\beta_G}\,\mathrm{d}p
\ge\frac{1}{E_G^2(1-|j_G|^2)}
\ge\frac{1}{E_G^2}. ∫ R 3 ⟨ p ⟩ β G G d p ≥ E G 2 ( 1 − ∣ j G ∣ 2 ) 1 ≥ E G 2 1 .
Using this inequality in the definition of Φ G \Phi_G Φ G gives (18).
Remark. The upper bound in (18) is sharp for every fixed energy E G > 1 E_G>1 E G > 1 . Indeed, it can be approached by spherically symmetric probability densities concentrated near the sphere ⟨ p ⟩ = E G \langle p\rangle=E_G ⟨ p ⟩ = E G .
Lemma 8. Every probability density G G G with finite first energy satisfies
E G 2 − ∣ P G ∣ 2 > 1. (22) E_G^2-|P_G|^2>1. \tag*{(22)} E G 2 − ∣ P G ∣ 2 > 1. (22)
Proof. The function
p ↦ ⟨ p ⟩ = 1 + ∣ p ∣ 2 p \mapsto\langle p\rangle= \sqrt{1+|p|^2} p ↦ ⟨ p ⟩ = 1 + ∣ p ∣ 2
is strictly convex. Jensen’s inequality gives
E G ≥ 1 + ∣ ∫ R 3 p G ( p ) d p ∣ 2 = 1 + ∣ P G ∣ 2 . E_G \ge\sqrt{1+\left|\int_{\mathbb{R}^3}p\,G(p)\,dp\right|^2}=\sqrt{1+|P_G|^2}. E G ≥ 1 + ∫ R 3 p G ( p ) d p 2 = 1 + ∣ P G ∣ 2 .
Equality in Jensen’s inequality would imply that p p p is constant almost everywhere with respect to the probability distribution, which would make the corresponding probability measure a Dirac mass. Since G G G is an L 1 L^1 L 1 probability density, this is impossible. Therefore the inequality is strict.
We continue with a simple calculus lemma.
Lemma 9. Let v , w ∈ R 3 v,w\in\mathbb{R}^3 v , w ∈ R 3 satisfy ∣ v ∣ < 1 |v|<1 ∣ v ∣ < 1 and ∣ w ∣ < 1 |w|<1 ∣ w ∣ < 1 . Then
0 ≤ 1 − v ⋅ w − ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) ≤ ∣ v − w ∣ . (23) 0\le1-v\cdot w-\sqrt{(1-|v|^2)(1-|w|^2)}\le|v-w|. \tag*{(23)} 0 ≤ 1 − v ⋅ w − ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) ≤ ∣ v − w ∣. (23) Proof. Set
r = ∣ v ∣ , s = ∣ w ∣ . r=|v|,\qquad s=|w|. r = ∣ v ∣ , s = ∣ w ∣. Choose z ∈ [ − 1 , 1 ] z\in[-1,1] z ∈ [ − 1 , 1 ] such that
v ⋅ w = r s z . v\cdot w=rsz. v ⋅ w = r sz . When r s = 0 rs=0 r s = 0 , any z ∈ [ − 1 , 1 ] z\in[-1,1] z ∈ [ − 1 , 1 ] may be chosen. Set
A = 1 − r s z − ( 1 − r 2 ) ( 1 − s 2 ) . A=1-rsz-\sqrt{(1-r^2)(1-s^2)}. A = 1 − r sz − ( 1 − r 2 ) ( 1 − s 2 ) . The first inequality in (23) follows from
1 − r s z ≥ 1 − r s ≥ ( 1 − r 2 ) ( 1 − s 2 ) . 1-rsz\ge1-rs\ge\sqrt{(1-r^2)(1-s^2)}. 1 − r sz ≥ 1 − r s ≥ ( 1 − r 2 ) ( 1 − s 2 ) . To prove A ≤ ∣ v − w ∣ A\le|v-w| A ≤ ∣ v − w ∣ we calculate
∣ v − w ∣ 2 − A 2 = r 2 s 2 ( 1 − z 2 ) + 2 ( 1 − r 2 ) ( 1 − s 2 ) [ 1 − r s z − ( 1 − r 2 ) ( 1 − s 2 ) ] . |v-w|^2-A^2=r^2s^2(1-z^2)+2\sqrt{(1-r^2)(1-s^2)}\left[1-rsz-\sqrt{(1-r^2)(1-s^2)}\right]. ∣ v − w ∣ 2 − A 2 = r 2 s 2 ( 1 − z 2 ) + 2 ( 1 − r 2 ) ( 1 − s 2 ) [ 1 − r sz − ( 1 − r 2 ) ( 1 − s 2 ) ] . Both terms on the right hand side are nonnegative, hence A 2 ≤ ∣ v − w ∣ 2 A^2\le|v-w|^2 A 2 ≤ ∣ v − w ∣ 2 , which proves the second inequality in (23) .
Recall that the energy tail of a probability density G G G is defined by (6) .
Proposition 10. Let E > 1 E>1 E > 1 and P ∈ R 3 P\in\mathbb{R}^3 P ∈ R 3 be given such that
Δ : = E 2 − ∣ P ∣ 2 − 1 > 0. \Delta:=E^2-|P|^2-1>0. Δ := E 2 − ∣ P ∣ 2 − 1 > 0. Then every probability density G G G satisfying ( E G , P G ) = ( E , P ) (E_G,P_G)=(E,P) ( E G , P G ) = ( E , P ) satisfies, for every R > 1 R>1 R > 1 ,
∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p ≥ ( Δ − 2 E ( 1 + γ E ) T R ( G ) ) + 2 4 E 2 R 2 . (24) \int_{\mathbb{R}^3}|v(p)-j_G|^2G(p)\,dp\ge\frac{\left(\Delta-2E(1+\gamma_E)\mathcal{T}_R(G)\right)_+^2}{4E^2R^2}. \tag*{(24)} ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p ≥ 4 E 2 R 2 ( Δ − 2 E ( 1 + γ E ) T R ( G ) ) + 2 . (24)
Proof. For p , q ∈ R 3 p,q \in\mathbb{R}^{3} p , q ∈ R 3 , write
x = ⟨ p ⟩ , y = ⟨ q ⟩ , v = v ( p ) , w = v ( q ) . x=\langle p\rangle,\qquad y=\langle q\rangle,\qquad v=v(p),\qquad w=v(q). x = ⟨ p ⟩ , y = ⟨ q ⟩ , v = v ( p ) , w = v ( q ) .
Since p = x v p=xv p = xv , we have
E 2 − ∣ P ∣ 2 = ∬ R 3 × R 3 x y ( 1 − v ⋅ w ) G ( p ) G ( q ) d p d q . E^{2}-|P|^{2}=\iint_{\mathbb{R}^{3}\times\mathbb{R}^{3}}xy(1-v\cdot w)G(p)G(q)\,\mathrm{d}p\,\mathrm{d}q. E 2 − ∣ P ∣ 2 = ∬ R 3 × R 3 x y ( 1 − v ⋅ w ) G ( p ) G ( q ) d p d q .
Moreover,
1 − ∣ v ∣ 2 = 1 x 2 , 1 − ∣ w ∣ 2 = 1 y 2 , 1-|v|^{2}=\frac{1}{x^{2}},\qquad1-|w|^{2}=\frac{1}{y^{2}}, 1 − ∣ v ∣ 2 = x 2 1 , 1 − ∣ w ∣ 2 = y 2 1 ,
and therefore
x y ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) = 1. xy\sqrt{(1-|v|^{2})(1-|w|^{2})}=1. x y ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) = 1.
Since G G G is a probability density,
1 = ∬ R 3 × R 3 G ( p ) G ( q ) d p d q . 1=\iint_{\mathbb{R}^{3}\times\mathbb{R}^{3}}G(p)G(q)\,\mathrm{d}p\,\mathrm{d}q. 1 = ∬ R 3 × R 3 G ( p ) G ( q ) d p d q .
Consequently,
Δ = ∬ R 3 × R 3 x y [ 1 − v ⋅ w − ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) ] G ( p ) G ( q ) d p d q . \Delta=\iint_{\mathbb{R}^{3}\times\mathbb{R}^{3}}xy\left[1-v\cdot w-\sqrt{(1-|v|^{2})(1-|w|^{2})}\right]G(p)G(q)\,\mathrm{d}p\,\mathrm{d}q. Δ = ∬ R 3 × R 3 x y [ 1 − v ⋅ w − ( 1 − ∣ v ∣ 2 ) ( 1 − ∣ w ∣ 2 ) ] G ( p ) G ( q ) d p d q .
By Lemma 9,
Δ ≤ ∬ R 3 × R 3 x y ∣ v − w ∣ G ( p ) G ( q ) d p d q . \Delta\leq\iint_{\mathbb{R}^{3}\times\mathbb{R}^{3}}xy|v-w|G(p)G(q)\,\mathrm{d}p\,\mathrm{d}q. Δ ≤ ∬ R 3 × R 3 x y ∣ v − w ∣ G ( p ) G ( q ) d p d q .
Using
∣ v − w ∣ ≤ ∣ v − j G ∣ + ∣ w − j G ∣ , |v-w|\leq|v-j_{G}|+|w-j_{G}|, ∣ v − w ∣ ≤ ∣ v − j G ∣ + ∣ w − j G ∣ ,
we obtain
Δ ≤ 2 E ∫ R 3 ⟨ p ⟩ ∣ v ( p ) − j G ∣ G ( p ) d p . \Delta\leq2E\int_{\mathbb{R}^{3}}\langle p\rangle|v(p)-j_{G}|G(p)\,\mathrm{d}p. Δ ≤ 2 E ∫ R 3 ⟨ p ⟩ ∣ v ( p ) − j G ∣ G ( p ) d p .
Set
V G = ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p . V_{G}=\int_{\mathbb{R}^{3}}|v(p)-j_{G}|^{2}G(p)\,\mathrm{d}p. V G = ∫ R 3 ∣ v ( p ) − j G ∣ 2 G ( p ) d p .
Splitting the last integral into ⟨ p ⟩ ≤ R \langle p\rangle\leq R ⟨ p ⟩ ≤ R and ⟨ p ⟩ > R \langle p\rangle>R ⟨ p ⟩ > R , and using ∣ v − j G ∣ ≤ 1 + γ E |v-j_{G}|\leq1+\gamma_{E} ∣ v − j G ∣ ≤ 1 + γ E , which follows from Lemma 6, gives
∫ R 3 ⟨ p ⟩ ∣ v − j G ∣ G d p ≤ R ∫ R 3 ∣ v − j G ∣ G d p + ( 1 + γ E ) T R ( G ) ≤ R V G + ( 1 + γ E ) T R ( G ) . \begin{aligned}
\int_{\mathbb{R}^{3}}\langle p\rangle|v-j_{G}|G\,\mathrm{d}p\leq R\int_{\mathbb{R}^{3}}|v-j_{G}|G\,\mathrm{d}p+(1+\gamma_{E})T_{R}(G) \\
&\leq R\sqrt{V_{G}}+(1+\gamma_{E})T_{R}(G).
\end{aligned} ∫ R 3 ⟨ p ⟩ ∣ v − j G ∣ G d p ≤ R ∫ R 3 ∣ v − j G ∣ G d p + ( 1 + γ E ) T R ( G ) ≤ R V G + ( 1 + γ E ) T R ( G ) .
Hence
Δ ≤ 2 E R V G + 2 E ( 1 + γ E ) T R ( G ) . \Delta\leq2ER\sqrt{V_{G}}+2E(1+\gamma_{E})T_{R}(G). Δ ≤ 2 E R V G + 2 E ( 1 + γ E ) T R ( G ) .
Rearranging yields
V G ≥ Δ − 2 E ( 1 + γ E ) T R ( G ) 2 E R , \sqrt{V_{G}}\geq\frac{\Delta-2E(1+\gamma_{E})T_{R}(G)}{2ER}, V G ≥ 2 E R Δ − 2 E ( 1 + γ E ) T R ( G ) ,
which proves (24).
Corollary 11. Let G ( t , ⋅ ) G(t,\cdot) G ( t , ⋅ ) , t ≥ 0 t \ge0 t ≥ 0 , be a family of probability densities satisfying
E G ( t ) = E , P G ( t ) = P , E_{G(t)}=E,\qquad P_{G(t)}=P, E G ( t ) = E , P G ( t ) = P , and Δ = E 2 − ∣ P ∣ 2 − 1 > 0 \Delta=E^2-\lvert P\rvert^2-1>0 Δ = E 2 − ∣ P ∣ 2 − 1 > 0 . Assume that there exists R ∗ > 1 R_* > 1 R ∗ > 1 such that
sup t ≥ 0 T R ∗ ( G ( t ) ) < Δ 2 E ( 1 + γ E ) . (25) \sup_{t\ge0}\mathcal{T}_{R_*}(G(t))<\frac{\Delta}{2E(1+\gamma_E)}. \tag*{(25)} t ≥ 0 sup T R ∗ ( G ( t )) < 2 E ( 1 + γ E ) Δ . (25) Then
inf t ≥ 0 Φ G ( t ) > 0. (26) \inf_{t\ge0}\Phi_{G(t)}>0. \tag*{(26)} t ≥ 0 inf Φ G ( t ) > 0. (26) In particular, (26) holds if
lim R → ∞ sup t ≥ 0 T R ( G ( t ) ) = 0. \lim_{R\to\infty}\sup_{t\ge0}\mathcal{T}_R(G(t))=0. R → ∞ lim t ≥ 0 sup T R ( G ( t )) = 0. Proof. By Proposition 10, the inequality (25) implies
inf t ≥ 0 ∫ R 3 ∣ v − j G ( t ) ∣ 2 G ( t , p ) d p > 0. \inf_{t\ge0}\int_{\mathbb{R}^3}\lvert v-j_{G(t)}\rvert^2G(t,p)\,dp>0. t ≥ 0 inf ∫ R 3 ∣ v − j G ( t ) ∣ 2 G ( t , p ) d p > 0. The conclusion (26) then follows from Lemma 7.
To conclude this section we prove a Lipschitz continuity result for the coefficient functionals j G j_G j G and Φ G \Phi_G Φ G . We first recall the definition of the 1-Wasserstein distance [9 ] . Let G G G and H H H be probability densities on R 3 \mathbb{R}^3 R 3 with finite energy. Then
W 1 ( G , H ) : = inf π ∈ Π ( G , H ) ∫ R 3 × R 3 ∣ p − q ∣ d π ( p , q ) , W_1(G,H):=\inf_{\pi\in\Pi(G,H)}\int_{\mathbb{R}^3\times\mathbb{R}^3}\lvert p-q\rvert\,d\pi(p,q), W 1 ( G , H ) := π ∈ Π ( G , H ) inf ∫ R 3 × R 3 ∣ p − q ∣ d π ( p , q ) ,
where Π ( G , H ) \Pi(G,H) Π ( G , H ) denotes the set of couplings of the measures ν G = G ( p ) d p \nu_G=G(p)\,dp ν G = G ( p ) d p and ν H = H ( p ) d p \nu_H=H(p)\,dp ν H = H ( p ) d p , i.e., the set of probability measures on R 3 × R 3 \mathbb{R}^3\times\mathbb{R}^3 R 3 × R 3 having ν G \nu_G ν G and ν H \nu_H ν H as first and second marginals, respectively. By the Kantorovich–Rubinstein duality theorem,
W 1 ( G , H ) = sup Lip ( f ) ≤ 1 ∣ ∫ R 3 f ( p ) ( G ( p ) − H ( p ) ) d p ∣ . W_1(G,H)=\sup_{\operatorname{Lip}(f)\le1}\left\lvert\int_{\mathbb{R}^3}f(p)(G(p)-H(p))\,dp\right\rvert. W 1 ( G , H ) = Lip ( f ) ≤ 1 sup ∫ R 3 f ( p ) ( G ( p ) − H ( p )) d p .
Lemma 12. For every 0 ≤ γ < 1 0\le\gamma<1 0 ≤ γ < 1 there exists a constant C γ C_\gamma C γ , depending also on the fixed parameter μ \mu μ , such that, whenever G G G and H H H are probability densities with finite energy satisfying
∣ j G ∣ , ∣ j H ∣ ≤ γ , \lvert j_G\rvert,\lvert j_H\rvert\le\gamma, ∣ j G ∣ , ∣ j H ∣ ≤ γ , one has
∣ j G − j H ∣ + ∣ Φ G − Φ H ∣ ≤ C γ W 1 ( G , H ) . (27) \lvert j_G-j_H\rvert+\lvert\Phi_G-\Phi_H\rvert\le C_\gamma W_1(G,H). \tag*{(27)} ∣ j G − j H ∣ + ∣ Φ G − Φ H ∣ ≤ C γ W 1 ( G , H ) . (27)
Proof. A direct calculation gives
∂ p j v i = δ i j − v i v j ⟨ p ⟩ , \partial_{p_j}v_i=\frac{\delta_{ij}-v_i v_j}{\langle p\rangle}, ∂ p j v i = ⟨ p ⟩ δ ij − v i v j ,
and hence
∇ v ( p ) = 1 ⟨ p ⟩ ( I − v ( p ) ⊗ v ( p ) ) . \nabla v(p)=\frac{1}{\langle p\rangle}\left(I-v(p)\otimes v(p)\right). ∇ v ( p ) = ⟨ p ⟩ 1 ( I − v ( p ) ⊗ v ( p ) ) .
The eigenvalues of ∇ v ( p ) \nabla v(p) ∇ v ( p ) are ⟨ p ⟩ − 1 \langle p\rangle^{-1} ⟨ p ⟩ − 1 , ⟨ p ⟩ − 1 \langle p\rangle^{-1} ⟨ p ⟩ − 1 , ⟨ p ⟩ − 3 \langle p\rangle^{-3} ⟨ p ⟩ − 3 . Therefore, denoting by ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ the operator norm,
∥ ∇ v ( p ) ∥ = 1 ⟨ p ⟩ ≤ 1. \lVert\nabla v(p)\rVert=\frac{1}{\langle p\rangle}\le1. ∥ ∇ v ( p )∥ = ⟨ p ⟩ 1 ≤ 1.
For every e ∈ R 3 e\in\mathbb{R}^{3} e ∈ R 3 with ∣ e ∣ = 1 |e|=1 ∣ e ∣ = 1 , the function f ( p ) = e ⋅ v ( p ) f(p)=e\cdot v(p) f ( p ) = e ⋅ v ( p ) is therefore 1-Lipschitz. By the Kantorovich–Rubinstein formula,
∣ ∫ R 3 f ( p ) ( G ( p ) − H ( p ) ) d p ∣ = ∣ e ⋅ ( j G − j H ) ∣ ≤ W 1 ( G , H ) . \left|\int_{\mathbb{R}^{3}} f(p)(G(p)-H(p))\,dp\right|=|e\cdot(j_G-j_H)|\le W_1(G,H). ∫ R 3 f ( p ) ( G ( p ) − H ( p )) d p = ∣ e ⋅ ( j G − j H ) ∣ ≤ W 1 ( G , H ) .
Taking the supremum over ∣ e ∣ = 1 |e|=1 ∣ e ∣ = 1 gives
∣ j G − j H ∣ ≤ W 1 ( G , H ) . (28) |j_G-j_H|\le W_1(G,H). \tag*{(28)} ∣ j G − j H ∣ ≤ W 1 ( G , H ) . (28)
Set
F ( p , a ) = 1 ⟨ p ⟩ 2 ( 1 − a ⋅ v ( p ) ) . F(p,a)=\frac{1}{\langle p\rangle^2(1-a\cdot v(p))}. F ( p , a ) = ⟨ p ⟩ 2 ( 1 − a ⋅ v ( p )) 1 .
Then
Φ G − Φ H = μ 3 ∫ R 3 ( F ( p , j H ) H − F ( p , j G ) G ) d p . \Phi_G-\Phi_H=\frac{\mu}{3}\int_{\mathbb{R}^{3}}\left(F(p,j_H)H-F(p,j_G)G\right)\,dp. Φ G − Φ H = 3 μ ∫ R 3 ( F ( p , j H ) H − F ( p , j G ) G ) d p .
For ∣ a ∣ ≤ γ |a|\le\gamma ∣ a ∣ ≤ γ , there holds 1 − a ⋅ v ≥ 1 − γ 1-a\cdot v\ge1-\gamma 1 − a ⋅ v ≥ 1 − γ , and direct differentiation gives
sup ∣ a ∣ ≤ γ ∥ ∇ p F ( ⋅ , a ) ∥ ∞ + sup ∣ a ∣ ≤ γ ∥ ∇ a F ( ⋅ , a ) ∥ ∞ ≤ C γ . \sup_{|a|\le\gamma}\lVert\nabla_pF(\cdot,a)\rVert_\infty+\sup_{|a|\le\gamma}\lVert\nabla_aF(\cdot,a)\rVert_\infty\le C_\gamma. ∣ a ∣ ≤ γ sup ∥ ∇ p F ( ⋅ , a ) ∥ ∞ + ∣ a ∣ ≤ γ sup ∥ ∇ a F ( ⋅ , a ) ∥ ∞ ≤ C γ .
Hence
∣ ∫ R 3 F ( p , j G ) G d p − ∫ R 3 F ( p , j H ) H d p ∣ ≤ ∣ ∫ R 3 F ( p , j G ) ( G − H ) d p ∣ + ∫ R 3 ∣ F ( p , j G ) − F ( p , j H ) ∣ H d p ≤ C γ W 1 ( G , H ) + C γ ∣ j G − j H ∣ . \begin{aligned}
\left|\int_{\mathbb{R}^{3}}F(p,j_G)G\,dp-\int_{\mathbb{R}^{3}}F(p,j_H)H\,dp\right|
&\le\left|\int_{\mathbb{R}^{3}}F(p,j_G)(G-H)\,dp\right|\\
&\quad+\int_{\mathbb{R}^{3}}|F(p,j_G)-F(p,j_H)|H\,dp\\
&\le C_\gamma W_1(G,H)+C_\gamma|j_G-j_H|.
\end{aligned} ∫ R 3 F ( p , j G ) G d p − ∫ R 3 F ( p , j H ) H d p ≤ ∫ R 3 F ( p , j G ) ( G − H ) d p + ∫ R 3 ∣ F ( p , j G ) − F ( p , j H ) ∣ H d p ≤ C γ W 1 ( G , H ) + C γ ∣ j G − j H ∣.
Together with (28), this proves (27). □
Remark. The coefficient functionals are also locally Lipschitz with respect to the L 1 L^1 L 1 distance. Indeed, if ∣ j G ∣ , ∣ j H ∣ ≤ γ < 1 |j_G|,|j_H|\le\gamma<1 ∣ j G ∣ , ∣ j H ∣ ≤ γ < 1 , then
∣ j G − j H ∣ + ∣ Φ G − Φ H ∣ ≤ C γ ∥ G − H ∥ L 1 . |j_G-j_H|+|\Phi_G-\Phi_H|\le C_\gamma\lVert G-H\rVert_{L^1}. ∣ j G − j H ∣ + ∣ Φ G − Φ H ∣ ≤ C γ ∥ G − H ∥ L 1 .
The use of the 1-Wasserstein distance in the preceding lemma is nevertheless important for the fixed point argument in the proof of local existence.
The frozen Fokker–Planck equation
In this section we study the linear equation obtained by freezing the nonlinear coefficients Φ G \Phi_G Φ G and j G j_G j G . Let
ϕ ∈ C ( [ 0 , ∞ ) ) , a ∈ C ( [ 0 , ∞ ) ; R 3 ) \phi\in C([0,\infty)), \qquad a \in C([0,\infty);\mathbb{R}^{3}) ϕ ∈ C ([ 0 , ∞ )) , a ∈ C ([ 0 , ∞ ) ; R 3 )
be such that, for every T > 0 T>0 T > 0 , there exist constants δ T , Λ T , γ T \delta_T,\Lambda_T,\gamma_T δ T , Λ T , γ T satisfying
0 < δ T ≤ ϕ ( t ) ≤ Λ T , ∣ a ( t ) ∣ ≤ γ T < 1 , 0 ≤ t ≤ T . (29) 0<\delta_T\leq\phi(t)\leq\Lambda_T,\qquad|a(t)|\leq\gamma_T<1,\qquad0\leq t\leq T. \tag*{(29)} 0 < δ T ≤ ϕ ( t ) ≤ Λ T , ∣ a ( t ) ∣ ≤ γ T < 1 , 0 ≤ t ≤ T . (29)
We consider the linear Fokker–Planck equation
∂ t G = ϕ ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V a G ) , (30) \partial_tG=\phi\nabla_p\cdot(D\nabla_pG)+\nabla_p\cdot(V_aG), \tag*{(30)} ∂ t G = ϕ ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V a G ) , (30)
where
V a ( p ) = μ ⟨ p ⟩ ( p − a β a ( p ) ) , β a ( p ) = ⟨ p ⟩ ( 1 − a ⋅ v ( p ) ) . V_a(p)=\frac{\mu}{\langle p\rangle}\left(p-\frac{a}{\beta_a(p)}\right),\qquad\beta_a(p)=\langle p\rangle(1-a\cdot v(p)). V a ( p ) = ⟨ p ⟩ μ ( p − β a ( p ) a ) , β a ( p ) = ⟨ p ⟩ ( 1 − a ⋅ v ( p )) .
The equation admits a useful stochastic representation. Since
∑ j = 1 3 ∂ p j D i j ( p ) = 3 v i ( p ) , \sum_{j=1}^{3}\partial_{p_j}D_{ij}(p)=3v_i(p), j = 1 ∑ 3 ∂ p j D ij ( p ) = 3 v i ( p ) ,
we may rewrite (30) as
∂ t G = − ∇ p ⋅ ( Q ϕ , a G ) + ϕ ∑ i , j = 1 3 ∂ p i ∂ p j ( D i j G ) , \partial_tG=-\nabla_p\cdot(Q_{\phi,a}G)+\phi\sum_{i,j=1}^{3}\partial_{p_i}\partial_{p_j}(D_{ij}G), ∂ t G = − ∇ p ⋅ ( Q ϕ , a G ) + ϕ i , j = 1 ∑ 3 ∂ p i ∂ p j ( D ij G ) ,
where
Q ϕ , a ( p ) = 3 ϕ v ( p ) − V a ( p ) . (31) Q_{\phi,a}(p)=3\phi\,v(p)-V_a(p). \tag*{(31)} Q ϕ , a ( p ) = 3 ϕ v ( p ) − V a ( p ) . (31)
Consider now the stochastic differential equation
d P t = Q ϕ ( t ) , a ( t ) ( P t ) d t + σ ϕ ( t ) ( P t ) d B t , (32) dP_t=Q_{\phi(t),a(t)}(P_t)\,dt+\sigma_{\phi(t)}(P_t)\,dB_t, \tag*{(32)} d P t = Q ϕ ( t ) , a ( t ) ( P t ) d t + σ ϕ ( t ) ( P t ) d B t , (32)
where B t B_t B t is a standard three dimensional Brownian motion and
σ ϕ ( p ) = 2 ϕ D 1 / 2 ( p ) . \sigma_\phi(p)=\sqrt{2\phi}\,D^{1/2}(p). σ ϕ ( p ) = 2 ϕ D 1/2 ( p ) .
An explicit positive square root of D D D is
D 1 / 2 ( p ) = ⟨ p ⟩ − 1 / 2 [ I + p ⊗ p ⟨ p ⟩ + 1 ] . D^{1/2}(p)=\langle p\rangle^{-1/2}\left[I+\frac{p\otimes p}{\langle p\rangle+1}\right]. D 1/2 ( p ) = ⟨ p ⟩ − 1/2 [ I + ⟨ p ⟩ + 1 p ⊗ p ] .
The stochastic differential equation (32) and the Fokker–Planck equation (30) have the following (standard) relation.
Suppose that P t P_t P t is a solution of (32) with initial data P 0 P_0 P 0 having probability distribution ν 0 = G 0 ( p ) d p \nu_0 = G_0(p)\,dp ν 0 = G 0 ( p ) d p , and denote by ν t \nu_t ν t the probability distribution of P t P_t P t . The infinitesimal generator associated with (32) is
L t ψ ( p ) = Q ϕ ( t ) , a ( t ) ( p ) ⋅ ∇ p ψ ( p ) + ϕ ( t ) D ( p ) : ∇ p 2 ψ ( p ) , (33) L_t\psi(p) = Q_{\phi(t),a(t)}(p) \cdot\nabla_p\psi(p) + \phi(t)D(p) : \nabla_p^2\psi(p), \tag*{(33)} L t ψ ( p ) = Q ϕ ( t ) , a ( t ) ( p ) ⋅ ∇ p ψ ( p ) + ϕ ( t ) D ( p ) : ∇ p 2 ψ ( p ) , (33)
where
D : ∇ p 2 ψ = ∑ i , j = 1 3 D i j ∂ p i ∂ p j ψ . D : \nabla_p^2\psi= \sum_{i,j=1}^{3} D_{ij}\partial_{p_i}\partial_{p_j}\psi. D : ∇ p 2 ψ = i , j = 1 ∑ 3 D ij ∂ p i ∂ p j ψ .
By (31) and
∇ p ⋅ ( D ∇ p ψ ) = D : ∇ p 2 ψ + 3 v ⋅ ∇ p ψ , \nabla_p \cdot(D\nabla_p\psi) = D : \nabla_p^2\psi+ 3v \cdot\nabla_p\psi, ∇ p ⋅ ( D ∇ p ψ ) = D : ∇ p 2 ψ + 3 v ⋅ ∇ p ψ ,
we have
L t ψ = ϕ ( t ) ∇ p ⋅ ( D ∇ p ψ ) − V a ( t ) ⋅ ∇ p ψ . (34) L_t\psi= \phi(t)\nabla_p \cdot(D\nabla_p\psi) - V_{a(t)} \cdot\nabla_p\psi. \tag*{(34)} L t ψ = ϕ ( t ) ∇ p ⋅ ( D ∇ p ψ ) − V a ( t ) ⋅ ∇ p ψ . (34)
Let T > 0 T > 0 T > 0 and ψ ∈ C c ∞ ( [ 0 , T ) × R 3 ) \psi\in C^\infty_c([0,T) \times\mathbb{R}^3) ψ ∈ C c ∞ ([ 0 , T ) × R 3 ) . Applying Itô’s formula to ψ ( t , P t ) \psi(t,P_t) ψ ( t , P t ) gives
ψ ( T , P T ) = ψ ( 0 , P 0 ) + ∫ 0 T [ ∂ t ψ ( t , P t ) + L t ψ ( t , P t ) ] d t + ∫ 0 T ∇ p ψ ( t , P t ) ⋅ σ ϕ ( t ) ( P t ) d B t . \psi(T,P_T) = \psi(0,P_0) + \int_0^T [\partial_t\psi(t,P_t) + L_t\psi(t,P_t)]\,dt + \int_0^T \nabla_p\psi(t,P_t) \cdot\sigma_{\phi(t)}(P_t)\,dB_t. ψ ( T , P T ) = ψ ( 0 , P 0 ) + ∫ 0 T [ ∂ t ψ ( t , P t ) + L t ψ ( t , P t )] d t + ∫ 0 T ∇ p ψ ( t , P t ) ⋅ σ ϕ ( t ) ( P t ) d B t .
Since ψ ( T , ⋅ ) = 0 \psi(T,\cdot)=0 ψ ( T , ⋅ ) = 0 , taking expectations yields
∫ R 3 ψ ( 0 , p ) d ν 0 ( p ) + ∫ 0 T ∫ R 3 [ ∂ t ψ ( t , p ) + L t ψ ( t , p ) ] d ν t ( p ) d t = 0. (35) \int_{\mathbb{R}^3} \psi(0,p)\,d\nu_0(p) + \int_0^T \int_{\mathbb{R}^3} [\partial_t\psi(t,p) + L_t\psi(t,p)]\,d\nu_t(p)\,dt = 0. \tag*{(35)} ∫ R 3 ψ ( 0 , p ) d ν 0 ( p ) + ∫ 0 T ∫ R 3 [ ∂ t ψ ( t , p ) + L t ψ ( t , p )] d ν t ( p ) d t = 0. (35)
The stochastic integral has zero expectation because ∇ p ψ \nabla_p\psi ∇ p ψ has compact support and σ ϕ ( t ) \sigma_{\phi(t)} σ ϕ ( t ) is bounded on compact subsets of [ 0 , T ] × R 3 [0,T] \times\mathbb{R}^3 [ 0 , T ] × R 3 . Consequently, the probability measures ν t \nu_t ν t solve the forward Kolmogorov equation
∂ t ν t = L t ∗ ν t (36) \partial_t\nu_t = L_t^*\nu_t \tag*{(36)} ∂ t ν t = L t ∗ ν t (36)
in the distributional sense. If ν t \nu_t ν t admits a density G ( t , p ) G(t,p) G ( t , p ) , then, using (34) , we obtain
∫ 0 T ∫ R 3 G ( t , p ) [ ∂ t ψ + ϕ ( t ) ∇ p ⋅ ( D ∇ p ψ ) − V a ( t ) ⋅ ∇ p ψ ] d p d t + ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p = 0. (37) \int_0^T \int_{\mathbb{R}^3} G(t,p) \left[\partial_t\psi+ \phi(t)\nabla_p \cdot(D\nabla_p\psi) - V_{a(t)} \cdot\nabla_p\psi\right]\,dp\,dt + \int_{\mathbb{R}^3} G_0(p)\psi(0,p)\,dp = 0. \tag*{(37)} ∫ 0 T ∫ R 3 G ( t , p ) [ ∂ t ψ + ϕ ( t ) ∇ p ⋅ ( D ∇ p ψ ) − V a ( t ) ⋅ ∇ p ψ ] d p d t + ∫ R 3 G 0 ( p ) ψ ( 0 , p ) d p = 0. (37)
This is precisely the weak formulation of the frozen Fokker–Planck equation (30) . Thus, whenever the time marginal distributions of a solution to (32) admit densities, these densities solve (30) in the weak sense.
Next we show that the coefficients of (32) satisfy the usual bounds in the Cauchy theory for stochastic differential equations [7 ] . For a matrix A = ( A i j ) A = (A_{ij}) A = ( A ij ) , we denote by
∥ A ∥ H S = ( ∑ i , j ∣ A i j ∣ 2 ) 1 / 2 \lVert A\rVert_{\mathrm{HS}} = \left(\sum_{i,j} |A_{ij}|^2\right)^{1/2} ∥ A ∥ HS = ( i , j ∑ ∣ A ij ∣ 2 ) 1/2
its Hilbert–Schmidt norm.
Lemma 13. Let 0 < δ ≤ ϕ , ψ ≤ Λ 0 < \delta\le\phi,\psi\le\Lambda 0 < δ ≤ ϕ , ψ ≤ Λ , ∣ a ∣ , ∣ b ∣ ≤ γ < 1 |a|,|b| \le\gamma< 1 ∣ a ∣ , ∣ b ∣ ≤ γ < 1 .
Then
∣ Q ϕ , a ( p ) ∣ ≤ C , ∥ σ ϕ ( p ) ∥ H S ≤ C ⟨ p ⟩ 1 / 2 (38) |Q_{\phi,a}(p)| \le C,\qquad\|\sigma_{\phi}(p)\|_{\mathrm{HS}} \le C\langle p\rangle^{1/2} \tag*{(38)} ∣ Q ϕ , a ( p ) ∣ ≤ C , ∥ σ ϕ ( p ) ∥ HS ≤ C ⟨ p ⟩ 1/2 (38) and
∣ Q ϕ , a ( p ) − Q ψ , b ( q ) ∣ ≤ C ( ∣ p − q ∣ + ∣ ϕ − ψ ∣ + ∣ a − b ∣ ) , (39) |Q_{\phi,a}(p)-Q_{\psi,b}(q)| \le C\left(|p-q|+|\phi-\psi|+|a-b|\right), \tag*{(39)} ∣ Q ϕ , a ( p ) − Q ψ , b ( q ) ∣ ≤ C ( ∣ p − q ∣ + ∣ ϕ − ψ ∣ + ∣ a − b ∣ ) , (39) ∥ σ ϕ ( p ) − σ ψ ( q ) ∥ H S 2 ≤ C ( ∣ p − q ∣ 2 + ∣ ϕ − ψ ∣ 2 ⟨ q ⟩ ) . (40) \|\sigma_{\phi}(p)-\sigma_{\psi}(q)\|_{\mathrm{HS}}^2 \le C\left(|p-q|^2+|\phi-\psi|^2\langle q\rangle\right). \tag*{(40)} ∥ σ ϕ ( p ) − σ ψ ( q ) ∥ HS 2 ≤ C ( ∣ p − q ∣ 2 + ∣ ϕ − ψ ∣ 2 ⟨ q ⟩ ) . (40) The constant C C C depends only on δ , Λ , γ \delta,\Lambda,\gamma δ , Λ , γ and μ \mu μ . Finally, the coefficients Q ϕ , a Q_{\phi,a} Q ϕ , a and σ ϕ \sigma_{\phi} σ ϕ are smooth in p p p , and for every integer m ≥ 1 m\ge1 m ≥ 1 there exists a constant C m C_m C m , depending only on m , δ , Λ , γ m,\delta,\Lambda,\gamma m , δ , Λ , γ and μ \mu μ , such that
sup p ∈ R 3 ( ∣ ∇ p m Q ϕ , a ( p ) ∣ + ∥ ∇ p m σ ϕ ( p ) ∥ ) ≤ C m . (41) \sup_{p\in\mathbb{R}^3}\left(|\nabla_p^m Q_{\phi,a}(p)|+\|\nabla_p^m\sigma_{\phi}(p)\|\right)\le C_m. \tag*{(41)} p ∈ R 3 sup ( ∣ ∇ p m Q ϕ , a ( p ) ∣ + ∥ ∇ p m σ ϕ ( p ) ∥ ) ≤ C m . (41)
Proof. Since
1 − a ⋅ v ( p ) ≥ 1 − γ , 1-a\cdot v(p)\ge1-\gamma, 1 − a ⋅ v ( p ) ≥ 1 − γ ,
the functions
p ⟼ a ⟨ p ⟩ β a ( p ) = a ⟨ p ⟩ 2 ( 1 − a ⋅ v ( p ) ) p\longmapsto\frac{a}{\langle p\rangle\beta_a(p)}
=\frac{a}{\langle p\rangle^2(1-a\cdot v(p))} p ⟼ ⟨ p ⟩ β a ( p ) a = ⟨ p ⟩ 2 ( 1 − a ⋅ v ( p )) a
and their first derivatives with respect to p p p and a a a are uniformly bounded for ∣ a ∣ ≤ γ |a|\le\gamma ∣ a ∣ ≤ γ . Since in addition v v v is bounded and Lipschitz continuous, the bound on Q ϕ , a Q_{\phi,a} Q ϕ , a in (38) and the estimate (39) follow. To study the diffusion coefficient σ ϕ ( p ) = 2 ϕ D 1 / 2 ( p ) \sigma_{\phi}(p)=\sqrt{2\phi}\,D^{1/2}(p) σ ϕ ( p ) = 2 ϕ D 1/2 ( p ) , we compute
∂ p k ( D 1 / 2 ) i j ( p ) = − p k δ i j 2 ⟨ p ⟩ 5 / 2 + δ i k p j + p i δ j k ⟨ p ⟩ 1 / 2 ( ⟨ p ⟩ + 1 ) − ( 3 ⟨ p ⟩ + 1 ) p i p j p k 2 ⟨ p ⟩ 5 / 2 ( ⟨ p ⟩ + 1 ) 2 . \partial_{p_k}(D^{1/2})_{ij}(p)
=-\frac{p_k\delta_{ij}}{2\langle p\rangle^{5/2}}
+\frac{\delta_{ik}p_j+p_i\delta_{jk}}{\langle p\rangle^{1/2}(\langle p\rangle+1)}
-\frac{(3\langle p\rangle+1)p_ip_jp_k}{2\langle p\rangle^{5/2}(\langle p\rangle+1)^2}. ∂ p k ( D 1/2 ) ij ( p ) = − 2 ⟨ p ⟩ 5/2 p k δ ij + ⟨ p ⟩ 1/2 (⟨ p ⟩ + 1 ) δ ik p j + p i δ j k − 2 ⟨ p ⟩ 5/2 (⟨ p ⟩ + 1 ) 2 ( 3 ⟨ p ⟩ + 1 ) p i p j p k .
Consequently,
∣ ∂ p k ( D 1 / 2 ) i j ( p ) ∣ ≤ C ⟨ p ⟩ − 1 / 2 ≤ C . |\partial_{p_k}(D^{1/2})_{ij}(p)|\le C\langle p\rangle^{-1/2}\le C. ∣ ∂ p k ( D 1/2 ) ij ( p ) ∣ ≤ C ⟨ p ⟩ − 1/2 ≤ C .
Thus, D 1 / 2 D^{1/2} D 1/2 is globally Lipschitz. Since the eigenvalues of D ( p ) D(p) D ( p ) are ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ , ⟨ p ⟩ − 1 \langle p\rangle^{-1} ⟨ p ⟩ − 1 , ⟨ p ⟩ − 1 \langle p\rangle^{-1} ⟨ p ⟩ − 1 , we have
∥ D 1 / 2 ( p ) ∥ H S 2 = Tr D ( p ) = ⟨ p ⟩ + 2 ⟨ p ⟩ ≤ 3 ⟨ p ⟩ , \|D^{1/2}(p)\|_{\mathrm{HS}}^2=\operatorname{Tr}D(p)=\langle p\rangle+\frac{2}{\langle p\rangle}\le3\langle p\rangle, ∥ D 1/2 ( p ) ∥ HS 2 = Tr D ( p ) = ⟨ p ⟩ + ⟨ p ⟩ 2 ≤ 3 ⟨ p ⟩ ,
and therefore the second bound in (38) follows. Moreover,
∣ ϕ − ψ ∣ ≤ ∣ ϕ − ψ ∣ 2 δ . |\sqrt{\phi}-\sqrt{\psi}|\le\frac{|\phi-\psi|}{2\sqrt{\delta}}. ∣ ϕ − ψ ∣ ≤ 2 δ ∣ ϕ − ψ ∣ .
Writing
σ ϕ ( p ) − σ ψ ( q ) = 2 ϕ ( D 1 / 2 ( p ) − D 1 / 2 ( q ) ) + 2 ( ϕ − ψ ) D 1 / 2 ( q ) \sigma_{\phi}(p)-\sigma_{\psi}(q)
=\sqrt{2\phi}\left(D^{1/2}(p)-D^{1/2}(q)\right)
+\sqrt{2}\left(\sqrt{\phi}-\sqrt{\psi}\right)D^{1/2}(q) σ ϕ ( p ) − σ ψ ( q ) = 2 ϕ ( D 1/2 ( p ) − D 1/2 ( q ) ) + 2 ( ϕ − ψ ) D 1/2 ( q )
and using the preceding estimates gives (40) . The bounds (41) follow by differentiating the explicit formulas for Q ϕ , a Q_{\phi,a} Q ϕ , a and D 1 / 2 D^{1/2} D 1/2 , using 1 − a ⋅ v ≥ 1 − γ 1-a\cdot v\ge1-\gamma 1 − a ⋅ v ≥ 1 − γ .
Theorem 14 (Global well-posedness of the frozen equation). Let ϕ \phi ϕ and a a a satisfy (29) , and let G 0 ∈ L 1 ( R 3 ) G_0 \in L^1(\mathbb{R}^3) G 0 ∈ L 1 ( R 3 ) satisfy
G 0 ≥ 0 a.e. , ∫ R 3 G 0 d p = 1 , E G 0 < ∞ . G_0 \ge0\ \text{a.e.},\qquad\int_{\mathbb{R}^3} G_0\,\mathrm{d}p = 1,\qquad E_{G_0}<\infty. G 0 ≥ 0 a.e. , ∫ R 3 G 0 d p = 1 , E G 0 < ∞. Then the Cauchy problem
{ ∂ t G = ϕ ( t ) ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V a ( t ) G ) , G ( 0 , p ) = G 0 ( p ) (42) \begin{cases}
\partial_t G=\phi(t)\nabla_p\cdot(D\nabla_pG)+\nabla_p\cdot(V_{a(t)}G),\\
G(0,p)=G_0(p) \tag*{(42)}
\end{cases} { ∂ t G = ϕ ( t ) ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V a ( t ) G ) , G ( 0 , p ) = G 0 ( p ) (42) has a weak solution
G ϕ , a ∈ C ( [ 0 , ∞ ) ; L 1 ( R 3 ) ) G^{\phi,a}\in C([0,\infty);L^1(\mathbb{R}^3)) G ϕ , a ∈ C ([ 0 , ∞ ) ; L 1 ( R 3 )) such that, for every T > 0 T>0 T > 0 ,
sup 0 ≤ t ≤ T E G ϕ , a ( t ) ≤ E G 0 + ( 3 Λ T + μ ) T . (43) \sup_{0\le t\le T}E_{G^{\phi,a}}(t)\le E_{G_0}+(3\Lambda T+\mu)T. \tag*{(43)} 0 ≤ t ≤ T sup E G ϕ , a ( t ) ≤ E G 0 + ( 3Λ T + μ ) T . (43) The solution is a time dependent probability density:
G ϕ , a ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ϕ , a ( t , p ) d p = 1 , t ≥ 0. (44) G^{\phi,a}(t,\cdot)\ge0\ \text{a.e.},\qquad\int_{\mathbb{R}^3}G^{\phi,a}(t,p)\,\mathrm{d}p=1,\qquad t\ge0. \tag*{(44)} G ϕ , a ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ϕ , a ( t , p ) d p = 1 , t ≥ 0. (44) For every T > 0 T>0 T > 0 , the solution is unique among nonnegative weak solutions in C ( [ 0 , T ] ; L 1 ( R 3 ) ) C([0,T];L^1(\mathbb{R}^3)) C ([ 0 , T ] ; L 1 ( R 3 )) with unit mass and bounded energy on [ 0 , T ] [0,T] [ 0 , T ] . Moreover, for every t ≥ 0 t\ge0 t ≥ 0 , the probability measure G ϕ , a ( t , p ) d p G^{\phi,a}(t,p)\,\mathrm{d}p G ϕ , a ( t , p ) d p is the distribution of the solution P t P_t P t of (32) whose initial data P 0 P_0 P 0 has distribution G 0 ( p ) d p G_0(p)\,\mathrm{d}p G 0 ( p ) d p . Finally, the frozen evolution is contractive in L 1 L^1 L 1 . More precisely, if G ϕ , a G^{\phi,a} G ϕ , a and H ϕ , a H^{\phi,a} H ϕ , a correspond to initial probability densities G 0 G_0 G 0 and H 0 H_0 H 0 with finite energy, then
∥ G ϕ , a ( t ) − H ϕ , a ( t ) ∥ L 1 ≤ ∥ G 0 − H 0 ∥ L 1 , t ≥ 0. (45) \left\|G^{\phi,a}(t)-H^{\phi,a}(t)\right\|_{L^1}\le\left\|G_0-H_0\right\|_{L^1},\qquad t\ge0. \tag*{(45)} G ϕ , a ( t ) − H ϕ , a ( t ) L 1 ≤ ∥ G 0 − H 0 ∥ L 1 , t ≥ 0. (45)
Proof. We divide the proof into several steps, using standard results from the theory of stochastic differential equations and stochastic flows.
Step 1. Construction by the stochastic differential equation.
Fix T > 0 T>0 T > 0 . By Lemma 13 and (29) , the coefficients Q ϕ ( t ) , a ( t ) ( p ) Q_{\phi(t),a(t)}(p) Q ϕ ( t ) , a ( t ) ( p ) and σ ϕ ( t ) ( p ) \sigma_{\phi(t)}(p) σ ϕ ( t ) ( p ) are globally Lipschitz in p p p , uniformly for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , and satisfy the linear growth condition. Therefore, by the standard Cauchy theory for stochastic differential equations [7 ] , (32) has a unique nonexplosive strong solution for every almost surely finite initial random variable P 0 P_0 P 0 independent of the Brownian motion. Since T > 0 T>0 T > 0 is arbitrary, the solution exists globally. Moreover, the regularity of the coefficients established in Lemma 13 allows us to apply the theory of differentiable stochastic flows [5 Theorem 4.6.5] . Consequently, the solutions of (32) generate a stochastic flow X s , t X_{s,t} X s , t of C 1 C^1 C 1 diffeomorphisms of R 3 \mathbb{R}^3 R 3 . More precisely, let P t s , p P_t^{s,p} P t s , p denote the solution of (32) at time t t t with deterministic initial condition P s s , p = p P^{s,p}_s=p P s s , p = p . Then, for each s ≤ t s\leq t s ≤ t and almost every sample path of the Brownian motion, the map
X s , t : R 3 → R 3 , X s , t ( p ) = P t s , p , X_{s,t}:\mathbb{R}^{3}\to\mathbb{R}^{3},\qquad X_{s,t}(p)=P^{s,p}_t, X s , t : R 3 → R 3 , X s , t ( p ) = P t s , p ,
is a C 1 C^{1} C 1 diffeomorphism of R 3 \mathbb{R}^{3} R 3 .
Choose P 0 P_0 P 0 with probability distribution ν 0 = G 0 ( p ) d p \nu_0=G_0(p)\,dp ν 0 = G 0 ( p ) d p , independently of the Brownian motion, and denote the distribution of P t P_t P t by ν t \nu_t ν t . The pushforward of G 0 ( p ) d p G_0(p)\,dp G 0 ( p ) d p under X 0 , t X_{0,t} X 0 , t is the random probability measure defined by
( X 0 , t ) # ( G 0 d p ) ( A ) = ∫ X 0 , t − 1 ( A ) G 0 ( p ) d p , A ⊂ R 3 Borel . (X_{0,t})_{\#}(G_0\,dp)(A)=\int_{X_{0,t}^{-1}(A)}G_0(p)\,dp,\qquad A\subset\mathbb{R}^{3}\text{ Borel}. ( X 0 , t ) # ( G 0 d p ) ( A ) = ∫ X 0 , t − 1 ( A ) G 0 ( p ) d p , A ⊂ R 3 Borel .
Since P 0 P_0 P 0 is independent of the Brownian motion, the distribution ν t \nu_t ν t of P t P_t P t satisfies
ν t ( A ) = E [ ( X 0 , t ) # ( G 0 d p ) ( A ) ] , A ⊂ R 3 Borel . \nu_t(A)=\mathbb{E}[(X_{0,t})_{\#}(G_0\,dp)(A)],\qquad A\subset\mathbb{R}^{3}\text{ Borel}. ν t ( A ) = E [( X 0 , t ) # ( G 0 d p ) ( A )] , A ⊂ R 3 Borel .
By the change of variables formula, the pushforward measure has density
g t ( p ) = G 0 ( X 0 , t − 1 ( p ) ) ∣ det ∇ X 0 , t − 1 ( p ) ∣ , g_t(p)=G_0(X_{0,t}^{-1}(p))\left|\det\nabla X_{0,t}^{-1}(p)\right|, g t ( p ) = G 0 ( X 0 , t − 1 ( p )) det ∇ X 0 , t − 1 ( p ) ,
and therefore
ν t ( A ) = E [ ∫ A g t ( p ) d p ] = ∫ A E [ g t ( p ) ] d p \nu_t(A)=\mathbb{E}\left[\int_Ag_t(p)\,dp\right]=\int_A\mathbb{E}[g_t(p)]\,dp ν t ( A ) = E [ ∫ A g t ( p ) d p ] = ∫ A E [ g t ( p )] d p
for every Borel set A ⊂ R 3 A\subset\mathbb{R}^{3} A ⊂ R 3 . Thus, ν t \nu_t ν t has density G ϕ , a ( t , p ) = E [ g t ( p ) ] G^{\phi,a}(t,p)=\mathbb{E}[g_t(p)] G ϕ , a ( t , p ) = E [ g t ( p )] for every t > 0 t>0 t > 0 . We set G ϕ , a ( 0 , ⋅ ) = G 0 G^{\phi,a}(0,\cdot)=G_0 G ϕ , a ( 0 , ⋅ ) = G 0 . By construction,
G ϕ , a ( t , p ) ≥ 0 a.e. , ∫ R 3 G ϕ , a ( t , p ) d p = 1 , t ≥ 0. G^{\phi,a}(t,p)\geq0\ \text{a.e.},\qquad\int_{\mathbb{R}^{3}}G^{\phi,a}(t,p)\,dp=1,\qquad t\geq0. G ϕ , a ( t , p ) ≥ 0 a.e. , ∫ R 3 G ϕ , a ( t , p ) d p = 1 , t ≥ 0.
Finally, applying Itô’s formula to ψ ( t , P t ) \psi(t,P_t) ψ ( t , P t ) , with ψ ∈ C c ∞ ( [ 0 , T ) × R 3 ) \psi\in C^\infty_c([0,T)\times\mathbb{R}^{3}) ψ ∈ C c ∞ ([ 0 , T ) × R 3 ) , and taking expectations shows, as established in (35) , that G ϕ , a G^{\phi,a} G ϕ , a satisfies the weak formulation (37) . The strong continuity of G ϕ , a G^{\phi,a} G ϕ , a in L 1 ( R 3 ) L^1(\mathbb{R}^{3}) L 1 ( R 3 ) will be established in Step 4.
Energy estimate
Let
h ( p ) = ⟨ p ⟩ h(p)=\langle p\rangle h ( p ) = ⟨ p ⟩
and recall the definition (34) of the backward generator L t L_t L t . A direct calculation gives
L t h = 3 ϕ ( t ) − μ + μ ⟨ p ⟩ β a ( t ) ( p ) . (46) L_th=3\phi(t)-\mu+\frac{\mu}{\langle p\rangle\beta_{a(t)}(p)}. \tag*{(46)} L t h = 3 ϕ ( t ) − μ + ⟨ p ⟩ β a ( t ) ( p ) μ . (46)
Moreover,
1 ⟨ p ⟩ β a ( p ) = 1 − ∣ v ∣ 2 1 − a ⋅ v ≤ 1 − ∣ v ∣ 2 1 − ∣ v ∣ = 1 + ∣ v ∣ ≤ 2. \frac{1}{\langle p\rangle\beta_a(p)}
=\frac{1-|v|^2}{1-a\cdot v}
\leq\frac{1-|v|^2}{1-|v|}
=1+|v|\leq2. ⟨ p ⟩ β a ( p ) 1 = 1 − a ⋅ v 1 − ∣ v ∣ 2 ≤ 1 − ∣ v ∣ 1 − ∣ v ∣ 2 = 1 + ∣ v ∣ ≤ 2.
Consequently, for 0 ≤ t ≤ T 0 \le t \le T 0 ≤ t ≤ T ,
− μ ≤ L t h ≤ 3 Λ T + μ . -\mu\le L_t h \le3\Lambda_T+\mu. − μ ≤ L t h ≤ 3 Λ T + μ .
For R > 0 R>0 R > 0 , define the stopping time
τ R = inf { t ≥ 0 : ∣ P t ∣ ≥ R } , \tau_R=\inf\{t\ge0: |P_t|\ge R\}, τ R = inf { t ≥ 0 : ∣ P t ∣ ≥ R } ,
with the convention inf ∅ = ∞ \inf\varnothing=\infty inf ∅ = ∞ . Since the solution of (32) is global and nonexplosive, we have τ R → ∞ \tau_R\to\infty τ R → ∞ almost surely as R → ∞ R\to\infty R → ∞ . Applying Itô’s formula to h ( P t ∧ τ R ) h(P_{t\wedge\tau_R}) h ( P t ∧ τ R ) and taking expectations, we obtain
E h ( P t ∧ τ R ) = E h ( P 0 ) + E ∫ 0 t 1 { s < τ R } L s h ( P s ) d s . \mathbb{E}h(P_{t\wedge\tau_R})=\mathbb{E}h(P_0)+\mathbb{E}\int_0^t \mathbf{1}_{\{s<\tau_R\}}L_sh(P_s)\,\mathrm{d}s. E h ( P t ∧ τ R ) = E h ( P 0 ) + E ∫ 0 t 1 { s < τ R } L s h ( P s ) d s .
The stopped stochastic integral has zero expectation because its integrand is bounded on the stopped region, and E G 0 < ∞ \mathbb{E}G_0<\infty E G 0 < ∞ . Using the upper bound on L s h L_sh L s h , we deduce
E h ( P t ∧ τ R ) ≤ E G 0 + ( 3 Λ T + μ ) t . \mathbb{E}h(P_{t\wedge\tau_R})\le\mathbb{E}G_0+(3\Lambda_T+\mu)t. E h ( P t ∧ τ R ) ≤ E G 0 + ( 3 Λ T + μ ) t .
Letting R → ∞ R\to\infty R → ∞ and using Fatou’s lemma gives
E h ( P t ) ≤ E G 0 + ( 3 Λ T + μ ) t , 0 ≤ t ≤ T . \mathbb{E}h(P_t)\le\mathbb{E}G_0+(3\Lambda_T+\mu)t,\qquad0\le t\le T. E h ( P t ) ≤ E G 0 + ( 3 Λ T + μ ) t , 0 ≤ t ≤ T .
Since the distribution of P t P_t P t has density G ϕ , a ( t , ⋅ ) G^{\phi,a}(t,\cdot) G ϕ , a ( t , ⋅ ) , we conclude that
E G ϕ , a ( t ) ≤ E G 0 + ( 3 Λ T + μ ) t , 0 ≤ t ≤ T . \mathbb{E}G^{\phi,a}(t)\le\mathbb{E}G_0+(3\Lambda_T+\mu)t,\qquad0\le t\le T. E G ϕ , a ( t ) ≤ E G 0 + ( 3 Λ T + μ ) t , 0 ≤ t ≤ T .
This proves (43).
The L 1 L^1 L 1 contraction property.
For 0 ≤ s ≤ t 0\le s\le t 0 ≤ s ≤ t , let P t s , q P_t^{s,q} P t s , q denote the solution of (32) at time t t t with deterministic initial condition P s s , q = q P_s^{s,q}=q P s s , q = q . Define the transition probability
K s , t ( q , A ) = P ( P t s , q ∈ A ) , K_{s,t}(q,A)=\mathbb{P}(P_t^{s,q}\in A), K s , t ( q , A ) = P ( P t s , q ∈ A ) ,
for every Borel set A ⊂ R 3 A\subset\mathbb{R}^3 A ⊂ R 3 . The stochastic flow constructed in Step 1 shows that, for every f ∈ L 1 ( R 3 ) f\in L^1(\mathbb{R}^3) f ∈ L 1 ( R 3 ) , the measure
A ⟼ ∫ R 3 K s , t ( q , A ) f ( q ) d q A\longmapsto\int_{\mathbb{R}^3}K_{s,t}(q,A)f(q)\,\mathrm{d}q A ⟼ ∫ R 3 K s , t ( q , A ) f ( q ) d q
is absolutely continuous with respect to Lebesgue measure. Thus, we may define the forward propagator U ( t , s ) U(t,s) U ( t , s ) by
∫ A [ U ( t , s ) f ] ( p ) d p = ∫ R 3 K s , t ( q , A ) f ( q ) d q . (47) \int_A[U(t,s)f](p)\,\mathrm{d}p=\int_{\mathbb{R}^3}K_{s,t}(q,A)f(q)\,\mathrm{d}q. \tag*{(47)} ∫ A [ U ( t , s ) f ] ( p ) d p = ∫ R 3 K s , t ( q , A ) f ( q ) d q . (47)
By the Markov property,
U ( t , r ) U ( r , s ) = U ( t , s ) , 0 ≤ s ≤ r ≤ t . U(t,r)U(r,s)=U(t,s),\qquad0\le s\le r\le t. U ( t , r ) U ( r , s ) = U ( t , s ) , 0 ≤ s ≤ r ≤ t .
In particular,
G ϕ , a ( t , ⋅ ) = U ( t , s ) G ϕ , a ( s , ⋅ ) . G^{\phi,a}(t,\cdot)=U(t,s)G^{\phi,a}(s,\cdot). G ϕ , a ( t , ⋅ ) = U ( t , s ) G ϕ , a ( s , ⋅ ) .
By positivity and linearity, ∣ U ( t , s ) f ∣ ≤ U ( t , s ) ∣ f ∣ |U(t,s)f|\le U(t,s)|f| ∣ U ( t , s ) f ∣ ≤ U ( t , s ) ∣ f ∣ . Moreover, conservation of total probability gives
∫ R 3 U ( t , s ) ∣ f ∣ d p = ∫ R 3 ∣ f ∣ d p . \int_{\mathbb{R}^{3}}U(t,s)|f|\,\mathrm{d}p=\int_{\mathbb{R}^{3}}|f|\,\mathrm{d}p. ∫ R 3 U ( t , s ) ∣ f ∣ d p = ∫ R 3 ∣ f ∣ d p .
Hence
∥ U ( t , s ) f ∥ L 1 ≤ ∥ f ∥ L 1 . \|U(t,s)f\|_{L^{1}}\le\|f\|_{L^{1}}. ∥ U ( t , s ) f ∥ L 1 ≤ ∥ f ∥ L 1 .
Applying this inequality to f = G 0 − H 0 f=G_{0}-H_{0} f = G 0 − H 0 proves (45) .
Strong continuity in L 1 L^{1} L 1
We prove that, for every f ∈ L 1 ( R 3 ) f\in L^{1}(\mathbb{R}^{3}) f ∈ L 1 ( R 3 ) , the map t ↦ U ( t , 0 ) f t\mapsto U(t,0)f t ↦ U ( t , 0 ) f is continuous from [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) into L 1 ( R 3 ) L^{1}(\mathbb{R}^{3}) L 1 ( R 3 ) . Let X s , t X_{s,t} X s , t be the stochastic flow of C 1 C^{1} C 1 diffeomorphisms constructed in Step 1. For f ∈ L 1 ( R 3 ) f\in L^{1}(\mathbb{R}^{3}) f ∈ L 1 ( R 3 ) , define
f 0 , t ( p ) = f ( X 0 , t − 1 ( p ) ) ∣ det ∇ X 0 , t − 1 ( p ) ∣ . (48) f_{0,t}(p)=f\left(X_{0,t}^{-1}(p)\right)\left|\det\nabla X_{0,t}^{-1}(p)\right|. \tag*{(48)} f 0 , t ( p ) = f ( X 0 , t − 1 ( p ) ) det ∇ X 0 , t − 1 ( p ) . (48)
By the change of variables formula, f 0 , t ( p ) d p f_{0,t}(p)\,\mathrm{d}p f 0 , t ( p ) d p is the pushforward of f ( p ) d p f(p)\,\mathrm{d}p f ( p ) d p under X 0 , t X_{0,t} X 0 , t . In particular,
∥ f 0 , t ∥ L 1 = ∥ f ∥ L 1 (49) \|f_{0,t}\|_{L^{1}}=\|f\|_{L^{1}} \tag*{(49)} ∥ f 0 , t ∥ L 1 = ∥ f ∥ L 1 (49)
almost surely. Moreover, the definition of the transition probabilities and Fubini’s theorem give
U ( t , 0 ) f = E [ f 0 , t ] (50) U(t,0)f=\mathbb{E}[f_{0,t}] \tag*{(50)} U ( t , 0 ) f = E [ f 0 , t ] (50)
as an equality in L 1 ( R 3 ) L^{1}(\mathbb{R}^{3}) L 1 ( R 3 ) .
Suppose first that f ∈ C c ∞ ( R 3 ) f\in C_{c}^{\infty}(\mathbb{R}^{3}) f ∈ C c ∞ ( R 3 ) . By the regularity theory of differentiable stochastic flows [5 Theorem 4.6.5] , we may choose a version of the flow such that, almost surely, X 0 , t X_{0,t} X 0 , t , its inverse, and their Jacobian matrices are continuous in ( t , p ) (t,p) ( t , p ) . Consequently, for almost every sample path of the Brownian motion,
f 0 , t ( p ) → f 0 , s ( p ) f_{0,t}(p)\to f_{0,s}(p) f 0 , t ( p ) → f 0 , s ( p )
locally uniformly in p p p as t → s t\to s t → s . Let K = supp f K=\operatorname{supp}f K = supp f . Since ( t , p ) ↦ X 0 , t ( p ) (t,p)\mapsto X_{0,t}(p) ( t , p ) ↦ X 0 , t ( p ) is continuous and K K K is compact, for almost every path of the Brownian motion, there exist ε > 0 \varepsilon>0 ε > 0 and R > 0 R>0 R > 0 such that X 0 , t ( K ) ⊂ B R X_{0,t}(K)\subset B_{R} X 0 , t ( K ) ⊂ B R whenever t ≥ 0 t\ge0 t ≥ 0 and ∣ t − s ∣ < ε |t-s|<\varepsilon ∣ t − s ∣ < ε . Thus, f 0 , t f_{0,t} f 0 , t and f 0 , s f_{0,s} f 0 , s are supported in the same bounded set B R B_{R} B R for all such t t t . It follows that
∥ f 0 , t − f 0 , s ∥ L 1 = ∫ B R ∣ f 0 , t ( p ) − f 0 , s ( p ) ∣ d p ≤ ∣ B R ∣ sup p ∈ B R ∣ f 0 , t ( p ) − f 0 , s ( p ) ∣ → 0 \|f_{0,t}-f_{0,s}\|_{L^{1}}=\int_{B_{R}}|f_{0,t}(p)-f_{0,s}(p)|\,\mathrm{d}p\le|B_{R}|\sup_{p\in B_{R}}|f_{0,t}(p)-f_{0,s}(p)|\to0 ∥ f 0 , t − f 0 , s ∥ L 1 = ∫ B R ∣ f 0 , t ( p ) − f 0 , s ( p ) ∣ d p ≤ ∣ B R ∣ p ∈ B R sup ∣ f 0 , t ( p ) − f 0 , s ( p ) ∣ → 0
almost surely as t → s t\to s t → s . Since (49) implies ∥ f 0 , t − f 0 , s ∥ L 1 ≤ 2 ∥ f ∥ L 1 \|f_{0,t}-f_{0,s}\|_{L^{1}}\le2\|f\|_{L^{1}} ∥ f 0 , t − f 0 , s ∥ L 1 ≤ 2∥ f ∥ L 1 , the dominated convergence theorem gives
E ∥ f 0 , t − f 0 , s ∥ L 1 → 0. \mathbb{E}\|f_{0,t}-f_{0,s}\|_{L^{1}}\to0. E ∥ f 0 , t − f 0 , s ∥ L 1 → 0.
Using (50) , we obtain
∥ U ( t , 0 ) f − U ( s , 0 ) f ∥ L 1 ≤ E ∥ f 0 , t − f 0 , s ∥ L 1 → 0. \lVert U(t,0)f-U(s,0)f\rVert_{L^1}\leq\mathbb{E}\lVert f_{0,t}-f_{0,s}\rVert_{L^1}\to0. ∥ U ( t , 0 ) f − U ( s , 0 ) f ∥ L 1 ≤ E ∥ f 0 , t − f 0 , s ∥ L 1 → 0.
Finally, the result extends to every f ∈ L 1 ( R 3 ) f\in L^1(\mathbb{R}^3) f ∈ L 1 ( R 3 ) by the density of C c ∞ ( R 3 ) C^\infty_c(\mathbb{R}^3) C c ∞ ( R 3 ) in L 1 ( R 3 ) L^1(\mathbb{R}^3) L 1 ( R 3 ) and the contraction property established in Step 3. In particular, taking f = G 0 f=G_0 f = G 0 proves
G ϕ , a ∈ C ( [ 0 , ∞ ) ; L 1 ( R 3 ) ) , G^{\phi,a}\in C([0,\infty);L^1(\mathbb{R}^3)), G ϕ , a ∈ C ([ 0 , ∞ ) ; L 1 ( R 3 )) ,
as required.
Uniqueness
The solution G = G ϕ , a G=G^{\phi,a} G = G ϕ , a constructed in Step 1 is the probability density of the solution P t P_t P t of (32), whose initial distribution has density G 0 G_0 G 0 . In particular, P t P_t P t solves the martingale problem associated with the generator (33) . We recall that a continuous stochastic process P t P_t P t solves the martingale problem associated with L t L_t L t if, for every f ∈ C c ∞ ( R 3 ) f\in C^\infty_c(\mathbb{R}^3) f ∈ C c ∞ ( R 3 ) , the process
M t f = f ( P t ) − f ( P 0 ) − ∫ 0 t L s f ( P s ) d s M_t^f=f(P_t)-f(P_0)-\int_0^t L_s f(P_s)\,\mathrm{d}s M t f = f ( P t ) − f ( P 0 ) − ∫ 0 t L s f ( P s ) d s
is a martingale with respect to the natural filtration of P t P_t P t . To prove uniqueness in Theorem 14, we use the superposition principle of [8 ] to show that any given weak solution G ~ \widetilde{G} G of (42) in the class stated in the theorem gives rise to a solution of the martingale problem.
Assume G ~ ( 0 , ⋅ ) = G 0 \widetilde{G}(0,\cdot)=G_0 G ( 0 , ⋅ ) = G 0 . Fix T > 0 T>0 T > 0 and set
ν ~ t ( d p ) = G ~ ( t , p ) d p . \widetilde{\nu}_t(\mathrm{d}p)=\widetilde{G}(t,p)\,\mathrm{d}p. ν t ( d p ) = G ( t , p ) d p .
Recall that the frozen equation can be written in the Fokker–Planck–Kolmogorov form
∂ t G ~ = ∑ i , j = 1 3 ∂ p i ∂ p j ( A i j G ~ ) − ∇ p ⋅ ( Q G ~ ) , \partial_t\widetilde{G}=\sum_{i,j=1}^{3}\partial_{p_i}\partial_{p_j}(A_{ij}\widetilde{G})-\nabla_p\cdot(Q\widetilde{G}), ∂ t G = i , j = 1 ∑ 3 ∂ p i ∂ p j ( A ij G ) − ∇ p ⋅ ( Q G ) ,
where
A ( t , p ) = ϕ ( t ) D ( p ) , Q ( t , p ) = Q ϕ ( t ) , a ( t ) ( p ) . A(t,p)=\phi(t)D(p),\qquad Q(t,p)=Q_{\phi(t),a(t)}(p). A ( t , p ) = ϕ ( t ) D ( p ) , Q ( t , p ) = Q ϕ ( t ) , a ( t ) ( p ) .
By Lemma 13,
∥ A ( t , p ) ∥ ≤ C T ⟨ p ⟩ , ∣ Q ( t , p ) ∣ ≤ C T . \lVert A(t,p)\rVert\leq C_T\langle p\rangle,\qquad|Q(t,p)|\leq C_T. ∥ A ( t , p )∥ ≤ C T ⟨ p ⟩ , ∣ Q ( t , p ) ∣ ≤ C T .
Since G ~ \widetilde{G} G has unit mass and locally bounded energy,
∫ 0 T ∫ R 3 ( ∥ A ( t , p ) ∥ + ∣ Q ( t , p ) ∣ ) d ν ~ t ( p ) d t < ∞ . \int_0^T\int_{\mathbb{R}^3}\bigl(\lVert A(t,p)\rVert+|Q(t,p)|\bigr)\,\mathrm{d}\widetilde{\nu}_t(p)\,\mathrm{d}t<\infty. ∫ 0 T ∫ R 3 ( ∥ A ( t , p )∥ + ∣ Q ( t , p ) ∣ ) d ν t ( p ) d t < ∞.
Moreover, the L 1 L^1 L 1 continuity of G ~ \widetilde{G} G implies that t ↦ ν ~ t t\mapsto\widetilde{\nu}_t t ↦ ν t is narrowly continuous. We may therefore apply the superposition principle [8 Theorem 2.5] . It follows that there exists a continuous stochastic process P ~ t \widetilde{P}_t P t which solves the martingale problem associated with the generator
L t f = Q ⋅ ∇ p f + A : ∇ p 2 f , L_t f = Q \cdot\nabla_p f + A : \nabla_p^2 f, L t f = Q ⋅ ∇ p f + A : ∇ p 2 f ,
and whose time marginal distributions are ν ~ t \widetilde{\nu}_t ν t , 0 ≤ t ≤ T 0 \le t \le T 0 ≤ t ≤ T . Since the coefficients of (32) are globally Lipschitz in p p p and satisfy the linear growth condition, the associated martingale problem is well posed. Since P ~ 0 \widetilde{P}_0 P 0 and P 0 P_0 P 0 have the same distribution G 0 ( p ) d p G_0(p)\,dp G 0 ( p ) d p , the processes P ~ t \widetilde{P}_t P t and P t P_t P t have also the same distribution. Consequently,
ν ~ t ( d p ) = G ( t , p ) d p , 0 ≤ t ≤ T . \widetilde{\nu}_t(dp) = G(t,p)\,dp, \qquad0 \le t \le T. ν t ( d p ) = G ( t , p ) d p , 0 ≤ t ≤ T .
Since T > 0 T>0 T > 0 is arbitrary, uniqueness follows.
The following Wasserstein estimates will be used in the proof of local existence for the nonlinear equation. The first estimate controls the time continuity of the nonlinear functionals ( Φ G , j G ) (\Phi_G,j_G) ( Φ G , j G ) , while the second provides the stability with respect to the frozen coefficients needed to establish contraction of the fixed-point map.
Lemma 15 (Frozen Wasserstein estimates). Let T > 0 T>0 T > 0 and assume that ϕ , ψ ∈ C ( [ 0 , T ] ) \phi,\psi\in C([0,T]) ϕ , ψ ∈ C ([ 0 , T ]) and a , b ∈ C ( [ 0 , T ] ; R 3 ) a,b\in C([0,T];\mathbb{R}^3) a , b ∈ C ([ 0 , T ] ; R 3 ) satisfy
0 < δ ≤ ϕ , ψ ≤ Λ , ∣ a ∣ , ∣ b ∣ ≤ γ < 1. 0<\delta\le\phi,\psi\le\Lambda,\qquad|a|,|b|\le\gamma<1. 0 < δ ≤ ϕ , ψ ≤ Λ , ∣ a ∣ , ∣ b ∣ ≤ γ < 1. $Let G 0 ∈ L 1 ( R 3 ) G_0\in L^1(\mathbb{R}^3) G 0 ∈ L 1 ( R 3 ) be a nonnegative probability density with E G 0 < ∞ E G_0<\infty E G 0 < ∞ , and denote by G ϕ , a G^{\phi,a} G ϕ , a and G ψ , b G^{\psi,b} G ψ , b the corresponding solutions of the frozen Fokker–Planck equation (30) . There exists a constant C T C_T C T , depending only on T T T , δ , Λ , γ , μ \delta,\Lambda,\gamma,\mu δ , Λ , γ , μ and E G 0 E G_0 E G 0 , such that
W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ C T ∣ t − s ∣ 1 / 2 , 0 ≤ s ≤ t ≤ T , (51) W_1\left(G^{\phi,a}(t),G^{\phi,a}(s)\right)\le C_T|t-s|^{1/2},\qquad0\le s\le t\le T, \tag*{(51)} W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ C T ∣ t − s ∣ 1/2 , 0 ≤ s ≤ t ≤ T , (51) and
W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C T t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ , 0 ≤ t ≤ T . (52) W_1\left(G^{\phi,a}(t),G^{\psi,b}(t)\right)\le C_T\sqrt{t}\left\|(\phi,a)-(\psi,b)\right\|_\infty,\qquad0\le t\le T. \tag*{(52)} W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C T t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ , 0 ≤ t ≤ T . (52)
Proof. We use the stochastic representation established in Theorem 14 . Throughout the proof we use the coefficient estimates of Lemma 13 , with constants depending only on δ , Λ , γ \delta,\Lambda,\gamma δ , Λ , γ and μ \mu μ .
Let P t P_t P t be the solution of (32) corresponding to ( ϕ , a ) (\phi,a) ( ϕ , a ) , with initial data P 0 P_0 P 0 having probability distribution ν 0 = G 0 ( p ) d p \nu_0=G_0(p)\,dp ν 0 = G 0 ( p ) d p . The energy estimate in Theorem 14 gives
sup 0 ≤ t ≤ T E ⟨ P t ⟩ ≤ K T , \sup_{0\le t\le T} E\langle P_t\rangle\le K_T, 0 ≤ t ≤ T sup E ⟨ P t ⟩ ≤ K T ,
where
K T = E G 0 + ( 3 Λ + μ ) T . K_T=E G_0+(3\Lambda+\mu)T. K T = E G 0 + ( 3Λ + μ ) T .
For 0 ≤ s ≤ t ≤ T 0\le s\le t\le T 0 ≤ s ≤ t ≤ T , the SDE (32) gives
P t − P s = ∫ s t Q ϕ ( r ) , a ( r ) ( P r ) d r + ∫ s t σ ϕ ( r ) ( P r ) d B r . (53) P_t-P_s=\int_s^t Q_{\phi(r),a(r)}(P_r)\,dr+\int_s^t \sigma_{\phi(r)}(P_r)\,dB_r. \tag*{(53)} P t − P s = ∫ s t Q ϕ ( r ) , a ( r ) ( P r ) d r + ∫ s t σ ϕ ( r ) ( P r ) d B r . (53)
By the Cauchy–Schwarz inequality and Itô’s isometry,
E ∣ P t − P s ∣ 2 ≤ 2 ( t − s ) ∫ s t E ∣ Q ϕ ( r ) , a ( r ) ( P r ) ∣ 2 d r + 2 ∫ s t E ∥ σ ϕ ( r ) ( P r ) ∥ H S 2 d r ≤ C T ∣ t − s ∣ . \mathbb{E}|P_t-P_s|^2 \le2(t-s)\int_s^t \mathbb{E}|Q_{\phi(r),a(r)}(P_r)|^2\,\mathrm{d}r+2\int_s^t \mathbb{E}\|\sigma_{\phi(r)}(P_r)\|_{\mathrm{HS}}^2\,\mathrm{d}r\le C_T|t-s|. E ∣ P t − P s ∣ 2 ≤ 2 ( t − s ) ∫ s t E ∣ Q ϕ ( r ) , a ( r ) ( P r ) ∣ 2 d r + 2 ∫ s t E ∥ σ ϕ ( r ) ( P r ) ∥ HS 2 d r ≤ C T ∣ t − s ∣.
This estimate does not require E ∣ P 0 ∣ 2 < ∞ \mathbb{E}|P_0|^2<\infty E ∣ P 0 ∣ 2 < ∞ , since it concerns only the increment P t − P s P_t-P_s P t − P s . Since the joint distribution of ( P t , P s ) (P_t,P_s) ( P t , P s ) is a coupling of G ϕ , a ( t ) G^{\phi,a}(t) G ϕ , a ( t ) and G ϕ , a ( s ) G^{\phi,a}(s) G ϕ , a ( s ) ,
W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ E ∣ P t − P s ∣ ≤ ( E ∣ P t − P s ∣ 2 ) 1 / 2 ≤ C T ∣ t − s ∣ 1 / 2 . W_1\left(G^{\phi,a}(t),G^{\phi,a}(s)\right)\le\mathbb{E}|P_t-P_s|\le\left(\mathbb{E}|P_t-P_s|^2\right)^{1/2}\le C_T|t-s|^{1/2}. W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ E ∣ P t − P s ∣ ≤ ( E ∣ P t − P s ∣ 2 ) 1/2 ≤ C T ∣ t − s ∣ 1/2 .
This proves (51). We next prove (52). Let P t P_t P t and P ~ t \widetilde{P}_t P t be the solutions corresponding to ( ϕ , a ) (\phi,a) ( ϕ , a ) and ( ψ , b ) (\psi,b) ( ψ , b ) , respectively. We construct them with the same initial random variable, having distribution G 0 ( p ) d p G_0(p)\,\mathrm{d}p G 0 ( p ) d p , and driven by the same Brownian motion. Set
Z t = P t − P ~ t , Z 0 = 0 , Z_t=P_t-\widetilde{P}_t,\qquad Z_0=0, Z t = P t − P t , Z 0 = 0 ,
and
η = ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ . \eta=\|(\phi,a)-(\psi,b)\|_\infty. η = ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ .
For R > 0 R>0 R > 0 , define the stopping time
τ R = inf { t ≥ 0 : ∣ P t ∣ + ∣ P ~ t ∣ ≥ R } . \tau_R=\inf\left\{t\ge0:|P_t|+|\widetilde{P}_t|\ge R\right\}. τ R = inf { t ≥ 0 : ∣ P t ∣ + ∣ P t ∣ ≥ R } .
Applying Itô’s formula to ∣ Z t ∧ τ R ∣ 2 |Z_{t\wedge\tau_R}|^2 ∣ Z t ∧ τ R ∣ 2 , taking expectations, and using the coefficient estimates above together with Young’s inequality, we obtain
E ∣ Z t ∧ τ R ∣ 2 ≤ C ∫ 0 t E ∣ Z s ∧ τ R ∣ 2 d s + C η 2 ∫ 0 t ( 1 + E ⟨ P ~ s ⟩ ) d s . \mathbb{E}|Z_{t\wedge\tau_R}|^2\le C\int_0^t\mathbb{E}|Z_{s\wedge\tau_R}|^2\,\mathrm{d}s+C\eta^2\int_0^t\left(1+\mathbb{E}\langle\widetilde{P}_s\rangle\right)\,\mathrm{d}s. E ∣ Z t ∧ τ R ∣ 2 ≤ C ∫ 0 t E ∣ Z s ∧ τ R ∣ 2 d s + C η 2 ∫ 0 t ( 1 + E ⟨ P s ⟩ ) d s .
The energy estimate gives
sup 0 ≤ s ≤ T E ⟨ P ~ s ⟩ ≤ K T . \sup_{0\le s\le T}\mathbb{E}\langle\widetilde{P}_s\rangle\le K_T. 0 ≤ s ≤ T sup E ⟨ P s ⟩ ≤ K T .
Consequently,
E ∣ Z t ∧ τ R ∣ 2 ≤ C ∫ 0 t E ∣ Z s ∧ τ R ∣ 2 d s + C T t η 2 . \mathbb{E}|Z_{t\wedge\tau_R}|^2\le C\int_0^t\mathbb{E}|Z_{s\wedge\tau_R}|^2\,\mathrm{d}s+C_Tt\eta^2. E ∣ Z t ∧ τ R ∣ 2 ≤ C ∫ 0 t E ∣ Z s ∧ τ R ∣ 2 d s + C T t η 2 .
By Gronwall’s inequality,
E ∣ Z t ∧ τ R ∣ 2 ≤ C T t η 2 . \mathbb{E}|Z_{t\wedge\tau_R}|^2\le C_Tt\eta^2. E ∣ Z t ∧ τ R ∣ 2 ≤ C T t η 2 .
Since both SDE solutions are global and nonexplosive, τ R → ∞ \tau_R\to\infty τ R → ∞ almost surely as R → ∞ R\to\infty R → ∞ . Fatou’s lemma therefore yields
E ∣ Z t ∣ 2 ≤ C T t η 2 . \mathbb{E}|Z_t|^2\le C_Tt\eta^2. E ∣ Z t ∣ 2 ≤ C T t η 2 .
Finally, the joint distribution of ( P t , P ~ t ) (P_t,\widetilde{P}_t) ( P t , P t ) is a coupling of the two frozen solutions. Hence
W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ E ∣ P t − P ~ t ∣ ≤ ( E ∣ Z t ∣ 2 ) 1 / 2 ≤ C T t η . W_1\left(G^{\phi,a}(t),G^{\psi,b}(t)\right)\le\mathbb{E}|P_t-\widetilde{P}_t|\le\left(\mathbb{E}|Z_t|^2\right)^{1/2}\le C_T\sqrt{t}\,\eta. W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ E ∣ P t − P t ∣ ≤ ( E ∣ Z t ∣ 2 ) 1/2 ≤ C T t η .
This proves (52).
Proof of Theorem 1
We are now ready to prove the first main theorem of this paper. We split the proof into several subsections.
Local existence
We first use the frozen problem studied in § 3 to construct a local solution of the nonlinear equation. Recall that
j G = ∫ R 3 v ( p ) G ( p ) d p j_G = \int_{\mathbb{R}^3} v(p)G(p)\,\mathrm{d}p j G = ∫ R 3 v ( p ) G ( p ) d p
and
Φ G = μ 3 [ 1 − ∫ R 3 G ( p ) ⟨ p ⟩ β G ( p ) d p ] , β G ( p ) = ⟨ p ⟩ ( 1 − j G ⋅ v ( p ) ) . \Phi_G = \frac{\mu}{3}\left[1-\int_{\mathbb{R}^3}\frac{G(p)}{\langle p\rangle\beta_G(p)}\,\mathrm{d}p\right], \qquad\beta_G(p)=\langle p\rangle\left(1-j_G\cdot v(p)\right). Φ G = 3 μ [ 1 − ∫ R 3 ⟨ p ⟩ β G ( p ) G ( p ) d p ] , β G ( p ) = ⟨ p ⟩ ( 1 − j G ⋅ v ( p ) ) .
Let G 0 G_0 G 0 satisfy
G 0 ∈ L 1 ( R 3 ) , G 0 ≥ 0 , ∫ R 3 G 0 d p = 1 , E G 0 < ∞ . G_0\in L^1(\mathbb{R}^3), \qquad G_0\geq0, \qquad\int_{\mathbb{R}^3}G_0\,\mathrm{d}p=1, \qquad EG_0<\infty. G 0 ∈ L 1 ( R 3 ) , G 0 ≥ 0 , ∫ R 3 G 0 d p = 1 , E G 0 < ∞.
Set
a 0 = j G 0 , ϕ 0 = Φ G 0 . a_0=j_{G_0}, \qquad\phi_0=\Phi_{G_0}. a 0 = j G 0 , ϕ 0 = Φ G 0 .
By Lemmas 6 and 7,
∣ a 0 ∣ < 1 , 0 < ϕ 0 ≤ μ 3 . |a_0|<1, \qquad0<\phi_0\leq\frac{\mu}{3}. ∣ a 0 ∣ < 1 , 0 < ϕ 0 ≤ 3 μ .
We define
δ = ϕ 0 2 , γ = 1 + ∣ a 0 ∣ 2 , Λ = μ 3 . (54) \delta=\frac{\phi_0}{2}, \qquad\gamma=\frac{1+|a_0|}{2}, \qquad\Lambda=\frac{\mu}{3}. \tag*{(54)} δ = 2 ϕ 0 , γ = 2 1 + ∣ a 0 ∣ , Λ = 3 μ . (54)
Thus
0 < δ < ϕ 0 ≤ Λ , ∣ a 0 ∣ < γ < 1. 0<\delta<\phi_0\leq\Lambda, \qquad|a_0|<\gamma<1. 0 < δ < ϕ 0 ≤ Λ , ∣ a 0 ∣ < γ < 1.
For 0 < T ≤ 1 0<T\leq1 0 < T ≤ 1 , let
K T = { ( ϕ , a ) ∈ C ( [ 0 , T ] ; R × R 3 ) : δ ≤ ϕ ( t ) ≤ Λ , ∣ a ( t ) ∣ ≤ γ , ϕ ( 0 ) = ϕ 0 , a ( 0 ) = a 0 } . (55) K_T=\left\{(\phi,a)\in C([0,T];\mathbb{R}\times\mathbb{R}^3):\delta\leq\phi(t)\leq\Lambda,\quad|a(t)|\leq\gamma,\quad\phi(0)=\phi_0,\quad a(0)=a_0\right\}. \tag*{(55)} K T = { ( ϕ , a ) ∈ C ([ 0 , T ] ; R × R 3 ) : δ ≤ ϕ ( t ) ≤ Λ , ∣ a ( t ) ∣ ≤ γ , ϕ ( 0 ) = ϕ 0 , a ( 0 ) = a 0 } . (55)
We equip K T K_T K T with the norm
∥ ( ϕ , a ) ∥ ∞ = sup 0 ≤ t ≤ T ( ∣ ϕ ( t ) ∣ + ∣ a ( t ) ∣ ) . \|(\phi,a)\|_\infty=\sup_{0\leq t\leq T}\left(|\phi(t)|+|a(t)|\right). ∥ ( ϕ , a ) ∥ ∞ = 0 ≤ t ≤ T sup ( ∣ ϕ ( t ) ∣ + ∣ a ( t ) ∣ ) .
Then K T K_T K T is a nonempty closed subset of the Banach space C ( [ 0 , T ] ; R × R 3 ) C([0,T];\mathbb{R}\times\mathbb{R}^3) C ([ 0 , T ] ; R × R 3 ) and is therefore complete.
For every ( ϕ , a ) ∈ K T (\phi,a) \in\mathcal{K}_T ( ϕ , a ) ∈ K T , let G ϕ , a G^{\phi,a} G ϕ , a denote the unique solution on [ 0 , T ] [0,T] [ 0 , T ] of the frozen equation
∂ t G ϕ , a = ϕ ( t ) ∇ p ⋅ ( D ∇ p G ϕ , a ) + ∇ p ⋅ ( V a G ϕ , a ) , G ϕ , a ( 0 ) = G 0 . (56) \partial_t G^{\phi,a}=\phi(t)\nabla_p\cdot\left(D\nabla_pG^{\phi,a}\right)+\nabla_p\cdot\left(V_aG^{\phi,a}\right),\qquad G^{\phi,a}(0)=G_0. \tag*{(56)} ∂ t G ϕ , a = ϕ ( t ) ∇ p ⋅ ( D ∇ p G ϕ , a ) + ∇ p ⋅ ( V a G ϕ , a ) , G ϕ , a ( 0 ) = G 0 . (56)
By Theorem 14,
G ϕ , a ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) , G ϕ , a ≥ 0 , ∫ R 3 G ϕ , a ( t , p ) d p = 1. G^{\phi,a}\in C([0,T];L^1(\mathbb{R}^3)),\qquad G^{\phi,a}\geq0,\qquad\int_{\mathbb{R}^3}G^{\phi,a}(t,p)\,\mathrm{d}p=1. G ϕ , a ∈ C ([ 0 , T ] ; L 1 ( R 3 )) , G ϕ , a ≥ 0 , ∫ R 3 G ϕ , a ( t , p ) d p = 1.
We define
F ( ϕ , a ) ( t ) = ( Φ G ϕ , a ( t ) , j G ϕ , a ( t ) ) . (57) \mathcal{F}(\phi,a)(t)=\left(\Phi_{G^{\phi,a}(t)},j_{G^{\phi,a}(t)}\right). \tag*{(57)} F ( ϕ , a ) ( t ) = ( Φ G ϕ , a ( t ) , j G ϕ , a ( t ) ) . (57)
We first record estimates whose constants can be chosen independently of T ≤ 1 T\leq1 T ≤ 1 . Indeed, the constants in Theorem 14 and Lemma 15 depend only on the upper bound for the time interval and on the parameters δ , Λ , γ , μ , E G 0 \delta,\Lambda,\gamma,\mu, E_{G_0} δ , Λ , γ , μ , E G 0 . Extend ϕ \phi ϕ and a a a constantly from T T T to 1. The extended coefficients satisfy the same bounds δ ≤ ϕ ≤ Λ \delta\leq\phi\leq\Lambda δ ≤ ϕ ≤ Λ and ∣ a ∣ ≤ γ |a|\leq\gamma ∣ a ∣ ≤ γ . Hence, applying Theorem 14 and Lemma 15 with upper time horizon 1, there are constants K K K and C 0 C_0 C 0 , independent of T ≤ 1 T\leq1 T ≤ 1 and of ( ϕ , a ) ∈ K T (\phi,a)\in\mathcal{K}_T ( ϕ , a ) ∈ K T , such that
sup 0 ≤ t ≤ T E G ϕ , a ( t ) ≤ K , (58) \sup_{0\leq t\leq T}E_{G^{\phi,a}(t)}\leq K, \tag*{(58)} 0 ≤ t ≤ T sup E G ϕ , a ( t ) ≤ K , (58)
where one may take
K = E G 0 + 2 μ , K=E_{G_0}+2\mu, K = E G 0 + 2 μ ,
and
W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ C 0 ∣ t − s ∣ 1 / 2 . (59) W_1\left(G^{\phi,a}(t),G^{\phi,a}(s)\right)\leq C_0|t-s|^{1/2}. \tag*{(59)} W 1 ( G ϕ , a ( t ) , G ϕ , a ( s ) ) ≤ C 0 ∣ t − s ∣ 1/2 . (59)
Moreover, for ( ϕ , a ) , ( ψ , b ) ∈ K T (\phi,a),(\psi,b)\in\mathcal{K}_T ( ϕ , a ) , ( ψ , b ) ∈ K T ,
W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C 0 t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ . (60) W_1\left(G^{\phi,a}(t),G^{\psi,b}(t)\right)\leq C_0\sqrt{t}\,\|(\phi,a)-(\psi,b)\|_\infty. \tag*{(60)} W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C 0 t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ . (60)
By Lemma 6 and (57),
∣ j G ϕ , a ( t ) ∣ ≤ 1 − 1 E G ϕ , a ( t ) 2 ≤ 1 − 1 K 2 = : γ ∗ < 1. |j_{G^{\phi,a}(t)}|\leq\sqrt{1-\frac{1}{E_{G^{\phi,a}(t)}^2}}\leq\sqrt{1-\frac{1}{K^2}}=:\gamma_*<1. ∣ j G ϕ , a ( t ) ∣ ≤ 1 − E G ϕ , a ( t ) 2 1 ≤ 1 − K 2 1 =: γ ∗ < 1.
Consequently, all the probability densities occurring in the fixed point argument satisfy a common strict bound
∣ j G ∣ ≤ γ ∗ < 1. |j_G|\leq\gamma_*<1. ∣ j G ∣ ≤ γ ∗ < 1.
By Lemma 12, there exists a constant C 1 C_1 C 1 , depending only on γ ∗ \gamma_* γ ∗ and μ \mu μ , such that
∣ Φ G − Φ H ∣ + ∣ j G − j H ∣ ≤ C 1 W 1 ( G , H ) (61) |\Phi_G-\Phi_H|+|j_G-j_H|\leq C_1W_1(G,H) \tag*{(61)} ∣ Φ G − Φ H ∣ + ∣ j G − j H ∣ ≤ C 1 W 1 ( G , H ) (61)
for all probability densities G , H G,H G , H arising in the construction.
We can now prove local existence.
Theorem 16 (Local existence). Let
G 0 ∈ L 1 ( R 3 ) , G 0 ≥ 0 a.e. , ∫ R 3 G 0 d p = 1 , E G 0 < ∞ . G_0 \in L^1(\mathbb{R}^3), \qquad G_0 \ge0\ \text{a.e.}, \qquad\int_{\mathbb{R}^3} G_0\,\mathrm{d}p = 1, \qquad E_{G_0} < \infty. G 0 ∈ L 1 ( R 3 ) , G 0 ≥ 0 a.e. , ∫ R 3 G 0 d p = 1 , E G 0 < ∞. Then there exists T 0 > 0 T_0 > 0 T 0 > 0 such that the Cauchy problem for (1) with initial data G 0 G_0 G 0 has a weak solution on [ 0 , T 0 ] [0,T_0] [ 0 , T 0 ] . Moreover,
G ∈ C ( [ 0 , T 0 ] ; L 1 ( R 3 ) ) , G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , G \in C([0,T_0];L^1(\mathbb{R}^3)), \qquad G(t,\cdot) \ge0\ \text{a.e.}, \qquad\int_{\mathbb{R}^3} G(t,p)\,\mathrm{d}p = 1, G ∈ C ([ 0 , T 0 ] ; L 1 ( R 3 )) , G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 , and
sup 0 ≤ t ≤ T 0 E G ( t ) < ∞ . \sup_{0\le t\le T_0} E_{G(t)} < \infty. 0 ≤ t ≤ T 0 sup E G ( t ) < ∞. In addition, if, for some constants
0 < δ 0 ≤ μ 3 , 0 ≤ γ 0 < 1 , 1 ≤ K 0 < ∞ , 0 < \delta_0 \le\frac{\mu}{3}, \qquad0 \le\gamma_0 < 1, \qquad1 \le K_0 < \infty, 0 < δ 0 ≤ 3 μ , 0 ≤ γ 0 < 1 , 1 ≤ K 0 < ∞ , there holds
Φ G 0 ≥ δ 0 , ∣ j G 0 ∣ ≤ γ 0 , E G 0 ≤ K 0 , \Phi_{G_0} \ge\delta_0, \qquad|j_{G_0}| \le\gamma_0, \qquad E_{G_0} \le K_0, Φ G 0 ≥ δ 0 , ∣ j G 0 ∣ ≤ γ 0 , E G 0 ≤ K 0 , then T 0 T_0 T 0 may be chosen depending only on δ 0 \delta_0 δ 0 , γ 0 \gamma_0 γ 0 , K 0 K_0 K 0 and μ \mu μ .
Proof. We claim that, for T > 0 T>0 T > 0 sufficiently small, the map F \mathcal{F} F defined by (56) is a contraction of K T \mathcal{K}_T K T into itself. We first prove that F \mathcal{F} F maps K T \mathcal{K}_T K T into itself. Since G ϕ , a ( 0 ) = G 0 G^{\phi,a}(0)=G_0 G ϕ , a ( 0 ) = G 0 , we have
F ( ϕ , a ) ( 0 ) = ( ϕ 0 , a 0 ) . \mathcal{F}(\phi,a)(0)=(\phi_0,a_0). F ( ϕ , a ) ( 0 ) = ( ϕ 0 , a 0 ) .
Moreover, by (60) and (58),
∣ Φ G ϕ , a ( t ) − ϕ 0 ∣ + ∣ j G ϕ , a ( t ) − a 0 ∣ ≤ C 1 W 1 ( G ϕ , a ( t ) , G 0 ) ≤ C 1 C 0 t . (62) \left|\Phi_{G^{\phi,a}(t)}-\phi_0\right|+\left|j_{G^{\phi,a}(t)}-a_0\right| \le C_1W_1\left(G^{\phi,a}(t),G_0\right) \le C_1C_0\sqrt{t}. \tag*{(62)} Φ G ϕ , a ( t ) − ϕ 0 + j G ϕ , a ( t ) − a 0 ≤ C 1 W 1 ( G ϕ , a ( t ) , G 0 ) ≤ C 1 C 0 t . (62)
In particular,
Φ G ϕ , a ( t ) ≥ ϕ 0 − C 1 C 0 T . \Phi_{G^{\phi,a}(t)} \ge\phi_0-C_1C_0\sqrt{T}. Φ G ϕ , a ( t ) ≥ ϕ 0 − C 1 C 0 T .
Choose T T T so small that
C 1 C 0 T ≤ min { ϕ 0 2 , γ − ∣ a 0 ∣ } . (63) C_1C_0\sqrt{T}\le\min\left\{\frac{\phi_0}{2},\gamma-|a_0|\right\}. \tag*{(63)} C 1 C 0 T ≤ min { 2 ϕ 0 , γ − ∣ a 0 ∣ } . (63)
Then
Φ G ϕ , a ( t ) ≥ ϕ 0 2 = δ \Phi_{G^{\phi,a}(t)} \ge\frac{\phi_0}{2}=\delta Φ G ϕ , a ( t ) ≥ 2 ϕ 0 = δ
and
∣ j G ϕ , a ( t ) ∣ ≤ ∣ a 0 ∣ + C 1 C 0 T ≤ γ . |j_{G^{\phi,a}(t)}| \le|a_0|+C_1C_0\sqrt{T}\le\gamma. ∣ j G ϕ , a ( t ) ∣ ≤ ∣ a 0 ∣ + C 1 C 0 T ≤ γ .
On the other hand, Lemma 7 gives
Φ G ϕ , a ( t ) ≤ μ 3 = Λ . \Phi_{G^{\phi,a}(t)} \le\frac{\mu}{3}=\Lambda. Φ G ϕ , a ( t ) ≤ 3 μ = Λ.
Thus, the pointwise bounds defining K T \mathcal{K}_T K T are satisfied. Moreover, again by (61) and (59) ,
∣ Φ G ϕ , a ( t ) − Φ G ϕ , a ( s ) ∣ + ∣ j ˙ G ϕ , a ( t ) − j ˙ G ϕ , a ( s ) ∣ ≤ C 1 C 0 ∣ t − s ∣ 1 / 2 . \left|\Phi_{G^{\phi,a}}(t)-\Phi_{G^{\phi,a}}(s)\right|+\left|\dot{j}_{G^{\phi,a}}(t)-\dot{j}_{G^{\phi,a}}(s)\right|\le C_1C_0|t-s|^{1/2}. ∣ Φ G ϕ , a ( t ) − Φ G ϕ , a ( s ) ∣ + j ˙ G ϕ , a ( t ) − j ˙ G ϕ , a ( s ) ≤ C 1 C 0 ∣ t − s ∣ 1/2 .
Hence, the functions defining the image of F \mathcal{F} F are continuous and therefore F ( ϕ , a ) ∈ K T \mathcal{F}(\phi,a)\in\mathcal{K}_T F ( ϕ , a ) ∈ K T . We next prove that F \mathcal{F} F is a contraction. Let
( ϕ , a ) , ( ψ , b ) ∈ K T . (\phi,a),(\psi,b)\in\mathcal{K}_T. ( ϕ , a ) , ( ψ , b ) ∈ K T .
By (61) and (60) ,
∣ Φ G ϕ , a ( t ) − Φ G ψ , b ( t ) ∣ + ∣ j ˙ G ϕ , a ( t ) − j ˙ G ψ , b ( t ) ∣ ≤ C 1 W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C 1 C 0 t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ . \begin{aligned}
\left|\Phi_{G^{\phi,a}}(t)-\Phi_{G^{\psi,b}}(t)\right|+\left|\dot{j}_{G^{\phi,a}}(t)-\dot{j}_{G^{\psi,b}}(t)\right|\le C_1W_1\left(G^{\phi,a}(t),G^{\psi,b}(t)\right) \\
&\le C_1C_0\sqrt{t}\left\|(\phi,a)-(\psi,b)\right\|_\infty.
\end{aligned} ∣ Φ G ϕ , a ( t ) − Φ G ψ , b ( t ) ∣ + j ˙ G ϕ , a ( t ) − j ˙ G ψ , b ( t ) ≤ C 1 W 1 ( G ϕ , a ( t ) , G ψ , b ( t ) ) ≤ C 1 C 0 t ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ .
Taking the supremum over 0 ≤ t ≤ T 0\le t\le T 0 ≤ t ≤ T gives
∥ F ( ϕ , a ) − F ( ψ , b ) ∥ ∞ ≤ C 1 C 0 T ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ . \left\|\mathcal{F}(\phi,a)-\mathcal{F}(\psi,b)\right\|_\infty\le C_1C_0\sqrt{T}\left\|(\phi,a)-(\psi,b)\right\|_\infty. ∥ F ( ϕ , a ) − F ( ψ , b ) ∥ ∞ ≤ C 1 C 0 T ∥ ( ϕ , a ) − ( ψ , b ) ∥ ∞ .
Shrinking T T T further, if necessary, so that
C 1 C 0 T < 1 , (64) C_1C_0\sqrt{T}<1, \tag*{(64)} C 1 C 0 T < 1 , (64)
the map F \mathcal{F} F is a strict contraction on the complete metric space K T \mathcal{K}_T K T . Banach’s fixed-point theorem therefore gives a unique ( ϕ , a ) ∈ K T (\phi,a)\in\mathcal{K}_T ( ϕ , a ) ∈ K T such that
F ( ϕ , a ) = ( ϕ , a ) . \mathcal{F}(\phi,a)=(\phi,a). F ( ϕ , a ) = ( ϕ , a ) .
If G = G ϕ , a G=G^{\phi,a} G = G ϕ , a , then, by the definition of F \mathcal{F} F ,
ϕ ( t ) = Φ G ( t ) , a ( t ) = j G ( t ) . \phi(t)=\Phi_{G}(t),\qquad a(t)=j_G(t). ϕ ( t ) = Φ G ( t ) , a ( t ) = j G ( t ) .
Consequently (56) becomes
∂ t G = Φ G ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V j G G ) , \partial_tG=\Phi_G\nabla_p\cdot(D\nabla_pG)+\nabla_p\cdot(V_{j_G}G), ∂ t G = Φ G ∇ p ⋅ ( D ∇ p G ) + ∇ p ⋅ ( V j G G ) ,
which is precisely the nonlinear equation (1) . The properties
G ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) , G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1 G\in C([0,T];L^1(\mathbb{R}^3)),\qquad G(t,\cdot)\ge0\ \text{a.e.},\qquad\int_{\mathbb{R}^3}G(t,p)\,dp=1 G ∈ C ([ 0 , T ] ; L 1 ( R 3 )) , G ( t , ⋅ ) ≥ 0 a.e. , ∫ R 3 G ( t , p ) d p = 1
follow from Theorem 14 , while
sup 0 ≤ t ≤ T E G ( t ) ≤ K \sup_{0\le t\le T}E_{G(t)}\le K 0 ≤ t ≤ T sup E G ( t ) ≤ K
follows from (58) . This proves local existence. Finally, suppose that
ϕ 0 = Φ G 0 ≥ δ 0 , ∣ a 0 ∣ = ∣ j G 0 ∣ ≤ γ 0 < 1 , E G 0 ≤ K 0 . \phi_0=\Phi_{G_0}\ge\delta_0,\qquad|a_0|=|j_{G_0}|\le\gamma_0<1,\qquad E_{G_0}\le K_0. ϕ 0 = Φ G 0 ≥ δ 0 , ∣ a 0 ∣ = ∣ j G 0 ∣ ≤ γ 0 < 1 , E G 0 ≤ K 0 .
Then, by (54) ,
δ = ϕ 0 2 ≥ δ 0 2 , γ − ∣ a 0 ∣ = 1 − ∣ a 0 ∣ 2 ≥ 1 − γ 0 2 . \delta= \frac{\phi_0}{2} \ge\frac{\delta_0}{2}, \qquad\gamma- |a_0| = \frac{1-|a_0|}{2} \ge\frac{1-\gamma_0}{2}. δ = 2 ϕ 0 ≥ 2 δ 0 , γ − ∣ a 0 ∣ = 2 1 − ∣ a 0 ∣ ≥ 2 1 − γ 0 .
Moreover, the constants K K K and C 0 C_0 C 0 above can be chosen to depend only on δ 0 \delta_0 δ 0 , γ 0 \gamma_0 γ 0 , K 0 K_0 K 0 and μ \mu μ . In particular, we may take
K = K 0 + 2 μ , K = K_0 + 2\mu, K = K 0 + 2 μ ,
so that
γ ∗ = 1 − 1 K 2 < 1 \gamma_* = \sqrt{1-\frac{1}{K^2}} < 1 γ ∗ = 1 − K 2 1 < 1
and the constant C 1 C_1 C 1 can also be chosen uniformly in terms of δ 0 \delta_0 δ 0 , γ 0 \gamma_0 γ 0 , K 0 K_0 K 0 and μ \mu μ . Therefore, the smallness conditions (62) and (63) are satisfied for a common positive time T 0 T_0 T 0 depending only on δ 0 \delta_0 δ 0 , γ 0 \gamma_0 γ 0 , K 0 K_0 K 0 and μ \mu μ .
Global existence and uniqueness will be established below.
Conservation laws
We next prove conservation of energy and momentum. For smooth solutions of (1), this property was established in [3 ] . The argument below extends this result to weak solutions with finite initial energy.
Let G G G be a weak solution on [ 0 , T ] [0,T] [ 0 , T ] with initial data G 0 G_0 G 0 satisfying
∫ R 3 ⟨ p ⟩ G 0 ( p ) d p < ∞ . \int_{\mathbb{R}^3} \langle p\rangle G_0(p)\,dp < \infty. ∫ R 3 ⟨ p ⟩ G 0 ( p ) d p < ∞.
For smooth functions f ( p ) f(p) f ( p ) , introduce the nonlinear operator
L G f = Φ G ∇ p ⋅ ( D ∇ p f ) − W G ⋅ ∇ p f . L_G f = \Phi_G \nabla_p \cdot(D\nabla_p f) - W_G \cdot\nabla_p f. L G f = Φ G ∇ p ⋅ ( D ∇ p f ) − W G ⋅ ∇ p f .
Formally, the weak equation gives
d d t ∫ R 3 f ( p ) G ( t , p ) d p = ∫ R 3 L G f ( p ) G ( t , p ) d p . \frac{d}{dt}\int_{\mathbb{R}^3} f(p)G(t,p)\,dp = \int_{\mathbb{R}^3} L_G f(p)G(t,p)\,dp. d t d ∫ R 3 f ( p ) G ( t , p ) d p = ∫ R 3 L G f ( p ) G ( t , p ) d p .
We justify this identity for the unbounded observables ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ and p i p_i p i .
Conservation of energy
Set
e ( p ) = ⟨ p ⟩ . e(p) = \langle p\rangle. e ( p ) = ⟨ p ⟩ .
A direct calculation gives
L G e = 3 Φ G − μ + μ ⟨ p ⟩ β G . (65) L_G e = 3\Phi_G-\mu+\frac{\mu}{\langle p\rangle\beta_G}. \tag*{(65)} L G e = 3 Φ G − μ + ⟨ p ⟩ β G μ . (65)
By (14) and (17),
∣ L G e ( t , p ) ∣ ≤ 2 μ , 0 ≤ t ≤ T , p ∈ R 3 . (66) \lvert L_G e(t,p)\rvert\le2\mu,\qquad0\le t\le T,\quad p\in\mathbb{R}^3. \tag*{(66)} ∣ L G e ( t , p )∣ ≤ 2 μ , 0 ≤ t ≤ T , p ∈ R 3 . (66)
To justify the use of e e e as a test function, we first establish a bound on the energy of G G G .
Choose a smooth, nondecreasing, concave function η : [ 0 , ∞ ) → [ 0 , ∞ ) \eta:[0,\infty)\to[0,\infty) η : [ 0 , ∞ ) → [ 0 , ∞ ) satisfying
η ( r ) = r ( 0 ≤ r ≤ 1 ) , η ′ ( r ) = 0 ( r ≥ 2 ) , 0 ≤ η ′ ≤ 1. \eta(r)=r\quad(0\le r\le1),\qquad\eta'(r)=0\quad(r\ge2),\qquad0\le\eta'\le1. η ( r ) = r ( 0 ≤ r ≤ 1 ) , η ′ ( r ) = 0 ( r ≥ 2 ) , 0 ≤ η ′ ≤ 1.
For R ≥ 1 R\ge1 R ≥ 1 , define
e R ( p ) = R η ( e ( p ) R ) . e_R(p)=R\eta\left(\frac{e(p)}{R}\right). e R ( p ) = R η ( R e ( p ) ) .
Then
0 ≤ e R ≤ e , e R → e 0\le e_R\le e,\qquad e_R\to e 0 ≤ e R ≤ e , e R → e
pointwise as R → ∞ R\to\infty R → ∞ . We compute
L G e R = η ′ ( e R ) L G e + Φ G R η ′ ′ ( e R ) ∇ p e ⋅ D ∇ p e . L_Ge_R=\eta'\left(\frac{e}{R}\right)L_Ge+\frac{\Phi_G}{R}\eta''\left(\frac{e}{R}\right)\nabla_pe\cdot D\nabla_pe. L G e R = η ′ ( R e ) L G e + R Φ G η ′′ ( R e ) ∇ p e ⋅ D ∇ p e .
Since η ′ ′ ≤ 0 \eta''\le0 η ′′ ≤ 0 , Φ G ≥ 0 \Phi_G\ge0 Φ G ≥ 0 , and
∇ p e ⋅ D ∇ p e = ∣ p ∣ 2 ⟨ p ⟩ ≥ 0 , \nabla_pe\cdot D\nabla_pe=\frac{\lvert p\rvert^2}{\langle p\rangle}\ge0, ∇ p e ⋅ D ∇ p e = ⟨ p ⟩ ∣ p ∣ 2 ≥ 0 ,
we obtain
L G e R ≤ 2 μ . L_Ge_R\le2\mu. L G e R ≤ 2 μ .
Moreover, L G e R L_Ge_R L G e R is bounded uniformly in R R R . Indeed, η ′ ′ ( e / R ) \eta''(e/R) η ′′ ( e / R ) vanishes outside the region R ≤ e ≤ 2 R R\le e\le2R R ≤ e ≤ 2 R , where
1 R ∇ p e ⋅ D ∇ p e ≤ 2. \frac{1}{R}\nabla_pe\cdot D\nabla_pe\le2. R 1 ∇ p e ⋅ D ∇ p e ≤ 2.
Although e R e_R e R is bounded, it is not compactly supported. Let χ ∈ C ∞ ( [ 0 , ∞ ) ) \chi\in C^\infty([0,\infty)) χ ∈ C ∞ ([ 0 , ∞ )) satisfy
0 ≤ χ ≤ 1 , χ ( r ) = 1 ( r ≤ 1 ) , χ ( r ) = 0 ( r ≥ 2 ) , 0\le\chi\le1,\qquad\chi(r)=1\quad(r\le1),\qquad\chi(r)=0\quad(r\ge2), 0 ≤ χ ≤ 1 , χ ( r ) = 1 ( r ≤ 1 ) , χ ( r ) = 0 ( r ≥ 2 ) ,
and set
χ N ( p ) = χ ( ∣ p ∣ N ) . \chi_N(p)=\chi\left(\frac{\lvert p\rvert}{N}\right). χ N ( p ) = χ ( N ∣ p ∣ ) .
For fixed R R R and N N N sufficiently large, the function e R e_R e R is constant on the support of ∇ p χ N \nabla_p\chi_N ∇ p χ N . Expanding the generator gives
L G ( χ N e R ) = χ N L G e R + Φ G e R ∇ p ⋅ ( D ∇ p χ N ) − e R W G ⋅ ∇ p χ N , L_G(\chi_Ne_R)=\chi_NL_Ge_R+\Phi_Ge_R\nabla_p\cdot(D\nabla_p\chi_N)-e_RW_G\cdot\nabla_p\chi_N, L G ( χ N e R ) = χ N L G e R + Φ G e R ∇ p ⋅ ( D ∇ p χ N ) − e R W G ⋅ ∇ p χ N ,
where the cross term vanishes. Since χ N \chi_N χ N is radial,
D ∇ p χ N = ⟨ p ⟩ ∇ p χ N , D\nabla_p\chi_N=\langle p\rangle\nabla_p\chi_N, D ∇ p χ N = ⟨ p ⟩ ∇ p χ N ,
and
∣ ∇ p χ N ∣ ≤ C N , ∣ ∇ p ⋅ ( D ∇ p χ N ) ∣ ≤ C N . \lvert\nabla_p\chi_N\rvert\le\frac{C}{N},\qquad\left\lvert\nabla_p\cdot(D\nabla_p\chi_N)\right\rvert\le\frac{C}{N}. ∣ ∇ p χ N ∣ ≤ N C , ∣ ∇ p ⋅ ( D ∇ p χ N ) ∣ ≤ N C .
Moreover, (14) implies
∣ W G ( t , p ) ∣ ≤ 3 μ , \lvert W_G(t,p)\rvert\le3\mu, ∣ W G ( t , p )∣ ≤ 3 μ ,
while 0 ≤ e R ≤ C R 0 \le e_R \le CR 0 ≤ e R ≤ C R . Consequently,
∣ L G ( χ N e R ) − χ N L G e R ∣ ≤ C μ R N . \lvert L_G(\chi_N e_R)-\chi_N L_G e_R\rvert\le C\mu\frac{R}{N}. ∣ L G ( χ N e R ) − χ N L G e R ∣ ≤ C μ N R .
Using χ N e R \chi_N e_R χ N e R in the integrated weak formulation and letting N → ∞ N \to\infty N → ∞ , we obtain
∫ R 3 e R ( p ) G ( t , p ) d p = ∫ R 3 e R ( p ) G 0 ( p ) d p + ∫ 0 t ∫ R 3 L G e R ( p ) G ( s , p ) d p d s . \int_{\mathbb{R}^3} e_R(p)G(t,p)\,dp
=
\int_{\mathbb{R}^3} e_R(p)G_0(p)\,dp
+
\int_0^t\int_{\mathbb{R}^3} L_G e_R(p)G(s,p)\,dp\,ds. ∫ R 3 e R ( p ) G ( t , p ) d p = ∫ R 3 e R ( p ) G 0 ( p ) d p + ∫ 0 t ∫ R 3 L G e R ( p ) G ( s , p ) d p d s .
Since L G e R ≤ 2 μ L_G e_R \le2\mu L G e R ≤ 2 μ and G G G has unit mass,
∫ R 3 e R ( p ) G ( t , p ) d p ≤ ∫ R 3 e R ( p ) G 0 ( p ) d p + 2 μ t . \int_{\mathbb{R}^3} e_R(p)G(t,p)\,dp
\le
\int_{\mathbb{R}^3} e_R(p)G_0(p)\,dp+2\mu t. ∫ R 3 e R ( p ) G ( t , p ) d p ≤ ∫ R 3 e R ( p ) G 0 ( p ) d p + 2 μ t .
Letting R → ∞ R \to\infty R → ∞ , Fatou’s lemma and dominated convergence yield
∫ R 3 ⟨ p ⟩ G ( t , p ) d p ≤ ∫ R 3 ⟨ p ⟩ G 0 ( p ) d p + 2 μ t , 0 ≤ t ≤ T . (67) \int_{\mathbb{R}^3} \langle p\rangle G(t,p)\,dp
\le
\int_{\mathbb{R}^3} \langle p\rangle G_0(p)\,dp+2\mu t,
\qquad0\le t\le T. \tag*{(67)} ∫ R 3 ⟨ p ⟩ G ( t , p ) d p ≤ ∫ R 3 ⟨ p ⟩ G 0 ( p ) d p + 2 μ t , 0 ≤ t ≤ T . (67)
Thus, the energy remains finite in any compact interval of time. We can now establish energy conservation. For χ R ( p ) = χ ( ∣ p ∣ / R ) \chi_R(p)=\chi(|p|/R) χ R ( p ) = χ ( ∣ p ∣/ R ) , we have
L G ( χ R e ) = χ R L G e + Φ G e ∇ p ⋅ ( D ∇ p χ R ) + 2 Φ G ∇ p χ R ⋅ D ∇ p e − e W G ⋅ ∇ p χ R . (68) L_G(\chi_R e)=\chi_R L_G e+\Phi_G e\,\nabla_p\cdot(D\nabla_p\chi_R)+2\Phi_G\nabla_p\chi_R\cdot D\nabla_p e-eW_G\cdot\nabla_p\chi_R. \tag*{(68)} L G ( χ R e ) = χ R L G e + Φ G e ∇ p ⋅ ( D ∇ p χ R ) + 2 Φ G ∇ p χ R ⋅ D ∇ p e − e W G ⋅ ∇ p χ R . (68)
Since D ∇ p e = p D\nabla_p e=p D ∇ p e = p , the cutoff terms are supported in
A R = { R ≤ ∣ p ∣ ≤ 2 R } . A_R=\{R\le|p|\le2R\}. A R = { R ≤ ∣ p ∣ ≤ 2 R } .
On this annulus,
∣ ∇ p χ R ∣ ≤ C R , ∣ ∇ p ⋅ ( D ∇ p χ R ) ∣ ≤ C R , |\nabla_p\chi_R|\le\frac{C}{R},
\qquad
|\nabla_p\cdot(D\nabla_p\chi_R)|\le\frac{C}{R}, ∣ ∇ p χ R ∣ ≤ R C , ∣ ∇ p ⋅ ( D ∇ p χ R ) ∣ ≤ R C ,
and ⟨ p ⟩ ≤ C R \langle p\rangle\le CR ⟨ p ⟩ ≤ C R . Using Φ G ≤ μ / 3 \Phi_G\le\mu/3 Φ G ≤ μ /3 and ∣ W G ∣ ≤ 3 μ |W_G|\le3\mu ∣ W G ∣ ≤ 3 μ , we obtain
∣ L G ( χ R e ) − χ R L G e ∣ ≤ C μ 1 A R ( p ) , |L_G(\chi_R e)-\chi_R L_G e|\le C\mu\mathbf{1}_{A_R}(p), ∣ L G ( χ R e ) − χ R L G e ∣ ≤ C μ 1 A R ( p ) ,
where C C C is independent of R R R and T T T . These errors vanish after integration against G G G over [ 0 , t ] × R 3 [0,t]\times\mathbb{R}^3 [ 0 , t ] × R 3 as R → ∞ R\to\infty R → ∞ , by dominated convergence. Moreover, (66) ensures that the energy is finite, while (65) provides an integrable bound for the generator term. Passing to the limit in the integrated weak formulation yields
∫ R 3 e ( p ) G ( t , p ) d p = ∫ R 3 e ( p ) G 0 ( p ) d p + ∫ 0 t [ 3 Φ G ( s ) − μ + μ ∫ R 3 G ( s , p ) ⟨ p ⟩ β G ( s , p ) d p ] d s . \int_{\mathbb{R}^3}e(p)G(t,p)\,dp
=
\int_{\mathbb{R}^3}e(p)G_0(p)\,dp
+
\int_0^t\left[3\Phi_G(s)-\mu+\mu\int_{\mathbb{R}^3}\frac{G(s,p)}{\langle p\rangle^\beta G(s,p)}\,dp\right]\,ds. ∫ R 3 e ( p ) G ( t , p ) d p = ∫ R 3 e ( p ) G 0 ( p ) d p + ∫ 0 t [ 3 Φ G ( s ) − μ + μ ∫ R 3 ⟨ p ⟩ β G ( s , p ) G ( s , p ) d p ] d s .
By the definition of Φ G \Phi_G Φ G , the integrand in square brackets vanishes identically. Hence
∫ R 3 ⟨ p ⟩ G ( t , p ) d p = ∫ R 3 ⟨ p ⟩ G 0 ( p ) d p = : E , 0 ≤ t ≤ T . (69) \int_{\mathbb{R}^3}\langle p\rangle G(t,p)\,dp
=
\int_{\mathbb{R}^3}\langle p\rangle G_0(p)\,dp
=:E,
\qquad0\le t\le T. \tag*{(69)} ∫ R 3 ⟨ p ⟩ G ( t , p ) d p = ∫ R 3 ⟨ p ⟩ G 0 ( p ) d p =: E , 0 ≤ t ≤ T . (69)
Conservation of momentum
Since ∣ p ∣ ≤ ⟨ p ⟩ |p| \le\langle p \rangle ∣ p ∣ ≤ ⟨ p ⟩ , energy conservation implies that the momentum
P G ( t ) = ∫ R 3 p G ( t , p ) d p P_G(t) = \int_{\mathbb{R}^3} p\,G(t,p)\,\mathrm{d}p P G ( t ) = ∫ R 3 p G ( t , p ) d p
is well defined. For i = 1 , 2 , 3 i = 1,2,3 i = 1 , 2 , 3 , we have
L G p i = ( 3 Φ G − μ ) v i + μ ( j G ) i ⟨ p ⟩ β G . L_G p_i = (3\Phi_G-\mu)v_i+\frac{\mu(j_G)_i}{\langle p\rangle\beta_G}. L G p i = ( 3 Φ G − μ ) v i + ⟨ p ⟩ β G μ ( j G ) i .
By (14) and (17) ,
∣ L G p i ∣ ≤ 3 μ . |L_Gp_i| \le3\mu. ∣ L G p i ∣ ≤ 3 μ .
Using χ R p i \chi_Rp_i χ R p i as a test function, the same cutoff estimates as in the energy argument give
∣ L G ( χ R p i ) − χ R L G p i ∣ ≤ C μ 1 A R ( p ) , \left|L_G(\chi_Rp_i)-\chi_RL_Gp_i\right|\le C\mu1_{A_R}(p), ∣ L G ( χ R p i ) − χ R L G p i ∣ ≤ C μ 1 A R ( p ) ,
where C C C is independent of R R R and T T T . The cutoff errors vanish after integration against G G G over [ 0 , t ] × R 3 [0,t]\times\mathbb{R}^3 [ 0 , t ] × R 3 as R → ∞ R\to\infty R → ∞ . Moreover, energy conservation and the bound on L G p i L_Gp_i L G p i justify passing to the limit in the remaining terms. We obtain
∫ R 3 p i G ( t , p ) d p = ∫ R 3 p i G 0 ( p ) d p + ∫ 0 t [ ( 3 Φ G ( s ) − μ ) ( j G ( s ) ) i + μ ( j G ( s ) ) i ∫ R 3 G ( s , p ) ⟨ p ⟩ β G ( s , p ) d p ] d s . \int_{\mathbb{R}^3}p_iG(t,p)\,\mathrm{d}p=\int_{\mathbb{R}^3}p_iG_0(p)\,\mathrm{d}p
+\int_0^t\left[(3\Phi_G(s)-\mu)(j_G(s))_i+\mu(j_G(s))_i\int_{\mathbb{R}^3}\frac{G(s,p)}{\langle p\rangle\beta_G(s,p)}\,\mathrm{d}p\right]\mathrm{d}s. ∫ R 3 p i G ( t , p ) d p = ∫ R 3 p i G 0 ( p ) d p + ∫ 0 t [ ( 3 Φ G ( s ) − μ ) ( j G ( s ) ) i + μ ( j G ( s ) ) i ∫ R 3 ⟨ p ⟩ β G ( s , p ) G ( s , p ) d p ] d s .
By the definition of Φ G \Phi_G Φ G , the integrand in square brackets vanishes identically. Therefore
∫ R 3 p G ( t , p ) d p = ∫ R 3 p G 0 ( p ) d p = : P , 0 ≤ t ≤ T . (70) \int_{\mathbb{R}^3}pG(t,p)\,\mathrm{d}p=\int_{\mathbb{R}^3}pG_0(p)\,\mathrm{d}p=:P,\qquad0\le t\le T. \tag*{(70)} ∫ R 3 pG ( t , p ) d p = ∫ R 3 p G 0 ( p ) d p =: P , 0 ≤ t ≤ T . (70)
Thus every weak solution with finite initial energy conserves energy and momentum.
Propagation of energy tails
We now prove that the energy tails of weak solutions remain uniformly integrable on every finite time interval.
For ϕ ≥ 0 \phi\ge0 ϕ ≥ 0 and ∣ a ∣ < 1 |a|<1 ∣ a ∣ < 1 , let L ϕ , a L_{\phi,a} L ϕ , a denote the frozen generator on the right hand side of (34) , that is,
L ϕ , a ψ = ϕ ∇ p ⋅ ( D ∇ p ψ ) − V a ⋅ ∇ p ψ . (71) L_{\phi,a}\psi=\phi\nabla_p\cdot(D\nabla_p\psi)-V_a\cdot\nabla_p\psi. \tag*{(71)} L ϕ , a ψ = ϕ ∇ p ⋅ ( D ∇ p ψ ) − V a ⋅ ∇ p ψ . (71)
The following lemma provides an estimate for the frozen generator that will be used to control the energy tails of solutions to the nonlinear equation. An important feature of the estimate is that it is uniform in a a a , even as ∣ a ∣ |a| ∣ a ∣ approaches one.
Lemma 17. Let η ∈ C ∞ ( [ 0 , ∞ ) ) \eta\in C^{\infty}([0,\infty)) η ∈ C ∞ ([ 0 , ∞ )) be nondecreasing and satisfy
0 ≤ η ≤ 1 , η ( s ) = 0 ( s ≤ 1 ) , η ( s ) = 1 ( s ≥ 2 ) . 0 \le\eta\le1,\qquad\eta(s)=0\quad(s\le1),\qquad\eta(s)=1\quad(s\ge2). 0 ≤ η ≤ 1 , η ( s ) = 0 ( s ≤ 1 ) , η ( s ) = 1 ( s ≥ 2 ) . For R > 1 R>1 R > 1 , define
H R ( p ) = ⟨ p ⟩ η ( ⟨ p ⟩ R ) . H_R(p)=\langle p\rangle\eta\left(\frac{\langle p\rangle}{R}\right). H R ( p ) = ⟨ p ⟩ η ( R ⟨ p ⟩ ) . Then
⟨ p ⟩ 1 { ⟨ p ⟩ ≥ 2 R } ≤ H R ( p ) ≤ ⟨ p ⟩ 1 { ⟨ p ⟩ > R } . (72) \langle p\rangle\mathbf{1}_{\{\langle p\rangle\ge2R\}}\le H_R(p)\le\langle p\rangle\mathbf{1}_{\{\langle p\rangle>R\}}. \tag*{(72)} ⟨ p ⟩ 1 {⟨ p ⟩ ≥ 2 R } ≤ H R ( p ) ≤ ⟨ p ⟩ 1 {⟨ p ⟩ > R } . (72) Moreover, for every Λ > 0 \Lambda>0 Λ > 0 , there exists a constant C > 0 C>0 C > 0 , depending only on η \eta η , such that
∣ L ϕ , a H R ( p ) ∣ ≤ C ( Λ + μ ) 1 { ⟨ p ⟩ > R } , (73) \left|L_{\phi,a}H_R(p)\right|\le C(\Lambda+\mu)\mathbf{1}_{\{\langle p\rangle>R\}}, \tag*{(73)} ∣ L ϕ , a H R ( p ) ∣ ≤ C ( Λ + μ ) 1 {⟨ p ⟩ > R } , (73) for all 0 ≤ ϕ ≤ Λ 0\le\phi\le\Lambda 0 ≤ ϕ ≤ Λ and ∣ a ∣ < 1 |a|<1 ∣ a ∣ < 1 . The constant C C C is independent of R R R , ϕ \phi ϕ , a a a and p p p .
Proof. Set h = ⟨ p ⟩ h=\langle p\rangle h = ⟨ p ⟩ . A direct computation gives
L ϕ , a h = 3 ϕ − μ + μ h β a . L_{\phi,a}h=3\phi-\mu+\frac{\mu}{h\beta_a}. L ϕ , a h = 3 ϕ − μ + h β a μ .
Moreover,
1 h β a = 1 − ∣ v ∣ 2 1 − a ⋅ v ≤ 1 − ∣ v ∣ 2 1 − ∣ v ∣ = 1 + ∣ v ∣ ≤ 2. \frac{1}{h\beta_a}=\frac{1-|v|^2}{1-a\cdot v}\le\frac{1-|v|^2}{1-|v|}=1+|v|\le2. h β a 1 = 1 − a ⋅ v 1 − ∣ v ∣ 2 ≤ 1 − ∣ v ∣ 1 − ∣ v ∣ 2 = 1 + ∣ v ∣ ≤ 2.
Consequently,
∣ L ϕ , a h ∣ ≤ 3 Λ + μ . |L_{\phi,a}h|\le3\Lambda+\mu. ∣ L ϕ , a h ∣ ≤ 3Λ + μ .
Write H R ( p ) = g R ( h ) H_R(p)=g_R(h) H R ( p ) = g R ( h ) , g R ( h ) = h η ( h / R ) g_R(h)=h\eta(h/R) g R ( h ) = h η ( h / R ) . Then
g R ′ ( h ) = η ( h / R ) + h R η ′ ( h / R ) , g_R'(h)=\eta(h/R)+\frac{h}{R}\eta'(h/R), g R ′ ( h ) = η ( h / R ) + R h η ′ ( h / R ) ,
and
g R ′ ′ ( h ) = 2 R η ′ ( h / R ) + h R 2 η ′ ′ ( h / R ) . g_R''(h)=\frac{2}{R}\eta'(h/R)+\frac{h}{R^2}\eta''(h/R). g R ′′ ( h ) = R 2 η ′ ( h / R ) + R 2 h η ′′ ( h / R ) .
There exists a constant C > 0 C>0 C > 0 , independent of R R R , such that
∣ g R ′ ( h ) ∣ ≤ C , ∣ g R ′ ′ ( h ) ∣ ≤ C R 1 { R < h < 2 R } . |g_R'(h)|\le C,\qquad|g_R''(h)|\le\frac{C}{R}\mathbf{1}_{\{R<h<2R\}}. ∣ g R ′ ( h ) ∣ ≤ C , ∣ g R ′′ ( h ) ∣ ≤ R C 1 { R < h < 2 R } .
Moreover, both derivatives vanish for h ≤ R h\le R h ≤ R . Hence,
L ϕ , a H R = g R ′ ( h ) L ϕ , a h + ϕ g R ′ ′ ( h ) ( h − 1 h ) . L_{\phi,a}H_R=g_R'(h)L_{\phi,a}h+\phi g_R''(h)\left(h-\frac{1}{h}\right). L ϕ , a H R = g R ′ ( h ) L ϕ , a h + ϕ g R ′′ ( h ) ( h − h 1 ) .
Since h ≤ 2 R h\le2R h ≤ 2 R on the support of g R ′ ′ g_R'' g R ′′ , it follows that
∣ L ϕ , a H R ∣ ≤ C ( Λ + μ ) 1 { h > R } , |L_{\phi,a}H_R|\le C(\Lambda+\mu)\mathbf{1}_{\{h>R\}}, ∣ L ϕ , a H R ∣ ≤ C ( Λ + μ ) 1 { h > R } ,
which proves (72). The bounds (71) follow directly from the definition of η \eta η .
Proposition 18. Let G G G be a weak solution on [ 0 , T ] [0,T] [ 0 , T ] with initial data G 0 G_0 G 0 satisfying E G 0 < ∞ E_{G_0}<\infty E G 0 < ∞ . Then there exists a constant C > 0 C>0 C > 0 , independent of G G G , T T T , and R R R , such that
sup 0 ≤ t ≤ T T 2 R ( G ( t ) ) ≤ T R ( G 0 ) + C μ E T R , R > 1 , (74) \sup_{0\le t\le T}\mathcal{T}_{2R}(G(t))\le\mathcal{T}_{R}(G_0)+\frac{C\mu ET}{R},\qquad R>1, \tag*{(74)} 0 ≤ t ≤ T sup T 2 R ( G ( t )) ≤ T R ( G 0 ) + R C μ E T , R > 1 , (74) where E = E G 0 E=E_{G_0} E = E G 0 . Consequently,
lim R → ∞ sup 0 ≤ t ≤ T T R ( G ( t ) ) = 0. (75) \lim_{R\to\infty}\sup_{0\le t\le T}\mathcal{T}_{R}(G(t))=0. \tag*{(75)} R → ∞ lim 0 ≤ t ≤ T sup T R ( G ( t )) = 0. (75)
Proof. Let H R H_R H R be the function introduced in Lemma 17 . Since
0 < Φ G ( t ) ≤ μ 3 , ∣ j G ( t ) ∣ < 1 , 0<\Phi_G(t)\le\frac{\mu}{3},\qquad|j_G(t)|<1, 0 < Φ G ( t ) ≤ 3 μ , ∣ j G ( t ) ∣ < 1 ,
the generator estimate (73) , applied with Λ = μ / 3 \Lambda=\mu/3 Λ = μ /3 , gives
∣ L G H R ( p ) ∣ ≤ C μ 1 { ⟨ p ⟩ > R } . |L_GH_R(p)|\le C\mu\mathbf{1}_{\{\langle p\rangle>R\}}. ∣ L G H R ( p ) ∣ ≤ C μ 1 {⟨ p ⟩ > R } .
We first justify the use of H R H_R H R in the integrated weak formulation. Let χ ∈ C ∞ ( [ 0 , ∞ ) ) \chi\in C^\infty([0,\infty)) χ ∈ C ∞ ([ 0 , ∞ )) satisfy
0 ≤ χ ≤ 1 , χ ( r ) = 1 ( r ≤ 1 ) , χ ( r ) = 0 ( r ≥ 2 ) , 0\le\chi\le1,\qquad\chi(r)=1\quad(r\le1),\qquad\chi(r)=0\quad(r\ge2), 0 ≤ χ ≤ 1 , χ ( r ) = 1 ( r ≤ 1 ) , χ ( r ) = 0 ( r ≥ 2 ) ,
and set
χ N ( p ) = χ ( ∣ p ∣ N ) . \chi_N(p)=\chi\left(\frac{|p|}{N}\right). χ N ( p ) = χ ( N ∣ p ∣ ) .
For fixed R R R and N N N sufficiently large,
H R ( p ) = ⟨ p ⟩ H_R(p)=\langle p\rangle H R ( p ) = ⟨ p ⟩
on the support of ∇ p χ N \nabla_p\chi_N ∇ p χ N . We compute
L G ( χ N H R ) = χ N L G H R + Φ G H R ∇ p ⋅ ( D ∇ p χ N ) + 2 Φ G ∇ p χ N ⋅ D ∇ p H R − H R W G ⋅ ∇ p χ N . \begin{aligned}
L_G(\chi_NH_R)={}&\chi_NL_GH_R+\Phi_GH_R\nabla_p\cdot(D\nabla_p\chi_N)\\
&+2\Phi_G\nabla_p\chi_N\cdot D\nabla_pH_R-H_RW_G\cdot\nabla_p\chi_N.
\end{aligned} L G ( χ N H R ) = χ N L G H R + Φ G H R ∇ p ⋅ ( D ∇ p χ N ) + 2 Φ G ∇ p χ N ⋅ D ∇ p H R − H R W G ⋅ ∇ p χ N .
On the support of ∇ p χ N \nabla_p\chi_N ∇ p χ N , we have
D ∇ p H R = p , ∣ ∇ p χ N ∣ ≤ C N , ∣ ∇ p ⋅ ( D ∇ p χ N ) ∣ ≤ C N . D\nabla_pH_R=p,\qquad|\nabla_p\chi_N|\le\frac{C}{N},\qquad|\nabla_p\cdot(D\nabla_p\chi_N)|\le\frac{C}{N}. D ∇ p H R = p , ∣ ∇ p χ N ∣ ≤ N C , ∣ ∇ p ⋅ ( D ∇ p χ N ) ∣ ≤ N C .
Moreover, (14) implies ∣ W G ( t , p ) ∣ ≤ 3 μ |W_G(t,p)|\le3\mu ∣ W G ( t , p ) ∣ ≤ 3 μ . Since ⟨ p ⟩ / N \langle p\rangle/N ⟨ p ⟩ / N is bounded on the support of ∇ p χ N \nabla_p\chi_N ∇ p χ N , it follows that
∣ L G ( χ N H R ) − χ N L G H R ∣ ≤ C μ 1 { N ≤ ∣ p ∣ ≤ 2 N } . |L_G(\chi_NH_R)-\chi_NL_GH_R|\le C\mu\mathbf{1}_{\{N\le|p|\le2N\}}. ∣ L G ( χ N H R ) − χ N L G H R ∣ ≤ C μ 1 { N ≤ ∣ p ∣ ≤ 2 N } .
Using χ N H R \chi_NH_R χ N H R in (13) and letting N → ∞ N\to\infty N → ∞ , we obtain
∫ R 3 H R ( p ) G ( t , p ) d p = ∫ R 3 H R ( p ) G 0 ( p ) d p + ∫ 0 t ∫ R 3 L G H R ( p ) G ( s , p ) d p d s . \int_{\mathbb{R}^3}H_R(p)G(t,p)\,dp
=
\int_{\mathbb{R}^3}H_R(p)G_0(p)\,dp
+
\int_0^t\int_{\mathbb{R}^3}L_GH_R(p)G(s,p)\,dp\,ds. ∫ R 3 H R ( p ) G ( t , p ) d p = ∫ R 3 H R ( p ) G 0 ( p ) d p + ∫ 0 t ∫ R 3 L G H R ( p ) G ( s , p ) d p d s .
Indeed, the cutoff errors vanish by dominated convergence, using conservation of mass. The other terms converge by dominated convergence, since H R ≤ ⟨ p ⟩ H_R \le\langle p\rangle H R ≤ ⟨ p ⟩ , the energy is finite, and L G H R L_GH_R L G H R is bounded. By (73) ,
∫ R 3 H R G ( t ) d p ≤ ∫ R 3 H R G 0 d p + C μ ∫ 0 t ∫ ⟨ p ⟩ > R G ( s , p ) d p d s . \int_{\mathbb{R}^3} H_R G(t)\,dp \le\int_{\mathbb{R}^3} H_R G_0\,dp + C_\mu\int_0^t \int_{\langle p\rangle>R} G(s,p)\,dp\,ds. ∫ R 3 H R G ( t ) d p ≤ ∫ R 3 H R G 0 d p + C μ ∫ 0 t ∫ ⟨ p ⟩ > R G ( s , p ) d p d s .
Since the energy is conserved,
∫ ⟨ p ⟩ > R G ( s , p ) d p ≤ E R , \int_{\langle p\rangle>R} G(s,p)\,dp \le\frac{E}{R}, ∫ ⟨ p ⟩ > R G ( s , p ) d p ≤ R E ,
and hence
∫ R 3 H R G ( t ) d p ≤ T R ( G 0 ) + C μ E t R . \int_{\mathbb{R}^3} H_R G(t)\,dp \le\mathcal{T}_R(G_0) + \frac{C_\mu Et}{R}. ∫ R 3 H R G ( t ) d p ≤ T R ( G 0 ) + R C μ E t .
Finally, by (72) ,
T 2 R ( G ( t ) ) ≤ ∫ R 3 H R G ( t ) d p . \mathcal{T}_{2R}(G(t)) \le\int_{\mathbb{R}^3} H_R G(t)\,dp. T 2 R ( G ( t )) ≤ ∫ R 3 H R G ( t ) d p .
Taking the supremum over 0 ≤ t ≤ T 0 \le t \le T 0 ≤ t ≤ T proves (74) . Since G 0 G_0 G 0 has finite energy,
lim R → ∞ T R ( G 0 ) = 0. \lim_{R\to\infty} \mathcal{T}_R(G_0)=0. R → ∞ lim T R ( G 0 ) = 0.
Letting R → ∞ R\to\infty R → ∞ in (74) therefore proves (75) .
Completion of the proof
We now complete the proof of Theorem 1 . We first show that the diffusion coefficient remains uniformly positive on every finite time interval. We then establish uniqueness and use the resulting a priori estimates to continue the local solution globally in time.
Positivity of the diffusion coefficient
We start by showing that the nonlinear diffusion coefficient cannot degenerate on a finite time interval.
Let G G G be a weak solution of (1) on [ 0 , T ] [0,T] [ 0 , T ] , where T > 0 T>0 T > 0 . By conservation of energy and momentum,
∫ R 3 ⟨ p ⟩ G ( t , p ) d p = E , ∫ R 3 p G ( t , p ) d p = P , 0 ≤ t ≤ T . \int_{\mathbb{R}^3} \langle p\rangle G(t,p)\,dp=E,\qquad\int_{\mathbb{R}^3} pG(t,p)\,dp=P,\qquad0\le t\le T. ∫ R 3 ⟨ p ⟩ G ( t , p ) d p = E , ∫ R 3 pG ( t , p ) d p = P , 0 ≤ t ≤ T .
Moreover, by Lemma 8,
Δ : = E 2 − ∣ P ∣ 2 − 1 > 0. \Delta:= E^2-|P|^2-1>0. Δ := E 2 − ∣ P ∣ 2 − 1 > 0.
By Proposition 10 , the inequality 1 + γ E ≤ 2 1+\gamma_E\leq2 1 + γ E ≤ 2 , and the lower bound (17) , for every R > 1 R>1 R > 1 ,
Φ G ( t ) ≥ 9 μ 128 E 2 R 2 ( Δ − 4 E T R ( G ( t ) ) ) + 2 , 0 ≤ t ≤ T . \Phi_G(t)\geq\frac{9\mu}{128E^2R^2}\left(\Delta-4E\mathcal{T}_R(G(t))\right)_+^2,\qquad0\leq t\leq T. Φ G ( t ) ≥ 128 E 2 R 2 9 μ ( Δ − 4 E T R ( G ( t )) ) + 2 , 0 ≤ t ≤ T .
which is equivalent to (8) . By Proposition 18 , we may choose R T > 1 R_T>1 R T > 1 , depending only on G 0 G_0 G 0 , E E E , P P P , μ \mu μ and T T T , such that
sup 0 ≤ t ≤ T T R T ( G ( t ) ) ≤ Δ 8 E . \sup_{0\leq t\leq T}\mathcal{T}_{R_T}(G(t))\leq\frac{\Delta}{8E}. 0 ≤ t ≤ T sup T R T ( G ( t )) ≤ 8 E Δ .
Consequently,
Δ − 4 E T R T ( G ( t ) ) ≥ Δ 2 , 0 ≤ t ≤ T . \Delta-4E\mathcal{T}_{R_T}(G(t))\geq\frac{\Delta}{2},\qquad0\leq t\leq T. Δ − 4 E T R T ( G ( t )) ≥ 2 Δ , 0 ≤ t ≤ T .
Hence
Φ G ( t ) ≥ 9 μ Δ 2 512 E 2 R T 2 = : δ T > 0 , 0 ≤ t ≤ T . (76) \Phi_G(t)\geq\frac{9\mu\Delta^2}{512E^2R_T^2}=:\delta_T>0,\qquad0\leq t\leq T. \tag*{(76)} Φ G ( t ) ≥ 512 E 2 R T 2 9 μ Δ 2 =: δ T > 0 , 0 ≤ t ≤ T . (76)
Thus the nonlinear diffusion coefficient remains uniformly positive on every finite time interval.
Uniqueness
Next we prove uniqueness in the class of weak solutions appearing in Theorem 1 .
Let G G G and H H H be two weak solutions of (1) on [ 0 , T ] [0,T] [ 0 , T ] with the same initial data G 0 G_0 G 0 . By the conservation laws,
∫ R 3 ⟨ p ⟩ G ( t , p ) d p = ∫ R 3 ⟨ p ⟩ H ( t , p ) d p = E \int_{\mathbb{R}^3}\langle p\rangle G(t,p)\,dp=\int_{\mathbb{R}^3}\langle p\rangle H(t,p)\,dp=E ∫ R 3 ⟨ p ⟩ G ( t , p ) d p = ∫ R 3 ⟨ p ⟩ H ( t , p ) d p = E
and
∫ R 3 p G ( t , p ) d p = ∫ R 3 p H ( t , p ) d p = P . \int_{\mathbb{R}^3}pG(t,p)\,dp=\int_{\mathbb{R}^3}pH(t,p)\,dp=P. ∫ R 3 pG ( t , p ) d p = ∫ R 3 p H ( t , p ) d p = P .
Consequently, by Lemma 6,
∣ j G ( t ) ∣ , ∣ j H ( t ) ∣ ≤ γ E = 1 − 1 E 2 < 1. (77) |j_G(t)|,|j_H(t)|\leq\gamma_E=\sqrt{1-\frac{1}{E^2}}<1. \tag*{(77)} ∣ j G ( t ) ∣ , ∣ j H ( t ) ∣ ≤ γ E = 1 − E 2 1 < 1. (77)
Since
G , H ∈ C ( [ 0 , T ] ; L 1 ( R 3 ) ) , G,H\in C([0,T];L^1(\mathbb{R}^3)), G , H ∈ C ([ 0 , T ] ; L 1 ( R 3 )) ,
the coefficient functions are continuous. Indeed,
∣ j G ( t ) − j G ( s ) ∣ ≤ ∥ G ( t ) − G ( s ) ∥ L 1 , |j_G(t)-j_G(s)|\leq\|G(t)-G(s)\|_{L^1}, ∣ j G ( t ) − j G ( s ) ∣ ≤ ∥ G ( t ) − G ( s ) ∥ L 1 ,
and similarly for H H H . Moreover, by (77) and the definitions of Φ G \Phi_G Φ G and Φ H \Phi_H Φ H ,
∣ Φ G ( t ) − Φ G ( s ) ∣ ≤ C ∥ G ( t ) − G ( s ) ∥ L 1 , |\Phi_G(t)-\Phi_G(s)|\leq C\|G(t)-G(s)\|_{L^1}, ∣ Φ G ( t ) − Φ G ( s ) ∣ ≤ C ∥ G ( t ) − G ( s ) ∥ L 1 ,
and similarly for Φ H \Phi_H Φ H . By (76) , there exists δ T > 0 \delta_T>0 δ T > 0 , depending only on G 0 G_0 G 0 , μ \mu μ and T T T , such that
δ T ≤ Φ G ( t ) , Φ H ( t ) ≤ μ 3 , 0 ≤ t ≤ T . (78) \delta_T \leq\Phi_G(t),\Phi_H(t) \leq\frac{\mu}{3}, \qquad0\leq t\leq T. \tag*{(78)} δ T ≤ Φ G ( t ) , Φ H ( t ) ≤ 3 μ , 0 ≤ t ≤ T . (78)
Hence
( Φ G , j G ) , ( Φ H , j H ) ∈ C ( [ 0 , T ] ; R × R 3 ) . (\Phi_G,j_G),(\Phi_H,j_H)\in C([0,T];\mathbb{R}\times\mathbb{R}^3). ( Φ G , j G ) , ( Φ H , j H ) ∈ C ([ 0 , T ] ; R × R 3 ) .
Thus G G G and H H H may be regarded as solutions of the frozen Fokker–Planck equations corresponding respectively to the coefficient functions ( Φ G , j G ) (\Phi_G,j_G) ( Φ G , j G ) and ( Φ H , j H ) (\Phi_H,j_H) ( Φ H , j H ) . These coefficient functions are continuous and satisfy the bounds (77) –(78) . Extending them constantly beyond T T T , we may apply Theorem 14 and Lemma 15 on [ 0 , T ] [0,T] [ 0 , T ] .
Let P t P_t P t and Q t Q_t Q t solve the stochastic differential equations corresponding respectively to ( Φ G , j G ) (\Phi_G,j_G) ( Φ G , j G ) and ( Φ H , j H ) (\Phi_H,j_H) ( Φ H , j H ) , driven by the same Brownian motion, and with P 0 = Q 0 P_0=Q_0 P 0 = Q 0 almost surely. Let G 0 ( p ) d p G_0(p)\,dp G 0 ( p ) d p be the probability distribution of P 0 P_0 P 0 and Q 0 Q_0 Q 0 . By the uniqueness statement in Theorem 14 , which applies to all weak solutions in the stated class, the time marginal distributions of P t P_t P t and Q t Q_t Q t are respectively G ( t , p ) d p G(t,p)\,dp G ( t , p ) d p and H ( t , p ) d p H(t,p)\,dp H ( t , p ) d p . Set
Z t = P t − Q t . Z_t=P_t-Q_t. Z t = P t − Q t .
The coefficient estimates in Lemma 13 , together with E ⟨ P t ⟩ = E ⟨ Q t ⟩ = E \mathbb{E}\langle P_t\rangle=\mathbb{E}\langle Q_t\rangle=E E ⟨ P t ⟩ = E ⟨ Q t ⟩ = E , give
E ∣ Z t ∣ 2 ≤ C T ∫ 0 t E ∣ Z s ∣ 2 d s + C T ∫ 0 t ( ∣ Φ G ( s ) − Φ H ( s ) ∣ 2 + ∣ j G ( s ) − j H ( s ) ∣ 2 ) d s . (79) \mathbb{E}|Z_t|^2\leq C_T\int_0^t\mathbb{E}|Z_s|^2\,ds+C_T\int_0^t\left(|\Phi_G(s)-\Phi_H(s)|^2+|j_G(s)-j_H(s)|^2\right)\,ds. \tag*{(79)} E ∣ Z t ∣ 2 ≤ C T ∫ 0 t E ∣ Z s ∣ 2 d s + C T ∫ 0 t ( ∣ Φ G ( s ) − Φ H ( s ) ∣ 2 + ∣ j G ( s ) − j H ( s ) ∣ 2 ) d s . (79)
Here the estimate is justified by localization, as in the proof of Lemma 15 . In particular, it does not require finite second moments of the individual processes. Since the joint distribution of ( P t , Q t ) (P_t,Q_t) ( P t , Q t ) is a coupling of G ( t ) G(t) G ( t ) and H ( t ) H(t) H ( t ) ,
W 1 ( G ( t ) , H ( t ) ) ≤ E ∣ P t − Q t ∣ ≤ ( E ∣ Z t ∣ 2 ) 1 / 2 . (80) W_1(G(t),H(t))\leq\mathbb{E}|P_t-Q_t|\leq\left(\mathbb{E}|Z_t|^2\right)^{1/2}. \tag*{(80)} W 1 ( G ( t ) , H ( t )) ≤ E ∣ P t − Q t ∣ ≤ ( E ∣ Z t ∣ 2 ) 1/2 . (80)
By Lemma 12 and (77) ,
∣ Φ G ( t ) − Φ H ( t ) ∣ + ∣ j G ( t ) − j H ( t ) ∣ ≤ C T W 1 ( G ( t ) , H ( t ) ) . |\Phi_G(t)-\Phi_H(t)|+|j_G(t)-j_H(t)|\leq C_TW_1(G(t),H(t)). ∣ Φ G ( t ) − Φ H ( t ) ∣ + ∣ j G ( t ) − j H ( t ) ∣ ≤ C T W 1 ( G ( t ) , H ( t )) .
Hence
∣ Φ G ( t ) − Φ H ( t ) ∣ 2 + ∣ j G ( t ) − j H ( t ) ∣ 2 ≤ C T W 1 ( G ( t ) , H ( t ) ) 2 . |\Phi_G(t)-\Phi_H(t)|^2+|j_G(t)-j_H(t)|^2\leq C_TW_1(G(t),H(t))^2. ∣ Φ G ( t ) − Φ H ( t ) ∣ 2 + ∣ j G ( t ) − j H ( t ) ∣ 2 ≤ C T W 1 ( G ( t ) , H ( t ) ) 2 .
Using (80) , we obtain from (79)
E ∣ Z t ∣ 2 ≤ C T ∫ 0 t E ∣ Z s ∣ 2 d s . \mathbb{E}|Z_t|^2\leq C_T\int_0^t\mathbb{E}|Z_s|^2\,ds. E ∣ Z t ∣ 2 ≤ C T ∫ 0 t E ∣ Z s ∣ 2 d s .
Since Z 0 = 0 Z_0=0 Z 0 = 0 , Gronwall’s inequality gives E ∣ Z t ∣ 2 = 0 \mathbb{E}|Z_t|^2=0 E ∣ Z t ∣ 2 = 0 for 0 ≤ t ≤ T 0\leq t\leq T 0 ≤ t ≤ T . Therefore
P t = Q t almost surely , P_t=Q_t \qquad\text{almost surely}, P t = Q t almost surely ,
and hence G ( t , p ) = H ( t , p ) G(t,p)=H(t,p) G ( t , p ) = H ( t , p ) for almost every p ∈ R 3 p\in\mathbb{R}^3 p ∈ R 3 and every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
Global continuation
We finally prove that the local solution in Theorem 16 extends globally in time.
By local existence and uniqueness, the solution can be continued to a unique maximal interval [ 0 , T ∗ ) [0,T_*) [ 0 , T ∗ ) , where 0 < T ∗ ≤ ∞ 0<T_*\leq\infty 0 < T ∗ ≤ ∞ . Suppose, by contradiction, that T ∗ < ∞ T_*<\infty T ∗ < ∞ . By conservation of energy,
E G ( t ) = E , 0 ≤ t < T ∗ . E_{G(t)}=E,\qquad0\leq t<T_*. E G ( t ) = E , 0 ≤ t < T ∗ .
Consequently, by Lemma 6,
∣ j G ( t ) ∣ ≤ γ E = 1 − 1 E 2 < 1 , 0 ≤ t < T ∗ . \lvert j_{G}(t)\rvert\leq\gamma_E=\sqrt{1-\frac{1}{E^2}}<1,\qquad0\leq t<T_*. ∣ j G ( t )∣ ≤ γ E = 1 − E 2 1 < 1 , 0 ≤ t < T ∗ .
By Proposition 18 , for every t < T ∗ t<T_* t < T ∗ and R > 1 R>1 R > 1 ,
T 2 R ( G ( t ) ) ≤ T R ( G 0 ) + C μ E T ∗ R . \mathcal{T}_{2R}(G(t))\leq\mathcal{T}_{R}(G_0)+\frac{C\mu E T_*}{R}. T 2 R ( G ( t )) ≤ T R ( G 0 ) + R C μ E T ∗ .
Since G 0 G_0 G 0 has finite energy, we may choose R ∗ > 1 R_*>1 R ∗ > 1 , depending only on G 0 , E , P , μ G_0,E,P,\mu G 0 , E , P , μ and T ∗ T_* T ∗ , such that
T R ∗ ( G ( t ) ) ≤ Δ 8 E , 0 ≤ t < T ∗ , \mathcal{T}_{R_*}(G(t))\leq\frac{\Delta}{8E},\qquad0\leq t<T_*, T R ∗ ( G ( t )) ≤ 8 E Δ , 0 ≤ t < T ∗ ,
where
Δ = E 2 − ∣ P ∣ 2 − 1 > 0. \Delta=E^2-\lvert P\rvert^2-1>0. Δ = E 2 − ∣ P ∣ 2 − 1 > 0.
Proposition 10 and Lemma 7 therefore give
Φ G ( t ) ≥ δ ∗ : = 9 μ Δ 2 512 E 2 R ∗ 2 > 0 , 0 ≤ t < T ∗ . \Phi_{G}(t)\geq\delta_*:=\frac{9\mu\Delta^2}{512E^2R_*^2}>0,\qquad0\leq t<T_*. Φ G ( t ) ≥ δ ∗ := 512 E 2 R ∗ 2 9 μ Δ 2 > 0 , 0 ≤ t < T ∗ .
Moreover,
Φ G ( t ) ≤ μ 3 . \Phi_{G}(t)\leq\frac{\mu}{3}. Φ G ( t ) ≤ 3 μ .
Fix t 0 < T ∗ t_0<T_* t 0 < T ∗ . We restart the local fixed-point argument at time t 0 t_0 t 0 , taking G ( t 0 ) G(t_0) G ( t 0 ) as initial data. The estimates above give
Φ G ( t 0 ) ≥ δ ∗ > 0 , ∣ j G ( t 0 ) ∣ ≤ γ E < 1 , E G ( t 0 ) = E . \Phi_{G}(t_0)\geq\delta_*>0,\qquad\lvert j_{G}(t_0)\rvert\leq\gamma_E<1,\qquad E_{G(t_0)}=E. Φ G ( t 0 ) ≥ δ ∗ > 0 , ∣ j G ( t 0 )∣ ≤ γ E < 1 , E G ( t 0 ) = E .
By the uniform statement in Theorem 16 , there exists τ ∗ > 0 \tau_*>0 τ ∗ > 0 , depending only on δ ∗ , γ E , E \delta_*,\gamma_E,E δ ∗ , γ E , E and μ \mu μ , and hence independent of t 0 < T ∗ t_0<T_* t 0 < T ∗ , such that the local existence argument with initial data G ( t 0 ) G(t_0) G ( t 0 ) produces a solution on [ t 0 , t 0 + τ ∗ ] [t_0,t_0+\tau_*] [ t 0 , t 0 + τ ∗ ] . Choose t 0 < T ∗ t_0<T_* t 0 < T ∗ sufficiently close to T ∗ T_* T ∗ that t 0 + τ ∗ > T ∗ t_0+\tau_*>T_* t 0 + τ ∗ > T ∗ . By uniqueness, the restarted solution agrees with G G G on their common interval of existence. By Lemma 5 , the two solutions can be concatenated to give a weak solution extending G G G beyond T ∗ T_* T ∗ , contradicting the maximality of T ∗ T_* T ∗ . Hence T ∗ = ∞ T_*=\infty T ∗ = ∞ .
This proves global existence and completes the proof of Theorem 1 .
Proof of Theorem 2
We now prove Theorem 2 . Fix T > 0 T > 0 T > 0 and set
ϕ ( t ) = Φ G ( t ) , a ( t ) = j G ( t ) . \phi(t) = \Phi_G(t), \qquad a(t) = j_G(t). ϕ ( t ) = Φ G ( t ) , a ( t ) = j G ( t ) .
By Theorem 1 , there exists a constant δ T > 0 \delta_T > 0 δ T > 0 such that
0 < δ T ≤ ϕ ( t ) ≤ μ 3 , ∣ a ( t ) ∣ ≤ γ E , 0 ≤ t ≤ T , 0 < \delta_T \le\phi(t) \le\frac{\mu}{3}, \qquad|a(t)| \le\gamma_E, \qquad0 \le t \le T, 0 < δ T ≤ ϕ ( t ) ≤ 3 μ , ∣ a ( t ) ∣ ≤ γ E , 0 ≤ t ≤ T ,
where
γ E = 1 − 1 E 2 < 1. \gamma_E = \sqrt{1 - \frac{1}{E^2}} < 1. γ E = 1 − E 2 1 < 1.
The diffusion matrix of the Fokker–Planck equation is ϕ ( t ) D ( p ) \phi(t)D(p) ϕ ( t ) D ( p ) . By (11), for ∣ p ∣ ≤ R |p| \le R ∣ p ∣ ≤ R and ξ ∈ R 3 \xi\in\mathbb{R}^3 ξ ∈ R 3 ,
δ T 1 + R 2 ∣ ξ ∣ 2 ≤ ϕ ( t ) ξ T D ( p ) ξ ≤ μ 1 + R 2 3 ∣ ξ ∣ 2 . (81) \frac{\delta_T}{\sqrt{1+R^2}}|\xi|^2 \le\phi(t)\xi^T D(p)\xi\le\frac{\mu\sqrt{1+R^2}}{3}|\xi|^2. \tag*{(81)} 1 + R 2 δ T ∣ ξ ∣ 2 ≤ ϕ ( t ) ξ T D ( p ) ξ ≤ 3 μ 1 + R 2 ∣ ξ ∣ 2 . (81)
Thus, the PDE (1) is uniformly parabolic on every bounded momentum set, uniformly on finite time intervals.
We divide the proof of Theorem 2 into several subsections.
An algebra of admissible observables
Fix 0 < γ < 1 0 < \gamma< 1 0 < γ < 1 . For ∣ a ∣ ≤ γ |a| \le\gamma ∣ a ∣ ≤ γ , set
z = 1 − a ⋅ v . z = 1-a\cdot v. z = 1 − a ⋅ v .
Then
z ≥ 1 − γ > 0. z \ge1-\gamma> 0. z ≥ 1 − γ > 0.
For k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 , let A k \mathcal{A}_k A k be the finite linear span of functions of the form
f ( a , p ) = ⟨ p ⟩ − k P ( a , v ) z ℓ , ℓ ∈ N 0 , f(a,p) = \langle p\rangle^{-k}\frac{P(a,v)}{z^\ell}, \qquad\ell\in\mathbb{N}_0, f ( a , p ) = ⟨ p ⟩ − k z ℓ P ( a , v ) , ℓ ∈ N 0 ,
where P P P is a polynomial in ( a , v ) (a,v) ( a , v ) . We call finite sums of elements of the classes A k \mathcal{A}_k A k admissible observables. Every function in A k \mathcal{A}_k A k , together with all its derivatives with respect to a a a , is bounded uniformly on { ∣ a ∣ ≤ γ } × R 3 \{|a| \le\gamma\} \times\mathbb{R}^3 { ∣ a ∣ ≤ γ } × R 3 . Moreover,
A k A m ⊂ A k + m , k , m ∈ N 0 . (82) \mathcal{A}_k\mathcal{A}_m \subset\mathcal{A}_{k+m}, \qquad k,m \in\mathbb{N}_0. \tag*{(82)} A k A m ⊂ A k + m , k , m ∈ N 0 . (82)
Since ⟨ p ⟩ − 2 = 1 − ∣ v ∣ 2 \langle p\rangle^{-2} = 1-|v|^2 ⟨ p ⟩ − 2 = 1 − ∣ v ∣ 2 is a polynomial in v v v , we also have
A k + 2 ⊂ A k , k ∈ N 0 . (83) \mathcal{A}_{k+2} \subset\mathcal{A}_k, \qquad k \in\mathbb{N}_0. \tag*{(83)} A k + 2 ⊂ A k , k ∈ N 0 . (83)
The identities
∂ p j ⟨ p ⟩ − k = − k ⟨ p ⟩ − k − 1 v j , ∂ p j v i = ⟨ p ⟩ − 1 ( δ i j − v i v j ) , \partial_{p_j}\langle p\rangle^{-k}=-k\langle p\rangle^{-k-1}v_j,\qquad\partial_{p_j}v_i=\langle p\rangle^{-1}(\delta_{ij}-v_iv_j), ∂ p j ⟨ p ⟩ − k = − k ⟨ p ⟩ − k − 1 v j , ∂ p j v i = ⟨ p ⟩ − 1 ( δ ij − v i v j ) ,
and
∂ p j z = − ⟨ p ⟩ − 1 ( a j − ( a ⋅ v ) v j ) \partial_{p_j}z=-\langle p\rangle^{-1}(a_j-(a\cdot v)v_j) ∂ p j z = − ⟨ p ⟩ − 1 ( a j − ( a ⋅ v ) v j )
give
∂ p j A k ⊂ A k + 1 , ∂ a j A k ⊂ A k . (84) \partial_{p_j}\mathcal{A}_k\subset\mathcal{A}_{k+1},\qquad\partial_{a_j}\mathcal{A}_k\subset\mathcal{A}_k. \tag*{(84)} ∂ p j A k ⊂ A k + 1 , ∂ a j A k ⊂ A k . (84)
We shall also use observables depending polynomially on an additional scalar parameter ϕ \phi ϕ . More precisely, let A \mathcal{A} A denote the finite linear span of functions of the form
f ( ϕ , a , p ) = ϕ r g ( a , p ) , r ∈ N 0 , g ∈ A k f(\phi,a,p)=\phi^r g(a,p),\qquad r\in\mathbb{N}_0,\qquad g\in\mathcal{A}_k f ( ϕ , a , p ) = ϕ r g ( a , p ) , r ∈ N 0 , g ∈ A k
for some k ≥ 0 k\geq0 k ≥ 0 . We shall refer to the elements of A \mathcal{A} A as extended admissible observables. By (82) , A \mathcal{A} A is an algebra under pointwise multiplication. On sets where ∣ a ∣ ≤ γ < 1 |a|\leq\gamma<1 ∣ a ∣ ≤ γ < 1 and ϕ \phi ϕ remains bounded, every extended admissible observable, together with all its derivatives with respect to ϕ \phi ϕ and a a a , is uniformly bounded.
Recall that the generator of the frozen Fokker–Planck equation is
L ϕ , a f = ϕ ∇ p ⋅ ( D ∇ p f ) − V a ⋅ ∇ p f , L_{\phi,a}f=\phi\nabla_p\cdot(D\nabla_p f)-V_a\cdot\nabla_p f, L ϕ , a f = ϕ ∇ p ⋅ ( D ∇ p f ) − V a ⋅ ∇ p f ,
where
D ( p ) = I + p ⊗ p ⟨ p ⟩ , V a ( p ) = μ ( v − a ⟨ p ⟩ 2 z ) , z = 1 − a ⋅ v . D(p)=\frac{I+p\otimes p}{\langle p\rangle},\qquad V_a(p)=\mu\left(v-\frac{a}{\langle p\rangle^2z}\right),\qquad z=1-a\cdot v. D ( p ) = ⟨ p ⟩ I + p ⊗ p , V a ( p ) = μ ( v − ⟨ p ⟩ 2 z a ) , z = 1 − a ⋅ v .
Lemma 19. For every k ≥ 0 k\geq0 k ≥ 0 and each fixed ϕ \phi ϕ ,
L ϕ , a A k ⊂ A k + 1 . (85) L_{\phi,a}\mathcal{A}_k\subset\mathcal{A}_{k+1}. \tag*{(85)} L ϕ , a A k ⊂ A k + 1 . (85) The coefficients appearing in this inclusion depend polynomially on ϕ \phi ϕ . In particular, the class A \mathcal{A} A is preserved by L ϕ , a L_{\phi,a} L ϕ , a , ∂ ϕ \partial_\phi ∂ ϕ , and ∇ a \nabla_a ∇ a .
Proof. Let f ∈ A k f\in\mathcal{A}_k f ∈ A k . By (84) ,
∂ p i f ∈ A k + 1 , ∂ p i ∂ p j f ∈ A k + 2 . \partial_{p_i}f\in\mathcal{A}_{k+1},\qquad\partial_{p_i}\partial_{p_j}f\in\mathcal{A}_{k+2}. ∂ p i f ∈ A k + 1 , ∂ p i ∂ p j f ∈ A k + 2 .
Since D i j = ⟨ p ⟩ − 1 δ i j + ⟨ p ⟩ v i v j D_{ij}=\langle p\rangle^{-1}\delta_{ij}+\langle p\rangle v_iv_j D ij = ⟨ p ⟩ − 1 δ ij + ⟨ p ⟩ v i v j , we have D i j ∂ p i ∂ p j f ∈ A k + 3 + A k + 1 D_{ij}\partial_{p_i}\partial_{p_j}f\in\mathcal{A}_{k+3}+\mathcal{A}_{k+1} D ij ∂ p i ∂ p j f ∈ A k + 3 + A k + 1 , and thus, by (83) ,
D i j ∂ p i ∂ p j f ∈ A k + 1 . D_{ij}\partial_{p_i}\partial_{p_j}f\in\mathcal{A}_{k+1}. D ij ∂ p i ∂ p j f ∈ A k + 1 .
Moreover, since ∂ p i D i j = 3 v j \partial_{p_i}D_{ij}=3v_j ∂ p i D ij = 3 v j , we have ( ∂ p i D i j ) ∂ p j f ∈ A k + 1 (\partial_{p_i}D_{ij})\partial_{p_j}f\in\mathcal{A}_{k+1} ( ∂ p i D ij ) ∂ p j f ∈ A k + 1 . It follows that
∇ p ⋅ ( D ∇ p f ) ∈ A k + 1 . \nabla_p\cdot(D\nabla_p f)\in\mathcal{A}_{k+1}. ∇ p ⋅ ( D ∇ p f ) ∈ A k + 1 .
Next, using
V a = μ ( v − a ⟨ p ⟩ 2 z ) , V_a=\mu\left(v-\frac{a}{\langle p\rangle^2z}\right), V a = μ ( v − ⟨ p ⟩ 2 z a ) ,
we obtain
V a ⋅ ∇ p f ∈ A k + 1 + A k + 3 ⊂ A k + 1 . V_a \cdot\nabla_p f \in A_{k+1} + A_{k+3} \subset A_{k+1}. V a ⋅ ∇ p f ∈ A k + 1 + A k + 3 ⊂ A k + 1 .
Combining these inclusions gives
L ϕ , a f ∈ A k + 1 , L_{\phi,a}f \in A_{k+1}, L ϕ , a f ∈ A k + 1 ,
which proves (84). Finally, the operator L ϕ , a L_{\phi,a} L ϕ , a depends linearly on ϕ \phi ϕ , while its remaining coefficients belong to the classes of admissible observables. Therefore L ϕ , a L_{\phi,a} L ϕ , a maps extended admissible observables into extended admissible observables. The invariance of A A A under ∂ ϕ \partial_\phi ∂ ϕ and ∇ a \nabla_a ∇ a follows directly from its definition and (83).
First derivatives of the nonlinear coefficients
We first establish differentiability of a a a and ϕ \phi ϕ directly from the weak equation.
Recall that
a ( t ) = j G ( t ) = ∫ R 3 v ( p ) G ( t , p ) d p . a(t) = j_G(t) = \int_{\mathbb{R}^3} v(p)G(t,p)\,dp. a ( t ) = j G ( t ) = ∫ R 3 v ( p ) G ( t , p ) d p .
A direct calculation gives
D ∇ p v i = e i ⟨ p ⟩ 2 , ∇ p ⋅ ( D ∇ p v i ) = − 2 v i ⟨ p ⟩ 3 , D\nabla_p v_i = \frac{e_i}{\langle p\rangle^2}, \qquad\nabla_p \cdot(D\nabla_p v_i) = -\frac{2v_i}{\langle p\rangle^3}, D ∇ p v i = ⟨ p ⟩ 2 e i , ∇ p ⋅ ( D ∇ p v i ) = − ⟨ p ⟩ 3 2 v i ,
and
V a ⋅ ∇ p v i = μ ( v i − a i ) ⟨ p ⟩ 3 ( 1 − a ⋅ v ) . V_a \cdot\nabla_p v_i = \frac{\mu(v_i-a_i)}{\langle p\rangle^3(1-a\cdot v)}. V a ⋅ ∇ p v i = ⟨ p ⟩ 3 ( 1 − a ⋅ v ) μ ( v i − a i ) .
Using suitable cutoff approximations of v i v_i v i in the integrated weak formulation, we obtain
a ′ ( t ) = − 2 ϕ ( t ) ∫ R 3 v ⟨ p ⟩ 3 G ( t , p ) d p − μ ∫ R 3 v − a ( t ) ⟨ p ⟩ 3 ( 1 − a ( t ) ⋅ v ) G ( t , p ) d p . (86) a'(t) = -2\phi(t)\int_{\mathbb{R}^3}\frac{v}{\langle p\rangle^3}G(t,p)\,dp-\mu\int_{\mathbb{R}^3}\frac{v-a(t)}{\langle p\rangle^3(1-a(t)\cdot v)}G(t,p)\,dp. \tag*{(86)} a ′ ( t ) = − 2 ϕ ( t ) ∫ R 3 ⟨ p ⟩ 3 v G ( t , p ) d p − μ ∫ R 3 ⟨ p ⟩ 3 ( 1 − a ( t ) ⋅ v ) v − a ( t ) G ( t , p ) d p . (86)
By conservation of energy and Lemma 6, ∣ a ( t ) ∣ ≤ γ E < 1 |a(t)|\le\gamma_E<1 ∣ a ( t ) ∣ ≤ γ E < 1 , t ≥ 0 t\ge0 t ≥ 0 . The integrands in (85) are therefore bounded continuous functions of ( a , p ) (a,p) ( a , p ) , uniformly for ∣ a ∣ ≤ γ E |a|\le\gamma_E ∣ a ∣ ≤ γ E . Since G ∈ C ( [ 0 , ∞ ) ; L 1 ( R 3 ) ) G\in C([0,\infty);L^1(\mathbb{R}^3)) G ∈ C ([ 0 , ∞ ) ; L 1 ( R 3 )) and a , ϕ a,\phi a , ϕ are continuous, the right hand side of (85) is continuous. Hence,
a ∈ C 1 ( ( 0 , ∞ ) ) . a\in C^1((0,\infty)). a ∈ C 1 (( 0 , ∞ )) .
Next set
F a ( p ) = 1 ⟨ p ⟩ 2 ( 1 − a ⋅ v ) . F_a(p)=\frac{1}{\langle p\rangle^2(1-a\cdot v)}. F a ( p ) = ⟨ p ⟩ 2 ( 1 − a ⋅ v ) 1 .
Then F a ∈ A 2 F_a\in A_2 F a ∈ A 2 and
ϕ ( t ) = μ 3 [ 1 − ∫ R 3 F a ( t ) ( p ) G ( t , p ) d p ] . \phi(t)=\frac{\mu}{3}\left[1-\int_{\mathbb{R}^3}F_{a(t)}(p)G(t,p)\,dp\right]. ϕ ( t ) = 3 μ [ 1 − ∫ R 3 F a ( t ) ( p ) G ( t , p ) d p ] .
Since a ∈ C 1 ( ( 0 , ∞ ) ) a \in C^1((0,\infty)) a ∈ C 1 (( 0 , ∞ )) , we may use cutoff approximations to the time-dependent observable F a ( t ) F_{a(t)} F a ( t ) in the integrated weak formulation. The chain rule in the parameter a a a gives
ϕ ′ ( t ) = − μ 3 [ ∫ R 3 L ϕ ( t ) , a ( t ) F a ( t ) ( p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a F a ( t ) ( p ) G ( t , p ) d p ] . (87) \phi'(t)=-\frac{\mu}{3}\left[\int_{\mathbb{R}^3}L_{\phi(t),a(t)}F_{a(t)}(p)G(t,p)\,\mathrm{d}p+a'(t)\cdot\int_{\mathbb{R}^3}\nabla_aF_{a(t)}(p)G(t,p)\,\mathrm{d}p\right]. \tag*{(87)} ϕ ′ ( t ) = − 3 μ [ ∫ R 3 L ϕ ( t ) , a ( t ) F a ( t ) ( p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a F a ( t ) ( p ) G ( t , p ) d p ] . (87)
where
∇ a F a = v ⟨ p ⟩ 2 ( 1 − a ⋅ v ) 2 . \nabla_aF_a=\frac{v}{\langle p\rangle^2(1-a\cdot v)^2}. ∇ a F a = ⟨ p ⟩ 2 ( 1 − a ⋅ v ) 2 v .
By Lemma 19,
L ϕ , a F a ∈ A 3 . L_{\phi,a}F_a\in\mathcal{A}_3. L ϕ , a F a ∈ A 3 .
The functions L ϕ , a F a L_{\phi,a}F_a L ϕ , a F a and ∇ a F a \nabla_aF_a ∇ a F a are therefore uniformly bounded and continuous in their parameters on
{ 0 ≤ ϕ ≤ μ 3 , ∣ a ∣ ≤ γ E } . \left\{0\leq\phi\leq\frac{\mu}{3},\ |a|\leq\gamma_E\right\}. { 0 ≤ ϕ ≤ 3 μ , ∣ a ∣ ≤ γ E } .
It follows that the right hand side of (86) is continuous. Hence,
ϕ ∈ C 1 ( ( 0 , ∞ ) ) . \phi\in C^1((0,\infty)). ϕ ∈ C 1 (( 0 , ∞ )) .
Differentiation of admissible moments
For an extended admissible observable f = f ( ϕ , a , p ) f=f(\phi,a,p) f = f ( ϕ , a , p ) , define
M f ( t ) = ∫ R 3 f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p . \mathcal{M}_f(t)=\int_{\mathbb{R}^3}f(\phi(t),a(t),p)G(t,p)\,\mathrm{d}p. M f ( t ) = ∫ R 3 f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p .
Lemma 20. Suppose that a , ϕ ∈ C 1 ( I ) a,\phi\in C^1(I) a , ϕ ∈ C 1 ( I ) on a compact interval I ⊂ ( 0 , ∞ ) I\subset(0,\infty) I ⊂ ( 0 , ∞ ) and that f f f is an extended admissible observable. Then M f ∈ C 1 ( I ) \mathcal{M}_f\in C^1(I) M f ∈ C 1 ( I ) and
d d t M f ( t ) = ∫ R 3 L ϕ ( t ) , a ( t ) f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + ϕ ′ ( t ) ∫ R 3 ∂ ϕ f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p . (88) \frac{\mathrm{d}}{\mathrm{d}t}\mathcal{M}_f(t)=\int_{\mathbb{R}^3}L_{\phi(t),a(t)}f(\phi(t),a(t),p)G(t,p)\,\mathrm{d}p
+\phi'(t)\int_{\mathbb{R}^3}\partial_\phi f(\phi(t),a(t),p)G(t,p)\,\mathrm{d}p
+a'(t)\cdot\int_{\mathbb{R}^3}\nabla_af(\phi(t),a(t),p)G(t,p)\,\mathrm{d}p. \tag*{(88)} d t d M f ( t ) = ∫ R 3 L ϕ ( t ) , a ( t ) f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + ϕ ′ ( t ) ∫ R 3 ∂ ϕ f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p . (88)
Proof. Let
χ ∈ C c ∞ ( [ 0 , ∞ ) ) , 0 ≤ χ ≤ 1 , \chi\in C^\infty_c([0,\infty)),\qquad0\leq\chi\leq1, χ ∈ C c ∞ ([ 0 , ∞ )) , 0 ≤ χ ≤ 1 ,
satisfy
χ ( r ) = 1 for 0 ≤ r ≤ 1 , χ ( r ) = 0 for r ≥ 2 , \chi(r)=1\quad\text{for }0\leq r\leq1,\qquad\chi(r)=0\quad\text{for }r\geq2, χ ( r ) = 1 for 0 ≤ r ≤ 1 , χ ( r ) = 0 for r ≥ 2 ,
and set
χ R ( p ) = χ ( ∣ p ∣ R ) , R > 1. \chi_R(p)=\chi\left(\frac{|p|}{R}\right),\qquad R>1. χ R ( p ) = χ ( R ∣ p ∣ ) , R > 1.
Then
∣ ∇ p χ R ∣ ≤ C R , |\nabla_p\chi_R| \le\frac{C}{R}, ∣ ∇ p χ R ∣ ≤ R C ,
and all derivatives of χ R \chi_R χ R are supported in
A R = { R ≤ ∣ p ∣ ≤ 2 R } . A_R = \{R \le|p| \le2R\}. A R = { R ≤ ∣ p ∣ ≤ 2 R } .
Since χ R \chi_R χ R is radial and the radial eigenvalue of D ( p ) D(p) D ( p ) is ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ ,
D ( p ) ∇ p χ R = ⟨ p ⟩ ∇ p χ R . D(p)\nabla_p\chi_R = \langle p\rangle\nabla_p\chi_R. D ( p ) ∇ p χ R = ⟨ p ⟩ ∇ p χ R .
A direct calculation therefore gives
∣ D ∇ p χ R ∣ ≤ C , ∣ ∇ p ⋅ ( D ∇ p χ R ) ∣ ≤ C R . (89) |D\nabla_p\chi_R| \le C,\qquad|\nabla_p\cdot(D\nabla_p\chi_R)| \le\frac{C}{R}. \tag*{(89)} ∣ D ∇ p χ R ∣ ≤ C , ∣ ∇ p ⋅ ( D ∇ p χ R ) ∣ ≤ R C . (89)
Suppose first that f ∈ A k f \in A_k f ∈ A k . Uniformly for ∣ a ∣ ≤ γ < 1 |a| \le\gamma< 1 ∣ a ∣ ≤ γ < 1 ,
∣ f ( a , p ) ∣ ≤ C ⟨ p ⟩ − k , ∣ ∇ p f ( a , p ) ∣ ≤ C ⟨ p ⟩ − k − 1 . |f(a,p)| \le C\langle p\rangle^{-k},\qquad|\nabla_p f(a,p)| \le C\langle p\rangle^{-k-1}. ∣ f ( a , p ) ∣ ≤ C ⟨ p ⟩ − k , ∣ ∇ p f ( a , p ) ∣ ≤ C ⟨ p ⟩ − k − 1 .
Since the largest eigenvalue of D ( p ) D(p) D ( p ) is ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ ,
∣ D ∇ p f ∣ ≤ C ⟨ p ⟩ − k . |D\nabla_p f| \le C\langle p\rangle^{-k}. ∣ D ∇ p f ∣ ≤ C ⟨ p ⟩ − k .
Hence,
L ϕ , a ( χ R f ) − χ R L ϕ , a f = ϕ f ∇ p ⋅ ( D ∇ p χ R ) + 2 ϕ ∇ p χ R ⋅ D ∇ p f − f V a ⋅ ∇ p χ R . (90) L_{\phi,a}(\chi_R f)-\chi_RL_{\phi,a}f
=\phi f\nabla_p\cdot(D\nabla_p\chi_R)
+2\phi\nabla_p\chi_R\cdot D\nabla_p f
-fV_a\cdot\nabla_p\chi_R. \tag*{(90)} L ϕ , a ( χ R f ) − χ R L ϕ , a f = ϕ f ∇ p ⋅ ( D ∇ p χ R ) + 2 ϕ ∇ p χ R ⋅ D ∇ p f − f V a ⋅ ∇ p χ R . (90)
Using the estimates
∣ a ( t ) ∣ ≤ γ E < 1 , 0 < ϕ ( t ) ≤ μ 3 |a(t)| \le\gamma_E < 1,\qquad0<\phi(t)\le\frac{\mu}{3} ∣ a ( t ) ∣ ≤ γ E < 1 , 0 < ϕ ( t ) ≤ 3 μ
in Lemmas 6 and 7, we obtain ∣ V a ( t ) ( p ) ∣ ≤ C E |V_{a(t)}(p)| \le C_E ∣ V a ( t ) ( p ) ∣ ≤ C E . It follows from (89) that
∣ L ϕ , a ( χ R f ) − χ R L ϕ , a f ∣ ≤ C I R ⟨ p ⟩ − k 1 A R ( p ) . |L_{\phi,a}(\chi_Rf)-\chi_RL_{\phi,a}f|
\le\frac{C_I}{R}\langle p\rangle^{-k}\mathbf{1}_{A_R}(p). ∣ L ϕ , a ( χ R f ) − χ R L ϕ , a f ∣ ≤ R C I ⟨ p ⟩ − k 1 A R ( p ) .
Hence
∫ R 3 ∣ L ϕ , a ( χ R f ) − χ R L ϕ , a f ∣ G ( t , p ) d p ≤ C I R , \int_{\mathbb{R}^3}|L_{\phi,a}(\chi_Rf)-\chi_RL_{\phi,a}f|G(t,p)\,dp
\le\frac{C_I}{R}, ∫ R 3 ∣ L ϕ , a ( χ R f ) − χ R L ϕ , a f ∣ G ( t , p ) d p ≤ R C I ,
uniformly for t ∈ I t\in I t ∈ I . By linearity, the same estimate holds for extended admissible observables, with a constant depending also on the observable f f f . Thus, all cutoff commutators vanish as R → ∞ R\to\infty R → ∞ . Since a , ϕ ∈ C 1 ( I ) a,\phi\in C^1(I) a , ϕ ∈ C 1 ( I ) , the function
( t , p ) ⟼ χ R ( p ) f ( ϕ ( t ) , a ( t ) , p ) (t,p)\longmapsto\chi_R(p)f(\phi(t),a(t),p) ( t , p ) ⟼ χ R ( p ) f ( ϕ ( t ) , a ( t ) , p )
belongs to C 1 , 2 ( I × R 3 ) C^{1,2}(I \times\mathbb{R}^{3}) C 1 , 2 ( I × R 3 ) and has compact momentum support uniformly on I I I . It is therefore an admissible test function in the integrated weak formulation of Lemma 5 . We compute
∂ t f = ϕ ′ ( t ) ∂ ϕ f + a ′ ( t ) ⋅ ∇ a f . \partial_t f = \phi'(t)\partial_\phi f + a'(t) \cdot\nabla_a f. ∂ t f = ϕ ′ ( t ) ∂ ϕ f + a ′ ( t ) ⋅ ∇ a f .
For t 1 , t 2 ∈ I t_1,t_2 \in I t 1 , t 2 ∈ I , with t 1 < t 2 t_1 < t_2 t 1 < t 2 , the integrated weak formulation yields
∫ R 3 χ R f ( ϕ ( t 2 ) , a ( t 2 ) , p ) G ( t 2 , p ) d p − ∫ R 3 χ R f ( ϕ ( t 1 ) , a ( t 1 ) , p ) G ( t 1 , p ) d p = ∫ t 1 t 2 ∫ R 3 χ R [ L ϕ , a f + ϕ ′ ∂ ϕ f + a ′ ⋅ ∇ a f ] G d p d t + ∫ t 1 t 2 ∫ R 3 [ L ϕ , a ( χ R f ) − χ R L ϕ , a f ] G d p d t , \begin{aligned}
&\int_{\mathbb{R}^{3}} \chi_R f\bigl(\phi(t_2),a(t_2),p\bigr)G(t_2,p)\,dp \\
&\quad- \int_{\mathbb{R}^{3}} \chi_R f\bigl(\phi(t_1),a(t_1),p\bigr)G(t_1,p)\,dp \\
&= \int_{t_1}^{t_2}\int_{\mathbb{R}^{3}} \chi_R\left[L_{\phi,a}f+\phi'\partial_\phi f+a'\cdot\nabla_a f\right]G\,dp\,dt \\
&\quad+ \int_{t_1}^{t_2}\int_{\mathbb{R}^{3}}\left[L_{\phi,a}(\chi_R f)-\chi_RL_{\phi,a}f\right]G\,dp\,dt,
\end{aligned} ∫ R 3 χ R f ( ϕ ( t 2 ) , a ( t 2 ) , p ) G ( t 2 , p ) d p − ∫ R 3 χ R f ( ϕ ( t 1 ) , a ( t 1 ) , p ) G ( t 1 , p ) d p = ∫ t 1 t 2 ∫ R 3 χ R [ L ϕ , a f + ϕ ′ ∂ ϕ f + a ′ ⋅ ∇ a f ] G d p d t + ∫ t 1 t 2 ∫ R 3 [ L ϕ , a ( χ R f ) − χ R L ϕ , a f ] G d p d t ,
where the coefficients and observables in the time integrals are evaluated at ( t , p ) (t,p) ( t , p ) . The last term tends to zero as R → ∞ R \to\infty R → ∞ by the commutator estimate. Moreover, Lemma 19 shows that
L ϕ , a f , ∂ ϕ f , ∇ a f L_{\phi,a}f,\qquad\partial_\phi f,\qquad\nabla_a f L ϕ , a f , ∂ ϕ f , ∇ a f
are extended admissible observables and hence are uniformly bounded on the relevant parameter set. Since a ′ , ϕ ′ a',\phi' a ′ , ϕ ′ are bounded on I I I , dominated convergence allows us to pass to the limit in the remaining terms. We obtain
M f ( t 2 ) − M f ( t 1 ) = ∫ t 1 t 2 B f ( t ) d t . \mathcal{M}_f(t_2)-\mathcal{M}_f(t_1)=\int_{t_1}^{t_2}B_f(t)\,dt. M f ( t 2 ) − M f ( t 1 ) = ∫ t 1 t 2 B f ( t ) d t .
where
B f ( t ) = ∫ R 3 L ϕ ( t ) , a ( t ) f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + ϕ ′ ( t ) ∫ R 3 ∂ ϕ f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p . \begin{aligned}
B_f(t)={}&\int_{\mathbb{R}^{3}}L_{\phi(t),a(t)}f\bigl(\phi(t),a(t),p\bigr)G(t,p)\,dp \\
&+\phi'(t)\int_{\mathbb{R}^{3}}\partial_\phi f\bigl(\phi(t),a(t),p\bigr)G(t,p)\,dp \\
&+a'(t)\cdot\int_{\mathbb{R}^{3}}\nabla_a f\bigl(\phi(t),a(t),p\bigr)G(t,p)\,dp.
\end{aligned} B f ( t ) = ∫ R 3 L ϕ ( t ) , a ( t ) f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + ϕ ′ ( t ) ∫ R 3 ∂ ϕ f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p + a ′ ( t ) ⋅ ∫ R 3 ∇ a f ( ϕ ( t ) , a ( t ) , p ) G ( t , p ) d p .
The function B f B_f B f is continuous on I I I , because all the observables involved are bounded and continuous in their parameters, while
G ∈ C ( I ; L 1 ( R 3 ) ) , a , ϕ ∈ C 1 ( I ) . G \in C\bigl(I;L^{1}(\mathbb{R}^{3})\bigr),\qquad a,\phi\in C^{1}(I). G ∈ C ( I ; L 1 ( R 3 ) ) , a , ϕ ∈ C 1 ( I ) .
Therefore M f ∈ C 1 ( I ) \mathcal{M}_f \in C^{1}(I) M f ∈ C 1 ( I ) and
M f ′ ( t ) = B f ( t ) , \mathcal{M}'_f(t)=B_f(t), M f ′ ( t ) = B f ( t ) ,
which proves (88) .
Smoothness of the nonlinear coefficients
Proposition 21. The nonlinear coefficient functions satisfy
a , ϕ ∈ C ∞ ( ( 0 , ∞ ) ) . (91) a,\phi\in C^{\infty}((0,\infty)). \tag*{(91)} a , ϕ ∈ C ∞ (( 0 , ∞ )) . (91)
Proof. We have already proved that
a , ϕ ∈ C 1 ( ( 0 , ∞ ) ) . a,\phi\in C^{1}((0,\infty)). a , ϕ ∈ C 1 (( 0 , ∞ )) .
Fix a compact interval I ⊂ ( 0 , ∞ ) I \subset(0,\infty) I ⊂ ( 0 , ∞ ) . By conservation of energy and Lemma 6, ∣ a ( t ) ∣ ≤ γ E < 1 |a(t)| \le\gamma_E < 1 ∣ a ( t ) ∣ ≤ γ E < 1 . We prove by induction on m ≥ 1 m \ge1 m ≥ 1 that
a , ϕ ∈ C m ( I ) , M f ∈ C m ( I ) a,\phi\in C^{m}(I), \qquad\mathcal{M}_f \in C^{m}(I) a , ϕ ∈ C m ( I ) , M f ∈ C m ( I )
for every extended admissible observable f f f . The case m = 1 m=1 m = 1 follows from the preceding subsections and Lemma 20. Suppose that the assertion holds for some m ≥ 1 m \ge1 m ≥ 1 . By (86) , the function a ′ a' a ′ is a finite linear combination of extended admissible moments. Hence
a ∈ C m + 1 ( I ) . a \in C^{m+1}(I). a ∈ C m + 1 ( I ) .
Using (87) , the regularity of a ′ a' a ′ , and the induction hypothesis for admissible moments, we similarly obtain
ϕ ′ ∈ C m ( I ) , ϕ ∈ C m + 1 ( I ) . \phi' \in C^{m}(I), \qquad\phi\in C^{m+1}(I). ϕ ′ ∈ C m ( I ) , ϕ ∈ C m + 1 ( I ) .
It remains to improve the regularity of the admissible moments. By Lemma 20,
d d t M f = M L ϕ , a f + ϕ ′ M ∂ ϕ f + a ′ ⋅ M ∇ a f . \frac{d}{dt}\mathcal{M}_f
= \mathcal{M}_{L_{\phi,a}f}
+ \phi'\mathcal{M}_{\partial_\phi f}
+ a'\cdot\mathcal{M}_{\nabla_a f}. d t d M f = M L ϕ , a f + ϕ ′ M ∂ ϕ f + a ′ ⋅ M ∇ a f .
By Lemma 19, all the observables appearing on the right hand side are extended admissible. Their moments belong to C m ( I ) C^{m}(I) C m ( I ) by the induction hypothesis. Since a ′ , ϕ ′ ∈ C m ( I ) a',\phi' \in C^{m}(I) a ′ , ϕ ′ ∈ C m ( I ) , it follows that
M f ∈ C m + 1 ( I ) . \mathcal{M}_f \in C^{m+1}(I). M f ∈ C m + 1 ( I ) .
This completes the induction. Therefore
a , ϕ ∈ C ∞ ( I ) . a,\phi\in C^{\infty}(I). a , ϕ ∈ C ∞ ( I ) .
Since I ⊂ ( 0 , ∞ ) I \subset(0,\infty) I ⊂ ( 0 , ∞ ) was arbitrary, the result follows.
Interior smoothness
We now prove that the weak solution is smooth for positive times. Fix
0 < τ < T , 0 < R < R ′ , 0 < \tau< T, \qquad0 < R < R', 0 < τ < T , 0 < R < R ′ ,
and set
Q ′ = ( τ / 2 , T ) × B R ′ . Q' = (\tau/2,T) \times B_{R'}. Q ′ = ( τ /2 , T ) × B R ′ .
On Q ′ Q' Q ′ , the function G G G is a distributional solution of
∂ t G = ∂ i ( ϕ ( t ) D i j ( p ) ∂ j G + V a ( t ) i ( p ) G ) . (92) \partial_t G = \partial_i\left(\phi(t)D^{ij}(p)\partial_jG + V^i_{a(t)}(p)G\right). \tag*{(92)} ∂ t G = ∂ i ( ϕ ( t ) D ij ( p ) ∂ j G + V a ( t ) i ( p ) G ) . (92)
Set
A ( t , p ) = ϕ ( t ) D ( p ) . A(t,p) = \phi(t)D(p). A ( t , p ) = ϕ ( t ) D ( p ) .
By (80), the matrix A A A is uniformly positive definite on Q ′ Q' Q ′ . Moreover, Proposition 21 implies that all the coefficients in (91) are C ∞ C^\infty C ∞ on Q ′ Q' Q ′ . Let
S ( t , p ) = A ( t , p ) 1 / 2 . S(t,p) = A(t,p)^{1/2}. S ( t , p ) = A ( t , p ) 1/2 .
Since A A A is smooth and positive definite on Q ′ Q' Q ′ , its positive square root S S S is smooth. Define the vector fields
X k = ∑ i = 1 3 S i k ( t , p ) ∂ p i , k = 1 , 2 , 3. X_k = \sum_{i=1}^{3}S_{ik}(t,p)\partial_{p_i}, \qquad k=1,2,3. X k = i = 1 ∑ 3 S ik ( t , p ) ∂ p i , k = 1 , 2 , 3.
Since S S T = A SS^T=A S S T = A , we have
∑ k = 1 3 X k 2 = A i j ∂ p i ∂ p j + b j ( t , p ) ∂ p j , \sum_{k=1}^{3}X_k^2 = A^{ij}\partial_{p_i}\partial_{p_j}+b^j(t,p)\partial_{p_j}, k = 1 ∑ 3 X k 2 = A ij ∂ p i ∂ p j + b j ( t , p ) ∂ p j ,
where
b j ( t , p ) = ∑ i , k = 1 3 S i k ( t , p ) ∂ p i S j k ( t , p ) b^j(t,p)=\sum_{i,k=1}^{3}S_{ik}(t,p)\partial_{p_i}S_{jk}(t,p) b j ( t , p ) = i , k = 1 ∑ 3 S ik ( t , p ) ∂ p i S j k ( t , p )
are smooth functions on Q ′ Q' Q ′ . Expanding (91), we see that G G G satisfies
in the sense of distributions on Q ′ Q' Q ′ , where
P = − ∂ t + A i j ∂ p i ∂ p j + ( ∂ p i A i j + V a ( t ) j ) ∂ p j + ∂ p i V a ( t ) i . P=-\partial_t+A^{ij}\partial_{p_i}\partial_{p_j}+\left(\partial_{p_i}A^{ij}+V^j_{a(t)}\right)\partial_{p_j}+\partial_{p_i}V^i_{a(t)}. P = − ∂ t + A ij ∂ p i ∂ p j + ( ∂ p i A ij + V a ( t ) j ) ∂ p j + ∂ p i V a ( t ) i .
Consequently,
P = ∑ k = 1 3 X k 2 + X 0 + c , P=\sum_{k=1}^{3}X_k^2+X_0+c, P = k = 1 ∑ 3 X k 2 + X 0 + c ,
where
X 0 = − ∂ t + Y , X_0=-\partial_t+Y, X 0 = − ∂ t + Y ,
with
Y = ∑ j = 1 3 ( ∂ p i A i j + V a ( t ) j − b j ) ∂ p j . Y=\sum_{j=1}^{3}\left(\partial_{p_i}A^{ij}+V^j_{a(t)}-b^j\right)\partial_{p_j}. Y = j = 1 ∑ 3 ( ∂ p i A ij + V a ( t ) j − b j ) ∂ p j .
and
c = ∂ p i V a ( t ) i . c = \partial_{p_i}V^i_{a(t)}. c = ∂ p i V a ( t ) i .
Thus X 0 , X 1 , X 2 , X 3 X_0,X_1,X_2,X_3 X 0 , X 1 , X 2 , X 3 are smooth vector fields on Q ′ Q' Q ′ , and c c c is a smooth function. Since A A A is positive definite, S S S is invertible. Therefore
span { X 1 , X 2 , X 3 } = span { ∂ p 1 , ∂ p 2 , ∂ p 3 } . \operatorname{span}\{X_1,X_2,X_3\} = \operatorname{span}\{\partial_{p_1},\partial_{p_2},\partial_{p_3}\}. span { X 1 , X 2 , X 3 } = span { ∂ p 1 , ∂ p 2 , ∂ p 3 } .
Since X 0 X_0 X 0 has a nonzero component in the time direction, it follows that
span { X 0 , X 1 , X 2 , X 3 } = T ( t , p ) ( R × R 3 ) \operatorname{span}\{X_0,X_1,X_2,X_3\} = T_{(t,p)}(\mathbb{R} \times\mathbb{R}^3) span { X 0 , X 1 , X 2 , X 3 } = T ( t , p ) ( R × R 3 )
at every ( t , p ) ∈ Q ′ (t,p) \in Q' ( t , p ) ∈ Q ′ . Thus, Hörmander’s condition is satisfied. Finally,
G ∈ C ( [ 0 , ∞ ) ; L 1 ( R 3 ) ) G \in C([0,\infty);L^1(\mathbb{R}^3)) G ∈ C ([ 0 , ∞ ) ; L 1 ( R 3 ))
implies G ∈ L l o c 1 ( Q ′ ) G \in L^1_{\mathrm{loc}}(Q') G ∈ L loc 1 ( Q ′ ) , and hence G G G defines a distribution on Q ′ Q' Q ′ . Hörmander’s hypoellipticity theorem [4 Theorem 1.1] therefore implies that G G G is smooth in the interior of Q ′ Q' Q ′ . In particular,
G ∈ C ∞ ( ( τ , T ) × B R ) . G \in C^\infty((\tau,T)\times B_R). G ∈ C ∞ (( τ , T ) × B R ) .
Since τ , T , R \tau,T,R τ , T , R are arbitrary, we conclude that
G ∈ C ∞ ( ( 0 , ∞ ) × R 3 ) . (93) G \in C^\infty((0,\infty)\times\mathbb{R}^3). \tag*{(93)} G ∈ C ∞ (( 0 , ∞ ) × R 3 ) . (93)
Strict positivity
We prove that G G G is strictly positive by using the parabolic strong maximum principle. Fix t > 0 t>0 t > 0 and p ∗ ∈ R 3 p_* \in\mathbb{R}^3 p ∗ ∈ R 3 , and set s = t / 2 s=t/2 s = t /2 . Since G G G has unit mass and is smooth by (92), there exists p 0 ∈ R 3 p_0 \in\mathbb{R}^3 p 0 ∈ R 3 such that
G ( s , p 0 ) > 0. G(s,p_0)>0. G ( s , p 0 ) > 0.
Choose R > max { ∣ p 0 ∣ , ∣ p ∗ ∣ } R>\max\{|p_0|,|p_*|\} R > max { ∣ p 0 ∣ , ∣ p ∗ ∣ } and set Ω = B R \Omega=B_R Ω = B R . On [ s / 2 , 2 t ] × Ω ‾ [s/2,2t]\times\overline{\Omega} [ s /2 , 2 t ] × Ω , the equation is uniformly parabolic with smooth coefficients. Writing
ϕ ( τ ) = Φ G ( τ ) , a ( τ ) = j G ( τ ) , \phi(\tau)=\Phi_G(\tau), \qquad a(\tau)=j_G(\tau), ϕ ( τ ) = Φ G ( τ ) , a ( τ ) = j G ( τ ) ,
the equation takes the form
∂ τ G = ϕ D : ∇ p 2 G + ( 3 ϕ v + V a ) ⋅ ∇ p G + ( ∇ p ⋅ V a ) G . \partial_\tau G=\phi D:\nabla_p^2G+(3\phi v+V_a)\cdot\nabla_pG+(\nabla_p\cdot V_a)G. ∂ τ G = ϕ D : ∇ p 2 G + ( 3 ϕ v + V a ) ⋅ ∇ p G + ( ∇ p ⋅ V a ) G .
Since the coefficients are smooth, we may choose C > 0 C>0 C > 0 such that
∇ p ⋅ V a ( τ ) ( p ) − C < 0 \nabla_p\cdot V_{a(\tau)}(p)-C<0 ∇ p ⋅ V a ( τ ) ( p ) − C < 0
on [ s / 2 , 2 t ] × Ω ‾ [s/2,2t]\times\overline{\Omega} [ s /2 , 2 t ] × Ω . Define
H ( τ , p ) = e − C τ G ( τ , p ) . H(\tau,p)=e^{-C\tau}G(\tau,p). H ( τ , p ) = e − C τ G ( τ , p ) .
Then H ≥ 0 H \ge0 H ≥ 0 and
∂ τ H = ϕ D : ∇ p 2 H + ( 3 ϕ v + V a ) ⋅ ∇ p H + ( ∇ p ⋅ V a − C ) H . \partial_{\tau} H = \phi D:\nabla_{p}^{2}H + (3\phi v + V_a)\cdot\nabla_p H + (\nabla_p\cdot V_a - C)H. ∂ τ H = ϕ D : ∇ p 2 H + ( 3 ϕ v + V a ) ⋅ ∇ p H + ( ∇ p ⋅ V a − C ) H .
Suppose that G ( t , p ∗ ) = 0 G(t,p_*)=0 G ( t , p ∗ ) = 0 . Then H ( t , p ∗ ) = 0 H(t,p_*)=0 H ( t , p ∗ ) = 0 , so − H -H − H attains its maximum value zero at ( t , p ∗ ) (t,p_*) ( t , p ∗ ) . By the parabolic strong maximum principle [6 Chapter II, Theorem 2.7] , H H H must vanish throughout ( s / 2 , t ] × Ω (s/2,t]\times\Omega ( s /2 , t ] × Ω . In particular, H ( s , p 0 ) = 0 H(s,p_0)=0 H ( s , p 0 ) = 0 , contradicting G ( s , p 0 ) > 0 G(s,p_0)>0 G ( s , p 0 ) > 0 . Therefore G ( t , p ∗ ) > 0 G(t,p_*)>0 G ( t , p ∗ ) > 0 . Since t > 0 t>0 t > 0 and p ∗ ∈ R 3 p_*\in\mathbb{R}^3 p ∗ ∈ R 3 were arbitrary, we conclude that
G ( t , p ) > 0 , t > 0 , p ∈ R 3 . G(t,p)>0,\qquad t>0,\quad p\in\mathbb{R}^3. G ( t , p ) > 0 , t > 0 , p ∈ R 3 .
This completes the proof of Theorem 2.
Proof of Theorem 3
In this section we prove Theorem 3. The main ingredient is a comparison principle for the particle energy distributions of even solutions.
Assume that the initial data is an even probability density. Since the PDE (1) is invariant under p ↦ − p p\mapsto-p p ↦ − p , uniqueness implies that
G ( t , p ) = G ( t , − p ) , for all t > 0 G(t,p)=G(t,-p),\qquad\text{for all }t>0 G ( t , p ) = G ( t , − p ) , for all t > 0
and therefore
j G ( t ) = 0 , for all t ≥ 0. j_G(t)=0,\qquad\text{for all }t\ge0. j G ( t ) = 0 , for all t ≥ 0.
For an even solution G G G , define
ρ G ( t , z ) = z z 2 − 1 ∫ S 2 G ( t , z 2 − 1 ω ) d ω , z > 1. \rho_G(t,z)=z\sqrt{z^2-1}\int_{\mathbb{S}^2}G\left(t,\sqrt{z^2-1}\,\omega\right)\,d\omega,\qquad z>1. ρ G ( t , z ) = z z 2 − 1 ∫ S 2 G ( t , z 2 − 1 ω ) d ω , z > 1.
The function ρ G \rho_G ρ G is the probability density of the particle energy z = ⟨ p ⟩ z=\langle p\rangle z = ⟨ p ⟩ . Indeed, passing to spherical coordinates and using ∣ p ∣ = z 2 − 1 |p|=\sqrt{z^2-1} ∣ p ∣ = z 2 − 1 , we obtain
∫ R 3 f ( ⟨ p ⟩ ) G ( t , p ) d p = ∫ 1 ∞ f ( z ) ρ G ( t , z ) d z \int_{\mathbb{R}^3}f(\langle p\rangle)G(t,p)\,dp=\int_1^\infty f(z)\rho_G(t,z)\,dz ∫ R 3 f (⟨ p ⟩) G ( t , p ) d p = ∫ 1 ∞ f ( z ) ρ G ( t , z ) d z
for every smooth function f f f such that the integral in the left hand side is finite. In particular, taking f = 1 f=1 f = 1 and f = z f=z f = z gives
∫ 1 ∞ ρ G ( t , z ) d z = 1 , ∫ 1 ∞ z ρ G ( t , z ) d z = E . (94) \int_1^\infty\rho_G(t,z)\,dz=1,\qquad\int_1^\infty z\rho_G(t,z)\,dz=E. \tag*{(94)} ∫ 1 ∞ ρ G ( t , z ) d z = 1 , ∫ 1 ∞ z ρ G ( t , z ) d z = E . (94)
The tail distribution of ρ G \rho_G ρ G is the decreasing function
F G ( t , z ) = ∫ z ∞ ρ G ( t , s ) d s , z ≥ 1. F_G(t,z)=\int_z^\infty\rho_G(t,s)\,ds,\qquad z\ge1. F G ( t , z ) = ∫ z ∞ ρ G ( t , s ) d s , z ≥ 1.
By (94) and Tonelli’s theorem,
F G ( t , 1 ) = 1 , ∫ 1 ∞ F G ( t , z ) d z = E − 1. (95) F_G(t,1)=1,\qquad\int_1^\infty F_G(t,z)\,\mathrm{d}z=E-1. \tag*{(95)} F G ( t , 1 ) = 1 , ∫ 1 ∞ F G ( t , z ) d z = E − 1. (95)
Consequently,
Φ G ( t ) = μ 3 ∫ 1 ∞ ( 1 − z − 2 ) ρ G ( t , z ) d z . \Phi_G(t)=\frac{\mu}{3}\int_1^\infty(1-z^{-2})\rho_G(t,z)\,\mathrm{d}z. Φ G ( t ) = 3 μ ∫ 1 ∞ ( 1 − z − 2 ) ρ G ( t , z ) d z .
Integration by parts gives
Φ G ( t ) = 2 μ 3 ∫ 1 ∞ z − 3 F G ( t , z ) d z . (96) \Phi_G(t)=\frac{2\mu}{3}\int_1^\infty z^{-3}F_G(t,z)\,\mathrm{d}z. \tag*{(96)} Φ G ( t ) = 3 2 μ ∫ 1 ∞ z − 3 F G ( t , z ) d z . (96)
A direct calculation using (1) shows that ρ G \rho_G ρ G satisfies
∂ t ρ G = Φ G ( t ) ∂ z 2 [ ( z − z − 1 ) ρ G ] − ∂ z [ ( 3 Φ G ( t ) − μ ( 1 − z − 2 ) ) ρ G ] . (97) \partial_t\rho_G=\Phi_G(t)\partial_z^2\left[(z-z^{-1})\rho_G\right]-\partial_z\left[\left(3\Phi_G(t)-\mu(1-z^{-2})\right)\rho_G\right]. \tag*{(97)} ∂ t ρ G = Φ G ( t ) ∂ z 2 [ ( z − z − 1 ) ρ G ] − ∂ z [ ( 3 Φ G ( t ) − μ ( 1 − z − 2 ) ) ρ G ] . (97)
Integrating this equation from z z z to infinity yields
∂ t F G = L Φ G ( t ) F G , (98) \partial_tF_G=\mathcal{L}_{\Phi_G(t)}F_G, \tag*{(98)} ∂ t F G = L Φ G ( t ) F G , (98)
where
L ϕ f = ϕ ( z − z − 1 ) f ′ ′ + [ μ ( 1 − z − 2 ) − ϕ ( 2 − z − 2 ) ] f ′ . (99) \mathcal{L}_{\phi}f=\phi(z-z^{-1})f^{\prime\prime}+\left[\mu(1-z^{-2})-\phi(2-z^{-2})\right]f^{\prime}. \tag*{(99)} L ϕ f = ϕ ( z − z − 1 ) f ′′ + [ μ ( 1 − z − 2 ) − ϕ ( 2 − z − 2 ) ] f ′ . (99)
These equations hold without assuming spherical symmetry of G G G .
We first establish the uniform integrability of the energy tails when 1 < E < 2 1<E<2 1 < E < 2 .
Lemma 22. Let G 0 G_0 G 0 be an even probability density with finite energy satisfying 1 < E < 2 1<E<2 1 < E < 2 , and let G G G be the global solution of (1) with initial data G 0 G_0 G 0 . Then
lim R → ∞ sup t ≥ 0 T R ( G ( t ) ) = 0. (100) \lim_{R\to\infty}\sup_{t\geq0}\mathcal{T}_R(G(t))=0. \tag*{(100)} R → ∞ lim t ≥ 0 sup T R ( G ( t )) = 0. (100) Consequently,
inf t ≥ 0 Φ G ( t ) > 0. \inf_{t\geq0}\Phi_G(t)>0. t ≥ 0 inf Φ G ( t ) > 0. Proof. By (18) and conservation of energy,
0 < Φ G ( t ) ≤ ϕ ∗ , ϕ ∗ = μ 3 ( 1 − E − 2 ) . 0<\Phi_G(t)\leq\phi_*,\qquad\phi_*=\frac{\mu}{3}(1-E^{-2}). 0 < Φ G ( t ) ≤ ϕ ∗ , ϕ ∗ = 3 μ ( 1 − E − 2 ) . Set
α = μ ϕ ∗ = 3 1 − E − 2 . \alpha=\frac{\mu}{\phi_*}=\frac{3}{1-E^{-2}}. α = ϕ ∗ μ = 1 − E − 2 3 . Since 1 < E < 2 1<E<2 1 < E < 2 , we have α > 4 \alpha>4 α > 4 . Define
π ( z ) = z 1 − α z 2 − 1 , S ( z ) = ∫ z ∞ π ( s ) d s . \pi(z)=z^{1-\alpha}\sqrt{z^2-1},\qquad S(z)=\int_z^\infty\pi(s)\,\mathrm{d}s. π ( z ) = z 1 − α z 2 − 1 , S ( z ) = ∫ z ∞ π ( s ) d s .
A direct calculation using (99) gives L ϕ ∗ S = 0 \mathcal{L}_{\phi_*}S=0 L ϕ ∗ S = 0 . Moreover, for every 0 < ϕ ≤ ϕ ∗ 0<\phi\le\phi_* 0 < ϕ ≤ ϕ ∗ ,
L ϕ S = ϕ ϕ ∗ L ϕ ∗ S + μ ( 1 − ϕ ϕ ∗ ) ( 1 − z − 2 ) S ′ ( z ) = − μ ( 1 − ϕ ϕ ∗ ) ( 1 − z − 2 ) π ( z ) ≤ 0. \mathcal{L}_{\phi}S=\frac{\phi}{\phi_*}\mathcal{L}_{\phi_*}S+\mu\left(1-\frac{\phi}{\phi_*}\right)(1-z^{-2})S'(z)=-\mu\left(1-\frac{\phi}{\phi_*}\right)(1-z^{-2})\pi(z)\le0. L ϕ S = ϕ ∗ ϕ L ϕ ∗ S + μ ( 1 − ϕ ∗ ϕ ) ( 1 − z − 2 ) S ′ ( z ) = − μ ( 1 − ϕ ∗ ϕ ) ( 1 − z − 2 ) π ( z ) ≤ 0.
Fix τ > 0 \tau>0 τ > 0 and write
ρ τ ( z ) = ρ G ( τ , z ) , F τ ( z ) = F G ( τ , z ) . \rho_\tau(z)=\rho_G(\tau,z),\qquad F_\tau(z)=F_G(\tau,z). ρ τ ( z ) = ρ G ( τ , z ) , F τ ( z ) = F G ( τ , z ) .
Define
A τ ( z ) = ∫ 1 z ρ τ ( s ) S ( s ) d s , F ‾ τ ( z ) = F τ ( z ) + S ( z ) A τ ( z ) . A_\tau(z)=\int_1^z\frac{\rho_\tau(s)}{S(s)}\,ds,\qquad\overline{F}_\tau(z)=F_\tau(z)+S(z)A_\tau(z). A τ ( z ) = ∫ 1 z S ( s ) ρ τ ( s ) d s , F τ ( z ) = F τ ( z ) + S ( z ) A τ ( z ) .
Since τ > 0 \tau>0 τ > 0 , and by the regularity of G G G proved in Theorem 2 , the functions ρ τ \rho_\tau ρ τ and F τ F_\tau F τ are smooth on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) , and consequently F ‾ τ ∈ C ∞ ( ( 1 , ∞ ) ) \overline{F}_\tau\in C^\infty((1,\infty)) F τ ∈ C ∞ (( 1 , ∞ )) . By construction,
F ‾ τ ( 1 ) = 1 , F ‾ τ ( z ) ≥ F τ ( z ) ≥ 0. \overline{F}_\tau(1)=1,\qquad\overline{F}_\tau(z)\ge F_\tau(z)\ge0. F τ ( 1 ) = 1 , F τ ( z ) ≥ F τ ( z ) ≥ 0.
Differentiating, we obtain
F ‾ τ ′ ( z ) = − π ( z ) A τ ( z ) ≤ 0 \overline{F}_\tau'(z)=-\pi(z)A_\tau(z)\le0 F τ ′ ( z ) = − π ( z ) A τ ( z ) ≤ 0
and
L ϕ F ‾ τ = A τ L ϕ S − ϕ ( z − z − 1 ) π ( z ) S ( z ) ρ τ ( z ) ≤ 0 \mathcal{L}_{\phi}\overline{F}_\tau=A_\tau\mathcal{L}_{\phi}S-\phi(z-z^{-1})\frac{\pi(z)}{S(z)}\rho_\tau(z)\le0 L ϕ F τ = A τ L ϕ S − ϕ ( z − z − 1 ) S ( z ) π ( z ) ρ τ ( z ) ≤ 0
for every 0 < ϕ ≤ ϕ ∗ 0<\phi\le\phi_* 0 < ϕ ≤ ϕ ∗ . Thus, F ‾ τ \overline{F}_\tau F τ is a smooth stationary supersolution of (98) on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) . We now apply the classical parabolic comparison principle. Fix T > τ T>\tau T > τ and 1 + ε < L 1+\varepsilon<L 1 + ε < L , and set
w ( t , z ) = F G ( t , z ) − F ‾ τ ( z ) . w(t,z)=F_G(t,z)-\overline{F}_\tau(z). w ( t , z ) = F G ( t , z ) − F τ ( z ) .
Since Φ G ( t ) ≤ ϕ ∗ \Phi_G(t)\le\phi_* Φ G ( t ) ≤ ϕ ∗ , we have
∂ t w − L Φ G ( t ) w = L Φ G ( t ) F ‾ τ ≤ 0. \partial_t w-\mathcal{L}_{\Phi_G(t)}w=\mathcal{L}_{\Phi_G(t)}\overline{F}_\tau\le0. ∂ t w − L Φ G ( t ) w = L Φ G ( t ) F τ ≤ 0.
By Theorem 1 , Φ G \Phi_G Φ G is bounded away from zero on [ 0 , T ] [0,T] [ 0 , T ] , hence the equation is uniformly parabolic on [ τ , T ] × [ 1 + ε , L ] [\tau,T]\times[1+\varepsilon,L] [ τ , T ] × [ 1 + ε , L ] . At the initial time, w ( τ , z ) ≤ 0 w(\tau,z)\le0 w ( τ , z ) ≤ 0 . At the left boundary, since F G ( t , 1 ) = 1 F_G(t,1)=1 F G ( t , 1 ) = 1 ,
w ( t , 1 + ε ) ≤ 1 − F ‾ τ ( 1 + ε ) . w(t,1+\varepsilon)\le1-\overline{F}_\tau(1+\varepsilon). w ( t , 1 + ε ) ≤ 1 − F τ ( 1 + ε ) .
At the right boundary, conservation of energy gives
w ( t , L ) ≤ F G ( t , L ) ≤ E L . w(t,L)\le F_G(t,L)\le\frac{E}{L}. w ( t , L ) ≤ F G ( t , L ) ≤ L E .
The classical parabolic comparison principle implies
w ( t , z ) ≤ max { 1 − F ‾ τ ( 1 + ε ) , E L } . w(t,z)\le\max\left\{1-\overline{F}_\tau(1+\varepsilon),\frac{E}{L}\right\}. w ( t , z ) ≤ max { 1 − F τ ( 1 + ε ) , L E } .
for τ ≤ t ≤ T \tau\le t \le T τ ≤ t ≤ T and 1 + ε ≤ z ≤ L 1+\varepsilon\le z \le L 1 + ε ≤ z ≤ L . Since F ‾ τ \overline{F}_{\tau} F τ is continuous at z = 1 z=1 z = 1 and F ‾ τ ( 1 ) = 1 \overline{F}_{\tau}(1)=1 F τ ( 1 ) = 1 , letting ε → 0 \varepsilon\to0 ε → 0 and L → ∞ L \to\infty L → ∞ yields
F G ( t , z ) ≤ F ‾ τ ( z ) , t ≥ τ , z ≥ 1. F_G(t,z) \le\overline{F}_{\tau}(z), \qquad t \ge\tau,\quad z \ge1. F G ( t , z ) ≤ F τ ( z ) , t ≥ τ , z ≥ 1.
We now let τ → 0 \tau\to0 τ → 0 . Since G ( τ ) → G 0 G(\tau) \to G_0 G ( τ ) → G 0 in L 1 ( R 3 ) L^1(\mathbb{R}^3) L 1 ( R 3 ) , we have
ρ τ → ρ 0 in L 1 ( ( 1 , ∞ ) ) , \rho_{\tau} \to\rho_0 \quad\text{in } L^1((1,\infty)), ρ τ → ρ 0 in L 1 (( 1 , ∞ )) ,
where ρ 0 \rho_0 ρ 0 is the particle energy probability density associated with G 0 G_0 G 0 . In particular,
F τ ( z ) → F 0 ( z ) F_{\tau}(z) \to F_0(z) F τ ( z ) → F 0 ( z )
for every z ≥ 1 z \ge1 z ≥ 1 . Moreover, since S S S is strictly positive on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , for every fixed z > 1 z>1 z > 1 ,
A τ ( z ) → A 0 ( z ) : = ∫ 1 z ρ 0 ( s ) S ( s ) d s . A_{\tau}(z) \to A_0(z) := \int_1^z \frac{\rho_0(s)}{S(s)}\,\mathrm{d}s. A τ ( z ) → A 0 ( z ) := ∫ 1 z S ( s ) ρ 0 ( s ) d s .
Consequently,
F ‾ τ ( z ) → F ‾ ( z ) : = F 0 ( z ) + S ( z ) A 0 ( z ) . \overline{F}_{\tau}(z) \to\overline{F}(z) := F_0(z) + S(z)A_0(z). F τ ( z ) → F ( z ) := F 0 ( z ) + S ( z ) A 0 ( z ) .
We conclude that
F G ( t , z ) ≤ F ‾ ( z ) , t ≥ 0 , z ≥ 1. (101) F_G(t,z) \le\overline{F}(z), \qquad t \ge0,\quad z \ge1. \tag*{(101)} F G ( t , z ) ≤ F ( z ) , t ≥ 0 , z ≥ 1. (101)
It remains to prove that F ‾ \overline{F} F has uniformly integrable energy tails. Define
ρ ‾ ( z ) = − F ‾ ′ ( z ) = π ( z ) ∫ 1 z ρ 0 ( s ) S ( s ) d s . \overline{\rho}(z) = -\overline{F}'(z) = \pi(z)\int_1^z \frac{\rho_0(s)}{S(s)}\,\mathrm{d}s. ρ ( z ) = − F ′ ( z ) = π ( z ) ∫ 1 z S ( s ) ρ 0 ( s ) d s .
By Tonelli’s theorem,
∫ 1 ∞ ρ ‾ ( z ) d z = ∫ 1 ∞ ρ 0 ( s ) S ( s ) ( ∫ s ∞ π ( z ) d z ) d s = 1. \int_1^\infty\overline{\rho}(z)\,\mathrm{d}z
= \int_1^\infty\frac{\rho_0(s)}{S(s)}
\left(\int_s^\infty\pi(z)\,\mathrm{d}z\right)\mathrm{d}s
= 1. ∫ 1 ∞ ρ ( z ) d z = ∫ 1 ∞ S ( s ) ρ 0 ( s ) ( ∫ s ∞ π ( z ) d z ) d s = 1.
Since F ‾ ( 1 ) = 1 \overline{F}(1)=1 F ( 1 ) = 1 , this implies
F ‾ ( z ) = ∫ z ∞ ρ ‾ ( s ) d s . \overline{F}(z) = \int_z^\infty\overline{\rho}(s)\,\mathrm{d}s. F ( z ) = ∫ z ∞ ρ ( s ) d s .
Furthermore,
∫ 1 ∞ z ρ ‾ ( z ) d z = ∫ 1 ∞ ρ 0 ( s ) S ( s ) ( ∫ s ∞ z π ( z ) d z ) d s . \int_1^\infty z\overline{\rho}(z)\,\mathrm{d}z
= \int_1^\infty\frac{\rho_0(s)}{S(s)}
\left(\int_s^\infty z\pi(z)\,\mathrm{d}z\right)\mathrm{d}s. ∫ 1 ∞ z ρ ( z ) d z = ∫ 1 ∞ S ( s ) ρ 0 ( s ) ( ∫ s ∞ z π ( z ) d z ) d s .
Since π ( z ) ∼ z 2 − α \pi(z) \sim z^{2-\alpha} π ( z ) ∼ z 2 − α as z → ∞ z \to\infty z → ∞ and α > 4 \alpha>4 α > 4 , there exists a constant C α < ∞ C_{\alpha}<\infty C α < ∞ such that
1 S ( s ) ∫ s ∞ z π ( z ) d z ≤ C α s , s ≥ 1. \frac{1}{S(s)}\int_s^\infty z\pi(z)\,\mathrm{d}z \le C_{\alpha}s,\qquad s \ge1. S ( s ) 1 ∫ s ∞ z π ( z ) d z ≤ C α s , s ≥ 1.
Consequently,
∫ 1 ∞ z ρ ‾ ( z ) d z ≤ C α ∫ 1 ∞ s ρ 0 ( s ) d s = C α E < ∞ . \int_1^\infty z\overline{\rho}(z)\,\mathrm{d}z
\le C_{\alpha}\int_1^\infty s\rho_0(s)\,\mathrm{d}s
= C_{\alpha}E < \infty. ∫ 1 ∞ z ρ ( z ) d z ≤ C α ∫ 1 ∞ s ρ 0 ( s ) d s = C α E < ∞.
Finally, by (101) ,
T R ( G ( t ) ) = R F G ( t , R ) + ∫ R ∞ F G ( t , z ) d z ≤ R F ‾ ( R ) + ∫ R ∞ F ‾ ( z ) d z = ∫ R ∞ z ρ ‾ ( z ) d z . \mathcal{T}_{R}(G(t))=RF_{G}(t,R)+\int_{R}^{\infty}F_{G}(t,z)\,\mathrm{d}z\leq R\overline{F}(R)+\int_{R}^{\infty}\overline{F}(z)\,\mathrm{d}z=\int_{R}^{\infty}z\overline{\rho}(z)\,\mathrm{d}z. T R ( G ( t )) = R F G ( t , R ) + ∫ R ∞ F G ( t , z ) d z ≤ R F ( R ) + ∫ R ∞ F ( z ) d z = ∫ R ∞ z ρ ( z ) d z .
The right hand side is independent of time and tends to zero as R → ∞ R\to\infty R → ∞ . Therefore
lim R → ∞ sup t ≥ 0 T R ( G ( t ) ) = 0. \lim_{R\to\infty}\sup_{t\geq0}\mathcal{T}_{R}(G(t))=0. R → ∞ lim t ≥ 0 sup T R ( G ( t )) = 0.
The uniform positivity of Φ G \Phi_{G} Φ G now follows from Corollary 11 .
We next show that radial monotonicity is preserved by the equation.
Lemma 23. Let H 0 ∈ L 1 ( R 3 ) H_{0}\in L^{1}(\mathbb{R}^{3}) H 0 ∈ L 1 ( R 3 ) be a radially nonincreasing probability density with finite energy, and let H H H be the corresponding solution of (1) . Then
H ( t , p ) = h ( t , ∣ p ∣ ) , H(t,p)=h(t,|p|), H ( t , p ) = h ( t , ∣ p ∣ ) , where
∂ r h ( t , r ) ≤ 0 , t > 0 , r > 0. \partial_{r}h(t,r)\leq0,\qquad t>0,\ r>0. ∂ r h ( t , r ) ≤ 0 , t > 0 , r > 0.
Proof. Rotational invariance and uniqueness imply that H ( t , p ) = h ( t , r ) H(t,p)=h(t,r) H ( t , p ) = h ( t , r ) , where r = ∣ p ∣ r=|p| r = ∣ p ∣ . In particular, j H ( t ) = 0 j_{H}(t)=0 j H ( t ) = 0 . Set
z = 1 + r 2 , ϕ ( t ) = Φ H ( t ) . z=\sqrt{1+r^{2}},\qquad\phi(t)=\Phi_{H}(t). z = 1 + r 2 , ϕ ( t ) = Φ H ( t ) .
Then h h h satisfies
∂ t h = ϕ ( t ) z ∂ r 2 h + [ ϕ ( t ) ( 2 z r + r z ) + μ r z ] ∂ r h + μ ( 2 z + 1 z 3 ) h . (102) \partial_{t}h=\phi(t)z\partial_{r}^{2}h+\left[\phi(t)\left(\frac{2z}{r}+\frac{r}{z}\right)+\frac{\mu r}{z}\right]\partial_{r}h+\mu\left(\frac{2}{z}+\frac{1}{z^{3}}\right)h. \tag*{(102)} ∂ t h = ϕ ( t ) z ∂ r 2 h + [ ϕ ( t ) ( r 2 z + z r ) + z μ r ] ∂ r h + μ ( z 2 + z 3 1 ) h . (102)
Set q = ∂ r h q=\partial_{r}h q = ∂ r h . Differentiating (102) gives
∂ t q = ϕ z ∂ r 2 q + [ ϕ ( 2 z r + 2 r z ) + μ r z ] ∂ r q + [ ϕ ( 1 z 3 − 2 z r 2 ) + μ ( 2 z + 2 z 3 ) ] q − μ r ( 2 z 3 + 3 z 5 ) h . (103) \begin{aligned}
\partial_{t}q={}&\phi z\partial_{r}^{2}q+\left[\phi\left(\frac{2z}{r}+\frac{2r}{z}\right)+\frac{\mu r}{z}\right]\partial_{r}q\\
&+\left[\phi\left(\frac{1}{z^{3}}-\frac{2}{zr^{2}}\right)+\mu\left(\frac{2}{z}+\frac{2}{z^{3}}\right)\right]q-\mu r\left(\frac{2}{z^{3}}+\frac{3}{z^{5}}\right)h. \tag*{(103)}
\end{aligned} ∂ t q = ϕ z ∂ r 2 q + [ ϕ ( r 2 z + z 2 r ) + z μ r ] ∂ r q + [ ϕ ( z 3 1 − z r 2 2 ) + μ ( z 2 + z 3 2 ) ] q − μ r ( z 3 2 + z 5 3 ) h . (103)
Fix T > 0 T>0 T > 0 . By Theorem 1 ,
0 < δ T ≤ ϕ ( t ) ≤ μ 3 , 0 ≤ t ≤ T . 0<\delta_{T}\leq\phi(t)\leq\frac{\mu}{3},\qquad0\leq t\leq T. 0 < δ T ≤ ϕ ( t ) ≤ 3 μ , 0 ≤ t ≤ T .
Moreover, Lemmas 15 and 12 , applied to H H H with frozen coefficients ( ϕ , 0 ) (\phi,0) ( ϕ , 0 ) , imply that ϕ \phi ϕ is Hölder continuous on [ 0 , T ] [0,T] [ 0 , T ] . Throughout the following approximation argument, we keep this coefficient function ϕ \phi ϕ fixed.
We first prove preservation of radial monotonicity for the frozen equation with smooth, compactly supported initial data. Let f ∈ C c ∞ ( R 3 ) f \in C^\infty_c(\mathbb{R}^3) f ∈ C c ∞ ( R 3 ) be a nonnegative, radially nonincreasing probability density, and let u u u be the solution of
∂ t u = ϕ ( t ) ∇ p ⋅ ( D ∇ p u ) + μ ∇ p ⋅ ( v u ) , u ( 0 , p ) = f ( p ) . \partial_t u = \phi(t)\nabla_p \cdot(D\nabla_p u) + \mu\nabla_p \cdot(vu), \qquad u(0,p) = f(p). ∂ t u = ϕ ( t ) ∇ p ⋅ ( D ∇ p u ) + μ ∇ p ⋅ ( v u ) , u ( 0 , p ) = f ( p ) .
For R R R sufficiently large that supp f ⊂ B R \operatorname{supp} f \subset B_R supp f ⊂ B R , let u R u^R u R solve the corresponding Dirichlet problem
{ ∂ t u R = ϕ ( t ) ∇ p ⋅ ( D ∇ p u R ) + μ ∇ p ⋅ ( v u R ) , ( t , p ) ∈ ( 0 , T ] × B R , u R ( t , p ) = 0 , ( t , p ) ∈ [ 0 , T ] × ∂ B R , u R ( 0 , p ) = f ( p ) , p ∈ B R . \begin{cases}
\partial_t u^R = \phi(t)\nabla_p \cdot(D\nabla_p u^R) + \mu\nabla_p \cdot(vu^R), & (t,p) \in(0,T] \times B_R,\\
u^R(t,p) = 0, & (t,p) \in[0,T] \times\partial B_R,\\
u^R(0,p) = f(p), & p \in B_R.
\end{cases} ⎩ ⎨ ⎧ ∂ t u R = ϕ ( t ) ∇ p ⋅ ( D ∇ p u R ) + μ ∇ p ⋅ ( v u R ) , u R ( t , p ) = 0 , u R ( 0 , p ) = f ( p ) , ( t , p ) ∈ ( 0 , T ] × B R , ( t , p ) ∈ [ 0 , T ] × ∂ B R , p ∈ B R .
Classical linear parabolic theory gives a nonnegative solution whose spatial gradient is continuous up to the initial time and the boundary. Here the compatibility conditions hold because f f f vanishes near ∂ B R \partial B_R ∂ B R . Rotational invariance and uniqueness give
u R ( t , p ) = k R ( t , ∣ p ∣ ) . u^R(t,p) = k^R(t,|p|). u R ( t , p ) = k R ( t , ∣ p ∣ ) .
Set
q R ( t , r ) = ∂ r k R ( t , r ) . q^R(t,r) = \partial_r k^R(t,r). q R ( t , r ) = ∂ r k R ( t , r ) .
The function q R q^R q R satisfies (103) , with h h h replaced by k R k^R k R and q q q replaced by q R q^R q R . Its forcing term is nonpositive because k R ≥ 0 k^R \ge0 k R ≥ 0 . The coefficient multiplying q R q^R q R satisfies
ϕ ( t ) ( 1 z 3 − 2 z r 2 ) + μ ( 2 z + 2 z 3 ) ≤ ϕ ( t ) + 4 μ ≤ 13 μ 3 . \phi(t)\left(\frac{1}{z^3} - \frac{2}{zr^2}\right) + \mu\left(\frac{2}{z} + \frac{2}{z^3}\right) \le\phi(t) + 4\mu\le\frac{13\mu}{3}. ϕ ( t ) ( z 3 1 − z r 2 2 ) + μ ( z 2 + z 3 2 ) ≤ ϕ ( t ) + 4 μ ≤ 3 13 μ .
Smoothness at the center gives
q R ( t , 0 ) = 0. q^R(t,0) = 0. q R ( t , 0 ) = 0.
At the outer boundary, nonnegativity of k R k^R k R and k R ( t , R ) = 0 k^R(t,R) = 0 k R ( t , R ) = 0 imply
q R ( t , R ) ≤ 0. q^R(t,R) \le0. q R ( t , R ) ≤ 0.
Moreover,
q R ( 0 , r ) ≤ 0 q^R(0,r) \le0 q R ( 0 , r ) ≤ 0
by the radial monotonicity of f f f . Apply the maximum principle to e − 5 μ t q R e^{-5\mu t}q^R e − 5 μ t q R . Its zeroth-order coefficient is strictly negative, and its forcing term is nonpositive. A positive maximum cannot occur at r = 0 r=0 r = 0 , at r = R r=R r = R , or at the initial time. At any r > 0 r>0 r > 0 , the coefficients of (103) are regular, and the equation gives a contradiction. Consequently,
∂ r k R ( t , r ) ≤ 0 , 0 ≤ t ≤ T , 0 < r < R . \partial_r k^R(t,r) \le0, \qquad0 \le t \le T,\qquad0 < r < R. ∂ r k R ( t , r ) ≤ 0 , 0 ≤ t ≤ T , 0 < r < R .
Extending u R u^R u R by zero outside B R B_R B R therefore gives a radially nonincreasing function on R 3 \mathbb{R}^3 R 3 .
We next pass to the whole space. Let P t P_t P t solve the frozen stochastic differential equation with coefficients ( ϕ , 0 ) (\phi,0) ( ϕ , 0 ) and initial distribution f ( p ) d p f(p)\,dp f ( p ) d p , and define
τ R = inf { s ≥ 0 : ∣ P s ∣ ≥ R } . \tau_R=\inf\{s\ge0:|P_s|\ge R\}. τ R = inf { s ≥ 0 : ∣ P s ∣ ≥ R } .
The Dirichlet solution is the density of this diffusion killed upon leaving B R B_R B R :
∫ A u R ( t , p ) d p = P ( P t ∈ A , t < τ R ) \int_A u^R(t,p)\,dp=\mathbb{P}(P_t\in A,\ t<\tau_R) ∫ A u R ( t , p ) d p = P ( P t ∈ A , t < τ R )
for every Borel set A ⊂ R 3 A\subset\mathbb{R}^3 A ⊂ R 3 , where u R u^R u R denotes its zero extension. Since u ( t , ⋅ ) u(t,\cdot) u ( t , ⋅ ) is the density of P t P_t P t , we have 0 ≤ u R ( t , ⋅ ) ≤ u ( t , ⋅ ) 0\le u^R(t,\cdot)\le u(t,\cdot) 0 ≤ u R ( t , ⋅ ) ≤ u ( t , ⋅ ) almost everywhere, and hence
∥ u ( t ) − u R ( t ) ∥ L 1 = 1 − P ( t < τ R ) = P ( τ R ≤ t ) . \|u(t)-u^R(t)\|_{L^1}=1-\mathbb{P}(t<\tau_R)=\mathbb{P}(\tau_R\le t). ∥ u ( t ) − u R ( t ) ∥ L 1 = 1 − P ( t < τ R ) = P ( τ R ≤ t ) .
The diffusion is global and nonexplosive. Its continuous paths therefore imply τ R → ∞ \tau_R\to\infty τ R → ∞ almost surely as R → ∞ R\to\infty R → ∞ . Consequently,
sup 0 ≤ t ≤ T ∥ u ( t ) − u R ( t ) ∥ L 1 ≤ P ( τ R ≤ T ) ⟶ 0. \sup_{0\le t\le T}\|u(t)-u^R(t)\|_{L^1}\le\mathbb{P}(\tau_R\le T)\longrightarrow0. 0 ≤ t ≤ T sup ∥ u ( t ) − u R ( t ) ∥ L 1 ≤ P ( τ R ≤ T ) ⟶ 0.
The cone of nonnegative, radially nonincreasing functions is closed in L 1 ( R 3 ) L^1(\mathbb{R}^3) L 1 ( R 3 ) . Thus u ( t , ⋅ ) u(t,\cdot) u ( t , ⋅ ) is radially nonincreasing for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] .
Finally, choose nonnegative, radially nonincreasing probability densities f n ∈ C c ∞ ( R 3 ) f_n\in C^\infty_c(\mathbb{R}^3) f n ∈ C c ∞ ( R 3 ) such that
f n ⟶ H 0 in L 1 ( R 3 ) . f_n\longrightarrow H_0\quad\text{in }L^1(\mathbb{R}^3). f n ⟶ H 0 in L 1 ( R 3 ) .
Such an approximation is obtained by radial truncation, convolution with a smooth radially nonincreasing mollifier, and normalization. Let u n u_n u n solve the frozen equation with coefficients ( ϕ , 0 ) (\phi,0) ( ϕ , 0 ) and initial data f n f_n f n . The preceding argument shows that u n ( t , ⋅ ) u_n(t,\cdot) u n ( t , ⋅ ) is radially nonincreasing. Since ȷ ˙ H = 0 \dot{\jmath}_H=0 ˙ H = 0 , the given nonlinear solution H H H also solves this same frozen equation, with initial data H 0 H_0 H 0 . The contraction estimate (45) therefore gives
∥ u n ( t ) − H ( t ) ∥ L 1 ≤ ∥ f n − H 0 ∥ L 1 , 0 ≤ t ≤ T . \|u_n(t)-H(t)\|_{L^1}\le\|f_n-H_0\|_{L^1},\qquad0\le t\le T. ∥ u n ( t ) − H ( t ) ∥ L 1 ≤ ∥ f n − H 0 ∥ L 1 , 0 ≤ t ≤ T .
Passing to the limit shows that H ( t , ⋅ ) H(t,\cdot) H ( t , ⋅ ) is radially nonincreasing. By Theorem 2 , H H H is smooth for positive times, so this implies
∂ r h ( t , r ) ≤ 0 , 0 < t ≤ T , r > 0. \partial_r h(t,r)\le0,\qquad0<t\le T,\quad r>0. ∂ r h ( t , r ) ≤ 0 , 0 < t ≤ T , r > 0.
Since T > 0 T>0 T > 0 was arbitrary, the conclusion follows. □ \square □
We now establish the comparison principle for two even solutions.
Lemma 24. Let G G G and H H H be global weak solutions corresponding to even probability densities G 0 G_0 G 0 and H 0 H_0 H 0 with finite first moments. Assume that H 0 H_0 H 0 is radially nonincreasing and that
F G ( 0 , z ) ≥ F H ( 0 , z ) , z ≥ 1. (104) F_G(0,z) \ge F_H(0,z), \qquad z \ge1. \tag*{(104)} F G ( 0 , z ) ≥ F H ( 0 , z ) , z ≥ 1. (104) Then
F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1. (105) F_G(t,z) \ge F_H(t,z), \qquad t \ge0,\ z \ge1. \tag*{(105)} F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1. (105) Consequently,
Φ G ( t ) ≥ Φ H ( t ) , t ≥ 0. (106) \Phi_G(t) \ge\Phi_H(t), \qquad t \ge0. \tag*{(106)} Φ G ( t ) ≥ Φ H ( t ) , t ≥ 0. (106)
Proof. Set
U ( t , z ) = F G ( t , z ) − F H ( t , z ) . U(t,z) = F_G(t,z) - F_H(t,z). U ( t , z ) = F G ( t , z ) − F H ( t , z ) .
Recall the definition (98) of the operator L ϕ L_\phi L ϕ . Subtracting the equations (97) for F G F_G F G and F H F_H F H , we obtain
∂ t U = L Φ G ( t ) U + ( Φ G ( t ) − Φ H ( t ) ) K H , (107) \partial_t U = L_{\Phi_G(t)}U + \left(\Phi_G(t) - \Phi_H(t)\right)K_H, \tag*{(107)} ∂ t U = L Φ G ( t ) U + ( Φ G ( t ) − Φ H ( t ) ) K H , (107)
where
K H ( t , z ) = ( z − z − 1 ) ∂ z 2 F H − ( 2 − z − 2 ) ∂ z F H . K_H(t,z) = \left(z-z^{-1}\right)\partial_z^2F_H-\left(2-z^{-2}\right)\partial_zF_H. K H ( t , z ) = ( z − z − 1 ) ∂ z 2 F H − ( 2 − z − 2 ) ∂ z F H .
By Lemma 23,
H ( t , p ) = h ( t , ∣ p ∣ ) , ∂ r h ( t , r ) ≤ 0. H(t,p)=h(t,|p|), \qquad\partial_r h(t,r) \le0. H ( t , p ) = h ( t , ∣ p ∣ ) , ∂ r h ( t , r ) ≤ 0.
Writing r = z 2 − 1 r=\sqrt{z^2-1} r = z 2 − 1 , we have
ρ H ( t , z ) = 4 π z r h ( t , r ) , ∂ z F H = − ρ H . \rho_H(t,z)=4\pi zr h(t,r), \qquad\partial_zF_H=-\rho_H. ρ H ( t , z ) = 4 π z r h ( t , r ) , ∂ z F H = − ρ H .
A direct calculation gives
K H ( t , z ) = − 4 π z r 2 ∂ r h ( t , r ) ≥ 0. K_H(t,z)=-4\pi zr^2\partial_rh(t,r)\ge0. K H ( t , z ) = − 4 π z r 2 ∂ r h ( t , r ) ≥ 0.
Moreover,
∫ 1 ∞ K H ( t , z ) d z = − 4 π ∫ 0 ∞ r 3 ∂ r h ( t , r ) d r = 12 π ∫ 0 ∞ r 2 h ( t , r ) d r = 3. \int_1^\infty K_H(t,z)\,dz=-4\pi\int_0^\infty r^3\partial_rh(t,r)\,dr=12\pi\int_0^\infty r^2h(t,r)\,dr=3. ∫ 1 ∞ K H ( t , z ) d z = − 4 π ∫ 0 ∞ r 3 ∂ r h ( t , r ) d r = 12 π ∫ 0 ∞ r 2 h ( t , r ) d r = 3.
The boundary terms vanish because h h h is nonnegative, radially nonincreasing, and integrable. Define
U − ( t , z ) = max { − U ( t , z ) , 0 } , A ( t ) = ∫ 1 ∞ z − 3 U − ( t , z ) d z . U_-(t,z)=\max\{-U(t,z),0\}, \qquad A(t)=\int_1^\infty z^{-3}U_-(t,z)\,dz. U − ( t , z ) = max { − U ( t , z ) , 0 } , A ( t ) = ∫ 1 ∞ z − 3 U − ( t , z ) d z .
By (95),
Φ G ( t ) − Φ H ( t ) = 2 μ 3 ∫ 1 ∞ z − 3 U ( t , z ) d z , \Phi_G(t)-\Phi_H(t)=\frac{2\mu}{3}\int_1^\infty z^{-3}U(t,z)\,dz, Φ G ( t ) − Φ H ( t ) = 3 2 μ ∫ 1 ∞ z − 3 U ( t , z ) d z ,
and therefore
( Φ H ( t ) − Φ G ( t ) ) + ≤ 2 μ 3 A ( t ) . \left(\Phi_H(t)-\Phi_G(t)\right)_+\le\frac{2\mu}{3}A(t). ( Φ H ( t ) − Φ G ( t ) ) + ≤ 3 2 μ A ( t ) .
We next prove the inequality
∂ t U − ≤ L Φ G ( t ) U − + 2 μ 3 A ( t ) K H ( t , z ) (108) \partial_t U_- \le\mathcal{L}_{\Phi_G(t)} U_- + \frac{2\mu}{3} A(t) K_H(t,z) \tag*{(108)} ∂ t U − ≤ L Φ G ( t ) U − + 3 2 μ A ( t ) K H ( t , z ) (108)
in the distributional sense on ( 0 , ∞ ) × ( 1 , ∞ ) (0,\infty) \times(1,\infty) ( 0 , ∞ ) × ( 1 , ∞ ) . For ε > 0 \varepsilon> 0 ε > 0 , set
β ε ( s ) = s 2 + ε 2 − s 2 . \beta_\varepsilon(s) = \frac{\sqrt{s^2+\varepsilon^2}-s}{2}. β ε ( s ) = 2 s 2 + ε 2 − s .
Then β ε → ( ⋅ ) − \beta_\varepsilon\to(\cdot)_- β ε → ( ⋅ ) − , − 1 ≤ β ε ′ ≤ 0 -1 \le\beta_\varepsilon' \le0 − 1 ≤ β ε ′ ≤ 0 , and β ε ′ ′ ≥ 0 \beta_\varepsilon'' \ge0 β ε ′′ ≥ 0 . By (106), and using K H ≥ 0 K_H \ge0 K H ≥ 0 ,
( ∂ t − L Φ G ( t ) ) β ε ( U ) = β ε ′ ( U ) ( Φ G − Φ H ) K H − Φ G ( t ) ( z − z − 1 ) β ε ′ ′ ( U ) ∣ ∂ z U ∣ 2 ≤ ( Φ H − Φ G ) + K H ≤ 2 μ 3 A ( t ) K H . \begin{aligned}
(\partial_t-\mathcal{L}_{\Phi_G(t)})\beta_\varepsilon(U)
=\beta_\varepsilon'(U)(\Phi_G-\Phi_H)K_H-\Phi_G(t)(z-z^{-1})\beta_\varepsilon''(U)|\partial_zU|^2 \\
&\le(\Phi_H-\Phi_G)_+K_H\le\frac{2\mu}{3}A(t)K_H.
\end{aligned} ( ∂ t − L Φ G ( t ) ) β ε ( U ) = β ε ′ ( U ) ( Φ G − Φ H ) K H − Φ G ( t ) ( z − z − 1 ) β ε ′′ ( U ) ∣ ∂ z U ∣ 2 ≤ ( Φ H − Φ G ) + K H ≤ 3 2 μ A ( t ) K H .
Letting ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 in the sense of distributions on ( 0 , ∞ ) × ( 1 , ∞ ) (0,\infty)\times(1,\infty) ( 0 , ∞ ) × ( 1 , ∞ ) yields (107).
Let L ϕ ∗ \mathcal{L}_\phi^* L ϕ ∗ denote the formal adjoint of L ϕ \mathcal{L}_\phi L ϕ with respect to Lebesgue measure. A direct calculation gives
L ϕ ∗ ( z − 3 ) = 3 μ z 2 − 5 μ − 15 ϕ z 6 ≤ 3 μ z − 3 . \mathcal{L}_\phi^*(z^{-3})=\frac{3\mu z^2-5\mu-15\phi}{z^6}\le3\mu z^{-3}. L ϕ ∗ ( z − 3 ) = z 6 3 μ z 2 − 5 μ − 15 ϕ ≤ 3 μ z − 3 .
For positive times, smoothness of G G G and H H H implies
U − ( t , z ) = O ( ( z − 1 ) 3 / 2 ) as z ↓ 1 , U_-(t,z)=O((z-1)^{3/2})\qquad\text{as }z\downarrow1, U − ( t , z ) = O (( z − 1 ) 3/2 ) as z ↓ 1 ,
locally uniformly in time. Together with 0 ≤ U − ≤ 1 0\le U_-\le1 0 ≤ U − ≤ 1 , this justifies integration of (107) against z − 3 z^{-3} z − 3 by standard localization. Thus, for 0 < s < t 0<s<t 0 < s < t ,
A ( t ) ≤ A ( s ) + 3 μ ∫ s t A ( τ ) d τ + 2 μ 3 ∫ s t A ( τ ) ( ∫ 1 ∞ z − 3 K H ( τ , z ) d z ) d τ . A(t)\le A(s)+3\mu\int_s^t A(\tau)\,\mathrm{d}\tau+\frac{2\mu}{3}\int_s^t A(\tau)\left(\int_1^\infty z^{-3}K_H(\tau,z)\,\mathrm{d}z\right)\mathrm{d}\tau. A ( t ) ≤ A ( s ) + 3 μ ∫ s t A ( τ ) d τ + 3 2 μ ∫ s t A ( τ ) ( ∫ 1 ∞ z − 3 K H ( τ , z ) d z ) d τ .
Using K H ≥ 0 K_H\ge0 K H ≥ 0 and
∫ 1 ∞ K H ( τ , z ) d z = 3 , \int_1^\infty K_H(\tau,z)\,\mathrm{d}z=3, ∫ 1 ∞ K H ( τ , z ) d z = 3 ,
we obtain
A ( t ) ≤ A ( s ) + 5 μ ∫ s t A ( τ ) d τ . (109) A(t)\le A(s)+5\mu\int_s^t A(\tau)\,\mathrm{d}\tau. \tag*{(109)} A ( t ) ≤ A ( s ) + 5 μ ∫ s t A ( τ ) d τ . (109)
Finally, A A A is continuous on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Indeed,
∣ A ( t ) − A ( s ) ∣ ≤ 1 2 ( ∥ G ( t ) − G ( s ) ∥ L 1 + ∥ H ( t ) − H ( s ) ∥ L 1 ) . |A(t)-A(s)|\le\frac{1}{2}\left(\lVert G(t)-G(s)\rVert_{L^1}+\lVert H(t)-H(s)\rVert_{L^1}\right). ∣ A ( t ) − A ( s ) ∣ ≤ 2 1 ( ∥ G ( t ) − G ( s ) ∥ L 1 + ∥ H ( t ) − H ( s ) ∥ L 1 ) .
By (103), A ( 0 ) = 0 A(0)=0 A ( 0 ) = 0 . Letting s ↓ 0 s\downarrow0 s ↓ 0 in (108) and applying Gronwall’s inequality yields
A ( t ) = 0 , t ≥ 0. A(t)=0,\qquad t\ge0. A ( t ) = 0 , t ≥ 0.
Hence U − ( t , z ) = 0 U_-(t,z)=0 U − ( t , z ) = 0 almost everywhere in z z z , for every t ≥ 0 t\geq0 t ≥ 0 . By continuity of the tail distributions,
F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1. F_G(t,z)\geq F_H(t,z),\qquad t\geq0,\ z\geq1. F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1.
Finally, (96) gives
Φ G ( t ) − Φ H ( t ) = 2 μ 3 ∫ 1 ∞ z − 3 ( F G ( t , z ) − F H ( t , z ) ) d z ≥ 0. \Phi_G(t)-\Phi_H(t)=\frac{2\mu}{3}\int_1^\infty z^{-3}\bigl(F_G(t,z)-F_H(t,z)\bigr)\,dz\geq0. Φ G ( t ) − Φ H ( t ) = 3 2 μ ∫ 1 ∞ z − 3 ( F G ( t , z ) − F H ( t , z ) ) d z ≥ 0.
This completes the proof.
We can now complete the proof of the theorem.
Proof of Theorem 3. The conclusion for 1 < E < 2 1<E<2 1 < E < 2 is proved in Lemma 22. We now establish uniform positivity of Φ G \Phi_G Φ G for arbitrary finite energy E > 1 E>1 E > 1 . Let G 0 ∗ G_0^* G 0 ∗ be the symmetric decreasing rearrangement of G 0 G_0 G 0 . Then G 0 ∗ G_0^* G 0 ∗ is a radially nonincreasing probability density satisfying
∫ ∣ p ∣ ≤ r G 0 ∗ ( p ) d p ≥ ∫ ∣ p ∣ ≤ r G 0 ( p ) d p , r > 0. (110) \int_{|p|\leq r}G_0^*(p)\,dp\geq\int_{|p|\leq r}G_0(p)\,dp,\qquad r>0. \tag*{(110)} ∫ ∣ p ∣ ≤ r G 0 ∗ ( p ) d p ≥ ∫ ∣ p ∣ ≤ r G 0 ( p ) d p , r > 0. (110)
Since ⟨ p ⟩ \langle p\rangle ⟨ p ⟩ is radially increasing, it follows that
∫ R 3 ⟨ p ⟩ G 0 ∗ ( p ) d p ≤ E . (111) \int_{\mathbb{R}^3}\langle p\rangle G_0^*(p)\,dp\leq E. \tag*{(111)} ∫ R 3 ⟨ p ⟩ G 0 ∗ ( p ) d p ≤ E . (111)
For 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 , define
H 0 ( p ) = ε − 3 G 0 ∗ ( p / ε ) . H_0(p)=\varepsilon^{-3}G_0^*(p/\varepsilon). H 0 ( p ) = ε − 3 G 0 ∗ ( p / ε ) .
Then H 0 H_0 H 0 is a radially nonincreasing probability density with finite first moment. By (109),
∫ ∣ p ∣ ≤ r H 0 ( p ) d p = ∫ ∣ p ∣ ≤ r / ε G 0 ∗ ( p ) d p ≥ ∫ ∣ p ∣ ≤ r G 0 ( p ) d p . \int_{|p|\leq r}H_0(p)\,dp=\int_{|p|\leq r/\varepsilon}G_0^*(p)\,dp\geq\int_{|p|\leq r}G_0(p)\,dp. ∫ ∣ p ∣ ≤ r H 0 ( p ) d p = ∫ ∣ p ∣ ≤ r / ε G 0 ∗ ( p ) d p ≥ ∫ ∣ p ∣ ≤ r G 0 ( p ) d p .
Equivalently,
F G ( 0 , z ) ≥ F H ( 0 , z ) , z ≥ 1. (112) F_G(0,z)\geq F_H(0,z),\qquad z\geq1. \tag*{(112)} F G ( 0 , z ) ≥ F H ( 0 , z ) , z ≥ 1. (112)
Let
E H = ∫ R 3 ⟨ p ⟩ H 0 ( p ) d p . E_H=\int_{\mathbb{R}^3}\langle p\rangle H_0(p)\,dp. E H = ∫ R 3 ⟨ p ⟩ H 0 ( p ) d p .
Since ⟨ ε p ⟩ ≤ 1 + ε ( ⟨ p ⟩ − 1 ) \langle\varepsilon p\rangle\leq1+\varepsilon(\langle p\rangle-1) ⟨ εp ⟩ ≤ 1 + ε (⟨ p ⟩ − 1 ) , we have
E H = ∫ R 3 ⟨ ε p ⟩ G 0 ∗ ( p ) d p ≤ 1 + ε ∫ R 3 ( ⟨ p ⟩ − 1 ) G 0 ∗ ( p ) d p ≤ 1 + ε ( E − 1 ) . E_H=\int_{\mathbb{R}^3}\langle\varepsilon p\rangle G_0^*(p)\,dp\leq1+\varepsilon\int_{\mathbb{R}^3}(\langle p\rangle-1)G_0^*(p)\,dp\leq1+\varepsilon(E-1). E H = ∫ R 3 ⟨ εp ⟩ G 0 ∗ ( p ) d p ≤ 1 + ε ∫ R 3 (⟨ p ⟩ − 1 ) G 0 ∗ ( p ) d p ≤ 1 + ε ( E − 1 ) .
Since H 0 H_0 H 0 is a probability density, we have E H > 1 E_H>1 E H > 1 . Choosing ε > 0 \varepsilon>0 ε > 0 sufficiently small, we obtain 1 < E H < 2 1<E_H<2 1 < E H < 2 . Let H H H be the global weak solution with initial data H 0 H_0 H 0 . By Lemma 22, there exists δ H > 0 \delta_H>0 δ H > 0 such that
Φ H ( t ) ≥ δ H , t ≥ 0. (113) \Phi_H(t)\geq\delta_H,\qquad t\geq0. \tag*{(113)} Φ H ( t ) ≥ δ H , t ≥ 0. (113)
Moreover, Lemma 23 shows that H ( t , ⋅ ) H(t,\cdot) H ( t , ⋅ ) remains radially nonincreasing. Therefore the assumptions of Lemma 24 are satisfied, and
F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1. F_G(t,z) \ge F_H(t,z), \qquad t \ge0,\ z \ge1. F G ( t , z ) ≥ F H ( t , z ) , t ≥ 0 , z ≥ 1.
Consequently,
Φ G ( t ) ≥ Φ H ( t ) , t ≥ 0. \Phi_G(t) \ge\Phi_H(t), \qquad t \ge0. Φ G ( t ) ≥ Φ H ( t ) , t ≥ 0.
Combining this inequality with (113) , we conclude that
Φ G ( t ) ≥ δ H > 0 , t ≥ 0. \Phi_G(t) \ge\delta_H > 0, \qquad t \ge0. Φ G ( t ) ≥ δ H > 0 , t ≥ 0.
The constant δ H \delta_H δ H depends only on H 0 H_0 H 0 and μ \mu μ , and hence only on G 0 G_0 G 0 and μ \mu μ . This proves (9) and completes the proof.