The core of the demonstration of this paper is to interpret the forward propagation process of machine learning as a parameter estimation problem of nonlinear dynamical systems. This process is to establish a connection between the Recurrent Neural Network and the discrete differential equation, so as to construct a new network structure: ODE-RU. At the same time, under the inspiration of the theory of ordinary differential equations, we propose a new forward propagation mode. In a large number of simulations and experiments, the forward propagation not only shows the trainability of the new architecture, but also achieves a low training error on the basis of main-taining the stability of the network. For the problem requiring long-term memory, we specifically study the obstacle shape reconstruction problem using the backscattering far-field features data set, and demonstrate the effectiveness of the proposed architecture using the data set. The results show that the network can effectively reduce the sensitivity to small changes in the input feature. And the error generated by the ordinary differential equation cyclic unit network in inverting the shape and position of obstacles is less than 10−2.
Citation: Pinchao Meng, Xinyu Wang, Weishi Yin. ODE-RU: a dynamical system view on recurrent neural networks[J]. Electronic Research Archive, 2022, 30(1): 257-271. doi: 10.3934/era.2022014
[1] | Senhao Duan, Yue MA, Weidong Zhang . Conformal-type energy estimates on hyperboloids and the wave-Klein-Gordon model of self-gravitating massive fields. Communications in Analysis and Mechanics, 2023, 15(2): 111-131. doi: 10.3934/cam.2023007 |
[2] | Yuxuan Chen . Global dynamical behavior of solutions for finite degenerate fourth-order parabolic equations with mean curvature nonlinearity. Communications in Analysis and Mechanics, 2023, 15(4): 658-694. doi: 10.3934/cam.2023033 |
[3] | Shuai Xi . A Kato-type criterion for the inviscid limit of the nonhomogeneous NS equations with no-slip boundary condition. Communications in Analysis and Mechanics, 2024, 16(4): 896-909. doi: 10.3934/cam.2024039 |
[4] | Shuyue Ma, Jiawei Sun, Huimin Yu . Global existence and stability of temporal periodic solution to non-isentropic compressible Euler equations with a source term. Communications in Analysis and Mechanics, 2023, 15(2): 245-266. doi: 10.3934/cam.2023013 |
[5] | Farrukh Dekhkonov, Wenke Li, Weipeng Wu . On the time-optimal control problem for a fourth order parabolic equation in the three-dimensional space. Communications in Analysis and Mechanics, 2025, 17(2): 413-428. doi: 10.3934/cam.2025017 |
[6] | Jiangbo Han, Runzhang Xu, Chao Yang . Continuous dependence on initial data and high energy blowup time estimate for porous elastic system. Communications in Analysis and Mechanics, 2023, 15(2): 214-244. doi: 10.3934/cam.2023012 |
[7] | Yang Liu . Global attractors for a nonlinear plate equation modeling the oscillations of suspension bridges. Communications in Analysis and Mechanics, 2023, 15(3): 436-456. doi: 10.3934/cam.2023021 |
[8] | Chunyou Sun, Junyan Tan . Attractors for a Navier–Stokes–Allen–Cahn system with unmatched densities. Communications in Analysis and Mechanics, 2025, 17(1): 237-262. doi: 10.3934/cam.2025010 |
[9] | Henryk Żołądek . Normal forms, invariant manifolds and Lyapunov theorems. Communications in Analysis and Mechanics, 2023, 15(2): 300-341. doi: 10.3934/cam.2023016 |
[10] | Haifa El Jarroudi, Mustapha El Jarroudi . Asymptotic behavior of a viscous incompressible fluid flow in a fractal network of branching tubes. Communications in Analysis and Mechanics, 2024, 16(3): 655-699. doi: 10.3934/cam.2024030 |
The core of the demonstration of this paper is to interpret the forward propagation process of machine learning as a parameter estimation problem of nonlinear dynamical systems. This process is to establish a connection between the Recurrent Neural Network and the discrete differential equation, so as to construct a new network structure: ODE-RU. At the same time, under the inspiration of the theory of ordinary differential equations, we propose a new forward propagation mode. In a large number of simulations and experiments, the forward propagation not only shows the trainability of the new architecture, but also achieves a low training error on the basis of main-taining the stability of the network. For the problem requiring long-term memory, we specifically study the obstacle shape reconstruction problem using the backscattering far-field features data set, and demonstrate the effectiveness of the proposed architecture using the data set. The results show that the network can effectively reduce the sensitivity to small changes in the input feature. And the error generated by the ordinary differential equation cyclic unit network in inverting the shape and position of obstacles is less than 10−2.
The motion dynamics of incompressible isotropic elastodynamics is characterized as a wave system in Lagrangian coordinates, which inherently satisfies the null condition. Based on this structure, a series of studies have been conducted to establish the global well-posedness of classical solutions to this system; see[1,2]. However, these studies reveal a certain time growth for the highest-order generalized energy. In this paper, we investigate the time growth of the Sobolev norm for classical solutions to three-dimensional inhomogeneous incompressible isotropic elastodynamics with small initial perturbation and establish the uniform bound for the highest-order energy.
Before presenting the main result of this paper, we briefly review related known results. For three-dimensional elastic waves, John [3] proved that the genuine nonlinearity condition leads to singularity formation even for arbitrarily small spherically symmetric displacement. We also refer readers to [4] regarding large displacement singularity. The existence of almost global solutions was established in [5,6] for three-dimensional quasilinear wave equations with sufficiently small initial data. Significant contributions toward global existence were independently made by Sideris [7,8] and Agemi [9] under the assumption that nonlinearity satisfies the null condition in three dimensions. In terms of three-dimensional incompressible elastodynamics, the only waves presented in the isotropic systems are shear waves, which are linearly degenerate. The global existence of a solution was demonstrated by Sideris and Thomases in [1,2] through two different methods. It is more challenging to obtain the global existence for the two-dimensional incompressible elastodynamics due to the weaker dispersive decay. In [10], the authors proved almost global existence for a two-dimensional incompressible system in Eulerian coordinates. By introducing the concept of strong null condition and observing that the incompressible elastodynamics automatically satisfies such strong null structure in Lagrangian coordinates, Lei [11] successfully proved the global well-posedness for two-dimensional incompressible elastodynamics by the method of Klainerman and Alinhac's ghost weight method [12]. We also see [13] for a different approach using the spacetime resonance method. All the aforementioned works considered the homogeneous fluids. In [14], the authors established the global well-posedness for the three-dimensional inhomogeneous incompressible elastodynamics in Lagrangian coordinates. It is noteworthy that the upper bound of the highest-order generalized energy in those studies depends on time. Utilizing the Klainerman's generalized energy method, an analysis of the inherent structure of the system and the ghost weight method, [15,16] established the uniform bound for the highest-order generalized energy estimates for two-dimensional and three-dimensional incompressible elastodynamics, respectively. Based on the above foundational works, it is natural to verify the uniform bound for the highest-order generalized energy for three-dimensional inhomogeneous incompressible isotropic elastodynamics. To establish the time growth of the Sobolev norm of classical solution, two novel methods are presented in this paper. First, based on the Sobolev embedding inequality and the structure of the system, the refined decay rates were derived for the solution in the domain away from the light cone. Second, we apply the KSS-type estimate to overcome the difficulties posed by insufficient time decay resulting from density perturbation.
This paper is organized as follows. In Section 2, we introduce the system of three-dimensional inhomogeneous isotropic elastodynamics and define the notations utilized throughout this paper. Besides, the main result along with several useful lemmas are presented in this section. The energy estimates are discussed in Section 3.
We first formulate the inhomogeneous isotropic elastodynamics and denote some notations that are used frequently in this paper.
For any given smooth flow map X(t,x), we call it incompressible if
∫Ωdx=∫ΩtdX,Ωt={X(t,x)|x∈Ω} |
for any smooth bounded connected domain Ω, which yields that
det(∇X)=1. |
Denote
X(t,x)=x+v(t,x). |
Simple calculation shows that the incompressible condition is equivalent to
∇⋅v+12[(tr∇v)2−tr(∇v)2]+det(∇v)=0. | (2.1) |
Without loss of generality, we assume that the density of fluid is a small perturbation around the constant state 1, that is, ρ(x)=1+η(x). For the inhomogeneous isotropic material, the motion of the elastic fluid in the Lagrangian coordinate is determined by
L(X;T,Ω)=∫T0∫Ω(12ρ(x)|∂tX|2−W(∇X)+p(t,x)[det(∇X)−1])dxdt. | (2.2) |
Here W(∇X)∈C∞ is the strain energy function. p(t,x) is a Lagrangian multiplier that is used to force the flow maps to be incompressible. To simplify the presentation, we only study the typical Hookean case for which the strain energy functional is given by
W(∇X)=12|∇X|2. |
By calculating the variation of (2.2), we obtain the equation
ρ∂2tv−Δv=−(∇X)−T∇p. | (2.3) |
Now, we introduce the following derivative vector fields
∂t=∂0,∇=(∂1,∂2,∂3)and∂=(∂0,∂1,∂2,∂3). |
The scaling operator is denoted by
S=t∂t+r∂r. |
Here, the radial derivative is defined by ∂r=xr⋅∇, r=|x|. The angular momentum operators are denoted by
Ω=x∧∇. |
In the application, we usually use the modified rotational operators and scaling operator; that is, for any vector v and scalar p and ρ, we set
˜Sp=Sp,˜Sρ=Sρ,˜Sv=(S−1)v,˜Ωip=Ωip,˜Ωiρ=Ωiρ,˜Ωiv=Ωiv+Uiv,i=1,2,3, |
where
U1=e2⊗e3−e3⊗e2,U2=e3⊗e1−e1⊗e3,U3=e1⊗e2−e2⊗e1. |
Let
Γ=(Γ1,⋯,Γ8)=(∂,˜Ω,˜S) |
and for any multi-index α=(α1,α2,⋯,α8)∈N8, we denote
Γα=Γα11⋯Γα88. |
We apply the derivatives Γα to the equations (2.1) and (2.3), and then the three-dimensional inhomogeneous isotropic elastodynamics can be written as
ρ∂2tΓαv−ΔΓαv=−(∇X)−T∇Γαp−∑β+γ=α|β|≠|α|CβαΓγη∂2tΓβv−∑β+γ=α|β|≠|α|Cβα(∇X)−T(∇Γγv)T⋅(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)=:Nα | (2.4) |
with the incompressible condition
∇⋅Γαv+∑β+γ=αi,j=1,2,3,i<jCβα(∂iΓβvi∂jΓγvj−∂iΓβvj∂jΓγvi)+∑β+γ+ι=αCβαCγα−β|∂1Γβv1∂2Γβv1∂3Γβv1∂1Γβv2∂2Γβv2∂3Γβv2∂1Γβv3∂2Γβv3∂3Γβv3|=0. | (2.5) |
Here the binomial coefficient Cβα is given by
Cβα=α!β!(α−β)!. |
We denote the Klainerman's generalized energy by
Eκ(t)=∑|α|≤κ−1∫R3(|∂tΓαv|2+|∇Γαv|2)dx. |
We also define the weighted energy norm
Xκ(t)=∑|α|≤κ−2∫R3⟨t−r⟩2|∂2Γαv|2dx. |
To describe the space of initial data, we introduce the time-independent analogue of Γ as
Λ=(∇,r∂r−1,˜Ω) |
and the space of initial data is defined by
HκΛ={(f,g):∑|α|≤κ−1(‖Λαf‖L2+‖∇Λαf‖L2+‖Λαg‖L2)<∞}. |
As the first step to investigate our problem, we introduce the following lemma, which helps us to solve the additional terms resulting from density perturbation. Let
LEκ(T)=∑|α|≤κ−1∫T0∫Rnr−1+2μ⟨r⟩−2μ′(|∂Γαv|+|Γαv|r)2dxdt |
with μ∈(0,12) and μ′>μ. Without loss of generality, we choose μ=14 and μ′=12 in this paper. In the case of μ′=μ, we see LEκ(T) is the KSS norm, and we denote it by KSSκ(T).
Lemma 1. Let f0=[r/(1+r)]2μ, fk=r/(r+2k) with k≥1, μ∈(0,1/2), and v be the solution to the equation ∂2tv−c2Δv+hab∂a∂bv=N in [0,T]×Rn with hab=hba, ∑0≤a,b≤n|hab|≤min(1,c2)/2 for any integer n≥3. Then there exists a positive constant C0 that depends only on the dimension n such that
sup0≤t≤T∫Rn|∂v|2(t)dx+LE1(T)+(ln(2+T))−1KSS1(T)≤C0∫Rn|∂v|2(0)dx+C0∫T0∫Rn[(|∂h|+|h|r1−2μ⟨r⟩2μ)|∂v|(|∂v|+|v|r)]dxdt+C0|∫T0∫Rn∂tv⋅Ndxdt|+C0supk≥0|∫T0∫Rnfk(∂rv+n−12rv)⋅Ndxdt|, |
where |h|=∑na,b=0|hab| and |∂h|=∑na,b,c=0|∂chab|.
This lemma can be found in [17]. See also [18,19] and references therein.
Based on the previous statement, we are ready to show the main result of this paper.
Theorem 1. Let W(∇X)=12|∇X|2 be an isotropic Hookean strain energy function and (v0,∂tv0)∈HκΛ with κ≥12. Let C0>0 be given constant in Lemma 1. Suppose v0 satisfies the structural constraint condition (2.1) and
Eκ(0)=∑|α|≤κ−1(‖∂tΛαv0‖2L2+‖∇Λαv0‖2L2)≤ε. |
If
‖⟨r⟩Λαη‖L2≤δfor|α|≤κ, |
then there exist two sufficiently small constants, ε0,δ0 and constant C1, that depend only on κ and C0 such that if ε≤ε0 and δ≤δ0, the system (2.3) has a unique global classical solution that satisfies
Eκ(t)≤C1ε |
uniformly for all t∈[0,+∞).
In this part, we establish several lemmas that are crucial for the energy estimates. Throughout this paper, we denote ⟨⋅⟩=(1+|⋅|2)12. The notation f≲g stands for f≤Cg for some generic constant C>0, which may vary from line to line. In the process of deriving the energy estimates, we usually separate the whole integration domain R3 into two parts:
R={x∈R3:r≤⟨t⟩/8},Rc={x∈R3:r>⟨t⟩/8}. |
We first recall the Sobolev-type inequalities, which were justified in [8].
Lemma 2. For any v∈C∞0(R3)3, r=|x| and ˜r=|y|, we have
‖⟨r⟩12v(x)‖L∞≲∑|α|≤1‖∇˜Ωαv‖L2,‖⟨r⟩v(x)‖L∞≲∑|α|≤1‖∂r˜Ωαv‖12L2(|y|≥r)‖˜Ωαv‖12L2(|y|≥r),‖⟨r⟩⟨t−r⟩v(x)‖L∞≲∑|α|≤1‖⟨t−˜r⟩∂˜r˜Ωαv‖L2(|y|≥r)+∑|α|≤2‖⟨t−˜r⟩˜Ωαv‖L2(|y|≥r). |
The following lemma concerns the dispersive decay of solutions in the domain away from the light cone.
Lemma 3. For any v∈H2(R3), there holds
⟨t⟩‖v‖L∞(r≤⟨t⟩/8)≲‖v‖L2(r≤⟨t⟩/4)+‖⟨t−r⟩∇v‖L2(r≤⟨t⟩/4)+‖⟨t−r⟩∇2v‖L2(r≤⟨t⟩/4). |
Proof. The proof can be found in Lemma 4.3 in [16].
Let s=8r⟨t⟩. We introduce a radial cutoff function ξ(s)∈C∞0 that satisfies
ξ(s)={1,s≤1,0,s≥2. |
It is easy to observe from Lemmas 2 and 3 that
⟨t⟩‖∂v‖L∞≲⟨t⟩‖∂v‖L∞(R)+‖⟨r⟩(1−ξ(s))∂v‖L∞≲‖∂v‖L2(r≤⟨t⟩/4)+‖⟨t−r⟩∇∂v‖L2(r≤⟨t⟩/4)+‖⟨t−r⟩∇2∂v‖L2(r≤⟨t⟩/4)+∑ι1,|ι2|≤1‖∂ι1r˜Ωι2[(1−ξ(s))∂v]‖L2≲E123(t)+X123(t). | (2.6) |
To control the weighted energy norm by the generalized energy norm, the pressure must be estimated via the system (2.4)-(2.5). A similar proof can be found in [14]. For a self-contained presentation, we include its proof below.
Lemma 4. For any integer κ≥6 and multi-index α satisfying |α|≤κ−2, if E[κ/2]+3(t) and ‖⟨r⟩Λαη‖L2 are small. Then there holds
⟨t⟩‖∇Γαp‖L2+⟨t⟩‖ρ∂2tΓαv−ΔΓαv‖L2≲δX12κ(t)+E12κ(t)X12[κ/2]+3(t)+E12[κ/2]+3(t)X12κ(t). |
Proof. It observes from (2.4) that
∇Γαp=−(∇X)T(ρ∂2tΓαv−ΔΓαv)−∑β+γ=α|β|≠|α|Cβα(∇X)TΓγη∂2tΓβv−∑β+γ=α|β|≠|α|Cβα(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v). | (2.7) |
We apply ∇Δ−1∇⋅ to the above equality and take the L2 norm. By the L2 boundness of the Riesz operator, one has
‖∇Γαp‖L2≲‖∇Δ−1∇⋅(ρ∂2tΓαv−ΔΓαv)‖L2+‖(∇v)T(ρ∂2tΓαv−ΔΓαv)‖L2+∑β+γ=α|β|≠|α|‖(∇X)TΓγη∂2tΓβv‖L2+∑β+γ=α|β|≠|α|‖(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2. | (2.8) |
Special attention is paid to the first term on the right-hand side of (2.8). We apply the derivative operator (∂2t−Δ) to the incompressible condition (2.5). Without loss of generality, we assume at least one order derivative operator of (∂2t−Δ) works on the first component of the velocity field in the second line of (2.5). Then we obtain
(∂2t−Δ)(∇⋅Γαv)=−∑β+γ=αi,j=1,2,3,i<jCβα(∂2t−Δ)(∂iΓβvi∂jΓγvj−∂iΓβvj∂jΓγvi)−∑β+γ+ι=αCβαCγα−β(∂2t−Δ)|∂1Γβv1∂2Γβv1∂3Γβv1∂1Γβv2∂2Γβv2∂3Γβv2∂1Γβv3∂2Γβv3∂3Γβv3|=∑β+γ=αm+n=1,i=0,1,2,3Cβα∇⋅(∂m+1iΓβv1∂2∂niΓγv2+∂m+1iΓβv1∂3∂niΓγv3−∂2∂miΓβv1∂n+1iΓγv2−∂3∂miΓβv1∂n+1iΓγv3−∂m+1iΓβv1∂1∂niΓγv2+∂m+1iΓβv2∂3∂niΓγv3+∂1∂miΓβv1∂n+1iΓγv2−∂3∂miΓβv2∂n+1iΓγv3−∂m+1iΓβv2∂2∂niΓγv3−∂m+1iΓβv1∂1∂niΓγv3+∂2∂miΓβv2∂n+1iΓγv3+∂1∂miΓβv1∂n+1iΓγv3)+∑β+γ+ι=αm2+m2+m3=1,i=0,1,2,3CβαCγα−β∇⋅[∂m1+1iΓβv1(∂2∂m2iΓγv2∂3∂m3iΓιv3−∂3∂m2iΓγv2∂2∂m3iΓιv3∂3∂m2iΓγv2∂1∂m3iΓιv3−∂1∂m2iΓγv2∂3∂m3iΓιv3∂1∂m2iΓγv2∂2∂m3iΓιv3−∂2∂m2iΓγv2∂1∂m3iΓιv3)]. |
Based on the above equality, we handle the first term on the right-hand side of (2.8) as follows
‖∇Δ−1∇⋅(ρ∂2tΓαv−ΔΓαv)‖L2≲‖∇Δ−1[(ρ∂2t−Δ)(∇⋅Γαv)]‖L2+∑i=1,2,3‖∇Δ−1(∂η∂2tΓαvi)‖L2≲(1+‖∇v‖L∞)‖∇v‖L∞‖ρ∂2tΓαv−ΔΓαv‖L2+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j‖∂2Γβvi∂Γγvj‖L2+∑β+γ+ι=α,|β|≠|α|i,j,k=1,2,3,i≠j≠k‖∂2Γβvi∂Γγvj∂Γιvk‖L2+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j‖∇Δ−1(∂η∂2Γβvi∂Γγvj)‖L2+∑β+γ+ι=α,|β|≠|α|i,j,k=1,2,3,i≠j≠k‖∇Δ−1(∂η∂2Γβvi∂Γγvj∂Γιvk)‖L2+∑i=1,2,3‖∇Δ−1(∂η∂2tΓαvi)‖L2+∑i,j=1,2,3i≠j‖∇Δ−1(∂η∂2tΓαvi∂vj)‖L2+∑i,j,k=1,2,3i≠j≠k‖∇Δ−1(∂η∂2tΓαvi∂vj∂vk)‖L2. | (2.9) |
By (2.8), (2.9), the Sobolev embedding inequality, and the smallness of ‖∇η‖L3, we have
‖∇Γαp‖L2≲(1+‖∇v‖L∞)‖∇v‖L∞‖ρ∂2tΓαv−ΔΓαv‖L2+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j‖∂2Γβvi∂Γγvj‖L2+∑β+γ+ι=α,|β|≠|α|1,j,k=1,2,3,i≠j≠k‖∂2Γβvi∂Γγvj∂Γιvk‖L2+∑β+γ=α|β|≠|α|‖Γγη∂2tΓβv‖L2+∑β+γ=α|β|≠|α|‖(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2+∑i=1,2,3‖∂η∂2tΓαvi‖L65+∑i,j=1,2,3i≠j‖∂η∂2tΓαvi∂vj‖L65+∑i,j,k=1,2,3i≠j‖∂η∂2tΓαvi∂vj∂vk‖L65. |
By (2.4), the above inequality and the smallness of E[κ/2]+3(t), one obtains
‖∇Γαp‖L2+‖ρ∂2tΓαv−ΔΓαv‖L2≲∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j‖∂2Γβvi∂Γγvj‖L2+∑β+γ+ι=α,|β|≠|α|i,j,k=1,2,3,i≠j≠k‖∂2Γβvi∂Γγvj∂Γιvk‖L2+∑β+γ=α|β|≠|α|‖Γγη∂2tΓβv‖L2+∑β+γ=α|β|≠|α|‖(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2+∑i=1,2,3‖∂η∂2tΓαvi‖L65+∑i,j=1,2,3i≠j‖∂η∂2tΓαvi∂vj‖L65+∑i,j,k=1,2,3i≠j≠k‖∂η∂2tΓαvi∂vj∂vk‖L65. |
We deduce from Lemma 2 that
⟨t⟩‖∇Γαp‖L2+⟨t⟩‖ρ∂2tΓαv−ΔΓαv‖L2≲∑β+γ=α|γ|<|β|<|α|‖⟨t−r⟩∂2Γβv‖L2‖⟨r⟩∂Γγv‖L∞+∑β+γ=α|β|≤|γ|‖⟨r⟩⟨t−r⟩∂2Γβv‖L∞‖∂Γγv‖L2+∑β+γ+ι=α|γ|,|ι|<|β|<|α|‖⟨t−r⟩∂2Γβv‖L2‖⟨r⟩∂Γγv‖L∞‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|ι|≤|γ|‖⟨r⟩⟨t−r⟩∂2Γβv‖L∞‖∂Γγv‖L2‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|γ|≤|ι|‖⟨r⟩⟨t−r⟩∂2Γβv‖L∞‖∂Γγv‖L∞‖∂Γιv‖L2+∑β+γ=α|γ|<|β|<|α|‖⟨r⟩Γγη‖L∞‖⟨t−r⟩∂2tΓβv‖L2+∑β+γ=α|β|≤|γ|‖Γγη‖L2‖⟨r⟩⟨t−r⟩∂2tΓβv‖L∞+‖⟨r⟩∇η‖L3(1+‖∇v‖L∞)2‖⟨t−r⟩∂2tΓαv‖L2+∑β+γ=α|γ|<|β|<|α|‖⟨r⟩∇Γγv‖L∞‖⟨t−r⟩(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2+∑β+γ=α|β|≤|γ|‖∇Γγv‖L2‖⟨r⟩⟨t−r⟩(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L∞≲δX12κ(t)+E12κ(t)X12[κ/2]+3(t)+E12[κ/2]+3(t)X12κ(t). |
It completes the proof.
As an application of the above result, we establish the estimate of weighted energy.
Lemma 5. Let v∈HκΓ(R3) be the solution to the system (2.3) with the constraint condition (2.1). For any integer κ≥6, if Eκ(t) and ‖⟨r⟩Λαη‖L2 are small, then we have
Xκ(t)≲Eκ(t). |
Proof. For any multi-index α satisfying |α|≤κ−2, we apply Γα to Lemma 3.3 in [11] to get
Xκ(t)≲Eκ(t)+t‖∂2tΓαv−ΔΓαv‖2L2. |
It follows from the above inequality and Lemma 4 that
Xκ(t)≲Eκ(t)+δ2Xκ(t)+E[κ/2]+3(t)Xκ(t)+Eκ(t)X[κ/2]+3(t). |
By the smallness of δ and Eκ(t), we arrive at the lemma.
In preparation for the energy estimates, more detailed analysis of pressure is needed. In what follows, we always assume that Eκ(t) and ‖⟨r⟩Λαη‖L2 are small.
Lemma 6. For any integer κ≥8 and multi-index α satisfying |α|≤κ−1, we have
‖∇Γαp‖L2≲E12κ(t) | (2.10) |
and
⟨t⟩‖∇Γαp‖L2(Rc)+⟨t⟩‖ρ∂2tΓαv−ΔΓαv‖L2(Rc)≲E12κ(t). | (2.11) |
Proof. Following the calculations in Lemma 4, we arrive at (2.9). Special attention is paid to the last three terms on the right-hand side of (2.9). From (2.4), one has
∂2tΓαv=ρ−1ΔΓαv−ρ−1(∇X)−T∇Γαp−∑β+γ=α|β|≠|α|Cβαρ−1Γγη∂2tΓβv−∑β+γ=α|β|≠|α|Cβαρ−1(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v). |
We use the above equality and the Sobolev embedding inequality to solve the last three terms on the right-hand side of (2.9) by
‖∇Δ−1[ρ−1∂η(1+∂v)2ΔΓαv]‖L2+‖∇Δ−1[ρ−1∂η(1+∂v)2(∇X)−T∇Γαp]‖L2+∑β+γ=α|β|≠|α|‖∇Δ−1[ρ−1∂η(1+∂v)2(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2∂2tΓι1v)]‖L2+∑β+γ=α|β|≠|α|‖∇Δ−1[ρ−1∂η(1+∂v)2Γγη∂2tΓβv]‖L2≲‖∂η(1+∂v)2∇Γαv‖L2+∑i+j+k=1‖∂i(ρ−1)∂j∂η∂k(1+∇v)2∇Γαv‖L65+‖∂η(1+∂v)2(∇X)−T∇Γαp‖L65+∑β+γ=α|β|≠|α|‖∂η(1+∂v)2(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L65+∑β+γ=α|β|≠|α|‖∂η(1+∂v)2Γγη∂2tΓβv‖L65. |
Utilizing (2.4), (2.8), the above estimate, and the Sobolev embedding inequality, we arrive at
‖∇Γαp‖L2+‖ρ∂2tΓαv−ΔΓαv‖L2≲∑β+γ=α|γ|<|β|<|α|‖∂2Γβv‖L2‖∂Γγv‖L∞+∑β+γ=α|β|≤|γ|‖∂2Γβv‖L∞‖∂Γγv‖L2+∑β+γ+ι=α|γ|,|ι|<|β|<|α|‖∂2Γβv‖L2‖∂Γγv‖L∞‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|ι|≤|γ|‖∂2Γβv‖L∞‖∂Γγv‖L2‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|γ|≤|ι|‖∂2Γβv‖L∞‖∂Γγv‖L∞‖∂Γιv‖L2+∑β+γ=α|γ|<|β|<|α|‖Γγη‖L∞‖∂2tΓβv‖L2+∑β+γ=α|β|≤|γ|‖Γγη‖L2‖∂2tΓβv‖L∞+∑1≤|ι|≤3‖∇ιη‖L2‖∇Γαv‖L2+∑β+γ=α|γ|<|β|<|α|‖∇Γγv‖L∞‖∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v‖L2+∑β+γ=α|β|≤|γ|‖∇Γγv‖L2‖∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v‖L∞≲δE12κ(t)+E12κ(t)E12[κ/2]+4(t). | (2.12) |
The smallness of δ and Eκ(t) leads to (2.10).
To verify (2.11), we use (2.12), Lemma 2, and the smallness of Eκ(t) to get
⟨t⟩‖∇Γαp‖L2(Rc)+⟨t⟩‖ρ∂2tΓαv−ΔΓαv‖L2(Rc)≲∑β+γ=α|γ|<|β|<|α|‖∂2Γβv‖L2‖⟨r⟩∂Γγv‖L∞+∑β+γ=α|β|≤|γ|‖⟨r⟩∂2Γβv‖L∞‖∂Γγv‖L2+∑β+γ+ι=α|β|,|ι|≤|γ|‖∂2Γβv‖L∞‖∂Γγv‖L2‖⟨r⟩∂Γιv‖L∞+∑β+γ+ι=α|β|,|γ|≤|ι|‖∂2Γβv‖L∞‖⟨r⟩∂Γγv‖L∞‖∂Γιv‖L2+∑β+γ+ι=α|γ|,|ι|<|β|<|α|‖∂2Γβv‖L2‖⟨r⟩∂Γγv‖L∞‖∂Γιv‖L∞+∑β+γ=α|γ|<|β|<|α|‖⟨r⟩Γγη‖L∞‖∂2tΓβv‖L2+∑β+γ=α|β|≤|γ|‖Γγη‖L2‖⟨r⟩∂2tΓβv‖L∞+∑1≤|ι|≤3‖⟨r⟩∇ιη‖L2‖∇Γαv‖L2+∑β+γ=α|γ|<|β|<|α|‖⟨r⟩∇Γγv‖L∞‖∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v‖L2+∑β+γ=α|β|≤|γ|‖∇Γγv‖L2‖⟨r⟩(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L∞≲E12κ(t)E12[κ/2]+4(t)+δE12κ(t). |
The smallness of δ and Eκ(t) implies (2.11).
In the subsequent part, we present the improved decay properties for the third-order spatial derivatives of unknown variables in the domain away from the light cone.
Lemma 7. For any integer κ≥10 and multi-index α satisfying |α|≤[κ/2], it holds that
⟨t⟩2‖∇3Γαv‖L2(R)≲E12[κ/2]+5(t). |
Proof. We apply the derivative operator ∇ to the equation (2.4) to get
∇ΔΓαv=∇η∂2tΓαv+ρ∂2t∇Γαv+∇(∇X)−T∇Γαp+(∇X)−T∇∇Γαp+∑β+γ=α|β|≠|α|Cβα∇Γγη∂2tΓβv+∑β+γ=α|β|≠|α|CβαΓγη∂2t∇Γβv+∑β+γ=α|β|≠|α|Cβα∇[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]. |
By multiplying the above equality by t2ξ(s) and taking the L2 inner product, one has
∫R3|t2ξ(s)∇ΔΓαv|2dx≤7∫R3|t2ξ(s)∇η∂2tΓαv|2dx+7∫R3|t2ξ(s)ρ∂2t∇Γαv|2dx+7∫R3|t2ξ(s)∇(∇X)−T∇Γαp|2dx+7∫R3|t2ξ(s)(∇X)−T∇2Γαp|2dx+∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)∇Γγη∂2tΓβv|2dx+∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)Γγη∂2t∇Γβv|2dx+∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)∇[(∇X)−T(∇Γγv)T⋅(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]|2dx. | (2.13) |
Since for any f∈H2(R3), the following Sobolev embedding inequality holds true
‖f‖L∞≲‖∇f‖12L2‖∇2f‖12L2. | (2.14) |
By Lemma 2, the first term on the right-hand side of (2.13) is estimated by
7∫R3|t2ξ(s)∇η∂2tΓαv|2dx≲∫R3|trξ(s)∇η∂r∂tΓαv|2dx+∫R3|tξ(s)∇η∂t˜SΓαv|2dx≲(‖ξ(s)⟨r⟩⟨t−r⟩∂r∂tΓαv‖2L∞+‖tξ(s)∂t˜SΓαv‖2L∞)‖∇η‖2L2≲(‖⟨r⟩⟨t−r⟩∂r∂tΓαv‖2L∞+‖∇(tξ(s)∂t˜SΓαv)‖L2‖∇2(tξ(s)∂t˜SΓαv)‖L2)‖∇η‖2L2≲[‖⟨r⟩⟨t−r⟩∂r∂tΓαv‖2L∞+(‖ξ′(s)∂t˜SΓαv‖L2+‖ξ(s)⟨t−r⟩∂t∇˜SΓαv‖L2)⋅(⟨t⟩−1‖ξ″(s)∂t˜SΓαv‖L2+‖ξ′(s)∂t∇˜SΓαv‖L2+‖ξ(s)⟨t−r⟩∂t∇2˜SΓαv‖L2)]‖∇η‖2L2≲δ2E[κ/2]+4(t). | (2.15) |
By a similar argument, we handle the second term on the right-hand side of (2.13) as follows
7∫R3|t2ξ(s)ρ∂2t∇Γαv|2dx≤7∫R3|trξ(s)ρ∂r∂t∇Γαv|2dx+C∫R3(|tξ(s)ρ∂t∇˜SΓαv|2+|tξ(s)ρ∂t∇Γαv|2)dx≤7∫R3|r2ξ(s)ρ∂2r∇Γαv|2dx+C∫R3(|tξ(s)ρ∂t∇˜SΓαv|2+|tξ(s)ρ∂t∇Γαv|2+|ξ(s)ρr∂r∇˜SΓαv|2+|ξ(s)ρr∂r∇Γαv|2)dx≤7∫R3|r2ξ(s)ρ∇3Γαv|2dx+C(‖⟨t−r⟩∂∇Γαv‖2L2(R)+‖⟨t−r⟩∂∇˜SΓαv‖2L2(R))≤716∫R3|t2ξ(s)∇3Γαv|2dx+CE[κ/2]+3(t). | (2.16) |
In view of (2.7), the smallness of δ and Eκ(t), one has
‖∇Γαp‖L2(R)+⟨t⟩‖∇Γαp‖L2(R)≲‖(1+⟨t⟩)(∇X)T(∂2tΓαv−ΔΓαv)‖L2(R)+∑β+γ=α‖(1+⟨t⟩)(∇X)TΓγη∂2tΓβv‖L2(R)+∑β+γ=α‖(1+⟨t⟩)(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2(R)≲E12[κ/2]+2(t). |
We substitute the above inequality into the third term on the right-hand side of (2.13) to get
7∫R3|t2ξ(s)∇(∇X)−T∇Γαp|2dx≲⟨t⟩2‖∇2v‖2L∞(R)⟨t⟩2‖∇v‖2L∞(R)‖∇Γαp‖2L2(R)+⟨t⟩2‖∇2v‖2L∞(R)⟨t⟩2‖∇Γαp‖2L2(R)≲E[κ/2]+2(t). |
We come back to the fourth term on the right-hand side of (2.13). We apply the divergence operator to the equality (2.7) and multiply t2 on both sides of the resulting equality. By taking the L2 inner product, one has
‖t2ξ(s)ΔΓαp‖L2≲‖t2ξ(s)∇⋅[(∇v)T(ρ∂2tΓαv−ΔΓαv)]‖L2+∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅[(∇v)TΓγη∂2tΓβv]‖L2+‖t2ξ(s)∇⋅(ρ∂2tΓαv−ΔΓαv)‖L2+∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅(Γγη∂2tΓβv)‖L2+∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅[(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]‖L2. | (2.17) |
For the first two terms on the right-hand side of (2.17), we deduce from Lemma 3 that
‖t2ξ(s)∇⋅[(∇v)T(ρ∂2tΓαv−ΔΓαv)]‖L2+∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅[(∇v)TΓγη∂2tΓβv]‖L2≲∑i+j=1⟨t⟩‖∇i(∇v)T‖L∞(R)‖⟨t−r⟩∇j(ρ∂2tΓαv−ΔΓαv)‖L2(R)+∑β+γ=α,|β|≤|γ| i+j+k=1⟨t⟩‖∇i(∇v)T‖L∞(R)‖∇jΓγη‖L2(R)⟨t⟩‖∂2t∇kΓβv‖L∞(R)+∑β+γ=α,|γ|<|β|<|α| i+j+k=1⟨t⟩‖∇i(∇v)T‖L∞(R)‖∇jΓγη‖L∞(R)‖⟨t−r⟩∂2t∇kΓβv‖L2(R)≲E124(t)E12[κ/2]+3(t)+δE124(t)E12[κ/2]+5(t). |
For the third and fourth terms on the right-hand side of (2.17), we have
‖t2ξ(s)∇⋅(ρ∂2tΓαv−ΔΓαv)‖L2+∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅(Γγη∂2tΓβv)‖L2≲‖t2ξ(s)∇η⋅∂2tΓαv‖L2+‖t2ξ(s)[(ρ∂2t−Δ)(∇⋅Γαv)]‖L2+∑β+γ=α,|β|≠|α|i+j=1‖t2ξ(s)∇iΓγη∂2t(∇j⋅Γβv)‖L2. | (2.18) |
By the definition of ˜S and the Sobolev embedding inequality, we solve the first term on the right-hand side of (2.18) by
‖t2ξ(s)∇η⋅∂2tΓαv‖L2≲‖∇η‖L3‖tξ(s)∂t˜SΓαv‖L6+‖trξ(s)∇η∂r∂tΓαv‖L2≲‖∇η‖L3(‖∂t˜SΓαv‖L2(R)+‖⟨t−r⟩∂t∇˜SΓαv‖L2(R))+‖⟨r⟩∇η‖L∞‖⟨t−r⟩∂r∂tΓαv‖L2(R)≲δE12[κ/2]+3(t). |
In terms of (2.5) and the definition of ˜S, the last two terms on the right-hand side of (2.18) are estimated by
‖t2ξ(s)[(ρ∂2t−Δ)(∇⋅Γαv)]‖L2+∑β+γ=α,|β|≠|α|i+j=1‖t2ξ(s)∇iΓγη∂2t(∇j⋅Γβv)‖L2≲∑β+γ=αi+j=1‖⟨t−r⟩∂i∂∇Γβv‖L2(R)⟨t⟩‖∂j∇Γγv‖L∞(R)+∑β+γ+ι=αi+j+k=1‖⟨t−r⟩∂i∂∇Γβv‖L2(R)⟨t⟩‖∂j∇Γγv‖L∞(R)‖∂k∇Γιv‖L∞(R)+∑β+γ+ι=αj+k=1‖Γγη‖L∞(R)‖⟨t−r⟩∂j∂∇Γβv‖L2(R)⟨t⟩‖∂k∇Γιv‖L∞(R)+∑β+γ+ι=αi+j+k=1‖Γγη‖L∞(R)‖⟨t−r⟩∂i∂∇Γβv‖L2(R)⟨t⟩‖∂j∇Γγv‖L∞(R)‖∂k∇Γιv‖L∞(R)+∑β+γ=α|β|≠|α|‖ξ(s)∇Γγη‖L3⟨t⟩‖ξ(s)∂t˜SΓβv‖L6+∑β+γ=α|β|≠|α|‖⟨r⟩∇Γγη‖L∞(R)‖⟨t−r⟩∂t∂rΓβv‖L2(R)≲(1+δ)E12[κ/2]+3(t)(E12[κ/2]+4(t)+E[κ/2]+4(t))+δE12[κ/2]+3(t). |
Along the same line, the last term on the right-hand side of (2.17) is dealt with by
∑β+γ=α|β|≠|α|‖t2ξ(s)∇⋅[(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]‖L2≲∑β+γ=α|β|≠|α|‖⟨t−r⟩∇2Γγv‖L2(R)⟨t⟩‖∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v‖L∞(R)+∑β+γ=α|β|≠|α|⟨t⟩‖∇Γγv‖L∞(R)‖⟨t−r⟩∇(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2(R)≲E[κ/2]+4(t). |
Combining all the estimates, Lemma 4, and the fact
‖t2ξ(s)∇2Γαp‖L2≲‖t2ξ(s)ΔΓαp‖L2+‖tξ(s)∇Γαp‖L2, |
the fourth term on the right-hand side of (2.13) is estimated by
7∫R3|t2ξ(s)(∇X)−T∇2Γαp|2dx≲⟨t⟩2‖∇v‖2L∞(R)⟨t⟩2‖∇v‖2L∞(R)‖∇2Γαp‖2L2(R)+⟨t⟩2‖∇v‖2L∞(R)⟨t⟩2‖∇2Γαp‖2L2(R)+‖t2∇2Γαp‖2L2(R)≲E[κ/2]+5(t). |
We employ the similar method as (2.15) and (2.16) to estimate the fifth and sixth terms on the right-hand side of (2.13) by
∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)∇Γγη∂2tΓβv|2dx+∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)Γγη∂2t∇Γβv|2dx≲∑β+γ=α|β|≠|α|∫R3(|tξ(s)∇Γγη∂t˜SΓβv|2dx+|tξ(s)r∇Γγη∂r∂tΓβv|2+|tξ(s)Γγη∂t∇˜SΓβv|2+|tξ(s)Γγη∂t∇Γβv|2+|tξ(s)Γγηr∂r∂t∇Γβv|2)dx≲∑β+γ=α|β|≠|α|(‖∇Γγη‖2L3‖tξ(s)∂t˜SΓβv‖2L6+‖⟨r⟩ξ(s)∇Γγη‖2L∞‖⟨t−r⟩∂∇Γβv‖2L2(R)+‖Γγη‖2L∞(R)‖⟨t−r⟩∂∇˜SΓβv‖2L2(R)+‖Γγη‖2L∞(R)‖⟨t−r⟩∂t∇Γβv‖2L2(R)+‖⟨r⟩ξ(s)Γγη‖2L∞‖⟨t−r⟩∂∇2Γβv‖2L2(R))≲δ2E[κ/2]+3(t). |
For the last term on the right-hand side of (2.13), we have
∑β+γ=α|β|≠|α|7∫R3|Cβαt2ξ(s)∇[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]|2dx≲∑β+γ=α,|β|≠|α|i+j≤1⟨t⟩2‖∇i(∇Γγv)T‖2L∞(R)‖⟨t−r⟩∇j(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖2L2(R)≲E[κ/2]+4(t). |
Collecting the above estimates and the fact
∫R3|t2ξ(s)∇3Γαv|2dx≲∫R3|t2ξ(s)∇ΔΓαv|2dx+∫R3|tξ(s)∇2Γαv|2dx, |
we complete the proof.
By Lemma 7, we obtain the following estimate.
Lemma 8. For any integer κ≥12 and multi-index α satisfying |α|≤[κ/2], there holds
⟨t⟩32‖∂Γαv‖L∞(R)≲E12[κ/2]+5(t), | (2.19) |
⟨t⟩2‖∂2Γαv‖L∞(R)≲E12[κ/2]+6(t). | (2.20) |
Proof. By the definition of ˜S and Lemma 7, we have
⟨t⟩2‖∂t∇2Γαv‖L2(R)≲‖⟨t−r⟩∂t∇2Γαv‖L2(R)+‖⟨t−r⟩∇2˜SΓαv‖L2(R)+‖⟨t−r⟩∇2Γαv‖L2(R)+⟨t⟩2‖∇3Γαv‖L2(R)≲E12[κ/2]+5(t). | (2.21) |
The inequalities (2.14), (2.21), and Lemma 7 yield that
⟨t⟩32‖∂Γαv‖L∞(R)≲⟨t⟩32‖∇(ξ(s)∂Γαv)‖12L2‖∇2(ξ(s)∂Γαv)‖12L2≲(‖ξ′(s)∂Γαv‖12L2+‖ξ(s)⟨t−r⟩∇∂Γαv‖12L2) (‖ξ″(s)∂Γαv‖12L2+‖ξ′(s)⟨t−r⟩∇∂Γαv‖12L2+⟨t⟩‖ξ(s)∇2∂Γαv‖12L2)≲E12[κ/2]+5(t). |
To consider (2.20), the definition of ˜S, combined with inequalities (2.14) and (2.21) and Lemmas 3 and 7, implies that
⟨t⟩2‖∂2tΓαv‖L∞(R)+⟨t⟩2‖∂∇Γαv‖L∞(R)≲⟨t⟩‖∂2tΓαv‖L∞(R)+⟨t⟩‖∂tΓαv‖L∞(R)+⟨t⟩‖∂t˜SΓαv‖L∞(R)+⟨t⟩‖r∂r∂tΓαv‖L∞(R)+⟨t⟩2‖∇(ξ(s)∂∇Γαv)‖12L2‖∇2(ξ(s)∂∇Γαv)‖12L2≲E12[κ/2]+4(t)+∑i,|ι|≤1‖⟨t−r⟩∂ir˜Ωι(ξ(s)∂r∂tΓαv)‖L2+(‖ξ′(s)⟨t−r⟩∂∇Γαv‖12L2+‖⟨t⟩2ξ(s)∂∇2Γαv‖12L2)⋅(‖ξ″(s)∂∇Γαv‖12L2+‖⟨t⟩ξ′(s)∂∇2Γαv‖12L2+‖⟨t⟩2ξ(s)∂∇3Γαv‖12L2)≲E12[κ/2]+6(t), |
which implies the desired.
Before concluding this section, we formulate the following two lemmas, which are utilized in the process of deriving energy estimates.
Lemma 9. For any integer κ≥12 and multi-index α satisfying |α|≤[κ/2], we have
⟨t⟩2‖ρ∂2tΓαv−ΔΓαv‖L∞≲E12[κ/2]+5(t). |
Proof. We separate two cases to consider this lemma. For the case x∈R, we use the Sobolev embedding inequality (2.14) to get
‖ξ(s)(ρ∂2tΓαv−ΔΓαv)‖L∞≲‖∇[ξ(s)(ρ∂2tΓαv−ΔΓαv)]‖12L2‖∇2[ξ(s)(ρ∂2tΓαv−ΔΓαv)]‖12L2. |
For the integer i satisfying 1≤i≤2, one has
⟨t⟩2‖∇i[ξ(s)(ρ∂2tΓαv−ΔΓαv)]‖L2≲∑j+k=i1≤j≤i⟨t⟩2‖∇jξ(s)∇k(ρ∂2tΓαv−ΔΓαv)‖L2+∑1≤i≤2⟨t⟩2‖ξ(s)∇i(ρ∂2tΓαv−ΔΓαv)‖L2. | (2.22) |
We estimate the first term on the right-hand side of (2.22) by
∑j+k=i1≤j≤i⟨t⟩2‖∇jξ(s)∇k(ρ∂2tΓαv−ΔΓαv)‖L2≲⟨t⟩‖ξ′(s)∇(ρ∂2tΓαv−ΔΓαv)‖L2+‖ρ∂2tΓαv−ΔΓαv‖L2≲E12[κ/2]+3(t). |
By (2.4), we solve the second term on the right-hand side of (2.22) as follows
∑1≤i≤2⟨t⟩2‖ξ(s)∇i(ρ∂2tΓαv−ΔΓαv)‖L2≲∑1≤i≤2j+k=i,1≤j⟨t⟩2‖ξ(s)∇j(∇X)−T∇k∇Γαp‖L2+∑1≤i≤2⟨t⟩2‖ξ(s)(∇X)−T∇i∇Γαp‖L2+∑β+γ=α,|β|≠|α| j+k=i,1≤i≤2⟨t⟩2‖ξ(s)∇jΓγη∂2t∇kΓβv‖L2+∑β+γ=α,|β|≠|α|1≤i≤2⟨t⟩2‖ξ(s)∇i[(∇X)−T(∇Γγv)T⋅(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]‖L2. | (2.23) |
For the first term on the right-hand side of (2.23), by Lemmas 3 and 4, we have
∑1≤i≤2j+k=i,1≤j⟨t⟩2‖ξ(s)∇j(∇X)−T∇k∇Γαp‖L2≲∑1≤i≤2j+k=i,1≤j⟨t⟩‖∇j(∇X)−T‖L∞(R)⟨t⟩‖∇k∇Γαp‖L2(R)≲X125(t)(E[κ/2]+5(t)+δE12[κ/2]+4(t)). |
Adopting the same method as was used in (2.17), we estimate the second term on the right-hand side of (2.23) by
∑1≤i≤2⟨t⟩2‖ξ(s)(∇X)−T∇i∇Γαp‖L2≲∑1≤i≤2⟨t⟩‖∇v‖L∞(R)(1+‖∇v‖L∞(R))⟨t⟩‖∇i∇Γαp‖L2(R)+∑1≤i≤2⟨t⟩2‖∇i∇Γαp‖L2(R)≲δE12[κ/2]+5(t)+E[κ/2]+5(t). |
For the third term on the right-hand side of (2.23), by the definition of ˜S and (2.6), we obtain
∑β+γ=α,|β|≠|α| j+k=i,1≤i≤2⟨t⟩2‖ξ(s)∇jΓγη∂2t∇kΓβv‖L2≲∑β+γ=α,|β|≠|α| j+k=i,1≤i≤2⟨t⟩(‖ξ(s)∇jΓγη∂2t∇kΓβv‖L2+‖ξ(s)∇jΓγηr∂r∂t∇kΓβv‖L2)+∑β+γ=α,|β|≠|α|1≤i≤2⟨t⟩‖ξ(s)∇iΓγη∂t˜SΓβv‖L2+∑β+γ=α,|β|≠|α| j+k=i,1≤k≤i≤2⟨t⟩(‖ξ(s)∇jΓγη∂t∇˜S∇k−1Γβv‖L2+‖ξ(s)∇jΓγη∂t∇kΓβv‖L2)≲∑β+γ=α,|β|≠|α| j+k=i,1≤i≤2‖⟨r⟩ξ(s)∇jΓγη‖L∞‖⟨t−r⟩∂t∂∇kΓβv‖L2(R)+∑β+γ=α,|β|≠|α|1≤i≤2‖∇iΓγη‖L2(R)⟨t⟩‖∂t˜SΓβv‖L∞(R)+∑β+γ=α,|β|≠|α| j+k=i,1≤k≤i≤2‖∇jΓγη‖L∞(R)(‖⟨t−r⟩∂t∇˜S∇k−1Γβv‖L2(R)+‖⟨t−r⟩∂t∇kΓβv‖L2(R))≲δE12[κ/2]+4(t). |
The last term on the right-hand side of (2.23) is estimated by
∑β+γ=α,|β|≠|α|1≤i≤2⟨t⟩2‖ξ(s)∇i[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]‖L2≲∑β+γ=α,|β|≠|α|1≤i≤2,j+k+l=i‖∇l(∇X)−T‖L∞(R)⟨t⟩‖∇j(∇Γγv)T‖L∞(R)‖⟨t−r⟩∇k(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)‖L2(R)≲E[κ/2]+5(t). |
We verify the case x∈Rc. By Lemmas 2 and 6, we have
⟨t⟩2‖ρ∂2tΓαv−ΔΓαv‖L∞(Rc)≲⟨t⟩‖⟨r⟩(1−ξ(s))(ρ∂2tΓαv−ΔΓαv)‖L∞≲∑i,|ι|≤1⟨t⟩‖∂ir˜Ωι[(1−ξ(s))(ρ∂2tΓαv−ΔΓαv)]‖L2≲∑i+j+k≤1|ι1|+|ι2|+|ι3|≤1≤j+|ι2|‖∂ir˜Ωι1(1−ξ(s))‖L∞(Rc)‖⟨r⟩∂jr˜Ωι2ρ‖L∞(Rc)‖∂kr˜Ωι3∂2tΓβv‖L2(Rc)+∑i+j≤1|ι1|+|ι2|≤1‖∂ir˜Ωι1(1−ξ(s))‖L∞(Rc)‖⟨t⟩(ρ∂2t∂jr˜Ωι2Γβv−Δ∂jr˜Ωι2Γβv)‖L2(Rc)≲E12[κ/2]+4(t). |
Collecting all of the estimates together, we justify the lemma.
Lemma 10. For any integer κ≥12 and multi-index α satisfying |α|≤[κ/2], one has
∑β+γ=α⟨t⟩2‖Γγη∂2tΓβv‖L∞≲δE12[κ/2]+4(t). |
Proof. We deduce from the definition of ˜S that
∑β+γ=αΓγη∂2tΓβv=∑β+γ=α11+t(Γγη∂2tΓβv+Γγη∂t˜SΓβv−Γγηr∂r∂tΓβv). |
By Lemma 2 and (2.6), one has
∑β+γ=α⟨t⟩2‖Γγη∂2tΓβv‖L∞≲∑β+γ=α‖Γγη‖L∞‖⟨r⟩⟨t−r⟩∂2tΓβv‖L∞+∑β+γ=α‖Γγη‖L∞⟨t⟩‖∂t˜SΓβv‖L∞+∑β+γ=α‖⟨r⟩Γγη‖L∞‖⟨r⟩⟨t−r⟩∂r∂tΓβv‖L∞≲δE12[κ/2]+4(t). |
This completes the proof.
This section is devoted to the energy estimates. For any integer κ≥12 and multi-index α∈N8 satisfying |α|≤κ−1, we apply Lemma 1 to the system (2.4)-(2.5) to get
sup0≤t≤T∫R3|∂Γαv|2(t)dx+LEκ(T)≤C0∫R3|∂Γαv|2(0)dx+C0∫T0∫R3[(|∂η|+|η|r12⟨r⟩12)|∂Γαv|(|∂Γαv|+|Γαv|r)]dxdt+C0|∫T0∫R3∂tΓαv⋅Nαdxdt|+C0supk≥0|∫T0∫R3fk(∂rΓαv+Γαvr)⋅Nαdxdt|. | (3.1) |
For the second term on the right-hand side of (3.1), by Lemma 2, one has
C0∫T0∫R3[(|∂η|+|η|r12⟨r⟩12)|∂Γαv|(|∂Γαv|+|Γαv|r)]dxdt≲∫T0(‖⟨r⟩r∂η‖L∞+‖⟨r⟩η‖L∞)‖r−14⟨r⟩−12|∂Γαv|‖L2‖r−14⟨r⟩−12(|∂Γαv|+|Γαv|r)‖L2dt≲∑i,j,k,|ι|≤1∫T0‖ri∂jr˜Ωι∂kη‖L2‖r−14⟨r⟩−12|∂Γαv|‖L2‖r−14⟨r⟩−12(|∂Γαv|+|Γαv|r)‖L2dt≲δLEκ(T). |
By utilizing (2.4), we formulate the third term on the right-hand side of (3.1) as follows
∫T0∫R3∂tΓαv⋅Nαdxdt=−∫T0∫R3∂tΓαv⋅(∇X)−T∇Γαpdxdt−∑β+γ=α|β|≠|α|Cβα∫T0∫R3∂tΓαv⋅Γγη∂2tΓβvdxdt−∑β+γ=α|β|≠|α|Cβα∫T0∫R3∂tΓαv⋅[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]dxdt. | (3.2) |
We handle term by term on the right-hand side of (3.2). For the first term on the right-hand side of (3.2), since (∇X)−T is composed of elements of the form ∂lvj∂mvk, ∂jvj, ∂ivj and the constant 1, where j,k,l,m=1,2,3, it follows that
−∫T0∫R3∂tΓαv⋅(∇X)−T∇Γαpdxdt≲∑i,j,k,l,m,n=1,2,3i≠j≠k∫T0∫R3|∂tΓαvi∂lvj∂mvk∂nΓαp|dxdt+∑i,j=1,2,3i≠j∫T0∫R3|∂tΓαvi∂jvj∂iΓαp|dxdt+∑i,j=1,2,3i≠j∫T0∫R3|∂tΓαvi∂ivj∂jΓαp|dxdt+|∫T0∫R3∂tΓαv⋅∇Γαpdxdt|. | (3.3) |
For the first term on the right-hand side of (3.3), by (2.6) and Lemma 6, we have
−∑i,j,k,l,m,n=1,2,3i≠j≠k∫T0∫R3|∂tΓαvi∂lvj∂mvk∂nΓαp|dxdt≲∫T0⟨t⟩−2‖∂tΓαv‖L2⟨t⟩2‖∇v‖2L∞‖∇Γαp‖L2dt≲∫T0⟨t⟩−2Eκ(t)E[κ/2]+2(t)dt. |
To handle the second and third terms on the right-hand side of (3.3), we apply (2.11) and (2.19) to show
∑i,j=1,2,3i≠j∫T0∫R3|∂tΓαvi∂jvj∂iΓαp|dxdt+∑1,j=1,2,3i≠j∫T0∫R3|∂tΓαvi∂ivj∂jΓαp|dxdt≲∫T0⟨t⟩−32‖∂tΓαv‖L2(R)⟨t⟩32‖∇v‖L∞(R)‖∇Γαp‖L2(R)dt+∫T0⟨t⟩−2‖∂tΓαv‖L2(Rc)‖⟨r⟩(1−ξ(s))∇v‖L∞(Rc)⟨t⟩‖∇Γαp‖L2(Rc)dt≲∫T0⟨t⟩−32E32κ(t)dt. |
Along the same line, the last term on the right-hand side of (3.3) is estimated by
|∫T0∫R3∂tΓαv⋅∇Γαpdxdt|=|∫T0∫R3∂t(∇⋅Γαv)⋅Γαpdxdt|=|∫T0∫R3∇⋅[∑β+γ+ι=α∂tΓβv1(∂2Γγv2∂3Γιv3−∂3Γγv2∂2Γιv3∂3Γγv2∂1Γιv3−∂1Γγv2∂3Γιv3∂1Γγv2∂2Γιv3−∂2Γγv2∂1Γιv3)−∑β+γ=α(∂tΓβv1∂2Γγv2+∂tΓβv1∂3Γγv3−∂2Γβv1∂tΓγv2−∂3Γβv1∂tΓγv3−∂tΓβv1∂1Γγv2+∂tΓβv2∂3Γγv3+∂1Γβv1∂tΓγv2−∂3Γβv2∂tΓγv3−∂tΓβv2∂2Γγv3−∂tΓβv1∂1Γγv3+∂2Γβv2∂tΓγv3+∂1Γβv1∂tΓγv3)]Γαpdxdt|≲∑β+γ+ι=α|γ|,|ι|≤|β|∫T0‖∂Γβv‖L2‖∂Γγv‖L∞(1+‖∂Γιv‖L∞)‖∇Γαp‖L2dt≲∑β+γ+ι=α|γ|,|ι|≤|β|∫T0⟨t⟩−32‖∂Γβv‖L2(1+‖∂Γιv‖L∞)(⟨t⟩32‖∂Γγv‖L∞(R)‖∇Γαp‖L2(R)+‖⟨r⟩(1−ξ(s))∂Γγv‖L∞⟨t⟩‖∇Γαp‖L2(Rc))dt≲∫T0⟨t⟩−32Eκ(t)E12[κ/2]+3(t)dt. |
For the second term on the right-hand side of (3.2), by the definition of ˜S and (2.6), one has
−∑β+γ=α|β|≠|α|Cβα∫T0∫R3∂tΓαv⋅Γγη∂2tΓβvdxdt≲∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−1‖r−14⟨r⟩−12|∂tΓαv|‖L2‖⟨r⟩Γγη‖L∞‖∂t˜SΓβv‖L2dt+∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−2‖∂tΓαv‖L2‖⟨r⟩2Γγη‖L∞(‖⟨t−r⟩∂2tΓβv‖L2+‖⟨t−r⟩∂t∂rΓβv‖L2)dt+∑β+γ=α|β|≤|γ|∫T0⟨t⟩−2‖∂tΓαv‖L2‖⟨r⟩Γγη‖L2(⟨t⟩‖∂2tΓβv‖L∞+⟨t⟩‖∂t˜SΓβv‖L∞+⟨t⟩‖∂t∂rΓβv‖L∞)dt≲δLEκ(T)+δ∫T0⟨t⟩−2Eκ(t)dt. |
It is left to estimate the last term on the right-hand side of (3.2). Two cases are considered. By Lemmas 2, 4, and (2.6), we solve the case |γ|<|β|<|α| by
−∑β+γ=α|γ|<|β|<|α|Cβα∫T0∫R3∂tΓαv⋅[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]dxdt≲∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−2‖∂tΓαv‖L2⟨t⟩‖∇Γγv‖L∞⟨t⟩‖ρ∂2tΓβv−ΔΓβv‖L2dt+∑β+γ=α,ι1+ι2=β|γ|<|β|<|α|,|ι1|≤|ι2|∫T0⟨t⟩−2‖∂tΓαv‖L2⟨t⟩‖∇Γγv‖L∞‖Γι2η‖L2‖⟨r⟩⟨t−r⟩∂2tΓι1v‖L∞dt+∑β+γ=α,ι1+ι2=β|γ|<|β|<|α|,|ι2|<|ι1|∫T0⟨t⟩−2‖∂tΓαv‖L2⟨t⟩‖∇Γγv‖L∞‖⟨r⟩Γι2η‖L∞‖⟨t−r⟩∂2tΓι1v‖L2dt≲∫T0⟨t⟩−2E32κ(t)dt. |
Utilizing Lemmas 9 and 10, we solve the case |β|≤|γ| by
−∑β+γ=α|β|≤|γ|Cβα∫T0∫R3∂tΓαv⋅[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]dxdt≲∑β+γ=α|β|≤|γ|∫T0⟨t⟩−2‖∂tΓαv‖L2‖∇Γγv‖L2⟨t⟩2‖ρ∂2tΓβv−ΔΓβv‖L∞dt+∑β+γ=α,ι1+ι2=β|β|≤|γ|∫T0⟨t⟩−2‖∂tΓαv‖L2‖∇Γγv‖L2⟨t⟩2‖Γι2η∂2tΓι1v‖L∞dt≲∫T0⟨t⟩−2E32κ(t)dt. |
By summing up the above estimates, we deduce that
|∫T0∫R3∂tΓαv⋅Nαdxdt|≲∫T0⟨t⟩−32E32κ(t)dt+δ∫T0⟨t⟩−32Eκ(t)dt+δLEκ(T). |
We continue to handle the last term on the right-hand side of (3.1). The identity (2.4) yields that
∫T0∫R3fk(∂rΓαv+Γαvr)⋅Nαdxdt=−∑β+γ=α|β|≠|α|Cβα∫T0∫R3fk(∂rΓαv+Γαvr)⋅Γγη∂2tΓβvdxdt−∑β+γ=α|β|≠|α|Cβα∫T0∫R3fk(∂rΓαv+Γαvr)⋅[(∇X)−T(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]dxdt−∫T0∫R3fk(∂rΓαv+Γαvr)⋅(∇X)−T∇Γαpdxdt. | (3.4) |
We use Lemmas 9 and 10 to handle the first two terms on the right-hand side of (3.4) by
∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−1‖r−14⟨r⟩−12(|∂rΓαv|+|Γαv|r)‖L2‖⟨r⟩2Γγη‖L∞‖⟨t−r⟩∂2tΓβv‖L2dt+∑β+γ=α|β|≤|γ|∫T0⟨t⟩−1‖r−14⟨r⟩−12(|∂rΓαv|+|Γαv|r)‖L2‖⟨r⟩Γγη‖L2‖⟨r⟩⟨t−r⟩∂2tΓβv‖L∞dt+∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2⟨t⟩‖∇Γγv‖L∞(⟨t⟩‖∂2tΓβv−ΔΓβv‖L2+∑ι1+ι2=β‖⟨r⟩⟨t−r⟩Γι2η∂2tΓι1v‖L2)dt+∑β+γ=α|β|≤|γ|∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2‖∇Γγv‖L2(⟨t⟩2‖ ρ∂2tΓβv−ΔΓβv‖L∞+∑ι1+ι2=β⟨t⟩2‖Γι2η∂2tΓι1v‖L∞)dt≲δLEκ(T)+δ∫T0⟨t⟩−2Eκ(t)dt+∫T0⟨t⟩−2E32κ(t)dt. |
We rewrite the last term on the right-hand side of (3.4) as follows
−∫T0∫R3fk(∂rΓαv+Γαvr)⋅(∇X)−T∇Γαpdxdt≲∑i,j=1,2,3i≠j∫T0∫R3(|fk(∂rΓαvi+Γαvir)∂jvj∂iΓαp|+|fk(∂rΓαvi+Γαvir)∂ivj∂jΓαp|)dxdt+∑j,l,m,n,s=1,2,3i≠j≠l∫T0∫R3|fk(∂rΓαvi+Γαvir)∂mvj∂nvl∂sΓαp|dxdt+|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Γαpdxdt|. | (3.5) |
For the first two terms on the right-hand side of (3.5), the Lemma 6 and (2.19) imply that
∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2(1+‖∇v‖L∞)‖⟨r⟩(1−ξ(s))∇v‖L∞⟨t⟩‖∇Γαp‖L2(Rc)dt+∫T0⟨t⟩−32‖|∂rΓαv|+|Γαv|r‖L2(1+‖∇v‖L∞)⟨t⟩32‖∇v‖L∞(R)‖∇Γαp‖L2(R)dt≲∫T0⟨t⟩−32E32κ(t)dt. |
We use (2.7) to formulate the last term on the right-hand side of (3.5) by
|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Γαpdxdt|≲∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(∇v)T(ρ∂2tΓαv−ΔΓαv)]|dxdt+∑β+γ=α|β|≠|α|∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(∇X)TΓγη∂2tΓβv]|dxdt+∑β+γ=α|β|≠|α|∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(∇Γγv)T(∂2tΓβv−ΔΓβv+∑ι1+ι2=βCι1βΓι2η∂2tΓι1v)]|dxdt+|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(ρ∂2tΓαv−ΔΓαv)dxdt|. | (3.6) |
By (2.19) and Lemma 6, we estimate the first term on the right-hand side of (3.6) by
∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(∇v)T(ρ∂2tΓαv−ΔΓαv)]|dxdt≲∫T0⟨t⟩−32‖|∂rΓαv|+|Γαv|r‖L2⟨t⟩32‖∇v‖L∞(R)‖ρ∂2tΓαv−ΔΓαv‖L2(R)dt+∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2‖⟨r⟩(1−ξ(s))∇v‖L∞ ⟨t⟩‖ρ∂2tΓαv−ΔΓαv‖L2(Rc)dt≲∫T0⟨t⟩−32E32κ(t)dt. | (3.7) |
The second and third terms on the right-hand side of (3.6) can be solved using the same method employed by the first two terms of (3.4).
For the last term on the right-hand side of (3.6), we observe that
|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(ρ∂2tΓαv−ΔΓαv)dxdt|≲|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Δ−1(ρ∂2t∇⋅Γαv−Δ∇⋅Γαv)dxdt|+∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2tΓαv)|dxdt. | (3.8) |
In view of (2.5), we write the first term on the right-hand side of (3.8) by
∑i,j=1,2,3i≠j∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(ρ∂2tΓαvi−ΔΓαvi)∂vj]|dxdt+∑i,j,l=1,2,3i≠j≠l∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅[(ρ∂2tΓαvi−ΔΓαvi)∂vj∂vl]|dxdt+∑β+γ+ι=α,|β|≠|α|i,j,l=1,2,3,i≠j≠l∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(∂2Γβvi∂Γγvj∂Γιvl)|dxdt+∑β+γ+ι=α,|β|≠|α|i,j,l=1,2,3,i≠j≠m∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∂η∂2Γβvi∂Γγvj∂Γιvl)|dxdt+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j|∫T0∫R3fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(∂2Γβvi∂Γγvj)dxdt|+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∂η∂2Γβvi∂Γγvj)|dxdt. | (3.9) |
The first two terms on the right-hand side of (3.9) are dealt with by the same method as (3.7). By (2.6), the third and fourth terms on the right-hand side of (3.9) are estimated by
∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2(1+‖∇η‖L3)(∑β+γ+ι=α|γ|,|ι|<|β|<|α|‖∂2Γβv‖L2⟨t⟩‖∂Γγv‖L∞⟨t⟩‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|ι|≤|γ|⟨t⟩‖∂2Γβv‖L∞‖∂Γγv‖L2⟨t⟩‖∂Γιv‖L∞+∑β+γ+ι=α|β|,|γ|≤|ι|⟨t⟩‖∂2Γβv‖L∞⟨t⟩‖∂Γγv‖L∞‖∂Γιv‖L2)dt≲∫T0⟨t⟩−2E32κ(t)dt. |
To consider the fifth term on the right-hand side of (3.9), we separate two cases to consider it. For the case x∈R, by (2.20), we have
∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j|∫T0∫R3ξ(s)fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(∂2Γβvi∂Γγvj)dxdt|≲∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0‖r−14⟨r⟩−12(|∂rΓαv+|Γαv|r)‖L2‖⟨t−r⟩∂2Γβvi∂Γγvj‖L2dt≲∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−1‖r−14⟨r⟩−12(|∂rΓαv+|Γαv|r)‖L2‖⟨t−r⟩∂2Γβv‖L2⟨t⟩‖∂Γγv‖L∞dt+∑β+γ=α|β|≤|γ|∫T0⟨t⟩−1‖r−14⟨r⟩−12(|∂rΓαv+|Γαv|r)‖L2(‖⟨r⟩⟨t−r⟩(1−ξ(s))∂2Γβv‖L∞+⟨t⟩2‖∂2Γβv‖L∞(R))‖∂Γγv‖L2dt≲sup0≤t≤TE12κ(t)LEκ(T)+∫T0⟨t⟩−2E32κ(t)dt. |
To consider the case x∈Rc, we use the integration by parts to get
∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j|∫T0∫R3[1−ξ(s)]fk(∂rΓαv+Γαvr)⋅∇Δ−1∇⋅(∂2Γβvi∂Γγvj)dxdt|≲∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j|∫T0∫R3∇⋅[[1−ξ(s)]fk(∂rΓαv+Γαvr)]⋅Δ−1∇⋅(∂2Γβvi∂Γγvj)dxdt|≲∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0⟨t⟩−1‖|∂rΓαv|+|Γαv|r‖L2‖Δ−1∇⋅(∂2Γβvi∂Γγvj)‖L2dt+∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0‖∇⋅Γαv‖L2‖∂2Γβvi∂Γγvj‖L2dt. | (3.10) |
For the first term on the right-hand side of (3.10), we have
∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0⟨t⟩−1‖|∂rΓαv|+|Γαv|r‖L2‖Δ−1∇⋅(∂2Γβvi∂Γγvj)‖L2dt≲∑β+γ=α|β|≤|γ|∫T0⟨t⟩−43‖|∂rΓαv|+|Γαv|r‖L2⟨t⟩13‖∂2Γβv‖13L∞‖∂2Γβv‖23L2‖∂Γγv‖L2dt+∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−43‖|∂rΓαv|+|Γαv|r‖L2‖∂2Γβv‖L2‖∂Γγv‖23L2⟨t⟩13‖∂Γγv‖13L∞dt≲∫T0⟨t⟩−43E32κ(t)dt. |
We use (2.5) to estimate the second term on the right-hand side of (3.10) by
∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0‖∇⋅Γαv‖L2‖∂2Γβvi∂Γγvj‖L2dt≲∑β+γ=α,˜β+˜γ+˜ι=α|β|≤|γ|,|˜γ|,|˜ι||≤|˜β|∫T0⟨t⟩−2‖∂Γ˜βv‖L2⟨t⟩‖∂Γ˜γv‖L∞(1+‖∂Γ˜ιv‖L∞)⟨t⟩‖∂2Γβv‖L∞‖∂Γγv‖L2dt+∑β+γ=α,˜β+˜γ+˜ι=α|γ|<|β|<|α|,|˜γ|,|˜ι|≤|˜β|∫T0⟨t⟩−2‖∂Γ˜βv‖L2⟨t⟩‖∂Γ˜γv‖L∞(1+‖∂Γ˜ιv‖L∞)‖∂2Γβv‖L2⟨t⟩‖∂Γγv‖L∞dt≲∫T0⟨t⟩−2E2κ(t)dt. |
For the last term on the right-hand side of (3.9), it follows from the Sobolev embedding inequality, (2.6), and (2.19), that
∑β+γ=α,|β|≠|α|i,j=1,2,3,i≠j∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∂η∂2Γβvi∂Γγvj)|dxdt≲∑β+γ=α|γ|<|β|<|α|∫T0⟨t⟩−2‖|∂rΓαv|+|Γαv|r‖L2‖⟨r⟩∇η‖L3‖⟨t−r⟩∂2Γβv‖L2⟨t⟩‖∂Γγv‖L∞dt+∑β+γ=α|β|≤|γ|∫T0⟨t⟩−32‖|∂rΓαv|+|Γαv|r‖L2‖∂Γγv‖L2(‖∇η‖L3(R)⟨t⟩32‖∂2Γβv‖L∞(R)+‖⟨r⟩(1−ξ(s))∇η‖L3‖⟨r⟩(1−ξ(s))∂2Γβv‖L∞)dt≲∫T0⟨t⟩−32E32κ(t)dt. |
To solve the second term on the right-hand side of (3.8), we rewrite it as follows
∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2tΓαv)|dxdt≲∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇Δ−1∇⋅Γαv)|dxdt+∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇⊥Δ−1∇⊥⋅Γαv)|dxdt. | (3.11) |
In view of (2.5), we formulate the first term on the right-hand side of (3.11) by
∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇Δ−1∇⋅Γαv)|dxdt≲∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇Δ−1(∂lΓβvi∂mΓγvj))|dxdt+∑β+γ+ι=αi,j,l=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇Δ−1(∂iΓβv1∂jΓγv2∂lΓιv3))|dxdt. | (3.12) |
Since the second term has analogous estimates to the first term of (3.12), it suffices to concentrate on the first term. We observe that
∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∂2t∇Δ−1(∂lΓβvi∂mΓγvj))|dxdt≲∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂2t∂lΓβvi∂mΓγvj))|dxdt+∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂t∂lΓβvi∂t∂mΓγvj))|dxdt+∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂lΓβvi∂2t∂mΓγvj))|dxdt. | (3.13) |
Here, we restrict our analysis to the first term on the right-hand side of (3.13), as the remaining two terms have similar estimates.
∑β+γ=αi,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂2t∂lΓβvi∂mΓγvj))|dxdt≲∑β+γ=α,|γ|<|β|i,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂2tΓβvi∂l∂mΓγvj))|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1∂l(∂2tΓβvi∂mΓγvj))|dxdt+∑β+γ=α,|β|≤|γ|i,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂2t∂lΓβvi∂mΓγvj))|dxdt. | (3.14) |
For the first term on the right-hand side of (3.14), we deduce from (2.4) that
∑β+γ=α,|γ|<|β|i,jl,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(∂2tΓβvi∂l∂mΓγvj))|dxdt≲∑β+γ=α,|γ|<|β|i,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(ρ−1ΔΓβvi∂l∂mΓγvj))|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(∇η⋅∇Δ−1(ρ−1Nβi∂l∂mΓγvj))|dxdt. | (3.15) |
We formulate the first term on the right-hand side of (3.15) as follows
∑β+γ=α,|γ|<|β|i,j,l,m,n=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇(η∇Δ−1∂n(ρ−1∂nΓβvi∂l∂mΓγvj))|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m,n=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∂n(ρ−1η∂nΓβvi∂l∂mΓγvj)|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m,n=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1∇(η∇Δ−1[∂nΓβvi∂n(ρ−1∂l∂mΓγvj)])|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m,n=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(ρ−1∂nη∂nΓβvi∂l∂mΓγvj)|dxdt+∑β+γ=α,|γ|<|β|i,j,l,m,n=1,2,3∫T0∫R3|fk(∂rΓαv+Γαvr)⋅∇Δ−1(η∂nΓβvi∂n(ρ−1∂l∂mΓγvj))|dxdt. | (3.16) |
The first three terms on the right-hand side of (3.16) can be estimated by
We utilize (2.19) to solve the last two terms on the right-hand side of (3.16) by
Along the same line, the second term on the right-hand side of (3.15) can be handled by
The same estimates hold for the last two terms on the right-hand side of (3.14).
Applying (2.4), we formulate the second term on the right-hand side of (3.11) as follows
(3.17) |
The first term on the right-hand side of (3.17) is solved by
The calculations in (2.12) imply that
(3.18) |
Substituting (3.18) into the second term on the right-hand side of (3.17), we have
(3.19) |
For the third term on the right-hand side of (3.17), we have
(3.20) |
We use (2.4) to formulate the first term on the right-hand side of (3.20) as follows
(3.21) |
For the first term on the right-hand side of (3.21), we have
The second term on the right-hand side of (3.21) can be solved as (3.19). We employ the analogous method utilized for the first two terms on the right-hand side of (3.4) to solve the last two terms on the right-hand side of (3.21).
For the second term on the right-hand side of (3.20), we have
For the last term on the right-hand side of (3.17), by Lemmas 2 and 6, we have
Combining all the estimates, we conclude that
where is some positive constant. By the smallness of , , and the standard continuity method, we arrive at the main result.
All authors contributed equally.
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
The work of the first author was partially supported by the National Natural Science Foundation of China under Grants 12401278. The work of the second author was partially supported by the RFS grant and GRF grants from the Research Grants Council (Project Nos. PolyU 11302021, 11310822, and 11302523). The authors would like to thank the research center for nonlinear analysis at PolyU for the opportunity of discussions and encouragement.
The authors declare there is no conflict of interest.
[1] | J. Collins, J. Sohl-Dickstein, D. Sussillo, Capacity and trainability in recurrent neural networks, Paper presented at 5th International Conference on Learning Representations, 2017. |
[2] |
S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput., 9 (1997), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 doi: 10.1162/neco.1997.9.8.1735
![]() |
[3] |
K. Cho, B. V. Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, et al., Learning phrase representations using RNN encoder-decoder for statistical machine translation, Comput. Sci., (2014), 1723–1734. https://doi.org/10.3115/v1/D14-1179 doi: 10.3115/v1/D14-1179
![]() |
[4] |
L. Bottou, F. E. Curtis, J. Nocedal, Optimization methods for large-scale machine learning, SIAM, 60 (2018), 223–231. https://doi.org/10.1137/16M1080173 doi: 10.1137/16M1080173
![]() |
[5] | B. Chang, M. Chen, E. Haber, E. H. Chi, AntisymmetricRNN: a dynamical system view on recurrent neural networks, Paper presented at 7th International Conference on Learning Representations, 2019. |
[6] | E. Haber, K. Lensink, T. Eran, L. Ruthotto, IMEXnet: A forward stable deep neural network, Paper ppresented at: Proceedings of the 36th International Conference on Machine Learning, 2019. |
[7] |
W. E, A proposal on machine learning via dynamical systems, Commun. Math. Stat., 1 (2017), 1–11. http://doi.org/10.1007/s40304-017-0103-z doi: 10.1007/s40304-017-0103-z
![]() |
[8] | Y. Lu, A. Zhong, Q. Li, B. Dong, Beyond finite layer neural networks: bridging deep architectures and numerical differential equations, Ppaer pressented at: Proceedings of Machine Learning Research Proceedings of the 35th International Conference on Machine Learning, 2018. |
[9] | R. T. Q. Chen, Y. Rubanova, J. Bettencourt, D. Duvenaud, Neural ordinary differential equations, Paper present at : Annual Conference on Neural Information Processing Systems, 2018. |
[10] |
P. Meng, L. Su, W. Yin, S. Zhang, Solving a kind of inverse scattering problem of acoustic waves based on linear sampling method and neural network, Alexandria Eng. J., 59 (2020), 1451–1462. http://doi.org/10.1016/j.aej.2020.03.047 doi: 10.1016/j.aej.2020.03.047
![]() |
[11] |
W. Yin, W. Yang, H. Liu, A neural network scheme for recovering scattering obstacles with limited phaseless far-field data, J. Comput. Phys., 417 (2020). http://doi.org/10.1016/j.jcp.2020.109594 doi: 10.1016/j.jcp.2020.109594
![]() |
[12] |
Y. Guo, D. Hoemberg, G. Hu, J. Li, H. Liu, A time domain sampling method for inverse acoustic scattering problems, J. Comput. Phys., 314 (2016), 647–660. http://doi.org/10.1016/j.jcp.2016.03.046 doi: 10.1016/j.jcp.2016.03.046
![]() |
[13] |
D. Zhang, F. Sun, L. Lu, Y. Guo, A harmonic polynomial method with a regularization strategy for the boundary value problems of Laplace's equation, Appl. Math. Lett, 79 (2018), 100–104. http://doi.org/10.1016/j.aml.2017.12.003 doi: 10.1016/j.aml.2017.12.003
![]() |
[14] |
H. Liu, M. Petrini, L. Rondi, J. Xiao, Stable determination of sound-hard polyhedral scatterers by a minimal number of scattering measurements, J. Differ. Equations, 262 (2018), 1631–1670. http://doi.org/10.1016/j.jde.2016.10.021 doi: 10.1016/j.jde.2016.10.021
![]() |
[15] |
H. Liu, X. Liu, Recovery of an embedded obstacle and its surrounding medium from formally determined scattering data, Inverse Probl., 33 (2017), 1–20. http://doi.org/10.1088/1361-6420/aa6770 doi: 10.1088/1361-6420/aa6770
![]() |
[16] |
H. Liu, X. Liu, X. Wang, Y. Wang, On a novel inverse scattering scheme using resonant modes with enhanced imaging resolution, Inverse Probl., 35 (2019). http://doi.org/10.1088/1361-6420/ab2932 doi: 10.1088/1361-6420/ab2932
![]() |
[17] |
W. Yin, J. Ge, P. Meng, F. Qu, A neural network method for the inverse scattering problem of impenetrable cavities, Electron. Res. Arch., 28 (2020), 1123–1142. http://doi.org/10.3934/era.2020062 doi: 10.3934/era.2020062
![]() |
[18] | J. Xie, Y. chen, A numerical analysis method of fixed points and their stability in a complex dynamical systems, CCAMMS, (2019), 59–62. |
[19] |
W. Stephen, D. S. Mazel, Introduction to applied nonlinear dynamical systems and chaos, Name J., 33 (1991), 81. http://doi.org/10.1063/1.4822950 doi: 10.1063/1.4822950
![]() |
[20] | G. Wang, Ordinary differential equations, Higher Education Press, (2013), 250–260. |
[21] |
E. Haber, L. Ruthotto, Stable architectures for deep neural networks, Inverse Probl., 34 (2018). http://doi.org/10.1088/1361-6420/aa9a90 doi: 10.1088/1361-6420/aa9a90
![]() |
[22] | C. Bo, M. Chen, E. Haber, E. Chi, Antisymmetricrnn: a dynamical system view on recurrent neural networks, Paper present at: 7th International Conference on Learning Representations, 2019. |
[23] | U. Ascher, L. Petzold, Computer methods for ordinary differential equations and differential-algebraic equations, in Science press, 2009. |
[24] |
D. Colton, Inverse acoustic and electromagnetic scattering theory, Inverse Prob., 47 (2003), 67–110. http://doi.org/10.1007/978-3-662-03537-5 doi: 10.1007/978-3-662-03537-5
![]() |
[25] |
D. Colton, J. Coyle, P. Monk, Recent developments in inverse acoustic scattering theory, SIAM, 42 (2000), 369–414. http://doi.org/10.1137/S0036144500367337 doi: 10.1137/S0036144500367337
![]() |
[26] |
A. Kirsch, New characterizations of solutions in inverse scattering theory, Appl. Anal., 76 (2000), 319–350. http://doi.org/10.1080/00036810008840888 doi: 10.1080/00036810008840888
![]() |
[27] | A. Kirsch, N. Grinberg, The factorization method for inverse problems, Oxford University Press, 2008. |
[28] |
F. Zeng, P. Suarez, J. Sun, A decomposition method for an interior inverse scattering problem, Inverse Probl. Imaging, 7 (2013), 291–303. http://doi.org/10.3934/ipi.2013.7.291 doi: 10.3934/ipi.2013.7.291
![]() |
[29] |
J. Li, H. Liu, W. Tsui, X. Wang, An inverse scattering approach for geometric body generation: a machine learning perspective, Math. Eng., 1 (2019), 800–823. http://doi.org/10.3934/mine.2019.4.800 doi: 10.3934/mine.2019.4.800
![]() |
[30] |
J. Li, H. Liu, J. Zou, Locating multiple multiscale acoustic scatterers, Multiscale Model. Simul., 12 (2014), 927–952. http://doi.org/10.1137/13093409X doi: 10.1137/13093409X
![]() |
[31] |
H. Liu, J. Zou, Uniqueness in an inverse acoustic obstacle scattering problem for both sound-hard and sound-soft polyhedral scatterers, Inverse Probl., 22 (2016), 515–524. http://doi.org/10.1088/0266-5611/22/2/008 doi: 10.1088/0266-5611/22/2/008
![]() |
[32] | Q. V. Le, N. Jaitly, G. E. Hinton, A simple way to initialize recurrent networks of rectified linear units, Comput. Sci., preprint, arXiv: 1504.00941. |