Loading [MathJax]/jax/output/SVG/jax.js

Estimating nonstationary inputs from a single spike train based on a neuron model with adaptation

  • Because every spike of a neuron is determined by input signals, a train of spikes may contain information about the dynamics of unobserved neurons. A state-space method based on the leaky integrate-and-fire model, describing neuronal transformation from input signals to a spike train has been proposed for tracking input parameters represented by their mean and fluctuation [11]. In the present paper, we propose to make the estimation more realistic by adopting an LIF model augmented with an adaptive moving threshold. Moreover, because the direct state-space method is computationally infeasible for a data set comprising thousands of spikes, we further develop a practical method for transforming instantaneous firing characteristics back to input parameters. The instantaneous firing characteristics, represented by the firing rate and non-Poisson irregularity, can be estimated using a computationally feasible algorithm. We applied our proposed methods to synthetic data to clarify that they perform well.

    Citation: Hideaki Kim, Shigeru Shinomoto. Estimating nonstationary inputs from a single spike train based on a neuron model with adaptation[J]. Mathematical Biosciences and Engineering, 2014, 11(1): 49-62. doi: 10.3934/mbe.2014.11.49

    Related Papers:

    [1] Junjie Wang, Yaping Zhang, Liangliang Zhai . Structure-preserving scheme for one dimension and two dimension fractional KGS equations. Networks and Heterogeneous Media, 2023, 18(1): 463-493. doi: 10.3934/nhm.2023019
    [2] Tingting Ma, Yayun Fu, Yuehua He, Wenjie Yang . A linearly implicit energy-preserving exponential time differencing scheme for the fractional nonlinear Schrödinger equation. Networks and Heterogeneous Media, 2023, 18(3): 1105-1117. doi: 10.3934/nhm.2023048
    [3] Fengli Yin, Dongliang Xu, Wenjie Yang . High-order schemes for the fractional coupled nonlinear Schrödinger equation. Networks and Heterogeneous Media, 2023, 18(4): 1434-1453. doi: 10.3934/nhm.2023063
    [4] Yves Achdou, Victor Perez . Iterative strategies for solving linearized discrete mean field games systems. Networks and Heterogeneous Media, 2012, 7(2): 197-217. doi: 10.3934/nhm.2012.7.197
    [5] Wen Dong, Dongling Wang . Mittag-Leffler stability of numerical solutions to linear homogeneous time fractional parabolic equations. Networks and Heterogeneous Media, 2023, 18(3): 946-956. doi: 10.3934/nhm.2023041
    [6] Farman Ali Shah, Kamran, Dania Santina, Nabil Mlaiki, Salma Aljawi . Application of a hybrid pseudospectral method to a new two-dimensional multi-term mixed sub-diffusion and wave-diffusion equation of fractional order. Networks and Heterogeneous Media, 2024, 19(1): 44-85. doi: 10.3934/nhm.2024003
    [7] Jin Cui, Yayun Fu . A high-order linearly implicit energy-preserving Partitioned Runge-Kutta scheme for a class of nonlinear dispersive equations. Networks and Heterogeneous Media, 2023, 18(1): 399-411. doi: 10.3934/nhm.2023016
    [8] Yaxin Hou, Cao Wen, Yang Liu, Hong Li . A two-grid ADI finite element approximation for a nonlinear distributed-order fractional sub-diffusion equation. Networks and Heterogeneous Media, 2023, 18(2): 855-876. doi: 10.3934/nhm.2023037
    [9] Jinhu Zhao . Natural convection flow and heat transfer of generalized Maxwell fluid with distributed order time fractional derivatives embedded in the porous medium. Networks and Heterogeneous Media, 2024, 19(2): 753-770. doi: 10.3934/nhm.2024034
    [10] Caihong Gu, Yanbin Tang . Global solution to the Cauchy problem of fractional drift diffusion system with power-law nonlinearity. Networks and Heterogeneous Media, 2023, 18(1): 109-139. doi: 10.3934/nhm.2023005
  • Because every spike of a neuron is determined by input signals, a train of spikes may contain information about the dynamics of unobserved neurons. A state-space method based on the leaky integrate-and-fire model, describing neuronal transformation from input signals to a spike train has been proposed for tracking input parameters represented by their mean and fluctuation [11]. In the present paper, we propose to make the estimation more realistic by adopting an LIF model augmented with an adaptive moving threshold. Moreover, because the direct state-space method is computationally infeasible for a data set comprising thousands of spikes, we further develop a practical method for transforming instantaneous firing characteristics back to input parameters. The instantaneous firing characteristics, represented by the firing rate and non-Poisson irregularity, can be estimated using a computationally feasible algorithm. We applied our proposed methods to synthetic data to clarify that they perform well.


    This paper mainly focuses on constructing and analyzing an efficient energy-preserving finite difference method (EP-FDM) for solving the nonlinear coupled space-fractional Klein-Gordon (KG) equations:

    $ uttκ2dk=1αkxku+a1u+b1u3+c1uv2=0, $ (1.1)
    $ vttκ2dk=1αkxkv+a2v+b2v3+c2u2v=0, $ (1.2)

    with $ (\boldsymbol{x}, t) \in \Omega \times[0, T] $ and the following widely used boundary and initial conditions

    $ (u(x,t),v(x,t))=(0,0),(x,t)Ω×[0,T], $ (1.3)
    $ (u(x,0),v(x,0))=(ϕ1(x),ϕ2(x)),xˉΩ, $ (1.4)
    $ (ut(x,0),vt(x,0))=(φ1(x),φ2(x)),xˉΩ. $ (1.5)

    Here, $ \boldsymbol{x} = (x_{1}, ..., x_{d})^{T}(d = 1, 2, 3) \in \Omega\subset R^{d} $, $ \partial \Omega $ is the boundary of $ \Omega $, $ \bar{\Omega} = \Omega\bigcup\partial \Omega $, $ \kappa $ is a constant and $ a_{i}, b_{i}, c_{i} $ are all positive constants. $ \phi_1, \phi_2, \varphi_1, \varphi_2 $ are all known sufficiently smooth functions. $ u(\boldsymbol{x}, t) $, $ v(\boldsymbol{x}, t) $ are interacting relativistic fields of masses, $ \partial_{x_k}^{\alpha_k}u $ and $ \partial_{x_k}^{\alpha_k}v $ stand for the Riesz fractional operator with $ 1 < \alpha_{k}\leq 2, (k = 1, ..., d) $ in $ x_{k} $ directions, which are well defined as follows

    $ αkxku(x,t)=12cos(αkπ/2)[Dαkxku(x,t)+xkDαk+u(x,t)], $ (1.6)

    where $ -\infty D_{x_k}^{\alpha_k} u(\boldsymbol{x}, t) $ and $ { }_{x_k} D_{+\infty}^{\alpha_k} u(\boldsymbol{x}, t) $ are the left and right Riemann-Liouville fractional derivative.

    Plenty of physical phenomena, such as the long-wave dynamics of two waves, are represented by the system (1.2). For example, these equations are used to study a number of issues in solid state physics, relativistic mechanics, quantum mechanics, and classical mechanics [1,2,3,4].

    Especially, when $ \alpha_{k} $ tends to 2, the fractional derivative $ \partial_{x_k}^{\alpha_k} $ would converge to the second-order Laplace operator, and thus Eqs (1.1) and (1.2) reduce to the classical system of multi-dimensional coupled KG equations [5,6,7]. The system has the following conserved energy, which is mentioned in detail in [11],

    $ E(t) = \frac{1}{2} \int_{\Omega}\left[\frac{1}{c_{1}}\left(u_t\right)^2+\frac{\kappa^{2}}{c_{1}}|\nabla u|^2+\frac{1}{c_{2}}\left(v_t\right)^2+\frac{\kappa^{2}}{c_{2}}|\nabla v|^2+2 G(u, v)\right] d {\Omega} = E(0), $

    where

    $ G(u, v) = \frac{b_1}{4 c_1} u^4+\frac{b_2}{4 c_2} v^4+\frac{a_1}{2 c_1} u^2+\frac{a_2}{2 c_2} v^2+\frac{1}{2} u^2 v^2. $

    The coupled KG equations is initially introduced in [8] and is applied to model the usual motion of charged mesons within a magnetic field. There have been many works for solving the classical KG equations. Tsutsumi [9] considered nonrelativistic approximation of nonlinear KG equations and proved the convergence of solutions rigorously. Joseph [10] obtained some exact solutions for these systems. Deng [11] developed two kinds of energy-preserving finite difference methods for the systems of coupled sine-Gordon (SG) equations or coupled KG equations in two dimensions. He [12] analyzed two kinds of energy-preserving finite element approximation schemes for a class of nonlinear wave equation. Zhu [13] developed the finite element method and the mesh-free deep neural network approach in a comparative fashion for solving two types of coupled nonlinear hyperbolic/wave partial differential equations. Deng [14] proposed a two-level linearized compact ADI method for solving the nonlinear coupled wave equations. More relevant and significant references can be found in [15,16,17].

    However, it has been found that fractional derivatives can be used to describe some physical problems with the spatial non-locality of anomalous diffusion. Therefore, more attentions have been paid to fractional KG equations. There are also some related numerical methods for the related models. These methods may be applied to solve the fractional KG systems. For example, Cheng [18] constructed a linearized compact ADI scheme for the two-dimensional Riesz space fractional nonlinear reaction-diffusion equations. Wang [19] proposed Fourier spectral method to solve space fractional KG equations with periodic boundary condition. Liu [20] presented an implicit finite difference scheme for the nonlinear time-space-fractional Schrödinger equation. Cheng [21] constructed an energy-conserving and linearly implicit scheme by combining the scalar auxiliary variable approach for the nonlinear space-fractional Schrödinger equations. Similar scalar auxiliary variable approach can also be found in [22,23]. Wang et al. [24,25] developed some energy-conserving schemes for space-fractional Schrödinger equations. Meanwhile, the equations are also investigated by some analytical techniques, such as the Fourier transform method [26], the Mellin transform method [27] and so on. Besides, the spatial disccretization of the KG equations usually gives a system of conservative ordinary differential equations. There are also some energy-conserving time discretizations, such as the implicit midpoint method [28], some Runge–Kutta methods [28,29], relaxation methods [30,31,32] and so on [33,34]. To the best of our knowledge, there exist few reports on numerical methods for coupled space-fractional KG equations. Most references focus on the KG equations rather than the coupled systems.

    The main purpose of this paper is to develop an EP-FDM for the system of nonlinear coupled space-fractional KG equations. Firstly, we transform the coupled systems of KG equations into an equivalent general form and provide energy conservation for the new system. Secondly, we propose a second-order consistent implicit three-level scheme by using the finite difference method to solve problems (1.1) and (1.2). Thirdly, we give the proof of the discrete energy conservation, boundedness of numerical solutions and convergence analysis in discrete $ L^2 $ norm. More specifically, the results show that the proposed schemes are energy-conserving. And the schemes have second-order accuracy in both the temporal and spatial directions. Finally, numerical experiments are presented to show the performance of our proposed scheme in one and two dimensions. They confirm our obtained theoretical results very well.

    The rest of the paper is organized as follows. Some denotations and preliminaries are given in Section 2. An energy-preserving scheme is constructed in Section 3. The discrete conservation law and boundedness of numerical solutions are given in Section 4. The convergence results are given in Section 5. Several numerical tests are offered to validate our theoretical results in Section 6. Finally, some conclusions are given in Section 7.

    Throughout the paper, we set $ C $ as a general positive constant that is independent of mesh sizes, which may be changed under different circumstances.

    We first rewrite Eqs (1.1) and (1.2) into an equivalent form

    $ αuttβdk=1αkxku+Gu(u,v)=0, $ (2.1)
    $ γvttσdk=1αkxkv+Gv(u,v)=0, $ (2.2)

    with the widely used boundary and initial conditions

    $ (u(x,t),v(x,t))=(0,0),(x,t)Ω×[0,T], $ (2.3)
    $ (u(x,0),v(x,0))=(ϕ1(x),ϕ2(x)),xˉΩ, $ (2.4)
    $ (ut(x,0),vt(x,0))=(φ1(x),φ2(x)),xˉΩ, $ (2.5)

    where $ G(u, v) = \frac{b_1}{4 c_1} u^4+\frac{b_2}{4 c_2} v^4+\frac{a_1}{2 c_1} u^2+\frac{a_2}{2 c_2} v^2+\frac{1}{2} u^2 v^2 $, and $ \alpha = 1/c_{1} $, $ \beta = \kappa^{2}/c_{1} $, $ \gamma = 1/c_{2} $, $ \sigma = \kappa^{2}/c_{2} $. A similar treatment is mentioned in [11]. The definition of operator $ \partial_{x_k}^{\alpha_k} $ is already presented in Eq (1.6), where the left and right Riemann-Liouville fractional derivatives in space of order $ \alpha $ are defined as

    $ Dαxu(x,t)=1Γ(2α)2x2xu(ξ,t)(xξ)α1dξ,(x,t)Ω,xDα+u(x,t)=1Γ(2α)2x2xu(ξ,t)(ξx)α1dξ,(x,t)Ω. $

    Theorem 1. Let $ {u(\boldsymbol{x}, t), v(\boldsymbol{x}, t)} $ be the solutions of this systems (2.1)–(2.5), the energy conservation law is defined by

    $ E(t)=12[αut2L2+βdk=1αk/2xku2L2+γvt2+σdk=1αk/2xkv2L2+2G(u,v),1]. $ (2.6)

    Namely, $ E(t) = E(0) $, where $ \|u(\cdot, t)\|_{L^2}^2 = \int_{\Omega}|u(\boldsymbol{x}, t)|^2 d \boldsymbol{x} $ and $ \langle G(u, v), 1 \rangle = \int_{\Omega}G(u, v)d \boldsymbol{x} $.

    Proof. Taking inner product of Eqs (2.1) and (2.2) with $ u_{t} $ and $ v_{t} $, then summing the obtained equations, and finally applying a integration over the time interval $ [0, t] $, it yields the required result.

    The finite difference method is used to achieve spatial and temporal discretization in this paper. We now denote temporal step size by $ \tau $, let $ \tau = T/N $, $ t_{n} = n\tau $. For a list of functions $ \{w^n\} $, we define

    $ wˉn=wn+1+wn12,δtwn=wn+1wnτ,μtwn=wn+1+wn2,Dtwn=wn+1wn12τ=δtwn+δtwn12,δ2twn=wn+12wn+wn1τ2=δtwnδtwn1τ. $

    Let $ \Omega = (a_{1}, b_{1})\times\cdots (a_{d}, b_{d}) $, with the given positive integers $ M_{1}, \cdots, M_{d} $, for the convenience of subsequent proofs, we have set it uniformly to M, so we get $ h_{k} = (b_{k}-a_{k}) / M_{k} = h \ (k = 1, \cdots, d) $ be the spatial stepsizes in $ x_{k} $-direction, then the spatial mesh is defined as $ \bar{\Omega}_{h} = \{(x_{k_{1}}, x_{k_{2}}, \cdots, x_{k_{d}}) \mid 0 \leq k_{s} \leq M_{s}, s = 1, \cdots, d \}$, where $ x_{k_{s}} = a_{s}+k_{s} h_{s} $.

    Moreover, we define the space $ \mathcal{V}_{h}^{0} $ as follows by using the grid function on $ \Omega_{h} $,

    $ \mathcal{V}_{h}^{0}: = \{v = v^{n}_{k_{1}\cdots k_{d}} \mid v_{k_{1}\cdots k_{d}}^n = 0\; \; \text{for} \; \; (k_{1}, \cdots, k_{d}) \in \partial \Omega_{h}\}, $

    where $ 1\le k_{s}\le M_{s}-1, \ s = 1, \cdots, d, \ 0\leq n\leq N $. Then we write $ \delta_{x_{1}} u_{k_{1}\cdots k_{d}} = \frac{u_{k_{1}+1\cdots k_{d}}-u_{k_{1}\cdots k_{d}}}{h} $. Notations $ \delta_{x_{s}} u_{k_{1}\cdots k_{d}} \ (s = 2, \cdots, d) $ are defined similarly.

    We then introduce the discrete norm, respectively. For $ u, v \in \mathcal{V}_{h}^{0} $, denote

    $ (u,v)=hdM11k1=1Md1kd=1uk1kdvk1kd,u=(u,u),|U|2H1=ds=1δxsU2,us=[hdM11k1=1Md1kd=1(uk1kd)s]1s. $

    Based on the definitions, we give the following lemmas which are important for this paper.

    Lemma 1. ([35]) Suppose $ p(x) \in L_1(\mathbb{R}) $ and

    $ p(x) \in \mathcal{C}^{2+\alpha}(\mathbb{R}): = \left\{p(x)\mid\int_{-\infty}^{+\infty}(1+|k|)^{2+\alpha}\left| \hat{p}(k) \right| d k < \infty\right\}, $

    where $ \hat{p}(k) $ is the Fourier transformation of $ p(x) $, then for a given $ h $, it holds that

    $ Dαxp(x)=1hα+k=0w(α)kp(x(k1)h)+O(h2),xDα+p(x)=1hα+k=0w(α)kp(x+(k1)h)+O(h2), $

    where $ w_k^{(\alpha)} $ are defined by

    $ {w(α)0=λ1g(α)0,w(α)1=λ1g(α)1+λ0g(α)0,w(α)k=λ1g(α)k+λ0g(α)k1+λ1g(α)k2,k2, $ (2.7)

    where $ \lambda_1 = \left(\alpha^2+3 \alpha+2\right) / 12, \quad \lambda_0 = \left(4-\alpha^2\right) / 6, \quad \lambda_{-1} = \left(\alpha^2-3 \alpha+2\right) / 12 $ and $ g_k^{(\alpha)} = (-1)^k\left(αk\right) $.

    In addition, we arrange in this section some of the lemmas that are necessary for the demonstration of later theorems in this paper.

    Lemma 2. ([36]) For any two grid functions $ u, v\in \mathcal{V}_{h}^{0} $, there exists a linear operator $ \Lambda^{\alpha} $ such that $ -(\delta_{x}^{\alpha}u, v) = (\Lambda^{\frac{\alpha}{2}}u, \Lambda^{\frac{\alpha}{2}}v) $, where the difference operator $ \Lambda^{\frac{\alpha}{2}} $ is defined by $ \Lambda^{\frac{\alpha}{2}} u = {\bf L} u $, and matrix $ {\bf L} $ satisfying $ {\bf C} = {\bf L}^{T}{\bf L} $ is the cholesky factor of matrix $ {\bf C} = 1/(2h^{\alpha}\cos(\alpha\pi/2))({\bf P}+{\bf P}^{T}) $ with

    $ {\bf P} = \left[w(α)1w(α)0w(α)2w(α)1w(α)0w(α)2w(α)1w(α)M2w(α)0w(α)M1w(α)M2w(α)2w(α)1\right]_{(M-1) \times(M-1)}. $

    While for multi-dimensional case, we give a further lemma.

    Lemma 3. ([18]) For any two grid functions $ u, v\in \mathcal{V}_{h}^{0} $, there exists a linear operator $ \Lambda_{k}^{\frac{\alpha_{k}}{2}} $ such that $ -(\delta_{x_{k}}^{\alpha_{k}}u, v) = (\Lambda_{k}^{\frac{\alpha_{k}}{2}}u, \Lambda_{k}^{\frac{\alpha_{k}}{2}}v), \ k = 1, \cdots, d, $ where $ \Lambda_{k}^{\frac{\alpha_{k}}{2}} $ is defined by $ \Lambda_{k}^{\frac{\alpha_{k}}{2}} u = [2\cos(\alpha_{k}\pi/2)h^{\alpha_{k}}]^{-1/2}{\bf L_{k}} u $, and matrix $ {\bf L_{k}} $ is given by $ -{\bf I} \otimes \cdots {\bf D}_{\alpha_{{\bf k}}} \otimes {\bf I} = $ $ \left[2 \cos \left(\alpha_k \pi / 2\right) h^{\alpha_k}\right]^{-1} {\bf L}_{{\bf k}}^T {\bf L}_{{\bf k}} $. $ {\bf I} $ is a unit matrix and matrix $ {\bf D}_{\alpha_{{\bf k}}} $ is given by $ {\bf D}_{\alpha_{{\bf k}}} = -1 /\left(2 \cos \left(\alpha_k \pi / 2\right) h^{\alpha_k}\right) $ $ \left({\bf P}_{{\bf k}}+{\bf P}_{{\bf k}}^T\right) $, $ {\bf P}_{{\bf k}} $ is the matrix $ {\bf P} $ in the case $ \alpha = \alpha_{k} $ as defined in Lemma 2.

    Lemma 4. ([11]) Let $ g(x) \in \mathcal{C}^4(I) $, then $ \forall x_0 \in I, x_0+\Delta x \in I $, we have

    $ g(x0+Δx)2g(x0)+g(x0Δx)Δx2=g(x0)+Δx2610[g(4)(x0+λΔx)+g(4)(x0λΔx)](1λ)3dλ,g(x0+Δx)+g(x0Δx)2=g(x0)+Δx210[g(x0+λΔx)+g(x0λΔx)](1λ)dλ. $

    Lemma 5. ([11]) Let $ u(\boldsymbol{x}, t), v(\boldsymbol{x}, t) \in \mathcal{C}^{4, 4}(\Omega \times[0, T]) $, and $ G(u, v) \in \mathcal{C}^{4, 4}\left(R^1 \times R^1\right) $. Then we have

    $ G(un+1,vn)G(un1,vn)un+1un1=Gu(u(x,tn),v(x,tn))+O(τ2),G(un,vn+1)G(un,vn1)vn+1vn1=Gv(u(x,tn),v(x,tn))+O(τ2). $

    Lemma 6. For any grid function $ u \in \mathcal{V}_h^0 $, it holds that

    $ \|u\|_p \leq C\|u\|^{C_{p_1}}\left(C_{p_2}|u|_{H^1}+\frac{1}{l}\|u\|\right)^{C_{p_3}}, \quad 2 \leq p < \infty, $

    where $ C_{p_1}, C_{p_2}, C_{p_3} $ are constants related to p, $ l = \min \left\{l_1, \cdots, l_d\right\} $, and d is the dimension of space $ \mathcal{V}_h^0 $.

    Specially, for two-dimensional case, the parameters $ C_{p_1} = \frac{2}{p} $, $ C_{p_2} = \max \left\{2 \sqrt{2}, \frac{p}{\sqrt{2}}\right\} $ and $ C_{p_3} = 1-\frac{2}{p} $ are shown in [37,38].

    While in the case of three dimensions, $ C_{p_1} = \frac{p+6}{4p} $, $ C_{p_2} = \max \left\{2 \sqrt{3}, \frac{p}{\sqrt{3}}\right\} $ and $ C_{p_3} = \frac{3p-6}{4p} $, the proof is given in Appendix.

    Lemma 7. ([39]) For $ M \geq 5, 1 \leq \alpha \leq 2 $ and any $ v \in \mathcal{V}_h^0 $, there exists a positive constant $ C_1 $, such that

    $ \|v\|^2 < \frac{\cos (\alpha \pi / 2)}{C_1 \ln 2}\left(\delta_x^\alpha v, v\right) = -\frac{\cos (\alpha \pi / 2)}{C_1 \ln 2}\left\|\Lambda^{\frac{\alpha}{2}} v\right\|^2 . $

    Specially, for multi-dimensional case, it can be written as $ \|v\|^2 < C \sum_{k = 1}^d\left\|\Lambda_{k}^{\frac{\alpha_{k}}{2}} v\right\|^2 $, where $ C $ is a positive constant.

    Lemma 8. ([40]) Assume that $ \left\{g^n \mid n \geq 0\right\} $ is a nonnegative sequence, $ \psi^0 \geq 0 $, and the nonnegative sequence $ \left\{G^n \mid n \geq 0\right\} $ satisfies

    $ G^n \leq \psi^0+\tau \sum\limits_{l = 0}^{n-1} G^l+\tau \sum\limits_{l = 0}^n g^l, \quad n \geq 0 . $

    Then it holds that

    $ G^n \leq e^{n \tau}\left(\psi^0+\tau \sum\limits_{l = 0}^n g^l\right), \quad n \geq 0. $

    Lemma 9. For any grid function $ u \in V_h^0 $, $ V_h^0 $ is defined in Section 2 for the the three-dimensional case, let $ p \leq r \leq q, \alpha \in[0, 1] $ satisfying $ \frac{1}{r} = \frac{\alpha}{p}+\frac{1-\alpha}{q} $, then

    $ \|u\|_r \leq\|u\|_p^\alpha \cdot\|u\|_q^{1-\alpha} . $

    Proof. By using Hölder inequality, we have

    $ h1h2h3M11i=1M21j=1M31k=1|uijk|r=h1h2h3M11i=1M21j=1M31k=1|uijk|αr+(1α)r(h1h2h3M11i=1M21j=1M31k=1|uijk|αrpαr)αrp(h1h2h3M11i=1M21j=1M31k=1|uijk|(1α)rq(1α)r)(1α)rq=urαpur(1α)q. $

    This completes the proof.

    Now we are ready to construct the fully-discrete numerical scheme for systems (2.1) and (2.2).

    With the help of Lemma 1 and for clarity of description, we will denote the space fractional operator under one-dimensional case firstly.

    $ δαx,+vnj=1hαjk=0w(α)kvnjk+1,δαx,vnj=1hαMjk=0w(α)kvnj+k1,δαxvnj=1/(2cos(απ/2))(δαx,+vnj+δαx,vnj), $

    where $ w_{k}^{(\alpha)} $ is given in Eq (2.7). In the multi-dimensional case, the definitions of $ \delta_{x_{k}}^{\alpha_{k}} $ are similar to it.

    For numerically solving systems (2.1)–(2.5), we propose a three-level scheme. We firstly define the following approximations.

    Let $ u^{n}_{k_{1}\cdots k_{d}} = u(\boldsymbol{x}, t_{n}) $ and $ v^{n}_{k_{1}\cdots k_{d}} = v(\boldsymbol{x}, t_{n}) $, for ease of presentation, we shall henceforth write $ u^{n}_{k_{1}\cdots k_{d}} $ for $ u^{n} $. Denote numerical solutions of $ u^{n} $ and $ v^{n} $ by $ U^{n} $ and $ V^{n} $, respectively.

    With the definition of $ G(u, v) $ in systems (2.1) and (2.2) and by using Lemma 5, then we have

    $ G(un+1,vn)G(un1,vn)un+1un1=Gu(u(x,tn),v(x,tn))+O(τ2), $ (3.1)
    $ G(un,vn+1)G(un,vn1)vn+1vn1=Gv(u(x,tn),v(x,tn))+O(τ2), $ (3.2)

    which is given in [11]. Further, using the space fractional operator which is already introduced above and second-order centered finite difference operator to approximate at node $ (\boldsymbol{x}, t_{n}) $, it holds that

    $ αδ2tunβdi=1δαixiuˉn+G(un+1,vn)G(un1,vn)un+1un1=Rn1,2nN1, $ (3.3)
    $ γδ2tvnσdi=1δαixivˉn+G(un,vn+1)G(un,vn1)vn+1vn1=Rn2,2nN1, $ (3.4)

    and

    $ u1=ϕ1(x)+τφ1(x)+τ22α[βdi=1δαixiϕ1(x)Gu(ϕ1(x),ϕ2(x))]+R11, $ (3.5)
    $ v1=ϕ2(x)+τφ2(x)+τ22γ[σdi=1δαixiϕ2(x)Gv(ϕ1(x),ϕ2(x))]+R12, $ (3.6)

    where $ R_1^n $ and $ R_2^n $ are the truncation errors.

    Let $ u(\boldsymbol{x}, t), v(\boldsymbol{x}, t) \in \mathcal{C}^{4, 4}(\Omega \times[0, T]) $. Combining Lemma 4 with Eqs (3.1) and (3.2), the truncation errors can be estimated as follows.

    $ max1nN1{Rn12,Rn22}C(τ2+h21++h2d)2, $ (3.7)

    where $ C $ is a positive constant and $ d $ means the dimension of space.

    Omitting the truncation errors in Eqs (3.3)–(3.6), we can get the three-level EP-FDM:

    $ αδ2tUnβdi=1δαixiUˉn+G(Un+1,Vn)G(Un1,Vn)Un+1Un1=0, $ (3.8)
    $ γδ2tVnσdi=1δαixiVˉn+G(Un,Vn+1)G(Un,Vn1)Vn+1Vn1=0, $ (3.9)

    and

    $ Un=Vn=0,  xΩh,  0nN, $ (3.10)
    $ U1=ϕ1(x)+τφ1(x)+τ22α[βdi=1δαixiϕ1(x)Gu(ϕ1(x),ϕ2(x))], $ (3.11)
    $ V1=ϕ2(x)+τφ2(x)+τ22γ[σdi=1δαixiϕ2(x)Gv(ϕ1(x),ϕ2(x))], $ (3.12)

    where $ U^{1} $ and $ V^{1} $ are obtained by applying Taylor expansion to expand $ u(\boldsymbol{x}, \tau) $ and $ v(\boldsymbol{x}, \tau) $ at $ (\boldsymbol{x}, 0) $, and by Eq (2.4) we know that $ U^0 = \phi_1\left(\boldsymbol{x}\right) $, $ V^0 = \phi_2\left(\boldsymbol{x}\right) $.

    For contrast, by doing explicit treatment of nonlinear terms $ \frac{\partial G}{\partial u} $ and $ \frac{\partial G}{\partial v} $, we introduce an explicit scheme as follows

    $ αδ2tUnβdi=1δαixiUˉn+Gu(Un,Vn)=0, $ (3.13)
    $ γδ2tVnσdi=1δαixiVˉn+Gv(Un,Vn)=0, $ (3.14)
    $ Un=Vn=0,  xΩh,  0nN, $ (3.15)
    $ U1=ϕ1(x)+τφ1(x)+τ22α[βdi=1δαixiϕ1(x)Gu(ϕ1(x),ϕ2(x))], $ (3.16)
    $ V1=ϕ2(x)+τφ2(x)+τ22γ[σdi=1δαixiϕ2(x)Gv(ϕ1(x),ϕ2(x))], $ (3.17)

    which will be used in Section 6 later.

    In this section, we give the energy conservation of the fully-discrete schemes (3.8)–(3.12) and boundedness of numerical solutions. Here, the lemmas given in Section 2 are applied.

    Now, we present the energy conservation of the EP-FDMs (3.8)–(3.12).

    Theorem 2. Let $ U^n, V^n\in \mathcal{V}_{h}^{0} $ be numerical solutions of the three-level FDMs (3.8)–(3.12). Then, the energy, which is defined by

    $ En=α2δtUn2+β2dk=1μtΛαk2kUn2+γ2δtVn2+σ2dk=1μtΛαk2kVn2+12hdM11k1=1Md1kd=1[G(Un+1k1kd,Vnk1kd)+G(Unk1kd,Vn+1k1kd)] $ (4.1)

    is conservative. Namely, $ E^n = E^0 $, for $ n = 1, \cdots, N-1, $ where $ \Lambda_{k}^{\frac{\alpha_{k}}{2}} $ is already introduced by Lemma 3.

    Proof. Multiplying $ h^dD_{t}U_{k_{1}\cdots k_{d}}^n $ to both sides of Eq (3.8), summing them over $ \Omega_{h} $, by using Lemma 3, we obtain

    $ α2τ(δtUn2δtUn12)+β4τdk=1(Λαk2kUn+12Λαk2kUn12)+12τhdM11k1=1Md1kd=1[G(Un+1k1kd,Vnk1kd)G(Un1k1kd,Vnk1kd)]=0, $ (4.2)

    where the second term can be reduced to

    $ \|\Lambda_{k}^{\frac{\alpha_{k}}{2}}U^{n+1}\|^2-\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}U^{n-1}\|^2 = 2\left(\mu_{t}\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}U^{n}\|^2-\mu_{t}\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}U^{n-1}\|^2\right), $

    then Eq (4.2) turned into

    $ α2τ(δtUn2δtUn12)+β2τdk=1(μtΛαk2kUn2μtΛαk2kUn12)+12τhdM11k1=1Md1kd=1[G(Un+1k1kd,Vnk1kd)G(Un1k1kd,Vnk1kd)]=0. $ (4.3)

    Similarly, multiplying $ h^dD_{t}V_{k_{1}\cdots k_{d}}^n $ to both sides of Eq (3.9), summing them over $ \Omega_{h} $, by using Lemma 3, we obtain

    $ γ2τ(δtVn2δtVn12)+σ2τdk=1(μtΛαk2kVn2μtΛαk2kVn12)+12τhdM11k1=1Md1kd=1[G(Unk1kd,Vn+1k1kd)G(Unk1kd,Vn1k1kd)]=0. $ (4.4)

    Adding up Eqs (4.3) and (4.4) yields that $ (E^n-E^{n-1})/\tau = 0 $, which infers that $ E^n = E^{n-1} $.

    By Theorem 2, we present the following estimation.

    Theorem 3. Let $ U^n, V^n\in \mathcal{V}_{h}^{0} $ be numerical solutions of the EP-FDMs (3.8)–(3.12). Then, the following estimates hold:

    $ max1nN{δtUn,δtVn,Un,Vn,Λαk2kUn,Λαk2kVn}C, $ (4.5)

    where $ C $ is a positive constant independent of $ \tau $ and $ h $ and $ 1\leq \alpha_{k}\leq 2. $ Specially, when $ \alpha_{k} = 2, $ it holds that $ |U^n|_{H_{1}}\leq C $, $ |V^n|_{H_{1}}\leq C $.

    Proof. It follows from Theorem 2, there exists a constant $ C $ such that

    $ En=α2δtUn2+β2dk=1μtΛαk2kUn2+γ2δtVn2+σ2dk=1μtΛαk2kVn2+12hdM11k1=1Md1kd=1[G(Un+1k1kd,Vnk1kd)+G(Unk1kd,Vn+1k1kd)]=E0=C, $

    then, we obtain

    $ \|\delta_{t}U^n\|\leq C, \quad \|\delta_{t}V^n\|\leq C, \quad \|\Lambda_{k}^{\frac{\alpha_{k}}{2}}U^n\|\leq C, \quad\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}V^n\|\leq C. $

    By $ \|\delta_{t}U^n\|\leq C $, we have $ \|U^{n+1}-U^n\|\leq C\tau $, then it is easy to check that

    $ \|U^n\| = \|U^0+\tau\sum\limits_{i = 0}^{n-1}\delta_{t}U^i\|\leq\|U^0\|+\tau\sum\limits_{i = 0}^{n-1}\|\delta_{t}U^i\|\leq C. $

    This completes the proof.

    In this section, the convergence analysis of the proposed scheme is given, which is based on some important lemmas presented in Section 2.

    We first give the error equations of the EP-FDMs (3.8) and (3.9). Let $ e^n = u^n-U^n $, $ \theta^n = v^n-V^n $ and for more readability we denote

    $ ε1(un+1,Un+1)=G(un+1,vn)G(un1,vn)un+1un1G(Un+1,Vn)G(Un1,Vn)Un+1Un1, $ (5.1)
    $ ε2(vn+1,Vn+1)=G(un,vn+1)G(un,vn1)vn+1vn1G(Un,Vn+1)G(Un,Vn1)Vn+1Vn1. $ (5.2)

    By deducting Eqs (3.8) and (3.9) from Eqs (3.3) and (3.4), we have

    $ αδ2tenβdi=1δαixieˉn+ε1(un+1,Un+1)=Rn1,1nN1, $ (5.3)
    $ γδ2tθnσdi=1δαixiθˉn+ε2(vn+1,Vn+1)=Rn2,1nN1, $ (5.4)
    $ en=θn=0,xΩh, 1nN1, $ (5.5)
    $ e0=θ0=0,xˉΩh, $ (5.6)
    $ e1c1τ3, xΩh, $ (5.7)
    $ θ1c2τ3, xΩh. $ (5.8)

    Before giving a proof of convergence, we provide the following estimates for Eqs (5.1)–(5.2).

    Lemma 10. On $ \bar{\Omega}_h $, we have

    $ (ε1(un+1,Un+1),Dten)C(dk=1Λαk2kθn2+dk=1Λαk2ken+12+dk=1Λαk2ken12+δten2+δten12), $ (5.9)
    $ (ε2(vn+1,Vn+1),Dtθn)C(dk=1Λαk2ken2+dk=1Λαk2kθn+12+dk=1Λαk2kθn12+δtθn2+δtθn12), $ (5.10)

    where $ C > 0 $ is a constant, independent of grid parameters $ \tau, h_1, \cdots, h_d $.

    Proof. Recalling the definition of $ G(u, v) $, we can obtain

    $ ε1(un+1,Un+1)=b12c1{[(un+1)2+(un1)2]uˉn[(Un+1)2+(Un1)2]Uˉn}+[(vn)2(uˉn)(Vn)2Uˉn]+a1c1eˉn=3k=1Qk. $

    Noting that $ U^k = u^k-e^k $ and $ V^k = v^k-\theta^k\ (k = n-1, n, n+1) $, then we get

    $ Q1=b12c1[2un+1en+1(en+1)2+2un1en1(en1)2]uˉn+b12c1[(un+1)22un+1en+1+(en+1)2+(un1)22un1en1+(en1)2]eˉn, $ (5.11)
    $ Q2=2uˉnvnθnuˉn(θn)2+(Vn)2eˉn. $ (5.12)

    When $ d = 2 $, combining Theorem 3, Lemma 6 with Lemma 7, we can get the estimation of $ \|e^m\|_{4}^{4} $, $ \|e^m\|_{6}^{6} $, $ \|e^m\|_{8}^{8} $, that is

    $ em44em2(2|em|H1+1lem)2em2[8(|um|2H1+|Um|2H1)+2l2(um2+Um2)]Cem2Cdk=1Λαk2kem2. $ (5.13)

    The same reasoning can be used to prove that

    $ em66Cdk=1Λαk2kem2, em88Cdk=1Λαk2kem2, $ (5.14)

    Smilarly, when $ d = 3 $ the results can be found in the same way.

    By using Cauchy-Schwarz inequality and the widely used inequality $ [(a+b)/2]^s\leq(a^s+b^s)/2 \ (a\geq0, b\geq0, s\geq1) $, multiplying both sides of Eq (5.11) by $ h^dD_{t}e^n $, then summing it on whole $ \Omega_h $, it follows that

    $ (Q1,Dten)b14c1[5M22(en+12+en12)+(3M+14)(en+144+en144)+12(en+166+en166)+18(en+188+en188)]+b14c1(5M2+6M+1)Dten2Cdk=1(Λαk2ken+12+Λαk2ken12)+b18c1(5M2+6M+1)(δten2+δten12), $ (5.15)

    the last inequality is derived by inequalities (5.13) and (5.14), similarly, we can also obtain

    $ (Q2,Dten)M2θn2+M2θn44+M24(en+12+en12)+(3M24+M4)Dten2Cdk=1(Λαk2kθn2+Λαk2ken+12+Λαk2ken12)+(3M24+M4)(δten2+δten12), $ (5.16)
    $ (Q3,Dten)a21C4c21dk=1(Λαk2ken+12+Λαk2ken12)+14(δten2+δten12), $ (5.17)

    combine inequalities (5.15)–(5.17), then we get inequality (5.9) is proved. We can demonstrate that inequality (5.10) is likewise true using techniques similar to inequality (5.9). This completes the proof.

    Now we further investigate the accuracy of the proposed scheme with the help of the above lemmas, see Theorem 4.

    Theorem 4. Assume that $ u(\boldsymbol{x}, t), v(\boldsymbol{x}, t) \in \mathcal{C}^{4, 4}(\Omega \times[0, T]) $ are exact solutions of systems (2.1)–(2.5), let $ u_{k_{1}\cdots k_{d}}^n = u(\boldsymbol{x}, t) $ and $ v_{k_{1}\cdots k_{d}}^n = v(\boldsymbol{x}, t) $, denote numerical solutions by $ U_{k_{1}\cdots k_{d}}^n $ and $ V_{k_{1}\cdots k_{d}}^n $, define $ e^n = u^n-U^n $, $ \theta^n = v^n-V^n (1\leq n\leq N) $. Then suppose that $ \tau $ is sufficiently small. The error estimates of the EP-FDM are

    $ dk=1Λαk2ken2C(τ2+h21++h2d)2,enC(τ2+h21++h2d),dk=1Λαk2kθn2C(τ2+h21++h2d)2,θnC(τ2+h21++h2d), $

    where $ C $ is a positive constant, independent of grid parameters $ \tau, h_1, \cdots, h_d $.

    Proof. Noting that at every time level, the systems defined in Eqs (3.8) and (3.9) is a linear PDE. Obviously, the existence and uniqueness of the solution can be obtained.

    For ease of expression, we write

    $ I^n = \alpha\|\delta_{t}e^n\|^2+\beta\sum\limits_{k = 1}^d\mu_{t}\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}e^n\|^2 +\gamma\|\delta_{t}\theta^n\|^2+\sigma\sum\limits_{k = 1}^d\mu_{t}\|\Lambda_{k}^{\frac{\alpha_{k}}{2}}\theta^n\|^2. $

    Apparently, we have that $ I^1\leq C(\tau^2+h_1^2+\cdots+h_d^2)^2 $.

    Multiplying $ h^dD_{t}e^n $ and $ h^dD_{t}\theta^n $ to both sides of Eqs (5.3) and (5.4), then summing it over the whole $ \Omega_h $ respectively. Then adding up the obtained results, it follows that

    $ InIn12τ+(ε1(un+1,Un+1),Dten)+(ε2(vn+1,Vn+1),Dtθn)=(Rn1,Dten)+(Rn2,Dtθn), $ (5.18)

    by using Cauchy-Schwarz inequality, we have

    $ InIn12τ|(ε1(un+1,Un+1),Dten)|+|(ε2(vn+1,Vn+1),Dtθn)|+12Rn12+14(δten2+δten12)+12Rn22+14(δtθn2+δtθn12), $ (5.19)

    multiplying $ 2\tau $ to both sides of inequality (5.19), and using Lemma 10, then we get

    $ InIn12Cτ(In+In1)+τRn12+τRn22. $ (5.20)

    Thus, $ \forall K (2\leq n\leq K\leq N-1) $, summing $ n $ from $ 2 $ to $ K $, we get

    $ (12Cτ)IKI1+4CτK1n=1In+Kn=2τ(Rn12+Rn22), $ (5.21)

    when $ C\tau\leq\frac{1}{3} $, inequality (5.21) is turned into

    $ IK3I1+12CτK1n=1In+3τKn=2(Rn12+Rn22), $ (5.22)

    then by using Lemma 8 and inequality (3.7), we obtain

    $ IKenτ(3I1+3τKn=2(Rn12+Rn22))C(τ2+h21++h2d)2. $ (5.23)

    By the definition of $ I $, it is easy to conclude that

    $ dk=1Λαk2ken2C(τ2+h21++h2d)2,δtenC(τ2+h21++h2d),dk=1Λαk2kθn2C(τ2+h21++h2d)2,δtθnC(τ2+h21++h2d), $

    furthermore, we have

    $ \|e^n\| = \|e^0+\tau\sum\limits_{i = 0}^{n-1}\delta_{t}e^i\|\leq\tau\sum\limits_{i = 0}^{n-1}\|\delta_{t}e^i\|\leq C(\tau^2+h_1^2+\cdots+h_d^2). $

    Similarly, $ \|\theta^n\|\leq C(\tau^2+h_1^2+\cdots+h_d^2). $ This completes the proof.

    We carry out several numerical examples to support the theoretical results in this section. All computations are performed with Matlab. Throughout the experiments, the spatial domain is divided into $ M $ parts in every direction uniformly, that is, in the 1D case, we set $ M_{1} = M $, while in the 2D case, we set $ M_{1} = M_{2} = M $, and the time interval $ [0, T] $ is also divided uniformly into $ N $ parts. Then we use the discrete $ L^\infty $-norm to measure the global error of the scheme, namely,

    $ Eu(M,N)=UNu(T),Ev(M,N)=VNv(T), $

    Example 1. Consider the following one-dimensional coupled KG model

    $ uttκ2αxu+a1u+b1u3+c1uv2=g,(x,t)Ω×[0,T],vttκ2αxv+a2v+b2v3+c2u2v=g,(x,t)Ω×[0,T], $

    with $ \Omega = [0, 1] $. The initial and boundary conditions are determined by the exact solutions

    $ u(x,t)=x4(1x)4et,v(x,t)=x5(1x)5cos(1+t), $

    as well as the source term $ g $. Here, we take $ a_1 = a_2 = 1 $, $ b_1 = -1 $, $ b_2 = -2 $, $ c_1 = 1 $, $ c_2 = 0.5 $ and $ \kappa = 1 $.

    The precision of the scheme in spatial direction is first tested by fixing $ N = 1000 $. We compute the global errors at $ T = 1 $ with different mesh sizes, and the numerical results with $ \alpha = 1.2, 1.5, 1.8 $ are listed in Table 1 and Table 2. As can be seen in the table, the proposed scheme can have second order convergence in space, which confirms the results of theoretical analysis in Theorem 4. To track the evolution of the discrete energy, we preserve the initial value condition in this case and set the source term to $ g = 0 $. Additionally, for the terminal time $ T = 50 $, we fix $ h = 0.05 $ and $ \tau = 0.05 $. The evolutionary trend image for scheme1 (3.8)–(3.12) and explicit scheme2 (3.13)–(3.17) with various $ \alpha $ are displayed in Figure 1. Then we further verify that the proposed scheme1 (3.8)–(3.12) preserves the discrete energy very well but scheme2 (3.13)–(3.17) does not.

    Table 1.  $ L^\infty $ error and spatial convergence rates of scheme1 (3.8)–(3.12) for Example 1.
    $ \alpha $=1.2 $ \alpha $=1.5 $ \alpha $=1.8
    $ M $ $ E_u(M, N) $ $ order(u) $ $ E_u(M, N) $ $ order(u) $ $ E_u(M, N) $ $ order(u) $
    32 1.51e-05 * 5.86e-06 * 4.30e-06 *
    64 3.69e-06 2.03 1.46e-06 2.01 1.09e-06 1.98
    128 9.18e-07 2.01 3.66e-07 2.00 2.74e-07 1.99
    256 2.28e-07 2.01 9.08e-08 2.01 6.75e-08 2.02

     | Show Table
    DownLoad: CSV
    Table 2.  $ L^\infty $ error and spatial convergence rates of scheme1 (3.8)–(3.12) for Example 1.
    $ \alpha $=1.2 $ \alpha $=1.5 $ \alpha $=1.8
    $ M $ $ E_v(M, N) $ $ order(v) $ $ E_v(M, N) $ $ order(v) $ $ E_v(M, N) $ $ order(v) $
    32 1.61e-06 * 3.33e-06 * 2.21e-06 *
    64 4.11e-07 1.97 8.14e-07 2.03 5.30e-07 2.06
    128 1.04e-07 1.99 2.02e-07 2.01 1.31e-07 2.01
    256 2.60e-08 2.00 5.06e-08 2.00 3.28e-08 2.00

     | Show Table
    DownLoad: CSV
    Figure 1.  The long time discrete energy of Example 1 with $ h = 0.05 $, $ \tau = 0.05 $ for scheme1 (3.8)–(3.12) and explicit scheme2 (3.13)–(3.17).

    Example 2. Consider the following two-dimensional coupled KG model

    $ uttκ2α1xuκ2α2yu+a1u+b1u3+c1uv2=g,(x,y,t)Ω×[0,T],vttκ2α1xvκ2α2yv+a2v+b2v3+c2u2v=g,(x,y,t)Ω×[0,T], $

    with $ \Omega = [0, 2] \times[0, 2] $. The initial and boundary conditions are determined by the exact solutions

    $ u(x,y,t)=x2(2x)2y2(2y)2et,v(x,y,t)=x4(2x)4y4(2y)4sin(1+t), $

    as well as the source term $ g $. Here, we take $ a_1 = a_2 = 1 $, $ b_1 = -1 $, $ b_2 = -2 $, $ c_1 = 1 $, $ c_2 = 0.5 $ and $ \kappa = 1 $.

    Similar to Example 1, we verify the convergence orders of the scheme in spatial direction at $ T = 1 $. For spatial convergence order, we still set $ N = 1000 $ and thus the temporal error of the scheme can be negligible. The numerical results are presented in Table 3 and Table 4 with different values of $ \alpha_1 $ and $ \alpha_2 $ which are in the x and y directions, respectively. The second-order accuracy of the scheme is achieved. Moreover, for the terminal time $ T = 100 $, Figure 2 shows the evolution of discrete energy for scheme1 (3.8)–(3.12) and explicit scheme2 (3.13)–(3.17) when $ g(x, y, t) = 0 $. The figure indicate that the discrete conservation law holds very well if the proposed scheme1 (3.8)–(3.12) are used. In contrast, scheme2 (3.13)–(3.17) cannot preserve the discrete energy. Both tables and figure further confirm the theoretical results.

    Table 3.  $ L^\infty $ error and spatial convergence rates of scheme1 (3.8)–(3.12) for Example 2.
    $ \alpha_{1} $=1.3, $ \alpha_{2} $=1.6 $ \alpha_1 $=1.5, $ \alpha_2 $=1.5 $ \alpha_1 $=1.7, $ \alpha_2 $=1.2
    $ M $ $ E_u(M, N) $ $ order(u) $ $ E_u(M, N) $ $ order(u) $ $ E_u(M, N) $ $ order(u) $
    8 3.76e-02 * 3.76e-02 * 3.82e-02 *
    16 9.44e-03 2.00 9.28e-03 2.02 9.63e-03 1.99
    32 2.32e-03 2.03 2.30e-03 2.01 2.37e-03 2.02
    64 5.70e-04 2.02 5.65e-04 2.03 5.85e-04 2.02

     | Show Table
    DownLoad: CSV
    Table 4.  $ L^\infty $ error and spatial convergence rates of scheme1 (3.8)–(3.12) for Example 2.
    $ \alpha_{1} $=1.3, $ \alpha_{2} $=1.6 $ \alpha_1 $=1.5, $ \alpha_2 $=1.5 $ \alpha_1 $=1.7, $ \alpha_2 $=1.2
    $ M $ $ E_v(M, N) $ $ order(v) $ $ E_v(M, N) $ $ order(v) $ $ E_v(M, N) $ $ order(v) $
    8 2.98e-01 * 3.02e-01 * 2.77e-01 *
    16 6.16e-02 2.28 6.13e-02 2.30 5.72e-02 2.28
    32 1.46e-02 2.08 1.45e-02 2.08 1.36e-02 2.08
    64 3.60e-03 2.02 3.56e-03 2.02 3.34e-03 2.02

     | Show Table
    DownLoad: CSV
    Figure 2.  The long time discrete energy of Example 2 with $ h = 0.1 $, $ \tau = 0.05 $ for scheme1 (3.8)–(3.12) and explicit scheme2 (3.13)–(3.17).

    Example 3. Consider the following two-dimensional coupled KG model

    $ uttκ2α1xuκ2α2yu+a1u+b1u3+c1uv2=0,(x,y,t)Ω×[0,T],vttκ2α1xvκ2α2yv+a2v+b2v3+c2u2v=0,(x,y,t)Ω×[0,T], $

    and

    $ (u(x,y,t),v(x,y,t))=(0,0),(x,y,t)Ω×[0,T],(u(x,y,0),v(x,y,0))=(u0(x,y),v0(x,y)),(x,y)ˉΩ,(ut(x,y,0),vt(x,y,0))=(0,0),(x,y)ˉΩ, $

    with $ \Omega = [0, 1] \times[0, 1] $.

    Here, we take

    $ u_0(x, y) = 2[1-\cos (2 \pi x)][1-\cos (2 \pi y)] \operatorname{sech}(x+y), $
    $ v_0(x, y) = 4 \sin (\pi x) \sin (\pi y) \tanh (x+y) $

    and

    $ a_1 = 10, a_2 = 4, b_1 = 6, b_2 = 5, c_1 = 2, c_2 = 3, \kappa = 1. $

    The scheme1 (3.8)–(3.12) with

    $ \tau = h = 0.05, \alpha_1 = \alpha_2 = 1.5 $

    are used to Example 3. Figure 3 and Figure 4 show the surfaces of $ U_{ij}^n $ and $ V_{ij}^n $ at different times, respectively. The significant dynamical evolutionary features of the numerical solutions $ U_{ij}^n $ and $ V_{ij}^n $, such as radiation and oscillation, can be found in Figure 3 and Figure 4.

    Figure 3.  Surfaces of $ U_{ij}^n $ at different times of Example 3 with $ \alpha_1 = \alpha_2 = 1.5 $ for scheme1 (3.8)–(3.12).
    Figure 4.  Surfaces of $ V_{ij}^n $ at different times of Example 3 with $ \alpha_1 = \alpha_2 = 1.5 $ for scheme1 (3.8)–(3.12).

    In this paper, the three-level energy-preserving scheme is proposed for the space-fractional coupled KG systems. The scheme is derived by using the finite difference method. The discrete conservation law, boundedness of numerical solutions and the global error of the scheme are further discussed. It is shown that the scheme can have second order convergence in both temporal direction and spatial direction. Several numerical examples are performed to support the theoretical results in the paper. Moreover, due to the nonlocal derivative operator and considering that the implicit methods involve Toeplitz matrices, fast methods are fairly meaningful to reduce the computational cost of the proposed scheme; refer to the recent work [41,42] for this issue.

    This work is supported by NSFC (Grant Nos. 11971010, 12001067).

    The authors declared that they have no conflicts of interest to this work.

    In the following, we present the proof of Lemma 6.

    Proof of Lemma 6: Obviously, the result holds for $ p = 2 $. We prove the conclusion for $ p > 2 $.

    For any $ m, s = 1, 2, \ldots, M_1-1 $, and $ m > s $, using mean value theorem, we have

    $ \left|u_{m j k}\right|^{\frac{p}{3}}-\left|u_{s j k}\right|^{\frac{p}{3}} = \sum\limits_{i = s}^{m-1}\left(\left|u_{i+1, j k}\right|^{\frac{p}{3}}-\left|u_{i j k}\right|^{\frac{p}{3}}\right) = \frac{p}{3} \sum\limits_{i = s}^{m-1}\left(\left|u_{i+1, j k}\right|-\left|u_{i j k}\right|\right) \xi_{i j k}^{\frac{p}{3}-1}, $

    where

    $ \xi_{i j k} \in\left(\min \left\{\left|u_{i j k}\right|, \left|u_{i+1, j k}\right|\right\}, \max \left\{\left|u_{i j k}\right|, \left|u_{i+1, j k}\right|\right\}\right) . $

    Then,

    $ |umjk|p3|usjk|p3p3m1i=s|ui+1,jkuijk|(|uijk|p31+|ui+1,jk|p31)=ph1m1i=s|δx1uijk||uijk|p31+|ui+1,jk|p312p(h1M11i=1|uijk|2p32)12(h1M11i=1|δx1uijk|2)12. $

    It is easy to verify the above inequality also holds for $ m \leq s $. Thus, we have

    $ \left|u_{m j k}\right|^{\frac{p}{3}} \leq p\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|u_{i j k}\right|^{\frac{2p}{3}-2}\right)^{\frac{1}{2}}\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|\delta_{x_1} u_{i j k}\right|^2\right)^{\frac{1}{2}}+\left|u_{s j k}\right|^{\frac{p}{3}}, \quad \forall 1 \leq m, s \leq M_1-1. $

    Multiplying the above inequality by $ h_1 $ and summing up for $ s $ from 1 to $ M_1-1 $, we have

    $ l_1\left|u_{m j k}\right|^{\frac{p}{3}} \leq l_1 p\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|u_{i j k}\right|^{\frac{2p}{3}-2}\right)^{\frac{1}{2}}\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|\delta_{x_1} u_{i j k}\right|^2\right)^{\frac{1}{2}}+h_1 \sum\limits_{i = 1}^{M_1-1}\left|u_{i j k}\right|^{\frac{p}{3}} . $

    Dividing the result by $ l_1 $, and noticing that the above inequality holds for $ m = 1, 2, \ldots, M_1-1 $, we have

    $ \max _{1 \leq m \leq M_1-1}\left|u_{m j k}\right|^{\frac{p}{3}} \leq p\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|u_{i j k}\right|^{\frac{2p}{3}-2}\right)^{\frac{1}{2}}\left(h_1 \sum\limits_{i = 1}^{M_1-1}\left|\delta_{x_1} u_{i j k}\right|^2\right)^{\frac{1}{2}}+\frac{1}{l_1} h_1 \sum\limits_{i = 1}^{M_1-1}\left|u_{i j k}\right|^{\frac{p}{3}} . $

    Multiplying the above inequality by $ h_2h_3 $ and summing over $ j, k $, then applying the Cauchy-Schwarz inequality, we obtain

    $ h2h3M21j=1M31k=1max1iM11|uijk|p3ph2h3M21j=1M31k=1(h1M11i=1|uijk|2p32)12(h1M11i=1|δx1uijk|2)12+1l1(up3)p3p(h2h3M21j=1M31k=1h1M11i=1|uijk|2p32)12(h2h3M21j=1M31k=1h1M11i=1|δx1uijk|2)12+1l1(up3)p3=p(u2p32)p31δx1u+1l1(up3)p3. $ (A.1)

    Multiply both sides of inequality (A.1) by $ (h_2h_3)^{\frac{1}{2}} $, it follows easily that there exists a constant C such that $ (h_2h_3)^{\frac{1}{2}}\leq C $, we obtain

    $ (h2h3)12M21j=1M31k=1max1iM11|uijk|p3Cp(u2p32)p31δx1u+Cl1(up3)p3. $ (A.2)

    Similarly to the previous analysis, we have

    $ (h1h3)12M11i=1M31k=1max1jM21|uijk|p3Cp(u2p32)p31δx2u+Cl2(up3)p3. $ (A.3)
    $ (h1h2)12M11i=1M21j=1max1kM31|uijk|p3Cp(u2p32)p31δx3u+Cl3(up3)p3. $ (A.4)

    Using the Cauchy-Schwarz inequality, we have

    $ \left(\|u\|_{\frac{p}{3}}\right)^{\frac{p}{3}} = h_1 h_2 h_3 \sum\limits_{i = 1}^{M_1-1} \sum\limits_{j = 1}^{M_2-1}\sum\limits_{k = 1}^{M_3-1}\left|u_{i j k}\right|^{\frac{p}{3}} = h_1 h_2 h_3\sum\limits_{i = 1}^{M_1-1} \sum\limits_{j = 1}^{M_2-1}\sum\limits_{k = 1}^{M_3-1}\left|u_{i j k}\right| \cdot\left|u_{i j k}\right|^{\frac{p}{3}-1} \leq\|u\| \cdot\left(\|u\|_{\frac{2p}{3}-2}\right)^{\frac{p}{3}-1} . $

    Substituting the above inequality into inequalities (A.2)–(A.4), we have

    $ (h2h3)12M21j=1M31k=1max1iM11|uijk|p3C(u2p32)p31(pδx1u+1l1u). $ (A.5)
    $ (h1h3)12M11i=1M31k=1max1jM21|uijk|p3C(u2p32)p31(pδx2u+1l2u). $ (A.6)
    $ (h1h2)12M11i=1M21j=1max1kM31|uijk|p3C(u2p32)p31(pδx3u+1l3u). $ (A.7)

    We now estimate $ \|u\|_p^p $,

    $ upp=h1h2h3M11i=1M21j=1M31k=1|uijk|p=(h1h2)12M11i=1M21j=1((h1h3)12(h2h3)12M31k=1|uijk|2p3|uijk|p3)(h1h2)12M11i=1M21j=1(max1kM31|uijk|p3(h1h3)12(h2h3)12M31k=1|uijk|2p3)((h1h2)12M11i=1M21j=1(max1kM31|uijk|p3))(M11i=1M21j=1(h1h3)12(h2h3)12M31k=1|uijk|2p3)((h1h2)12M11i=1M21j=1(max1kM31|uijk|p3))((h2h3)12M21j=1M31k=1(max1iM11|uijk|p3))   ((h1h3)12M11i=1M21j=1M31k=1|uijk|p3)((h1h2)12M11i=1M21j=1(max1kM31|uijk|p3))((h2h3)12M21j=1M31k=1(max1iM11|uijk|p3))   ((h1h3)12M11i=1M31k=1(max1jM21|uijk|p3))C3(u2p32)p3(pδx1u+1l1u)(pδx2u+1l2u)(pδx3u+1l3u), $ (A.8)

    the last inequality is obtained by inequalities (A.5)–(A.7).

    In addition, we set $ l = \min \left\{l_1, l_2, l_3\right\} $, by using mean value inequality then we have

    $ (pδx1u+1l1u)(pδx2u+1l2u)(pδx3u+1l3u)(pδx1u+1lu)(pδx2u+1lu)(pδx3u+1lu)p3δx1uδx2uδx3u+pl2u2(δx1u+δx2u+δx3u)+p2lu(δx1uδx3u+δx2uδx3u+δx1uδx2u)+1l3u3(p3)3(δx1u2+δx2u2+δx3u2)32+3pl2u2(δx1u2+δx2u2+δx3u2)12+p2lu(δx1u2+δx2u2+δx3u2)+1l3u3=(p3)3|u|3H1+3pl2|u|H1u2+p2l|u|2H1u+1l3u3=(p3|u|H1+1lu)3. $ (A.9)

    Combining inequalities (A.8) and (A.9) yields

    $ (up)pC3(u2p32)p3(p3|u|H1+1lu)3. $ (A.10)

    We consider the case $ p \geq 6 $, applying Lemma 9 for $ p \geq 6 $, it holds

    $ \left(\|u\|_{\frac{2p}{3}-2}\right)^{p-3} \leq\|u\|^{\frac{p+6}{p-2}}\left(\|u\|_p\right)^{\frac{p(p-6)}{p-2}} . $

    Substituting the above inequality into inequality (A.10), we get

    $ \left(\|u\|_p\right)^{\frac{4 p}{p-2}} \leq C^3\|u\|^{\frac{p+6}{p-2}} \cdot\left(\frac{p}{\sqrt{3}}|u|_{H^1}+\frac{1}{l}\|u\|\right)^3 , $

    that is

    $ upC3up+64p(p3|u|H1+1lu)3p64p. $ (A.11)

    Thus, we have proved the result for $ p \geq 6 $. Taking $ p = 6 $ in inequality (A.11) yields

    $ u6C3u12(23|u|H1+1lu)12. $ (A.12)

    When $ 2 < p < 6 $, using Lemma 9 and inequality (A.12), we have

    $ upu6p2pu3(p2)4p6C3u6p2p[u12(23|u|H1+1lu)12]3p64p=C3up+64p(23|u|H1+1lu)3p64p. $

    This completes the proof.

    [1] Exp. Brain Res., 210 (2011), 353-371.
    [2] J. Neurophysiol., 104 (2010), 2806-2820.
    [3] Adv. Appl. Probab., 19 (1987), 784-800.
    [4] Methuen & Co., Ltd., London; John Wiley & Sons, Inc., New York, 1966.
    [5] Phys. Rev. E, 71 (2005), 011907, 9 pp.
    [6] Phys. Rev. E, 79 (2009), 021905.
    [7] J. Neurosci., 26 (2006), 448-457.
    [8] Biol. Cybern., 73 (1995), 209-221.
    [9] Biol. Cybern., 77 (1997), 289-295.
    [10] in "Selected tables in mathematical statistics, Vol. III," Amer. Math. Soc., Providence, RI, (1975), 233-327.
    [11] Phys. Rev. E, 86 (2012), 051903.
    [12] Biol. Cybern., 56 (1987), 19-26.
    [13] Biol. Cybern., 99 (2008), 253-262.
    [14] Revised edition, Dover Publications, Inc., New York, 1972.
    [15] J. Theor. Biol., 232 (2005), 505-521.
    [16] J. Comput. Neurosci., 10 (2001), 25-45.
    [17] J. Neurosci., 10 (1990), 1415-1428.
    [18] J. Neurosci., 11 (1991), 72-84.
    [19] J. Neurophysiol., 54 (1985), 782-806.
    [20] J. Comput. Neurosci., 24 (2008), 179-194.
    [21] Neurocomput., 70 (2007), 1717-1722.
    [22] Neural Comput., 16 (2004), 2533-2561.
    [23] J. Comput. Neurosci., 24 (2008), 69-79.
    [24] $2^{nd}$ edition, Cambridge University Press, Cambridge, 1992.
    [25] J. Appl. Prob., 25 (1988), 43-57.
    [26] Neural Netw., 12 (1999), 1181-1190.
    [27] Neural Comput., 21 (2009), 1931-1951.
    [28] Neural Comput., 11 (1999), 935-951.
    [29] Neural Comput., 15 (2003), 965-991.
    [30] J. Neurosci., 13 (1993), 334-350.
    [31] Nature, 423 (2003), 288-293.
    [32] Nat. Neurosci., 1 (1998), 210-217.
    [33] Neural Comput., 9 (1997), 971-983.
    [34] Cambridge Studies in Mathematical Biology, No. 8, Cambridge University Press, Cambridge, 1988.
    [35] J. Theor. Biol., 257 (2009), 90-99.
    [36] Nature, 426 (2003), 442-446.
    [37] Neural Comput., 21 (2009), 3079-3105.
  • This article has been cited by:

    1. Dongdong Hu, Fully decoupled and high-order linearly implicit energy-preserving RK-GSAV methods for the coupled nonlinear wave equation, 2024, 445, 03770427, 115836, 10.1016/j.cam.2024.115836
  • Reader Comments
  • © 2014 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2457) PDF downloads(494) Cited by(12)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog