Large Time-Stepping Spectral Methods for the Semiclassical Limit of the Defocusing Nonlinear Schrödinger Equation

Abstract

We analyze a class of large time-stepping Fourier spectral methods for the semiclassical limit of the defocusing Nonlinear Schrödinger equation and provide highly stable methods which allow much larger time step than for a standard implicit-explicit approach. An extra term, which is consistent with the order of the time discretization, is added to stabilize the numerical schemes. Meanwhile, the first-order and second-order semi-implicit schemes are constructed and analyzed. Finally the numerical experiments are performed to demonstrate the effectiveness of the large time-stepping approaches.

1. Introduction

Consider the initial-value problem for the semiclassical limit of defocusing nonlinear Schrödinger equation (DFNLS) in one space dimension 𝑖𝜀𝜕𝑢+𝜀𝜕𝑡2²𝑢_𝑥𝑥−𝜆|𝑢|²𝑢𝑢=0(𝑥∈Ω,𝑡>0),₀(𝑥)=𝐴₀(𝑥)𝑒^{𝑖𝑆0(𝑥)/𝜀}, (1.1) where 𝑖²=−1, the domain Ω is an open set in ℝ, 𝑢(𝑥,𝑡) is the complex-valued wave function, 0<𝜀≪1 is the scaled Planck constant, and 𝜆≥0 is a real constant. 𝐴₀(𝑥) is given with the initial amplitude and 𝑆₀(𝑥) is the real initial phase function with 𝐴₀(𝑥) and 𝑆^′₀(𝑥) decaying rapidly as |𝑥|→∞. To avoid boundary effects and obtain enough information of statistics, we compute the solution in a sufficiently large interval [−𝑀,𝑀].

The DFNLS (1.1) has infinitely many conservation laws (see, e.g., [1–3]), two of which are

𝜕_𝑡𝜌+𝜕_𝑥𝜕𝐽=0,_𝑡𝐽+𝜕_𝑥(𝐽²𝜌1)+2𝜕_𝑥(𝜌²𝜀)=²4𝜕_𝑥(𝜌𝜕²_𝑥log𝜌), (1.2) where 𝜌 (the position density) and 𝐽 (the current density) are primary physical quantities and can be computed from the wave function 𝑢(𝑥,𝑡):

𝜌(𝑥,𝑡)=|𝑢(𝑥,𝑡)|²,𝐽(𝑥,𝑡)=𝜀𝐼𝑚(𝑢(𝑥,𝑡)𝑢_𝑥(𝑥,𝑡)), (1.3) where “−“ denotes complex conjugation. The two conservation laws can also be denoted by the 𝐿² norm and energy density conservation laws

‖𝑢(⋅,𝑡)‖²=_𝑅|𝑢(𝑥,𝑡)|²𝑑𝑥=‖𝑢₀(𝑥)‖²,𝜀𝐸(⋅,𝑡)=²2_𝑅|𝑢_𝑥(𝑥,𝑡)|²𝜆𝑑𝑥+2_𝑅|𝑢(𝑥,𝑡)|⁴𝑑𝑥, (1.4) where ‖⋅‖ is the standard 𝐿²-norm in ℝ.

The DFNLS is connected with many applications in science and technology. One of the most important applications of the semiclassical limit of (1.1) arises in nonlinear optics [4], in particular regarding the production and propagation of laser pulses in fiber optics. Here 𝜀 depends on the dimensions on the problem (amplitude of the initial pulse, dimensions of the fiber, etc.). Another application appears in the study of the so-called Bose-Einstein condensate [5, 6], where 𝜀 now does represent the Planck constant.

The limit 𝜀→0 is called the semiclassical limit and considerable attention has been given recently to the investigation of its existence and structure [7–11]. The dynamics of the limit is an open problem. Wright et al. have provided the exact solution of the geometric optics approximation of the defocusing nonlinear Schrödinger equation in [12]. Much progress has been made recently in understanding semiclassical limits of the linear Schrödinger equation. Particularly by the introduction of tools from microlocal analysis, such as defect measures [13], H-measures [14], and Wigner measures [15, 16].

Numerically this is a notoriously difficult problem. The oscillatory nature of the solutions of the nonlinear Schrödinger equation with small 𝜀 provides severe numerical burdens. Even for stable numerical approximations (or under mesh size and time step restrictions which guarantee stability), the oscillations may very well pollute the solution in such a way that the quadratic macroscopic quantities and other physical observables come out completely wrong unless the spatial-temporal oscillations are fully resolved numerically. For a long-time considerable consideration has been given to this problem by many researchers. In [17, 18], Markowich et al. utilized the Wigner measure in analyzing the semiclassical limit for the the linear Schrödinger equation with small 𝜀, to study the finite difference approximation. In [19, 20], Bao et al. simulated the solution by the time-splitting spectral approximation for the linear and nonlinear Schrödinger equation. Hector D. Ceniceros presented the semiclassical limit of the focusing nonlinear Schrödinger equation by moving mesh method [21] and Ceniceros and Tian studied it by nonlinear Fourier filtering method in [22]. But there are few research on DFNLS.

The purpose of this paper is to provide an effective numerical method to solve the DFNLS with small parameter 𝜀. In space direction, we use Fourier spectral method, and in time direction we use difference method. In the case that the space direction has spectral accuracy, we expect to improve the computational efficiency by increasing the time step, which is beneficial to the numerical simulation and the discussion on the long-time behavior of this problem. In numerical computation, for one thing, because of the strong nonlinear character of (1.1), the numerical stability cannot be completely ensured by the traditional numerical methods. For another thing, because of the oscillation of high frequency of the solution of this equation, an ideal discretization of space and time should decompose bordering surface precisely. Hence a high resolution ratio of space is needed, and very small time step must be used to satisfy the demands of computational stability if we take standard explicit or semi-implicit scheme. When 𝜀 is very small, the explicit treatment for nonlinear item 𝜆|𝑢²|𝑢 usually causes strict restriction on time step, and it is not practical to have a long-time computation in this case. To relax restriction of this kind of problem, we provide a semi-implicit difference Fourier spectral method. That is to say, to improve the stability of algorithm, we add a routine item to the semi-implicit scheme where a larger time step is allowed in the numerical experiment [23–25]. First, for the treatment of time discretization, the semi-implicit scheme is used. Second, for the second-order item in (1.1), the implicit scheme is employed to reduce the restriction on time step. Third, for the nonlinear item, the explicit scheme is used to avoid solving systems of nonlinear equations at each time step as this explicit treatment can be computed by FFT. We expect that the computation will be easy, large time step is allowed in the solution procedure, and the numerical stability can be guaranteed at the same time.

The organization of this paper is as follows. In Section 2, a theoretical analysis for the semi-implicit method is provided, and the first-order time-stepping method stability properties are investigated. The second-order semi-implicit methods are mainly investigated in Section 3. Full discretization schemes and accuracy tests are presented in Section 4. Finally, some computational results are given in Section 5.

2. Stability Analysis for First-Order Time Discretization for (1.1)

In this section, we present Fourier spectral approximations of the problem (1.1) with periodic boundary conditions.

We define periodic Sobolev space as

𝐻^𝑠(Ω)={𝑣(𝑥,𝑡)∈𝐻^𝑠_loc(Ω)∣𝑣(𝑥+𝐿,𝑡)=𝑣(𝑥,𝑡)}. (2.1)

Define inner product and norm as

(𝑢,𝑣)=_Ω𝑢𝑣𝑑𝑥,‖𝑣‖²=(𝑣,𝑣). (2.2)

For any 1≤𝑝≤∞, let 𝐿^𝑝(Ω)={𝑣∣‖𝑣‖_𝐿^𝑝<∞}, where

‖𝑣‖_𝐿^𝑝=(_Ω|𝑣|^𝑞𝑑𝑥)^1/𝑝,1≤𝑝<∞,esssup_𝑥∈Ω|𝑣(𝑥)|,𝑝=∞. (2.3)

First we consider the classical first-order semi-implicit method, the scheme is obtained by combining first-order backward difference (BD1) for 𝜕𝑢/𝜕𝑡 and a first-order extrapolation (EP1) for the nonlinear item at the same time 𝑢𝑖𝜀^𝑛+1−𝑢^𝑛+𝜀▵𝑡2²𝑢^𝑛+1_𝑥𝑥−𝜆|𝑢^𝑛|²𝑢^𝑛=0(𝑛>0), (2.4) where ▵𝑡 is the time-step and 𝑡_𝑛=𝑛▵𝑡, 𝑢^𝑛 is an approximation to 𝑢(𝑥,𝑡_𝑛).

We expect that the implicit treatment for the second-order item 𝑢_𝑥𝑥 in (2.4) can relax the restriction on time step. And we notice that it will not increase the complexity of the algorithm through the spectral method because the second-order item here is complete linear. In Section 4, it is showed that large step is not allowed in scheme (2.4) when 𝜀 is small. That is to say, to guarantee the stability, a very small time step must be used to get the precise solution. To overcome this shortage, we add an item to the scheme and have the following modified scheme (BD1/EP1):

𝑢𝑖𝜀^𝑛+1−𝑢^𝑛+𝜀▵𝑡2²𝑢^𝑛+1_𝑥𝑥−𝐴(𝑢^𝑛+1−𝑢^𝑛)−𝜆|𝑢^𝑛|²𝑢^𝑛=0(𝑛>0), (2.5) where 𝐴 is an undetermined positive constant, which will be given in the following results. The purpose of adding an item is to improve the stability of the original scheme, and a larger time step can be allowed in the computation.

As we can see, (2.5) can be denoted as the following weak form:

𝑢𝑖𝜀(^𝑛+1−𝑢^𝑛𝜀▵𝑡,𝑣)−²2(𝑢_𝑥^𝑛+1,𝑣_𝑥)−𝐴(𝑢^𝑛+1−𝑢^𝑛,𝑣)−𝜆(|𝑢^𝑛|²𝑢^𝑛,𝑣)=0,∀𝑣∈𝐻¹(Ω). (2.6)

Theorem 2.1. If 𝐴 in (2.5) is sufficiently large, the following energy inequality holds: 𝐸(𝑢^𝑛+1)≤𝐸(𝑢^𝑛), (2.7) where 𝐸(𝑢^𝑛𝜀)=²2‖𝑢^𝑛_𝑥‖²+𝜆2‖𝑢^𝑛‖⁴_𝐿⁴. (2.8) More precisely, the positive constant 𝐴 satisfies 𝜆𝐴≥2|𝑢^𝑛|²+𝜆4|𝑢^𝑛+1+𝑢^𝑛|². (2.9)

Proof. Let 𝑣=2(𝑢^𝑛+1−𝑢^𝑛) in (2.6), take the real part of the result 𝜀²Re(𝑢_𝑥^𝑛+1,𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥)+2Re(𝐴(𝑢^𝑛+1−𝑢^𝑛)+𝜆|𝑢^𝑛|²𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛)=0. (2.10)
Now we estimate the terms in (2.10), respectively,
2Re𝐴(𝑢^𝑛+1−𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛)=2𝐴‖𝑢^𝑛+1−𝑢^𝑛‖²,𝜀²Re(𝑢_𝑥^𝑛+1,𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥)=𝜀²𝑢Re(^𝑥𝑛+1+𝑢^𝑛_𝑥2+𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥2,𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥)=𝜀²2‖𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥‖²+𝜀²2(‖𝑢_𝑥^𝑛+1‖²−‖𝑢^𝑛_𝑥‖²),2𝜆Re(|𝑢^𝑛|²𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛)=2𝜆Re(|𝑢^𝑛|²,(𝑢^𝑛+1+𝑢^𝑛2−𝑢^𝑛+1−𝑢^𝑛2)(𝑢^𝑛+1−𝑢^𝑛))=𝜆(|𝑢^𝑛|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²−|𝑢^𝑛+1−𝑢^𝑛|²)=𝜆(|𝑢^𝑛+1|²+|𝑢^𝑛|²2−|𝑢^𝑛+1|²−|𝑢^𝑛|²2,|𝑢^𝑛+1|²−|𝑢^𝑛|²)−𝜆_Ω|𝑢^𝑛|²|𝑢^𝑛+1−𝑢^𝑛|²=𝜆𝑑𝑥2(‖𝑢^𝑛+1‖⁴_𝐿⁴−‖𝑢^𝑛‖⁴_𝐿⁴𝜆)−2_Ω(|𝑢^𝑛+1|²−|𝑢^𝑛|²)²𝑑𝑥−𝜆_Ω|𝑢^𝑛|²|𝑢^𝑛+1−𝑢^𝑛|²𝑑𝑥. (2.11)
From the inequality
_Ω(|𝑢^𝑛+1|²−|𝑢^𝑛|²)²𝑑𝑥≤_Ω|𝑢^𝑛+1+𝑢^𝑛|²|𝑢^𝑛+1−𝑢^𝑛|²𝑑𝑥 (2.12) and the above estimations, we obtain 𝐸(𝑢^𝑛+1𝜀)+²2‖𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥‖²+_Ω𝜆[2𝐴−2|𝑢^𝑛+1+𝑢^𝑛|²−𝜆|𝑢^𝑛|²]|𝑢^𝑛+1−𝑢^𝑛|²𝑑𝑥≤𝐸(𝑢^𝑛). (2.13)
The last term mentioned previously can be made nonnegative if the following inequality holds:
𝜆𝐴≥2|𝑢^𝑛|²+𝜆4|𝑢^𝑛+1+𝑢^𝑛|². (2.14)

Remark 2.2. From the proof of Theorem 2.1, we know that the selection of 𝐴 in (2.14) depends on solution of the equation in the (𝑛+1)th step. But essentially it is not difficult to choose a suitable 𝐴. Actually, we can determine it approximately from the following expression: 𝐴≥max_𝑥∈Ω{3𝜆2|𝑢^𝑛(𝑥)|²}. (2.15)

Remark 2.3. In numerical computation, condition (2.14) can be guaranteed by a posterior method; that is, when computation of each step finishes, we check whether the condition (2.14) is satisfied to guarantee the decrement or the conservation of discrete energy.

3. Higher-Order Methods for Semi-Implicit Time Discretization

3.1. Second-Order Scheme: BD2/EP2

Similar to the first-order scheme, Higher-Order time discretization can be constructed. For example, by combining the second-order backward difference (BD2) for 𝜕𝑢/𝜕𝑡, and the second-order extrapolation (EP2) for the explicit treatment of nonlinear item, we have the following second-order scheme:

𝑖𝜀3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1+𝜀2Δ𝑡2²𝑢^𝑛+1_𝑥𝑥−𝜆|2𝑢^𝑛−𝑢^𝑛−1|²(2𝑢^𝑛−𝑢^𝑛−1)=0, (3.1) here we add an 𝐴-item similar to the item added in the first-order scheme, and expect that this item can be consistent with the global order of the scheme and keep the scheme stable. Therefore an 𝑂(▵𝑡²𝜕_𝑡𝑡𝑢) item is added in the above scheme, and we arrive at the following modified second-order scheme (BD2/EP2):

𝑖𝜀3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1+𝜀2Δ𝑡2²𝑢^𝑛+1_𝑥𝑥−𝐴(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1)−𝜆|2𝑢^𝑛−𝑢^𝑛−1|²(2𝑢^𝑛−𝑢^𝑛−1)=0. (3.2) As usual, to start the iteration, 𝑢₀(𝑥) is given by the initial condition and 𝑢¹(𝑥) is computed by the first-order scheme (2.5).

Equation (3.2) can also be denoted by the following weak form: 𝑖𝜀(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1𝜀2Δ𝑡,𝑣)−²2(𝑢_𝑥^𝑛+1,𝑣_𝑥)−𝐴(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1,𝑣)−𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²(2𝑢^𝑛−𝑢^𝑛−1),𝑣)=0,∀𝑣∈𝐻¹(Ω), (3.3) where 𝐴 is a positive constant.

Lemma 3.1. One has Re(2𝑢^𝑛+1,3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1)=‖𝑢^𝑛+1‖²−‖𝑢^𝑛‖²+‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²+‖𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1‖². (3.4)

Proof. We have Re(2𝑢^𝑛+1,3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1)=Re(2𝑢^𝑛+1,𝑢^𝑛+1−𝑢^𝑛+(2𝑢^𝑛+1−𝑢^𝑛)−(2𝑢^𝑛−𝑢^𝑛−1))=Re(2𝑢^𝑛+1,𝑢^𝑛+1−𝑢^𝑛)+Re(2𝑢^𝑛+1,(2𝑢^𝑛+1−𝑢^𝑛)−(2𝑢^𝑛−𝑢^𝑛−1)),Re(2𝑢^𝑛+1,𝑢^𝑛+1−𝑢^𝑛)=Re(𝑢^𝑛+1+𝑢^𝑛+𝑢^𝑛+1−𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛)=‖𝑢^𝑛+1‖²−‖𝑢^𝑛‖²+Re(𝑢^𝑛+1−𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛),Re(2𝑢^𝑛+1,(2𝑢^𝑛+1−𝑢^𝑛)−(2𝑢^𝑛−𝑢^𝑛−1))=Re((2𝑢^𝑛+1−𝑢^𝑛)+(2𝑢^𝑛−𝑢^𝑛−1)−𝑢^𝑛+𝑢^𝑛−1,(2𝑢^𝑛+1−𝑢^𝑛)−(2𝑢^𝑛−𝑢^𝑛−1))=‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²+Re(−𝑢^𝑛+𝑢^𝑛−1,2𝑢^𝑛+1−3𝑢^𝑛+𝑢^𝑛−1)=‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²+Re(−𝑢^𝑛+𝑢^𝑛−1,2(𝑢^𝑛+1−𝑢^𝑛)−(𝑢^𝑛−𝑢^𝑛−1))=‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²−2Re(𝑢^𝑛−𝑢^𝑛−1,𝑢^𝑛+1−𝑢^𝑛)+Re(𝑢^𝑛−𝑢^𝑛−1,𝑢^𝑛−𝑢^𝑛−1), (3.5) then Re(2𝑢^𝑛+1,3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1)=‖𝑢^𝑛+1‖²−‖𝑢^𝑛‖²+‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²+Re(𝑢^𝑛+1−𝑢^𝑛,𝑢^𝑛+1−𝑢^𝑛)−2Re(𝑢^𝑛−𝑢^𝑛−1,𝑢^𝑛+1−𝑢^𝑛)+Re(𝑢^𝑛−𝑢^𝑛−1,𝑢^𝑛−𝑢^𝑛−1)=‖𝑢^𝑛+1‖²−‖𝑢^𝑛‖²+‖2𝑢^𝑛+1−𝑢^𝑛‖²−‖2𝑢^𝑛−𝑢^𝑛−1‖²+‖𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1‖². (3.6)

Lemma 3.2. One has Re(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1,3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1)=‖𝑢^𝑛+1−𝑢^𝑛‖²−‖𝑢^𝑛−𝑢^𝑛−1‖²+2‖𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1‖². (3.7)

Theorem 3.3. Let {𝑢^𝑛(𝑥),𝑛=1,2,…} be the solution of the semidiscrete problem (3.2). The corresponding generalized energy 𝐸^𝑛 is defined as follows: 𝐸^𝑛=𝜀²2(‖𝑢^𝑛_𝑥‖²+‖2𝑢^𝑛_𝑥−𝑢_𝑥^𝑛−1‖²)+2𝐴‖𝑢^𝑛−𝑢^𝑛−1‖²+𝜆2(‖𝑢^𝑛‖⁴_𝐿⁴+‖2𝑢^𝑛−𝑢^𝑛−1‖⁴_𝐿⁴). (3.8) Then, for large enough 𝐴, the following discrete energy inequality holds 𝐸^𝑛+1≤𝐸^𝑛. (3.9)

Proof. Let 𝑣=−2(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1) in (3.3), taking the real part of the result we have 𝜀²2Re(3𝑢_𝑥^𝑛+1−4𝑢^𝑛_𝑥+𝑢_𝑥^𝑛−1,2𝑢_𝑥^𝑛+1)+2𝐴Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1)+2𝜆Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,|2𝑢^𝑛−𝑢^𝑛−1|²(2𝑢^𝑛−𝑢^𝑛−1))=0. (3.10) Using Lemmas 3.1 and 3.2, the first term of the left of (3.10) is 𝜀²2Re(3𝑢_𝑥^𝑛+1−4𝑢^𝑛_𝑥+𝑢_𝑥^𝑛−1,2𝑢_𝑥^𝑛+1)=𝜀²2(‖𝑢_𝑥^𝑛+1‖²−‖𝑢^𝑛_𝑥‖²+‖2𝑢_𝑥^𝑛+1−𝑢^𝑛_𝑥‖²−‖2𝑢^𝑛_𝑥−𝑢_𝑥^𝑛−1‖²+‖𝑢_𝑥^𝑛+1−2𝑢^𝑛_𝑥+𝑢_𝑥^𝑛−1‖²),2𝐴Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1)=2𝐴(‖𝑢^𝑛+1−𝑢^𝑛‖²−‖𝑢^𝑛−𝑢^𝑛−1‖²+2‖𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1‖²), (3.11) the third term of the left of (3.10) is 2𝜆Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,|2𝑢^𝑛−𝑢^𝑛−1|²(2𝑢^𝑛−𝑢^𝑛−1))=𝜆Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,|2𝑢^𝑛−𝑢^𝑛−1|²2𝑢^𝑛+1)−2𝜆Re(3𝑢^𝑛+1−4𝑢^𝑛+𝑢^𝑛−1,|2𝑢^𝑛−𝑢^𝑛−1|²(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1))=𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²+|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²+|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²)−2𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|𝑢^𝑛+1−𝑢^𝑛|²−|𝑢^𝑛−𝑢^𝑛−1|²+2|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²)=𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²+|𝑢^𝑛+1|²−|𝑢^𝑛|²−3|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²)−2𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|𝑢^𝑛+1−𝑢^𝑛|²−|𝑢^𝑛−𝑢^𝑛−1|²)=𝐼₁+𝐼₂+𝐼₃+𝐼₄,𝐼₁=𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²)=𝜆2(|2𝑢^𝑛+1−𝑢^𝑛|²+|2𝑢^𝑛−𝑢^𝑛−1|²,|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²)−𝜆2(|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²,|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²)=𝜆2(‖2𝑢^𝑛+1−𝑢^𝑛‖⁴_𝐿⁴−‖2𝑢^𝑛−𝑢^𝑛−1‖⁴_𝐿⁴𝜆)−2_Ω(|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²)²𝐼𝑑𝑥,₂=𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²)=𝜆(|𝑢^𝑛+1|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²)+(|2𝑢^𝑛−𝑢^𝑛−1|²−|𝑢^𝑛+1|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²)=𝜆2(|𝑢^𝑛+1|²+|𝑢^𝑛|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²𝜆)+2(|𝑢^𝑛+1|²−|𝑢^𝑛|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²)+𝜆(|2𝑢^𝑛−𝑢^𝑛−1|²−|𝑢^𝑛+1|²,|𝑢^𝑛+1|²−|𝑢^𝑛|²)=𝜆2(‖𝑢^𝑛+1‖⁴_𝐿⁴−‖𝑢^𝑛‖⁴_𝐿⁴𝜆)+2_Ω(|𝑢^𝑛+1|²−|𝑢^𝑛|²)²𝑑𝑥+𝜆_Ω(|2𝑢^𝑛−𝑢^𝑛−1|²−|𝑢^𝑛+1|²)(|𝑢^𝑛+1|²−|𝑢^𝑛|²𝐼)𝑑𝑥,₃=−2𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²(|𝑢^𝑛+1−𝑢^𝑛|²−|𝑢^𝑛−𝑢^𝑛−1|²𝐼)𝑑𝑥,₄=−3𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²𝑑𝑥, (3.12)(3.13) from Re[−2|𝑎|²+2|𝑏|²−2|𝑎−𝑏|²]=−4Re[𝑎(𝑎−𝑏)], (3.14) we have 𝐼₃+𝐼₄=−2𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²(|𝑢^𝑛+1−𝑢^𝑛|²−|𝑢^𝑛−𝑢^𝑛−1|²)𝑑𝑥−3𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²𝑑𝑥=−4𝜆Re_Ω|2𝑢^𝑛−𝑢^𝑛−1|²(𝑢^𝑛+1−𝑢^𝑛)(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1)𝑑𝑥−𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²𝑑𝑥. (3.15)(1)(2)(3) Using the definition of 𝐸^𝑛 and from the estimation of (3.11), (3.12), then 𝐸^𝑛+1+𝜀²2‖𝑢_𝑥^𝑛+1−2𝑢^𝑛_𝑥+𝑢_𝑥^𝑛−1‖²+𝑟=𝐸^𝑛, (3.16) where 𝑟=4𝐴‖𝛿_𝑡𝑡𝑢^𝑛‖²+𝜆2_Ω(|𝑢^𝑛+1|²−|𝑢^𝑛|²)²𝜆𝑑𝑥−2_Ω(|2𝑢^𝑛+1−𝑢^𝑛|²−|2𝑢^𝑛−𝑢^𝑛−1|²)²𝑑𝑥−𝜆_Ω(|𝑢^𝑛+1|²−|2𝑢^𝑛−𝑢^𝑛−1|²)(|𝑢^𝑛+1|²−|𝑢^𝑛|²)𝑑𝑥−4𝜆Re_Ω|2𝑢^𝑛−𝑢^𝑛−1|²(𝑢^𝑛+1−𝑢^𝑛)(𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1)𝑑𝑥−𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1|²𝑑𝑥. (3.17) Obviously, (3.9) holds provided that 𝑟>0. In the following we will show that 𝑟 can be made nonnegative if 𝐴 is chosen large enough. To simplify we denote 𝑢^𝑛+1−2𝑢^𝑛+𝑢^𝑛−1 by 𝛿_𝑡𝑡𝑢^𝑛. By simple calculation and applying Young's inequality 𝑎𝑏≤𝜖𝑎²+𝑏²/4𝜖(𝜖>0), to the suitable term, we get 𝑟≥4𝐴‖𝛿_𝑡𝑡𝑢^𝑛‖²+𝜆2_Ω(|𝑢^𝑛+1|²−|𝑢^𝑛|²)²−𝜆𝑑𝑥2Re_Ω|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−𝑢^𝑛+𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥−𝜆Re_Ω(𝑢^𝑛+1+2𝑢^𝑛−𝑢^𝑛−1)(𝑢^𝑛+1+𝑢^𝑛)(𝑢^𝑛+1−𝑢)𝛿_𝑡𝑡𝑢^𝑛𝑑𝑥−4𝜆Re_Ω|2𝑢^𝑛−𝑢^𝑛−1|²(𝑢^𝑛+1−𝑢^𝑛)𝛿_𝑡𝑡𝑢^𝑛𝑑𝑥−𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥≥4𝐴‖𝛿_𝑡𝑡𝑢^𝑛‖²+𝜆2_Ω|(𝑢^𝑛+1+𝑢^𝑛)(𝑢^𝑛+1−𝑢^𝑛)|²−𝜆𝑑𝑥4_Ω|(𝑢^𝑛+1+𝑢^𝑛)(𝑢^𝑛+1−𝑢^𝑛)|²𝑑𝑥−𝜆_Ω|𝑢^𝑛+1+2𝑢^𝑛−𝑢^𝑛−1|²|𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥−𝜆𝜖_1Ω|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−𝑢^𝑛|²−𝜆𝑑𝑥4𝜖_1Ω|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²|𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥−𝜆𝜖_2Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝑢^𝑛+1−𝑢^𝑛|²−𝑑𝑥4𝜆𝜖_2Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥−𝜆_Ω|2𝑢^𝑛−𝑢^𝑛−1|²|𝛿_𝑡𝑡𝑢^𝑛|²𝑑𝑥, (3.18)(4)(5)(6)(7)(8)(9)(10)(11)(12)(13)(14) that is, 𝑟≥{4𝐴−𝜆max_𝑥∈Ω|𝑢^𝑛+1+2𝑢^𝑛−𝑢^𝑛−1|²−max_𝑥∈Ω𝜆4𝜖₁|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²−4𝜆𝜖₂max_𝑥∈Ω|2𝑢^𝑛−𝑢^𝑛−1|²−𝜆max_𝑥∈Ω|2𝑢^𝑛−𝑢^𝑛−1|²}‖𝛿_𝑡𝑡𝑢^𝑛‖²+𝜆_Ω(14|𝑢^𝑛+1+𝑢^𝑛|²−𝜖₁|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²−𝜖₂|2𝑢^𝑛−𝑢^𝑛−1|²)|𝑢^𝑛+1−𝑢^𝑛|²𝑑𝑥. (3.19)
We choose 𝜖₁,𝜖₂ such that the integral term in the above estimate is nonnegative, then constant 𝐴 holds
4𝐴≥𝜆max_𝑥∈Ω{|𝑢^𝑛+1+2𝑢^𝑛−𝑢^𝑛−1|²+14𝜖₁|2𝑢^𝑛+1+𝑢^𝑛−𝑢^𝑛−1|²+4𝜖₂|2𝑢^𝑛−𝑢^𝑛−1|²+|2𝑢^𝑛−𝑢^𝑛−1|²}. (3.20)
Under condition (3.20) on 𝐴, we have
𝐸^𝑛+1≤𝐸^𝑛. (3.21) This completes the proof of this theorem.

3.2. Third-Order Scheme: BD3/EP3

𝐴 third-order scheme for solving the semiclassical limit of the defocusing nonlinear Schrödinger equation can be constructed in a similar manner as used in the subsection. Specifically, we can obtain the BD3/EP3 scheme in the following form:

𝑖𝜀11𝑢^𝑛+1−18𝑢^𝑛+9𝑢^𝑛−1−2𝑢^𝑛−2+𝜀6Δ𝑡2²𝑢^𝑛+1_𝑥𝑥−𝐴(𝑢^𝑛+1−3𝑢^𝑛+3𝑢^𝑛−1−𝑢^𝑛−2)−𝜆|3𝑢^𝑛−3𝑢^𝑛−1+𝑢^𝑛−2|²(3𝑢^𝑛−3𝑢^𝑛−1+𝑢^𝑛−2)=0(𝑛≥2), (3.22) where, in order to start the iteration, 𝑢¹,𝑢² are calculated via a first- and second-order scheme, respectively. The stability analysis of the scheme (3.22) requires some very detailed energy estimates and will not be presented here. The numerical results obtained in the next section indicate that the third-order time discretization of type (3.22) is also stable as long as the constant 𝐴 is sufficiently large. But comparing with the lower-order scheme, the influence on the stability from 𝐴-item in the third-order scheme is not very notable.

4. Full Discretization Scheme and Accuracy Tests

A complete numerical algorithm also requires a discretization strategy in space. Since Fourier spectral method is one of the most suitable spatial approximation methods for periodic problems [26–28], here we get the space discretization of semi-implicit scheme from different time iteration schemes by this method.

Let 𝑆_𝑁=span{𝜑_𝑘𝑁(𝑥)∣−2𝑁≤𝑘≤2−1}. (4.1)

We use the following Fourier transformations:

1𝑢(𝑘,𝑡)=𝑁^𝑁−1_𝑗=0𝑢(𝑥_𝑗,𝑡)𝑒^{−𝑖𝑘𝑥𝑗}𝑁,−2𝑁≤𝑘≤2𝑢−1,_𝑁(𝑥_𝑗,𝑡)=^𝑁/2−1_𝑘=𝑁/2𝑢(𝑘,𝑡)𝑒^{𝑖𝑘𝑥𝑗},0≤𝑗≤𝑁−1. (4.2)

The expansion of (2.5) is required to satisfy the following weak form:

𝑢𝑖𝜀(^𝑁𝑛+1−𝑢^𝑛_𝑁𝜀▵𝑡,𝑣)−²2(𝑢^𝑛+1_𝑁𝑥,𝑣_𝑥)−𝐴(𝑢_𝑁^𝑛+1−𝑢^𝑛_𝑁,𝑣)−𝜆(|𝑢^𝑛_𝑁|²𝑢^𝑛_𝑁,𝑣)=0,∀𝑣∈𝑆_𝑁. (4.3)

Similarly, we have the full discretization Fourier spectral approximation of problem (3.2):

𝑖𝜀(3𝑢_𝑁^𝑛+1−4𝑢^𝑛_𝑁+𝑢_𝑁^𝑛−1𝜀2Δ𝑡,𝑣)−²2(𝑢^𝑛+1_𝑁𝑥,𝑣_𝑥)−𝐴(𝑢_𝑁^𝑛+1−2𝑢^𝑛_𝑁+𝑢_𝑁^𝑛−1,𝑣)−𝜆(|2𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1|²(2𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1),𝑣)=0,∀𝑣∈𝑆_𝑁. (4.4)

The space discetization of BD3/EP3 (3.22) has a similar form.

It deserves to mention that the corresponding spectral coefficient 𝑢^𝑛+1 can be fully solved because all implicit items about 𝑢_𝑁^𝑛+1 in (4.3) and (4.4) are linear.

Choosing test function 𝑣 as the primary function 𝑒^{−𝑖𝑘𝑥} of 𝑆_𝑁, we obtain a system of linear equations for each module 𝑘 in Fourier space:

𝑢𝑖𝜀^𝑛+1(𝑘)−𝑢^𝑛(𝑘)−𝜀▵𝑡2²𝑘²𝑢^𝑛+1(𝑘)−𝐴𝑢^𝑛+1(𝑘)=𝐴𝑢^𝑛(𝑘)+𝜆|𝑢^𝑛(𝑘)|²𝑢^𝑛(𝑘), (4.5) The corresponding second-order BD2/EP2 scheme is as follows:

3𝑢𝑖𝜀^𝑛+1(𝑘)−4𝑢^𝑛(𝑘)+𝑢^𝑛−1(𝑘)−𝜀2Δ𝑡2²𝑘²𝑢^𝑛+1(𝑘)−𝐴𝑢^𝑛+1𝑢(𝑘)=𝐴(2^𝑛(𝑘)−𝑢^𝑛−1(𝑘))+𝜆|2𝑢^𝑛(𝑘)−𝑢^𝑛−1(𝑘)|²(2𝑢^𝑛(𝑘)−𝑢^𝑛−1(𝑘))=0. (4.6) Visibly solving (4.5) and (4.6) is efficient because we can solve the system of linear equations determined by (4.5) and (4.6) through obtaining the inverse of a simple diagonal matrix. For the problem of fully discretization (4.5) and (4.6) we derive two theorems similar to the energy inequalities in Theorems 2.1 and 3.3 (its proof will be omitted here).

Theorem 4.1. Consider the numerical scheme (4.3), if 𝜆𝐴≥2|𝑢^𝑛_𝑁|²+𝜆4|𝑢_𝑁^𝑛+1+𝑢^𝑛_𝑁|², (4.7) the solution of (4.3) satisfies 𝐸(𝑢_𝑁^𝑛+1)≤𝐸(𝑢^𝑛_𝑁),∀𝑛≥0, (4.8) where 𝐸(𝑢^𝑛_𝑁𝜀)=²2‖𝑢^𝑛_𝑁𝑥‖²+𝜆2‖𝑢^𝑛_𝑁‖⁴_𝐿4. (4.9)

Theorem 4.2. If 𝜆𝐴≥4max_𝑥∈Ω{|𝑢_𝑁^𝑛+1+2𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1|²+𝑐₁|2𝑢_𝑁^𝑛+1+𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1|²+𝑐₂|2𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1|²}, (4.10) then the solution of (4.4) satisfies 𝐸(𝑢_𝑁^𝑛+1)≤𝐸(𝑢^𝑛_𝑁),∀𝑛≥0, (4.11) where 𝑐₁,𝑐₂ is positive constant and 𝐸^𝑛=𝜀²2(‖𝑢^𝑛_𝑁𝑥‖²+‖2𝑢^𝑛_𝑁𝑥−𝑢^𝑛−1_𝑁𝑥‖²)+2𝐴‖𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1‖²+𝜆2(‖𝑢^𝑛_𝑁‖⁴_𝐿4+‖2𝑢^𝑛_𝑁−𝑢_𝑁^𝑛−1‖⁴_𝐿4). (4.12)

Then we investigate the stability properties of different schemes by numerical experiments, where special attention will be paid to the improvements of stability caused by the 𝐴-item. In the computation we choose Ω=[−4,4] or Ω=[−8,8] as the domain of computation, and choose random number as the initial condition. First, we define Δ𝑡_𝑐 as the largest time step which makes the computation stable; that is, when the time step is larger than Δ𝑡_𝑐 the numerical solution will “blow up." For different 𝐴, Table 1 lines up the values of Δ𝑡_𝑐 in the Fourier spectral schemes of (2.5), (3.2), and (3.22), respectively, and the Fourier module used in the computation is 𝑁=256. From observing the data in Table 1, we draw the following conclusion.

Table 1 

Stability comparison with different 𝐴 and 𝜀 when 𝜆=𝜀.

When 𝜆=𝜀 is large, we see that the Δ𝑡_𝑐 is also large. Since all standard semi-implicit schemes are stable, small 𝐴-item is not necessarily required. When 𝜆=𝜀 is small, smaller time step is needed to satisfy the requirement of stability if we use a standard semi-implicit approach. This phenomenon is prominent especially for the case of higher-order or very small 𝜀. For the first-order scheme BD1/EP1, just as proved in Theorem 2.1, when 𝐴 is large enough the unconditional stability of the computational process can be guaranteed. For the second-order scheme, 𝐴-item improves the stability obviously. Comparing with the case of 𝐴=0, about 10 times time-step is allowed when we use 𝐴-item. For the third-order scheme, although the effect of adding 𝐴-item is not very obvious, adding the 𝐴-item scheme is still effective in improving the stability and enlarging the time step to a certain extent.

Now we turn to time accuracy comparison. Since the exact solution of (1.1) is unknown, we use numerical results of the BD2/EP2 with Δ𝑡=10⁻⁵, 𝜀=0.0025 and 𝑁=1024 as the “exact" solution. We discuss it in the following two cases.

Case 1 (weak 𝑂(𝜀) cubic defocusing nonlinearity). We set 𝜆=𝜀, the final time 𝑡=10. Table 2 shows the 𝐿²-errors obtained, at the same time we derive the rate of convergence about time by the following formulation: log(𝐸_𝑖/𝐸_𝑖+1)log(Δ𝑡_𝑖/Δ𝑡_𝑖+1), (4.13) where 𝐸_𝑖 is the 𝐿²-error when using time-step Δ𝑡_𝑖, the rate may be verified experimentally by solving a problem on a sequence of finer and finer partitions and approximation. From Table 2, we can easily calculate 𝑟(𝐵𝐷1/𝐸𝑃1)≈1, 𝑟(𝐵𝐷2/𝐸𝑃2)≈2, 𝑟(𝐵𝐷3/𝐸𝑃3)≈3, which is coincident with our theoretical conclusion.

Table 2 

Numerical errors obtained by using the BD1/EP1, BD2/EP2, and BD3/EP3 with 𝜆=𝜀=0.0025 and 𝑁=256.

Case 2 (strong 𝑂(1) cubic defocusing nonlinearity). Figure 1 shows the position density 𝜌 and the current density 𝐽 for 𝜀=0.0025, Ω=[−8,8] with different values of (𝐴,Δ𝑡)=(0,0.001),(0.1,0.01);(0,0.002),(0.1,0.02) and in different time 𝑡=0.5,1.5,10,20. It is observed that there is a good agreement between the numerical results obtained by using standard semi-implicit time-stepping method (i.e., 𝐴=0) with small Δ𝑡 and the modified method (3.2) with larger Δ𝑡.

(a)

(b)

(c)

(d)

Figure 1 

Numerical solutions for the strong O(1) defocusing nonlinearity by using ED2/EP2. 𝜆=1, 𝜀=0.0025, 𝑁=256. “−“: (𝐴,Δ𝑡)=(0,0.001),(0,0.002), “+ + +", “***": (𝐴,Δ𝑡)=(0.1,0.01),(0.1,0.02). (a) 𝑡=0.5, (b) 𝑡=1.5, (c) 𝑡=10, (d) 𝑡=20.

It is seen that once the methods are stable, the expected order of convergence (in time) is obtained, and larger time-steps can be used by adding an 𝐴-term.

5. Numerical Examples

In this section, we present the numerical results by simulating the semiclassical limit of defocusing nonlinear Schrödinger equations. The simulations are carried out in the domain Ω=[−4,4] or Ω=[−8,8], where periodic boundary conditions are used in the spatial directions. The second-order schemes (BD2/EP2), that is, (3.2) are used in our simulations. In the following computations we let 𝑁=256. The initial condition (1.1) is always chosen in the classical WKB form: 𝑢₀(𝑥,𝑡=0)=𝑢₀(𝑥)=𝐴₀(𝑥)𝑒^{𝑖𝑆0(𝑥)/𝜀}=𝜌₀(𝑥)𝑒^{𝑖𝑆0(𝑥)/𝜀}. (5.1)

The analytic solutions of the semiclassical limit are available from [29] and are used for numerical simulations.

Example 5.1 (weak 𝑂(𝜀) cubic defocusing nonlinearity). The initial condition is taken as 𝐴₀(𝑥)=𝑒^−𝑥2,𝑆₀𝑥(𝑥)=−²2+𝜀𝑒^−2𝑥21𝑙𝑛𝜀,𝑥∈ℝ. (5.2)
We choose these initial data for this example such that the weak limits of the position density 𝜌 and current density 𝐽 can be obtained analytically [29]. The weak limits 𝜌⁰, 𝐽⁰ of 𝜌, 𝐽, respectively, as 𝜀→0, are given in [29], for example, before breaking
𝜌⁰1(𝑥,𝑡)=𝑒1−𝑡^{−2𝑥2/1−𝑡},𝐽⁰𝑥(𝑥,𝑡)=−(1−𝑡)²𝑒^{−2𝑥2/1−𝑡},0≤𝑡<1. (5.3) When 𝑡→1, they are singular distributions (“𝛿-like"). Figures 2, 3, 4, and 5 show variations of the numerical results about the position density 𝜌 and current density 𝐽 for different 𝜀=0.04,0.01,0.0025,0.000625 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking) and at 𝑡=10.

(a)

(b)

(c)

Figure 2 

Numerical solutions of Example 5.1 by ED2/EP2 for the weak 𝑂(𝜀) defocusing nonlinearity (𝜀=0.04). (a) Position density 𝜌 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (b) Current density 𝐽 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (c) Evolution of the position density 𝜌 (left) and the current density 𝐽 (right).

(a)

(b)

(c)

Figure 3 

Numerical solutions of Example 5.1 by ED2/EP2 for the weak 𝑂(𝜀) defocusing nonlinearity (𝜀=0.01). (a) Position density 𝜌 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (b) Current density 𝐽 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (c) Evolution of the position density 𝜌 (left) and the current density 𝐽 (right).

(a)

(b)

(c)

Figure 4 

Numerical solutions of Example 5.1 by ED2/EP2 for the weak 𝑂(𝜀) defocusing nonlinearity (𝜀=0.0025).

(a)

(b)

Figure 5 

Numerical solutions of Example 5.1 by ED2/EP2 for the weak 𝑂(𝜀) defocusing nonlinearity (𝜀=0.000625).

Example 5.2 (weak O(𝜀^3/2) cubic defocusing nonlinearity, with 𝜆=𝜀^3/2). The initial condition is taken as 𝐴₀(𝑥)=𝑒^−𝑥2,𝑆₀𝑥(𝑥)=−²2,𝑥∈ℝ. (5.4) The semiclassical limits of the position density 𝜌 and current density 𝐽 are given analytically in [29]. For convenience, we only calculate the variations about the position density 𝜌 and current density 𝐽 for 𝜀=0.04,0.01,0.0025,0.000625 at 𝑡=0.5 (before breaking), 𝑡=1.5 (after breaking), and 𝑡=10, respectively, (Figures 6, 7, 8, and 9). The computation scheme we used is BD2/EP2 with 𝑁=256,Δ𝑡=0.0001,𝐴=0.05.

(a)

(b)

(c)

Figure 6 

Numerical solutions of Example 5.2 by ED2/EP2 for the weak 𝑂(𝜀3/2) defocusing nonlinearity (𝜀=0.04). (a) Position density 𝜌 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (b) Current density 𝐽 at 𝑡=0.5 (before breaking), at 𝑡=1.5 (after breaking), at 𝑡=10. (c) Evolution of the position density 𝜌 (left) and the current density 𝐽 (right).

(a)

(b)

(c)

Figure 7 

Numerical solutions of Example 5.2 by ED2/EP2 for the weak 𝑂(𝜀3/2) defocusing nonlinearity (𝜀=0.01).

(a)

(b)

Figure 8 

Numerical solutions of Example 5.2 by ED2/EP2 for the weak 𝑂(𝜀3/2) defocusing nonlinearity (𝜀=0.0025).

(a)

(b)

Figure 9 

Numerical solutions of Example 5.2 by ED2/EP2 for the weak 𝑂(𝜀3/2) defocusing nonlinearity (𝜀=0.000625).

Example 5.3 (strong O(1) cubic defocusing nonlinearity, 𝜆=1). The initial condition is taken as 𝐴₀(𝑥)=𝑆1−|𝑥|,|𝑥|<1,0,otherwise,⁰(𝑥)=−𝑙𝑛(𝑒^𝑥+𝑒^−𝑥),𝑥∈ℝ. (5.5) These initial data are not analytic at 𝑥=0 and 𝑥=±1. Resolving the solution for nonanalytic conditions is a challenging task. For numerical study of NLSs with cubic defocusing nonlinearity and analytic initial data, we refer to [17, 18, 20, 30]. To test the numerical method, for each fixed 𝜀=0.04,0.01,0.0025,0.000625 (Figures 10, 11, 12, and 13), we compute the numerical solution with 𝑁=512,Δ𝑡=0.0001,𝐴=0.005.

(a)

(b)

Figure 10 

Numerical solutions of Example 5.3 by ED2/EP2 for the strong 𝑂(1) cubic defocusing nonlinearity (𝜆=1),𝜀=0.04.

(a)

(b)

(c)

Figure 11 

Numerical solutions of Example 5.3 by ED2/EP2 for the strong 𝑂(1) cubic defocusing nonlinearity (𝜆=1),𝜀=0.01.

(a)

(b)

Figure 12 

Numerical solutions of Example 5.3 by ED2/EP2 for the strong 𝑂(1) cubic defocusing nonlinearity (𝜆=1),𝜀=0.0025.

(a)

(b)

Figure 13 

Numerical solutions of Example 5.3 by ED2/EP2 for the strong 𝑂(1) cubic defocusing nonlinearity (𝜆=1),𝜀=0.000625.

6. Conclusions

The semiclassical limit of the defocusing nonlinear Schrödinger equation presents a great computational challenge. Not only is the numerical method required to resolve the solution accurately, but also required to have high resolution ratio in temporal and spatial directions due to its high oscillation of solutions. In this work, a large time-stepping spectral approximation for the semiclassical limit of the defocusing nonlinear schrödinger equation is studied. It is demonstrated that the classical semi-implicit method can be further improved by simply adding a linear term consistent with the truncation errors in time direction. This treatment can be used to increase the time-step size and computational stability. Our numerical results show that: this method is very effective in capturing oscillatory solutions of the NLS for small 𝜀. We believe that the method presented here is a tool of great value and can be used to learn more about the structure of the solutions of limiting behavior for the semiclassical limit of the defocusing nonlinear schrödinger equation and some other numerical simulations.

Acknowledgments

The research is supported by Fujian National Science Foundation of China Grant 2008J0198. The author would like to thank Professor Chuanju Xu for insightful discussions regarding the nonlinear Schrödinger equation and Professor Yinnian He for helpful suggestions and comments.