# Analysis and application of the discontinuous Galerkin method to the RLW equation

- 2.1k Downloads
- 1 Citations

**Part of the following topical collections:**

## Abstract

In this work, our main purpose is to develop of a sufficiently robust, accurate and efficient numerical scheme for the solution of the regularized long wave (RLW) equation, an important partial differential equation with quadratic nonlinearity, describing a large number of physical phenomena. The crucial idea is based on the discretization of the RLW equation with the aid of a combination of the discontinuous Galerkin method for the space semi-discretization and the backward difference formula for the time discretization. Furthermore, a suitable linearization preserves a linear algebraic problem at each time level. We present error analysis of the proposed scheme for the case of nonsymmetric discretization of the dispersive term. The appended numerical experiments confirm theoretical results and investigate the conservative properties of the RLW equation related to mass, momentum and energy. Both procedures illustrate the potency of the scheme consequently.

**PACS Codes:**02.70.Dh, 02.60 Cb, 02.60.Lj, 03.65.Pm, 02.30.-f.

**MSC:**65M60, 65M15, 65M12, 65L06, 35Q53.

## Keywords

discontinuous Galerkin method regularized long wave equation backward Euler method linearization semi-implicit scheme*a priori*error estimates solitary and periodic wave solutions experimental order of convergence

## 1 Introduction

We are concerned with a proposal of a sufficiently robust, accurate and efficient numerical method for the solution of scalar nonlinear partial differential equations. As a model problem, we consider a regularized long wave (RLW) equation firstly introduced by Peregrine (in [1]) to provide an alternative description of nonlinear dispersive waves to the Korteweg-de Vries (KdV) equation. As a consequence of this, the RLW can be observed as a special class of a family of KdV equations.

The RLW equation contains a quadratic nonlinearity and exhibits pulse-like solitary wave solutions or periodic waves; see [2]. It governs various physical phenomena in disciplines such as nonlinear transverse waves in shallow water, ion-acoustic waves in plasma or magnetohydrodynamics waves in plasma. Since the RLW equation can be solved by analytical means in special cases, the proposed numerical methods can be easily verified. Several numerical studies of the RLW equation and its modified variant have been introduced in the literature, from finite difference methods [3], over collocation methods [4, 5], to finite element approaches [6, 7], or Galerkin methods [8], and in references cited therein.

In this paper, we present a semi-implicit scheme for the numerical solution of the RLW equation based on an alternative approach to the commonly used methods. The discontinuous Galerkin (DG) methods have become a very popular numerical technique for the solution of nonlinear problems. DG space semi-discretization uses higher-order piecewise polynomial discontinuous approximation on arbitrary meshes; for a survey, see [9, 10] and [11]. Among several variants of DG methods, we deal with the nonsymmetric variant interior penalty Galerkin discretizations; see [12]. The discretization in time coordinate is performed with the aid of linearization and the backward Euler method, sidetracking the time step restriction well known from the explicit schemes, proposed in [13]. Consequently, we extend the results from [13], and the attention is paid to the *a priori* error analysis of the method with the aid of standard techniques introduced in [14] and [15].

The rest of the paper is organized as follows. The problem formulation and its variational reformulation are given in Section 2. Discretization, including space semi-discretization and fully time space discretization, is considered in Section 3. The Section 4 is devoted to *a priori* error analysis. Finally, in Section 5, the theoretical results are illustrated by numerical tests on a propagation of a single solitary wave and experimental orders of convergence are computed for piecewise linear approximations together with invariant quantities of the RLW equation.

## 2 Regularized long wave equation

*i.e.*,

The whole system (1)-(4) was found to have single solitary or periodic traveling wave solutions; for details, see [2].

**Remark 1** In the case of a single solitary wave propagation, the homogeneous Dirichlet boundary conditions in (2) arise from the asymptotic behavior of the exact solution *u*, and the endpoints *a* and *b* are chosen large enough so that the boundaries do not affect the single solitary wave during its propagation up to final time *T*.

*X*and the spaces ${C}^{k}([0,T];X)$ of

*k*-times continuously differentiable mappings of the interval [0, T] with values in

*X*. By ${H}_{0}^{1}(\mathrm{\Omega})$ we denote the subspace of all functions $v\in {H}^{1}(\mathrm{\Omega})$ satisfying $v(a)=v(b)=0$. To this end, we use the following notation for a scalar product in ${L}^{2}(\mathrm{\Omega})$ by

for a norm in ${L}^{2}(\mathrm{\Omega})$ by ${\parallel \cdot \parallel}_{2}=\sqrt{(\cdot ,\cdot )}$, for a seminorm in ${H}^{1}(\mathrm{\Omega})$ by ${|\cdot |}_{1,2}=\sqrt{(\mathrm{\nabla}\cdot ,\mathrm{\nabla}\cdot )}$ and for a norm in ${H}^{1}(\mathrm{\Omega})$ by ${\parallel \cdot \parallel}_{1,2}=\sqrt{(\cdot ,\cdot )+(\mathrm{\nabla}\cdot ,\mathrm{\nabla}\cdot )}$. It is a known fact that ${|\cdot |}_{1,2}$ is a norm on ${H}_{0}^{1}(\mathrm{\Omega})$ equivalent to ${\parallel \cdot \parallel}_{1,2}$

*classical solution*. Now, we are ready to introduce the concept of

*weak formulation*. Firstly, we recall the definition of a bilinear dispersion form $a(\cdot ,\cdot )$ and a nonlinear convection form ${b}^{\epsilon}(\cdot ,\cdot )$ from [13],

*i.e.*,

where symbol $u(t)$ stands for the function on Ω such that $u(t)(x)$, $x\in \mathrm{\Omega}$ and function $f(u)$ in (7) represents the physical flux.

**Definition 1**We say that

*u*is a weak solution of problem (1)-(3) if $u\in {L}^{2}(0,T;{H}^{1}(\mathrm{\Omega}))\cap {L}^{\mathrm{\infty}}({Q}_{T})$ such that $\frac{\partial u}{\partial t}\in {L}^{2}(0,T;{H}^{1}(\mathrm{\Omega}))$ and the following conditions are satisfied:

**Remark 2** In order to unify the definition of the weak solution (8), we consider nonhomogeneous Dirichlet boundary conditions instead of the second parallel analysis of periodic-type solutions with the aid of Sobolev spaces of periodic functions ${H}_{\lambda}^{k}((a,b))$ with period $\lambda >0$ satisfying $mod(b-a,\lambda )=0$.

Further, to carry out the error analysis later, we need to specify additional assumptions on the regularity of a solution of continuous problem (1)-(3). Therefore, we assume that the weak solution *u* is sufficiently regular, namely

**Assumptions (R)**

## 3 Discretization

Let ${\mathcal{T}}_{h}$ ($h>0$) be a family of the partitions of the closure $\overline{\mathrm{\Omega}}=[a,b]$ of the domain Ω into *N* closed mutually disjoint subintervals ${I}_{k}=[{x}_{k-1},{x}_{k}]$ with length ${h}_{k}:={x}_{k}-{x}_{k-1}$ and the symbol $\mathcal{J}$ stands for an index set $\{1,\dots ,N\}$. Then we call ${\mathcal{T}}_{h}=\{{I}_{k},k\in \mathcal{J}\}$ a partition with a spatial step $h:={max}_{k\in \mathcal{J}}({h}_{k})$ and interval ${I}_{k}$ an element. By ${\mathcal{E}}_{h}$ we denote the set of all partition nodes of Ω, *i.e.*, ${\mathcal{E}}_{h}=\{{x}_{0}=a,{x}_{1},\dots ,{x}_{N-1},{x}_{N}=b\}$. Further, we label by ${\mathcal{E}}_{h}^{I}$ the set of all inner nodes. Obviously, ${\mathcal{E}}_{h}={\mathcal{E}}_{h}^{I}\cup \{a,b\}$.

We additionally assume that the partitions satisfy the following condition.

**Assumption (M)**${\mathcal{T}}_{h}$

*are locally quasi*-

*uniform*:

The condition (12) in fact allows to control a level of the mesh refinement if adapted meshes are used.

where ${\parallel \cdot \parallel}_{{H}^{{s}_{k}}({I}_{k})}$ and ${|\cdot |}_{{H}^{{s}_{k}}({I}_{k})}$ denote the standard norm and the seminorm on the Sobolev space ${H}^{{s}_{k}}({I}_{k})$, ${I}_{k}\in {\mathcal{T}}_{h}$.

where ${P}_{{p}_{k}}({I}_{k})$ denotes the space of all polynomials of degree $\le {p}_{k}$ on ${I}_{k}$, ${I}_{k}\in {\mathcal{T}}_{h}$.

*v*at inner points ${x}_{k}\in {\mathcal{E}}_{h}^{I}$ of the domain Ω by

By convention, we also extend the definition of jump and mean value for endpoints of Ω, *i.e.*, $[v({x}_{0})]=-v({x}_{0}^{+})$, $\u3008v({x}_{0})\u3009=v({x}_{0}^{+})$, $[v({x}_{N})]=v({x}_{N}^{-})$ and $\u3008v({x}_{N})\u3009=v({x}_{N}^{-})$. In case that ${x}_{k}\in {\mathcal{E}}_{h}$ are arguments of $v({x}_{k}^{-})$ or $v({x}_{k}^{+})$, we usually omit these arguments ${x}_{k}^{-}$, ${x}_{k}^{+}$ and write simply ${v}^{-}$ and ${v}^{+}$, respectively.

### 3.1 DG semi-discrete formulation

*i.e.*,

*i.e.*,

*i.e.*, outer normal is equal to one):

where $A={f}^{\prime}(\frac{u({x}^{-})+u({x}^{+})}{2})$ and the choice of $f(u({x}_{0}^{-}))$ and $f(u({x}_{N}^{+}))$ for boundary points has to satisfy the prescribed Dirichlet boundary conditions; for more details, see [14].

In what follows, we shall assume that the numerical flux $H:{\mathbb{R}}^{2}\to \mathbb{R}$ has the following properties.

**Assumptions (H)**

*is Lipschitz*-

*continuous with respect to*

*u*,

*v*:

*is consistent*:

*is conservative*:

One can see that the numerical flux *H* given by (21) satisfies conditions (H2) and (H3) and is Lipschitz-continuous on any bounded subset of ${\mathbb{R}}^{2}$.

A particular attention should be also paid to the treatment of the dispersion terms, which include an artificially added stabilization in the form ${\sum}_{x\in {\mathcal{E}}_{h}}\u3008{v}^{\prime}\u3009[u(t)]$, in order to guarantee the stability of the numerical scheme. In our case, where this stabilization is added with a positive sign (+), we speak of the nonsymmetric interior penalty Galerkin method.

which replaces the inter-element discontinuities and guarantees the fulfillment of the prescribed boundary conditions.

which is bilinear due to (19) and (25). Consequently, we can here define the semi-discrete solution ${u}_{h}$ of problem (8).

**Definition 2**We say that ${u}_{h}$ is a semidiscrete solution of problem (8) if ${u}_{h}\in {C}^{1}(0,T;{S}_{h\mathsf{p}})$ and the following conditions are satisfied:

### 3.2 Semi-implicit linearized DG scheme

In order to obtain the discrete solution, it is necessary to equip the scheme (28) with suitable solvers for the time integration. In [13], we have proposed a semi-implicit time discretization based on the backward Euler scheme with the linearized convection form ${b}_{h}^{\epsilon}$ which is suitable for avoiding the strong time step restriction of explicit time schemes as well as for the preservation of linear algebraic problems at each time level.

*f*is treated in spirit of [13] as

where ${H}_{L}(\cdot ,\cdot ,\cdot ,\cdot )$ and ${H}_{N}(\cdot ,\cdot )$ represent the corresponding linearized and nonlinear parts of the original numerical flux $H(\cdot ,\cdot )$ given by (21); for more details, see [13]. One can easily observe that the form ${b}_{hL}^{\epsilon}(\cdot ,\cdot ,\cdot )$ is linear in its second and third argument and the form ${b}_{hN}^{\epsilon}(\cdot ,\cdot )$ is in fact an original convection form (20) with half the amount of the physical flux but from the previous time level.

The fully discrete solution of problem (18) via the aforementioned semi-implicit approach is defined in following way.

**Definition 3**Let $0={t}_{0}<{t}_{1}<\cdots <{t}_{r}=T$ be a partition of the time interval $[0,T]$ and ${\tau}_{l}\equiv {t}_{l+1}-{t}_{l}$, $l=0,1,\dots ,M$. We define the approximate solution of problem (8) as functions ${u}_{h}^{l}\in {S}_{h\mathsf{p}}$, $t\in [0,T]$, $l=0,\dots ,M$, satisfying the following conditions:

Discrete problem (33) is equivalent to a system of linear algebraic equations at each time instant ${t}_{l}\in [0,T]$. In what follows, we shall be concerned with the analysis of method (33).

**Lemma 1** *Discrete problem* (33) *has a unique solution*.

*Proof*We rewrite problem (33) in the following way. For ${u}_{h}^{l}\in {S}_{h\mathsf{p}}$, ${\tau}_{l}$ and ${t}_{l+1}\in [0,T]$, we find ${u}_{h}^{l+1}$ such that

*i.e.*,

Hence, equation (34) has a unique solution ${u}_{h}^{l+1}\in {S}_{h\mathsf{p}}$. □

## 4 *A priori* error analysis

For error analysis and in experiments, we consider ${p}_{k}=p$ for all $k\in \mathcal{J}$. Thus we denote ${S}_{hp}={S}_{h\mathsf{p}}$. Now we would like to analyze the error estimates of the approximate solution ${u}_{h}^{l}$, $l=0,1,\dots $ , obtained by method (33). For simplicity, we consider a uniform partition ${t}_{l}=l\tau $, $l=0,1,\dots ,M$, of the time interval $[0,T]$ with time step $\tau =T/M$, where $M>1$ is an integer.

*cf.*[18]) for all $v\in {H}^{p+1}({I}_{k})$, ${I}_{k}\in {\mathcal{T}}_{h}$,

*h*and

*v*. We set

*τ*, subtracting from (42) and using again the linearity of the form ${\mathcal{A}}_{h}^{\mu}$, we get

For next estimates, we use the following lemmas.

**Lemma 2**

*Under assumptions*(R)

*for*${t}_{l},{t}_{l+1}\in [0,T]$,

*the following hold*:

*where* *c* *is a generic constant independent of* *h* *and* *τ*.

*Proof* The proof of these standard estimates can be found, for instance, in [15]. □

**Lemma 3**

*Under assumptions*(R), (H)

*and for*${t}_{l},{t}_{l+1}\in [0,T]$,

*the following hold*:

*where* *c* *is a generic constant independent of* *h* *and* *τ*.

*Proof* Again, one can find the proof of these estimates in [15]. □

In order to finish our estimates, we require a fulfillment of the following technical assumption.

**Assumption (T)**

(T1) *There exists* $\theta \in (0,1)$ *such that* $0<\tau <\theta /{C}_{L}$.

If assumption (T) is fulfilled, then $\tau <\frac{1}{{C}_{L}}\le \frac{\mu}{\mu {\nu}_{12}+2/{\nu}_{7}}\le \frac{{\nu}_{7}}{2}\mu $. Thus we can also reformulate assumption (T) so that $\tau =\mathcal{O}(\mu )$.

*i.e.*,

where we used a straightforward estimate ${\parallel {\xi}_{h}^{0}\parallel}_{\mu}^{2}\le \tilde{C}(\mu {h}^{2p}+{h}^{2(p+1)})$. We can notice that due to the presence of the factor *μ* in front of the function $q(\tau ,h,\mu )$ on the left-hand side of (65), we lost *μ* in denominators of $q(\tau ,h,\mu )$.

Now we are ready to formulate the main theorem.

**Theorem 1**

*Let assumptions*(M), (H), (R)

*and*(T)

*be satisfied*,

*then there exists a constant*$C=C(\mu )$

*such that*

*where* $\u2980\cdot \u2980$ *is defined by* (60).

*Proof*Since ${\parallel {e}_{h}^{l}\parallel}_{\mu}\le {\parallel {\xi}_{h}^{l}\parallel}_{\mu}+{\parallel {\eta}_{h}^{l}\parallel}_{\mu}$, the statement of the theorem comes from (65) and the fact that ${\parallel {\eta}_{h}^{l}\parallel}_{\mu}\le \tilde{c}({h}^{p+1}+\mu {h}^{p})$. Then we set

□

**Remark 3** Theorem 1 implies that the error of our method is $\mathcal{O}({h}^{p}+\tau )$ in both energy and ${L}^{2}$-norm. However, as we will see in the next section, the error estimate in the ${L}^{2}$-norm is suboptimal with respect to *h*.

**Remark 4** The dependency *C* on *μ* in the expression (68) (choice of *θ* depends on *μ*) can be removed by applying the so-called continuous mathematical induction mentioned in [19]. This is useful namely in the cases when convection terms dominate, *i.e.*, $\mu \to 0+$. Consequently, in these cases assumption (T) can be weakened to a CFL-like condition $\tau =\mathcal{O}({h}^{\alpha})$ for suitable $\alpha >0$.

## 5 Numerical experiments

In this section we shall numerically verify the theoretical *a priori* error estimates of the proposed semi-implicit method (33) for the cases of propagation of both a single solitary wave and periodic waves.

*τ*and mesh size

*h*. The computational errors are evaluated at certain time instants $t=l\tau $ during all computations in the corresponding norms,

*i.e.*,

*lτ*and ${u}_{h}^{l}$ is the numerical solution at time level

*lτ*obtained by the semi-implicit scheme (33) with constant time step

*τ*on the uniform grid with mesh size

*h*. We suppose that errors behave according to the formula

*τ*and $\stackrel{\u02c6}{{D}_{n}}$, $n=0,1$, are independent of

*h*. The values ${a}_{n}$, ${b}_{n}$, $n=0,1$, are the orders of accuracy of the method in the corresponding considered norms. We define the experimental order of convergence (EOC) by

### 5.1 Single solitary case

which represents a single solitary wave of amplitude 3*c*, traveling with the velocity $v=1+\epsilon c$ in a positive *x*-direction and located initially at the point $\overline{x}$. The initial condition is extracted from the exact solution (74) and homogeneous Dirichlet boundary conditions are set.

*T*for a piecewise linear approximation with time step $\tau =0.001$ and mesh size $h=0.05$. The similar plots are also obtained for another combination of

*τ*and

*h*as we consider below.

#### 5.1.1 Convergence with respect to *h*

First, we investigate the convergence of the method with respect to *h*. In order to restrain the discretization errors with respect to time step *τ*, we use a sufficiently small time step $\tau ={10}^{-3}$. Numerical experiments are carried out with the use of piecewise linear (${P}^{1}$) approximations on five consecutive uniformly refined meshes having 125, 250, 500, 1,000 and 2,000 elements.

**Single solitary case: Computational errors in the** ${\mathit{L}}^{\mathbf{2}}$ **-norm and experimental orders of convergence for** ${\mathit{P}}^{\mathbf{1}}$ **approximation on a consequence of meshes at time instances** **t****(** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{5.0}$ | $\mathit{t}\mathbf{=}\mathbf{10.0}$ | $\mathit{t}\mathbf{=}\mathbf{15.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||||
---|---|---|---|---|---|---|---|---|

${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | |

0.80 | 2.643E-03 | - | 3.507E-03 | - | 4.376E-03 | - | 5.255E-03 | - |

0.40 | 6.732E-04 | 1.973 | 8.762E-04 | 2.001 | 1.099E-03 | 1.993 | 1.298E-03 | 2.017 |

0.20 | 1.634E-04 | 2.043 | 2.232E-04 | 1.973 | 2.605E-04 | 2.077 | 3.284E-04 | 1.983 |

0.10 | 4.061E-05 | 2.009 | 5.763E-05 | 1.953 | 6.561E-05 | 1.989 | 8.061E-05 | 2.026 |

0.05 | 1.004E-05 | 2.016 | 1.355E-05 | 2.089 | 1.638E-05 | 2.002 | 2.107E-05 | 1.936 |

**Single solitary case: Computational errors in the energy norm and experimental orders of convergence for** ${\mathit{P}}^{\mathbf{1}}$ **approximation on a consequence of meshes at time instances** **t****(** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{5.0}$ | $\mathit{t}\mathbf{=}\mathbf{10.0}$ | $\mathit{t}\mathbf{=}\mathbf{15.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||||
---|---|---|---|---|---|---|---|---|

${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | |

0.80 | 7.082E-03 | - | 7.096E-03 | - | 7.108E-03 | - | 7.123E-03 | - |

0.40 | 3.540E-03 | 1.000 | 3.550E-03 | 0.999 | 3.603E-03 | 0.980 | 3.673E-03 | 0.956 |

0.20 | 1.778E-03 | 0.993 | 1.791E-03 | 0.987 | 1.810E-03 | 0.993 | 1.891E-03 | 0.958 |

0.10 | 8.866E-04 | 1.004 | 8.943E-04 | 1.002 | 9.049E-04 | 1.000 | 9.099E-04 | 1.055 |

0.05 | 4.441E-04 | 0.997 | 4.492E-04 | 0.993 | 4.533E-04 | 0.997 | 4.573E-04 | 0.993 |

Further, the results for EOC in the energy norm are in a quite good agreement with derived theoretical estimates; in other words, this technique produces an optimal order of convergence $\mathcal{O}({h}^{p})$. Finally, both estimates in Theorem 1 confirm the well-know attribute of DG schemes from the class of convection-diffusion problems, *cf.* [14] and [15].

#### 5.1.2 Convergence with respect to *τ*

Secondly, we verify experimentally the convergence of the method in the ${L}^{2}$-norm and the energy norm with respect to time step *τ*. In order to restrain the discretization errors with respect to *h*, we use a fine mesh with 2,000 elements with piecewise linear approximation.

*τ*, see Table 3. The computational error is evaluated at final time $t=T$ in the ${L}^{2}$-norm and the energy norm, respectively. We observe that both computational errors have EOC of order $\mathcal{O}(\tau )$ in the corresponding norms, which is again in a good agreement with derived theoretical results.

**Single solitary case: Computational errors in the** ${\mathit{L}}^{\mathbf{2}}$ **-norm and the energy norm for** ${\mathit{P}}^{\mathbf{1}}$ **approximation with respect to time step (** $\mathit{h}\mathbf{=}\mathbf{0.05}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{20.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||
---|---|---|---|---|

${\mathbf{err}}_{\mathit{\tau}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{\tau}}^{\mathbf{1}}$ | EOC | |

0.2000 | 5.852E-02 | - | 8.647E-03 | - |

0.1000 | 2.977E-02 | 0.975 | 4.384E-03 | 1.008 |

0.0500 | 1.445E-02 | 1.046 | 2.160E-03 | 0.994 |

0.0250 | 7.228E-03 | 0.996 | 1.097E-03 | 0.977 |

0.0125 | 3.601E-03 | 1.005 | 5.300E-04 | 1.049 |

#### 5.1.3 Invariant conservation quantities

assessing the accuracy of the method by measuring the difference between the numerical and analytic solutions ${u}_{h}$ and *u*, respectively.

**Single solitary case: Computed invariant quantities and errors in the** ${\mathit{l}}^{\mathbf{\infty}}$ **-norm and the** ${\mathit{L}}^{\mathbf{2}}$ **-norm (** $\mathit{h}\mathbf{=}\mathbf{0.05}$ **,** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

Method | Time | ${\mathbf{err}}^{\mathbf{\infty}}$ | ${\mathbf{err}}^{\mathbf{0}}$ | ${\mathit{I}}_{\mathit{M}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ | ${\mathit{I}}_{\mathit{P}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ | ${\mathit{I}}_{\mathit{E}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ |
---|---|---|---|---|---|---|

present method | 0.0 | - | - | 3.9799 | 0.8105 | 2.5790 |

5.0 | 2.325E-05 | 1.004E-05 | 3.9799 | 0.8104 | 2.5787 | |

10.0 | 4.518E-05 | 1.355E-05 | 3.9799 | 0.8103 | 2.5785 | |

15.0 | 6.901E-05 | 1.638E-05 | 3.9799 | 0.8101 | 2.5782 | |

20.0 | 9.322E-05 | 2.107E-05 | 3.9799 | 0.8100 | 2.5780 | |

ref. meth. [6] ( | 20.0 | 6.843E-04 | 1.757E-03 | 3.9800 | 0.8104 | 2.5792 |

ref. meth. [7] ( | 20.0 | 1.501E-03 | 1.480E-05 | 3.96467 | 0.80462 | 2.56972 |

ref. meth. [20] ( | 20.0 | 6.660E-05 | 1.820E-04 | 3.97992 | 0.81046 | 2.57901 |

ref. meth. [13] ( | 20.0 | 3.960E-03 | 9.092E-03 | 3.9800 | 0.8105 | 2.5791 |

ref. meth. [5] ( | 20.0 | 7.805E-05 | 2.069E-04 | 3.97988 | 0.81046 | 2.57901 |

analytical val. ( | - | - | - | 3.979949 | 0.810462 | 2.579007 |

### 5.2 Periodic case

*cf.*[20])

*v*in a positive

*x*-direction. The spatial period ${\omega}_{k}$ and time period ${T}_{k}$ for each wave are defined by

where $K(k)$ is a complete elliptic integral of the first kind, see [21]. The limit $k\to 1$ implies that the periodic behavior reduces to the propagation of a single solitary wave.

In order to compute the periodic case on approximately the same space-time domain as in the single solitary case, we again set the parameter values $c=0.1$, $\overline{x}=0.0$, $\epsilon =\mu =1.0$ and the parameter *k* is experimentally set up as $k=0.63048$ to have periods ${T}_{k}\doteq 20.0$ and ${\omega}_{k}\doteq 22.0$.

In what follows, we shall proceed similarly as in Section 5.1 to verify the convergence and preservation of studied invariant quantities.

#### 5.2.1 Convergence with respect to *h*

The *h*-convergence in the periodic case is investigated on a sequence of five successive refined grids partitioning the considered problem domain $[-44,44]$. The choice of time step is again small enough to suppress the influence of time discretization errors, and the computations are performed by piecewise linear approximations, subsequently.

*i.e.*, ${\text{err}}_{h}^{0}=\mathcal{O}({h}^{2})$ and ${\text{err}}_{h}^{1}=\mathcal{O}(h)$, for piecewise linear approximations and confirm the spatially suboptimal

*a priori*error estimates (66) with respect to the ${L}^{2}$-norm and spatially optimal estimates (67) in the energy norm, respectively. The influence of discretization errors on computations with a long time domain can be better eliminated by using the Crank-Nicolson numerical scheme instead of the backward Euler method.

**Periodic case: Computational errors in the** ${\mathit{L}}^{\mathbf{2}}$ **-norm and experimental orders of convergence for** ${\mathit{P}}^{\mathbf{1}}$ **approximation on a consequence of meshes at time instances** **t****(** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{5.0}$ | $\mathit{t}\mathbf{=}\mathbf{10.0}$ | $\mathit{t}\mathbf{=}\mathbf{15.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||||
---|---|---|---|---|---|---|---|---|

${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{0}}$ | EOC | |

0.80 | 2.790E-02 | - | 3.589E-02 | - | 4.531E-02 | - | 6.290E-02 | - |

0.40 | 7.016E-03 | 1.992 | 9.085E-03 | 1.982 | 1.156E-02 | 1.971 | 1.612E-03 | 1.964 |

0.20 | 1.860E-03 | 1.915 | 2.308E-03 | 1.977 | 3.005E-03 | 1.943 | 4.281E-03 | 1.912 |

0.10 | 4.952E-04 | 1.909 | 6.179E-04 | 1.901 | 8.084E-04 | 1.894 | 1.204E-04 | 1.830 |

0.05 | 1.357E-04 | 1.868 | 1.724E-04 | 1.841 | 2.314E-04 | 1.804 | 3.534E-04 | 1.768 |

**Periodic case: Computational errors in the energy norm and experimental orders of convergence for** ${\mathit{P}}^{\mathbf{1}}$ **approximation on a consequence of meshes at time instances** **t****(** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{5.0}$ | $\mathit{t}\mathbf{=}\mathbf{10.0}$ | $\mathit{t}\mathbf{=}\mathbf{15.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||||
---|---|---|---|---|---|---|---|---|

${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | ${\mathbf{err}}_{\mathit{h}}^{\mathbf{1}}$ | EOC | |

0.80 | 9.029E-03 | - | 9.004E-03 | - | 9.115E-03 | - | 9.251E-03 | - |

0.40 | 4.615E-03 | 0.968 | 4.658E-03 | 0.951 | 4.767E-03 | 0.935 | 4.862E-03 | 0.928 |

0.20 | 2.370E-03 | 0.961 | 2.488E-03 | 0.905 | 2.597E-03 | 0.876 | 2.770E-03 | 0.812 |

0.10 | 1.275E-03 | 0.894 | 1.363E-03 | 0.868 | 1.502E-03 | 0.790 | 1.601E-03 | 0.790 |

0.05 | 6.906E-04 | 0.885 | 7.533E-04 | 0.856 | 8.569E-04 | 0.809 | 9.154E-04 | 0.807 |

#### 5.2.2 Convergence with respect to *τ*

*τ*-convergence is experimentally verified by the computations on the finest spatial grid having 1,760 elements with piecewise linear approximation. The computations are performed by five different time steps

*τ*and monitored at final time of one period ${T}_{k}$. The theoretical results are in accordance with the observations listed in Table 7,

*i.e.*, ${\text{err}}_{\tau}^{0}=\mathcal{O}(h)$ and ${\text{err}}_{\tau}^{1}=\mathcal{O}(h)$.

**Periodic case: Computational errors in the** ${\mathit{L}}^{\mathbf{2}}$ **-norm and the energy norm for** ${\mathit{P}}^{\mathbf{1}}$ **approximation with respect to time step (** $\mathit{h}\mathbf{=}\mathbf{0.05}$ **)**

| $\mathit{t}\mathbf{=}\mathbf{20.0}$ | $\mathit{t}\mathbf{=}\mathbf{20.0}$ | ||
---|---|---|---|---|

${\mathbf{err}}_{\mathit{\tau}}^{\mathbf{0}}$ | EOC | ${\mathbf{err}}_{\mathit{\tau}}^{\mathbf{1}}$ | EOC | |

0.2000 | 7.690E-02 | - | 2.287E-02 | - |

0.1000 | 4.061E-02 | 0.921 | 1.221E-02 | 0.905 |

0.0500 | 2.078E-02 | 0.967 | 6.393E-03 | 0.934 |

0.0250 | 1.044E-02 | 0.993 | 3.453E-03 | 0.889 |

0.0125 | 5.514E-03 | 0.921 | 1.831E-03 | 0.915 |

From the presented numerical results in Sections 5.1.1-5.1.2 and 5.2.1-5.2.2, we see that the quality of approximate solutions obtained for a single solitary case and a periodic case is quite comparable.

#### 5.2.3 Invariant conservation quantities

Similarly as in Section 5.2.3, we monitor the preservation of invariants of mass, momentum and energy defined by (75), (76) and (77), respectively. During the whole period of time, in the course of which the waves propagate inside the periodic domain $[-2{\omega}_{k},2{\omega}_{k}]$, all these three invariants of motion remain conserved and equal to their original values that are well-determined analytically at $t=0$.

**Periodic case: Computed invariant quantities and errors in the** ${\mathit{l}}^{\mathbf{\infty}}$ **-norm and the** ${\mathit{L}}^{\mathbf{2}}$ **-norm (** $\mathit{h}\mathbf{=}\mathbf{0.05}$ **,** $\mathit{\tau}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{3}}$ **)**

Method | Time | ${\mathbf{err}}^{\mathbf{\infty}}$ | ${\mathbf{err}}^{\mathbf{0}}$ | ${\mathit{I}}_{\mathit{M}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ | ${\mathit{I}}_{\mathit{P}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ | ${\mathit{I}}_{\mathit{E}}\mathbf{(}{\mathit{u}}_{\mathit{h}}^{\mathit{l}}\mathbf{)}$ |
---|---|---|---|---|---|---|

present method | 0.0 | - | - | 16.5051 | 3.3180 | 10.6008 |

5.0 | 2.855E-05 | 1.357E-04 | 16.5056 | 3.3180 | 10.6005 | |

10.0 | 3.063E-05 | 1.724E-04 | 16.5057 | 3.3179 | 10.6003 | |

15.0 | 4.604E-05 | 2.314E-04 | 16.5057 | 3.3178 | 10.6000 | |

20.0 | 7.854E-05 | 3.534E-04 | 16.5060 | 3.3178 | 10.6001 | |

analytical val. ( | - | - | - | 16.50560 | 3.318064 | 10.60086 |

## 6 Conclusion

We have presented and theoretically analyzed an efficient numerical method for the solution of the RLW equation, which is based on the space dicretization by the discontinuous Galerkin method and a semi-implicit time discretization with suitable linearization of convective terms. Under some additional assumptions, we have derived *a priori* error estimates, namely $\mathcal{O}({h}^{p}+\tau )$ in the ${L}^{2}$-norm and in the energy norm. On the other hand, the presented numerical experiments for single solitary as well as periodic cases signal a better behavior of the experimental $({L}^{2})$-order of convergence, which is expected to be asymptotically $\mathcal{O}({h}^{2}+\tau )$ for piecewise linear approximations with a nonsymmetric variant of interior penalty Galerkin discretizations. In the case of the energy norm, we obtain the optimal experimental order of convergence.

The obtained results confirm that the proposed scheme is a powerful and reliable method for the numerical solution of a nonstationary nonlinear partial differential equation such as the RLW equation.

## Notes

### Acknowledgements

Dedicated to Professor Hari M Srivastava.

The authors would like to express their sincere gratitude to the referees for valuable comments and helpful suggestions. JH also would like to thank P. Červenková for her assistance with elaboration of numerical experiments. This work was partly supported by the ESF Project No. CZ.1.07/2.3.00/09.0155 ‘Constitution and improvement of a team for demanding technical computations on parallel computers at TU Liberec’ and by SGS Project ‘Modern numerical methods’ financed by TU Liberec.

## Supplementary material

## References

- 1.Peregrine DH: Calculations of the development of an undular bore.
*J. Fluid Mech.*1966, 25: 321-330. 10.1017/S0022112066001678CrossRefGoogle Scholar - 2.Pava JA Mathematical Surveys and Monographs 156. In
*Nonlinear Dispersive Equations: Existence and Stability of Solitary and Periodic Travelling Wave Solutions*. Am. Math. Soc., Providence; 2009.CrossRefGoogle Scholar - 3.Esen A, Kutluay S: A finite difference solution of the regularized longwave equation.
*Math. Probl. Eng.*2006., 2006: Article ID 85743Google Scholar - 4.Haq F, Islam S, Tirmizi IA: A numerical technique for solution of the MRLW equation using quartic
*B*-splines.*Appl. Math. Model.*2010, 34(12):4151-4160. 10.1016/j.apm.2010.04.012MathSciNetCrossRefGoogle Scholar - 5.Islam S, Haq S, Ali A: A meshfree method for the numerical solution of the RLW equation.
*J. Comput. Appl. Math.*2009, 223(2):997-1012. 10.1016/j.cam.2008.03.039MathSciNetCrossRefGoogle Scholar - 6.Chen Y, Mei L: Explicit multistep method for the numerical solution of RLW equation.
*Appl. Math. Comput.*2012, 218: 9547-9554. 10.1016/j.amc.2012.03.050MathSciNetCrossRefGoogle Scholar - 7.Dag I, Ozer MN: Approximation of the RLW equation by the least square cubic
*B*-spline finite element method.*Appl. Math. Model.*2001, 25: 221-231. 10.1016/S0307-904X(00)00030-5CrossRefGoogle Scholar - 8.Esen A, Kutluay S: Application of a lumped Galerkin method to the regularized long wave equation.
*Appl. Math. Comput.*2006, 174(2):833-845. 10.1016/j.amc.2005.05.032MathSciNetCrossRefGoogle Scholar - 9.Cockburn B: Discontinuous Galerkin methods for convection dominated problems. Lecture Notes in Computational Science and Engineering 9. In
*High-Order Methods for Computational Physics*. Edited by: Barth TJ, Deconinck H. Springer, Berlin; 1999:69-224.CrossRefGoogle Scholar - 10.Cockburn B, Karniadakis GE, Shu C-W (Eds):
*Discontinuous Galerkin Methods*. Springer, Berlin; 2000.Google Scholar - 11.Rivière B: Discontinuous Galerkin methods for solving elliptic and parabolic equations: theory and implementation. In
*Frontiers in Applied Mathematics*. SIAM, Philadelphia; 2008.Google Scholar - 12.Arnold DN, Brezzi F, Cockburn B, Marini LD: Unified analysis of discontinuous Galerkin methods for elliptic problems.
*SIAM J. Numer. Anal.*2002, 39(5):1749-1779. 10.1137/S0036142901384162MathSciNetCrossRefGoogle Scholar - 13.Hozman J: Discontinuous Galerkin method for numerical solution of the regularized long wave equation.
*AIP Conf. Proc.*2012, 1497: 118-125.CrossRefGoogle Scholar - 14.Dolejší V, Feistauer M, Hozman J: Analysis of semi-implicit DGFEM for nonlinear convection-diffusion problems.
*Comput. Methods Appl. Mech. Eng.*2007, 196: 2813-2827. 10.1016/j.cma.2006.09.025CrossRefGoogle Scholar - 15.Dolejší V, Feistauer M, Sobotíková V: Analysis of the discontinuous Galerkin method for nonlinear convection-diffusion problems.
*Comput. Methods Appl. Mech. Eng.*2005, 194: 2709-2733. 10.1016/j.cma.2004.07.017CrossRefGoogle Scholar - 16.Feistauer M, Felcman J, Straškraba I:
*Mathematical and Computational Methods for Compressible Flow*. Oxford University Press, Oxford; 2003.Google Scholar - 17.Dolejší V, Hozman J,
*et al*.: A priori error estimates for DGFEM applied to nonstationary nonlinear convection-diffusion equation. In*Numerical Mathematics and Advanced Applications, ENUMATH 2009*. Edited by: Kreiss G. Springer, Berlin; 2010:459-468.Google Scholar - 18.Ciarlet PG:
*The Finite Elements Method for Elliptic Problems*. North-Holland, Amsterdam; 1979.Google Scholar - 19.Kučera V: On
*ε*-uniform error estimates for singularly perturbed problems in the DG method.*Numerical Mathematics and Advanced Applications 2011*2013, 368-378.Google Scholar - 20.Djidjeli K, Price WG, Twizell EH, Cao Q: A linearized implicit pseudo-spectral method for some model equations: the regularized long wave equations.
*Commun. Numer. Methods Eng.*2003, 19: 847-863. 10.1002/cnm.635MathSciNetCrossRefGoogle Scholar - 21.Abramowitz M, Stegun IA:
*Handbook of Mathematical Functions*. Dover, New York; 1965.Google Scholar

## Copyright information

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.