Hindawi

Journal of Applied Mathematics

Volume 2023, Article ID 4941679, 21 pages
https://doi.org/10.1155/2023/4941679

Research Article
Exact Nonlinear Spectral Analysis of Nonsmooth
WENO-JS Solutions

1,2
Tamer H. M. A. Kasem
1
Smart Engineering Systems Research Center (SESC), Nile University, Giza 12588, Egypt
2
Cairo University, Egypt

Correspondence should be addressed to Tamer H. M. A. Kasem; tamer.kasem65@gmail.com

Received 14 February 2023; Revised 8 April 2023; Accepted 20 April 2023; Published 10 May 2023

Academic Editor: Oluwole D. Makinde

Copyright © 2023 Tamer H. M. A. Kasem. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

The fifth-order accurate Weighted Essentially Nonoscillatory space discretization developed by Jiang and Shu (WENO-JS) is
studied theoretically. An exact Nonlinear Spectral Method (NSM) is developed based on an innovative yet simple
methodology. The NSM explains the behaviour of nonsmooth solutions because it is valid for arbitrary modified wave
numbers (MWN). The NSM clarifies the effects of the time integration methods and the Courant number. The mode isolation
assumption, extensively used to analyse WENO-JS, is elucidated, and several novel findings are presented. The improved
performance of the combination of WENO-JS with the forward Euler time integration method, compared to the Linear Fifth-
Order Upwind discretization, is illustrated. The overdamping of the combination of WENO-JS with the popular third-order
total variation diminishing Runge-Kutta method is discovered. Thus, the NSM covers several gaps in the current literature.

1. Introduction main categories: Geometric Stability Analysis (GSA) and

Generally, high-order accurate discretizations suffer from spu-
rious oscillations near high gradients due to Gibbs oscillations
[1]. The fifth-order accurate Weighted Essentially Nonoscilla-
tory method was developed by Jiang and Shu (WENO-JS) to
achieve high-order accuracy while avoiding Gibbs oscillations
[2]. The design of WENO-JS is based on a clever nonlinear
modification of the fifth-order Linear Upwind (LUW5) discre-
tization. The relatively wide five-point stencil of LUW5 can be
subdivided into three small substencils. In WENO-JS, each
smooth substencil is assigned a higher weight and vice versa.
Hence, the constant weights adopted by LUW5 are replaced
by nonlinear rational weights in WENO-JS.
A wide variety of WENO-JS improvements were devel-
oped, and a sample of the recent developments was provided
in [3]. On the contrary, theoretical studies of WENO-JS are
relatively rare. Spectral analysis of WENO-JS is a theoretical
methodology that quantifies the solution in terms of dissipa-
tion and dispersion [4–7]. The spectral analysis methodolo-
gies used to analyse WENO-JS can be classified into two
Approximate Dispersion Relation (ADR).
The pioneering study of [5] adopted GSA to study the Lin-
ear Stability of WENO-JS and concluded that the forward
Euler time integration method is linearly unstable. The results
of [5] were revised by [6] who illustrated how linear stability
can be satisfied because the discretization spectrum is equal
to a finite set. However, the GSA provided by [5, 6] adopted
linearized approximations of WENO-JS. For instance, many
results were derived by [5] using Taylor expansions in terms
of the modified wave number (MWN). Hence, these results
were limited to small MWN that describe smooth solutions.
This is a major disadvantage because the main goal of
WENO-JS is modelling non-smooth solutions coexistent with
high MWN. Moreover, the GSA of [6] was mainly based on
LUW5. Hence, both [5, 6] did not identify the discrepancy
between WENO-JS and LUW5 at high MWN.
The ADR is an alternative methodology based on dis-
crete Fourier transform (DFT) and was originally developed
by [4]. Within ADR, an initial condition defined on an arbi-
trary grid is advanced numerically in time. Next, DFT is
2 Journal of Applied Mathematics

applied to the solution to obtain the spectrum. The ADR and will be presented briefly. The reader may refer to [1] for
algorithm accounts for nonlinearity and is valid for arbitrary more details. The main task is to obtain a high-order finite dif-
MWN. The ADR was used to analyse the cost efficiency of ference discretization of the flux space derivative ∂f ðuÞ/∂x . To
WENO-JS [4], and a wide application was presented in [8]. simplify notations, let f stand for f ðuÞ. The space derivative of f
The ADR was further developed by [7] and applied to ana- at xi will be written as
lyse synthetic turbulent energy spectra [9, 10]. The design
of the ADR algorithm requires adopting an arbitrarily small ∂f f i+1/2 − f i−1/2 À Á
time step to minimize the effects of time integration. There = + O h5 : ð4Þ
∂x xi h
are no definite criteria to select the numerical grid parame-
ters within the ADR. In addition, neither the effects of the
Courant number (CN) nor those of the numerical integra- The attention is shifted towards the calculation of f i−1/2 .
tion method were provided in [4, 7–10]. This quantity is commonly termed as the numerical flux at
A cornerstone assumption that was adopted in the past point xi−1/2 = ðxi + xi−1 Þ/2. Generally, f i−1/2 is calculated based
GSA and ADR studies was ignoring the higher-order on the five-point stencil shown in Figure 1. This five-point sten-
modes induced due to WENO-JS nonlinearity. However, cil includes three smaller substencils as also illustrated in
theoretical validation of this assumption was not provided Figure 1. The three substencils are given by
yet. This assumption will be denoted by mode isolation in È É
this work. An exact methodology is proposed in this study S j = xp : xp ∈ X, p ∈ fi − 3 + j, i − 2 + j, i − 1 + jg, j ∈ f0, 1, 2g :
to address the challenges mentioned above. A novel Non-
ð5Þ
linear Spectral Method (NSM) is introduced to analyse
WENO-JS. The benefits of the application of spectral
methods to nonlinear problems are well explained in Various choices exist to calculate f i−1/2 . Consider the choice
of ̂f i−1/2 that yields a third-order approximation given by
j
[11]. The development of the novel NSM is based on a
continuous presentation of the discrete solution. This sub-
tle and key idea does not compromise accuracy and proves ̂f j ̂j À Á
∂f i+1/2 − f i−1/2
to be effective within the analysis. Unlike [5, 6], NSM = + O h3 : ð6Þ
includes the full nonlinear rational weights of WENO-JS, ∂x xi h
and the results are valid for the full range of MWN. The
NSM does not adopt any arbitrary grids, and the effects Here, the superscript j corresponds to the substencil S j
of CN and various time integration methods are described.
selected to calculate ̂f i−1/2 .
j
A quantitative analysis of the mode isolation assumption is
provided. To the author’s best knowledge, these contribu- The linear combination of f i that is used to calculate
tions are provided for the first time. ̂f j is given by
i−1/2

2. Numerical Discretization i−1+j

̂f j 〠 cp f p :
j
ð7Þ
i−1/2 =
The numerical algorithms studied in this work will be p=i−3+j
detailed. Consider the scalar conservation law given by
j
The values of the constants cp are provided explicitly
∂u ∂f ðuÞ
+ = 0, ðx, t Þ ∈ Ωx × Ωt , ð1Þ below for each S j .
∂t ∂x
8 0 1 7 11
subject to the suitable boundary and initial conditions. Here, u > ̂
= uðx, tÞ stands for a conserved quantity, and f ðuÞ denotes > f i−1/2 = 3 f i−3 − 6 f i−2 + 6 f i−1 ,
>
>
>
>
the corresponding flux. Space and time domains are defined <
̂f j
̂f 1 1 5 1
as Ωx = ½0, LŠ and Ωt = ½0,∞Þ, respectively. To obtain a numer- = i−1/2 = − f i−2 + f i−1 + f i , ð8Þ
i−1/2
>
> 6 6 3
ical solution of Equation (1) using the finite difference method, >
>
>
> 1 5 1
the space domain Ωx is discretized using the grid X given by : ̂f 2 f + f − f :
i−1/2 =
3 i−1 6 i 6 i+1
X = fxi : xi = ih, i ∈ f0, 1, ⋯, N gg: ð2Þ
The three choices for ̂f i−1/2 can be combined to provide
j

Here, h = L/N stands for the grid spacing, and N ∈ ℕ a fifth-order accurate flux using the weighted sum given by
stands for the total number of grid subdivisions. The numerical
solution is defined as j=2
f i−1/2 = 〠 C j ̂f i−1/2 :
j
ð9Þ
u i ≈ u ðx i , t Þ: ð3Þ j=0

2.1. Fifth-Order Space Discretization. The fifth-order accurate The selection of the weights C j will be detailed in the
space discretization algorithm is explained in numerous sources next subsections.
Journal of Applied Mathematics 3

S2 ðmÞ
Here, q j is the mth derivative of the polynomial used to
S0 interpolate f at S j . This definition of β j in Equation (13)
yields the following expressions:
i−3 i−2 i−1 i i+1
xi−3 xi−2 xi−1 i−1/2 xi xi+1 8 13 1
>
> β0 = ð f − 2f i−2 + f i−1 Þ2 + ð f i−3 − 4f i−2 + 3f i−1 Þ2 ,
>
> 12 i−3
>
>
4
<
S1 13 1
β j = β1 = ð f − 2f i−1 + f i Þ2 + ð f i−2 − f i Þ2 ,
>
> 12 i−2 4
Figure 1: The full five-point stencil used to calculate f i−1/2 and the >
>
>
> 13 1
substencils S j where j ∈ f0, 1, 2g. :β = ð f i−1 − 2f i + f i+1 Þ2 + ð3f i−1 − 4f i + f i+1 Þ2 :
2
12 4
ð14Þ

2.1.1. Fifth-Order Linear Upwind Discretization. Generally, The modified weights are calculated as
f i are interpolated using a high-order polynomial. The fol-
lowing constants are defined [2, 12]:
dj
αj =  2 : ð15Þ
ϵ + βj
d 0 = 0:1, d 1 = 0:6, d2 = 0:3: ð10Þ

Here, ϵ is an arbitrary small number introduced to avoid

The same values of d j are provided by [1] in the opposite division by zero. The normalized weights are given by
order. The reason for this slight difference is that [1] listed
the formulas to calculate the flux at point xi+1/2 instead of αj
xi−1/2 . For the fifth-order Linear Upwind (LUW5) method, ωj = s=2
: ð16Þ
∑s=0 αs
C j are given by

Hence, for WENO-JS, the weights C j are given by

C j = d j: ð11Þ

C j = ωj, ð17Þ
Hence, the fifth-order accurate linear numerical flux
f
i−1/2 is given by and WENO-JS flux is given by

j=2 j=2
f ̂j
i−1/2 = 〠 d j f i−1/2 : ð12Þ ~f ̂j
i−1/2 = 〠 ω j f i−1/2 : ð18Þ
j=0 j=0

The algorithm of LUW5 is linear because constant values Substituting f i−1/2 by f i−1/2 (given by Equation (12)) or
are assigned to C j , and accurate results may be obtained in ~f
i−1/2 (given by Equation (18)) in Equation (4) results in
smooth regions. However, the high-order polynomial used LUW5 or WENO-JS space discretizations, respectively.
in LUW5 may produce spurious oscillations near high gradi-
ents due to the Gibbs phenomenon. 2.2. Numerical Time Integration. The operator Lðui Þ is
introduced to simplify notations. Referring to Equation (4),
2.1.2. Fifth-Order Weighted Essentially Nonoscillatory Lðui Þ is defined as
Discretization. The fifth-order WENO-JS method is
designed to overcome the disadvantages of LUW5 [1],  
and C j are designed to be higher for smooth S j and vice ∂f ðuÞ
Lðui Þ = f i−ð1/2Þ − f i+ð1/2Þ = −h × + O h6 : ð19Þ
versa. Similar to [5, 6], the current analysis will focus on ∂x xi
WENO-JS. Smoothness is quantified using the smoothness
indicators (β j ) based on a sub-stencil S j . The L2 norm The operator Lðui Þ encapsulates the space discretization
squared was used to calculate β j in [2] using the following algorithm. The time integration algorithm is based on the
formula: method of lines [1, 4, 5]. A semidiscrete approximation of
Equation (1) is given by
2 ð xi+1/2  
ðmÞ 2 dui Lðui Þ
βj = 〠 h2m−1 q j dx: ð13Þ = : ð20Þ
m=1 xi−1/2 dt h
4 Journal of Applied Mathematics

To perform the numerical integration of Equation (20), The space periodic boundary condition uð0, tÞ = uðL, tÞ
the time domain Ωt is discretized using T given by is adopted. Consider the initial condition uðx, t = 0Þ = FðxÞ.
È É The exact solution ue ðx, tÞ is obtained by the method of
T = t n : n ∈ ℕ, t n+1 > t n : ð21Þ characteristics as follows:

The time step Δ is defined as Δ = t n+1 − t n . Numerical ue ðx, t Þ = F ðx − t Þ: ð27Þ

integration can be done using a variety of standard methods.
The single-stage, first-order Euler forward solution of Equa- Hence, the exact solution corresponds to a right-
tion (20) is written as follows [1]: travelling wave whose form is identical to the initial condi-
tion and speed equal to 1.
Δ
un+1
i = uni + Lðuni Þ: ð22Þ 3.1. Spectral Analysis of Numerical Solutions. Any periodic
h intermediate solution uni is exactly presented as a summation
Here, uni = uðxi , t n Þ and un+1 = uðxi , t n+1 Þ are the solution of Fourier modes.
i
vectors at t and t , respectively. The symbol λ = Δ/h is
n n+1
N−1   N−1  x 
introduced. And Equation (22) is rewritten as follows: uni = 〠 a j sin θ j i + ϕ j = 〠 a j sin θ j i + ϕ j : ð28Þ
j=0 j=0
h
un+1
i = uni + λLðuni Þ: ð23Þ
The total number of modes is equal to the grid size N.
For a wave speed of unity, λ is commonly termed the The modified wave number (MWN) termed as θ j is given by
Courant number (CN). The first-order Euler forward
method can be regarded as a member of the set of j
Runge-Kutta (RK) methods [13]. Hence, the method θ j = 2π : ð29Þ
N
defined by Equation (23) will be denoted by RK1.
Higher-order multistage time integration algorithms can The operation defined by Equation (28) is termed the
be implemented using the intermediate solution vectors discrete Fourier transform (DFT). An extended explanation
ui ð1Þ and ui ð2Þ . The second-order explicit total variation of DFT is provided in Appendix A. Generally, the range of
diminishing Runge-Kutta (TVDRK2) solution of Equation θ j is θ j ∈ ½0, 2πŠ. However, for any real-valued periodic func-
(20) is given by tion, the total number of unique modes is N/2 and the rest of
the modes are obtained by Hermitian symmetry [14]. Hence,
ui ð1Þ = uni + λLðuni Þ, the effective range of θ j is reduced to be θ j ∈ ½0, πŠ as adopted
1 n 1 ð1Þ 1  ð1Þ  ð24Þ by [4, 8]. The Hermitian symmetric property of DFT of real
un+1 = u + u + λL ui : functions is derived in Appendix B.
i
2 i 2 i 2
The numerical solution un+1 i of any numerical algorithm
The third-order explicit total variation diminishing can be written as
Runge-Kutta (TVDRK3) solution of Equation (20) is given by
N−1  
ui ð1Þ
= uni + λLðuni Þ, un+1
i = 〠 G j × a j sin θ j i + ϕ j + Φ j : ð30Þ
j=0
3 n 1 ð1Þ 1  ð1Þ 
ui ð2Þ =u + u + λL ui , ð25Þ
4 i 4 i 4 The spectral parameters G j and Φ j stand for the magni-
1 2 ð2Þ 2  ð2Þ  fication factor and the phase shift, respectively. Generally, G j
un+1 = uni + u + λL ui :
i
3 3 i 3 and Φ j are expressed as

The combination of WENO-JS space discretization with À Á

G j = G j θ j , λ, a0 , ⋯, aN−1 , ϕ0 , ⋯, ϕN−1 , ð31Þ
Equations (23), (24), and (25) will be denoted by WENO-
RK1, WENO-TVDRK2, and WENO-TVDRK3, respectively. À Á
Similarly the combinations of LUW5 with Equations (23), Φ j = Φ j θ j , λ, a0 , ⋯, aN−1 , ϕ0 , ⋯, ϕN−1 : ð32Þ
(24), and (25) will be denoted by LUW5-RK1, LUW5-
The effect of the numerical algorithm on each mode is
TVDRK2, and LUW5-TVDRK3, respectively.
fully described in terms of G j and Φ j . A value of G j > 1 or G j
3. Analysis < 1 indicates whether mode j is magnified or damped, respec-
tively. The significance of Φ j will be clarified as follows.
Similar to [4–6, 9], the analysis will focus on the linear flux Rewrite Equation (30) in terms of xi , and factor out θ j .
f ðuÞ = u. Hence, Equation (1) is rewritten as
!!
N−1
xi ϕ j Φ j
∂u ∂u un+1 = 〠 G j × a j sin θj + + : ð33Þ
+ = 0: ð26Þ i
h θj θj
∂t ∂x j=0
Journal of Applied Mathematics 5

Each mode j in uni can be regarded as a wave, that travels at (43) approach those provided by Equations (40) and (41),
a speed V j . Hence, un+1
i is obtained by shifting uni by a distance respectively.
V j Δ. Hence, Equation (33) is rewritten in the form of the trav- So far, no limitations have been imposed on the pre-
elling wave solution: sented analysis. However, it will be useful to introduce an
assumption that was a cornerstone in [4–6]. This assump-
!! tion will be denoted by the Mode Isolation or MI for abbre-
N−1
xi − V j Δ ϕ j
ui = 〠 G j × a j sin θ j
n+1
+ : ð34Þ viation. This assumption is stated as follows.
j=0
h θj
Assumption 1. Let uni be a monochromatic intermediate dis-
Comparing the arguments of the sin function in Equations crete solution described by Equation (36). Then, un+1 i result-
(33) and (34), we get ing from Equation (23) is well approximated as
Φ j = −V j θ j λ: ð35Þ  
un+1
i ≈ G j × a j sin θ j i + ϕ j + Φ j : ð44Þ
Hence, Φ j is proportional to V j .
In other words, when a monochromatic uni is considered,
3.1.1. Mode Isolation. Consider a monochromatic intermedi- the numerical solution un+1 can be equated to a single mode.
i
ate solution uni that is described by a single mode [4] and Hence, all the summation terms in Equation (30) are
defined as equated to zero, except that corresponding to the principal
  mode. The validity of Assumption 1 will be thoroughly clar-
uni = a j sin θ j i + ϕ j : ð36Þ ified in the next sections.
For this monochromatic case, let us denote all wave param- 3.2. Linear von Neumann Analysis of LUW5. In this subsec-
eters with subscript j (e.g., θ j ) as the principal mode parameters. tion, the Linear von Neumann (VN) analysis will be pre-
It will be useful to write explicitly the exact solution for this sented by application to the LUW5-RK1 algorithm. The
case using Equation (27). First, Equation (36) is rewritten as VN analysis is generally used to study finite domain
 x  problems with periodic boundary conditions [15]. Adopt-
uni = a j sin θ j i + ϕ j : ð37Þ ing Equation (11) yields a linear scheme, and the operator
h Lðui Þ is given by [6]
Next, the exact solution is obtained by replacing xi with 1 1 1 1 1
xi − Δ and using the definition of Equation (2). Lðui Þ = ui−3 − ui−2 + ui−1 − ui − ui+1 + ui+2 :
30 4 3 2 20
    ð45Þ
À Á Δ
ue xi , t n+1 = a j sin θ j i − + ϕj : ð38Þ
h Alternatively, Equation (45) is rewritten in the follow-
ing compact form.
Finally, Equation (38) is rewritten in terms of the spectral 2
parameters, introduced in Equation (30). Lðui Þ = 〠 Dr ui+r : ð46Þ
À Á   r=−3
ue x i , t n+1
= G j a j sin θ j i + ϕ j + Φ j : ð39Þ
Here, Dr are constants whose values are listed explicitly
in Equation (45). The numerical solution of LUW5-RK1
Upon comparison of Equations (38) and (39), the spec- based on Equation (23) can be written as
tral parameters of the exact solution are given by
2
G j = 1, ð40Þ un+1
i = uni + λ 〠 Dr uni+r : ð47Þ
r=−3

Φ j = −θ j λ: ð41Þ The coefficients of ui+r on the right-hand side of

Equation (47) are constants. Hence, Equation (47) is lin-
Let us consider the numerical solution un+1 i . Based on ear in terms of uni+r , and the analysis can be simplified as
Equation (36), the expression given by Equation (30) is unal- follows. The VN analysis considers a discrete Fourier
tered. However, the expressions given by Equations (31) and Mode [16] that corresponds to the intermediate solution
(32) are simplified to be given by Equation (36). Although one mode is considered
  out of the multiple modes in Equation (28), the analysis
G j = G j θ j , λ, a j , ϕ j , ð42Þ
is still general [17]. Also, adopting the complex notation
  is allowed,
Φ j = Φ j θ j , λ, a j , ϕ j : ð43Þ pffiffiffiffiffiffi and Equation (36) can be rewritten in terms
of i = −1 as follows:
As the accuracy of the numerical discretization is
improved, the expressions provided by Equations (42) and uni = a j eiðθ j i+ϕ j Þ : ð48Þ
6 Journal of Applied Mathematics

The expression of Equation (48) is substituted into based on adopting uni given by Equation (48). To analyse
Equation (47), and we get LUW5-TVDRK2, Equation (24) is rewritten as
À Á
2
ui ð1Þ = uni + λLðuni Þ = H θ j , λ uni ,
un+1 = a j eiðθ j i+ϕ j Þ + λ 〠 Dr eiθ j ði+rÞ+iϕ j : ð49Þ
1  ð1Þ  
i
r=−3 1
un+1
i = uni + ui + λL ui ð1Þ
2 2
To simplify Equation (49), the term a j eiðθ j i+ϕ j Þ is fac- 1 n 1 À Á 
tored out to obtain = ui + H θ j , λ ui ð1Þ ð57Þ
2 2
1 n 1 À Á À Á
! = ui + H θ j , λ × H θ j , λ uni
2
À Á 2 2 
iðθ j i+ϕ j Þ
un+1 = aje 1 + λ 〠 Dr e iθ j r
= uni H θ j , λ : ð50Þ 1 1 2À Á
i
r=−3
= uni + H θj, λ :
2 2

Here, Hðθ j , λÞ = 1 + λ∑2r=−3 Dr eiθ j r , and the second Hence, for LUW5-TVDRK2, the spectral parameters are
equality is obtained by substitution by Equation (48). defined as
The spectral parameters are calculated as
1 1 2À Á
À Á Gj = + H θj, λ ,
Gj = H θj, λ , ð51Þ 2 2
  ð58Þ
À À ÁÁ 1 1 2À Á
Φ j = arg H θ j , λ : ð52Þ Φ j = arg + H θj, λ :
2 2
Hence, the numerical solution of LUW5-RK1 is writ-
ten as Similarly, LUW5-TVDRK3 is analysed using Equation
(25). The derivation follows the same ideas listed in Equa-
un+1 = G j eiΦ j uni : ð53Þ tion (57). To avoid distraction, the derivation is provided
i
in a relatively brief format.
The following points should be emphasized regarding À Á
the application of the VN method to LUW5-RK1 and ui ð1Þ = uni + λLðuni Þ = H θ j , λ uni ,

3 n 1  ð1Þ  
other linear discretizations.
The expressions for G j and Φ j provided by Equations (51) ui ð2Þ = ui + ui + λL ui ð1Þ
4 4 
and (52) are exactly valid for arbitrary values of h and Δ.
n 3 1 2À Á
= ui + H θj, λ ,
(i) As illustrated by Equation (50), Assumption 1 is 4 4
1 n 2  ð2Þ  
exactly satisfied by LUW5-RK1 and all linear
discretizations un+1
i = ui + ui + λL ui ð2Þ
3 3 
(ii) Referring to Equations (51) and (52), the spectral 1 1 À Á 1 À Á
parameters G j and Φ j are independent of a j and ϕ j . = uni + H θ j , λ + H3 θ j , λ :
3 2 6
Hence, Equations (42) and (43) are simplified to be ð59Þ
À Á Hence, for LUW5-TVDRK3, the spectral parameters are
Gj = Gj θj, λ , ð54Þ
defined as
À Á
Φj = Φj θj, λ : ð55Þ 1 1 À Á 1 À Á
Gj = + H θ j , λ + H3 θ j , λ ,
3 2 6
3.2.1. Extension to High-Order Time Integration. The analy-   ð60Þ
sis based on VN can be extended to high-order multistage 1 1 À Á 1 À Á
Φ j = arg + H θ j , λ + H3 θ j , λ :
time integration methods. It should be noted that the algo- 3 2 6
rithms described in Equations (24) and (25) adopt Equation
(23) as a building block. Hence, Hðθ j , λÞ will be used fre- 3.3. Nonlinear Spectral Analysis of WENO-JS. It should be
quently. Since LUW5 is linear, the following simplification noted that the numerical solution of Equation (26) using
of expressions including L is justified: WENO-JS is nonlinear. The reason is that C j used in
WENO-JS are not constants and are based on rational
LðA × uni Þ = A × Lðuni Þ: ð56Þ expressions in terms of uni . Hence, the steps explained in
Section 3.2 cannot be applied directly to WENO-JS. Still,
The symbol A in Equation (56) stands for any expression the nonlinear features of WENO-JS could be analysed using
that does not depend explicitly on i. The analysis will be a slight modification of the VN method. This modification is
Journal of Applied Mathematics 7

based on a continuous presentation of WENO-JS. This pre-
sentation’s main benefit is clarifying two main solution
properties: periodicity and homogeneity.
As explained in Section 3.2, the VN method is based
on analysing Fourier modes for periodic finite domains.
It will be shortly illustrated how the solution periodicity
will be used to develop a nonlinear analysis that enjoys
several advantages of the VN method. For instance, the
grid parameters h and N will be eliminated and the anal-
ysis will depend only on λ and θ j . The homogeneity of
WENO-JS weights and flux will eliminate the dependence
of the numerical solution on a j if the MI is adopted. It
could be shown that the periodicity and the homogeneity
properties are satisfied by several other nonlinear numeri-
cal algorithms. However, only WENO-JS will be consid-
ered in the current work.
3.3.1. Continuous Presentation of WENO-JS. The algorithm
of WENO-JS is designed originally to deal with discrete
quantities. However, WENO-JS algorithm steps can be
applied to continuous quantities as follows. The continuous
version of β j is obtained by substituting with the linear flux
f ðuÞ = u into Equation (14) and replacing f i−k by uðx − khÞ,
where k ∈ f−1, 0, 1, 2, 3g. Hence, based on Equation (14) the
continuous version of β j is given by

8 13 1
>
> βc0 = ðuðx − 3hÞ − 2uðx − 2hÞ + uðx − hÞÞ2 + ðuðx − 3hÞ − 4uðx − 2hÞ + 3uðx − hÞÞ2 ,
>
>
>
>
12 4
<
13 1
βcj ðx ; hÞ = βc1 = ðuðx − 2hÞ − 2uðx − hÞ + uðxÞÞ2 + ðuðx − 2hÞ − uðxÞÞ2 , ð61Þ
>
> 12 4
>
>
>
> 13 1
: βc = ðuðx − hÞ − 2uðxÞ + uðx + hÞÞ2 + ð3uðx − hÞ − 4uðxÞ + uðx + hÞÞ2 :
2
12 4

The superscript c in βcj ðx ; hÞ stands for continuous. The Lemma 1. Let uðxÞ be a space periodic function with period
reader should note the following: P ∈ ℝ. Then, βcj ðx ; hÞ is also space periodic with period P.

(i) Although t was skipped in Equation (61), time Proof. By definition,

dependence is still included. In fact uðx − khÞ = uðx
− kh, tÞ. This symbolic simplification is justified
since Lðui Þ operates only in space uðxÞ = uðx + PÞ: ð64Þ

(ii) βcj ðx ; hÞ is defined for a continuous range of x ∈ Substituting by x + P into Equation (62), we get
½0, LŠ. This is in contrast to β j defined for the dis-
crete points in S j defined by Equation (5)    
β β
βcj ðx + P ; hÞ = g j ∘ u ðx + PÞ = g j ∘ u ðxÞ = βcj ðx ; hÞ:
(iii) βcj ðx ; hÞ is equal to a second-order homogeneous
polynomial [18] in terms of uðx − khÞ. Hence, βcj ð65Þ
ðx ; hÞ can be regarded as a composite homoge-
neous function The second equality is obtained based on Equation (64).
Hence, βcj ðx ; hÞ is also periodic with period P.
Based on the discussion above, the following equation is
To proceed further, Equation (16) is rewritten as
introduced:
 2
 
β d j/ ϵ + βj
βcj ðx ; hÞ = gj ∘ u ðx Þ: ð62Þ ωj = : ð66Þ
∑s=2
s=0 d s /ðϵ + βs Þ
2

β
Here, ∘ stands for the composite function, and g j stands
for the polynomials defined in the right-hand side of Equa- Following the same direction, a continuous version of ω j
β is defined as
tion (61). Since g j ðuÞ is a second-order homogeneous func-
tion, the following equation is satisfied:  2
d j / ϵ + βcj ðx ; hÞ
β β ωcj ðx ; hÞ = c
: ð67Þ
g j ða × u Þ = a 2 × g j ðu Þ: ð63Þ ∑s=2
s=0 d s /ðϵ + βs ðx ; hÞÞ
2
8 Journal of Applied Mathematics

A composite function form can be also used to describe ~ c is applied to a × u

To clarify homogeneity, the flux F
ωcj ðx; hÞ as follows. Substituting Equation (62) into Equation as follows:
(67), we get
2
  2
β
d j / ϵ + g j ∘ u ðx Þ   gFða × uÞ = 〠 ωcj ðx ; hÞF̂cj ðx ; hÞ
ωcj ðx ; hÞ =  
ω
  2 = g j ∘ u ðx Þ:
j=0
β
∑s=2
s=0 d s / ϵ + gs ∘ u ðxÞ
2 i−1+j
= 〠 gωj ða × uÞ 〠 cp a × uðx + ðp − iÞhÞ
j

ð68Þ j=0 p=i−3+j

2 i−1+j
= a × 〠 gωj ðuÞ 〠 cp uðx + ðp − iÞhÞ = a × gFðuÞ:
j

j=0 p=i−3+j
Corollary 2. Let uðxÞ be a space periodic function with period
P ∈ ℝ. Then, ωcj ðx ; hÞ is also space periodic with period P. ð73Þ

Corollary 2 can be deduced using the same steps used to The second equality is based on Equations (68) and
prove Lemma 1.
(70). Hence, gFðuÞ is a first-order homogeneous function.
The value of ϵ should be proportional to the machine
precision [12]. Hence, ϵ will be neglected and assigned to
Corollary 3. Let uðxÞ be a space periodic function with period
zero within the current exact analysis. Consequently, the
P ∈ ℝ. Then, F~ c ðx ; hÞ is also space periodic with period P.
expression for gωj ða × uÞ is written based on Equation
(68) as
Corollary 3 can be deduced using the same steps used to
  prove Lemma 1 and Corollary 2.
 2  2 ~ c ðx ; hÞ, a continuous operator Lc ðuðxÞÞ is
β 4
d j/ a ×
β
g j ðuÞ Based on F
d j/ g j ða × uÞ
gωj ða × uÞ =  defined as

β
2 =  2 
∑s=2 4 × gβ ðuÞ
s=0 d s / gs ða × uÞ ∑s=2
s=0 d s / a s

 2
β
d j / g j ðuÞ ~ c ð x ; hÞ − F
Lc ðuðxÞÞ = F ~ c ð x + h ; hÞ
= 
ω
 2 = g j ð uÞ :
    ð74Þ
∑s=2
β = gF ∘ u ðxÞ − gF ∘ u ðx + hÞ:
s=0 sd / gsð Þ
u

ð69Þ
Corollary 4. Let uðxÞ be a space periodic function with period
The second equality in Equation (69) is based on P ∈ ℝ. Then, Lc ðuðxÞÞ based on WENO-JS is also space peri-
Equation (63). Hence, gωj ðuÞ is a zero-order homogeneous odic with period P.
function as illustrated by Equation (69) that. A continuous
The proof is straightforward and is based on Corollary 3.
version of ̂f i−1/2 as defined by Equation (7) is given by
j
It can be illustrated that Lc ðuðxÞÞ is a first-order homo-
geneous composite function as follows:
i−1+j
F̂cj ðx ; hÞ =
j
〠 cp uðx + ðp − iÞhÞ: ð70Þ    
p=i−3+j Lc ða × uðxÞÞ = a × gF ∘ u ðxÞ − a × gF ∘ u ðx + hÞ
= a × Lc ðuðxÞÞ:
Similarly, a continuous version of ~f i−1/2 defined by
ð75Þ
Equation (18) is given by

2 The first equality is justified by Equation (73).

~ c ðx ; hÞ = 〠 ωc ðx ; hÞFc ðx ; hÞ:
F ð71Þ The main theorem will be presented through the analysis
j ̂j
j=0 of the relatively simple WENO-RK1 algorithm.

The flux F ~ c ðx ; hÞ can be regarded as the composite Theorem 5. Consider the initial value problem of Equation
function given by (26), subject to space periodic boundary conditions. Let the
initial condition uðx, 0Þ be a space periodic function with
period P ∈ ℝ. Then, the numerical solution based on
 
~ c ð x ; hÞ = g F ∘ u ð x Þ :
F ð72Þ WENO-JS space discretization and RK1 time integration is
also space periodic with period P.
Journal of Applied Mathematics 9

Proof. The proof proceeds by induction. Let un ðxÞ = uðx, t n Þ θ j . As illustrated in Theorem 5, the subsequent solutions will
be a space periodic function with period P. Applying Equa- be also exactly periodic with the same period P j . The variable
tion (23), we get ic ∈ R is defined as a continuous version of index i to be
un+1 ðxÞ = un ðxÞ + λLc ðun ðxÞÞ: ð76Þ x
ic = : ð84Þ
h
c
The periodicity of λL ðu ðxÞÞ is clarified by Corollary 4,
n

and un ðxÞ is periodic by assumption. Hence, the whole ~n ðic Þ

right-hand side of Equation (76) is periodic with period P.
The proof is straightforward for the base case since u0 ðxÞ = u
ðx, 0Þ is periodic with period P by assumption.
Both variables x and ic are interchangeable. Hence, u
is defined as
~ n ði c Þ = u n ðx = h × i c Þ:
u ð85Þ

Corollary 6. Let un ðxÞ = uðx, t n Þ be a space periodic function Also, the numerical solution un+1 ðxÞ defined in Equation
with period P. Consider the numerical solution un+1 ðxÞ based (76) can be written as
on WENO-JS space discretization and RK1 time integration.
If un+1 ðxÞ is C0 continuous, then it can be expressed using ~n+1 ðic Þ = un+1 ðx = h × ic Þ:
u ð86Þ
the following Fourier series:
Using the continuous presentation, Equation (37) is
~a ∞  x ~  x rewritten in terms of x instead of xi as
un+1 ðxÞ = 0 + 〠 ~aα cos 2πα + bα sin 2πα  x 
2 α=1 P P
∞
ð77Þ un ðxÞ = a j sin θ j + ϕ j : ð87Þ
i2παðx/PÞ
h
= 〠 ~cα e :
α=−∞ Also, Equation (37) is rewritten in terms of ic as
 
The coefficients ~a0 , ~aα , ~bα , and ~cα are the Fourier expan- ~n ðic Þ = a j sin θ j ic + ϕ j :
u ð88Þ
sion coefficients defined as follows:
ðP The periodicity of un ðxÞ and u ~n ðic Þ are defined as
2
~a0 = u n+1
ðxÞdx, ð78Þ À Á
P 0 u n ðx Þ = u n x + P j ,
ðP À Á ð89Þ
2 x  u j :
~ n ic + P
~ n ði c Þ = u
~aα = un+1 ðxÞ × cos 2πα dx, ð79Þ
P 0 P
ðP The periods are defined as
 x
~b = 2 u n+1
ðxÞ × sin 2πα dx, ð80Þ 2π
α
P P Pj = h × , ð90Þ
0 θj
ðP
1  j = 2π : ð91Þ
~cα = un+1 ðxÞ × e−i2παðx/PÞ dx: ð81Þ P
θj
P 0

Proof. By definition, The main analysis step is substituting Equation (88) into
Equation (76). Based on Corollary 6, un+1 ðxÞ can be
u n ðx Þ = u n ðx + P Þ: ð82Þ expressed as a series expansion of Equation (77). In principle,
the coefficients ~a0 , ~aα , ~bα , and ~cα can be obtained from Equa-
Hence, from Theorem 5, tions (78), (79), (80), and (81). However, equivalent formulas
can be written in terms of ic as follows:
un+1 ðxÞ = un+1 ðx + PÞ: ð83Þ
ð P j
2
~a0 = ~n+1 ðic Þdic ,
u ð92Þ
Consequently, un+1 ðxÞ can be expressed as the infinite j
P 0
trigonometric series provided by Equation (77). The series ð P j !
expression is provided by Equations (3.7, 3.10) in [19], and 2 ic
~aα = ~ ðic Þ × cos 2πα
u n+1
dic , ð93Þ
provided by Equation (77). The series coefficients ~a0 , ~aα , ~bα , 
Pj 0
j
P
and ~cα are provided by Equations (3.5, 3.6, 3.9) in [19] and !
ð P j
given by Equations (78)–(81). The condition of C 0 continuity ~b = 2 ic
~ ðic Þ × sin 2πα
u n+1
dic , ð94Þ
is included to assure the series convergence. For more details, α j 0
P j
P
the reader may refer to Theorem 4.3 in [19].
ð P j
1
~n+1 ðic Þ × e−i2παðic /P j Þ dic :

3.3.2. Fourier Analysis of a Monochromatic Solution. Con- ~cα = u ð95Þ

Pj
sider a monochromatic uni with a single period P j and MWN 0
10 Journal of Applied Mathematics

Although un ðxÞ and u ~n ðic Þ refer to the same quantity, eliminated. Hence, the limitations of the ADR
the latter is used in Fourier analysis. The specific advan- developed by [4] are avoided
tage of u~n ðic Þ as defined in Equation (88) is that h is elim-
(ii) The expressions resulting from plugging Equation
inated. The solution behaviour is fully clarified considering
 j g, instead of considering the whole domain (88) into Equation (76) are mightily lengthy.
only ic ∈ f0, P
Hence, this work uses the computer algebra system
i ∈ f0, Ng. MAXIMA [20]
Although un ðxÞ in Equation (87) is monochromatic,
the numerical solution un+1 ðxÞ includes multiple modes (iii) The integrands of Equations (93) and (94) include
as described in Equation (77). In other words, a mono- rational expressions in terms of trigonometric
chromatic initial condition induces high-order modes. functions. This is attributed to the form of the
Assumption 1 will be revisited within the continuous nonlinear weights of WENO-JS. Still, the definite
presentation provided. As implied by MI, zero and high integrals can be calculated up to arbitrary accu-
order modes can be neglected and Equation (77) can be racy using numerical quadrature. The reader could
rewritten as follows: refer to [21] or other standard textbooks for the
details of numerical integration using the trapezoi-
un+1 ðxÞ = u~n+1 ðic Þ ≈ u
~a n+1 ðic Þ dal approximation
! !
x x (iv) The nonlinear nature of the WENO-JS algorithm
= ~a1 cos 2π + ~b1 sin 2π should be well captured by the present analysis.
Pj Pj ð96Þ
The rational expressions of ω j are adopted, unlike
1
[6] where ω j were approximated as constants
= 〠 ~cα ei2παðx/P j Þ :
α=−1,α≠0
(v) Past theoretical studies [5, 22] adopted the assump-
tion of small MWN (θ j ) because Taylor expansion
The spectral parameters are defined based on Fourier
expansion coefficients as follows: was used. Such limitations are avoided, and the
NSM is valid for arbitrary values of θ j
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~a1 2 + ~b1
2
2j~c j ð97Þ
3.3.3. Mode Isolation Error. The error introduced by
Gj = = 1 , Assumption 1 will be quantified. Consider the well-known
aj aj
Parseval identity:
 
a1 ð P j
Φ j = tan−1 = arg ð~c1 Þ: ð98Þ 1 n+1 2 1 À Á2
b1 ~ ði c Þ
u = ~n+1 ðic Þ dic
u

Pj 
Pj 0
Using G j and Φ j , Equation (96) is rewritten as
~a0 1 ∞
2  ∞
+ 〠 ~aα 2 + ~bα = 〠 j~cα j2 :
2
! =
4 2 α=1 α=−∞
x
u n+1
~
ðx Þ = u n+1
~a
ði c Þ ≈ u n+1
ðic Þ = G j a j sin 2πα + Φ j : ð100Þ
Pj
ð99Þ It is straightforward to verify that

The parameter ϕ j describes a shift in the waveform of un ð P j

~a 2 + ~b1
2
1 n+1 2 1 À Á2
ðxÞ. This shift does not affect neither ~c1 nor ~c−1 [19]. In addi- ~ ð ic Þ
u = ~a n+1 ðic Þ dic = 1
u = j~c−1 j2 + j~c1 j2 :
j
P j
P 0 2
tion, ~c1 and ~c−1 depend linearly on a j as illustrated by Equation
ð101Þ
(75). Consequently, the values a j = 1 and ϕ j = 0 can be
adopted without loss of generality. Hence, G j and Φ j can be Hence, the percentage error normalized with respect to
expressed using Equations (54) and (55), respectively. 2
un+1 ðic Þk is defined as
k~
The analysis described in Equations (97) and (98) will
be denoted by the Nonlinear Spectral Method, abbreviated 2 2
as NSM. Upon introduction of MI, the focus is limited to À Á u~n+1 ðic Þ − u ~n+1
a ði c Þ
E θ j , λ = 100 × 2
calculating only the principal mode spectral parameters, ~ ðic Þ
u n+1

namely, ~a1 and ~b1 . Still, it should be emphasized that the À Á  

aα 2 + ~bα
2
 j ~a0 2 /4 + ð1/2Þ∑∞
P α=2 ~
error introduced by MI does not affect the accuracy of
= 100 × :
NSM. In fact the principal mode parameters ~a1 and ~b1 ~n+1 ðic Þ
u
2

are still recovered exactly. The following features of the

NSM should be noted: ð102Þ

(i) The results of the NSM are valid for any h and Δ The error Eðθ j , λÞ is proportional to the magnitude of zero
similar to VN, and the computational effects are and higher-order modes truncated by u ~a n+1 ðic Þ. The reader is
Journal of Applied Mathematics 11

𝜃j = 0.1𝜋
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

Cj
Cj

Cj
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
ic ic ic

𝜃j = 0.3𝜋
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

Cj
Cj

Cj

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6
ic ic ic

𝜃j = 0.5𝜋
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

Cj
Cj

Cj

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
ic ic ic

𝜃j = 0.9𝜋
1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

Cj
Cj

Cj

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2 0 0.5 1 1.5 2
ic ic ic

Figure 2: Plots of weights used to calculate the numerical flux. The linear weights d0 , d1 , d 2 used for LUW5 are plotted with solid lines. The
nonlinear weights used in WENO-JS are plotted as (~ ωc0 •, ω
~ c1 ♦, ω
~ c2 ◼). Four values of θ j are adopted (0:1π, 0:3π, 0:5π, 0:9π).

reminded that MI is exactly satisfied by LUW5. Hence, a con- WENO-TVDRK2 and WENO-TVDRK3 with slight modifi-
siderable discrepancy may exist between the results of cations. The continuous versions of Equations (24) and (25)
WENO-JS and LUW5 due to Eðθ j , λÞ, even though the spectra are defined as follows:
of both methods may be similar in some cases.
uð1Þ ðxÞ = un ðxÞ + λLc ðun ðxÞÞ,
1 n 1 1   ð103Þ
3.3.4. Extension to High Order Time Integration. The algo- un+1 ðxÞ = u ðxÞ + uð1Þ ðxÞ + λLc uð1Þ ðxÞ :
rithm described for WENO-RK1 can be applied to 2 2 2
12 Journal of Applied Mathematics

WENO-RK1 WENO-TVDRK2 WENO-TVDRK3

10 10 10

1 1 1

0.1 0.1 0.1

𝜀 0.01 𝜀 0.01 𝜀 0.01

0.001 0.001 0.001

0.0001 0.0001 0.0001

1 × 10−5 1 × 10−5 1 × 10−5

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
𝜃j 𝜃j 𝜃j

𝜆 = 0.025
𝜆 = 0.1
𝜆 = 0.4

Figure 3: Eðθ j , λÞ at λ = 0:025,0:1,0:4, for WENO-RK1, WENO-TVDRK2, and WENO-TVDRK3 schemes.

uð1Þ ðxÞ = un ðxÞ + λLc ðun ðxÞÞ, For any numerical time integration method, the Linear
Stability Domain and its boundary are given by
3 n 1 1  
uð2Þ ðxÞ =
u ðxÞ + uð1Þ ðxÞ + λLc uð1Þ ðxÞ , ð104Þ
4 4 4 S t = f̂z : jgð̂z Þj ≤ 1g,
  ð110Þ
1 2 2 ∂S t = f̂z : jgð̂z Þj = 1g:
un+1 ðxÞ = un ðxÞ + uð2Þ ðxÞ + λLc uð2Þ ðxÞ :
3 3 3
The algorithm described in Section 3.3.2 is modified by The Geometric Linear Stability Limit (GLSL) is defined
replacing Equation (76) with Equations (103) and (104) to ana- ~ such that
as the largest λ
lyse WENO-TVDRK2 and WENO-TVDRK3, respectively.
~ ⊆S :
λS ð111Þ
t
3.4. Geometric Stability Analysis. The spectral methods
developed by [5, 6] termed the Geometric Stability Analysis It was illustrated by [5] that the GLSL exists for nei-
(GSA) will be briefly presented. The spectral properties of ther WENO-RK1 nor WENO-RK2, and both were con-
the spatial discretization operator are defined in terms of z cluded to be linearly unstable. However, the grid X given
ðθ j Þ given by by Equation (2) was used in [6] to define a discrete spec-
trum S that consists of the set of eigenvalues given by
À Á L ðu n Þ  
z θ j = − ni : ð105Þ À Á j
ui S = −z θ j : θ j = 2π , j ∈ f0, 1, ⋯, N g : ð112Þ
N
Here, uni is given by Equation (48) because MI is a
cornerstone in GSA [5, 6]. The continuous spectrum S is Hence, Equation (111) should be replaced by
given by
λS ⊆ S t : ð113Þ
È
À Á É
S = −z θ j : θ j ∈ ½0, 2πŠ : ð106Þ
It was illustrated by [6] that Equation (113) can be sat-
isfied by both WENO-RK1 and WENO-RK2. Given a spe-
Generally, numerical time integration methods are cific value of λ, define the intersection point between λS
characterized in terms of ̂z given by and ∂S t to be ̂z ∗ ðλÞ. The nonzero eigenvalue of S that
À Á is closest to the origin −zðθ j = 2πhÞ is equated to ̂z ∗ ðλÞ/λ
̂z = −λz θ j : ð107Þ = −zðθ∗j ðλÞÞ. Hence, the linear stability condition can be
given by
The amplification factor gð̂z Þ for a numerical scheme
is given by θ∗j ðλÞ
h≥ : ð114Þ
2π
un+1
i = gð̂zÞun+1
i : ð108Þ
For LUW5-RK1, the following formula of θ∗j ðλÞ that is
The relation between gð̂z Þ and the spectral parameters valid for small MWN was derived by [6]
in the current work is given by
 
À
À ÁÁ 15λ 1/4
θ∗j ðλÞ = 2 : ð115Þ
gð̂z Þ = g −λz θ j = G j eiΦ j : ð109Þ 8
Journal of Applied Mathematics 13

WENO-RK1
0
1
−0.2
0.9
−0.4
0.8
−0.6

𝛷j
Gj

0.7 −0.8

0.6 −1

0.5 −1.2

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

𝜃j θj

(a) (b)
WENO-TVDRK2
0
1
−0.2

0.9 −0.4

−0.6
𝛷j
Gj

0.8
−0.8
0.7
−1
0.6
−1.2

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

θj θj
(c) (d)
WENO-TVDRK3
0
1
−0.2
0.9
−0.4

0.8 −0.6
Gj

𝛷j

−0.8
0.7
−1
0.6
−1.2

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

𝜃j θj
(e) (f)

Gj = 1 FFT 𝜆 = 0.025 𝛷j = −0.025 𝜃j 𝜆 = 0.4

NSM 𝜆 = 0.025 𝜆 = 0.1 𝛷j = −0.1 𝜃j FFT 𝜆 = 0.025
𝜆 = 0.1 𝜆 = 0.4 𝛷j = −0.4 𝜃j 𝜆 = 0.1
𝜆 = 0.4
𝜆 = 0.4 NSM 𝜆 = 0.025
𝜆 = 0.1

Figure 4: Various solutions spectra are plotted for the exact solution (- -), NSM (—), and FFT numerical experiments (⋄, ∘, ◻). The
corresponding curves for λ = 0:025,0:1, and 0:4 are plotted in black, red, and blue colours, respectively.
14 Journal of Applied Mathematics

0
–0.2
–0.4 0 × 100
𝜆 = 0.001
–1 × 10–4
Im (-z (𝜃j))

–0.6

Im (-z (𝜃j))
–0.8 –2 × 10–4
–3 × 10–4

⌒
–1
–1.2 –4 × 10–4
–1.4 –5 × 10–4
–6 × 10–4

–4 × 10–7

–3.5 × 10–7

–3 × 10–7

–2.5 × 10–7

–2 × 10–7

–1.5 × 10–7

–1 × 10–7

–5 × 10–8

0
–2 –1.5 –1 –0.5 0
Re (-z (𝜃j))

WENO-JS, 𝜃j ⋲ (0, 𝜋)
LUW5, 𝜃j ⋲ (0, 𝜋)
⌒
Re (z (𝜃j))
RK1 𝜕St
WENO-JS, 𝜃j ⋲ (0, 0.75𝜋) LUW5, 𝜃j > 0.135 × 𝜋
WENO-JS, 𝜃j > 0.075𝜋 RK1 𝜕St
LUW5, 𝜃j ∈ (0, 0.135𝜋)

(a) (b)

Figure 5: The GSA of WENO-RK1 and LUW5-RK1.

The function zðθ j Þ plays a key role in the GSA. How- neglected. Hence, the magnification factor and the phase
ever, linearised procedures were used to calculate zðθ j Þ in shift are calculated as
[5, 6]. For instance, the GSA of [6] was mainly based on
Equation (50) which describes LUW5 instead of WENO-
2i n+1
JS. The GSA provided by [5] was mainly based on Taylor Gj = U ,
expansion and was limited to small θ j . The proposed NSM N j
  ð118Þ
can be used to calculate zðθ j Þ exactly as follows. Define ~aL1 2i n+1
Φ j = arg U :
and ~b as
L N j
1

ð P j !
2 c ic The spectral parameters are calculated using the FFT
~aL1 = ~n ðic ÞÞ × cos
L ðu 2π dic ,
j
P 0
j
P of the solution obtained using a specific grid size N.
! ð116Þ Hence, the procedure will be sensitive to the machine’s
ð P j
~bL 2 ic precision. A unity domain size L = 1:0 was adopted for
= Lc ðu
~n ðic ÞÞ × sin 2πα dic :
1 
Pj 0
j
P this paper’s numerical experiments.

4. Results
Hence, zðθ j Þ is given by
4.1. Significance of the Nonlinear Analysis
À Á
z θ j = −~b1 − i~aL1 :
L
ð117Þ 4.1.1. Linear and Nonlinear Weights. To clarify the signifi-
cance of the NSM, the weights of both LUW5 (d j ) and
3.5. Analysis Based on Numerical Experiments. The results WENO-JS (ω j ) were plotted. The curves are provided in
provided above can be further verified using numerical Figure 2 for four values of θ j = 0:1π, 0:3π, 0:5π, and 0:9π,
experiments. The methodology follows closely the proce-
dure outlined by [4]. Also, complete derivations are pro- covering small and large MWN. A low MWN corresponds
vided in Appendix A. For given values of N, L, and λ, a to smooth solutions at which the performance of LUW5
grid xi is adopted based on Equation (2). A value of θ j is and WENO-JS should be close. Hence, for θ j = 0:1π, the dif-
selected, where j ∈ f0, 1, ⋯, N/2g. A monochromatic ference between d j and ω j is small. As the value of MWN
intermediate condition uni = un ðxi Þ is calculated based on increases the solution smoothness decreases and the influ-
Equation (87). The solution un+1 is obtained by advanc- ence of nonlinear WENO-JS weights should be stronger.
i
For instance, at θ j = 0:5π, we get P j = 4 based on Equation
ing one time step using one of Equations (23), (24),
and (25). Finally, Fast Fourier transform (FFT) procedure (91), and a complete wave is located within 4 grid spacings
h. Consequently, for large θ j , the difference between d j and
is applied to un+1 to get the spectrum U n+1 w = Fðui Þ,
n+1
i
where w ∈ f0, 1, ⋯, N/2g. According to the MI assump- ω j cannot be ignored. This can be easily observed in
tion, only U n+1
j is considered and other modes are Figure 2 for θ j = 0:3π, 0:5π, 0:9π.
Journal of Applied Mathematics 15

𝜆 = 0.3
1.05 0
1 –0.2
0.95 –0.4
0.9 –0.6
0.85

Фj
Gj

–0.8
0.8 –1
0.75 –1.2
0.7 –1.4
0.65
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
𝜃j 𝜃j
(a) (b)
𝜆 = 0.7
1 0
0.8 –0.5
0.6 –1

Фj
Gj

0.4 –1.5
0.2 –2

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

𝜃j 𝜃j
(c) (d)
𝜆 = 0.9
1 0
–0.5
0.8
–1
0.6
Фj
Gj

–1.5
0.4 –2
0.2 –2.5

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

𝜃j 𝜃j
(e) (f)
𝜆 = 1.1
1 0
0.8 –0.5
–1
0.6 –1.5
Gj

Фj

0.4 –2
–2.5
0.2 –3

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

𝜃j 𝜃j
(g) (h)

Figure 6: Solutions spectra based on TVDRK3 time integration are plotted for the exact solution - -, WENO-JS —, and LUW5 ―. The
results of WENO-JS and LUW5 are provided using NSM and VN, respectively. The corresponding curves for λ = 0:3,0:7,0:9, and 1:1 are
plotted in magenta, black, red, and blue colours, respectively.

4.1.2. Mode Isolation Error Results. The error induced by
Assumption 1 for WENO-RK1, WENO-TVDRK2, and
WENO-TVDRK3 is plotted in Figure 3 for three values of
λ = 0:025, 0:1, and 0:4. This range of λ covers both small
and moderate values. Several features can be observed:
The value of Eðθ j , λÞ vanishes as θ j ⟶ 0. This corre-
sponds to the trivial case of a constant uni , where Lðuni Þ = 0.
(i) On the other hand, Eðθ j , λÞ also vanishes as θ j
⟶ π. As already illustrated, this corresponds to
the maximum θ j sampled by the grid. Hence, the
solution is described exactly using a single mode
and the MI is exactly satisfied
(ii) The value of Eðθ j , λÞ is almost negligible for the
interval θ j < π/2. For this range of θ j , modes extend
over more than 5 grid points and the solution will
be relatively smooth. Hence, WENO-JS accuracy
should be relatively higher and MI is well satisfied.
This is the opposite for θ j > π/2, where modes
extend over less than 5 grid points and uni exhibits
strong variation
16 Journal of Applied Mathematics

LUW5-TVDRK3

1 1

0.8
1 × 10–5
0.6
u

1 × 10–10

Gj
0.4

0.2 1 × 10–15

0
1 × 10–20
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j
(a) (b)
WENO-TVDRK3
1

0.8 0.1
0.01
0.6 0.001
u

Gj 0.0001
0.4
1 × 10–5
0.2 1 × 10–6
1 × 10–7
0
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j
(c) (d)
Exact 𝜆 = 0.9 Initial 𝜆 = 0.9
𝜆 = 0.3 𝜆 = 1.1 𝜆 = 0.3 𝜆 = 1.1
𝜆 = 0.7 𝜆 = 0.7

Figure 7: Spectra of the exact (initial) solution - - and numerical solutions — using TVDRK3 time integration based on Equation (119). The
numerical parameters are provided in the text. The corresponding curves for λ = 0:3,0:7,0:9, and 1:1 are plotted in magenta, black, red, and
blue colours, respectively.

(iii) As λ increases, Eðθ j , λÞ increases and MI becomes
less satisfied. This is expected since the numerical
accuracy is generally inversely proportional to λ
(iv) The value of Eðθ j , λÞ is generally similar and less
than 1% for the three methods WENO-RK1,
WENO-TVDRK2, and WENO-TVDRK3 at λ =
0:025,0:1. However, at λ = 0:4, the value of Eðθ j , λÞ
is considerably higher for WENO-RK1 than both
WENO-TVDRK2 and WENO-TVDRK3
Based on the above discussion, it is expected that the
discrepancy between WENO-JS and LUW5 results should
be high for solutions including modes of θ j > π/2 and
using λ ≥ 0:4.
4.2. Numerical Solution Spectra. The spectra of WENO-JS
are plotted for λ = 0:025,0:1, and 0:4, and for the three
time integration methods RK1, TVDRK2, and TVDRK3
in Figure 4. Hence, the effects of both λ and the time inte-
gration methods are clarified. For each value of λ, three
spectra are provided based on the following: the NSM
explained in Section 3.3.2, FFT numerical experiments
explained in Section 3.5, and the exact solution spectrum
given by Equations (40) and (41). The results of both
FFT and NSM are almost identical, except for some minor
discrepancies. Hence, the validity of NSM at the full range
of θ j ∈ ð0, πŠ is illustrated. The deviation between the spec-
tra of WENO-JS and the exact solution is proportional to
λ. This inverse proportionality between accuracy and λ is
expected. Most of the spectra of WENO-JS are close to
the spectra of the exact solution at θ j ∈ ð0, 1Þ. However,
G j of WENO-RK1 is considerably high for λ = 0:4 at θ j
∈ ð0, 1Þ, as shown in Figure 4(a).
Due to the current limited scope, the GSA will be applied
to RK1 only. The results of the GSA of WENO-RK1 and
LUW5-RK1 are shown in Figure 5. In Figure 5(a), the red
unit circle represents ∂S t of RK1 [13], and the blue and
black curves represent S of LUW5 and WENO-JS,
respectively. Both spectra are calculated exactly using Equa-
tions (50) and (117) for LUW5 and WENO-JS, respectively.
The full range of the MWN θ j ∈ ð0, πŠ is included in
Figure 5(a). For small θ j in the neighbourhood of the imag-
inary axis, WENO-JS and LUW5 have very close S. How-
ever, S of LUW5 is closer to the imaginary axis than that
Journal of Applied Mathematics 17

LUW5-TV DRK3
1
1
0.5
1 × 10–5
0
u

1 × 10–10

Gj
–0.5 1 × 10–15
–1 1 × 10–20
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j

(a) (b)
WENO-TV DRK3
1
0.5 1
0 1× 10–5
u

1 × 10–10

Gj
–0.5
1 × 10–15
–1 1 × 10–20
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j
(c) (d)
Exact 𝜆 = 0.9 Initial 𝜆 = 0.9
𝜆 = 0.3 𝜆 = 1.1 𝜆 = 0.3 𝜆 = 1.1
𝜆 = 0.7 𝜆 = 0.7

Figure 8: Spectra of the exact (initial) solution - - and numerical solutions — using TVDRK3 time integration based on Equation (120). The
numerical parameters are provided in the text. The corresponding curves for λ = 0:3,0:7,0:9, and 1:1 are plotted in magenta, black, red, and
blue colours, respectively.

of WENO-JS. Also, a strong discrepancy can be observed for
large θ j away from the imaginary axis.
To explain how Equation (113) is satisfied, ̂z adopting
λ = 0:001 is plotted for both WENO-JS and LUW5 in
Figure 5(b). Similar results can be obtained for other
small values of λ. The region of Figure 5(b) is very close
to the imaginary axis and corresponds to low MWN.
Two very close values of θ∗j ð0:001Þ for LUW5-RK1 are
obtained based on Figure 5(b) and Equation (115) to be
≈0:135π and 0:132π, respectively. Hence, the relation
with the GSA of [6] is illustrated. Also, it is shown in
Figure 5(b) that θ∗j ð0:001Þ ≈ 0:075π for WENO-RK1.
Hence, the intersection between λS and S t is higher for
WENO-RK1 than that for LUW5-RK1 at λ = 0:001. Also,
the grid spacing is restricted to h ≥ 0:075/2 and h ≥
0:135/2 for WENO-RK1 and LUW5-RK1, respectively.
Hence, WENO-RK1 admits a wider range of h compared
to LUW5-RK1.
In practice, WENO-RK1, LUW5-RK1, WENO-RK2, and
LUW5-RK2 require prohibitively small Δ [6]. Hence, the
next results will focus on WENO-TVDRK3 because it is
considered the most popular combination [23].
The spectra are plotted for WENO-TVDRK3 and
LUW5-TVDRK3 using moderate and high values of λ in
Figure 6. Generally, the magnitudes of G j and Φ j of
WENO-TVDRK3 are considerably lower than those of
LUW5-TVDRK3 for θ j ∈ ðπ/2, πÞ. The low values of G j
and jΦ j j correspond to more dissipation and deceleration
of high MWN modes. This is expected because WENO-
JS is designed to filter out high MWN modes. Specifically,
the weighting algorithm of WENO-JS is focused on the
solution behaviour within the five-point stencil. The width
of this stencil is 4h, and it is covered by the MWN modes
of θ j ≥ π/2. Hence, the effect of WENO-JS should be
stronger on the MWN modes of θ j ≥ π/2. The spectra of
WENO-TVDRK3 and LUW5-TVDRK3 show somehow
similar trends for λ = 0:3,0:7. However, the spectra differ
considerably at higher λ. Specifically, two small peaks
can be observed for WENO-TVDRK3 at λ = 0:9 in the
range θ j ∈ ð1:5,2:5Þ as shown in Figure 6(e). These two
peaks merge into a single spike at λ = 1:1 located at the
high MWN θ j ≈ 2:1 as shown in Figure 6(g). This spike
contradicts the main requirement of WENO-JS, which is
damping high MWN modes. In addition, the effect of this
spike will extend to other modes due to the high Eðθ j , λÞ
at θ j ≈ 2:1 (Figure 3). Hence, the convergence of WENO-
TVDRK3 solutions at high λ will deteriorate.
4.3. Numerical Experiments. The significance of the NSM
results is clarified using numerical experiments. All the
numerical experiments are done using ε = 1 × 10−40 , domain
size L = 1, grid spacing h = 1/128, and λ = 0:3,0:7,0:9,1:1.
The solutions are provided at a relatively large time t = 20
18 Journal of Applied Mathematics

LUW5-TVDRK3
1
1
0.5
1 × 10–5
0
u

1 × 10–10

Gj
–0.5 1 × 10–15
–1 1 × 10–20
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j

(a) (b)
WENO-TVDRK3
1

0.5 1

0 1 × 10–5
u

1 × 10–10

Gj
–0.5
1 × 10–15
–1 1 × 10–20
0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 2 2.5 3
x 𝜃j
(c) (d)
Exact 𝜆 = 0.9 Initial 𝜆 = 0.9
𝜆 = 0.3 𝜆 = 1.1 𝜆 = 0.3 𝜆 = 1.1
𝜆 = 0.7 𝜆 = 0.7

Figure 9: Spectra of the exact (initial) solution - - and numerical solutions — using TVDRK3 time integration based on Equation (121). The
numerical parameters are provided in the text. The corresponding curves for λ = 0:3,0:7,0:9, and 1:1 are plotted in magenta, black, red, and
blue colours, respectively.

to elucidate any long-term effects. Three experiments are
performed adopting the initial conditions uðx, t = 0Þ = FðxÞ
given by
( three lobes in the range θ j ∈ ½0, 1Š as shown in Figure 7(b).
1, if x ∈ ð0:4,0:6Þ,
F ðx Þ =
0, if x ∈ ½0,0:4Š ∪ ½0:6,1Š,
F ðxÞ = sin ð2πxÞ,
1 ∈ ½0, 1Š for LUW5-TVDRK3 spectra presented in Figure 6.
F ðxÞ = sin ð2πxÞ + sin ð90πxÞ:
100
The numerical and exact solutions based on Equations
(119), (120), and (121) are provided in Figures 7–9, respec-
tively. The results of LUW5-TVDRK3 and WENO-
TVDRK3 are provided in the upper and lower parts of
Figures 7–9, respectively. The solutions uðx, t = 20Þ and the
magnification factors are provided in the left and right parts
of Figures 7–9, respectively. Based on Equation (27), the
solutions plotted at t = 20 correspond to traversing outside
and inside the domain and returning to the original position
20 times. Hence, the exact and the initial conditions are
coincident and plotted using the same dashed line.
The initial condition of Equation (119) includes a sharp
discontinuity and is selected to clarify the performance of
WENO-JS for nonsmooth solutions. For the exact solution,
G j approximates a Sinc function with multiple lobes. This
behaviour is expected and is explained in [14] and other
standard textbooks.
Considering LUW5-TVDRK3, G j approximates the first
However, G j almost vanishes for high MWN in the range θ j
ð119Þ
∈ ð1, πŠ and the spectrum is almost truncated. This truncation
leads to an oscillatory behaviour and Gibbs oscillations can be
observed in the solution plotted in Figure 7(a). This truncated
ð120Þ
spectrum is expected since G j was close to 1 in the range θ j
ð121Þ The behaviours of LUW5-TVDRK3 solutions are almost sim-
ilar for λ = 0:3,0:7,0:9,1:1.
For WENO-TVDRK3, the solution behaviour is strongly
affected by λ. Considering low and moderate CN (λ = 0:3,0:7),
multiple lobes are included in G j and cover the whole MWN
range as shown in Figure 7(d). This distribution of G j
approximates the product of two Sinc functions that
describes a relatively smeared but non-oscillatory solution.
Hence, WENO-TVDRK3 solutions at λ = 0:3,0:7 approxi-
mate the exact solution avoiding Gibbs oscillations as shown
in Figure 7(c). Inverse proportionality between G j and θ j is
observed in Figure 7(c). However, G j does not vanish at high
MWN in the range θ j ∈ ð1, πŠ. This behaviour is attributed to
the nonlinear nature of WENO-JS, where low MWN modes
contribute to high MWN modes as illustrated in Section
3.3.2 and quantified by Eðθ j , λÞ.
Journal of Applied Mathematics 19

The distribution of G j for WENO-TVDRK3 considerably
deteriorates at the high CN (λ = 1:1). The Sinc distribution
disappears and a strong peak is observed at θ j ≈ 2:1 as shown
in Figure 7(d). This peak is expected and corresponds to the
spike that is clarified in Figure 6(g). As already mentioned in
Section 4.2, this spike has an adverse effect on the solution.
The spike tends to magnify high MWN modes leading to non-
smooth and irregular solutions. On the other hand, the
WENO-JS mechanism tends to smooth the solution and
remove the irregularity. As a result, the solution of WENO-
TVDRK3 for λ = 1:1 shown in Figure 7(c) is overdamped
and infested by small yet distinguishable spurious oscillations.
The detection of this overdamping phenomenon is gen-
erally difficult using numerical experiments. For instance,
consider the solutions based on Equation (120) plotted in
Figure 8. This initial condition, given by Equation (120),
includes only one low MWN mode (θ j = π/64). Both
WENO-TVDRK3 and UW5-TVDRK3 solutions are almost
indistinguishable, and the effect of λ disappears. The reason
for this is that overdamping occurs due to the influence of
high MWN modes. A numerical experiment was done by
[5] at λ = 1:4 using a low MWN initial condition. Hence,
overdamping was hidden, and it was erroneously concluded
that the performance of WENO-TVDRK3 is generally satis-
factory at this high CN.
Finally, the sensitivity of WENO-TVDRK3 to high wave
numbers is illustrated by considering the initial condition of
Equation (121). This condition consists of a high amplitude
low MWN mode (θ j = π/64), and a low amplitude high
MWN mode (θ j ≈ 2:2). The amplitude of the high MWN
mode (1/100) is almost negligible compared to that of the
low MWN mode. The influence of θ j ≈ 2:2 mode is negligi-
ble for UW5-TVDRK3 solution plotted in Figure 9(a), and
it is effectively filtered in UW5-TVDRK3 spectra plotted in
Figure 9(b). However, the solution of WENO-TVDRK3 suf-
fers overdamping at λ = 1:1 as illustrated in Figure 9(c). A
The NSM can reproduce the spectra of numerical exper-
iments while being independent of any grid parameters. The
effects of both the CN and the numerical time integration
methods are elucidated. At low CN, the performance of the
combination of WENO-JS with the forward Euler time inte-
gration method is better than that of LUW5. At high CN, the
combination of WENO-JS with the third-order variation
diminishing Runge-Kutta method suffers extra damping
compared to LUW5. This peculiar phenomenon was over-
looked in past studies because it occurs due to a spike
located at the high MWN spectra, and its detection via
numerical experiments is difficult. The results of the current
study should be extended to include other time integration
algorithms and nonlinear discretizations.
Appendix
A Discrete Fourier Transform
Consider the periodic function uðxÞ of period L ∈ ℝ. Samples of
uðxÞ are taken at N ∈ ℕ points xi , where i ∈ f0, 1, ⋯, N − 1g.
The samples are obtained at equally spaced points xi .
L
xi = i = ih: ðA:1Þ
N
Hence, the sequence ui = uðxi Þ is obtained. The sequence
can be written as the following sum [14]:
1 N−1 1 N−1
ui = 〠 U w eið2wπ/N Þi = 〠 U w eið2wπ/LÞxi : ðA:2Þ
N w=0 N w=0
pffiffiffiffiffiffi
Here, i = −1. The coefficients U w are termed as the dis-
crete Fourier transform or DFT as an abbreviation. The operator
F is used to define DFT. This is defined as
N−1 N−1

small yet distinguishable peak located at θ j ≈ 2 can be U w = F ðui Þ = 〠 ui e−ið2wπ/N Þi = 〠 ui e−ið2wπ/LÞxi : ðA:3Þ
i=0 i=0
observed in the WENO-TVDRK3 spectrum for λ = 1:1 in
Figure 9(d). This peak corresponds to the spike that is clar- Similarly, the inverse operation of DFT can be defined using
ified in Figure 6(g). the operator notation:

5. Conclusions ui = F −1 ðU w Þ: ðA:4Þ
The nonlinear Weighted Essentially Nonoscillatory method
For a better presentation, the wave number is defined as
developed by Jiang and Shu (WENO-JS) is analysed to clar-
ify its spectral properties. An exact Nonlinear Spectral 2wπ
Method (NSM) is developed based on the Fourier expansion kw = , w ∈ f0, 1, ⋯, N − 1g: ðA:5Þ
of a continuous presentation of WENO-JS solutions. The L
nonlinear rational formula of WENO-JS weights is fully And the modified wave number (MWN) is defined as
included. Unlike other theoretical methods, the NSM is valid
for arbitrary modified wave numbers (MWN). The mode 2πw
isolation (MI) assumption, which was used extensively in θw = kw h = : ðA:6Þ
N
the literature, is quantified for the first time, and its impact
on the performance of WENO-JS is explained. The MI error Hence, Equations (A.2) and (A.3) can be rewritten as
increases for high MWN and moderate Courant numbers
(CN), and the discrepancy between the results of WENO- 1 N−1 1 N−1
JS and those of the fifth-order Linear Upwind (LUW5) dis- ui = F −1 ðU w Þ = 〠 U w eikw xi = 〠 U w eiθw i : ðA:7Þ
cretization cannot be neglected. N w=0 N w=0
20 Journal of Applied Mathematics

N−1 N−1 verified by substitution by Equations (A.12) and (A.14)

U w = F ðui Þ = 〠 ui e−ikw xi = 〠 ui e−iθw i : ðA:8Þ into Equation (A.7).
i=0 i=0
 
1 N−1 ið2wπi/N Þ A N iϕ ið2jπi/N Þ N −iϕ ið2ðN−jÞπi/N Þ
〠U e = e e − e e
B DFT of a Monochromatic Wave N m=0 w N 2i 2i
A iϕ ið2jπi/N Þ A −iϕ −ið2jπi/N Þ
Consider the real monochromatic data series, with an ampli- = e e − e e
tude A, a wave number k j , and an arbitrary phase shift ϕ. 2i  2i
2jπi À Á
À Á À Á = A × sin + ϕ = A × sin θ j i + ϕ :
uðxi Þ = ui = A × sin k j xi + ϕ = A × sin θ j i + ϕ : ðA:9Þ N
ðA:15Þ
The discrete Fourier transform is obtained as
Finally, referring to Equations (A.12) and (A.14), it is
N−1
ÁÀ deduced that
U w = 〠 A × sin k j xi + ϕ e−ikw xi
i=0 N
A 
N−1 U j = U N−j = A ,
2
= 〠 eiðk j xi +ϕÞ − e−iðk j xi +ϕÞ e−ikw xi À Á π  À Á
ðA:16Þ
2i π
i=0 arg U j = ϕ − = − − ϕ = − arg U N−j :
2 2
A N−1 iϕ iðk j −kw Þxi −iϕ −iðk j +kw Þxi
= 〠e e −e e ,
2i i=0 The conclusions are summarized in the following
points:
A N−1 iϕ ið j−wÞð2πi/N Þ −iϕ −ið j+wÞð2πi/N Þ
Uw = 〠e e −e e :
2i i=0 (i) The amplitude of mode j is recovered from the DFT
coefficient U j as A = jð2i/NÞU j j
ðA:10Þ
(ii) The phase shift of mode j is recovered from the
Consider the case w = j: DFT coefficient U j as ϕ = arg ðð2i/NÞU j Þ

A N−1À iϕ −iϕ i Á (iii) The range of θw as defined in Equation (A.6) is θw

Uj = 〠 e −e r : ðA:11Þ ∈ ½0, 2πŠ. However, the effective range of θw is
2i i=0
reduced to θw ∈ ½0, πŠ due to the property of Hermi-
tian Symmetry, as illustrated by Equations (A.16)
Here, r = e−ið2jÞð2π/NÞ . The terms e±iϕ are constants w.r.t. and (A.17)
the summation. Using the geometric series sum formula
N−1 i
∑i=0 r = ðr N − 1Þ/ðr − 1Þ, it is concluded that ∑i=0 N−1 i
r = (iv) The results above were derived for a monochro-
−ið2jÞ2π −ið2jÞ2π matic signal. However, due to the linearity of DFT
ðe − 1Þ/ðr − 1Þ = 0 since e = 1. Hence,
[14], the results still hold for any periodic signal.
N iϕ Consider the general case:
Uj = A e : ðA:12Þ
2i
N−1   N−1  
Consider the case w = N − j: uðxi Þ = ui = 〠 A j sin k j xi + ϕ j = 〠 A j sin θ j i + ϕ j :
j=0 j=0
A N−1 ðA:17Þ
U N−j = 〠 eiϕ eið j−N+jÞð2πi/N Þ − e−iϕ e−ið j+N−jÞð2πi/N Þ
2i i=0
ðA:13Þ The DFT can be obtained as shown in Equation (A.3).
A N−1
= 〠 eiϕ~r i − e−iϕ e−i2πi : Hence, the amplitude and phase shift of mode j can be
2i i=0 recovered back from the DFT as follows:

Similarly, ~r is defined as ~r = eið2j−NÞð2π/NÞ . Following the 2i

same results of the geometric series, it is concluded that Aj = U ,
N j
~r vanishes. Hence,
N−1 i
∑i=0   ðA:18Þ
2i
ϕ j = arg Uj :
N −iϕ N
U N−j = −A e : ðA:14Þ
2i
As mentioned in [6], N can be assumed even without
Following the same steps, it can be easily deduced loss of generality. Based on the above discussion, a discrete
that that U n = 0 for n∈fj, N − jg. These results can be mesh yields only a finite number of distinct modes. These
Journal of Applied Mathematics
modes can be represented using k j = 2jπ/L, or θ j = 2jπ/N
where j ∈ f0, 1, ⋯, N/2g.
Data Availability
Data will be provided upon reasonable request.
Conflicts of Interest
The author declares no conflicts of interest. The research
was performed as a part of the employment of the author
at Nile University.
Acknowledgments
The author is grateful for the administrative support and
encouragement provided by Nile University. Also, the
author would like to thank his mentors and colleagues for
their suggestions during the manuscript preparation.
References
[1] C.-W. Shu, “Essentially non-oscillatory and weighted essen-
tially non-oscillatory schemes for hyperbolic conservation
laws,” in Advanced Numerical Approximation of Nonlinear
Hyperbolic Equations: Lectures Given at the 2nd Session of the
Centro Internazionale Matematico Estivo (C.I.M.E.) held in
Cetraro, Italy, June 23–28, 1997, A. Quarteroni, Ed., pp. 325–
432, Springer, Berlin, Heidelberg, 1998.
[2] G.-S. Jiang and C.-W. Shu, “Efficient implementation of
weighted ENO schemes,” Journal of Computational Physics,
vol. 126, no. 1, pp. 202–228, 1996.
[3] M. L. Bittencourt, N. A. Dumont, and J. S. Hesthaven, Eds.,
“Spectral and high order methods for partial differential equa-
tions,” in ICOSAHOM 2016: Selected Papers from the ICOSA-
HOM Conference, Rio de Janeiro, Brazil, 2016volume 119 of
Lecture Notes in Computational Science and Engineering.
Springer International Publishing, Cham, 2017.
[4] S. Pirozzoli, “On the spectral properties of shock-capturing
schemes,” Journal of Computational Physics, vol. 219, no. 2,
pp. 489–497, 2006.
[5] R. Wang and R. J. Spiteri, “Linear instability of the fifth-order
WENO method,” SIAM Journal on Numerical Analysis,
vol. 45, no. 5, pp. 1871–1901, 2007.
[6] M. Motamed, C. B. Macdonald, and S. J. Ruuth, “On the linear
stability of the fifth-order WENO discretization,” Journal of
Scientific Computing, vol. 47, no. 2, pp. 127–149, 2011.
[7] D. Fauconnier and E. Dick, “On the spectral and conservation
properties of nonlinear discretization operators,” Journal of
Computational Physics, vol. 230, no. 12, pp. 4488–4518, 2011.
[8] G. Zhao, M. Sun, A. Memmolo, and S. Pirozzoli, “A general
framework for the evaluation of shock-capturing schemes,”
Journal of Computational Physics, vol. 376, pp. 924–936, 2019.
[9] D. Fauconnier and E. Dick, “Spectral analysis of nonlinear
finite difference discretizations,” Journal of Computational
and Applied Mathematics, vol. 246, pp. 113–121, 2013.
[10] F. Jia, Z. Gao, and W. S. Don, “A spectral study on the dissipa-
tion and dispersion of the WENO schemes,” Journal of Scien-
tific Computing, vol. 63, no. 1, pp. 49–77, 2015.
[11] J. S. Hesthaven, S. Gottlieb, and D. Gottlieb, “Spectral methods
for time-dependent problems,” in Cambridge Monographs on
21
Applied and Computational Mathematics, Cambridge Univer-
sity Press, Cambridge, 2009.
[12] A. K. Henrick, T. D. Aslam, and J. M. Powers, “Mapped
weighted essentially non-oscillatory schemes: achieving opti-
mal order near critical points,” Journal of Computational Phys-
ics, vol. 207, no. 2, pp. 542–567, 2005.
[13] J. C. Butcher, Numerical Methods for Ordinary Differential
Equations, John Wiley & Sons, Ltd, Chichester, England, sec-
ond edition edition, 2008.
[14] I. Amidror, “Mastering the discrete Fourier transform in one,
two or several dimensions: pitfalls and artifacts,” in Volume
43 of Computational Imaging and Vision, Springer, London,
2013.
[15] R. J. LeVeque, Finite Difference Methods for Ordinary and Par-
tial Differential Equations, Society for Industrial and Applied
Mathematics, Philadelphia, first edition edition, 2007.
[16] J. W. Thomas, “Numerical partial differential equations: finite
difference methods,” in Texts in Applied Mathematics,
Springer, New York, 2013.
[17] K. W. Morton and D. F. Mayers, Numerical Solution of Partial
Differential Equations: An Introduction, Cambridge University
Press, New York, second edition edition, 2012.
[18] D. A. Cox, J. B. Little, and D. O’Shea, Using Algebraic Geome-
try. Number 185 in Graduate Texts in Mathematics, Springer,
New York, second edition edition, 2006.
[19] R. J. Beerends, H. G. ter Morsche, J. C. van den Berg, and E. M.
van de Vrie, Fourier and Laplace Transforms, Cambridge Uni-
versity Press, New York, first edition edition, 2012.
[20] Maxima, Maxima, a Computer Algebra System. Version 5.46.0,
2022.
[21] E. Kreyszig, H. Kreyszig, and E. J. Norminton, Advanced Engi-
neering Mathematics, John Wiley, Hoboken, NJ, tenth edition
edition, 2011.
[22] T. Kasem, “1 error analysis of nonlinear weno schemes using
modified equation,” in Recent Advances in Engineering Math-
ematics and Physics, M. H. Farouk and M. A. Hassanein, Eds.,
pp. 33–46, Springer International Publishing, Cham, 2020.
[23] C.-W. Shu, “Essentially non-oscillatory and weighted essen-
tially non-oscillatory schemes,” Acta Numerica, vol. 29,
pp. 701–762, 2020.