Advances in Applied Mathematics and Mechanics DOI: 10.4208/aamm.

OA-2019-0124
Adv. Appl. Math. Mech., Vol. 12, No. 1, pp. 1-29 February 2020

High-Order Accurate Entropy Stable Finite Difference

Schemes for One- and Two-Dimensional Special
Relativistic Hydrodynamics
Junming Duan1 and Huazhong Tang1,2,∗
1 Center for Applied Physics and Technology, HEDPS, and LMAM, School of
Mathematical Sciences, Peking University, Beijing 100871, China
2 School of Mathematics and Computational Science, Xiangtan University, Xiangtan,

Hunan 411105, China

Received 27 April 2019; Accepted (in revised version) 12 July 2019

Abstract. This paper develops the high-order accurate entropy stable finite difference
schemes for one- and two-dimensional special relativistic hydrodynamic equations.
The schemes are built on the entropy conservative flux and the weighted essentially
non-oscillatory (WENO) technique as well as explicit Runge-Kutta time discretiza-
tion. The key is to technically construct the affordable entropy conservative flux of
the semi-discrete second-order accurate entropy conservative schemes satisfying the
semi-discrete entropy equality for the found convex entropy pair. As soon as the en-
tropy conservative flux is derived, the dissipation term can be added to give the semi-
discrete entropy stable schemes satisfying the semi-discrete entropy inequality with
the given convex entropy function. The WENO reconstruction for the scaled entropy
variables and the high-order explicit Runge-Kutta time discretization are implemented
to obtain the fully-discrete high-order entropy stable schemes. Several numerical tests
are conducted to validate the accuracy and the ability to capture discontinuities of our
entropy stable schemes.
AMS subject classifications: 65M10, 78A48
Key words: Entropy conservative scheme, entropy stable scheme, high order accuracy, finite
difference scheme, special relativistic hydrodynamics.

1 Introduction
This paper is concerned with the high-order accurate numerical schemes for the one-
and two-dimensional special relativistic hydrodynamic (RHD) equations, which in the

∗ Corresponding author.
Emails: [email protected] (J. M. Duan), [email protected] (H. Z. Tang)

http://www.global-sci.org/aamm 1 c 2020 Global Science Press

2 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

laboratory frame, can be cast in the divergence form

d
∂U ∂F (U )
+∑ ` = 0, (1.1)
∂t `=1 ∂x`

where U and F` are respectively the conservative vector and the flux vector in the x` -
direction and defined by

U = ( D,m1 , ··· ,md ,E)T , (1.2a)

F` = ( Du` ,m1 u` + pδ1,` , ··· ,md u` + pδd,` ,m` )T , ` = 1, ··· ,d, (1.2b)

with the mass density D = ρW, the momentum density m = (m1 , ··· ,md )T = DhWu, and
the energy density E = DhW − p. Here, d = 1 or 2, ρ, p and u = (u1 , ··· ,ud )T denote the
rest-mass
q density, the kinetic pressure, and the fluid velocity, respectively. Moreover,
W = 1/ 1 −(u21 +···+ u2d ) is the Lorentz factor and h is the specific enthalpy defined by
h = 1 + e + p/ρ with units in which the speed of light is equal to one, and the specific
internal energy e. The system (1.1)-(1.2) should be closed by using the equation of state
(EOS). This paper will only consider the ideal-fluid EOS

p = (Γ − 1)ρe

with the adiabatic index Γ ∈(1,2]. Because there is no explicit expression for the primitive
variables (ρ,uT , p) and the flux F` in terms of U, in order to recover the values of the
primitive variables and the flux from the given U, a nonlinear algebraic equation such as

Γ
E + p = DW + pW 2
Γ−1
has to be numerically solved to obtain the pressure p, and then the rest-mass density ρ,
the specific enthalpy h, and the velocity u can be orderly calculated by

D Γp m
ρ= , h = 1+ , u= .
W ( Γ − 1) ρ Dh

The relativistic description for the dynamics of the fluid (gas) at nearly the speed of light
should be considered when we investigate the astrophysical phenomena from stellar
to galactic scales, e.g., coalescing neutron stars, core collapse supernovae, active galac-
tic nuclei, superluminal jets, the formation of black holes, and gamma-ray bursts etc.
Due to the relativistic effect, the nonlinearity of the system (1.1)-(1.2) becomes much
stronger than the non-relativistic case so that its analytic treatment is extremely diffi-
cult and challenging. Numerical simulation is a primary way to help us understand
the physical mechanisms in the RHD. It can be traced back to the artificial viscosity
method for the RHD equations in the Lagrangian coordinates [29, 30] and in the Eule-
rian coordinates [40]. After those, the modern shock-captured methods for the RHD
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 3

equations would not be noticed until later. Here listed are some works: the Harten-Lax-
van Leer method [37], the two-shock solvers [1, 8], the Roe solver [10], the essentially
non-oscillatory (ENO) and the weighted ENO (WENO) methods [9, 53], the piecewise
parabolic methods [26, 32], the adaptive mesh refinement method [54], the Runge-Kutta
discontinuous Galerkin methods with WENO limiter [55], the direct Eulerian general-
ized Riemann problem schemes [44, 49–51], the adaptive moving mesh methods [18, 19],
the two-stage fourth-order accurate time discretizations [52] and so on. The readers are
referred to the early review articles [15, 27, 28] and the references therein. Recently, the
properties of the admissible state set and the physical-constraints-preserving (PCP) nu-
merical schemes were well studied for the RHD, see [41, 46, 48] and [35], and for the
special relativistic magnetohydrodynamics [45, 47]. The PCP schemes satisfy that both
the rest-mass density and the kinetic pressure are positive and the magnitude of the fluid
velocity is less than the speed of light. Motivated by [45, 47], the positivity-preserving
schemes for the non-relativistic ideal magnetohydrodynamics were successfully studied
in [42, 43].
It is well known that the weak solution of the quasi-linear hyperbolic conservation
laws might not be unique so that the entropy conditions are needed to single out the
physical relevant solution among all weak solutions.

Definition 1.1 (Entropy function). A strictly convex function η (U ) is called an entropy

function for the system (1.1)-(1.2) if there are associated entropy fluxes q` (U ) such that

q0` (U ) = V T F`0 (U ), ` = 1, ··· ,d, (1.3)

where V = η 0 (U )T is called the entropy variables, and (η,q` ) is an entropy pair.

For the smooth solutions of (1.1)-(1.2), multiplying (1.1) by V T gives the entropy iden-
tity
d
η (U )t + ∑ ql (U ) xl = 0.
l =1

However, if the solutions contain discontinuity, then the above identity does not hold.

Definition 1.2 (Entropy solution). A weak solution U of (1.1) is called an entropy solution
if for all entropy functions η, the inequality

d
η (U )t + ∑ ql (U ) xl ≤ 0, (1.4)
l =1

holds in the sense of distributions.

Formally, integrating (1.4) in Rspace and imposing a periodic or no-inflow boundary

d
conditions gets the inequality dt R
η (U )dx ≤ 0, which can be converted into an a priori
estimate on the solution of (1.1) in a suitable L p space if η is convex [7, 12]. The entropy
4 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

conditions are of great importance in the well-posedness of hyperbolic conservation laws,

thus it is reasonable to seek the entropy stable schemes of (1.1), satisfying a discrete or
semi-discrete version of the entropy inequality (1.4). For the smooth solutions of the
special RHD equations (1.1)-(1.2), the thermodynamic entropy

S = ln( p)− Γln(ρ),

satisfies
d
∂(ρWS) ∂(ρu` WS)
+∑ = 0,
∂t `=1
∂x`
thus an entropy pair of (1.1)-(1.2) can be defined by

−ρWS −ρu` WS
η (U ) = , q ` (U ) = , ` = 1, ··· ,d,
Γ−1 Γ−1
and the corresponding entropy variables V = η 0 (U )T can be explicitly given by
T
Γ − S ρ ρWuT ρW

V= + , ,− , (1.5)
Γ−1 p p p

which gives the “potential” ψ` := V T F` (U )− q` (U ) = ρWu` , ` = 1, ··· ,d. It can be verified

∂F` ∂U
that, for d = 1,2, the matrices ∂U ∂U
∂V and ∂U ∂V are symmetric and ∂V is positive definite so
that (1.1)-(1.2) can be symmetrized with the above entropy pair. This is useful in design-
ing the dissipation term for the entropy stable schemes in Subsection 2.2, where a set of
particular scaled eigenvectors is used, which requires that Eqs. (1.1)-(1.2) is symmetriz-
able [31].
For the scalar conservation laws, the conservative monotone schemes were shown
that they were nonlinearly stable and satisfied the discrete entropy conditions so that
they could converge to the entropy solution [6,17]. A class of so-called E-schemes satisfy-
ing the entropy condition for any convex entropy were studied in [33,34]. Those schemes
are only first-order accurate. Due to the fact that it is basically impossible to show that
the high-order schemes of the scalar conservation laws and the schemes for the system
of hyperbolic conservation laws satisfy the entropy inequality for any convex entropy
pair, the researcher tries to study the high-order accurate entropy stable schemes, which
satisfy the entropy inequality for a given entropy pair. The second-order entropy conser-
vative schemes (satisfying the discrete entropy identity) were built in [38, 39], and their
higher-order extension was introduced in [23]. It is known that an entropy conserva-
tive scheme may become oscillatory near the shock wave, thus the additional dissipation
terms has to be added to the entropy conservative schemes to obtain the entropy stable
schemes. Combining the entropy conservative flux of the entropy conservative schemes
with the “sign” property of the ENO reconstructions, the arbitrary high-order entropy
stable schemes were constructed by using high-order dissipation terms [12]. Some en-
tropy stable schemes based on the DG framework were studied, such as [2, 20] in the
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 5

space-time DG formulation, the entropy stable nodal DG schemes using suitable quadra-
ture rules [5], and its extension to magnetohydrodynamics equations [25]. The entropy
stable schemes based on summation-by-parts (SBP) operators were developed for the
Navier-Stokes equations [11]. The existing works do not address the entropy stable
schemes for the special or general RHD equations.
The paper aims at constructing the high-order accurate entropy stable schemes for
one- and two-dimensional special RHD equations (1.1)-(1.2). The key is to technically
construct the affordable entropy conservative flux of the semi-discrete second-order ac-
curate entropy conservative schemes satisfying the semi-discrete entropy equality for the
found convex entropy pair. The paper is organized as follows. Section 2 introduces the
entropy conservative fluxes, entropy conservative schemes, and entropy stable schemes
for the one-dimensional special RHD equations. Section 3 introduces our schemes for
the two-dimensional special RHD equations. Several one- and two-dimensional numer-
ical tests are conducted in Section 4 to validate the effectiveness of our schemes. Some
conclusions are summarized in Section 5.

2 One-dimensional schemes
This section considers the one-dimensional special RHD equations, i.e., (1.1)-(1.2) with
d = 1. For the sake of convenience, the notations F1 , u1 , m1 and x1 are replaced with F,u,m
and x, respectively, so that the flux, the entropy pair and the potential become

−ρWS −ρuWS
F = ( Du,mu + p,m)T , η= , q= and ψ = ρuW,
Γ−1 Γ−1
respectively.
Let us consider a uniform mesh x1 < x2 < ··· < x N with the step size ∆x = xi − xi−1 ,
i = 2, ··· , N and the semi-discrete conservative finite difference scheme
d 1  
Ui (t) = − F̂i+ 1 (t)− F̂i− 1 (t) , (2.1)
dt ∆x 2 2

where Ui (t) approximates the point value of U ( xi ,t) and F̂i+ 1 is the numerical flux ap-
2
proximating F at xi+ 1 = xi + ∆x/2.
2

2.1 Entropy conservative fluxes

Definition 2.1 (Entropy conservative scheme). The scheme (2.1) is called entropy conser-
vative scheme if its solution satisfies a semi-discrete entropy equality

d 1  
η (Ui (t)) = − q̃ 1 (t)− q̃i− 1 (t) (2.2)
dt ∆x i+ 2 2

for some numerical entropy flux q̃i+ 1 consistent with q.

2
6 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

Theorem 2.1 (Tadmor [38]). If a consistent numerical flux F̃i+ 1 satisfies

2

JV KTi+ 1 F̃i+ 1 = JψKi+ 1 (2.3)

2 2 2

with JaKi+ 1 := ai+1 − ai and {{ a}}i+ 1 := 12 ( ai + ai+1 ), then the scheme (2.1) with F̃i+ 1 is second-
2 2 2
order accurate and entropy conservative. The corresponding numerical entropy flux is q̃i+ 1 =
2
{{V }}Ti+ 1 F̃i+ 1 −{{ψ}}i+ 1 .
2 2 2

For the scalar equation, solving (2.3) can uniquely give F̃i+ 1 , but it is not clear for a
2
general system. In [38], a solution of (2.3) was constructed by the path integral in the
phase space
Z 1
F̃i+ 1 = F (Vi + ξ (Vi+1 − Vi ))dξ, (2.4)
2 0
which might be very hard to calculate except in some special cases [13]. An explicit
solution of (2.3) was given in [39], but it was both expensive and numerically unstable.
Some explicit algebraic solutions of (2.3) were constructed in the literature for the specific
systems, such as the linear symmetric system, the shallow water equations [13], the Euler
equations [21], and the multiclass Lighthill-Whitham-Richards traffic model [4]. Herein,
we construct the affordable entropy conservative flux for the one-dimensional special
RHD equations using the strategy introduced in [36]. The key is to use the identity

JabK = {{ a}}JbK +{{b}}JaK,

where JaK and {{ a}} denote the jump and mean of a, respectively, and rewrite the jumps
of the entropy variables V and the potential ψ as some linear combinations of the jump of
a specially chosen parameter vector. To be specific, we first deal with the Lorentz factor
and omit the subscript i for simplicity. Because

1 1 (u + u )(u − u L )
q −q =q q L Rq R q ,
1 − u2R 1 − u2L 1 − u2L 1 − u2R 1 − u2L + 1 − u2R

where the subscripts L and R denote the left and right values of the variables used to
calculate the entropy conservative flux, one can define the “Lorentz mean” by

JWK = {{u}}Lor JuK,

where
uL + uR
{{u}}Lor = q q q q . (2.5)
1 − u2L 1 − u2R 1 − u2L + 1 − u2R

If choosing the parameter vector

z = (z1 ,z2 ,z3 )T = (ρ,ρ/p,u)T ,

J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 7

then
1

Jz1 K Jz2 K
 JV1 K = + + Jz2 K,
Γ − 1 {{z2 }}ln
 ln



 {{z1 }}
JV2 K = {{uW }}Jz2 K +{{z2 }}{{W }}Jz3 K +{{z2 }}{{z3 }}{{z3 }}Lor Jz3 K, (2.6)
 JV3 K = −{{W }}Jz2 K −{{z2 }}{{z3 }}Lor Jz3 K,




JψK = {{uW }}Jz1 K +{{z1 }}{{W }}Jz3 K +{{z1 }}{{z3 }}{{z3 }}Lor Jz3 K,


where {{ a}}ln = JaK/JlnaK is the logarithmic mean introduced in [21], where its stable
numerical implementation can be found. If assuming that the entropy conservative flux
is F̃ = ( F̃1 , F̃2 , F̃3 )T , and substituting Eqs. (2.6) in (2.3), then

F̃1
= {{uW }},






 {{z1 }}ln
F̃1


+ F̃1 +{{uW }} F̃2 −{{W }} F̃3 = 0,

 ( Γ − 1 ){{z2 }}ln
F̃2 {{z2 }}({{W }}+{{z3 }}{{z3 }}Lor )− F̃3 {{z2 }}{{z3 }}Lor





= {{z }}({{W }}+{{z }}{{z }}Lor ).


1 3 3

Solving the above equations can obtain the entropy conservative flux for the one-
dimensional special RHD equations as follows

F̃1 = {{z1 }}ln {{uW }},




 " 
1



 F̃2 = Q − 1 1+ {{z2 }}{{z3 }}Lor F̃1 +{{z1 }}{{W }}2
Γ


( − 1 ){{ z }} ln
2





 #

Lor


 +{{z1 }}{{z3 }}{{W }}{{z3 }} ,
(2.7)

 "


F̃3 = Q−1 {{z1 }}{{W }}{{uW }}+{{z1 }}{{z3 }}{{uW }}{{z3 }}Lor







 #
  
1


F̃1 {{z2 }}{{W }}+{{z2 }}{{z3 }}{{z3 }}Lor ,


 + 1+
(Γ − 1){{z2 }}ln



where Q = {{z2 }}{{W }}2 +{{z2 }}{{z3 }}{{W }}{{z3 }}Lor −{{z2 }}{{uW }}{{z3 }}Lor .
Remark 2.1. It is worth noting that Q in the above expressions is positive so that
our entropy conservative flux is well defined. In fact, if u L = u R and WL = WR , then
{{z3 }}Lor = {{u}}Lor = u L WL3 due to (2.5), thus Q = {{z2 }}WL2 > 0; otherwise, {{z3 }}Lor =
{{u}}Lor = JWK/JuK, thus
(WL + WR )2 u + u R WL + WR u L WL + u R WR WL − WR
   
Q ={{z2 }} + L −
4 2 2 2 uL − uR
={{z2 }}WL WR > 0.
8 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

Remark 2.2. It is also easy to verify that the entropy conservative flux (2.7) is consistent
with the flux F. If letting (ρ L ,u L , p L ) = (ρ R ,u R , p R ) = (ρ,u, p), then

F̃1 = ρuW,
  2 2 4
p ρ u W
1 + ( Γ −1) ρ p + ρW 2 + ρu2 W 4
F̃2 = = ρhW 2 u2 + p,
ρW 2 /p
  2 2 2 4 3
p ρ W u+ρ W u
ρW 2 u + ρW 4 u3 + 1 + (Γ−1)ρ p
F̃3 = 2
= ρhW 2 u.
ρW /p
The scheme (2.1) with the entropy conservative flux (2.7) is only second-order accu-
rate. However, if using that entropy conservative flux as a building block, then one can
obtain an entropy conservative flux of the 2pth-order (p ∈ N+ ) accurate scheme by using
the linear combinations of the “second-order accurate” entropy conservative fluxes [23].
Here only presents the specific expressions for the “6th-order accurate” entropy conser-
vative flux
3 3
F̃i6th


+ 21
= F̃ (Ui ,Ui+1 )− F̃ (Ui−1 ,Ui+1 )+ F̃ (Ui ,Ui+2 )
2 10
1

+ F̃ (Ui−2 ,Ui+1 )+ F̃ (Ui−1 ,Ui+2 )+ F̃ (Ui ,Ui+3 ) . (2.8)
30
The readers are referred to [12, 23] for more details on constructing the “high-order accu-
rate” entropy conservative flux.

2.2 Entropy stable fluxes

The entropy of hyperbolic conservation laws is conserved only if the solution is smooth.
In other words, the entropy is not conserved if the discontinuity such as the shock wave
appears in the solution. It is well-known that an entropy conservative scheme may be-
come oscillatory near the shock wave, thus we expect to construct an entropy stable
scheme by adding a dissipation term in the original entropy conservative scheme. This
section will first introduce the entropy stable flux and its high-order extension of Tad-
mor and his collaborators via the ENO reconstruction, and then go to the low dissipative
entropy stable flux by using a switch function in the dissipation term [3].
Theorem 2.2 (Tadmor [38]). If assuming that Di+ 1 is a symmetric positive semi-definite matrix
2
and F̃i+ 1 is an entropy conservative flux, then the scheme (2.1) with the following numerical flux
2

1
F̂i+ 1 = F̃i+ 1 − Di+ 1 JV Ki+ 1 (2.9)
2 2 2 2 2

is entropy stable, i.e., satisfying the semi-discrete entropy inequality

d 1  
η (Ui (t))+ q̂i+ 1 (t)− q̂i− 1 (t) ≤ 0
dt ∆x 2 2
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 9

for some numerical entropy flux function q̂i+ 1 consistent with q.

2

The above theorem holds for any positive semi-definite matrix Di+ 1 , which is usually
2
chosen as
Di+ 1 = Ri+ 1 |Λi+ 1 | RTi+ 1 ,
2 2 2 2

where R is a scaled matrix of right eigenvectors, whose existence can be ensured by the
eigenvector scaling theorem in [31], and satisfies
∂F
= RΛR−1 , UV = RRT .
∂U
For the one-dimensional special RHD equations, after some algebraic manipulations, the
scaled matrix R is
 12
ρW (1 − ucs )

  0 0
1 1 1  2Γ 
(Γ − 1)ρW
 
 (u − cs )Wh uW (u + cs )Wh  
 , (2.10)

0 0
Γ
 
(1 − ucs )Wh W (1 + ucs )Wh 
 
ρW (1 + ucs )

0 0
2Γ
where cs = Γp/(ρh) is the sound speed. Because the scaled matrix R is mainly defined
p

at xi+ 1 , one has to use some “averaged” values of the variables to calculate it. This paper
2
chooses
{{ρ}}ln
i + 12
ln
ρ̄ = {{ρ}}i+ 1 , ū = {{u}}i+ 1 , p̄ = ,
2 2 {{ρ/p}}ln i+ 1 2

to replace the variables ρ,u, p in (2.10) to obtain the scaled matrix Ri+ 1 .
2
For the choice of |Λ|, one can use the Roe type dissipation term

|Λ| = diag{|λ1 |, |λ2 |, |λ3 |},

∂F
where λ1 , λ2 , λ3 are three eigenvalues of ∂U , or the Lax-Friedrichs type dissipation term

|Λ| = max{|λ1 |, ··· , |λm |} I.

If only Vi and Vi+1 are used to calculate JV Ki+ 1 , then the scheme (2.1) with the entropy
2
stable flux (2.9) is only first-order accurate even if a “high-order accurate” entropy con-
servative flux is used. In [12], the arbitrary high-order entropy stable schemes are con-
structed by applying the ENO reconstruction to the scaled entropy variables w = RT V .
More specifically, apply the kth-order accurate ENO reconstruction to w to obtain the left
and right limit values at xi+ 1 , denoted by wi−+ 1 and wi++ 1 , and
2 2 2

hhwiii+ 1 = wi++ 1 − wi−+ 1 ,

2 2 2
10 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

then replace the second-order entropy conservative flux (2.7) with 2pth-order entropy
conservative flux F̃ 2pth , where p = k/2 for even k and p = (k + 1)/2 for odd k, and replace
the dissipation term with 21 Ri+ 1 |Λi+ 1 |hhwiii+ 1 . Finally, one has the “kth-order accurate”
2 2 2
entropy stable flux
2pth 1
F̂i+ 1 = F̃i+ 1 − Ri+ 1 |Λi+ 1 |hhwiii+ 1 . (2.11)
2 2 2 2 2 2

The semi-discrete numerical schemes (2.1) with above high-order flux is entropy stable if
the reconstruction satisfies the following “sign” property [12]

sign(hhwiii+ 1 ) = sign(JwKi+ 1 ),
2 2

which is satisfied by the ENO reconstructions [14].

Certainly, one can also obtain higher-order accuracy with the WENO reconstruction
instead of the ENO reconstruction if the same number of candidate points values are
used, but a general WENO reconstruction may not satisfy the “sign” property. Borrowing
the idea from [3], we add a switch function in the dissipation term as follows

2pth 1
F̂i+ 1 = F̃i+ 1 − Si+ 1 Ri+ 1 |Λi+ 1 |hhwiii+ 1 , (2.12)
2 2 2 2 2 2 2

where   
1, if sign hhwiiil+ 1 = sign(JwKil+ 1 ) 6= 0,
Sil+ 1 = 2 2
2 0, otherwise,

here the superscript l denotes the l-th entry of the diagonal matrix Si+ 1 or the l-th compo-
2
nent of the jump of w. When the WENO reconstruction does not satisfy the “sign” prop-
erty, corresponding dissipation term becomes zero, and thus the semi-discrete numerical
scheme with the flux (2.12) is entropy stable according to Theorem 2.2. Meanwhile, com-
pared to the entropy stable flux using the ENO reconstruction, the above flux using the
WENO reconstruction leads to less dissipation because the switch function is not active
at all locations.
This paper uses the fifth-order accurate WENO reconstruction in [22] and the follow-
ing third-order accurate Runge-Kutta time discretization for the time derivatives in (2.1)

U (1) = U n + ∆tL(U n ),
3 1 
U (2) = U n + U (1) + ∆tL(U (1) ) ,
4 4
1 2 
U n+1 = U n + U (2) + ∆tL(U (2) ) ,
3 3

where [ L(U )]i denotes the right-hand side term of (2.1).

J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 11

3 Two-dimensional schemes
The two-dimensional finite difference scheme for solving (1.1) with d = 2 can be done
in a dimension-by-dimension fashion. For simplicity, the notations F1 , F2 , u1 , u2 , m1 ,
m2 and x1 , x2 are replaced with F, G, u, v, m x , my and x, y respectively, and thus F =
( Du,m x u + p,my u,m x )T , G = ( Dv,m x v,my v + p,my )T . The two-dimensional entropy pair
and potential are

−ρWS −ρuWS −ρvWS

η= , qx = , qy = , ψx = ρuW, ψy = ρvW.
Γ−1 Γ−1 Γ−1

Consider a uniform Cartesian mesh with the spatial stepsizes ∆x, ∆y. The solution U is
approximated at ( xi ,y j ), i = 1, ··· , Nx , j = 1, ··· , Ny , and the x- and y-directional numerical
fluxes are defined at ( xi+ 1 ,y j ) and ( xi ,y j+ 1 ), respectively. Then a semi-discrete conserva-
2 2
tive finite difference scheme can be expressed as

d 1   1  
Ui,j (t) = − F̂i+ 1 ,j − F̂i− 1 ,j − Ĝi,j+ 1 − Ĝi,j− 1 , (3.1)
dt ∆x 2 2 ∆y 2 2

where the numerical fluxes F̂i+ 1 ,j , Ĝi,j+ 1 are defined by

2 2

2pth 1
F̂i+ 1 ,j = F̃i+ 1 ,j − Si+ 1 ,j Rix+ 1 ,j |Λix+ 1 ,j |hhw x iii+ 1 ,j ,
2 2 2 2 2 2 2

2pth 1 y y
Ĝi,j+ 1 = G̃i,j+ 1 − Si,j+ 1 Ri,j+ 1 |Λi,j+ 1 |hhwy iii,j+ 1 ,
2 2 2 2 2 2 2

with w x = R x V , wy = Ry V . Here F̃ 2pth , G̃2pth , R x , Ry , and Λ x , Λy are the “high-order

accurate” entropy conservative fluxes, the scaled matrices of right eigenvectors, and the
diagonal Roe type or Lax-Friedrichs type dissipation terms in x- and y-directions, re-
spectively, which will be given below, and S is the same switch function as in the one-
dimensional case.
Motivated by the one-dimensional case, one can define two “Lorentz mean” by

JWK = {{u,v}}Lorx JuK +{{u,v}}Lory JvK,

where
uL + uR
{{u,v}}Lorx = q q q q ,
1 − u2L − v2L 1 − u2R − v2R 1 − u2L − v2L + 1 − u2R − v2R
vL + vR
{{u,v}}Lory = q q q q .
1 − u2L − v2L 2 2
1 − uR − vR 2 2 2 2
1 − uL − vL + 1 − uR − vR
12 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

If taking the parameter vector z = (z1 ,z2 ,z3 ,z4 )T = (ρ,ρ/p,u,v)T and following the same
procedure in the one-dimensional case, then one can obtain the entropy conservative flux
F̃ = ( F̃1 , F̃2 , F̃3 , F̃4 )T in the x-direction

F̃1 = {{z1 }}ln {{uW }},




 n
−1 α {{ z }}{{ z ,z }}Lorx F̃ +{{ z }}({{W }}2 −{{ vW }}{{ z ,z }}Lory )

F̃ = Q



 2 2 3 4 1 1 3 4

 o

+{{z1 }}{{W }}({{z3 }}{{z3 ,z4 }}Lorx +{{z4 }}{{z3 ,z4 }}Lory ) ,
 n o
−1 α {{ z }}{{ z ,z }}Lory F̃ +{{ z }}{{ uW }}{{ z ,z }}Lory ,

F̃ = Q

3 2 3 4 1 1 3 4





  
 F̃4 = {{W }}−1 α F̃1 +{{uW }} F̃2 +{{vW }} F̃3 ,


and the entropy conservative flux G̃ = ( G̃1 , G̃2 , G̃3 , G̃4 )T in the y-direction

G̃1 = {{z1 }}ln {{vW }},




 n o
−1 α {{ z }}{{ z ,z }}Lorx G̃ +{{ z }}{{ vW }}{{ z ,z }}Lorx ,

G̃ = Q



 2 2 3 4 1 1 3 4

 n
G̃3 = Q−1 α{{z2 }}{{z3 ,z4 }}Lory G̃1 +{{z1 }}({{W }}2 −{{uW }}{{z3 ,z4 }}Lorx )


 o

+{ { z }}{{ W }}({{ z }}{{ z ,z }} Lorx +{{ z }}{{ z ,z }}Lory ) ,
1 3 3 4 4 3 4





  
 G̃4 = {{W }}−1 α G̃1 +{{uW }} G̃2 +{{vW }} G̃3 ,


where
1
α = 1+ , (3.2a)
(Γ − 1){{z2 }}ln

Q = {{z2 }}{{W }}2 +{{z2 }} {{z3 }}{{W }}{{z3 ,z4 }}Lorx −{{uW }}{{z3 ,z4 }}Lorx

+{{z4 }}{{W }}{{z3 ,z4 }}Lory −{{vW }}{{z3 ,z4 }}Lory . (3.2b)

Similarly, Q in (3.2) is positive, because if WL = WR , then

uL + uR 3 vL + vR 3
{{z3 ,z4 }}Lorx = WL , {{z3 ,z4 }}Lory = WL ,
2 2
and thus Q = {{z2 }}WL2 > 0; otherwise,

(u L + u R )JWK (v L + v R )JWK
{{z3 ,z4 }}Lorx = , {{z3 ,z4 }}Lorx = ,
Ju2 K + Jv2 K Ju2 K + Jv2 K

and thus one can simplify Q as {{z2 }}WL WR , which is positive. The above two-
dimensional entropy conservative fluxes are also consistent after some algebraic simpli-
fication.
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 13

For the two-dimensional special RHD equations, the scaled matrix R x in x-direction
is
1
B x −C x

2
0 0 0

 2 

(Γ − 1)ρW 3

1 1/W Wv 1  
x λx
hW A− u 2hW 2 uv x λx  
hW A+
 0 0 0 
− + Γ

 ,
h(1 + 2W 2 v2 )

 hWv v hWv   p 
x x
 0 0 0 
hW A− 1 2hW 2 v hW A+ 
 W (1 − u2 ) h 

 B x +C x 
0 0 0
2
√
x= u(1−c2s )±cs /W 1−u2 −v2 c2s x = 1− u2 ,
where λ± 1−(u2 +v2 )c2s
are two eigenvalues in the x-direction, and A± 1−uλ± x
√ 2 22
2
ρW (1−u −v cs )2 2 ρuc 1 − u − v c
Bx = Γ (1− u2 )
, C x = s Γ(1−u2 ) s . The scaled matrix Ry in y-direction is
1
B y −C y

2
0 0 0
 2 
 p

1 Wu 1/W 1

0 0 0
 
h(1 + 2W 2 u2 ) W (1 − v2 ) h
 
 hWu u hWu   
 ,
hW Ay λy 2hW 2 uv
y y
  
− − v hW A+ λ+   (Γ − 1)ρW 3 
y y
 0 0 0 
hW A− 2hW 2 u 1 hW A+   Γ


 B y +C y 
0 0 0
2
√
y v(1−c2s )±cs /W 1−v2 −u2 c2s y 2
where λ± = 1−(u2 +v2 )c2s
are two eigenvalues in the y-direction, and A± = 11−−vλ
v
y ,
√ ±
ρW (1−v2 −u2 c2s ) ρvcs 1−v2 −u2 c2s
By = Cy =
Γ (1− v2 )
, Γ (1− v2 )
.
The dissipation terms Λ x and Λy are
similar to the one-dimensional case except for
that the eigenvalues used in the dissipation terms are λ− x , u, u, λ x for Λ x , and λy , v, v,
+ −
y
λ+ for Λy , respectively. In order to obtain the jumps hhw x ii and hhwy ii, one just needs
to perform the WENO reconstructions in x- and y-directions independently. Similar to
the one-dimensional case, the third-order Runge-Kutta scheme is also used for the time
derivatives in (3.1). This completes the description of the two-dimensional entropy stable
finite difference scheme for the special RHD equations.

4 Numerical results
This section presents some numerical results to validate the performance of our entropy
stable schemes for the special RHD equations (1.1) with d = 1,2. All the tests take the CFL
number as 0.4, Γ = 5/3, and the Lax-Friedrichs type dissipation terms unless otherwise
stated. For the one- and two-dimensional tests, the time stepsizes are respectively chosen
as
CFL∆x
∆t = ,
max |λ x (Ui )|
i
14 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

and
CFL
∆t = ,
max |λ x (Ui,j )|/∆x + max |λy (Ui,j )|/∆y
i,j i,j

where |λ x | and |λy | are the maximum absolute values of all the eigenvalues in x- and
y-directions, respectively, but for the accuracy tests, ∆t is taken as the minimum between
the above choices and CFL∆x5/3 (resp. CFLmin(∆x,∆y)5/3 ) for the one-dimensional
(resp. two-dimensional) test to make the spatial error dominate.

4.1 One-dimensional results

Example 4.1 (Accuracy test). This test is used to verify the accuracy. The initial condition
is
(ρ,u, p) = (1 + 0.2sinx, 0.2, 1), x ∈ [0,2π ],

with the periodic boundary condition. The exact solutions can be given by

(ρ,u, p) = (1 + 0.2sin( x − 0.2t), 0.2, 1).

Table 1 lists the errors and the orders of convergence in ρ at t = 0.1 obtained by using
our 1D scheme. It is seen that our scheme gets the fifth-order accuracy as expected.

Table 1: Example 4.1: Errors and orders of convergence of in ρ at t = 0.1.

N `1 error order `2 error order `∞ error order

20 5.475e-06 - 6.741e-06 - 1.453e-05 -
40 1.615e-07 5.08 1.966e-07 5.10 3.979e-07 5.19
80 2.692e-09 5.91 3.450e-09 5.83 7.490e-09 5.73
160 7.791e-11 5.11 1.054e-10 5.03 2.622e-10 4.84
320 2.448e-12 4.99 3.331e-12 4.98 8.297e-12 4.98

Example 4.2 (Riemann problem 1). The initial data of the first 1D Riemann problem are
(
(10, 0, 40/3), x < 0.5,
(ρ,u, p) =
(1, 0, 10−6 ), x > 0.5.

As the time increases, the initial discontinuity will be decomposed into a left-moving
rarefaction wave, a contact discontinuity, and a right-moving shock wave. The rest-mass
density, the velocity, and the pressure at t = 0.4 obtained by the entropy stable schemes
with 400 points are shown in Fig. 1. One can see that the numerical solutions are in good
agreement with the exact solutions, and the shock, the rarefaction wave, and the contact
discontinuity are well captured without obvious oscillations.
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 15

(a) ρ (b) u (c) p

Figure 1: Example 4.2: Circles and the solid lines denote the numerical and exact solutions, respectively.
N = 400, t = 0.4.

(a) ρ (b) u (c) p

Figure 2: Example 4.3: Circles and the solid lines denote the numerical and exact solutions, respectively.
N = 400, t = 0.4.

Example 4.3 (Riemann problem 2). The initial data of the second 1D Riemann problem
are
(
(1, 0, 103 ), x < 0.5,
(ρ,u, p) = − 2
(4.1)
(1, 0, 10 ), x > 0.5.

The flow pattern is similar to that of the first Riemann problem, but more extreme
and difficult with a heavily curved profile for the rarefaction fan. The region between the
contact discontinuity and the right-moving shock wave is very narrow so that resolving
those waves is very challenging. The rest-mass density, the velocity, and the pressure at
t = 0.4 obtained by the entropy stable scheme with 400 points are shown in Fig. 2. It can be
seen that our scheme can still resolve the waves well, especially for the rest-mass density
profile, even though small undershoot appears in the narrow region between the contact
discontinuity and the right-moving shock wave.
16 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

(a) ρ (b) u (c) p

Figure 3: Example 4.4: Circles and the solid lines denote the numerical and exact solutions, respectively.
N = 400, t = 0.4.

Example 4.4 (Riemann problem 3). The initial data of the third 1D Riemann problem are
(
(1, 0.9, 1), x < 0.5,
(ρ,u, p) = (4.2)
(1, 0, 10), x > 0.5,

with Γ = 4/3. The solutions will contain a slowly left-moving shock wave, a contact dis-
continuity, and a fast right-moving shock wave, see Fig. 3, where numerical solutions at
t = 0.4 are obtained by our entropy stable scheme with 400 points. Our numerical solu-
tions are in agreement with the exact solutions, although small oscillations are observed
behind the left-moving shock wave like many shock-capturing methods, but they can be
improved by using the adaptive moving mesh method, see [18].

Example 4.5 (Riemann problem 4). The initial data of the fourth 1D Riemann problem
are (
(1, − 0.7, 20), x < 0.5,
(ρ,u, p) = (4.3)
(1, 0.7, 20), x > 0.5.

The solution of this Riemann problem consists of a left-moving rarefaction wave, a con-
tact discontinuity, and a right-moving rarefaction wave. The rest-mass density, the ve-
locity, and the pressure at t = 0.4 obtained by the entropy stable scheme with 400 points
are shown in Fig. 4. It is seen that our entropy stable scheme can well capture the wave
pattern, although in the profile of the rest-mass density, there is undershoot similar to the
results in [50].

Example 4.6 (Density perturbation problem). This is a more general Cauchy problem
obtained by including a rest-mass density perturbation in the initial data of correspond-
ing Riemann problem in order to test the ability of shock-capturing schemes to resolve
small scale flow features, which may give a good indication of the numerical (artificial)
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 17

(a) ρ (b) u (c) p

Figure 4: Example 4.5: Circles and the solid lines denote the numerical and exact solutions, respectively.
N = 400, t = 0.4.

(a) ρ (b) u (c) p

Figure 5: Example 4.6: Circles and the solid lines denote the numerical and reference solutions, respectively.
N = 400, t = 0.35.

viscosity of the scheme. The initial data are given by

(
(5, 0, 50), x < 0.5,
(ρ,u, p) = (4.4)
(2 + 0.3sin(50x ), 0, 5), x > 0.5.

Fig. 5 plots the numerical results at t = 0.35 obtained by using our entropy stable scheme
with 400 points. The reference solution is obtained by using the first-order local Lax-
Friedrichs scheme with 10000 uniform cells. We can see that the shock wave is moving
into a sinusoidal rest-mass density field, and then some smooth but complex structures
are generated at the left when the shock wave interacts with the sine wave; and our
entropy stable scheme can capture them well.

Example 4.7 (Blast wave interaction). This test describes the collision of blast waves and
is used to evaluate the performance of our entropy stable scheme for the flow with strong
18 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

(a) ρ (b) u (c) p

Figure 6: Example 4.7: Circles and the solid lines denote the numerical and exact solutions, respectively.
N = 4000, t = 0.43.

discontinuities. The initial condition is


3
(1, 0, 10 ),
 x < 0.1,
(ρ,u, p) = (1, 0, 10−2 ), 0.1 < x < 0.9, (4.5)

(1, 0, 102 ), x > 0.9,


with inflow and outflow boundary conditions and Γ = 1.4.

Fig. 6 gives close-up of the solutions at t = 0.43 obtained by using the entropy stable
scheme with 4000 uniform cells. The solutions at this time within the interval [0.5,0.53]
consists of two shock waves and two contact discontinuities. It can be seen that our
scheme can well resolve those discontinuities and clearly capture the complex relativistic
wave configuration, except for slight overshoot and undershoot of the rest-mass density
and the pressure near x = 0.517.

4.2 Two-dimensional results

Example 4.8 (Accuracy test). It is a 2D relativistic isentropic vortex problem to test the
accuracy, whose detailed construction can be found in [24]. The initial rest-mass density
and pressure are

p = ρΓ ,
2 1
ρ( x,y) = (1 − C1 e1−r ) Γ−1 ,

where

(Γ − 1)/Γ 2
q
C1 = e , r = x02 + y20 ,
8π 2
γ−1 γ−1 1
x0 = x + ( x + y ), y0 = y + ( x + y ), γ= √ ,
2 2 1 − w2
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 19

Table 2: Example 4.8: Errors and orders of convergence in the rest-mass density at t = 20.
N `1 error order `2 error order `∞ error order
20 1.704e-02 - 4.982e-02 - 4.276e-01 -
40 2.886e-03 2.56 8.947e-03 2.48 7.352e-02 2.54
80 1.781e-04 4.02 6.750e-04 3.73 1.300e-02 2.50
160 4.973e-06 5.16 2.962e-05 4.51 9.425e-04 3.79
320 1.026e-07 5.60 8.240e-07 5.17 3.048e-05 4.95

and the initial velocities are

γw2
 
1 u0 w
u= √ −√ + ( u0 + v0 ) ,
1 − w ( u0 + v0 ) / 2 γ 2 2( γ + 1)
γw2
 
1 v0 w
v= √ −√ + ( u0 + v0 ) ,
1 − w ( u0 + v0 ) / 2 γ 2 2( γ + 1)

with
s
2
C2 2ΓC1 e1−r
(u0 ,v0 ) = (−y0 ,x0 ) f , f= , C2 = 2.
1 + C2 r2 2Γ − 1 − ΓC1 e1−r

This test describes a relativistic vortex moves with a constant speed of magnitude w in
(−1, −1) direction. √
The test is computed in the domain [−5,5]2 with w = 0.5 2, e = 5, and periodic bound-
ary conditions. The output time is t = 20 so that the vortex travels and returns to the
original position after a period. Table 2 lists the errors of the rest-mass density and or-
ders of convergence. It can be clearly seen that our entropy stable scheme achieves fifth-
order accuracy. Fig. 7 plots the contours of the rest-mass density, and the velocities with
30 equally spaced contour lines. The results show that due to the Lorentz contraction,
the vortex becomes elliptic and the velocity in x- (resp. y-) direction is not symmetric
respect to y = 0 (resp. x = 0) anymore. Fig. 8 presents the change of the total entropy
∑i,j η (Ui,j )∆x∆y with respect to the time obtained by the entropy conservative scheme
and the entropy stable scheme with Nx = Ny = 320. We can see that for the entropy con-
servative scheme, the total entropy almost keeps conservative and for the entropy stable
scheme, the total entropy decays as expected.

Example 4.9 (Riemann problem 1). The initial data are




 (0.5, 0.5, −0.5, 5), x > 0.5, y > 0.5,

(1, 0.5, 0.5, 5), x < 0.5, y > 0.5,
(ρ,u,v, p) =


 (3, −0.5, 0.5, 5), x < 0.5, y < 0.5,
(1.5, −0.5, −0.5, 5), x > 0.5,

y < 0.5.
20 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

(a) ρ (b) u (c) v

Figure 7: Example 4.8: 30 equally spaced contour lines. Nx = Ny = 320, t = 20.

Figure 8: Example 4.8: The change of the total entropy with respect to t. Circles and deltas denote the results
obtained by the entropy conservative scheme and the entropy stable scheme, respectively. Nx = Ny = 320.

It describes the interaction of four contact discontinuities (vortex sheets) with the same
sign (the negative sign).
Fig. 9 shows the contours of the rest-mass density and pressure logarithms with 30
equally spaced contour lines. We can see that the four initial vortex sheets interact each
other to form a spiral with the low rest-mass density around the center of the domain as
time increases, which is the typical cavitation phenomenon in gas dynamics.
Example 4.10 (Riemann problem 2). The initial data are



 (1, 0, 0, 1), x > 0.5, y > 0.5,

(0.5771, −0.3529, 0, 0.4), x < 0.5, y > 0.5,
(ρ,u,v, p) =


 (1, −0.3529, −0.3529, 1), x < 0.5, y < 0.5,
(0.5771, 0, −0.3529, 0.4), x > 0.5,

y < 0.5,
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 21

Figure 9: Example 4.9: Nx = Ny = 400, left: lnρ, right: ln p, 30 equally spaced contour lines.

Figure 10: Example 4.10: Nx = Ny = 400, left: lnρ, right: ln p, 30 equally spaced contour lines.

which is about the interaction of four rarefaction waves.

Fig. 10 plots the contours of the rest-mass density and pressure logarithms with 30
equally spaced contour lines. The results show that those four initial discontinuities first
evolve as four rarefaction waves and then interact each other and form two (almost par-
allel) curved shock waves perpendicular to the line y = x as time increases.

Example 4.11 (Riemann problem 3). The initial data are




 (0.035145216124503, 0, 0, 0.162931056509027), x > 0.5, y > 0.5,

(0.1, 0.7, 0, 1), x < 0.5, y > 0.5,
(ρ,u,v, p) =


 (0.5, 0, 0, 1), x < 0.5, y < 0.5,

(0.1, 0, 0.7, 1), x > 0.5, y < 0.5,
22 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

Figure 11: Example 4.11: Nx = Ny = 400, left: lnρ, right: ln p, 30 equally spaced contour lines.

where the left and bottom discontinuities are two contact discontinuities and the top and
right are two shock waves.
We show the contours of the rest-mass density and pressure logarithms with 30
equally spaced contour lines in Fig. 11. The initial discontinuities interact each other,
and form a “mushroom cloud” around the point (0.5,0.5).

Example 4.12 (Shock-bubble interaction problems). This example considers two shock-
bubble interaction problems within the computational domain [0,325]×[−45,45]. The
detailed setup can be found in [18]. For the first problem, the initial left and right states
of the planar shock wave moving left are
(
(1, 0, 0, 0.05), x < 265,
(ρ,u,v, p) =
(1.865225080631180, −0.196781107378299, 0, 0.15), x < 265,

and the state of the bubble is

q
(ρ,u,v, p) = (0.1358, 0, 0, 0.05), ( x − 45)2 +(y − 45)2 ≤ 25.

The setup of the second problem is the same except that the initial state of the bubble is
replace with (ρ,u,v, p) = (0.1358, 0, 0, 0.05).
Fig. 12 shows the schlieren images of the rest-mass density ρ of the first shock-bubble
interaction problem at t = 90,180,270,360,450 (from top to bottom) with Nx = 650, Ny =
180. Fig. 13 gives the schlieren images of the rest-mass density ρ of the second shock-
bubble interaction problem at t = 100,200,300,400,500 (from top to bottom) with Nx = 650,
Ny = 180. Those plots clearly show the dynamics of the interaction between the shock
waves and the bubbles, and the discontinuities and some small wave structures are also
captured well by our entropy stable scheme.
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 23

Figure 12: The first problem of Example 4.12: the schlieren images of ρ at t = 90,180,270,360,450 from top to
bottom.

Example 4.13 (Shock-vortex interaction). The final example is about the interaction be-
tween a shock wave and a vortex. The computational domain is [−17,3]×[−5,5], with
reflective boundary conditions at y =±5, inflow and outflow boundary conditions at x = 3
24 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

Figure 13: The second problem of Example 4.12: the schlieren images of ρ at t = 100,200,300,400,500 from top
to bottom.

and x = −17, respectively, and the adiabatic index is Γ = 1.4. We put a similar isentropic
vortex initially centered at (0,0) as in Example 4.8, except that the vortex here is mov-
ing left with magnitude of w = 0.9. A planar stationary shock wave with Mach number
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 25

Figure 14: Example: 4.13: 50 equally spaced contours lines from 0 to 1 of log10 (1 +|∇ρ|) at t = 19. Ms = 1.5.
Nx = 800, Ny = 400. Left: entropy stable scheme; right: fifth order WENO-LLF.

Ms = 1.5 is placed at x =−6, which is initially far away from the vortex, thus the pre-shock
state is a const state (ρ,u,v, p) = (1, −0.9, 0, 1). Then from the jump condition and the Lax
shock condition, we can obtain the post-shock state as

(ρ,u,v, p) = (4.891497310766981, −0.388882958251919, 0, 11.894863258311670).

Fig. 14 plots the contours with 50 equally spaced contour lines from 0 to 1 of
log10 (1 +|∇ρ|), obtained with our entropy stable scheme and the fifth-order finite dif-
ference WENO scheme with the local Lax-Friedrichs splitting. The computation is per-
formed until t = 19 with Nx = 800, Ny = 400. We can see that, after the interaction of
the vortex and the shock, the shock wave is still located at x = −6, and many linear and
non-linear waves, and sound waves generate and propagate in the domain. Our entropy
stable scheme can capture the subtle details better than the fifth-order finite difference
WENO scheme with the local Lax-Friedrichs splitting.

5 Conclusions
For the special relativistic hydrodynamic (RHD) equations, the schemes satisfying the
discrete entropy condition for a convex entropy function have not been considered be-
fore. This paper has presented the high-order accurate entropy stable finite difference
schemes for one- and two-dimensional special RHD equations. Those schemes are built
on the entropy conservative flux and the weighted essentially non-oscillatory (WENO)
technique as well as explicit Runge-Kutta time discretization. The key is to technically
construct the affordable entropy conservative flux of the semi-discrete second-order ac-
curate entropy conservative schemes satisfying the semi-discrete entropy equality for the
found convex entropy pair. The entropy conservative schemes may become oscillatory
26 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

near the shock wave, thus as soon as the entropy conservative flux is derived, the dissi-
pation term can be added to give the semi-discrete entropy stable schemes satisfying the
semi-discrete entropy inequality with the given convex entropy function. The WENO
reconstruction for the scaled entropy variables and the high-order explicit Runge-Kutta
time discretization are implemented to obtain the fully-discrete high-order schemes. Sev-
eral numerical tests are conducted to validate the accuracy and the ability to capture
discontinuities of our entropy stable schemes. Especially, the shock-vortex interaction
problem is designed for the first time. The results show that our schemes can achieve
designed accuracy, and can well resolve the discontinuities and subtle details. In future,
it will be interesting to study the physical-constraint-preserving property of the entropy
stable schemes, or extend them to the relativistic magnetohydrodynamic equations.

Acknowledgements
The authors were partially supported by the Special Project on High-performance Com-
puting under the National Key R&D Program (No. 2016YFB0200603), Science Chal-
lenge Project (No. TZ2016002), the National Natural Science Foundation of China
(Nos. 91630310 and 11421101), and High-Performance Computing Platform of Peking
University.

References
[1] D. S. B ALSARA , Riemann solver for relativistic hydrodynamics, J. Comput. Phys., 114 (1994), pp.
284–297.
[2] T. J. B ARTH , Numerical methods for gas-dynamics systems on unstructured meshes, in An Intro-
duction to Recent Developments in Theory and Numerics of Conservation Laws, edited by
D. Kroner, M. Ohlberger and C. Rohde, Lecture Notes in Comput. Sci. Eng., volume 5, Berlin:
Springer, 1999, pp. 195–285.
[3] B. B ISWAS AND R. D UBEY, Low dissipative entropy stable schemes using third order WENO and
TVD reconstructions, Adv. Comput. Math., 44 (2018), pp. 1153–1181.
[4] R. B ÜRGER , H. T ORRES AND C. A. V EGA , An entropy stable scheme for the multiclass Lighthill-
Whitham-Richards traffic model, Adv. Appl. Math. Mech., 11 (2019), pp. 1022–1047.
[5] T. C HEN AND C.-W. S HU , Entropy stable high order discontinuous Galerkin methods with suitable
quadrature rules for hyperbolic conservation laws, J. Comput. Phys., 345 (2017), pp. 427–461.
[6] M. G. C RANDALL AND A. M AJDA , Monotone difference approximations for scalar conservation
laws, Math. Comput., 34 (1980), pp. 1–21.
[7] C. D AFERMOS , Hyperbolic Conservation Laws in Continuum Physics, Berlin: Springer,
2000.
[8] W. L. D AI AND P. R. W OODWARD , An iterative Riemann solver for relativistic hydrodynamics,
SIAM J. Sci. Stat. Comput., 18 (1997), pp. 982–995.
[9] A. D OLEZAL AND S. S. M. W ONG , Relativistic hydrodynamics and essentially non-oscillatory
shock capturing schemes, J. Comput. Phys., 120 (1995), pp. 266–277.
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 27

[10] F. E ULDERINK AND G. M ELLEMA , General relativistic hydrodynamics with a Roe solver, As-
tron. Astrophys. Suppl. Ser., 110 (1995), pp. 587–623.
[11] T. C. F ISHER AND M. H. C ARPENTER , High-order entropy stable finite difference schemes for
nonlinear conservation laws: Finite domains, J. Comput. Phys., 252 (2013), pp. 518–557.
[12] U. S. F JORDHOLM , S. M ISHRA AND E. TADMOR , Arbitrarily high-order accurate entropy stable
essentially nonoscillatory schemes for systems of conservation laws, SIAM J. Numer. Anal., 50
(2012), pp. 544–573.
[13] U. S. F JORDHOLM , S. M ISHRA AND E. TADMOR , Energy preserving and energy stable schemes
for the shallow water equations, in Proceedings of Found. of Computational Mathematics,
Hong Kong, 2008, edited by F. Cucker, A. Pinkus, and M. Todd, London Math. Soc. Lec-
ture Notes Ser. 363, London: Cambridge University Press, 2009, 93–139.
[14] U. S. F JORDHOLM , S. M ISHRA AND E. TADMOR , ENO reconstruction and ENO interpolation
are stable, Found. Comput. Math., 13 (2013), pp. 139–159.
[15] J. A. F ONT, Numerical hydrodynamics and magnetohydrodynamics in general relativity, Living
Rev. Relativ., 11 (2008), pp. 7.
[16] E. G ODLEWSKI AND P.-A. R AVIART, Numerical Approximation of Hyperbolic Systems of
Conservation Laws, New York: Springer, 1996.
[17] A. H ARTEN , On a class of high-resolution TVD stable finite difference schemes, SIAM. J. Nu-
mer. Anal., 21 (1984), pp. 1–23.
[18] P. H E AND H. Z. TANG , An adaptive moving mesh method for two-dimensional relativistic hydro-
dynamics, Commun. Comput. Phys., 11 (2012), 114–146.
[19] P. H E AND H. Z. TANG , An adaptive moving mesh method for two-dimensional relativistic mag-
netohydrodynamics, Comput. Fluids, 60 (2012), pp. 1–20.
[20] A. H ILTEBRAND AND S. M ISHRA , Entropy stable shock capturing space–time discontinuous
Galerkin schemes for systems of conservation laws, Numer. Math., 126 (2014), pp. 103–151.
[21] F. I SMAIL AND P. L. R OE , Affordable, entropy-consistent Euler flux functions II: Entropy produc-
tion at shocks, J. Comput. Phys., 228 (2009), pp. 5410–5436.
[22] G.-S. J IANG AND C.-W. S HU , Efficient implementation of weighted ENO schemes, J. Com-
put. Phys., 126 (1996), pp. 202–228.
[23] P. G. L EFLOCH , J.-M. M ERCIER AND C. R OHDE , Fully discrete, entropy conservative schemes of
arbitrary order, SIAM J. Numer. Anal., 40 (2002), pp. 1968–1992.
[24] D. L ING , J. M. D UAN AND H. Z. TANG , Physical-constraints-preserving Lagrangian finite vol-
ume schemes for one- and two-dimensional special relativistic hydrodynamics, J. Comput. Phys.,
396 (2019), pp. 507–543.
[25] Y. L IU , C.-W. S HU AND M. Z HANG , Entropy stable high order discontinuous Galerkin methods
for ideal compressible MHD on structured meshes, J. Comput. Phys., 354 (2018), pp. 163–178.
[26] J. M. M ART Í AND E. M ÜLLER , Extension of the piecewise parabolic method to one-dimensional
relativistic hydrodynamics, J. Comput. Phys., 123 (1996), pp. 1–14.
[27] J. M. M ART Í AND E. M ÜLLER , Numerical hydrodynamics in special relativity, Living Rev. Rela-
tiv., 6 (2003), pp. 7.
[28] J. M. M ART Í AND E. M ÜLLER , Grid-based methods in relativistic hydrodynamics and magnetohy-
drodynamics, Living Rev. Comput. Astrophys., 1 (2015), pp. 3.
[29] M. M. M AY AND R. H. W HITE , Hydrodynamics calculations of general-relativistic collapse,
Phys. Rev., 141 (1966), pp. 1232–1241.
[30] M. M. M AY AND R. H. W HITE , Stellar dynamics and gravitational collapse, in Methods Com-
put. Phys., Vol. 7, edited by B. Alder, S. Fernbach, and M. Rotenberg, New York: Academic,
1967, pp. 219–258.
28 J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29

[31] M. L. M ERRIAM , An entropy-based approach to nonlinear stability, NASA-TM-101086, 1989.

[32] A. M IGNONE , T. P LEWA AND G. B ODO , The piecewise parabolic method for multidimensional
relativistic fluid dynamics, Astron. Astrophys. Suppl. Ser., 160 (2005), pp. 199–219.
[33] S. O SHER , Riemann solvers, the entropy condition, and difference approximations, SIAM J. Nu-
mer. Anal., 21 (1984), pp. 217–235.
[34] S. O SHER AND E. TADMOR , On the convergence of difference approximations to scalar conserva-
tion laws, Math. Comput., 50 (1988), pp. 19–51.
[35] T. Q IN , C.-W. S HU AND Y. YANG , Bound-preserving discontinuous Galerkin methods for rela-
tivistic hydrodynamics, J. Comput. Phys., 315 (2016), pp. 323–347.
[36] H. R ANOCHA , Comparison of some entropy conservative numerical fluxes for the Euler equations,
J. Sci. Comput., 76 (2018), pp. 216–242.
[37] V. S CHNEIDER , U. K ATSCHER , D. H. R ISCHKE , B. WALDHAUSER , J. A. M ARUHN AND
C. D. M UNZ , New algorithms for ultra-relativistic numerical hydrodynamics, J. Comput. Phys.,
105 (1993), pp. 92–107.
[38] E. TADMOR , The numerical viscosity of entropy stable schemes for systems of conservation laws, I,
Math. Comput., 49 (1987), pp. 91–103.
[39] E. TADMOR , Entropy stability theory for difference approximations of nonlinear conservation laws
and related time-dependent problems, Acta. Numer., (2004), pp. 451–512.
[40] J. R. W ILSON , Numerical study of fluid flow in a Kerr space, Astrophys. J., 173 (1972), pp.
431–438.
[41] K. L. W U , Design of provably physical-constraint-preserving methods for general relativistic hydro-
dynamics, Phys. Rev. D, 95 (2017), 103001.
[42] K. L. W U , Positivity-preserving analysis of numerical schemes for ideal magnetohydrodynamics,
SIAM J. Numer, Anal., 56 (2018), pp. 2124–2147.
[43] K. L. W U AND C.-W. S HU , A provably positive discontinuous Galerkin method for multidimen-
sional ideal magnetohydrodynamics, SIAM J. Sci. Comput., 40 (2018), pp. B1302–B1329.
[44] K. L. W U AND H. Z. TANG , A direct Eulerian GRP scheme for spherically symmetric general
relativistic hydrodynamics, SIAM J. Sci. Comput., 38 (2016), pp. B458–B489.
[45] K. L. W U AND H. Z. TANG , Admissible states and physical constraints preserving numerical
schemes for special relativistic magnetohydrodynamics, Math. Mod. Meth. Appl. Sci., 27 (2017),
pp. 1871–1928.
[46] K. L. W U AND H. Z. TANG , High-order accurate physical-constraints-preserving finite difference
WENO schemes for special relativistic hydrodynamics, J. Comput. Phys., 298 (2015), pp. 539–564.
[47] K. L. W U AND H. Z. TANG , On physical-constraints-preserving schemes for special relativistic
magnetohydrodynamics with a general equation of state, Z. Angew. Math. Phys., 69 (2018), pp.
84.
[48] K. L. W U AND H. Z. TANG , Physical-constraints-preserving central discontinuous Galerkin meth-
ods for special relativistic hydrodynamics with a general equation of state, Astrophys. J. Suppl. Ser.,
228 (2017), pp. 3.
[49] K. L. W U , Z. C. YANG AND H. Z. TANG , A third-order accurate direct Eulerian GRP scheme for
one-dimensional relativistic hydrodynamics, East Asian J. Appl. Math., 4 (2014), pp. 95–131.
[50] Z. C. YANG , P. H E AND H. Z. TANG , A direct Eulerian GRP scheme for relativistic hydrodynam-
ics: one-dimensional case, J. Comput. Phys., 230 (2011), pp. 7964–7987.
[51] Z. C. YANG AND H. Z. TANG , A direct Eulerian GRP scheme for relativistic hydrodynamics:
two-dimensional case, J. Comput. Phys., 231 (2012), pp. 2116–2139.
[52] Y. H. Y UAN AND H. Z. TANG , Two-stage fourth-order accurate time discretizations for 1D and
2D special relativistic hydrodynamics, J. Comput. Math. (2019), accepted.
J. M. Duan and H. Z. Tang / Adv. Appl. Math. Mech., 12 (2020), pp. 1-29 29

[53] L. D. Z ANNA AND N. B UCCIANTINI , An efficient shock-capturing central-type scheme for multi-
dimensional relativistic flows, I: hydrodynamics, Astron. Astrophys., 390 (2002), pp. 1177–1186.
[54] W. Q. Z HANG AND A. I. M AC FADYEN , RAM: A relativistic adaptive mesh refinement hydrody-
namics code, Astrophys. J. Suppl. Ser., 164 (2006), pp. 255–279.
[55] J. Z HAO AND H. Z. TANG , Runge-Kutta discontinuous Galerkin methods with WENO limiter for
the special relativistic hydrodynamics, J. Comput. Phys., 242 (2013), pp. 138–168.