Acta Numerica (2003), pp.

451–512 c Cambridge University Press, 2003

DOI: 10.1017/S0962492902000156 Printed in the United Kingdom

Entropy stability theory

for difference approximations
of nonlinear conservation laws
and related time-dependent problems
Eitan Tadmor∗
Department of Mathematics,
Center for Scientific Computation
and Mathematical Modeling (CSCAMM )
and
Institute for Physical Science & Technology (IPST ),
University of Maryland,
College Park, MD 20742, USA
E-mail: [email protected]

We study the entropy stability of difference approximations to nonlinear hy-

perbolic conservation laws, and related time-dependent problems governed
by additional dissipative and dispersive forcing terms. We employ a com-
parison principle as the main tool for entropy stability analysis, comparing
the entropy production of a given scheme against properly chosen entropy-
conservative schemes.
To this end, we introduce general families of entropy-conservative schemes,
interesting in their own right. The present treatment of such schemes extends
our earlier recipe for construction of entropy-conservative schemes, introduced
in Tadmor (1987b). The new families of entropy-conservative schemes offer
two main advantages, namely, (i) their numerical fluxes admit an explicit,
closed-form expression, and (ii) by a proper choice of their path of integration
in phase space, we can distinguish between different families of waves within
the same computational cell; in particular, entropy stability can be enforced
on rarefactions while keeping the sharp resolution of shock discontinuities.
A comparison with the numerical viscosities associated with entropy-conserv-
ative schemes provides a useful framework for the construction and analysis
of entropy-stable schemes. We employ this framework for a detailed study of
entropy stability for a host of first- and second-order accurate schemes. The
comparison approach yields a precise characterization of the entropy stability
of semi-discrete schemes for both scalar problems and systems of equations.

∗
Research was supported by NSF grants DMS01-07917 and DMS01-07428 and by ONR
grant N00014-91-J-1076.
452 E. Tadmor

We extend these results to fully discrete schemes. Here, spatial entropy dissi-
pation is balanced by the entropy production due to time discretization with
a sufficiently small time-step, satisfying a suitable CFL condition. Finally,
we revisit the question of entropy stability for fully discrete schemes using a
different approach based on homotopy arguments. We prove entropy stability
under optimal CFL conditions.

CONTENTS
1 Introduction 452
2 The entropy variables 456
3 Entropy-conservative and entropy-stable schemes 463
4 The scalar problem 464
5 Systems of conservation laws 470
6 Entropy-conservative schemes revisited 482
7 Entropy stability of fully discrete schemes 488
8 Entropy stability by the homotopy approach 494
9 Higher-order extensions 500
References 502
Appendix: Entropy stability of Roe-type schemes 507

1. Introduction
We discuss the stability of difference approximations to conservation laws
and related time-dependent problems. The related problems we have in
mind are governed by additional dissipative and dispersive forcing terms.
Our main focus, however, is devoted to nonlinear convection governed by
hyperbolic systems of conservation laws. In the linear hyperbolic framework,
L2-stability is sought as a discrete analogue for the a priori energy estimates
available in the differential set-up, e.g., Richtmyer and Morton (1967) and
Gustafsson, Kreiss and Oliger (1995); consult the recent Acta Numerica
review by Kreiss and Lorenz (1998). In the present context of nonlinear
problems dominated by nonlinear convection, we seek entropy stability as a
discrete analogue for the corresponding statement in the differential set-up.
The prototype one-dimensional problem consists of systems of conservation
laws, ut +f (u)x = 0. A distinctive feature of this problem is the spontaneous
formation of shock discontinuities. The entropy condition plays a decisive
role in the theory and numerics of such problems (Lax 1972, Smoller 1983,
Dafermos 2000). It requires u to satisfy the additional inequality, U (u)t +
F (u)x ≤ 0, for all admissibleR entropy pairs (U (u), F (u)). It follows that
the total amount of entropy, U (u(·, t) dx, does not increase in time. This
is a generalization of the (weighted) L2 -energy bound encountered in the
 Entropy stability 453

linear case. The possibility of strict inequality reflects entropy decay due to
concentration along shock discontinuities.
We consider difference approximations of the general conservative form,
d fν+ 1 − fν− 1
2 2
uν (t) + = 0.
dt ∆xν
Here uν (t) is the numerical solution computed at discrete grid lines (xν , t),
and fν+ 1 ∼ f is a numerical flux based on a stencil of neighbouring grid
2
values, uν−p+1 , . . . , uν+p . We enquire when such schemes are entropy-stable
in the sense of satisfying the corresponding discrete entropy inequality,
d Fν+ 1 − Fν− 1
2 2
U (uν (t)) + ≤ 0.
dt ∆xν
So far we have specified semi-discrete schemes based on spatial differenc-
ing. We will address the question of entropy stability for the semi-discrete
as well as the fully discrete case, taking into account additional temporal
discretization. The extension to the multidimensional set-up and a host of
related problems with additional dissipative and dispersive terms can be
handled in a straightforward manner.
We distinguish between three main tools of the trade in the analysis of
entropy stability: comparison arguments, a homotopy approach and ki-
netic formulations. We will discuss the first two and refer the reader to
Bouchut (2002), Makridakis and Perthame (2003) and the references therein
for recent contributions regarding the third. Most of our discussion will
be devoted to the main approach, based on a comparison principle: we
compare the amount of entropy dissipation produced by a given scheme
against a properly chosen entropy-stable reference. The entropy stability
of solutions to monotone schemes, for example, Harten, Hyman and Lax
(1976), is carried out by a comparison with the (entropy-stable) constant
solution (Crandall and Majda 1980). The class of entropy-stable E-schemes
(Osher 1984) is characterized by having more numerical viscosity than the
entropy-stable Godunov scheme (Tadmor 1984b). And we mention in pass-
ing the kinetic approach presented in Makridakis and Perthame (2003),
which is based on comparison of the corresponding pseudo-Maxwellians.
In Tadmor (1987b), the question of entropy stability was addressed by the
construction of certain entropy-conservative schemes, interesting for their
own sake. We begin, in Section 3, with the construction of these entropy-
conservative schemes. There are two main ingredients: (i) the use of entropy
variables, outlined in Section 2, and (ii) the choice of certain paths of inte-
gration in phase space of these entropy variables. In the scalar case, the nu-
merical fluxes are path-independent, and entropy-conservative schemes are
unique (for a given entropy pair). In Section 4 we study a host of instruc-
tive scalar examples whose entropy stability is verified by comparison with
454 E. Tadmor

entropy-conservative ones. These include the first-order Engquist–Osher,

the optimality of the Godunov scheme and the second-order Lax–Wendroff,
as well as other centred schemes. In Section 5 we turn our attention to
systems. Here we revisit the construction of entropy-conservative schemes
in terms of numerical fluxes which are integrated along straight line paths in
phase space. A comparison of numerical viscosities provides a detailed study
of entropy stability for Rusanov, Lax–Friedrichs and the family of Roe-type
schemes, as well as second-order extensions. In Section 6 we present the gen-
eral framework, introducing new families of entropy-conservative fluxes sub-
ject to the choice of path of integration in phase space. These new entropy-
conservative schemes offer two main advantages: (i) their numerical fluxes
admit an explicit, closed-form expression, and, more importantly, (ii) by a
proper choice of the path of integration (aligned with the eigen-directions
of fu ), one can distinguish between different families of waves within the
same cell, [xν , xν+1 ]. In particular, entropy stability can be enforced on
rarefactions while keeping the sharp resolution of shock discontinuities. In
Section 7 we extend our discussion to fully discrete schemes,
∆t h 1 1
i
un+1
ν ≡ uν (tn + ∆t) = unν − fν+ 1 (un+ 2 ) − fν− 1 (un+ 2 ) .
∆xν 2 2

There are three prototype examples. In the fully implicit case where we
1
set un+ 2 := un+1 , additional entropy dissipation is introduced by the time
discretization and hence this implicit backward Euler scheme is entropy-
stable whenever the semi-discrete scheme is. In the case of Crank–Nicolson
time discretization, a proper (possibly nonlinear) choice of intermediate val-
1
ues un+ 2 inherits the same unconditional entropy stability properties of
the semi-discrete problem associated with the numerical flux fν+ 1 ; finally,
2
1
the fully explicit case, un+ 2 := un , yields entropy production which needs
to be balanced by entropy dissipation on the spatial part. This balance
is achieved for a mesh ratio satisfying a suitable Courant–Friedrichs–Lewy
∆t
(CFL) condition, ∆x kfu k ≤ Const.
In Section 8 we revisit the question of entropy stability using a completely
different approach, based on homotopy arguments. The results apply to
semi- and fully discrete approximations of scalar and systems of conservation
laws. We prove the entropy stability for a large class of first-order schemes,
this time under an optimal CFL condition. For second-order scalar exten-
sions we refer to Nessyahu and Tadmor (1990, Appendix). The homotopy
argument was introduced by Lax (1971) in the context of the Lax–Friedrichs
scheme.
The entropy stability study is based on comparison with entropy-con-
servative schemes. The entropy-stable schemes discussed so far were lim-
ited by the use of second-order accurate entropy-conservative schemes as a
 Entropy stability 455

reference for a comparison. We conclude, in Section 9, with higher-order

extensions. We recast the original entropy-conservative schemes in their
piecewise linear finite element formulation (Tadmor 1986b). Higher orders
with larger stencils follow from piecewise polynomials of higher degrees. A
general framework for such high-order entropy-conservative schemes was re-
cently introduced in LeFloch, Mercier and Rohde (2002), and should serve
as the starting point for the corresponding higher-order entropy stability
analysis.
Our discussion on entropy stability theory is tied to several topics which
we were unable to explore in the present framework, and we conclude this
Introduction by mentioning a few of the items that were left out.

• Entropy-conservative schemes play an essential role in our discussion

below as the main reference for calibrating entropy stability. Entropy-
conservative schemes are interesting in their own right in the con-
text of zero dispersion limits, and completely integrable systems (con-
sult, for example, Lax, Levermore and Venakidis (1993) and Deift and
McLaughlin (1998)), with much recent renewed interest (Abramov,
Kovačič and Majda 2003, Abramov and Majda 2002). Entropy-con-
servative schemes are also sought in the context of energy conservation
for long-term shock-free integration: for example, Arakawa (1966). Let
us mention the related class of completely conservative schemes devel-
oped by the school of A. A. Samarskii and co-workers: consult, for
example, Moskalkov (1980) and the references therein.
• Entropy stability serves as an essential guideline in the design of new
computationally reliable difference schemes. Much of our discussion
below is devoted to the development of a general framework for prov-
ing the entropy stability of such schemes. As an alternative approach,
we mention the design of entropy corrections for existing schemes. For
early numerical simulations with entropy corrections along these lines,
we refer to ? and Kaddouri (1993), for example. The new class of
entropy-conservative/entropy-stable schemes explored in Section 6 of-
fers a challenging new set-up for revisiting numerical simulations with
entropy corrected schemes.
• Entropy variables are essential for symmetrization, and hence for the
sense of ordering required for the comparison approach in verifying
entropy stability. Entropy variables are essential for the weak finite
element formulation as briefly outlined in Section 9. We refer to the
streamline diffusion of Hughes, Johnson and collaborators (Hughes,
Franca and Mallet 1986, Johnson and Szepessy 1986) as an example
of a successful class of entropy variables-based finite element methods
(FEM) for treating convection-dominated problems.
456 E. Tadmor

• Compensated compactness. Quantifying the amount of entropy dissi-

pation (consult Corollary 5.1 below) enables us to convert the entropy
stability statement into a convergence proof by compensated compact-
ness arguments (Tartar 1975, DiPerna 1983, Chen 2000). This method-
ology was applied to different classes of discrete methods: for instance,
FEM streamline diffusion (Johnson, Szepessy and Hansbo 1990), the
spectral viscosity method (Tadmor 1989), and multidimensional finite
volume methods (LeVeque 2002). The present framework should pave
the way for a systematic development of a convergence theory for a
large class of entropy-stable finite difference approximations (for scalar
and 2 × 2 systems).
• Nonclassical shocks and a host of nonlinear phenomena are governed
by a borderline balance between dissipative and dispersive forces: we
refer, for example, to the recent phase transitions studies of LeFloch
and co-workers (LeFloch 2002). The numerical simulation in those
regimes becomes possible by carefully tuning the amount of entropy
dissipation/dispersion added to the entropy-conservative schemes.
• Boundary conditions. Once the entropy-conservative schemes are intro-
duced, the question of entropy stability is answered by summation by
parts, carried out in phase space of entropy variables. This reveals the
skew-symmetry of the spatial operators (Tadmor 1984a), while retain-
ing the conservative form. Consequently, summation by parts along
these lines should in principle enable us to treat the question of en-
tropy stability in the presence of boundaries: consult Olsson (1995),
for example.

2. The entropy variables

We consider systems of conservation laws of the form
∂ ∂
u+ f (u) = 0, (x, t) ∈ R × [0, ∞), (2.1)
∂t ∂x
where f (u) = (f1 (u), . . ., fN (u))> are smooth flux functions of the N -vector
of conservative variables1 u(x, t) = (u1(x, t), . . ., uN (x, t))>. We assume
that system (2.1) is equipped with a convex entropy function, U (u), such
that
Uuu A = [UuuA]> , A(u) := fu(u). (2.2)
Thus, the Hessian of an entropy function symmetrizes the system (2.1)
upon multiplication ‘on the left’ (Friedrichs and Lax 1971). An alternative
1
Here and below, scalars are distinguished from vectors, which are denoted by bold
letters.
 Entropy stability 457

procedure, which respects both strong and weak solutions of (2.1), is to sym-
metrize ‘on the right’, where (2.2) is replaced by the equivalent statement
 >
A(Uuu)−1 = A(Uuu)−1 . (2.3)
To this end, Mock (1980) (see also Godunov (1961)) suggested the follow-
ing procedure. Define the entropy variables
v ≡ v(u) := ∇u U (u). (2.4)
Thanks to the convexity of U (u), the mapping u → v is one-to-one and
hence we can make the change of variables u = u(v), which puts the system
(2.1) into its equivalent symmetric form
∂ ∂
u(v) + g(v) = 0, g(v) := f (u(v)). (2.5)
∂t ∂x
Here, u(·) and g(·) become the temporal and spatial fluxes in the indepen-
dent entropy variables, v, and the system (2.5) is symmetric in the sense
that the Jacobians of these fluxes are, namely
H(v) := uv (v) = H > (v) > 0 and B(v) := gv (v) = B > (v). (2.6)
Indeed, (2.2) holds if and only if there exists an entropy flux function, F =
F (u), such that the following compatibility relation holds:
Uu> fu = Fu> . (2.7)
Consequently, we have
u(v) = ∇v φ(v), φ(v) := hv, u(v)i − U (u(v)) (2.8)
g(v) = ∇v ψ(v), ψ(v) := hv, g(v)i − F (u(v)), (2.9)
where h·, ·i denotes the usual Euclidean inner product. Hence the Jacobians
H(v) and B(v) in (2.6) are symmetric, being the Hessians of φ(v) and ψ(v).
The latter, so-called potential functions, φ(v) and ψ(v), are significant tools
in our discussion below. Observe that the symmetry of B = AH amounts
to the symmetrization ‘on the right’ indicated in (2.3).
Entropy functions play an important role in the stability theory of PDEs
dominated by the nonlinear convection of the type (2.1). We provide be-
low a brief overview and refer the reader to a detailed account in Volpert
(1967), Kružkov (1970), Friedrichs and Lax (1971), Lax (1972), Tartar
(1975), DiPerna (1983), Smoller (1983), Majda (1984), Serre (1999), Dafer-
mos (2000) and LeFloch (2002). We first recall that ‘physically relevant’ so-
lutions of (2.1), are those arising as vanishing viscosity limits, u = lim↓0 u ,
where
ut + f (u )x = (P ux )x. (2.10)
Here P = P (u, ux ) is any admissible viscosity matrix which is H-symmetric
458 E. Tadmor

(compare (2.3)), that is,

 >
P H = P H ≥ 0, H = (Uuu)−1, (2.11)
so that integration of (2.10) against Uu> yields
 
∂  ∂  > ∂ ∂
U (u ) + F (u ) = Uu , u + f (u)
∂t ∂x ∂t ∂x
= −hUuu ux , P ux i


= − (H −1ux ), P H(H −1ux ) ≤ 0. (2.12)
Passing to the limit we obtain the entropy inequality/}
∂ ∂
U (u) + F (u) ≤ 0. (2.13)
∂t ∂x
The passage to the limit on the right of (2.12) is understood weakly, in
the sense of measures; the passage inside the nonlinear terms on the left,
however, requires strong limits: consult the recent breakthrough of Bianchini
and Bressan (Bianchini and Bressan 2003, Bressan 2003). The possibility
of a strictly negative measure on the left of (2.13) is due to concentration of
entropy dissipation along shock discontinuities on the right of (2.12).
The entropy inequality (2.13) is necessary in order to single out a unique,
‘physically relevant’ solution among the possibly many weak solutions of
(2.1). In this context it is important whether (2.1) is endowed with a suf-
ficiently ‘rich’ family of entropy pairs, (U, F ): consult, for example, Serre
(1991). How ‘rich’ is the family of such entropy functions? In the scalar
case, N = 1, scalar Jacobians are symmetric and hence every convex U
serves as an entropy function. This is the starting point for the L1-stability
theory of Kružkov (1970) for general scalar equations; we postpone this dis-
cussion to the end of this section. If the N × N system happens to be
symmetric to begin with, then we can use the identity as a symmetrizer in
(2.2), Uuu = IN , and hence the usual ‘energy’, U (u) = |u|2/2, is an en-
tropy function (Godunov 1961). In this case, integration of (2.13) yields the
entropy bound
Z Z
U (u(x, t)) dx ≤ U (u(x, 0)) dx, (2.14)
x x
which is the usual L2 -stability statement familiar from the linear theory
of symmetric hyperbolic systems. Thus, entropy stability could be viewed
as a nonlinear extension of the L2 linear stability set-up to general, non-
symmetric N × N systems. For 2 × 2 systems, the symmetrizing require-
ment from an entropy function, (2.2), amounts to a second-order linear
hyperbolic equation and Lax (1971) has shown how to construct a family of
entropy functions in this case. For general N ≥ 3 equations, (2.2) is over-
determined. Nevertheless, most physically relevant systems are equipped with
 Entropy stability 459

(at least one) entropy pair. The canonical example is of course the following
one.

Example 2.1. (Euler equations) We consider entropy solutions, u =

(ρ, m, E)> of the Euler equations
   
ρ m
∂  ∂
m +  qm + p  = 0. (2.15)
∂t E ∂x q(E + p)

These equations govern inviscid polytropic gas dynamics, asserting the con-
servation of the density ρ, the momentum m, and the total energy E.
Here q and p are, respectively, the velocity q := m
ρ and the pressure p =
 m2

(γ − 1) · E − 2ρ (where γ is the adiabatic exponent). Harten has shown
that this system of equations is equipped with a family of entropy pairs,
(U, F ). These pairs take the form
U (u) = −ρh(S), F (u) = −mh(S) (2.16)
(Harten 1983b; consult also Tadmor (1986a)). Here S stands for the non-
dimensional specific entropy
S = `n(pρ−γ ), (2.17)
and h = h(S) is any scalar function satisfying
h0 − γh00 > 0, h0 > 0, (2.18)
so that the requirement for U (u) to be convex is met (Harten 1983b). The
corresponding entropy variables are given by
   
v1 0
p
E + γ−1 ( hh(S)
0 (S) − γ − 1)
h (S) 
v ≡  v2  = (1 − γ) · · −m , (2.19)
v3 p
ρ

with the corresponding potential pairs, (φ, ψ) = (γ − 1)h0(S)(ρ, m). A par-

S
γ+1
ticularly convenient form to work with is determined by h(S) = γ−1 · e γ+1 .
With this choice we find the entropy pair
γ+1 1 γ+1 1
U (u) = · (ρp) γ+1 , F (u) = q · (ρp) γ+1 , (2.20)
1−γ 1−γ
with the corresponding entropy variables, v = v(u), given by
   
v1 E
− γ
v ≡  v2  = −(ρp) γ+1 ·  −m . (2.21)
v3 ρ
460 E. Tadmor

The inverse mapping, v → u, is easily obtained as

   
ρ γ
v3
u =  m  = −(ρp) γ+1  −v2 , (2.22)
E v1
where
   1+γ
v2 1−γ
ρp = (γ − 1) v1 v3 − 2 . (2.23)
2
Godunov has studied the special choice h(S) = S, which leads to the
canonical ‘physical’ entropy pair
U (u) = −ρS, F (u) = −mS. (2.24)
Expressed in terms of the absolute temperature, T , the entropy variables in
this case read
   2 
v1
c T (S − γ) + q2 p
v
v ≡  v2  = −  −q , T := (γ − 1) · cv · , (2.25)
v3 T ρ
1
and the inverse mapping v → u can be found in Harten (1983b). We con-
clude this example with several remarks.

(1) The Euler equations (2.15) provide us with an example which shows
how the ‘richness’ of the entropy pairs can be used for a stability state-
ment: using the one-parameter family2 (−ρ(S − c)− , −m(S − c)− ),
which is admissible by (2.18), we obtain a minimum entropy principle,
S(x, t) ≥ miny S(y, 0) (consult Tadmor (1986a)).
(2) We note that the family of admissible entropy pairs, (2.16), (2.17),
(2.18), becomes smaller once we seek further symmetrization of the
viscous Navier–Stokes terms along the lines of (2.11) (consult Hughes
et al. (1986)).
(3) Finally, we call attention to the fact that, with the particular choice of
entropy pair (U, F ) = (−ρS, −mS) (consult (2.24), (2.25)), the corre-
sponding potential pair (φ, ψ) turns out to be the density and momen-
tum components of the flow, (φ(v), ψ(v)) = (γ − 1)(ρ, m). Hence, in
view of (2.8), (2.9), Euler equations can be rewritten in the intriguing
form
∂ ∂
[∇v ρ] + [∇v m] = 0. (2.26)
∂t ∂x

2
The superscript + (respectively −
) denotes the positive (respectively negative) part of
the indicated scalar.
 Entropy stability 461

We close the section with the promised discussion on the entropy stability
of the scalar case. We start with the following result, extending the pen-
etrating scalar arguments of Kružkov (1970), which demonstrates how the
‘richness’ of the family of entropy pairs is converted into a stability state-
ment.

Theorem 2.2. (Tadmor 1997, Theorem 2.1) Assume the system (2.1)
is endowed with an N -parameter family of entropy pairs, (U (u; c), F (u; c)),
c ∈ RN , satisfying the symmetry property
U (u; c) = U (c; u), F (u; c) = F (c; u). (2.27)
Let u1 , u2 be two entropy solutions of (2.1). Then the following a priori
estimate holds:
Z Z
U (u (x, t); u (x, t)) dx ≤ U (u1 (x, 0); u2(x, 0)) dx.
1 2
(2.28)
x x

Sketch of proof. Let u1 (x, t) be an entropy solution of (2.1) satisfying

the entropy inequality (2.13). We employ the latter with the entropy pair
(U (u; c), F (u; c)) parametrized with c = u2 (y, τ ). This tells us that u1 (x, t)
satisfies
∂t U (u1 (x, t); u2(y, τ )) + ∂x F (u1 (x, t); u2(y, τ )) ≤ 0. (2.29)
Let ϕδ denote a symmetric C0∞ unit mass mollifier that converges to Dirac
mass in R as δ ↓ 0; set φδ (x − y, t − τ ) := ϕδ ( x−y t−τ
2 )ϕδ ( 2 ) as an approxi-
mate Dirac mass in R × R+ . ‘Multiplication’ of the (distributional) entropy
inequality (2.13) by φδ (x − y, t − τ ) yields
∂t (φδ U (u1; u2)) + ∂x (φδ F (u1 ; u2))
≤ (∂tφδ )U (u1 ; u2) + (∂xφδ )F (u1 ; u2). (2.30)
A dual manipulation, this time with (y, τ ) as the primary integration vari-
ables of u2(y, τ ) and (x, t) parametrizing c = u1 (x, t), yields
∂τ (φδ U (u2; u1)) + ∂y (φδ F (u2 ; u1))
≤ (∂τ φδ )U (u2 ; u1) + (∂y φδ )F (u2 ; u1). (2.31)
We now add the last two inequalities: by the symmetry property (2.27), the
sum of the right-hand sides of (2.30) and (2.31) vanishes; whereas by sending
δ to zero, the sum of the left-hand sides of (2.30) and (2.31) amounts to
∂t U (u1(x, t); u2(x, t)) + ∂x F (u1 (x, t); u2(x, t)) ≤ 0.
The result follows by spatial integration.
462 E. Tadmor

Let us point out that the elegance of the last result is confronted with the
difficulty of satisfying the symmetry property (2.27). Thus, for example,
N × N symmetric hyperbolic systems are endowed with the N -parameter
family of entropies U (u, c) = |u−c|2/2, but (2.27)
Ru fails for the corresponding
entropy fluxes, F (u, c) = hu − c, f (u)i − hf (w), dwi. The favourable
situation occurs in the scalar case where each convex U serves as an entropy
function. In particular, Kružkov (1970) set the one-parameter family

U (u) = |u − c|, F (u) = sgn(u − c)(f (u) − f (c)).

The symmetry requirement (2.27) holds and (2.28) leads to the following
L1-stability estimate.

Corollary 2.3. (Kružkov 1970) If u1, u2 are two entropy solutions of

the scalar conservation law (2.1) subject to L1 initial data, then

ku2(·, t) − u1 (·, t)kL1(x) ≤ ku2(·, 0) − u1 (·, 0)kL1(x). (2.32)

Thus there exists a unique (entropy) solution operator associated with the
scalar conservation law (2.1), S(t) : u(·, 0) 7→ u(·, t), which is conservative
and, according to Corollary 2.3, is also L1 -contractive, and hence by the
Crandall–Tartar lemma (Crandall and Tartar 1980), S is order-preserving,
u2(·, 0) ≥ u1 (·, 0) =⇒ S(t)u2(·, 0) ≥ S(t)u1(·, 0). There is a parallel discrete
theory for so-called monotone schemes which respect a similar discrete prop-
erty of order preserving. The entropy stability of such schemes goes back to
the pioneering work of Harten et al. (1976). We will not be able to expand
on the details in the limited framework of this review, but let us mention the
elegant approach of Crandall and Majda (1980), which clarified the entropy
stability of monotone schemes in terms of a comparison with the constant
solution. Sanders (1983) generalized the result to variable grids and we re-
fer to Godlewski and Raviart (1996), Kröner (1997), Tadmor (1998) and
LeVeque (2002) and the references therein for a series of later works, with
particular emphasis on multidimensional extensions. Monotone schemes are
at most first-order accurate (Harten et al. 1976); indeed, being entropy-stable
with respect to all convex entropies, monotone schemes are necessarily lim-
ited to first-order accuracy (Osher and Tadmor 1988). This limitation led
to systematic development of high-resolution schemes which circumvent this
first-order limitation. For a brief overview of the convergence analysis of
such schemes, we refer to Tadmor (1998). Our discussion below focuses on
the question of entropy stability of such first-order as well as higher-order
resolution schemes, in the context of both scalar and systems of conservation
laws.
 Entropy stability 463

3. Entropy-conservative and entropy-stable schemes

We consider semi-discrete conservative schemes of the form
d 1 h i
uν (t) = − fν+ 1 − fν− 1 , (3.1)
dt ∆xν 2 2

serving as consistent approximations to (2.1). Here, uν (t) denotes the dis-

crete solution along the grid line (xν , t) with ∆xν := 12 (xν+1 − xν−1 ) being
the variable meshsize, and fν+ 1 being the Lipschitz-continuous numerical
2
flux consistent with the differential flux, that is,
fν+ 1 = f (uν−p+1 , . . . , uν+p ), f (u, u, . . . , u) ≡ f (u). (3.2)
2

The numerical flux, f (·, ·, . . ., ·), involves a stencil of 2p neighbouring grid

values, and as such could be clearly distinguished from the (same notation
of ) the differential flux, f (·). The difference schemes (3.1) and (3.2) are
conservative in the sense of Lax and Wendroff (1960), namely, the change
of total mass,
XR
uν (t)∆xν ,
ν=−L

is solely due to the flux through the local neighbourhoods of the arbitrary
boundaries at x−L and xR .
We are concerned here with the question of entropy stability of such
schemes. To this end, let (U, F ) be an entropy pair associated with the
system (2.1). We ask whether the scheme (3.1) is entropy-stable with re-
spect to such a pair, in the sense of satisfying a discrete entropy inequality
analogous to (2.13), that is,
d 1 h i
U (uν (t)) + Fν+ 1 − Fν− 1 ≤ 0. (3.3)
dt ∆xν 2 2

Here, Fν+ 1 is a consistent numerical entropy flux

2

Fν+ 1 = F (uν−p+1 , . . . , uν+p ), F (u, u, . . . , u) = F (u). (3.4)

2

If, in particular, equality holds in (3.3), we say that the scheme (3.1) is
entropy-conservative.
The answer to this question of entropy stability provided in Tadmor (1987b)
consists of two main ingredients: (i) the use of the entropy variables and (ii)
the comparison with appropriate entropy-conservative schemes. We con-
clude this section with a brief overview.
By making the changes of variables uν = u(vν ), the scheme (3.1) recasts
into the equivalent form
d 1 h i
uν (t) = − gν+ 1 − gν− 1 , uν (t) = u(vν (t)), (3.5)
dt ∆xν 2 2
464 E. Tadmor

with a numerical flux

gν+ 1 = g(vν−p+1 , . . . , vν+p ) := f (u(vν−p+1 ), . . . , u(vν+p)), (3.6)
2

consistent with the differential flux, that is,

g(v, v, . . ., v) = g(v) ≡ f (u(v)). (3.7)
Define
1D E 1 
Fν+ 1 := [vν + vν+1 ], gν+ 1 − ψ(vν ) + ψ(vν+1) . (3.8)
2 2 2 2
Then the following identity holds:
d 1 h i
U (uν (t)) + Fν+ 1 − Fν− 1 (3.9)
dt ∆xν 2 2

1 hD E i 1 hD E i
= ∆vν+ 1 , gν+ 1 − ∆ψν+ 1 + ∆vν− 1 , gν− 1 − ∆ψν− 1
2 2 2 2 2 2 2 2

(Tadmor 1987b, Section 4; see also Osher (1984)). Here ∆ψν+ 1 := ψ(vν+1)−
2
ψ(vν ) denotes the difference of entropy flux potential, (2.9), of two neigh-
bouring grid values vν and vν+1 . Thanks to (2.9), Fν+ 1 is a consistent
2
entropy flux and this brings us to the next result.

Theorem 3.1. (Tadmor 1987b, Theorem 5.2) The conservative scheme

(3.5) is entropy-stable (respectively, entropy-conservative) if, and for three-
point schemes (p = 1) only if,
D E
∆vν+ 1 , gν+ 1 ≤ ∆ψν+ 1 , (3.10)
2 2 2

and, respectively,
D E
∆vν+ 1 , gν+ 1 = ∆ψν+ 1 . (3.11)
2 2 2

4. The scalar problem

We discuss the entropy stability of scalar schemes of the form (see (3.5))
d 1 h i
uν (t) = − gν+ 1 − gν− 1 , uν (t) ≡ u(vν (t)). (4.1)
dt ∆xν 2 2

For a more convenient formulation, let us define for ∆vν+ 1 6= 0

2

f (uν ) + f (uν+1 ) − 2gν+ 1

2
Qν+ 1 = , ∆vν+ 1 := vν+1 − vν . (4.2)
2 ∆vν+ 1 2
2
 Entropy stability 465

Our scheme recasts into the equivalent viscosity form

d 1 h i
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Qν+ 1 ∆vν+ 1 − Qν− 1 ∆vν− 1 , (4.3)
2∆xν 2 2 2 2

which reveals the role of Qν+ 1 as the numerical viscosity coefficient (e.g.,
2
Tadmor (1984b)).
According to (3.11), scalar entropy-conservative schemes are uniquely de-
∗
termined by the numerical flux gν+ 1 = gν+ 1 , that is,
2 2

∆ψν+ 1 Z 1  
2
∗ 2
gν+ 1 := ≡ g vν+ 1 (ξ) dξ,
2 ∆vν+ 1 ξ=− 12 2
2
1
vν+ 1 (ξ) := (vν + vν+1 ) + ξ∆vν+ 1 . (4.4)
2 2 2

Noting that
Z 1  
∗
2 d
gν+ 1 = (ξ) · g vν+ 1 (ξ) dξ, (4.5)
2 ξ=− 12 dξ 2

we find upon integration by parts that entropy-conservative schemes admit

the viscosity form (4.3), with a viscosity coefficient Qν+ 1 = Q∗ν+ 1 given by3
2 2

Z 1  
2
Q∗ν+ 1 = 2ξg 0 vν+ 1 (ξ) dξ. (4.6)
2 ξ=− 12 2

The entropy-conservative scheme then takes the form

d 1 h ∗ ∗
i
uν (t) = − gν+ 1 − gν− 1
dt ∆xν 2 2

1  
=− f (uν+1 ) − f (uν−1 )
2∆xν
1 h ∗ i
+ Qν+ 1 ∆vν+ 1 − Q∗ν− 1 ∆vν− 1 . (4.7)
2∆xν 2 2 2 2

The entropy stability portion of Theorem 3.1 can now be restated in the
following form.
Corollary 4.1. (Tadmor 1987b, Theorem 5.1) The conservative scheme
(4.7) and (4.3) is entropy-stable, if – and for three-point schemes (p = 1)
only if – it contains more viscosity than the entropy-conservative one (4.6),

3
We use primes to indicate differentiation with respect to primary dependent variables,
e.g., g0 = gv (v), f 00 = fuu (u), etc.
466 E. Tadmor

that is,
Q∗ν+ 1 ≤ Qν+ 1 . (4.8)
2 2

The rest of this section is devoted to examples demonstrating applications

of the last corollary.
Example 4.2. (Entropy-conservative schemes) We begin with two ex-
amples of entropy-conservative schemes, interesting in their own right, which
played a significant role in studying zero dispersion phenomena: see, e.g.,
Lax (1986) and collaborators.
We consider the inviscid Burgers’ equation, ut +( 12 u2 )x = 0, and we seek a
semi-discrete scheme that conserves the logarithmic entropy U (u) = − ln u.
The entropy flux in this case is F (u) = −u. Using the entropy variable
v(u) = −1/u, we compute the entropy flux potential
1
ψ(v) = vf (u(v)) − F (u(v)) = − ,
2v
which in turn yields the entropy-conservative flux
∗ ψ(vν+1) − ψ(vν ) 1 1 1
gν+ 1 = = = uν uν+1 .
2 vν+1 − vν 2 vν vν+1 2
This yields the entropy-conservative centred schemes
d uν+1 (t) − uν−1 (t)
uν (t) = uν (t) ,
dt 2∆xν
studied in Goodman and Lax (1988), Hou and Lax (1991) and Levermore
and Liu (1996), among others.
Next, we consider ut + (eu )x = 0 and we seek the semi-discrete scheme
that conserves the exponential entropy, U (u) = eu . The entropy flux is
F (u) = 12 e2u . Using the corresponding entropy variable v(u) = eu , we
compute the entropy flux potential
1
ψ(v) = vf (u(v)) − F (u(v)) = v 2 ,
2
which in turn yields the entropy-conservative flux
∗ ψ(vν+1 ) − ψ(vν ) 1 1 
gν+ 1 = = (vν + vν+1 ) = euν + euν+1 .
2 vν+1 − vν 2 2
This yields the entropy-conservative centred schemes
d euν+1 (t) − euν−1 (t)
uν (t) =
dt 2∆xν
associated with Toda flow: consult Lax et al. (1993), Levermore and Liu
(1996), Deift and McLaughlin (1998), and the references therein.
 Entropy stability 467

We continue with a series of entropy-stable examples.

Example 4.3. (Engquist and Osher 1980) Using the estimate
Z 1
2  0  Z
 0    du(vν+ 1 (ξ)) 
∗
Qν+ 1 ≤  
g vν+ 1 (ξ) dξ = f u vν+ 1 (ξ)   2 
2 2 ∆vν+ 1 
2 ξ=− 12
2
 Z uν+1 
1
= |f 0(u)| du =: QEOν+ 12
, (4.9)
∆vν+ 1 uν
2

we obtain an upper bound QEO

ν+ 1
, which is the viscosity coefficient associated
2
with the entropy-stable Engquist–Osher (EO) scheme (Engquist and Osher
1980).
The quantity inside the brackets on the right-hand side of (4.9) is inde-
pendent of different choices for entropy variables. Consequently, the entropy
stability of the EO scheme is uniform with respect to all admissible entropy
pairs, (U, F ). This raises the question of the minimal amount of viscosity
required to maintain such uniformity.
Example 4.4. (Godunov 1959) We rewrite the second term on the right-
hand side of the schemes (4.3) as
" ! ! #
1 ∆vν+ 1 ∆vν− 1
2 2
Qν+ 1 ∆uν+ 1 − Qν− 1 ∆uν− 1 ,
2∆xν 2 ∆u
ν+ 1 2 2 ∆u
ν− 1
2
2 2

thus normalizing their viscous part by using the conservative variables as

our fixed scale. Since in the scalar case all convex functions, U (u), are ad-
missible entropy functions, it follows that, for an entropy stability which is
uniform with respect to every such U , we need to
maximize the correspond-
ing entropy viscous factors Q∗ν+ 1 ∆vν+ 1 /∆uν+ 1 ,
2 2 2
" ∗ # Z
f (uν ) + f (uν+1 ) − 2gν+ 1
1
2


2 ∗
sup , gν+ 1 = f u vν+ 1 (ξ) dξ,
v ∆uν+ 1 2 ξ=− 12 2
2

where the supremum is taken over all increasing v = v(u). This yields
Godunov’s viscosity coefficient (Osher 1985, Tadmor 1984b)
" #
f (u ν ) + f (u ν+1 ) − 2f (u)
QGν+ 12
= max . (4.10)
min(uν ,uν+1 )≤u≤max(uν ,uν+1 ) ∆uν+ 1
2

Thus, the scalar schemes which are uniformly entropy-stable with respect
to all convex entropies are precisely those that contain at least as much
numerical viscosity as the Godunov scheme does. These so-called E-schemes
were first identified in Osher (1984); see also Tadmor (1984b).
468 E. Tadmor

The E-schemes are only first-order accurate: consult, e.g., Lemma 4.5
below. Corollary 4.1 enables us to verify the entropy stability of second-
order accurate schemes as well. To this end we recall from (4.6) that
Z 1  
2 1
Q∗ν+ 1 = 2ξg 0 vν+ 1 (ξ) dξ, vν+ 1 (ξ) = (vν + vν+1 ) + ξ∆vν+ 1 .
2 ξ=− 1 2 2 2 2
2

Integration by parts yields

Z  
1 0 
1
2 d
Q∗ν+ 1 = 2
ξ − g vν+ 1 (ξ) dξ
2 ξ=− 12 dξ 4 2

Z  
d 0 
1
2 1
= − ξ2 g vν+ 1 (ξ) dξ, (4.11)
ξ=− 12 4 dξ 2

and hence the entropy-conservative viscosity coefficient Q∗ν+ 1 takes the form
2

Z 1    
2 1
Q∗ν+ 1 = − ξ g 00 vν+ 1 (ξ) dξ · ∆vν+ 1 .
2
(4.12)
2 ξ=− 12 4 2 2

Thus, the viscosity coefficients of the entropy-conservative schemes are in

fact of order O(|∆vν+ 1 |), and this implies their second-order accuracy in
2
view of the following lemma.

Lemma 4.5. Consider the conservative
schemes (4.3) with viscosity coef-
ficient, Qν+ 1 , such that Qν+ 1 /∆vν+ 1 is Lipschitz-continuous. Then these
2 2 2
schemes are second-order accurate, in the sense that their local truncation
error is of the order
h i
O |xν+1 − xν |2 + |xν − xν−1 |2 + |xν+1 − 2xν + xν−1 | .

Verification of this lemma is straightforward and therefore omitted.

Example 4.6. (Second-order accurate schemes) Using the simple up-

per bound
1  
Q∗ν+ 1 ≤ max |g 00(v)| · ∆vν+ 1 , (4.13)
2 6 min(vν ,vν+1 )≤v≤max(vν ,vν+1 ) 2

we obtain a viscosity coefficient on the right of (4.13) which, according to

Corollary 4.1 and Lemma 4.5, maintains both entropy stability and second-
order accuracy. Viscosity terms similar to this were previously derived in
a number of special cases, dealing with the entropy stability question of
second-order schemes, such as (generalized) van Leer’s MUSCL scheme (van
Leer 1977, Osher 1985, Lions and Souganidis 1985, Yang 1996a) as well as
 Entropy stability 469

other high-resolution schemes (Majda and Osher 1978, Majda and Osher
1979, Harten 1983a, Harten and Hyman 1983, Nessyahu and Tadmor 1990).
We remark that the careful calculations required in those derivations are
due to the delicate balance of the cubic order of entropy loss, which should
match the third-order dissipation in this case.

An instructive example of using the above arguments of entropy stability

is provided in the genuinely nonlinear case, where f (u) is, say, convex. A
quadratic entropy stability is sufficient in this case, to single out the unique
physically relevant solution (Szepessy 1989, Chen 2000). In particular, the
choice of the quadratic entropy function U (u) = 12 u2 leads to entropy vari-
ables that coincide with the conservative ones, g(v) = f (u). The last three
examples of this section deal with this important special case.

Example 4.7. (Lax and Wendroff 1960) By convexity, the entropy-con-

servative viscosity coefficient in (4.12),
Z 1    
∗
2 1 2 00
Qν+ 1 = − ξ f uν+ 1 (ξ) dξ · ∆uν+ 1
2 ξ=− 1 4 2 2
2

is negative whenever ∆uν+ 1 is negative, and hence numerical viscosity is

2
required only in the case of rarefactions where ∆uν+ 1 > 0. To see how
2
much viscosity is required in this case, we use the fact that the integrand on
the right of Q∗ is positive, leading to the upper bound
Z 1  
1 2 1 +
Q∗ν+ 1 ≤ f 00 uν+ 1 (ξ) dξ · ∆uν+ 1 = a(uν+1 ) − a(uν ) . (4.14)
2 4 ξ=− 1 2 2 4
2

The resulting viscosity coefficient on the right is the second-order accurate

viscosity originally proposed by Lax and Wendroff (1960),
1 +
QLW
ν+ 1
= a(uν+1 ) − a(uν ) , a(u) = f 0 (u). (4.15)
2 4
Example 4.8. (Centred schemes) According to (2.9) with g(v) = f (u),
the entropy flux potential is given by the primitive of f (·), and by (3.10),
entropy stability is guaranteed if
Z uν+1
∆uν+ 1 · fν+ 1 ≤ f (u) du.
2 2
uν

In the rarefaction case, ∆uν+ 1 > 0, the integral on the right approximated
2
from below by the midpoint rule; in the case of a shock, ∆uν+ 1 < 0, signs
2
are reversed and we can instead use the trapezoidal rule. Thus we derive a
second-order accurate entropy stable scheme (4.1), whose simple numerical
470 E. Tadmor

flux is given by the centred numerical flux

  uν + uν+1 
f , ∆uν+ 1 > 0
 2 2

gν+ 1 = fν+ 1 =  (4.16)
2 2
f (uν ) + f (uν+1 )
, ∆uν− 1 ≤ 0.
2 2

We conclude this section with the following example.

Example 4.9. A well-known ‘trick’ for deriving a quadratic entropy-con-

servative approximation in the particular case of the inviscid Burgers’ equa-
tion  
∂ ∂ 1 2
u+ u = 0, (4.17)
∂t ∂x 2
is based on centred differencing of its equivalent skew-adjoint form (Tadmor
1984a)
∂ 1 ∂ 2 1 ∂
u+ [u ] + u [u] = 0,
∂t 3 ∂x 3 ∂x
which yields
d 1 1  2  1 1  
uν (t) = − uν+1 − u2ν−1 − uν uν+1 − uν−1 .
dt 3 2∆xν 3 2∆xν
In fact, there is more than just a ‘trick’ here: the resulting scheme is
simply a special case of our entropy-conservative recipe (4.6)
 
d 1 1 2 1 2
uν (t) = − u − u
dt 2∆xν 2 ν+1 2 ν−1
  2 1  2 
1 1
+ ∆uν+ 1 − ∆uν− 1 . (4.18)
2∆xν 6 2 6 2

If we exclude negative viscosity, however, then according to (4.12) the least

viscous entropy-stable approximation of Burgers’ equation (4.17) is given by
 
d 1 1 2 1 2
uν (t) = − u − u
dt 2∆xν 2 ν+1 2 ν−1
  + 
1 1 1 +
+ ∆uν+ 1 ∆uν+ 1 − ∆uν− 1 ∆uν− 1 . (4.19)
2∆xν 6 2 2 6 2 2

5. Systems of conservation laws

We study the entropy stability of the semi-discrete schemes that are consis-
tent with the system of conservation laws (2.5). The schemes assume the
 Entropy stability 471

following viscosity form:

d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Qν+ 1 ∆vν+ 1 − Qν− 1 ∆vν− 1 . (5.1)
2∆xν 2 2 2 2

Difference schemes that admit the viscosity form (5.1) are precisely the
so-called essentially three-point schemes (Harten (1983a), Tadmor (1987b,
Lemma 5.1)), namely, difference schemes whose numerical flux, f (·, ·, · · ··)
satisfies the restricted consistency relation
f (uν−p+1 , . . . , uν = uν+1 = u, . . . , uν+p ) = f (u).
This is the case of (5.1), with
1  1
f (. . . , uν , uν+1 , . . .) = f (uν ) + f (uν+1 ) − Qν+ 1 (vν+1 − vν ),
2 2 2

vν = v(uν ).
A couple of remarks are in order.

(1) The class of essentially three-point schemes includes classical schemes

based on three-point stencils (p = 1), as well as most modern high-
resolution schemes (van Leer 1977, Harten 1983a); consult Godlewski
and Raviart (1996), Kröner (1997), LeVeque (1992, 2002) and the ref-
erences therein.
(2) The use of essentially three-point stencils in this section is linked to
the specific second-order entropy-conservative schemes discussed below.
Extensions to higher orders and larger stencils were carried out by
LeFloch and Rohde (2000, Section 4); consult Section 9 below.

To extend our scalar entropy stability analysis to systems of conservation

laws we proceed as before, by comparison with certain entropy-conservative
schemes. Unlike the scalar problem, however, we now have more than one
way to meet the entropy conservation requirement (3.11). The various ways
differ in their choice of the path of integration in phase space. In this sec-
tion, we restrict our attention to the simplest choice along the straight path
vν+ 1 (ξ) = 12 (vν + vν+1 ) + ξ∆vν+ 1 . The corresponding entropy-conservative
2 2
flux is given by
Z 1  
2
∗
gν+ 1 = g vν+ 1 (ξ) dξ,
2
2 ξ=− 12

1
vν+ 1 (ξ) := (vν + vν+1 ) + ξ∆vν+ 1 . (5.2)
2 2 2
472 E. Tadmor

Indeed, the entropy conservation requirement (3.1) is fulfilled in this case,

since, in view of (2.9),
D E Z 1 D  E
2
∗
∆vν+ 1 , gν+ 1 = ∆vν+ 1 , g vν+ 1 (ξ) dξ
2 2 ξ=− 12 2 2
Z vν+1
= h dv, g(v)i = ∆ψν+ 1 .
2
vν

The entropy-conservative flux (5.2) was introduced in Tadmor (1986b,

1987b). As before (see (4.5), (4.6)), we integrate by parts to find
Z 1  
∗
2 d
gν+ 1 = (ξ)g vν+ 1 (ξ) dξ
2 ξ=− 12 dξ 2

  1 Z 1   dvν+ 1 (ξ)
2 2
2
= ξg vν+ 1 (ξ)  − ξg v v ν+ 1 (ξ) dξ
2 ξ=− 12 ξ=− 12 2 dξ
Z 1  
1  2
= f (uν ) + f (uν+1 ) − ξB vν+ 1 (ξ) dξ∆vν+ 1 . (5.3)
2 ξ=− 1 2 2
2

Thus, the entropy-conservative scheme (5.2) admits the equivalent viscosity

form
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h ∗ i
+ Qν+ 1 ∆vν+ 1 − Q∗ν− 1 ∆vν− 1 , (5.4)
2∆xν 2 2 2 2

with a numerical viscosity matrix coefficient, Q∗ν+ 1 , given by

2

Z 1  
2
Q∗ν+ 1 := 2ξB vν+ 1 (ξ) dξ, B(v) = gv (v). (5.5)
2 ξ=− 12 2

The entropy stability portion of Theorem 3.1 can now be conveniently inter-
preted as follows.

Corollary 5.1. The conservative scheme (5.1) is entropy-stable if – and

for three-point schemes (p = 1) only if – it contains more viscosity than the
entropy-conservative one (5.4), (5.5), that is,
D E D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 ≤ ∆vν+ 1 , Qν+ 1 ∆vν+ 1 . (5.6)
2 2 2 2 2 2

Indeed, we can provide a precise measure for the amount of entropy dissi-
pation in terms of the dissipation matrix Dν+ 1 ≡ Dν+ 1 (v(t)) := Qν+ 1−Q∗ν+ 1
2 2 2 2
 Entropy stability 473

(Tadmor 1987b, Theorem 5.2)

d 1 h i
U (uν (t)) + Fν+ 1 − Fν− 1 (5.7)
dt ∆xν 2 2

1 hD E 1D Ei
=− ∆vν− 1 , Dν− 1 ∆vν− 1 + ∆vν+ 1 , Dν+ 1 ∆vν+ 1
4∆xν 2 2 2 4 2 2 2

Here, Fν+ 1 stands for the entropy flux (see (3.8))

2

1D ∗
E 1 
Fν+ 1 = vν + vν+1 , gν+ 1 − ψ(vν ) + ψ(vν+1)
2 2 2 2
1 D E
− vν + vν+1 , Dν+ 1 ∆vν+ 1 . (5.8)
4∆xν 2 2

The entropy-conservative flux (5.2), and likewise its corresponding vis-

cosity coefficient in (5.5), cannot be evaluated in a closed form. How-
ever, Corollary 5.1 enables us to verify entropy stability by comparison,
Q∗ν+ 1 ≤ Re Qν+ 1 , with the usual ordering between symmetric matrices. We
2 2
note in passing that Q∗ν+ 1 is symmetric (since B(·) is) and that, in the
2
generic case, the viscosity coefficient Qν+ 1 is also symmetric. The follow-
2
ing examples demonstrate this point.

Example 5.2. (Rusanov 1961, Lax 1954) We seek a scalar viscosity

coefficient, pν+ 1 IN , which guarantees the entropy stability of the scheme
2

d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ pν+ 1 ∆uν+ 1 − pν− 1 ∆uν− 1 . (5.9)
2∆xν 2 2 2 2

Using (2.6) we have

Z 1 Z 1
2 d   2
 
∆uν+ 1 = u vν+ 1 (ξ) dξ = H vν+ 1 (ξ) dξ · ∆vν+ 1 ,
2
ξ=− 12 dξ 2
ξ=− 12 2 2

(5.10)
and hence the viscous part of the scheme (5.9) can be interpreted in terms
of the entropy variables (rather than the conservative ones), as
pν+ 1 ∆uν+ 1 = Qν+ 1 ∆vν+ 1 , (5.11)
2 2 2 2

where
Z 1  
2
Qν+ 1 = pν+ 1 H(ξ) dξ, H(ξ) ≡ H vν+ 1 (ξ) . (5.12)
2 2
ξ=− 12 2
474 E. Tadmor

Recalling (5.5), we conclude that entropy stability is guaranteed by Corol-

lary 5.1, provided pν+ 1 is chosen so that the inequality
2
  1 1
2ξB(ξ) ≤ pν+ 1 H(ξ), B(ξ) ≡ B vν+ 1 (ξ) , − ≤ ξ ≤ ,
2 2 2 2
1
holds. To this end, multiply both sides by H − 2 (ξ); by congruence, we end
up with the equivalent inequality4
 1 1 
2ξ sup λ H − 2 (ξ)B(ξ)H − 2 (ξ) ≤ pν+ 1 IN . (5.13)
2
λ
We recall that B = gv = fu uv = AH, that is,
  
B(ξ) = A(ξ)H(ξ), A(ξ) ≡ A u vν+ 1 (ξ) . (5.14)
2

Hence (5.13) holds and entropy stability follows, for any scalar pν+ 1 satis-
2
fying
 h i
 1 1 
pν+ 1 ≥ max 2ξλ H − 2 (ξ)A(ξ)H 2 (ξ) 
2 λ,|ξ|≤ 12
 h   i
 
= max λ A u vν+ 1 (ξ) . (5.15)
1 2
λ,|ξ|≤ 2

The cell-dependent viscosity factor on the right corresponds to the Rusanov

scheme (Rusanov 1961; see also Richtmyer and Morton (1967, Section 2)),
while a uniform viscosity factor, satisfying
pν+ 1 ≡ p ≥ max |λ[A(u)]|,
2 λ,u

corresponds to a Lax–Friedrichs viscosity (Friedrichs 1954, Lax 1954). Both

schemes are entropy-stable with respect to any entropy pair associated with
equation (2.1).
The last example was restricted to first-order accurate schemes. Yet Corol-
lary 5.1 can be used to maintain both entropy stability and second-order ac-
curacy, as was done in the scalar case. To this end, we proceed as follows.
Using (5.14) we can rewrite the quantity on the left of (5.6) as
D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 (5.16)
2 2 2
Z 1 D 1 E
2 1 1 1
= 2ξ H 2 (ξ)∆vν+ 1 , H − 2 (ξ)A(ξ)H 2 (ξ) · H 2 (ξ)∆vν+ 1 dξ.
2 2
ξ=− 12

Let {ak (ξ), rk (ξ)}N

k=1 be the eigenpairs of A(ξ), that is,
     
(k) (k)
ak (ξ) ≡ aν+ 1 u vν+ 1 (ξ) , rk (ξ) ≡ rν+ 1 u vν+ 1 (ξ) .
2 2 2 2

4
Here and below, λk [·] denotes the kth eigenvalue of a matrix.
 Entropy stability 475
1 1 1
Since H − 2 (ξ)rk (ξ) are the eigenvectors of the matrix H − 2 (ξ)A(ξ)H 2 (ξ),
and since, by (5.14),
1 1 1 1
 
H − 2 (ξ)A(ξ)H 2 (ξ) ≡ H − 2 (ξ)B(ξ)H − 2 (ξ), B(ξ) ≡ B vν+ 1 (ξ) , (5.17)
2

1
is a symmetric matrix, it follows after normalization that {H − 2 (ξ)rk (ξ)}
form an orthonormal system, that is,
D 1 1
E
H − 2 (ξ)rk (ξ), H − 2 (ξ)rj (ξ) = δjk . (5.18)
1
We expand H 2 (ξ)∆vν+ 1 and substitute the expansion into the right-hand
2
side of (5.16) to find
D E N Z
D E
1
X 2 2
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 = 2ξak (ξ)·  rk (ξ), ∆vν+ 1  dξ. (5.19)
2 2 2
2
k=1 ξ=− 12

Finally, we integrate by parts along the lines of (4.11), arriving at

D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1
2 2 2
N Z 1  
X 2 1 d D E
2
= −ξ 2
ak (ξ) ·  rk (ξ), ∆vν+ 1  dξ
1 4 dξ 2
k=1 ξ=− 2
N Z 1  
X 2 1 d D E
2
+ 2
− ξ ak (ξ) · rk (ξ), ∆vν+ 1  dξ. (5.20)
ξ=− 1 4 dξ 2
k=1 2

We compute
d D E
2
D E D E
2
ak (ξ) ·  rk (ξ), ∆vν+ 1  = ∇v ak (ξ), ∆vν+ 1 ·  rk (ξ), ∆vν+ 1 
dξ 2 2 2

and
d D E
2
D E D E
rk (ξ), ∆vν+ 1  = 2 · rk (ξ), ∆vν+ 1 · ∆vν+ 1 , ∇v rk (ξ)∆vν+ 1 ,
dξ 2 2 2 2

 3

and since both terms are of order O ∆vν+ 1  , it follows that the quantity
2
on the right of (5.20) does not exceed
D E  
 ∆v 1 , Q∗ 1 ∆v 1  ≤ C 1 · ∆v 1 3. (5.21)
ν+ ν+
2
ν+ 2
ν+ ν+ 2 2
2

Thus, the entropy-conservative schemes (5.2) dissipate entropy at a cubic

rate and are therefore second-order accurate: consult Lemma 4.5. Compar-
ison of this in light of Corollary 5.1 yields the following entropy stability
criterion which respects second-order accuracy.
476 E. Tadmor

Theorem 5.3. The conservative scheme

d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Qν+ 1 ∆vν+ 1 − Qν− 1 ∆vν− 1 (5.22)
2∆xν 2 2 2 2

is entropy-stable if the eigenvalues of (the symmetric part of) its viscosity

coefficient matrix, Re Qν+ 1 , satisfy
2
h i  
min λ Re Qν+ 1 ≥ Cν+ 1 · ∆vν+ 1 . (5.23)
λ 2 2 2

Next, we would like to convert this entropy stability criterion to difference

schemes which are written in the standard form
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Pν+ 1 ∆uν+ 1 − Pν− 1 ∆uν− 1 . (5.24)
2∆xν 2 2 2 2

Thus, the viscous part is expressed entirely in terms of the conservative

variables, u, instead of the entropy variables, v, used in (5.22). From the
corresponding differential set-up, (2.10), we already know that the admissible
−1
viscosity matrices for this formulation are those P s for which P Uuu are
>
symmetric positive definite, (2.11), P H −HP = 0; consequently, a discrete
analogue should hold, at least to leading order, that is,
 
P 1 H 1 − H 1 P > 1 ≤ δ 1 ∆u 1 . (5.25)
ν+ ν+
2
ν+ ν+
2
ν+ 2
ν+ 2 2
2

Here, Hν+ 1 can be any first-order symmetric approximation to the inverse

2
−1 ,
Hessian, H = Uuu
Z    
1
2
Hν+ 1 = H vν+ 1 (ξ) dξ + O ∆uν+ 1  , (5.26)
2
ξ=− 12 2 2

1
0< · IN ≤ Hν+ 1 ≤ K · IN .
K 2

Theorem 5.4. Consider the conservative difference scheme (5.24) with

numerical viscosity coefficient Pν+ 1 , which is essentially H-symmetric (5.25),
2
(5.26). The scheme is entropy-stable if the eigenvalues of its viscosity coef-
ficient matrix λ(Pν+ 1 ) satisfy
2
h i  
min λ Pν+ 1 ≥ γν+ 1 ∆uν+ 1 . (5.27)
λ 2 2 2
 Entropy stability 477

Here γν+ 1 is a suitably large constant depending on (5.21), (5.25), (5.26)

2
for which  
γν+ 1 ≥ K Cν+ 1 + δν+ 1 . (5.28)
2 2 2

Proof. Using (5.10), the scheme (5.24) meets the desired form in (5.22)
with
Z 1  
2
Qν+ 1 = Pν+ 1 H, H ≡ Hν+ 1 = H vν+ 1 (ξ) dξ.
2 2 2
ξ=− 12 2

Next, we invoke the identity

> !
  Pν+ 1 H − HPν+ 1
− 12 1
− 12 − 12 − 12 1
H−2 .
2 2
H Pν+ 1 H = H2 Re Qν+ 1 H +H
2 2 2

According to (5.27), the eigenvalues of the matrix on the left are bounded
from below by γν+ 1 · |∆uν+ 1 |. Hence, by (5.25), (5.26), the same is true for
2 2
the eigenvalues of the first matrix on the right; more precisely, we have
h 1
  1
i    
λ H − 2 Re Qν+ 1 H − 2 ≥ γν+ 1 − Kδν+ 1 · ∆uν+ 1 .
2 2 2 2

1
Multiplying on both sides by H we find, on account of (5.28),
2

h i 1 
   
λ Re Qν+ 1 ≥ γν+ 1 − δν+ 1 · ∆uν+ 1  ≥ Cν+ 1 · ∆vν+ 1 ,
2 K 2 2 2 2 2

and entropy stability follows from Theorem 5.3.

Equipped with Theorem 5.4 we turn to the following example.

Example 5.5. (Second-order accurate scalar numerical viscosity)

We re-examine Example 5.2, considering the case of scalar viscosity in (5.9),
where pν+ 1 = pν+ 1 IN . By Theorem 5.4, any scalar satisfying
2 2
 
pν+ 1 ≥ KCν+ 1 · ∆uν+ 1  (5.29)
2 2 2

will guarantee entropy stability as well as maintain second-order accuracy.

Example 5.6. (Roe-type schemes) We consider the class of schemes

based on Roe’s decomposition (Roe 1981). To this end, we introduce a
Lipschitz-continuous averaged Jacobian, the so-called Roe matrix, Aν+ 1 ,
2
satisfying
∆fν+ 1 ≡ Aν+ 1 ∆uν+ 1 , ∆fν+ 1 := f (uν+1 ) − f (uν ), (5.30)
2 2 2 2

and having a complete real eigensystem. Roe (1981) constructed such a

478 E. Tadmor

matrix for the Euler equations (2.15); Harten and Lax (1981) have shown
its existence in the general case, namely
−1
Aν+ 1 = Bν+ 1 · Hν+ 1, (5.31)
2 2 2

where Bν+ 1 and Hν+ 1 are defined by the cell averages

2 2
Z 1   Z 1  
2 2
Bν+ 1 = B vν+ 1 (ξ) dξ, Hν+ 1 = H vν+ 1 (ξ) dξ. (5.32)
2 2 2 2
ξ=− 12 ξ=− 12

Given a Roe matrix, the viscosity coefficient in (5.24), Pν+ 1 , is then set
2
to be  
Pν+ 1 = p Aν+ 1 . (5.33)
2 2

Here p(·) is an appropriate viscosity function which is computed according

to the spectral decomposition of Aν+ 1 , namely,
2
 
  p(a1 )
 ..  −1
p Aν+ 1 = Rν+ 1 ·  .  · Rν+ 1 . (5.34)
2 2 2
p(aN )
where {(a)N
1 , Rν+ 21 } is the eigensystem of A, that is,
 
a1 h i
 ..  −1
Aν+ 1 = Rν+ 1 ·  .  · Rν+ 1 , ak := λk Aν+ 1 .
2 2 2 2
aN
For a given system, the possibly various choices of a Roe matrices are
within O(|∆uν+ 1 |) of each other; since the set-up of Theorem 5.4 is in-
2
variant under such perturbations we can discuss without restriction the one
choice given in (5.31)–(5.32). With this choice of a Roe matrix we have
  1
 Z 1   
2
− 21 2 − 12 −1
Pν+ 1 = p Aν+ 1 = Hν+ 1 · p Hν+ 1 · B vν+ 1 (ξ) dξ · Hν+ 1 · Hν+21 ,
2 2 2 2 ξ=− 12 2 2 2

and hence, by the symmetry of B, it follows that P is H-symmetric, so that

(5.25) holds with δν+ 1 = 0. Theorem 5.4 applies and we are led to the
2
following.
Theorem 5.7. The conservative Roe-type scheme (5.24), (5.33), (5.34) is
entropy-stable, provided that its viscosity function p(·) satisfies
  h i
p(ak ) ≥ KCν+ 1 · ∆uν+ 1 , ak := λk Aν+ 1 . (5.35)
2 2 2

There are various ways of choosing a viscosity function p(·) satisfying

(5.35), which give rise to either first- or second-order accurate entropy-
 Entropy stability 479

stable schemes. We start by discussing the pros and cons of some first-order
choices.
The original choice of Roe (1981) employs the viscosity function
p(ak ) = |ak |. (5.36)
It has the desirable property that discrete steady shocks are perfectly resolved
on the grid. With this choice, the entropy stability requirement (5.35) reads
 
|ak | ≥ KCν+ 1 · ∆uν+ 1 ,
2 2

and it is fulfilled as long as we are away from sonic points. Yet this require-
ment may be violated in sonic neighbourhoods where ak ≈ 0; indeed, Roe’s
scheme is the canonical example of an entropy-unstable scheme, for it admits
steady expansion shocks. Theorem 5.7 suggests a simple modification – first
proposed by Osher (1985, Theorem 3.3) – in order to avoid such instability.
Example 5.8. (First-order entropy fix of the Roe scheme) The Roe
scheme (5.24), (5.33), (5.34), is entropy-stable with a viscosity function
n  o
p(ak ) = max |ak |, KCν+ 1 · ∆uν+ 1  . (5.37)
2 2

The slightly more viscous modification of Harten (1983a) takes the form
  
p(ak ) = max |ak |, ε , ∆uν+ 1   1. (5.38)
2

In these cases entropy stability is achieved by adding viscosity near sonic

points, regardless of whether they occur in rarefaction or shock waves. This
is done at the expense of destroying the sharp steady shock resolution of
Roe’s original scheme (5.36).
However, we can do better with regard to Roe-type schemes, by sharpening
the general sufficient entropy stability condition (5.27) which led us to The-
orem 5.7. To this end, we first note that the eigensystem of a Roe matrix
in (5.30), Aν+ 1 , is within O(|∆uν+ 1 |2) from the eigensystem of the exact
2 2
mid-value Jacobian, say A(ξ = 0), for example,
 2
|rk − rk (ξ = 0)| + |ak − ak (ξ = 0)| ≤ Const∆uν+ 1  . (5.39)
2

By virtue of (5.39) we can obtain rather detailed information about the en-
tropy dissipation rate of the Roe-type schemes (5.24), (5.33), (5.34).
Let ∆ak (uν ) denote the jump in the kth eigenvalue
∆ak (uν ) = λk (A(uν+1 )) − λk (A(uν )).
In the Appendix we prove the following theorem.
Theorem 5.9. The conservative Roe-type scheme (5.24), (5.33), (5.34) is
480 E. Tadmor

entropy-stable if its viscosity function, p(·), satisfies

   
1 |ak |  2
p(ak ) ≥ ∆ak (uν ) + εk |ak | + 1 +  
Const ∆uν+ 1 . (5.40)
6 εk 2

Here εk > 0 are arbitrary parameters at our disposal.

Remark. We note that the essential ingredient of Theorem 5.9 is not
the Roe averaging property (5.30), but the requirement that the eigensystem
 2

of Aν+ 1 be within O ∆uν+ 1  of the eigensystem of A(ξ = 0), (5.39).
2 2
Hence, Theorem 5.9 and its consequences applyto other Roe averages,
 for
1 1
example, Aν+ 1 = A[ 2 (uν + uν+1 )] or Aν+ 1 = 2 A(uν ) + A(uν+1 ) .
2 2

We shall apply Theorem 5.9 to hyperbolic systems which contain either

genuinely nonlinear (GNL) or linearly degenerate waves. Lax (1957) has
shown that any two nearby states in such systems, uν and uν+1 , can be
connected by a certain continuous path in phase space; the jump from uν to
uν+1 is resolved into a succession of k-waves, k = 1, 2, . . . , N , each of which
is either a k-shock, a k-contact or a k-rarefaction, depending on whether ak
increases, remains constant or decreases, respectively, along the correspond-
ing kth subpath. In this section we answer the entropy stability question
by comparison with the entropy-conservative schemes (5.2) which are based
upon integration along a simple straight path in phase space. Therefore,
these schemes do not resolve the full structure in phase space of the solu-
tion path to the Riemann problem described above. Instead, we shall confine
ourselves to identify each cell with one dominant k-wave. This fact of one
dominant wave per cell is certainly the case with GNL scalar problems and,
as observed by Harten (1983a), is also valid in actual computations with the
gas dynamics system (2.15). We will refine our stability analysis in Sec-
tion 6 below, in terms of new entropy conservative schemes which do take
into account different subpaths in phase space.
Choosing εk = 6 in (5.40), then the following entropy-stable modification
of the first-order Roe-type scheme is obtained.
Example 5.10. (Modified Roe scheme revisited) The conservative
Roe-type scheme (5.24), (5.33), (5.34) is entropy-stable with viscosity func-
tion
 
1  2 +

p(ak ) = |ak | + ∆ak (uν ) + Const ∆uν+ 1  . (5.41)
6 2

This choice of viscosity function was suggested in Harten and Hyman (1983,
Appendix A) and numerical simulations were carried out in Kaddouri (1993),
for example. To gain a better insight into this choice, we shall distinguish
between three different cases.
   2

Case I. ∆ak (uν ) ≤ −Const∆u 1  + O ∆u 1  .
ν+ 2 ν+ 2
 Entropy stability 481

In this case the jump ∆ak is dominated by a k-shock, and with sufficiently
small variation, no additional viscosity is required in (5.36), i.e., (5.41) is
reduced to p(ak ) = |ak |. Thus, (5.41) retains the perfect resolution of (suf-
ficiently weak) discrete steady shocks.
 2

Case II. ∆ak (uν ) = O ∆uν+ 1  .
2
In this case the jump ∆ak is essentially due to the k-contact field and/or
the balance between the other fields. Here, a minimal
 amount
2 of viscosity
 
is required near sonic points p(ak ) = |ak | + Const ∆uν+ 1 .
2

Case III. Finally, in all other cases we shall identify  the jump
∆ak (uν )
as dominated by a k-rarefaction, and as expected, O ∆uν+ 1  amount of
2 
dissipation is required near sonic points, p(ak ) = |ak | + Const∆u 1 . ν+ 2

This concludes our discussion of the first-order accurate

 Roe-type
 schemes
and we turn to the second-order case. Choosing εk ∼ ∆uν+ 1  in (5.40), we

2
find p(ak ) = ∆ak (uν ) + Const |ak |∆uν + |∆uν |2 . Thus, if we set
 
p(ak ) = Const∆uν+ 1 , (5.42)
2

then the resulting Roe-type scheme (5.24), (5.33), (5.34) is second-order

accurate by Lemma 4.5, and it is entropy-stable for sufficiently large Const ≥
KCν+ 1 : consult Theorem 5.7.
2
The examples studied so far are based on a priori (positive) bounds for the
entropy-conservative viscosity. We close this section with the following ex-
ample, which shows how to enforce entropy stability a posteriori by carefully
removing any viscosity production. We start with an essentially three-point
scheme in its conservative variables formulation (5.24),
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Pν+ 1 ∆uν+ 1 − Pν− 1 ∆uν− 1 , (5.43)
2∆xν 2 2 2 2

and we compare it with the corresponding formulation of the conservative

scheme (5.4),
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h ∗ ∗
i
+ Pν+ 1 ∆uν+ 1 − Pν− 1 ∆uν− 1 ,
2∆xν 2 2 2 2
Z 1  
2
∗ ∗ −1 ∗ −1 ∗ −1
where Pν+ 1 := Q
ν+ 1 (H ) , (H ) := H u ν+ 1 (ξ) dξ.
2 2 ξ=− 12 2
482 E. Tadmor

According to (5.7) which we express in terms of the conservative variables,

the entropy production of (5.43) is quantified by
D E
∗
eν+ 1 := ∆vν+ 1 , Eν+ 1 ∆uν+ 1 , Eν+ 1 := Pν+ 1 − Pν+ 1 ,
2 2 2 2 2 2 2

and we arrive at the following example.

Example 5.11. (Khalfallah and Lerat 1988) The scheme (5.43) be-
comes entropy-stable if we add a minimal amount of (scalar) viscosity cor-
rection, pcν+ 1 , replacing Pν+ 1 with
2 2
(eν+ 1 )+
Pν+ 1 −→ Pν+ 1 + pν+ 1 IN ×N , pc :=

c 2
. (5.44)
2 2 2 ∆vν+ 1 , uν+ 1
2 2

We note that the quantity on the right is well defined since the denominator
does not vanish h(H −1)∗ u, ui > 0. The correction preserves second-order
accuracy, and Lerat and his co-workers (Khalfallah and Lerat 1989) report
on successful applications of such entropy correction in numerical simula-
tions of fluid dynamics problems. Let us point out two limitations to the
present approach: (i) we need to compute the entropy-conservative term
P ∗ u = Q∗ v, which might not be readily available, and (ii), as before, the
entropy correction does not distinguish between different waves within the
same cell. Both points are addressed in the context of the new entropy-
conservative schemes introduced in the next section: consult (6.12) below,
for example.

6. Entropy-conservative schemes revisited

Our study of entropy stability is based on comparison with entropy-con-
servative schemes. In the scalar case, entropy-conservative schemes are
unique (for a given entropy pair). For systems, there are various choices for
numerical fluxes which meet the entropy conservation requirement (3.11). In
Section 5 we restricted our attention to just one such choice. In this section
we present the general framework.
The entropy-conservative schemes treated in Section 5 are based on inte-
gration along a straight path in phase space. Consequently, the corresponding
entropy stability analysis, for instance, Example 5.10, took into account only
one dominant wave per cell. In contrast, in this section we introduce a new
general family of entropy-conservative schemes which are based on differ-
ent paths in phase space. This enables us to enforce entropy stability by
fine-tuning the amount of numerical viscosity along each subpath carrying
different intermediate waves. Moreover, the straight path integration of the
entropy-conservative flux (5.2) does not admit a closed form, whereas the
new family of entropy-conservative schemes enjoys an explicit, closed-form
formulation. To this end, at each cell consisting of two neighbouring values
 Entropy stability 483
 N
vν and vν+1 , we let rjν+ 1 j=1 be an arbitrary set of N linearly indepen-
 2 N
dent N -vectors, and let `jν+ 1 j=1 denote the corresponding orthogonal set,

j k
 2
 j N
`ν+ 1 , rν+ 1 = δjk . Next, we introduce the intermediate states, vν+ 1 j=1 ,
2 2 2
1
starting with vν+ 1 = vν , and followed by
2
D E
j+1 j j
vν+ 1 = vν+ 1 + `ν+ 1
, ∆v ν+ 1 rjν+ 1 , j = 1, 2, . . ., N, (6.1)
2 2 2 2 2

thus defining a path in phase space, connecting vν to vν+1 ,

N D
X E
j
N +1
vν+ 1 = v 1
ν+ 1 + ` ν+ 1
, ∆v ν+ 1 rjν+ 1 = vν + ∆vν+ 1 ≡ vν+1 . (6.2)
2 2 2 2 2 2
j=1

Since the mapping u 7→ v is one-to-one, the path is mirrored in the usual


N +1
phase space of conservative variables, ujν+ 1 := u vν+
j
1 j=1
, starting with
2 2
u1ν+ 1 = uν and ending with uN +1
ν+ 1
= uν+1 . Equipped with this notation we
2 2
turn to our next result.
Theorem 6.1. The conservative scheme
d 1 h ∗ ∗
i
uν (t) = − gν+ 1 − gν− 1 ,
dt ∆xν 2 2

∗
with a numerical flux gν+ 1 given by
2

j+1
j

XN ψ v 1 − ψ v 1
∗ ν+ ν+ 2 j
gν+ 1 =
j 2  `ν+ 1 , (6.3)
2
j=1 `ν+ 1 , ∆vν+ 1 2
2 2

is an entropy-conservative approximation consistent with (2.5). Here, ψ is

the entropy flux potential associated with the conserved entropy pair (U, F ).
Remark. We note that the quantities on the right of (6.3) are well defined:
consult (6.6) below.
Proof. The entropy conservation requirement (3.11) follows directly from
(6.2) for
j+1
j

D E X N ψ v
ν+ 1 − ψ v ν+ 21 D E
∗
∆vν+ 1 , gν+ 1 =
j 2  `jν+ 1 , ∆vν+ 1
2 2
j=1 `ν+ 1 , ∆vν+ 1 2 2
2 2

N
X    
j+1 j
= ψ vν+ 1 − ψ v ν+ 1 2 2
j=1
   
N +1 1
= ψ vν+ 1 − ψ vν+ 1 = ∆ψν+ 1 .
2 2 2
484 E. Tadmor

It remains to verify the consistency relation (3.7). Let

j+ 1 1 j j+1
 D
j
E
j 1 1
vν+ 21 (ξ) := vν+ 1 +vν+ 1 +ξ ` ν+ 12
, ∆v ν+ 2 rν+ 1 ,
1 − ≤ ξ ≤ , (6.4)
2 2 2 2 2 2 2
j j+1
denote the straight subpath connecting vν+ 1 and v
ν+ 1
; then we can use
2 2
(2.9) to express the ψ-potential jump between two consecutive intermediate
states as
    Z 12 d  j+ 12 
j+1 j
ψ vν+ 1 − ψ v ν+ 12
= ψ vν+ 1 (ξ) dξ
2 ξ=− 12 dξ 2

Z 1  j+ 1  D E
2
j
= 2
g vν+ 1 (ξ) dξ, rν+ 1 `jν+ 1 , ∆vν+ 1 .
ξ=− 12 2 2 2 2

(6.5)
Inserting this into (6.3), we find that the entropy-conservative flux can be
equivalently written as
XN Z 1  j+ 1  
2
∗
gν+ 1 = g vν+ 1 (ξ) dξ, rν+ 1 `jν+ 1 ,
2 j
(6.6)
2
j=1 ξ=− 12 2 2 2

and consistency is now obvious:

N D
X E
g∗(v, v) = g(v), rjν+ 1 `jν+ 1 = g(v). (6.7)
2 2
j=1

n j+ 12
oN +1
Remark. We note that if we let rν+ 1 collapse into the same direc-
2 j=1
tion of ∆vν+ 1 , then the new entropy-conservative flux (6.5) collapses into
2
the entropy-conservative flux of the ‘first kind’ studied earlier in Section 5.
As before, the new entropy-conservative schemes admit a viscosity form,
subject to the phase space path. Considering a typical subpath factor on the
right of (6.6), we integrate by parts along the lines of (5.3), to obtain
Z 1 D  j+ 1  E
2 d 2 j
(ξ) g vν+ 1 (ξ) dξ, rν+ 1
ξ=− 12 dξ 2 2

1D  j  
j+1

j
E
= f uν+ 1 + f uν+ 1 , rν+ 1
2 2 2 2
Z 1 D   ED E
2 j+ 1
+ ξ rjν+ 1 , B vν+ 21 (ξ) rjν+ 1 `jν+ 1 , ∆vν+ 1 .
ξ=− 12 2 2 2 2 2

This yields the following family of entropy-conservative schemes.

 Entropy stability 485

Corollary 6.2. Given a complete path in phase space,

n  oN +1
ujν+ 1 := u vν+
j
1 ,
2 2 j=1


associated with left and right orthogonal sets `jν+ 1 , rkν+ 1 = δjk , where
2 2
rjν+ 1 is in the direction of vν+
j+1
1 −v
j
ν+ 1
. Then we have the following entropy-
2 2 2
conservative scheme:

1 X D  j    E
N
d
uν (t) = − f uν+ 1 + f uj+1
ν+ 12
, r j
ν+ 21
`jν+ 1
dt 2∆xν 2 2
j=1

XN D     E
− f ujν− 1 + f uj+1
ν− 1
, rjν− 1 `jν− 1 
2 2 2 2
j=1

1 X D j ED E
N
j+ 1 ,∗
+ rν+ 1 , Qν+ 21 rjν+ 1 `jν+ 1 , ∆vν+ 1 `jν+ 1
2∆xν 2 2 2 2 2 2
j=1

N D
X ED E
j+ 1 ,∗
− rjν− 1 , Qν− 21 rjν− 1 `jν− 1 , ∆vν− 1 `jν− 1  ,
2 2 2 2 2 2
j=1
Z 1  j+ 1 
j+ 1 ,∗ 2
Qν+ 21 := 2ξB vν+ 21 (ξ) dξ. (6.8)
2 ξ=− 12 2

The viscosity form of the entropy-conservative scheme outlined in Corol-

lary 6.2 is a refinement of the entropy-conservative schemes (5.4). In par-
ticular, we can revisit the examples of entropy-stable recipes outlined in Sec-
tion 5, using the two ingredients of (i) comparison with entropy-conservative
schemes, and (ii) a proper choice of path in phase space. We continue with
a discussion of these two ingredients.

j+ 1
(i) Comparison. We seek appropriate viscosity amplitudes, qν+ 21 , which
2
upper-bound the amount of entropy-conservative viscosities on each subpath
j+ 1
in phase space, vν+ 21 (ξ), so that (compare Corollary 5.1)
2

D j+ 1 ,∗
E j+ 1
rjν+ 1 , Qν+21 rjν+ 1 ≤ qν+ 21 . (6.9)
2 2 2 2

A straightforward argument along the lines of our previous results yields the
following result.
486 E. Tadmor

Theorem 6.3. The semi-discrete scheme


1 X D  j    E
N
d
uν (t) = − f uν+ 1 + f uj+1
ν+ 1 , r j
ν+ 1 `
j
ν+ 12
dt 2∆xν 2 2 2
j=1

D     E
− f uj 1 + f uj+11 , rj 1 `j 1 
ν− 2 ν− 2 ν− 2 ν− 2


1 X j+ 12 D j E
N
+ qν+ 1 `ν+ 1 , ∆vν+ 1 `jν+ 1
2∆xν 2 2 2 2
j=1

N
X j+ 1
D E
− qν− 21 `jν− 1 , ∆vν− 1 `jν− 1 , (6.10)
2 2 2 2
j=1

is entropy-stable if it contains more numerical viscosity than the entropy-

conservative one in the sense that (6.9) holds.

(ii) Choice of path. The new ingredient here is the choice of a proper sub-
path in phase space. We demonstrate the advantage of using such a subpath
in the context of second-order accurate reformulation of the conservative
schemes outlined in Corollary 6.2. Let
n  j+ 1 o
wk (v(ξ)) = wk vν+ 21 (ξ)
2

 j+ 1 
be the orthonormal eigensystem of the symmetric B = B vν+ 21 (ξ) ,
2

j+ 1
   j+ 1 
B vν+ 21 (ξ) wk (v(ξ)) = bk (v(ξ))wk(v(ξ)), bk (v(ξ)) := λk B vν+ 21 .
2 2

j P
j 
Expanding rν+ 1 = k wk (v(ξ)), rν+ 1 wk (v(ξ)), we rewrite the amount of
2 2
entropy-conservative viscosity corresponding to a typical subpath on the left
of (6.9)
D j+ 1 ,∗
E Z 1
2
D  j+ 1  E
rjν+ 1 , Qν+ 21 rjν+ 1 = 2ξ rjν+ 1 , B vν+ 21 (ξ) rjν+ 1 dξ
2 2 2 ξ=− 12 2 2 2

N Z
X 1 D E2
2
= 2ξbk (v(ξ)) wk (v(ξ)), rjν+ 1 dξ.
k=1 ξ=− 12 2

Simple upper bounds, for instance, 2ξbk (v(ξ)) ≤ supξ |bk (v(ξ))|, characterize
the first-order Roe-type schemes. For second-order accuracy, we perform one
 Entropy stability 487

more integration by parts along the lines of (5.20):

D j+ 1 ,∗
E
rjν+ 1 , Qν+ 21 rjν+ 1
2 2 2
N Z
X 1 
hD ED E2
2 1
= − ξ2 ∇v bk (v(ξ)), rjν+ 1 wk (v(ξ)), rjν+ 1 dξ
1 4 2 2
k=1 ξ=− 2
D Ei
+ 2bk (v(ξ)) rjν+ 1 , ∇v wk (v(ξ))rjν+ 1 dξ. (6.11)
2 2

Here,
 second-order

accuracy is reflected by viscosity amplitudes of order
O ∆vν+ 1  along each subpath (being entropy-conservative, the amount of
2
entropy dissipation is zero). How should we choose an appropriate subpath?
To simplify matters we consider the symmetric case where the entropy and
 N
conservative variables coincide, B(v) = A(u). We let ujν+ 1 j=1 be the
2
breakpoints along the path of (approximate) solutions to the Riemann prob-
lem. It is well known (Lax 1957) that each subpath is directed along the


eigensystem of A ujν+ 1 , that is, uj+1
ν+ 1
− ujν+ 1 ∼ rjν+ 1 , so wk ∼ rkν+ 1 is the
2 2 2 2 2
normalized eigensystem of A. With this choice, all but one of the terms on
the right of (6.11) vanish to higher order (in |∆uν+ 1 |) and the leading term
2
governing entropy dissipation is given by
D j+ 1 ,∗
E Z 12  1 D  j+ 1  E
rjν+ 1 , Qν+21 rjν+ 1 ≈ − ξ 2 ∇uaj uν+ 21 (ξ) , rjν+ 1 dξ.
2 2 2 ξ=− 1 4 2 2
2

The last expression captures the essence of the entropy-conservative schemes

that balance between entropy dissipation along j-shocks, where
D E
∇uaj (u(ξ)), rjν+ 1 > 0,
2

and the entropy production along j-rarefactions, where

D E
∇v aj (u(ξ)), rjν+ 1 < 0.
2

To enforce entropy stability, we need to increase the amount of numerical

viscosity. The use of different subpaths allows us to stabilize rarefactions
while avoiding spurious entropy dissipation with shocks. A detailed study
for the general nonsymmetric case requires lengthy calculations, and can be
carried out along the lines of the Appendix. Here we note a simple entropy-
stable correction by turning off the entropy production along the rarefactions,
j+ 1
leading to viscosity amplitude, qν+ 21 , acting along the j-wave,
2
Z 1  D  j+ 1  E+
j+ 1 2 1
qν+ 21 = −ξ 2
∇u aj uν+ 21 (ξ) , rjν+ 1 dξ. (6.12)
2 ξ=− 12 4 2 2
488 E. Tadmor

We conclude this section with the following two corollaries.

Corollary 6.4. The difference scheme (6.10), (6.12) is a second-order ac-
curate entropy-stable approximation of (2.1). No artificial dissipation is
added in shocks and, in particular, it has the desirable property of keeping
the sharpness of shock profiles.
Next, we note that if the path connecting ujν+ 1 and uj+1
ν+ 1
is chosen along
2 2
the (approximate) Riemann solution, then the integrand on the right of
(6.12) does not change sign. A simple upper bound of the entropy-conservative
amplitude on the right of (6.12) along the lines of Example 4.7 yields an
entropy-stable Lax–Wendroff-type viscosity


+
Z 1  D E+ aj uj+1 − aj ujν+ 1
2 1 1 ν+ 1
− ξ 2 ∇u aj (u(ξ)), rjν+ 1 dξ ≤
j2 2 .
ξ=− 1 4 2 4 ` 1 , ∆vν+ 1
2 ν+ 2 2
(6.13)
This yields our next result.
Corollary 6.5. The following Lax–Wendroff-type difference scheme is a
second-order accurate entropy-stable approximation of (2.1):

1 X D  j    E
N
d
uν (t) = − f uν+ 1 + f uj+1
ν+ 12
, r j
ν+ 21
`jν+ 1
dt 2∆xν 2 2
j=1

D     E
− f uj 1 + f uj+11 , rj 1 `j 1 
ν− 2 ν− 2 ν− 2 ν− 2


1 X h  j+1   i+
N
+ aj uν+ 1 − aj ujν+ 1 `jν+ 1
8∆xν 2 2 2
j=1

N h 
X   i+
− aj uj+1
ν− 1 − a j u j
ν− 1 `jν− 1  . (6.14)
2 2 2
j=1

No artificial dissipation is added in shocks and in particular, it has the

desirable property of keeping the sharpness of shock profiles.

7. Entropy stability of fully discrete schemes

In this section we study the time discretizations of the semi-discrete entropy-
stable schemes
d 1 h i
uν (t) = − gν+ 1 − gν− 1 (7.1)
dt ∆xν 2 2
 Entropy stability 489

with essentially three-point numerical flux

1 1
gν+ 1 (v(t)) = [f (uν (t)) + f (uν+1 (t))] − Qν+ 1 ∆vν+ 1 . (7.2)
2 2 2 2 2

According to Corollary 5.1, the semi-discrete scheme is entropy-stable

if it contains more numerical viscosity than entropy-conservative schemes,
namely Q∗ν+ 1 ≤ Re Qν+ 1 . We recall (5.7), which allows us to measure the
2 2
amount of entropy dissipation in (7.1) in terms of the dissipation matrices
Dν+ 1 ≡ Dν+ 1 (v(t)) := Qν+ 1 − Q∗ν+ 1 . (7.3)
2 2 2 2

The spatial part of (7.1) satisfies

D h iE h i 1 (x)
vν (t), gν+ 1 − gν− 1 = Fν+ 1 − Fν− 1 + E (v(t)), (7.4)
2 2 2 2 ∆xν ν
(x)
with Fν+ 1 being the entropy flux specified by (5.8) and Eν denoting the
2
amount of entropy dissipation due to spatial discretization in (7.1), given by
1D E 1D E
Eν(x) := ∆vν− 1 , Dν− 1 ∆vν− 1 + ∆vν+ 1 , Dν+ 1 ∆vν+ 1 ≥ 0. (7.5)
4 2 2 2 4 2 2 2

We note in passing that use of the dissipation matrix Dν+ 1 is restricted here
2
to entropy-conservative schemes of the ‘first kind’ discussed in Section 5,
and we can use a similar, refined argument with the entropy-conservative
schemes of the ‘second kind’ in Section 6, leading to the corresponding gen-
eralization of the fully discrete entropy stability analysis presented below.
1
To discretize in time, we introduce a local time step, tn+1 = tn + ∆tn+ 2 .
We shall use superscripts to denote dependence on the time level, for in-
stance, unν = u(v(xν , tn )), gν+
n n
1 = gν+ 1 (v(t )), etc. To simplify notation,
2 2
we suppress the variability of the time step and grid cell width, abbreviating
1
∆tn+ 2 /∆xν = ∆x
∆t
. We shall study the entropy stability of the fully discrete
schemes in terms of three prototype examples, which demonstrate the balance
between the entropy dissipation from spatial stencil vs. the entropy dissipa-
tion/production due to the time discretization. We begin with the following.

Example 7.1. (Implicit backward Euler (BE) time discretization)

We discretize (7.1) by the backward Euler scheme
∆t h i
un+1
ν = u n
ν − g 1 (v n+1
) − g 1 (v n+1
) , vn+1 = v(u(tn+1)). (7.6)
∆x ν+ 2 ν− 2

We claim that the fully implicit time discretization in (7.6) is unconditionally

entropy-stable. Indeed, implicit time discretization is responsible for addi-
tional entropy dissipation. For a quantitative measure of this statement, we
490 E. Tadmor

invoke the identity

Z
d   n+ 12 
1
2
U (u(vνn+1)) − U (u(vνn)) ≡ U u vν (ξ) dξ
ξ=− 12 dξ
Z 1 D n+ 1  n+ 1  n+ 1 E
2
= vν 2 (ξ), H vν 2 (ξ) ∆vν 2 dξ, (7.7)
ξ=− 12

where the following abbreviation is used:

n+ 1 1
n+ 1 n+ 1
vν 2 (ξ) = vνn+1 + vνn + ξ∆vν 2 , ∆vν 2 := vνn+1 − vνn. (7.8)
2
Rearranging the last term on the right of (7.7), we find that time discretiza-
tion yields

n+1 n+1 

1

vν , uν − unν = U un+1 ν − U unν + EνBE vn+ 2 , (7.9)
where EνBE measures the entropy dissipation due to time discretization by
backward Euler differencing:
Z 1  D
BE n+ 12

2 1 n+ 1 n+ 1
n+ 1 E
Eν v := − ξ ∆vν 2 , H vν 2 (ξ) ∆vν 2 dξ ≥ 0.
ξ=− 12 2
(7.10)
Returning to (7.6), we multiply by vνn+1 and obtain entropy dissipation
from both the spatial discretization (7.4), (7.5) and time discretization (7.9),
(7.10)

∆t h n+1 i
U un+1
ν − U (unν ) + n+1
Fν+ 1 − Fν− 1
∆x 2 2

n+1 n+1 n
 ∆t D n+1 h

iE
= vν , uν − uν + vν , gν+ 1 vn+1 − g vn
∆x 2

∆t (x) n+1
1

− E v − EνBE vn+ 2
∆x ν
∆t (x) n+1
1

=− Eν v − EνBE vn+ 2 ≤ 0. (7.11)
∆x
Entropy stability is enhanced by fully implicit time discretization. In con-
trast, explicit time discretization, discussed in the next example, leads to
entropy production. Thus, the entropy stability of explicit schemes hinges
on a delicate balance between temporal entropy production and spatial en-
tropy dissipation.
Example 7.2. (Explicit forward Euler (FE) time discretization)
We discretize (7.1) by the forward Euler scheme
∆t h i
un+1
ν = u n
ν − g 1 (v n
) − g 1 (v n
) . (7.12)
∆x ν+ 2 ν− 2
 Entropy stability 491

Now, the identity (7.7), (7.8) can be put into the equivalent form

n n+1 

1

vν , uν − unν = U un+1 ν − U unν − EνF E vn+ 2 , (7.13)
1
with entropy production EνF E (vn+ 2 ) given by
Z 1  D
FE n+ 12

2 1 n+ 1 n+ 1
n+ 1 E
Eν v := + ξ ∆vν 2 , H vν 2 (ξ) ∆vν 2 dξ ≥ 0.
ξ=− 12 2
(7.14)
We multiply (7.12) by vνn , and together with the spatial dissipation of en-
tropy quantified in (7.5), we arrive at

∆t h n i ∆t (x) n 1

U un+1
ν − U (u n
ν ) + F ν+ 1 − F n
ν− 1 =− Eν (v ) + EνF E vn+ 2 .
∆x 2 2 ∆x
(7.15)
To study the entropy stability of (7.12), we therefore need to upper-bound
the entropy production EνF E , in terms of the spatial dissipation matrices
Dν± 1 , which are responsible for the entropy dissipation in (7.5). We proceed
2
as follows. From (7.14) we have
Z 1  D
1
2 1 n+ 1 n+ 1
n+ 1 E
EνF E vn+ 2 ≤ + ξ ∆vν 2 , H vν 2 (ξ) ∆vν 2 dξ
ξ=− 1 2
2

K  n+ 1 2 K 3  n+ 12 2
≤ ∆vν 2  ≤ ∆uν , (7.16)
2 2
where K 2 is the condition number of H: see (5.26). To upper-bound the time
n+ 12 n ) − g(v n ) = B n
differences, ∆uν , we recall g(vν+1 ν ν+ 12 ∆vν+ 1 with Bν+ 12
R n

2
given in (5.32) as Bν+ 1 = B vν+ 1 (ξ) dξ. This enables us to rewrite the
2 2
discrete forward Euler scheme (7.12) in the equivalent incremental form
∆t h


un+1
ν − u n
ν = g v n
ν+1 − g v n
ν
n
+ Qν+ 1 ∆vν+ 1
2∆x 2 2


 i
+ g vνn − g vν−1 n n
+ Qν− 1 ∆vν− 1
2 2

∆t h 
n
 
n
i
= Bν+ 1 + Qν+ 1 ∆vν+ 1 + Bν− 1 + Qν− 1 ∆vν− 1 .
2∆x 2 2 2 2 2 2

Finally, we recall the viscosity matrix Qν+ 1 = Q∗ν+ 1 + Dν+ 1 . This enables
2 2 2
us to rewrite the last expression as
∆t h e   
e 1 + D 1 ∆vn 1 ,
i
uνn+1 − unν = n
Bν+ 1 + Dν+ 1 ∆vν+ 1 + B ν− 2 ν− 2 ν− 2
2∆x 2 2 2
(7.17)
492 E. Tadmor

where
Z    
1
2
e := B + Q∗ =
B n
(1 + 2ξ)B vν+ 1 dξ = Bν+ 1 + O ∆vν+ 1  .
ξ=− 12 2 2 2

Squaring (7.17) we find

 
 n+ 1 2 1 ∆t 2hD  2 E
∆uν 2  ≤ ∆vν+ 1 , B e 1 + D 1 ∆v 1
ν+ 2 ν+ 2 ν+ 2
2 ∆x 2
D  2 Ei
+ ∆vν− 1 , B e 1 + D 1 ∆v 1 . (7.18)
2
ν− ν− 2
ν− 2 2

Compared with the spatial entropy dissipation in (7.5), we find that the
1
∆t (x) n
forward Euler scheme is entropy-stable, − ∆x Eν (v ) + EνF E (vn+ 2 ) ≤ 0,
provided D is sufficiently large that
 2 2
3 ∆t e 1 + D 1 ≤ ∆t D 1 .
K B ν+ 2 ν+ 2 (7.19)
∆x ∆x ν+ 2
We consider the two prototype examples of centred and upwind schemes.
∆x
If we set Dν+ 1 = 2∆t IN ×N we obtain the centred modified Lax–Friedrichs
2
(MLxF) scheme (e.g., Tadmor (1984b))
1 n
∆t 


un+1
ν = uν+1 + 2unν + unν−1 + f unν+1 − f unν−1 . (7.20)
4 2∆x
To simplify matters, we consider the symmetric case, where the Bs are
turned into As, and (7.19) with condition number K = 1 yields the entropy
stability of the MLxF for sufficiently small CFL number
√
∆t e 2−1
max |λ(A)| ≤ .
∆x λ 2
e 1 | leads to the
Similarly, the viscosity coefficient matrix, Dν+ 1 = |A
2
ν+ 2
upwind scheme
∆t 


un+1
ν = unν − f unν+1 − f unν−1
2∆x
1 h  
+ Aν+ 1 + Q∗ν+ 1  + Q∗ν+ 1 ∆vν+n
1
2∆xν 2 2 2 2
   i
 ∗ 
− A 1 + Q 1 + Q 1 ∆v 1 .∗ n
(7.21)
ν− 2 ν+ 2 ν− 2 ν− 2

According to (7.19), entropy stability follows under the CFL condition

∆t   
max λ A e 1  ≤ 1/4, A
ν+ 2
e := A + Q∗ .
∆x λ,ν
 Entropy stability 493

We conclude this fully explicit example with several remarks.

(1) CFL optimality. In both examples of the centred and upwind schemes,
entropy stability is obtained under less than optimal CFL conditions,
which is due to less than optimal bounds on the entropy production
rate, EνF E . In particular, the resulting entropy stability condition (7.19)
excludes the entropy stability of second-order fully discrete schemes,
which are identified with Lax–Wendroff (LxW) dissipation matrices of
∆t 2
order Dν+ 1 ∼ ∆x Ãν+ 1 .
2 2

(2) Entropy stability of Lax–Wendroff scheme. For first-order accurate

schemes, sharp CFL entropy stability conditions would follow from an
alternative approach discussed in Section 8 below. The question of
entropy stability for second-order fully discrete schemes, however, is
more delicate. It would be desirable to refine the above arguments
to obtain an improved CFL condition, which in particular entertains
the second-order case. For a systematic approach to enforcing entropy
stability of the second-order scalar LxW scheme we refer to Majda
and Osher (1978, 1979). We note the limitation that entropy stabil-
ity places on fully discrete forward Euler time discretization, namely,
higher-order accuracy requires spatial stencils with more than three
points (Schonbek 1982). Part of the difficulty is due to lack of fully
discrete entropy-conservative schemes (LeFloch and Rohde 2000, The-
orem 6.1). This requires entropy production bounds of the kind dis-
cussed in the current example. Sharp entropy production bounds in
the scalar case can be found in Chalons and LeFloch (2001b).

(3) Entropy stability with distinguished waves. Finally, we remark that an

extension based on entropy-conservative schemes of the ‘second kind’
discussed in Section 6 would lead to an entropy stability statement
under a refinement of the CFL statement (7.19), similar to the semi-
discrete discussion in Section 6.

Example 7.3. (Crank–Nicolson time discretization) The fully ex-

plicit Euler time discretization does not conserve entropy except in the
case of linear fluxes (LeFloch and Rohde 2000). Consequently, both the
fully explicit and fully implicit Euler differencing do not respect (nonlin-
ear) entropy conservation, independent of the spatial discretization. Fully
discrete entropy conservation is offered by Crank–Nicolson time differenc-
ing. In its standard version, for example, Richtmyer and Morton (1967)
and Gustafsson et al. (1995), time is replaced by divided differences centred
1
at tn+ 2 := 12 (tn + tn+1 ) and spatial terms are evaluated at the mid-value,
1 n n+1
2 (v + v ). In the present nonlinear context, the mid-value should be
494 E. Tadmor

weighted by the specific entropy function we are dealing with. We set

Z 1  
n+ 12
2 1 n n+1

n+ 12 1
v := v u +u + ξ∆u dξ, ∆un+ 2 = un+1 − un ,
1
ξ=− 2 2
(7.22)
and we discretize (7.1) by the (generalized) Crank–Nicolson scheme
∆t h n+ 12

n+ 12

i
un+1
ν = u n
ν − g 1 v − g 1 v . (7.23)
∆x ν+ 2 ν− 2

1 
Noting that vn+ 2 , un+1 − un = U (un+1 ) − U (un ), we conclude the follow-
ing.

Corollary 7.4. The Crank–Nicolson scheme (7.22), (7.23) is entropy-stable

(and, respectively, entropy-conservative), if and only if the semi-discrete
scheme associated with the numerical flux g(·) is entropy-stable (respec-
tively, entropy-conservative).
1
Observe that in the symmetric case, vn+ 2 = 12 (un + un+1 ), and (7.23)
1
recovers the standard differencing centred around tn+ 2 .

We conclude this section by referring the reader to the recent work of

LeFloch and his co-workers (LeFloch and Rohde 2000, LeFloch et al. 2002)
for a general framework along these lines for entropy stability of fully discrete
schemes.

8. Entropy stability by the homotopy approach

We study the cell entropy inequality for general difference schemes written
in their viscosity form corresponding to (5.24):
∆t 


un+1
ν = unν − f unν+1 − f unν−1
2∆x
∆t 


+ Pν+ 1 unν+1 − unν − Pν− 1 unν − unν−1 . (8.1)
2∆x 2 2


We decompose un+1
ν = un+1
ν+ 1
+ un+1
ν− 1
/2 where
2 2

∆t 

 ∆t

un+1
ν+ 1
:= unν − f unν+1 − f unν + Pν+ 1 unν+1 − unν ,
2 ∆x ∆x 2

∆t 

 ∆t

un+1
ν− 1
:= unν − f unν − f unν−1 − Pν+ 1 unν − unν−1 ,
2 ∆x ∆x 2

and we study the entropy inequality for each term. This decomposition
 Entropy stability 495

into left- and right-handed stencils in the context of cell entropy inequal-
ity was first introduced in Tadmor (1984b). We begin by considering un+1
ν+ 1
.
2
To this end, we set unν+ 1 (s) := unν +s(unν+1 −unν ) and the following inequality
2
is sought (here and below, ∆u := unν+1 − unν ):
 
∆t 


I+ := U un+1
ν+ 12
− U unν + F unν+1 − F unν
∆x
Z 1 D   E
∆t
− U 0 unν+ 1 (s) , Pν+ 1 ∆u ds ≤ 0. (8.2)
∆x s=0 2 2

We refer to the last statement as a quasi-cell entropy inequality since the

last expression on the right is not conservative. To verify (8.2) we proceed
as follows. We set
∆t h  n  i ∆t  
un+1
ν+ 12
(s) := u n
ν − f u ν+ 12
(s) − f (u n
ν ) + P 1 u n
1 (s) − u n
ν .
∆x ∆x ν+ 2 ν+ 2
Noting that un+1
ν+ 1
(0) = unν and un+1
ν+ 1
(1) = un+1
ν+ 1
, we compute
2 2 2
 
U un+1
ν+ 1
− U (unν )
2
Z
d  n+1 
1
= U uν+ 1 (s) ds
s=0 ds 2
Z 1     ∆t   ∆t  
0 n+1 n+1
= U uν+ 1 (s) , − A uν+ 1 (s) + P 1 ∆u ds
s=0 2 ∆x 2 ∆x ν+ 2
and
Z D  
∆t 

 ∆t 1
E
F unν+1 − F unν = F 0 unν+ 1 (s) , unν+1 − unν ds
∆x ∆x s=0 2

Z 1 D     E
∆t
= U 0 unν+ 1 (s) A unν+ 1 (s) , ∆u ds.
∆x s=0 2 2

Adding the last two equalities yields

Z 1 D    
I+ = U 0 un+1
ν+ 1 (s) − U
0
u n
ν+ 1 (s) ,
s=0 2 2

∆t    E
− Pν+ 1 − A unν+ 1 (s) ∆u ds. (8.3)
∆x 2 2

Next, we introduce
 

unν+ 1 (r, s) := unν+ 1 (s)+r unν −unν+ 1 (s) ≡ unν +s(1−r) unν+1 −unν , (8.4)
2 2 2
496 E. Tadmor

and we set
∆t   n   
un+1
ν+ 1
(r, s) = unν+ 1 (r, s) − f uν+ 1 (s) − f unν+ 1 (r, s)
2 2 ∆x 2 2

∆t  
+ P 1 un 1 (s) − unν+ 1 (r, s)
∆x ν+ 2 ν+ 2 2

so that un+1
ν+ 1
(0, s) = unν+ 1 (s) and un+1
ν+ 1
(1, s) = un+1
ν+ 1
(s). This then yields
2 2 2 2

    Z 1
d 0  n+1 
U 0 un+11 (s) − U
ν+ 2
0
u n
ν+1 (s) = U u 1 (r, s) dr
ν+ 2
r=0 dr
Z 1    ∆t   ∆t 
= −s U 00 un+1
ν+ 12
(r, s) dr I + A u n
ν+ 12
(r, s) − P 1 ∆u.
r=0 ∆x ∆x ν+ 2
Inserting the last expression into the right-hand side of (8.3) we end up with
Z 1  
∆t  n  ∆t
I+ = − s I+ A uν+ 1 (r, s) − P 1 ∆u,
r,s=0 ∆x 2 ∆x ν+ 2
  ∆t   ∆t  
00 n+1 n
U uν+ 1 (r, s) − A uν+ 1 (s) + P 1 ∆u dr ds. (8.5)
2 ∆x 2 ∆x ν+ 2
To continue, we focus our attention on two prototype cases.
(i) The scalar case. The positivity of the last expression on the right of
(8.5) follows from a CFL condition
∆t  n  ∆t ∆t  n 
a uν+ 1 (s) ≤ pν+ 1 ≤ I + a uν+ 1 (r, s) . (8.6)
∆x 2 ∆x 2 ∆x 2

In a similar manner, the CFL condition

∆t  n  ∆t ∆t  n 
− a uν− 1 (s) ≤ pν− 1 ≤ I − a uν− 1 (r, s)
∆x 2 ∆x 2 ∆x 2

yields the quasi-cell entropy inequality

 
∆t 


I− := U un+1ν− 12
− U u n
ν + F u n
ν − F u n
ν−1
∆x
Z 1 D  
∆t
E
+ U 0 unν− 1 (s) , pν− 1 unν − unν−1 ds ≤ 0. (8.7)
∆x s=0 2 2

Again, the last expression is nonconservative, but together with (8.2) we end
up with the cell entropy inequality.

Corollary 8.1. Consider the fully discrete scalar scheme (8.1) and assume
the CFL condition
∆t   n  ∆t ∆t   n 
a uν+ 1 (s)  ≤ pν+ 1 ≤ I − a uν+ 1 (r, s)  (8.8)
∆x 2 ∆x 2 ∆x 2
 Entropy stability 497

is fulfilled. Then the following cell entropy inequality holds:

1   n+1  
n+1

U un+1
ν ≤ U u 1
ν+ 2
+ U u ν− 21
2

∆t 


≤ U unν − F unν+1 − F unν−1
"∆xZ 1 D 
∆t 
E
+ U 0 unν+ 1 (s) , pν+ 1 unν+1 − unν ds
2∆x s=0 2 2

Z 1 D  #

E
0 n n n
− U uν− 1 (s) , pν− 1 uν − uν−1 ds .
2 2
s=0

Next, we extend our discussion to systems of conservation laws.

(ii) Symmetric systems of conservation laws with the quadratic entropy,

U (u) = |u|2/2. We start by setting C(s) := Pν+ 1 − A(unν+ 1 (s)), and noting
2 2
the (r, s)-variables in (8.4), we find that
 
A unν+ 1 (r, s) − Pν+ 1 = −C((1 − r)s)).
2 2

Change of variables, t := (1 − r)s, in (8.5) then yields

Z 1 Z s   
∆t ∆t
I+ = I− C(t) ∆u, C(s)∆u dt ds. (8.9)
s=0 t=0 ∆x ∆x
We now make the first requirement of positivity, assuming C(·) ≥ 0; then
the positivity of I+ follows if and only if the corresponding eigenvalues satisfy
 ∆t

 ∆t

1
λ C(s) I − ∆x C(t) ≥ 0. But C(s) I − ∆x C(t) is similar to C 2 (s) I −
∆t

1
∆x C(t) C , which is congruent to, and hence by Sylvester’s theorem has the
2
∆t
same number of nonnegative eigenvalues as, I − ∆x C(t). This leads to the
∆t


second requirement, ∆x λ(C(·)) ≤ 1. Recall that C(s) = Pν+ 1 − A unν+ 1 (s)
2 2
is symmetric, and hence the last two requirements amount to the same CFL
condition we met earlier in connection with the scalar case (8.6):
∆t  n  ∆t ∆t  
A uν+ 1 (s) ≤ Pν+ 1 ≤ I + A uν+ 1 (r, s) .
∆x 2 ∆x 2 ∆x 2

In a similar manner, the CFL condition

∆t  n  ∆t ∆t  
− A uν+ 1 (s) ≤ Pν+ 1 ≤ I − A uν+ 1 (r, s)
∆x 2 ∆x 2 ∆x 2

n+1
yields the quasi-cell entropy inequality for uν− 1 , and the following conclu-
2
sion.
498 E. Tadmor

Corollary 8.2. Consider the fully discrete scheme (8.1) consistent with
the symmetric system (2.1) and assume the CFL condition5
∆t   n  ∆t ∆t   n 
A uν+ 1 (s)  ≤ Pν+ 1 ≤ I − A uν+ 1 (r, s)  (8.10)
∆x 2 ∆x 2 ∆x 2

is fulfilled. Then the following cell entropy

R u inequality holds for the quadratic
entropy pair U (u) = |u|2/2, F (u) = f (w) dw − hu, f (u)i:

1   n+1   
U un+1ν ≤ U uν+ 1 + U un+1 1
ν− 2
2 2

∆t



≤ U unν − F unν+1 − F unν−1
2∆x
Z 1 D  
∆t
E
+ U 0 vν+
n
1 (s) , P 1
ν+ 2 u n
ν+1 − u n
ν ds
2∆x s=0 2
Z 1 D   
0 n n n

E
− U uν− 1 (s) , Pν− 1 uν − uν−1 ds .
2 2
s=0
We demonstrate the application of Corollaries 8.1 and 8.2 with two pro-
totype examples of centred and upwind schemes.
Example 8.3. (Modified Lax–Friedrichs scheme) Here we set Pν+ 1 =
2
∆x
2∆t IN ×N , leading to the modified Lax–Friedrichs scheme (7.20)
1 n
∆t 


un+1
ν = uν−1 − 2unν + unν+1 + f unν+1 − f unν−1 .
4 2∆x
The modified Lax–Friedrichs scheme is entropy-stable with respect to the
quadratic entropy function (for symmetric systems) and for all convex en-
tropies (for scalar equations), provided the CFL condition (8.8), (8.10) holds,
which amounts to
∆t     1
sup λ A uν+ 1 (s)  ≤ .
∆x s,λ 2 2
A linearized von Neumann stability analysis reveals that this CFL condition
is sharp.
Remark. The original homotopy argument in this context of entropy sta-
bility is due to Lax (1971), where he proves the entropy stability of the Lax–
Friedrichs (LxF) scheme, corresponding to Pν+ 1 = ∆x ∆t IN ×N . In this special
2

5
We recall (consult (5.34)) that |A| stands for the absolute value of A, defined by its
spectral decomposition
 
|a1 |
 ..  −1
|A| = R  . R .
|aN |
 Entropy stability 499

case of a two-point LxF stencil, we can apply the homotopy argument on

the full stencil of the scheme. In the general case of essentially three-point
schemes, (8.1), we follow the decomposition into left- and right-handed sten-
cils and, as in Tadmor (1984b), this restricts the maximal viscosity coeffi-
cient to that of the modified LxF scheme. For a recent extension to entropy
stability under an optimal CFL condition in the scalar case, consult Makri-
dakis and Perthame (2003).

We now turn to discussion of entropy stability with the minimal amount

of viscosity.



Example 8.4. (Upwind scheme) We set Pν+ 1 = p A unν+ 1 (s) with
2 2
p(·) being any viscosity function satisfying p(·) ≥ | · |: consult Example 5.6.
The typical example is the upwind scheme
∆t 


un+1
ν = unν − f unν+1 − f unν−1 (8.11)
2∆x
∆t h           i
+ sup A unν+ 1 (s)  ∆unν+ 1 − sup A unν− 1 (s) ∆unν− 1 .
2∆x s 2 2 s 2 2

We find that the upwind scheme is entropy-stable for the quadratic en-
tropy function (for symmetric systems) and for all convex entropies (for
scalar equations), provided the CFL condition (8.8), (8.10) holds, which
amounts to
∆t     1
sup λ A uν+ 1 (s)  ≤ . (8.12)
∆x s,λ 2 2

Again, a linearized von Neumann stability analysis reveals that the CFL
condition is sharp.

We conclude this section with several remarks.

(1) Comparison with Roe scheme. Consider the numerical viscosity of the
Roe-type scheme (5.24), (5.33), Pν+ 1 = |Aν+ 1 |. A comparison with
2  2

the upwind scheme (8.11), P 1 = sups A un 1 (s) , reveals that the
ν+ 2 ν− 2
entropy stability of the latter is explained by taking into account all in-
termediate values between two neighbouring values, uν and uν+1 . An
alternative entropy correction discussed in Example 5.10 adds the ad-
ditional term of order O(|∆u|) to compensate for the missing interme-
diate values.
(2) Extensions. The entropy stability results in this section are based on
a homotopy approach. Our initial point was the essentially three-point
scheme (8.1). The same homotopy approach refinement, starting with
500 E. Tadmor

the entropy-conservative schemes of the ‘second kind’ in Theorem 6.1,

would yield a refinement of the above entropy stability statements. In
particular, (i) the intermediate values sought in the upwind viscosities,
 N
(8.12), would be confined to each subpath, ujν+ 1 j=1 . This leads to
2
an entropy stability criterion which distinguishes between shock, rar-
efaction and contact waves; and (ii) starting the present homotopy ap-
proach with the entropy variables, rather than the entropy-conservative
formulation in (8.1), enables us to extend the above entropy stability
statements to the general nonsymmetric case.
(3) Second-order accuracy. The CFL conditions (8.8) and (8.10) are re-
stricted to first-order accurate schemes. A similar, more careful com-
putation along these lines enables us to treat the entropy stability of
second-order accurate schemes. For a scalar entropy stability analysis
along the lines of the second-order accurate Nessyahu–Tadmor central
scheme, we refer to Nessyahu and Tadmor (1990, Appendix).
(4) The scalar case. More could be said on the scalar case, and we should
mention a considerable amount of work in this direction. In particular,
the entropy stability of fully discrete second-order schemes was system-
atically analysed in Osher and Tadmor (1988) (see the follow-up in
Aiso (1993)). A key to enforcing entropy stability in this case is the
use of all intermediate values within critical cells: consult Coquel and
LeFloch (1995), Bouchut, Bourdarias and Perthame (1996), Johnson
and Szepessy (1986) and Yang (1996c). Otherwise, entropy stability
is enforced for a single entropy, as in Osher and Tadmor (1988), for
example, or high-order accuracy should be given up (Yang 1996b).

9. Higher-order extensions
We generalize the construction of second-order entropy-conservative schemes
to higher orders. To this end, we revisit the original derivation of the second-
order entropy-conservative schemes (Tadmor 1986b), using finite element
discretization. We begin with the weak formulation of the systems of con-
servation laws (2.6),
Z   Z  
∂ ∂
w(x, t), u(v) = w(x, t), g(v) dx dt, Ω ⊂ R × (0, T ),
Ω ∂t Ω ∂x
(9.1)
where w(·) is an arbitrary C0∞ (Ω) test function. The key point is the use
of the entropy variables, which enables us to use the standard finite element
framework where both the primary computed solution v and the test function
w belong to the same finite-dimensional scale of spaces. In particular, let the
 Entropy stability 501
P
trial solution v̂ = µ vµ (t)Ĥµ (x) be chosen from the typical finite element
space spanned by the C 0 ‘hat functions’
 x−x
µ−1
xµ −xµ−1 , xµ−1 ≤ x ≤ xµ ,
Ĥµ (x) =  x −x
µ+1
xµ+1 −xµ , xµ ≤ x ≤ xµ+1 .

Testing (9.1) against w(x) = w(x) = Ĥν (x), the right-hand side of (9.1)
yields
Z xν+1 !
∂ X
Ĥν (x)g vµ (t)Ĥµ(x) dx dt =
xν−1 ∂x µ
"Z 1 Z 1 #
2
  2
 
− g vν+ 1 (ξ) dξ − g vν− 1 (ξ) dξ , (9.2)
2 2
ξ=− 12 ξ=− 12

where we employed a change of variables, expressed in terms of the usual

vν+ 1 = 12 (vν + vν+1 ) + ξ∆vν+ 1 . A second-order mass lumping on the left
2 2
of (9.1) leads to
Z xν+1 !
∂ X
Ĥν u vµ (t)Ĥµ (x) dx dt
xν−1 ∂t µ
d  2
= ∆xν u(vν (t)) + O vν+ 1  . (9.3)
dt 2

Equating (9.2) and (9.3) while neglecting the quadratic error term, we end
up with the entropy-conservative scheme (5.2):
Z 1
d 1 h ∗ ∗
i
∗
2
 
uν (t) = − gν+ 1 − gν− 1 , gν+ 1 = g vν+ 1 (ξ) dξ.
dt ∆xν 2 2 2 ξ=− 1 2
2

Mass lumping preserves the entropy conservation induced by (and, in a

sense, built into) the weak formulation (9.1), upon choosing ŵ(x, t) = v̂(x, t),
Z    
d ∂
0= v̂(x, t), u(v̂) − v̂(x, t)g(v̂(x, t)) dx dt
Ω dt ∂x
Z  
∂ ∂
= U (u(v̂(x, t))) + F (u(v̂(x, t))) dx dt. (9.4)
Ω ∂t ∂x
Using higher-order piecewise polynomial finite element building blocks will
lead to entropy-conservative schemes of any desired order. We note in pass-
ing that, with the increased order, the size of the stencil increases. For exam-
ple, piecewise quadratic splines would lead to five-point entropy-conservative
stencils of the following form.
502 E. Tadmor

Theorem 9.1. (LeFloch and Rohde 2000, Section 3) Consider the

semi-discrete scheme
d 1 h ∗∗ ∗∗
i
uν (t) = − gν+ 1 − gν− 1 , (9.5)
dt ∆xν 2 2

∗∗
with a numerical flux, gν+ 1 = g(vν−1 , vν , vν+1 , vν+2 ), given by
2

Z 1  
2
∗∗
gν+ 1 = g vν+ 1 (ξ) dξ
2 ξ=− 12 2

1 h ∗∗ i
− Qν+ 3 (vν+2 − vν+1 ) − Q∗∗ 1 (vν − vν−1 ) .
ν− 2
(9.6)
12 2

Here, Q∗∗
ν+ 1
is a secondary viscosity coefficient depending on
2

Q∗∗
ν+ 1
= Q∗∗(vν−1 , vν , vν+1 ) (9.7)
2

The resulting five-point scheme (9.5), (9.6), (9.7) is entropy-conservative,

and it is of (at least) third-order accuracy provided Q∗∗(v, v, v) = B(v).
We observe that the higher-order accuracy is intimately linked to the wider
stencil, beyond the essentially three-point schemes discussed in Section 5.
Once more, these wider stencils could serve as the starting point for an en-
tropy stability theory for higher (than second)-order entropy-stable schemes.

REFERENCES
R. Abramov and A. Majda (2002), ‘Discrete approximations with additional con-
served quantities: Deterministic and statistical behavior’, preprint.
R. Abramov, G. Kovačič and A. Majda (2003), ‘Hamiltonian structure and statis-
tically relevant conserved quantities for the truncated Burgers–Hopf equation’,
Comm. Pure Appl. Math. 56, 1–46.
H. Aiso (1993), ‘Admissibility of difference approximations for scalar conservation
laws’, Hiroshima Math. J. 23, 15–61.
A. Arakawa (1966), ‘Computational design for long-term numerical integration of
the equations of fluid motion: Two-dimensional incompressible flow, Part I’,
J. Comput. Phys. 1, 119–143.
S. Bianchini and A. Bressan (2003), ‘Vanishing viscosity solutions of nonlinear
hyperbolic systems’, Ann. Math. To appear.
F. Bouchut (2002), ‘Entropy satisfying flux vector splitting and kinetic BGK mod-
els’, Numer. Math., DOI http://dx.doi.org/10.1007/s00211-002-0426-9.
F. Bouchut, Ch. Bourdarias and B. Perthame (1996), ‘A MUSCL method satisfying
all the numerical entropy inequalities’, Math. Comp. 65, 1439–1461
A. Bressan (2003), Viscosity solutions of nonlinear hyperbolic systems, in Hyper-
bolic Problems: Theory, Numerics and Applications, Proc. 9th Hyperbolic
Conference (T. Hou and E. Tadmor, eds), Springer. To appear.
 Entropy stability 503

C. Chalons and P. LeFloch (2001a), ‘High-order entropy-conservative schemes and

kinetic relations for van der Waals fluids’, J. Comput. Phys. 168, 184–206.
C. Chalons and P. LeFloch (2001b), ‘A fully discrete scheme for diffusive-dispersive
conservation law’, Numer. Math. 89, 493–509.
G.-Q. Chen (2000), ‘Compactness methods and nonlinear hyperbolic conservation
laws: Some current topics on nonlinear conservation laws’, in Vol. 15 of
AMS/IP Stud. Adv. Math., Amer. Math. Soc., Providence, RI, pp. 33–75.
F. Coquel and P. LeFloch (1995), ‘An entropy satisfying MUSCL scheme for sys-
tems of conservation laws’, CR Acad. Sci. Paris, Série 1 320, 1263–1268.
M. G. Crandall and A. Majda, (1980) ‘Monotone difference approximations for
scalar conservation laws’, Math. Comp. 34, 1–21.
M. G. Crandall and L. Tartar (1980), ‘Some relations between non-expansive and
order preserving mapping’, Proc. Amer. Math. Soc. 78, 385–390.
C. Dafermos (2000) Hyperbolic Conservation Laws in Continuum Mechanics,
Springer.
P. Deift and K. T. R. McLaughlin, (1998) ‘A continuum limit of the Toda lattice’,
Mem. Amer. Math. Soc. 131, 1–216.
R. DiPerna (1979), ‘Uniqueness of solutions to hyperbolic conservation laws’, In-
diana University Math. J. 28, 137–188.
R. J. DiPerna (1983), ‘Convergence of approximate solutions to conservation laws’,
Arch. Rational Mech. Anal. 82, 27–70.
B. Engquist and S. Osher (1980), ‘Stable and entropy condition satisfying approx-
imations for transonic flow calculations’, Math. Comp. 34, 44–75.
L. P. Franca, I. Harari, T. J. R. Hughes, M. Mallet, F. Shakib, T. E. Spelce,
F. Chalot and T. E. Tezduyar (1986), A Petrov–Galerkin finite element
method for the compressible Euler and Navier–Stokes equations, in Numer-
ical Methods for Compressible Flows: Finite Difference, Element and Volume
Techniques (T. E. Tezduyar and T. J. R. Hughes, eds), Amer. Soc. Mech.
Eng., AMD, Vol. 78, pp. 19–44.
K. O. Friedrichs (1954), ‘Symmetric hyperbolic linear differential equations’,
Comm. Pure Appl. Math. 7, 345–392.
K. O. Friedrichs and P. D. Lax (1971), ‘Systems of conservation laws with a convex
extension’, Proc. Nat. Acad. Sci. USA 68, 1686–1688.
E. Godlewski and P.-A. Raviart (1996), Numerical Approximation of Hyperbolic
Systems of Conservation Laws, Springer.
S. K. Godunov (1959), ‘A difference scheme for numerical computation of discon-
tinuous solutions of fluid dynamics’, Mat. Sb. 47, 271–306.
S. K. Godunov (1961), ‘An interesting class of quasilinear systems’, Dokl. Acad.
Nauk. SSSR 139, 521–523.
J. Goodman and P. D. Lax (1981), ‘On dispersive difference schemes, I’, Comm.
Pure Appl. Math. 41, 591–613.
B. Gustafsson, H.-O. Kreiss and J. Oliger (1995), Time Dependent Problems and
Difference Methods, Wiley-Interscience.
A. Harten (1983a), ‘High-resolution scheme for hyperbolic conservation laws’,
J. Comput. Phys. 49, 357–393.
A. Harten (1983b), ‘On the symmetric form of systems of conservation laws with
entropy’, J. Comput. Phys. 49, 151–164.
504 E. Tadmor

A. Harten and J. M. Hyman (1983), ‘Self-adjusting grid method for one-

dimensional hyperbolic conservation laws’, J. Comput. Phys. 50, 235–269.
A. Harten and P. D. Lax (1981), ‘A random choice finite difference scheme for
hyperbolic conservation laws’, SIAM J. Numer. Anal. 18, 289–315.
A. Harten, J. M. Hyman, and P. D. Lax (1976), ‘On finite difference approxi-
mations and entropy conditions for shocks’, Comm. Pure Appl. Math., 29,
297–322.
A. Harten, B. Engquist, S. Osher, and S. Chakravarthy (1987), ‘Uniformly high-
order accurate essentially non-oscillatory schemes, III’, J. Comput. Phys. 71,
231–303.
T. Hou and P. Lax (1991), ‘Dispersive approximations in fluid dynamics’, Comm.
Pure Appl. Math. 44, 1–40.
T. J. R. Hughes, L. P. Franca, and M. Mallet (1986), ‘A new finite-element for-
mulation for computational fluid dynamics, I: Symmetric forms of the com-
pressible Euler and Navier–Stokes equations and the second law of thermody-
namics’, Comput. Methods Appl. Mech. Eng. 54, 223–234.
G.-S. Jiang and C. -W. Shu (1994), ‘On a cell entropy inequality for discontinuous
Galerkin method’, Math. Comp. 62, 531–538.
C. Johnson and A. Szepessy (1986), ‘A shock-capturing streamline diffusion finite
element method for a nonlinear hyperbolic conservation laws’, Technical Re-
port 1986-V9, Mathematics Department, Chalmers University of Technology,
Goteborg.
C. Johnson, A. Szepessy, and P. Hansbo (1990), ‘On the convergence of shock-
capturing streamline diffusion finite element methods for hyperbolic conserva-
tion laws’ Math. Comp. 54, 107–130.
L. Kaddouri (1993), ‘Une méthode d’éléments finis discontinus pour les équations
d’Euler des fluides compressibles’, Thesis, Université Paris 6.
K. Khalfallah and A. Lerat (1989), ‘Correction d’entropie pour des schémas
numeriques approchant un système hyperbolique’, Note CR Acad. Sci. 308
II, 815–820.
H.-O. Kreiss and J. Lorenz (1998), ‘Stability for time-dependent differential equa-
tions’, in Acta Numerica, Vol. 7, Cambridge University Press, pp. 203–286.
D. Króner (1997), Numerical Schemes for Conservation Laws, Wiley-Teubner,
Stuttgart.
S. N. Kružkov (1970), ‘First order quasilinear equations in several independent
variables’, Math. USSR Sbornik 10, 217–243.
P. D. Lax (1954), ‘Weak solutions of non-linear hyperbolic equations and their
numerical computations’, Comm. Pure Appl. Math. 7, 159–193.
P. D. Lax (1957), ‘Hyperbolic systems of conservation laws, II’, Comm. Pure Appl.
Math. 10, 537–566.
P. D. Lax (1971), Shock waves and entropy, in Contributions to Nonlinear Func-
tional Analysis, (E. A. Zarantonello, ed.), Academic Press, New York,
pp. 603–634.
P. D. Lax (1972), Hyperbolic Systems of Conservation Laws and the Mathematical
Theory of Shock Waves, Vol. 11 of SIAM Regional Conference Lectures in
Applied Mathematics.
 Entropy stability 505

P. D. Lax (1986), ‘On dispersive difference schemes’, Physica 18D, 250–254.

P. D. Lax and B. Wendroff (1960), ‘Systems of conservation laws’, Comm. Pure
Appl. Math. 13, 217–237.
P. D. Lax, D. Levermore and S. Venakidis (1993), The generation and propagation
of oscillations in dispersive IVPs and their limiting behavior, in Important
Developments in Soliton Theory 1980–1990 (T. Fokas and V. E. Zakharov,
eds), Springer, Berlin.
B. van Leer (1977), ‘Towards the ultimate conservative difference scheme, III:
Upstream-centered finite-difference schemes for ideal compressible flow’,
J. Comput. Phys. 23, 263–275.
P. LeFloch (2002), Hyperbolic Systems of Conservation Laws: The Theory of Clas-
sical and Nonclassical Shock Waves, Springer, Lectures in Mathematics, ETH
Zürich.
P. LeFloch and C. Rohde (2000), ‘High-order schemes, entropy inequalities and
nonclassical shocks’, SIAM J. Numer. Anal. 37, 2023–2060.
P. LeFloch, J. M. Mercier and C. Rohde (2002), ‘Fully discrete, entropy conser-
vative schemes of arbitrary order’, SIAM J. Numer. Anal. 40, 1968–1992.
R. LeVeque (1992), Numerical Methods for Conservation Laws, Birkhäuser, Basel,
Lectures in Mathematics.
R. LeVeque (2002), Finite Volume Methods for Hyperbolic Problems, Cambridge
University Press, Texts in Applied Mathematics.
D. Levermore and J.-G. Liu (1996), ‘Large oscillations arising from a dispersive
numerical scheme’, Phys. D 99, 191–216.
P. L. Lions and Souganidis (1995), ‘Convergence of MUSCL and filtered schemes
for scalar conservation laws and Hamilton–Jacobi equations’, Numer. Math.
69, 441–470.
A. Majda (1984), Compressible Fluid Flow and Systems of Conservation Laws in
Several Space Variables, Springer, New York.
A. Majda and S. Osher (1978), ‘A systematic approach for correcting nonlinear
instabilities: The Lax–Wendroff scheme for scalar conservation laws’, Numer.
Math. 30, 429–452.
A. Majda and S. Osher (1979), ‘Numerical viscosity and the entropy condition’,
Comm. Pure Appl. Math. 32, 797–838.
C. Makridakis and B. Perthame (2003), ‘Sharp CFL, discrete kinetic formulation
and entropy schemes for scalar conservation laws’, SIAM J. Numer. Anal.
To appear.
M. Merriam (1989), ‘Towards a rigorous approach to artificial dissipation’, AIAA-
89-0471, Reno, Nevada.
M. S. Mock (1978), ‘Some higher order difference schemes enforcing an entropy
inequality’, Michigan Math. J. 25, 325–344.
M. S. Mock (1980), ‘Systems of conservation of mixed type’, J. Diff. Eqns 37,
70–88.
M. Moskalkov (1980), ‘Completely conservative schemes for gas dynamics’, USSR
Comput. Maths. Math. Phys. 20, 162–170.
H. Nessyahu and E. Tadmor (1990), ‘Non-oscillatory central differencing for hy-
perbolic conservation laws’. J. Comput. Phys. 87, 408–463.
506 E. Tadmor

J. von Neumann and R. D. Richtmyer (1950), ‘A method for the numerical calcu-
lation of hydrodynamic shocks’, J. Appl. Phys. 21, 232–237.
P. Olsson (1995), ‘Summation by parts, projections, and stability, III’, RIACS
Technical report, 95.06.
S. Osher (1984), ‘Riemann solvers, the entropy condition, and difference approxi-
mations’, SIAM J. Numer. Anal. 21, 217–235.
S. Osher (1985), ‘Convergence of generalized MUSCL schemes’, SIAM J. Numer.
Anal. 22, 947–961.
S. Osher and S. Chakravarthy (1984), ‘High resolution schemes and the entropy
condition’, SIAM J. Numer. Anal. 21, 955–984.
S. Osher and F. Solomon (1982), ‘Upwind difference schemes for hyperbolic con-
servation laws’, Math. Comp. 38, 339–374.
S. Osher and E. Tadmor (1988), ‘On the convergence of difference approximations
to scalar conservation laws’, Math. Comp. 50, 19–51.
R. D. Richtmyer and K.W. Morton (1967), Difference Methods for Initial-Value
Problems, Interscience, 2nd edn.
P. L. Roe (1981), ‘Approximate Riemann solvers, parameter vectors and difference
schemes’, J. Comput. Phys. 43, 357–372.
V. V. Rusanov (1961), ‘Calculation of interaction of non-steady shock-waves with
obstacles’, J. Comput. Math. Phys., USSR 1, 267–279.
R. Sanders (1983), ‘On convergence of monotone finite difference schemes with
variable spatial differencing’, Math. Comp. 40, 91–106.
D. Serre (1991), ‘Richness and the classification of quasilinear hyperbolic systems’,
Multidimensional Hyperbolic Problems and Computations, Minneapolis, MN,
Inst. Math. Appl., Vol 29, Springer, pp. 315–333.
D. Serre (1999), Systems of Conservation Laws, 1: Hyperbolicity, Entropies, Shock
Waves (English translation), Cambridge University Press.
M. Schonbek (1982), ‘Convergence of solutions to nonlinear dispersive equations’,
Comm. PDEs 7, 959–1000.
J. Smoller (1983), Shock Waves and Reaction Diffusion Equations, Springer,
Berlin.
T. Sonar (1992), ‘Entropy production in second-order three-point schemes’, Numer.
Math. 62, 371–390.
A. Szepessy (1989), ‘An existence result for scalar conservation laws using measure
valued solutions’, Comm. PDEs 14, 1329–1350.
E. Tadmor (1984a), ‘Skew-adjoint form for systems of conservation form’, J. Math.
Anal. Appl. 103, 428–442.
E. Tadmor (1984b), ‘Numerical viscosity and the entropy condition for conservative
difference schemes’, Math. Comp. 43, 369–381.
E. Tadmor (1986a), ‘A minimum entropy principle in the gas dynamics equations’,
Appl. Numer. Math. 2, 211–219.
E. Tadmor (1986b), Entropy conservative finite element schemes, in Numerical
Methods for Compressible Flows: Finite Difference Element and Volume Tech-
niques, Proc. Winter Annual Meeting of the Amer. Soc. Mech. Eng. AMD,
Vol. 78 (T. E. Tezduyar and T. J. R. Hughes, eds), pp. 149–158.
E. Tadmor (1987a), ‘Entropy functions for symmetric systems of conservation
laws’, J. Math. Anal. Appl. 121, 355–359.
 Entropy stability 507

E. Tadmor (1987b), ‘The numerical viscosity of entropy stable schemes for systems
of conservation laws, I’, Math. Comp. 49, 91–103.
E. Tadmor (1989), ‘Convergence of spectral methods for nonlinear conservation
laws’, SIAM J. Numer. Anal. 26, 30–44.
E. Tadmor (1997), Approximate solution of nonlinear conservation laws and related
equations, in Recent Advances in Partial Differential Equations and Applica-
tions, Proc. 1996 Venice Conference in Honor of Peter D. Lax and Louis
Nirenberg on their 70th Birthday (R. Spigler and S. Venakides, ed.), AMS
Proceedings Symp. Appl. Math. Vol. 54, Providence, RI, pp. 321–368.
E. Tadmor (1998), Approximate solutions of nonlinear conservation laws, in Ad-
vanced Numerical Approximation of Nonlinear Hyperbolic Equations, Lecture
Notes from CIME Course Cetraro, Italy, 1997 (A. Quarteroni, ed.), Vol. 1697
of Lecture Notes in Mathematics, Springer, pp. 1–150.
L. Tartar (1975), Compensated compactness and applications to partial differential
equations, in Nonlinear Analysis and Mechanics, Heriot-Watt Symposium,
Vol. 4 (R. J. Knopps, ed.), Vol. 39 of Research Notes in Mathematics, Pitman
Press, pp. 136–211.
A. I. Vol’pert (1967), ‘The spaces BV and quasilinear equations’, Math. USSR-Sb.
2, 225–267.
H. Yang (1996a), ‘On wavewise entropy inequalities for high-resolution schemes, I:
The semidiscrete case’ Math. Comp. 65, 45–67. Supplement Math. Comp. 65
S1–S13.
H. Yang (1996b), ‘Convergence of Godunov type schemes’ Appl. Math. Letters 9,
63–67.
H. Yang (1998), ‘On wavewise entropy inequalities for high-resolution schemes, II:
Fully discrete MUSCL schemes with exact evolution in small time’, SIAM J.
Numer. Anal. 36, 1–31.

Appendix: Entropy stability of Roe-type schemes

We consider Roe-type schemes of the form
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 ) (A.1)
dt 2∆xν
1 h i  
+ pν+ 1 ∆uν+ 1 − pν− 1 ∆uν− 1 , pν+ 1 = p Aν+ 1 .
2∆xν 2 2 2 2 2 2

Once more (consult Example 5.2), we use (5.10) to rewrite the viscous part
of (A.1) in terms of the entropy variables, obtaining
d 1  
uν (t) = − f (uν+1 ) − f (uν−1 )
dt 2∆xν
1 h i
+ Qν+ 1 ∆vν+ 1 − Qν− 1 ∆vν− 1 , (A.2)
2∆xν 2 2 2 2
508 E. Tadmor

where
Z 1    
2
Qν+ 1 = p Aν+ 1 H(ξ) dξ, H(ξ) ≡ H vν+ 1 (ξ) . (A.3)
2
ξ=− 12 2 2

Corollary 5.1 suggests that the entropy dissipation of these schemes should
be measured by the quantity h∆vν+ 1 , Qν+ 1 ∆vν+ 1 i. A lower bound for the
2 2 2
latter is provided in the following lemma.
Lemma A1. Let {rk , ak } be the eigensystem of Aν+ 1 , and assume that
2
 2

|rk − rk (ξ = 0)| + |ak − ak (ξ = 0)| ≤ Const ∆vν+ 1  . (A.4)
2

Then we have
D E
∆vν+ 1 , Qν+ 1 ∆vν+ 1
2 2 2
N h Z
2 i D E
1
X  2
≥ p(ak ) − Const∆vν+ 1   rk (ξ), ∆v 2 dξ. (A.5)
2
ν+ 21
k=1 ξ=− 12
1
Proof. Using the orthonormal system {H − 2 (ξ)rk (ξ)} in (5.18), we can ex-
pand the right-hand side of the equality (A.3), which we rewrite as
D E Z 12 D   1 1
E
∆vν+ 1 , Qν+ 1 ∆vν+ 1 = ∆vν+ 1 , p Aν+ 1 H 2 (ξ)·H 2 (ξ)∆vν+ 1 dξ,
2 2 2 2 2 2
ξ=− 12

and find
D E XN Z 1 D   E
2
∆vν+ 1 , Qν+ 1 ∆vν+ 1 = ∆vν+ 1 , p Aν+ 1 rk (ξ) αk (ξ) dξ,
2 2 2 2 2
k=1 ξ=− 12

 (A.6)
where αk (ξ) abbreviates αk (ξ) := rk (ξ), ∆vν+ 1 .
2
Consider the quantities on the right of (A.6): their dependence on ξ is
reflected through their dependence on vν+ 1 (ξ) = 12 (vν + vν+1 ) + ξ∆vν+ 1 ;
2 2
for such quantities we have
 s 
 d X(ξ)
 ds  


s 
 .
 dξ s ≡ X v ν+ 12 (ξ)  = O ∆v ν+ 12 (A.7)
dξ s
By Taylor’s theorem,
ξ2
rk (ξ) = rk (0) + ξ ṙk (0) + r̈k (θξ), for some θ ∈ [0, 1].
2
Here and below, (˙) denotes ξ-differentiation, dξ
d
( ). In view of (A.7), |r̈k | ≤
 2
Const∆v 1  , and together with assumption (A.4) we have
ν+ 2
 2
rk (ξ) = rk + ξ ṙk (0) + Jk , |Jk | ≤ Const∆vν+ 1  . (A.8)
2
 Entropy stability 509

Inserting this into (A.6) we conclude that

D E
∆vν+ 1 , Qν+ 1 ∆vν+ 1
2 2 2
N Z
X 1 D   E
2
= ∆vν+ 1 , p Aν+ 1 rk αk (ξ) dξ
1 2 2
k=1 ξ=− 2
N Z
X 1 D   E
2
+ ξ ∆vν+ 1 , p Aν+ 1 ṙk (0) αk (ξ) dξ
1 2 2
k=1 ξ=− 2
XN Z 1 D   E
2
+ ∆vν+ 1 , p Aν+ 1 Jk αk (ξ) dξ. (A.9)
2 2
k=1 ξ=− 12

Finally, taking into account (A.8), we have

 




p Aν+ 1 rk = p ak rk = p ak rk (ξ) − ξp ak ṙk (0) − p ak Jk ;
2

we substitute the last three terms into three summations on the right of
(A.9), respectively, and end up with
D E
∆vν+ 1 , Qν+ 1 ∆vν+ 1
2 2 2
N
X Z N
= p(ak ) · α2k (ξ) dξ
k=1 ξ=− 12
N Z
X 1 D h   i E
2
+ ξ ∆vν+ 1 , p Aν+ 1 − p(ak ) · IN ṙk (0) αk (ξ) dξ
1 2 2
k=1 ξ=− 2
XN Z 1 D h   i E
2
+ ∆vν+ 1 , p Aν+ 1 − p(ak ) · IN Jk αk (ξ) dξ
2 2
k=1 ξ=− 12
N
X  Z 1
2 D E
= p ak  rk (ξ), ∆v 2 dξ + II + III . (A.10)
ν+ 21
k=1 ξ=− 12

 3

Since Jk αk (ξ) is of order O ∆vν+ 1  , we have
2
 4
|III | ≤ ConstIII · ∆vν+ 1  ; (A.11)
2

integration by parts of the second summation on the right of (A.10) yields

N Z  
d nD h   i E o
1
X 2 1 2
II = −ξ ∆vν+ 1 , p Aν+ 1 −p(ak )IN ·rk (0) αk (ξ) dξ.
1 4 dξ 2 2
k=1 ξ=− 2
510 E. Tadmor
 2

Since ṙk (0)αk (ξ) is of order O ∆vν+ 1  , the expansion inside the curly
 3
2

brackets is O ∆vν+ 1  , and its ξ-derivative gives us (consult (A.7))

2
 4
|II | ≤ ConstII · ∆vν+ 1  . (A.12)
2

The result (A.5) now follows, noting that H(ξ) ≤ K · IN (see (5.26)), and
hence
Z 1
   1 
∆v 1 2 ≤ 1
2
H 2 (ξ)∆v 1 2 dξ
ν+ 2 ν+
K ξ=− 1 2
2
N Z 1
1 X 2 D E
2
≤ rk (ξ), ∆vν+ 1  dξ; (A.13)
K ξ=− 1 2
k=1 2

consequently, (A.11), (A.12), and (A.13) imply

II + III ≥
Z 1
N
X 1  2 2 D E
−  
(ConstII + ConstIII ) · ∆vν+ 1 ·  rk (ξ), ∆v 1 2 dξ
ν+ 2
K 2
ξ=− 1
k=1 2

and together with (A.10), this amounts to having (A.5).

Next we turn to bounding the viscosity part of the entropy-conservative
scheme (5.4), (5.5) from above, in the spirit of Lemma A1.
Lemma A2. Let Q∗ν+ 1 be the viscosity matrix associated with the entropy-
2
conservative scheme (5.2), (5.3). Let
∆ak (uν ) ≡ ak (uν+1 ) − ak (uν )
denote the jump of the kth eigenvalue, λk (A(u)), from the state on the left,
uν , to its right neighbour uν+1 . Then, for arbitrary εk > 0, we have
D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1
2 2 2

X N    
1 |ak |  2
≤ ∆ak (uν ) + εk |ak | + 1 + 
Const ∆vν+ 1 
6 εk 2
k=1
Z 1
2 D E
×  rk (ξ), ∆v 1 2 dξ. (A.14)
ν+ 2
ξ=− 12

Proof. According to (5.19) we have

D E N Z
X 1
2
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 = 2ξak (ξ)α2k (ξ) dξ.
2 2
2
k=1 ξ=− 12
 Entropy stability 511

Changing variables, ξ → −ξ, and averaging, we can rewrite this as

D E X N Z 1
∗
2  
∆vν+ 1 , Qν+ 1 ∆vν+ 1 = ξ ak (ξ)α2k (ξ) − ak (−ξ)α2k (−ξ) dξ.
2 2
2
k=1 ξ=− 12
(A.15)
By Taylor’s expansion,
ξ 2 d2 2
α2k (±ξ) = α2k (0) ± ξαk (0)α̇k (0) + (α (ξ)). (A.16)
2 dξ 2 k
 4

In view of (A.7), dξ d2 2 
2 (αk (ξ)) is of order O ∆vν+ 1
 , and therefore
2
 
ak (ξ)α2k (ξ) − ak (−ξ)α2k (−ξ)
 
= ak (ξ) − ak (−ξ) α2k (0)
   4 
+ ak (ξ) + ak (−ξ) 2ξαk (0)α̇k (0) + O ∆vν+ 1  . (A.17)
2

Moreover, we have
ξ2  2
αk (±ξ) = αk (0) ± ξ ȧk (0) + äk , |äk | ≤ Const∆vν+ 1  ;
2 2

we substitute this into the right-hand side of (A.17): since ȧk (0), α2k (0) and
 
 2
 3

αk (0)α̇k (0) are of order O ∆vν+ 1  , O ∆vν+ 1  , and O ∆vν+ 1  re-
2 2 2
spectively, we obtain
 
ak (ξ)α2k (ξ) − ak (−ξ)α2k (−ξ)
 4 
= 2ξ ȧk (0)α2k (0) + 4ξak (0)αk (0)α̇k (0) + O ∆vν+ 1  . (A.18)
2

Inserting the latter expression into (A.15), we find after integration that
D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1
2 2 2
N
X 1   4
≤ ȧk (0)α2k (0) + 2ak (0)αk (0)α̇k (0) + Const ∆vν+ 1  .
6 2
k=1

In view of (A.4), we can replace ak (0) by ak and end up with

D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 (A.19)
2 2 2
N
X 1   4
≤ ȧk (0)α2k (0) + 2ak αk (0)α̇k (0) + Const ∆vν+ 1  .
6 2
k=1
We upper-bound the second term by Cauchy–Schwartz:
1
2ak αk (0)α̇k (0) ≤ εk |ak |α2k (0) + |ak |(α̇k (0))2,
εk
512 E. Tadmor
 4
and since (α̇k (0))2 ≤ Const∆vν+ 1  , (A.19) gives us
2
D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1 (A.20)
2 2 2
  
X N
1  2 |ak | 4 
≤ ȧk (0) + εk |ak | αk (0) + 1 + Const ∆vν+ 1  .
6 εk 2
k=1

Finally, by (A.16) we have

Z 1
2 D E   
α2k (0) =  rk (ξ), ∆v 1 2 dξ + O ∆v 1 4 ,
ν+ ν+
ξ=− 12 2 2

and, according to (A.13),

N Z 1
  X D E
∆v 1 2 ≤ 1
2
 rk (ξ), ∆v 1 2 dξ.
ν+ 2 ν+ 2
K ξ=− 1
k=1 2

Using this together with (A.20) implies

D E
∆vν+ 1 , Q∗ν+ 1 ∆vν+ 1
2 2 2

X1N    
|ak |  2
≤ ȧk (0) + εk |ak | + 1 + Const∆vν+ 1 
6 εk 2
k=1
Z 1
2 D E
×  rk (ξ), ∆v 1 2 dξ, (A.21)
ν+ 2
ξ=− 12
 2

and the result (A.14) follows, noting that ȧk (0) = ∆ak (uν ) + O ∆vν+ 1  .
2

We close this section with the following proof.

Proof of Theorem 5.9. Comparing (A.5) and (A.14), we conclude from
Corollary 5.1 that the Roe-type scheme (A.1), (A.4) is entropy-stable pro-
vided that
   
1 |ak |  2
p(ak ) ≥ ∆ak (uν ) + εk |ak | + 1 + Const∆vν+ 1  , (A.22)
6 εk 2

and since, by (5.10), (5.26),

1      
∆vν+ 1  ≤ ∆uν+ 1 ≡ Hν+ 1 ∆vν+ 1  ≤ K ∆vν+ 1 ,
K 2 2 2 2 2

it follows that (A.22) is equivalent to (5.40).