SIAM J. SCI. COMPUT.

c 2009 Society for Industrial and Applied Mathematics


Vol. 31, No. 3, pp. 1926–1945

ON A CLASS OF UNIFORMLY ACCURATE IMEX RUNGE–KUTTA

Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

SCHEMES AND APPLICATIONS TO HYPERBOLIC SYSTEMS

WITH RELAXATION∗
SEBASTIANO BOSCARINO† AND GIOVANNI RUSSO†

Abstract. In this paper we consider hyperbolic systems with relaxation in which the relaxation
time ε may vary from values of order one to very small values. When ε is very small, the relaxation
term becomes very strong and highly stiff, and underresolved numerical schemes may produce spu-
rious results. In such cases it is important to have schemes that work uniformly with respect to ε.
IMplicit-EXplicit (IMEX) Runge–Kutta (R-K) schemes have been widely used for the time evolu-
tion of hyperbolic partial differential equations but the schemes existing in literature do not exhibit
uniform accuracy with respect to the relaxation time. We develop new IMEX R-K schemes for hy-
perbolic systems with relaxation that present better uniform accuracy than the ones existing in the
literature and in particular produce good behavior with high order accuracy in the asymptotic limit,
i.e., when ε is very small. These schemes are obtained by imposing new additional order conditions
to guarantee better accuracy over a wide range of the relaxation time. We propose the construction
of new third-order IMEX R-K schemes of type CK [S. Boscarino, SIAM J. Numer. Anal., 45 (2008),
pp. 1600–1621]. In several test problems, these schemes, with a fixed spatial discretization, exhibit
for all range of the relaxation time an almost uniform third-order accuracy.

Key words. Runge–Kutta methods, stiff problems, hyperbolic systems with relaxation, order
conditions

AMS subject classifications. 65C20, 65L06

DOI. 10.1137/080713562

1. Introduction. Several physical problems of great relevance for applications

are described by stiff system in the form:
1
(1) ∂t U = F (U ) + G(U ),
ε
where U = U (t) ∈ RM , F , G : RM → RM , and ε > 0 is the stiffness parameter.
The development of numerical schemes for systems (1) attracted a lot of attention
in recent years. Systems of such form often arise from the discretization of partial
differential equations, such as convection-diffusion equations and hyperbolic systems
with relaxation. In this work we consider the latter case which in recent years has
been a very active field of research, due to its great impact on applied sciences. In
fact, relaxation is important in many physical situations, for example, it arises in
discrete kinetic theory of rarefied gases, hydrodynamical models for semiconductors,
linear and nonlinear waves, viscoelasticity, traffic flows, shallow water, etc. ([7], [16],
[17], [18], [14], [6], [21], [8], [13]).
In one space dimension these problems can be described mathematically by the
system
1
(2) ∂t u + ∂x F (u) = R(u), u = u(x, t) ∈ RN , x ∈ R, t ∈ R+
ε
∗ Received by the editors January 16, 2008; accepted for publication (in revised form) November

10, 2008; published electronically March 27, 2009. This work was supported by Italian PRIN 2006
project Modelli cinetici e del continuo per il trasporto di particelle nei gas e nei semiconduttori:
aspetti analitici e computazionali. Prot. 2006012132-003.
http://www.siam.org/journals/sisc/31-3/71356.html
† Department of Mathematics and Computer Science, University of Catania, viale A. Doria 6,

95125, Italy ([email protected], [email protected]).

1926

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1927

from which system (1) can be obtained by suitable finite difference discretization
in space (methods of lines) [20]. System (2) is assumed to be hyperbolic; i.e., the
Jacobian matrix F  (u) has real eigenvalues and admits a basis of eigenvectors for every
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

u ∈ RN . We will call system (2) the relaxation system and ε is called the relaxation
time, which is small in many physical situations. Here we use the term relaxation in
the sense of Whitham [21] and Liu ([8], [15]); i.e., the operator R : RN → RN is said
relaxation operator, and consequently (2) defines a relaxation system, if there exists
a constant n × N matrix Q, with rank(Q)= n < N such that QR(u) = 0 for u ∈ RN .
This yields n independent conserved quantities v = Qu. In addition, we assume
that each such v uniquely determines a local equilibrium value u = E(v) satisfying
R(E(v)) = 0 and such that QE(v) = v for all v. The image of E represents the
manifold of local equilibria of R. Associated with Q are n conservation laws satisfied
by every solution of (2) and that take the form ∂t (Qu) + ∂x (QF ) = 0. These can be
closed as a reduced system for v = Qu if we take the local relaxation approximation
for u, namely u = E(v), ∂t v + ∂x H(v) = 0, where H is defined by H(v) = QF (E(v)).
Typical examples of such systems are discrete velocity models in kinetic theory, such
as the Broadwell model [14, 6], which in one space dimension has the structure of
a semilinear hyperbolic system of three equations that degenerates to a quasilinear
hyperbolic system with two equations, as ε → 0.
The development of efficient numerical schemes for such systems is challenging,
since in many applications the relaxation time varies from values of order one to very
small values if compared to the time scale determined by the characteristic speeds of
the system. In this second case the hyperbolic system with relaxation is said to be
stiff.
Here we will focus on the development of schemes that work for all ranges of
the relaxation time and we are interested in high order numerical schemes for the
stiff relaxation system which are able to capture the correct physical behavior with
high order accuracy. Underresolved numerical schemes may yield spurious numerical
solutions that are unphysical and, in general for hyperbolic systems with stiff terms,
high order schemes may also reduce to lower order when the mesh fails to resolve the
small relaxation time.
In this article we will present some recent results on the development of high-order
implicit-explicit (IMEX) Runge–Kutta (R-K) schemes suitable for time-dependent
partial differential systems that work better than other IMEX R-K schemes existing
in the literature. Classical high-order IMEX R-K schemes fail to maintain the high-
order accuracy in time in the whole range of the relaxation time and in particular in
the asymptotic limit, i.e., when the relaxation time is not temporally resolved. We
introduce here a new class of IMEX R-K schemes that are able to handle the stiffness
of the system (1) in the whole range of the relaxation time.
This work is motivated by the study of the global error for different types of
IMEX R-K methods existing in literature [2]. This error analysis is accompanied by
an order reduction phenomenon where the observed convergence rates of the methods
drop the classical order of accuracy in time. The development of these schemes is
aided by the knowledge of extra order conditions in addition to the classical ones
[7, 16] which ensure accuracy in time at the various order in the stiffness parameter
ε [3, 4]. We remark that in the classical literature IMEX R-K schemes do not satisfy
these additional order conditions.
This approach allows the construction of new IMEX R-K schemes producing an
error which is more uniform in the stiffness parameter. In particular, such conditions
are used to improve some existing scheme (see also [2]). Our analysis is based on the

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1928 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

smoothness assumption of the solution and also applies to the stiff case when Δt  ε
(see [2], [3]).
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The outline of the paper is as follows. In the next section we present the general
structure of IMEX Runge–Kutta schemes. In section 3 we report the new order
conditions up to third-order derived from [3]. Section 4 is devoted to the development
of uniformly accurate IMEX R-K schemes. Important properties are proved and new
assumptions are considered. In section 5 we perform several test problems to check the
accuracy of new schemes for various values of the stiffness parameter and we discuss
the result. Conclusions are drawn and work in progress is mentioned. Appendices are
also included.
2. IMEX Runge–Kutta schemes and classification. An IMEX R-K scheme
applied to (1) has the following form:
s
 s
 1
(3) Un+1 = Un + h b̃i F (tn + c̃i h, U i ) + h bi G(tn + ci h, U i ),
i=1 i=1
ε
with internal stages given by
i−1
 i

i i 1
(4) U = Un + h ãij F (tn + c̃i h, U ) + aij G(tn + ci h, U i ).
j=1 j=1
ε

The matrices (ãij ), with ãij = 0 for j ≥ i, and (aij ) are s × s matrices such that
the resulting method is explicit in F , and implicit in G. Using a diagonally implicit
method (aij = 0, for j > i) for G gives a sufficient condition to guarantee that
F is always evaluated explicitly; furthermore, this choice is usually preferred over
full implicit schemes for efficiency reasons. The coefficient vectors c̃ = (c̃1 , . . . , c̃s )T ,
b̃ = (b̃1 , . . . , b̃s )T , c = (c1 , . . . , cs )T , b = (b1 , . . . , bs )T complete the characterization
of the schemes. They can be represented by a double tableau in the usual Butcher
notation,
c̃ Ã c A
.
b̃ T bT
We assume that the coefficients c̃ and c, used for the treatment of nonautonomous
systems, are given by the relation
i−1
 i

(5) c̃i = ãij , ci = aij .
j=1 j=1

The great number of IMEX R-K methods presented in literature lead us to clas-
sify them in three different types characterized by the structure of the matrix A =
(aij )si,j=1 of the implicit scheme [2].
Definition 1. We call an IMEX R-K scheme of type A, (see [18]), if the matrix
A ∈ Rs×s is invertible.
Definition 2. We call an IMEX R-K scheme of type CK, (see [7]), if matrix
A ∈ Rs×s can be written as
 
0 0
A=
a Â
with the submatrix  ∈ R(s−1)×(s−1) invertible.
Remark 1. IMEX R-K schemes, called of type ARS (see [1]), are a special case
of the type CK with the vector a = 0.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1929

3. Additional order conditions. IMEX Runge–Kutta methods can be viewed

as a particular class of partitioned Runge–Kutta methods. Therefore, their order
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

conditions can be derived from the general theory of partitioned methods [12]. In
particular, for a detailed account of the classical order conditions for the IMEX R-K
schemes see, for example, [7] and [16].
In [3, 4] we derived additional order conditions that together with the classical
ones ensure accuracy at the various order in the stiffness parameter ε.
In his paper [2] Boscarino studied the global error behaviour of some popular
IMEX R-K schemes, showing a degradation of the order of accuracy for small values
of ε.
By applying to IMEX R-K schemes a technique similar to the one used in [11]
Chapt. VII in the case of implicit R-K schemes, additional order conditions are
obtained by imposing that the numerical solution and the exact solution agree to
various order in an expansion of ε. This analysis explains the degradation in the
order of accuracy of existing IMEX R-K schemes, and allowed the construction of new
schemes with better uniform accuracy in ε. For a detailed analysis we refer [3], [4].
Let us write explicitly the additional order conditions up to third order. Then,
the index 1 and 2 order conditions for type A schemes read:
Index 1 Order Conditions
  
2
(6) bi ωij c̃j = 1, i,j bi ωij c̃j = 1, i,j,k bi ωij ãjk c̃k = 1/2,
i,j

Index 2 Order Conditions

 
bi ωij ωjk c̃2k = 2, 2
i,j,k bi ωij ωjk ck = 2,
i,j,k
 
(7) bi ωij ωjk c̃k ck = 2, i,j,k,l bi ωij ωjk ãkl c̃l = 1,
i,j,k

bi ωij ωjk ãkl cl = 1,
i,j,k,l

where ωij are the elements of the inverse matrix of A. Observe that we deduced these
order conditions for the type A; for the type CK (consequently for type ARS) we can
rewrite the same ones replacing ωij with ŵij (the elements of the inverse of the matrix
Â) with i, j, k from 2 to s.
Then, it may be advantageous to have schemes that satisfy these new order con-
ditions. Although the goal of this paper is to present a class of numerical schemes
that work with uniform accuracy in the whole range of the stiffness parameter ε, we
will show that the use of these additional order conditions improves existing schemes.
In [2, 5], we introduced two IMEX R-K schemes, called MARK3(2)4L[2]SA and
MARS(3,4,3), that are an improvement of the existing schemes, ARK3(2)4L[2]SA [7]
and ARS(3,4,3) [1]. These schemes use the whole set of classical order conditions
plus conditions (6) and show a good accuracy behavior with respect to the increasing
stiffness in the limit of ε = 0 where the system becomes an index 1 differential algebraic
system (see [2]). In section 5 we compare the different performances of these schemes.
4. Construction of a particular uniformly accurate IMEX R-K scheme.
In [7], [16], [18], [17], [1], and [14] IMEX R-K schemes were introduced. As we saw in
[2] and [4] none of them are suitable to preserve high-order accuracy uniformly with
respect to the wide range of stiffness parameter ε when applied to stiff system (1).

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1930 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

Here we construct a third-order uniformly accurate IMEX Runge–Kutta scheme with

respect to the stiffness parameter ε through several assumptions and additional order
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

conditions. First of all, we require the following assumptions:

0. bi = b̃i , ci = c̃i for i = 1, . . . , s,
i
1. aij ck−1
j = cki /k for i = 2, . . . , s and k = 1, 2,
j=1
i−1
2. ãij ck−1
j = cki /k for i = 3, . . . , s and k = 1, 2,
j=1
s s s 2
3. i=1 bi = 1, i=1 bi ci = 1/2, i=1 bi ci = 1/3,
4. b̃2 = b2 = 0.
1 and 2 are the simplifying assumptions C(2) for the implicit and explicit part with
ci = c̃i for i = 1, . . . , s, which will be helpful in designing IMEX R-K schemes because
they simplify the classical order conditions and conditions (6), (7) (see [11, 7]). The
technical condition 4 is necessary because the assumption 2 cannot be satisfied for
i = 2 in the explicit scheme. Furthermore, the classical order conditions for IMEX
R-K schemes ([7], [16]) simplify a lot if assumption 0 is satisfied, so that the standard
order p conditions for the explicit and implicit schemes are enough to guarantee the
same order p for the IMEX R-K schemes, for p ≤ 3. Notice that if assumptions 0 and
1 for k = 1, 2, are satisfied, then classical order conditions for the two R-K schemes
imply that coupled order conditions are automatically satisfied for schemes up to third
order.
Now, among the different types of IMEX schemes considered in [2], we choose
to construct schemes of type CK because of their good properties. Concerning these
type CK schemes, we consider schemes that are in the implicit part stiffly accurate
(asi = bi , for i = 1, . . . , s) and singly diagonally implicit (SDIRK) with aii = γ > 0,
for all i = 2, . . . , s and a11 = 0. Stiffly accurate guarantees that an A-stable scheme
is also L-stable (see [11]), and the SDIRK assumption here is used to simplified the
analysis. Moreover, the implicit part of this scheme differs from the classical SDIRK
one because aii = 0. In [7] these schemes are called explicit singly diagonally implicit
R-K schemes (ESDIRK). A consequence to set a11 = 0 is that we have the possibility
to guarantee a higher stage-order for ESDIRK scheme than the classical one for a
SDIRK scheme, i.e., 1. For instance, a stage order of two for the scheme, i.e., C(2) for
every i-stage, is obtained by imposing the assumption 1 for i = 2, .., s − 1 with k = 2.
We remark that if the scheme is stiffly accurate condition 1 for i = s and k = 2 is
equivalent to the second condition in 3.
Moreover, the additional order conditions (6), (7) can be simplified using 2 and
1 written in the form
 i
(8) ω̂ij ckj = kck−1
i for i = 2, . . . , s and k = 1, 2,
j=2

where ω̂i,j are the elements of the inverse of Â.

Now, we have already combined the properties of the type CK IMEX scheme,
order conditions (6), (7), and assumptions 0, 1, 2, 3, 4, but we have to require that
other conditions are satisfied.
We first impose a stability condition on the implicit part of the scheme. Let us
denote by R(z) the stability function of the implicit part of the scheme. R(z) is the
numerical solution obtained after one time step Δt by applying a R-K scheme to the
equation
(9) y  = λy, y(0) = 1,
with λ ∈ C and z = λΔt.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1931

We know that a standard SDIRK scheme with nonsingular A matrix that satisfies
the condition asi = bi for all i has R(∞) = 0 and this makes an A-stable scheme L-
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

stable [11].
However, for IMEX R-K schemes of type CK with a11 = 0, where the matrix
A has the structure given in definition 2, the only stiffly accurate condition is not
enough to guarantee that R(∞) = 0, and, then an additional condition is required.
Lemma 3. If

(10) −eTs Â−1 a = ω̂sj aj1 = 0,
j≥2

then R(∞) = 0, where es = (0, . . . , 0, 1)T and the elements ω̂ij are the elements of
the inverse of Â.
In order to obtain this result, we apply one step of the implicit part to (9), which
yields for the internal stages
(11) Y = 1s + zAY
with 1s = (1, . . . , 1)T . Now, if we set Y = (Y1 , U T ), the expression (11) becomes
Y1 = 1,
(12)
U = 1s−1 + za + z ÂU,

where a = (a21 , . . . , as1 )T . Therefore, we get

(13) U = (I − z Â)−1 (1s−1 + za).
Now, if the scheme is stiffly accurate, the numerical solution is equal to the last internal
stage of the scheme and, therefore, when z → ∞, R(∞) = −eTs Â−1 a, assumed Â
invertible. Then we can state that if the type CK scheme is stiffly accurate and the
condition (10) is satisfied, we get R(∞) = 0. However, thus is not sufficient for the
L-stability because the implicit part of the type CK scheme should also be A-stable.
If we want this implicit part of the type CK schemes to be A-stable, we have to
consider again the stability function R(z). We know that the stability function R(z)
for a SDIRK scheme has the following form ([11], Sect. IV.6)
P (z)
(14) R(z) = ,
Q(z)
where Q(0) = 1 and both P (z) and Q(z) are polynomials of degree at most s. For
P (z)
ESDIRK schemes with a11 = 0 and a22 = · · · = ass = γ, we have R(z) = (1−γz) s−1 .

In particular, R(∞) = 0 implies that the degree of P (z) has to be less than s − 1;
consequently, the polynomial P (z) has degree at most s − 2, and if the order of the
scheme is at least p ≥ s − 2, it follows from (1 − zγ)s−1 ez = P (z) + Cz s−1 + · · · that
the coefficients of P (z) are uniquely determined in terms of γ and we have
s−2
 (s−1−j) 1
(15) P (z) = (−1)s−1 Ls−1 ( )(γz)j
j=0
γ

with C = (−1)s−1 Ls−1 (1/γ)γ s−1 , where

s−1
  
s − 1 xj
j
(16) Ls−1 (x) = (−1) ,
j=0
j j!

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1932 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

2
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1.5

1 γ = 0.57281606

0.5

γ = 0.3
0

−0.5
γ = 0.5

−1

γ = 0.4
−1.5

−2
−1.5 −1 −0.5 0 0.5

Fig. 1. Stability domains for increasing values of γ in the interval (25).

(k)
is the s-degree Laguerre polynomial and Ls denotes the kth derivative (see [11],
Sect. IV.6). Now, provided that γ > 0, it follows that the stability function (14) is
analytic in C − and we obtain A-stability if there is stability in the imaginary axis
(I-stability, see [11]). It is equivalent to the fact that the polynomial
2 2
(17) E(y) = |Q(iy)| − |P (0, iy)|

satisfies E(y) ≥ 0 for all y ∈ R. Then, if E(y) ≥ 0 for all y the implicit part of the
type CK IMEX R-K scheme is A-stable, and we can state that the IMEX R-K scheme
is L-stable.
We remark that it is straightforward to provide that the stability function R(z̃, z)
for a IMEX R-K scheme of type CK has the form
s−2

P (z̃, z) (s−1−j) 1
(18) R(z̃, z) = , P (z̃, z) = (−1)s−1 Ls−1 (z̃, )(γz)j
(1 − γz)s−1 j=0
γ

with error constant C(z̃) = (−1)s−1 Ls−1 (z̃, γ1 )(γ)s−1 , where

  s−1   j−k
(k) 1 j s − 1 (1/γ)
(19) Ls−1 z̃, = (−1) P̃ (j−k) (z̃),
γ j (j − k)!
j=k

(k)
and P̃ (z̃) = 1 + z̃/1! + z̃ 2 /2! + · · · + z̃ s /s! . Ls−1 denotes the kth derivative.
Several stability domains for IMEX R-K schemes of type CK are given in Figure
1 for different values of γ.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1933

Finally, a careful consideration follows from the proof of the Theorem 6.2 in [2]
about the type CK. In fact, we require that the formula
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

s

(20) bi ω̂i,j ω̂j,2 = 0,
j=2

for i = 2, . . . , s, is included among the assumptions in order to have at least second-

order accuracy in time in the algebraic variable. This is due to the fact that without
(20) the accuracy of the stiffness (or algebraic) variable drops to first order. Hence
an extra care must be taken to properly obtain a third-order type CK IMEX R-K
scheme. The derivation of condition (20) is reported in Appendix 1.
 Remark 2. It is noteworthy that if the method is stiffly accurate the assumption
i,j bi ω̂ij ω̂j2 = 0 is equivalent to ω̂s2 = 0 and in particular it follows from (10) that
j≥3 ω̂sj aj1 = 0.
Now, in the construction of a third-order type CK IMEX R-K scheme, we first
search schemes with four internal stages. In order to be able to design such a method
we require to solve the system of equations from 1 to 4 with the condition (10) and
the assumption (20). Unfortunately there are not third-order, four-stage schemes
satisfying the previous assumptions. This is formally stated in the following.
Proposition 1. Consider IMEX Runge–Kutta scheme of type CK, stiffly ac-
curate (i.e., asi = bi for i = 1, . . . , s) in the implicit scheme, with bi = b̃i , ci = c̃i
for all i and a11 = 0. Then there exist no third-order, four-stage schemes satisfying
assumptions 1–4 with condition (10) and assumption (20).
Proof. By remark 2 for s = 4 we get ω̂42 = 0 and ω̂43 a31 + ω̂44 b1 = 0. Then, by
assumptions 1 with k = 2, 3, and 4 the resulting reduced system of equations can be
explicitly written as

a32 b3
(21a) = 0,
γ3
b3 a31 b1
(21b) + = 0,
γ2 γ
(21c) c2 = 2γ,
(21d) 2a32 γ + γc3 = c23 /2,
(21e) b3 c3 + γ = 1/2,
(21f) b3 c23 + γ = 1/3

with a21 = c2 −γ and a31 = c3 −a32 −γ by 1 with k = 1, and b1 = 1−b3 −γ. In order to
state the proposition one can prove easily by trivial computational that it is possible
to obtain several contradictions. There is a total of five coefficients to determine:
γ, c2 , c3 , b3 , and a32 . We require that the conditions (21a), (21b) are satisfied
simultaneously and as consequence from (21a) we consider a32 = 0 with a31 = c3 − γ.
The resulting reduced system (21) can be computed as follows:
(a)

1/3 − γ 1/2 − γ
(22) c3 = , b3 =
1/2 − γ c3

from (21e) and (21f);

(b) by (21b) we compute γ and then b1 .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1934 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

(c) Using (21d) we get c3 = 0 or c3 = 2γ and from (a) we have c3 = (1/3 −

γ)/(1/2 − γ).
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Hence, if we substitute the values of γ computed above, we obtain contradictions

about the values of c3 . Notice that if c3 = 0, b3 is not defined. Furthermore, it can
be verified again by trivial computations that by choosing b3 = 0 in (21a) one obtains
other contradictions in (a).
Then, a third-order type CK IMEX R-K scheme is designed in five stages, i.e.,
s = 5. Making use of conditions 0, 4, and 1, 2 with k = 1, the Butcher tableau of a
IMEX R-K scheme of type CK becomes
0 0 0 0 0 0 0 0 0 0 0
c2 c2 0 0 0 0 c2 − γ γ 0 0 0
c3 c3 − ã32 ã32 0 0 0 c3 − a32 − γ a32 γ 0 0
.
c4 c4 − ã42 − ã43 ã42 ã43 0 0 c4 − a42 − a43 a42 − γ a43 γ 0
1 b1 0 b3 b4 γ b1 0 b3 b4 γ
b1 0 b3 b4 γ b1 0 b3 b4 γ
Then, there is a total of 16 coefficients to be determined: the 10 coefficients, c2 , c3 ,
c4 , a32 , a42 , a43 , b1 , b2 , b3 , b4 , γ for the implicit part and 6 coefficients ã32 , ã42 , ã43 ,
ã52 , ã53 , ã54 for the explicit one.
In order to determine the 10 coefficients of the implicit scheme, we use the fol-
lowing conditions.
1. Assumption 3.
b1 = 1 − b3 − b4 − γ,
(23a) b3 c3 + b4 c4 + γ = 1/2,
b3 c23 + b4 c24 + γ = 1/3.
2. Assumption 2 for i = 2, 3, 4 with k = 2,
γc2 = c22 /2,
(23b) a32 c2 + γc3 = c23 /2,
a42 c2 + a43 c3 + γc4 = c24 /2.

3. By remark 2 condition (10) is j≥3 ω̂sj aj1 = 0, i.e.,
b4 a43 a31 b3 a31 b4 a41 b1
(23c) 3
− 2
− 2
+ = 0.
γ γ γ γ
with a31 = c3 − a32 − γ and a41 = c4 − a42 − a43 − γ.
4. Analogously considering remark 2, additional condition (20) for s = 5 is
ω̂52 = 0, i.e.,
b3 a32 b4 a42 b4 a43 a32
(23d) 3
+ 3
− = 0.
γ γ γ4
Now, from the first and second formula in (23b) we get c2 = 2γ and a32 = c3 (c2c 3 −2γ)
2
.
In particular, we have a21 = c2 − γ = γ. Using the third formula in (23b), the second
and third one in (23a) and (23d), we compute b3 , b4 , a42 , and a43 as functions of c3
and c4 and b1 by (23a).
Finally, by using the condition (23c) and substituting the quantities computed
above b1 , ai1 , for i = 3, 4 and a43 , by algebraic manipulations, we obtain
2(6γ 2 − 6γ + 1)
(24) c3 =
3(2γ 2 − 4γ + 1)
in function of γ. Then we have only two free parameters γ and c4 .

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1935

Now, in order to determine the optimal values for a good choice of the parameter
γ, we, explicitly, give the formula for the polynomial (17), with s = 5. Then, we have
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

with s − 1 = 4 and order p ≥ s − 2 = 3,

E(y) = y 4 (1/12 − 4γ/3 + 6γ 2 − 8γ 3 + 2γ 4 )

+ y 6 (−1/36 + 2γ/3 − 6γ 2 + 76γ 3 /3 − 52γ 4 + 48γ 5 − 12γ 6 ) + y 8 γ 8 .

Therefore, A-(and hence L-)stability means that all the coefficients of E(y) must be
nonnegative and, then, we obtain the following interval (see [11], Table 6.4, p. 98):

(25) γ1 ≤ γ ≤ γ2 ,

where γ1 = 0.22364780 . . . and γ2 = 0.57281606 . . . . Notice that for different values

of γ in the interval, c4 has been computed to minimize the fourth-order error constant
([10]).
We now turn our attention to the evaluation of the six coefficients of the explicit
part of the scheme. Therefore, assumption 2 for i = 3, 4, 5, with k = 2 and the
technical condition 4 give

ã32 c2 = c23 /2,

(26) ã42 c2 + ã43 c3 = c24 /2,
ã52 c2 + ã53 c3 + ã54 c4 = 1/2.

Then, by (26) and c2 = 2γ we get

ã32 = c23 /(4γ),

(27) ã43 = c24 /(2c3 ) − 2γã42 /c3 ,
ã54 = 1/(2c4 ) − (2γã52 /c4 + ã53 c3 /c4 ).

Thus, we have only to determine the following coefficients: ã42 , ã43 , ã52 , and ã53 .
In order to compute these coefficients we make the following assumption:

(28) bi ãij ãjk ck = 1/24,
i,j,k

which is one of the order conditions for the explicit scheme. When applied (28) to a
linear system with constant coefficients, such condition provides fourth-order accuracy.
 It is noteworthy
2
sthe assumption 2 with k = 2, (28) implies
that, considering
i,j bi ãij cj = 1/12, if and only if i=3 bi ãi2 = 0. Then, we here consider the
following equation:

(29) b3 ã32 + b4 ã42 + γã52 = 0.

Hence ifwe choose ã42 = 0 we can immediately compute ã52 by (29), ã43 by (27), and
ã53 by i,j bi ãij c2j = 1/12. Finally, we get ã31 , ã41 , ã51 .
5. Test problems. In this section we investigate numerically the convergence
rate for a wide range of parameter ε. To this aim we apply to several test problems
the IMEX R-K scheme introduced in section 3 and the new schemes introduced in
section 4.
Numerical convergence rate is calculated by the formula

(30) p = log2 (EΔt1 /EΔt2 ),

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1936 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

where EΔt1 and EΔt2 are the global errors computed with step Δt1 and Δt2 = Δt1 /2.
The convergence rates for the different schemes are reported in the tables.
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Before introducing some numerical examples, we start to suppose that the initial
data of our test problems lie on a suitable manifold that allows smooth solutions even
in the limit of infinite stiffness, and that the step size Δt  ε. In fact, arbitrary
initial data introduce in the solution a fast transient. One possibility to overcome
this difficulty is simply to ensure that the numerical scheme resolves the transient
phase by taking time step Δt  ε in the first few steps. Then, the following results
are obtained assuming that the transient phase is over. A simple way to avoid the
presence of the so-called initial layer is to suitably choose the initial data in such
a way that the initial layer does not form. This is obtained by using the so-called
well-prepared data, as outlined below.
Consider a linear system with stiff relaxation source term [19]

(31a) ∂t u + ∂x v = 0
(31b) ∂t v + ∂x u = (au − v)/ε,

with ε > 0 and a constant. An expansion in ε can be performed directly at the level
of the system (31), by using the so-called Chapman–Enskog expansion.
Consider a formal expansion of the solution to system (31) in the form

u = u0 + εu1 + ε2 u2 + · · ·
(32)
v = v0 + εv1 + ε2 v2 + · · ·

and insert it in (31).

The leading term approximation to (31) is v0 = au0 and ∂t u0 + a∂x u0 = 0. In
fact, the second equation (31b), to order ε−1 , gives

v0 = au0 ,

and using this relation in equation (31a) to order ε0 gives

∂t u0 + a∂x u0 = 0.

This equation is formally obtained as ε → 0 and is called relaxed equation.

The first-order correction to the leading term approximation is obtained as follows.
The equation (31b), to zero the order in ε, gives

∂t v0 + ∂x u0 = au1 − v1 ,

which gives

v1 = au1 − (1 − a2 )∂x u0 .

The equation (31a), to first order, gives

∂t u1 + a∂x u1 = ε∂x ((1 − a2 )∂x u0 ).

By keeping first-order terms in the expansion (32) and neglecting second and higher
order terms, one has

u = u0 + εu1 , v = v0 + εv1

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1937

and obtain
v = au − (1 − a2 )∂
 x u,
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(33)
∂t u + a∂x u = ε∂x (1 − a2 )∂x u).
The second equation in (33) is a convection-diffusion equation with viscosity coefficient
ν = ε(1 − a2 ). This equation is dissipative if ν ≥ 0, i.e., |a| ≤ 1 (subcharacteristc
condition of Liu [15] for (31)).
Motivated by this analysis we perform an accuracy test for the test problems
considering well-prepared initial data. Then, concerning the test problem (31), we
have considered a periodic smooth solution and well-prepared initial data by u(x, 0) =
sin(2πx) and v(x, 0) = v0 (x, 0) + v1 (x, 0), where v0 (x, 0) = au(x, 0) and v1 (x, 0) =
(a2 − 1)∂x u0 . We set a = 0.5 and the final time is t = 0.2. The system is integrated
for x ∈ [0, 2].
The second test problem is the Broadwell model equations:
∂t ρ + ∂x m = 0,
(34) ∂t m + ∂x z = 0,
∂t z + ∂x m = 1ε (ρ2 + m2 − 2ρz).
Here ε represents the mean free path of particles. The dynamical variables ρ and m
are the density and the momentum, respectively, while z represents the flux of the
momentum. As ε → 0, z is given by a local Maxwellian distribution
1 2
(35) z = zE (ρ, m) = (ρ + m2 ),
2ρ
and we are in the fluid dynamic limit, satisfying the equation
∂t ρ + ∂x (ρu) = 0,
1
∂t (ρu) + ∂x ( 2ρ (ρ2 + m2 )) = 0,

with u = m/ρ. We have considered a periodic smooth solution with initial well-
prepared data given by
ρ(x, 0) = ρ0 (x), u(x, 0) = u0 (x), m(x, 0) = m0 (x) = ρ0 (x)u0 (x),
z(x, 0) = zE (ρ0 (x), m0 (x)) + εz1 (ρ0 (x), m0 (x)).
   2πx 
where ρ0 (x) = 1 + aρ sin 2πx 1
L , u0 (x) = 2 + au sin L , and
−H(ρ0 ,m0 )
z1 (ρ0 , m0 ) = ρ0

with H(ρ0 , m0 ) = ((1 − ∂ρ zE + (∂m zE )2 )∂x m0 + (∂ρ zE ∂m zE )∂x ρ0 ).

Periodic boundary conditions are imposed. In our computations we used the
parameters aρ = 0.3, au = 0.1, L = 20 and we integrate the equation for t ∈ [0, 30].
Finally, we present other numerical results obtained concerning situation in which
hyperbolic system with relaxation plays a major role in applications. In fact, we
consider as third test problem that was used by Shi Jin in [13]:
∂t h + ∂x w = 0
(36)
∂t w + ∂x (h + 0.5h2 ) = − 1ε (w − 0.5h2 ),
with the well-prepared initial data given by h(0, x) = 1 + 0.2 sin(8πx) and w(0, x) =
w0 +εw1 with w0 = f (h(0, x)) = 0.5h2 (0, x) and w1 = (f  (h(0, x))−p (h(0, x)))∂x h(0, x)
where p(h) = (h + 0.5h2 ). The results have been obtained for fixed Δx.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1938 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

Table 1
Convergence rate for v in L∞ -norm.
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Schemes ε=1 ε = 10−1 ε = 10−3 ε = 10−4 ε = 10−5 ε = 10−6 ε = 10−8

ARS 2.982 2.970 2.471 2.386 2.041 2.003 1.999
MARS 2.976 2.741 1.905 1.142 2.889 2.980 2.997
ARK3 3.012 2.833 2.443 2.108 2.012 2.002 2.002
MARK3 2.989 2.740 1.905 1.144 2.118 2.980 2.996

Our test problems are computed with coarse grids (Δx, Δt  ε) that do not
resolve the small scales, and high accuracy in space is obtained by finite difference
discretization with weighted essentially nonoscillatory (WENO) reconstruction [20],
[17].
Discussion. Concerning our numerical experiments, we have constructed several
IMEX R-K schemes of type CK considering different values of γ in interval (25) and
we have obtained several schemes that qualitatively have shown a good accuracy
behavior. Different values of the parameter γ in (25) with a reasonable large stability
domain have been considered, and we report only the most relevant values for the
different test problems that guaranteed a good behavior about the accuracy on the
whole range of the parameters ε. These values are reported in the different tables
presented in this article. In particular, we have chosen the values of γ = γ2 and
γ = γ̃ with γ̃ = 0.43586652150845 because γ2 guarantees the largest stability domain
corresponding to the stability function (18), (see Figure 1), whereas γ = γ̃ is the same
value used in the IMEX R-K schemes introduced in section 3. All these values are a
good choice because the SDIRK scheme is L-stable. We report in the appendix only
the values of the coefficients for the schemes with γ = γ2 and γ = γ̃. In order to
identify the schemes derived in this article we name this new scheme BHR(s, ρ, p),
where the triplet (s, ρ, p) characterizes the number s of the stages of the implicit
part, ρ the number of the stages of the explicit one, and p the order of the scheme.
The coefficients of these schemes are displayed in the Appendix 2.
First we compare MARS(3,4,3) and MARK3(2)4L[2]SA schemes, with the
ARS(3,4,3), [1], and ARK3(2)4L[2]SA, [7] (see Table 1). Tables show the corre-
sponding convergence rate in L∞ -norm. Different norms give essentially the same
results.
We have obtained an improvement for the convergence of the different stiff com-
ponents. In fact, for the first problem we have increased the convergence rate for
sufficiently stiff parameters (ε < 10−4 ), namely when ε → 0. These results show a
third-order accuracy for small and large values of ε and note that for intermediate
values of the parameter ε (10−4 < ε < 10−2 ) we have a slight deterioration of the
accuracy.
In a similar way, it is possible to improve the convergence of variable z in the
Broadwell model (see Figures 2 and 3) and for the w variable in the shallow water.
Now we show that the new third order scheme have better uniform accuracy in
ε. Investigating the numerical convergence rate of this scheme for a whole range of ε.
As it is evident, from Figures 4 and 5 and from Tables 2 and 3, the BHR(5,5,3)
with γ = γ̃ verifies a third-order accuracy in the whole range of ε. Furthermore, we
note that in a neighborhood of γ̃ (from γ = 0.43 to γ = 0.45) we have obtained sev-
eral schemes with a good accuracy behavior (Tables 2, 3, and 4). Instead, BHR(5,5,3)
schemes, with γ = γ2 and γ = 0.5, show for intermediate values of ε a slight deterio-
ration of the accuracy.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1939

8
ARS(3,4,3) 8
MARS(3,4,3)
u−component u−component
v−component v−component
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

7 7

6 6

convergence rate
convergence rate

5 5

4 4

3 3

2 2

1 1

0 0
−6 −5 −4 −3 −2 −1 0 −6 −5 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10 10 10 10 10
epsilon epsilon

Fig. 2. Convergence rate vs ε for the density ρ (◦) (differential component) and the flux of the
momentum z (∗) (stiff component). On the left ARS(3,4,3) scheme, and on the right MARS(3,4,3)
scheme.

8
ARK3(2)4L[2]SA MARK3(2)4L[2]SA
8
u−component u−component
v−component v−component
7 7

6 6
convergence rate

convergence rate

5 5

4 4

3 3

2 2

1 1

0 0
−6 −5 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0

Epsilon epsilon

Fig. 3. Convergence rate vs ε for the density ρ (◦) (differential component) and the flux of the
momentum z (∗) (stiff component). On the left ARK3(2)4L[2]SA, and on the right ARK3(2)4L[2]SA
scheme.

In the rest of the section we show that the lack of accuracy of the schemes for
intermediate values of the stiffness parameter is consistent with a theoretical anal-
ysis. In order to obtain these results in a general setting, we consider the singular
perturbation problem

y  = f (y, z),
(37)
εz  = g(y, z),

with 0 < ε  1 where f and g are sufficiently differentiable. We are mainly

interested in smooth solutions of (37) which are of the form

y(t) = y0 (t) + εy1 (t) + ε2 y2 (t) + · · ·

(38)
z(t) = z0 (t) + εz1 (t) + ε2 z2 (t) + · · · ,

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1940 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

8
BHR(5,5,3) 8
BHR(5,5,3)
u−component u−component
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

v−component v−component

7 7

6 6
convergence rate

convergence rate
5 5

4 4

3 3

2 2

1 1

0 0
−8 −7 −6 −5 −4 −3 −2 −1 0 −8 −7 −6 −5 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
epsilon epsilon

Fig. 4. Convergence rate vs ε for the u-component (◦), (differential component) and v-
component (∗), (stiff component). On the left BHR(5,5,3) scheme with γ = γ2 , and on the right
BHR(5,5,3) scheme with γ = γ̃.

8
BHR(5,5,3) 8
BHR(5,5,3)
ρ−component ρ−component
z−component z−component

7 7

6 6
convergence rate
convergence rate

5 5

4 4

3 3

2 2

1 1

0 0
−8 −7 −6 −5 −4 −3 −2 −1 0 −8 −7 −6 −5 −4 −3 −2 −1 0
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
epsilon epsilon

Fig. 5. Convergence rate vs ε for the density ρ (◦) (differential component) and the flux of the
momentum z (∗) (stiff component). On the left BHR(5,5,3) scheme with γ = γ2 , and on the right
BHR(5,5,3) scheme with γ = γ̃.

Table 2
Convergence rate for v in L∞ -norm for different values of γ.

γ ε=1 ε = 10−1 ε = 10−2 ε = 10−3 ε = 10−4 ε = 10−6 ε = 10−8

BHR (γ = 0.43) 3.130 3.186 3.306 3.363 3.049 3.007 3.007
BHR (γ̃) 3.657 3.352 3.003 3.407 3.053 3.000 3.000
BHR (γ = 0.44) 3.112 3.101 3.132 3.371 3.051 3.008 3.007
BHR (γ = 0.45) 2.857 3.022 2.451 3.379 3.053 3.009 3.009
BHR (γ = 0.5) 3.257 2.882 1.596 3.370 3.062 3.009 3.008
BHR (γ2 ) 3.070 2.678 1.405 3.096 3.010 3.000 3.000

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1941
Table 3
Convergence rate for z-component in L∞ -norm for different values of γ.
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Schemes ε=1 ε = 10−1 ε = 10−2 ε = 10−3 ε = 10−4 ε = 10−6 ε = 10−8

BHR (γ = 0.43) 3.055 2.916 2.670 2.776 3.218 3.019 3.017
BHR (γ = 0.44) 3.050 2.921 2.780 3.539 3.200 3.019 3.016
BHR (γ = 0.45) 2.896 2.975 2.520 2.911 3.172 2.986 2.984
BHR (γ = 0.5) 3.177 3.000 2.708 2.388 3.069 3.064 3.063

Table 4
Convergence rate for w in L∞ -norm for different values of γ.

Schemes ε=1 ε = 10−1 ε = 10−2 ε = 10−3 ε = 10−4 ε = 10−6 ε = 10−8

ARS 3.582 3.117 2.962 2.858 3.340 2.140 2.113
ARK3 2.963 3.013 2.982 2.860 2.482 2.060 2.044
BHR (γ̃) 3.028 3.026 2.998 3.072 3.525 3.210 3.185
BHR (γ = 0.45) 3.119 2.994 2.930 3.117 3.146 3.211 3.187
BHR (γ = 0.5) 3.186 3.024 2.970 2.730 2.070 3.036 3.012
BHR (γ2 ) 3.545 3.317 2.930 2.542 2.165 3.071 3.050

where yi (t) and zi (t) are ε-independent functions. We suppose that the initial values
lie on a suitable manifold that allows smooth solutions even in the limit of infinite
stiffness. We remark that a sequence of differential algebraic systems arises in the
study of problem (37). In [2], Boscarino showed that the problem (37) gives us the
possibility of studying the dependence of the global error on the stiffness parameter
ε, valid as ε < Δt. In fact, we may expand the global error in power of ε and show
that the coefficients of the global error are the global errors of the IMEX R-K schemes
applied to a differential algebraic system of different index.
Concerning (37), the global error satisfies

Δyn = Δyn0 + εΔyn1 + ε2 Δyn2 + O(ε3 )

(39)
Δzn = Δzn0 + εΔzn1 + ε2 Δzn2 + O(ε3 /Δt)

with Δyn = yn − y(tn ) and Δzn = zn − z(tn ) where Δyn0 = yn0 − y0 (tn ), Δzn0 =
zn0 − z0 (tn ), . . . are the global errors of the scheme applied to a differential-algebraic
system of different index whereas O(ε3 ), O(ε3 /Δt) are the estimates on the remainder.
By imposing additional order conditions obtained by matching the exact solution
and the numerical solution at various orders in ε, we produce an error which is more
uniform in ε. To do that, in [3], we imposed that the IMEX R-K scheme is of a given
order p for the hierarchy of differential-algebraic systems (e.g., indexes 1 and 2).
In order to explain the lack of accuracy, we start observing that, for ε > Δt,
classical analysis can be used to estimate the error. In fact, the classical global error
is of order O(Δt/ε)p . As both estimates about the global error have to be satisfied,
we can write, for example, for the z-component
  p 
Δt
(40) |error| < min C , Δzn0 + εΔzn1 + ε2 Δzn2 + O(ε3 /Δt)
ε

(see Figure 6).

We put the order of the scheme p = 3. Now, the balance in the arguments of the
right-hand side of equation (40) gives that:

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1942 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

−7
x 10
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1.6

1.4
Taylor expansion error
1.2
ERROR

0.8
ε −expansion error

0.6
numerical error
0.4

0.2

0
−6 −5 −4 −3 −2 −1
10 10 10 10 10 10
EPSILON

Fig. 6. Numerical global error (∗) for z-component (stiff component) using BHR(5,5,3) IMEX
R-K scheme and (◦) theoretical global error.

1. about the algebraic variable z, the dominant term, in the worse case when
Δt  ε, is O(ε3 /Δt) and then we have
 3
(41) C Δt ε ≈ O(ε3 /Δt),
2
which gives ε ≈ Δt 3 , so to get the maximum error O(Δt).
2. In a similar way, for the differential variable y, the dominant term is O(ε3 ),
3
which gives a maximum error O(Δt) 2 . These results state that the third-order
schemes can degrade up to first order in the worst case, for example, for the z-
component (such case is observed in Figures 3 and 4). Notice, however, that this is
a lower bound for the convergence rate. In many case a better behavior is observed.
For example, for BHR(5,5,3) scheme, in a neighborhood of γ = γ̃, the error is more
uniform.
6. Conclusion. IMEX R-K schemes are introduced for application to hyperbolic
systems with relaxation. Additional conditions up to the third order are provided to
improve the accuracy of several schemes with respect to the stiffness parameter ε. For
instance, MARK3(2)4L[2]SA and MARS(3,4,3) schemes are slightly more efficient in
the limit of ε = 0 than the ARK3(2)4L[2]SA and ARS(3,4,3) ones. Construction of
a third-order accurate IMEX R-K scheme is presented and numerical tests on several
hyperbolic systems with relaxation term reveal better behaviors for this new scheme
called BHR(5,5,3) than other ones presented in the literature. As it is evident from
the figures and tables, the BHR(5,5,3) schemes with γ = γ2 is third order accurate for

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1943

small and large values of ε, and there is a slight deterioration of the accuracy in the
intermediate regime. This lack of the accuracy for intermediate values of the stiffness
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

has been explained through a theoretical analysis.

A scalar stability analysis has also shown that the value of γ = γ2 has the largest
stability domain with respect to the other values in the interval (25). However, nu-
merical stability is not the central issue here. We remark that a stability analysis for
IMEX schemes applied to stiff systems has been performed by Higueras et al., see [9],
using contractivity and monotonicity properties and concept of algebraic stability for
additive R-K methods. Finally, concerning the asymptotic limit as ε → 0, this new
scheme becomes a good (high order) approximation of the equilibrium equations; i.e.,
it has the correct asymptotic limit.

Appendix 1. To derive the condition (20), here we just outline the main ideas.
To this end, the condition (20) is an immediate consequence of Theorem 6.2 in [2].
We denoted the difference to the exact solution values by

Δynν = ynν − yν (tn ), Δznν = znν − zν (tn ),

ν ν ν ν
(42) ΔYni = Yni − yν (tn + c̃i h), ΔZni = Zni − Zν (tn + ci h),
 
ν ν
Δkni = kni − yν (tn + c̃i h), Δ νni ν
= ni − zν (tn + ci h).

ν = 0 or ν = 1 in (42) correspond, respectively, to the numerical solution of the index

1 or index 2 problem, (see [11, 2]).
Furthermore, from Theorem 5.2 in [2], we deduced that

Δyn0 = O(hq̃+2 + hp ), ΔYni

0
= O(hq̃i +1 ), Δkni
0
= O(hq̃i +1 ),
(43)
Δzn0 = O(hq̃+1 + hp ), ΔZni
0
= O(hq̃i +1 ),

with q̃ = min {q̃s , q̃i + 1, i = 2, . . . , s − 1} where q̃i the stage order of the ith stage of
the explicit part of the IMEX R-K scheme.
Moreover, we obtained that
   
(44) Δ 0
ni = O(Δyn1  + Δzn1 ) + h−1 ω̂i2 (ΔZn2
0
− Δzn0 ) + O(hq̃i ) + O(hqi ),

with qi the stage order of the ith stage of the implicit part, for i = 2, . . . , s. Conse-
quently, evaluated (see [2])

1 0 1
(45) ΔZni = C1 Δ ni + C2 ΔYni + O(hq̃i +1 )

and

1
(46) ΔYni = Δyn1 + Ch(||Δyn1 || + ||Δzn1 ||) + O(h2 ),

substituting the formula (44) and (46) in the expression (45) we get
       
1
ΔZni = O(Δyn1  + Δzn1 ) + Ch(Δyn1  + Δzn1 )
(47) + C1 h−1 ω̂i2 (ΔZn2
0
− Δzn0 ) + O(h2 ).

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 1944 SEBASTIANO BOSCARINO AND GIOVANNI RUSSO

By the properties of the type CK IMEX scheme, (see [2]), we get the formula
 1      
Δzn+1  ≤ |ρ| Δzn1  + |bi ω̂ij | ΔZnj
1 
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

+ O(hqi +1 )
i,j

with |ρ| = |R(∞)| < 1. Then, substituting (47), it follows that

 1   1     
Δzn+1  ≤ Δyn  + (|ρ| + Ch) Δzn1  + h−1 C1 |bi ω̂ij ω̂i2 | ΔZn2
0
− Δzn0  + O(h2 )
i,j

0
By (43) and by the fact that q̃2 = 1, we have ΔZn2 = O(h2 ) and Δzn0 = O(h2 ). Then,
if (20) holds, we get
 1   1  
Δzn+1  ≤ Δyn  + (|ρ| + Ch) Δzn1  + O(h2 ).

Analogously, a direct estimation of Δyn1 is given and applying the Lemma 5.2 in [2] we
have Δyn1 = O(h2 ) for the differential component and Δzn1 = O(h2 ) for the algebraic
component. Again, we can conclude that if (20) is satisfied we have a second order
accuracy in time for the algebraic variable otherwise it drops to first order and we get
again the same estimates given in [2].
Appendix 2.
1. BHR(5,5,3) IMEX R-K method with c2 = 2γ, γ = γ̃ ≈ 424782/974569.

0 0 0 0 0 0 0 0 0 0 0 0
c2 2γ 0 0 0 0 c2 γ γ 0 0 0
c3 ã31 ã32 0 0 0 c3 a31 a32 γ 0 0
c4 ã41 0 ã43 0 0 c4 a41 a42 a43 γ 0
1 ã51 ã52 ã53 ã54 0 1 b1 0 b2 b3 γ
b1 0 b2 b3 γ b1 0 b3 b4 γ

ã31 γ,
=
ã32 ã31 ,
=
ã41 −475883375220285986033264/594112726933437845704163,
=
ã42 =
0,
ã43 =
1866233449822026827708736/594112726933437845704163,
ã51 =
62828845818073169585635881686091391737610308247/176112910684412105319781630311686343715753056000,
ã52 −b3 ,
=
ã53 =
262315887293043739337088563996093207/297427554730376353252081786906492000,
ã54 −987618231894176581438124717087/23877337660202969319526901856000,
=
a31 ã31 ,
=
a32 −31733082319927313/455705377221960889379854647102,
=
a41 −3012378541084922027361996761794919360516301377809610/
=
45123394056585269977907753045030512597955897345819349,
a42 = −62865589297807153294268/102559673441610672305587327019095047,
a43 = 418769796920855299603146267001414900945214277000/212454360385257708555954598099874818603217167139,
b1 = 487698502336740678603511/1181159636928185920260208,
b2 = 0,
b3 = 302987763081184622639300143137943089/1535359944203293318639180129368156500,
b4 = −105235928335100616072938218863/2282554452064661756575727198000.
c3 = 902905985686/1035759735069, c4 = 2684624/1147171.
2. BHR(5,5,3) IMEX R-K method with γ = γ2 ≈ 2051948/3582211, c2 = 2γ.

ã31 =
473447115440655855452482357894373/1226306256343706154920072735579148,
ã32 =
129298766034131882323069978722019/1226306256343706154920072735579148,
ã41 =
37498105210828143724516848/172642583546398006173766007,
ã42 =
0,
ã43 =
76283359742561480140804416/172642583546398006173766007,
ã51 =
(−3409975860212064612303539855622639333030782744869519)/
5886704102363745137792385361113084313351870216475136,
ã52 = (−237416352433826978856941795734073)/554681702576878342891447163499456,
ã53 = 4298159710546228783638212411650783228275/2165398513352098924587211488610407046208,
ã54 = 6101865615855760853571922289749/272863973025878249803640374568448,
a31 = 259252258169672523902708425780469319755/4392887760843243968922388674191715336228,
a32 = (−172074174703261986564706189586177)/1226306256343706154920072735579148,
a41 = 1103202061574553405285863729195740268785131739395559693754/
9879457735937277070641522414590493459028264677925767305837,
a42 = (−103754520567058969566542556296087324094)/459050363888246734833121482275319954529,
a43 = 3863207083069979654596872190377240608602701071947128/
19258690251287609765240683320611425745736762681950551,
b1 = (−2032971420760927701493589)/38017147656515384190997416,
b2 = 0,
b3 = 2197602776651676983265261109643897073447/945067123279139583549933947379097184164,
b4 = (−128147215194260398070666826235339)/69468482710687503388562952626424.
c3 = 12015769930846/24446477850549, c4 = 3532944/5360597.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.

 UNIFORMLY ACCURATE IMEX RUNGE–KUTTA SCHEMES 1945

Acknowledgments. The authors wish to thank the anonymous referees for sug-
gesting many improvements.
Downloaded 08/17/24 to 109.171.238.173 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

REFERENCES

[1] U. Ascher, S. Ruuth, and R. J. Spiteri, Implicit-explicit Runge-Kutta methods for time
dependent Partial Differential Equations, Appl. Numer. Math., 25, (1997), pp. 151–167.
[2] S. Boscarino, Error analysis of IMEX Runge-Kutta methods derided from differential alge-
braic systems, SIAM J. Numer. Anal., 45 (2008), pp. 1600–1621.
[3] S. Boscarino, On an accurate third order implicit-explicit Runge-Kutta method for stiff prob-
lems, Appl. Numer. Math., DOI 10.1016/j.apnum.2008.10.003.
[4] S. Boscarino, On the Uniform Accuracy of Implicit-Explicit Runge-Kutta Methods, Ph.D.
Thesis in Mathematics for the Technology, Department of Mathematics and Computer
Science, University of Catania, Italy, 2005.
[5] S. Boscarino and G. Russo, On the uniform accuracy of IMEX Runge-Kutta schemes and
applications to hyperbolic systems with relaxation, in Proceedings of SIMAI2006 VIII Con-
vegno SIMAI Ragusa (Italy), May 2006. Published in Communications to SIMAI Confer-
ences, 2006, DOI: 10.1685/CSC06028, ISSN 1827-9015, Vol. 2, 2007.
[6] R. E. Caflisch, S. Jin, and G. Russo, Uniformly accurate schemes for hyperbolic systems
with relaxation, SIAM J. Numer. Anal., 34 (1997), pp. 246–281.
[7] M. H. Carpenter and C. A. Kennedy, Additive Runge-Kutta schemes for convection-
diffusion-reaction equations, Appl. Numer. Math., 44 (2003), pp. 139–181.
[8] C. Q. Chen, C. D. Levermore, and T. P. Liu, Hyperbolic conservation laws with relaxation
terms and entropy, Comm. Pure Appl. Math., 47 (1994), pp. 787–830.
[9] B. Garcia-Celayeta, I. Higueras, and T. Roldan, Contractivity/monotonicity for additive
Runge-Kutta methods: Inner product norms, Appl. Numer. Math., 56 (2006), pp. 862–878.
[10] E. Hairer, S. P. Nørsett, and G. Wanner, Solving Ordinary Differential Equation I: Non-
stiff problems. Springer Series in Comput. Mathematics, Vol. 8, Springer-Verlag, Berlin,
1987, second revised edition, 1993.
[11] E. Hairer and G. Wanner, Solving Ordinary Differential Equation II: Stiff and Differential
Algebraic Problems. Springer Series in Comput. Mathematics, Vol. 14, Springer-Verlag,
Berlin, 1991, second revised edition, 1996.
[12] E. Hairer, Ch. Lubich, and G. Wanner, Geometric Numerical Integration: Structure-
Preserving Algorithms for Ordinary differential Equations, Springer series in Computa-
tional Mathematics, Vol. 31, Springer-Verlag, Berlin, 2002.
[13] S. Jin, Runge-Kutta methods for hyperbolic conservation laws with stiff relaxation terms, J.
Comput. Phys., 122 (1995), pp. 51–67.
[14] S. F. Liotta, V. Romano, and G. Russo, Central schemes for balance laws of relaxation type,
SIAM J. Numer. Anal., 38 (2000), pp. 1337–1356.
[15] T. P. Liu, Hyperbolic conservation laws with relaxation, Comm. Math. Phys., 108 (1987)
pp. 153–175.
[16] L. Pareschi and G. Russo, Implicit-Explicit Runge-Kutta schemes for stiff systems of dif-
ferential equations, Recent trends in numerical analysis, Adv. Theory Comput. Math., 3,
Nova Sci. Publ., Huntington, NY, 2001, pp. 241–251.
[17] L. Pareschi and G. Russo, Implicit-Explicit Runge-Kutta schemes and applications to hyper-
bolic systems with relaxations, J. Sci. Comput., 25 (2005), pp. 129–155
[18] L Pareschi and G. Russo, High order asymptotically strong-stability-preserving methods for
hyperbolic systems with stiff relaxation, Hyperbolic Problems: Theory, Numerics, Applica-
tions, Springer, Berlin, 2003, pp. 241–251.
[19] S. Qamar and G Warnecke, A Space-Time Conservative Method for Hyperbolic System with
Stiff and Non Stiff Terms, Commun. Comput. Phys., 1, pp. 449–478.
[20] C. W. Shu, Essentially Non Oscillatory and Weighted Essentially Non Oscillatory Schemes
for Hyperbolic Conservation Laws, in Advance Numerical Approximation of Nonlinear
Hyperbolic Equations, Lecture Notes in Math. 1697, (2000).
[21] G. B. Whitham, Linear and non-linear waves, Wiley, New York, 1974.

Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.