Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
STRONG STABILITY OF EXPLICIT RUNGE–KUTTA TIME1
DISCRETIZATIONS∗2
ZHENG SUN† AND CHI-WANG SHU‡3
Abstract. Motivated by studies on fully discrete numerical schemes for linear hyperbolic con-4servation laws, we present a framework on analyzing the strong stability of explicit Runge–Kutta5(RK) time discretizations for semi-negative autonomous linear systems. The analysis is based on the6energy method and can be performed with the aid of a computer. Strong stability of various RK7methods, including a sixteen-stage embedded pair of order nine and eight, has been examined under8this framework. Based on numerous numerical observations, we further characterize the features of9strongly stable schemes. A both necessary and sufficient condition is given for the strong stability of10RK methods of odd linear order.11
Key words. Runge–Kutta methods, strong stability, energy method, hyperbolic problems,12conditional contractivity.13
AMS subject classifications. 65M12, 65M20, 65L0614
1. Introduction. Explicit Runge–Kutta (RK) methods have been commonly15
used for time integration of hyperbolic conservation laws. In this paper, we study16
the strong stability of the methods in such context. We focus on autonomous linear17
ordinary differential equation (ODE) systems18
(1.1)d
dtu = Lu,19
which are obtained from method of lines schemes for linear hyperbolic problems.20
u ∈ RN and L is an N ×N real constant matrix, where N is the degrees of freedom21
for the spatial discretization. If the semidiscrete scheme honors the (weighted) L222
stability of the partial differential equation (PDE), then for certain symmetric and23
positive definite matrix H,24
(1.2) L>H +HL ≤ 025
is a semi-negative definite matrix. Here H can be related with both the symmetrizer of26
the PDE [16] and the mass matrix or quadrature weights of a Galerkin or collocation27
type spatial discretization. If (1.2) holds, then we say L is semi-negative and (1.1)28
satisfies the energy decay law29
(1.3)
d
dt‖u‖2H = 〈 d
dtu, u〉H + 〈u, d
dtu〉H
= 〈Lu,Hu〉+ 〈u,HLu〉 = 〈u, (L>H +HL)u〉 ≤ 0.30
Here 〈·, ·〉H = 〈·, H·〉 with 〈·, ·〉 being the usual l2 inner product in RN and ‖u‖H =31 √〈u, u〉H . We are concerned with whether this property is preserved at the discrete32
∗Submitted to the editors on Nov 25, 2018. This work is mainly done during the symposium heldat ICERM in Providence RI, for celebrating 75 years of Mathematics of Computation. We wouldlike to thank the committee members, speakers and staffs for this inspiring workshop. ZS especiallywants to express his appreciation for the travel and lodging support from ICERM.†Department of Mathematics, The Ohio State University, Columbus, OH 43210
([email protected]).‡Division of Applied Mathematics, Brown University, Providence, RI 02912
([email protected]). Research supported by ARO grant W911NF-16-1-0103 and NSF grantDMS-1719410.
1
2 Z. SUN AND C.-W. SHU
level, namely, whether33
(1.4) ‖un+1‖H ≤ ‖un‖H ,34
holds after applying an explicit RK time integrator under a suitably restricted time35
step. The time step constraint is referred to as the Courant–Friedrichs–Lewy (CFL)36
condition for numerical conservation laws. We would say the explicit RK method is37
strongly stable, if (1.4) is satisfied when discretizing (1.1) under the condition (1.2).38
Note this definition of stability is stronger than the usual one, which states that the39
norm of the numerical solution is bounded by the norm of the initial data up to a40
constant (typically dependent on the total time) [16]. We remark that the concept41
of strong stability in our context is also connected with the conditional contractivity42
from the ODE community [28, 18], while our analysis does not rely on the circle43
condition, which is usually assumed in the previous works.44
Solving hyperbolic conservation laws with explicit time discretizations can be45
dated back to early days. They seem to be the most natural choice to avoid the46
inversion of large nonlinear systems resulted from spatial discretizations. Many of47
the first order monotone schemes, which preserves various properties of the contin-48
uous PDE, adopt Euler forward method for time marching [19]. This flavor has49
been pushed further with the development of strong-stability-preserving (or total-50
variation-diminishing) high order time discretizations [25, 26, 14, 15], especially the51
RK methods in this family. The strong-stability-preserving RK (SSPRK) methods52
can be formulated as convex combinations of Euler forward steps, hence automati-53
cally preserve many properties that are achieved by the Euler forward discretization.54
One of the successes is to use them for integrating method of lines scheme obtained55
from discontinuous Galerkin (DG) spatial discretizations for hyperbolic conservation56
laws. The method is referred to as the RKDG method, which was developed by57
Cockburn et al in a series of papers [9, 8, 7, 6, 10]. In the RKDG method, a lim-58
iter can be applied after each Euler forward stage to control total variation. The59
non-increasing total variation semi-norm is then preserved after a full time step with60
SSPRK methods. Another application of SSPRK methods is for designing high order61
positivity-preserving or bound-preserving schemes [43, 45]. The approach, which is62
based on a similar methodology, has received intensive attentions in recent years and63
has been successfully applied to various problems [44, 40, 39, 22, 30, 31]. It should64
also be noted, although progress has been made on preserving positivity with the Eu-65
ler backward method [21], due to the nonexistence of second or higher order methods66
as the positive combinations of Euler backward steps [15], it is a nontrivial task to67
design high order implicit positivity-preserving schemes.68
Along with the popularity of explicit RK methods for hyperbolic problems, a69
growing attention has been paid to the role that the time integrator has played in70
a fully discrete scheme. One of the major issues is on the stability: whether the L271
stability achieved by method of lines schemes would be preserved after the explicit72
time stepping. Indeed, this issue has been raised during the initial development of the73
RKDG method. Even for linear advection equation, it is reported in [3] that under74
the usual CFL condition, the Euler forward time stepping coupled with the linear75
DG method is unstable, even in the sense of the weaker stability. The second order76
RK methods are stable only if the spatial discretization uses at most piecewise linear77
elements [11]. From then on, analyses particularly on fully discrete DG methods78
have been performed, such as [42, 41, 36, 37, 32, 46]. But it seems that an universal79
approach of analyzing the stability is missing.80
STRONG STABILITY OF RK TIME DISCRETIZATIONS 3
One of the attempts is to perform eigenvalue analysis. If the matrix L is normal,81
then it can be diagonalized with an orthogonal matrix. Hence the ODE system is82
transformed into decoupled scalar equations. The method is strongly stable if and83
only if the amplifier of each decoupled equation falls within the absolute stability84
region of the time integrator. However, when L is not normal, which is generic for85
L obtained from method of lines schemes, eigenvalue analysis gives only necessary86
but not sufficient conditions on strong stability. In such situation, eigenvalue analysis87
should be avoided for the following reasons. Firstly, if L can be diagonalized, the88
analysis ensures the strong stability of the transformed system. But in practice, the89
diagonalizing matrix can be ill-conditioned, the resulting stability would have very90
weak controls on the actual solution after the backward transformation [20]. Secondly,91
to ensure stability, we need the matrix norm of the amplifier to be properly controlled,92
but the eigenvalue analysis only examines the spectral radius, which is strictly smaller93
than the matrix norm for non-normal cases. Hence the actual time step constraint94
should be stricter than that from the eigenvalue analysis. The result solely relying on95
eigenvalue analysis can be misleading. We refer to [17] for a particular example.96
A more sound way of analyzing stability is to use the energy method. The study97
of (1.1) through this approach initiates from problems with coercive matrices L>H +98
HL ≤ −ηL>HL, where η is a positive constant. In [20], Levy and Tadmor proved that99
for coercive matrices, the classic third order and fourth order explicit RK methods100
are strongly stable, under the time step constraint τ ≤ λη with λ = 350 for the third101
order method and λ = 162 for the fourth order method. Later, a simpler proof was102
discovered based on the strong-stability-preserving RK formulation and the fact that103
Euler forward method is strongly stable for the coercive problems [15]. This approach104
gives a significantly relaxed time step τ ≤ η and extends the result to linear RK105
methods of arbitrary order. These analyses coincide with the earlier contractivity106
analysis in [28] and [18]. In their study of contractivity, or strong stability in our107
context, a circle condition is assumed, which is essentially equivalent to the strong108
stability assumption for the Euler forward steps.109
The coercivity condition typically arises from diffusive problems, but is uncommon110
in numerical approximations of hyperbolic problems. This motivates us to remove111
this assumption and consider the general semi-negative case. The difficulty is that112
the operator becomes less dissipative and the Euler forward step is no longer strongly113
stable. In [34], Tadmor proved the third order RK method is strongly stable for semi-114
negative L. Then in [33], we found a counter example showing that the fourth order115
RK method can not preserve the strong stability, although the fourth order method116
satisfies the necessary condition in the eigenvalue analysis and is hence strongly stable117
if L is normal. Furthermore, in the same paper, the fourth order RK method is proven118
to be strongly stable in two steps. In other words, successively applying the method119
for two steps yields a strongly stable method with eight stages. Recently in [23], it is120
shown that the low storage SSPRK method of order four with ten stages also admits121
strong stability.122
As one can see from [33] and [23], although the computation for analyzing RK123
methods with many stages looks complicated at the first glance, it only involves124
elementary algebraic manipulations. A single technique, which corresponds to inte-125
gration by parts for the PDE, has been repeatedly used. Inspired by the proofs in126
[33] and [23], as well as many previous analyses on DG methods, we develop a unified127
framework on analyzing the strong stability of explicit RK methods. The main idea128
is based on an induction procedure. With the aid of a computer, we can easily obtain129
the energy equality, which equates the energy at the next time steps with that at the130
4 Z. SUN AND C.-W. SHU
current time level plus terms of specific forms. Then a sufficient condition, generalized131
from Lemma 2.4 in [33], is used to justify the strong stability. A necessary condition132
is also provided to exclude some methods that are not strongly stable. With this133
framework, we can easily examine strong stability of various RK methods, including134
a ninth order and eighth order embedded pair with sixteen stages. Finally, base on135
numerous observations, we summarize patterns in the energy equality to further char-136
acterize the strongly stable RK methods. In particular, we give a both necessary and137
sufficient condition for the RK methods of odd linear order to be strongly stable. We138
remark that the framework has its limitation, for example, it can not determine if the139
classic fourth order method is strongly stable.140
The rest of the paper is organized as follows. In section 2, we present our frame-141
work on analyzing the strong stability of explicit RK methods of any order with142
arbitrary stages. In section 3, the strong stability of various methods has been ex-143
amined using the framework, including linear RK methods, the classic fourth order144
methods (strong stability in multiple steps), several SSPRK methods and the em-145
bedded pairs used in the commercial software Mathematica. Then in section 4, we146
further investigate the energy equality to characterize the structure of strongly stable147
methods. Numerical dissipation of these methods for energy conserving systems has148
also been discussed. Finally in section 5, conclusions are given.149
2. Stability analysis: a framework. Let us drop the superscript n in un.150
Consider an explicit RK time discretization for the linear autonomous system (1.1).151
The scheme is of the form152
(2.1) un+1 = Rsu,153
where154
(2.2) Rs =
s∑k=0
αs(τL)k, α0 = 1, αs 6= 0.155
Here τ is the time step and s is the number of stages. The coefficients {αk}sk=0 depend156
solely on the scheme itself. For an s-stage method, it is of linear order p if and only if157
the first p+ 1 terms in the summation (2.2) coincide with the truncated Taylor series158
of eτL. In particular, p ≤ s [2]. We would like to examine the strong stability of (2.1)159
under the usual CFL condition: if there exists a constant λ, such that160
(2.3) ‖Rsu‖2H ≤ ‖u‖2H ,161
for all τ‖L‖H ≤ λ and all inputs u. This is equivalent to162
(2.4) ‖Rs‖H ≤ 1,163
under the prescribed condition and ‖Rs‖H is the matrix norm of Rs.164
A natural attempt is to adopt the following expansion to compare ‖Rsu‖2H with165
‖u‖2H .166
(2.5)
‖Rsu‖2H =
s∑i,j=0
αiαjτi+j〈Liu, Lju〉H = ‖u‖2H +
∑1≤max{i,j}≤s
αiαjτi+j〈Liu, Lju〉H .167
However, each term 〈Liu, Lju〉H may not necessarily have a sign. The idea for over-168
coming the difficulty is to convert 〈Liu, Lju〉H into linear combinations of terms of169
STRONG STABILITY OF RK TIME DISCRETIZATIONS 5
the form ‖Lku‖2H , JLkuK2H and [Liu, Lju]H . Here170
(2.6) [v, w]H = −〈v, (L>H +HL)w〉, v, w ∈ RN171
is a semi inner product and172
(2.7) JvKH =√
[v, v]H173
defines the induced semi-norm. Indeed, this can be achieved through the following174
induction procedure.175
Proposition 2.1 (Integration by parts). Suppose j ≥ i, then176
(2.8) 〈Liu, Lju〉H =
‖Liu‖2H , j = i,− 1
2JLiuK2H , j = i+ 1,−〈Li+1u, Lj−1u〉H − [Liu, Lj−1u]H , otherwise.
177
Proof. The case j = i can be justified with the definition of ‖·‖H . When j = i+1,178
(2.9)〈Liu, Li+1u〉H =
1
2〈Liu, (HL)Liu〉+
1
2〈L>HLiu, Liu〉
=1
2〈Liu, (L>H +HL)Liu〉 = −1
2JLiuK2H .
179
When j > i+ 1,180
(2.10)
〈Liu, Lju〉H = 〈Liu, (HL)Lj−1u〉= −〈Liu, L>HLj−1u〉+ 〈Liu, (L>H +HL)Lj−1u〉= −〈Li+1u, Lj−1u〉H − [Liu, Lj−1u]H .
181
In the context of approximating the spatial derivative ∂x for periodic functions182
with L, Proposition 2.1 is the discrete version of integration by parts. Since L may183
not preserve the exact anti-symmetry of ∂x, namely L>H + HL 6= 0, an extra term184
[Liu, Lj−1u]H is produced. Furthermore, − 12JLiuK2H is usually the numerical dissipa-185
tion from the spatial discretization. In particular, J·KH is the jump semi-norm in the186
DG method.187
Furthermore, one can repeat the induction procedure to obtain the following188
expansion.189
Corollary 2.2. For j ≥ i,190
(2.11) 〈Liu, Lju〉H = ζi,j −b j−i
2 c−1∑k=0
(−1)k[Li+ku, Lj−1−ku]H ,191
where192
(2.12) ζi,j =
{(−1)
j−i+12
12JL
i+j−12 uK2H , i+ j odd,
(−1)j−i2 ‖L
i+j2 u‖2H , i+ j even.
193
Based on Corollary 2.2, we have the following energy equality.194
Lemma 2.3 (Energy equality). Given H and Rs =∑sk=0 αk(τL)k with α0 = 1.195
There exists a unique set of coefficients {βk}sk=0 ∪ {γi,j}s−1i,j=0, such that for all u and196
L satisfying L>H +HL ≤ 0,197
(2.13) ‖Rsu‖2H =
s∑k=0
βkτ2k‖Lku‖2H +
s−1∑i,j=0
γi,jτi+j+1[Liu, Lju]H , γi,j = γj,i.198
6 Z. SUN AND C.-W. SHU
Proof. The existence of such expansion can be justified by (2.5) and Corollary 2.2.199
It remains to show that the set of coefficients is unique. Each term in the summation200
should be considered as a polynomial of τ , elements in L and elements in u. We are201
going to show these polynomials are linearly independent.202
(i) It suffices to analyze the case H being the identity matrix, otherwise we can203
consider L =√HL√H−1
and u =√Hu instead.204
(ii) It suffices to show {τ i+j〈Liu, Lju〉}0≤i≤j≤s are linearly independent. Since205
elements in the set can be expressed as linear combinations of {τ2k‖Lku‖2}sk=0 ∪206
{τ i+j [Liu, Lju]}0≤i≤j≤s−1 due to Corollary 2.2. Since the two sets have the same207
cardinality, the linear independence of the previous set implies that of the latter one.208
Particularly, we take L =
(cos θ − sin θsin θ cos θ
)with θ ∈ (π2 ,
3π2 ) (hence L> + L ≤ 0)209
and u =
(10
). Noting that L is a rotation matrix, which is orthogonal, one can obtain210
(2.14)
0 =∑
0≤i≤j≤s
αi,jτi+j〈Liu, Lju〉 =
2s∑m=0
bm2 c∑k=max{0,m−s}
αk,m−k〈Lku, Lm−ku〉
τm
=
2s∑m=0
bm2 c∑k=max{0,m−s}
αk,m−k cos((m− 2k)θ)τm.
211
Note {m − 2k}bm2 c
k=max{0,m−s} are distinct non-negative integers. Due to linear inde-212
pendence of cos((m−2k)θ)τm, {αk,m−k}bm2 ck=max{0,m−s} are all zeros for each m. Hence213
{τ i+j〈Liu, Lju〉}0≤i≤j≤s are linearly independent.214
Remark 2.4. The uniqueness is not used in the framework. But it facilitates our215
analysis in section 4. Note that the uniqueness is nontrivial. For example, if we216
restrict ourselves to a small subset {L : L>H + HL = −ηL>HL, 0 > η ∈ R}, then217
one can certainly obtain different linear combinations.218
To facilitate our discussion, we introduce the following definitions.219
Definition 2.5. The leading index of Rs, denoted as k∗, is the positive integer220
such that βk∗ 6= 0 and βk < 0 for all 1 ≤ k < k∗. The coefficient βk∗ is called the221
leading coefficient. The k∗-th order principal submatrix Γ∗ = (γi,j)0≤i,j≤k∗−1 is called222
the leading submatrix.223
Note that k∗ is well-defined since βs = α2s 6= 0, which implies k∗ ≤ s. For small224
τ‖L‖H , βk∗τ2k∗‖Lku‖2H and
∑k∗−1i,j≥0 γi,jτ
i+j+1[Liu, Lju]H become dominant terms in225
the energy equality. Hence the strong stability would be closely related with the226
negativity of βk∗ and Γ∗. In particular, we have the following necessary condition and227
sufficient condition.228
Theorem 2.6 (Necessary condition). The method is not strongly stable if βk∗ >229
0. More specifically, if βk∗ > 0, then there exists a constant λ, such that ‖Rs‖H > 1230
if 0 < τ‖L‖H ≤ λ and L>H +HL = 0.231
Proof. With L>H+HL = 0, the latter summation of terms [Liu, Lju]H in (2.13)232
STRONG STABILITY OF RK TIME DISCRETIZATIONS 7
is zero. Hence233
(2.15)
‖Rsu‖2H = ‖u‖2H + βk∗τ2k∗‖Lk
∗u‖2H +
s∑k=k∗+1
βkτ2k‖Lku‖2H
≥ ‖u‖2H +
(βk∗ −
s∑k=k∗+1
|βk|(τ‖L‖H)2(k−k∗)
)τ2k
∗‖Lk
∗u‖2H .
234
Therefore,235
(2.16) ‖Rs‖2H ≥ 1 + (βk∗ − β)τ2k‖Lk∗‖2H > 1,236
if β =∑sk=k∗+1 |βk|λ2(k−k
∗) < βk∗ .237
Theorem 2.7 (Sufficient condition). If βk∗ < 0 and Γ∗ is negative definite, then238
there exists a constant λ such that ‖Rs‖H ≤ 1 if τ‖L‖H ≤ λ.239
Proof. Let −ε to be the largest eigenvalue of Γ∗. Then Γ∗ + εI is negative semi-240
definite. From Lemma 2.3 in [33],∑k∗−1i,j=0(γi,j + εδi,j)τ
i+j+1[Liu, Lju]H ≤ 0, where241
δi,j is the Kronecker delta function. Hence242
(2.17)
‖Rsu‖2H ≤‖u‖2H + βk∗τ2k∗‖Lk
∗u‖2H +
s∑k=k∗+1
βkτ2k‖Lku‖2H
− εk∗−1∑k=0
τ2k+1JLkuK2H +∑
k∗≤max{i,j}≤s−1
γi,jτi+j+1[Liu, Lju]H .
243
Note that ‖Lku‖H ≤ ‖L‖k−k∗
H ‖Lk∗u‖H and τ‖L‖H ≤ λ. Hence we have244
(2.18)
s∑k=k∗+1
βkτ2k‖Lku‖2H ≤
(s∑
k=k∗+1
|βk|λ2(k−k∗)
)τ2k
∗‖Lk
∗u‖2H .245
Using the fact246
(2.19)
[Liu, Lju]H ≤ JLiuKHJLjuKH , τJLjuK2H ≤ 2λ‖Lju‖2H ≤ 2λ‖L‖2(j−k∗)
H ‖Lk∗u‖2H ,247
together with the arithmetic-geometric mean inequality, one can obtain248
(2.20)
τ i+j+1[Liu, Lju]H ≤
{ετ2i+1
2βJLiuK2H + βλ2(j−k∗)+1
ε τ2k∗‖Lk∗u‖2H , i < k∗, j ≥ k∗,
2λi+j+1−2k∗τ2k∗‖Lk∗u‖2H , i, j ≥ k∗.
249
Therefore,250
(2.21)
‖Rsu‖2H ≤‖u‖2H +
βk∗ +
s∑k=k∗+1
|βk|λ2(k−k∗) + 2
s∑i,j≥k∗
|γi,j |λi+j+1−2k∗
+2β
ε
s∑j=k∗
k∗−1∑i=1
|γi,j |λ2(j−k∗)+1
τ2k∗‖Lk
∗u‖2H
− εk∗−1∑k=0
(1−
∑sj=k∗ |γk,j |β
)τ2k+1JLkuK2H .
251
8 Z. SUN AND C.-W. SHU
It suffices to take β = max{∑sj=k∗ |γk,j |}
k∗−1k=0 and then choose λ sufficiently small so252
that the second coefficient on the right is negative.253
Given an s-stage RK scheme Rs, we expand ‖Rsu‖2H with Lemma 2.3 and then254
use the necessary condition in Theorem 2.6 and sufficient condition in Theorem 2.7 to255
examine its strong stability. Note that {βk}sk=0 and {γi,j}s−1i,j=0 can be obtained from256
Algorithm 2.1, which is based on Proposition 2.1. Since Γ∗ is symmetric, one only257
needs to check its largest eigenvalue to determine if Γ∗ is negative definite.258
Algorithm 2.1 Obtain coefficients in Lemma 2.3
Input: β = (β0, · · · , βs) = 0, Γ = (γi,j)s−1i,j=0 = 0
for i← 0 to s doβi ← βi + α2
i
for j ← i+ 1 to s doα← 2αiαjk ← il← jwhile l > k do
switch l docase k do
βk ← βk + αendcase k + 1 do
γk,k ← γk,k − α/2endotherwise do
γk,l−1 ← γk,l−1 − αα← −α
end
endk ← k + 1l← l − 1
end
end
end
3. Applications. In this section, we examine strong stability of various RK259
methods using the framework in the previous section. In tables that will be provided260
later, the last column “SS” refers to the strong stability property. A question mark261
will be put into the entry if strong stability of the corresponding scheme can not be262
determined. “no*” means particular counter examples can be constructed.263
3.1. Linear RK methods. For general nonlinear systems, to admit accuracy264
order higher than four, RK methods must have more stages than its order [2]. How-265
ever, for autonomous linear systems, the desired order of accuracy can be achieved266
with the same number of stages. All such methods would be equivalent to the Taylor267
series method268
(3.1) Rp = Pp =
p∑k=0
(τL)k
k!.269
STRONG STABILITY OF RK TIME DISCRETIZATIONS 9
In Table 1 and Table 2, we document the leading indexes and coefficients of linear270
RK methods from first order to twelfth order. The leading submatrices for low order271
RK methods are also given in Table 1. We report that the third order, seventh order272
and eleventh order methods are strongly stable. The fourth order, eighth order and273
twelfth order methods can not be judged with the framework. All other methods are274
not strongly stable.
p k∗ βk∗ Γ∗ λ(Γ∗) SS
1 1 1 −(
1)
−1.00000 no
2 2 14
−(
1 12
12
12
)−1.30902−1.90983 × 10−1 no
3 2 − 112
−(
1 12
12
13
)−1.26759−6.57415 × 10−2 yes
4 3 − 172
−
1 12
16
12
13
18
16
18
124
−1.30128−7.93266 × 10−2
+5.60618 × 10−3no*
5 3 1360
−
1 12
16
12
13
18
16
18
120
−1.30150−8.07336 × 10−2
−1.10151 × 10−3no
6 4 12880
−
1 1
216
124
12
13
18
130
16
18
120
172
124
130
172
1240
−1.30375−8.21871 × 10−2
−1.40529 × 10−3
−1.60133 × 10−4
no
7 4 − 120160
−
1 1
216
124
12
13
18
130
16
18
120
172
124
130
172
1252
−1.30375−8.21836 × 10−2
−1.36301 × 10−3
−7.86229 × 10−6
yes
8 5 − 1201600
−
1 1
216
124
1120
12
13
18
130
1144
16
18
120
172
1336
124
130
172
1252
11152
1120
1144
1336
11152
23120960
−1.30384−8.22588 × 10−2
−1.38580 × 10−3
−9.32706 × 10−6
+2.24989 × 10−6
?
Table 1Linear RK methods: from first order to eighth order.
275
3.2. The classic fourth order method. The classic fourth order method with276
four stages, which is widely used in practice due to its stage and order optimality, is277
unfortunately not covered under the framework. In [33], we found a counter example278
to show that the method is not strongly stable, but successively applying the method279
for two steps yields a strongly stable method with eight stages.280
Proposition 3.1 (Sun and Shu, 2018). The fourth order RK method with four281
stages is not strongly stable. More specifically, for H = I and L = −
1 2 20 1 20 0 1
,282
we have ‖P4‖ > 1, if τ‖L‖H > 0 is sufficiently small.283
Theorem 3.2 (Sun and Shu, 2018). The fourth order RK method with four stages284
is strongly stable in two steps. In other words, there exists a constant λ, such that285
‖(P4)2‖H ≤ 1 if τ‖L‖H ≤ λ.286
Here we examine multi-step strong stability of the fourth order method using our287
framework. Note the derivation using this framework is slightly different from that288
10 Z. SUN AND C.-W. SHU
p k∗ βk∗ λ(Γ∗) SS
9 5 11814400
−1.30384−8.22588 × 10−2
−1.38585 × 10−3
−9.75366 × 10−6
−3.11800 × 10−8
no
10 6 1221772800
−1.30384−8.22613 × 10−2
−1.38688 × 10−3
−9.91006 × 10−6
−4.70638 × 10−8
−1.63872 × 10−8
no
11 6 − 1239500800
−1.30384−8.22613 × 10−2
−1.38688 × 10−3
−9.90966 × 10−6
−3.87351 × 10−8
−7.87018 × 10−11
yes
12 7 − 13353011200
−1.30384−8.22614 × 10−2
−1.38691 × 10−3
−9.91617 × 10−6
−3.93334 × 10−8
+1.45458 × 10−10
−8.54170 × 10−11
?
Table 2Linear RK methods: from ninth order to twelfth order.
in [33]. The relevant quantities for strong stability are given in Table 3. Note that289
the method is both two-step and three-step strongly stable (with the same time step290
size), which means the norm of the solution after the first step is always bounded by291
the initial data, if sufficiently small uniform time steps are used.
(P4)m k∗ βk∗ Γ∗ λ(Γ∗) SS
(P4)2 3 − 136
−
2 2 43
2 83
243
2 1912
−5.73797−4.99093 × 10−1
−1.29329 × 10−2yes
(P4)3 3 − 124
−
3 92
92
92
9 818
92
818
978
−2.28380 × 101
−1.21069−7.62892 × 10−2
yes
Table 3The classic fourth order method: multi-step strong stability.
292
Theorem 3.3. The four-stage fourth order RK method has the following property.293
With uniform time steps such that τ‖L‖H ≤ λ for sufficiently small λ, ‖un‖H ≤294
‖u0‖H for all n > 1.295
3.3. SSPRK methods. In this section, we study the strong stability of several296
SSPRK methods. The (explicit) SSPRK methods are a class of RK methods that can297
be formulated as combinations of Euler forward steps. The second order method with298
two stages, and the third order method with three stages are equivalent to the linear299
STRONG STABILITY OF RK TIME DISCRETIZATIONS 11
RK methods for autonomous linear systems, which have been discussed. For fourth300
order methods, in order to avoid backward-in-time steps and negative coefficients,301
at least five stages should be used [14, 18], which are denoted as SSPRK(5,4). We302
specially consider the method that is independently discovered in [18] and [29]. When303
applied on (1.1), the method takes the form304
(3.2) SSPRK(5,4) = P4 + 4.477718303076007× 10−3(τL)5.305
We will study its strong stability (in multiple steps) using our framework. Besides306
SSPRK(5,4), we will also consider two commonly used low storage SSPRK methods,307
a third order method with four stages SSPRK(4,3) and a fourth order method with308
ten stages SSPRK(10,4) [18, 29]. The two methods are309
(3.3) SSPRK(4,3) = P3 +1
48(τL)4,310
and311
(3.4)SSPRK(10,4) =P4 +
17
2160(τL)5 +
7
6480(τL)6 +
1
9720(τL)7
+1
155520(τL)8 +
1
4199040(τL)9 +
1
251942400(τL)10,
312
when applied on (1.1). We remark that the strong stability of SSPRK(10,4) has been313
proved by Ranocha and Offner in [23]. This is a reexamination using our framework.314
From Table 4, we are able to conclude the follow results.
SSPRK k∗ βk∗ Γ∗ λ(Γ∗) SS
(4,3) 2 − 124
−(
1 12
12
13
)−1.26759−6.57415 × 10−2 yes
(10,4) 3 − 13240
−
1 12
16
12
13
18
16
18
1072160
−1.30149−8.06493 × 10−2
−7.35115 × 10−4yes
(5,4) 3 −4.93345 × 10−3 −
1 12
16
12
13
18
16
18
124
−1.30140−8.00541 × 10−2
+1.97309 × 10−3no*
(5,4)2 3 −9.86690 × 10−3 −
2 2 43
2 83
243
2 1.5923
−5.74021−5.01739 × 10−1
−1.70056 × 10−2yes
(5,4)3 3 −1.48004 × 10−2 −
3 92
92
92
9 818
92
818
12.138
−2.28450 × 101
−1.21415−7.93174 × 10−2
yes
Table 4SSPRK methods: strong stability and multi-step strong stability.
315
Theorem 3.4. SSPRK(4,3) and SSPRK(10,4) are strongly stable.316
Theorem 3.5. The property stated in Theorem 3.3 also holds for SSPRK(5,4).317
The behavior of SSPRK(5,4) is very similar to that of the classic fourth order318
method, since it is almost the four-stage method except for a small fifth order per-319
turbation. Although the method can not be judged within this framework, one can320
indeed use the same counter example in Proposition 3.1 to disprove its strong stability.321
The proof would be similar to that in [33].322
12 Z. SUN AND C.-W. SHU
3.4. Embedded RK methods in NDSolve. Finally, we consider embedded323
RK pairs that are used in NDSolve, a function for numerically solving differentiable324
equations in the commercial software Mathematica. Embedded RK methods are pairs325
of RK methods sharing the same stages. The notation p(p) is commonly used, if two326
methods in the pair are of order p and order p respectively. Their Butcher tableau327
has the following form.328
(3.5)
0c2 a2,1c3 a3,1 a3,2...
......
. . .
cs as,1 as,2 · · · as,s−1b1 b2 · · · bs−1 bsb1 b2 · · · bs−1 bs
329
For ddtu = f(t, u), the tableau gives two solutions330
(3.6) un+1 = un + τ
s∑i=1
biki, un+1 = un + τ
s∑i=1
biki,331
where332
(3.7) ki = f(t+ ciτ, un + τ
s∑j=1
ai,jkj).333
Then the difference un+1 − un+1 can be used for local error estimates for time step334
adaption.335
We examine strong stability of all such pairs used in Mathematica from order336
2(1) to order 9(8). These methods are chosen with several desired properties being337
considered, including the FSAL (First Same As Last) strategy and stiffness detection338
capability [38]. Tableaux of 2(1), 3(2) and 4(3) pairs [27] are given in Table 5, Table 6339
and Table 7. For the 5(4) pair [1, 24] and higher order pairs [35], the tableaux can
01 11 1
212
12
12 0
1 − 16
16
Table 5Tableau of embedded RK 2(1).
340be obtained through the Mathematica command341
(3.8) NDSolve`EmbeddedExplicitRungeKuttaCoefficients[p, Infinity].342
The output takes the form343
(3.9) {A, b, c, b− b}344
STRONG STABILITY OF RK TIME DISCRETIZATIONS 13
012
12
1 -1 21 1
623
16
16
23
16 0
22−√82
72
√82+1436
√82−4144
16−√82
48
Table 6Tableau of embedded RK 3(2).
025
25
35 − 3
2034
1 1944 − 15
441011
1 1172
2572
2572
1172
1172
2572
2572
1172 0
12515158970912
37101058970912
25196958970912
611058970912
119041747576
Table 7Tableau of embedded RK 4(3).
and the corresponding tableau is345
(3.10)
c Ab
b
.346
The stability results are documented in Table 8 and Table 9.347
4. Characterization of strongly stable methods. By numerically examining348
various RK methods, we recognize certain patterns of the coefficients in Lemma 2.3,349
which will be proved in this section.350
To characterize the coefficients, we would like to split the RK operator as a351
truncated exponential and a high order perturbation.352
(4.1) Rs = Pp + (τL)p+1Qs−(p+1),353
where354
(4.2) Pp =
p∑k=0
(τL)k
k!, Qs−(p+1) =
s−(p+1)∑k=0
αk+p+1(τL)k, αp+1 6=1
(p+ 1)!.355
Note that an RK method of order p for general nonlinear system may achieve higher356
order accuracy when applied to linear autonomous problems. Without special clarifi-357
cation, the order p in this section refers to the linear order.358
4.1. Coefficients in the energy equality. Coefficients in Lemma 2.3 have the359
following pattern.360
Lemma 4.1. For an s-stage RK method of linear order p, β0 = 1. Furthermore,361
(i) if p is odd, then362
(4.3) k∗ =p+ 1
2, βk∗ = (−1)k
∗2
(αp+1 −
1
(p+ 1)!
),363
14 Z. SUN AND C.-W. SHU
Methods s p/p k∗ βk∗ λ(Γ∗) SS
2(1) 3 2 2 14
−1.30902−1.90983 × 10−1 no
1 1 1 −1.00000 no
3(2) 4 3 2 − 112
−1.26759−6.57415 × 10−2 yes
2 2 112
−1.28130−1.11257 × 10−1 no
4(3) 5 4 3 − 172
−1.30128−7.93266 × 10−2
+5.60618 × 10−3no*
3 2 − 1190414485456
−1.26759−6.57415 × 10−2 yes
5(4) 8 5 3 − 436209280
−1.3015−8.07336 × 10−2
−1.10151 × 10−3yes
4 3 51767367590960
−1.30150−8.07430 × 10−2
−1.14174 × 10−3no
6(5) 9 6 4 790072560896000
−1.30375−8.21839 × 10−2
−1.36689 × 10−3
−2.38718 × 10−5
no
5 3 12334679027158400
−1.30150−8.07336 × 10−2
−1.10151 × 10−3no
7(6) 10 7 4 2961560506338967665360400000
−1.30375−8.21836 × 10−2
−1.36301 × 10−3
−7.86229 × 10−6
no
6 4 − 202029199011855603112400000
−1.30375−8.21833 × 10−2
−1.35985 × 10−3
+5.49402 × 10−6
?
*The fourth order method in the 4(3) pair is exactly the classic four-stage fourth order method for
autonomous linear systems.Table 8
Embedded RK pairs: from 2(1) to 7(6) pairs.
364
(4.4) γi,j = − 1
i!j!(i+ j + 1), ∀0 ≤ i, j ≤ k∗ − 1;365
(ii) if p is even, then366
367
(4.5) k∗ ≥ p
2+ 1, β p
2+1 = (−1)p2+12
(αp+2 − αp+1 +
1
p!(p+ 2)
),368
369
(4.6) γi,j = − 1
i!j!(i+ j + 1)+ ιi,j,p ∀0 ≤ i, j ≤ p
2,370
STRONG STABILITY OF RK TIME DISCRETIZATIONS 15
Methods s p k∗ βk∗ λ(Γ∗) SS
8(7) 13 8 5 −3.21308 × 10−7
−1.30384−8.22588 × 10−2
−1.38584 × 10−3
−9.71236 × 10−6
+1.43671 × 10−7
?
7 4 −2.39706 × 10−6
−1.30375−8.21836 × 10−2
−1.36301 × 10−3
−7.86229 × 10−6
yes
9(8) 16 9 5 −8.95352 × 10−9
−1.30384−8.22588 × 10−2
−1.38585 × 10−3
−9.75366 × 10−6
−3.11800 × 10−8
yes
8 5 −5.46447 × 10−7
−1.30384−8.22588 × 10−2
−1.38585 × 10−3
−9.78641 × 10−6
−1.64476 × 10−7
yes
Table 9Embedded RK pairs: 8(7) pair and 9(8) pair.
where371
(4.7) ιi,j,p =
{(−1)
p2+1
(αp+1 − 1
(p+1)!
), i = j = p
2 ,
0, otherwise.372
The proof of this lemma is postponed to the end of the section.373
4.2. Criteria for strong stability. With Lemma 4.1, one can obtain the fol-374
lowing theorem regarding strong stability of linear RK methods with Rs = Pp. Note375
that the case p ≡ 0 (mod 4) is not covered. The difficulty for analyzing this class of376
methods has already been recognized in [33].377
Theorem 4.2. Consider a linear RK method of order p with p stages.378
(i) The method is not strongly stable if p ≡ 1 (mod 4) or p ≡ 2 (mod 4).379
(ii) The method is strongly stable if p ≡ 3 (mod 4).380
Proof. Note αp+1 = αp+2 = 0. βk∗ > 0 if p ≡ 1 (mod 4) or p ≡ 2 (mod 4). The381
method can not preserve strong stability due to Theorem 2.6. On the other hand,382
for p ≡ 3 (mod 4), the leading submatrix Γ∗ can be written as −ΛMΛ, where Λ =383
diag(1/0!, 1/1!, 1/2!, · · · , 1/(k∗ − 1)!) and M is the Hilbert matrix of order k∗. Γ∗ is384
negative definite since the Hilbert matrix is positive definite. Also note βk∗ = − 2(p+1)!385
in such cases. The strong stability can be proved using Theorem 2.7.386
Remark 4.3. This modulo pattern with periodicity 4 is closely related with the387
fact that i4 = 1, in which i is the imaginary unit. One can get a flavor by considering388
the scalar ODE ddtu = (iω)u, ω ∈ R.389
For a method with non-zero Qs−(p+1), noting that only limited number of stages390
would affect the leading coefficient and submatrix, one can conclude the following391
criteria for strong stability. We highlight that the condition for methods of odd linear392
order is both necessary and sufficient.393
16 Z. SUN AND C.-W. SHU
Theorem 4.4. An RK method of odd linear order p is strongly stable if and only394
if395
(4.8) (−1)p+12
(αp+1 −
1
(p+ 1)!
)< 0.396
Proof. Since Γ∗ = −ΛMΛ is always negative definite, the method is strongly397
stable if and only if βk∗ < 0, which corresponds to the prescribed condition in the398
theorem.399
For RK methods with even linear order, we can only obtain a sufficient condition,400
which is given in Theorem 4.5. A similar condition has also been discussed in [23] for401
p = 4.402
Theorem 4.5. An RK method of even linear order p is strongly stable if403
(4.9) (−1)p2+1
(αp+2 − αp+1 +
1
p!(p+ 2)
)< 0,404
and405
(4.10) (−1)p2+1
(p2
!)2(
αp+1 −1
(p+ 1)!
)< ε.406
Here ε is the smallest eigenvalue of the Hilbert matrix of order p2 + 1.407
Proof. From Lemma 4.1, (4.9) implies k∗ = p2 +1 and βk∗ < 0. Note Γ∗ is negative408
definite with (4.10). Strong stability then follows from Theorem 2.7.409
We remark that, constraints in Theorem 4.4 and Theorem 4.5 can be used with410
order conditions for designing strongly stable RK methods.411
4.3. Regarding energy conserving systems. Problems for wave propaga-412
tions are usually featured with a conserved L2 energy. This conservation is also413
expected numerically to maintain the accurate shape and phase of the waves in long414
time simulations. Suppose an energy-conserving spatial discretization is used, for ex-415
ample [5, 4, 12, 13], the resulting method of lines scheme would satisfy L>H+HL = 0.416
Hence ddt‖u‖
2H = 0. While a strongly stable RK time discretization may not preserve417
this equality, we would like know how the total energy is dissipated. The following418
interpretation can be obtained based on Lemma 2.3. While one can also perform419
eigenvalue analysis alternatively, since L is normal in 〈·, ·〉H for energy conserving420
systems.421
With L>H +HL = 0, [·, ·]H = 0. The energy equality in Lemma 2.3 would then422
become423
(4.11) ‖Rsu‖2H = ‖u‖2H +
s∑k=k∗
βkτ2k‖Lku‖2H .424
Since βk∗ < 0 for strongly stable RK methods, we have dissipative energy unless425
L = 0.426
Proposition 4.6. With a suitably restricted time step, a strongly stable explicit427
RK method is energy conserving, if and only if L = 0.428
At the final time T = nτ , ‖un‖2H = ‖u0‖2H + O(τ2k∗−1). The total numerical429
dissipation due to the time integrator would then be of order 2k∗ − 1. We refer to430
2k∗ − 1 as the energy accuracy of the RK time integrator. With k∗ determined in431
Lemma 4.1, we have the following proposition.432
STRONG STABILITY OF RK TIME DISCRETIZATIONS 17
Proposition 4.7. Consider an RK method of linear order p applied on an energy433
conserving system.434
(i) The order of energy accuracy is p if p is odd.435
(ii) The order of energy accuracy is at least p + 1 if p is even. It achieves higher436
energy accuracy if and only if αp+2 = αp+1 − 1p!(p+2) .437
One can see from Proposition 4.7, the linearly even order RK methods achieve438
at least one degree higher order of energy accuracy than we usually expect. Hence439
compared with the odd order methods, they may be more suitable for integrating sys-440
tems modeling wave propagations. This also reveals the fact that even order methods441
introduce less numerical dissipation, and explains why it is harder to achieve strong442
stability.443
4.4. Proof of Lemma 4.1. We now prove Lemma 4.1. Instead of using the444
mathematical induction, we provide a more motivated proof below. Let Es(τ) =445
‖Rs(τ)‖2H . The idea is to use dm
dτm E(0) to determine coefficients in the energy equality.446
For clearness of the presentation, we first consider the case Qs−(p+1) = 0 and then447
move on to general RK methods. As a convention, matrices and coefficients with448
negative or fractional indexes are considered as 0.449
Proof. Step 1: (Rs = Pp.)450
It is easy to check β0 = α20 = 1. For m > 0, using the fact that451
(4.12)dm
dτm〈v, w〉H =
m∑k=0
(m
k
)〈 d
k
dτkv,
dm−k
dτm−kw〉H452
and453
(4.13)d
dτPp = Pp−1L,454
we have455
(4.14)dm
dτmEp(τ) =
m∑k=0
(m
k
)〈Pp−kLku, Pp−(m−k)Lm−ku〉H .456
Noting that Pp−k(0) = I if k ≤ p, for 1 ≤ m ≤ p+ 2, one can obtain457
(4.15)dm
dτmEp(0) = −2µ+
m∑k=0
(m
k
)〈Lku, Lm−ku〉H ,458
where459
(4.16) µ =
0, 1 ≤ m ≤ p,〈u, Lp+1u〉H , m = p+ 1,
〈u, Lp+2u〉H + (p+ 2)〈Lu,Lp+1u〉H , m = p+ 2.460
18 Z. SUN AND C.-W. SHU
Furthermore, since461
(4.17)
m∑k=0
(m
k
)〈Lku, Lm−ku〉H
=
m∑k=1
(m− 1
k − 1
)〈Lku, Lm−ku〉H +
m−1∑k=0
(m− 1
k
)〈Lku, Lm−ku〉H
=
m−1∑k=0
(m− 1
k
)(〈Lk+1u, Lm−(k+1)u〉H + 〈Lku, Lm−ku〉H
)=−
m−1∑k=0
(m− 1
k
)[Lku, Lm−1−ku]H ,
462
(4.15) can be written as463
(4.18)dm
dτmEp(0) = −2µ−
m−1∑k=0
(m− 1
k
)[Lku, Lm−1−ku]H , 1 ≤ m ≤ p+ 2.464
On the other hand, by differentiating the expansion in Lemma 2.3, we obtain465
(4.19)dm
dτmEp(0) = βm
2m!‖Lm
2 u‖2H +
p−1∑k=0
γk,m−1−km![Lku, Lm−1−ku]H .466
Due to the uniqueness of the expansion, coefficients in (4.18) and (4.19) must be the467
same. With m = 1, · · · , p, one can get468
(4.20) βm2
= 0, ∀1 ≤ m ≤ p469
and470
(4.21) γi,j = − 1
(i+ j + 1)!
(i+ j
i
)= − 1
i!j!(i+ j + 1), i+ j ≤ p− 1, i, j ≥ 0.471
Case I: (p is odd.) If p is odd, then472
(4.22) βk = 0, k = 1, 2, · · · , p− 1
2.473
We need to compute dp+1
dτp+1 Ep(0) to determine β p+12
. With m = p+1, (4.18) and (4.19)474
imply the following identity.475
(4.23)
dp+1
dτp+1Ep(0) = β p+1
2(p+ 1)!‖L
p+12 u‖2H +
p−1∑k=0
γk,p−k(p+ 1)![Lku, Lp−ku]H
= −2〈u, Lp+1u〉H −p∑k=0
(p
k
)[Lku, Lp−ku]H .
476
From Corollary 2.2, we have477
(4.24) 〈u, Lp+1u〉H = (−1)p+12 ‖L
p+12 u‖2H −
p−12∑
k=0
(−1)k[Lku, Lp−ku]H .478
STRONG STABILITY OF RK TIME DISCRETIZATIONS 19
Hence (4.23) together with (4.24) gives479
(4.25) β p+12
= −(−1)p+12
2
(p+ 1)!6= 0,480
which implies k∗ = p+12 . For 0 ≤ i, j ≤ k∗−1, we have i+ j ≤ 2k∗−2 = p−1. Hence481
the leading submatrix Γ∗ is completely determined by (4.21).482
Case II: (p is even). For even p,483
(4.26) βk = 0, k = 1, 2, · · · , p2.484
Note β p2+1 appears in the expansion of dp+2
dτp+2 Ep(0). With m = p + 2, one can derive485
from (4.18) and (4.19) that486
(4.27)
dp+2
dτp+2Ep(0)
=β p2+1(p+ 2)!‖L
p2+1u‖2H +
p−1∑k=0
γk,p+1−k(p+ 2)![Lku, Lp+1−ku]H
=− 2〈u, Lp+2u〉H − 2(p+ 2)〈Lu,Lp+1u〉H −p+1∑k=0
(p+ 1
k
)[Lku, Lp+1−ku]H .
487
From Proposition 2.1, we have488
(4.28) 〈u, Lp+2u〉H = (−1)p2+1‖L
p2+1u‖2H −
p2∑
k=0
(−1)k[Lku, Lp+1−ku]H489
and490
(4.29) 〈Lu,Lp+1u〉H = (−1)p2 ‖L
p2+1u‖2H −
p2−1∑k=0
(−1)k[Lk+1u, Lp−ku]H .491
After identifying the coefficients of ‖Lp2+1u‖2H in (4.27) with (4.28) and (4.29), one492
can get493
(4.30) β p2+1 = (−1)
p2+1 2
p!(p+ 2)6= 0.494
Hence k∗ = p2 + 1.495
Except for γ p2 ,
p2, all other γi,j with i + j ≤ p − 1 are clarified in (4.21). To496
determine γ p2 ,
p2, we need to consider dp+1
dτp+1 Ep(0). Note that (4.23) holds regardless of497
the parity of p. Furthermore, with p being even,498
(4.31) 〈u, Lp+1u〉H = (−1)p2+1 1
2JL
p2 uK2H −
p2−1∑k=0
(−1)k[Lku, Lp−ku]H .499
Hence by comparing the coefficient of JLp2 uK2H in (4.23) with that in (4.31), we get500
(4.32) γ p2 ,
p2
= − 1
(p2 )!(p2 )!(p+ 1)− (−1)
p2+1 1
(p+ 1)!.501
20 Z. SUN AND C.-W. SHU
Now we have proved Lemma 4.1 for Rs = Pp.502
Step 2: Rs = Pp + (τL)p+1Qs−(p+1).503
The proof of the general case is based on the fact that not all high order stages504
will contribute to k∗ and Γ∗. Note that505
(4.33) ‖Rsu‖2H = ‖Ppu‖2H +∑
max(i,j)>p
αiαjτi+j〈Liu, Lju〉H .506
If p is odd, from Corollary 2.2, only 2αp+1τp+1〈L0u, Lp+1u〉H (since α0 = 1) in the507
second term would affect β p+12
. With this additional term being considered, instead508
of (4.23), we obtain509
(4.34)
dp+1
dτp+1Es(0) = β p+1
2(p+ 1)!‖L
p+12 u‖2H +
p−1∑k=0
γk,p−k(p+ 1)![Lku, Lp−ku]H
= (2αp+1(p+ 1)!− 2)〈u, Lp+1u〉H −p∑k=0
(p
k
)[Lku, Lp−ku]H .
510
After expanding 〈u, Lp+1u〉 with (4.24), (4.34) implies511
(4.35) β p+12
= (−1)p+12 2
(αp+1 −
1
(p+ 1)!
).512
Since αp+1 6= 1(p+1)! , we have β p+1
26= 0 and k∗ = p+1
2 . Also note that 〈Liu, Lju〉H513
can only produce [Li′u, Lj
′u]H terms with i′ + j′ = i+ j − 1. Hence if max{i, j} > p,514
〈Liu, Lju〉H can not affect values of γi′,j′ with i′ + j′ ≤ 2(k∗ − 1) = p − 1. In other515
words, the leading submatrix Γ∗ is unchanged.516
Similarly, for even p, we still have β1 = · · · = β p2
= 0 after adding extra high517
order terms, which implies k∗ ≥ p2 + 1. When computing dp+2
dτp+2 E(0) to obtain β p2+1,518
an extra term 2τp+2(αp+2〈L0u, Lp+2〉H + αp+1〈L1u, Lp+1u〉H
)should be considered.519
Then we have520
(4.36)
dp+2
dτp+2Es(0)
=β p2+1(p+ 2)!‖L
p2+1u‖2H +
p−1∑k=0
γk,p+1−k(p+ 2)![Lku, Lp+1−ku]H
=2 (αp+2(p+ 2)!− 1) 〈u, Lp+2u〉H + 2 (αp+1(p+ 2)!− (p+ 2)) 〈Lu,Lp+1u〉H
−p+1∑k=0
(p+ 1
k
)[Lku, Lp+1−ku]H .
521
After substituting (4.28) and (4.29) into (4.36), we get522
(4.37) β p2+1 = (−1)
p2+12
(αp+2 − αp+1 +
1
p!(p+ 2)
).523
As for the corresponding principle submatrix, only γ p2 ,
p2
will be changed due to524
2αp+1τp+1〈L0u, Lp+1u〉H . Then with (4.34) and (4.31), one can obtain525
(4.38) γ p2 ,
p2
= − 1
(p2 )!(p2 )!(p+ 1)+ (−1)
p2+1
(αp+1 −
1
(p+ 1)!
),526
STRONG STABILITY OF RK TIME DISCRETIZATIONS 21
which completes the proof.527
Remark 4.8 (Uniqueness in Lemma 2.3). Without showing the uniqueness in528
Lemma 2.3, one can still prove Lemma 4.1 with induction and then derive other529
results in section 4. The reason for justifying the uniqueness, is to exclude the possi-530
bility that the leading submatrices for even order RK methods are negative definite531
under another expansion, which may result in an if and only if condition, as that for532
the methods of odd linear order in Theorem 4.4.533
5. Conclusions. In this paper, we present a framework on analyzing the strong534
stability of explicit RK methods for solving semi-negative linear autonomous systems,535
which are typically obtained from stable method of lines schemes for hyperbolic prob-536
lems. With this framework, strong stability in one step or multiple steps of various537
RK methods are examined. Finally, we analyze the coefficients in the energy equality,538
based on which, corollaries regarding the criteria of strong stability are derived.539
REFERENCES540
[1] P. Bogacki and L. F. Shampine, An efficient Runge–Kutta (4, 5) pair, Computers & Math-541ematics with Applications, 32 (1996), pp. 15–28.542
[2] J. C. Butcher, Numerical Methods for Ordinary Differential Equations, John Wiley & Sons,5432016.544
[3] G. Chavent and B. Cockburn, The local projection p0 − p1-discontinuous-Galerkin finite545element method for scalar conservation laws, ESAIM: Mathematical Modelling and Nu-546merical Analysis, 23 (1989), pp. 565–592.547
[4] Y. Cheng, C.-S. Chou, F. Li, and Y. Xing, L2 stable discontinuous Galerkin methods for one-548dimensional two-way wave equations, Mathematics of Computation, 86 (2017), pp. 121–549155.550
[5] E. T. Chung and B. Engquist, Optimal discontinuous Galerkin methods for the acoustic551wave equation in higher dimensions, SIAM Journal on Numerical Analysis, 47 (2009),552pp. 3820–3848.553
[6] B. Cockburn, S. Hou, and C.-W. Shu, The Runge–Kutta local projection discontinuous554Galerkin finite element method for conservation laws. IV. the multidimensional case,555Mathematics of Computation, 54 (1990), pp. 545–581.556
[7] B. Cockburn, S.-Y. Lin, and C.-W. Shu, TVB Runge–Kutta local projection discontinu-557ous Galerkin finite element method for conservation laws III: one-dimensional systems,558Journal of Computational Physics, 84 (1989), pp. 90–113.559
[8] B. Cockburn and C.-W. Shu, TVB Runge–Kutta local projection discontinuous Galerkin560finite element method for conservation laws. II. general framework, Mathematics of com-561putation, 52 (1989), pp. 411–435.562
[9] B. Cockburn and C.-W. Shu, The Runge–Kutta local projection P 1-discontinuous-Galerkin563finite element method for scalar conservation laws, ESAIM: Mathematical Modelling and564Numerical Analysis, 25 (1991), pp. 337–361.565
[10] B. Cockburn and C.-W. Shu, The Runge–Kutta discontinuous Galerkin method for conser-566vation laws V: multidimensional systems, Journal of Computational Physics, 141 (1998),567pp. 199–224.568
[11] B. Cockburn and C.-W. Shu, Runge–Kutta discontinuous Galerkin methods for convection-569dominated problems, Journal of scientific computing, 16 (2001), pp. 173–261.570
[12] G. Fu and C.-W. Shu, Optimal energy-conserving discontinuous Galerkin methods for linear571symmetric hyperbolic systems, arXiv preprint arXiv:1804.10307, (2018).572
[13] G. Fu and C.-W. Shu, An energy-conserving ultra-weak discontinuous Galerkin method for573the generalized Korteweg–de Vries equation, Journal of Computational and Applied Math-574ematics, 349 (2019), pp. 41–51.575
[14] S. Gottlieb and C.-W. Shu, Total variation diminishing Runge–Kutta schemes, Mathematics576of Computation, 67 (1998), pp. 73–85.577
[15] S. Gottlieb, C.-W. Shu, and E. Tadmor, Strong stability-preserving high-order time dis-578cretization methods, SIAM review, 43 (2001), pp. 89–112.579
[16] B. Gustafsson, H.-O. Kreiss, and J. Oliger, Time Dependent Problems and Difference580Methods, John Wiley & Sons, 1995.581
22 Z. SUN AND C.-W. SHU
[17] A. Iserles, A First Course in the Numerical Analysis of Differential Equations, Cambridge582University Press, 2009.583
[18] J. F. B. M. Kraaijevanger, Contractivity of Runge–Kutta methods, BIT Numerical Mathe-584matics, 31 (1991), pp. 482–528.585
[19] R. J. LeVeque, Numerical Methods for Conservation Laws, Birkhauser, 1992.586[20] D. Levy and E. Tadmor, From semidiscrete to fully discrete: Stability of Runge–Kutta587
schemes by the energy method, SIAM review, 40 (1998), pp. 40–73.588[21] T. Qin and C.-W. Shu, Implicit positivity-preserving high-order discontinuous Galerkin meth-589
ods for conservation laws, SIAM Journal on Scientific Computing, 40 (2018), pp. A81–590A107.591
[22] T. Qin, C.-W. Shu, and Y. Yang, Bound-preserving discontinuous Galerkin methods for592relativistic hydrodynamics, Journal of Computational Physics, 315 (2016), pp. 323–347.593
[23] H. Ranocha and P. Offner, L2 stability of explicit Runge–Kutta schemes, Journal of Scien-594tific Computing, (2018), pp. 1–17.595
[24] L. F. Shampine, Numerical Solution of Ordinary Differential Equations, Routledge, 2018.596[25] C.-W. Shu, Total-variation-diminishing time discretizations, SIAM Journal on Scientific and597
Statistical Computing, 9 (1988), pp. 1073–1084.598[26] C.-W. Shu and S. Osher, Efficient implementation of essentially non-oscillatory shock-599
capturing schemes, Journal of computational physics, 77 (1988), pp. 439–471.600[27] M. Sofroniou and G. Spaletta, Construction of explicit Runge-Kutta pairs with stiffness601
detection, Mathematical and Computer Modelling, 40 (2004), pp. 1157–1169.602[28] M. Spijker, Contractivity in the numerical solution of initial value problems, Numerische603
Mathematik, 42 (1983), pp. 271–290.604[29] R. J. Spiteri and S. J. Ruuth, A new class of optimal high-order strong-stability-preserving605
time discretization methods, SIAM Journal on Numerical Analysis, 40 (2002), pp. 469–491.606[30] Z. Sun, J. A. Carrillo, and C.-W. Shu, A discontinuous Galerkin method for nonlinear607
parabolic equations and gradient flow problems with interaction potentials, Journal of Com-608putational Physics, 352 (2018), pp. 76–104.609
[31] Z. Sun, J. A. Carrillo, and C.-W. Shu, An entropy stable high-order discontinuous Galerkin610method for cross-diffusion gradient flow systems, arXiv preprint arXiv:1810.03221, (2018).611
[32] Z. Sun and C.-W. Shu, Stability analysis and error estimates of Lax–Wendroff discontinu-612ous Galerkin methods for linear conservation laws, ESAIM: Mathematical Modelling and613Numerical Analysis, 51 (2017), pp. 1063–1087.614
[33] Z. Sun and C.-W. Shu, Stability of the fourth order Runge–Kutta method for time-dependent615partial differential equations, Annals of Mathematical Sciences and Applications, 2 (2017),616pp. 255–284.617
[34] E. Tadmor, From semidiscrete to fully discrete: Stability of Runge–Kutta schemes by the618energy method. II, Collected Lectures on the Preservation of Stability under Discretization,619Lecture Notes from Colorado State University Conference, Fort Collins, CO, 2001 (D. Estep620and S. Tavener, eds.), Proceedings in Applied Mathematics, SIAM, 109 (2002), pp. 25–49.621
[35] J. H. Verner, Numerically optimal Runge–Kutta pairs with interpolants, Numerical Algo-622rithms, 53 (2010), pp. 383–396.623
[36] H. Wang, C.-W. Shu, and Q. Zhang, Stability and error estimates of local discontinuous624Galerkin methods with implicit-explicit time-marching for advection-diffusion problems,625SIAM Journal on Numerical Analysis, 53 (2015), pp. 206–227.626
[37] H. Wang, C.-W. Shu, and Q. Zhang, Stability analysis and error estimates of local discon-627tinuous Galerkin methods with implicit-explicit time-marching for nonlinear convection-628diffusion problems, Applied Mathematics and Computation, 272 (2016), pp. 237–258.629
[38] S. Wolfram, Advanced numerical differential equation solving in Mathematica, Wolfram630Mathematica Tutorial Collection. Wolfram Research, Champaign, (2008).631
[39] K. Wu and H. Tang, High-order accurate physical-constraints-preserving finite difference632WENO schemes for special relativistic hydrodynamics, Journal of Computational Physics,633298 (2015), pp. 539–564.634
[40] Y. Xing, X. Zhang, and C.-W. Shu, Positivity-preserving high order well-balanced discontin-635uous Galerkin methods for the shallow water equations, Advances in Water Resources, 33636(2010), pp. 1476–1493.637
[41] Q. Zhang and F. Gao, A fully-discrete local discontinuous Galerkin method for convection-638dominated Sobolev equation, Journal of Scientific Computing, 51 (2012), p. 107–134.639
[42] Q. Zhang and C.-W. Shu, Stability analysis and a priori error estimates of the third order640explicit Runge–Kutta discontinuous Galerkin method for scalar conservation laws, SIAM641Journal on Numerical Analysis, 48 (2010), pp. 1038–1063.642
[43] X. Zhang and C.-W. Shu, On maximum-principle-satisfying high order schemes for scalar643
STRONG STABILITY OF RK TIME DISCRETIZATIONS 23
conservation laws, Journal of Computational Physics, 229 (2010), pp. 3091–3120.644[44] X. Zhang and C.-W. Shu, On positivity-preserving high order discontinuous Galerkin schemes645
for compressible Euler equations on rectangular meshes, Journal of Computational Physics,646229 (2010), pp. 8918–8934.647
[45] X. Zhang and C.-W. Shu, Maximum-principle-satisfying and positivity-preserving high-order648schemes for conservation laws: survey and new developments, Proceedings of the Royal So-649ciety of London A: Mathematical, Physical and Engineering Sciences, 467 (2011), pp. 2752–6502776.651
[46] L. Zhou, Y. Xia, and C.-W. Shu, Stability analysis and error estimates of arbitrary652Lagrangian–Eulerian discontinuous Galerkin method coupled with Runge–Kutta time-653marching for linear conservation laws, ESAIM: Mathematical Modelling and Numerical654Analysis, (2018, to appear).655