Parallel Linear System Solvers for

Runge-Kutta Methods
P.J. van der Houwen & J.J.B. de Swart
CWI
P.O. Box 94079, 1090 GB Amsterdam, The Netherlands

Abstract
If the nonlinear systems arising in implicit Runge-Kutta methods like the Radau IIA methods are iterated
by (modified) Newton, then we have to solve linear systems whose matrix of coefficients is of the form
I - A⊗hJ with A the Runge-Kutta matrix and J an approximation to the Jacobian of the righthand side
function of the system of differential equations. For larger systems of differential equations, the solution
of these linear systems by a direct linear solver is very costly, mainly because of the LU-decompositions.
We try to reduce these costs by solving the linear systems by a second (inner) iteration process. This inner
iteration process is such that each inner iteration again requires the solution of a linear system. However,
the matrix of coefficients in these new linear systems is of the form I - B⊗hJ where B is similar to a
diagonal matrix with positive diagonal entries. Hence, after performing a similarity transformation, the
linear systems are decoupled into s subsystems, so that the costs of the LU-decomposition are reduced to
the costs of s LU-decompositions of dimension d. Since these LU-decompositions can be computed in
parallel, the effective LU-costs on a parallel computer system are reduced by a factor s 3 . It will be shown
that matrices B can be constructed such that the inner iterations converge whenever A and J have their
eigenvalues in the positive and nonpositive halfplane, respectively. The theoretical results will be
illustrated by a few numerical examples. A parallel implementation on the four-processor Cray-C98 /
4256 shows a speed-up ranging from at least 2.4 until at least 3.1 with respect to RADAU5 applied in
one-processor mode.

CR Subject Classification (1991): G.1.7

Keywords and Phrases: numerical analysis, convergence of iteration methods, Runge-Kutta methods,
parallelism.

Note: The research reported in this paper was partly supported by the Technology Foundation (STW) in
the Netherlands.

1. Introduction
Suppose that we integrate the IVP

dy
(1.1) dt = f(y), y(t0 ) = y 0 , y, f ∈ R d

by an implicit step-by-step method. In general, this requires in each step the solution of a nonlinear
system of the form

(1.2) R(Yn) = 0, R(Y) := Y - h(A⊗I)F(Y) - Wn-1,

2
where A denotes an s-by-s matrix (assumed to be nondefective), I is the d-by-d identity matrix, Wn-1
contains information from preceding steps, h is the stepsize tn - tn-1, and ⊗ denotes the Kronecker
product. It will always be assumed that A is nondefective and has its eigenvalues in the positive
halfplane. The s components Yni of the sd-dimensional solution vector Yn represent s numerical
approximations to the s exact solution vectors y(tn-1 + cih); here, c = (ci) denotes the abscissa vector
whose components ci are assumed distinct. Furthermore, for any vector Yn = (Y ni), F(Yn) contains
the derivative values (f(Y ni)). In the following, we shall use the notation I for any identity matrix.
However, its order will always be clear from the context. The solution Yn of (1.2) will be called the
stage vector, yn the step point value, s the number of stages, and A the Runge-Kutta matrix.
Usually, the nonlinear system (1.2) is solved by modified Newton iteration. This leads to linear
systems whose matrix of coefficients is of the form I - A⊗hJ with J an approximation to the Jacobian
of the righthand side function f. The solution of these systems may be extremely costly. For example,
if a direct solver is used, then in general the LU-decomposition requires 23 s3d3 arithmetic operations
which is considerable, even for moderate values of d (say d ≈ 10). Moreover, there is only a limited
intrinsic parallelism in building the LU decomposition of the matrix I - A⊗hJ.

1.1. Reduction of computational costs

We briefly survey various approaches to reduce the computational costs associated with the solution
of the Newton systems using parallel computer systems. Firstly, one may look for special methods in
which A is a triangular matrix with positive diagonal entries like the DIRK type methods. Then,
confining our considerations to the costs of the LU decomposition, we see that the effective LU-costs
on s processors reduce to 32 d3 operations, a factor s3 less than those needed for the Newton process.
However, these DIRK type methods also have disadvantages. In the case of one-step DIRKs
available in the literature, the step point order is at most 4 and they have a relatively low stage order
which may be a disadvantage in certain classes of stiff IVPs. Higher step point orders and stage
orders can be obtained in the class of multistep RK methods (cf. Burrage and Chipman [3]), but they
have the disadvantage of quite large abscissae values ci (much larger than 1).
More sophisticated than the DIRK methods are methods characterized by matrices A with only
positive eigenvalues such as the one-step RK methods of Nørsett [16], Burrage [1] and Orel [17]. By
performing a similarity transformation (or Butcher transformation [5]), the linear systems can be
decoupled into s subsystems of dimension d. Again, the effective LU-costs reduce by a factor s 3 , and
moreover, the stage order and step point order are much higher than for DIRK methods. However, a
possible disadvantage of these methods is the lack of superconvergence at the step points.
Finally, one may choose the classical RK methods possessing both a high stage order and a high step
point order, but also one or more complex eigenvalues. Again, applying a similarity transformation,
the Newton system is transformed to block-diagonal form with (real) diagonal blocks, each block
corresponding to an eigenvalue of A. If an eigenvalue is real, then the associated diagonal block is of
order d, otherwise it has order 2d. The LU-costs of these blocks are reduced to 23 d3 and 16 d3
3
operations, so that effectively the LU-costs are 16 d3 operations, irrespective the value of s (the code
3
RADAU5 of Hairer and Wanner [7] uses such a transformation).
 3

1.2. Iterative solution of the linear systems

Instead of using direct solution methods, one may also look for iterative linear solvers, such as
GMRES or preconditioned GMRES (see e.g. Burrage [2] where further references are given).
In this paper, we shall follow an approach that is a mixture of an iterative and a direct approach. It
allows A to have complex eigenvalues (in the positive halfplane), so that the superconvergent RK
methods like the Radau IIA methods are included. The linear systems arising in the modified Newton
method are solved by an iterative method (the inner iteration process), which needs itself LU
decompositions of matrices, but these matrices are only of dimension d. In fact, the linear systems to
be solved have a matrix of coefficients of the form I - B⊗hJ where B is similar to a diagonal matrix
with positive diagonal entries. Hence, after performing a similarity transformation, the effective LU
costs are 32 d3 operations like the methods of Burrage and Orel. We shall refer to this inner iteration
process by PILSRK (Parallel Iterative Linear System solver for RK methods). The combination of the
modified Newton and the PILSRK method will be called the Newton-PILSRK method.
There are several options for choosing the matrix B. The most simple approach chooses B = D where
D is a diagonal matrix (with positive entries), so that the sd-dimensional system can directly be split
into s uncoupled subsystems of dimension d which can be solved concurrently. In fact, we can
employ the same matrices D as used in the Parallel Diagonal-implicitly Iterated RK methods (PDIRK
methods) analysed in [10]. The PDIRK method is also an iterative method, but unlike the PILSRK
method it is a nonlinear system solver and directly iterates on the nonlinear system (1.2). Using
results derived by Lioen [13] for PDIRK matrices, it can be shown that for the first eight Radau IIA
correctors, the PILSRK methods are A-convergent, that is, it converges if J has its eigenvalues in the
nonpositive halfplane. Furthermore, these PDIRK matrices have the property that the stiff
components are removed from the iteration error within s iterations. However, a disadvantage of the
PDIRK matrices is the poor convergence (or even divergence) of the PILSRK method in the first few
iterations which is worse as the number of stages of the underlying RK corrector increases. Such a
convergence behaviour is highly undesirable if we want to apply step-parallel iteration, where the
iteration process is already started at the next step point tn+1 before the iterates at tn have converged. A
poor initial convergence implies that no accurate predictor value is available for starting the iteration
process at tn+1. A substantial improvement in the initial phase of the convergence of the PILSRK
method is obtained by employing the matrices L used in the Parallel Triangular-implicitly Iterated RK
methods (PTIRK methods) constructed in [11] (like the PDIRK methods, the PTIRK methods are
nonlinear system solvers). The PTIRK matrices L are defined by the lower triangular factor of the
Crout decomposition LU of the RK matrix A. By virtue of results obtained by Hoffmann and De
Swart [8], it can be shown that for all RK correctors that are based on collocation with positive,
distinct abscissae, the matrix L has positive diagonal entries and that the PILSRK method is A-
convergent. Furthermore, like the PDIRK matrices, the PTIRK matrices have the property that the
stiff components are removed from the iteration error within s iterations. After performing a similarity
transformation, the effective LU-costs are reduced by a factor s 3 . A preliminary parallel
4
implementation of the Newton-PILSRK method based on the one-step 4-stage Radau IIA formula and
using the PTIRK matrix showed on the four-processor Cray-C98 / 4256 speed-up factors ranging
from at least 2.4 until at least 3.1 with respect to RADAU5 in one-processor mode (cf. [11]).

1.3. Outline of the paper

The aim of the present paper is to find matrices B that are still more effective than the PTIRK matrices
L. Our starting point is the representation B = QTQ-1 with T a lower triangular matrix with positive
diagonal entries and with Q a nonsingular transformation matrix such that Q-1AQ is lower block-
triangular. It will be shown that matrices T and Q exist such that:

(i) B is nondefective and has positive eigenvalues,

(ii) the PILSRK method is A-convergent whenever A has its eigenvalues in the positive halfplane,
(iii) the stiff components are removed from the iteration error in the second iteration,
(iv) the spectral radius of the iteration-error-amplification matrix is minimized in the left halfplane.

The difficult part is the construction of matrices Q such that the iteration-error-amplification matrix has
a sufficiently small norm. In this paper, we construct transformation matrices so that Q-1AQ is block-
diagonal (in a forthcoming paper, we shall deal with alternative families of transformation matrices).
For the 4-stage and 8-stage Radau IIA correctors, matrices Q will be constructed such that the
Euclidean norm of powers of the iteration-error-amplification matrix are satisfactorily small.
As soon as T and Q, and hence B, are obtained, we can compute the diagonalizing similarity
transformation, to obtain a highly parallel linear system solver.
In this paper, we have restricted our analysis of the Newton-PILSRK method to the case where (1.2)
represents the class of one-step Radau IIA methods, that is, A is the Radau IIA matrix and
W n-1 := (E⊗I)Yn-1 with E = (0, ... , 0, e), e being an s-dimensional vector with unit entries. These
methods are of particular interest because of their high step point order p = 2s-1 and high stage order
q = s, their stiff accuracy and their excellent stability properties. The Newton-PILSRK methods were
applied to a few problems taken from the literature. The results show a considerable improvement of
the convergence in the first few outer iterations. Recalling that a parallel implementation of the
Newton-PILSRK method using the PTIRK matrices already shows a speed-up factor of at least 2.4
with respect to RADAU5, we expect that using the new matrices B = QTQ-1 will yield a further
speed-up. The parallel implementation of the new methods will be subject of future research.
Finally, we remark that it may well be that the class of multistep RK methods of Radau type (cf.
Hairer and Wanner [7, p. 293]) is a better choice for the corrector equation (1.2) than the one-step
Radau methods. For nonstiff IVPs, Burrage and Suhartanto [4] have investigated the use of parallel
iteration methods for such correctors and they report promising results. This indicates that applying
the PILSRK approach of this paper to the Newton systems arising in multistep Radau methods may
lead to quite effective parallel IVP methods.
 5

2. The parallel iterative linear system solver

Consider the modified Newton iteration scheme for solving the corrector equation (1.2):

(2.1) (I - A⊗hJ)(Y(j) - Y(j-1)) = - R(Y(j-1)), j = 1, 2, ... , m,

where J = ∂f/∂y is evaluated at tn-1, Y(0) is the initial iterate to be provided by some predictor
formula, and where Y(m) is adopted as the solution Y n of the corrector equation (1.2). Each iteration
with (2.1) requires the solution of an sd-dimensional linear system for the Newton correction
Y(j) - Y(j-1). As already observed, direct solution of this Newton system can be extremely costly and
transformation to block-diagonal form reduces computational costs considerably. In order to achieve a
still greater reduction of the computational complexity we follow an alternative approach by applying
an iterative linear solver to the Newton systems in (2.1). This solver again requires the solution of
linear systems, but these systems are only of dimension d. It is tuned to the RK structure of the
systems in (2.1) and possesses a lot of intrinsic parallelism. This Parallel Iterative Linear System
solver for RK methods (PILSRK method) is defined by

(I - B⊗hJ)(Y(j,ν) - Y(j,ν-1)) = - (I - A⊗hJ)Y(j,ν-1) + C(j-1),

(2.2) ν = 1, 2, ... , r,
C(j-1) := (I - A⊗hJ)Y(j-1) - R(Y(j-1)),

where Y(j,0) = Y(j-1,r) and where Y(m,r) is accepted as the solution Yn of the corrector equation (1.2).
The matrix B is assumed to be nondefective and to have positive eigenvalues. Note that C(j-1) does
not depend on ν, so that the application of the inner iteration process requires only one evaluation of
the function R. The processes (2.1) and (2.2) may be considered as the outer and inner iteration
processes.
In order to construct a suitable matrix B, we observe that the condition on the spectrum of B implies
that we can write B = QTQ-1 with Q an arbitrary real, nonsingular matrix and T a lower triangular
~
matrix with positive diagonal entries. Hence, by performing the transformation Y(j,ν) = (Q⊗I)Y(j,ν) ,
we obtain

~ ~ ~ ~
(2.3) (I - T⊗hJ)(Y(j,ν) - Y(j,ν-1) ) = - (I - A ⊗hJ)Y(j,ν-1) + (Q-1⊗I)C (j-1), ν = 1, 2, ... , r,
~ ~ ~
where A = Q-1AQ and Y(j,0) = (Q-1⊗I)Y(j-1). If for a given j, the transformed inner iterates Y(j,ν)
~
converge to a vector Y(j,∞) , then the Newton iterate defined by (2.1) can be obtained from
~
Y(j) = (Q⊗I)Y(j,∞) . Given the matrix A, the PILSRK method (2.3) is completely defined by the
matrix pair (T,Q) and will be denoted by PILSRK(T,Q). The representation (2.3) will be the starting
point for the construction of the matrix B.
Before discussing the computational costs of the actual implementation of the Newton-PILSRK
method {(2.1),(2.3)}, we should specify the matrix B. This will be the subject of Section 3. Details
on the computational complexity can be found in Section 4.2.
6

Remark 2.1. In the first Newton iterations, it seems a waste to perform many inner iterations with the
PILSRK method, because there is no point in computing a very accurate approximation to Y(j), as
long as Y(j) is itself a poor approximation to Y n. Likewise, in later outer iterations, we expect that
only a few inner iterations suffice to solve Y(j) from (2.1). In the extreme case, only one inner
iteration is performed in each outer iteration. In such an iteration strategy, the Newton-PILSRK
iteration method {(2.1),(2.2)} simplifies to

(2.4) (I - B⊗hJ)(Y(j) - Y(j-1)) = - R(Y(j-1)), j = 1, 2, ... , m.

However, this process may converge very slowly in the first few outer iterations, and it is
recommended, either to use highly accurate predictor formulas for Y(0) or to introduce a dynamic
iteration strategy so that when necessary, sufficiently many inner iterations in the first few outer
iterations are performed.
Notice also that the iterative method obtained from (2.1) by using a splitting of A into B and A-B is
identical with the iteration method (2.4). ♦

3. Construction of the matrix B

Given the matrix A, the PILSRK method (2.2) is completely determined by the matrix B = QTQ-1. In
the construction of B, the region of convergence and the averaged amplification factors for the
iteration errors play a central role.

3.1. Convergence region of the PILSRK method

In order to analyse the region of convergence for the PILSRK method, we consider the error
recursion

(3.1) Y(j,ν) - Y(j) = M(Y(j,ν-1) - Y(j)), M := (I - B⊗hJ)−1 ((A - B)⊗hJ).

We have convergence if the powers Mν of the amplification matrix M tend to zero as ν → ∞, that is,
if the spectral radius ρ(M) of M is less than 1. The eigenvalues of M are given by the eigenvalues of
the matrix

(3.2) Z(z) := z (I - zB)-1 (A - B), z := hλ,

where λ runs through the eigenvalues of J. We call Γ := {z: ρ(Z(z)) < 1} the region of convergence
of the PILSRK method. Thus, the method converges if the eigenvalues of hJ lie in Γ . If Γ contains
the whole nonpositive halfplane, then the method will be called A-convergent.
We shall call Z(z) the amplification matrix at the point z and ρ(Z(z)) the (asymptotic) amplification
factor at z. The maximum of ρ(Z(z)) in the left halfplane Re(z) ≤ 0 will be denoted by ρ.
In [10] and [11] where the PDIRK and PTIRK methods were analysed, it turned out that strong
damping of the stiff error components, that is, small amplification factors for error components
 7
corresponding to eigenvectors of J with eigenvalues λ of large magnitude, is crucial for a fast overall
convergence. This leads us to require the matrix B to be such that ρ(Z(∞)) = ρ(I - B-1A) vanishes. If
we succeed in finding such matrices B, then Zs(∞) = O, so that within s iterations, the components
corresponding to |λ| = ∞ are removed from the iteration error (this can be verified by considering the
Schur decomposition of Zs(∞)).
As an example, let Q = I and let T be a diagonal matrix D, so that B = D. Lioen [13] showed that for
the s-stage Radau IIA correctors with s ≤ 8, it is possible to construct diagonal matrices D satisfying
ρ(I - D-1A) = 0 such that the generated PILSRK(D,I) method is A-convergent. These matrices are
also used in the PDIRK methods studied in [10], and will therefore be called PDIRK matrices.
The next theorem defines a family of PILSRK(T,Q) methods automatically satisfying the condition
ρ(I - B-1A) = 0.

Theorem 3.1. Let Q be an arbitrary, nonsingular matrix and let B = QTQ-1, where T is the lower
~
triangular factor in the Crout-decomposition of A := Q-1AQ. Then, the asymptotic amplification factor
vanishes at infinity.
~
Proof. Let TU represent the Crout-decomposition of A . Then Q-1Z(∞)Q = I - Q-1B-1AQ =
~
I - T-1A = I - U is strictly upper triangular. Hence, ρ(Q-1Z(∞)Q) = ρ(Z(∞)) = 0. ♦

The matrix B in the PILSRK methods characterized by this theorem does not necessarily have positive
eigenvalues and hence, does not automatically generate A-convergent methods. This requires special
transformation matrices Q. Let us again consider the case where Q = I. Then, B equals the lower
triangular factor in the Crout-decomposition of A, that is, B equals the PTIRK matrix L derived in
[11]. In [8], Hoffmann and De Swart were able to prove that the PTIRK matrix L possesses positive
diagonal entries for all collocation-based RK correctors with positive, distinct abscissas, so that B has
positive eigenvalues as required. Furthermore, numerical computations in [11] showed the A-
convergence for a large number of RK correctors based on Gaussian quadrature formulas.
The aim of this paper is to derive A-convergent methods with ρ(I - B-1A) = 0 for more general pairs
(T,Q) than the PTIRK pair (L,I), and to find pairs (T,Q) such that we can a priori prove both the
positiveness of the eigenvalues of B and the A-convergence of the generated iteration method.
~ ~
Let us choose Q such that A := Q-1AQ = (A kl) is a (real) σ-by-σ lower block-triangular matrix, of
~
which the diagonal blocks A kk are either one-by-one or two-by-two matrices. If ξk is a real
~
eigenvalue of A, then A kk = ξk, and if ξk ± iηk is a complex eigenvalue pair of A, then

~ ak bk
(3.3) A kk =   , b = - c -1(a 2 - 2ξ a + α 2), c ≠ 0, α := ξk2 + η k2 ,
 ck 2ξ k - a k  k k k k k k k k

where ak and ck are free parameters. In the following, K will denote the set of integers with the
property that ηk ≠ 0 whenever k ∈ K.
A natural choice for T now is
8

 ~A 
T11 O O O ...

T22 O O ...  uk 0 
 if k ∈ K, Tkk = ξk otherwise,
21 , Tkk := 

 A31 ... 
(3.4a) T :=
~ ~  vk wk 
.
A32
.
T33 O
. . . 
where uk, vk and wk are free parameters with uk and wk assumed to be positive.
~ ~
Theorem 3.2. Let A have its eigenvalues in the positive halfplane, let A := Q-1AQ = (A kl) be lower
block-triangular, let the diagonal blocks be defined by (3.3) and let T = T(γ) be defined by (3.4a) with

(γ 2 - 1)a k - 2γ 2 ξ k + 2γα k αk
(3.4b) u k = γα k , vk = ck α k , wk = ,
γ(a k 2 - 2ξ k a k + α k 2 ) γ

where γ is a positive parameter. Then, for all ak and ck the following assertions hold:

(i) The asymptotic amplification factor vanishes at infinity.

(ii) B has positive eigenvalues and if γ ≠ 1 it is nondefective.
(iii) The PILSRK(T(γ),Q) method is A-convergent with ρ = max{|1 - 2γ(γ2 + 1)-1ξkαk-1|: k ∈ K}.

Proof. Let

~ ~
(3.5) Z(z) := Q-1Z(z)Q = z (I - zT)-1 (A - T), z := hλ.
~
If T is of the form (3.4a), then the value of ρ(Z(z)) = ρ(Z (z) ) equals the maximum of the spectral
~
radius ρ(Z kk(z)) of the diagonal blocks

~ ~
(3.6) Z kk := z (I - zTkk)-1 (A kk - Tkk)
~ ~
of Z . Here, Z kk vanishes if the underlying eigenvalue of A is real. Hence, in order to achieve
ρ(Z(∞)) = 0, we choose the Tkk with k ∈ K such that the spectral radius of the corresponding
diagonal blocks Zkk(z) vanishes at infinity.
~
We derive from (3.3) and (3.6) that the eigenvalues ζk of Z kk satisfy the characteristic equation

 (ak-u k)z - ζ k(1-zuk) bkz 

(3.7) det   = 0.
 
 (ck-v k)z + ζ kv kz (2ξk-ak-w k)z - ζk(1-zw k) 

It is easily verified that ζk = ζk(z) vanishes at infinity if uk, vk and wk are defined according to (3.4b).
~
Hence, ρ(Z kk(z)) vanishes at infinity which proves part (i) of the theorem.
Since the eigenvalues of B are given by (uk, wk) for k ∈ K and by ξk for k ∉ K, and because we
assumed γ > 0, (3.4b) also implies that B has positive eigenvalues and if γ ≠ 1 it is nondefective,
proving part (ii).
The characteristic equation (3.7) is solved by
 9

(2ξ - u k - w k ) z
(3.8) ζk = 0, ζk = (1 - kzu )(1
k - zw k ) ,
~
so that ρ(Z (z) ) equals the maximum of the values |ζk(z)|. Since ζk(z) is regular in left half plane
(provided that uk and wk are positive), its maximum in the left halfplane Re(z) ≤ 0, to be denoted by
ρk, is assumed on the imaginary axis. It is easily verified that

~ |2ξ k - u k - w k | |y|
(3.9) ρ(Z kk(iy)) = |ζk(iy)| =
( 1 + u k 2 y 2 )( 1 + w k 2 y 2 )

assumes an absolute maximum at y = y0 := (ukwk)-1/2 and that the maximum value ρk of ρ(Zkk(iy))
is given by ρk = |1 - 2ξk(uk + wk)-1| = |1 - 2γ(γ2 + 1)-1ξkαk-1|, which is less than 1 whenever
γξk > 0. This proves part (iii) of the theorem. ♦

The asymptotic amplification factor ρ is minimized for γ = 1 and assumes the minimal value
ρ = max{1 - ξkαk-1: k ∈ K}. However, then the matrices Tkk are defective (because uk = w k).
Hence, T cannot be diagonalized, and although the effective LU-costs are still reduced by a factor s ,
the Newton-PILSRK(T(1),Q) method should be considered as a σ-processor method, rather than an
s-processor method. Fortunately, the asymptotic amplification factor varies slowly with γ, so that we
can remove the defectness of T at the cost of a slight increase of ρ. For example, for the method
defined by (3.4) we find for γ = 78 ,

(3.10) ρ = max{1 - 112 ξ α -1: k ∈ K},

113 k k

which is only slightly larger than the minimal value. For a detailed discussion of the computational
complexity of an implementation of the Newton-PILSRK(T(γ),Q) method, we refer to Section 4.2.

Remark 3.1. When faced with the problem of choosing a matrix T such that the eigenvalues of the
matrix Z(z) are of small magnitude, it is tempting to minimize the magnitude of the matrix factor
~ ~
A - T occurring in the matrix Z (z) defined by (3.5). Since

~ ~ ~ ~  a -u bk 
A - T = diag (A 11 - T11, ..., A σσ - Tσσ), A kk - Tkk =  k k 
 ck-vk 2ξk-ak-wk 

and because for given ak, the magnitude of the entry bk = - ck-1(ak2 - 2ξkak + αk2) can be made as
~
small as we want, we are led to zero the other three entries of A kk - Tkk by setting uk = ak, vk = ck
and wk = 2ξk - ak. This still leaves ak as a free parameter which can be used to minimize bk for given
ck, to obtain ak = ξk and bk = -ηk2ck-1. However, substitution of the parameters uk, vk, w k, ak and
b k into the characteristic equation (3.7) reveals that the nonzero eigenvalue is given by
ζk = (ξk2 - αk2)z2(1 - zξk)-2, which assumes the extreme value -(ηkξk-1)2 at infinity. Thus, we have
no A-convergence when A has eigenvalues whose imaginary part exceeds its real part. Since many
10
RK methods based on Gaussian quadrature do have imaginary parts that exceed the real parts, the
~
approach of minimizing the magnitude of A - T is the wrong way to go.♦

Remark 3.2. The family of matrices T defined by (3.4) contains the special case where T is defined by
~
the lower triangular factor in the Crout-decomposition of A := Q-1AQ (see Theorem 3.1):

 ak 0 
(3.11) Tkk :=   if k ∈ K, Tkk = ξk otherwise.
 ck αk2ak-1 

This expression is also obtained from (3.4) by setting γ = akαk-1. ♦

We conclude this section with listing values of ρk for a few Radau IIA correctors and for the iteration
strategy PILSRK(T(7/8),Q) defined by Theorem 3.2. In addition, we list the values of ρ for
PILSRK(D,I) with the PDIRK matrix D and for PILSRK(L,I) with the PTIRK matrix L. The figures
in Table 3.1 show that on the basis of the asymptotic amplification factors, the PILSRK(T(7/8),Q)
approach is superior to PILSRK(D,I) and PILSRK(L,I).

Table 3.1. Values of ρk for Radau IIA methods.

---------------------------------------------------------------------------------
Iteration k s=2 s=3 s=4 s=6 s=8
---------------------------------------------------------------------------------
PILSRK(D,I) - 0.26 0.40 0.52 0.72 0.90
PILSRK(L,I) - 0.18 0.37 0.51 0.70 0.86
PILSRK(T(7/8),Q) 1 0.19 0.35 0.45 0.57 0.64
2 0.06 0.21 0.33
3 0.03 0.12
4 0.02
---------------------------------------------------------------------------------

3.2. Averaged amplification factors

Because the matrix M in (3.1) is not expected to be a normal matrix, the asymptotic amplification
factor ρ discussed in the preceding section only gives an indication of the speed of convergence after a
quite large number of iterations and does not give insight into the convergence behaviour in the initial
phase of the iteration process. In fact, for large ν we have the estimate || Mν || ≤ κ(S) [ρ(M)]ν, where
S represents the eigensystem of M, κ(S) := ||S || ||S-1|| is the condition number of S, and where we
assumed that M has eigenvalues of multiplicity 1 (cf. Varga [20]). In order to analyse the convergence
rate in the first few iterations, one may use to the pseudo-eigenvalue analysis of Trefethen (see e.g.
[18]). Alternatively, we may resort to a well-known theorem of Von Neumann. We shall follow the
latter approach.
Let the logarithmic matrix norm µ[S] associated with the Euclidean norm be defined by
µ[S] = 12 λmax(S+SH), where SH is the complex transposed of S and λmax(.) denotes the
algebraically largest eigenvalue. Then, we have:
 11
Theorem 3.3. If µ[J] ≤ 0, then || Mν|| ≤ max{|| Zν(z)|| : Re(z) ≤ 0}.

Proof. The proof is based on a generalization of a theorem of Von Neumann. Von Neumann's
theorem states that, given a matrix J and a rational function R of z which has a bounded maximum
norm || R ||∞ in the left halfplane, then ||R(J)|| ≤ || R ||∞, provided that µ[J] ≤ 0 (see e.g. [7, p.179]). A
matrix-valued version of Von Neumann's theorem, applying to the case where R(z) is a matrix with
entries that are rational functions of z, was proved by Nevanlinna [15] (see also [7, p.356]). Since
Mν can be considered as a matrix-valued function of J (see (3.1)), we apply the matrix-valued version
of Von Neumann's theorem with R(z) := Mν(z), where

ν
(3.12) Mν(z) = [(I - B⊗zI)−1((A - B)⊗zI)] = Zν(z)⊗I, z = hλ.

This leads to the assertion of the theorem.♦

This theorem motivates us to define the local averaged amplification factor at the point z = hλ and the
global averaged amplification factor by

ν
(3.13a) ρ(ν)(z) := || Zν(z)|| , ρ(ν) := max{ρ(ν)(z): Re(z) ≤ 0}.

Note that ρ(ν)(z) approximates the asymptotic amplification factor ρ(Z(z)) as ν → ∞. Since in the left
halfplane, ρ(ν)(z) assumes its maximum on the imaginary axis, we may restrict our considerations to
the imaginary axis, so that ρ(ν) := max{ρ(ν)(iy): y ≥ 0}.
Theorem 3.3 indicates that we may expect faster convergence as ρ(ν) is smaller. However, for small
numbers of iterations (say ν ≤ 5), ρ(ν) will give a rather conservative estimate of the speed of
convergence, because in some sense it is a 'worst case' estimate. In order to get insight into the
amplification of individual error components, one may use the local amplification factor ρ(ν)(z). Let
us consider error components of the form a ⊗ v, where a is an s-dimensional vector and v is an
eigenvector of J with eigenvalue λ. By observing that Mν(a ⊗ v) = (Zν(hλ)⊗I) (a ⊗ v), it follows
that ρ(ν)(hλ) characterizes the averaged convergence of the error component corresponding with hλ
and that only for larger values of ν, when the error component with maximal ρ(ν)(hλ) has become
dominant, ρ(ν) yields a quantitative estimate of the averaged convergence rate. In the first few
iterations, when all error components play their part, the L2 norm of the local amplification factor
ρ(ν)(z) provides more realistic estimates than the L∞ norm. This suggests to define a second global
amplification factor:

y=∞
(3.13b) σ(ν) := ( ⌠
⌡ [ρ(ν)(iy)]2 dy ) ) 1/2
.
y=0

We did not succeed in finding an approach which really minimizes ρ(ν). However, by considering the
estimate
12
ν ~ν ~ν
(3.14) || Z (z) || = || QZ (z)Q-1 || ≤ κ(Q)|| Z (z) || ,
~ν
we see that ρ(ν)(z) ≤ (κ(Q)|| Z (z) ||) 1/ν , which suggests the separate minimization of the factors
~ν
κ(Q) and || Z (z) || . We distinguish two approaches. In the first approach, we choose Q orthogonal,
~
so that κ(Q) = 1. This can be achieved by defining A := Q-1AQ by the real Schur decomposition of
A, leading to ak = ξk and ck = -ηk (see e.g. [6]). In a future paper, we shall elaborate on this case. In
~ν
the present paper, we analyse a second approach where first || Z (z) || is minimized and then κ(Q). We
~
shall do this for the case where A is block-diagonal.

3.3. The block-diagonal case

~
In the remainder of this section, we shall analyse the case where A := Q-1AQ is block-diagonal and
~ν
we shall use the still free parameters ak and ck for reducing the magnitude of || Z (z) ||. However, we
~
first justify our choice of a block-diagonal matrix A by considering the damping of the stiff error
components. The following theorem presents a result on the amplification of the stiff iteration errors.
~
Theorem 3.4. Let the conditions of Theorem 3.2 be satisfied and let A := Q-1AQ be block-diagonal.
Then, the averaged amplification factor ρ(ν)(z) = O(z(1-ν)/ν) as z → ∞ and the averaged global
amplification factor σ(ν) is finite if ν > 2.

Proof. For z → ∞, it follows from (3.2) that

Z(z) = (I - z-1B-1)-1 (I - B-1A) = Z(∞) + z-1B-1Z(∞) + O(z-2), Z(∞) = I - B-1A

(B may be assumed to be nonsingular because it is required to have positive eigenvalues). More

generally, we have that

∞
ν ceil [ν / i]
Z (z) = ∑ (Z(∞)) O(z1−i),
i=1

where for any real x, ceil[x] denotes the first integer greater than or equal to x. We first show that all
ν ~ν
integer powers of Z(∞) greater than 1 vanish. Since Z = QZ Q-1, we have to show that all integer
~
powers of Z (∞) greater than 1 vanish. Because Q-1AQ is block-diagonal, it follows from (3.4) that T
~ ~
is block-diagonal and from (3.5) that Z(z) is block-diagonal. Hence, Z(∞) is block-diagonal with
~ ~
diagonal blocks Z kk(∞). Since by virtue of Theorem 3.2, these blocks have a zero spectral radius, (Z
ν
kk(∞)) vanishes for ν ≥ 2 (this can easily be verified by considering their Schur decompositions).
~ν ν ν
Consequently, Z (∞) itself, and hence Z (∞), vanishes for ν ≥ 2. From the expansion of Z (z) we
now immediately obtain Zν(z) = O(z1-ν) as z → ∞. Substitution into (3.13) yields the result of the
theorem. ♦
 13
From this theorem it follows that the stiff error components may be considered as being removed
from the iteration error within two (inner) iterations.
If we only know that Z(∞) has a zero spectral radius, as in the case of the PDIRK and PTIRK
matrices D and L, then Zν(∞) vanishes for ν ≥ s. Hence, by virtue of (3.14) it is seen that for ν ≥ s
we have Zν(z) = O(z1−ceil [ν/(s-1)]) as z → ∞, so that ρ(ν)(z) = O(z(1−ceil [ν/(s-1)])/ν) as z → ∞ and
σ(ν) is finite only if 2(1 - ceil[ν / (s-1)]) / ν is less than -1, i.e. if s ≤ 2. Thus, by virtue of the block-
~
diagonality of the matrix A , the PILSRK(T,Q) has a much better stiff initial convergence than the
PILSRK(D,I) and PILSRK(L,I) methods.

~ν
3.3.1. Reduction of || Z (z) || in the left halfplane. We derive an estimate for the maximum norm of
~ ~
|| Zν(z)|| in the left halfplane by using the inequality (3.14). Since A := Q-1AQ is block-diagonal, Z
ν ~ ν
(z) is also block-diagonal with diagonal blocks Z kk (z) given by

~ (ak-uk)(1-wkz) bk(1-wkz)
(3.15) Z kk(z) = (1-u z)(1-w z)   .
z
k k  (a -u )v
k k k z + (c -v
k k )(1-u k z) b v
k k z + (2ξ -a -w
k k k )(1-u k z) 

Here, the parameters uk, vk and wk satisfy (3.4b). We first minimize the magnitude of
~ ν ~
|| Z kk (z)||. Note that this can be done independently of Q. Having found Z kk, we determine Q by
minimizing κ(Q). The representation (3.15) suggests setting ak = uk and ck = vk, to obtain for k ∈ K

~ ν ν (2γξ k - γ 2 α k - α k )z γ-α z
(3.16) Z kk (z) = ζk (z)  0 k1  , ζ k (z) :=
0 q (z)
, qk(z) := α k c k .
  (1 - γα k z)(γ - α k z ) k

Note that setting ak = uk in (3.4b) implies ck = vk.

Theorem 3.5. Let the conditions of Theorem 3.4 be satisfied, let ak = γαk, |ck| ≥ γ-1(1 + γ2)αk. Then,
with respect to the maximum norm, the averaged amplification factor satisfies ρ(ν) ≤ [κ(Q)]1/νρ,
where ρ = max{|1 - 2γ(γ2 + 1)-1ξkαk-1|: k ∈ K}.

Proof. Let for any matrix M(z) depending on the complex variable z, ||| M ||| denote the maximum
norm of the function || M(z) || in the left halfplane, where || . || denotes the maximum matrix norm. It is
easily seen that

~ 2γξ k - γ 2 α k - α k 2γξ k - γ 2 α k - α k
(3.17) ||| Z ||| = max {| (1 + γ 2 )α k
|,| γck
| : k ∈ K} .
~
By choosing |ck| ≥ γ-1(1 + γ2)αk, we find that ||| Z ||| equals the asymptotic amplification factor ρ as
~ν ~ ν ν
given in Theorem 3.2. Hence, ||| Z ||| ≤ ||| Z ||| = ρ . Obviously, we can never have strict inequality,
~ν ~ν
so that we conclude that ||| Z ||| = ρ ν . Finally, it follows from (3.14) that ||| Zν||| ≤ κ(Q)||| Z ||| =
κ(Q)ρν. Thus, the averaged amplification factor ρ(ν) is bounded by [κ(Q)]1/νρ. This completes the
proof of the theorem.♦
14
We remark that for ν → ∞, we have the estimate ρ(ν) ≤ ρ max{[κ(S(z))]1/ν: Re(z) ≤ 0 } , where S(z)
represents the eigensystem of Z(z) and where we assumed that Z(z) has distinct eigenvalues. The
advantage of the estimate in Theorem 3.5 is that it holds for all ν.

3.3.2. The transformation matrix Q. In this subsection, we assume that the PILSRK method satisfies
the conditions of Theorem 3.5. In order to obtain small amplification factors (ρ (ν), σ(ν)) as defined
by (3.13), we shall use the freedom left in choosing the transformation matrix Q. We specify our
approach for the case where all eigenvalues ξk ± iηk of A are complex (ηk ≠ 0), so that σ = s/2. Then,
the column vectors qj of Q are defined by

(3.18) (q2k-1, q2k) = (βkxk + δkyk, -δkxk + βkyk) Q k, k = 1, ... , 12 s,

where βk and δk are free parameters and x k ± iyk represent the normalized eigenvectors of A
corresponding with ξk ± iηk such that the first component of y k vanishes. Here, Q k is a
transformation matrix satisfying (cf. (3.3))

~  ξ ηk  ~  γαk
γ(γ 2 α k - 2γ ξ k + α k )

(3.19a) A kk = Qk-1  k  Qk,A kk :=
 -ηk ξk 
 1 + γ2
1 + γ2  .
- γ
αk 2ξ k - γα k 
It can be verified that the matrix

1  (1 + γ 2 )η k 0 
(3.19b) Qk =  
γ(γ 2 α k - 2γ ξ k + α k )  (1 + γ 2 ) ( γ α k - ξ k ) γ(γ 2 α k - 2γ ξ k + α k ) 
~
satisfies (3.19a). By means of (3.18) and (3.19) it is easily verified that we do obtain the matrix A =
Q-1AQ. The advantage of this approach is that the resulting matrix Q has real entries.
For a given value of γ, the equations (3.19) and (3.20) determine a family of transformation matrices
Q with free parameters vectors β = (βk), δ = (δk) and c = (ck), where |ck| ≥ γ-1(1 + γ2)αk.

Table 3.2. Global amplification factors ρ(ν) for PILSRK methods.

--------------------------------------------------------------------------------------------------------
4-stage Radau IIA corrector 8-stage Radau IIA corrector
ν PILSRK(D,I) PILSRK(L,I) PILSRK(T,Q) PILSRK(D,I) PILSRK(L,I) PILSRK(T,Q)
--------------------------------------------------------------------------------------------------------
1 3.60 0.59 1.95 19.83 1.03 3.51
2 2.48 0.54 0.98 11.52 0.94 1.88
3 1.64 0.53 0.76 7.74 0.91 1.30
4 1.16 0.53 0.66 5.55 0.90 1.19
5 0.96 0.52 0.61 4.08 0.89 1.09
... ... ... ... ... ... ...
9 0.72 0.51 0.53 1.86 0.88 0.87
10 0.69 0.51 0.52 1.72 0.88 0.85
11 0.67 0.51 0.51 1.61 0.88 0.83
... ... ... ... ... ... ...
∞ 0.52 0.51 0.45 0.90 0.86 0.64
--------------------------------------------------------------------------------------------------------
 15

By a numerical search, we found in the case of the 4-stage and 8-stage Radau IIA correctors for γ = 87
the values (3.20) yielding a sequence of satisfactory small amplification factors ρ(ν) (see Table 3.2):

s = 4:β = (5, -4), δ = (-1, -5), a = 78 (α 1, α 2) , c = - 113

56
a,
(3.20)
s = 8:β =(-0.9, -2, -2, 1.1), δ = (1.1, 0.3, 0.3, -1.9),
7
a= (α 1, α 2, α 3, α 4) , c = - 113
56
a.
8

Table 3.2 also lists amplification factors ρ(ν) for the PILSRK(L,I) and PILSRK(D,I) methods. This
table clearly shows that in terms of ρ(ν) values, the PILSRK(T,Q) methods are superior to the
PILSRK(D,I) method. With respect to PILSRK(L,I), the ρ(ν) values of PILSRK(T,Q) are smaller
only for large numbers of inner iterations. In fact, they become less than those associated with
PILSRK(L,I) only if ν is greater than about 10. However, in terms of the σ(ν) values, the
PILSRK(T,Q) methods are also superior to the PILSRK(L,I) method for small numbers of inner
iterations, because in the case of PILSRK(T,Q), σ(ν) becomes finite for ν > 2, whereas PILSRK(L,I)
has infinite σ(ν)-values for all ν.

4. The Newton-PILSRK iteration process

In actual application of the Newton-PILSRK iteration process {(2.1),(2.2)}, the inner iteration
process will not always be iterated to convergence, so that the Newton iterates are only approximately
computed. This will affect the convergence and stability behaviour and the computational costs of
integration method.

4.1. Overall convergence and stability

The overall convergence of the Newton-PILSRK process is determined by the total number of inner
iterations summed over all outer iterations in one step, that is, the effective amplification factors
associated with the total iteration error Y(j,ν) - Yn are approximately given by ρ(i) and σ(i), where i
denotes the total number of inner iterations needed to compute Y (j,ν), i.e. i = (j-1)r + ν, and where r
denotes the number of inner iterations per outer iteration. In order to see this, we define

(4.1) Y(j,0) := Y(j-1,r), G(∆) := F(Y + ∆) - F(Y) - (I⊗J)∆, N := (I - A⊗hJ)-1(A⊗I).

By a simple manipulation we find that

(4.2) Y(j,ν) - Yn = Mν(Y(j-1,r) - Yn) + h(I - Mr)NG(Y(j-1,r) - Yn), j = 1, ... , m,

where M is defined in (3.1). Ignoring second-order terms, we may set G(Y(j-1,r) - Yn) = 0, to obtain

(4.3) Y(j,ν) - Yn = Mi(Y(0,r) - Yn), i := (j-1)r + ν .

16
From this relation, we see that in a first approximation, the convergence behaviour of the Newton-
PILSRK iteration process is approximately characterized by the amplification factors. As a
consequence, Table 3.2 applies if we replace ν by i.
A second feature of the overall performance of the integration method is its stability if the Newton
iterates are not exactly computed. This aspect has been discussed in [12], where the number of
iterations needed to achieve sufficient stability was computed. The values of mr for which the method
becomes and remains L-stable depends on the predictor used. For the extrapolation (EPL) predictor
defined by Y(0) = (P⊗I)Yn-1, where P is such that Y(0) has maximal order q = s-1, and the four-stage
and eight-stage Radau IIA corrector, these stable mr-values are listed in Table 4.1. In the case of the
four-stage corrector, the stable mr-values are acceptable for all three iteration strategies, but for the
eight-stage corrector, only the Newton-PILSRK(T,Q) method possesses an acceptable stable mr-
value.

Table 4.1. Stable values of mr for γ = 87 .

-----------------------------------------------------------
Iteration method s=4 s=8
-----------------------------------------------------------
PILSRK(D,I) 7 > 61
PILSRK(L,I) 4 > 43
PILSRK(T,Q) 5 14
-----------------------------------------------------------

Summarizing, we conclude that with respect to the Newton-PILSRK(D,I)-based integration method,

the Newton-PILSRK(T,Q) method always generates an integration method that has a superior
convergence and stability behaviour. With respect to the Newton-PILSRK(L,I)-based integration
method, we conclude that the Newton-PILSRK(T,Q) method:

(i) damps the stiff error components much stronger for i < s (Theorem 3.4),
(ii) has a better overall convergence for larger values of i (Table 3.2, with ν replaced by i),
(iii) is much more stable for the 8-stage corrector (Table 4.1).

4.2. Computational costs

In an actual implementation of the linear solver (2.2), we diagonalize (2.2) by a transformation
Y(j,ν) = (S⊗I)X(j,ν) to obtain

(4.4) (I - S-1BS⊗hJ)(X(j,ν) - X(j,ν-1)) = - (I - S-1AS⊗hJ)X(j,ν-1) + (S-1⊗I)C(j-1),

where the matrix S-1BS is diagonal. For the PILSRK(L,I) and PILSRK(T(γ≠1),Q) methods, the
matrices S-1BS and S corresponding to the 4-stage and 8-stage Radau IIA correctors are given in the
Appendix to this paper. In this Appendix, we also give a computer-program type description of the
Newton-PILSRK iteration process {(2.1),(2.2),(4.4)} and a specification of the computational costs
of the most important steps of the algorithm. Here, we present in Table 4.2 the total costs per step for
s-stage correctors where s is even. In this table, Cf and CJ denote the average costs of one component
of f and its Jacobian J, respectively.
 17

Table 4.2. Total computational costs per step.

-----------------------------------------------------------------------------------------------------------------------
1
Method 1 processor s processors s processors
2
-----------------------------------------------------------------------------------------------------------------------
PILSRK(L,I) sd( 23 d2 + ds CJ + d + 2s) 2d( 23 d2 + ds CJ + d + 2s) d( 23 d2 + ds CJ + d + 2s)
s s s
& + 4mrsd2(1 + 2d ) + 8mrd2(1 + 2d ) + 4mrd2(1 + 2d )
PILSRK(T(γ≠1),Q) + msd(s + Cf - 2d) + 2md(2s + Cf - 2d) + md(s + Cf - 2d)
-------------------- ----------------------------- ------------------------------ ----------------------------
sd( 13 d2 + ds CJ + d + 2s) d( 23 d2 + 2d CJ + 2d + 4s) d( 23 d2 + ds CJ + d + 2s)
s
2s 2s s
PILSRK(T(1),Q) + 5mrsd2(1 + 5d ) + 10mrd2(1 + 5d ) + 8mrd2(1 + 4d )
+ msd(2s + Cf - 2d) + 2md(2s + Cf - 2d) + md(2s + Cf - 2d)
-----------------------------------------------------------------------------------------------------------------------

The following conclusions can be drawn:

(i) Newton-PILSRK(L,I) and Newton-PILSRK(T(γ≠1),Q) are equally expensive,

(ii) If mr is fixed and d > s + 21 Cf, then the costs are minimized for r = 1,
(iii) Newton-PILSRK(L,I) and Newton-PILSRK(T(γ≠1),Q) are to be preferred on s processors,
whereas Newton-PILSRK(T(γ=1),Q) is to be preferred on one or on σ processors.

5. Numerical illustration
In this section, we compare the new Newton-PILSRK(T(7/8),Q) method with the Newton-
PILSRK(L,I) method. In our experiments, we use the EPL predictor defined in the preceding section
and either the 4-stage or the 8-stage Radau IIA corrector with constant stepsizes. We integrated three
test problems taken from the CWI test set [14]. In these problems, the initial condition was adapted
such that the integration starts outside the transient phase. The first test problem is provided by a
problem of Schäfer (called the HIRES problem in [7, p.157]). It consists of 8 mildly stiff nonlinear
equations on the interval [5,305]. The second test example is the Pollution problem of Verwer [20].
The ODE system consists of 20 highly stiff nonlinear ODEs on the interval [5,60], originating from
an air pollution model. Our third test problem, the Ring Modulator originating from circuit analysis,
is a highly stiff system of 15 equations on the interval [0,10-3], and is due to Horneber [9].
The tables of results present the minimal number of correct digits cd of the components of y at the
end point of the integration interval (i.e. at the end point, the absolute errors are written as
10-cd). Negative cd-values are indicated with *. Table 5.1 leads us to the following conclusions:

(i) For fixed values of m ≥ 3, the Newton-PILSRK methods always converge and usually find the
Newton iterate with high accuracy within two inner iterations (in the case of the 4-stage
corrector, we even have convergence for m ≥ 1).
18
(ii) Comparing results for fixed values of mr reveals that r = 1 is usually preferable (however, in an
actual implementation, m and r should both be determined dynamically, see also Remark 2.1).
(iii) For r ≤ 2 the Newton-PILSRK(T,Q) method is more robust than Newton-PILSRK(L,I),
particularly for the eight-stage corrector, and approximates the Newton iterate usually much
better (the better cd-values produced by Newton-PILSRK(L,I) in the Pollution problem for r = 2
and m = 3, 4 is due to 'overshoot' and does not mean that Newton-PILSRK(L,I) produces a
better approximation to the corrector solution). The divergent behaviour is due to the
development of instabilities for small values of mr (see Table 4.1).

Finally, we remark that for the relatively difficult Ring modulator problem, a parallel implementation
of the Newton-PILSRK(L,I) method on the four-processor Cray-C98 / 4256 shows a speed-up
ranging from at least 2.4 until at least 3.1 with respect to RADAU5 in one-processor mode (cf. [11]).
Since Newton-PILSRK(T(γ≠1),Q) is equally expensive as Newton-PILSRK(L,I), the same speed-
ups are expected for Newton-PILSRK(T(γ≠1),Q).

Table 5.1. Newton-PILSRK applied to HIRES with h = 15.

----------------------------------------------------------------------------------------------------------------------
4-stage Radau IIA corrector 8-stage Radau IIA corrector
Solver r m=1 m=2 m=3 m=4 m=10 m=20 m=1 m=2 m=3 m=4 m=10 m=20
----------------------------------------------------------------------------------------------------------------------
PILSRK(L,I) 1 * 3.0 4.8 5.1 7.3 7.9 * * * * 8.2 9.9
PILSRK(T,Q) * 4.5 4.9 5.3 7.7 7.9 * * * * 9.3 10.8
PILSRK(L,I) 2 * 4.3 4.9 5.3 8.1 7.9 * * * * 9.2 10.1
PILSRK(T,Q) 3.9 4.4 4.9 5.4 8.2 7.9 * * 5.5 7.0 9.3 10.8
PILSRK(L,I) 10 3.8 4.4 4.9 5.4 8.2 7.9 * * 5.6 7.0 9.4 10.3
PILSRK(T,Q) 3.8 4.4 4.9 5.4 8.2 7.9 * * 5.6 7.0 9.3 10.8
----------------------------------------------------------------------------------------------------------------------

Table 5.2. Newton-PILSRK applied to Pollution problem with h = 11.

----------------------------------------------------------------------------------------------------------------------
4-stage Radau IIA corrector 8-stage Radau IIA corrector
Solver r m=1 m=2 m=3 m=4 m=10 m=20 m=1 m=2 m=3 m=4 m=10 m=20
----------------------------------------------------------------------------------------------------------------------
PILSRK(L,I) 1 2.0 3.7 6.3 7.0 10.9 10.9 * * * * 10.3 10.3
PILSRK(T,Q) 1.1 5.3 6.9 7.3 10.9 10.9 * * * 6.7 12.0 12.6
PILSRK(L,I) 2 4.6 5.7 7.5 8.5 10.9 10.9 * * * 8.0 10.3 10.7
PILSRK(T,Q) 4.9 5.7 6.7 7.9 10.9 10.9 * 2.9 6.6 7.8 12.6 12.3
PILSRK(L,I) 10 4.6 5.7 6.8 7.9 10.9 10.9 * 4.8 6.7 7.8 11.0 10.9
PILSRK(T,Q) 4.6 5.7 6.8 7.9 10.9 10.9 * 4.8 6.7 7.8 12.5 12.5
----------------------------------------------------------------------------------------------------------------------
 19

Table 5.3. Newton-PILSRK applied to the Ring Modulator with h = 1.2510-7.

----------------------------------------------------------------------------------------------------------------------
4-stage Radau IIA corrector 8-stage Radau IIA corrector
Solver r m=1 m=2 m=3 m=4 m=10 m=20 m=1 m=2 m=3 m=4 m=10 m=20
----------------------------------------------------------------------------------------------------------------------
PILSRK(L,I) 1 * 5.7 7.8 8.5 10.2 10.2 * * * * 8.6 9.1
PILSRK(T,Q) * 7.3 8.4 9.7 10.2 10.2 * * * 10.5 10.8 11.1
PILSRK(L,I) 2 5.5 7.5 8.7 10.2 10.2 10.2 * * * 8.9 9.4 9.3
PILSRK(T,Q) 5.7 7.4 8.8 10.0 10.2 10.2 * 8.5 10.4 10.9 10.5 11.3
PILSRK(L,I) 10 5.8 7.4 8.8 9.9 10.2 10.2 * 8.5 9.0 8.9 9.0 9.2
PILSRK(T,Q) 5.8 7.4 8.8 9.9 10.2 10.2 * 8.5 10.4 11.1 11.3 10.6
----------------------------------------------------------------------------------------------------------------------

References
[1] Burrage, K. [1978]: A special family of Runge-Kutta methods for solving stiff differential
equations, BIT 18, 22-41.
[2] Burrage, K. [1995]: Parallel and sequential methods for ordinary differential equations,
Clarendon Press, Oxford.
[3] Burrage, K. & Chipman, F.H. [1989]: Construction of A-stable diagonally implicit multivalue
methods, SIAM J. Numer. Anal. 26, 391-413.
[4] Burrage, K. & Suhartanto, H. [1997]: Parallel iterated methods based on multistep Runge-Kutta
methods of Radau type, this issue, p. xx-yy.
[5] Butcher, J.C. [1976]: On the implementation of implicit Runge-Kutta methods, BIT 16, 237-
240.
[6] Golub, G.H. & Van Loan, C.F. (1983): Matrix computations, North Oxford Academic,
Oxford.
[7] Hairer, E. & Wanner, G. [1991]: Solving ordinary differential equations, II. Stiff and
differential-algebraic problems, Springer-Verlag, Berlin.
[8] Hoffmann, W. & Swart, J.J.B. de [1995]: Approximating Runge-Kutta matrices by triangular
matrices, Preprint NM-R9517, CWI, Amsterdam.
[9] Horneber, E.H. [1976]: Analysis of nonlinear RCLÜ-circuits by means of a mixed potential
function with a systematic representation of the nonlinear dynamic circuit analysis (German),
PhD thesis, University of Kaiserslautern.
[10] Houwen, P.J. van der, & Sommeijer, B.P. [1991]: Iterated Runge-Kutta methods on parallel
computers, SIAM J. Sci. Stat. Comput. 12, 1000-1028.
[11] Houwen, P.J. van der & Swart, J.J.B. de [1996]: Triangularly implicit iteration methods for
ODE-IVP solvers, to appear in SIAM. J. Sci. Comput.
[12] Houwen, P.J. van der, & Sommeijer, B.P. [1996]: CWI contributions to the development of
parallel Runge-Kutta methods, to appear in APNUM.
[13] Lioen, W.M. [1996]: On the diagonal approximation of full matrices, to appear in J. Comput.
Appl. Math.
20
[14] Lioen, W.M., Swart, J.J.B. de, & Veen, W.A. van der [1995]: Test set for IVP solvers,
available via WWW at URL: http: //www.cwi.nl/cwi/projects/IVPtestset.shtml.
[15] Nevanlinna, O. [1985]: Matrix valued versions of a result of Von Neumann with an application
to time discretization, J. Comput. Appl. Math. 12 & 13, 475-489.
[16] Nørsett, S.P. [1976]: Runge-Kutta methods with a multiple real eigenvalue only, BIT 16, 388-
393.
[17] Orel, B. [1993]: Parallel Runge-Kutta methods with real eigenvalues, Appl. Numer. Math. 11,
241-250.
[18] Reichel, L. & Trefethen, L.N. (1992): Eigenvalues and pseudo-eigenvalues of Toeplitz
matrices,Linear Algebra Appl. 162 / 164,153-185.
[19] Varga, R.S. [1962]: Matrix iterative analysis, Prentice Hall, Englewood Cliffs, N.J.
[20] Verwer, J.G. [1994]: Gauss-Seidel iteration for stiff ODEs from chemical kinetics, SIAM. J .
Sci. Comput. 15, 1243-1250.
 

    !#"%$'&)(+*

,.-0/21436587:949<;=-4>43@?#A;!5B9<;=C=36DFE/214;GC=HI5B/25JHIDK/214;G36L9 M6;=LN;=-O/P7:/23FHI-45QHID)RQ,TSKUWVJXZY[8\^]`_a7b-4>0R,.S)UWVcXZYedfYhgi_j\lk!_jm,.-
n Hb/21L;o/214Hp>45/214;G36/2;oqj7:/2;=5c5P7:/23F52DFE/214;'qP;=C=r4qP5236HI-

Yh]tsvu w<xPyzu0{}|~_€Yhƒ‚ „P† ‡=ˆsƒƒ‚ „P† ‡%w<x‰ˆŠ_Œ‹ sGYh]svuw<xBŽ'u{|i~_€Yh‚ „P† ‡%w<x‰ˆKsƒ‚‘„=w<x‰ˆ._

s  ‚‘„=w<x‰ˆ’ |<YŠu w<x Ž}{“]`_.”ZY• ‚ „–w<x‰ˆ _ ’ Y—˜{v]W_‰z™
w<x–š
›J;oqj;b\z™ ‹œYŠu w<x {Z]W_.•™ \ ‚ „P† ž2ˆ ‹Ÿ ‚ „–w<x‰ˆ \I ‚@ž2ˆ ‹œYŠu w<x {Z]W_.#Y• ™ _j\ ‚‘„jˆ ‹Ÿ ‚‘„2† ¡2ˆ \•™t‹¢YŠuJ{Z]W_‰ ‚¤£ˆ \
w<x w<x w<x
ZYT¥l_>4;=-4Hb/2;=5K/214;Q9 qP;=>436Co/2Hbq)Hb9<;oqj7:/2Hbq–\b¦§/214;-pr4L n ;oq¨HID4HIr`/2;oq)3@/2;oqj7:/236HI-45)7b-4>©8/214; -r4L n ;oq¨HID43F-4-4;oq3@/2;oq€7:/236HI-45=m
ª4Hbq'RQ,TSKUWVJXZY[8\^]`_ 7b-4>«RQ,TSKUWVJXZYedYhg‹ ¬
_j\lk!_j\i/214;!L7:/BqP3¤?0u w<x yzu365J>43F7b­IHI- 7bM®\DeHbq¯R,.S)UWVcXZYedfY
_j\lk!_j\i3@/c365
n M6HpC2°>43F7b­IHI- 7bM®\`Ac3@/21 f±« MFH–Aa;oqc/BqP3F7b- ­Ir4MF7:q n M6HpC2°W58C=HI-/P7b36-436-4­z3F>4;=-O/23FC–7bM<>43F7b­IHI- 7bM¨;=-/BqP36;=5–m
²³;G36L9 M6;=LN;=-/2;=>z/214365 qj;=C=r`qP5236HI-07b5–´

µ¶·¸º¹¼»0½I¾¿¶–ÀÂÁ'¶·}¸hÃÄ4Å<»½¾µp¶
ÆhÇOÈ É ·}ÊÀTËIÀPÌjÌPÌBÀ.Í
¸eÎϾ ÐKѓ·³½cÒÓIÔÖÕ×O¸ÃÄ`Å2دè¾4»+Ù`Ú
µ¯Û ¶.Ü ·³ÝÞ¸eµß Ä`Å ¾
¸eÎoàp¾ ÁNÛ ¶TÜ ·¸hÃKÄ`Ũ»0½I¾®µ¯Û ¶TÜ
ÆhÇOÈ)á ·}ÊÀTËIÀPÌjÌPÌ2À‰â
¸ Ç Ï¾ ã·ƒÁNÛlä2Ä`Å Ü Ò0Ù`¸ÃÄ`Å2å«»0½I¾®æÞ¸ºµ¯Ûlä2Ä`Å Ü ¾Ò+¸ºç³»½¾hÁ ß Ä4Å
Ü Ü
¸ Ç àp¾ Á é Ûlä‰è Å ·ƒÁ é ۑä2Ä`Å Ò«¸eÐKÑ8¾ é Ä`Å ã é ¸6ê6ë:ì<í ëIÓÓ¾
Ü Ü Ü
¸ Çpî ¾ Á é Ûlä‰è Å ·ƒÁ é ۑä2Ä`Å Ò«¸eÐKÑ8¾ é Ä`Å ¸ºã é Òï é è é Ä`Å ÙWÚ`Á é Ûlä‰è Å ¾ ¸6ê6ë:ì<í4ðPñ%ðjò¾
Ä`Å
ÆhÇÈaó ·³ËÀ‰ôOÀPÌPÌjÌ2ÀŠõ
¸eö®ÏI¾ ÷ø·}¸F½8Ò0ÃÄ4ÅBåÃz»+Ù`Ú ¾B¸FÁNÛlä‰è ùoÄ4Å Ü ÒNÁNۑä2Ä`Å Ü ¾Òã
Ü Ü
¸eöeྠÁ é ۑä‰è ù ·ƒÁ é ۑä‰è ù=Ä`Å Ò+¸ºÐ)Ñ8¾ é Ä4Å ÷ é ¸6ê6ë:ì<í ëIÓÓ¾
Ü Ü Ü
¸eö î ¾ Á é ۑä‰è ù ·ƒÁ é ۑä‰è ù=Ä`Å Ò+¸ºÐ)Ñ8¾ é Ä4Å ¸e÷ é Òï é è é Ä`Å ÙWÚ`Á é ۑä.è ù ¾ ¸6ê6ë:ì<í4ðPñ%ðjò¾
Ä4Å
úbûü
Á Ûlä Ü ·³Á ۑä.è ý Ü
¸ Çþ ¾ µÞۑä Ü ·¸hÃz»0½¾hÁNۑä Ü
úbûü
µ ß ·³µ¯Û¤ÿ Ü ÀÂÁ ß ·ÁNÛ¤ÿ Ü
úbûü
›J;oqj;b\
365!/21 ;#-pr4L n ;oq!HID 36-/2;=­bqj7:/236HI-5B/2;o9 5=m814;O7bC=H n 3F7b-“365f7b5252r4LN;=>ƒ/2H n ;r`9<> 7:/2;=> ;;oqPE /236LN;z5B/2;o9¨m
JHb/23FC=;G/21 7:/¯DºHbq¯RQ,TSKUWVJXZY[8\^]`_ 7b-4>R,.S)UWVcX#YedYhg“‹ ¬
_j\lk!_8/214;GL7:/BqP3¤?0u w<x ytu}365c> 3F7b­IHI- 7bM\ 5PHz/21 7:/JHI- ;!C–7b-
HIL36/ Y _)7b-4>Y_KDeHbq¼/214365)C–7b52;b\p36D4HI-4;9<;oqPDeHbqPL5 Y _)7b-4>Y  _KDeHbq7bM6M2m¼ª4HbqRQ,TSKU`VcXZYedY
_j\lk_A; HI-4M6E!-4;=;=>
 94qPHpC=;=5252Hbqj5/2H!9i;oqjDºHbqPLø/214;c [ Š>4;=C=HIL9<HI523@/236HI-45)36-Z97:qj7bM6M6;=M®\Ac14;oqP;  365¼/21 ;c-pr4L n ;oq)HID<C=HIL9 M6;€?tC=HI-2r4­7:/2;=>
;=36­I;= - b7bM6r ;Þ97b3@qP5=m¼›¯;oqP;'Aa;!7b5252r L;J/21 7: / !3658; ;=-<\52H   ‹  m
814;DºHIMFM6H–Ac3F-4­z/P7 n M6;=5¯M63652/8/214;fC=HI5B/25ÞHID)/214;LNHI5B/c3FL#9<Hbq2/P7b-/ 5B/2;o9 5ÞHID)/214365Þ7bM6­IHbqP3@/21 L! m ¯5 n ;=DeHbqP;b# \ "N365J/214;
>436LN;=-45236HI-zHID/214;¯94qPH n M6;=L$ m 814;Þ%7 ;oqj7b­I;C=HI5B/258HIDKHI-4;ÞC=HIL#9<HI-4;=-/ HID/214;¯qP36­I1&/ Š1 7b-4' > Š5236>4;JDºr4- Co/236HI
- (7b-4>HI-4;
;=-O/BqPEtHID3@/2) 5 7bC=H n 3F7b-0~07:qj;8>4;=-4Hb/2;=> n + E * ( 7b-4 > *#,`\Oqj;=5B9<;=Co/2-3 ;=M@EI.
m 814
; 7bC=H n 3F7b-Z365 7b5252r4LN;=>f/2H n ; Dºr4MFMm,T-z/214;
 I

/ qP5B/aC=HIM6r4LN-z/214;ÞC=HIL9 r`/P7:/236HI-/21 7:/81 7b5/2H n ;J9i;oqjDºHbqPLN;=>Z365aM63F5B/2;=><m.814;¯52;=C=HI-4>C=HIM6r L-Z­I30;=5 /214;Þ-r L n ;oq HID

1 H7:/236-4­t9<HI36-/8Hb9<;oqj7:/236HI-458qP;%2r 3@qP;=>DeHbqc/214365cC=HIL#9 r4/P7:/236HI-36DHI- M@ENHI- ;'94qPHpC=;=5252Hbqc3F5c7%I7b36MF7 n MF;bm)814;G5P;%2r4;=-/23F7bM
C=HI5B/25 HID</214;¯C=HIL9 r 7:/236HI-HI-  7b-4>3J94qPHpC=;=5252Hbqj5C–7b- n ;¯DºHIr4- >#36-/214;¯/2143@qP>N7b-4>NDºHIr`qP/21NC=HIMFr4L-¨\IqP;=5B9<;=Co/23-;=M@EIm
ª4HbqqP;–7b52HI-45HID<5236L9 M636C=36/ŠEI\bA;J>436>Z- Hb/;€?W9 M6HI3@//214;cM6HA;oq/BqP3F7b- ­Ir4M4DeHbqPL HIDi/21 ;cL7:/BqP3¤?zu 36-ZR,.S)UWVcX#Y[8\^]W_j\W-4Hbq
/214; n M6HpC2°>43F7b­IHI-7bM¨DºHbqjL HID/21 ;GLN7:/Bqj3¤?u w<x Ž'u36-RQ,TSKU`VcXZYedYhg<_j\lk!_jm

46587!9;:=<?>B[@¤]BADCE4F5879G:<H>.dfYhg“‹ ¬
_I@FkJALK

M Ç NPO#QBRTSUR öÇ û M Ç ÎR J
Î VW ÇXO ÎY
ë%òÊ!Zì‰ë\[PðI]^]Šë:ì ë%òJ_=ZOìŠëU[2ðT]`]‰ë%ì^] ë%òJaZì‰ëU[2ðT]`]‰ë%ì^]
¸eÎϾ aTbc¸ dc fb e Åg!hi e³Êo¾ ËLbjc=¸ cd bke Åghi e³Êo¾ bc¸ cd bfe Åg!hi e³Êo¾
¸eÎoàp¾ LË aTclb maTb ËLaTb
¸ Ç Ï¾ aTOb ¸hËLane h ( ¾ Ë8b¸eËjaoe h ( ¾ bO¸hËLane h ( ¾
¸ Ç àp¾ ¸-p'a=¾ ËLaTbc mjbc Ë8bc
¸eö®ÏI¾ ËLaTb ¸qbres=a ¾ mjb¸qbresa=¾ Ë8b¸qbresa=¾
¸eöeàO¾ ¸0p'a=¾ ËLaTbc mjbc Ë8bc
¸ Çþ ¾ aTclb maTb ËLaTb
tvuUwqx8yzB{}|w0~€6{rIw{`z aTbO¸ dc bc‚e„ƒg h i ePbk e Ëja=¾ Ë8b¸ cd bjc‚e„ƒg h i ePbre«ËLa=¾ Ob ¸ cd bc‚e„ƒg h i ePbkeËja=¾
efaTb âN¸eËLbk e ôja‚e h ( e Ë8b%â¸hË8b)e3ma‚e h (
a e bâN¸eËLbke3ma‚e h (
r
e ¸ºõ Ò+Êo¾B¸ mbr«
J e ËL=a ¾Š¾ e ¸ºõ Ò+Êo¾B¸ mbkeËja=¾Š¾
J eJ¸ºõ Ò+Êo¾B¸ mbkeËja=¾Š¾

4F5879G:<H>TdY
I_ @Fk†ALK
M Ç NPO#QBRTSUR öÇ û M Ç ÎR J
Î VW ÇXO ÎY
ë%òÊ!Zì‰ë\[PðI]^]Šë:ì ë%òJ_=ZOìŠëU[2ðT]`]‰ë%ì^] ë%òJaZì‰ëU[2ðT]`]‰ë%ì^]
¸eÎϾ aTbc¸ dÅ bfe Åg hi e³Êo¾ ËLbjc=¸ dÅ bke Åg hi e³Êo¾ bc¸ cd bfe Åg hi e³Êo¾
¸eÎoàp¾ LË aTclb maTb ËLaTb
¸ Ç Ï¾ aTOb ¸hËLane h ( ¾ Ë8b¸eËjaoe h ( ¾ bO¸hËLane h ( ¾
¸ Ç àp¾ aTbjc Ë8bc Ë8bc
¸ Çpî ¾ ËLaTbc mjbc mjbc
¸eö®ÏI¾ ËLaTb ¸qbres=a ¾ mjb¸qbresa=¾ Ë8b¸qbresa=¾
¸eöeྠaTbjc Ë8bc Ë8bc
¸eö î ¾ ËLaTbc mjbc mjbc
¸ Çþ ¾ ËLa c b maTb ËLaTb
tvuUwqx8yzB{}|w0~€6{rIw{`z aTbO¸ dÅ bc‚e„ƒg h i ePbk e Ëja=¾ Ë8b¸ dÅ bjc‚e„ƒg h i ePbre«ËLa=¾ Ob ¸ cd bc‚e„ƒg h i ePbkeËja=¾
efaTb âN¸ºôjbke3ma‚e h ( e Ë8b%â¸eôLb)e3ma‚e h (
a e bâN¸q‡jbke3ma‚e h (
r
e ¸ºõ Ò+Êo¾B¸‰ˆLbr«
J e ËL=a ¾Š¾ e ¸ºõ Ò+Êo¾B¸‰ˆLbkeËja=¾Š¾
J eJ¸ºõ Ò+Êo¾B¸qŠjbkeËja=¾Š¾
 \‹

Œr `ŽQ¼œ)#‘G’˜`I“‘
,.-f/2143F5K7:949<;=-4> 3¤?'A;a5B9<;=C=36DºE'/214;aL;o/21 H>!97:qj7bLN;o/2;oqP5HID4/214;aRQ,TSKUWVJXZY[8\^]`_7b-4>tR,.S)UWVcXZYedfY•”U\–_j\lk!_KLN;o/214H> 5
DºHbJ q #˜‹ —«7b- ? > #‹™–pm²³;#M63652/'/214;ZL7:/BqP36C=;=5!u w<x yzu˜7b-4>vua\Ac1 36CP1³7:qP;#-4;=;=>4;=> DeHbq!/21 ;Z36L9 M6;=L#;=-/P7:/236HI-«HID
DºHbqjLzr4MF7ZY —Wšm —O_jfm ¯5c7b>4>43@/23FHI- 7bM<36-4DeHbqPL7:/236HI-ZAa;'94qPHLp3F>4;Gy\`/214;'LN7:/Bqj3¤?#/21 7:/J7:9 94qPH=?`36L7:/2;=5aŽ!m

46587!9;:=<?>B[@¤]BALK

J‹›—
œLŸžIU¡ ¢£¤`¥r¢“¦?§ ¨ ©Lªš«T«&¬T© ©8ª ­l®T©T¯ ©Lª ¬I©T°T¬ ©Lªš«T«²±I³s´
¶·· «Iª ©T©T©I© © © © l» ¼¼
¢µ§ ¹ «Iª ¬­I©T¯ «Iª ©I©T©I© © ©
¸ ­%ªš«}¯I®lº ­¹ ±%ª ­I­I³I¬
T T« ª ©I©T©T© © ½
«¹ T«&®Lª °T®I°T° ¹ ³T³8ª °T­I³T¯ ­8ª ¬8«}¯I° «Tª ©T©I©T©
·¶· ©8ªš«T«²¬I© © © © » ¼¼
¥¾§ ©8ª ­I¬IºIº ©8ª ­l®T©T¯ © ©
¸ ©8ª ­8«²³T± ©8ª ºT°T¬lº ©Lª ¬I©T°T¬ © ½
©8ª ­T­I©T¯ ©8ª ºT³T³I° ©Lª ºIºj«&º ©Lªš«T«²±I³

J‹¿–
œLŸžlU¡ ¢#£¤^¥r¢“¦H§ ¨ ©8ª ©­I°I° ©8ª ©I°T³T¯ ©Lªš«&¬Iº%¯ ©Lªš«²³T­lº ©8ªÀ«&³¯}º ©Lªš«&º­T± ©8ª ©T®±l³ ©Lª ©I¬T©T° ´
¶·· «Iª ©T©T©I© © © © © © © © l» ¼¼
·· ¹ «Iª ©I³T®lº «Tª ©I©T©T© © © © © © © ¼¼
·· «Iª ©Iº°I³ ¬¹ Lª ­l¬¯lº T« ª ©T©I©T© © © © © © ¼¼
·· © ¼¼
¢Á§ ·· ¹ «Iª ©T±8«&° ±8ª ±l³T¬T³ ¹ °Lªš«T«&©L« T« ª ©T©I©T© © © ©
· «Iªš«²°¯I­ ¹ «²®Lª ©T­lº© ³­8ª ©8«²°­ ¹ °I°Lªš«T«²±T¯ «Iª ©I©T©I© © © © ¼¼¼
¹ «Iª ºT°T°T± ³­%ª ±l³¯l³ ©¹ 8ªÀ«²±T­l©IÂ&º ¹ ©Lªš«I«&ºj«^Â&º T« «Iª ¬I³T®lº T« ª ©T©I©T© © ©
¸ ­%ª º%±I©I° ®¹ I©T°8ª ºT°T°I® ¹ ©8ª ®T¯T­l³IÂ&º ¹ ©Lª ºT©±I©lÂ&º ¬T®8ª ­I¯T±l¬ ºjª ±I©T­I° I« ª ©T©T©I© © ½
°¹ T°8ª ­8«}¯}º ¹ ­8ª ©T©T±I¬IÂ&¬ ¹ «Iª ¯I¯I®I©IÂ&º ¹ ©Lª ³I©T®±}Â&º ¯I°8ª ¬T±T¯%« ±8ª º³I®T® ­%ª ©T©­I± «Tª ©I©T©T©
¶·· ©8ª ©­I°I° © © © © © © © » ¼¼
·· ©8ª ©T³L«²± ©Lª ©I°T³¯ © © © © © © ¼¼
·· ©8ª ©¯T¯l¬ ©Lªš«²¯T¯I¬ ©Lªš«²¬lº%¯ © © © © © ¼¼
·· ©8ª ©¯I°I¬ ©Lªš«^º%­lº ©Lª ­T­l³L« ©Lªš«²³T­lº © © © © ¼¼
¥Ã§ ·· ©8ª ©¯I³T± ©Lªš«^º°T¬ ©Lª ­8«&©T³ ©Lª ­I³8«²® ©8ªÀ«&³¯}º © © © ¼¼¼
· ©8ª ©¯T±I¯ ©Lªš«^º%¯lº ©Lª ­8«²±8« ©Lª ­lº±8« ©8ª ­I¯T±I­ ©Lªš«&º­T± © ©
¸ ©8ª ©¯T±%« ©Lªš«^º³± ©Lª ­8«^ºTº ©Lª ­T¯I­T­ ©8ª ­}º³I© ©Lª ­8«²­lº ©8ª ©T®±l³ © ½
©8ª ©¯T±l¬ ©Lªš«^º³T¬ ©Lª ­8«²¯8« ©Lª ­T¯%«²© ©8ª ­}º°I¬ ©Lª ­I©T±I¬ ©8ªš«²¬T¬I° ©Lª ©I¬T©T°
 —

4F5879G:<H>.dfY•”U\–_I@6kJALK

J‹›—
œ8ÄžlU¡ ¢£X¤^¥k¢B¦H§ ¨ ©Lªš«}¯I­8« ©8ªÀ«&®T°I³ ©Lªš«}±l¬± ©8ª ­T­I³I® ´
¶·· ­8ª ®¯I­I³ ©8ª ¬8«}¯l® «Tª ¯I¬T­T¯ ©8ª ©­T±l³ » ¼¼
¢Á§ ¹ ±8ª ­I³I³T¬ ¹ ©8ª °T±T¯l³ ¹ «Tª ©¯I¯I¬ ¹ ©8ª ¬8«T«&¬
¸ ¬Lª º%­l©­ ©8ª ®lº®I¬ ¹ «&©Lª ±I®I®± ¹ ­%ªÀ«&¬IºT® ½
¬lºjª °T®T±I© ºLª ¬T±T¯l¬ ¹ º­8ª ®T©I¬T® ¹ ¯%ª °I®T³I©
¶·· ©Lªš«²©I®T³ ¹ ©8ª ©lº¬I© ©Lª ©­l³T° ¹ ©8ª ©I©T°I© » ¼¼
¥Ã§ ©Lª ­I©I°¯ ©8ª ¬I©T³lº ¹ ©Lª ©T³T±8« ©8ª ©­8«I«
¸ ©Lª ­lºT°Iº ©8ª ©I°­l¬ ©Lª ­T¯I±I¬ ¹ ©8ª ©8«&º­ ½
©Lª ­T¯l®T³ ¹ ©8ª ©T¯8«²¯ ©Lª º%­%«²® ©8ª ©±I°I©

J‹¿–
œLŸžIU¡ ¢ £X¤ ¥k¢“¦?§ ¨ ©Lª ©I³±I® ©Lª ©T°I°T³ ©8ª ©T±I³I° ©Lªš«²©I©T¬ ©8ª ©T°­l¬ ©Lªš«&©±lº ©8ª ©T°IºT® ©Lªš«I«²©T® ´
¶·· ©Lªš«&ºT¬T© ©8ª ©L«&ºT® ©Lª ©I©¯8« ¹ ©Lª ©T©8«²¬ ¹ ©8ª ©T­I©I° ¹ ©Lª ©T©T­I® ©8ª ©L«²°I© ¹ ©Lª ©I©T©L« » ¼¼
·· ¹ ©Lª ­I³I³± ¹ ©8ª ©­I°lº ¹ ©Lª ©I¬T©T³ ¹ ©Lª ©T©I©T³ ©8ª ©8«²®T¯ ©Lª ©T©I¬Iº ¹ ©8ª ©L«²°T­ ¹ ©Lª ©I©T©­ ¼¼
·· ©Lª º°lº° ©8ª ©¯lºT© ©Lª ©I®L«}¯ ©Lª ©T©I°T¬ ¹ ©8ª ©I©¯l© ¹ ©Lª ©T©8«²© ©8ª ©­I©T¯ ¹ ©Lª ©I©T©T° ¼¼
·· ©Lª °T°I°L« ©8ªš«²©T³T¯ ¼¼
¢µ§ ·· ¹ ¹ ¹ ©Lª ©I¬±T­ ¹ ©Lª ©T©I®T® ©8ª ©I®±I¯ ©Lª ©L«&©L« ¹ ©8ª ©L«T«²­ ¹ ©Lª ©I©±T­ ¼¼
· «Tªš«²¬T­I³ ©8ªš«²³­l° ¹ ©Lª ®I©Iº° ¹ ©Lª ©T®I®T³ ©8ª ­l¬Iº± ¹ ©Lª ©T©T¯I© ¹ ©8ªš«T«²©T­ ¹ ©Lª ©T¯T­T­ ¼
«Tª ³T³I©T¬ ©8ªš«T«²©T¯ ¹ ©Lª ­l³T°L« ©Lª ©T®I¬T¬ ¹ «Iª ©8«}­I¯ ¹ ©Lª ­lºT°L« ¹ «Iª ¬T°T¬lº ¹ ©Lª ­l°­I³
¸ ¹ ¯8ª ®T©T­T¯ ¹ ©8ª ±T¯I¬I® ±8ª ³8«²©T° «Tª ©­I¯lº ¹ ±%ª ¬lº³T± ¹ «Tª ©L«²­I° ¹ ³8ª ¬T¬T³T± ¹ ©Lª °I®T°L« ½
¹ °Lª ®T°T­I° ¹ ©8ª ®T®±l° «&ºLªÀ«&³T©I® «Tª ³T¬I³T© ¹ «&ºLªÀ«&°T°I³ ¹ «Tª ³±l¬T© ¹ «}­%ª ­I°L«&© ¹ «Tª ºT°T®±
¶·· ©Lª ©¯l©± ¹ ©8ª ©­I³lº ¹ ©Lª ©8«&º%± ¹ ©Lª ©T©T±T± ©8ª ©I©T³8« ¹ ©Lª ©T©I¬Iº ©8ª ©T©­I­ ¹ ©Lª ©I©T©T° »l¼¼
·· ©Lª ©­l®¯ ©8ª ©T°¯l³ ©Lª ©8«}¯I¬ ©Lª ©L«&³­ ¹ ©8ª ©8«²©lº ©Lª ©T©T¯I® ¹ ©8ª ©T©T¬T± ©Lª ©I©L«&º ¼¼
·· ©Lª ©¯%«²¬ ©8ªš«²¬±I­ ©Lª ©I®¯T­ ¹ ©Lª ©T¬8«&º ©8ª ©8«}±l© ¹ ©Lª ©T©I®T³ ©8ª ©T©¯l® ¹ ©Lª ©I©­T­ ¼¼
·· ©Lªš«²³I©L« ©8ª ©Iº%¯I¯ ©Lª ©I³T³­ ©Lªš«&º¯I° ¼¼
¥Ã§ ·· ¹ ©8ª ©I¬Iº­ ©Lª ©­l©L« ¹ ©8ª ©L«}­I± ©Lª ©I©Iº°
¼¼
· ©Lª ­I©T±T­ ©8ª ©­T¯l¬ ©Lª ©T¯I³T® ©Lª ©IºT³­ ©8ªÀ«^º³I© ¹ ©Lª ©T¬8«}­ ©8ª ©L«²¬8« ¹ ©Lª ©I©T¬Iº ¼
©Lª ­lºT®¯ ¹ ©8ª ©L«}¯%« ©Lª ©T¯I®T© ©Lª ©L«&°¯ ©8ªÀ«^º³8« ©Lª ©­l©­ ©8ª ©T³T¬lº ¹ ©Lª ©T­I³­
¸ ©Lª ­T¯l³T° ¹ ©8ª ©­I°8« ©Lª ©I®­I¬ ¹ ©Lª ©L«²¯I® ©8ª ©lº©T¯ ©Lª ©IºL«²° ©8ª ­I©T®T¯ ¹ ©Lª ©I³T°T° ½
©Lª ­I³T¯I¬ ¹ ©8ª ©T¬­I¯ ©Lª ©I°±I¬ ¹ ©Lª ©T®T­lº ©8ªÀ«&©T®T­ ©Lª ©IºT®T® ©8ª ­8«²®I© ¹ ©Lª ©I¬Iº©