Cambridge University Press & Assessment

978-0-521-80615-2 — Collocation Methods for Volterra Integral and

Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

1
The collocation method for ODEs:
an introduction

A collocation solution u h to a functional equation (for example an ordinary

differential equation or a Volterra integral equation) on an interval I is an
element from some finite-dimensional function space (the collocation space)
which satisfies the equation on an appropriate finite subset of points in I (the
set of collocation points) whose cardinality essentially matches the dimension
of the collocation space. If initial (or boundary) conditions are present then u h
will usually be required to fulfil these conditions, too.
The use of polynomial or piecewise polynomial collocation spaces for the
approximate solution of boundary-value problems has its origin in the 1930s.
For initial-value problems in ordinary differential equations such collocation
methods were first studied systematically in the late 1960s: it was then shown
that collocation in continuous piecewise polynomial spaces leads to an impor-
tant class of implicit (high-order) Runge–Kutta methods.

1.1 Piecewise polynomial collocation for ODEs

1.1.1 Collocation-based implicit Runge–Kutta methods
Consider the initial-value problem

y (t) = f (t, y(t)), t * I := [0, T ], y(0) = y0 , (1.1.1)

and assume that the (Lipschitz-) continuous function f : I × ¢ IR ³ IR is

such that (1.1.1) possesses a unique solution y * C 1 (I ) for all y0 * . Let

Ih := {tn : 0 = t0 < t1 < . . . < t N = T }

be a given (not necessarily uniform) mesh on I , and set Ãn := (tn , tn+1 ], Ã̄n :=
[tn , tn+1 ], with h n := tn+1 2 tn (n = 0, 1, . . . , N 2 1). The quantity

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

2 1 The collocation method for ODEs: an introduction

h := max{h n : 0 f n f N 2 1} will be called the diameter of the mesh

Ih ; in the context of time-stepping we will also refer to h as the stepsize. Note
that we have, in rigorous notation,
tn = tn(N ) , Ãn := Ãn(N ) , h n = h (N )
n (n = 0, 1, . . . , N 2 1), and h = h
(N )
.
However, we will usually suppress this dependence on N , the number of subin-
tervals corresponding to a given mesh Ih , except occasionally in the conver-
gence analyses where N ³ >, h = h (N ) ³ 0, so that N h (N ) remains uni-
formly bounded.
The solution y of the initial-value problem (1.1.1) will be approximated by
an element u h of the piecewise polynomial space
Sm(0) (Ih ) := {v * C(I ) : v|Ã̄n * Ãm (0 f n f N 2 1)}, (1.1.2)
where Ãm denotes the space of all (real) polynomials of degree not exceeding
m. It is readily verified that Sm(0) (Ih ) is a linear space whose dimension is
dim Sm(0) (Ih ) = N m + 1
(a description of more general piecewise polynomial spaces will be given in
Section 2.2.1). This approximation u h will be found by collocation; that is, by
requiring that u h satisfy the given differential equation on a given suitable finite
subset X h of I , and coincide with the exact solution y at the initial point t = 0.
It is clear that the cardinality of X h , the set of collocation points, will have to
be equal to N m, and the obvious choice of X h is to place m distinct collocation
points in each of the N subintervals Ã̄n . To be more precise, let X h be given by
X h := {t = tn + ci h n : 0 f c1 < . . . < cm f 1 (0 f n f N 2 1)}. (1.1.3)
For a given mesh Ih , the collocation parameters {ci } completely determine X h .
Its cardinality is
|X h |

Nm if 0 < c1 < . . . < cm f 1 (or 0 f c1 < . . . < cm < 1),
=
N (m 2 1) + 1 if 0 = c1 < c2 < . . . < cm = 1 (m g 2).
The collocation solution u h * Sm(0) (Ih ) for (1.1.1) is defined by the collocation
equation
u h (t) = f (t, u h (t)), t * X h , u h (0) = y(0) = y0 . (1.1.4)
If u h corresponds to a set of collocation points with c1 = 0 and cm = 1 (m g 2),
it lies (if it exists on I ) in the smoother space Sm(0) (Ih ) + C 1 (I ) =: Sm(1) (Ih ) of
dimension N (m 2 1) + 2 whenever the given function f in (1.1.1) is contin-
uous. This follows readily by considering the collocation equation (1.1.4) at

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

1.1 Piecewise polynomial collocation for ODEs 3

t = tn21 + cm h n21 =: tn2 and at t = tn + c1 h n =: tn+ : taking the difference and

using the continuity of f leads to

u h (tn+ ) 2 u h (tn2 ) = 0, n = 1, . . . , N 2 1,

and this is equivalent to u h being continuous at t = tn .

In order to obtain more insight into this piecewise polynomial collocation
method, and to exhibit its recursive nature, we now derive the computational
form of (1.1.4). This will reveal that the collocation equation (1.1.4) represents
the stage equations of an m-stage continuous implicit Runge–Kutta method for
the initial-value problem (1.1.1) (compare also the original papers by Guillou
and Soulé (1969), Wright (1970), or the book by Hairer, Nørsett and Wanner
(1993).
Here, and in subsequent chapters of the book, it will be convenient (and
natural) to work with the local Lagrange basis representations of u h . Since
u h |Ãn * Ãm21 , we have
m
u h (tn + vh n ) = L j (v)Yn, j , v * (0, 1], Yn, j := u h (tn + c j h n ), (1.1.5)
j=1

where the polynomials

m
v 2 ck
L j (v) := ( j = 1, . . . , m),
k= j
c j 2 ck

denote the Lagrange fundamental polynomials with respect to the (distinct)

collocation parameters {ci }. Setting yn := u h (tn ) and
 v
³ j (v) := L j (s)ds ( j = 1, . . . , m),
0

we obtain from (1.1.5) the local representation of u h * Sm(0) (Ih ) on Ã̄n , namely
m
u h (tn + vh n ) = yn + h n ³ j (v)Yn, j , v * [0, 1]. (1.1.6)
j=1

The unknown (derivative) approximations Yn,i (i = 1, . . . , m) in (1.1.6) are

defined by the solution of a system of (generally nonlinear) algebraic equations
obtained by setting t = tn,i := tn + ci h n in the collocation equation (1.1.4) and
employing the local representations (1.1.5) and (1.1.6). This system is
 
m
Yn,i = f tn,i , yn + h n ai, j Yn, j , (i = 1, . . . , m), (1.1.7)
j=1

where we have defined ai, j := ³ j (ci ).

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

4 1 The collocation method for ODEs: an introduction

We see that the equations (1.1.6) and (1.1.7) define, as asserted above, a
continuous implicit Runge–Kutta (CIRK) method for the initial-value prob-
lem (1.1.1): its m stage values Yn,i are given by the solution of the nonlinear
algebraic systems (1.1.7), and (1.1.6) defines the approximation u h for each
t * Ã̄n (n = 0, 1, . . . , N 2 1). This local representation may be viewed as the
natural interpolant in Ãm on Ã̄n for the data {(tn , yn ), (tn,i , Yn,i ) (i = 1, . . . , m)}.
It thus follows that such a continuous implicit RK method contains an em-
bedded ‘classical’ (discrete) m-stage implicit Runge–Kutta method for (1.1.1):
it corresponds to (1.1.6) with v = 1,
m
yn+1 := u h (tn + h n ) = yn + h n b j Yn, j (n = 0, 1, . . . , N 2 1), (1.1.8)
j=1

with b j := ³ j (1), and the stage equations (1.1.7).

If m g 2 and if the collocation parameters {ci } are such that
0 = c1 < c2 < . . . < cm = 1,
then tn,1 = tn implies Yn,1 = f (tn , yn ), and the CIRK method (1.1.6), (1.1.7)
reduces to
m
u h (tn + vh n ) = yn + h n ³1 (v) f (tn , yn ) + h n ³ j (v)Yn, j , v * [0, 1],
j=2
(1.1.9)
and
 
m
Yn,i = f tn,i , yn + h n ai,1 f (tn , yn ) + h n ai, j Yn, j (i = 2, . . . , m).
j=2
(1.1.10)
Moreover, since cm = 1, we obtain
 
m
Yn,m = f tn+1 , yn + h n b1 f (tn , yn ) + h n b j Yn, j .
j=2

Example 1.1.1 u h * S1(0) (Ih ) (m = 1), with c1 =: » * [0, 1]:

Since L 1 (v) c 1 and ³1 (v) = v (hence a1,1 = » and b1 = 1), (1.1.6) reduces to
u h (tn + vh n ) = yn + h n vYn,1 , v * [0, 1],
with Yn,1 defined by the solution of
Yn,1 = f (tn + » h n , yn + h n » Yn,1 ).
These equations may be combined into a single one (by setting v = 1 in the
expression for u h (tn + vh n ) and solving for Yn,1 ); the resulting method is the

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

1.1 Piecewise polynomial collocation for ODEs 5

continuous »-method for (1.1.1),

u h (tn + vh n ) = (1 2 v)yn + vyn+1 , v * [0, 1].

where

yn+1 = yn + h n f (tn + » h n , (1 2 »)yn + » yn+1 )

implicitly defines yn+1 .

This family of continuous one-stage Runge–Kutta methods contains the
continuous implicit Euler method (» = 1) and the continuous implicit mid-
point method (» = 1/2). For » = 0 we obtain the continuous explicit Euler
method. Due to its importance in the time-stepping of (spatially) semidiscre-
tised parabolic PDEs (or PVIDEs) we state the continuous implicit midpoint
method for the linear ODE

y (t) = a(t)y(t) + g(t), t * I,

with a and g in C(I ). Setting » = 1/2 we obtain

hn
yn+1 = yn + a(tn + h n /2)[yn + yn+1 ] + g(tn + h n /2)(n = 0, 1, . . . , N 21),
2
or, using the notation tn+1/2 := tn + h n /2,
hn hn
12 a(tn+1/2 ) yn+1 = 1 + a(tn+1/2 ) yn + h n g(tn+1/2 ). (1.1.11)
2 2
Observe the difference between (1.1.11) and the continuous trapezoidal
method: the latter corresponds to collocation in the space S2(0) (Ih ), with
c1 = 0, c2 = 1 being the Lobatto points; it is described in Example 1.1.2 below
(m = 2).

Example 1.1.2 u h * S2(0) (Ih ) (m = 2), with 0 f c1 < c2 f 1:

It follows from L 1 (v) = (c2 2 v)/(c2 2 c1 ), L 2 (v) = (v 2 c1 )/(c2 2 c1 ) that
v(2c2 2 v) v(v 2 2c1 )
³1 (v) = , ³2 (v) = .
2(c2 2 c1 ) 2(c2 2 c1 )
Hence, b1 = ³1 (1) = (2c2 21)/(2(c2 2c1 )), b2 = ³2 (1) = (122c1 )/(2(c2 2c1 )).
The resulting continuous two-stage Runge–Kutta method thus reads:

u h (tn + vh n ) = yn + h n {³1 (v)Yn,1 + ³2 (v)Yn,2 }, v * [0, 1],

where

Yn,i = f (tn,i , yn + h n {ai,1 Yn,1 + ai,2 Yn,2 }) (i = 1, 2).

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

6 1 The collocation method for ODEs: an introduction

We present three important special cases:

r Gauss points c = (3 2 :3)/6, c = (3 + :3)/6:
1 2
We obtain
: :
³1 (v) = v(1 + 3(1 2 v))/2, ³2 (v) = v(1 2 3(1 2 v))/2,

and
" : " " "
1 1 3 1
2
" "
b
4 4 6
b= 1 2
" "
A := ai, j = : , = 1
.
1
+ 3 1 b2
4 6 4 2

The discrete version of this two-stage implicit Runge–Kutta–Gauss method

(of order 4; cf. Section 1.1.3, Corollary 1.1.6) was introduced by Hammer and
Hollingsworth (1955) and generalised by Kuntzmann in 1961 (see Ceschino
and Kuntzmann (1963) for details).
r Radau II points c = 1/3, c = 1:
1 2
Here, we have

³1 (v) = 3v(2 2 v)/4, ³2 (v) = 3v(v 2 2/3)/4,

and
" " " "
5 1 3
12
2 12 4
A= 3 1
, b= 1
.
4 4 4

This represents the continuous two-stage Radau IIA method.

r Lobatto points c = 0, c = 1 (=ó u * S (1) (I )):
1 2 h 2 h
The continuous weights are now

³1 (v) = v(2 2 v)/2, ³2 (v) = v 2 /2,

and hence
" " " "
1
0 0 2
A= 1 1
, b= 1
.
2 2 2

This yields the continuous trapezoidal method: it can be written in the form
hn !
v(2 2 v)Yn,1 + v 2 Yn,2 , v * [0, 1],
!
u h (tn + vh n ) = yn +
2
with

Yn,1 = f (tn , yn ), Yn,2 = f (tn+1 , yn + (h n /2){Yn,1 + Yn,2 }).

(See also Hammer and Hollingsworth (1955).)

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

1.1 Piecewise polynomial collocation for ODEs 7

For the linear ODE y (t) = a(t)y(t) + g(t) the stage equation assumes the
form
h n a(tn+1 ) h n a(tn ) h n a(tn+1 )
12 Yn,2 = 1+ a(tn+1 )yn + g(tn ) + g)tn+1 ).
2 2 2
Remark Other examples of (discrete) RK methods based on collocation, in-
cluding methods corresponding to the Radau I points (c1 = 0, c2 = 2/3 when
m = 2), may be found for example in the books by Butcher (1987, 2003),
Lambert (1991), and Hairer and Wanner (1996).
There is an alternative way to formulate the above continuous implicit
Runge–Kutta method (1.1.6),(1.1.7). Setting
m
Un,i := yn + h n ai, j Yn, j (i = 1, . . . , m),
j=1

we obtain the symmetric formulation

m
u h (tn + vh n ) = yn + h n ³ j (v) f (tn, j , Un, j ), v * [0, 1], (1.1.12)
j=1

with
m
Un,i = yn + h n ai, j f (tn, j , Un, j ) (i = 1, . . . , m). (1.1.13)
j=1

Here, the unknown stage values Un,i represent aproximations to the solution
y at the collocation points tn,i (i = 1, . . . , m). For v = 1, (1.1.12) yields the
symmetric analogue of (1.1.8),
m
yn+1 = yn + h n b j f (tn, j , Un, j ); (1.1.14)
j=1

if cm = 1 we have yn+1 = Un,m .

For later reference, and to introduce notation needed later, we also write down
the above CIRK method (1.1.6), (1.1.7) for the linear initial-value problem
y (t) = a(t)y(t), t * I, y(0) = y0 ,
where a * C(I ). Setting A := (ai, j ) * L(IR m ), ³(v) := (³1 (v), . . . , ³m (v))T *
IR m , and Yn := (Yn,1 , . . . , Yn,m )T * IR m , the CIRK method can be written in
the form
u h (tn + vh n ) = yn + h n ³ T (v)Yn , v * [0, 1], (1.1.15)
with Yn given by the solution of the linear algebraic system
[Im 2 h n An ]Yn = diag(a(tn,i ))e · yn (n = 0, 1, . . . , N 2 1). (1.1.16)

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

8 1 The collocation method for ODEs: an introduction

Here, Im denotes the identity in L(IR m ), An := diag(a(tn,i ) )A, and e :=

(1, . . . , 1)T * IR m .
The derivation of the analogue of (1.1.15),(1.1.16) corresponding to the sym-
metric formulation (1.1.12),(1.1.13) of the CIRK method is left as an exercise
(Exercise 1.10.1).
The classical conditions for the existence and uniqueness of a solution y *
1
C (I ) to the initial-value problem (1.1.1) (see, e.g. Hairer, Nørsett and Wanner
(1993, Sections I.7–I.9) assure the existence and uniqueness of the collocation
solution u h * Sm(0) (Ih ) to (1.1.1) or its linear counterpart for all h := max(n) h n
in some interval (0, h̄), provided that f y is bounded (or a and g lie in C(I )
when the ODE is y = a(t)y + g(t)). In the latter case, the existence of such
an h̄ follows from the Neumann Lemma which states that (Im 2 h n An )21 is
uniformly bounded for all sufficiently small h n > 0, so that h n ||An || < 1 for
some (operator) matrix norm. We shall give the precise formulation of this
result in in Chapter 3 (Theorem 3.2.1) for VIDEs which contains the version
for ODEs as a special case.
It is clear that not every implicit Runge–Kutta method can be obtained by
collocation as described above (see, for example, Nørsett (1980), Hairer, Nørsett
and Wanner (1993)): a necessary condition is clearly that the parameters ci
are distinct. The framework of perturbed collocation (Nørsett (1980), Nørsett
and Wanner (1981); see also Section 1.2 below) encompasses all implicit)
Runge–Kutta methods. There is also an elegant connection between continuous
Runge–Kutta methods and discontinuous collocation methods (Hairer, Lubich
and Wanner (2002, pp. 31–34)). The following result (which can be found in
Hairer, Nørsett and Wanner (1993, p. 212)) characterises those implicit Runge–
Kutta methods that are collocation-based.

Theorem 1.1.1 The m-stage implicit Runge–Kutta method defined by (1.1.7)

and (1.1.8), with distinct parameters ci and order at least m, can be obtained
by collocation in Sm(0) (Ih ), as described above, if and only if the relations

m
ci¿
ai, j c¿21
j = , ¿ = 1, . . . , m (i = 1, . . . , m),
j=1
¿

hold.

The proof of this result is left as an exercise. Recall that a (discrete) Runge–
Kutta method for (1.1.1) is said to be of order p if

|y(t1 ) 2 y1 | f Ch p

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

1.1 Piecewise polynomial collocation for ODEs 9

for all sufficiently smooth f = f (t, y) in (1.1.1). The next section will reveal
that the collocation solution u h * Sm(0) (Ih ) to (1.1.1) is of global order p g m
on I .

1.1.2 Convergence and global order on I

Suppose that the collocation equation (1.1.4) defines a unique collocation so-
lution u h * Sm(0) (Ih ) for all sufficiently small mesh diameters h * (0, h̄). What
are the optimal values of p¿ and p¿7 (¿ = 0, 1) in the (global and local) error
estimates
||y (¿) 2 u (¿) (¿) (¿)
h ||> := sup |y (t) 2 u h (t)| f C ¿ h
p¿
(1.1.17)
t*I

and
7
||y (¿) 2 u (¿) (¿) (¿) p¿
h ||h,> := max |y (t) 2 u h (t)| f C ¿ h , (1.1.18)
t*Ih \{0}

respectively? These values depend of course on the regularity of the solution

y of the initial-value problem (1.1.1). For arbitrarily regular y we will refer
to the largest attainable p¿ (¿ = 0, 1) as the (optimal) orders of global (super-)
convergence (on the interval I ) of u h and u h , respectively, and the corresponding
p¿7 will be called the (optimal) orders of local superconvergence (at the mesh
points Ih \ {0}) of u h and u h , provided p¿7 > p¿ .
In order to introduce the essential ideas underlying the answer to the above
question regarding the optimal orders, we first present the result on global
convergence for the linear initial-value problem
y (t) = a(t)y(t) + g(t), t * I, y(0) = y0 . (1.1.19)
Theorem 1.1.2 Assume that
(a) the given functions in (1.1.19) satisfy a, g * C m (I );
(b) the collocation solution u h * Sm(0) (Ih ) for the initial-value problem (1.1.19)
corresponding to the collocation points X h is defined by (1.1.15), (1.1.16);
(c) h̄ > 0 is such that, for any h * (0, h̄), each of the linear systems (1.1.16)
has a unique solution.
Then the estimates
||y 2 u h ||> := max |y(t) 2 u h (t)| f C0 ||y (m+1) ||> h m (1.1.20)
t*I

and
||y 2 u h ||> := sup |y (t) 2 u h (t)| f C1 ||y (m+1) ||> h m , (1.1.21)
t*I

© in this web service Cambridge University Press & Assessment www.cambridge.org

Cambridge University Press & Assessment
978-0-521-80615-2 — Collocation Methods for Volterra Integral and
Related Functional Differential Equations
Hermann Brunner
Excerpt
More Information

10 1 The collocation method for ODEs: an introduction

hold for h * (0, h̄) and any X h with 0 f c1 < . . . < cm f 1. The constants C¿
depend on the collocation parameters {ci } but are independent of h, and the
exponent m of h cannot in general be replaced by m + 1.

Proof Assumption (a) implies that y * C m+1 (I ) and hence y * C m (I ). Thus

we have, using Peano’s Theorem (Corollary 1.8.2 with d = m) for y on Ã̄n ,
m
(1)
y (tn + vh n ) = L j (v)Z n, j + h m
n Rm+1,n (v), v * [0, 1], (1.1.22)
j=1

with Z n, j := y (tn, j ). The Peano remainder term and Peano kernel are given by
 1
(1)
Rm+1,n (v) := K m (v, z)y (m+1) (tn + zh n )dz, (1.1.23)
0

and
m
1
K m (v, z) := (v 2 z)m21
+ 2 L k (v)(ck 2 z)m21
+ , v * [0, 1].
(m 2 1)! k=1

Integration of (1.1.22) leads to

m
y(tn + vh n ) = y(tn ) + h n ³ j (v)Z n, j + h m+1
n Rm+1,n (v), v * [0, 1],
j=1
(1.1.24)
where
 v
(1)
Rm+1,n (v) := Rm+1,n (s)ds
0

(see also Exercise 1.10.3).

Recalling the local representation (1.1.6) of the collocation solution u h on
Ã̄n , and setting En, j := Z n, j 2 Yn, j , the collocation error eh := y 2 u h on Ã̄n
may be written as
m
eh (tn + vh n ) = eh (tn ) + h n ³ j (v)En, j + h m+1
n Rm+1,n (v), v * [0, 1],
j=1
(1.1.25)
while
m
(1)
eh (tn + vh n ) = L j (v)En, j + h m
n Rm+1,n (v), v * (0, 1], (1.1.26)
j=1

© in this web service Cambridge University Press & Assessment www.cambridge.org