two-point boundary value problems - computer science at rpi
TRANSCRIPT
Chapter
Fundamental Problems and Methods
Problems to be Solved
Several problems arising in science and engineering are modeled by dierential equations
that involve conditions that are specied at more than one point Some examples follow
PP
s θ
1
Figure Deformation of an elastica
r
u
u
u
z
Ω
θ
r
z
Figure Swirling ow above a rotating disk
Deformation of an Elastica The transverse deformation of a thin elastic inexten
sional rod subjected to an axial loading and clamped at its ends is governed by the
dierential systemd
ds P sin s
As shown in Figure the rod has unit length the magnitude of the loading
is P and is the angle that the deformed rod makes with the initial undeformed
axis This classical secondorder nonlinear twopoint boundary value problem is
called the elastica problem One solution is This solution however becomes
unstable as P increases and the rod bends into a deformed shape as shown in Figure
Hence this boundary value problem is also a dierential eigenvalue problem
that consists of determining theta and the critical load P for deformed shapes to
exist Once has been determined the Cartesian coordinates of a deformed point
on the rod can be determined as the solution of the initial value problems
dx
ds cos
dy
ds sin
x y
Swirling Flow The swirling ow of a viscous incompressible uid over a disk
spinning with speed Figure can be analyzed by solving the nonlinear
twopoint boundary value problem
df
dx f
df
dx
df
dx g
dg
dx f
dg
dx
df
dxg x
f df
dx g
limx
df x
dx lim
xg x
where is the Rosby number The dimensionless variable x is related to the axial
coordinate z Figure by
x z
r
where is the kinematic viscosity The radial tangential and axial components of
the velocity vector respectively are obtained from the functions f x and g x as
ur rdf x
dx u rg x uz f x
p
This problem involves a boundary value problem for a pair of secondorder nonlinear
ODEs An interesting feature of this problem is that one of the boundaries is at
innity
As indicated in these two examples boundary value problems BVPs have several
forms The two that will be most important to us are
A vector system of rstorder equations
y x f xy x
where y dydx and y and f are mdimensional vectors We have changed the
independent variable from t to x since it frequently corresponds to a spatial position
rather than time
A scalar m thorder dierential equation
um g x u u um
Naturally the higherorder scalar problem can be reduced to a rstorder vector
system as described in Section however it may sometimes be convenient to work
with the higherorder scalar problem
Focusing on the vector system for the moment if
f xy A xy b x
the ODE is linear otherwise it is nonlinear
If is to be solved on a x b then m conditions are needed to uniquely
determine the solution of the BVP These may be of the general form
g y ay b a
where g has dimension m For the most part we will consider simpler boundary condi
tions having the form
gL y a gR y b b
where gL has dimension l and gR has dimension r m l Conditions of the form
b are called separated while the more general form a are unseparated
Linear versions of a and b have the forms
Ly a Ry b c a
and
LLy a cL RRy b cR b
The matrices L and R are of dimension mm and the vector c is mdimensional For
the separated conditions b LL is lmdimensional RR is rmdimensional cL
is ldimensional and cR is rdimensional
There are three standard numerical approaches to solving twopoint boundary value
problems
Shooting An appropriate IVP is dened and solved by initial value techniques and
software The IVP is dened so that solutions iteratively converge to the solution
of the original BVP
Finite Dierences A mesh is introduced on a b and derivatives in the ODE are
replaced by nitedierences relative to the mesh This leads to a linear or nonlinear
algebraic problem which may be solved to produce a discrete approximation of the
solution of the BVP
Projections The solution of the BVP is approximated by simpler functions eg
piecewise polynomials and the dierential equations and boundary conditions are
satised approximately Collocation or nite element techniques often furnish these
approximations
In the next three Sections we illustrate the basic ideas of these methods by using
simple BVPs
Introduction to Shooting
Let us consider a secondorder nonlinear twopoint BVP
u x f x u u a x b u a A u b B
Writing the ODE as a rstorder system let us also consider the related IVP
y y y a A a
y f x y y y a b
In what follows well need to emphasize the dependence of the solution on the parameter
appearing in the initial conditions so well write the solution of as yk x
k
Solutions of satisfy the original dierential equation and the initial condition
at x a but fail to satisfy the terminal condition at x b Thus shooting consists of
repeatedly solving for dierent choices of until the terminal condition
y b B
is also satised Regarding as a nonlinear function of convergence to the
solution of the BVP can be enhanced by using an iterative strategy for nonlinear algebraic
equations Secant and Newton iteration are two possible procedures Let us illustrate
the simpler secant method rst
Solve the IVP for two choices of say and The corresponding
solutions y x and y x may appear as illustrated in Figure The
various choices of alter the initial slope y a y a Regarding the solution
y x as the trajectory of a projectile red from a cannon at x a y a A
the problem is to alter the initial angle of the cannon so that the projectile hits a
target at at x b y a B hence the name shooting
Assuming that y b is locally a linear function of we use the two values
y b and y b to compute the next value in the sequence thus is the
B
A
b x
y
a
y (b;
y (b;
1
1
1α )
α )1
0
α1
α0
Figure Solutions y x and y x of the IVP y (b;
α α
α)
0 1
1
y (b;1
y (b;1
α )
α )
0
1
B
xα2
Figure Secant method of using two guesses and to select another guess
such that y b B
solution of Figure
y bB
y b y b
Solving for yields
y b B
y b y b
This can be repeated to yield the general relation
y b B
y b y b
The iteration may be terminated when eg
y b B
B
for a prescribed value of Other termination criteria should be used when B
Remark If the ODE is linear then y b is a linear function of and y x is
the exact solution neglecting round o errors of the BVP
The nonlinear equation can also be solved by Newtons method If for exam
ple is a suciently close guess to the solution of then the next guess may
be generated as
y bByb
An expression for y b may be obtained by dierentiating the IVP with
respect to to obtain
y
y
y a
a
y
f x y y
y
y
f x y y
y
y
y a
b
These equations are linear in the partial derivatives y and y
An algorithm for performing shooting with Newtons method is shown in Figure
In order to simplify the notation let
z y
z y
In order to solve the IVP functions to evaluate fy and fy must be available
In contrast the secant method only requires knowledge of f In fact the secant method
can be viewed as an approximation of Newtons method with backward
procedure newtonbegin
Select an initial guess repeat
Solve the IVP for x a by y y a Ay f x y y y a z z z a z fy x y yz fy x y yz z a
if not converged then
begin
yb Bzb
end
until convergedend
Figure The shooting method for solving with Newton iteration
dierences replacing y b For secondorder BVPs Newtons method requires
the solution of a fourdimensional IVP while the secant method only requires a two
dimensional IVP
Convergence of Newtons method is generally secondorder quadratic ie
j j Cj j
where is the value of that satises the terminal condition Convergence of
the secant method is slightly slower typically
j j Cj j
Thus the secant method would generally be preferred to Newtons method This how
ever may not be the case with higherdimensional BVPs
Example Consider the solution of the clamped elastica problem
P sin
by shooting methods using Newton iteration Symmetry considerations have been used
to cut the domain of the problem illustrated in Figure in half
Letting
y y
we introduce the IVP
y y y
y P sin y y
Dierentiating this system with respect to yields
z z z
z Pz cos y z
where zk k satises Iterates are computed by the relation
y
z
Using a convergence test of
jy j
we found that Newtons method converged in ve iterations when P and
Introduction to Finite Dierence Methods
Well again use the secondorder scalar nonlinear twopoint boundary value problem
y x f x y y a x b y a A y b B
to describe the essential details of nite dierence methods
To begin we divide the domain a x b into N uniform subintervals of width
h b a
N a
as shown in Figure Restriction to uniform subintervals is not essential but is
introduced here for simplicity We also let
xi a ih i N b
x x
x
a = x
y
y = A
h
y(x )
yy
N0 1x = b
0
1
1
1N
y = B
i
Figure Domain discretization and notation used for nite dierence solutions of
In solving the BVP by nite dierences all derivatives are replaced by nite
dierence approximations These can be constructed from interpolating polynomials but
well illustrate a dierent approach using Taylors series expansions of the solution y x
Thus consider
y x y xi x xiy xi
x xi
y xi
x xik
kyk
where is between xi and x Specically choosing x xi xi h yields
y xi y xi hy xi h
y xi
h
y xi
hk
kyk i a
Similarly selecting x xi xi h produces
y xi y xi hy xi h
y xi h
y xi
hkk
yk i b
Setting k in a and solving for y xi yields
y xi y xi y xi
h h
y i a
Finite dierence approximations are obtained by neglecting the error term of the Taylors
series thus the rst forward nite dierence approximation of y xi is
yi yi yi
h b
and the local discretization error of this approximation is
i h
y i c
Subscripts on y denote nite dierence approximations hence yi denotes an approxima
tion of y xi etc
In a similar manner the rst backward dierence approximation of y xi is obtained
by setting k in b
yi yi yi
h i
h
y i
Notice however that a higherorder and symmetric dierence approximation can be
obtained by subtracting b from a and setting k to get
y xi y xi hy xi h
y i
Solving for y xi
yi yi yi
h i h
y i
The dierence formula is called the rst central dierence approximation of y xi
In Chapter we found that this approximation led to the leap frog scheme which had
poor stability characteristics Here with secondorder ODEs central dierences will
generally be preferred to either forward or backward dierences because of their higher
order local discretization errors
Remark The higherorder accuracy of relative to or only
occurs on a uniform mesh With nonuniform spacing the second derivative terms in
a and b would not cancel upon subtraction
A central dierence approximation of the second derivative y xi is obtained by
adding a and b while setting k to obtain
yi yi yi yi
h i h
yiv i
No further approximations are needed to solve by nite dierences however
we note that approximations of higher derivatives are obtained by using Taylors series
at more points For example consider evaluating the Taylors series at x xi
and xi to obtain
y xi y xi hy xi hy xi h
y xi
h
yiv xi a
and
y xi y xi hy xi hy xi h
y xi
h
yiv xi b
Subtracting b from a yields
y xi y xi hy xi h
y xi O h
A similar subtraction of b from a yields
y xi y xi hy xi h
y xi O h
Elimination of the rst derivative term yields a central dierence approximation of the
third derivative as
yi yi yi yi yi
h
The local discretization error i O h
Similar combinations of and yield an O h central dierence approxi
mation of the fourth derivative as
yivi yi yi yi yi yi
h
Now let us return to the task of solving by nite dierence approximations
Well try centraldierence approximations because of their higher order Thus evaluat
ing at x xi and replacing derivatives by central dierences using a and
a we obtain
yi yi yih
f xi yiyi yi
h a
Writing a at each interior mesh point i N and using the two
boundary conditions
y A yN B b
gives a system of N nonlinear algebraic equations in the N unknowns yi i
N This system is too complex for an introductory example so let us conne
our attention to linear problems with
f x y y p xy q xy r x
In this case the approximation a becomes
yi yi yih
piyi yi
h qiyi ri i N
where pi p xi etc Referring to this as the centraldierence approximation of the
ODE we dene the local discretization error as follows
Denition Consider an ODE in the form Ly x and let Lhy be a dierence
approximation of it with L and Lh being dierential and dierence operators The local
discretization error or the local truncation error at x xi is
i Lhy xi
Example The dierential and dierence operators for the linear ODE
satisfy
Ly x y p xy q xy r x
and
Lhy xi y xi y xi y xi
h p xi
y xi y i
h q xiy xi r xi
Using and we nd
i y xi h
yiv i p xiy
xi h
y i q xiy xi r xi
Using the dierential equation
i h
yiv i p xi
h
y i
Thus as we might have expected the local discretization of the central dierence ap
proximation of is O h
The algebraic system b is linear for the linear BVP b
and may be solved by eg Gaussian elimination Towards this end let us write
in the form
biyi aiyi ciyi hri i N a
where
bi h
pi ai hqi ci
h
pi b
The boundary conditions b may be used in conjunction with a to create
a system of dimension N or used to explicitly eliminate y and yN as unknowns from
a The latter approach is preferred for simple problems like this one Thus using
b with a when i we nd
ay cy hr bA a
Similarly using b with a when i N yields
bNyN aNyN hrN cNB b
Grouping the N equations a b and a i N
yields
Ay f a
where
A
a cb a c
bN aN cNbN aN
b
y
yy
yN
f
hr bAhr
hrNhrN cNB
c
The linear algebraic problem requires the solution of a tridiagonal system to
determine the N unknowns yi i N The basic solution strategy is
Gaussian elimination Pivoting is frequently unnecessary As seen from b A will
be diagonally dominant unless jp xj is large relative to jq xj or h is too small Pivoting
and other special treatment may be necessary in these exceptional situations Let us
proceed without pivoting and write A in the slightly more general form
A
a cb a c
bn an cnbn an
a
where in our case n N We factor A as
A LU b
where L is a lower triangular matrix and U is an upper triangular matrix Having
performed this factorization we write a as
LUy f
let
Uy z c
and solve
Lz f d
Since L is lower triangular d may be solved for z by forward substitution Know
ing z we determine y by solving c by backward substitution All that remains is
the determination of L andU Let us hypothesize that they have the following bidiagonal
forms
L
n
U
n nn
Using with ab we nd
a a
i bii b
i ai ii i n c
i ci i n d
Using a and d we nd the forward substitution step to be
z f a
zi fi izi i n b
Similarly using b and c the backward substitution step is
yn znn a
yi zi iyii i n n b
The solution procedure dened by is the basis of the famous tridiag
onal algorithm We state it as a pseudoPASCAL algorithm in Figure The version
implemented in Figure overwrites ai bi and ci with i i and i to reduce stor
age By counting we see that the algorithm requires approximately n multiplications or
divisions and n additions and subtractions The work required to factor a full matrix
by Gaussian elimination is approximately n Thus the ratio of the work to factor
a tridiagonal matrix to that of a full matrix is approximately n This is a signif
icant savings even for small matrices and one should never use a Gaussian elimination
procedure for full matrices to solve a tridiagonal system
Example Evidence from the Taylors series expansion would suggest that the
global error of the nite dierence solution of the BVP b has an
Procedure tridi n integer var ab c f y array n of realbegin
f Factorization gfor i to n do
begin
bi biaiai ai bici
end
f Forward substitution gy ffor i to n do yi fi biyi
f Backward substitution gyn ynanfor i n downto do yi yi ciyiai
end
Figure Tridiagonal algorithm
expansion in even powers of h beginning with O h terms Lets assume that this is the
case Then Richardsons extrapolation can be used to both estimate the global error
and to improve the solution To this end we calculate two solutions using dierent step
sizes of eg h and h In order to emphasize the dependence of the discrete solution
on step size let yhi denote the nite dierence solution yi at xi a ih obtained with
step size h With the assumed error dependence we have
y a ih yhi Ch O h
and
y a ih yhi C
h
O h
The variable yhi is the nite dierence solution at a i h a ih
Subtracting the two error equations to eliminate the exact solution yields
Ch
y
hi yhi O h
Using this result we estimate the error in the ner grid solution as
y a ih yhi y
hi yhi
Furthermore
yhi y
hi
yhi yhi
i N
is an O h approximation of the solution
Let us apply Richardsons extrapolation to the simple BVP
y y y y
This problem has the form of with p x r x and q x Thus the
elements of the tridiagonal system are
ai h i N
bi i N ci i N
Centraldierence solutions with h the solution by Richardsons extrapo
lation and the exact solution
y x sinhx
sinh
are shown in Table Using the error at x as a measure of accuracy we have
jy y j
jy y j
jy y j
jy y j
These results indicate that
the global error of the centered nite dierence solution is approximately O h
since decreasing h by onehalf quarters the error and
Richardsons extrapolation furnishes a good approximation of the error while also
improving accuracy
i xi y xi yhi yhi yhi
Table Solution of Example using central dierence approximations andRichardsons extrapolation
As a next step let us consider a linear BVP with a prescribed Robin boundary
condition eg
y p xy q xy r x a x b a
y a A b
y b Cy b B c
As distinct from the Dirichlet conditions used in y b is now an unknown We
could approximate y b by backward dierences and use the discrete version of the
terminal condition c to determine an approximation of y b however this has
some drawbacks If rstorder backward dierences were used to approximate y b
xa = x
y
y = A
h
0 1
0
x = bN
x
xxN+1N-1
Figure Domain and discretization used to approximate a Robin terminal condition
then the boundary condition c would only be accurate to O h while the discrete
approximation of the ODE a is accurate to O h If higherorder backward
dierences were used to approximate y b then the tridiagonal structure of the discrete
system would be lost
The usual strategy is to introduce a ctitious external point xN b h as shown
in Figure Extending the solution to this exterior point we use central dierences
to approximate the terminal condition c to O h as
yN yNh
CyN B a
This does little to solve the problem since weve introduced both another equation
a and another unknown yN The additional equation that we need is the central
dierence approximation of the ODE a at x xN Thus using a with
i N we have
bNyN aNyN cNyN hrN b
Once again It is common to eliminate yN by combining a and b to
obtain
bN cNyN aN hCcNyN hrN hcNB c
Observing that bN cN by use of b we obtain the tridiagonal system
a cb a c
bN aN cN aN hCcN
yyyN
hr bAhr
hrNhrN hcNB
which may be solved by the tridiagonal algorithm of Figure
Now let us return to the original nonlinear problem Most iterative schemes
for solving nonlinear algebraic equations can be used to determine the solution but well
illustrate the use of Newtons method which is the most popular To begin let us write
the nite dierence system a in the form
Fi y yi yi yi hf xi yiyi yi
h i N
subject to the Dirichlet boundary conditions b and the denition of the vector of
unknowns y given by c
Newtons iteration involves solving
Fy y y y F y a
where
F y
F yF y
FN y
Fy y
Fy
Fy
FyN
Fy
Fy
FyN
FN
y
FN
y FN
yN
b
Dierentiating the Jacobian Fy y is
Fi
yj
hfy
if j i
h fy if j i
hfy
if j i
otherwise
Letting
bi
h
f
y xi y
i
yi y
i
h a
ai h
f
y xi y
i
yi y
i
h b
ci h
f
y xi y
i
yi y
i
h c
gives
Fy y
a c
b a
c
bN a
N c
N
bN aN
d
Each Newton iteration requires the solution of a tridiagonal system The Jacobian of
this system need not be reevaluated and factored after each Newton step thus only the
forward and backward substitution steps of the tridiagonal algorithm shown in Figure
need be performed at each iterative step The derivatives fy and fy can
be approximated by nite dierences
Convergence of Newtons method is typically quadratic except at a bifurcation point
where it is often linear The use of nite dierence approximations in the Jacobian also
slows the convergence rate
Example Consider the elastica problem described in Section and repeated
here using the notation of this Section as
y P sin y y y
Hence
f x y y P sin y
f
y P cos y
f
y
and
bi c
i a
i hP cos y
i
Using the convergence criteria that
jjF yjj maxiN
jFi yj
and setting P we found the number of Newton iterations and solution at x
to be as recorded in Table The number of Newton iterations decreases as the mesh
becomes ner This is a result of the solution appearing to be smoother The convergence
rate seems to be nearly quadratic
h K i yi
Table Number of iterations K to reach convergence and the approximate solutionat x for Example
The development and description of nite dierence equations may be simplied by
introducing a set of nite dierence operators as shown in Table
The next several examples illustrate some applications of these nite dierence oper
ators
Example The centered dierence formula can be expressed in terms of
the central dierence and averaging operators and as
yih
yi yi
h
yi yih
Example An operator raised to a positive integer power is iterated eg
yi yi yi yi yi yi
Thus the centered second dierence approximation of the second derivative can
be written as
yi yih
Example Expanding y xi in a Taylors series about xi yields
y xi y xi hy xi h
y xi
h
y xi
Using the derivative operator D dened in Table
y xi hD h
D y xi
Operator Symbol Denition
Forward Dierence yi yi yi
Backward Dierence r ryi yi yi
Central Dierence yi yi yi
Average yi yi yi
Shift E Eyi yi
Derivative D Dyi yi
Table Denition of nite dierence operators
This suggests the shorthand operator notation
Ey xi y xi ehDy xi
where E is the shift operator Table We thus infer the identity between the shift
exponential and derivative operators
E ehD
Additional relationships can be obtained by noting that yi E yi which implies
that E or E Using this with gives
hD lnE ln
a
where the series expansion of ln x jxj has been used A similar relation in terms
of the backward dierence operator can be constructed by noting that r E thus
hD lnE ln r r
r
r b
These identities can be used to derive highorder nite dierence approximations of
rst derivatives For example suppose that we retain the rst two terms in a
ie
hDyi
yi
or
hDyi yi yi yi yi yi
or
Dyi yi yi yih
This formula can be veried as an O h approximation of y xi
Example Let us use with h replaced by h to obtain
y xi eh
Dy xi y xi e
h
Dy xi
Subtracting the two formulas and using the central dierence operator gives
y xi eh
D e
h
Dy xi sinh
h
Dy xi
Thus
sinhh
D
or
hD sinh
which can be used to construct central dierence approximations of y xi
Example We can square cube etc relations and to construct
approximations of second third etc derivatives For example squaring gives
hDy xi hy xi
y xi
At some point these formal manipulations would have to be veried as being correct
and estimates of their local discretization errors would have to be obtained Nevertheless
using the formal operators of Table provides us with a simple way of developing
highorder nite dierence approximations
Introduction to Collocation Methods
Unlike nite dierence methods projection methods such as collocation give a continuous
approximation of the solution as a function of x The basic idea is to approximate the
solution y x of a BVP by a simpler function Y x and then determine Y x so that it
is the best approximation of y x in some sense Two reasonable choices for Y x are
a discrete Fourier series
Y x MXk
ckeikx
and a polynomial
Y x MXk
ckxk
It is convenient to regard the BVP solution y x as an element of an innitedimensional
function space and Y x as an element of an M dimensional subspace of it Thus as
suming that y x has continuous second derivatives on a x b we would write
y x C a b which is read y x is an element of the space of functions that have
continuous second derivatives on a b Then Y x SM C a b where the space
SM consists of those C functions having a prescribed form The chosen functions eg
eikx or xk k M comprise a basis for SM
In order to introduce some concepts well again focus on the secondorder nonlinear
scalar BVP After selecting a basis the coordinates ck k M
can be determined by eg the least squares technique
minY sM
Z b
a
R xdx
where R x is the residual
R x Y f x Y Y
In this case it is clear that Y x is the best approximation of y x in the sense of
minimizing the square of the integral of the residual
Using Galerkins method we determine Y x so that the residualR x is orthogonal
to every function in SM ie
Z b
a
w xR xdx w x SM
The optimality of this procedure is not clear however since Galerkins method is pri
mariliy used with partial dierential equations we will not pursue it further
Collocation has been shown to be a successful procedure tor twopoint BVPs and is
the one on which we focus Collocation consists of satisfying
R i i M
with
a M b
The optimality of collocation is also not clear but well pursue this elsewhere
Global approximations such as the Fourier series and the polynomials introduced
above lead to illconditioned algebraic problems It is far better to use piecewise poly
nomial approximations that result in sparse and wellconditioned algebraic systems It
is also unwise to infer more continuity than necessary Discontinuous and continuous
piecewise polynomial approximations might have the forms shown in Figure The
discontinuous polynomial on the left has jumps at xi i N Thus the
rst derivative doesnt exist at these points and this would be an unsuitable function to
approximate the solution of a secondorder ODE The continuous approximation on the
right has jumps in its rst derivative at xi i N and its second derivative
doesnt exist at these points Hence minimally Y x C a b In this case the rst
derivative of Y x is continuous and the second derivative is piecewise continuous
x
Y(x) Y(x)
x x
x
x x x
x
0 1 0 1 NN
Figure Discontinuous left and continuous right piecewise polynomial functionY x having jumps left and jumps in its rst derivative right at xi i N
Perhaps the simplest way of satisfying the continuity requirements is to select a basis
for SM that includes them For example a basis for a space of piecewise constant
functions could be chosen as
i x
if x xi xi otherwise
i N
The approximation Y x would then be written in the form
Y x NXi
cii x
and is shown in Figure In this case the dimension of the subspace M and the
number of subintervals N are identical however this need not be so
x x0
x
xN
x x1 1ii-1
c
c
c1
2
Nφ (x)
10
1
Y
Figure Piecewise constant basis element i x and the resulting piecewise constantapproximation Y x
Using and we see that
Y x ci x xi xi
We may interpret ci as Y xi however this is not necessary As shown in Figure
the basis i x Ca b thus Y x Ca b
Of course the basis doesnt satisfy any continuity requirements however it
can be used to generate a continuous basis More generally a piecewisepolynomial basis
whose continuity increases with increasing degree can be constructed by integrating a
linear combination of basis elements of a piecewisepolynomial space having one degree
less than the desired degree For example let us construct a C piecewiselinear basis by
integrating i and i as
i x
Z x
x
i s i sds
The appropriate continuity will be automatically obtained by the integration The result
for this example is
i x
if x xi x xi if xi x xihi x xi if xi x xi
hi hi if xi x
where
hi xi xi
The parameters and are at our disposal Let us pick them so that the resulting
approximation has compact support ie so that
i x x xi
Let us furthermore normalize the basis so that
i xi
Then we nd
i
hi i
hi
and the nal result
i x
xxihi
if xi x xixixhi
if xi x xi
otherwise
The approximation Y x has the form
Y x NXi
cii x
x x0
x x1 1ii-1
1
x
xN
cN
c1
c0
xi+1
ci
φ (x)11
Y(x)
Y
Figure Piecewiselinear basis element i x and the resulting piecewiselinear approximation
The basis element i x and the piecewiselinear approximation Y x are shown in Figure
Using we see that i xj ij where ij is the Kronecker delta Using this
with yields Y xi ci i N The restriction of Y x to the subinterval
xi xi is the linear function
Y x cii x ci
i x x xi xi
Well continue by constructing a C piecewisequadratic polynomial basis as
i x
Z x
x
i s i sds
Proceeding in three steps we see that
Z x
x
i sds
if x xixxi
hi if xi x xi
hi hi
xix
hi if xi x xi
hi hi
if xi x
and
i x
if xi x
xxi
hi if xi x xi
hi hi xix
hi xxi
hi if xi x xi
hi hi hi hi xix
hi if xi x xi
hi hi hi hi if xi x
Enforcing the condition that i x x xi implies
hi hi
hi hi
Thus
i x
xxi
hihihi if xi x xi
xix
hihihi xxi
hihihi if xi x xi
xix
hihihi if xi x xi
otherwise
a
Normalizing the result so that i xi yields
hi
hi hi hi
hi hi
b
−1 −0.5 0 0.5 1 1.5 20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure Quadratic spline basis element i x relative to a mesh with xi xi xi and xi
The basis fi xgNi denes a quadratic spline The element i is shown in Figure
Evaluating ab at a node xj yields
i xj
hihihi
if j i hi
hihi if j i
otherwise
c
The basis simplies when the mesh is uniform thus if hi h i N
i x
xxih
if xi x xi xix
h xxi
h if xi x xi
xixh
if xi x xi
otherwise
d
The quadratic spline approximation is written in the form
Y x NXi
cii x a
and its restriction to xi xi is
Y x
c x c
x x x x
cii x ci
i x ci
i x x xi xi i N
cNN x cN
N x x xN xN
b
Thus three elements of the spline basis are nonzero on any interval except the rst and
the last where there are only two nonzero elements
Using with b we see that
Y xi
chhh
if i cihicihi
hihi if i N
cNhNhNhN
if i N
c
One could take h h and hN hN
When solving BVPs or interpolation and approximation problems its convenient to
introduce an extra basis element and unknown at each end of the domain and write Y x
as
Y x
NXi
cii x
Now there are three nonzero basis elements on every subinterval For interpolation prob
lems the coecients ci i N may be determined by satisfying
Y xi f xi i N
This yields N equations for the N unknowns The extra two equations could be
obtained by
interpolating f x at x and xN
prescribing f x at x and xN or
prescribing Y x Y xN
Regardless of the boundary prescription this interpolation problem requires the solution
of a tridiagonal linear algebraic problem to determine ci i N
Its possible to continue in this manner introducing more continuity with increasing
polynomial degree however usually well want approximations having the minimum al
lowable continuity Since higherdegree approximations increase the convergence rate of
smooth solutions we seek alternative spline constructions that do not increase smooth
ness with increasing polynomial degree Such procedures exist however for the
moment well examine piecewise cubic Hermite approximation Thus let us consider a
function of the form
Y x NXi
cii x di
i x
The basis fi x i xgNi is constructed to satisfy the following conditions
i x and i x are piecewise cubic polynomials on x xN
i x i x C x xN
i x and i x are nonzero only on xi xi and
i xi and i xi i N
These conditions imply that
i xj ij i xj i j N a
i xj
i xj ij i j N b
Together and give Y xi ci and Y xi di i N
Given these requirements the piecewise cubic Hermite basis is
i x
xxihi
xxihi
if xi x xi xxi
hi xxi
hi if xi x xi
otherwise
a
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
Figure Cubic Hermite polynomial basis elements i x solid and i x dashed
relative to a mesh with xi xi and xi
i x
x xi xxihi
if xi x xi x xi xxi
hi if xi x xi
otherwise
b
Representative basis elements are illustrated in Figure Examining we see
that there are four nontrivial basis elements i i i and i on the subinterval
xi xi i N
Having constructed appropriate piecewise polynomial approximations let us use them
to dene a collocation method by satisfying at a prescribed number of points per
subinterval and typically satisfying the boundary conditions Thus
Y ij f ij Y ij Y ij j J i N a
Y a A Y b B b
As indicated there are J collocation points ij j J per subinterval For
the piecewise quadratic spline approximation we would determine the N
unknowns ci i N by collocating at J point per subinterval and satisfying
the boundary conditions b With the piecewise cubic Hermite polynomial
we would determine the N unknowns ci di i N by collocating at J
point per subinterval and satisfying b
Example Well develop the collocation equations when piecewise quadratic
splines are applied to a linear problem with f x y y given by For
simplicity well assume that the mesh is uniform with spacing h Utilizing and
the boundary conditions b are
Y a A
c c Y b B
cN cN a
Selecting the sole collocation point
i xi xi xi
i N
at the center of each subinterval Figure and using and we have
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Figure Piecewise quadratic spline basis used for collocation on a mesh with spacingh on
Y xi
ci ci ci
Y xi
h ci ci
and
Y xi
h ci ci ci
Substituting these results into a while using we nd
h ci ci ci
pih
ci ci qi
ci ci ci ri
i N b
These dierence equations are similar but not identical to the central nite dierence
equations The relationship can be made more precise by rewriting in
terms of the nodal unknowns Y xi and comparing this result with Problem
The system may be written in a tridiagonal form as
N N N N N
cccNcN
ArrNB
a
where
N
i
h qi i N b
i
h pi
hqi
i N N
c
i
h
pih
qi
i N d
As a simple numerical example lets suppose p q r A and
B This is the problem of Example which has the exact solution
y x sinh x
sinh
Solution and pointwise at xi i N error data are presented in Table
for N The maximum pointwise errors for collocation solutions computed with
i xi Y xi y xi Y xi
E E E E E E E E E E E
Table Solution and pointwise errors for Example using collocation with piecewise quadratic splines at the center of each of ten subintervals
N kY yk NkY yk
Table Maximum pointwise errors for the solution of Example using collocationwith piecewise quadratic splines at the center of each of N subintervals
N and subintervals are presented in Table Solutions appear to be
converging as O h Pointwise errors are about half of those found using central nite
dierences Table
Example In the previous example it seemed natural to place the single collo
cation point at the center of each subinterval Were we to used piecewise cubic Hermite
approximations however we would have two collocation points per subinterval Placing
them at the ends of each subinterval is one possibility however our work with implicit
RungeKutta methods Section would suggest that the GaussLegendre points give
a higher rate of convergence DeBoor and Swartz showed that this is the case and we
will repeat this analysis in Chapter however for the moment let us assume it so and
consider the collocation solution of Chapter
y y
x
x x
y y
N jjY yjj
Table Maximum pointwise errors for the solution of Example using collocationwith piecewise cubic Hermite polynomials at two GaussLegendre points per subinterval
which has the exact solution
y x ln
x
The solution is smooth on x but the coecient p x x is unbounded at
x This would lead to problems with nite dierence or shooting methods but not
with collocation methods that collocate at points other than subinterval ends Ascher
et al solve this problem by a variety of techniques Well report their results using
piecewise cubic Hermite polynomials with collocation at the two GaussLegendre points
i
xi xi hp
i
xi xi
hp
per subinterval The maximum pointwise errors are presented in Table A simple
calculation veries that the solution is converging as O N
Problems
Construct the basis for a cubic spline approximation that is of class C and has
support on xi xi by integrating the quadratic splines i x and i x of
and imposing appropriate normalization and compact support conditions
For simplicity assume that the mesh is uniform with spacing h
Using c show that
Y xi
ci ci
on a uniform mesh of spacing h Use this to rewrite the quadraticspline collocation
equations in terms of Y xi i N instead of ci i N
Compare the results with the nite dierence equations