A modification of Talbot's method for the ... - ACM Digital Library

3 downloads 0 Views 901KB Size Report
A Modification of Talbot's Method for the. Simultaneous Approximation of Several. Values of the Inverse Laplace Transform. MARIAROSARIA. RIZZARDI.
A Modification of Talbot’s Method for the Simultaneous Approximation of Several Values of the Inverse Laplace Transform MARIAROSARIA

RIZZARDI

Istituto Universitario

Navale

In recent years many results have been obtained in the field of the numerical inversion of Laplace transforms. Among them, a very accurate and general method is due to Talbot: this method approximates the value of the inverse Laplace transform f(t), for t fixed, using the complex values of the Laplace transform IVs ) sampled on a suitable contour of the complex plane. On the basis of the interest raised by Talbot’s method implementation, the author has been induced to investigate more deeply the possibilities of this method and has been able to generahze Talbot’s method, to approximate simultaneously several values of f( t) using the same sampling values of the Laplace transform. In this way, the only unfavorable aspect of the classical Talbot method, that is, that of recomputing all of the samples of IVs) for each t, has been eliminated. .Aualysis]: General—error analysis; Subject Descriptors: G. 1.0 [Numerical G. 1.2 [Numerical Analysis]: Approxirnation-nonhnear approximation; G, 1.4 [Numerical Analysis]: Quadrature and Numerical Differentiation-equal interval integration; error analysis; G.1.9 [Numerical Analysis]: Integral Equations—Fredholm equations

Categories

numerical

General

and

algorithms;

Terms: Algorithms

Additional numerical

Key Words and Phrases: Complex method, TALBOT, trapezoidal rule

inversion

formula,

inverse

Laplace

transform,

1. INTRODUCTION The Laplace absolutely

transform integrable

of a function on any finite

F(s)

= /me-s’f(t) o

f(t), interval

dt,

defined (O, a],

on the interval is defined

(O, + CJ) and

as follows:

Res>aO,

where a. is the Laplace transform abscissa of convergence. The inverse f(t) from known values of ~(s). Laplace problem is that of reconstmcting

Author’s address: Istituto Universitario Navale, Facolta di Scienze Nautiche, Istituto di Matematicaj via A. De Gasperi, 5-80133 Napoli, Italy; email: [email protected]. Permission to make digital/hard copy of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage, the copyright notice, the title of the publication, and its date appear, and notice is given that copying is by permission of ACM. Inc. To copy otherwise, to republish. to post on servers, or to redistribute to lists, requires prior specific permission and/or a fee. 01995 ACM 0098-3500/95/12000-0347 $03.50 ACM Transactions

on Mathematical

Software,

Vol. 21, No. 4, December

1995, Pages 347-371.

348

Marlarosaria

.

The

Laplace

problems

in

Rlzzardi

transform the

represents

fields

of science

a very and

effective

tool

engineering;

for solving

however,

the

several

numerical

inversion of Laplace transforms still remains a very difficult problem to be solved in general. Among the methods of inversion based on the complex contour integration of the Riemann Inversion Formula, 1983; 1990; Talbot 1979] has proved Laplace

transform

to compute

functions

~(t).

computational pends If

very

need

to

transform

desirable

that,

mate just

a single

range

inefficient.

compute

the

of f(t),

same

for

may

also be stated

a particular

suitable possible? In this cation form

without

to Talbot’s

considerations

that

several

uses the values

in the

The Riemann

theory,

same

results

question, been chosen

for

any

Moreover,

of the

t in a

how

is it

a modifi-

Laplace

transform.

of

Some

transpractical

too.

BACKGROUND

Inversion

of its Laplace

Formula

transform

gives

an integral

function

F(s),

that

1

f(t)





27Ti

where

B is the Bromwich

to the

imaginary

ties

have

we introduce

are reported

be

any value

again

samples

inverse

to approxi-

previous

result?

of its inverse

and some numerical

2. THEORETICAL

to the

de-

it would

computed

to use them

loss of accuracy

method

method,

parameters

And what is the cost? article, starting from the underlying

to approximate

terms

answer

if Talbot’s

of t, is it possible

value

interval

as follows:

the

algorithm

f(tk ) of an

of 3’(s),

for

that

evaluations.

can also be used to approximate

when t belongs to a suitable interval. This work attempts to give a positive

which

values

to use Talbot’s

values

known

inversion function

several

and we want

for efficiency, value

In fact, it is well

transform

numerically

function,

we need

all of the computations

transform

of Laplace

Rizzardi range of

of t where

of values

to repeat

of a Laplace

on the number

Laplace

f(t)

it needs

may become complexity

heavily we

and to a wide

Unfortunately,

each t,and this

Talbot’s method [Murli and to be applicable both to a wide

axis.

\ -B

estl’(s)

contour The

ds,

from

contour

representation

i.-

t>o,

to the

in

(2.1)

y – i~ to y + i~, with

is located

for ~(t)

is,

right

y > u-., parallel

of all

the

singulari-

of F’(s).

In Talbot’s

method

the Bromwich

contour

is replaced

by an equivalent

one,

which retains a proper balance between the two following effects: the new contour is moved as much as possible to the left so as to reduce in magnitude of the factor e ‘~ in (2.1), but it must not be moved too close to singularities F(s),

as to do so will

Referring

to Murli

result

in peaks

and Rizzardi

in the integrand

[1983;

1990]

function.

and Talbot

[1979]

for further

details about Talbot’s method and its implementation, here \\-e only outline those aspects that seem to be relevant to the method to be introduced. ACM

TransactIons

on Mathematical

Software,

Vol

21 Xo

4 December

1595

A Modification Talbot’s

method

transform

requires

function

that

3’(s)

the convex

be known

IimlF(s)l ~+.

F(s)

applicable

needs

Talbot’s

contour

in

Re(s ) 0,

Iim El(t,

n) = O

(2.llC)

n) = O.

(2.lld)

n

Vt >0,

limllz(t, n

The last two properties tell us that, for any possible value Talbot’s method is uniformly convergent for t >0. Nevertheless, it consists of a suitable choice, as a function

of A, a,

and

P,

of t, of all the

Talbot parameters for assuring a high rate of convergence, and this implies that, if several values of f(t) have to be evaluated, Talbot’s method needs to On the contrary, our idea is to look at a repeat aH of the computations. suitable which

choice is the

of

domain

To the previous

Talbot’s

parameters

where

truncation

as

defined by Talbot as the roundoff error sion, the middle term of the trapezoidal

E,(t, ACM

TransactIons

on Mathematical

a function

of

the

f(t) has to be evaluated. errors, we must add the roundoff

n) =&’ Software,

(

whole

error,

in computing, to the machine sum approximating (2.4):

10-C

interval,

which

is

preci-

Q(O, t) n’

(2.12) )

VOI 21, No 4, December

1995

A Modification where

c is the

computations This

decimal

will

formal

precision

be carried

definition

of the

of Talbot’s Method

floating-point

351

.

system

where

the

out.

of the roundoff

error

is justified

by the fact

that,

in

the classical method, usually the middle term in the sum is the largest; so the roundoff introduced in the summation process will be negligible in respect to the final result. But, as we will modified method,

see later, so we will

the estimate (2.12) is no longer reliable need to examine its stability closely.

for the

3. ERROR ANALYSIS Let us choose Talbot’s

parameters

A, a,

v, and

n at t* and use them

also for

Vt G [tmin, tmax]. In the following

sections,

we will

(i.e., the method modification the

that repeats all modified method

fied method

be built

will

call Talbot’s of the (which

step-by-step

method

computations avoids that

by examining

the

classical

for each repetition).

the error

method

t) and our This modi-

components.

3.1 Convergence First,

we want

to derive

the convergence

the classical method. Indeed, this assured by the last two properties emphasizes the following result: THEOREM

fied

dependence

3.1.1.

method

PROOF.

If we

m

- f(t)

the

term

[ f(t”

of the

If the classical

is convergent

of the modified

method

is not necessary since the of (2.11), but the following method

Talbot

t and

on

method

n. Let

converges,

from

that

of

convergence is analysis well us

then

prove

the

the modi-

too.

write

= [m

-m]

) – f(t )1 does not

+ [mm

depend

- f(t”)]

the term

ACM

[~

Transactions

– ~]

- f(t)],

(3.1)

on n, but only on t and ~(t). The

term [~ – f(t* )1 is the error of Talbot’s for n - cc by the last on t,and it vanishes Let us examine we have

+ [ f(t”)

method at t*; two properties with

on Mathematical

it does not dePend in (2.11).

respect

to t and

n. From

Software,

Vol. 21, No. 4, December

(2.8)

1995.

352

.

Mariarosar!a

Rizzardr

then, for n + CO, e-n’ - () on Integral Theorem, we have that

=& and by (3.1), it follows lim

Ml

and

e-n’

+ m on

&f2;

)(a+As(2)) - I)cJz

jMQ(z,t*)(e(t-t*

So by

Cauchy’s

=f(~)

-f(~*)

that [~-~(t)]

V t > 0

uniformly

= O,

n-x since, 3.2

by the hypothesis,

Talbot’s

method



is convergent.

Stability

To emphasize the role played by roundoff in our modified method, recall from Murli and Rizzardi [1983; 1990] that the numerical Talbot’s method may be expressed as

we must result of

Ae[’t m= —-Tn(t), n

(3.2)

and (3J= j(n/n). Formula (3.2) has been obtained rule approximation to (2.4), after taking symmetry substitution kfz G M, so that

Talbot’s

contour

z=2i6,

in (2.3) is parameterized

s(O)

=

CT+

by applying the trapezoidal into account and after the

AsL,(@),

as

/3=(

-T,7r),

where s,,(o) All have

of Talbot’s an error

(3.2) and (3.3),

parameters that

have

= Ocoto+ been

rewritten

for

t’,

TransactIons

on Mathematical

at

t*,

accuracy

so f( t* ) is supposed requirement,

we have

Ifil ACM

chosen

does not exceed the input

ivo.

Software,

S ftt.und(t’, Vol

n),

21, No 4, December

1995

to

and from

A Modification

of Talbot’s Method

.

353

where

Now,

from

(3.2) and (3.3), it follows

that

~emt Iml

= —lT.(t)/ n

~(a

+ A)l +

and

n–1

:eAt*+A(t–t*) 2 “+:,( The have

sequence Vj:

eAt’6Jcot

6,+

A(t–t*)~,

cOttIJ

6J)F(a+ AS,,(OJ))I.

{61 cot @j}, j = 1,...,

l O; more

U.

inverses

t

— ERRT.lLO$

~

ERRT.i.

~

ERRT.,.X

—%-+?---+3-

Fig.1.

Error

-l

plots fort

precisely,

the

10

5

t

~[1,

lO].

Laplace

transform

test

functions

are

Fl(s)

= 1/s2

F2(S)

= 1/(s

F3(s)

= atan(l/s)

~4(S)

=

S2/(S3

– 1)

= t

f2(t)

= e’

f,(t)= sin(t)/t +

8)

The behavior of the errors defined of test functions. More precisely, ACM

f,(t)

Transactions

f,(t)

= (e- 2~ + 2etcos(t@))\3.

above is surprisingly looking at the tests on Mathematical

Software,

homogeneous for a lot in the macrointeruals Vol

21, No. 4, December

1995.

356

Manarosaria

.

Rlzzardl

F1

Test function

Test function F2

102

104

10°

102

,0-2

10°

/

8 k

a)

I

&

10“4

$10-2

,0.4

,0-6

,0-6

108

, ~.8

,0.10 o

20

40

60

o

20

40

t

60

t

Test function F3

Test function

F4

~~m 102

8 k

al

10° ,0.2 ,0.4

,(--’”~

I

o

20

40

O“60L

60

20

40

60

t

IL

I Fig.2,

Error

plOtsfort=

I

[10,50],

(Figures 1-3), we can observe that creases, as expected, but it becomes t + tmax, when compared to the other

ERR~~,X decreases and ERR~~,m . . . .. inquickly very large (often overflow) for errors. (Figures 4–5), the On the contrary, looking at plots in the micromtervals swap the roles between them, remaining of a similar magnitude; two errors but usually Talbot’s parameters chosen at the lower bound of the interval give more accuracy at a minor computational-effort: this can be explained by the instability of the method that computes f~~,X. According to the previous analysis, from “small” error ERR ~~,,( t ) always occurs for the Talbot error analysis’): this phenomenon ACM

Transactions

on Mathematical

Softww-e,

Vol

all

of these

tests

we see that

a

t < t* = tmax(incontradiction to is due to the fact that for t < t*

21, No 4, December

1995

A Modification

of Talbot’s Method

357

Test function F1 1020 ~ L

1o 10

15 10

05 10° 0.5

1 1

:

0.10 ~

500

o

1000

t Test function F3 1040

Fig. 3.

Error

plots for t ● [50, 1000].

1030 ~ 1020 L

$10’0 10°

,o-lo~~ o

1000

500

t

= we have chosen a contour unnecessarily close to the singularities of F(s), and this leads to a roundoff error in the trapezoidal sum that exceeds the E, in (2.12) as defined by Talbot. So, from results of the previous error analysis, we can conclude that the new method shows a similar performance (in terms of accuracy) over the whole

interval

[ t~,.,

t* such that

—we

choose

—we

maintain

The former

t~,X 1 if

the value

requirement

it minimizes

elt –‘* 1 and

of n A as small is obviously

as possible.

satisfied

by (3.7)

t* = tmld = :(tm,n + tma,). ACM Transactions

on Mathematical

Software,

Vol

21, No 4. December

1995

358

Manarosana

.

Rlzzardl

Test function F1 ‘0-4

Test function

~

’02

, ~.5

F2

~—

10°

,0-6

,0.2

z

s

:10”7

k

10”4

a)

,0-8

,0-6

,0-9

,0-8

,0-10

,0-10

2

1

3

4

I

2

1

3

t

4

t

Test function

F3

Test function

F4

105

104 ~ 102 10°

10° 8 k a)

,0.4 ,0.6

1

, ~.e 2

1

3

t

4

r== t .EzzzZ. I

ERRT,lLax

Fig.4

The latter into

is difficult

account

examine

them

4. ERROR Before

the

in

Error

pIotsfort=[l,4]

to be satisfied

other

the

error

next

because

components

as

the value well,

of n, in order

tends

to

large.

to take We

will

section.

ESTIMATION

introducing

components

the modified

of the classical

method,

method.

we need to examine

In this

way,

we will

Transactions

on Mathematical

Software.

Vol. 21, No 4, December

1995

closely

the error

be able to estimate

the number of function evaluations necessary to guarantee, method, the same accuracy the classical method has. ACM

be

in the

modified

A Modification

of Talbot’s Method

Test function F1

Test function

.

359

F2

1 , ~-5

1

~ 10-’ k al ,0.7

1

, ~-8

1 4

6

8

,0-9 ~ 4

6

4

1

t

Test function F3

Test function F4

8

‘ 0-3~

,0.4 ~ 10”5 k al ,0-6 ,0.7 ,&

~ 4

6

8

t

t

,

Fig.5.

Let us use the Talbot

Error plots fort

parameters

~[4,8].

A, u, v, and

n, chosen

at t*,for each t in

since the global error grows as t moves from t* to t~,n or the Talbot error components at the endpoints of the t~ax, let us consider t belongs to. We do not consider the roundoff error Er(t,n) interval that [%n>t

defined roundoff First,

~ax ], and

in (2.12) since, as shown in Section 3.2, Talbot’s estimation of error is no longer reliable. consider the error component E1(t, n) defined by (2.9). From (2.11)

the integrand

in (2.9) is regular

arc Ml may be any path from substituted by the parabola rl VZ

G

rl ACM

eZ

in the halfstrip

=2(7)

TransactIons

If

defined

in (2.6);

then

the

– 2 n i to 2 n i, and in particular, Ml may be from –2mi to 2wi with the vertex at z = p:

=x(T)

+ iy(~),

on Mathematical

Software,

Irl o

– 72),

Since rl is free to get deformed in the halfstrip (0, +~). Then, by the Darboux Theorem, we have Ilczl(t, Remembering on El,

Iength( rl ) ~m max z~r,

n)l


=x(T)

171s 1

and

But,

unlike

rl,

X(T)

= –p(l

y(~)

= 2rr7_.

rz cannot

be arbitrarily

be located on the right t+ +~, then p~O. Repeating function

the

of all

analysis

the

carried

deformed singularities

out

for

El,

o

in the halfstrip of

K; it must

I’( a + AsU(z)).

applied

in

this

So, for

case

to

the

eKc+As, (z)) e for sufficiently the error

large

Ez that

values

l~~(t,



the dependence

n)l < O(A, of p, this

–1

of n, we are able to establish

emphasizes

lE2(f, For smal

values

–nz

a, v) .e~[~-~p[u-

of this

error

lJ/~+~p/(e’-l)]

an upper on t and

bound

to

n:

-nP$

becomes

n)l < @(A, u, v) .e’iU+’-’~(~-l)/2]-~P,

(4.5)

where

p s such that r, is located, in the halfstrip K, on the right of all the singularities of 3’( a + ~sp( z )). Also for Ez, choose the Talbot parameters A, u, v, and n at t“, and use them also for t~,. and t~,,. Writing (4.5) and t equal to tm,n,t*, and t~,X, at tm,n. respectively, we see again that the error is bigger at tmax and smaller ACM

Transactions

on Mathematical

Software,

Vol. 21, No, 4, December

1995.

362

Marlarosarla

.

Rlzzardl

Test function ‘0-2

Test function

F1 ’00

~

F2

~ I

, “.2

10-8

,o-lo~ o

10

5

t

t Test function

Test function

F3

F4

10°,

10°

1

I

10

t

t

I

‘RR’’’’’” ~ Fig,6,

Then,

if we wish

n(z) ml n for

t~,.,

to find

such

a suitable

plots

fort

value

=[0,1,

lO],

of n, namely,

n(#X

for

t~~,

and

as

lE,(tmln, from

Error

(4.5), it suffices

ng:n)l

=

to choose ~(a

m] n

= ~ — (4.6)

which ACM

once again TransactIons

returns

on Mathematical

n(~~~ < n. Software,

Vol. 21, No 4, December

1995

A Modification

,

Test function FI

~-6

lD

z k

10”6

al

,0.9

~ 0.7

, ~-lo

,0-8

0

20

40

L

60



0

20

Test function F3

,0-4

Test function ‘ 0-2 ~

T

,0.3 r ~ 10-4[ k al ,0.5 ,0-6

,(--@~ o

-t

1O-’L

20

40

o

60

60

F4

x

3 20

40

60

t

t

F1g.7.

MODIFIED

40

t

t

5. THE

363

,0-5

10”8

.-

.

Test function F2

,0.4

, ~.7

8 k

of Talbot’s Method

TALBOT

Error

plots fort

~[10,50].

METHOD

Taking into account the previous considerations and both formulas for n&).X in (4.4) and (4.6), the simplest computational strategy consists of the following algorithm: —By

a call to the subroutine

rical parameters (according as stated in (3.7).

TAPAR

[Murli

to Talbot’s

and Rizzardi

method)

1990],

are computed

the geometfor

t = t~,d,

—A correction to the accuracy parameter n is then computed by choosing the formulas (4.4) and (4.6) when t“ ==tm,d. maximum value of n’~~, between ACM

TransactIons

on Mathematical

Software,

Vol

21, No. 4, December

1995

364

.

Marlarosaria

Rizzardl Test function F1 ‘0”2

~

,o-’”~ 1000

500

o

t Test function

F3

102 Fig.8.

Error

plots fort

G [50,1000]. 10°

,0.2 G k a) ,0.4

,0-6

10-8

500

0

1000

t

==J —Finally,

to compute

n~~X in (4.6),

we need

to know

a value

of p such

that

the corresponding parabola rz is located on the right side of all of the singularities of F( a + Asv( z )). A numerical value of p can be obtained, for example, by assuming that the image of the vertex of 17z, SU(—p), agrees with the midpoint between cro and m + A, the rightmost real part on the Talbot contour, i.e., mo+u+A ,o>O:u+As,,(-p)=

which

ACM

leads

TransactIons

to the following

on Mathematical

2’

nonlinear

Software,

Vol

equation:

21, No 4, December

1995

A Modification

of Talbot’s Method

Test function F1

.

365

Test function F2 ,0-2

,0-4

& k

al

10-6 ,0-6

, ~-lo .-

,()-lo~

0

5 t

o

10

Test function F3

,0.3

5 t

10

Test function F4

,0.2

[

,0.4 ,0-4

~ 10-5 S k

k al ,0-6

a

10“6 10-8

,0-7 ,0-8

,(yo~

0

5

10

5 t

o

t

Fig. 9.

This

equation

6. NUMERICAL Numerical assuring First, classical

the

of the radix

solved just

by a Newton

using

have

shown

inverse

that

Laplace

the

modified

transform

over

Talbot

that

and with ACM

Talbot’s

Transactions

method

parameters

on Mathematical

method

a relatively

the same accuracy of Talbot’s method almost we report on some of the results of comparing method

process

gives

a

a few iterations.

EXAMPLES

results

approximate

plots for t ● [1, 10].

Error

may be numerically

good approximation

10

Software,

large

everywhere. our method always Vol

is able

to

interval, with

chosen

the

at t~,d

21, No 4, December

1995.

.

366

1

Manarosaria

Rizzard!

Test function FI

~.4

Test function

F2

‘“”’~

~-5

,0.4

1

0-6

1

~ 10”5 k a ,0-6

L ~.7

:1 al

1 1

~-8

%

:;L4

~.9 ~.lo

1

o

20

40

o

60

20

t

40

60

t

Test function

Test function

F3

‘ 0“4 ~—–

‘ 0“2 ~~

,0-5

1

F4

0-6

m

10’

I

d

L

1

, ~.t?

1

o

40

20

Fig

60

0

Error

plots fort

in Figures

—ERR~,lbO,

method;

of the classical

—ERR~~O~,~ : the error

of our modified

—ERR~~,~

obtained

: the

error

method;

by using

60

=[10,50]

for each t G [t ~,~, t~,, ]. In particular, ing errors: : the error

40

t

e, 10.

20

6-8

we compare

the follow-

and

V t Talbot’s

parameters

chosen

at

tmid. P1ots show that ‘RR Tmod,f iS always less than ER%~,d: both use the same geometrical parameters h, c, and v, but they differ in the value of the accuracy parameter n (which in “Tmodif” is larger than for “Tmid”) to take into account the whole interval that t belongs to. ‘hese

ACM

TransactIons

on Mathematical

Software,

Vol

21, No 4. December

1995,

A Modlflcatlon

of Talbot’s Method

.

367

Test function F1

, ~.2

, ~-4

& k 10-6 al ,0-8

,0.10 0

500

1000

t Test function F3 Fig. 11.

Error

plots fort

e[50j

1000].

,o-J_____4 o

500

1000

t

Second,

in Figures

9– 11 we compare

the errors

ERR~.lbOt,

ERR~~O~l~,

and

> where the first two are the same as before, and where ERR~~., (defined in Section 3) is related to Talbot’s method with parameters chosen at t~ax and using Vt e [tm,n,tmax]. These results are very encouraging. As expected, in the middle part of the interval the aim is always achieved. Often, the accuracy is better than the classical Talbot’s method. induces a But, sometimes, when t is small, the closeness of singularities However, when t is small, then “case 1“2 holds, loss of accuracy near t = tm,n.

ERRTmax

and

in “case

1“ the

values

of n in the

classical

method

“’Case 1“ occurs, in Talbot’s method, when F’(s) has only real singularities complex singularities for moderate t (usually t < 10). ACM Transactions

on Mathematical

Software,

are

so small

(e.g.,

‘dt or when F(s) has

Vol. 21, No. 4, December

1995.

368

.

Mariarosana

Rlzzardi

Test function

Test function

FI

14! 13 I

1

18

I

16

.

F2

n n

.

.

s

~ 14 fl) K 12 ~ 10

b *

8

‘~

246810

246810

t

t Test function

Test function

F3

F4

35

20 L-3-e-J

30 ~ 25 > :20 ‘ z

.

/’.

n

n

o

it may

be more

L

15

246810

t

t 1

Fig. 12.

n s 11 for

a required

Number

accuracy

of function

of five

evaluations

decimal

fort

digits)

=[1,

lO]

that

convenient to repeat the whole computation for each value of t. Or a simple trick may be used: enlarge the interval [ t~,~, t~~, ] toward O when establishing the method’s parameters. Figures 12– 14 report the number of function evaluations (n) used by the three different strategies: respectively, the classical method, Talbot’s method at t~,,, and the modified method; more precisely, we report the number of function evaluations required to approximate just a single value of an inverse Laplace transform: if we need to compute, for example, 100 samples of ~1(t) 12 we get that the classical for t G (O, 10], then from the first plot in Figure method will require 1000 function evaluations; Talbot’s method at tmax will ACM

TransactIons

on Mathematical

Software,

Vol. 21, No 4, December

1995

A Modification

of Talbot’s Method

369

.

Test function F2

Test function F1 35

I

13

30 3“”

15 1, 10

I

20

30

40

U”””

”””3”

I

I

I

1.

50

10

20

30

40

t

t

Test function F3

l-est function F4

50

4’ ~ 40 t

I

200 150

n

n

250

—1 I

t

I

t

20

15 I

L

10

20

30

40

50

10

20

30

40

50

t

t ~Talbot

?aTn,odif



u



nTmax Fig. 13.

Number

of function

evaluations

fort

=[10,50].

require 10 function evaluations; and the modified method will require only 12 function evaluations (giving more accuracy). This means that the global number of function evaluations, in the classical method, is the sum of all the function evaluations carried out for each value of t, while in the other two strategies, this is constant. Pictures report the number of function evaluations of Talbot’s method only to give

an idea

of how

much

the value

of n has been

raised

by the modified

method, since it is meaningless to compare the classical method and our modification from a computational point of view. Finally, from these plots we can see that the modified method produces, for gives at tmax; small values of t, a value of n larger than the Talbot method ACM

Transactions

on Mathematical

Software,

Vol. 21, No 4, December

1995.

370

Marlarosaria

.

Rizzardi

F1

Test function

14 13 9

12 ~ 11 & & & A .k *

10 ~~~~

*-

9 8 200

400

600 t

800

1000

Test function F3 Fig ?!=

14

Number

of function

evaluations

for

[50, 1000].

~-200

400

600 t

800

1000

n~ajbo~ n’h.

d,f



i nTnl.x

this

occurs

when

the

methods.

Otherwise,

a minor

computational

7. CONCLUDING A modification

transform

method

is based

on an

required. In this also ACM

way,

TransactIons

method

that

function

evaluations

is

is able to produce

not only

function

is the high

strategy on Mathematical

has

formula

F(s)

for

accuracy

is reached Software,

approximates

~(t)

approximation

transform

an efficient

method

small more

for

both

accuracy

at

cost.

Talbot’s

Laplace

Laplace

of function

1

+--+-+

REMARKS of

inverse the

number

the modified



Vol

that

been that

all

several introduced. uses

of the of Talbot’s

applies

the

values

well

21, No. 4, December

values The

same

of ~(t)

method

of

samples

that

retained,

to multidimensional 1995

an

modified of

are but

A Modification Laplace

inversion

tional

efficiency

problems. of the

algorithm,

or it

transform,

otherwise

Moreover,

summation

is possible

not usable

it is possible algorithm,

to use

of Talbot’s Method

for

prescribed

to enhance example,

equally

by the classical

the computavia

spaced

371

.

a parallel

values

of the

method.

REFERENCES A. AND RIZZA~DI, M. 1983. Sull’implementazione numei-ica dells Trasformata di Laplace. Rep. SOFMAT

MURIJ,

del metodo di Talbot per la inversion 10.83, Progetto

Finalizzato

Informatica

del CNR, Rome, Italy. MURLI,

A. AND RIZZARDI,

problem.

M.

1990.

Algorithm

ACM Trans. J4ath. SoftLV. 16,

TALBOT, A. 1979.

The accurate

numerical

682:

Talbot’s

method

for

the

Laplace

inversion

2. inversion

of Laplace

transforms.

J.

Inst. Math. Appl.

23. Received May 1992; revised July

1993 and August

ACM Transactions

1994; accepted September

on Mathematical

1994

Software, Vol. 21, No. 4, December

1995.