You are on page 1of 9

Jour nal of Engi neer i ng Mat hemat i cs, Vol. 5, No.

1, J anuar y 1971
Wol t er s - Noor dhof f Publ i s hi ng- Gr oni ngen
Pr i nt ed in t he Net her l ands
Gaussian Quadrature Formulas for the Numerical Integration of
Bromwich's Integral and the Inversion of the Laplace Transform
R. P I E S S E NS
Institute for Applied Mathematics, University of Leuven, Heverlee, Belgium
(Received August 6, 1969 and in revi sed form May 15, 1970)
S UMMAR Y
An appr oxi mat e formul a for t he i nversi on of t he Lapl ace t r ansf or m F(p) is st udi ed. The f or mul a is exact whenever
Y(p) is a l i near combi nat i on of p-S+k, k=0, 1, 2, ..., 2 N- 1, wi t h s an ar bi t r ar y posi t i ve real number . The f or mul a is
~erived f r om a gaussi an i nt egr at i on f or mul a for Br omwi ch' s i nver si on i nt egral .
A numer i cal exampl e is given as i l l ust rat i on of t he use of t he appr oxi mat e i nver si on formul a.
1. Introduction
The Laplace t ransform is useful in solving some ordi nary and partial differential equat i ons and
integral equat i ons and arises in many areas of engineering mat hemat i cs. However, the exact
det ermi nat i on of the original function f ( t ) from its Lapl ace t ransform
is often a great difficulty. In many cases, numeri cal met hods must be used. A number of
interesting numerical met hods are known [1], [2], [3], [4], [5], [6], the maj ori t y of which,
unfort unat el y gives unsat i sfact ory results if F(p) has a branch point at itifinity.
In this paper, we shall consider an approxi mat e formul a for the inversion of the Laplace
transform, namely,
N
f ( t ) = t ~ 1 Z {Ak(Uk/t)~V(Uk/t)} (2)
k = l
where s is a positive real number t hat must be chosen so t hat pSF (p) is anal yt i c and has no
branch poi nt at infinity, t hus so t hat we can write
pSF(p) = ~ akP - k .
k=O
We shall give formulas for the coefficients Ak and we shall discuss a met hod for the Calculation
of the coefficients Uk. The coefficients A k and Uk are dependent on the value of s and N. Salzer
[6], [7] has al ready studied a formul a of the same form of (2), but only for the special value
s = 1. Kryl ov and Skoblya [8] and Skobl ya [9] have given tables for the coefficients A k and Uk
for several values of N and s, but with a very limited accuracy. Moreover, t hey have not given
the formulas for Ak and Uk, presented here.
2. Inversion of the Laplace Transform by Integration of Bromwich's Integral
The original function f ( t ) is given by
f ( t ) = ~ j ept F(p)dp (3)
where c is chosen so t hat the line Re (p) = c lies to the right of all singularities of F (p), but is
otherwise arbitrary.
Journal of Engineerin9 Math., Vol. 5 (1971) 1 9
2 R. Pi e s s e ns
Substituting
p t = u
and
F (u/t) = u - S G (u)
where s is the paramet er defined in section 1, (3) yields
1
f c ' + ~ e " u - ~ G( u ) d u (4)
t f ( t ) = ~ j .c'-jo~
where c' = tc.
Thus for the inversion of the Laplace transform, we need a formul a to calculate the integral
in the second member of (3). We consider the approxi mat e formul a
( c'+joo
1 t c , _ j o o e , u S G( u ) d u , . . ' ~ A k G( u k ) . (5)
2~j k=l
Indeed, substituting (5) in (4), we have the desired approxi mat e formul a (2) for the inversion
of the Laplace transform. We j ust have to give conveni ent formulas for the coefficients Ak
and uk.
We try to determine the abscissas uk and the weights A k in (5) so t hat the integration is exact
whenever G(u) is a pol ynomi al in u 1, of degree < 2 N- 1. Then we have a Gaussi an i nt egrat i on
formula, t hat has a degree of precision 2 N- 1. In section 3, we shall st udy ort hogonal pol yno-
mials, the roots of which are the coefficients uk. In section 6, we shall give formulas for the
coefficientsA k.
3. Orthogonal Pol ynomi al s Connected with the Gaussian Integration Formulas
Fr om the t heory of Gaussi an quadr at ur e formul ae (Davis and Rabi nowi t z [10]), it is well
known t hat the existence of a pol ynomi al PN,=(P-1) in the variable p-1, of degree N, with the
propert y
2 r c j e Pp - =PN, =( p - l ) p - r d p = 0 (6)
for r =0 , 1, 2 . . . . , N- - 1, where L is an arbi t rary vertical line in the positive half plane of the
complex plane, is necessary and sufficient for the existence of an N-poi nt i nt egrat i on formul a
with degree of precision 2 N- 1. Condi t i on (6) means t hat PN,~(P- 1) is a pol ynomi al of degree N,
which is ort hogonal to all pol ynomi al s of degree < N, with respect to the weight function
eP p - s.
If we set
cbr (t) = 5 ~ -1 {p-(s + ~)Pu,~ (P- ' ) } (7)
where S - , denotes the inverse Laplace transform, the condi t i on (6) is equivalent to
9 ~ ( i ) = 0
for r =0 , 1, 2 . . . . . N- 1 .
Using the convol ut i on propert y of the Laplace transform, we obt ai n
f 1 1 ~ ( 1 - u ) d u r CN + U~+r-~qN_ , = 0 , = 0 , 1 . . . . . N- - 1 (8)
0
where cN and qN-1,s(t) are defined by the relations
PN,s(p - 1 ) = A[ CN q- p - I QN_ I , s ( p - a)] ( 9 )
and
qN_ l , s(t ) = 5O- 1 { p - 1 Q N _ I, ~(p-1) }
Journal of Engineering Math., Vol. 5 (1971) 129
Gaussian quadrature formulas 3
wi t h QN- 1,~ a pol ynomi a l in p - 1 of degr ee N - 1 a n d A a n o r ma l i z a t i o n coeffi ci ent whi c h will be
de t e r mi ne d l at er.
Re l a t i on (8) i mpl i es
f l u s + r _ t ( l _ u ) q n _ l , s ( l _ u ) d u = O , r = 0 , . . . . . 1 N - 2 .
o
Thus, qN-1,~( 1 - u ) is o r t h o g o n a l on [0, 1] t o al l pol ynomi a l s of degr ee < N - 1 , wi t h r es pect
t o t he wei ght f unct i on w( u ) =u ~ - l ( 1 - u).
Hence, we ha ve
qN- 1,s(t) = p ~ s ? 1)(1 __ 2t) (10)
whe r e P~"P)(x) de not e s a J acobi pol ynomi a l of degr ee N (see [ 11] , sect i on 10.8).
Cons equent l y, we ha ve
qN-l,~(t) = 2 F I ( - N + 1, N + s ; 2; t)
a n d
f
l
c N : -- u s - 1 2 F I ( - N + l , N + s ; 2; 1 - u ) d u
JO
o r
2 s + l ' N ( N + s - 1 )
Subs t i t ut i ng (10) a nd (11) i n (9), we have
A A [ - N + I , N+s , 1 ~1
PN's(P-I)-- N ( N + s - - 1 ) + --p 3F1 2 ; . (12)
I f we choos e t he s t a nda r di z a t i on f act or as
A = ( - 1 ) N + I N ( N + s - 1 )
we obt ai n fi nal l y
PN,s(p- 1) = (--1)N 2Fo ( - N , N + s - 1 ; - ; p - l ) . (13)
F o r mu l a (13) is val i d f or all r eal s >0 .
F o r s > 1 however , t he de r i va t i on c a n be si mpl i fi ed a n d it is eas y t o d e mo n s t r a t e t h a t
PN,~(p-1)=p~{p~O'~-2)(1--2t)} ( S > I ) . (14)
An o t h e r i nt er es t i ng f o r m of PN,s ( P- 1) is t he c onf l ue nt h y p e r g e o me t r i c f unc t i on
PN,s(P-1) = p - N ( U + s _ 1)N l V l ( - U ; 2 - 2 U - s ; - p ) (15)
whe r e
F( q + N) ( Po c h h a mme r ' s s ymbol )
( q ) N = - r( q) "
4. Properties of the Orthogonal Polynomials
4.1. Differential Equation.
I t is well k n o wn [11] t ha t t he gener al i zed h y p e r g e o me t r i c f unc t i on
U : pFq [ C t l ' 0(2, 0(3, " ' " G(p 1
[ f l l , f 1 2 , f 1 3 , . . - , f l ~ ; x
satisfies t he di f f er ent i al e q u a t i o n
(16)
Journal of Enoineerin9 Math., V o l . 5 ( 1 9 7 1 ) 1 - 9
aN, s =
a nd
{6 (J q - f l ~ - 1). . . ( 5 + f l a - 1 ) - x ( b + ~1). . . ( ( ~ A V ( ~ p ) } U = 0
R. Piessens
(17)
whe r e
d
( ~X- - .
dx
Thus, u = PN,s(X) is a s ol ut i on of
x Z u " +( s x - 1)u' = U ( U + s - 1 ) u . (18)
The di fferent i al e q u a t i o n (18) is a speci al case of
x e u" +( a x + b ) u ' = U ( N + a - 1)u (19)
whi ch is s t udi ed by Kr al l a nd Fr i nk [ 12] . Sol ut i ons of (19) are t he gener al i zed Bessel pol y-
nomi a l s YN(X, a, b). Thus, t he p o l y n o mi a l s PN,~ (x) ar e speci al cases of t he gener al i zed Bessel
p o l y n o mi a l s
PN,~(x) = ( - - 1) N YN(x, s, -- 1) . (20)
The pr ope r t i e s 4.2, 4.3 a nd 4.4 ar e easi l y der i ved f r om t he c o r r e s p o n d i n g pr oper t i es of t he
gener al i zed Bessel pol ynomi a l s . Pr ope r t i e s 4.3 a nd 4.4 can al s o be o b t a i n e d f r om k n o wn
pr ope r t i e s of conf l uent h y p e r g e o me t r i c f unct i ons.
4.2. Rodri gues' Formul a
d N
PN,~(P- 1) = (__ 1)Ne-ppN+s-1 dP N (ePp-N-s+ 1) (21)
4.3. Recurrence Rel at i on
The mo s t i mp o r t a n t r ecur r ence r el at i on for t he p o l y n o mi a l s PN,~(x) is
PN,~ (x) = (aN,~ x + b N,~) PN- 1,s (X) + CN,s PN- 2,~ (x) f or N > 2 (22)
wher e
( 2N + s - 3) ( 2N + s - 2) ( 2N + s - 3)(2 - s) ( 2N + s - 2) ( N - 1)
U + s - 2 bN,~ = ( U + s - 2) ( 2N + s - - 4) CU,s = ( N + s - - Z)(ZN + s - - 4)
eo, ( x ) = 1 , el, (x) = s x - 1 .
4.4. Expressi on f or the Derivative
d p - - NP- +
N
2 N + s - 2
1 , ( p ) . ( 2 3 ) PN -1
N)
2 N + s - 2 PN's(P-1)
4.5. l nverse Laplace TransJorm
Us i ng (13), we obt a i n
t a - 1
5 f - 1 {p- apN, s( p- 1) } _- ( _ 1)N F ~ 2F1 ( - N , N + s - 1; a ; t)
wher e a is an a r bi t r a r y real pos i t i ve numbe r .
4.6. Mome nt s
We defi ne t he r-t h mo me n t as
(24)
Journal of Engineering Math., VoL 5 (1971) 1 9
Gaussian quadrature formulas 5
MN,r= 2rcj Leep-~PN,s(p-1)p-~dP.
It is obvi ous t hat
MN,,.=O for r = 0 , 1 , 2 . . . . . N- 1 .
Fur t her , we have
(-- 1)U(N+k)!
MN'N+k = F( 2N+k+s) k! for
To demonst r at e (26), we r emar k t hat
MN, r = ~, (1)
where ~r(t)is defined as in (7).
Setting t = 1 in (24), we obt ai n
k =O, 1, 2 . . . .
M N r - - ( - - 1 ) N 2F I (-N,N-- s-1 9 r - }- s; 1 )
r (r + s)
o r
( - 1 ) N r ( r + s ) r ( r + l )
M , , , , - r ( r + s ) r (~ + s + N ) r ( r + l - N) "
Thus, f or mul a (25) is proved.
(25)
(26)
Corrollary
The coefficient of x u in PN,s(x) is
r ( ZN+s - 1 )
F ( N + s - 1 )
Using (26) with k = 0, and (27), we obt ai n t he i mpor t ant expressi on
1 F ( - 1 ) NN!
hu's= 2~j JL e"P-s[P~v's(P- 1)]2dp = ( 2 N+s - 1 ) F ( N+s - 1 ) "
(27)
(28)
5. Abscissas of the Gaussian Integration Formulas
The abscissas of t he Gaussi an i nt egr at i on f or mul a of or der N are the zeros of PN,s(P-1).
Grosswal s [-13] has st udi ed cert ai n propert i es concer ni ng t he zeros of t he simple Bessel
pol ynomi al s. Some of these pr oper t i es are also valid for t he general i zed Bessel pol ynomi al s,
[-14]. The quadr at ur e formul as of even or der have no real abscissas, t hose of odd or der have
exactly one real abscissa. All the abscissas have positive real part s and occur in compl ex
conj ugat ed pairs.
The zeros of PN(u-1) must be found by a numeri cal process of solving a pol ynomi al for its
root s, for instance Newt on- Raphs on' s i t erat i on process. Fi ndi ng an appr oxi mat e zero as
st art i ng value for t he i t erat i on process is facilitated by a cert ai n regul ari t y in t he di st ri but i on of
t he zeros. For fixed N and s, t he zeros lie ver y nearl y on a circle wi t h cent er on the negat i ve real
axis. The radius of this circle is appr oxi mat el y an i ncreasi ng linear funct i on of N and s. See
fig. 1 and fig. 2. For fixed N and s, t he angul ar di st ance bet ween t wo consecut i ve abscissas is
nearl y const ant (see [15]).
6. Weights of the Gaussian Integration Formulas
We shall give t wo formul as for the weights. These formul as can be pr oved in t he same way as
for the classical Gaussi an i nt egrat i on formulas.
Journal of Engineering Math., Vol. 5 (1971) 1-9
6 R. Piessens
I m (p)
2 5
N = 1 5 , , , ,
\ \
\ \
\ \
"~,\\
2 0 N = 1 2 %
N\
\k
\ \ x
\ \ ~ \ \
N = 1 0 % " \
1 5 ~ \ \ \ \ \
N = 9 % \ \ \\\ \ ~
\ \ \ x - k m \
10 N = 7 , , , \ \ \ \ ~ \ \ \ " \ .
\ \ Xx - - \ \ \ \ \ \
\ X x \ \ \ \ \ \
N =S % X ~\ \ ~ , ~, !
5 - N = 4 , \ \ \ ~ \ , t \ ' t
N = 3 ~ \ ~ t \ ~ \ ~ ~. ~ ]
I . I I . [ I I [ [ I
0 ~ ) 5 1 0 1 5 2 0 P,o(p)
Fi gur e 1. Di s t r i but i on of t he Ga u s s i a n absci ssas i n t he first qua dr a nt of t he compl ex pl ane for s = 1 a nd var i ous val ues
of N.
[ m (p)
2 0
15
10
N=5 s = 4 . 0
s = 2 . 0 " ,
I s = 1 . 0 o \ \ \ \
I s : 0 . 1 . < ' < \ \ \ \ \ \
5 1 - k\ \ \\ \
N=12 s =4. 0, t
S = 2 . 0 , " \ ,
S = 1 . 0 , ~ \ , " \ ,
\ \ \
\ \ \ \ \ "1~
\ \ \ \ ~ \
" ~ , " \ \ \ \ ~ \ k
k \ \ \
\ k \ \ \ \ \
3 ~ \ \ \ \ \ \ \
\ \ \
\ \ \ ~\
,\ \ ~\
'\\ ~\
15
\
t
i
' R e f p )
Fi gur e 2. Di s t r i but i on of t he Ga us s i a n absci ssas in t he first qua dr a nt of t he compl ex pl ane for N = 5 a nd N = 12 a nd
var i ous val ues o f s .
(i) Fi rst f ormul a Jot the k- t h weight A k
A k = [P M ,s ( u ~ - t ) ] (29)
whe r e hM, S is gi ven by (28) a n d whe r e uk is t he k- t h abscissa.
(ii) Second f ormul a f o r the k-t h weight
Us i ng t he r el at i ons (23) a n d (26) a nd t he Ch r i s t o f f e l - Da r b o u x f or mul a
N pM,.~(p-1)pM,s(y-1) pN+l , ~( p- 1) pN, s ( y- 1) _pN+l , s ( y- 1) pN, s ( p- 1)
M~=o hM,s au + l,~hu,s(p- t - - y - ~)
whe r e hu,s is gi ven by (28) a nd a s + 1,~ is t he coeffi ci ent whi c h ari ses in t he r e c ur r e nc e r el at i on
Journal o f Engineering Math. , Vol. 5 (1971) 1 9
Gaussian quadrature f ormul as 7
(22), formula (29) can be t ransformed in
( N- I ) ! 2f e~ + s _ 2 ]2
ak = (-- 1)u-1 F ( N + s - 1 ) N u ~ - - u C i "
1, s( k ) J
(30)
7. Some Practical Remarks
1. No convenient rule exists for det ermi ni ng at the outset the order N so t hat the desired
accuracy is obtained. The practical procedure is the use of a series of formul as (2) with increasing
order. If, for the given value of t, agreement occurs of two successive approxi mat i ons to wi t hi n
the desired accuracy, the last comput ed value is ret ai ned as definitive result. This procedure
has the di sadvant age of using different values of the argument s of the Laplace t ransform for
different values of the order of formul a (2). Increase in the order of the formul a makes no use of
previous evaluations of the Laplace t ransform F(p). This di sadvant age can be avoided if, in
combi nat i on with the Gaussi an rules, new quadr at ur e formul as are used, given by Piessens [ 16].
2. When the order N of the formul a (2) is large, the modul i of the coefficients A k are also
large. This may lead, from the numerical poi nt of view, to large losses of significance by can-
cellation.
3. Several aut hors have calculated tables of the coefficients Ak and Uk. We give here a survey
of these tables. Kryl ov and Skoblya give in [8] the coefficients Ak and Uk to 5 to 7 significant
figures for N = 1 (1)9 and for values of s which are integer multiples of 88 and 89 between the limits
88 and 3. In [19], t hey give for s = 1(1)5 and N = 1(1)15 the Ak'S and Uk'S to 20 significant figures
and for s=0.01(0.01)3.00 and N= 1(1)10 to 7 to 8 significant figures.
Skoblya [9] gives the Uk'S to 8 significant figures and the Ak'S to 7 significant figures for
N= 1(1)10 and s=0.1(0.1)3.0.
Salzer [6] gives the coefficients Uk, their inverses, and the product s Uk Ak to between 4 and 8
significant figures for N= 1(1)8 and s = 1.
In anot her paper, Salzer [7] gives the same quantities to between 12 and 15 significant
figures for N= 1(1)16 and s = 1.
St roud and Secrest [17] give the abscissas Uk and the weights Ak to 30 significant figures for
N=2( 1) 24 and s = 1.
Piessens [-18] gives the Uk'S and the Ak'S to 16 significant figures for N=6( 1) 12 and s=0. 1
10
(0.1)3.0, 3.5, 4.0, ~, ~, 4, ~, 7, ~, 7, 88 3, ~, -], 9, ~, 1, ~, 9~.
8. A Numerical Example
The first step in using formul a (2) is the det ermi nat i on of the value of s. In most cases, there is
no difficulty. However, for some Lapl ace transforms, there exists no conveni ent value of s.
Sometimes, F(p) can t hen be wri t t en as a sum of two or more Lapl ace t ransforms t hat can be
inverted separately, each wi t h a different value of s.
For instance
F ( p ) = p O.Sexp(_p-O.5)
can be written as
F (p) = p-O.S cosh (p-O.5)_p-O.5 sinh(p-O.5).
(31)
(32)
Here, the first t erm of the second member can be inverted wi t h s = 0.5 and the second t erm wi t h
s =l . 0.
The exact original funct i on is
1 f u exp ( - u2/4t) Jo (2x/u) du.
f ( O- o
Using the approxi mat e sol ut i on
Jour nal o f Engineering Ma t h . , Vol. 5 (1971) 1 9
8 R. Pi e s s e n s
6 6
f ( t ) = t - ~ Z A ? ' 5 ) G I ( u ? "5)) - A(ka'~176 (33)
k=l k =l
where
G i (x) = cosh [ 4 ( t / x ) ]
and
G 2 (x) = x / ( x / t ) sinh [ 4 ( t / x ) ]
and where A(k ~) and U(k ~) are t he wei ght s and absci ssas for s = a and N = 6, given in [18], we obt ai n
the fol l owi ng errors (by er r or we mean J t rue v a l u e - a p p r o x i ma t e val ue I)
t exact original function error
1.0 -0. 01072342858155 < 1.10 - 14
10.0 - 0. 02478598394520 1.10 i4
20.0 -0. 00308187970796 5.10 - 14
50.0 0.00271995009362 7.10 - 14
100.0 0.00021092905918 4.10 -12
Acknowledgement
The aut hor wishes t o express his t hanks t o t he referee for val uabl e comment s , par t i cul ar l y
concer ni ng t he der i vat i on of t he or t hogonal pol ynomi al s of sect i on 3. The cal cul at i ons of t he
numer i cal exampl e are carri ed out on t he I BM 360/44 of the Comput i ng Cent r e of t he Uni -
versi t y of Leuven.
R E F E R E NC E S
[ I ] R. E. Bellman, R. E. Kalaba and J. Lockett, Numerical Inversion oJ'the Laplace Transform, American Elsevier
Publishing Company, New York (1966).
[2] C. Lanczos, Applied Analysis, Prentice Hall, Englewood Cliffs, N.J. (1956).
[3] A. Papoulis, A New Method of Inversion of the Laplace Transform. Quart. Appl. Math., 14 (1956) 405-414.
[4] W. T. Weeks, Numerical Inversion of Laplace Transforms using Laguerre Functions, Journal of the A.C.M. 13
(1966) 419426.
[5] H. Dubner and J. Abate, Numerical Inversion of Laplace Transforms and the Finite Cosine Transform, Journal
of the A.C.M. 15 (1968) 115-123.
[6] H. E. Salzer, Orthogonal Polynomials arising in the Numerical Evaluation of Inverse Laplace Transforms,
M.T.A.C. 9, (1955) 164-177.
[7] H. E. Salzer, Additional Formulas and Tables for Orthogonal Polynomials originating from Inversion Integrals,
J. Math. Phys., 40 (1961) 72 86.
[8] V. I. Krylov and N. S. Skoblya, On the numerical Inversion of the Laplace Transform, Inzh.-Fiz. Zh. 4 (1961)
85 101 (Russian).
[9] N. S. Skoblya, Tables for the Numerical Inversion of the Laplace Transform, Minsk. Izdat. Akad. Nauk BSSSR
(1964) (Russian).
[10] P. J. Davis and P. Rabinowitz, Numerical Integration, Blaisdell Publishing Co., Waltham (1967).
[11] A. Erd61yi, et al., Higher Transcendental Functions, McGraw-Hill Book Company, New York (1953).
[12] H. L. Krall and O. Frink, A new Class of orthogonal Polynomials: the Bessel Polynomials, Trans. Amer. Math.
Soc., 65 (1949) 10(~115.
[13] E. Grosswald, On some algebraic properties of the Bessel polynomials, Trans. Amer:Math. Soc., (1951) 197-210.
[14] A. Van Rossum, A note on the location of the zeros of generalized Bessel polynomials and totally positive
polynomials. Nieuw Archi ef voor Wiskunde 17 (1969) 142 149.
[15] R. Piessens, Numerieke metodes voor de inversie van de Laplace Transformatie, Doctoraatstesis, Katholieke
Universiteit Leuven, 1970.
[16] R. Piessens, New Quadrature Formulas for the Numerical Inversion of the Laplace Transform, BIT, 9 (1969)
351-361.
Journal of Engineering Math., Vol. 5 (1971) 1 9
Gaussian quadrature formulas 9
[17] A. Stroud and D. Secrest, Gaussian Quadrature Formulas, Prentice-Hall, Englewood Cliffs, N.J. (1966).
[18] R. Piessens, Gaussian Quadrature formulas for the Numerical Integration of Bromwich's Integral and the
Inversion of the Laplace Transform, Inst. of Applied Math., University of Leuven, (1969).
[19] V. I. Krylov and N. S. Skoblya, Handbook of Numerical Inversion of Laplace Transforms. Israel Program for
Scientific Translations, Jerusalem (1969) (translated from Russian).
Journal of Engineering Math., Vol. 5 (1971) 1-9