Professional Documents
Culture Documents
ch09 PDF
ch09 PDF
”
The Transforms and Applications Handbook: Second Edition.
Ed. Alexander D. Poularikas
Boca Raton: CRC Press LLC, 2000
9
The Hankel Transform
ABSTRACT Hankel transforms are integral transformations whose kernels are Bessel functions. They
are sometimes referred to as Bessel transforms. When we are dealing with problems that show circular
symmetry, Hankel transforms may be very useful. Laplace’s partial differential equation in cylindrical
coordinates can be transformed into an ordinary differential equation by using the Hankel transform.
Because the Hankel transform is the two-dimensional Fourier transform of a circularly symmetric
function, it plays an important role in optical data processing.
x 2 y ′′ + xy ′ + (x 2 − p 2 )y = 0 (9.1)
where p is a parameter.
Equation (9.1) can be solved using series expansions. The Bessel function Jp(x) of the first kind and
of order p is defined by
The Bessel function Yp(x) of the second kind and of order p is another solution that satisfies
J (x) Y p ( x ) 2
W ( x ) = det p = .
J ′p ( x ) Y p′ ( x ) πx
Properties of Bessel function have been studies extensively (see References 7, 22, and 26).
Elementary properties of the Bessel functions are
1. Asymptotic forms.
2 1 1
J p (x ) ~ cos x − p π − π , x →∞. (9.3)
πx 2 4
2. Zeros. Jp(x) and Yp(x) have an infinite number of real zeros, all of which are simple, with the
possible exception of x = 0. For nonnegative p the sth positive zero of Jp(x) is denoted by jp,s . The
distance between two consecutive zeros tends to π: lim ( jp,s+1 – jp,s ) = π.
s→∞
3. Integral representations.
p
1
x
2 π
Jp(x) =
π1/ 2 Γ ( p + 1 / 2) ∫0
cos( x cos θ )sin 2 p θ d θ . (9.4)
∫
1
J p (x ) = cos(x sin θ − p θ)dθ
π 0
(9.5)
j −n π
=
π ∫ 0
e jx cos θ cos(p θ)dθ .
4. Recurrence relations.
2p
J p−1 ( x ) − J ( x ) + J p+1 ( x ) = 0 (9.6)
x p
p
J ′p ( x ) = J p−1 ( x ) − J (x) (9.8)
x p
p
J ′p ( x ) = − J p+1 ( x ) + J (x) . (9.9)
x p
∫ 0
f (r ) r dr
∞ ∞
∫ ∫
1
s ds f (r) J p (sr) J p (su)r dr = [ f (u +) + f (u −)] (9.10)
0 0 2
provided that ƒ(r) satisfies certain Dirichlet condtions.
For a proof, see Reference 26. The reader should also refer to Chapter 1, Section 5.6 for more
information regarding Bessel functions.
∞
Fν (s ) ≡ ν { f (r)} ≡
∫0
r f (r) J ν (sr)dr . (9.11)
If v > –1/2, Hankel’s repeated integral immediately gives the inversion formula
∞
f (r ) = −ν1{Fν (s )} ≡
∫ 0
sFν (s )J ν (sr )ds . (9.12)
The most important special cases of the Hankel transform correspond to v = 0 and v = 1. Sufficient but
not necessary conditions for the validity of (9.11) and (9.12) are
1. f (r) = O(r –k), r → ∞ where k > 3/2.
2. f ′ (r) is piecewise continuous over each bounded subinterval of [0, ∞).
3. f (r) is defined as [ f (r+) + f (r–)]/2.
These conditions can be relaxed.
∞ ∞
∫ ∫
1
Φ(ζ, η) = f (x , y )e − j( x ζ , y η )dx dy . (9.13)
2π −∞ −∞
x = r cos θ , y = r sin θ
and
ζ = s cos ϕ , η = s sin ϕ .
∞ 2π
∫ ∫
1
φ (s cos ϕ, s sin ϕ) ≡ F (s , ϕ) = r dr e − j r s cos(θ−ϕ ) f (r)d θ
2π 0 0
∞ 2π
∫ ∫
1
= rf (r)dr e − j r s cos α d α
2π 0 0
∞
=
∫ 0
rf (r) J 0 (rs )d r .
This result shows that F(s,ϕ) is independent of ϕ, so that we can write F(s) instead of F(s, ϕ). Thus, the
two-dimensional Fourier transform of a circularly symmetric function is, in fact, a Hankel transform of
order zero.
This result can be generalized: the N-dimensional Fourier transform of a circularly symmetric function
of N variables is related to the Hankel transform of order N/2 – 1. If ƒ(r,θ) depends on θ, we can expand
it into a Fourier series
f (r , θ) = ∑ f (r )e
n =−∞
n
j nθ
(9.14)
and, similarly
∞ ∞ ∞
∫ ∫ ∑ F (s)e
1
F (s , ϕ) = r dr e − j r s cos(θ−ϕ ) f (r, θ)d θ = jn ϕ
(9.15)
2π 0 0
n =−∞
n
where
2π
∫
1
f n(r ) = f (r , θ)e − j n θd θ (9.16)
2π 0
and
2π
∫
1
Fn(s ) = F (s , ϕ)e − j n ϕd ϕ . (9.17)
2π 0
2π 2π ∞
∫ ∫ ∫
1
Fn (s ) = e − jn ϕ dϕ dθ f (r, θ)e j s r cos(θ−ϕ )r dr
(2 π) 2 0 0 0
2π ∞ 2π ∞
∫ ∫ ∫ ∑f
1
= e − jn ϕ dϕ r dr e j s r cos(θ−ϕ )dθ × (r)e jm θ
(2 π) 2
m
0 0 0
m=−∞
∞ 2π
∫ ∫
1
= r dr e − jn α e j s r cos α f n (r)dα
(2 π) 0 0
∞
=
∫0
rf n (r) J n (sr)dr
= n { f n (r)}.
f n (r ) = n {Fn (s )} . (9.18)
Fν (s ) = ν { f (x )} .
Then
ν +1 ν −1
G ν (s ) = ν { f ′(x )} = s Fν−1(s ) − Fν+1(s ) . (9.19)
2ν 2ν
Proof
∞
Gν ( s) =
∫ xf ′( x ) J ( sx )dx
0
ν
∫
d
= [ xf ( x ) Jν ( sx )]∞0 − f (x) [ xJ ( sx )]dx .
0 dx ν
d sx
[ xJ ( sx )] = [(ν + 1) Jν−1 ( sx ) − (ν − 1) Jν+1 ( sx )] .
dx ν 2ν
2 2
d 2 1 d ν 1 d d ν
∆ν ≡ 2 + − = r −
dr r dr r r dr dr r
∂2 1 ∂ 1 ∂2 ∂2
∇2 = + + +
∂r 2 r ∂r r 2 ∂θ 2 ∂z 2
ν {∆ ν f (r )} = −s 2 ν { f (r )} . (9.20)
∞ d df ν 2
ν {∆ ν f (r)} =
∫
0
r
dr dr
−
r
f (r) J ν (sr)dr
∞ 2 ν2
∫
s
= s J ν′′ (sr) + J ν′ (sr) − 2 J ν (sr) f (r)r dr
0 x r
∞
= −s 2
∫0
rf (r) J ν (rs )dr
= −s 2 ν { f (r)}.
This property is the principal one for applications of the Hankel transforms to solving differential
equations. See References 2, 3, 24, 25, and 28.
3. Similarity.
1 s
ν { f (ar )} = Fν . (9.21)
a2 a
4. Division by r.
s
ν {r −1 f (r )} = [F (s ) + Fν+1(s )] . (9.22)
2 ν ν−1
5.
d
ν r ν−1 [r 1− ν f (r)] = −sFν−1(s ) . (9.23)
dr
6.
d
ν r − ν−1 [r ν+1 f (r )] = sFν+1(s ) . (9.24)
dr
7. Parseval’s theorem. Let
Fν (s ) = ν { f (r )}
and
G ν (s ) = ν {g (r )} .
Then
∫
0
Fν(s )G ν(s )s ds =
∫ 0
Fν(s )s ds
∫ r g (r )J (sr )dr
0
ν
∞ ∞
=
∫ 0
r g (r )dr
∫ sF (s)J (sx)ds
0
ν ν (9.25)
∞
=
∫ r g (r )f (r )dr .
0
Example 9.1
1 − s 2 / 4a
{e −ar } =
2
e , a >0 .
2a
Example 9.2
a a
From the relationship ∫ 0 r J 0 (sr) dr = ∫ 0 (1 / s )(d / dr)[rJ 1(sr)] = [aJ1(as)]/s (see Chapter 1, Section 5.6),
we conclude that
aJ 1(as )
0 {pa (r )} =
s
Example 9.3
∞
From the identity ∫ 0 J 0 (sr) dr = 1 / s , s > 0 (see Chapter 1, Section 5.6), we obtain
1 1
0 = .
r s
Example 9.4
∞
Since ∫ 0 r δ(r − a ) J 0 (sr) dr = aJ 0 (as ) (see Chapter 1, Section 2.4), we obtain
Convolution Identity
Let ƒ1(r) and ƒ2(r) have Hankel transforms F1(s) and F2(s), respectively. From Section 2.4.1 above, we have
Hence, we have
1
0 { f1(r)★★f 2 (r)} = { f (r)★★f 2 (r)} = 2 πF1(s )F2 (s ) .
2π ( 2) 1
Example 9.5
If ƒ1(r) = ƒ2(r) = [J1(ar)]/r then from the convolution identity above, we obtain
J 2 (ar) 1
0 2 π 1 2 = 2 pa (s )★★pa (s )
r a
where
s s s2
pa (s )★★pa (s ) = 2 cos −1 − 1− 2 a2 .
2a a 4a
Hence,
J 2 (ar) s s s2
0 2 π 1 2 = 2 cos −1 − 1 − 2 p2a (s )
r 2a a 4a
1 | s | ≤ 2a
p2a (s ) =
0 otherwise .
Example 9.6
From the definition, the Hankel transform of r vh(a – r), a > 0 is given by
∫ ∫
a as
1
ν {r νh (a − r )} = r ν+1J ν (sr )dr = x ν+1J ν (x )dx
0 s ν+ 2 0
since h(a – r) is the unit step function with value equal to 1 for r ≤ a and 0 for r > a. But ∫ t vJv–1(t) dt =
t vJv(t) + C (see Chapter 1, Section 5.6, Table 5.6.1) and hence,
∞ ∞ a
∫ ∫
1 − t
ν {r ν−1e − ar } = r νe − ar J ν(sr )dr = t ν J ν(t )e s
dt
0 s ν+1 0
1 a
= ν+1
L t ν J ν(t ); p =
s s
where we set t = rs and L is the Laplace transform operator (see also Chapter 5). But
∞
(−1)n t 2n+2 ν
t ν J ν(t ) = ∑n=0
n ! Γ(n + ν + 1)22n+ ν
and, hence,
∞
(−1) n
ν
{t J ν (t ); p} = ∑
n= 0
n! Γ(n + ν + 1)2 2n+ ν
L{t 2n+ 2 ν ; p}
∞
(−1) n Γ(2n + 2 ν + 1)
= ∑ n! Γ(n + ν + 1)2
n= 0
2 n + ν 2 n + 2 ν+1
p
.
From Chapter 1, Section 2.5, the duplication formula of the gamma function gives the relationship
Γ(2n + 2 ν + 1) 1 2n+ 2 ν 1
= 2 Γ n + ν +
Γ(n + ν + 1) π 2
1
∞ (−1)n Γ n + ν + n
ν
2 1
∑
2
{t ν J ν (t ); p} = 2 .
π p 2 ν+1 n= 0
n! p
The last series can be summed by using properties of the binomial series
∞ ∞
− b n (−1) n Γ(n + b) n
(1 + x )− b = ∑
n= 0
x =
n ∑
n= 0
n ! Γ(b)
x , |x| < 1
π + 1
s
and, hence,
1 1
2 ν Γ ν + s ν 2 ν Γ ν +
1 2 2 1
ν {r ν−1e −ar } = ν+1 1
= 1
, ν>− .
s ν+ ν+ 2
a 2 2
π (a 2 + s )
2 2
π + 1
s
If we set v = 0 and a = 0 in the above equation, we obtain the results of Example 9.3. If we set v = 0, we
obtain
1
0 {r −1e −ar } = , a >0 .
a +s2
2
Example 9.8
The Hankel transform 0{e –ar} is given by
d 2 −
1
0 {e − ar } = {r J 0 (sr); r → a } = − (s + a 2 ) 2
da
a
= , a >0
[s + a 2]3/ 2
2
d 2 rn ( x ) 1 drn ( x )
2
+ + rn ( x ) = 2nrn+1 ( x ) ,
dx x dx
(1 − s 2 )R n ( s ) = 2nR n+1 ( s )
or
1 − s2 (1 − s 2 )n
Rn+1(s ) = Rn(s ) = L = n R1(s ) .
2n 2 n!
J (r ) (1 − s 2 )n−1
0 n n = n−1 p1(s )
r 2 (n − 1)!
G ( s ) = 2πF ( s )H ( s ) .
Since 0{J0(ar)} = [δ(s – a)]/a (see Example 9.4) and ϕ(s)δ(s – a) = ϕ(a)δ(s – a), we conclude that if
the input is ƒ(r) = J0(ar), then
2π 2πH (a)
G ( s) = δ ( s − a) H ( s ) = δ ( s − a) .
a a
9.5 Applications
∂ 2υ 1 ∂υ ∂ 2υ
∇ 2υ ≡ + + =0. (9.26)
∂r 2 r ∂r ∂z 2
∂υ
(r, 0) = 0, r >1 . (9.28)
∂z
In (9.27), υ0 is the potential of the disc. Condition (9.28) arises from the symmetry about the plane z = 0.
Let
V (s , z ) = 0 {υ (r, z )}
so that
∂ 2V
0 {∇ 2υ } = −s 2V (s , z ) + (s , z ) = 0 .
∂z 2
V ( s , z ) = A ( s )e− sz + B ( s )e sz
where A and B are functions that we have to determine using the boundary conditions.
Because the potential vanishes as z tends to infinity, we have B(s) 0. By inverting the Hankel
transform, we have
∞
υ (r , z ) =
∫ 0
sA(s )e − sz J 0 (sr)ds . (9.29)
∞
υ (r, 0) =
∫ 0
sA(s ) J 0 (rs )ds = υ 0 , 0 ≤ r <1 (9.30)
∂υ ∞
∂z
(r, 0) =
∫ 0
s 2 A(s ) J 0 (rs )ds = 0, r >1 . (9.31)
Using entries (9.8) and (9.9) of Table 9.1 (see Section 9.11), we see that A(s) = sin s/s 2 so that
2υ 0 ∞
∫
sin s − sz
υ (r , z ) = e J 0 (sr)ds . (9.32)
π 0 s
In Figure 9.1, the graphical representation of υ (r,z) for υ0 = 1 is depicted on the domain 0 ≤ r ≤ 2, 0 ≤
z ≤ 1. The evaluation of υ (r,z) requires numerical integration.
Equations (9.30) and (9.31) are special cases of the more general pair of equations
∫
0
f (t )t 2α Jν ( xt )dt = a( x ), 0≤ x <1 (9.33)
∫ 0
f (t ) Jν ( xt )dt = 0, x >1 (9.34)
2− α x 1−α
∫ ∫ a( t ) t
1 s
d
f (x) = s − ν−α Jν+α ( xs ) ν+1
( s 2 − t 2 )α dt ds , − 1 < α < 0 (9.35)
Γ (α + 1) 0 ds 0
(2 x )1−α
∫ ∫
1 s
f (x) = s − ν−α +1 Jν+α ( xs ) a(t )t ν+1 ( s 2 − t 2 )α −1 dt ds , 0 < α < 1 . (9.36)
Γ (α ) 0 0
If a(x) = x β, and α < 1, 2α + β > –3/2, α + v > –1, v > –1, then
β + ν −( 2α +β+1)
Γ 1 + x
2
∫
x
f (x ) = t α +β+1 J ν+α (t )dt . (9.37)
β + ν 0
2 α Γ 1 + α +
2
Γ (ν + 1)
f (x) = Jν+α +1 ( x ) . (9.38)
(2 x )α Γ (ν + α + 1)
∂υ (r, z )
−K = Q, r < a, z = 0
∂z (9.39)
= 0, r > a, z = 0 .
∂2V
( s, z ) − s 2V ( s, z ) = 0 . (9.40)
∂z 2
We can now transform also the boundary condition, using formula (3) in Table 9.1:
V ( s , z ) = A ( s )e− sz .
A ( s ) = Qa J1 (as )/(Ks 2 ) .
∫
Qa
υ (r , z ) = e − sz J 1(as ) J 0 (rs )s −1ds . (9.42)
K 0
∂ 2υ 1 ∂υ ∂ 2υ
+ + = 0, z > 0, 0 < r < ∞
∂r 2 r ∂r ∂z 2 (9.43)
υ (r, 0) = f (r).
∂2V
( s, z ) − s 2V ( s, z ) = 0
∂z 2
and
∞
V ( s , 0) =
∫ rf (r ) J ( sr )dr .
0
0
The solution is
∞
V ( s , z ) = e− sz
∫ rf (r ) J ( sr )dr
0
0
so that
∞ ∞
υ (r , z ) =
∫0
s e − sz J 0 (sr)ds
∫ 0
p f (p) J 0 (sp)dp . (9.44)
f (r ) = h(a − r )
∞
υ (r , z ) = a
∫
0
e − sz J 0 (sr) J 1(as )ds . (9.45)
where ϕ(r,z) satisfies Laplace’s Equation (9.26). The boundary conditions are
∂ 2Φ
( s , z ) − s 2Φ ( s , z ) = 0 (9.48)
∂z 2
qe− sl
Φ( s , ± l ) = − (9.49)
s
A ( s )e− sz + B ( s )e sz
q e − sl
A(s )e + sl + B(s )e − sl = −
s
q e − sl
A(s )e − sl + B(s )e sl = − .
s
Hence,
q e − sl
A(s ) = B(s ) = −
2s cosh(sl)
and
q e− sl cosh( sz )
Φ( s , z ) = − .
s cosh( sl )
Hence,
∫e
q − sl cosh( sz )
ϕ(r , z ) = −q J ( sr )ds . (9.50)
r +z
2 2 0 cosh( sl ) 0
∫ rf (r ) J (αr )dr .
1
Fν (α ) = H ν { f , α } = ν (9.51)
0
hJ ν(x ) + x J ν′ (x ) = 0
The transformation (9.51) with α = j v,s , s = 1, 2, … is the finite Hankel transform. It maps the function
ƒ(r) into the vector (Fv (jv,1), Fv(jv, 2), Fv(j v,3) …). The inversion formula can be obtained from the well-
known theory of Fourier-Bessel series
∑J
Fν ( jν , s )
f (r ) = 2 2
J ν ( jν , s r ) . (9.52)
s =1 ν+1 ( jν , s )
The transformation (9.51) with α = βv ,s , s = 1, 2, … is the modified finite Hankel transform. The inversion
formula is
∞
β 2ν , s Fν (β ν , s ) Jν (β ν , s r )
f (r ) = 2 ∑ h +β
s =1
2 2
ν,s − ν 2 Jν2 (β ν , s )
. (9.53)
∂ 2υ 1 ∂υ ∂υ
+ = , 0 ≤ r < 1, t > 0 (9.54)
∂r 2 r ∂r ∂t
υ (r, 0) = 1, 0 ≤ r ≤1 . (9.55)
The radiation at the surface of the cylinder is described by the mixed boundary condition
∂υ
(1, t ) = − hυ (1, t ) (9.56)
∂r
dV
(β , t ) = −β 20, s V (β 0, s , t ) (9.57)
dt 0, s
where
∫
1
V (α, t ) = rυ (r, t ) J 0 (αr)dr
0
so that
J1 (α )
∫ rJ (αr )dr =
1
V (α , 0) = . (9.58)
0
0
α
J1 (β 0, s ) −β20 , s t
V (β 0, s , t ) = e .
β 0, s
∞
β 0, s J 1(β 0, s ) J 0 (β 0, s r)
υ (r , t ) = 2 ∑e
j=1
− β 20 , s t
h 2 + β 20, s J 02 (β 0, s )
. (9.59)
Z ν ( s , r ) = Jν ( sr )Yν ( s ) − Yν ( sr ) Jν ( s ) . (9.60)
Using this function as a kernel, we can construct the following transform pair:
∞
Fν ( s ) =
∫ rf (r )Z ( s, r )dr
1
ν (9.61)
∫ sF ( s) J ( s) + Y ( s) ds
Zν ( s, r )
f (r ) = ν 2 2
(9.62)
0
ν ν
The inversion formula follows immediately from Weber’s integral theorem (see Watson26):
∞ ∞
∫ ∫
1
u du f ( s )Z ν (r , u)Z ν ( s , u) s ds = [ Jν2 (r ) + Yν2 (r )][ f (r + ) + f (r − )] . (9.63)
1 0 2
For this reason, we will refer to (9.61) and (9.62) as the Weber transform. This transform has the following
important property:
If
1 ν2
f ( x ) = g ′′ ( x ) + g′( x ) − 2 g( x ) (9.64)
x x
then
2
Fν ( s ) = − s 2 G ν ( s ) − g (1) . (9.65)
π
We may expect that this transform is useful for solving Laplace’s equation in cylindrical coordinates, with
a boundary condition at r = 1.
Example
We want to compute the steady-state temperature u(r,z) in a horizontal infinite homogeneous slab of
thickness 2, through which there is a vertical circular hole of radius 1. The horizontal faces are held at
temperature zero and the circular surface in the hole is at temperature T0.
The mathematical model is
∂2u 1 ∂u ∂2u
+ + =0
∂r 2 r ∂r ∂z 2
u (r , l) = u (r , − l) = 0 (9.66)
u (1, z ) = T0 .
∂ 2 U0 2
( s , z ) − s 2 U0 ( s , z ) = T .
∂z 2
π 0
The solution of this ordinary differential equation, satisfying the boundary condition, is
Consequently, we have
∞
1 cosh sz Z 0(s , r )
∫
2T0
u (r , z ) = −1 (r > 1) .
s cosh sl J 02(s ) + Y02(s )
ds (9.67)
π 0
∫ J ( px ) f ( x )dx
a
I (a, p, ν) = ν (9.68)
0
where a is a positive real number or infinite. In most cases, analytical integration of (9.68) is impossible,
and numerical integration is necessary. But integrals of type (9.68) are difficult to evaluate numerically if
1. The product ap is large.
2. a is infinite.
3. ƒ (x) shows a singular or oscillatory behavior.
In cases 1 and 2, the difficulties arise from the oscillatory behavior of Jv(x) and they grow when the
oscillations become stronger.
We give here a survey of numerical methods that are especially suited for the evaluation of I (a, p, v)
when ap is large or a is infinite. We restrict ourselves to cases where ƒ(x) is smooth, or where ƒ(x) =
x α g (x) where g (x) is smooth and α is a real number.
∫x
1
I (a, p, ν) = aα +1 α
Jν (apx ) g (ax )dx . (9.69)
0
We assume that α + v > – 1. If (α+v) is not an integer, then there is an algebraic singularity of the
integrand at x = 0. If ap is large, then the integrand is strongly oscillatory.
If (α + v) is an integer and ap is small, classical numerical integration methods, such as Romberg
integration, Clenshaw-Curtis integration of Gauss-Legendre integration (see Davis and Rabinowitz 4 ) are
applicable. If (α + v) is not an integer and ap is small, the only difficulty is the algebraic singularity at
x = 0, and Gauss-Jacobi quadrature or Iri-Moriguti-Takesawa (IMT) integration 4 can be used. If ap is
large, special methods should be applied that take into account the oscillatory behavior of the integrand.
We describe two methods here.
∑ ∫
a
I (a , p , ν) = (−1)k +1I k + J ν( px )f (x )dx (9.70)
j ν, N p
k =1
where
∫
j ν, k p
Ik = | J ν( px )| f (x )dx (9.71)
j ν , k −1 p
and where N is the largest natural number for which jv,N ≤ ap. This means that N is large when ap is large.
Using a transformation attributed to Longman,10 the summation in Equation (9.70) can be written as
S= ∑(−1)
k =1
k +1 1
2
1
4
1
I k = I 1 − ∆I 1 + ∆2I 1 + L + (−1) p−12− p ∆p−1I 1
8
1 1 1
+ (−1)N −1 I N + ∆I N −1 + ∆2I N −2 + L + 2− p ∆p−1I n− p+1
2 4 8
Assuming now that N and p are large and that high-order differences are small, the last bracket may be
neglected and
1 1 1
S I 1 − ∆I 1 + ∆2I 1 − L
2 4 8
(9.72)
N −1 1 1 1 2
+ (−1) 2 I N + 4 ∆I N −1 + 8 ∆ I N −2 + L .
The summations in (9.72) may be truncated as soon as the terms are small enough. For the evaluation
of Ik , k = 1, 2, …, classical integration methods (e.g., Lobatto’s rule) can be used, but special Gauss
quadrature formulas (see Piessens16) are more efficient. If the integral I1 has an algebraic singularity at
x = 0, then the Gauss-Jacobi rules or the IMT rule are recommended.
+1
I=
∫ −1
w ( x ) g ( x )dx . (9.73)
g (x ) ∑′ c T (x),
k =0
k k −1≤ x ≤ 1 . (9.74)
Here the symbol Σ′ indicates that the first term in the sum must be halved. For the computation of the
coefficients c k in Equation (9.74) several good algorithms, based on the Fast Fourier Transform, are
available.
The integral in (9.73) can now be approximated by
I ∑′c M
k =0
k k (9.75)
where
+1
Mk =
∫ −1
w ( x )Tk ( x )dx
∫x
1
α
I= Jν (ωx ) g ( x )dx . (9.76)
0
If
g (x ) ∑ ′ c T ′(x)
k =0
k k (9.77)
then
I ∑ ′ c M (ω, ν, α)
k =0
k k (9.78)
where
∫ x J (ωx)T
1
α
M k (ω, ν, α) = ν k
★
(x )dx . (9.79)
0
These modified moments satisfy the following homogeneous, linear, nine-term recurrence relation:
3ω 2
− 2(k 2 − 4) + 6(ν 2 − α 2 ) − 2(2α − 1) − M (9.80)
8 k
ω2 ω2
+ (k − 3)(k − 3 − 2α) + α 2 − ν 2 − M + M =0.
4 k − 2 16 k − 4
Because of the symmetry of the recurrence relation of the shifted Chebyshev polynomials, it is convenient
to define
T−★k (x ) = Tk★ (x ), k = 1, 2, 3, …
and consequently
M − k (ω , ν, α ) = M k (ω , ν, α ) .
To start the recurrence relation with k = 0, 1, 2, 3, … we need only M 0, M 1, M 2, and M 3 . Using the
explicit expressions of the shifted Chebyshev polynomials, we obtain
M 0 = G (ω , ν, α )
M1 = 2G (ω , ν, α + 1) − G (ω , ν, α )
(9.81)
M 2 = 8G (ω , ν, α + 2) − 8G (ω , ν, α + 1) + G (ω , ν, α )
where
∫x
1
α
G (ω , ν, α ) = Jν (ωx )dx . (9.82)
0
Because
ω 2 G (ω , ν, α + 2) = [ν 2 − (α + 1) 2 ]G (ω , ν, α )
(9.83)
+ (α + ν + 1) Jν (ω ) − ωJν−1 (ω )
ν + α + 1
α
Γ
2 2 2
G (ω , ν, α ) = − ( g1 cos θ + g 2 sin θ ) (9.85)
ω α +1 ν − α + 1 πω 3
Γ
2
where
θ = ω − νπ / 2 + π / 4
and
g1 ~ ∑(−1) a
k =0
k
2k ω −2k , ω→∞
g2 ~ ∑(−1) a
k =0
k
2 k +1 ω −2k −1, ω→∞
(1 / 2 − ν)k (1 / 2 + ν)k
ak = bk
2k k !
b0 = 1
2(k + 1)(α − k − 1 / 2)
bk +1 = 1 + b .
(ν − k − 1 / 2)(ν + k + 1 / 2) k
ν−1
∫ ∫ ∑J
1 1
2
J 2 ν(ωx )dx = J 0(ωx )dx − 2 k +1 (ω)
0 0 ω k =0
ν
1 − J 0(ω) 2
∫ ∑J
1
J 2 ν+1(ωx )dx = − (ω) .
0 ω ω k =1
2k
∫ J (ωx )dx
1
0
0
y1, k ~ k −2
y 2, k ~ k −4
y 3, k ~ k −2(α +1)−2 ν
~ k −2(α+1)lnk , if ν = 0
k
ω
y 5, k ~ y 6, k ~ e kk α
4k
k
4k
y 7, k ~ y 8, k ~ e − kk α
ω
and
1
M k (ω, ν, α) ~ − J (ω)k −2
2 ν
(9.87)
ων
+ (−1) 2
k −3 ν− 2α −1
Γ(ν + 1)
[ ]
cos π(α + 1) Γ(2α + 2)k −2α − 2 ν− 2 .
The asymptotically dominant solutions are y7,k and y8,k . The asymptotically minimal solutions are y5,k and
y6,k. We may conclude that forward and backward recursion are asymptotically unstable. However, the
instability of forward recursion is less pronounced if k ≤ ω/2. Indeed, practical experiments demonstrate
that Mk(ω, v, α) can be computed accurately using forward recursion for k ≤ ω/2. For k > ω/2 the loss
of significant figures increases and forward recursion is no longer applicable. In that case, Oliver’s
algorithm14 has to be used. This means that (9.80) has to be solved as a boundary value problem with
six initial values and two end values. The solution of this boundary value problem requires the solution
of a linear system of equations having a band structure.
An important advantage of the MCC method is that the function evaluations of g, needed for the
computation of the coefficients ck of the Chebyshev series expansion, are independent of the value of ω.
Consequently, the same function evaluations may be used for different values of ω, and have to be
computed only once.
Numerical examples can be found in References 19 and 20.
∞
I ( p, ν) =
∫0
Jν ( px ) f ( x )dx . (9.88)
I ( p, ν) = ∑ (−1)
k =1
k +1
Ik (9.89)
where
∫
j ν,k p
Ik = J ν( px ) f (x )dx . (9.90)
j ν , k −1 p
1 1 1 1
I ( p, ν) = I1 − ∆I1 + ∆2 I1 − ∆3 I1 + L . (9.91)
2 4 8 16
It is not always desirable to start the convergence acceleration with I1, but with some later term, say Im,
so that
1
∫
j ν, m −1 p
1 1
I ( p , ν) = J ν( px ) f (x )dx + (−1)m −1 I m − ∆I m + ∆2I m − L .
0 2 4 8
Other convergence accelerating methods, for example the -algorithm,23 are also applicable (for an
example, see Reference 21).
( x / 2) ν
∫ (1 − t )
1
2 ν−1/ 2
Jν ( x ) = 2 cos( xt )dt (9.92)
Γ (ν + 1 / 2) π 0
2( p / 2) ν
∫ (1 − t )
1
2 ν−1/ 2
I ( p, ν) = F (t )dt (9.93)
Γ (ν + 1 / 2) π 0
where
∞
F (t ) =
∫x 0
ν
f ( x )cos( pxt )dx . (9.94)
We assume that the integral in (9.94) is convergent. If we want to evaluate (9.93) using an N-point Gauss-
Jacobi rule, then we have to compute the Fourier integral (9.94) for N values of t. Because F(t) shows a
peaked or even a singular behavior especially when ƒ(x) is slowly decaying, a large enough N has to be
chosen.
This method is closely related to Linz’s method,8 which is based on the Abel transformation of I (p,v).
∫ J ( px ) f ( x )dx + R ( p, a)
a
I ( p, ν) = ν (9.95)
0
where
∞
R ( p, a) =
∫
a
Jν ( px ) f ( x )dx . (9.96)
The first integral in the right side of (9.95) can be computed using the methods of Section 9.9.
If a is sufficiently large and ƒ is strongly decaying, then we may neglect R(p,a).
If
c1 c 2
f (x) ~ + +L (9.97)
x x2
is an asymptotic series approximation which is sufficiently accurate in the interval [a, ∞), then
∑ ∫
J ν( px )
R( p , a) ck dx . (9.98)
k
a xk
Longman9 has tabulated the values of the integrals in (9.98) for some values of v and ap.
Using Hankel’s asymptotic expansion,1 for x → ∞
2
Jν ( px ) ~ [P ( px )cos χ − Qν ( px )sin χ ] (9.99)
πpx ν
where χ = px – (v/2+1/4)π, and where Pv(x) and Qv(x) can be expressed as a well-known asymptotic
series, R(p,a) can be written as the sum of two Fourier integrals.
Especially, if a = (8 + v/2 + 1/4)π/p, we have
∫
2
R( p , a) = P (p(u + a))f (u + a)cos pu du
0 πp(u + a) ν
(9.100)
∞
∫
2
− Q (p(u + a))f (u + a)sin pu du .
0 πp(u + a) ν
For the computation of the Fourier integrals in (9.100), tailored methods are available.4