See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/45876010

On the Computation and Applications of Bessel Functions with Pure Imaginary

Indices

Article · October 2009

Source: arXiv

CITATION READS
1 80

2 authors, including:

Edwin Fohtung
New Mexico State University & Los Alamos National Laboratory
29 PUBLICATIONS 117 CITATIONS

SEE PROFILE

All content following this page was uploaded by Edwin Fohtung on 26 May 2014.

The user has requested enhancement of the downloaded file.

 On the Computation and Applications of Bessel Functions with Pure Imaginary
Indices
A. A. Matyshev1 and E. Fohtung2, ∗
1
Department of Physical Electronics, St Petersburg State Technical University,
Politekhnitcheskaya-Str 29, 195251 St Petersburg, Russia
2
Institute for Synchrotron radiation, Forschungszentrum Karlsruhe, 76344 Eggenstein-Leopoldshafen, Germany
Bessel functions with pure imaginary index (order) play an important role in corpuscular optics
where they govern the dynamics of charged particles in isotrajectory quadrupoles. Recently they
were found to be of great importance in semiconductor material characterization as they are man-
ifested in the strain state of crystalline material. A new algorithm which can be used for the
computation of the normal and modified Bessel functions with pure imaginary index is proposed.
The developed algorithm is very fast to compute and for small arguments converges after a few
arXiv:0910.0365v1 [math-ph] 2 Oct 2009

iterations.
PACS numbers:

I. INTRODUCTION

Bessel functions occur in many branches of mathematical physics as solutions of differential equations when boundary
conditions such as the Dirichlet or Neumann are imposed on various space domains. The Bessel functions with index
(order) ν can be represented as the solutions of the following differential equation [1]:

x2 y ′′ + xy ′ + x2 − v 2 y = 0,


(1)

where y ′ and y ′′ are the first and second derivatives (respectively) with respect to x while ν is a complex constant.
The solutions of eqn.(1) can be expressed as an absolute converging series that is defined in the entire complex plane:

∞
 x v X  x 2n
n 1
Jv (z) = (−1) . (2)
2 n=0
n!Γ (v + n + 1) 2

In a great variety of applications such as diffraction, electrical induction, etc, only Bessel functions of the zeroth
and first orders, namely, J0 (x) and J1 (x) occur while in other physical applications (such as solutions of Kepler’s
equation), the entire indices n are used. Bessel functions with complex indices were considered to have limited areas
of applicability in natural and applied sciences till recently.
The normal and modified Bessel function with pure imaginary index was shown to be the solution governing the
motion of charged particles in isotrajectory quadrupoles [2]. Recently these functions have been found to be present as
solutions of the Lame’s equation which characterizes the displacement (and strain) field distribution in semiconductor
nanostructures [3].
In order to compute the solutions of equation (1) for the case of purely imaginary indices iν, it is required to obtain
the sum of a series with each coefficient being a complex number and to compute the complex value function Γ(iν)
which on its own is a very laborious operation. In reality, for the case of a purely imaginary index iν and for a natural
number n we have:

n
Y
Γ(iν + n + 1) = Γ(iν) (iν + m) (3)
m=0

and thereafter Bessel function (2) may be re-expressed in the form:

∞  x 2n
1  x iν X 1
Jiν (x) = (−1)n n , (4)
Γ(iν) 2 Q 2
n=0 n! (iν + m)
m=0

∗ [email protected]
 2

where ν and x are real and complex numbers respectively. The solution of (1) expressed in the terms of (4) represents
a ”difficult-to-compute” complex solution of a differential equation with the real coefficients. To the best of our
knowledge there is no concrete algorithm that can enable us to compute Bessel functions with purely imaginary order
as they have had no applications in areas of natural sciences till date. Only a hand full of pure mathematicians made
attempts to investigate such functions.
The author of one of the most detailed treatment on Bessel functions [1] thought that these functions were of no
interest though they were noticed by Lommel [4] who defined the function Jν+iµ (x) in the form

ν+iµ
(x/2)
Jν+iµ (x) = [Kν, µ (x) + i Sν, µ (x)] (5)
Γ(ν + iµ + 1/2)Γ(1/2)
with K and S being real valued functions. Lommel was motivated by the following differential equation:

x2 y ′′ + (2α − 2βν + 1)xy ′ + α(α − 2βν) + β 2 γ 2 x2β y = 0 .

 
(6)

If any real β 6= 0 then equation (6) has a solution

y = xβν−α AJν (γxβ ) + BJ−ν (γxβ ) .

 
(7)

Comparing equation (7) with the equation

x2 y ′′ + axy ′ + (b + cx2β )y = 0 , (8)

Lommel found a relationship between the coefficients of equation (7) and those of equation (9) and expressed them
in the form:
βν − α = −(a − 1)/2, (9)
√
q
2
βν = [(a − 1)/2] − b, βγ = c. (10)

It turned out that the real valued coefficients (a, b, c) produced in a general case solution of equation (6) with a
complex index ν. At the end Lommel obtained a very cumbersome solution of equation (6) as multiples of a complex
valued functions. He did not propose a method for the calculation of the real valued functions K and S from (5).
Another mathematician Bocher encountered the modified Bessel function with pure imaginary index while he ponder-
ered over the solutions of Laplace’s equation using the method of variable separation in cylindrical coordinate system
[5]. As opposed to Lommel [4], Bocher proposed real valued solutions with the aid of real valued series without
studying the convergence of these series. Each of the term of Bocher’s series was found to be a ratio of polynomial
functions in ν, with coefficients that increased n the order n!.This possed a major problem for the calculation and
computation via his representation.
Finally the McDonald’s function Kiτ (x) also having pure imaginary index was extensively used in its integral form
to obtain solutions of the Laplace and wave equations for different boundary value conditions by the Soviet mathe-
maticians M. Kantorovitch and N. Lebedev [6]. They used the integral representation

Z∞
Kiτ (x) = exp (−x cosh t) cos (τ t)dt, x > 0. (11)
0

This representation are convenient only for sufficiently large values of the argument x. Unfortunately, for small values
of x this integral representation is not quite effective, and this is the case where Bessel functions with pure imaginary
indices become applicable and useful in charge particle dynamics [2] and strain field investigations in nanostructures
[3].
The above summary shows that there is no effective algorithm for the computation of Bessel functions with pure
imaginary indices. In the next sections, we provide this algorithm.

II. METHOD AND CALCULATIONS

So, Bessel functions with pure imaginary indices are solutions of the equation

x2 y ′′ + xy ′ + (x2 + ν 2 )y = 0 , (12)
where ν is a real number. The pure mathematician George Boole was not interested in Bessel functions but was the
first person who, over a century and a half ago developed an operational method for solving differential equations.
 3

Nevertheless in a small paragraph of his article [7], equation (12) was mentioned and the substitution that facilitated
the solving of the problem of calculating Bessel functions with purely imaginary order was used. So expression (2)
suggests the form of the solutions of equation (12):

y = A (x) cos (ν ln x) + B (x) sin (ν ln x) , (13)

where A (x) and B (x) are series with the coefficients as shown:

A (x) = a0 + a1 x + a2 x2 + · · · + an xn + · · ·,
(14)
B (x) = b0 + b1 x + b2 x2 + · · · + bn xn + · · ·.

Substituting equation (13) into equation (12) and equating the coefficients before the linearly independent functions
cos (ν ln x) and sin (ν ln x) to zeros, firstly we obtain the relation

a1 = b1 = 0, (15)

and secondly ∀n ≥ 2 the recurrent relationships

nan−2 − 2νbn−2
an = − , (16)
n (n2 + 4ν 2 )
−2νan−2 + nbn−2
bn = − . (17)
n (n2 + 4ν 2 )

It is clearly seen from equations (16) and (17) that odd term coefficients take on zero values while those with even
terms can be obtained via a recurrent relations using the values a0 , b0 . The solution of the Bessel equation (12) in the
form (13) was firstly introduced by Boole in [7]. However, he did not analyze the convergence of the series (14). Boole’s
recurrent relationship was completely forgotten although they can be easily simplified and used to effectively calculate
the functions under examination. Since only coefficients with even indices occur, we can re-define the functions A (x)
and B (x) in the form:

∞
X  x 2n ∞
X  x 2n
A (x) = a2n 22n = A2n , (18)
n=0
2 n=0
2
∞
X  x 2n X∞  x 2n
B (x) = b2n 22n = B2n . (19)
n=0
2 n=0
2

Now for n ≥ 1 the recurrent relationship for the coefficients A2n , B2n can be simplified:

nA2n−2 − νB2n−2
A2n = − , (20)
n (n2 + ν 2 )
νA2n−2 + nB2n−2
B2n =− . (21)
n (n2 + ν 2 )

III. CONSTRUCTION OF COMPUTABLE SOLUTIONS AND PROOF OF CONVERGENCE

In order to investigate and numerically compute the strain field distribution in nanostructures, or to numerically
implement these Bessel functions in other areas of physics, we need to be able to demonstrate that the solutions to
this differential equation exist and above of all converges in our domain of definition and interest. We now show that
the series (16), (17) converges and does so absolutely for any complex valued argument x and order v. As a series
will converge absolutely if and only if the absolute value of the nth term (which we shall refer to as the majorant)
converges. Let the majorant M2n be defined such that

M2n = |A2n | + |B2n | . (22)

It follows from (20), (21), (22) that

 4

 
1 |v|
M2n ≤ + M2n−2
n2 + v 2 n (n2 + v 2 )
 
1 |v|
≤ 2
+ 3 M2n−2 . (23)
n n

The last inequality provides us the possibility of studying the order of which the majorant M2n decays (or grows). It
can be easily shown that

n|ν|
M2n ≤ C 2, (24)
(n!)

where C is a positive constant dependent on the zero-th terms of the sequence. Using the d’Alembert’s test for
convergence the proof of absolute convergence for all values of x and ν, is completed.
Now it is possible to form, without lost of generality, two linearly independent solutions of (12). Let us choose two
pairs of the values for A0 , B0 that generates the sequences A2n , B2n and its corresponding functions (18) and (19).
For brevity, these two pairs provide two solutions (12) that can be easily computed numerically for a few number of
iterations. In [2], it was shown that the solution spawned from the pair

(A0 , B0 ) = (0, 1) , (25)

represented as Sf ν (x) and the ones that are spawned from the pair

(A0 , B0 ) = (1, 0) , (26)

was also represented as Cf ν (x). Then an analytical expression of these functions can be easily obtained:

 
1  x 2
Sf ν (x) = 1 − + · · · sin (ν ln x)
1 + ν2 2
 
ν  x 2
+ + · · · cos (ν ln x) , (27)
1 + ν2 2

 
1  x 2
Cf ν (x) = 1 − + · · · cos (ν ln x)
1 + ν2 2
 
ν  x 2
+ + · · · sin (ν ln x) . (28)
1 + ν2 2

For x → 0 we have two equivalent relationships

Sf ν (x) ∼ sin (ν ln x) , (29)

Cf ν (x) ∼ cos (ν ln x) . (30)

Now the general solution of the differential equation (12) with real coefficients may be explicitly written out via the
real valued functions

y = c1 Sf ν (x) + c2 Cf ν (x) , (31)

where c1 and c2 are any real constants. Provided that equation (12) has a complex valued solution in the form of
the Bessel function (4). Therefore the latter must be the linear combination of the functions Sf ν (x) and Cf ν (x). It is
easy to show that

Jiν (x) = Cf ν (x) + iSf ν (x) . (32)

 5

In other words,the functions Sf ν (x) and Cf ν (x) are the real and imaginary parts of Bessel function (4). It follows
from the following equivalent relationships:

1  x iν  1  x 2

Jiν (x) = 1− + ··· ∼
Γ(iν + 1) 2 iν + 1 2

1
∼ [cos(ν ln x) + i sin(ν ln x)] ∼ [cos(ν ln x) + i sin(ν ln x)] .
iνΓ(iν)2iν

The first equivalent relationship is valid as long as x → 0 and the second[9] one is valid so long as ν → 0.

x2 y ′′ + xy ′ + −x2 + ν 2 y = 0.


(33)

As

 πν   πν 
cos (ν ln ix) = cosh cos (ν ln x) − sinh sin (ν ln x) , (34)
2 2

 πν   πν 
sin (ν ln ix) = cosh sin (ν ln x) + sinh cos (ν ln x) , (35)
2 2
and as

∞
X  x 2n
C (x) = A (ix) = C2n , (36)
n=0
2
X∞  x 2n
D (x) = B (ix) = D2n , (37)
n=0
2

it is possible to construct two linearly independent solutions for this case in the form

y1 = C (x) cos (ν ln x) + D (x) sin (ν ln x) , (38)

y2 = D (x) cos (ν ln x) − C (x) sin (ν ln x) , (39)

where as earlier mentioned, for n ≥ 1 the recurrent equations for C2n , D2n takes the expected form
nC2n−2 − νD2n−2
C2n = , (40)
n (n2 + ν 2 )
νC2n−2 + nD2n−2
D2n = . (41)
n (n2 + ν 2 )

As the series C(x) and D(x) have the same majorant (24), these series will thus converge absolutely for any x and ν.
One real valued solution of (12) spawns two real valued solutions for (33). Thus by choosing only one pair of values
C0 , D0 it is possible to obtain two real valued functions being solutions of equation (33). Let

(C0 , D0 ) = (0, 1) . (42)

Substituting (36) and (37) into (38) and (39) it is possible to obtain two functions which was denoted in [2] as Sdν (x)
and Cdν (x):

 
1  x 2
Sdν (x) = 1 + + · · · sin (ν ln x)
1 + ν2 2
 
ν  x 2
+ − + · · · cos (ν ln x) , (43)
1 + ν2 2
 6

 
1  x 2
Cdν (x) = 1 + + · · · cos (ν ln x)
1 + ν2 2
 
ν  x 2
− − + · · · sin (ν ln x) . (44)
1 + ν2 2
As x → 0 we also have two equivalent relationships

Cdν (x) ∼ cos (v ln x) , (45)

Sdν (x) ∼ sin (v ln x) . (46)
It is obvious that

Jiν (ix) = Cdν (x) + iSdν (x) , (47)

where x and ν are any real numbers.

IV. WRONSKIANS OF FUNCTIONS

It can be easily shown that the Wronskian W of the two linearly independent solutions of the equation

1 ′
y ′′ + y + f (x) y = 0 (48)
x
has the form W = a/x, where the constant a depends on the pair of concrete solutions. From the equivalent
relationships (29), (30) and (45), (46) it follows that

′ ν ′
,
Cf ν (x) [Sf ν (x)] − Sf ν (x) [Cf ν (x)] = (49)
x
′ ′ ν
Cdν (x) [Sdν (x)] − Sdν (x) [Cdν (x)] = . (50)
x
It is clearly seen that the Wronskians (49) and (50) vanishes for ν = 0. As ν → 0 we have

Cf ν (x) → J0 (x) , (51)

Sf ν (x) → 0 (52)
and
Cdν (x) → I0 (x) , (53)
Sdν (x) → 0 . (54)

V. ACCURACY OF PROPOSED COMPUTATIONAL ALGORITHM

To obtain a measure of the computational error for the proposed procedure, we require a detailed study of the majorant
(remainder) M2n . Let

n|ν|
M2n = 2 m2n . (55)
(n!)
Now it is possible to obtain an estimate of remainder m2n for n ≥ 2

  |ν|
m2n |ν| 1
≤ 1+ 1− . (56)
m2n−2 n n
 7

The Lagrange’s formula for remainder of a Taylor’s series gives for n ≥ 2

|ν|−2
ν2

m2n |ν| (|ν| − 1) θ
≤1− 2 + 1− ×
m2n−2 n 6n3 n
"   |ν|−3 #
θ θ
× 3 |ν| + (|ν| − 2) 1 + |ν| 1− , (57)
n n

where 0 ≤ θ ≤ 1. The last inequality may be improved by removing from the right hand side the quantity θ:

m2n ν2 F
≤1− 2 + 3 , (58)
m2n−2 n n

where


 ν 2 ||ν| − 1| 2(1−|ν|) |ν| ≤ 2 ,
3−|ν|

F = |ν| ||ν| − 1| 3|ν| + (|ν| − 2)(1 + |ν|/2)2 /6 , 2 < |ν| ≤ 3 , (59)
2


|ν| ||ν| − 1| ν /2 + 3|ν| − 2 /6 , |ν| > 3 .


As for ν 6= 0

m2n ν2 F ν2 F
≤ 1 − 2 + 3 < 1 + 2 + 3, (60)
m2n−2 n n n n

then it is possible to obtain

ν2 F ν2 F
ln (m2n ) − ln (m2n−2 ) < ln(1 + + ) < + 3.
n2 n3 n2 n
Using the well known sums

∞
X 1 π2
2
= − 1 = 0.6449 ,
n=2
n 6
∞
X 1
3
= ζ(3) − 1 = 0.2021
n=2
n

and also the values

m0 = 1 ,
1+|ν|
m2 = 1+ν 2,

it is possible finally to obtain the inequality

m2n
ln < 0.6449ν 2 + 0.2021F . (61)
m2n−2

From the last inequality it is easy to obtain a final one simplified in the form

1 + |ν|
m2n < exp(0.6449v 2 + 0.2021F ) = m(ν) . (62)
1 + ν2
Now we have
 8

n|ν|
M2n < m(ν) (63)
(n!)2

and it is possible to conclude that the sum of the remaining terms for the series of functions A (x) , B (x) , C (x) , D (x)
cannot exceed the value

∞
X n|ν|  x 2n
ε = m(ν) 2 . (64)
(n!) 2
n=N

Since the remainder (64) depends on the value of ν, let us consider for example the case |ν| ≤ 2. Then

∞
x2 X 1  x 2n−2
εN ≤ m(ν) =
4
n=N +1
[(n − 1)!]2 2
∞
x2 X 1  x 2n
m(ν) 2 2 . (65)
4
n=N (n!)

The sum at the right hand side of equation (65) is simply the remainder of the series for the modified Bessel function
I0 (x). This remainder may be evaluated further as shows below:

∞ ∞
X 1  x 2n 1  x 2N X (N !)2  x 2n
2 = 2 2 2 =
n=N (n!) 2 (N !) 2 n=0 [(n + N )!]
(
1  x 2N 1  x 2
= 2 2 1+ 2 +
(N !) (1 + N ) 2
)
1  x 2n
···+ 2 + ··· <
[(1 + N ) (2 + N ) · · · (n + N )] 2
 
1  x 2N 1  x 2 1  x 2n
< 2 1+ + ··· + + ··· =
(N !) 2 1·2 2 n! (n + 1)! 2
1  x 2N 2 1  x 2N −1
= 2 I1 (x) = 2 I1 (x) . (66)
(N !) 2 x (N !) 2

Now for any real x and |ν| < 2 we have

1 + |ν| 2

1  x 2N +1
εN ≤ exp 0.6449ν + 0.2021F 2 I1 (x) . (67)
1 + ν2 (N !) 2

Elimination of |ν| in the last inequality gives an estimation for any x and |ν| < 2:

1  x 2N +1
εN ≤ 15 I1 (x) .
(N !)2 2

For instance for x ≤ 2 inequality (68) reduces to

24
εN ≤ 2. (68)
(N !)

The last inequality shows that for the functions A (x) , B (x) , C (x) , D (x) and for x ≤ 2 and |ν| ≤ 2 a computational
error less then ε8 < 1.5 × 10−16 occurs after the computation of only eight terms of the corresponding series.
 9

VI. CONCLUSION

A new algorithm for computing the real valued solutions of some special functions namely the Gamma function with
pure imaginary argument, the Bessel (and modified Bessel) function of purely imaginary orders was systematically
declared. We have used these developed Bessel functions for the investigation and calculation of strain fields in
semiconductor structures [3, 8]. In the above applied case the values of the arguments of the Bessel functions lie
in the vicinity of unity. The algorithm for their computations is shown to be rapidly converging and gives a very
favorable accuracy for as few as eight iterations.
The algorithm described may give an alternate approach to solving the ground state of the screened coulombs
potential problem. It is also promising in solving the inverse scattering problem which is mostly approached via the
inverse Green’s function. Further studies of the asymptotic behavior of these functions are still warranted as they
play a great role in a wide class of boundary value problems and integral and differential problems of wave scattering.

[1] G. Watson, A treatise on the theory of Bessel functions (Cambridge Mathematical Library, 1945).
[2] A. Matyshev, Isotrajectory Corpuscular Optics (SPb, Science (in Russian), 2000).
[3] E. Fohtung et al, in preparation (2009).
[4] E. Lommel, Math. Ann. 3, 475 (1871).
[5] M. Bocher, Annals of Math. 6, 137 (1892).
[6] G. G. Abramovitch, Selected Problems of Mathematical Theory of Electromagnetic Phenomena (Acad. of Sciences Pub.,
Moscow-Leningrad (in Russian), 1948).
[7] G. Boole, Phil. Trans. of the Roy. Soc. of London 134, P2, 225 (1844).
[8] M. Riotte, E. Fohtung, A.A. Minkevich, et al, in preparation (2009).
[9] To prove the second equivalent relationship it is necessary to use the identity
π
Γ(iν)Γ(1 − iν) ≡ .
sin iν

View publication stats