Survival Guide to Bessel Functions

December 4, 2013

(Original by Mike Herman; edits and additions by Paul A.)

1 The Problem
For cylindrical boundary conditions, Laplace’s equation is:

1 ∂2Φ ∂2Φ
 
1 ∂ ∂Φ
s + 2 + =0 (1)
s ∂s ∂s s ∂φ2 ∂z 2
The way s mixes throughout the equation means that the z and φ behavior will have a combined nontrivial
influence on the very equation for the function of s in the mode expansion found by separating variables.
The resulting functions of s are called Bessel functions, and the usual notation to denote them shows the z
and φ behavior in different places: the φ behavior gives the order of the Bessel function and appears in the
subscript, while the z behavior appears in the argument of the Bessel function, multiplied by s.
Choices: It will turn out that the type of solution encountered depends crucially upon the domain being
considered, including the type of boundary conditions assumed. A “flow chart” method is sketched in the
following paragraphs in order to help decide which functions are appropriate.
Behavior with z: The first choice concerns the z behavior. In the separation of variables technique, we
need the eigenvalues of the sub-operators corresponding to the different coordinates to all sum to zero, in each
of the basic solutions used to construct the general solution. This leaves a choice of sign for the eigenvalue of
(∂/∂z)2 , and the sign chosen (to satisfy the necessary boundary conditions) determines whether the ordinary
or modifield Bessels appear for the radial part. For a positive eigenvalue (exponential behavior in z), the
ordinary Bessels (Jν and Yν ) appear. For a negative eigenvalue (oscillating behavior in z), we find the
modified Bessels (Iν and Kν ).
Behavior with φ: The next domain to consider is that of φ. The sign ambiguity usually doesn’t arise
in the eigenvalue of (∂/∂φ)2 : we usually choose its eigenvalue to be negative (and equal to −ν 2 ) because
φ is periodic. (It is conceivable that the positive eigenvalues will appear for special situations, such as zero
potential on the surface φ = 0 and a fixed nonzero potential at another surface φ = φ0 , giving sinh(νφ)
behavior and thus a positive eigenvalue. These cases lead to Bessels of purely imaginary order, and are
relatively rare.) When the entire range of φ is in the domain where Laplace’s equation holds, this gives
ν = m = an integer. In other cases, ν will be found by fitting the boundary conditions. For instance, if the
domain is limited to a “slice of pie” of angular width β, ν will not be restricted to integers, since we don’t
require periodicity of length 2π in that case. Instead, ν = mπ β , with m = 1, 2, 3, . . ., for Dirichlet boundary
conditions on the boundaries. For Neumann boundary conditions, the same values of ν apply, but m = 0 is
now allowed (since we have cosines rather than sines). For mixed boundary condtions (Dirichlet on one side,
and Neumann on the other), we must fit an odd number of quarter-wavelengths into φ, so ν = (2m+1)π 2β in
that case, with m = 0, 1, 2, . . ..
Resulting equations for s: With the above behavior in φ assumed, we can replace φ derivatives with
im and z derivatives with +k 2 (case A) or −k 2 (case B) in the cylindrical Laplacian (eq. 1 above) to arrive at
an equation for S(s). These choices correspond to passing to a Fourier series in φ and either doing a Laplace
transform in z (case A) or Fourier transform in z (case B). Case A gives the Bessel equation, which is
eq. (6) below, and that for case B gives the modified Bessel equation, which replaces k 2 by −k 2 in the

1
equation. Recall that we may write nonlinear functions only of dimensionless variables, so both of these can
be put into a standard form by switching to the dimensionless variable x = ks:
   2 
1 d dS ν
x − ∓1 S =0 (2)
x dx dx x2

where the − sign gives the ordinary Bessel equation, and the + sign gives the modified Bessel equation.
Some of the properties given below are in terms of s and some in terms of the dimensionless x instead, for
convenience. (Be careful especially when doing derivatives to watch factors of k when translating from s to
x or back).

2 Solutions
Assuming a solution of (sums of terms of) the form

Φ(s, φ, z) = S(s)Ψ(φ)Z(z), (3)

we find for the two general cases:

2.1 Case A: Non-Cyclic/Exponential in Z (k ∈ R, k > 0)

d2 Z
− k2 Z = 0 (4)
dz 2

d2 Ψ
+ ν2Ψ = 0 (5)
dφ2
giving

d2 S ν2
 
1 dS 2
+ + k − 2 S=0 (6)
ds2 s ds s
Note that k is determined by either z or s boundary conditions; e. g., if cylinder is grounded at top
(z = L) and bottom (z = 0), then the z dependence determines k: Z(z) ∝ sin(kn z), where kn = nπ L . If
the cylinder is grounded (Dirichlet boundary conditions) at its radial edge (s = a) instead, then k will be
restricted to kn = xνn /a, n = 1, 2, 3, . . ., where xνn is the nth zero of Jν : Jν (xνn ) = 0. (These are instead
called zνn in Cahill.) For Neumann boundary conditions, the zeroes of the derivatives of Jν appear (called
x′νn and/or yνn in Jackson): [(d/dx)Jν (x)]x=x′νn = 0.
Bessel functions of the first kind of order ±ν are the solution for “Case A”, where:
(
ekz , e−kz
Z(z) ∝
sinh(kz), cosh(kz)
This solution is used when you have non-cyclic behavior in z, as in the case where Φ(z = 0) = 0 and
Φ(z = L) = V (s, φ).
In this case, the solutions to the radial equation are:

S(s) = J±ν (ks)

However, the two solutions (for the two signs of ν) are linearly dependent when ν = m ∈ Z, so we need
another solution:

Jν (ks) cos(νπ) − J−ν (ks)

S(s) = Yν (ks) =
sin(νπ)
(with a limit ν → integer m understood for integer ν.)
The above solutions Jν and Yν are always linearly independent (even when ν ∈
/ Z). (Note that Yν is often
instead denoted Nν in some texts, such as Jackson.) The domain in question determines which of these is

2
used: if s = 0 is included in the domain, only the Jν appear. (This follows from the s → 0 behavior listed
below: the Yν (s) blow up as s → 0.) If s = 0 is not included, we must keep a linear combination of both the
Jν and the Yν in a general solution.
Limiting forms of these functions are given for x = ks → 0 as:
1  x ν
Jν (x) →
Γ(ν + 1) 2
( 
2
ln x2 + 0.5772 , ν = 0


Yν (x) → π
2 ν
− Γ(ν)


π x , ν 6= 0
where Γ(ν + 1) is the Gamma function: when ν = m = an integer, Γ(m + 1) = m!, the factorial function.
Plugging in, we see that J0 (0) = 1 and Jν6=0 (0) = 0, while Yν (x → 0) → ∞.
For the behavior at large argument, we have
r
2  νπ π 
Jm (x → ∞) −→ cos x − −
πx 2 4
r
2  νπ π 
Ym (x → ∞) −→ sin x − −
πx 2 4
Because of this asymptotic behavior, it is sometimes convenient to introduce instead as the two indepen-
dent solutions the Hankel functions H 1 and H 2 , defined by

Hν(1) (ks) = Jν (ks) + iYν (ks)

Hν(2) (ks) = Jν (ks) − iYν (ks)

√
which will thus have e±ix / x behavior as x → ∞.

2.1.1 Orthogonality and Completeness Relations

On the infinite domain 0 ≤ s < ∞, the orthogonality relation is
Z ∞
δ(k − k ′ )
s Jν (ks) Jν (k ′ s) ds =
0 k
and the completeness relation is
∞
δ(s − s′ )
Z
k Jν (ks) Jν (ks′ ) dk = .
0 s
Note the appearance of s ds as the integration measure; this comes naturally from the volume element
in cylindrical coordinates. (You’ve likely noticed that these are basically the same integral with variables
switched; recall that orthogonality and completeness also gave the same integral as one another with variables
switched for the modes in Fourier transforms, where the domain was also infinite.)
When the domain is instead the interior of a finite cylinder of radius a with Dirichlet conditions at s = a,
orthogonality becomes instead
Z a
a2 2
s Jν (zνn′ s/a) Jν (zνn s/a) ds = J (zνn ) δnn′ .
0 2 ν+1
where zνn is the nth zero of Jν , as mentioned above. Note that this making the domain finite has given
us discrete modes (“quantized” numbers) instead, just as it did when changing from Fourier transforms for
infinite domain to Fourier series on a finite domain.

2.1.2 Generating Function

hx i ∞
X
exp (t − 1/t) = tm Jm (x)
2 m=−∞

3
2.1.3 Recursion Relations
All solutions of the ordinary Bessel equation (S can be J, Y , H (1) , or H (2) ) obey
x
[Sν−1 (x) + Sν+1 (x)] = νSν (x)
2
d
Sν−1 (x) − Sν+1 (x) = 2 Sν (x)
dx

2.2 Case B: Solutions Cyclic in Z

Modified Bessel functions are the solutions for “Case B,” where:
(
eikz , e−ikz
Z(z) ∝
sin(kz), cos(kz)
This solution is used when you have cyclic behavior in z, as in the case where Φ(z = 0) = 0 and
Φ(z = L) = 0. These are equivalent to making k imaginary above, but it is more convenient to introduce
explicit real-valued solutions. Real-valued linearly independent solutions are:

Iν (ks) = i−ν Jν (iks) (7)

π ν+1 (1)
Kν (ks) = i Hν (iks) (8)
2

2.2.1 Properties
As x → 0, we have
1  x ν
Iν (x → 0) →
Γ(ν + 1) 2
( 
− ln x + 0.5772 , ν = 0


Kν (x → 0) → Γ(ν) 22
ν
2 x , ν 6= 0
and so the Iν behave the same as Jν for small x: I0 (0) = 1 and Iν6=0 (0) = 0, while the Kν blow up as x goes
to zero. The similarity with ordinary Bessels is not shared at large x, where:
1
Iν (x → ∞) −→ √ ex
2πx
r
π −x
Kν (x → ∞) −→ e
2x

2.2.2 Generating Function

hx i ∞
X
exp (t + 1/t) = tm Im (x)
2 m=−∞

2.2.3 Recursion Relations

Solutions of the modified Bessel equation obey slightly different relations:
x
[Iν−1 (x) − Iν+1 (x)] = νIν (x)
2
d
Iν−1 (x) + Iν+1 (x) = 2 Iν (x)
dx

4
for the Iν , but an extra minus sign for the Kν :
x
[Kν−1 (x) − Kν+1 (x)] = −νKν (x)
2
d
Kν−1 (x) + Kν+1 (x) = −2 Kν (x)
dx

3 A few other useful integrals

(See also orthogonality and completeness integrals given above.)
Z ∞
Jν (ks) ds = 1/k
0

1 2 ν2
 
1
Z
2
Jν2 (ks) s ds = s − 2 Jν2 (ks) + s2 J ′ ν (ks)
2 k 2
When J vanishes on the boundaries, this reduces to the J ′2 term only, and we can use the recurrence formulas
2
to replace this by Jν+1 , as is done in the orthogonality integrals.
For integer m, we have the integral representation

1 π
Z
Jm (x) = cos [mφ − x sin φ] dφ (9)
π 0

which generalizes to arbitrary order ν by

1 π sin(νπ) ∞
Z Z
Jν (x) = cos [νφ − x sin φ] dφ − exp[−νt − x sinh t] dt . (10)
π 0 π 0

Another integral which appears surprisingly often in physics is

Z ∞
Kν (x) = e−x cosh t cosh(νt) dt (11)
0