Appendix Dirac Delta Function
Appendix Dirac Delta Function
T
v = [ v1 , v2 , . . . , vN ] = î vi ,
where the set { î } forms the basis vectors. The scalar product of these vectors is
given by,
u · v = î · ĵ ui vj .
In the important special case where the basis vectors are orthogonal unit vectors,
or, orthonormal vectors, we have,
1, if i = j ;
î · ĵ =
0, j.
if i =
u · v = ui vj δij = ui [ vj δij ] = ui vi .
The role of δij , in other words, is to select one of the terms of the summation, i.e.,
vi = vj δij (B.2)
The Kronecker delta is useful in the general problem of the expansion of any function
or quantity in terms of its normal modes, eigenfunctions, basis vectors, etc.
To introduce the Dirac delta function, it is convenient to first define the unit
step function, U(x), or Heaviside function,∗∗ where (cf. Fig. B.1),
0, for x < 0 ;
U(x) ≡ 1/2 , for x = 0 ; (B.4)
1, for x > 0 .
7(N)
1.0
One of the utilities of the unit step function U(x), is the ability to express any
integral over finite limits, e.g., a ≤ x ≤ b, as an integral over the whole x-axis, i.e.,
b ∞
f (x) dx = [ U(x − a) − U(x − b) ] f (x) dx . (B.5)
a −∞
Consider now the integral over the interval −h/2 ≤ x ≤ h/2 of a function f (x), i.e.,
h/2
f (x) dx .
−h/2
This can be expressed in terms of the window function (cf. Fig. B.2),
h h
W(x; h) ≡ U(x + ) − U(x − ) , (B.6)
2 2
i.e.,
h/2 ∞
f (x) dx = W(x; h) f (x) dx .
−h/2 −∞
9(N; D)
1.0
N - D/2 N + D/2 N
for some ξ,
h h
− ≤ ξ ≤ ,
2 2
or, ∞
1
W(x; h) f (x) dx = f (ξ) .
−∞ h
We can now define the Dirac delta function, δD (x), as the limit,
1 1 h h
δD (x) ≡ lim W(x; h) = lim U(x + ) − U(x − ) . (B.8)
h→0 h h→0 h 2 2
with,
0, for x = 0 ;
δD (x) = (B.10)
∞, for x = 0
(see also Prob. B.1), but such that (cf. Eq. B.9),
∞
δD (x) dx = 1 . (B.11)
−∞
It is the combination of Eqs. B.10 and B.11 that make the Dirac delta function a
“generalized” function. The integral of an ordinary function that is everywhere zero
except at a point, or at a number of points (of measure zero), will be zero.
Proof : From the definition of the Fourier transform and its inverse, i.e.,
∞
FT { f (x) } ≡ f (x) e−ikx dx = F (k) (B.14a)
−∞
and
∞
−1 1
FT { F (k) } ≡ F (k) eikx dk = f (x) , (B.14b)
2π −∞
we have,
∞
1
FT { δD (x) } = lim W(x; h) e−ikx dx
−∞ h→0 h
∞
1
= lim W(x; h) e−ikx dx
h→0 −∞ h
h/2
1
= lim e−ikx dx
h→0 h −h/2
2 kh
= lim sin .
h→0 kh 2
The proof is completed by taking the inverse Fourier transform, i.e.,
∞ ∞
1 2 kh ikx 1
δD (x) = lim sin e dk = eikx dk .
2π −∞ h→0 kh 2 2π −∞
where the xi are the roots of g(x), i.e., g(xi ) = 0, in the interval (a, b). Note
the absolute value in this expression (omitted in some references)! It is this latter
property that is the claim to fame of the Dirac delta function. Some important
special cases are: ∞
f (x) δD (x) dx = f (0) , (B.16a)
−∞
∞
f (x) δD (x − x0 ) dx = f (x0 ) , (B.16b)
−∞
∞
f (0)
f (x) δD (αx) dx = . (B.16c)
−∞ |α|
The similarity between the Dirac delta function δD (x − x0 ) and the Kronecker
delta function δij should not escape unnoticed. We see that it is the property of
δD (x − x0 ) to select the value of f (x) at x = x0 , when integrating over x, i.e.,
∞
f (x0 ) = f (x) δD (x − x0 ) dx .
−∞
fi = fj δij .
The two are equivalent. Indeed, it would have been possible to define δD (x − x0 )
in terms of δij by discretizing the integral of Eq. B.16b and taking the limit of
infinitesimally spaced points in the summation.
with analogous provisos if the point at r = a resides on the bounding surface of the
volume V . We can see that, at least in Cartesian coordinates, the function that
does the job is given by,
δD (r − a) = δD (x − ax ) δD (y − ay ) δD (z − az ) , (B.20a)
where,
T
a = [ ax , ay , az ] . (B.20b)
7N
N
A
@,N
@,\N
@,\\N
One can get a feeling for what δD (x) and δD (x) look like by trying to draw
them. From the defining Eq. B.8 for δD (x) and the definition of the derivative of a
function we also have,
dU(x)
δD (x) = U (x) = . (B.23)
dx
9Nov02 13:34 Dimotakis em.Br
The Kronecker & Dirac delta functions: The Dirac delta function B.9
Imagine now a U(x) with slightly rounded corners so that the transition from 0 to
1 is achieved over an x-interval across the origin. Differentiating this graphically,
we have a picture for δD (x), i.e., a spike at x = 0. This can be differentiated, in
turn, to get δD (x), and again to get δD (x). See Fig. B.3 and Prob. B.4.
Derivatives of the Dirac delta function can be used to expand a function that
is localized around some point (say, the origin) into a moment expansion, i.e.,
1
f (x) = f0 δD (x) − f1 δD (x) + f2 δD (x) − · · · ,
2
or,
fn (n)
f (x) = (−1)n δ (x) , (B.24a)
n
n! D
where, the fn are constants given by,
∞
f0 = f (x) dx
−∞
i.e., the nth moment of f (x) about x = 0. The expansion suggested by Eq. B.24
is in general possible if the moments exist as, for example, would be the case for
a function f (x) that is identically equal to zero for |x| > a, or for a function that
decays to zero away from x = 0 at least as fast as an exponential. Both the left-
and right-hand side of Eq. B.24a have the same moments, as can easily be verified.
The validity of this expansion rests on the usually, but not always, valid contention
that two (localized) functions, with all their moments equal, are equal.†
† This depends on the degree of localization of the function. See, for example, Orzag (1970) and
Monin & Yaglom (1975, footnote, p. 640), who discuss the Orzag reference.
To demonstrate the utility of this formalism, consider the problem of the com-
putation of the integral,
∞
I ≡ f (x) G(x) dx , (B.25)
−∞
where,
0, for |x| ≥ a ;
f (x) =
< K , for |x| < a ,
with a function G(x) that is slowly varying across the interval −a < x < a, and
possesses a maximum at, say, x = x0 > 0, where a/x0 1, as is indicated in
Fig. B.4 below.
∞
fn
(n)
I = (−1)n δ (x) G(x) dx
−∞ n
n! D
∞
n fn (n)
= (−1) δD (x) G(x) dx ,
n
n! −∞
and, therefore,
fn
I = G(n) (0) , (B.26a)
n
n!
where,
(n) dn G(x)
G (0) ≡ , (B.26b)
dxn x=0
G(x)
f(x)
x0
-a a x
If we can assume that, for large distances from its maximum at x = x0 , G(x)
goes to zero like,
G0
G(x) ∼ ,
|x − x0 |α
where α ≥ 1, then, for large |x − x0 |, we have,
(n) n!
G (x) = O
|x − x0 |n+α
and, therefore,
(n) n!
G (0) = O .
x0n+α
Using the assumed properties for f (x), we can also make an estimate for the mo-
ments fn , i.e.,
∞
fn = xn f (x) dx
−∞
a
= xn f (x) dx
−a
1
n+1 2 K an+1
= a ξ n f (aξ) dξ <
−1 n+1
n+1
a
= O .
n+1
This means that the nth term in the series goes like,
n+α
1 a
∼ ,
n+1 x0
and the series converges since it is bounded term by term by the series expansion
of,
α−1
a a
ln 1 − ,
x0 x0
provided a/x0 < 1, which we assumed for the purposes of this discussion.
Retaining the first few terms of our series expansion (Eq. B.26a) for our integral
and letting α = 1 (the worst case), we have,
1 a 4
I = f0 G(0) + f1 G (0) + f2 G (0) + O .
2 x0
For x0 a few times a, we have a very good approximation to our integral (e.g., if
x0 = 10 a, the error after three terms is of the order of 10−4 ). In addition, it is easy
to see that the exact details of f (x) do not affect the solution and are not needed.
Note also that if f (x) is an even function of x, then all the odd moments vanish
identically, and we have,
1 a 5
I = f0 G(0) + f2 G (0) + O ,
2 x0
The success of this scheme is the basis for the utility of multipole expansions,
commonly used in electromagnetic theory, fluid mechanics, and other disciplines.
B.R References
Courant, R. & Hilbert, D. 1962 Methods of Mathematical Physics II (John
Wiley & Sons), Appendix to Ch. VI.
Lighthill, M. J. 1964 Fourier Analysis and Generalized Functions (Cambridge
University Press, London).
Monin, A. S. & Yaglom, A. M. 1975 Statistical Fluid Mechanics: Mechanics of
Turbulence II (Ed. J. Lumley, MIT Press, Cambridge, MA).
Orzag, S. A. 1970 Indeterminacy of the moment problem for intermittent turbu-
lence. Phys. Fluids 13(9), 2211–2212.
B.P Problems
Make sure that your arguments for Part (a) are consistent with Part (b).
B.2 The unit step function and the sign(x) function.
B.3 Note that while the Dirac delta function is unique, in addition to the scaled
window function (Eq. B.6), it can be expressed as the limit of several other
functions.
a. Triangular function. If > 0, and
1 |x|
1− , for |x| < ;
Λ(x; ) ≡ (B.28a)
0, for |x| ≥ ,
show that
lim { Λ(x; ) } = δD (x) . (B.28b)
→0
This function is useful in deriving Poisson process (shot noise) statistical
properties.
b. Lorentzian: Show that
lim { L(x; ) } = sign() δD (x) (B.29a)
→0
where
1 2 2
G(x; σ) ≡ √ e− x /2σ . (B.30b)
2π σ
Hint: Consider the Fourier transform of the window function W(x; h),
Eq. B.6, p. B.3.
B.4 Derivatives of delta functions. Use the definition of δD (x) as the limit of a
normalized gaussian (Eq. B.30),
a. to express
(n) dn δD (x)
δD (x) ≡ ,
dxn
b. Use the definition of the Dirac delta function to prove Eq. B.15. Note
that it is the absolute value of g (xi ) that enters.
B.6 Spherical coordinate representation of the delta function. Express the Dirac
delta function in a way that is appropriate for integration in three-dimensional
spherical coordinates. Hint: For spherical coordinates, require that,
∞
δD (r) dr = 1
0
B.7 Show that the moment expansion of Eq. B.24 in x-space is equivalent to a
Taylor expansion of F (k), the Fourier transform of f (x), about k = 0.
B.8 Using the properties of the delta function, prove the Fourier transform pair
Eqs. B.14. In other words, show that if F (k) = FT { f (x) }, then f (x) =
FT −1 { F (k) }.
B.9 A problem by U. Frisch. Solve for f (x) in the equation, x f (x) = 1.