A. Fractional Calculus
A. Fractional Calculus
A. Fractional Calculus
(1 + x)2 = 1 + 2x +
1
2! 2(2
1
2! 3(3
1)x2
1
(1 + x)3 = 1 + 3x +
1)x2 + 3!
3(3 1)(3 2)x3
..
.
p
k
1
(1 + x) = 1 +
k! p(p 1)(p 2) (p [k 1])x
(1)
k=1
z = 0, 1, 2, . . .
d 2f (x)
1
dx 2
or, more generally, to dpf (x)/dxp where p is any number (positive or negative,
real or complex), then you labor deprived of a powerful resource.
Early in the present century physicists learned, under the leadership of
people like P. W. Debye and A. Sommerfeld, to escape the tyranny of the real
line, to do their physics on the complex plane, and only at the end of their
calculations to let z become real. More recently they have learned to escape the
tyranny of dimensional integrality; though fractal dimension comes instantly
to mind, I am thinking here more particularly of the many papers (especially
papers treating statistical mechanical and eld-theoretic topics) that adhere to
the pattern
dicult physics in 3 dimensions
= lim relatively easy physics in 3 dimensions
0
dx
dx
We recall in this connection that
d
integer
d
integer
dx
dx
was achieved already by the Fundamental Theorem of the Calculus, and that
(1 + x)integer (1 + x)real or complex
entailed invention of the concept of innite series; in the latter context as in
the context that will concern us, a kind of interpolation is going on, but it
is interpolation in the exponent. In both contexts, innity intrudes. Given a
Introduction
f
f
f f f f f
and seek to interpolate between the points thus marked out in function space.
An element of ambiguity attaches, of course, to all interpolation/extrapolation
schemes; one looks for the scheme that is most natural in the sense most
empowering. It is in that always-somewhat-vague sense that (for example)
one defends the claim that (n + 1) provides the most natural interpolation
amongst the discrete numbers n!.1 In just that same sense, one develops the
strong convictionbut cannot explicitly provethat the fractional calculus
does in fact proceed optimally to its interpolative goal.
The vision of a fractional calculus was evident already to the founding
fathers of the ordinary calculus (whichbut for the double meaning of a word
which becomes intolerable in this contextwe are tempted to call the integral
calculus). Leibnizwho
was, after all, the inventor of both the dn f /dxn
notation and of f (x)dx wrote in September of to his friend lHospital
as follows:2
Jean Bernoulli seems to have told you of my having mentioned to
him a marvelous analogy which makes it possible to say in
a way that successive dierentials are in geometric progression.
One can ask what would be a dierential having as its exponent
a fraction. You see that the result can be expressed by an innite
series, although this seems removed from Geometry, which does not
yet know of such fractional exponents. It appears that one day these
paradoxes will yield useful consequences, since there is hardly a
paradox without utility. Thoughts that mattered little in themselves
may give occasion to more beautiful ones.
Thirty-ve years later, Euler expressed a similar thought, and took explicit note
of the fact that a kind of interpolation theory comes necessarily into play:
Concerning transcendental progressions whose terms cannot be
given algebraically: when n is a positive integer, the ratio dn f /dxn
can always be expressed algebraically. Now it is asked: what kind of
1
(p) =
xp1 ex dx
xm1 (1 x)n1 dx =
B(m, n) =
0
(m) (n)
(m + n)
which play such a central role in this storydid not appear until publication
of Institutiones calculi integralis (3 volumes, ). Notable contributions
to the eld were made successively by Laplace, Fourier, Abel, Liouville,
Riemann, Heaviside andin the present centuryby Bateman, Hardy, Weyl,
Riesz and Courant, as well as by many pure and applied mathematicians of
lesser reknown. Our subject can claim, therefore, to be well upwards of years
old, and its foundations have been securely in place for more than a century.
Yet the rst book-length account of the eld did not appear until , when
Keith B. Oldham & Jerome Spanier published The Fractional Calculus: Theory
& Applications of Dierentiation & Integration to Arbitrary Order.3 Though it
retains its place as the primary reference in the eld, this monograph was joined
recently by K. S. Miller & B. Ross An Introduction to the Fractional Calculus
and Fractional Dierential Equations (). Both books are very accessibly
written, and so far as concerns matters of design and emphasis they are neatly
complementary; when you undertake study of this eld you will want to have
both within easy reach. Both books provide brief but usefully detailed accounts
of the history of the fractional calculus (as well as valuable bibliographic data),
and it is from them that, in the absence of explicit indication to the contrary,
my own historical remarks have been taken.
1. Elementary preliminaries. Let us agree (pending future renements) to write
d
dx
f ()d = f (x)
(1)
a
3
Elementary preliminaries
(2)
These familiar statements suggest a range of meanings that might with equal
plausibility be assigned to the operator D1 . I will call a the ducial point,4
and write
x
1
aDx f (x)
f ()d
(3)
We then have
D aDx1 f (x) = f (x)
1
aDx D f (x) = f (x) f (a)
(4.1)
(4.2)
x
2
f ()dd
aDx f (x)
a a
x
(5)
3
f ()dd d
aDx f (x)
a a a
..
.
These operators all feedeach in its own characteristic wayon the same
innite set of f (x)-data.
To discuss the higher order implications of (4) I adopt (as frequently I shall
in what follows) an abbreviated notation
m
aDx
Dm
Taking Dp as our seed and proceeding recursively with the aid of (4.1) we
obtain
Dp f = Dp (D D1 )f = Dp+1 D1 f = Dp+1 (D D1 )D1 f
= Dp+2 D2 f
..
.
= Dp+n Dn f
Later we will have reason to set a = 0 (RiemannLiouville), else
(Liouville), else + (Weyl), but it will in general be our practice to leave
the ducial point unspecied until we encounter some specic reason to do
otherwise; ultimately we will, in one specialized connection, allow a (in the
sense limax ) to join the variables of the theory.
4
dierentiation dierentiation
dierentiation integration
obey the law of exponents
integration integration
in this sense:
mn
(dierentiation)
if m n
m
n
(dierentiation) (integration) =
(6.1)
nm
if m n
(integration)
(7)
(6.2)
and to kill the extra terms must constrain f (x) to satisfy conditions of the form
f (a) = f (a) = f (a) = = 0. Similar complications arise typically when one
looks to operators of mixed type Dm Dn Dp Dq but so, on occasion, do
some surprising simplications; look, for example, to the case
D5 D2 D3 D2 f = D5 D2 D1 f = D3 D1 f = D4 f
Elementary preliminaries
+2
+1
D
D
D
D
0
1
2
3
f (x) = f (a)
f (x) = f (a)
at x = a
at x = a
at x = a
f (x) = f (a) at x = a
f (x) = 0
at x = a
f (x) = 0
at x = a
f (x) = 0
..
.
at x = a
t
0
g( ) d d
The rst substantive step toward the creation of a fractional calculus was
taken in when S. F. Lacroixquite casually, and with no evident practical
intentremarked that the familiar formula
Dm xp = p(p 1)(p 2) (p m + 1)xpm
which we notate Dm xp =
case be notated
p!
(pm)!
Dm xp =
(p + 1)
xpm
(p m + 1)
(8.1)
and that (8.1) makes formal sense even when m is not an integer. The fact
that
Dm xp = 0 when m and p are integers with m > p
can then be attributed to the circumstance that (0), (1), (2), . . . are
singular. Proceeding in the other direction, one has
x
1
D1 xp 0Dx1 xp
p d = (p+1)
xp+1
D
2 p
x =
0
p+2
1
(p+2)(p+1) x
..
.
Dn xp =
p+n
1
(p+n)(p+2)(p+1) x
p!
p+n
(p+n)! x
(p + 1)
xp+n
(p + n + 1)
(8.2)
Dm Dn xp =
= Dmn xp
(9)
D
: any number, real or complex
fp xp
p
and did not fail to note that such a calculus would give surprising results even
in the simplest cases; one has, for example, the semiderivatives
1
x
(2) 1
2 1
2 = x2 = 2
D2x =
x
( 32 )
Grunwalds
construction
( 12 )
1
2
1
x
(10)
f (x) =
1
2
f () d
cos p(x a) dp
1
2
f () d
p cos p(x a) + 12 dp
f (x) f (x h)
h
1 ehD
f (x)
h
(11)
(12)
h0
10
(xh)f (x2h)]
= lim [f (x)f (xh)][f
h2
h0
(x2h)f (x4h)
D f (x) = lim f (x)3f (xh)+3f
h3
3
D f (x) = lim
h0
..
.
h0
(13)
The operators Dm , as (13) makes explicitly clear, feed on nite f (x)-data sets
of ascending size, and are in this sense local operators.
In view of the structure of (12) it becomes entirely natural to relax the
requirement that m be an integer, writing (for example)
1
1
D 2 = lim 1 12 ehD + 18 e2hD +
h0
h
and
D1 = lim 1 + ehD + e2hD + e3hD h
h0
(14)
(15)
= lim
k=0
h0
ekhD f (x) h
(16)
f (x kh) h
k=0
f ()d
=Dx1 f (x)
f ()d
a
N
1
x a
x a
= lim
f xk
N
N
N
k=0
(17)
11
Grunwalds
construction
h=(xa)/N
m
aDx f (x) = lim
(18)
k=0
At m = 1 the preceding formula, for all of its seeming complexity, yields the
simple result
f (x) f (x xa
1
N )
aDx f (x) = lim
xa
N
When we insert
m
k
(m + 1)
(k + 1) (m + 1 k)
m
aDx f (x) lim
which assigns natural meaning to the derivative operators aDx of all (integral/
non-integral non-negative) orders 0.5 If m is a negative integer, then
n m is a positive integer, and we have
()k
m
k
n(n + 1)(n + 2) (n + k 1)
k!
(n + k 1)! n + k 1
=
=
k
k!(n 1)!
(k + n)
=
(k + 1) (n)
12
n
aDx f (x) lim
(19.2)
k=0
aDx
Those operators feed in every case on the same f (x)-data, which however they
weight in distinct waysvery simple distinct ways, as will soon emerge.
3. The Riemann-Liouville construction. In the case n = 2 the denition (19.2)
gives
2
aDx f (x) lim
1
x a 2 N
k=0
x a
(k + 1)f x k
N
(20)
13
w(x, y)
=1
x
and
(x y)f (y) dy =
a
f (y) dy
a
y)
f
(y)
dy
=
2(x y)f (y) dy
dx2 a
dx a
x
=2
f (y) dy
a
dn
dxn
..
.
(x y)n f (y) dy = n!
a
f (y) dy
n = 0, 1, 2, . . .
(22.1)
My sources are Oldham & Spanier, 2.7 and R. Courant, Dierential &
Integral Calculus (), Volume II, p. 221. Courants, by the way, is the only
text known to me that even mentions the existence of a fractional calculus; see
his Chapter IV, 7. The topic is not mentioned in Volume I of R. Courant &
D. Hilbert, Methods of Mathematical Physics (), but is mentioned twice in
their Volume II; specic citations will be given later.
8
14
y
a a
y
y
x
a
x yn
yn1
y2
(x y) f (y) dy
a
f (y) dydy dy =
a
f (y) dydy =
..
.
f (y1 ) dy1 dy2 dyn =
a
a a a
n-fold iterated
(x y)2 f (y) dy
1
2!
1
(n1)!
which Oldham & Spanier attribute to Cauchy. Bringing (22.2) to (5), we have
x
n
n1
1
D
f
(x)
=
f (y) dy : n = 1, 2, 3, . . .
(23)
a x
(n1)! (x y)
0
Upon (23) hangs the computational essence of the fractional calculusas will
emerge. But the question immediately before usassured, as we now are, that
(5) (23)is this: How does (20) relate to (23)? The answer is that they
are identical , as I now show. Write
1
a
xa
yk x (k + 1)
= (1 N )x + N x at k = 0
N
a at k = N 1
Then x yk = (k + 1)(x a)/N , y yk+1 yk = (x a)/N and (20) can
be notated
N
1
2
aDx f (x) = lim
k=0
x
x
(x yk )f yk
(y)
N
(x y)f (y)dy
=+
0
1
x a n 1 N
(k + n)
x a
= lim
f xk
N
N
(n)
(k + 1)
N
k=0
x
1
(x y)n1 f (y) dy
=
(n) a
15
aDx f (x)
= lim
x a
1
=
()
N 1
x a
1 (k + )
f xk
()
(k + 1)
N
(24.1)
k=0
(x y)1 f (y) dy
(24.2)
it does not follow that F (z) = G(z) : all z; this is the interpolative ambiguity
problem mentioned previously. In a more careful account of the fractional
calculus one would have to describe the conditionsconditions on the structure
of f (x)under which the right sides of (24) are in fact equal. This is work
which I am happy to leave to the mathematicians;10 as a physicist, I know
myself to be protected from major faux pas by the well-constructedness of
Nature; it is my habit to look closely to my informal tools only when they
have led me to an implausible result. In practice, (24.1) and (24.2) seldom lead
to outright contradiction for the simple reason that (24.1) is, except in trivial
cases, computationally unworkable. In practice, one usually treats (24.2) as a
stand-alone denition:
x
1
D
f
(x)
C
1
f () d
( z)m+1
Such an approach to the fractional calculus (of analytic functions) was explored
by A. V. Litnikov, N. Ya. Sonin and H. Laurent in the s and s. For
discussion, see Miller & Ross, p. 28 or Oldham & Spanier, p. 54.
16
= Dm aDx f (x)
4
3
aDx
aDx
aDx
f (y) dy
dx ( 12 ) a
xy
1
x
1
1
dy
xy
a
2 xa
(x a)
and at a = 0 gives back Lacroix result (10). Is this result surprising? Not, I
think, when viewed in context: writing D 0Dx and making free use of
the identity z (z) = (z + 1), we obtain12
x
1
1
D 1 u(x; ) =
(x y)1 dy =
x
() 0
()
1
=
:
>0
x
(1 + )
(1 )
D+ 1 D1 D(1) 1 = D u(x; 1 ) =
x
(2 )
1
x : 0 < < 1
=
(1 )
Curiously, this variant of the same argument
D+ 1 = D2 D(2) 1 = D2 u(x; 2 ) =
=
12
1
x
(1 )
(2 )(1 )
x
(3 )
0<<2
17
Further elaborations
D 1 U (x; ) =
(26)
and to attribute the familiar (but from this point of view remarkable) fact that
Dm 1 = 0 (m = 1, 2, 3, . . .) to the circumstance that (1 ) has poles at
precisely those integral points.
Returning again to (25), we set a = 0 (this can always be achieved by a
simple change of variable) and obtain
x
1
D f (x) =
(x y)1 f (y) dy : > 0
(27)
() 0
Riemann-Liouville fractional integral transform of f (x)
The fractional integral transform is somewhat reminiscent of the
Hilbert transform[f (x)]
P
(x y)1 f (y) dy
primitive statements as
linearity: D(f + g) = Df + Dg
product rule : D(f g) = Df g + f Dg
chain rule : Df (g(x)) =
df
dg
dg
dx
18
the ordinary calculus tend generally within the fractional calculus to become
innite sums; Leibniz formula, for example, becomes (in one of its variant
forms)
D (f g) =
Dk f Dk g
k
k=0
then13
()n (n) (s) = L[xn f (x)]
sn (s) sn1 f (0) sn2 f (0) f (n1) (0) = L[f (n) (x)]
which are in some respects strongly reminiscent of (7). The expressions on the
right/left/right/left sides of the preceding equations invite relaxation of any
presumption that n be an integer. Special interest attaches to the observation
that if (s) = L[f (x)] and (s) = L[g(x)] then
x
(s) (s) = L[
f (y)g(x y) dy]
(28)
0
Since the right side of (28) involves precisely such a convolution integral as
appears on the right side of (27), it would appear to be quite easy to discuss
the Laplace transform properties of fractional integrals; specically, we have
x
1
L[Dm f (x)] =
L[
(x y)m1 f (y) dy]
(m)
0
1
=
L[xm1 ] L[f (x)]
(m)
= sm
But Laplace transform methods, it should be borne in mind, are applicable only
to a restricted subset of the set of functions susceptible to the more general
methods of the factional calculus.14
13
19
ILLUSTRATIVE APPLICATIONS
According to Miller & Ross, the fractional calculus nds use in many elds
of science and engineering, including uid ow,15 rheology, diusive transport
theory,16 electrical networks, electromagnetic theory, probability and statistics
. . . viscoelasticity andof all subjectsthe electrochemistry of corrosion.
And, though the vision of a fractional calculus came to Leibniz and others in
moments of idle speculation, it does appear to the be case that concrete progress
in the eld was accomplished mainly by persons who drew their inspiration
from specic problems of an applied nature. It is perhaps a measure only of my
own limitations that I nd most interesting the applications to mathematics
itself, and to problems derived from physics. In following paragraphs I discuss
two physical applications of historic (but continuing) interest, several derived
from my own work, and one derived from the most recent issue of the journal
CHAOS.
5. Abels solution of the tautochrone problem. A mass m slides, under inuence
of gravity, along a frictionless wire, of which x(y) serves to describe the gure.
For convenience we place the bottom end of the wire at the origin: x(0) = 0. If
the mass is released at heighth > 0, then its speed v(y) when is has descended
to height y (0 y h) isby energy conservation, and irrespective of the
gure of the wiregiven by
v 2 = 2g(h y)
Let s(y) denote arc length, as measured along the wire from the origin to the
point [x(y), y] = [x(s), y(s)]:
s(y)
!
1+
dx 2
dy
dy
ds
= s (y)y =
dt
2g(h y)
(h) time of descent =
0
15
1
2g(h y)
s (y) dy
20
The tautochrone problem asks for the design of the curve with the property
that (h) is in fact independent of h.17 To ask for such a curve C is, by notational
adjustment of the result just obtained, to ask for the function s(y) such that
y
1
2g T =
(y z) 2 s (z)dz
(29)
0
2g T / ( 12 )
2g T 2 / T, a constant of prescribed value
By integration
with A
2g (T /)2
s(y) = 2Ay 2
(30)
(31)
since s(0) = 0. Extraction of a cycloid from (31) entails only some relatively
uninteresting analytical geometry; backing up to (30) we have
!
dx
2
1
1+
= Ay 2
dy
giving
dx
=
dy
A2 y 1 1
A2 z 1 1 dz
by x(0) = 0
arcsin Ay
= 2A
cos2 d by z A sin2
x(y) =
= A( +
17
1
2
sin 2)
21
(32.1)
y() = A sin2 21
= R(1 cos )
(32.2)
yield with equal ease to the line of argument which brought us to (31). We note
(see below) that the success of that line of argument hinges critically on the
linearity of the gravitational potential U (y) = mgy. And that the tautochrone
problem had been posed and solvedby other meanslong before Abel entered
the picture; Huygens, by , had discovered the tautochronous property of
the cycloid and made it the basis of a famous horological invention.18
I digress here to pose this arcane question: Does tautochronicity imply
harmonicity? The question derives from the familiar fact that if a mass m
attached to a spring of strength k is removed from the origin to the point
x = A and then released, it returns to the origin in A-independent time
T = 14 (period)
Thats what me mean when we say that the oscillator is harmonic. Evidently
harmonicity tautochronicity. To ask (as we now do) Is also true? is,
in eect, to ask For what potentials U (x) does
T (x)
0
"
1
2
m [U (x)
U (y)]
dy
d
have the property that dx
T (x) = 0? The problem, thus formulated, appears
on its face to lieexcept in the gravitational case U (x) = a + bx considered
by Abelbeyond the purview of the fractional calculus, so I set it aside for
another day; that the motion s(t) of Abels cycloidally constrained particle is
18
22
in fact harmonic is, however, quite easy to demonstrate: returning with (31) to
the energy conservation equation that served as our point of departure, one has
= E 12 m 2 s2
2
1
2 ms
2
=
2T
period
(x, t) = 0
x
t
a>0
= aD2
t
e(ax
+2bx+c)
dx =
b2 ac
exp
a
a
(a) > 0
1
=
4at
e 4at eD d
1
(1/4at) > 0
(33)
Evidently
t (x) =
1
4at
e 4at 0 (x ) d
1
by Taylors theorem
g(, t)0 (x ) d
g(x, t)
19
2
1
1
e 4at x
4at
(34)
(35)
23
One easily establishes that g(x, t) is itself a solution of the heat equation, and
has these special properties:
g(x, t) dx = 1
i m (x y)2
m
exp
2it
2
t
with b2
1
a
(36)
then gives
t (x) = Aebx
= Aebx
+ Be+bx
= b pt
x t
20
with b > 0
(37)
24
Writing
Aebx
(bx)n d
12 n
A
n!
dt
n=0
(bx)n d
12 n
=A+
A
n!
dt
=A
n odd
by
d
non-zero integer
dt
A=0
(bx)2m+1 d m d 12
A
(2m + 1)! dt
dt
m=0
=A 1
A
=
t
by Lecroix (10)
()m
(bx)2m+1
m! (2m + 1)22m tm+ 12
m=0
bx/2t
=2
2m d
0
= A erfc( 12 bx/ t)
(38)
where I have appealed to the denitions of the error funtion erf(x) and its
complement erfc(x). The preceding remarks have been adapted from Miller
& Ross somewhat disparaging account of a line of argument rst advanced by
Oliver Heaviside () in connection with the theory of transmission lines, and
have much in commonboth in spirit and in detailwith 8 & 9 of Widders
Chapter III . More recent work in this same ancient tradition has been concerned
with the application of the fractional calculus to (for example) the study of the
fractional diusion equation and diusion on fractal domains.21
21
= bp 1/ t
x t
where is the so-called anomalous diusion exponent; in some applications
(random walk on a Cantor set) turns out to be closely related to fractal
dimension. A more recent reference (for which I am indebted to Oz Bonm) is
B. J. West et al , Fractional diusion and Levy stable processes, Phys. Rev. E
55, 99 (1997).
25
() that if (x, 0) and t (x, 0) are the values assumed initially (i.e., at
time t = 0) by a eld (x, t) and its time-derivative, then the solution of the
one-dimensional wave equation
=0
x2
1 2
u2 t
1
2
(x + ut, 0) + (x ut, 0) +
1
2u
x+ut
t (y, 0) dy
(39)
xut
(x, t) =
22
(y, 0)Gt (x y, t) + t (y, 0)G(x y, t) dy
(40)
The constructive argument runs as follows: write (as one invariably can)
(x, t) = f (x + ut) + g(x ut)
Then
(x, 0) = f (x) + g (x)
1
u t (x, 0)
1
u
= f (x) g (x)
give
f (x) =
g(x) =
1
2
1
2
(x, 0) +
(x, 0)
1
u
1
u
t (y, 0) dy
t (y, 0) dy
from which (39) follows at once. Or one can simply verify that (39) does in fact
possess the stated properties.
26
where
G(x y, t)
=
1
2u y (x ut) y (x + ut)
1
2u (y x + ut)) (y x ut)
2 2
2
(41)
= (u t (y x) ) according as t 0
Gt (x y, t) = 12 (y x + ut)) + (y x ut)
The (x, t) that appears on the left side of (40) satises the wave equation
because G(x y, t) does (and so also, therefore, does Gt (x y, t)); within
the population of solutions, the particular solution G(x y, t) possesses these
distinguishing/dening features:
G (x y, 0) = 0
#
(42)
Gt (x y, 0) = (x y)
The function G(x y, t) describes the eld that results when the quiescent eld
is given an initial kick at the point x = y, and (40) describes how general
solutions (x, t) are to be assembled by superposition of such special solutions.
A wonderful feature of (40) is that it is structurally so robust; an argument
of famous elegance23 shows that the solutions of virtually any sensible wave
equation can, in terms of prescribed Cauchy data, be described by an equation
of type (40); all that changes, when one moves from wave system to wave
x y , t).
system, is the precise meaning assigned to the Greens function G(x
In the winter of / I was motivated by Richard Crandalls then
on-going experimental eort to measure the mass of a photon24 to study
the 3-dimensional wave system
(
+ 2 ) = 0
(x2 + y2 + z2 )
1 2
u2 t
(43)
23
For an account of that argument, see relativistic classical fields
(), p. 160 et seq. The argument hinges on a generalization of Greens
theorem
dn x =
d
R
R
which was invented () for essentially this purpose, and is itself usually
considered to be a corollary of Stokes theorem (which, however, came later).
24
See R. E. Crandall, Photon mass experiment, AJP 51, 698 (1983);
R. E. Crandall & N. A. Wheeler, Klein-Gordon radio and the problem of
photon mass, Il Nuovo Cimento 80B, 231 (1984) and R. Leavitt, A photon
mass experiment: an experimental verication of Gausss Law, (Reed College,
1983).
27
which gives back the 3-dimensional wave equation 3 = 0 in the limit 0.25
I was, at the time, teaching the rudiments of elementary wave theory to a class
of sophomores and savoring the splendid little monograph The Mathematical
Theory of Huygens Principle by B. B. Baker & E. T. Copson (2nd edition ),
who were themselves strongly inuenced by M. Riesz then fairly recent success
in clarifying (by appeal to a generalization of the fractional calculus) the work
of J. Hadamard;26 I found it therefore natural to study (43) in a context which
considers dimension to be a variable, and to include the results of my research
in my sophomore notes: introduction to the analytical methods
of physics (), where in all their extravagant detail they can be found on
pp. 366433. The wave systems treated there are the free-eld Klein-Gordon
systems
(
+ 2 ) = 0
N
n2
1
u2 t
(44)
n=1
=0
and
x y , t) =
GN (x
1
(2)N
sin ut
u
k 2 + 2 ik r
dk1 dk2 dkN
e
k 2 + 2
(46)
28
now in closer detail to the expression interior to the curly brackets (where all
the r-dependence originates) we have
etc. =
1 J 1 (kr)
(kr) 2 2
1
=
J (kr) k
(kr)0 0
1
= k2
1 J 1 (kr)
(kr) 2 2
1
= k2
J (kr) k
(kr)1 1
1
= k4
3 J 3 (kr)
(kr) 2 2
1
= k4
J (kr) k
(kr)2 2
..
.
when N = 1
when N = 2
when N = 3
when N = 4
when N = 5
when N = 6
1 n 12
k 2n kr
Jn 12 (kr)
2n 1 n
k
Jn (kr) k
kr
when N = 2n + 1 is odd
when N = 2n + 2 is even
But29
1 d
n 1
J
(z)
=
J
(z)
z n+ n+
z dz
z
1 d n
n
and trivially kr
= k 2n 1r r
, so we have
d(kr)
1
1 n
r r
2 "
etc. =
= 2 cos kr
1 n
r r J0 (kr) k
when N = 2n + 2 is even
(48)
(49.1)
(49.2)
29
where
1
x y , t) G1 (r, t) =
G1 (x
k 2 + 2
sin ut
u
k 2 + 2
cos kr dk
sin ut k 2 + 2
1
x y , t) G2 (r, t) =
G2 (x
J0 (kr)k dk
2 0
u k 2 + 2
1
sin(ut) 2 + k 2
=
k
J0 (kr) kr dk
2u r 0
2 + k 2
(50.1)
(50.2)
The Greens functions G1 (r, t) and G2 (r, t) acquire special importance from the
circumstance that, according to (49), they are the seeds from which arise the
parallel constructions
#
G1 (r, t) G3 (r, t) G5 (r, t) G7 (r, t)
(51)
G2 (r, t) G4 (r, t) G6 (r, t) G8 (r, t)
The integrals which at (50) serve to dene G1 (r, t) and G2 (r, t) are tabulated;
consulting Volume I of A. Erdelyi et al , Tables of Integral Transforms ()
we nd (at 1.7.30, p.26 in the table of Fourier cosine transforms) that
1 J0 ( (ut)2 r2 ) if (ut)2 r2 0
2u
G1 (r, t) =
0
otherwise
1
= (s2 )
with s (ut)2 r2
(52.1)
J (s)
2u 0
according as t 0, while in Volume II of that same work (at 8.7.20, p. 35 in
the table of Hankel transforms30 ) we nd that
r
1
2
(s )
G2 (r, t) =
J 1 (s)
2s 2
2u r
2
=
cos s
s
1 cos s
= (s2 )
(52.2)
2u s
Equations (52) owe their analytical simplicity in part to the denition
s
30
(ut)2 r2
f (x)
f (x)J (xy) xy dx : y > 0
0
30
1
J ( )
2u 0
1 cos
G2 () = ()
2u
G1 () = ()
(55.1)
(55.2)
G02 () = ()
2u
G01 () = ()
(56.1)
(56.2)
which are remarkable for their simplicity. Equation (56.1) reproduces precisely
(41), while by straightforward extension of a line of argument familiar from
p. 16 we have
1
1
1 1 1
0
2
D G1 () = D
( ) 2 ( ) d
2u
0
$
2
if > 0
=
0
otherwise
1
= ()
2u
31
(57)
1
1
D 2 J0 (a x) = D1 D 2 J0 (a x)
x
1
1
= D1
(x y) 2 J0 (a y) dy
0
2
sin(a
x)
1
=D
according to Mathematica
a
cos(a x)
=
x
to obtain
1
12
G1 () = G2 ()
from which (57) can be recovered as a limiting case. Conversely,
(58)
1
12
1
1
1 cos
G2 () =
( ) 2 ( )
d
2u
0
1
2
12 cos
= () 3
( )
d
2 2 u 0
= J0 ( )
= G1 ()
so we have
1
1
G1 ()
1
1 1
12
=
G2 ()
G3 () =
1
+ 12
The implication of the results now in hand is that (compare (53)) it makes sense
to write
1
N21
GN () =
G1 ()
(59)
and that in renement of (54) we have
G1 () G2 () G3 () G4 () G5 () G6 ()
(60)
32
GN (rr, t) drN =
N
(2)
u k 2 + 2
N
1
n
N
dk1 dk2 dkN drN
exp i
kn r + kN r
n=1
1
(2)N 1
sin ut
u
k 2 + 2
k 2 + 2
1
2
eiKR dR
= (K)
1
N
exp i
kn rn dk1 dk2 dkN 1 dK
n=1
= GN 1 (rr, t)
(61)
J (k r2 + s2 )
J 12 (kr)
ds
=
1
2k r 12
(r2 + s2 ) 2
0
of Sonines formula.33 Hadamards method achieves
G1 () G2 () G3 () G4 () G5 () G6 ()
(62)
33
1
()
2u
1
1
1
0
()
=
G1 () =
2u
1
2
1
G05 () =
G01 () = 2 ()
2 u
..
.
1
n
1
0
0
G2n+1 () =
G1 () = n (n) ()
2 u
G03 ()
(63.1)
1 1
2 ()
2u
1
1
3
1 1
G04 () =
G02 () = 2 2 () 12 2 ()
2 u
1
2
3
5
1 1
0
0
G6 () =
G2 () = 3 2 () 2 () + 34 2 ()
2 u
..
.
n
1
n
(np)
n 1 (p)
1
0
0
G2n+2 () =
2
G2 () = n+1
(63.2)
()
2
u p=0 p
1 (p)
(np)
12 p
where 2
= ()p 2(2p)!
and ()
= (np1) () : 0 p < n.
2p p!
From (63) we learn that
In all cases, GN0 () vanishes outside the lightcone;
G0odd3 () is singular on the lightcone, but vanishes inside;
G0even () is singular on the lightcone, butowing to the
presence of a dangling -functionfails to vanish inside
(as also does G01 ()); it follows that radiative events in odddimensional spacetimes have persistent local eects. This
is in sharp contrast to the situation in spacetimes of even
dimension N + 1 4.
Amongst the cases G0odd3 () the case G03 ()which refers, of course, to
the world we physically inhabitis special, in a sense which we are in position
34
1
rN 1
N 3
2
sin utk JN 2 (kr) kr dk
(64)
J 12 (kr) kr =
2
cos kr
2
J+ 12 (kr) kr =
sin kr
(
)
2 1
J+ 32 (kr) kr =
sin kr cos kr cos kr
kr
..
.
2 2n1 1 d
n1 sin z
J 2n1 (kr) kr =
z 2
2
z dz
z
n = 1, 2, 3, . . .
z=kr
1
u
k 1 sin
utk cos rk dk
= 12 sin k(r + ut) sin k(r ut)
1
2 2 ur
0
k 0 sin
utk sin rk dk
= 12 cos k(r + ut) + cos k(r ut)
35
sin
rk
sin
utk
cos
rk
dk
4 3 u 0
r3
r2
= superposition of running waves with distinct attenuation factors
1 0
=
G (r, t)
2r r 3
1
2 1 1
=
(65)
G0 (r, t)
2 r r r r 1
Wave propagation is, in this and higher-dimensional cases, dispersive because
some r
-operators see 1r -factors standing to their right. The implication is that
non-dispersive telegraphy is possible only
in 2-dimensional spacetime (case N = 1), which is arguably
too simple to support physicists, and
in 4-dimensional spacetime (case N = 3), which manifestly
is not
It is dicult to escape the feeling that the remarkable fact thus exposed must
have something to do with why space is 3-dimensional. The even-dimensional
cases are non-contenders for reasons that we have traced to (10), i.e., to the
curious fact that
1
the semiderivative (with respect to x) of unity =
x
Remarks which are in many respects qualitatively similar pertain to the
Klein-Gordon Greens functions G N (r, t) : > 0. All K-G systems are,
however, dispersive, and the analytical details tend (as we have seen) to be
more intricate.34
8. Application to the mensuration of hyperspheres. It is well known that the
when N = 2n is even
p! r
N
N
VN (r) =
=
r
2 n
2n
(1 + N2 )
2n+1
when N = 2n + 1 is odd
135(2n+1) r
34
36
1
1
()
( y)1 dy =
0
(1 + )
N
( )N
(1 + N2 )
= VN ( )
VN ()
N D 2 1 =
N
VN () =
(66)
and then set = r2 . Equation (66) isexcept for the reversed sign in the
exponentstrongly reminiscent of (59), and gives rise to a Riesz construction
V1 () V2 () V3 () V4 () V5 () V6 ()
(67)
VN (r) =
N
rN
(1 + N2 )
(68)
37
One has
V1 (1) = 2.00000 = 2
V2 (1) = 3.14159 =
V3 (1) = 4.18879 = 43
V4 (1) = 4.93480 = 12 2
V5 (1) = 5.26379 =
V6 (1) = 5.16771 =
V7 (1) = 4.72477 =
V8 (1) = 4.05871 =
V9 (1) = 3.29851 =
V10 (1) = 2.55016 =
..
.
V = 0.00000
8 2
15
1 3
6
16 3
105
1 4
24
32 4
945
5
1
120
d
Looking to dN
log VN (1), we nd that VN (1) is maximal when ( N2+2 ) = log ,
d
where (z) dz
log (z) = (z)/ (z) denes the digamma function. With
assistance from Mathematica, I discover that
did not have fractals in mind, or the outreaches of 20th Century physics, when
he undertook to prove by counterexample that
continuity does not imply dierentiability
This he did by exhibiting (at a meeting of the Berlin Academy in ) the curve
that now bears his name. Related work had been done by Bernard Bolzano in
1834, by Bernard Riemann in the early 1860s and simultaneously by the Swiss
mathematician Charles Cellerierall of whom elected (as also did Weierstrass)
not to publish their ndings.36
For many years Weierstrass creation lived in legend as a mathematical
curiosity that physicists were happy not to have to worry about. But with
36
R(t) =
n2 cos n2 t
1
and turned out to be only nearly but not quite nowhere dierentiable; see
J.Gerver, The dierentiability of the Riemann function at certain rational
multiples of , Amer. J. of Math. 92, 33 (1970).
38
recognition of the fractal aspects of the natural world all that has changed; one
of the best recent accounts of the properties of the Weierstrass curve and its
cognates is, in fact, the work of a pair of physicists, who study the function37
W (t) =
[1 ei t ]ein
(2D)n
A(t) = [W (t)]
=0
1 cos n t
(2D)n
()n sin n t
(2D)n
U (x) = A/x
(x) = 0 at x =
with A >
1
4
n = 0, 1, 2, . . .
39
0<1
40
: 0<1
aDx f (x)
To evaluate what they call the local fractional derivative of f () at = x
they proceed as follows:
step one: Form the function F () f () f (x), which has by design the
property that F (x) = 0. We note that if f () is regular at = x
f () = f (x) +
1 (p)
f (x)( x)p
p!
p=1
F () =
1 (p)
f (x)( x)p
p!
p=1
then
D
F
(x)
=
D
(x y) [f (y) f (x)] dy
a x
(1 ) a
step three: Proceed to the limit (when it exists), writing
Dx f (x) lim aDx F (x)
ax
x
1
=
lim D (x y) [f (y) f (x)] dy
(1 ) ax
a
where D
d
dx
(69)
C = constant
= f (x)
41
all f (x)
aDx
1
C : 0 < a < x, [] < 1
(1 ) (x a)
= Lacroix curious equation (10) when = 12 and a = 0
= Dx C, which was just seen to vanish for all C
C=
If x is a regular point of f (x) (as it would, of course, be in the special case just
studied: f (x) = C) then
1 (p)
f (x)(y x)p dy
ax
p!
a
p=1
x
p
1
() (p)
=
(x y)p dy
lim D
f (x)
(1 ) ax p=1 p!
a
x
1
()p (p)
=
lim
f (x) (p )
(x y)p1 dy
(1 ) ax p=1 p!
a
(x a)p
=
p
p
1
() (p)
=
lim
f (x) (x a)p
(1 ) ax p=1 p!
Dx f (x) = lim D
1
(1 )
(x y)
=0
which gives back Dx f (x) = 0 in the case f (x) = C.43 Evidently
Dx f (x) = 0
(70)
sin k t
(2D)k
k=1
of Weierstrass A(t) function, and manage with relative ease to establish that
Dt W (t) = 0
43
(71)
The result just achieved is not to be confused (continued on the next page)
42
1 (p)
D
f
(x)
=
D
f (a)(x a)p
a x
a x
p!
p=0
x
1
1
f (p) (a)
D (x y) (y a)p dy
=
p!
(1 ) a
p=0
1 (p)
(p + 1) (1 ) (p + 1)
=
f (a)
(x a)p
p!
(1
)
(p
+
2)
p=0
Use (p + 1) =
=
(p + 2)
(p + 1)
1 (p)
(p + 1)
f (a)
(x a)p
p!
(p
+
1)
p=0
Compare (31)
44
43
Q=
f (x) dx
0
while the electrostatic energy is (in suitable units) described by the functional
1
f (x)f (y)
dydx
|x y|
0 0
x
1
1
=2
f (y) dy dx
f (x)
0
0 xy
E[f (x)] =
The expression interior to the brackets is in fact divergent, but only weakly so,
in this sense:
x
: [p] 1
1
dy
=
p
x1p : [p] < 1
0 (x y)
1p
It sits, in other words, right at the leading edge of the divergent regime.
Assuming the unit interval to be conductive (informally, a needle), we are
led to consider this tempered version of a problem previously studied (with
inconclusive results) by David Griths and his student, Ye Li:46 Find the f (x)
which minimizes
x
1
1
F [f ] lim
f (x)
f
(y)
dy
dx
(72)
1
0 0
0 (x y)
subject to the constraint
G[f ]
f (x) dx = 1
(73)
(74)
[x(1 x)] 3
G[f ] = A + gB
46
with g
( 23 )
1
3
2 ( 76 )
= 2.053390
44
B = B(A) 0.4869995(1 A)
Returning with these to (72)and suspending temporarily the operation 0
we obtain
F [f] = A2 p + ABq + B 2 r
where {p, q, r} are certain A-independent functions of which I will describe in
a moment. Dierentiation with respect to the adjustable parameter A gives
F [f] = 2Ap + (B + AB )q + 2BB r
with B = k
A=
(2k 2 r
(75)
Looking now to the construction of p(), q() and r() we nd that the rst of
those functions
1 x
1
1
1
p() =
dy
dx =
1
(1 + )
0
0 (x y)
is easy to evaluate, but becomes singular in the limit. We infer on the basis of
(75) that q() and/or r() must become similarly singular if a nite result is to
be achieved in the limit. Looking now to those, we have
x
1 x
1
1
1
q() =
(y) dy dx +
(x)
dy dx
1
1
0 (x y)
0 (x y)
0
0
r() =
0
(x)
0
q1 ()
x
q2 ()
1
(y) dy dx
(x y)1
( 34 )
7
5.299916
1 2 3 3 ( 76 )
r()
=
( 23 )
q1 () and q2 ()
45
Q with Q 2 ( 23 )/ ( 34 )
= 2.053390
7
7
2
R with R 2 3 3 ( 6 )/ ( 3 ) = 5.299916
1 kQ
= 3.89145(1kQ)
k(kR Q)
1
2 3 ( 76 ) ( 23 )
kQ =
= 1.000000
( 23 ) ( 43 )
Evidently kQ = 1 exactly, in conseqence of properties of the -function which
I will not linger to spell out. The implication is that = 0, therefore A = 1,
therefore B = 0, therefore thatself-energy apartthe least energy achievable
within the class of trial distributions f(x) is in fact achieved by the at
distribution
0
fmin (x) = 1 +
(76)
1
[x(1 x)] 3
This result is at seeming variance from the result reported by Griths & Li,47
but in recent conversation Griths has admitted to a growing suspicion that his
gures are probably best read as an illustration of the slowness of the approach
to an equilibrium distribution which he has come to suspect is in truth at.
In their paper, Griths & Li detect (in the behavior of models alternative to
their bead models) certain ambiguous hints of a atness which they cannot
absolutely exclude [as a] counterintuitive possibility, but seem more inclined
to read those as evidence that the [needle] problem is ill posed (in the sense
that the answer depends upon the model adopted).
The preceding discussion does serve to inspire condence in the reliability
of our regularization procedure 0, but cannot exclude the possibility that
there might conceivably exist a non-at distribution f (x) with lower energy than
the at distribution, for we have actually established only that (76) is optimal
within one specic population {f(x)} of test functions. To do better we must
have recourse to more general methods. As a rst step toward that objective,
we recall the Riemann-Liouville denition (25) of the fractional integration
operator
x
1
1
D f (x)
f (y) dy : > 0
() 0 (x y)1
47
46
and note that the energy functional (72) can in this notation be described
F [f ] = lim ()
0
f (x)D f (x) dx
I draw attention now to the fact that in (for example) the case
u(x) 1
one has
unit at distribution
1
F [u] = lim
0
1
dy
(x y)1
1
= lim
0 ( + 1)
=
I propose to resolve this instance of the self-energy problem not by subtraction
of an innite term, but by dividing out the factor responsible for the singularity.
Noting that48
2
1
2
= 1 + +
+
+
()
2
12
we construct the renormalized energy functional
F[f ]
1
F [f ] = lim
0
()
f (x)D f (x) dx
(77)
Transparently, F[u] = 1: regularization has (at least in this typical case) been
achieved, by a mechanism which which is natural to the fractional calculus.49
It is in this modied context that we do our calculus of variations.
Writing
F[f + g] = F[f ] + lim
0
f (x)D g(x) + g(x)D f (x) dx +
g(x) dx
0
48
f (x)D g(x) + g(x)D f (x) dx = 0
(78)
47
g(x) dx = 0
(79)
n=1
fn cos 2nx
gn cos 2nx
n=1
g0 = 0
F[f ]
= f0
W0n gn +
fm Wmn gn = 0
g
n=1
m=1 n=1
all {gn }
where
Wmn Vmn + Vnm
with
cos 2mx D cos 2nx dx
0
1
x
1
1
cos 2mx
cos 2ny dy dx
= lim
0 0
() 0 (x y)1
Vmn lim
1
1
cos 2ny dy = x() p Fq 1 , 12 + 2 , 1 + 2 , n2 2 x2 )
1
() 0 (x y)
V0n =
cos 2nx dx = 0
n = 1, 2, 3, . . .
m, n = 1, 2, 3, . . .
Vmn =
0
F[f ]
fn gn
=
g
m=1
=0
48
n = 0, 1, 2, . . .
D 2 ex
ey dy
1
xy
(2) 0
= ex erf( x)
1
1
DD 2 ex D+ 2 ex = ex erf( x) + 1x
3
1
D+ 2 ex = ex erf( x) + 1x 1 2x
5
1
D+ 2 ex = ex erf( x) + 1x 1 2x
+
..
.
(80)
3
4x2
49
Applications to analysis
D eax =
E(x; , a)
where (, ax) and (; ax) are variant forms of the incomplete gamma
function.50 The importance of the (nameless) function E(x; , a) has been
emphasized particularly by Miller & Ross, who list many of its properties in
their Appendix C. Drawing upon some of that material, one has (for n =
0, 1, 2, . . .)
Dn D eax = Dn eax = Dn E(x; , a)
= E(x; n, a)
Since D ex = ex unless n is a non-negative integer, the question arises:
What function e(x; ) does have the property that
D e(x; ) = e(x; )
>0
(81)
To approach the problem we look rst in an unfamiliar way to the familiar case
= 1; we construct
e(x; 1) = 1 + D1 + D2 + D3 + g(x; 1)
and impose upon the generator g(x; 1) the requirement
Dg(x; 1) = 0
so as to achieve
De(x; 1) = 0 + D0 + D1 + D2 + g(x; 1) = e(x; 1)
If, in particular, we take g(x; 1) 1 then we obtain
e(x; 1) = 1 + x +
1 2
2! x
1 3
3! x
+ = ex
e(x; 2) = 1 + D1 + D2 + D3 + g(x; 2)
D2 g(x; 2) = 0
50
50
Necessarily g(x; 2) = g0 + g1 x, so
e(x; 2) = g0 1 +
1 2
2! x
1 4
4! x
+ + g0 x +
1 3
3! x
1 5
5! x
= g0 cosh x + g1 sinh x
and we obviously have only to set g0 = g1 = 1 to recover ex . Proceeding
similarly from g(x; 3) = g0 + g1 x + 12 g2 x2 we obtain
e(x; 3) =
g0
k=0
x3k
x3k+1
x3k+2
+ g1
+ g2
(3k)!
(3k + 1)!
(3k + 2)!
(82)
xkn+p
Wp (x; n)
(kn + p)!
k=0
k=0
p = 0, 1, 2, . . . , n 1
1
xkn+p
(1 + kn + p)
(83.1)
(83.2)
n = 0, 1, 2, . . .
(84)
Wp (x; n) = ex
(85)
p=0
e(x; ) = 1 + D + D2 + D3 + g(x; )
(86)
(1)
1
g(x)
(1 )
0
1
g(y) dy = constant
(x y)
(87)
51
Applications to analysis
1 1
x
()
(all x)
1
x(k+1)1
((k + 1))
to (86) we obtain
E(x; )
1
xk1
(k)
(88)
k=1
1
x
x21
x31
+
+
+
() (2) (3)
() x01
(2) x21
(3) x31
+
+
+
(0) () (2 ) (2)
(3 ) (3)
0
E(x; )
as required. The function E(x; ) is, as Miller & Ross have remarked, strongly
reminiscent of a seldom-studied function52
E (z)
k=0
1
zk
(1 + k)
(89)
D 2 1x = 0
(90)
52
1
3
W1 (x; 2) = 12 (ex + ex )
W0 (x; 2) = DW1 (x; 2) = 12 (ex +
(91.1)
ex )
while if we take
2
3
2
3
2 )
we obtain
2
W2 (x; 3) = 13 (ex + ex + 2 e x )
2
W0 (x; 3) = D W2 (x; 3) =
1 x
3 (e
2 x
(91.2)
from which (see again (85)) W0 (x; 3) + W1 (x; 3) + W2 (x; 3) = ex follows now as
a consequence of the pretty cyclotomic condition
1 + + 2 = 0
(92)
53
Applications to analysis
when n is not prime), and provide an ecient starting point for many lines of
argument. For example, it follows immediately from (91) by (92) that
1 if p = 0
Wp (0; n) =
0 if p = 1, 2, . . . , n 1
Somewhat less trivially, we write
W0 (x; 2)
=
W1 (x; 2)
1
2
1
1
ex
ex
and obtain
W0 (x; 2)W0 (y; 2) + W1 (x; 2)W1 (y; 2) =
1
4
T
ex
2
0
ex
0
2
ex+y + e(x+y)
= W0 (x + y; 2)
1
2
ey
ey
(93)
T
ex
1 1
= 19 ex 1 2
2
1
e x
x
1
1 0 0
e
1 1
1
0 0 1 1 2 ex
2
0 1 0
2
1 2
e x
M
x T
x
e
3 0 0
e
= 19 ex 0 3 0 ex
2
2
0 0 3
e x
e x
2
= 13 ex+y + e(x+y) + e (x+y)
(94)
= W0 (x + y; 3)
54
profusion attach to the circular and hyperbolic functions are but the tip of an
unfamiliar iceburg. We stand, evidently, on the shore of an unexplored continent,
where the air is heavy with the scent of latent group theory.
Looking next to the illustrative (butfor obscure reasonsuniquely
tractable) fractional case = 12 , we are informed by Mathematica (who refuses
to sum the series, but can be tickled into reluctant cooperation) that
1
1
x 2 k1
1
(
k)
2
k=0
= ex + 1x 1 + 2x + 43 x2 +
E(x; 12 )
= ex +
1
x
8 3
15 x
16 4
105 x
32 5
945 x
64
6
10395 x
k
k!
(2k)! (4x)
k=0
1
x
+ e erfc( x )
x
(95)
This is consistent with a result remarked by Oldham & Spanier, who at p. 122
achieve (95) not by calculation but simply by observing, in the course of other
work, that53
1
D 2 1x = 0
1
D 2 ex erfc( x ) = 1x + ex erfc( x )
and that therefore
1
1
1
D 2 E(x; 12 ) = D 2 1x + D 2 ex erfc( x )
= 0 + 1x + ex erfc( x )
= E(x; 12 )
(93)
x 2 k2
1)
= ex erfc( x ) +
D2 E(x; 12 ) =
( 12 k
k=0
1
1
2x
1
1
2x
x 2 k3
2)
= ex erfc( x ) +
( 12 k
k=0
1
x
1
x
3
4x2
..
.
Dp E(x; 12 ) =
53
1
1
x 2 k(p+1)
1
(
k
p)
2
k=0
p = 0, 1, 2, . . .
The rst of the following equations reproduces (90), while the second is a
corollary of (80)
55
Applications to analysis
These results suggest strongly that something very like the integral theory
carries over into the non-integral domain. I have, however, had as yet no
opportunity to try to develop analytical methods strong enough to permit
systematic exploration of the details. A rst objective, in such an eort, would
be to develop a non-integral generalization of (91). As a rst step in that
direction I can report the following development:
I have remarked that Mathematica appears to have nothing useful to
say concerning (88) in the general case. Material tabulated on p. 1023 of
Abramowitz & Stegun led me to ask what Mathematica might have to say
concerning the Laplace transforms of the functions E(x; ) and their cognates;
I learn that
1
L[E(x; 1)] =
= L[W0 (x; 1)]
s 1
1
= L[W1 (x; 2)]
1
s
L[DE(x; 2)] = 2
= L[W0 (x; 2)]
s 1
L[E(x; 2)] =
s2
1
= L[W2 (x; 3)]
1
s
L[DE(x; 3)] = 3
= L[W1 (x; 3)]
s 1
s2
L[D2 E(x; 3)] = 3
= L[W0 (x; 3)]
s 1
L[E(x; 3)] =
s3
L[E(x; 12 )] =
1
s1
It is pretty to see the roots of unity emerge so naturally as locations of the poles
of a function so closely associated with E(x; ). Curiously (or perhaps not),
Mathematica is powerless to provide a description of L1 [1/(s 1)].
In classical analysis one makes fairly heavy use of shift rules of various
types, of which
(D + a)n = eax Dn eax
n = 0, 1, 2, . . .
56
provide the familiar simplest instance. The fractional analogs of such rules will
evidently be fairly intricatepartly for reasons developed above (replacement of
eax by E(ax; )), and partly because of the relative complexity of the fractional
generalization of Leibnizs formula.
Classical analysis is, of course, a subject of many parts, and provides a
wide and diverse range of opportunities to test the resources of the fractional
calculus. A systematic survey is out of the question; I conclude this discussion,
therefore, with a couple of semi-random remarks:
Analysis recommends to our attention a population of functions describable
by various instances of Rodrigues formula54
fn (x) =
1 d
n
w(x)[g(x)]n
w(x) dx
with L = 0, 12 , 1, 32 , 2, . . .
with m {L, (L 1), . . . , +(L 1), +L}
57
Conclusion
1
4a
e 4a eD 2 d
1
1
4a
e 4a eD 2 f (x) d
1
=
D 2 f (x) d
e
n!
4a
n even
n odd
1 + n
1 2
n
4a
=
d D 2 f (x)
e
4a n!
n even
+ 2m
1 2
=
e 4a d Dm f (x)
4a (2m)!
m=0
=
am m
D f (x)
m!
m=0
Evidently
2m e 4a d =
1
(2m)!
4a2m+1
m!
m = 0, 1, 2, . . .
A =
esA s1 ds
() 0
used to dene fractional powers of quite general operators; the little argument
just concluded is in that broad tradition.
Concluding remarks. The fractional calculus has a speculative history that
stretches back for more than years; the rst signicant applicationmarked
even then by an elegance which has been typicalwas published more than
years ago, and the modern foundations of the subject have beeen securely in
place for well more than years. Why, therefore, does the subject remain so
57
58
relatively little known? Why do modern authors feel an almost invariable need
to spell out the fundamentals before getting down to business, and why does
an apologetic air attach so often to their work? Several factors, it seems to me,
may be contributory:
There is, to my ear, something faintly o-putting about the very name of
the eld; fractional calculus sounds like a calculus that lives in the cracks
(like atonal music), a subject one can plausibly expect to get along well enough
without. The name broadcasts a signal seemingly calculated to engage the
passionate interest of disestablishmentarian iconoclasts, and to rub more polite
folk the wrong way. Nor is the term notable for its accuracy; when one writes
D one does not actually require that be rational, and one gives precise
meaning to the operator by appeal to the ordinary calculus. How dierent
might have been the recent history and present status of the subject had it been
called what it is: the theory of the Riemann-Liouville integral transform! Such
nomenclature is (rightly) evocative of nobel ancestory and workhorse utility.
And it suggests to anyone familiar with the rich interconnectedness present
within the theory of integral transforms that would in most contexts be
counterproductive to force the right foot into the left shoe, to attempt to force
the Riemann-Liouville transform to be what it (in general) isnta specialized
attachment to the theory of the Laplace transform.
Of course, people tend to have most lively interest in problems accessible
to the tools in their command, and to cultivate an interest in exotic tools only
when the urgency of otherwise inaccessible new problems enforces that interest.
Thus, in recent times, did tomography stimulate an interest in the pre-existing
theory of the Radon transform. It is my intuition that the on-going explosion of
interest in the fractal aspects of the world, in critical phenomena, in physics in
the near proximity of disasterthose things, but by no means only thosewill
soon spark a more general interest in the fractional calculus as a tool of choice,
and that some of the papers I have cited can be read as precursors of such a
development.
The syllabi of the Fractional Calculus 201 courses of the future will list
a number of topicssome quite basicI have had opportunity in these few
pages to mention only glancingly or not at all. Many of those are developed in
Oldham & Spanier and Miller & Ross, and in sources there cited; others appear
to await development. Theres opportunity aplenty within this eld for anyone
who feels there might reside a bit of Euler in his bones, or a bit of Kowalewski
in hers.
Miller & Ross, in 8 of their Chapter I: Historical Survey, remark that
Some of the still-open questions are intriguing. For example: Is it possible to
nd a geometric interpretation for a fractional derivative of noninteger order?
The question is ancientrecall Leibniz lament that the subject seems removed
from Geometry, which does not yet know of such fractional exponentsand it
is, in view of the diagram we traditionally draw when we explain what it means
to construct Df (x), quite natural. The geometrical meaning of D1 f (x)
is similarly direct. But when we look to Dn f (x) our geometrical intuition
59
Conclusion