i
THE MODERN APPROACH
1. INTRODUCTION
The reader who has followed the sometimes tortuous birth pangs of
the fractional calculus described in Chapter I is aware that more than
one version of the fractional integral exists. For convenience we shall
call
1 7 v-1
De "f(x) = To (#0) "f(¢) de (1.4)
the Riemann version and
if vol
-Dehs) = TEFL NF at (1.2)
the Liouville version. The case where c = 0 in (1.1), namely,
1 on 4)
Pe"fs) = The-N MO ae (13)
will be called the Riemann-Liouville fractional integral. Most of our
efforts in this book are centered on this version. We have also
222 THE MODERN APPROACH
introduced the Weyl fractional integral
—v. 1 7 eo di 1A
MoS) = Ty LOA at. oe)
This transform is studied more fully in Chapter VII. We have also
noted that the formula
n! net
D'flz) = a= [ (G-2) 9G) a (15)
in the complex domain (derived from the Cauchy integral formula)
bears a striking resemblance to our fractional integral definitions.
In (1.1) to (1.4) we assume that Re v > 0 and observe that the class
of functions to which f must belong is not necessarily the same in our
various versions. For example, if f(x) is a constant, then (1.1) and
(1.3) are meaningful, but (1.2) and (1.4) are not. (Recall our earlier
discussion in Section I-6 of functions of Riemann class and functions
of Liouville class.)
Our main objective in this chapter is to present various arguments
which should convince the reader that the definitions of the fractional
integrals that we have introduced are feasible entities. Equation (1.5)
also will be employed to arrive at this same conclusion.
So far we have mentioned only fractional integrals. Naturally, we
also shall define the fractional derivative of a function. More than one
such definition of the Riemann fractional derivative will be given. The
Weyl fractional derivative is defined in Chapter VII.
One trivial observation we may make even at this early stage is that
if D represents any of the operators in (1.1) to (1.4), then for
appropriate functions f and g, and any scalars @ and ,
D[af(t) + Bg(¢)] = aDf(t) + BDg(t).
That is, they are all linear operators. After we have defined the
fractional derivatives, it will be easy to see that they too are linear
operators.
However, before we embark on this program we should say a word
about notation. Various authors have used different notations. The
one we have adopted is due to H. T. Davis [8]. For example, as we
mentioned in Section I-4,
1 ey
Pets) = Toy LN Mae, Rev>0 (1.6)
(for x > c and suitable functions f) denotes integration to an arbi-
trary order » along the real axis from c¢ to x. Similarly, when weTHE ITERATED INTEGRAL APPROACH 23,
define the fractional derivative, we shall use the notation ,D? f(x),
Rev > 0, to denote differentiation to an arbitrary order v. Although
this notation is not free from ambiguity, it is sufficient for our
purposes.
The choice of a precise notation for the fractional calculus cannot
be minimized. For as we shall see, some of the power and elegance of
the fractional calculus rests in its simplified notation. The abridged
manner of representing these defining integrals may seem to be a
trivial matter; but the advantage of a simple notation has been the
source of many profound discoveries not obvious by other means.
2. THE ITERATED INTEGRAL APPROACH
The first argument that we shall give which leads to a definition of the
fractional integral begins with a consideration of the n-fold integral
De Wa) = [ae [def des LAO a. (2.1)
The function f in (2.1) will be assumed to be continuous on the
interval [c, b], where b > x. We assert that (2.1) may be reduced to a
single integral of the form
[RK Of at, (2.2)
where the kernel K,(x, ¢) is a function of n, x, and f. It will be shown
that K,(x,t) is a meaningful function even when n is not a positive
integer. Thus we shall define ,.D>"f(x) as
Dz'flx) = [K(x 0F(0 dt (2.3)
for all v with Rev > 0.
To prove this conjecture we start by recalling that if G(x, t) is
jointly continuous on [c,b] x [c,b], where b > x, then from the
elementary theory of functions we have
ff def Gx oat = [def G(x.) dx,. (24)24 ‘THE MODERN APPROACH
If, in particular,
G(x4,t) = f(4),
that is, if G(x,,1) is a function only of the variable t, then (2.4) may
be written as
[anf roam [rae fae,
= Le = t)f(t) de. (2.5)
Thus we have reduced the two-fold iterated integral to a single
integral.
If n = 3, then (2.1) becomes
Def(a) = [def def FO a
= fae fanf"roar].
If we apply the identity of (2.5) to the pair of integrals in brackets,
there results
Pra) = fae] fs - ontoarl,
Another application of (2.5) to the formula above leads to
Dex) = [1 def", — 9 ae,
- fro ( > Da.
Iterating this process n times reduces (2.1) to (2.2), where
(x=) 1
K,(x,t) = GapTHE DIFFERENTIAL EQUATION APPROACH — 25
Hence we may write .D;"f(x) as
7 1 * Fee
De"f (x) = Tal (x = 1)" 1 f(t) at. (2.6)
Clearly, the right-hand side of expression (2.6) is meaningful for any
number n whose real part is greater than zero. We shall call
1 * v-1
Fol e-9 f(t)dt, Rev>0
the fractional integral of f of order v and denote it by the symbol
Dz" f(x).
3. THE DIFFERENTIAL EQUATION APPROACH
We now show how the theory of linear differential equations may be
exploited to arrive at our fundamental definition
=) 1 * vol
0 (3.1)
of the fractional integral. Suppose, then, that
L=D"+p\x)D" +--+ +p,(x)D° (3.2)
is a linear differential operator whose coefficients p,(x) are continu-
ous on some interval J. Then if f(x) is continuous on J, and if c is
any point in I, we may consider the linear differential system
Ly(x) = f(x)
D'y(c)=0, OSksn-1. (33)
The unique solution of (3.3) for all x € J is given by
v(x) = [HO OE) ae, 4)26 THE MODERN APPROACH
where H is the one-sided Green’s function associated with L (see
[2sp.
If (¢,(x),...,,(2)} is any fundamental set of solutions of the
homogeneous equation Ly(x) = 0, the Green’s function H may be
written explicitly as
(x) bx) bal)
: $,(€) $2(€) os $(E)
HC y- Do(é) — Dbx{é) Db 6)
, W(é) | D?4(é) — D76x(E)-*- —— D*4,(E)
DagiO(G)) Dagse (2) a ee Damo .(2)
where
(8) 2(€) aa $,(E)
Do(é) Dé{é) +> Dba)
W(E)=| D’b(é) — D*b(€) ++ D6, (€)
D''6(é) D"-'b,(€) «++ D""16,(€)
is the Wronskian.
Now suppose that L is simply the nth-order derivative operator,
L=pD". (3.5)
Then (3.3) may be written as
D*y(x) = f(z)
D'y(c)=0, Osk 0. Again we are led to (3.1).
Those readers familiar with the elementary theory of the Laplace
transform will observe that if c = 0, the transform of (3.6) is
s"Y(s) = F(s),
where Y(s) and F(s) are the Laplace transforms of y(t) and f(t),
respectively. Thus
Y(s) =s""F(s),
and by the convolution theorem [7],
() = Fa Fa feo" eae.
Once again we are led to the definition (1.3) of the Riemann-Liouville
fractional integral. In later chapters we treat the Laplace transform in
a less cavalier fashion.
4, THE COMPLEX VARIABLE APPROACH
The Cauchy integral formula states that if f(z) is single-valued and
analytic in an open region A of the complex plane, and if A is an open
region interior to A bounded by a closed smooth curve C, then
1; £()
f2) = 75 D2
dg (4.1)
for any point z in A. From (4.1) it follows that
nt foun
DI) = Thay
(4.2)
(see, e.g., [23]).
In Section II-1 we remarked on the similarity between (4.2) and our
definitions of the fractional integral. We now attempt to convince the
reader that we can deduce (1.1) from (4.2).THE COMPLEX VARIABLE APPROACH 29.
Img
Branch cs
cu —=—_
___t Re
>t :
Figure La
If 7 is an arbitrary number, say v, we may replace n! by [(y + 1) in
(4.2). But if v is not an integer, the point z is now a branch point and
not a pole of the integrand of (4.2). The simple closed curve C is
therefore no longer an appropriate contour. To overcome this diffi-
culty we make a branch cut along the real axis from the point™z to
negative infinity. (See Fig. 1a, where we have assumed that z is a
positive real number, say x.) We are thus invited to define , DY f(x) as
the loop integral (x > c)
Te +1)
Qari
at eed Ty -vl
[Pea neds = EP pee peas.
2ri
(4.3)
Figure 1a may be redrawn as Fig. 1b—less colorful than Fig. 1a,
perhaps, but more convenient for computation. The loop -¥ is then
the union of L,, y, and L,, where y is a circle of radius r with center
at x and L, and L, are the line segment [c,x —r]. These line
segments coincide with a portion of the real axis in the ¢-plane but are
on different sheets of the Riemann surface for (¢ —x)~’~!. For
purposes of visualization we have drawn them as distinct.
If £ — x is a positive number, we define In(¢ — x) as a real number
(the arithmetical logarithm). Thus on y (see Fig. 16)
(fx) = ele Daina +18),30 THE MODERN APPROACH
Img
Figure 1b
Since @ = 7 on L,,
(f- eee = e(—Y— Dhln|s—x| +r)
= e(- PY MMinx 2) +m]
on L,, and since @ = —7 on L,,
(fh = ee Mn -im1
on Ly.
Returning to (4.3), we see that if Rev < 0,
[G9 ae= fff
= etre fe = 1) Gt) at
+f) ag
Y
eter f(x = UGA) dt,THE COMPLEX VARIABLE APPROACH 31
where ¢ = Re ¢. But
JEUNE) de = [ete F(x + rel) ire a)
and
|e ar nea] grterf” | f(x + re9)| a0.
Therefore, as r approaches zero, we have
fre =x) (Ode = [eet — eter] f(x <1) Y(t) dt
or
= + 1) ie
Dif (x) = (f — x) T(E) at
“reo + i v+1)7r
f (x=) 1F(t) de.
The reflection formula (B-2.8), p. 298, implies that
T(v + 1sin(v + 1) i
7 T(-»
Hence
1 7 -w-1
Difls) = Foy -9 f(t)dt, Rev <0. (44)
If c>0, we have Riemann’s version; if c= 0, we have the
Riemann-Liouville fractional integral; and if c= —*, we have
Liouville’s version.
“We now would like to make a brief digression and mention an
integral representation involving a more complicated contour than the
loop considered earlier in this section. Since we shall have no
occasion to return to complex variable theory in the remainder of this
book (see, however, Theorem 2 of Chapter IV, p. 90) this seems like
an appropriate juncture in our development to inject these results.
Hopefully, it will benefit those readers who wish to pursue the
*The remainder of this section may be omitted on a first reading.32 THE MODERN APPROACH
Figure 2
fractional calculus further from the complex variable point of view.
Since a detailed analysis would probably take us too far afield, we
shall content ourselves by merely stating certain important results.
Let @ and B be two branch points of F(z). Then a Pochhammer
contour # may be illustrated as in Fig. 2 (see [15, p. 255], (23, p. 163],
[48, p. 257). It is also sometimes called a double loop. A crucial
feature of this contour is that the multivalued function F(z) returns
to its original value after traversing 7.
The key formulas we wish to present are given below in Theorem 1.
Theorem 1. Let # be a simply connected open region in the complex
z-plane, containing the origin. Let @, be the region # with the origin
deleted. Then for z € #, v # —1, —2,..., and A not an integer,
een
D'z'f(z) = a [re-ar'oa (45)
and
D°z\(Inz)f(z) = TO De fae =z)" On (0 ae
4m sina Je
_Te +1)
Fanta LC t—z) 7’ f(t) dt (4.6)
4sin? 7A
where f(z) is analytic in @ and # is a Pochhammer contour with
respect to the branch points 0 and z.THE WEYL TRANSFORM — 33
__ This and other results, together with proofs, may be found in [15].
One of the noteworthy facts about the representations above is that
D” may be a fractional integral operator or a fractional differential
operator depending on whether Rev < 0 or Rev > 0 (provided that
v is not a negative integer.)
5. THE WEYL TRANSFORM
In previous sections of this chapter we have presented various argu-
ments that led to the definition
—w 1 7 vad
Dy f= FLe@-49 'f(t)dt, Rev>0 (5.1)
of the fractional integral. The Weyl fractional integral
We”; ir F(t) di Ri 0. 0 (5.2
7 mH THL ay 'fli)dt, — Rev>0, x>0 (52)
was introduced in Section I-6. We shall now attempt to convince the
reader that it is indeed a logical choice.
The arguments in Section II-3 that led to (5.1) started with a linear
differential operator
L=D" +p(x)D"! +++ +p,(x)D®. (5.3)
We now shall show that starting with the adjoint operator L*, we shall
be led to the definition (5.2) of the Weyl fractional integral.
The equation adjoint to Ly(x) = 0 is
L*y(x) = (-1)"D" + (-1)" 'D*""[p,(x)»(x)]
+o +(=1)"[p,(x)9(2)] = 0.
Then, of course, the solution of the nonhomogeneous equation
L*y(x) = f(x) with the initial conditions D*y(c) = 0,0