Derivatives of Noninteger Order

Author(s): Kenneth S. Miller

Source: Mathematics Magazine, Vol. 68, No. 3 (Jun., 1995), pp. 183-192
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/2691413
Accessed: 02-10-2017 16:51 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and

extend access to Mathematics Magazine

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 VOL. 68, NO. 3, JUNE 1995 183

Derivatives of Noninteger Order

KENNETH S. MILLER
25 Bonwit Road
Rye Brook, NY 10573

1. Introduction

The purpose of this paper is to introduce the reader to the subject of the fractional
calculus and fractional differential equations. We define integrals and derivatives of
noninteger order and study some of their elementary properties. In the final section
we examine fractional differential equations.
We shall indicate differentiation by the operator D. Thus

Dnf( t), n = 1, 2, .. .
r.epresents the nth derivative of f(t). The particular n-fold integral

ftdtldft2d dt3 f * l* f(() df (1.1)

of f(t) will be denoted by D-nf(t).
We shall give meaning to the symbol D"f(t) where v is not necessarily an integer.
Our results are true for any v that is not purely imaginary (and even this case may be
treated, r1]). For simplicity, and with little loss of generality, we shall assume v to be
real. The study of operators of the form D' constitutes the content of the subject
called fractional calculus (even though v need not be rational).
The higher transcendental function E,(v, a) introduced in Section 2 plays a
prominent role in the theory of fractional linear differential equations. We devote
Section 3 to an analysis of some of its elementary properties.
While the formula

Dn[Dmf(t)I =Dn+mf(t) = Dm[Dnf(t)I

is certainly true for suitable functions f in the case n and m are positive integers; it
may not be true if n and m are not positive integers. Certain aspects of this problem
are examined in Section 4.

2. Fractional Integrals and Derivatives

It is convenient to define first the fractional integral D-"f(t), v > 0, and then define
the fractional derivative DAf(t), A > 0, in terms of the fractional integral, Let I be
the closed interval [0, XI where X > 0, and let f be continuous on I. Then

D-lf(t) = If(() df (2.1)

exists for all t in I; see (1.1).

In our attempt to make a suitable definition o
the formula

D-nf(t) = t (t _ )n- lf() d (2.2)

(n Jo1)

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 184 MATHEMATICS MAGAZINE

for n = 1, 2,.... The proof is

assume that it is true for n = N. Then

D -(N+ I)f(t) D-1[ ] t(t-)-f) d(]

(N -(5 1)!1 x( ()(d

Interchanging the order of integration leads to

D-(N+l)f(t) = 'ft(t - ()Nf( ) d(, (2.3)

and our proof is complete.

We are prepared now to define the fractional integral. Equation (2.2) may be
written as

D-nf(t) = F(n (t)n- )lf(t() df, (2.4)

where we have replaced (n - 1)! by the equivalent representation in terms of the
Gamma function. Now the right-hand side of (2.4) is valid, not just for n a positive
integer, but for any positive number n. Thus if v > 0 and if f is continuous on
I' = (0, X] and integrable on every finite subinterval of I = [0, X], we call

D-vf(t) = ( ft(t - ,)V-'f(() d(, tEl' (2.5)

the fractional integral of f of order v.

Of course we could have defined the fractional integral of f o
right-hand side of (2.5). However, the analysis leading to (2.4) provides some
motivation for making this definition.
Let us consider an example. Suppose

f(t) = tA, A> - 1. (2.6)

Then (2.5) is essentially a Beta func

D-vtA= F(A + 1) A+v tel', A >-1, v>O. (2.7)

D F(A?+1?+v) t tE 27
Unfortunately, except for polynomials, the fractional integral of practica
other simple function we can think of leads to a nonelementary function. For
example, let us attempt to find the fractional integral of eat where a is a constant.
From (2.5)

1va - -
D_ ea = 1|(t e () la d(, v > O,

and if we make the trivial change of variable x = t -

Dveat= F eat txv- Ie-ax dx. (2.8)

We shall call the right-hand side of (2.8) Et(v, a) so that

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 VOL. 68, NO. 3, JUNE 1995 185

The function E,(v, a) is es

equations. In the next section we shall examine some properties of this important
function.
Let us turn now to the problem of defining the fractional derivative. We can not
use (2.5). That is, we may not write

Dgf (t) = t( t _- -'t f( (~) d(, > O

since the integral does not converge (unless f has compact support contained in
[0, t)). Therefore we resort to the following artifice. Let ,u > 0 and let m be the
smallest integer exceeding ,u. Then we define the fractional derivative of f of order
,u (if it exists) as

D'Af(t) = Dm[ D-(m-)f(t)]. (2.10)

For example, if f is given by (2.6), then from (2.7)

D-(m-A)t'k F (A ( + )_ A t+71-A
F(A + 1+m - .)

and thus

DA~tA - F(A + 1) D IltA?km+ m

F(A + 1 +m-/L)

= FA + 1) tA t E I'; A >-1 , > 0.

We may use (2.10) to verify that the fractional derivative of order ,u is the ordinary
derivative of order ,u when ,u is a positive integer. For suppose ,u = p (a positive
integer). Then m = p + 1, and if f has a pth derivative, (2.10) implies

DPf(t) = DP+'[ D-'f(t)];

see (2.1).

3. Some Special Functions

We already have used the Gamma function in the previous section. The reader recalls
that if v > 0, it has the integral representation
00

F(v) l v-le -x dx (3.1)

Now from (2.8) and (2.9) we may write E

Ej(v, a) = tleaty*(v, at) (3.2)

where

y* (v, at) - ] v lt V-le-6 dg v > O

a) F(v) (at)"

is the incomplete Gamma function [2, page 337]. A comparison with (3.1) demon-
strates the aptness of the name.

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 186 MATHEMATICS MAGAZINE

Let us look more close

oo k

y* (, t) =e k F(v?k?1) (33)
and is an entire function of both v and t. This form of y*(v
certain formulae involving Et(v, a). In particular some
Et(v, a) function we shall need are the recursion relation

Et(v, a) = aEt(v + 1, a) + + (3.4)

F(v + 1)

the differentiation formula

DPEt(v, a) = Et(v-p, a), p = 0,1, ... . (3.5)

and the special value

Et(O, a) = eat. (3.6)

To prove (3.4) we see from (3.3) that

t0 k 00 ti- 1
y*(v +t)=e-t E t e _______
y*(v +F1) e E (O +k + 2) 1F( +j + 1)

=e-tt1[J rF(;+j+l) - F(v+1) ]

=t-1[Y*(V,t) - F( +I) 1)

~ Et(v, a) =p - at______
and using (3.2) we get (3.4).
To prove (3.5) we see that

ap ap m) 0 a ktk k?m aktk?v-p

atp t(p F( k F(v+k?1)k- oF(v+k-p+l)

= Et(v -p, a),

which is (3.5). Trivially, see (3.3),

Et(O, a ) = e aty* (0, at ) = E0 at k - e at,

kOF(k + 1)

which is (3.6).

4. The Law of Exponents

If a and ,B are any real numbers, then for suitable functions we may compute
D a+f(t). We use (2.10) if a +,B > 0 and we use (2.5) if a +,B < 0. However, in
general, the law of exponents

DADP = DA+P (4.1)

is not necessarily valid without addit

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 VOL. 68, NO. 3, JUNE 1995 187

D'[ Dvf(t)] =D +vf(t)). For example, if f is continuous on I, then indeed

D[D-'f(t)] =f(t) (4.2)

(or symbolically DD -1 = D' -' =

Integral Calculus. But D -[ Df(t)] is not necessarily equal to f(t). For, let f(t) = 1.
Then D- lf(t) = t and D[ D- lf(t)] = 1; but D- 1[ Df(t)] = 0.
However, we shall show that

D-,[ D-`f(t) ] = D-(A+`)f(t) (4.3)

is true for all positive ,u and v, and continuous f. Fortuna
of the general law of exponents that we need for our pres
In an attempt to prove (4.3) we write

D -[ D`f(t)] = ( ) (t_) -[ D-vf( ) ] df

=-r(l ) tt(t - ;)A l d- X)V-1f( x) dff (4.4)

One is then tempted to interchange the order of integration as we did in (2.3) to

obtain

D A[D`f(t)] = ((t+v ) fo(t-xy)L+ lf(x) dx. (4.5)

We recognize the right-hand side of (4.5) as D-(A+v)f(t).

Of course the fly in the ointment is: How do we justify the interchange of order of
integration? If G({, x) is jointly continuous on I X I, then there is no problem. We
know that

ftd| fG((, x) dx = ftdx tG((, x) d(.

If G is not continuous, but if the integrals foGdx and f,XGGdf both exist as ordinary
or as improper integrals, then general conditions under which the order of integration
may be interchanged are difficult to obtain. And in our problem the integrand is not
continuous if A and/or v is less than one.
Fortunately, Dirichlet's formula may be used to justify the passage from (4.4) t
(4.5). A special case of this formula states that if g is continuous on I X I, and if
,u > 0, v > 0, then

ft(t - )' df| ({ - x) 1g({, x) dx

0 0

ftd ft(t - {)y'({ - x)v- g(, x) d{.

0 x

A proof may be found in [5, page 77

As an application of (4.3) we shall s

DPEt(v,a) =Et(v-p,a), v>O (4.6)

for all p, positive or negative. To do this we see fro

D1veat = Et(v, a), v > O. (4.7)

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 188 MATHEMATICS MAGAZINE

Thus (4.3) implies that for ,u > 0

D-AE,(v, a) = D-(A`V)eat. (4.8)

But from (4.7) we see that the right-hand side of (4.8) is Et
shown that

D-Et(v, a) = Et(l + v, a) (4.9)

Let us now calculate the fractional derivative of

DgEt(v, a) = D11[ D-("'-)Et(v, a)]

(where ,u > 0 and m is the smallest integer greater than ,u) is the fractional
derivative of E,(v, a) of order ,u. Since D-(7"-)Et(v, a) = Et(v + m - ,, a), by (4.9
we have

DgEt(v, a) = D ...[Et(v +m - .,ua)].

Equation (3.5) thus establishes the fundamental and useful result that

DgEt(v, a) = Et(v- ,a), > 0 (4.10)

is the fractional derivative of Et(v, a) of order ,

demonstrates the truth of (4.6).

5. Fractional Differential Equations

An ordinary differential equation is an equation involving derivatives of a function,

and the basic problem is to find a function that satisfies this equation. For example,

[D2+ aD + bD?]y(t) = 0 (5.1)

(where a and b are constants) is a second-order ordinary
with constant coefficients. The problem is to find nonid
that identically satisfy (5.1). Therefore it should come as
fractional differential equation as an equation involving fractional derivatives of a
function. In particular, if n and q are positive integers and v = 1/ q, then we call

9= D nv + ajD(`)v + ...+a Do (5.2)

a fractional differential operator of order (n, q). We shall assume that t
constants. Of course there exist more complicated fractional differential oper
(5.2) is more than sufficiently complex. Although we can solve the genera
&y(t) = 0, in this elementary treatment we shall focus our attention on equat
order (2, q), that is, on equations of the form

[D2v + aDv + bDO]y(t) = 0. (5.3)

Our problem, of course, is to find a function y(t) that

Green's function may be used to solve the nonhom
but we shall not discuss it here, see [3].)
In order to find a viable approach let us briefly revie
differential equation theory that may give us a hint a
end, we consider the ordinary differential equation
attempting to find a solution is to apply the different

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 VOL. 68, NO. 3, JUNE 1995 189

the exponential function ect

[D2+aD +bDO]ect= (c2+ac +b)ect.

Now the above expression will vanish if C2 + ac + b = 0. Thus we see that if c = a

where a is a zero of the indicial polynomial P,

P(A) = A2 + aA + b = (A-aa)(A-,f3), (5.4)

at~~~~~~~~~~~~~~~~~~~~~~~~~,
then ea' is a solution of (5.1). Of course, so is et, and if a # 1, we have two linearly
independent solutions. If a = 1, then a is a double root of P(A) = 0 and eat and teat
are linearly independent solutions of (5.1).
Let us now attempt to use the above arguments in attacking (5.3). We have seen in
(3.6) that

Et(O, c) = ect,

and using (4.6) we readily compute

DVEt(O, c) =Et((-v, c)
DvEt(-v, c) =Et((-2v, c)

DvEt - (q - 1)v, c)-=Et( -qv, c).

But qv = 1, and from (3.4)

Et(-1, c) =cEt(O, c) + J(o)

Since F(O) = oo we see that

Et( - 1, c) = cEt(O, c) .

The significance of these manipulations is that if we apply the operator Dv to

Et(O, c), Et(- v, c),...,Et(- (q - 1)v, c), we get a cyclic permutation of the same
functions. That is, no new functions are introduced.
Therefore we shall choose a linear combination of these functions as a candidate for
a solution of (5.3), say

y(t) = BoEt(O, c) + BjEt( -v, c) + B2Et( -2v, c) + ***+Bq_lEt( -(q - )v, c)

(5.5)

where Bo, B1, .. ., Bqi as well as c are arbitrary constants for the moment. From our
preceding arguments

DVy(t) = cBqilEt(O, c) + BoEt( -v, c) + BEt( -2v, c)

+ +Bq-2Et(-(q-1)v,c). (5.6)

Now if D2vy(t) has the same cyclic p

bD?]y(t). It will be a linear combinatio
whose coefficients are functions of th
B0, Bi, ..., Bq-, c such that the coefficien
we shall have a solution of (5.3).

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 190 MATHEMATICS MAGAZINE

Let us therefore start by calculating D2uy(t). Again from (4.6)

D2vy(t) = BoEt(-2v, c) + BIEt(-3v, c) + +Bq2 Et(-qv, c)

+ Bq_lE,( -(q+ 1)v, c).

Now, as before, E,( - qv, c) = E(- 1, 0) = cEt(O, c). Also

Et( -(q + 1)v, c) = Et( - 1-v, c). (5.7)

To simplify this expression we again invoke (3.4) with v = - 1 - v and a = c. Then

t-l-v

Et(-1-v,c) =cEt(-v,c) + F(-v)

and (5.7) becomes

t-l-v
Et (-(q + 1)v, c) = cEt( -v, c) + -v

So while we still have our cyclical property, that is, only terms of the form
Et(-kv, c), k = 0,1,...,q - 1, appear, we also have the unwanted term t-U' V/
R( -v). Well, let's worry about this term later. For the present we may write D2vy(t)
as

D2vy(t) = cBq2Et(O, c) + cBqEt(-v, c) + BoEt(-2v, c) +

t-1-v
+ Bq-3Et(-(q- 1)v, c) + Bq1 Fk -v) (5.8)
Now from (5.5), (5.6), and (5.8) we may compute [D2v + aDv + bD?]y(t). If we do
so, collecting coefficients of the Et( -kv, c) terms, we get

[ D2v + aDv + bD?]y(t) = [cBq_2 + acBq, + bBo] Et(0, c)

+ [cBqi1 +aBo+bBu]Et(-v,c)
q-3

? E [Bk + aBk+l + bBk+2]Et( - (k + 2)v, c)

k =o

t-l-v

* B -l(v (5.9)
Since a is a root of the indicial equation P(A) = 0 (see (5.4)),

a2+aa+b=O. (5.10)

Thus if we compare (5.10) wi

that if the Bk represent decr
us therefore choose Bk as

Bk =Aack

where A is an arbitrary nonzero factor independent of k. (For the moment we are

assuming that a # 0.) Then from (5.9)

Bk + aBk?, + bBk?2 =A(ac k + aa-k-1 + ba-k-2)

=Aa- k 2(a2 + aa + b)
= 0.

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 VOL. 68, NO. 3, JUNE 1995 191

We therefore have reduced (5

[ D2v + aDv + bD?]y(t)

+ A[ca-q+l + a + ba-']Et( -v, c)
t-l-v
+ Aa!-q+l
rF -v)

But the constant c is still at our disposal. If we let c = a', then the above
expression reduces to

t-l-v
[ D2v + aDv + bDO]y(t) = Aaq1(5
F( -v) (.1

Since A is arbitrary, we shall choose it as aq-1 so

side of (5.11) is independent of a.
With these choices of BO B1,..., Bq- 1, c, A we may write y(t) (see (5.5)) as
q-1

ea(t) = , aqk lEt(-kv, aq)

k=o

and (5.11) may be written as

t-l -v
[ D2v + aDv + bDO]ea(t) = F(-v)' (5.12)

where a is a zero of P(A). If a = 0 we choose ea(t) as

t-l+v
eo(t) = v

Of course ea(t) is not a solution of (5.3) since we still

on the right-hand side of (5.12). But we are getting cl
zeros. Let A =13 be the other zero. Then similar argum

t-l-v
[ D2v + aDv + b]e,(t) = _____

where of course

q-1

e,(t) = 8-k - lEt( -kv,pq).

k=o

Thus

q4(t) = ea()- ep(t) (5.13)

is the solution of (5.3).
If a #13 then (5.13) represents a nonidentically
a = f we have the trivial solution +f(t) 0. However, we recall the same phe-
nomenon in ordinary differential equation theory. If a = f, then teat is a solution of
(5.1).
Using a similar but more sophisticated argument one can show that if a = f8 # 0,
that is if the zeros of the indicial polynomial P are equal and nonzero, then

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms
 192 MATHEMATICS MAGAZINE

q-1

r( t ) ~E a k(q - Iki )D1-(k + 1)v (tea"t)

k =- (q-1)

is a nontrivial solution of (5.3).

If a = 0 = /, then (5.3) becomes

D2vy(t) = 0

whose solution is

)= Ft2v-1
2-
y(t)y( = (v

REFERENCES

1. E. R. Love, Fractional derivatives of imaginary order, J. London Math. Soc. (2) 3 (1971), 241-259.
2. W. Magnus, F. Oberhettinger and R. P. Soni, Formulas and Theorems for the Special Functions of
Mathematical Physics, Springer-Verlag New York, 1966.
3. K. S. Miller and B. Ross, Fractional Green's functions, Indian J. Pure Appl. Math. 22 (1991), 763-767.
4. K. B. Oldham and J. Spanier, The Fractional Calculus, Academic Press, New York, 1974.
5. E. T. Whittaker and G. N. Watson, A Course of Modern Analysis, Cambridge University Press,
Cambridge, 1965.

Proof without Words:

The Trapezoidal Rule (for Increasing Functions)
U

fl(x,,) 1Ys

f(x) -

x
aa m-O xI x,,_ x,,b

b ' b ba 1b-a
Jf(x) dx f(x) + [f(x,,) -f(x.)]
i=0

-JESUS URIAS
COMUNICACION OPTICA, UASLP
78000, SAN Luis POTOSI, SLP, ME'XICO

This content downloaded from 200.133.215.22 on Mon, 02 Oct 2017 16:51:24 UTC
All use subject to http://about.jstor.org/terms