J. Austral. Math. Soc. Ser.

B 40( 1998), 257-265

FRACTIONAL CONVOLUTION

DAVID MUSTARD1

(Received 23 January 1995; revised 23 April 1996)

Abstract

A continuous one-parameter set of binary operators on Z.2(K) called fractional convolution

operators and which includes those of multiplication and convolution as particular cases is
constructed by means of the Condon-Bargmann fractional Fourier transform. A fractional
convolution theorem generalizes the standard Fourier convolution theorems and a fractional
unit distribution generalizes the unit and delta distributions. Some explicit double-integral
formulas for the fractional convolution between two functions are given and the induced
operation between their corresponding Wigner distributions is found.

1. Introduction

The operation *, of convolution, on the space L2(M.) is the dual under the Fourier-
Plancherel operator & of the operation x, of multiplication [6]; that is, denning

&f(y) = f(y) = (2*)-* I e-i>xf{x)dx (1.1)

Jw.
and

f f(y-x)g(x)dx (1.2)

one gets for the convolution theorem the dual pair

fxg = f*'g and / *g = f x f . (1.3)

In the context of signal processing the multiplication of signal / by signal g corre-

sponds to a modulation of / by g whereas the convolution of / with g corresponds to
afilteringof / by thefilterwith spectral response g~[ 10]. It is of interest to see whether
these two operations can be extended to a one-parameter family of operations {*e} in
'School of Mathematics, University of New South Wales, Sydney, Australia 2052.
© Australian Mathematical Society, 1998, Serial-fee code 0334-2700/98
257

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 258 David Mustard [2]

which * and x appear just as particular cases. The operations *e between multiplica-
tion and convolution would then correspond to an influence between modulation and
filtering.
In this paper I construct such a set of "fractional convolution" operators {*e}eei
(where T = ^/2nZ), in which *0 = x and *7T/2 = *, by means of the Condon-
Bargmann fractional Fourier transform [1, 4, 8, 9] and investigate some of its formal
properties. The operator *e is commutative, associative and bilinear and obeys a
fractional convolution theorem that includes both results (1.3) as particular cases.
I find a "fractional unit" for *e that generalizes the units under x and * and a
deconvolution formula.
From the initial triple-integral construction of / *e g I get some other representa-
tions. Two are double-integral formulas and one shows / *s g as a product of certain
elementary operations and Fourier transform operations.
In another paper [9] I have shown the Radon-transform relationship between the
Condon-Bargmann fractional Fourier transform &9f and the Wigner distribution
[2,7, 11] Wf of / . The Wigner distributions Wfg and Wftg are equal to one-
dimensional convolutions of W{ and Wg in the directions of the two axes in the
Wigner plane. I define a 0—angled one-dimensional convolution, also denoted by *e,
between Wf and Wg, generalizing the two axial ones, and show that WfHg = Wf*g Wg,
which generalizes and extends the earlier results.

2. The fractional convolution operator *8

The integer powers of & form a cyclic group of order 4 of unitary operators on
L2([R) [6] in which the inner product and associated 2-norm are defined by

(/, g) = (2TT)- 1/2 / f(x)g(x) dx and

This finite discrete group can be imbedded in a continuous one-parameter group

of unitary operators, {^e)eej, the Condon-Bargmann group of fractional Fourier
transforms [1, 4, 8, 9], obeying

V 0,, 02 6 T J^, ^ = J^1+fl2 and V * e Z &k = &km. (2.1)

This one-dimensional fractional Fourier operator is defined by

^ef(y) = fe(y) = (2nyl/2 f K6{x, y)f{x)dx (9 e T), (2.2a)

where

/ ^ - y ) and Kn{x, y) = V2^ 8(x + y) (2.2b)

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 [3] Fractional convolution 259

(giving ^ o = ^ ° = ^, the identity operator, and &„ — &1 = 0P, the reversal
operator, defined by &f(x) = /(—x)) and where for 0 < \0\ < n

Ke{x, y) = Ag exp \ j ^ {- (*2 + y2) costf + 2xy]\ , (2.2c)

where

Ag = | sin0|-1/2exp | " - i ( | s g n 5 - 0 ) 1 . (2.2d)

If one rewrites (1.3) replacing * by "*^/2" and x by "* 0 " and using the notation of
(2.2a) they suggest a generalization to a theorem involving a fractional convolution
operator and a definition of *g provided that *e satisfies *n/2 = *-n/2 and *0 = *„.

DEFINITION 2.1. Let f,ge L\R) n L2(R) then the fractional convolution / *9 g is
defined by

/ *eg = (f-eg-e)g (9 € T). (2.3)

PROPOSITION 2.1.

V<p e T * 0 = * 0+;r . (2.4)

PROOF. From Definition 2.1 and (2.1)

/o *(p+n gO — {f-^+JT)8~(<t>+"))lf,+Jr — \^n (/-W+ff)g-(0+jr))j0 • (2-5)

But ^ = ^ , the reversal operator, for which &>(fg) = i&fK&g) - fngn, so

= =
tr-n \f-(<t>+n)g-(<t>+n)) /-(0+^)+n-^-(0+^)+^ f-<p8-<t>- (2-6)

Using (2.6) in (2.5):

/o *0+7r 50 = (/-0g_0) 0 i

that is, by (2.3),

/o *0+^ ^0 = /o *0 ^0,

which is just what the proposition claims.

COROLLARY.

*o = *n and *7!/2 = *-ni2\ (2.7)

so the definition of*e does satisfy the proviso made earlier.

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 260 David Mustard [4]

We are now ready for the fractional convolution theorem.

THEOREM 2.1. Let f,ge. L'(R) n L 2 (I) then

PROOF. From (2.3) and using (2.1)

(/o ** 8o)e = ([f-t8-*\)e = (f-*8-*)m (2-9)

and

fe *<p+e Se = (/fl-(0+e)ge-(0+6i))0+9 = (f-^g-^j^g',

that is, using (2.9), the required result, (2.8).

COROLLARY. Taking <f> = 0 and 9 = JT/2 (2.8) yields^ Jg = / * g~; f/?e« ta/fc/ng
0 = ^/2 a«d 0 = ^ / 2 a«<i wj/ng (2.7) (2.8) yields f * g = fg so both the standard
convolution results appear as particular cases of the fractional convolution theorem.

PROPOSITION 2.2. *e is associative; that is,

V0eT Vf,g,h f*e(g*eh) = (f*eg)*eh. (2.10)

PROOF. From the definition (2.3) and (2.1)

/ *e (g *eh) = f *e (g-8h-e)$ = (f-e (g-eh-e)0^)g = (f-gg-eh.e)g

from which the result is obvious.

REMARK. Generally for (0, - 92)/n <£ 1 f *9) (g *$2 h) ^ ( / *9l g) *9l h.

3. The fractional convolution unit le and deconvolution

DEFINITION 3.1. Define the distribution \e by

\6 = &e\ (Oel). (3.1)

The relations between the unit for x, 1, and the unit for * (as in (1.2)), V27r 8, are
now expressed lff/2 = V2TT S and V2n 5^/2 = 1 and so

(3.2)

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 [5] Fractional convolution 261

From (2.2) one gets immediately for 0 < \9\ < n

V2nSg(y) = \sine\-i/2exp\-j(^sgfie -6 - y2cot6U. (3.3)

Replacing 6 by n/2 + 0 in (3.3) and using (3.2) one gets for \0\ < n/2 the explicit
function

le(y) = | cos 0 r l / 2 e x p | -'- (-9 + / t a n 0)1 . (3.4)

PROPOSITION 3.1. lg is the unit under fractional convolution *g; that is,

V/eL'(l)flL2(K) h*ef = f- (3.5)

PROOF. From the definitions of *g and le in (2.3) and (3.1) and (2.1)

1# *e f = (\e-ef-e)e

Given f *$ g, where / is known, then to deconvolve it by a further fractional

convolution with some x, so as to recover g, means x must satisfy

Vg x*e(f*eg)=g. (3.6)

By Propositions 2.2 and 3.1 this means x must satisfy

x *e f = U\ (3.7)

that is, x is a convolutional inverse of / .

PROPOSITION 3.2. If(l/f-e)e exists then it is a convolutional inverse of f and solves

the deconvolution problem (3.6).

PROOF. Formally solve (3.7) for A: using (2.3), (3.1) and (2.1).

(If the given / is in L2(K) then l//_ 0 is not and so (l/f-o)e exists only in a
distributional sense.)

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 262 David Mustard [6]

4. Explicit formulas for f *g g

The calculation of a fractional convolution directly from Definition 2.1 involves

three integrations however a little manipulation yields double-integral formulas.
One gets

/ *„ g(t) = (2TT)-' / Ce{u, v; t)f(u)g(v) du dv, (4.1a)

where

Cg(u,v\t) = | sin #r'exp[/(u - t)(v - t)cot 0]l-9(u + v -t), (4.1b)

a double-integral formula symmetric in / and g.

Changing the variables in (4.1) by putting z — u + v — t and w = u — v leads to

,., f exp[/(^i) 2 cot 9]

f *e 8(0 = (2n)-]/2 J |sing|1/2 l-*(z)Q8(f,g;z+ Odz, (4.2a)

where

^ ,r N ~ \n f exp[—/ w2cot 6]
E
Q,(f, r. z> = o*r">J
a repeated integral, again symmetric in / and g.
To develop numerical algorithms to approximate / *e g it may be useful to represent
it as the result of elementary operations for which efficient algorithms are already
known.
Define the chirp and scaling groups of unitary operators {Ca}aeR and {5 a } asR .
(where Of = IR \ {0}) by Caf(t) = exp[iat2/2]f(t) and Saf(t) = \a\i/2f(at) then
the fractional Fourier transform ^e of (2.2) can be written

&B = <x6CcoieScsce^ Ccoie where ae = exp - / - {— sgn 0 - 6J .

Applying this representation of &g to the definition of / *g g yields

f *eg = A-gC<Me&~iC-cosgsiag& [(C_co,e/) * (C_ cote g)], (4.3)

where Ag is defined in (2.2d).

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 [7] Fractional convolution 263

5. Fractional convolution and the Wigner distribution

One member of the Cohen class {C/} of generalised phase-space distributions [3]
associated with a function / e L2((R) is the Wigner distribution [11] W : L2(R) ->•
Z.2(K2) where f \-+ Wf and

Wf{x) = {In)-"2 j e-wf (x, + | ) / (JC, - | ) dp. (5.1)

Many members of the Cohen class have marginal distributions along the two axes
given by

(2TT)- 1/2 / C/(x)rfx, = |f(x 2 )| 2 and {In)-"2 [ Cf(x)dx2 = |/(x,)| 2

JR JR

and this has the natural generalization that

VSeT (2;r)- 1/2 f Cf(x) dl = \fe(r)\2, (5.2)

Jl(r.6)

where dl is the element of Euclidean arc length along the line l(r, 9) whose equation
is X| cos 6 + x2 sin 9 = r. This generalization states that the Radon transform [5] of
Cf is the energy-density function \fe{r)\2 of the fractional Fourier transform regarded
as a function on K2 in polar coordinates r and 9. In another paper [9] I have shown
that this Radon-transform relationship (2) with the fractional Fourier transform holds
only for the Wigner distribution Wf.
One now naturally asks what is the operation between Wf and Ws that is induced
by fractional convolution under the map f v-* Wf. First define *' and *2 as the
one-dimensional convolutions in the Wigner plane with respect to the first and second
arguments.

DEFINITION 5.1.

Wf *' Wg(\) = (2TZT I / 2 / Wf{Xi - u, x2)Wg(u, x2) du (5.3a)

JR

and

Wf *2 Wg(x) = (2TZT1/2 f Wf(xl,x2-u)Wg(xuu)du. (5.3b)

It is easy to show the following relationships linking multiplication and convolution

between / and g to *' and *2 between Wf and Wg.

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 264 David Mustard [8]

PROPOSITION 5.1.

Wftg = Wf *' Wg and Wfg = Wf * 2 Wg. (5.4)

I now define "convolution in direction 9" on the Wigner plane and denote it also
by' V .

DEFINITION 5.2. Let

then define Wf *e Wg by

Wf *e Wg(x) = W-ej * 2 W_eg(Pex). (5.6)

One can show (for example, from the more general result of Proposition 4.28 in [7]
and using there srf = Pe from above) that

W0J = Wh. (5.7)

The generalization of the results of Proposition 5.1 and the answer to the question
raised earlier are contained in the following theorem.

THEOREM 5.1.

V0 € T, V/, g e L2(K) WfHg = Wf *B Wg. (5.8)

PROOF.

Wfteg(x) = W{f^g_eh(x) (by Definition 2.1)

= WUsg_e(Pex) (by (5.7) and (5.5))
= Wf_e * 2 Wg_e{Pex) (by (5.4))
= W-ej *2 W_e,g(Pex) (by (5.7))
= Wf *e Wg(x) (by Definition 5.2).

COROLLARY. Choosing 0 = 0 and 0 = n/2 one recovers as particular cases the

results of Proposition 5.1.

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509
 [9] Fractional convolution 265

Acknowledgements

I am grateful for study leave from the University of New South Wales and for the
generous hospitality of the Indian Statistical Institute (Bangalore and New Delhi) and
Murdoch University (Perth). I also thank the referee for several helpful suggestions.

References

[1J V. Bargmann, "On a Hilbert space of analytic functions and an associated integral transform. Part
\",Comm. Pure Appl. Math. 14(1961) 187-214.
[2] T. A. C. M. Claasen and W.F.G. Mecklenbrauker, "The Wigner distribution - a tool for time-
frequency signal analysis III", Philips J. Res. 35 (1980) 372-389.
[3] L. Cohen, "Generalized phase-space distribution functions", J. of Math. Phys. 7 (1966) 781-786.
[4] E. U. Condon, "Immersion of the Fourier transform in a continuous group of functional transfor-
mations", Proc. Nat. Acad. Sci.USA 23 (1937) 158-164.
[5] S. R. Deans, The Radon transform and some of its applications (John Wiley, New York, 1983).
[6] H. Dym and H.P. McKean, Fourier Series and Integrals (Academic Press, New York, 1972).
[7] G. B. Folland, "Harmonic analysis in phase space", in Annals of Mathematics Studies 122,
(Princeton Univ. Press, Princeton, 1989).
[8] D. Mustard, "The fractional Fourier transform and the Wigner distribution", J. Austral. Math. Soc.
Ser.B 38 (1996) 209-219.
[9] D. Mustard, "Lie group imbeddings of the Fourier transform and a new family of uncertainty
principles", Proceedings of the Centre for Mathematical Analysis 16 (1987) 211-222.
[10] A. Papoulis, Signal Analysis (McGraw-Hill, New York, 1977).
[11] E. Wigner, "On the quantum correction for thermodynamic equilibrium", Phys. Rev. 40 (1932)
749-759.

Downloaded from https://www.cambridge.org/core. IP address: 213.87.136.8, on 18 Dec 2019 at 06:37:40, subject to the Cambridge Core terms of
use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0334270000012509