Professional Documents
Culture Documents
The Fractional Fourier Transform and Time-Frequency Representations
The Fractional Fourier Transform and Time-Frequency Representations
The Fractional Fourier Transform and Time-Frequency Representations
Abstract-The fractional Fourier transform (FRFT), which is a The purpose of this paper is twofold: First, to briefly
generalization of the classical Fourier transform, was introduced introduce the fractional Fourier transform and its main
a number of years ago in the mathematics literature but appears properties and, second, to present the new results, including
to have remained largely unknown to the signal processing
community, to which it may, however, be potentially useful. The the FRFT’s interpretation as a rotation in the time-frequency
FRFT depends on a parameter cy and can be interpreted as a plane and the simple relationships that the FRFT has
rotation by an angle a in the time-frequency plane. An FRFT with several time-frequency re resentations that support
with a = n-/2 corresponds to the classical Fourier transform, and the interpretation as a rotation. We will also see that
an FRFT with Q = 0 corresponds to the identity operator. On the computing the FRFT of a signal corresponds to expressing
other hand, the angles of successively performed FRFT’s simply
add up, as do the angles of successive rotations. The FRFT of a it in terms of an orthonormal basis formed by chirps, i.e.,
signal can also be interpreted as a decomposition of the signal in complex exponentials with linearly varying instantaneous
terms of chirps. frequencies.
In this paper, we briefly introduce the FRFT and a number This paper is organized as follows. Section I1 presents
of its properties and then present some new results: the in- the FRFT, emphasizing its interpretation as a rotation oper-
terpretation as a rotation in the time-frequency plane, and the
FRFT’s relationships with time-frequency representations such ator, lists some of its properties, and gives some examples.
as the Wigner distribution, the ambiguity function, the short- Section I11 discusses the relationships with the Wigner dis-
time Fourier transform and the spectrogram. These relationships tribution, the ambiguity function, and other quadratic TFR’s.
have a very simple and natural form and support the FRFT’s Section IV discusses the relationships with the short-time
interpretation as a rotation operator. Examples of FRFT’s of Fourier transform and the spectrogram. Section V briefly
some simple signals are given. An example of the application of
the FRFT is also given. presents an example of an application. Section VI concludes
the paper.
Notes on the Formalism: We will represent by “j” the
I. INTRODUCTION imaginary unit and by a superscript asterisk * the complex
conjugation operation. We will often use square roots of
F OURIER analysis is one of the most frequently used
tools in signal processing and is frequently used in many
other scientific disciplines. Besides the Fourier transform itself,
complex numbers, and we will reserve the square root symbol
‘‘J” to denote the square root whose argument lies in the
time-frequency representations (TFR’s) of signals, such as the interval ] - 5 ~ / 2 , ~ / 2 ] .
Wigner distribution (WD) [ 11, [3], the ambiguity function [ 11,
[3] the short-time Fourier transform (STFT) [1]-[3] and the
11. THE TRANSFORM AND ITS PROPERTIES
spectrogram [ 11-[3] are often used, e.g., in speech processing,
radar, or quantum physics.
In the mathematics literature, a generalization of the Fourier A. Definition
transform known as the fractional Fourier transform (FRFT), In time-frequency representations, one normally uses a
was proposed some years ago [4], [SI. Although potentially plane with two orthogonal axes corresponding to time and
useful for signal processing applications, this transform ap- frequency, respectively (Fig. 1). If we consider a signal x ( t )
pears to have remained largely unknown to the signal process- represented along the time axis and its Fourier transform X ( w )
ing community. Recently, the FRFT has been independently represented along the frequency axis, we can view the Fourier
reinvented by a number of researchers, including this author transform operator,’ which we shall designate by F, as a
[6]-[9]. New results, concerning the FRFT’s interpretation as change in the representation of the signal corresponding to
a rotation in the time-frequency plane, and its relationships a counterclockwise axis rotation of ~ / rad.
2 This is consistent
with time-frequency transforms, have also been independently with the result of the repeated application of the F operator,
found, more or less simultaneously, by this author and by
some of the above referenced ones. ‘We consider the Fourier transform defined as S(UJ) & = JTm~ ( t )
x p - J d f d t . In the mathematics literature, the Fourier transform is normally
Manuscript received December 6 , 1992; revised January i8, 1994. The defined with a plus sign in the exponent. We prefer to use the engineering
associate editor coordinating the review of this paper and approving it for convention here, although this results in a number of sign differences between
publication was Prof. K. M. Buckley. the expressions of this paper and those of the references (which amounts to
The author is with INESCAST, Instituto Superior Tecnico, Lisbon, Portugal. a change in the sign of the angle ct in the expressions given ahead for the
IEEE Log Number 9403739. FRFT).
=i
el- cot a - p t c s r a
generalized function, its product by the chirp is also in L1, L2,
if cy is not a multiple of 7r or is a generalized function, respectively. Therefore, in these
(6) situations, the FRFT of z ( t ) exists in the same conditions in
6(t - U ) if N is a multiple of 271.
6(t + U ) if (Y+T is a multiple of%. which its Fourier transform exists.
It is easy to see that the FRFT, as defined in (13), satisfies
The square root factor that precedes the exponential, in this definition, the properties that were discussed in the introduction. Property
can be written in a number of equivalent forms, some of which appear in the
1 holds directly by definition. Property 2 is a consequence of
references. Note, however, that 141 contained a flaw in the definition of this
factor, a correction having been made in [SI. Another useful form for this the fact that for cy = 7r/2, K , coincides with the kernel of
factor is J 2 ~ ~ ~ " / ( 7 r s i i i c r ) . the Fourier transform. Property 3 is easily derived, using (1 1).
3086 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 42, NO. 11, NOVEMBER 1994
In fact, if we define
Y(U) = X,(u)
Signal Fractional Founer transform with angle a
II I
4 /(X(I8,dlr
seca e
-I-iana
if a -TI 1 2
2
IS
l X a ( z ) e ' ~ ' " a dz if a - x I 2 is not a multiple of x
z ( t )=
.II,X,(U)K_,(U,t)dU. (20)
From (20), we can see that the FRFT consists of expressing z ( t ) is even, X,(u) is also even; if z ( t ) is odd, X,(u) is
~ ( ton) a basis formed by the set of functions K-,(u,t) also odd.
(with u acting as a parameter for spanning the set of basis The scaling property, although somewhat complex, is im-
functions). This basis is orthonormal, according to (12). The portant since it shows the effect of a change of units (or a
basis functions are chirps, i.e., complex exponentials with scaling) of the independent variable t. In the classical Fourier
linear frequency modulation. For different values of U , they transform, the effect of such a change is only a corresponding
only differ by a time shift and by a phase factor that depends change of units (or scaling) of the frequency variable and a
on U : scaling of the amplitude. In the FRFT, the effects are a scaling
of the u variable by sin p / ( c sin a ) , a (complex) amplitude
K,(t,u) = e-+ano K,(t - useca.0). (21) scaling, a product by a chirp, and most important, a change
in the angle at which we compute the transform from (Y to
An interesting issue is whether this is the only transform that ,O = arctan(c2 t a n a ) . This angle change can be understood
satisfies properties 1-4. It is easy to see that if we would if we think of contracting the time axis, in the time-frequency
multiply the transform by e4jna with integer n, the four plane, by a factor c. As we know from the classical Fourier
properties would still hold. However, some of the results transform, we will then also have to expand the frequency axis
presented below would become more complex, and therefore, by c. With these two operations, the axis along which we were
the FRFT appears to be the most natural one among these. computing the transform, which was originally at an angle Q
We do not know if there is still any other transform, besides with the time axis, will move to a new position at the angle
these, obeying the four properties. However, most probably, if ,O given above.
such a transform exists, its relationships with time-frequency The Parseval relation
distributions will not be as simple as those derived below for W 00
the FRFT.
In summary, the FRFT is a linear transform, continuous
in the angle a, which satisfies the basic conditions for being
1, z(t)y*(t)dt =
.I,
is easily derived by expressing x ( t ) in the left-hand side of
X,(u)Y,*(u) du (22)
interpretable as a rotation in the time-frequency plane. this equation as the inverse transform of X , (U ) and then using
(9). A consequence of this equality is the energy-preserving
property of the FRFT
B. Further Properties
oc
L L
DCI
C. Examples
Table I1 gives the FRFT's of a number of common signals.
An expression for the FRFT of a rectangle can also be
obtained, but that expression is rather complex and does not and therefore
-
appear to be particularly instructive. Instead of presenting
it here, we prefer to give a feeling of the behavior of the
transform by showing the variation of the transform of a
X ( t . w ) = 2P2JWt
.
e-2'JtL
1"'1:
--x
..(T)X;t(-Z
'5lllOCOS~+2JZtSlll~
+ 2tcosn)
rectangle [ z ( t )= 1 for It1 < 2; otherwise, x ( t ) = 01 with the
angle a (Figs. 2 and 3). In Fig. 2, we can see some evidence x K, (T. z ) P J Y ' dr d z (28)
oc
of nonuniform convergence towards x ( t ) as cy i 0, with a
phenomenon similar to the so-called "Gibbs effect" of the
classical Fourier transform.
-
-
1,x ; ( - z +
2e2J.c't 2tcosa)
e - ~ ~s it nzn c o s n + 2 j z t s i r i n
111. RELATIONSHIPS
WITH THE WIGNER DISTRIBUTION,
THE
x s_, oc
z(T)e-2JwTKa(T. z ) d7 d z . (29)
AMBIGUITY FUNCTION,AND OTHER QUADRATICTFRS The inner integral is easily computed using property 2 of
The Wigner distribution (WD) [I], [3] is a very important Table I, and we obtain
time-frequency analysis tool, with applications in such diverse - E
Taking into account property 1 of Table I, we can express '11) = t cos cy + w sin (Y (32)
z*(2t - T ) as II = -tsin N +wcosa (33)
cc
z*(2t - T ) =
1,x;(-z+e-
2t c o s n )
2jtLsin n cos n + Z j z t sin n
K,(T. Z) (1%
we arrive at the following result after simplifying the
exponents:
(a) (b)
(e)
Fig. 2. FRFT of a rectangle, computed at various angles. Solid line: real part. Dashed line: imaginary part.
The right-hand side of this equation is the WD of X , X , is the WD of z, rotated by an angle -cy, or that it is
computed with arguments ( U , U ) . The left-hand side is the WD simply the WD of z expressed in the new set of coordinates
of II’ computed with arguments (t.U ) . The equation shows (U.U). The fact that the FRFT induces a simple rotation of
that the WD of X , coincides with the WD of s if we take the WD is remarkable and enhances our view of the FRFT as
into account the rotation that corresponds to the fact that we a rotation of the axis along which the signal is represented.
are using different axes on the left- and right-hand sides of On the other hand, this fact also shows that a rotated WD is
the equation. This is equivalent to saying that the WD of still a legal WD. It is well known that the WD has similar
ALMEIDA: FRACTIONAL FOURIER TRANSFORM 3089
5
L(T)lli*(t - T)e-JdTd7. (37)
Fig. 3 . Variation of magnitude of the FRFI’ of a rectangle with the angle 0.
From (37), the MSTFT can also be computed as
- 1 “Jt
X(t.w) = - e - j T X ( v ) W * ( w- v)eJutdv (38)
properties relative to time and frequency. This result shows G
simple rotation in the ambiguity function. The right-hand side of this equation is the MSTFT of X,
computed with window W , and with arguments (..TI). The
left-hand side is the MSTFT of 5 computed with window w
Iv. RELATIONSHIPS WITH THE SHORT-TIME and with arguments ( t . ~ )As . in the case of the WD, this
FOURIERTRANSFORM AND THE SPECTROGRAM equation shows that the MSTFT of X , is the same as the
The short-time Fourier transform (STFT) is another impor- MSTFT of z, again taking into account the rotation, i.e., that
tant time-frequency analysis tool. It is frequently used, e.g., it is simply a rotated version of the MSTFT of z or that it is
for speech processing. A related tool is the spectrogram, which the MSTFT of z expressed in the rotated axes ( U , w ) . Again, ~
corresponds to the squared magnitude of the STFT. The STFT this enhances our view of the FRFT as a rotation operator.
of a signal z ( t ) is defined as [l], [3] The spectrogram [l], [3] is simply the squared magnitude
of the STFT and, therefore, of the MSTFT as well. The
X ( t ,w ) = -
S”
-32
~ ( ~ ) w- * (7 t) e - J d Td.r ( 3 5 ) results we obtained on the MSTFT immediately lead us to
conclude that the effect of the fractional Fourier transform on
the spectrogram is identical to the one it has on the MSTFT:
where w ( t ) is a suitably chosen analysis win do^.^ The STFT
can also be computed in a simple way from the Fourier the spectrogram of X, computed with window W, is a rotated
version of the spectrogram of x computed with window IU.
transform of z ( t ) :
, rcc
X ( t , w ) = --&-jut X ( v ) W * ( w- v)eJutdv (36) V. APPLICATION
fi
In this section, we will briefly present, as an example, the
where X and W are the Fourier transforms of z and w , application of the FRFT to the study of swept-frequency filters.
respectively. This equation is similar to (35), except for the These filters are used, for example, in frequency analyzers for
presence of the exponential factor e - J w t . This is an asymmetry high-frequency signals. Swept-frequency filters are linear time-
between time and frequency that we wish to avoid since we varying systems that can be represented in the form shown in
want to deal with rotations in the time-frequency plane. We Fig. 4. They can also be represented by their time-varying
will therefore define a mod$ed short-time Fourier transform
‘When we compute the MSTFT of S,,we have to use the transformed
40ften, in the literature, the windows that are considered are real, and the window TI, as (39) shows. Note that even for a real-valued w, 147n will
complex conjugation of the window is not mentioned in the definition. Here, normally be complex. If we use the Gaussian window ~ ( t=) the
however, we prefer to adopt a more general viewpoint, allowing the window window’s FRFT is equal to to for any cr. Therefore, in this case, the
to be complex. FRFT induces a simple rotation of the MSTFT without any window change.
3090 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 42, NO. 1 I , NOVEMBER 1994
=.J 1-jcotcu
2T 1, O0 .(+j- cota
FRFT allows a treatment of swept-frequency filters that is very
similar to the classical treatment of shift-invariant filters with
the Fourier transform.
The work presented here opens several areas for further
or finally research. Among these is the study of discrete versions of this
transform. On the other hand, the relationships between the
FRFT and time-frequency transforms suggest that it can be a
where G is the (classical) Fourier transform of the impulse useful tool to further study the properties of these transforms
response g. Therefore, G(u csccr) can be called the transfer and to develop time-frequency transforms better suited to
function of the swept-frequency filter in the FRFT domain. specific applications.
The use of the FRFT and of this transfer function allows a
treatment of swept-frequency filters that is very similar to the
classical treatment of time-invariant filters with the Fourier APPENDIXA
transform. We shall prove the scaling property given in row 8
It is also worth mentioning that the FRFT is useful for of Table I. Let us consider two signals x and y, which
solving certain classes of differential equations that appear, are related by
for example, in quantum mechanics and optics, and that may
also be useful in signal processing applications (see [4], [SI,
[lo], [111). Y(t) =
(53)
(54)
ALMEIDA: FRACTIONAL FOURIER TRANSFORM 309 1