Professional Documents
Culture Documents
Convolution, Correlation, and Sampling Theorems For The Offset Linear Canonical Transform
Convolution, Correlation, and Sampling Theorems For The Offset Linear Canonical Transform
Convolution, Correlation, and Sampling Theorems For The Offset Linear Canonical Transform
DOI 10.1007/s11760-012-0342-0
ORIGINAL PAPER
123
SIViP
As a generalization of many other linear transforms, the A = (a, b, c, d, u 0 , ω0 ) is given by an OLCT with parame-
OLCT has found wide applications in optics and signal pro- ters A−1 = (d, −b, −c, a, bω0 −du 0 , cu 0 −aω0 ). The exact
cessing [1–3]. It is theoretically interesting and practically inverse OLCT expression is given by [1,2]:
useful to consider convolution and correlation theorems in −1
the OLCT domain. Convolution theorem for a linear integral f (t) = O LA [FA (u)](t)
transform can be formulated in several ways. Based on the ∞
expression for the generalized translation in the LCT domain, =C FA (u)h A−1 (u, t)du, (3)
the generalized convolution theorem has been derived in the −∞
LCT domain by Wei [16], which shows that the generalized where
convolution of two signals in time domain is equivalent to 1
C = e j 2 (cdu 0 −2adu 0 ω0 +abω0 ) .
2 2
simple multiplication of their LCTs in the LCT domains.
However, the generalized convolution structure in [16] is a This can be verified by using the definition (1). Some of
triple integral. So, it is complicated to reduce the expres- the special cases of the OLCT are listed in Table 1. Those
sion of the generalized convolution to a single integral form relations can be easily verified by substituting the specific
as in the ordinary convolution expression. In this paper, we parameters A in Eq. (2).
propose a new convolution structure for the OLCT, which is A signal f (t) is band-limited to UA in the OLCT domain,
different from the generalized convolution structure. It can be which means that:
expressed by a simple one-dimensional integral and easy to
implement in multiplicative filter design. In this work, based FA (u) = 0 for |u| > UA ,
on a new convolution operation for OLCT, we will discuss where UA is called the bandwidth of the signal f (t) in the
convolution and correlation theorems in the OLCT domain. OLCT domain with parameters A.
As an application, we also investigate the sampling theorem The OLCT has many properties [1–7,17]. We present the
and multiplicative filtering for the band-limited signal in the following important space shift and phase shift properties of
OLCT domain by using convolution theorems proposed here. OLCT [7,17], which can also be easily verified by using the
The rest of the paper is organized as follows. Section 2 definition (1).
gives a brief description of the OLCT along with some prop-
erties. In Sect. 3, based on a new convolution operation, Property 1: The space shift property [7,17]
we derive convolution and correlation theorems for OLCT
in detail. The sampling theorem for OLCT and discussions O LA [ f (t −τ )](u)
about multiplicative filter in the OLCT domain are presented acτ 2
in Sect. 4. Section 5 concludes the paper. = FA (u − aτ )e− j 2 +jcτ (u−u 0 )+jaτ ω0 . (4)
123
SIViP
Property 3: The space shift and phase shift properties [17] Definition 1 For any function f (t), h(t) ∈ W , let us define
A
a new convolution operation ⊗ for OLCT as follows [17]:
O LA [ f (t − τ )e jvt ](u)
acτ 2 +bdv 2 A
= FA (u − aτ − bv)e− j 2 z(t) = ( f ⊗ h)(t)
×e j (cτ +dv)(u−u 0 ) e− jbcτ v e j (aτ +bv)ω0 . ∞
aτ
= KA f (τ )h(t − τ )e− j b (t−τ ) dτ (7)
(6)
−∞
aτ + bv = 0 (9)
In the general framework of convolution theory (see [18,
Ch. 4]), it is known that to every integral transformation , According to Eq. (9), we get
one can, at least theoretically, associate with it a convolution
operation, , such that a
v=− τ (10)
b
( f g) = ( f ) · (g).
Substituting (10) into (6) and making use of ad − bc = 1,
Let be W that subspace of the space of all integrable func- Eq. (6) can be reduce to
tions with the property that f (t) ∈ W if and only if the FT
aτ aτ 2 τ
of f (t) is also in W [13,14]. Let us consider two functions, O LA f (t − τ )e− j b t (u) = FA (u)e− j 2b − j b (u−u 0 ) . (11)
f, h ∈ W , the conventional convolution operator for FT, is
given by [10] Then, using the definition of the OLCT, we have
∞
FA (u) · HA (u)
f (t)∗h(t) = f (τ )h(t − τ )dτ . ∞
j
K A e 2b [aτ +2τ (u 0−u)−2u(du 0 −bω0 )+du ] h(τ )dτ .
2 2
−∞ = FA (u)
And the convolution theorems for the FT are as follows: −∞
(12)
ψFT ( f (t)∗h(t)) = F(u)H (u),
where ψFT (·) denotes the FT operator and F(u), H (u) According to Eq. (11), we get
denotes the FT of f (t)andh(t), respectively. In this paper,
τ aτ 2 aτ
the FT is defined as follows [10] FA (u)e j b (u 0 −u) = e j 2b O LA f (t − τ )e− j b t (u). (13)
∞
F(u) = ψFT ( f (t)) = f (t)e− jut dt, Substituting (13) into (12), we can obtain
−∞
FA (u) · HA (u)
∞ ∞
1
f (t) = F(u)e jut du. =
j
K A e 2b [aτ −2u(du 0 −bω0 )+du ] h(τ )
2 2
2π
−∞
−∞
To obtain the convolution theorem for OLCT, we first define j aτ
2
− j aτ
a new convolution operation for OLCT. × e 2b O LA f (t − τ )e b t (u) dτ. (14)
123
SIViP
O LA [ f (t)h(t)](u)
Fig. 1 A new convolution operation for OLCT 1 j
e 2b [du −2u(du 0 −bω0 )]
2
=
2π
b
u
1
HA (u)e− j 2b [du
2 −2u(du
× 0 −bω0 )] ∗F . (19)
Using the definition of the OLCT, we rewrite (14) as b
FA (u) · HA (u) Proof Let y(t) = f (t)h(t), then
∞
j aτ 2
= K A2 e 2b [−2u(du 0 −bω0 )+du
2]
ej b h(τ ) YA (u) = O LA [y(t)](u)
−∞ ∞
⎛ j
K A e 2b [at +2t (u 0 −u)−2u(du 0−bω0 )+du ] f (t)h(t)dt.
2 2
∞ =
j
×⎝ e 2b [at
2 +2t (u
0 −u)−2u(du 0 −bω0 )+du
2]
−∞
−∞ (20)
aτ
× f (t − τ )e− j b t
dt dτ. (15)
According to the inverse OLCT expression (3), we get
According to (7), we get ∞
∞ h(t) = C HA (u)h A−1 (u, t)du
aτ
z(t) = K A f (τ )h(t − τ )e− j b (t−τ )
dτ −∞
∞
−∞ 1
HA (u)e− j 2b [du +2u(bω0−du 0−t)]
2
∞ = C K A−1
aτ
= KA h(τ ) f (t − τ )e− j b (t−τ )
dτ . (16) −∞
1
− j 2b [−2t[a(bω0 −du 0 )+b(cu 0 −aω0 )]+at 2 ]
−∞ ×e du
∞
1
HA (u)e− j 2b [du
2 +2u(bω
Substituting (16) into (15), we can obtain = C K A−1 0 −du 0 −t)]
j −∞
FA (u) · HA (u)e 2b [−du
2 +2u(du
0 −bω0 )]
1
− j 2b (2u 0 t+at 2 )
= O LA [z(t)](u) = Z A (u). (17) ×e du, (21)
123
SIViP
1 d 2 j 3.2 Correlation theorem for the OLCT
e− j 2b u 0 e 2b [du −2u(du 0 −bω0 )]
2
= KA
− j2π b
Correlation, which is similar to convolution, is another
∞
1 important operation in signal processing. Since the correla-
HA (v)e− j 2b [dv −2v(du 0 −bω0 )]
2
×
tion of two functions is no more than their convolution after
−∞ one of the two functions has been axis-reversed and complex
⎛ ⎞
∞ conjugated, the property of the new convolution results in the
×⎝ f (t)e− j b (u−v) dt ⎠ dv
t
property of the correlation. Making use of the new convolu-
−∞ tion structure (7) for OLCT, we can present a new correlation
1 j operation in the OLCT domain as following
e 2b [du −2u(du 0 −bω0 )]
2
=
2π
b
u
1
Definition 2 For any function f (t), h(t) ∈ W , let us define
HA (u)e− j 2b [du
2 −2u(du
× 0 −bω0 )] ∗F , A
b a new correlation operation ⊕ for OLCT as [17]
where F(·) denotes the FT of f (t). Thus, the proof of The- A
orem 2 is achieved. These results of the above-derived theo- z(t) = ( f ⊕ h)(t)
rems are extensions of convolution theorems of the FT and ∞
at
∗
LCT to OLCT domain. = KA KA f (τ )h ∗ (τ − t)e j b (τ −t) dτ , (22)
−∞
Corollary 1 If the parameter of OLCT changes to (a, b, c, d, d 2
− j 2b
u 0 , ω0 ) = (0, 1, −1, 0, 0, 0), then convolution formulas of where K A∗ = − j2π 1
be
u0
and superscript “∗” denotes
Theorems 1 and 2 reduce to conventional convolution for- complex conjugation, which is also used in the following
mulas in FT domain as follows: content of this paper. Further, based on the conventional con-
volution operator, we can express the correlation operator
FT
f (t) ∗ h(t) ←→ F(u)H (u), (22) in the OLCT domain as follow:
A
and z(t) = ( f ⊕ h)(t)
FT 1 ∞
f (t)h(t) ←→ F(u) ∗ H (u). = KA KA ∗ at
f (τ )h ∗ (τ − t)e j b (τ −t) dτ
2π
−∞
Replacing parameters A = (a, b, c, d, u 0 , ω0 ) with A =
(0, 1, −1, 0, 0, 0) in Eqs. (7), (8), and (19), respectively, the 2
∞
at aτ 2 a
K A K A∗ e− j 2b h ∗ (−(t −τ ))e−j 2b (t−τ ) dτ
2
proof of the Corollary 1 is achieved. = f (τ )e j 2b
−∞
Corollary 2 If the parameter of OLCT changes to (a, b, c, d, at 2 2
j at2b
2
− j at2b
= K A K A∗ e− j 2b f (t)e ∗
∗ h (−t)e (23)
u 0 , ω0 ) = (a, b, c, d, 0, 0), then convolution formulas of
Theorems 1 and 2 reduce to convolution formulas in LCT
domain [15] as follows: A
Theorem 3 Let z(t) = ( f ⊕ h)(t) and Z A (u), FA (u),
HA (u) denote the OLCT of z, f and h, respectively. Then
1 − j at 2 2
j at2b
2
j at2b
e 2b f (t)e ∗ h(t)e
j2π b O LA [z(t)](u) = Z A (u)
j
= K A FA (u) · HA∗ (u)e 2b [du
2 −2u(du
LC T du 2 0 −bω0 )] .
←→ F(a,b,c,d) (u) · H(a,b,c,d) (u)e− j 2b ,
(24)
and
u Proof Let us rewrite (23) in the form
LC T 1 j du 2 du 2
f (t)h(t) ←→ e 2b H(a,b,c,d) (u)e− j 2b ∗F ,
2πb b A
z(t) = ( f ⊕ h)(t)
where F(a,b,c,d) (u), H(a,b,c,d) (u) denote the LCT of f (t), 2
− j at2b
2
j at2b ∗ ∗
2 2
− j atb j at2b
h(t), respectively. The proof of the Corollary 2 is similar to = KAe f (t)e ∗ K A h (−t)e e .
the proof of the Corollary 1 and is omitted. (25)
In addition to the FT and LCT, convolution theorems for the And let g(t) be such that
specific OLCT cases FRST and FRFT shown in Table 1 can
at 2
also be obtained from the above-derived theorems [9,13,14]. g(t) = K A∗ h ∗ (−t)e− j b . (26)
123
SIViP
Substituting (26) into (25), we obtain domain is further investigated. In particular, the formulas of
A uniform sampling and low-pass reconstruction are obtained.
z(t) = ( f ⊕ h)(t) Firstly, we define the uniform sampled signal as follows:
at 2 at 2 at 2
= K A e− j 2b f (t)e j 2b ∗ g(t)e j 2b . ∞
fˆ(t) = f (t)sδ (t) = f (t) δ(t − nT )
According to Theorem 1, we get n=−∞
j ∞
Z A (u) = FA (u) · G A (u)e 2b [−du
2 +2u(du
0 −bω0 )] , (27) = f (nT )δ(t − nT )
where G A (u) denote the OLCT of g(t). Using the definition n=−∞
of the OLCT, we have
∞ where T is the sampling period and sδ (t) is the uniform
j
K A e 2b [at
2 +2t (u
0 −u)−2u(du 0 −bω0 )+du
2]
G A (u) = impulse train. Using the definition of the OLCT, we have
−∞
at 2 F̂A (u) = O LA [ fˆ(t)](u)
×K A∗ h ∗ (−t)e− j b dt
j
∞
= K A e− b [2u(du 0 −bω0 )−du ]
2 j
K A e 2b [at +2t (u 0 −u)−2u(du 0 −bω0 )+du ]
2 2
⎛ ∞ =
−∞
j
×⎝ K A e 2b [at +2t (u 0 −u)]
2
∞
−∞ × f (nT )δ(t − nT )dt
⎞∗ n=−∞
j
· e 2b [−2u(du 0 −bω0 )+du ] h(t)dt ⎠
2 j
= K A e 2b [du −2u(du 0 −bω0 )]
2
∞
j
f (nT )e 2b [a(nT ) +2nT (u 0 −u)] .
2
j × (29)
= K A e b [du
2 −2u(du
0 −bω0 )] HA∗ (u). (28) n=−∞
123
SIViP
the harmonic order n. When n = 0, we obtain So the output of the low-pass filter is as follows:
n=−∞
2UA
ω≥ . 1 T sin ωbc (t − nT )
b ×e− j b u 0 (t−nT ) . (32)
π(t − nT )
Thus, FA (u) can be recovered, and the other replicated spec-
trums are filtered out through a low-pass filter H̃A (u) having Equation (32) is the reconstruction formula of band-limited
the gain T and the cutoff frequency ωc in the OLCT domain, signal in the OLCT domain. Let ωc = UA in (31), the inter-
whose transfer function is shown by (31). The original signal polation formula (32) can be presented as follows:
f (t) can be reconstructed without any distortion by inverse
OLCT of the recovered FA (u). ∞
at 2 a(nT )2
f (t) = e− j 2b f (nT )e j 2b
123
SIViP
123
SIViP
that the major computation load of the former method is con- reconstruction are obtained. Moreover, the multiplicative fil-
volution, which can be done by the classical FFT, while the ter in the OLCT domain has also been discussed in this paper.
latter method needs to calculate OLCT twice, and there exists A practical method to achieve the multiplicative filter through
still no satisfactory fast digital algorithm for OLCT so far. In convolution in the time domain is proposed, which can be
the following paragraphs, we will analyze the computation realized by classical FFT and has the same capability, but
complexity of the multiplicative filter achieved in the time less computational complexity compared with the method
domain from (42) in detail. achieved in the OLCT domain.
According to the conventional convolution theorem,
Eq. (42) can be written as Acknowledgments The authors would like to thank the editor and the
anonymous reviewers for their valuable comments and suggestions that
improved the clarity and quality of this manuscript. This work was sup-
1 − j at 2 2
j at2b − j b1 u 0 t t
rout (t) = e 2b rin (t)e ∗ e h ported in part by the Fundamental Research Funds for the Central Uni-
b b versities, Southwest University for Nationalities (No.11NZYQN18),
1 at 2 and the National Natural Science Foundation of China (No.60672029).
= e− j 2b F −1 [F[ϕ1 ](u)F[ϕ2 ](u)] (t), (43)
b
References
2
j at2b − j b1 u 0 t
where ϕ1 (t) = rin (t)e , ϕ2 (t) = e h( bt ), F −1 (·)
1. Pei, S.C., Ding, J.J.: Eigenfunctions of the offset Fourier, frac-
denote the inverse FT operator. Since h(t) is the inverse FT tional Fourier, and linear canonical transforms. J. Opt. Soc. Am.
of H̃A (u), Eq. (43) can be expressed as A 20(3), 522–532 (2003)
2. Stern, A.: Sampling of compact signals in the offset linear canonical
1 at 2
domain. Signal Image Video Process. 1(4), 359–367 (2007)
rout (t) = e− j 2b F −1 F[ϕ1 ](u)b H̃A (bu + u 0 ) (t) 3. Pei, S.C., Ding, J.J.: Eigenfunctions of Fourier and fractional Fou-
b rier transforms with complex offsets and parameters. In: IEEE
at 2
= e− j 2b F −1 [ϕ3 ](t), (44) Trans. Circuits Syst. I 54(7), 1599–1611 (2007)
4. Abe, S., Sheridan, J.T.: Optical operations on wave functions as
the Abelian Subgroups of the special affine Fourier transforma-
where ϕ3 (u) = F[ϕ1 ](u) H̃A (bu + u 0 ). According to (44),
tion. Opt. Lett. 19(22), 1801–1803 (1994)
the FFT can be used to reduce the computation complexity 5. Moshinsky, M., Quesne, C.: Linear canonical transformations
of this filtering method. Since the computation complexity of and their unitary representations. J. Math. Phys. 12(8), 1772–
FFT is o(N log2N ), then we can draw the conclusion that the 1783 (1971)
6. Pei, S.C., Ding, J.J.: Relations between fractional operations and
computation complexity of the multiplicative filter achieved
time-frequency distributions,and their applications. In: IEEE Trans.
in the time domain can be reduced to o(N log2N ) for N point Signal Process. 49, 1638–1655 (2001)
of samples. 7. Sharma, K.K., Joshi, S.D.: Signal separation using linear canoni-
cal and fractional Fourier transforms. Opt. Commun. 265(2), 454–
460 (2006)
8. Almeida, L.B.: The fractional Fourier transform and time-
5 Conclusion frequency representations. In: IEEE Trans. Signal Pro-
cess. 42(11), 3084–3091 (1994)
9. James, D.F.V., Agarwal, G.S.: The generalized Fresnel trans-
In this paper, we first introduced the definition and some
form and its applications to optics. Opt. Commun. 126(5), 207–
properties of OLCT. Then, convolution theorems for OLCT, 212 (1996)
that is, Theorems 1 and 2 have been derived in detail, pre- 10. Ozaktas, H.M., Zalevsky, Z., Kutay, M.A.: The Fractional Fou-
senting the OLCT domain’s behavior of a new convolution as rier Transform with Applications in Optics and Signal Process-
ing. Wiley, New York (2000)
well as a product between two functions in the time domain.
11. Akay, O., Boudreaux, B.G.F.: Fractional convolution and correla-
We also obtain the expressions for OLCT of a new correla- tion via operator methods and application to detection of linear FM
tion operation of two functions based on Theorem 1. Since signals. In: IEEE Trans. Signal Process. 49, 979–993 (2001)
the correlation of two functions is no more than their convo- 12. Torres, R., Pellat, F.P., Torres, Y.: Fractional convolution, fractional
correlation and their translation invariance properties. Signal Pro-
lution after one of the two functions has been axis-reversed cess. 90, 1976–1984 (2010)
and complex-conjugated, the property of the new convolu- 13. Almeida, L.B.: Product and convolution theorems for the frac-
tion structure for OLCT results in the property of the correla- tional Fourier transform. In: IEEE Trans. Signal Proc. Lett. 4, 15–
tion for OLCT. All these results developed for OLCT can be 17 (1997)
14. Zayed, A.I.: A product and convoution theorems for the frac-
applied to a large class of signals and system outputs and can tional Fourier transform. In: IEEE Trans. Signal Proc. Lett. 5, 101–
be regarded as unified and extended versions of previously 103 (1998)
developed convolution and correlation theorems for special 15. Deng, B., Tao, R., Wang, Y.: Convolution theorems for the lin-
cases of the OLCT. Finally, as an application, utilizing con- ear canonical transform and their applications. Sci. China (Ser.E
Information Science) 49(5), 592–603 (2006)
volution theorems for OLCT derived in this paper, sampling 16. Wei, D.Y., Ran, Q.W., Li, Y.M., Ma, J., Tan, L.Y.: A convolution
of band-limited signals in the OLCT domain has been further and product theorem for the linear canonical transform. In: IEEE
investigated. The formulas of uniform sampling and low-pass Signal Process. Lett. 16(10), 853–856 (2009)
123
SIViP
17. Wei, D.Y., Ran, Q.W., Li,Y.M.: A convolution and correlation theo- 20. Erseghe, T., Kraniauskas, P., Cariolaro, G.: Unified fractional Fou-
rem for the linear canonical transform and its application. Circuits rier transform and sampling theorem. In: IEEE Trans. Signal Pro-
Syst. Signal Process. (2011). doi:10.1007/s00034-011-9319-4 cess. 47(12), 3419–3423 (1999)
18. Zayed, A.I.: Function and Generalized Function Transforma- 21. Li, B.Z., Tao, R., Wang, Y.: New sampling formulae related to
tions. CRC, Boca Raton, FL (1996) linear canonical transform. Signal Process. 87(5), 983–990 (2007)
19. Candan, C., Ozaktas, H.M.: Sampling and series expansion the- 22. Stern, A.: Sampling of linear canonical transformed signals. Signal
orems for fractional Fourier and other transform. Signal Pro- Process. 86(7), 1421–1425 (2006)
cess. 83(11), 2455–2457 (2003)
123