See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/306357739

Uncertainty Principle for Measurable Sets and Signal Recovery in Quaternion

Domains

Article · August 2016

CITATIONS READS

5 85

3 authors, including:

Kit-Ian Kou
University of Macau
66 PUBLICATIONS   907 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Fourier transform and its applications View project

All content following this page was uploaded by Kit-Ian Kou on 10 October 2016.

The user has requested enhancement of the downloaded file.

 Mathematical
Methods in the
Research Article Applied Sciences

Received XXXX

(www.interscience.wiley.com) DOI: 10.1002/sim.0000

MOS subject classification: XXX; XXX

Uncertainty Principle for Measurable Sets and

Signal Recovery in Quaternion Domains†
Kit Ian Kou, Yan Yang∗ and Cui-Ming Zou

The classical uncertainty principle of harmonic analysis states that a nontrivial function and its Fourier transform cannot
both be sharply localized. It plays an important role in signal processing and physics. This paper generalizes the uncertainty
principle for measurable sets from complex domain to hypercomplex domain using quaternion algebras, associated with
arXiv:1608.05543v1 [cs.IT] 19 Aug 2016

the Quaternion Fourier transform. The performance is then evaluated in signal recovery problems where there is an interplay
of missing and time-limiting data. Copyright c 2009 John Wiley & Sons, Ltd.

Keywords: signal recovery; uncertainty principle; Quaternion Fourier transform.

1. Introduction
The classical uncertainty principle (the continuous-time uncertainty principle)
R ∞states that if a function f (t) is essentially zero outside
an interval of length ∆t and its Fourier transform fˆ(ω) (defined by ( fˆ(ω) = −∞ f (t)e−2πitω dt) is essentially zero outside an interval of
length ∆ω , then
∆t ∆ω ≥ 1. (1.1)
That means a function and its Fourier transform cannot both be higher concentrated. It was recently generalized from intervals to
measurable sets [11]. If f (t) is practically zero outside a measurable set T and fˆ(ω) is practically zero outside a measurable set W ,
then
|T ||W| ≥ 1 − δ, (1.2)
where |t| and |W| denote the measures of the sets T and W , and δ is a small number bound.
The quaternion Fourier transform (QFT) plays a vital role in the representation of (hypercomplex) signals. It transforms a real (or
quaternionic) 2D signal into a quaternion-valued frequency domain signal. The four components of the QFT separate four cases
of symmetry into real signals instead of only two as in the complex FT. In [6, 32] the authors used the QFT to proceed color image
analysis. The paper [5] implemented the QFT to design a color image digital watermarking scheme. The authors in [4] applied the QFT
to image pre-processing and neural computing techniques for speech recognition. Recently, certain asymptotic properties of the QFT
were analyzed and a straightforward generalization of the classical Bochner-Minlos theorem to the framework of quaternion analysis
was derived [13]. In this paper, we study the uncertainty principle of measurable sets (1.2) associated with QFT, the generalization of
the 2D Fourier transform (FT) in the Hamiltonian quaternion algebra. The main motivation of the present study is to develop further
iterative methods for signal recovery problems and to investigate the corresponding problems in quaternion analysis setting. Further
investigations and extensions of this topic will be reported in a forthcoming paper.
The article is organized as follows. Section 2 gives a brief introduction to some general definitions and basic properties of
quaternion analysis. The uncertainty principle for measurable sets is generalized for the right-sided Quaternion Fourier transform
of quaternion-valued signals in Section 3. In Section 4, applications to signal recovery problems were studied, which can be used
to recover a bandlimited hypercomplex signal with missing data. The proposed algorithm for hypercomplex signal recovery problem
are given. We test the performance of the proposed algorithm on two different size of Lena images and compare their performances.
Experimental results demonstrate the advantages of the proposed algorithm in Section 5. Some conclusions are drawn in Section 6.

Department of Mathematics, Faculty of Science and Technology, University of Macau, Taipa, Macao, China. Email:kikou@umac.mo and
zoucuiming2006@163.com
School of Mathematics and Computational Science, Sun Yat-Sen University, Guangzhou, 510631 Guangdong, P.R. China.
mathyy@sina.com
∗
Correspondence to: Yan Yang, School of Mathematics and Computational Science, Sun Yat-Sen University, Guangzhou, 510631 Guangdong, P.R. China.
mathyy@sina.com

Math. Meth. Appl. Sci. 2009, 00 1–10 c 2009 John Wiley & Sons, Ltd.
Copyright
Prepared using mmaauth.cls [Version: 2009/09/02 v2.00]
Mathematical
Methods in the
Applied Sciences K. I. Kou, Y. Yang, C. Zou

2. Preliminaries
The quaternion algebra H was first invented by W. R. Hamilton in 1843 for extending complex numbers to a 4D algebra [31]. A
quaternion q ∈ H can be written in this form

q = q0 + q = q0 + iq1 + jq2 + kq3 , qk ∈ R, k = 0, 1, 2, 3,

where i, j, k satisfy Hamilton’s multiplication rules

i2 = j2 = k2 = −1, ij = −ji = k,

jk = −kj = i, ki = −ik = j.
Using the Hamilton’s multiplication rules, the multiplication of two quaternions p = p0 + p and q = q0 + q can be expressed as

pq = p0 q0 + p · q + p0 q + q0 p + p × q,

where p · q = −(p1 q1 + p2 q2 + p3 q3 ) and p × q = i(p3 q2 − p2 q3 ) + j(p1 q3 − p3 q1 ) + k(p2 q1 − p1 q2 ).

We define the conjugation of q ∈ H by q = q0 − iq1 − jq2 − kq3 . Clearly, qq̄ = q20 + q21 + q22 + q23 . So the modulus of a quaternion q
defined by q
|q| = qq̄ = q20 + q21 + q22 + q23 .
p

In this paper, we study the quaternion-valued signal f : R2 → H that can be expressed as

f (x) = f0 (x) + i f1 (x) + j f2 (x) + k f3 (x),

where x = x1 i + x2 j ∈ R2 and fk , k = 0, 1, 2, 3 are real-valued functions.

For 1 ≤ p < ∞, the quaternion modules L p (R2 , H) are defined as
Z
L p = L p (R2 , H) := { f | f : R2 → H, k f kLp p := | f (x)| p dx < ∞}.
R2

Let f ∈ L1 (R2 , H), the (right-sided) Quaternion Fourier transform (QFT) of f is defined by
Z
1
F { f }(ξ) := f (x)e−ix1 ξ1 e−jx2 ξ2 dx (2.3)
2π R2

and if in addition, F { f } ∈ L1 L2 (R2 , H), function f can be recovered by its QFT as

T

Z
1
f (x) = F { f }(ξ)ejx2 ξ2 eix1 ξ1 dξ.
2π R2

The inner product of f (x), g(x) ∈ L2 (R2 , H) is defined by

Z
< f (x), g(x) >:= f (x)g(x)dx.
R2

Clearly, k f k2L2 =< f, f >. In this paper, we consider unit energy signal for simplification. That is, k f kL2 = 1. By Parseval’s identity,
Z Z
| f (x)|2 dx = |F { f }(ξ)|2 dξ,
R2 R2

we have kF { f }kL2 = 1 as well. It means that the QFT preserves the energy of the quaternion-valued signal.

3. Uncertainty Principles
The uncertainty principle of harmonic analysis states that a non-trivial function and its FT cannot both be sharply localized. The
uncertainty principle plays an important role in signal processing [20, 10, 24, 33, 26, 19, 29, 9, 18, 36, 34, 7, 21, 38], and physics
[25, 15, 27, 16, 17, 30, 2, 8, 37, 35]. In quantum mechanics an uncertainty principle asserts that one cannot be certain of the position
and of the velocity of an electron (or any particle) at the same time. That is, increasing the knowledge of the position decreases the
knowledge of the velocity or momentum of an electron. In quaternion analysis some lectures combined the uncertainty relations and
the QFT [3, 14, 28]. In this section we generalize the uncertainty principle for measurable sets associated with QFT. To proceed, we
first define the ε−concentrated on a measurable set in the space and frequency domains.

2 Copyright c 2009 John Wiley & Sons, Ltd. Math. Meth. Appl. Sci. 2009, 00 1–10
Prepared using mmaauth.cls
 Mathematical
Methods in the
K. I. Kou, Y. Yang, C. Zou Applied Sciences

Definition 3.1 Let f : R2 → H be ε−concentrated on a measurable set T ⊆ R2 , if there is a function g : R2 → H vanishing outside
T such that
k f − gkL2 < ε.

Similarly,

Definition 3.2 If f ∈ L1 (R2 , H), then its QFT F { f } is ε−concentrated on a measurable set W ⊆ R2 if there is a function h : R2 → H
vanishing outside W with
kF { f } − hkL2 < ε.

Now we state the main result.

 
Theorem 3.1 Let T and W be measurable sets on R2 and suppose there is a Quaternion Fourier transform pair f (x), F { f }(ξ) , with
f and F { f } of unit norm, such that f is εT −concentrated on T and F { f } is εW −concentrated on W . Then we have

|T ||W| ≥ [1 − (εT + εW )]2 .

Here |T | and |W| are the measures of the sets T and W .

From Theorem 3.1, we can immediately obtain the following corollary.

 
Corollary 3.1 Let T and W be measurable sets on R2 and suppose that there is a Quaternion Fourier transform pairs f (x), F { f }(ξ) ,
with f and F { f } of unit norm (energy), such that f and F { f } are compact supports on the measurable sets T and W , respectively.
Then we have
|T ||W| ≥ 1.

Theorem 3.1 and Corollary generalizes the results from the complex case [11] to the quaternion algebra. Before to proceed the
proof of Theorem 3.1, we introduce two crucial operators on f : R2 → H , namely the space-limiting operator

(S T f )(x) := χT (x) f (x),

where
(
f (x), x ∈ T,
χT (x) :=
0, x < T,

and the frequency-limiting operator Z

1
(FW f )(x) := F { f }(ω)ejx2 ω2 eix1 ω1 dω.
2π W

Clearly, we have F {FW f }(ω) = χW (ω)F { f }(ω).

For all f ∈ L2 (R2 , H), given the kernel k : R2 × R2 → H which satisfies the following two conditions: f (·)k(t, ·) ∈ L1 (R2 , H) for almost
every t ∈ R2 and if
Z
Q f (x) := f (t)k(t, x)dt,
R2

then Q f ∈ L2 (R2 , H).

Then we define the norm of Q to be
kQ f kL2
kQk := sup f ∈L2 = sup f ∈L2 kQ f kL2
k f k L2
for any unit energy f and the Hilbert-Schmidt norm of Q to be

Z Z ! 12
kQkHS := 2
|k(t, x)| dtdx .
R2 R2

Using Cauchy-Schwarz inequality, we can easily obtain the following lemma.

Lemma 3.1 kQkHS ≥ kQk.

To begin the proof of Theorem 3.1, we disgress briefly to make the following observation.

Lemma 3.2 kS T FW kHS = kFW S T kHS .

Math. Meth. Appl. Sci. 2009, 00 1–10 c 2009 John Wiley & Sons, Ltd.
Copyright 3
Prepared using mmaauth.cls
Mathematical
Methods in the
Applied Sciences K. I. Kou, Y. Yang, C. Zou

Proof 1 Using the definition of Quaternion Fourier transform (2.3), we have

Z
1
(FW S T f )(x) = F {S T f }(ω)ejx2 ω2 eix1 ω1 dω, x ∈ R2
2π W
Z Z !
1 −it1 ω1 −jt2 ω2
= f (t)e e dt ejx2 ω2 eix1 ω1 dω
(2π)2 W T
Z Z !
1 −it1 ω1 −jt2 ω2 jx2 ω2 ix1 ω1
= f (t) e e e e dω dt
(2π)2 T W
Z
= f (t)k(t, x)dt, (3.4)
T

where
Z
1
k(t, x) := e−it1 ω1 e−jt2 ω2 ejx2 ω2 eix1 ω1 dω, t ∈ T and x ∈ R2 . (3.5)
(2π)2 W

While

(S T FW f )(x) = χT (x)(FW f )(x)

Z
1
= χT (x) F { f }(ω)ejx2 ω2 eix1 ω1 dω
2π W
Z Z !
1 −it1 ω1 −jt2 ω2
= χT (x) f (t)e e dt ejx2 ω2 eix1 ω1 dω
(2π)2 W R2
Z Z !
1 −it1 ω1 −jt2 ω2 jx2 ω2 ix1 ω1
= χT (x) f (t) e e e e dω dt
(2π)2 R2 W
Z
= χT (x) f (t)k(t, x)dt. (3.6)
R2

From (3.4) and (3.6), we have

Z Z ! 12
kFW S T kHS = |k(t, x)|2 dtdx
R2 T
and
Z Z ! 21
kS T FW kHS = |k(t, x)| dtdx2
.
T R2

From (3.5), we know that k(t, x) = k(x, t). Therefore, we have |k(t, x)|2 = |k(x, t)|2 . This completes the proof.

Remark 3.1 From (3.4) and (3.6), we found that the product of two operators S T and FW are not commute. Fortunately, we can prove
that the HS-norms of these operators S T and FW are commute, although Quaternion algebra is a non-commutative algebra.

From another application of Eq. (3.4), we have the following lemma.

√
Lemma 3.3 kFW S T kHS = |T ||W|.

Proof 2 Applying (3.4), we have Z Z

kFW S T kHS = kQk2HS = |k(t, x)|2 dtdx.
R2 T

Let gt (x) := k(t, x), where k(t, x) is given by (3.5). Note that F {gt }(ω) = χW (ω)e−it1 ω1 e−jt2 ω2 . By Parseval’s identity, we have
Z Z Z
|gt (x)|2 dx = |F {gt }(ω)|2 dω = 1dω = |W|.
R2 R2 W

Therefore, we have kFW S T k2HS = |T ||W|. This completes the proof.

Now, we proceed the proof of Theorem 3.1.

Proof 3 (Proof of Theorem 3.1.) Consider the operator FW S T , by Parseval’s equality and the triangle inequality, applying the
assumptions of f is εT -concentrated on T and F { f } is εW -concentrated on W , we have

k f − FW S T f k = kF { f } − F {FW S T f }k
≤ kF { f } − F {FW f }k + kF {FW f } − F {FW S T f }k
≤ εW + kFW f − FW S T f k
≤ εW + kFW kk f − S T f k
≤ εW + εT . (3.7)

4 Copyright c 2009 John Wiley & Sons, Ltd. Math. Meth. Appl. Sci. 2009, 00 1–10
Prepared using mmaauth.cls
 Mathematical
Methods in the
K. I. Kou, Y. Yang, C. Zou Applied Sciences

The last step of equation (3.7) use the fact that kFW k = 1. For

k f k − kFW S T f k ≤ k f − FW S T f k ≤ εT + εW ,

we have
kFW S T f k ≥ 1 − εT − εW .
Therefore,
kFW S T k ≥ 1 − εT − εW .
Here k f k = 1 is used. By Lemma 3.1 and Lemma 3.3, we complete the proof.

4. Signal recovery problem

Donoho and Stark [11] studied some examples which applied the generalized uncertainty principle (1.2) to show something
unexpected is possible. The recovery of a signal despite significant amounts of missing information. One example is: A signal
f ∈ L2 (R2 , H) is transmitted to a receiver who knows that f is bandlimited, meaning that f was synthesized using only frequencies
in a set W ∈ R2 . Now suppose that the receiver is unable to observe all the data of f , a certain subset T of x-values is unobserved.
Moreover, the observed signal is contaminated by observational noise n(x) ∈ L2 (R2 , H). Thus the received signal r(x) satisfies

f (x) + n(x),
(
x < T,
r(x) = (4.8)
0, x ∈ T,

where supp F ( f ) ∈ W .
The receiver’s aim is to reconstruct the transmitted signal f from the noisy received signal r. Although it may seem that information
of f about x ∈ T is unavailable, the uncertainty principles says that the recovery is possible provided that |T ||W| < 1. Donoho and
Stark [11] proved this result in the one dimensional case. We may derive the analogue result to quaternion-valued signals.

Theorem 4.1 If W and T ∈ R2 satisfy the condition |T ||W| < 1, then f can be uniquely reconstructed from r. That is, there exists a
√
linear operator Q and a constant C with C ≤ (1 − |T ||W|)−1 such that

k f − Qrk ≤ Cknk

for all f, r, and the noise n obeying (4.8).

Proof 4

Step 1. We first prove that f (x) is the unique signal which can be recovered from the observed signal r. Suppose that f1 can be
recovered from r. Let h(x) := f (x) − f1 (x), we have h(x) = 0, for all x < T . While FW f (x) = f (x) and FW f1 (x) = f1 (x), so that
FW h(x) = h(x). That means h(x) is bandlimited in W . Then h(x) must be zero function on R2 , otherwise it would be contradiction
with Theorem 3 (since the condition |T ||W| < 1). Thus f (x) = f1 (x) is unique.
√
Step 2. Let Q = (I − S T FW )−1 . Form Lemma 3.3, we have kFW S T kHS = |T ||W|. Using Lemma 3.1, Lemma 3.2 and the condition
|T ||W| < 1, we have kS T FW k ≤ kS T FW kHS = kFW S T kHS < 1. Then Q exists because the well-known argument that the linear
operator I − L is invertible if kLk < 1. We also have

k(I − L)−1 k ≤ (1 − kLk)−1 . (4.9)

Since (I − S T ) f (x) = (I − S T FW ) f (x) for every bandlimited f (x) and

f (x) − Qr(x) = f (x) − Q(I − S T ) f (x) − Qn(x)

= f (x) − (I − S T FW )−1 (I − S T FW ) f (x) − Qn(x)
= 0 − Qn(x),

so
p
k f − Qrk = kQnk ≤ kQkknk ≤ (1 − |T ||W|)−1 knk,

Eq. (4.9) is used in the last step. This complete the proof.

The operator
∞
X
Q = (I − S T FW )−1 = (S T FW )k
k=0

suggests an algorithm for computing Qr.

Math. Meth. Appl. Sci. 2009, 00 1–10 c 2009 John Wiley & Sons, Ltd.
Copyright 5
Prepared using mmaauth.cls
Mathematical
Methods in the
Applied Sciences K. I. Kou, Y. Yang, C. Zou

Theorem 4.2 (Algorithm for signal recovery by uncertainty principle) Suppose that f ∈ L2 (R2 , H) and f is W -bandlimited, i.e.,
supp F { f } ⊆ W ∈ R2 . Given the received signal r satisfies (4.8) with the observational noise n ∈ L2 (R2 , H), then the information of f
about x ∈ T can be recovered by the following algorithm

s(0) = r
s (1)
= r + S T FW s(0)
s (2)
= r + S T FW s(1)
···

and so on, where S T FW are given in equation (3.6), provided that |T ||W| < 1. Then s(n) → f as n → ∞.

Example 4.1 We first consider a simulation example for one-dimensional W = [−Ω, Ω] bandlimited signal f (t) = f (t) without noise
n(t) and missing a part of information, which is needed to be recovery. We have

s(0) (t) = f (t)

(1)
s (t) = f (t) + s(0) ∗ sinc(t)
···
s(n) (t) = f (t) + s(n−1) ∗ sinc(t).

In fact, from (3.5), we have

1 sin Ω(t − x)
Z
1
k(t, x) = e−itω ejxω dω = = sinc(t − x)
2π W π (t − x)

Using (3.6), we obtain

Z Z
S T FW f (x) = f (t)k(t, x)dt = f (t)sinc(x − t)dt = f ∗ sinc(x), x ∈ T.
R R

Here, the number 1 in the inequality |T ||W| < 1 of Theorem 4.1 is corresponding to the normalized signals. In real world, most of
signals are not unit energy signals. In Fig. 1, we construct a simulation signal with the sustained domain [−20, 20]. The original signal
is generated by the inverse Fourier transform of a rectangular function with band [−10, 10] and the W = 10 in Fig. 1(a). For this signal,
we consider the inequality T W < 20 and W = 10. That is to say, the condition of the limit of T equals to 2, i.e., the information missing
in the time domain is no more bigger than [−2, 2] in Fig. 1(b). The signal f (t) is recovered by the proposed algorithm in Fig. 1(c). In
order to get this recovered signal, we iterate 10000 times and show the different from the recovered signal to original signal in Fig.
1(d). We can find that the information is filled in the missing parts of the signal and most of the information for the recovered signal is
still the same as the original signal. Hence, this method is effective on this example.

Example 4.2 Now, if the original signal f (x) ( x = x1 i + x2 j) is bandlimited in W = [−Ω, Ω] × [−Ω, Ω], and we would like to recover it
from observed data r, then by applying the proposed algorithm,

s(0) = r
s (1)
= r + s(0) ∗ sinc(t1 , t2 )
···
s(n) = r + s(n−1) ∗ sinc(t1 , t2 ),

where
sin(Ωt1 ) sin(Ωt2 )
sincΩ (t1 , t2 ) := .
ßt1 ßt2
In fact, from (3.5), we have
Z
1
k(t, x) = e−it1 ω1 e−jt2 ω2 ejx2 ω2 ejx1 ω1 dω
(2π)2 W
Z "Z #
1 −it1 ω1 −jt2 ω2 jx2 ω2
= e e e dω2 ejx1 ω1 dω1
(2π)2 [−Ω,Ω] [−Ω,Ω]
1 sin Ω(t1 − x1 ) sin Ω(t2 − x2 )
=
π2 (t1 − x1 ) (t2 − x2 )
sin Ω(t1 − x1 ) sin Ω(t2 − x2 )
=
π(t1 − x1 ) π(t2 − x2 )
= sincΩ (t1 − x1 , t2 − x2 ) = sincΩ (x1 − t1 , x2 − t2 ).

6 Copyright c 2009 John Wiley & Sons, Ltd. Math. Meth. Appl. Sci. 2009, 00 1–10
Prepared using mmaauth.cls
 Mathematical
Methods in the
K. I. Kou, Y. Yang, C. Zou Applied Sciences

Figure 1. 1D rectangular signal with bandwidth W = [−10, 10] in Example 4.1. (a) original signal (b) the signal with missing information (c) recovered signal (d) the difference
between the original signal and recovered signal.

Using (3.6), we obtain

Z
PT PW f (x) = f (t)k(t, x)dt
2
ZR
= f (t1 , t2 )sinc(x1 − t1 , x2 − t2 )dt1 dt2
R2
= f ∗ sincΩ (x1 , x2 ), (x1 , x2 ) ∈ T.

5. Experiments
In this section, two experiments are carried out to test the performance of the proposed algorithm as example. The reconstruction
results are shown in Fig. 2-3.
Two bandlimited images are constructed by Lena, which size is size 400 × 400. Two bandlimited Lena are constructed with
bandwidth W = 80 and W = 40, which are the original images and shown in Fig. 2(a) and Fig. 3(a). Here, the condition of |T ||W| < 1
in Theorem 4.1 becomes T W < 400, since the whole domain is 400 × 400. That means for Lena with bandwidth W = 80, T cannot
exceed 5, i.e., the information missing block in this image is less than 5 × 5. As for the Lena with bandwidth W = 40, this missing region
must be less than 10 × 10. In Fig. 2(b) and Fig. 3(b), the two missing information Lena are shown, where the two black rectangular
block are the regions we are generated with no information. In Fig. 2(c) and Fig. 3(c), the two recovered images are shown. And the
errors between the recovered images and the original Lena images are shown in Fig. 2(d) and Fig. 3(d). The error for the Lena with
bandwidth W = 80 is smaller than the Lena with bandwidth W = 40. From the two error images we can found that, except the center of
image, the difference for original images and recovered images is black. That is to say, for these regions with no information missing,
there is little changes. For the regions with information missing, there is some information be filled. This makes sense of the proposed
method. Apart from that, we found that the bigger size of the bandwidth, it makes image sharper. That makes the images in Fig. 2 is
clearer than that in Fig. 3.

Math. Meth. Appl. Sci. 2009, 00 1–10 c 2009 John Wiley & Sons, Ltd.
Copyright 7
Prepared using mmaauth.cls
Mathematical
Methods in the
Applied Sciences K. I. Kou, Y. Yang, C. Zou

Figure 2. Example of bandlimited Lena with W = 80. (a) original bandlimited Lena (b) the bandlimited Lena with missing information (c) recovered bandlimited Lena (d) the
difference between the original bandlimited Lena and recovered bandlimited Lena.

6. Conclusions
In this paper, we have proposed the Quaternion Fourier transform (QFT) for signal recovery problems. The mathematical definitions
of QFT for measurable sets are first presented. Then two crucial operators namely space-limiting and frequency-limiting operators are
discussed. Applying their properties, the uncertainty principle for measurable sets associated with QFT are given. Finally, the image
representation capabilities are discussed by experiments on real images. Experimental results have demonstrated that the proposed
algorithms have achieved promising results. As future works, we will apply the proposed QFT in a variety of applications, such as
color image retrieval and color image watermarking. The generalized integral transformations namely Quaternion fractional Fourier
transform and Quaternion linear canonical transform will also be considered in the upcoming paper.

Acknowledgment
This work is supported by the National Natural Science Foundation of China (11401606), University of Macau (MYRG2015-00058-
L2-FST and MYRG099(Y1-L2)-FST13-KKI) and the Macao Science and Technology Development Fund (FDCT/094/2011/A and
FDCT/099/2012/A3).

References
1. J. Flusser, T. Suk, and B. Zitová, Moments and moment invariants in pattern recognition. Wiley, 2009.
2. O. Aytur and H.M. Ozaktas, “Non-orthogonal domains in phase space of quantum optics and their relation to fractional Fourier transform,” Opt.
Commun., vol. 120, pp. 166-170, 1995.
3. M. Bahri, E. Hitzer, A. Hayashi and R. Ashino, “An uncertainty principle for quaternion Fourier transform,” , vol. 56, pp. 2398-2410, 2008.
4. E. Bayro-Corrochano, N. Trujillo, M. Naranjo, “Quaternion Fourier descriptors for preprocessing and recognition of spoken words using images of
spatiotemporal representations,” Journal of Mathematical Imaging and Vision, vol. 28, pp. 179-190, 2007.
5. P. Bas, N. Le Bihan, J. M. Chassery, “Color image watermarking using quaternion Fourier transform,” in Proceedings of the IEEE International
Conference on Acoustics Speech and Signal and Signal Processing, ICASSP, Hong-kong, pp. 521-524, 2003.

8 Copyright c 2009 John Wiley & Sons, Ltd. Math. Meth. Appl. Sci. 2009, 00 1–10
Prepared using mmaauth.cls
 Mathematical
Methods in the
K. I. Kou, Y. Yang, C. Zou Applied Sciences

Figure 3. Example of bandlimited Lena with W = 40. (a) original bandlimited Lena (b) the bandlimited Lena with missing information (c) recovered bandlimited Lena (d) the
difference between the original bandlimited Lena and recovered bandlimited Lena.

6. T. Bülow, “Hypercomplex spectral signal representations for the processing and analysis of images,” Ph.D. Thesis, Institut für Informatik und
Praktische Mathematik, University of Kiel, Germany, 1999.
7. L. Chen, K.I. Kou, M. Liu, “Pitt’s inequality and the uncertainty principle associated with the quaternion Fourier transform,” Journal of Mathematical
Analysis and Applications, vol. 423, pp. 681-700, 2015.
8. L. Cohen, “The uncertainty principles of windowed wave functions,” Opt. Commun., vol. 179, pp. 221-229, 2000.
9. Z.X. Da, “Modern signal processing,” Tsinghua University Press, Beijing, 2nd edn, 2002.
10. A. Dembo and T.M. Cover, “Information theoretic inequalities,” IEEE Trans. Inform. Theory, vol. 37, pp. 1501-1508, 1991.
11. D. L. Donoho and P. B. Stark, “Uncertainty principles and signal recovery,” SIAM Journal on Applied Mathematics, vol. 49, pp. 906-931, 1989.
12. T. A. Ell, “Quaternion-fourier transfotms for analysis of two-dimensional linear time-invariant partial differential systems,” in: Proceeding of the
32nd Conference on Decision and Control, San Antonio, Texas, pp. 1830-1841, 1993.
13. S. Georgiev, J. Morais, K.I. Kou and W. Sprößig, “Bochner-Minlos Theorem and Quaternion Fourier Transform. In Eckhard Hitzer and Steve
Sangwine, Quaternion and Clifford Fourier Transforms and Wavelets,” Springer, Birkhauser Trends in Mathematics Series, pp. 105-120, 2013.
14. E. Hitzer, “Directional uncertainty principle for quaternion Fourier transform,” Advances in Applied Clifford Algebras, vol. 20, pp. 271-284, 2010.
15. B.B. Iwo, “Entropic uncertainty relations in quantum mechanics,” Lecture Notes in Mathematics 1136,(Springer, Berlin), 1985.
16. B.B. Iwo, “Formulation of the uncertainty relations in terms of the Rényi entropies,” Phys. Rev. A, 2006.
17. B.B. Iwo, “Rényi entropy and the uncertainty relations,” Foundations of probability and physics, pp. 52-62, 2007.
18. P.J. Loughlin and L. Cohen, “The uncertainty principle: global, local, or both?,” IEEE Trans. Signal Porcess., vol. 52, pp. 1218-1227, 2004.
19. G. Hardy, J.E. Littlewood and G. Polya, “Inequalities,” Press of University of Cambridge, 2nd edn, 1951.
20. H. Heinig and M. Smith, “Extensions of the Heisenberg-Weyl inequality,” Int. J. Math. Math. Sci., vol. 9, pp. 185-192, 1986.
21. K. I. Kou, J. Ou and J. Morais, “On uncertainty principle for quaternionic linear canonical transform,” Abstract and Applied Analysis, vol. 2013,
2013.
22. K.I. Kou, R. H. Xu and Y. H. Zhang, “Paley-Wiener theorems and uncertainty principles for the windowned linear canonical transform,” To appear
in Mathematical Methods in the Applied Sciences.
23. H. Maassen, “A discrete entropic uncertainty relation,Quantum probability and applications,” Lecture Notes in Mathematics (Springer,
Berlin/Heidelberg), pp. 263-266, 1990.
24. V. Majernik, M. Eva and S. Shpyrko, “Uncertainty relations expressed by Shannon-like entropies,” CEJP, vol. 3, pp. 393-420, 2003.
25. H. Maassen, “A discrete entropic uncertainty relation,Quantum probability and applications,” Lecture Notes in Mathematics (Springer,
Berlin/Heidelberg), pp. 263-266, 1990.
26. D. Mustard, “Uncertainty principle invariant under fractional Fourier transform,” J. Austral. Math. Soc. Ser. B, vol. 33, pp. 180-191, 1991.
27. H. Maassen and J.B.M. Uffink, “Generalized entropic uncertainty relations,” Phys. Rev. Lett. , vol. 60, pp. 1103-1106 , 1988.

Math. Meth. Appl. Sci. 2009, 00 1–10 c 2009 John Wiley & Sons, Ltd.
Copyright 9
Prepared using mmaauth.cls
 Mathematical
Methods in the
Applied Sciences K. I. Kou, Y. Yang, C. Zou

28. Keith E. Nicewarner and A. C. Sanderson, “A General Representation for Orientational Uncertainty Using Random Unit Quaternions,” In Proc.
IEEE International Conference on Robotics and Automation, pp. 1161–1168, 1994.
29. H. M. Ozaktas and O. Aytur, “Fractional Fourier domains,” Signal Process., vol. 46, pp. 119-124, 1995.
30. A. Rényi, “On measures of information and entropy,” Proc Fouth Berkeley Symp. on Mathematics, Statistics and Probability, 1960.
31. A. Sudbery, “Quaternionic analysis,” Math. Proc. Cambridge Phil. Soc., vol. 85, pp. 199-225, 1979.
32. S. J. Sangwine, T. A. Ell, “Hypercomplex Fourier transforms of color images,” IEEE Transactions on Image Processing, vol. 16, pp. 22-35, 2007.
33. S. Shinde and M.G. Vikram, “An uncertainty principle for real signals in the fractional Fourier transform domain,” IEEE Trans. Signal Process.,
vol. 49, pp. 2545-2548, 2001.
34. A. Stern, “Sampling of compact signals in offset linear canonical transform domains,” Signal, Image Video Process., vol. 1, pp. 259-367, 2007.
35. A. Stern, “Uncertainty principles in linear canonical transform domains and some of their implications in optics,” J. Opt. Soc. AM. A, vol. 25, pp.
647-652, 2008.
36. G.L. Xu, X.T. Wang and X.G. Xu, “Three uncertainty relations for real signals associated with linear canonical transform,” IET Signal Process.,
vol. 3, pp. 85-92, 2009.
37. K. Wódkiewicz, “Operational approach to phase-space measurements in quantum mechanics,” Phys. Rev. Lett., vol. 52, pp. 1064-1067, 1984.
38. Y. Yang and K. I. Kou, “Uncertainty Principles for Hypercomplex Signals in the Linear Canonical Transform Domains,” Signal Processing, vol. 95,
pp. 67-75, 2014.

10 Copyright c 2009 John Wiley & Sons, Ltd. Math. Meth. Appl. Sci. 2009, 00 1–10
Prepared using mmaauth.cls

View publication stats