Академический Документы
Профессиональный Документы
Культура Документы
manipulate such matrices. The good agreement between the theoreti- A Class of Scaled Bessel Sampling Theorems
cally-predicted and the empirically-obtained separation performance
was demonstrated in simulation. Luc Knockaert
A remaining key-question is, given the sources’ covariance matrices
and our analytic expressions for the performance prediction, whether
(and if so, how) the association-matrices can be chosen so as to opti- Abstract—Sampling theorems for a class of scaled Bessel unitary trans-
mize the performance (in some sense). This question will be addressed forms are presented. The derivations are based on the properties of the
in future work. generalized Laguerre functions. This class of scaled Bessel unitary trans-
forms includes the classical sine and cosine transforms, but also novel chirp
sine and modified Hankel transforms. The results for the sine and cosine
transform can also be utilized to yield a sampling theorem, different from
REFERENCES Shannon’s, for the Fourier transform.
[1] A. Yeredor, “On optimal selection of correlation matrices for matrix- Index Terms—Bessel functions, chirp transform, Hankel transform, sam-
pencil-based separation,” in Proc. 8th Int. Conf. Independent Compo- pling theorems.
nent Anal. Source Separation (ICA), 2009, pp. 187–194.
[2] L. Tong, V. C. Soon, Y.-F. Huang, and R. Liu, “AMUSE: A new blind
identification algorithm,” in Proc. ISCAS, 1990, pp. 1784–1787. I. INTRODUCTION
[3] L. Tong and R. Liu, “Blind estimation of correlated source signals,” in
Proc. 24th Asilomar Conf., 1990, pp. 258–262. The classical Shannon [1] sampling theorem is omnipresent in large
[4] J. F. Cardoso, “Source separation using higher-order moments,” in areas of signal processing. Numerous generalizations have been de-
Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), rived [2]–[5], while extensions of the sampling theorem have arisen in
1989, pp. 2109–2112. connection with wavelets [6], the fractional Fourier and quasi-Fourier
[5] A. Yeredor, “Blind source separation via the second characteristic func- transforms [7]–[10]. In this correspondence, we develop sampling the-
tion,” Signal Process., vol. 80, no. 5, pp. 897–902, 2000. orems for a class of scaled Bessel unitary transforms in a constructive
[6] A. M. Tomé, “The generalized eigendecomposition approach to the
blind source separation problem,” Digit. Signal Process., vol. 16, no.
way, based on the generalized Laguerre functions. This class includes
3, pp. 288–302, 2006. the sine and cosine transforms, but also novel chirp sine and modified
[7] L. Parra and P. Sajda, “Blind source separation via generalized eigen- Hankel transforms. The results for the sine and cosine transform can
value decomposition,” J. Mach. Learn. Res., vol. 4, pp. 1261–1269, also be utilized to yield a sampling theorem, different from Shannon’s,
2003. for the Fourier transform. Novel asymptotic truncation error expres-
[8] E. Ollila, H. Oja, and V. Koivunen, “Complex-valued ICA based on a sions are derived and pertinent examples are presented.
pair of generalized covariance matrices,” Comput. Statist. Data Anal.,
vol. 52, pp. 3789–3805, 2008.
[9] J. F. Cardoso and A. Souloumiac, “Blind beamforming for non II. MAIN RESULTS
Gaussian signals,” Proc. Inst. Electr. Eng.—F, vol. 140, no. 6, pp.
1
362–370, 1993. We begin by stating the following result due to Kramer [2], [4], [5].
Theorem 1: Let be a real unitary operator over L2 [0; ] with
f g
[10] L. Fêty and J.-P. V. Uffelen, “New methods for signal separation,” in
Proc. 14th Conf. HF Radio Syst. Tech., 1988, pp. 226–230. product kernel K (xt) such that the set K (x
k ) forms a complete or-
[11] L. Parra and C. Spence, “Convolutive blind source separation of non- thogonal basis in L2 [0; 1] for some countable set of strictly increasing
stationary sources,” IEEE Trans. Speech Audio Process., pp. 320–327, f g
positive constants
k . Suppose further that f (t) is -bandlimited to
2000.
B , i.e., the transform
[12] A. Belouchrani, K. Abed-Meraim, J.-F. Cardoso, and E. Moulines, “A
blind source separation technique using second-order statistics,” IEEE 1
Trans. Signal Process., vol. 45, no. 2, pp. 434–444, 1997. fx ^( ) = K xt f t dt
( ) ( )
0
[13] D.-T. Pham and J.-F. Cardoso, “Blind separation of instantaneous mix-
is in L2 [0; B ] such that f^(x) for x > B . Then, f t
tures of nonstationary sources,” IEEE Trans. Signal Process., vol. 49,
no. 9, pp. 1837–1848, 2001. = 0 ( ) admits the
[14] A. Belouchrani and M. G. Amin, “Blind source separation based on sampling representation
time-frequency signal representations,” IEEE Trans. Signal Process.., 1
vol. 46, no. 11, pp. 2888–2897, 1998.
[15] K. Abed-Meraim, Y. Xiang, J. H. Manton, and Y. Hua, “Blind source
ft ( ) = f
Bk k Bt ( ) (1)
k=1
separation using second-order cyclostationary statistics,” IEEE Trans.
Signal Process., vol. 49, no. 4, pp. 694–701, 2001. where the sampling kernel is
[16] A. Yeredor, “On using exact joint diagonalization for non-iterative ap-
proximate joint diagonalization,” IEEE Signal Process. Lett., vol. 12, 1
no. 9, pp. 645–648, 2005. k v
( ) =
1
k K uv K u
k du
( ) ( )
Manuscript received February 15, 2011; revised June 14, 2011; accepted June
14, 2011. Date of publication June 27, 2011; date of current version September
14, 2011. The associate editor coordinating the review of this manuscript and
approving it for publication was Prof. Xiang-Gen Xia. This work was supported
by a grant of the Research Foundation-Flanders (FWO-Vlaanderen).
The author is with the INTEC-IBCN-IBBT, Ghent University, B-9050 Gent,
Belgium (e-mail: luc.knockaert@intec.ugent.be).
Digital Object Identifier 10.1109/TSP.2011.2160634
with
are orthogonal in L2 [0; 1]. Completeness is guaranteed by a theorem
20 2 of Szász [13, p. 282].
k = k ; k = (k ) J+1 (k ) ; k = 1; 2; . . .
2 The sampling kernel k (t) can be explicitly written as
satisfy the requirements of Theorem 1. t 0
J (t )
Proof: Consider the generalized Laguerre functions k (t) =
2(
)
k J0 (k )
2 (
t )2 0 1
: (4)
n (x) =
n!
0(n + 1 + )
x
=2 e0x=2 L (x);
n > 0 1
0!
It can be verified that limt
k (t) = 1, implying the characteristic
interpolation property k (
l ) = k;l . The sampling representation (1)
where Ln (x) are the Laguerre polynomials. It is shown in [11] that the can be written as f (t) = fN (t) + RN (t), where fN (t) is the truncated
generalized Laguerre functions n (x) form a complete orthonormal sampling expression and RN (t) is the truncation error
1
basis for L2 [0; ]. Next consider the operator with kernel 1
k
1 RN (t) = k (Bt):
0
f
n B
U (x; t) = ( 1) n (x)n (t): k=N +1
n=0
Clearly, the operator is real and unitary with eigenfunctions n (t) The asymptotic behavior of the truncation error for N is the sub- 0! 1
and eigenvalues f0 g
1; 1 . Moreover, U (x; t) is a product kernel since
ject of the two following corollaries.
Corollary 1: Let the -bandlimited function f (t) be O(t0 ) for
we have t0! 1 with > 1=2. Then, for fixed t0
p
0! 1
1
U (x; t) = J xt : ( ) (102 )=2 01=2
2 RN (t0 ) = O N for N
and
This follows from the well-established formula [12, p. 189] 1
1 n! n 0
2
RN (t) dt = O N
(102 )=
for N 0! 1:
Ln (x)Ln (y )z
n=0 0(n + 1 + )
p Proof: See Appendix I.
0z 0
= (1 0z x 0 zy xyz 0= I 0xyzz ;
)
1
exp
1
+
( )
2
2
1
The next corollary proves that there are numerous -bandlimited
functions which satisfy the requirements of Corollary 1. We have
Corollary 2: Let f^(x) be in L2 [0; B ] such that f^(x) = O(x ) for
valid for jz j < , and letting z tend to 0 from inside the unit circle.
1
From the complete orthonormal basis n (t), we can construct another
1
x 0! 0 with 0
1=2. Then
We are left with one last, but major, requirement : K; ( k x) should f g A. Sine Transform
form a complete orthogonal basis in L2 [0; 1]. After some calculations, Here, we take = 1=2 and = 1. The product kernel is
we find that
1 K1=2;1 (xt) =
pxtJ = 1 2 (xt) =
2
sin(xt)
K; (
k x)K; (
l x)dx
0
01=2
1 p p
and
k = k = k . The sampling kernel is
= (
k
l ) J ( t
k )J ( t
l )dt: (2)
0 k
0 k 2
2 0 sin(t ) 2
k (t) = :
For orthogonality, expression (2) must vanish when k = l. This is re- 6 t 1 + t=k
B. Cosine Transform
Here, we take = 01=2 and = 1. The product kernel is
p 2 cos(xt)
K01=2;1 (xt) = xtJ01=2 (xt) =
and
k = k = k 0 =2. The sampling kernel is
k (t) = sin(t 0
k ) 2 2 :
t 0
k 1 + t=
k
Note that the sampling kernels for the cosine and sine transforms are
the same, except for a translation 0=2 in the
k coefficients. Fig. 1. F (t) versus F (t) for the Fourier transform.
C. Fourier Transform
p
Although the Fourier transform exp(0ixt)= 2 admits the usual
O(t01 ) sampling theorem
1
f (t) = f (n1) sinc [B (t 0 n1)];
n=01
1 0ixt 1
p1 e f (t)dt = 2 cos(xt)fe (t)dt
2 01 0
1
0 i 2 sin(xt)fo (t)dt:
0
B
F (t) = eixt sin(ix + 1) dx
0B Fig. 3. G(t) versus G (t) for the Hankel transform.
= 1 +2 t2 <e ieiBt (cos(1 + iB) 0 t sin(1 + iB)) ;
with its truncated sine plus cosine sampling reconstruction FN (t) for We compared the -bandlimited function
B = 8 and N = 30 over the range 0 t 4. The result is shown in p p p p
Fig. 1. F (t) = 1 B
J0 ( xt) dx = 2 BJ1 ( Bt)= t
2 0
D. Hankel Transform with its truncated sampling representation FN (t) for B = 10 and N =
We will restrict ourselves to the modified Hankel transform of order 20 over the range 0 t 100. The result is shown in Fig. 2. With the
0, in which case = 0 and = 1=2 The product kernel is same sampling conditions as above, another example is
p p
K0;1=2 (xt) = 1 J0 ( xt) G(t) = 1
B
2 2 0 xp(B 0 x)J0 ( xt) dx p
and
k = k2 ; with fk g the positive zeroes of J0 (z). The sampling = 4B2 J2 ( Bt)=t 0 16B3=2J3 ( Bt)=t3=2 :
For this example, we have G(t) = O(t05=4 ) and hence we may ex-
kernel is given by
p
k (t) = 2 2 J0( t) : pect a better truncated sampling approximation as compared to the first
k J1 (k ) 1 0 t=
k example. The result is shown in Fig. 3.
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 10, OCTOBER 2011 5085
We do not have to worry about J0 (k ) or for that case J+1 (k ) since
these values are asymptotically located on the 1= t envelope of the 6 p
Bessel function. Hence
1
RN (t0 ) = O k(102 )=203=2
k=N +1
= O N (102)=201=2 :
For the squared truncation error, we utilize the orthogonality of the
functions k (Bt). We have
1 1 1
Fig. 4. F (t) versus F (t) for the chirp sine transform. RN (t)2 dt = f (
k =B )2 k (Bt)2 dt
0 k=N +1 0
1 f (
k =B )2 :
= Bk
E. Chirp Sine Transform k=N +1
Here, we take = 1=2 and = 2. The product kernel is From (3), we deduce that
f (
k =B )2 = O k(102)=01 ;
K1=2;2 (xt) = 2 2xt sin((xt)2 ) Bk
and
k
p
= k . The sampling kernel is
leading to
1
p RN (t)2 dt = O N (102 )= :
sin(t2
0 k ) 2 t= k
k (t) = 2
t 0 k
2 1 + t2 =k : 0
S S
where is the Fresnel integral (z ) = 0 sin t2 =2 dt, with its
z
for x 0! 0 with 1=2. Then 0
truncated sampling representation FN (t) for B = 3 and N = 20 over
the range 0 t 2: The result is shown in Fig. 4. f (t) = O(t0(+1)=2 ) for t 0
! 1:
IV. CONCLUSION
Proof: Without loss of generality, we can take B = 1. Consider
the integral [15, p. 684]
Sampling theorems for a class of scaled Bessel unitary transforms 1
have been presented. This class of scaled Bessel unitary transforms in- x J (xt) dx = t001 [( + 0 1)tJ (t)
cludes the classical sine and cosine transforms, but also novel chirp sine 0
and modified Hankel transforms. Novel asymptotic truncation error ex- +S01;01(t) 0 tJ01 (t)S; (t)+ C ]:
pressions were derived and pertinent examples presented.
Here, 0
0; > 1; C is a constant and S; (t) is the Lommel
function. Since S; (t) = O(t01 ) for t , we readily obtain that0! 1
APPENDIX I
1
x J (xt) dx = O(t001=2 ); 0 1
PROOF OF COROLLARY 1
Corollary 1: Let the -bandlimited function f (t) be O(t0 ) for 0
! 1 with > 1=2. Then for fixed t0
t0 = O(t03=2 ); 1: (5)
f (
k =B ) = O k0= = O k0= : Condition (6) is not difficult to satisfy, i.e., taking n = n for n n0
is just fine. Completeness implies that f^(x) can be written as
From formula (4) we deduce that 1
k (Bt0 ) = O
k1=20 =k J0 (k ) = O k 1=2 03=2
: f^(x) = an x
0
5086 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 59, NO. 10, OCTOBER 2011
REFERENCES
I. INTRODUCTION
[1] C. E. Shannon, “Communication in the presence of noise,” in Proc.
IRE, 1949, vol. 37, pp. 10–21. Given positive integers m and n with m n and a measurement
[2] H. P. Kramer, “A generalized sampling theorem,” J. Math. Phys., vol. A
matrix of size m 2 n, we consider the problem of compressive sam-
38, pp. 68–72, 1959.
x
pling in recovering a compressible signal 2 n from its noisy mea-
[3] A. J. Jerri, “Some applications for Kramer’s generalized sampling the-
orem,” J. Eng. Math., vol. 3, no. 2, pp. 103–105, 1969. surements = z Ax n
+ via solving the following `q -minimization
[4] A. J. Jerri, “The Shannon sampling theorem—its various exten- problem:
sions and applications: A tutorial review,” Proc. IEEE, vol. 11, pp.
1565–1596, 1977.
[5] A. I. Zayed, Advances in Shannon’s Sampling Theory. Boca Raton, min kykqq subject to kAy 0 zkp (I.1)
FL: CRC Press, 1993.
[6] I. W. Selesnick, “Interpolating multiwavelet bases and the sampling where 0 < q 1, q p 1, 0, and the measurement noise n
theorem,” IEEE Trans. Signal Process., vol. 47, no. 6, pp. 1615–1621, n
satisfies k kp [1]–[8]. Here, k 1 kq , 0 < q 1, stands for the
“`q -norm” on the Euclidean space.
1999.
[7] X. G. Xia, “On bandlimited signals with fractional Fourier transform,”
IEEE Signal Process. Lett., vol. 3, no. 3, pp. 72–74, 1996. x
Given a subset S f1; . . . ; ng and a vector 2 n , denote by S x
[8] T. Erseghe, P. Kraniauskas, and G. Cariolaro, “Unified fractional the vector whose components on S are the same as those of the vector
Fourier transform and sampling theorem,” IEEE Trans. Signal x x
and vanish on the complement S c . A vector 2 n is said to be s-
Process., vol. 47, no. 12, pp. 3419–3423, 1999.
x x
sparse if = S for some subset S f1; . . . ; ng with its cardinality
#S less than or equal to s, where s 1. Denote by 6s the set of
[9] Q. Wang and L. Wu, “A sampling theorem associated with
x
quasi-Fourier transform,” IEEE Trans. Signal Process., vol. 48,
no. 3, pp. 895–895, 2000. all s-sparse vectors. Given a vector , its best s-sparse approximation
[10] D. Y. Wei, Q. W. Ran, and Y. M. Li, “Generalized sampling expansion x
vector s in `q is an s-sparse vector which has minimal distance to x
for bandlimited signals associated with the fractional Fourier trans- x x x x y
in `q ; i.e., k 0 s kq = s;q ( ) := inf y26 k 0 kq . For q = 1,
form,” IEEE Signal Process. Lett., vol. 17, no. 6, pp. 595–598, 2010.
[11] G. Szego, Orthogonal Polynomials. New York: Amer. Math. Soc. x x
we use s ( ) instead of s;1 ( ) for brevity.
In this correspondence, we introduce a new property of a measure-
A
Coll., 1939.
[12] A. Erdélyi, W. Magnus, F. Oberhettinger, and F. G. Tricomi, Higher ment matrix : there exist positive constants D and such that
Transcendental Functions. New York: McGraw-Hill, 1953, vol. 2.
[13] P. J. Davis, Interpolation and Approximation. New York: Dover,
1975.
[14] M. Abramowitz and I. Stegun, Handbook of Mathematical Func- kxs krq DkAxkpq + sq=r01 (s;q (x))q for all x2 n (I.2)
tions. New York: Dover, 1965.
[15] I. S. Gradshteyn and I. Ryzhik, Table of Integrals, Series and Prod-
ucts. New York: Academic, 1965.
[16] V. Operstein, “Full Müntz theorem in l [0; 1];,” J. Approx. Theory, vol. Manuscript received June 06, 2010; revised October 26, 2010 and April 13,
85, pp. 233–235, 1996. 2011; accepted June 28, 2011. Date of publication July 12, 2011; date of current
version September 14, 2011. The associate editor coordinating the review of this
manuscript and approving it for publication was Prof. Namrata Vaswani. Part of
this work was done when the author visited Ecole Polytechnique Federale de
Lausanne.
The author is with the Department of Mathematics, University of Central
Florida, Orlando, FL 32816 USA (e-mail: qiyu.sun@ucf.edu).
Digital Object Identifier 10.1109/TSP.2011.2161470