Bilinear Transform PDF
Bilinear Transform PDF
m
(t) =
2
1/4
2
m
m!
H
m
_
2t
_
e
t
2
, (1)
where H
m
(t) is the mth-order Hermite polynomial. Hermite-
Gaussian functions are eigenfunctions of the Fourier trans-
form
F {
m
}(t) = e
m/2
m
(t), (2)
where F is the Fourier transformation operator and
e
m/2
is the mth-order eigenvalue. An mth-order Hermite-
Gaussian function has m zero-crossings. The Hermite-
Gaussian functions are homogeneous solutions of the dif-
ferential equation, which is also known as the Hermite-
Gaussian generating dierential equation
d
2
f (t)
dt
2
4
2
t
2
f (t) = f (t),
(3)
with = 2m+1. The Hermite-Gaussian generating function
can be expressed by its operator equivalent as
_
D
2
+ F D
2
F
1
_
f (t) = f (t), (4)
where D
2
denotes the second derivative operator.
2.2. Commuting Matrix Generation. Let A and B be N N
square matrices. If AB = BA, then A and B are commuting
matrices. If A and B commute, they share at least one set of
common eigenvector sets [6].
Candan [8] showed that a DFT commuting matrix K can
be obtained for any arbitrary N N matrix L as
K = L + FLF
1
+ F
2
LF
2
+ F
3
LF
3
, (5)
where F is the N point DFT matrix which is dened as
(F)
n,m
=
1
N
exp
_
2
N
nm
_
, n, m = 0, 1, . . . , N 1.
(6)
In [10] it is shown that if L commutes with F
2
, (5) is
simplied to
K = L + FLF
1
. (7)
Theorem1. One can further extend this idea such that, if L is
circulant and symmetric the above equation is also valid.
Proof. Let C be a circulant and symmetric matrix, then the
eigenvalue decomposition of C is [15, pages 201-202]
C = F
1
C
F, (8)
where
C
= diag(
(n) + x
(n 1) =
c
t
(x(n) x(n 1)), (11)
where x
N is
the sampling period, N is the length of the signal x(n), and
c is a real scalar. Hence, the second-order discrete derivative
x
(n 1) + 2x
(n) + x
(n + 1)
=
_
c
t
_
2
(x(n 1) 2x(n) + x(n + 1)).
(12)
EURASIP Journal on Advances in Signal Processing 3
The bilinear transform maps the analog domain to the
discrete domain one-to-one. It maps points in the s-domain
with Re{s} = 0 ( j axis) to the unit circle in the z-plane
|z| = 1. However, the forward dierence method maps the
j to a circle of radius 0.5 and centered at the point z = 0.5
as shown in Figure 1. Bilinear transform maps every point in
the j-plane to the z-plane without aliasing.
We express (12) in matrix form as
B
1
X
=
_
c
t
_
2
E
2
X, (13)
where X
= [x
(0), x
(1), . . . , x
(N 1)]
T
, X =
[x(0), x(1), . . . , x(N 1)]
T
with
B
1
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
2 1 0 0 1
1 2 1 0
0 1 2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1 0 0 1 2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
, (14)
E
2
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
2 1 0 0 1
1 2 1 0
0 1 2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1 0 0 1 2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
. (15)
Hence, we conclude with an equivalent form of discrete
second derivative as
X
=
_
c
t
_
2
B
1
1
E
2
X, (16)
with the discrete second derivative operator D
2
=
(c/t)
2
B
1
1
E
2
.
3. Obtaining DFT Commuting Matrices
An easy and accurate way of obtaining Hermite-Gaussian-
like eigenvectors of the DFT matrix is to dene a better
commuting matrix, which imitates the Hermite-Gaussian
generating dierential equation given in (3) as a discrete
substitute. In this section we disclose an elegant way of
obtaining better commuting matrices by taking advantage of
the bilinear transform, which is a good discrete substitute for
the derivative operator.
The algorithm is straightforward; we substitute the
second derivative and the Fourier transform operators in
(3) with the matrix given in (16) and the DFT matrix,
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Real {z}
Figure 1: Image of j axis in the z-plane for bilinear transform and
the forward dierence method. Solid: bilinear transform, dashed:
forward dierence method.
respectively. Hence, DFT commuting matrix inspired by the
bilinear transform is given by
B = B
1
1
E
2
+ FB
1
1
E
2
F
1
. (17)
We omit the coecient (c/t)
2
, since it has no eect on the
eigenvectors of B.
Theorem2. B commutes with the DFT matrix.
Proof. As B
1
and E
2
are both circulant and symmetric, B
1
1
E
2
is symmetric and circulant also. We use Theorem 1 given in
Section 2.2, which states that any circulant and symmetric
matrix C can be used to generate a commuting matrix as in
(10). Since B
1
1
E
2
is both circulant and symmetric, the proof
is complete.
After generating the commuting matrix B, we nd its
eigenvectors. The eigenvectors are Hermite-Gaussian-like
eigenvectors with the number of zero-crossings equal to the
order of Hermite-Gaussian eigenvectors. In Section 4 we give
extensive simulations and results on these Hermite-Gaussian
like eigenvectors.
3.1. Stability. Stability of B can easily be proved when
B
1
is not singular. We can show this by the eigenvalue
decomposition of B
1
B
1
= F
1
B1
F, (18)
where
B1
is a diagonal matrix containing the eigenvalues
of B
1
. As B
1
is circulant, the eigenvalues are
B1
=
diag(
NFb
1
), where b
1
is the rst column of B
1
and N N
4 EURASIP Journal on Advances in Signal Processing
is the dimension of
B1
. As b
1
= [2, 1, 0, 0, . . . , 1]
T
, Fourier
transform of b
1
can be easily found and replaced to nd the
eigenvalues
B1
B1
= diag
_
2 + 2 cos
_
2n
N
__
, n = 0, 1, 2, . . . , N 1.
(19)
B1
is never zero for odd N, since diag(2 + 2 cos(2n/N)) >
0 for all n. However as N increases
B1
becomes poorly
conditioned. Besides, for even N, diag(2+2 cos(2n/N)) = 0
for n = N/2, which causes instability. We can add a small
> 0 in the diagonal of B
1
to overcome instability. Then
B1
is changed to
B1
= diag
_
2 + + 2 cos
_
2n
N
__
, n = 0, 1, 2, . . . , N 1
(20)
to preserve stability for even N. Consequently, to ensure
stability we substitute B
1
dened in (14) with
B
1
=
_
_
_
_
_
_
_
_
_
_
_
_
_
2 + 1 0 0 1
1 2 + 1 0
0 1 2 +
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1 0 0 1 2 +
_
_
_
_
_
_
_
_
_
_
_
_
_
. (21)
Adding a small value in the diagonal will not perturb the
eigenvectors of the commuting matrix.
3.2. Better Conditioned Bilinear Methods. Bilinear transform
can be considered as a trapezoidal approach to the derivative.
Hence, we can assure stability by using alternative B
1
matrices. We have found out that changing the diagonal
of B
1
by a constant k > 2 both ensures the stability and
increases the performance. Therefore we substitute B
1
with
B
1
, where we dene B
1
as
B
1
=
_
_
_
_
_
_
_
_
_
_
_
_
_
k 1 0 0 1
1 k 1 0
0 1 k
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1 0 0 1 k
_
_
_
_
_
_
_
_
_
_
_
_
_
. (22)
As k > 2, the commuting matrix is better conditioned. The
optimum value of k is found to be approximately 4.3, which
is given in Section 4.
3.3. Higher-order Bilinear Dierentiation Matrix Substitutes.
So far, we have used the bilineartransforminspired
matrices to nd a better discrete substitute for the sec-
ond derivative. To nd better denitions of dierentiation
matrices we suggest that a Taylor series-like approach to B
1
Table 1: Optimum a
i
coecients generated for B
14
.
a
i
Optimum Value
a
1
1.00
a
2
0.247634068038315
a
3
0.103839534211561
a
4
0.141176982675410
a
5
0.005956945393076
a
6
0.008133047918379
a
7
0.020103743248487
a
8
0.001866823892062
a
9
0.000336065416294
a
10
0.002383849560258
a
11
0.000725049220057
a
12
0.000698349278537
a
13
0.003339855815284
a
14
0.001759635742928
(1) Compute one of
B
1
, B
1
, or B
n
matrices.
(2) Replace the computed matrix in (17) as a substitute for
B
1
and compute the DFT-commuting matrix B.
(3) Find the eigenvectors of B, which are Hermite-
Gaussian-like orthonormal vectors.
Algorithm 1: Summary of the proposed algorithms.
will grant us higher-order bilinear dierentiation matrices.
Therefore, we dene higher-order bilinear dierentiation
matrices as
B
n
= a
1
B
1
+ a
2
_
B
1
_
2
+ + a
n
_
B
1
_
n
, (23)
where we name B
n
as nth-order bilinear approximation to
the second derivative, and a
i
are real scalars. The value of k =
4.3 is chosen for B
n
, as it is an optimum value with respect to
minimum total error norm which is discussed in Section 4.
We have not come up with an analytical expression of a
i
s
yet, however, genetic and/or pattern search algorithms may
be used to optimize the coecients.
We have used the genetic [16] and the pattern search
[17] algorithms and determined optimum a
i
coecients,
i = 1, 2, . . . , 14, which are given in Table 1. These coecients
are inserted in (23) to obtain B
14
. We have generated the
commuting matrix B by substituting B
1
with B
14
in (17).
When B
14
is employed, eigenvectors of B are found to be
very close to the samples of Hermite-Gaussian functions as
the performance is discussed in detail in the very Section 4.
So far, three dierent methods are proposed, which are
summarized in Algorithm 1. The rst method computes
B
1
, in which a small is added in the diagonal of B
1
to achieve stability. In the second method we alter the
diagonal of B
1
, with a value k > 2. Changing the diagonal
both improves the performance and ensures stability. In
the last proposed method we nd higher-order matrices,
using the B
1
and its weighted powers with k = 4.3
for a better denition of the commuting matrix. After-
wards, we replace the computed
B
1
, B
1
, or B
n
with B
1
in (17). The obtained matrix B is the DFT-commuting
EURASIP Journal on Advances in Signal Processing 5
5
10
15
20
25
30
35
40
45
50
55
2 3 4 5 6 7 8 9 10
(k)
T
o
t
a
l
n
o
r
m
o
f
e
r
r
o
r
N = 64
N = 56
N = 48
N = 40
N = 32
(a)
10
2
10
1
10
0
0 5 10 15 20 25 30
Number of zero crossing
l
o
g
1
0
(
n
o
r
m
o
f
e
r
r
o
r
)
B
1
, k = 2 +
B
1
, k = 3
B
1
, k = 4
B
1
, k = 4.3
(b)
Figure 2: (a) The total norm of error versus k in B
1
for N = 32,
40, 48, 56, and k = 4.3. The total norm of error is minimum when
k 4.3. (b) Error norms between the discrete Hermite-Gaussian
like eigenvectors and the samples of the Hermite-Gaussian
functions when B
1
for k = 3, 4, and k = 4.3 and
B
1
with N = 32.
matrix whose eigenvectors are the Hermite-Gaussian-like
orthonormal vectors.
4. Simulations and Results
We have proposed three dierent techniques for nding
Hermite-Gaussian-like eigenvectors of the DFT. In the rst
method we employ
B
1
dened in (21). As a second method
we use B
1
as dened in (22) for dierent values of k. Finally,
we employ B
n
given in (23). We replace each matrices in
(17) as a substitute to B
1
to generate commuting matrices B.
Afterwards, we nd eigenvectors of these commuting matri-
ces and nd the norm of error between samples of corre-
sponding Hermite-Gaussian functions and the eigenvectors.
First we compare total norms of errors between the
Hermite-Gaussian functions and the samples of Hermite-
Gaussian-like eigenvectors to determine optimum k for B
1
.
We dene the total normof error as sumof norms of error for
each eigenvector. Figure 2(a) shows the total norm of error
10
6
10
5
10
4
10
3
10
2
10
1
10
0
0 5 10 15 20 25 30
Number of zero crossings
l
o
g
1
0
(
n
o
r
m
o
f
e
r
r
o
r
)
B
1
S
2
S
6
S
16
(a)
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
0 10 20 30 40 50 60
Number of zero crossings
l
o
g
1
0
(
n
o
r
m
o
f
e
r
r
o
r
)
B
1
S
2
S
6
S
16
(b)
Figure 3: Error norms between the discrete Hermite-Gaussian like
eigenvectors and the samples of the Hermite-Gaussian functions of
B
1
method when k = 4.3 are compared with various other methods
for (a) N = 32 and (b) N = 64.
versus k for dierent values of N, and the best value for k is
approximately 4.3.
Comparison of errors between B
1
for k = 3, 4, and 4.3
and
B
1
with the dimensional N = 32 is given in Figure 2(b).
The error norm is measured as the norm of error between
the samples of Hermite-Gaussian functions and Hermite-
Gaussian like eigenvectors of B using these matrices. As it
is clear from the gures, the best overall performance is
obtained in the B
1
method when k = 4.3.
Figure 3(a) plots the norms of errors for dierent
methods dened in [8]. We compare the error norms of
S
2
, S
6
, and S
16
in between, which are of O(h
2
), O(h
6
), and
O(h
16
) Taylor approximations, respectively, as shown in [8],
with the B
1
method, k = 4.3 for N = 32. Figure 3(b) plots
6 EURASIP Journal on Advances in Signal Processing
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 5 10 15 20 25 30 35
Number of zero crossings
N
o
r
m
o
f
e
r
r
o
r
B
14
S
100
S
400
S
32
Figure 4: Comparison of error norms between the discrete
Hermite-Gaussian like eigenvectors and the samples of the
Hermite-Gaussian functions of B
14
and S
32
, S
100
, and S
400
methods
for N = 32.
the same comparison for N = 64. These plots show that
our proposed algorithm is slightly worse than some other
methods for small orders, but much better for higher-orders
of eigenvectors. As it is clear from the gures, our method
outperforms the other methods in total.
We compare the proposed higher-order B
14
method
with the other higher-order methods, S
32
, S
100
, and S
400
that employ higher-order Taylor approximations to the
second derivative as shown in [10]. Figure 4 presents the
performance of the proposed method together with the other
methods. Despite the fact that our method uses only the
14th order approximation, it is denitely better than these
methods, even better than S
400
.
5. Conclusions
As the eigenvectors that are closer to the samples of
continuous Hermite-Gaussian functions are important for
a better denition of DFrFT, we employ bilinear transform-
based methods to dene better commuting matrices. We
have proposed three dierent methods and analyzed their
stability issues. A stable method is proposed by inserting
a small in the diagonal of the bilinear matrix. Better
conditioned bilinear dierentiation matrices that have
better performance are also obtained. Besides, a method of
generating higher-order bilinear dierentiating matrices is
also suggested.
Simulation results show that the proposed methods
posess better eigenvectors when compared to the other
methods recently suggested.
Future works on this subject may include nding a
closed form expression for the coecients generating the
higher-order bilinear matrices, B
n
. Furthermore, B
n
matrices
may be used in linear combinations with other commuting
matrices, such as S
2k
.
References
[1] H. M. Ozaktas, Z. Zalevski, and M. A. Kutay, The Frac-
tional Fourier Transform with Applications in Optics and
Signal Processing, John Wiley & Sons, New York, NY, USA,
2001.
[2] M. A. Kutay, H. M. Ozaktas, O. Ankan, and L. Onural,
Optimal ltering in fractional Fourier domains, IEEE Trans-
actions on Signal Processing, vol. 45, no. 5, pp. 11291143,
1997.
[3] X.-G. Xia, On bandlimited signals with fractional Fourier
transform, IEEE Signal Processing Letters, vol. 3, no. 3, pp. 72
74, 1996.
[4] O. Akay and G. F. Boudreaux-Bartels, Fractional convolution
and correlation via operator methods and an application to
detection of linear FM signals, IEEE Transactions on Signal
Processing, vol. 49, no. 5, pp. 979993, 2001.
[5] B. Santhanam and J. H. McClellan, The discrete rotational
Fourier transform, IEEE Transactions on Signal Processing, vol.
44, no. 4, pp. 994998, 1996.
[6] C . Candan, M. A. Kutay, and H. M. Ozaktas, The discrete
fractional Fourier transform, IEEE Transactions on Signal
Processing, vol. 48, no. 5, pp. 13291337, 2000.
[7] S.-C. Pei, W.-L. Hsue, and J.-J. Ding, Discrete fractional
Fourier transform based on new nearly tridiagonal commut-
ing matrices, IEEE Transactions on Signal Processing, vol. 54,
no. 10, pp. 38153828, 2006.
[8] C . Candan, On higher order approximations for Hermite-
Gaussian functions and discrete fractional Fourier trans-
forms, IEEE Signal Processing Letters, vol. 14, no. 10, pp. 699
702, 2007.
[9] S.-C. Pei, J.-J. Ding, W.-L. Hsue, and K.-W. Chang, Gener-
alized commuting matrices and their eigenvectors for DFTs,
oset DFTs, and other periodic operations, IEEE Transac-
tions on Signal Processing, vol. 56, no. 8, pp. 38913904,
2008.
[10] S.-C. Pei, W.-L. Hsue, and J.-J. Ding, DFT-commuting
matrix with arbitrary or innite order second derivative
approximation, IEEE Transactions on Signal Processing, vol.
57, no. 1, pp. 390394, 2009.
[11] B. Santhanam and T. S. Santhanam, Discrete Gauss-Hermite
functions and eigenvectors of the centered discrete Fourier
transform, in Proceedings of the IEEE International Conference
on Acoustics, Speech and Signal Processing ( ICASSP 07), vol. 3,
pp. 13851388, 2007.
[12] B. W. Dickinson and K. Steiglitz, Eigenvectors and functions
of the discrete Fourier transform, IEEE Transactions on
Acoustics, Speech, and Signal Processing, vol. 30, no. 1, pp. 25
31, 1982.
[13] F. A. Gr unbaum, The eigenvectors of the discrete Fourier
transform: a version of the Hermite functions, Journal of
Mathematical Analysis and Applications, vol. 88, no. 2, pp. 355
363, 1982.
[14] A. Serbes and L. Durak, Ecient computation of DFT
commuting matrices by a closedform innite order approxi-
mation to the second dierentiation matrix, Signal Process. In
press.
[15] G. G. Golub and C. F. Van Load, Matrix Computations, The
Johns Hopkins University Press, London, UK, 3rd edition,
1993.
EURASIP Journal on Advances in Signal Processing 7
[16] D. E. Goldberg, Genetic Algorithms in Search, Optimzation
and Machine Learning, Addison-Wesley, Reading, Mass, USA,
1989.
[17] C. Audet and J. E. Dennis Jr., Analysis of generalized pattern
searches, SIAM Journal on Optimization, vol. 13, no. 3, pp.
889903, 2003.