1. Introduction
Linear canonical transform which is generated by second-order differential operators, is a four-parameter class of linear integral transform [
1,
2],
where
is LCT of the function
. The LCT integral kernels are Green functions of quadratic Hamiltonians that can be found in [
3]. The research on LCT was first proposed by Collins (1970) and Moshinsky (1971) [
1,
2]. It includes many special cases, such as, the Fourier transform (FT), the fractional Fourier transform (FRFT), the Fresnel transform, the Lorentz transform and scaling transform. The class of LCTs are important in signal processing [
4,
5], computational and applied mathematics [
6,
7], optics [
8] and quantum mechanics [
9]. Significant applications of LCT in signal processing include radar system analysis, filter design, pattern recognition, image watermarking and so on [
10,
11]. Basic theories of LCT have been developed that include convolution theorems, sampling theorems, and uncertainty principles [
12,
13,
14]. The numerical approximation of the LCT is of importance in modeling first-order optical systems and many signal processing applications. Therefore, the discrete and fast algorithms of the LCT are one of the most important issues in practical applications.
After the continuous LCT has been introduced, therefore, the definition and fast implementation of the discrete linear canonical transform (DLCT) have been widely considered by many researchers [
15,
16,
17,
18,
19,
20,
21,
22,
23,
24,
25,
26,
27,
28]. The existing algorithms can be divided into the following categories. The one is the operator decomposition type, which decomposes an arbitrary LCT operator into its special cases that have a fast algorithm [
17,
18,
29]. The second is the eigenvector decomposition type, which computes the LCT of a function by using the eigenfunctions of the LCT [
24,
27]. The third is split basis algorithm, which decomposes the discrete transform matrix of LCT into smaller matrices iteratively [
20,
23]. All these algorithms can effectively and rapidly calculate the entire spectrum in the LCT domain. In recent years, the new LCT algorithms have been proposed, which can realize local spectrum analysis of signals [
30,
31,
32]. In addition to the above-mentioned algorithms, various other fast algorithms have been proposed in [
33,
34,
35,
36,
37,
38,
39]. These research results provide a good basis for the further development of the DLCT toward meeting the requirements of practical applications. Many aspects of the fast methods of the DLCT still need to be studied. To the best of our knowledge, the existing discrete algorithms are required that both input and output data are uniform sampling. However, for certain applications, the input or output data is nonuniform. In these cases, the fast DLCT will be lost.
For overcome the aforementioned problems, in this paper, we present a set of fast algorithms for computing nonuniform DLCT, namely,
where
,
,
,
, and
. According to the sampling of the
and
, we will operate under the following assumptions
Uniform samples and non integer frequencies: In Equation (
1) the samples are equispaced, i.e.,
, and the frequencies
are non integer. This corresponds to evaluating a generalized linear canonical series at equispaced points.
Nonuniform samples and integer frequencies: In Equation (
1),
are nonequispaced points in
and the frequencies
are integers.
Nonuniform samples and non-integer frequencies: In Equation (
1),
are nonequispaced points in
and the frequencies
are non-integers. This is the fully nonuniform transform and corresponds to evaluating a generalize linear canonical series at nonequispaced points.
To develop various nonuniform fast LCT(NFLCT) for above issues, one has to exploit a nonzero working precision of and makes careful approximations. The approximation properties of the various approaches may be obtained by considering how they perform on the linear canonical modes.
The rest of the paper is organized as follows: In
Section 2, a brief review of the related preliminaries is presented which are used in the design of the algorithms. The
Section 3 is main results of this paper. In this section, we give an exact statement of the problem and introduce some notation that is used. The algorithms are derived. Some numerical examples are presented in
Section 4 to illustrate the preference of the schemes. Finally, conclusions are drawn in
Section 5.
2. Preliminary
The linear canonical series (LCS) is a generalized form of Fourier series (FS), which can reveal the mixed time and frequency components of signals. The basis function of LCS is defined as [
40]
where
. Thus,
construct an orthonormal basis. It can be observed that every basis function is a chirp function with chirp rate
, which is an aperiodic function. Therefore, the LCS is only applicable to finite-length function. The LCS expansion of the finite-length function
can be written as
where
and
are called LCS expansion coefficients with the parameter matrix
A.The LCS expansion coefficients are computed by the inner product of the function and chirp basis function. The relationship between LCS and LCT is that the LCS expansion coefficients are the sampled values of LCT, by
The well-known FS is just a special case of LCS for the parameter matrix
. In addition, we also presented some well results to be used in the remainder of the paper [
41].
Lemma 1. For any real and complex z, Lemma 2. For any real and , In the next section, the main results will be derived based on above facts.
4. Simulations
In this section, some numerical examples are given to support our theoretical analysis in the above section. All tests of numerical examples are implemented in Matlab R2016a. Two measures of precision are selected for each algorithm,
where
is the input data,
f is the result of a direct computation, and
is the result of computation by proposed methods.
Example 1. Here we consider the transformation of Problem 1, as defined Equation (8). In this example, , were randomly distributed on the interval , and were generated randomly on the unit square in the complex plane defined as We take ; , .
We applied Algorithm 1 and direct method to this problem respectively. For different
N, we loop Algorithm 1 20 times, the error mean are presented in
Table 1. The results show that the precision is almost independent the length of input data.
Example 2. The Problem 2 as defined Equation (9) is considered for . In this example, were randomly distributed on the interval , and , , were distributed randomly on the interval . We take the parameters
, the interpolate factors
, the terms of LCS
,
. For
, we loop our algorithm 20 times, the results are presented in
Figure 1. It shows that the Algorithm 2 has almost the same effective as the direct method. For different
N, the results of
and
are showed in
Table 2. It suggested that the precision of the Algorithm 2 is independent
N. For different
and
q, the errors are presented in
Table 3. It shows that the lager
and
q is, the more accuracy of the Algorithm 2 is, and the rate of accuracy improvement is decreasing. The complexity of the Algorithm 2 dependents
q and
N. For
and different
q, the calculation is plotted in
Figure 2. Therefore, we should not choose too big
q according to the accuracy and complexity.
Example 3. Here we consider the transformation of the Problem 3 as defined Equation (10). In this example, were randomly distributed on the interval , were randomly distributed on the interval , andwhere , ; interpolation factor ; the terms of the linear canonical series and the constant . For
, the
Figure 3 shows the amplitude and error for this problem by direct summation Equation (
10) and Algorithm 3, respectively. It suggested that the Algorithm 3 has the almost same performance as the direct method. For different
N, the
Table 4 shows the error
and
between direct method and Algorithm 3 by cyclic algorithms 20 times. The results show that the precision of Algorithm 3 is almost independent of the length of input data.
From the above three examples, the errors produced by Algorithms 1–3 are comparable with those produced by the corresponding direct methods. we can see that the numerical results coincide with the theoretical analyses which show the high efficiency of the new method.
The results of this paper can be generalized in the following ways: the algprithms of 1, 2 and 3 will allow the efficient application of linear transform
defined by
for
, and
For
, we will also consider the more general transformation
defined by the formula
The algorithms ofthis paper also assume that . Other distributions can be handle by partitioning the vectors u and t,treating each partition separately and finally combining the results.