1. Introduction
The autocorrelations of transmitted signals are of great significance for the signal processing in active sonars. A good autocorrelation property indicates that the signal is nearly uncorrelated with its own time-delay versions [
1,
2], which ensures that the active sonar can precisely extract the echo information from the interested time lags while suppressing interferences from other time lags [
3,
4]. One of the most widely used transmitted signals with good autocorrelation properties is the phase-coded sequence [
5,
6]. Due to practical constraints of sonar transducers such as the frequency response and the energy efficiency, it is desirable to design the transmitted signals with nearly a constant amplitude, which aroused a lot of effort to research on unimodular phase-coded sequences [
7,
8]. The early studies regarding phase-coded sequences mainly focused on binary sequences to reach the low autocorrelation sidelobes, for example, the Barker sequence [
9]. Lately, considering the fact that binary sequences are of low computational efficiency and difficult to be generated with long length, researches shifted to other polyphase sequences such as the Golomb sequences [
10,
11] and the Frank sequences [
12]. Correspondingly, the optimization for polyphase sequences becomes one of the major interests in the sequence designs. Most researchers optimize this problem with the Integrated Sidelobe Level (ISL) metric [
13,
14,
15] or the Weighted Integrated Sidelobe Level (WISL) metric [
11,
16,
17].
The ISL metric is the most commonly-used criterion to evaluate the autocorrelation properties of phase-coded sequences. Based on the singular value decomposition (SVD), Li first proposed the cyclic algorithm (CA) to design sequences satisfying the ISL metric [
18]. However, as the SVD operations are relatively computationally intensive, it might be difficult to generate the sequence with length more than
. Afterwards, Stoica developed two extensions of the CA, called ‘Cyclic Algorithm New’ (CAN) and ‘Periodic-correlation Cyclic Algorithm New’ (PeCAN) to generate sequences of length
in the aperiodic and periodic ISL case, respectively [
13,
14]. Compared with the CA, the CAN and the PeCAN utilize the Fast Fourier Transform (FFT) operations and reduce computational burdens effectively.
For a phase-coded sequence satisfying the ISL metric, it is difficult and time-consuming to suppress autocorrelation sidelobes at all time lags. As a result, the WISL metric is developed to reduce autocorrelation sidelobe levels at specific time lags [
13]. Hao He proposed the periodic CA (PeCA) to generate sequence sets with the periodic WISL metric based on the SVD operations [
19]. Stoica developed an algorithm called ‘Weighted Cyclic Algorithm New’ (WeCAN) for the design of sequences with the aperiodic WISL metric [
13]. Compared with sequences satisfying the ISL metric, the PeCA and the WeCAN sequences are flexible in practical applications.
Although the above algorithms, PeCA and WeCAN, can generate sequences with the WISL metric, the computational burdens are considerably intensive. In order to enhance the computational efficiency, the majorization–minimization (MM) strategy was applied to sequence design methods. Instead of optimizing the objective problem directly, the MM strategy constructs surrogate Equations to approximate the objective and gradually decomposes the
-dimensional equation into the sum of one-dimensional equations which can be minimized easily [
20]. This strategy has already been implemented to large-scale or non-convex optimization problems [
21]. Song dealt with the aperiodic ISL and WISL Equations through the MM strategy [
15,
17]. Compared with the aforementioned CA and CAN, the algorithm using MM strategy reaches a faster convergence speed and lower computational burden.
In this paper, we intended to apply the MM strategy for the periodic WISL Equation and proposed a real-time sequence design method. The main contributions and advantages of the paper can be summarized as: (1) We intended to reach low autocorrelation sidelobes at specific time lags (namely WISL metric), which is much easier to reach than the ISL metric in the complex underwater environment we focus on. (2) In this paper, we reduced the calculation amount and improved the real-time capacity of sequence design methods in periodic WISL case, since the recent methods generating sequences with the periodic WISL cannot be satisfactory [
19]. (3) For the first time, we proved that the periodic WISL Equation can be tackled through the MM strategy which has been applied in the aperiodic/periodic ISL and the aperiodic WISL equation as for example in [
15,
16,
17].
The rest of the paper is organized as follows. In
Section 2, we derived a phased-coded sequence design method based on the MM strategy. Some derivations of our method in periodic WISL case are different from the aperiodic case [
17]. For completeness and clearness, we presented all derivations of the method in this section. Also, an acceleration scheme was applied in the method in order to enhance the convergence speed. Simulations are presented in
Section 3 to evaluate the convergence performance of the proposed method. In addition, the matched filter performance of the transmitted sequence was evaluated since the matched filter is a standard echo processing of active sonars. Finally,
Section 4 proposed the conclusions of this paper.
Notation: In this paper, boldface upper case letters denote matrices, boldface lower case letters denote column vectors. Notations denote transpose, conjugate, and conjugate transpose. denotes the real part, denotes the trace of the matrix and denotes the column-wise vectorization. denotes the Euclidean norm. is a diagonal matrix formed with as its principal diagonal. is the identity matrix.
2. MM Based Phase-Coded Sequence Design Method
Let’s define
as a complex unimodular sequence with length
. It is well known that the periodic autocorrelation of the sequence can be defined as [
13]:
Here
is the modulo operation which denotes that:
where
is the largest integer smaller than or equal to
. The WISL metric of periodic autocorrelation can be written as:
where
represent the weights set of the WISL metric. Then the design method of the sequence
is considered as the following optimization problem:
Equation (4) expresses the optimization problem of both the aperiodic [
17] and the periodic WISL metric. To make the WISL metric expressed clearly, the basis matrices in the aperiodic case are defined as
Toeplitz matrices
:
where
in
kth diagonal are 1 and elsewhere are 0 [
17].
On the other hand, in periodic case, we define new basis matrices
as
circulant matrices:
where
in
kth diagonal and
in −(
N − k)th diagonal are all 1 and elsewhere are 0. To express Equation (4) as the symmetric form, let
be a null matrix and further define
. Through the expression of basis matrices
, Equation (4) with WISL metric of (3) can be rewritten as:
where
. Since
, Equation (7) can be expressed as:
Let’s define:
It is obvious that
is a Hermitian matrix. By the definition of
, Equation (8) can be rewritten as:
In the following parts, Equation (10) will be tackled by using the MM strategy. The key procedures of the MM strategy include: (1) Majorization: constructing majorization functions
as accurate as possible by designing an upperbound matrix
of the object matrix
. (2) Minimization: surrogating the original Equation (10) with the majorization function
and minimizing the surrogate equation. These two procedures will repeat several times until the closed-form solution can be reached. Before we use the MM strategy, a useful conclusion [
15] should be displayed so that it can be used later.
Lemma 1. are bothHermitian matrices and. For each, the majorization function of is , which ensures that .
According to Lemma 1, an upperbound matrix
which satisfying
needs to be designed in order to construct the majorization function. A simple choice is to define
where
is the maximum eigenvalue of
. The value of
is quantitatively expressed as:
Proof. Owing to the property of circulant matrices,
are mutually orthogonal so that:
Then both sides of Equation (9) are multiplied by
:
where the second equality results from the fact
and the third equality results from the fact
. According to Equation (13),
are non-zero eigenvalues of
with corresponding eigenvectors
. Then the maximum eigenvalue of
is given as Equation (11).
Giving
of the
pth iteration and choosing
. It is easy to see that both
and
are Hermitian matrices. According to Lemma 1, the majorization function which surrogates Equation (10) at
can be given as:
Considering the fact that
, the first term of Equation (14) is a constant and the third term depends on the
pth iteration only. Ignoring these immaterial terms, Equation (10) can be surrogated by
as:
Substituting Equation (9) into Equation (15), the equation becomes:
Considering
and
, then we rewrite Equation (16) as:
Here the matrix
has the form as follows:
According to the definition in [
22,
23],
,
, and
are all circulant matrices.
Through the above MM procedures, we decompose the quartic function of Equation (10) to the quadratic function of Equation (17). However, the quadratic function cannot get closed-form solution and needs to be further decomposed. As a result, we intend to use the MM strategy again and construct another majorization function to decompose the equation. In Lemma 1, both the object matrix and the surrogate matrix should be Hermitian matrix. To decompose Equation (17), we will introduce another lemma so that the surrogate function in Lemma 1 is also valid when the object matrix is a circulant matrix. □
Lemma 2. is thecirculant matrix,is theHermitian matrix, and. For each, the majorization function ofis similar with the conclusion in Lemma 1.
With Lemma 2, the key to construct the majorization function of Equation (17) is to design an upperbound matrix
so that
. Similar with Equation (11), a simple way is to choose the upperbound matrix
so that it can be expressed as:
Since
is a constant, we should focus on the eigenvalues of
. A column vector
is defined which is composed of the first row elements of matrix
. Also, the Fourier transform matrix is defined as:
Then all eigenvalues of
is solved as:
By Equations (20) and (21), the eigenvalues of
, and
can be obtained by inverse Fourier transforms. Moreover, the circulant matrix ensures the relation between the maximum eigenvalues of
, and
as [
24]:
Since
in Equation (19), Equation (22) can be expanded as:
Now we define
and
which is a Hermitian matrix. Since the matrix
is also circulant like
, we can now use Lemma 2 and the majorization function of Equation (17) can be yielded as:
Since
, the first term of Equation (24) is a constant and the third term depends on the
pth iteration only. Ignoring these constant terms, the original Equation can be surrogated by
as:
which can be simplified as follows:
Equation (26) has a closed-form solution as:
In order to improve the computation efficiency of MM procedures, we can express the WISL metric in Equation (3) and
in Equation (27) via the Fourier transform matrix. First of all, the periodic autocorrelation of a unimodular sequence can be written as follows [
25]:
where the Fourier matrix
is defined as Equation (20). Then the WISL metric can be obtained through Equation (3). Furthermore, the computation of
depends on the matrix
in Equation (18) which can be expressed as follows [
24]:
where
is the column vector composed of the first row elements of
.
The MM procedures of the method are presented as
Table 1. During the procedures, it may not be definite to reach the steepest descent in each iteration, which results in a slow convergence speed. Hence, we modify the gradient direction in each iteration with the so called squared iterative method (SQUAREM) in order to improve the MM strategy and accelerate the convergence. The SQUAREM is derived from the Cauchy–Barzilai–Borwein algorithm [
26] and originally used to accelerate the expectation–maximization (EM) algorithm in maximum likelihood estimates [
27]. Recently, the SQUAREM is proved to be applicable for the MM strategy in aperiodic WISL case [
17]. Since it is an ‘off-the-shelf’ acceleration scheme which needs nothing other than the parameter updating rules of MM procedures [
16], we apply this scheme straightforward to accelerate our method. The details of the scheme can be found in [
17]. Finally, the sequence design method based on MM procedures and the SQUAREM scheme is named as PeWISL method.