【论文翻译】（第三部分）The Fourier decomposition method for nonlinear and non-stationary time series analysis

3. The Fourier decomposition method
3. 傅里叶分解方法

We propose a class of functions termed as the FIBFs, belonging to C∞[a,b], here with the following formal definition.

Definition 3.1. Let x(t) be an arbitrary signal, defined in the interval [a,b], following the Dirichlet conditions. A set of functions { y i ( t ) : y i ( t ) ∈ C ∞ [ a , b ] , 1 ≤ i ≤ M } \{y_i(t):y_i(t)\in C^\infty[a,b], 1≤i≤M\} {yi​(t):yi​(t)∈C∞[a,b],1≤i≤M} is called a FIBF set of $x(t)$, if the following conditions are satisfied:

(1) $x(t)={\sum }_{i=1}^M{y}_i(t)+a_0$, where $a_0$ is the mean value of $x(t)$ ;
(2) the FIBFs are zero mean functions, that is ${\int }_a^b{y}_i(t)dt=0,\forall i$ ;
(3) the FIBFs are orthogonal functions, that is ${\int }_a^b{y}_i(t)y_l(t)dt=0$, for $i\ne l$ ;
(4) the FIBF admit the Analytic FIBFs (AFIBFs) representation: $y_i(t)+j{\hat{y}}_i(t)=a_i(t)e^{j{\varphi }_i(t)}$ , with IF ${\omega }_i(t)=(d/dt){\varphi }_i(t)\ge 0,\forall t$ and amplitude $a_i(t)\ge 0,\forall t$ , where ${\hat{y}}_i(t)$ is obtained by the complex exponential Fourier representation and it is equivalent to the HT of FIBF $y_i(t)$.

Thus, the AFIBFs are monocomponent signals because, physically , IF has meaning only for monocomponent signals consisting of a single-frequency component or a narrow range of frequencies varying as a function of time [17]. Hence, the FIBF is sum of zero mean sinusoidal functions of consecutive frequency bands.

The main objective of this study is to develop a novel and adaptive decomposition method, completely based on the Fourier theory , to obtain a unique representation of multicomponent signal as a sum of the mean-value and non-stationary monocomponent signals satisfying the properties presented in definition 3.1. The necessary conditions [3] for a set of basis vectors to represent a nonlinear and non-stationary time series are completeness, orthogonality , locality and adaptiveness. The FIBFs, intrinsically , follow all the necessary conditions by virtue of the proposed decomposition.

我们提出了一类称为FIBFs的函数，属于C∞[a，b]，下面是正式的定义。

定义3.1. 设x(t)是任意信号，定义在区间[a,b]内，遵循狄利克雷条件。一组满足下列条件的函数 { y i ( t ) : y i ( t ) ∈ C ∞ [ a , b ] , 1 ≤ i ≤ M } \{y_i(t):y_i(t)\in C^\infty[a,b], 1≤i≤M\} {yi​(t):yi​(t)∈C∞[a,b],1≤i≤M} 称为 $x(t)$ 的FIBF集:

(1) $x(t)={\sum }_{i=1}^M{y}_i(t)+a_0$, 其中 $a_0$ 是 $x(t)$ 的均值;
(2) FIBFs 是零均值函数，也就是说 ${\int }_a^b{y}_i(t)dt=0,\forall i$ ;
(3) FIBFs 是正交函数，也就是说 ${\int }_a^b{y}_i(t)y_l(t)dt=0$, for $i\ne l$ ;
(4) FIBF 承认解析 FIBFs (AFIBFs)表示: $y_i(t)+j{\hat{y}}_i(t)=a_i(t)e^{j{\varphi }_i(t)}$ , 其 IF （瞬时频率）为 ${\omega }_i(t)=(d/dt){\varphi }_i(t)\ge 0,\forall t$ ，振幅为 $a_i(t)\ge 0,\forall t$ , 其中 ${\hat{y}}_i(t)$ 由复指数傅里叶表示获得，相当于FIBF $y_i(t)$ 的HT（希尔伯特）变换。

因此，AFIBFs是单分量信号，因为在物理上，瞬时频率仅对由一个单频分量或一个以时间[17]为函数变化的窄范围的频率组成的单分量信号有意义。因此，FIBF是连续频带零均值正弦函数的和。

本研究的主要目的是开发一种新颖的自适应分解方法，完全基于傅里叶理论，以获得满足定义3.1性质的多分量信号的均值和非平稳单分量信号的唯一表示。用以表示非线性非平稳时间序列的一组基向量的必要条件[3]是完备性、正交性、局部性和自适应性。本质上，由于所提出的分解，FIBFs遵循所有的必要条件。

(a) Continuous time Fourier decomposition method
（a）连续时间傅里叶分解法

In this section, we propose a continuous time FDM (CT-FDM) which decomposes a signal into a set of FIBFs. The available data are usually of finite duration, non-stationary and generated from systems that may be nonlinear. Let $x(t)$ be a real-valued and time-limited signal in the interval $t_1\le t\le t_1+T_0$. We construct a periodic extension of this signal as $x_{T_0}(t)={\sum }_{k=-\infty }^{\infty }x(t-k{T}_0)$ such that $x(t)=x_{T_0}(t)w(t)$, where $w(t)=1$, for $t_1\le t\le t_1+T_0$ and zero otherwise. The Fourier series expansion of $x_{T_0}(t)$ is given by
$$x_{T_0}(t)=a_0+\sum _{k=1}^{\infty }[a_k\cos (k{\omega }_{0}t)+b_k\sin (k{\omega }_{0}t)](3.1)$$
where ${\omega }_0=2\pi /T_0$ rad s-1,
$$a_0=\frac{1}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)dt$$
$$a_k=\frac{2}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)\cos (k{\omega }_{0}t)dt$$
$$b_k=\frac{2}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)\sin (k{\omega }_{0}t)dt$$
Using the complex exponential representation of
$$\cos (k{\omega }_{0}t)=\frac{\exp (jk{\omega }_{0}t)+\exp (-jk{\omega }_{0}t)}{2}$$and$$\sin (k{\omega }_{0}t)=\frac{\exp (jk{\omega }_{0}t)-\exp (-jk{\omega }_{0}t)}{2}$$functions, we can wrote (3.1) as
$$x_{T_0}(t)=a_0+\frac{1}{2}\sum _{k=1}^{\infty }[c_k\exp (jk{\omega }_{0}t)+c_k^{\ast }\exp (-jk{\omega }_{0}t)](3.2)$$
where $c_k=(a_k-j{b}_k)$ and $c_k^{\ast }=a_k+j{b}_k$. From (3.2), it is clear that
x T 0 ( t ) = a 0 + R e { z T 0 ( t ) } ( 3.3 ) x_{T_0}(t)=a_0+Re\{z_{T_0} (t)\} \quad(3.3) xT0​​(t)=a0​+Re{zT0​​(t)}(3.3)
where the analytic function
$$z_{T_0}(t)\stackrel{\Delta }{=}\sum _{k=1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)(3.4)$$
is complex conjugate of ${\tilde{z}}_{T_0}(t)\stackrel{\Delta }{=}{\sum }_{k=1}^{\infty }{c}_k^{\ast }\exp (-jk{\omega }_{0}t)$ and Re{$z_{T_0}(t)$} denotes the real part of $z_{T_0}(t)$. In order to obtain a set of analytic FIBFs (AFIBFs), we write $z_{T_0}(t)$ as
$$z_{T_0}=\sum _{I=1}^M{a}_i(t)\exp (j{\varphi }_i(t))(3.5)$$
where, in forward search (low to high frequency scan (LTH-FS)) of AFIBFs,
$a_1(t)\exp (j{\varphi }_1(t))={\sum }_{k=1}^{N_1}{c}_k\exp (jk{\omega }_{0}t)$,$a_2(t)\exp (j{\varphi }_2(t))={\sum }_{k=N_1+1}^{N_2}{c}_k\exp (jk{\omega }_{0}t)$,…,$a_M(t)\exp (j{\varphi }_M(t))={\sum }_{k=N_M+1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)$. Thus, in general, we can write
$$a_i(t)\exp (j{\varphi }_i(t))=\sum _{k=N_{i-1}+1}^{N_i}{c}_k\exp (jk{\omega }_{0}t),\text{for}i=1,\dots ,M,(3.6)$$
with $N_0=0$ and $N_M=\infty$. The FIBFs are the real parts of AFIBFs presented in (3.6). In order to obtain a minimum number of AFIBFs in LTH-FS, for each i, start with ($N_{i-1}+1$) and append more terms until the maximum value of $N_i$ is reached such that ($N_{i-1}+1)\le N_i\le \infty$ and
$$a_i(t)\ge 0,{\omega }_i(t)=\frac{d{\varphi }_i(t)}{dt}\ge 0,\forall t(3.7)$$
where $a_i(t)$ and ${\omega }_i(t)=2\pi f_i(t)$ are the IA and IF o f ith FIBF, respectively . It is easy to observe that such a decomposition is always possible.

在这一节中，我们提出了一个连续时间FDM (CT-FDM)，它将一个信号分解成一组 FIBFs。可用数据通常是有限持续时间的、非平稳的，并且由可能是非线性的系统生成。设$x(t)$是区间$t_1\le t\le t_1+T_0$内的实值限时信号。 我们构造这个信号的周期延拓为$x_{T_0}(t)={\sum }_{k=-\infty }^{\infty }x(t-k{T}_0)$，并且有 $x(t)=x_{T_0}(t)w(t)$，其中 $w(t)=1，\text{for}{t}_1\le t\le t_1+T_0$并且在其他地方等于0。$x_{T_0}(t)$ 的傅里叶级数展开式为：
$$x_{T_0}(t)=a_0+\sum _{k=1}^{\infty }[a_k\cos (k{\omega }_{0}t)+b_k\sin (k{\omega }_{0}t)](3.1)$$
其中，${\omega }_0=2\pi /T_0$ rad s-1,$a_0=\frac{1}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)dt$,$a_k=\frac{2}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)\cos (k{\omega }_{0}t)dt$,$b_k=\frac{2}{T_0}{\int }_{t_1}^{t_1+T_0}{x}_{T_0}(t)\sin (k{\omega }_{0}t)dt$，使用$\cos (k{\omega }_{0}t)=\frac{\exp (jk{\omega }_{0}t)+\exp (-jk{\omega }_{0}t)}{2}$和$\sin (k{\omega }_{0}t)=\frac{\exp (jk{\omega }_{0}t)-\exp (-jk{\omega }_{0}t)}{2}$函数的复指数表示，我们可以将(3.1)写成
$$x_{T_0}(t)=a_0+\frac{1}{2}\sum _{k=1}^{\infty }[c_k\exp (jk{\omega }_{0}t)+c_k^{\ast }\exp (-jk{\omega }_{0}t)](3.2)$$
其中$c_k=(a_k-j{b}_k)$， $c_k^{\ast }=a_k+j{b}_k$. 根据式(3.2)，可以清楚的知道
x T 0 ( t ) = a 0 + R e { z T 0 ( t ) } ( 3.3 ) x_{T_0}(t)=a_0+Re\{z_{T_0} (t)\} \quad(3.3) xT0​​(t)=a0​+Re{zT0​​(t)}(3.3)
其中解析函数
$$z_{T_0}(t)\stackrel{\Delta }{=}\sum _{k=1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)(3.4)$$
是 ${\tilde{z}}_{T_0}(t)\stackrel{\Delta }{=}{\sum }_{k=1}^{\infty }{c}_k^{\ast }\exp (-jk{\omega }_{0}t)$ 的复共轭，而 Re{$z_{T_0}(t)$} 代表 $z_{T_0}(t)$ 的实部。为了得到一系列解析 FIBFs (AFIBFs)。我们将$z_{T_0}(t)$ 写成
$$z_{T_0}=\sum _{I=1}^M{a}_i(t)\exp (j{\varphi }_i(t))(3.5)$$
其中，在AFIBFs正向搜索(由低到高频率扫描(LTH-FS))中，$a_1(t)\exp (j{\varphi }_1(t))={\sum }_{k=1}^{N_1}{c}_k\exp (jk{\omega }_{0}t)$,$a_2(t)\exp (j{\varphi }_2(t))={\sum }_{k=N_1+1}^{N_2}{c}_k\exp (jk{\omega }_{0}t)$,…,$a_M(t)\exp (j{\varphi }_M(t))={\sum }_{k=N_M+1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)$。因此，一般来说，可以写成
$$a_i(t)\exp (j{\varphi }_i(t))=\sum _{k=N_{i-1}+1}^{N_i}{c}_k\exp (jk{\omega }_{0}t),\text{for}i=1,\dots ,M,(3.6)$$
其中 $N_0=0$ 且 $N_M=\infty$。FIBFs是(3.6)中给出的AFIBFs的实部。为了在LTH-FS中获得最小数量的AFIBFs，对于每一个 i，从($N_{i-1}+1$)开始，增加更多项，直到 达到 $N_i$ 的最大值即 ($N_{i-1}+1)\le N_i\le \infty$，并且
$$a_i(t)\ge 0,{\omega }_i(t)=\frac{d{\varphi }_i(t)}{dt}\ge 0,\forall t(3.7)$$
其中 $a_i(t)$ 和 ${\omega }_i(t)=2\pi f_i(t)$ 分别是 FIBF的第$i$个IA（瞬时振幅）和IF（瞬时频率）。很容易观察到这样的分解总是可能的。

Similarly , in reverse search (from high to low frequency scan (HTL-FS)) of AFIBFs, we obtain $a_1(t)\exp (j{\varphi }_1(t))={\sum }_{k=N_1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)$,$a_2(t)\exp (j{\varphi }_2(t))={\sum }_{k=N_2}^{N_1-1}{c}_k\exp (jk{\omega }_{0}t)$,…,$a_M(t)\exp (j{\varphi }_M(t))={\sum }_{k=1}^{N_{M-1}-1}{c}_k\exp (jk{\omega }_{0}t)$
and the lower and upper limits of sum in (3.6) would change to $k=N_i$ to $(N_{i-1}-1)$, respectively , with $N_0=\infty$ and $N_M=1$. Here, we start with $(N_{i-1}-1)$, decrease and select minimum value of $N_i$ such that $1\le N_i\le (N_{i-1}-1)$ and (3.7) is satisfied for $i=1,...,M$.

Similarly , in reverse search (from high to low frequency scan (HTL-FS)) of AFIBFs, we obtain$a_1(t)\exp (j{\varphi }_1(t))={\sum }_{k=N_1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)$,$a_2(t)\exp (j{\varphi }_2(t))={\sum }_{k=N_2}^{N_1-1}{c}_k\exp (jk{\omega }_{0}t)$,…,$a_M(t)\exp (j{\varphi }_M(t))={\sum }_{k=1}^{N_{M-1}-1}{c}_k\exp (jk{\omega }_{0}t)$ and the lower and upper limits of sum in (3.6) would change to $k=N_i$ to $(N_{i-1}-1)$, respectively , with $N_0=\infty$ and $N_M=1$. Here, we start with $(N_{i-1}-1)$, decrease and select minimum value of $N_i$ such that $1\le N_i\le (N_{i-1}-1)$ and (3.7) is satisfied for $i=1,...,M$.

同样，对AFIBFs进行反向搜索(从高频扫描到低频扫描(HTL-FS))，可以得到$a_1(t)\exp (j{\varphi }_1(t))={\sum }_{k=N_1}^{\infty }{c}_k\exp (jk{\omega }_{0}t)$,$a_2(t)\exp (j{\varphi }_2(t))={\sum }_{k=N_2}^{N_1-1}{c}_k\exp (jk{\omega }_{0}t)$,…,$a_M(t)\exp (j{\varphi }_M(t))={\sum }_{k=1}^{N_{M-1}-1}{c}_k\exp (jk{\omega }_{0}t)$ ，(3.6)中求和的下限和上限分别变为$k=N_i$ 到 $(N_{i-1}-1)$，其中 $N_0=\infty$ ， $N_M=1$。这里，我们从 $(N_{i-1}-1)$开始，减小并选择最小的 $N_i$ 值，使得 $1\le N_i\le (N_{i-1}-1)$ 并且在$i=1,...,M$ 时满足式(3.7) 。

We observe that (3.5) has precisely the form which in the literature [3] is termed as a generalized Fourier expansion. Moreover, this representation is complete, orthogonal, local, adaptive and purely Fourier based. Thus, we have presented a generalized signal specific Fourier expansion of a time series in (3.5) by the Fourier method itself. The variable amplitude and the IF have improved the efficiency of the expansion by expanding the signal into finite number of analytic FIBFs in (3.5), and enabled the expansion to accommodate non-stationary data. Thus, we have obtained a variable amplitude and frequency representation, whereas the classical Fourier expansion provides constant amplitudes and fixed-frequencies representation.

我们观察到（3.5）的形式在文献[3]中被称为广义傅里叶展开式。此外，这种表示是完全的、正交的、局部的、自适应的和纯粹基于傅立叶的。因此，我们在（3.5）中用Fourier方法给出了时间序列的广义信号特定Fourier展开式。在(3.5)中，可变幅度和瞬时频率（IF）通过将信号扩展到有限数量的解析FIBFs来提高扩展的效率，并使扩展能够适应非平稳数据。因此，我们得到了一个可变振幅和频率表示，而经典的傅立叶展开提供了恒定振幅和固定频率表示。

For each of the FIBFs, the amplitude $a_i(t)$ and frequency $f_i(t)$ are functions of time; therefore, we define the three-dimensional { t , f i ( t ) , a i ( t ) } \{t,f_i(t),a_i(t)\} {t,fi​(t),ai​(t)} TFD of amplitude as the Fourier–Hilbert spectrum (FHS) H($f,t$). The marginal Hilbert spectrum (MHS), derived from Hilbert spectrum, is defined in [3]. Similarly , we define the marginal FHS (MFHS) from the FHS as $H(f)={\int }_0^{T_0}H(f,t)dt$. The MFHS offers a measure of total IA (or energy) contribution from each frequency in an average sense. The frequency in either $H(f,t)$ or $H(f)$ has a different meaning from the Fourier spectral analysis [3]. The presence of energy at each frequency in MFHS $H(f)$ means that, in the total duration of the signal, there is a higher likelihood for such a wave (FIBF) to have appeared locally . The frequency in the MFHS indicates only the likelihood that an oscillation with such a frequency exists [3]. The exact occurrence time of that oscillation is given in the full FHS. The instantaneous energy (IE) density , which can be used to measure the variation of energy with time, can be defined as $E(t)={\int }_0^{f_M}{H}^2(f,t)df$, where $f_M$ is the maximum frequency of signal. From (3.1), we obtain the energy of signal $x(t)$ (or power of signal $x_{T_0}(t)$) by the Parseval’s theorem as $E_x=a_0^2+\frac{1}{2}{\sum }_{i=1}^{\infty }[a_n^2+b_n^2]$, and from (3.4) energy of the analytic signal (or power of signal $z_{T_0}(t)$) as $E_z={\sum }_{n=1}^{\infty }[a_n^2+b_n^2]$. Therefore, relation between $E_x$ and $E_z$ is given by $E_x=a_0^2+E_z/2$. Hence, the energy of zero mean signal is half of the energy of its analytic signal.

对于每个 FIBFs，振幅 $a_i(t)$和频率 $f_i(t)$ 是时间的函数；因此，我们把振幅的三维 { t , f i ( t ) , a i ( t ) } \{t,f_i(t),a_i(t)\} {t,fi​(t),ai​(t)} TFD 定义为傅里叶-希尔伯特谱(FHS) H($f,t$)。由希尔伯特谱导出的边际希尔伯特谱(MHS)在[3]中定义。同样，我们把FHS的边际FHS (MFHS)定义为 $H(f)={\int }_0^{T_0}H(f,t)dt$。MFHS提供了平均意义上每个频率的总IA（或能量）贡献的度量。 $H(f,t)$ 或者$H(f)$ 中的频率与傅里叶谱分析的含义不同[3]。MFHS $H(f)$ 表示的每个频率处能量的存在意味着，在信号的总持续时间内，这种波（FIBF）局部出现的可能性更高。MFHS中的频率仅表示存在这种频率的振荡的可能性[3]。振荡的确切发生时间在完整的FHS中给出。可以用来测量能量随时间变化的瞬时能量(IE)密度，可以定义为 $E(t)={\int }_0^{f_M}{H}^2(f,t)df$，其中 $f_M$ 是信号的最大频率。从(3.1)出发， 我们可以通过Parseval定理获得信号 $x(t)$ 的能量(或信号 $x_{T_0}(t)$ 的能量) 即$E_x=a_0^2+\frac{1}{2}{\sum }_{i=1}^{\infty }[a_n^2+b_n^2]$，并且从 (3.4)获得解析信号的能量 (即 $z_{T_0}(t)$ 信号的能量) 为 $E_z={\sum }_{n=1}^{\infty }[a_n^2+b_n^2]$。因此， $E_x$ 和 $E_z$ 的关系由 $E_x=a_0^2+E_z/2$ 给出。因此，零均值信号的能量是解析信号能量的一半。

We observe that the FDM follows a nonlinear time-invariant (NTI) system model, and hence it is a nonlinear time-invariant ZPF operations to decompose a signal into a set of FIBFs. The FDM is nonlinear because it does not follow the principle of superposition, that is there exist $n$ signals { x l } l = 1 n \{x_l\}^n_{l=1} {xl​}l=1n​: $\text{FDM}[{\sum }_{l=1}^n{x}_l]\ne {\sum }_{l=1}^n\text{FDM}[x_l]$. We present the following counterexample to prove that the FDM is a nonlinear system model.

我们观察到FDM遵循非线性时不变（NTI）系统模型，因此将信号分解为一组FIBFs是一种非线性时不变的ZPF操作。 FDM是非线性的，因为它不遵循叠加原理，即存在 $n$ 个信号 { x l } l = 1 n \{x_l\}^n_{l=1} {xl​}l=1n​: $\text{FDM}[{\sum }_{l=1}^n{x}_l]\ne {\sum }_{l=1}^n\text{FDM}[x_l]$。我们给出以下反例来证明FDM是一个非线性系统模型。

Example. Let $x_1(t)=\sin (10\pi t)$ and $x_2(t)=\sin (100\pi t)$ in time interval $0\le t\le 1$. Clearly , both signal $x_1(t)$ and $x_2(t)$ are FIBFs, $\text{FDM}[x_1(t)]=x_1(t)$ and $\text{FDM}[x_2(t)]=x_2(t)$, thus $\text{FDM}[x_1(t)]+\text{FDM}[x_2(t)]=x_1(t)+x_2(t)$. However, it is very easy to verify , by the FDM algorithms, that FDM [ x 1 ( t ) + x 2 ( t ) ] = { x 1 ( t ) , x 2 ( t ) } \text{FDM}[x_1(t) + x_2(t)]=\{x_1(t),x_2(t)\} FDM[x1​(t)+x2​(t)]={x1​(t),x2​(t)} which implies that $\text{FDM}[x_1(t)+x_2(t)]\ne \text{FDM}[x_1(t)]+\text{FDM}[x_2(t)]$. It is also easy to show that FDM is a time-invariant system model as $\text{FDM}[x_1(t-\tau )]=x_1(t-\tau ),\text{FDM}[x_2(t-\tau )]=x_2(t-\tau )$ and FDM [ x 1 ( t − τ ) + x 2 ( t − τ ) ] = { x 1 ( t − τ ) , x 2 ( t − τ ) } \text{FDM}[x_1(t − \tau) + x_2(t − \tau)]=\{x_1(t − τ),x_2(t − \tau )\} FDM[x1​(t−τ)+x2​(t−τ)]={x1​(t−τ),x2​(t−τ)}.

例子. 设 $x_1(t)=\sin (10\pi t)$ ，$x_2(t)=\sin (100\pi t)$ 在时间间隔$0\le t\le 1$上。显然，信号 $x_1(t)$ 和 $x_2(t)$ 是 FIBFs， $\text{FDM}[x_1(t)]=x_1(t)$ ，$\text{FDM}[x_2(t)]=x_2(t)$，因此 $\text{FDM}[x_1(t)]+\text{FDM}[x_2(t)]=x_1(t)+x_2(t)$。 然而，通过FDM算法很容易验证， FDM [ x 1 ( t ) + x 2 ( t ) ] = { x 1 ( t ) , x 2 ( t ) } \text{FDM}[x_1(t) + x_2(t)]=\{x_1(t),x_2(t)\} FDM[x1​(t)+x2​(t)]={x1​(t),x2​(t)} 这意味着 $\text{FDM}[x_1(t)+x_2(t)]\ne \text{FDM}[x_1(t)]+\text{FDM}[x_2(t)]$。这也很容易表明，FDM是一个时不变的系统模型，即 $\text{FDM}[x_1(t-\tau )]=x_1(t-\tau )，\text{FDM}[x_2(t-\tau )]=x_2(t-\tau )$ 且 FDM [ x 1 ( t − τ ) + x 2 ( t − τ ) ] = { x 1 ( t − τ ) , x 2 ( t − τ ) } \text{FDM}[x_1(t − \tau) + x_2(t − \tau)]=\{x_1(t − τ),x_2(t − \tau )\} FDM[x1​(t−τ)+x2​(t−τ)]={x1​(t−τ),x2​(t−τ)}。

(b) Discrete Fourier decomposition method
（b）离散傅里叶分解方法

In practice, the continuous time signals are, generally , discretized for further processing by a computing device; therefore, we present the FDM for discrete signal. Let, $x[n]$, be a discrete signal of length $N$. Using the discrete FT (DFT), we can write $x[n]$ as
$$x[n]=\sum _{k=0}^{N-1}X[k]\exp (\frac{j2\pi kn}{N})(3.8)$$
where $X[k]=(1/N){\sum }_{n=0}^{N-1}x[n]\exp (-j2\pi kn/N)$ is the DFT of the signal $x[n]$. Let $N$ be an even number (we can proceed in the similar fashion when $N$ is an odd number), then X[0] and X[N/2] are real numbers; and we can write $x[n]$ as
$$x[n]=X[0]+\sum _{k=1}^{N/2-1}X[k]\exp (\frac{j2\pi kn}{N})+X[\frac{N}{2}]\exp (j\pi n)+\sum _{k=N/2+1}^{N-1}X[k]\exp (\frac{j2\pi kn}{N})(3.9)$$

在实践中，连续时间信号通常被离散化，以便由计算设备进一步处理；因此，我们提出了用于离散信号的FDM。设 $x[n]$ 是长度为 $N$的离散信号。利用离散 FT (DFT)，可以将 $x[n]$ 写成
$$x[n]=\sum _{k=0}^{N-1}X[k]\exp (\frac{j2\pi kn}{N})(3.8)$$
其中 $X[k]=(1/N){\sum }_{n=0}^{N-1}x[n]\exp (-j2\pi kn/N)$ 是信号 $x[n]$ 的 DFT 。设$N$为偶数(当$N$为奇数时，我们可以用类似的方式进行)，则X[0]和X[N/2]为实数；我们可以将 $x[n]$ 写成
$$x[n]=X[0]+\sum _{k=1}^{N/2-1}X[k]\exp (\frac{j2\pi kn}{N})+X[\frac{N}{2}]\exp (j\pi n)+\sum _{k=N/2+1}^{N-1}X[k]\exp (\frac{j2\pi kn}{N})(3.9)$$

Since $x[n]$ is real, therefore, $z_1[n]\stackrel{\Delta }{=}{\sum }_{k=1}^{N/2-1}X[k]\exp (j2\pi kn/N)$ is complex conjugate of $z_2[n]\stackrel{\Delta }{=}{\sum }_{k=N/2+1}^{N-1}X[k]\exp (j2\pi kn/N)$ and we can write (3.9) as
x [ n ] = X [ 0 ] + 2 Re { z 1 [ n ] } + X [ N 2 ] ( − 1 ) n ( 3.10 ) x[n]=X[0]+2 \text{Re} \{z_1[n]\} + X[\frac N2] (-1)^n \quad (3.10) x[n]=X[0]+2Re{z1​[n]}+X[2N​](−1)n(3.10)

where Re{$z_1[n]$} denote the real part of $z_1[n]$. Now, we write analytic signal $z_1[n]$ as
$$\sum _{k=1}^{N/2-1}X[k]\exp (\frac{j2\pi kn}{N})=\sum _{i=1}^M{a}_i[n]\exp (j{\varphi }_i[n]),(3.11)$$
where, in forward search (LTH-FS) of AFIBFs, we obtain $a_1[n]\exp (j{\varphi }_1[n])={\sum }_{k=1}^{N_1}X[k]\exp (j2\pi kn/N)$,$a_2[n]\exp (j{\varphi }_2[n])={\sum }_{k=(N_1+1)}^{N_2}X[k]\exp (j2\pi kn/N)$,…,$a_M[n]\exp (j{\varphi }_M[n])={\sum }_{k=(N_{M-1}+1)}^{N/2-1}X[k]\exp (j2\pi kn/N)$ and, in general,
$$a_i[n]\exp (j{\varphi }_i[n])=\sum _{k=N_{i-1}+1}^{N_i}X[k]\exp (\frac{j2\pi kn}{N}),(3.12)$$
with $N_0=0$ and $N_M=(N/2-1)$. In order to obtain a minimum number of FIBFs in LTH-FS, for each i, we scan from $(N_{i-1}+1)$ to $(N/2-1)$, obtain a maximum value of $N_i$ such that $(N_{i-1}+1)\le N_i\le (N/2-1)$ and phase ${\varphi }_i[n]$ is a monotonically increasing function, that is an estimate of the discrete IF
$${\omega }_i[n]=({\varphi }_i[n+1]-{\varphi }_i[n])\ge 0,\forall n(3.13)$$
or
$${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n(3.14)$$
and amplitude $a_i[n]\ge 0,\forall n$ and for $i=1,...,M$. We observe that such a decomposition always exists. An estimation of the discrete IF by (3.14) has the following advantages: (i) it is unbiased and has zero group delay (GD) for linear frequency modulated signals [17], and (ii) it corresponds to the first moment in frequency of a number of TFDs [29–31].

既然 $x[n]$ 是实数，那么 $z_1[n]\stackrel{\Delta }{=}{\sum }_{k=1}^{N/2-1}X[k]\exp (j2\pi kn/N)$ 是 $z_2[n]\stackrel{\Delta }{=}{\sum }_{k=N/2+1}^{N-1}X[k]\exp (j2\pi kn/N)$ 的复共轭， 式 (3.9) 可以变为
x [ n ] = X [ 0 ] + 2 Re { z 1 [ n ] } + X [ N 2 ] ( − 1 ) n ( 3.10 ) x[n]=X[0]+2 \text{Re} \{z_1[n]\} + X[\frac N2] (-1)^n \quad (3.10) x[n]=X[0]+2Re{z1​[n]}+X[2N​](−1)n(3.10)

其中 Re{$z_1[n]$} 代表 $z_1[n]$ 的实部。现在，我们写出解析信号 $z_1[n]$ 为
$$\sum _{k=1}^{N/2-1}X[k]\exp (\frac{j2\pi kn}{N})=\sum _{i=1}^M{a}_i[n]\exp (j{\varphi }_i[n]),(3.11)$$
其中，在AFIBFs的前向搜索(LTH-FS)中，可以得到 $a_1[n]\exp (j{\varphi }_1[n])={\sum }_{k=1}^{N_1}X[k]\exp (j2\pi kn/N)$,$a_2[n]\exp (j{\varphi }_2[n])={\sum }_{k=(N_1+1)}^{N_2}X[k]\exp (j2\pi kn/N)$,…,$a_M[n]\exp (j{\varphi }_M[n])={\sum }_{k=(N_{M-1}+1)}^{N/2-1}X[k]\exp (j2\pi kn/N)$ 总的来说
$$a_i[n]\exp (j{\varphi }_i[n])=\sum _{k=N_{i-1}+1}^{N_i}X[k]\exp (\frac{j2\pi kn}{N}),(3.12)$$
其中 $N_0=0$ ， $N_M=(N/2-1)$。为了在 LTH-FS中获得最小数目的 FIBFs，对每一个 i，我们从$(N_{i-1}+1)$ 扫描到 $(N/2-1)$，获得一个最大的 $N_i$ 即 $(N_{i-1}+1)\le N_i\le (N/2-1)$ 并且相位 ${\varphi }_i[n]$ 是一个单调递增的函数，是离散IF（瞬时频率）的一个估计，即
$${\omega }_i[n]=({\varphi }_i[n+1]-{\varphi }_i[n])\ge 0,\forall n(3.13)$$
或者
$${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n(3.14)$$
并且当$i=1,...,M$，振幅 $a_i[n]\ge 0,\forall n$。 我们观察到这种分解总是存在的。用(3.14)估计离散IF（瞬时频率）有以下优点:(1)对线性调频信号[17]是无偏的，具有零群延迟(GD)，(2)对应多个TFDs的频率一阶矩[29-31]。

Similarly , in HTL-FS for FIBFs, we obtain $a_1[n]\exp (j{\varphi }_1[n])={\sum }_{k=N_1}^{N/2-1}X[k]\exp (j2\pi kn/N)$,$a_2[n]\exp (j{\varphi }_2[n])={\sum }_{k=N_2}^{N_1-1}X[k]\exp (j2\pi kn/N)$,…,$a_M[n]\exp (j{\varphi }_M[n])={\sum }_{k=1}^{(N_{M-1}-1)}X[k]\exp (j2\pi kn/N)$. The lower and upper limits of the sum in (3.12) would change to $k=N_i$ to $(N_{i-1}-1)$, respectively , with $N_0=N/2,N_M=1$. In this case, for each i, we scan from $(N_{i-1}-1)$ to 1, obtain the minimum value of $N_i$ such that $1\le N_i\le (N_{i-1}-1)$ and phase ${\varphi }_i[n]$ is a monotonically increasing function.

类似的，用 HTL-FS 获得 FIBFs，我们得到了 $a_1[n]\exp (j{\varphi }_1[n])={\sum }_{k=N_1}^{N/2-1}X[k]\exp (j2\pi kn/N)$,$a_2[n]\exp (j{\varphi }_2[n])={\sum }_{k=N_2}^{N_1-1}X[k]\exp (j2\pi kn/N)$,…,$a_M[n]\exp (j{\varphi }_M[n])={\sum }_{k=1}^{(N_{M-1}-1)}X[k]\exp (j2\pi kn/N)$。式(3.12) 中求和的上下限分别为 $k=N_i$ 和 $(N_{i-1}-1)$，其中 $N_0=N/2,N_M=1$。在这种情况下，对于每一个 i，我们从 $(N_{i-1}-1)$ 扫描到 1，获得一个最小的 $N_i$ 值即 $1\le N_i\le (N_{i-1}-1)$ 并且相位 ${\varphi }_i[n]$是单调递增函数。

It is interesting to observe that FDM provides two different views of TFE distribution, namely low to high frequency and high to low frequency views of a signal. Depending on the signal, both views may be the same or they may reveal two different kinds of features of the signal. The DT-FDM is summarized in tables 1 and 2, its implementation is presented in algorithms 1 and 2. The FDM with FT and FDM with discrete time FT (DTFT) are summarized in the electronic supplementary material.

有趣的是，FDM提供了TFE分布的两种不同视图，即信号的低频到高频和高频到低频视图。根据信号的不同，两个视图可能是相同的，也可能显示信号的两种不同的特征。表1和表2总结了DT-FDM算法，算法1和算法2介绍了它的实现。电子补充材料中总结了由FT得到的的FDM和由离散时间傅里叶变换（DTFT）得到的FDM。

Table 1. The DT-FDM algorithmic summary (LTH-FS) to obtain AFIBFs, for $i=1,...,M$ with $N_0=0,N_M=(N/2-1)$ (or $N_M=(N-1)/2$ if $N$ is odd).

STEP 1. Obtain X [ k ] = F F T { x [ n ] } X[k]=FFT \{ x[n] \} X[k]=FFT{x[n]}.
STEP 2. Set ${\text{AFIBF}}_i={\sum }_{k=(N_{i-1}+1)}^{N_i}X[k]\exp (j2\pi kn/N)=a_i[n]\exp (j{\varphi }_i[n])$, obtain maximum value of $N_i$ such that $(N_{i-1}+1)\le N_i\le (\frac{N}{2}-1)$ and phase ${\varphi }_i[n]$ of AFIBFi is a monotonically increasing function, that is, ${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n$. A simple implementation of this is presented in algorithm1.

表格 1. DT-FDM算法汇总(LTH-FS)以得到AFIBFs，对于 $i=1,...,M$ 有 $N_0=0,N_M=(N/2-1)$ (或者如果 $N$ 是奇数的话有 $N_M=(N-1)/2$ )。

步骤 1. 获得 X [ k ] = F F T { x [ n ] } X[k]=FFT \{ x[n] \} X[k]=FFT{x[n]}.
步骤 2. 设 ${\text{AFIBF}}_i={\sum }_{k=(N_{i-1}+1)}^{N_i}X[k]\exp (j2\pi kn/N)=a_i[n]\exp (j{\varphi }_i[n])$, 获得 $N_i$ 的最大值即 $(N_{i-1}+1)\le N_i\le (\frac{N}{2}-1)$ ，AFIBFi 的相位 ${\varphi }_i[n]$ 是一个单调递增的函数，即 ${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n$。算法1给出了一个简单的实现。

Table 2. The DT-FDM algorithmic summary (HTL-FS) to obtain AFIBFs, for $i=1,...,M$ with $N_0=N/2$ (or $N_0=(N+1)/2$ if $N$ is odd) and $N_M=1$.

STEP 1. Obtain X [ k ] = F F T { x [ n ] } X[k]=FFT \{ x[n] \} X[k]=FFT{x[n]}.
STEP 2. Set ${\text{AFIBF}}_i={\sum }_{k=N_i}^{N_{i-1}+1}X[k]\exp (j2\pi kn/N)=a_i[n]\exp (j{\varphi }_i[n])$, obtain minimum value of $N_i$ such that $1\le N_i\le (N_{i-1}+1)$ and phase ${\varphi }_i[n]$ of AFIBFi is a monotonically increasing function, that is, ${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n$. A simple implementation of this is presented in algorithm2.

表格 2. DT-FDM 算法总结 (HTL-FS) 以获得 AFIBFs，对 $i=1,...,M$ 有 $N_0=N/2$ (或者如果 $N$ 是奇数的话 $N_0=(N+1)/2$)，且$N_M=1$.

步骤 1. 获得 X [ k ] = F F T { x [ n ] } X[k]=FFT \{ x[n] \} X[k]=FFT{x[n]}.
步骤 2. 设 ${\text{AFIBF}}_i={\sum }_{k=N_i}^{N_{i-1}+1}X[k]\exp (j2\pi kn/N)=a_i[n]\exp (j{\varphi }_i[n])$, 获得 $N_i$ 的最小值即 $1\le N_i\le (N_{i-1}+1)$ ，AFIBFi的相位 ${\varphi }_i[n]$ 是一个单调递增函数，即 ${\omega }_i[n]=(\frac{{\varphi }_i[n+1]-{\varphi }_i[n-1]}{2})\ge 0,\forall n$。算法2给出了一个简单的实现。