 Research
 Open access
 Published:
Widely linear Markov signals
EURASIP Journal on Advances in Signal Processing volume 2012, Article number: 256 (2012)
Abstract
The insufficiency to guarantee the existence of a statespace representation of the classical widesense Markov condition for improper complexvalued signals is shown and a generalization is suggested. New characterizations for widesense Markov signals which are based either on secondorder properties or on statespace representations are studied in a widely linear setting. Moreover, the correlation structure of such signals is revealed and interesting results on modeling in both the forwards and backwards time directions are proved. As an application we give some recursive estimation algorithms obtained from the Kalman filter. The performance of the proposed results is illustrated in a numerical example in the areas of estimation and simulation.
1 Introduction
Markov signals are characterized by the condition that future development of these signals depends only on current states and not their history up to that time. In general, Markov processes are easier to model and analyze, and they do include interesting applications. Among others, estimation and detection are areas of signal processing where this kind of process has provided efficient solutions (see, e.g., [1, 2]). NonMarkov processes in which the future state of a process depends on its whole history are generally harder to analyze mathematically [3]. In linear minimummean square error (MMSE) estimation theory, when the processes under consideration are not Gaussian, the classes of stochastic processes which are of practical importance are widesense Markov (WSM) processes. The concept of WSM signal is easier to check than the condition of (strictly) Markov since it involves only secondorder characteristics [4]. In general, WSM processes (with the exception of Gaussian processes) are not Markov in the strict sense. The equivalence between the WSM condition and the statespace representation for the signal is really what makes WSM signals especially attractive in signal processing [1].
Widely linear (WL) processing is an emerging research area in the complexvalued signal analysis which gives significant performance gains with respect to strictly linear (SL) processing (excellent account of the topic and the literature can be found in [5, 6]). It has proved to be a more useful approach than SL processing since complexvalued random signals are in general improper (i.e., they are correlated with their complex conjugates). Thus, the improper nature of most signals forces us to consider the socalled augmented statistics to entirely describe their secondorder properties. Using augmented statistics means incorporating in the analysis the information supplied by the complex conjugate of the signal and examining properties of both the correlation and complementary correlation functions. SL processing operates ignoring this last function. Some areas of signal processing in which the treatment of improper signals by using a WL processing has proved to be beneficial are estimation [5–11], detection [12], modeling [8], and simulation [13].
A general characteristic of the articles devoted to studying WSM complexvalued signals is that they use a SL processing approach (see e.g., [1, 14–16]). We will show by means of simple examples that the classical definition and the associated characterizations of WSM signals are incorrect for improper signals. The examples then motivate the extension of the concept of WSM signal to a WL setting and the study of new characterizations. Specifically, we introduce the concept of widely linear Markov (WLM) signals and we give different characterizations based either on secondorder properties or on statespace representations from a WL processing point of view. The analysis is performed in both the forwards and backwards directions of time. We also provide a way to check the WLM condition, similar to the wellknown triangular property, based on augmented statistics and determine the correlation structure of WLM signals. The modeling part is the focus of this article. In this sense, WL forwards and backwards Markovian representations are suggested, the interrelation between them is studied and the connection with the WL autoregressive representations defined in [8] is established. These Markovian representations also become a starting point for the application of different recursive estimation algorithms. Thus, the application of the Kalman filter on the forwards and backwards representations yields different WL prediction, filtering and smoothing algorithms. The point, which is illustrated in an example, is that besides the wellknown performance gain of the WL approach we also get more realistic results in simulation and modeling.
The article is organized as follows. In Section 2, we present some background material on complexvalued Markov signals, illustrate the incapacity of the usual WSM condition in order to characterize the statespace representation for improper signals and suggest the concept of WLM signal. Some preliminary characterizations are also given. Section 3 studies the correlation structure of WLM signals. In Section 4, we discuss the modeling problem for WLM signals and analyze the stationary case. The estimation problem is treated in Section 5. We apply our results in the fields of signal simulation and estimation by considering a numerical example in Section 6. A Section of conclusions ends the article. To preserve continuity in our presentation, all proofs are deferred to an Appendix Appendix 1.
2 Preliminaries
In this section, we give the main definitions, notation and auxiliary results. We also present two examples which motivate the necessity of the new concept introduced.
Bold capital letters will be used to refer to matrices and bold lowercase letters will be used to refer to vectors. The row j of any matrix A(·) will be denoted by A_{[j]}(·), the nvector of zeros by 0_{ n } and the n×mmatrix of zeros by 0_{n×m}. Furthermore, the superscripts ∗, T, and H represent the complex conjugate, transpose, and complex transpose, respectively.
Let \{{x}_{t},t\in \mathbb{Z}\} be a zeromean complex random signal with correlation function r(t,s)=E\left[{x}_{t}{x}_{s}^{\ast}\right] and complementary correlation function c(t s)=E x_{ t }x_{ s }. Most of the results in this article are valid for nonstationary signals. However, for some of them the stationary condition is necessary. The signal x_{ t }is said to be of secondorder widesense stationary (SOS) if the functions r(t s) and c(t s) depend on t − s. A zeromean stochastic process w_{ t } is called a doubly white noise if E\left[{w}_{t}{w}_{s}^{\ast}\right]={e}_{1}\delta (ts) and E w_{ t }w_{ s }=e_{2}δ(t − s) with e_{2} ≤ e_{1}(see [8] for a complete study of their characteristics). The linear MMSE estimator of x_{ t } based on the set of observations \{{x}_{{t}_{1}},{x}_{{t}_{2}},\dots ,{x}_{{t}_{m}}\} will be denoted by \widehat{x}\left(t\right{t}_{1},{t}_{2},\dots ,{t}_{m}) and we will refer to it as the SL estimator.
The Markov condition on a signal \{{x}_{t},t\in \mathbb{Z}\} establishes the following identity for the conditional probability:
for all x and t > t_{1} > ⋯ > t_{ m }. Doob [4] introduced a weaker concept based on the SL estimator which has received great attention in the literature (e.g., [1, 14–16]). A signal x_{ t } is called WSM if \widehat{x}\left(t\right\tau \le s)=\widehat{x}(t\lefts\right) for any s < t. Such signals have remarkable properties. For example, Beutler [14] showed that a signal x_{ t } is WSM if, and only if, the function \stackrel{\u0304}{k}(t,s)=r(t,s){r}^{1}(s,s) has the triangular property, i.e.,
Another characterization in terms of socalled Markovian statespace models can be found in [1]. They showed that a signal {x_{ t }t ≥ 0} is WSM if, and only if, it has a statespace model of the form
where u_{ t } is a white noise uncorrelated with x_{0}. Doob’s definition was later generalized in [16] in the following sense: x_{ t } is a WSM signal of order n ≥ 1 if \widehat{x}\left(t\right\tau \le s)=\widehat{x}(ts,s1,\dots ,sn+1) for any s < t. The authors also studied the secondorder properties of such signals.
All these studies have a common characteristic: the information supplied by the complementary correlation function is ignored, i.e., the results are derived assuming implicitly that the signal is proper (c(t,s)=0). As noted above, nowadays, the research activity in the field of the complexvalued signal is more and more focused on the better performing and less familiar WL processing. In this setting the SL MMSE estimator is replaced by the WL MMSE estimator, denoted by {\widehat{x}}^{\mathit{WL}}\left(t\right{t}_{1},{t}_{2},\dots ,{t}_{m}), which uses the information of the augmented vector of observations {[{x}_{{t}_{1}},{x}_{{t}_{1}}^{\ast},{x}_{{t}_{2}},{x}_{{t}_{2}}^{\ast},\dots ,{x}_{{t}_{m}},{x}_{{t}_{m}}^{\ast}]}^{T}. The immediate question that arises is whether the classical concept of WSM signals remains valid in the WL processing approach. The following two examples give us the answer.
Example 1. Consider a signal {x_{ t },t ≥ 0}with correlation functionr(t,s)=\frac{1}{2}({e}^{3ts}+{e}^{ts})and complementary correlation functionc(t,s)=\frac{1}{2}({e}^{3ts}{e}^{ts}).It is easy to check that r(t,s) does not satisfy the triangular property (1) and then, the signal cannot be modeled by a representation of the form (2). However, as we will show later, it is possible to find a statespace representation for such a signal given by (26). Thus, the classical WSM condition is clearly insufficient in the improper case to find a statespace representation for the signal involved.
Example 2. Assume that {x_{ t },1 ≤ t ≤ 100}is a signal with correlation and complementary correlation functions given by r(t,s) = (t/100 + 1)^{1/6}(s/100)^{4} and c(t,s) = j(s/100)^{4}, for s ≤ t,respectively, withj=\sqrt{1}.Here, the triangular property (1) holds and then x_{ t }has the representation
with x_{ t }uncorrelated with u_{ t }.However, this model presents two important shortcomings in the WL processing framework: the noise u_{ t }is correlated with{x}_{t}^{\ast}and the information supplied by c(t,s)is ignored. Both problems can be avoided by considering a more competitive model for x_{ t }obtained with the additional information of{x}_{t}^{\ast}.In fact, we can write an alternative statespace representation for x_{ t }given by (27). An exhaustive study about the superiority of (27) against (3) is presented in Section 6.
From these two simple examples we extract the following consequences: the classical definition of a WSM signal must be extended to deal with improper signals, this new concept must be characterized adequately to avoid the drawback shown in Example 1 and new results about modeling are necessary to exploit the information available in both x_{ t } and {x}_{t}^{\ast} thus attaining better models for the signal as illustrated in Example 2. Next, we introduce such a definition in a WL processing setting.
Definition 1. A complexvalued signal \{{x}_{t},t\in \mathbb{Z}\} is said to be WLM of order n ≥ 1, briefly a WLM(n) signal, if the following condition holds
for any s < t.
Notice that this concept extends both the classical notion of WSM introduced by Doob in [4] and the later generalization given in [16].
In the rest of the section, we provide different characterizations of WLM(n) signals. For that, we need to introduce some additional notation. Denote the augmented forwards vector of order n ≥ 1 of x_{ t } as the 2nvector
and its correlation function by \mathit{R}(t,s)=E\left[{\mathbf{x}}_{t}{\mathbf{x}}_{s}^{H}\right]. From now on, we assume that det {R_{ t }} ≠ 0 with R_{ t }:=R(t,t). Moreover, we define the normalized correlation function as
Similarly, we define the augmented backwards vector of order n ≥ 1 of x_{ t } as the 2nvector
The following results establish the relation between the signals x_{ t } and their augmented forwards and backwards versions. We start first with the augmented forwards vector and we give a test similar to (1) for a signal being WLM(n).
Theorem 1. The following statements are equivalent:

1.
\{{x}_{t},t\in \mathbb{Z}\} is a WLM(n) signal.

2.
For s < t, the WL MMSE estimator of x _{ t } on the basis of the set \{{\mathbf{x}}_{\tau},{\mathbf{x}}_{\tau}^{\ast},\tau \le s\} is of the form
{\widehat{\mathbf{x}}}^{\mathit{WL}}\left(t\right\tau \le s)=\mathit{K}(t,s){\mathbf{x}}_{s}(5)

3.
For t ≥ τ ≥ s,
\mathit{K}(t,s)=\mathit{K}(t,\tau )\mathit{K}(\tau ,s)(6)
Now, we suggest a characterization based on the augmented backwards vector. This result also shows the independence from the time direction of the Markov condition.
Theorem 2. The following statements are equivalent:

1.
\{{x}_{t},t\in \mathbb{Z}\} is a WLM(n) signal.

2.
{\widehat{x}}^{\mathit{WL}}\left(t\right\tau \ge s)={\widehat{x}}^{\mathit{WL}}(ts,s+1,\dots ,s+n1) for any s > t.

3.
For s > t, the WL MMSE estimator of {\mathbf{x}}_{t}^{b} on the basis of the set \{{\mathbf{x}}_{\tau}^{b},{\mathbf{x}}_{\tau}^{{b}^{\ast}},\tau \ge s\} is of the form
{\widehat{\mathbf{x}}}^{{b}^{\mathit{WL}}}\left(t\right\tau \ge s)=\mathit{K}(t+n1,s+n1){\mathbf{x}}_{s}^{b}(7)
3 Correlation structure of WLM(n) signals
In this section, the secondorder properties of a WLM(n) signal \{{x}_{t},t\in \mathbb{Z}\} are analyzed. Specifically, we study the structure of the matrices R(t,s), K(t,s), R_{ t }, and K_{ t }:=K(t + 1,t).
Proposition 1. 1. The following relations hold:
2. The matrix K _{ t } is of the form
where k_{i,t} = k_{ i }(t + 1,t) for i = 1,…,2n and k_{ i }(t + 1,t) is defined in (28).
3. The matrices R(t,s) and K _{ t } satisfy the recursive equation
which has the solution
Moreover,
where Q_{ t } is a 2n×2nmatrix of the form
with
where a_{1,t} are real positive numbers and A_{ t } is nonnegative definite.
4 Modeling of WLM(n) signals
We aim to provide different ways of modeling for WLM(n) signals. The connection between stationary WLM(n) signals and the autoregressive representations defined in [8] is also established. First, we present a new characterization in which the equivalence between a WLM signal of order n and their forwards and backwards representations is given. Such representations show that a WLM(n) signal depends only on the n preceding or subsequent states and their conjugates.
Theorem 3. A signal {x_{ t },0 ≤ t ≤ m} is a WLM(n) if, and only if, it has the forwards and backwards representations
where k_{ t }, {\mathit{k}}_{t}^{b} are 2nvectors, and w_{ t }, {w}_{t}^{b} are doubly white noises such that
Now we state a parallel result to the classical one established for stationary WSM processes and autoregressive representations [16].
Corollary 1. If {x_{ t },0 ≤ t ≤ m} is a SOS WLM(n) signal, then x_{ t } is the solution of the WL system defined in [8]
where {g}_{1,i},{g}_{2,i}\in \mathbb{C}, i = 1,…,n − 1, and w_{ t } is a doubly white noise such that E\left[{w}_{t}{w}_{t}^{\ast}\right]={a}_{1} and E[w_{ t }w_{ t }] = a_{2}.
We summarize the previous results in the following steps which provides forwards and backwards models for a WLM(n) signal:

Step 1: Define the 2nvector k_{ t } such that {\mathit{k}}_{t}^{T} coincides with the first row of the matrix
{\mathit{K}}_{t}:=\mathit{R}(t+1,t){\mathit{R}}_{t}^{1}(21)Similarly, we define the 2nvector {\mathit{k}}_{t+1}^{b} such that {{\mathit{k}}^{b}}_{t+1}^{T} is equal to the 2n − 1 row of the matrix
{\mathit{K}}_{t+1}^{b}:=\mathit{K}(t+n1,t+n)=\mathit{R}(t+n1,t+n){\mathit{R}}_{t+n}^{1}(22) 
Step 2: Consider the matrices
\begin{array}{l}{\mathit{Q}}_{t}={\mathit{R}}_{t+1}{\mathit{K}}_{t}{\mathit{R}}_{t}{\mathit{K}}_{t}^{H}\phantom{\rule{2em}{0ex}}\end{array}(23)\begin{array}{l}{\mathit{Q}}_{t+1}^{b}={\mathit{R}}_{t+n1}{\mathit{K}}_{t+1}^{b}{\mathit{R}}_{t+n}{{\mathit{K}}^{b}}_{t+1}^{H}\phantom{\rule{2em}{0ex}}\end{array}(24) 
Step 3: The signal x_{ t }can be represented by the following forwards and backwards models:
\begin{array}{l}{x}_{t+1}={\mathit{k}}_{t}^{T}{\mathbf{x}}_{t}+{w}_{t},\phantom{\rule{3.75em}{0ex}}t\ge n1\\ \phantom{\rule{1.75em}{0ex}}{x}_{t}={{\mathit{k}}^{b}}_{t+1}^{T}{\mathbf{x}}_{t+1}^{b}+{w}_{t+1}^{b},\phantom{\rule{2em}{0ex}}t\le mn+1\end{array}where w_{ t } is a doubly white noise uncorrelated with x_{n − 1}for all t ≥ n − 1 and {w}_{t}^{b} is a doubly white noise uncorrelated with x_{m − n + 1}for all t ≤ m − n + 1. Moreover, E\left[{w}_{t}{w}_{t}^{\ast}\right] and E[w_{ t }w_{ t }] are the (1,1)element and (1,2)element of the matrix Q_{ t }, respectively. Similarly, E\left[{w}_{t}^{b}{{w}_{t}^{b}}^{\ast}\right] and E\left[{w}_{t}^{b}{w}_{t}^{b}\right] are the (2n−1,2n−1)element and (2n−1,2n)element of the matrix {\mathit{Q}}_{t}^{b}, respectively.
In certain situations we have a forwards model of the form (17) for the signal x_{ t }. It would be interesting to be able to obtain a backwards model directly from the forwards model. Next, we show a useful way to get our objective.
Proposition 2. Given a forwards model of the form
with w_{ t }a doubly white noise uncorrelated with x_{n − 1}, then {x_{ t },0 ≤ t ≤ m} has the backwards representation
where the 2nvector {\mathit{k}}_{t+1}^{b} satisfies that {{\mathit{k}}^{b}}_{t+1}^{T} is equal to the 2n−1 row of the matrix {\mathit{K}}_{t+1}^{b}={\mathit{R}}_{t+n1}{\mathit{K}}_{t+n1}^{H}{\mathit{R}}_{t+n}^{1} and {w}_{t}^{b} is a doubly white noise with the properties given in Step 3 above.
Example 1 (continued) It is not difficult to check that x_{ t }is a WLM(1) signal by using property (6). Hence, applying Steps 1–3 above, it has the statespace representation
with w_{ t } a doubly white noise uncorrelated with x_{0}and {x}_{0}^{\ast}. Moreover, as x_{ t } is also a SOS signal, this model is trivially its WL autoregressive representation.
Example 2 (continued) From Theorem 1 and Steps 1–3, it follows that x_{ t }is a WLM(1) signal and has the statespace representation
with w_{ t } a doubly white noise uncorrelated with x_{1}and {x}_{1}^{\ast}.
5 Estimation problem of WLM(n) signals
Once the modeling problem has been solved for WLM(n) signals, we address the MMSE estimation problem of such signals under a WL processing approach. The forwards and backwards representations given in Theorem 3 notably simplify the design of different recursive estimation algorithms. To this end, we use the Kalman recursions on the forwards representation to provide the solution for the prediction and filtering problems and on the backwards representation for the smoothing problem (see, e.g., [17, 18]).
Suppose that we observe a WLM(n) signal {x_{ t },0 ≤ t ≤ m} via the process
with v_{ t } a doubly white noise such that E\left[{v}_{t}{v}_{t}^{\ast}\right]={n}_{1,t} and E[v_{ t }v_{ t }]=n_{2,t} with n_{1,t} > n_{2,t}. Moreover, we assume that v_{ t } is uncorrelated with x_{ s }and {x}_{s}^{\ast} for all t,s.
Consider the 2vector {\mathbf{y}}_{t}={[{y}_{t},{y}_{t}^{\ast}]}^{T}, the 2×2n matrix
and the 2×2 matrix
5.1 Prediction and filtering cases
Denote the WL filtered estimator of x_{ t }by {\widehat{x}}_{t}^{\mathit{WL}} and the onestepahead predictor of x_{t + 1} by {\widehat{x}}_{t+1t}^{\mathit{WL}}, both obtained on the basis of the information provided by the set \{{y}_{0},{y}_{0}^{\ast},\dots ,{y}_{t},{y}_{t}^{\ast}\}, and consider their associated errors {p}_{t}=E\left[\right{x}_{t}{\widehat{x}}_{t}^{\mathit{WL}}{}^{2}] and {p}_{t+1t}=E\left[\right{x}_{t+1}{\widehat{x}}_{t+1t}^{\mathit{WL}}{}^{2}]. Also denote the estimate of x_{n−1} obtained from the information provided by {[{y}_{n1},{y}_{n1}^{\ast},\dots ,{y}_{0},{y}_{0}^{\ast}]}^{T} by {\widehat{\mathbf{x}}}_{n1} and its associated error by P_{n−1}. By combining the forwards representation (17) and the classical Kalman filter we present Algorithm 1 which provides these estimators in an efficient way.
Algorithm 1. WL filter and prediction Require: y_{ t }, H_{ t }, N_{ t }, K_{ t }, Q_{ t }, g =[1,0,…,0]^{T}, {\widehat{\mathbf{x}}}_{n1}, and P_{n−1}Ensure:{\widehat{x}}_{t+1t}^{\mathit{WL}}, {\widehat{x}}_{t+1}^{\mathit{WL}}, p_{t + 1t}, and p_{t + 1}

1:
? for t = n 1do

2:
{\widehat{\mathbf{x}}}_{t+1t}?{\mathit{K}}_{t}{\widehat{\mathbf{x}}}_{t}

3:
{\mathit{P}}_{t+1t}?{\mathit{K}}_{t}{\mathit{P}}_{t}{\mathit{K}}_{t}^{H}+{\mathit{Q}}_{t}

4:
\begin{array}{l}\mathit{F}{}_{t+1}?\\ {\mathit{P}}_{t+1t}{\mathit{H}}_{t+1}^{H}{[{\mathit{H}}_{t+1}{\mathit{P}}_{t+1t}{\mathit{H}}_{t+1}^{H}+{\mathit{N}}_{t+1}]}^{1}\end{array}

5:
{\widehat{\mathbf{x}}}_{t+1}?{\widehat{\mathbf{x}}}_{t+1t}+{\mathit{F}}_{t+1}[{\mathbf{y}}_{t+1}{\mathit{H}}_{t+1}{\widehat{\mathbf{x}}}_{t+1t}]

6:
P_{t + 1}?P_{t + 1t}F_{t + 1}H_{t + 1}P_{t + 1t}

7:
{\widehat{x}}_{t+1t}^{\mathit{WL}}?{\mathit{g}}^{T}{\widehat{\mathbf{x}}}_{t+1t}

8:
{\widehat{x}}_{t+1}^{\mathit{WL}}?{\mathit{g}}^{T}{\widehat{\mathbf{x}}}_{t+1}

9:
p_{t + 1t}?g^{T}P_{t + 1t}g

10:
p_{t + 1}?g^{T}P_{t + 1}g

11:
? end for
5.2 Smoothing case
Next, we compute two WL smoothing estimators of x_{ t }based on future data. The first smoother is obtained from the set of observations \{{y}_{t},{y}_{t}^{\ast},{y}_{t+1},{y}_{t+1}^{\ast},\dots ,{y}_{m},{y}_{m}^{\ast}\} and it will be denoted by {\widehat{x}}_{t}^{{b}^{\mathit{WL}}}. The second one is derived from the information supplied by the set \{{y}_{t+1},{y}_{t+1}^{\ast},\dots ,{y}_{m},{y}_{m}^{\ast}\} and we will refer to it as {\widehat{x}}_{tt+1}^{{b}^{\mathit{WL}}}. The errors of both estimators are {p}_{t}^{b}=E\left[\right{x}_{t}{\widehat{x}}_{t}^{{b}^{\mathit{WL}}}{}^{2}] and {p}_{tt+1}^{b}=E\left[\right{x}_{t}{\widehat{x}}_{tt+1}^{{b}^{\mathit{WL}}}{}^{2}], respectively. The initial condition {\widehat{\mathbf{x}}}_{mn+1}^{b} is the estimate of {\mathbf{x}}_{mn+1}^{b} obtained from the 2n + 2vector {[{y}_{mn+1},{y}_{mn+1}^{\ast},\dots ,{y}_{m},{y}_{m}^{\ast}]}^{T} and {\mathit{P}}_{mn+1}^{b} is its associated error. By applying the backwards Kalman recursions on the backwards model (18) we get Algorithm 2.
Algorithm 2. WL smoothing Require: y_{ t }, H_{ t }, N_{ t }, {\mathit{K}}_{t+1}^{b}, {\mathit{Q}}_{t+1}^{b}, l =[0,…,0,1,0]^{T}, {\widehat{\mathbf{x}}}_{mn+1}^{b}, and {\mathit{P}}_{mn+1}^{b}Ensure:{\widehat{x}}_{tt+1}^{{b}^{\mathit{WL}}}, {\widehat{x}}_{t}^{{b}^{\mathit{WL}}}, {p}_{tt+1}^{b}, and {p}_{t}^{b}

1:
? for t = m  n do

2:
{\widehat{\mathbf{x}}}_{tt+1}^{b}?{\mathit{K}}_{t+1}^{b}{\widehat{\mathbf{x}}}_{t+1}^{b}

3:
{\mathit{P}}_{tt+1}^{b}?{\mathit{K}}_{t+1}^{b}{\mathit{P}}_{t+1}^{b}{{\mathit{K}}^{b}}_{t+1}^{H}+{\mathit{Q}}_{t+1}^{b}

4:
{\mathit{F}}_{t}^{b}?{\mathit{P}}_{tt+1}^{b}{\mathit{H}}_{t}^{H}{[{\mathit{H}}_{t}{\mathit{P}}_{tt+1}^{b}{\mathit{H}}_{t}^{H}+{\mathit{N}}_{t}]}^{1}

5:
{\widehat{\mathbf{x}}}_{t}^{b}?{\widehat{\mathbf{x}}}_{tt+1}^{b}+{\mathit{F}}_{t}^{b}[{\mathbf{y}}_{t}{\mathit{H}}_{t}{\widehat{\mathbf{x}}}_{tt+1}^{b}]

6:
{\mathit{P}}_{t}^{b}?{\mathit{P}}_{tt+1}^{b}{\mathit{F}}_{t}^{b}{\mathit{H}}_{t}{\mathit{P}}_{tt+1}^{b}

7:
{\widehat{x}}_{tt+1}^{{b}^{\mathit{WL}}}?{\mathit{l}}^{T}{\widehat{\mathbf{x}}}_{tt+1}^{b}

8:
{\widehat{x}}_{t}^{{b}^{\mathit{WL}}}?{\mathit{l}}^{T}{\widehat{\mathbf{x}}}_{t}^{b}

9:
{p}_{tt+1}^{b}?{\mathit{l}}^{T}{\mathit{P}}_{tt+1}^{b}\mathit{l}

10:
{p}_{t}^{b}?{\mathit{l}}^{T}{\mathit{P}}_{t}^{b}\mathit{l}

11:
? end for
6 Numerical example
This section is devoted to showing the advantages of representation (27) (model 2) in relation to (3) (model 1) in two fields of signal processing: simulation and estimation. Firstly, we use such models to simulate trajectories of x_{ t } defined in Example 2. Specifically, 50,000 trajectories of both models have been generated via Montecarlo simulation. To assess the performance of the simulations we compare the true correlation and complementary correlation functions with the simulated ones. Figure 1a,b depicts the true correlation and complementary correlation functions of x_{ t }, Figure 1c,d the simulated simulated ones corresponding to model 1 and Figure 1e,f the simulated ones for model 2. We can see that the simulated trajectories of model 1 pick up adequately the behavior of the correlation function. However, these trajectories are unable to show the basic characteristics of the complementary correlation function. This shortcoming does not appear with model 2 whose simulated trajectories yield accurate representations of the secondorder moments of x_{ t }. For more detail, the 2D sections of the true complementary function and the simulated ones with models 1 and 2 for t=60 and t=90, respectively, are shown in Figure 2a,b.
Finally, we compare the SL smoother obtained with model 1 and the WL smoother derived in Algorithm 2 for model 2. For the particular case in which h_{ t }=1 and n_{1,t}=1, Figure 3a compares the error {p}_{t}^{b} obtained for n_{2,t}=0.25 (dotted line) and n_{2,t}=0.8 (solid line) with the counterpart SL error (dashed line). On the other hand, considering n_{2t}=n_{2}and denoting the errors of the improper and proper smoothers for every value of n_{2} by {p}_{t}^{b}\left({n}_{2}\right) and {\stackrel{\u0304}{p}}_{t}^{b}\left({n}_{2}\right), respectively, Figure 3b displays the mean of the difference between the SL and WL estimation errors, that is, \mathit{DE}\left({n}_{2}\right)=\frac{1}{100}\sum _{t=1}^{100}\left({\stackrel{\u0304}{p}}_{t}^{b}\right({n}_{2}){p}_{t}^{b}({n}_{2}\left)\right) with n_{2} varying within the interval [0,1). As expected, both figures show that WL estimation outperforms SL estimation, that is, they illustrate the better performance of the improper smoother in relation to the proper one. From Figure 3b, we also come to the conclusion that this gain in performance decreases as n_{2}reduces.
7 Conclusions
The limited utility of the classical WSM definition to characterize the existence of a statespace representation for improper random signals has been revealed. By means of two simple examples, we have shown that in some cases the triangular condition fails to hold for signals with a statespace representation or that there exist signals with autocorrelations satisfying the triangular property for which the associated statespace representations present drawbacks in relation to their WL counterparts. Thus, the definition of a WSM signal has been extended to deal with improper signals providing new characterizations for WLM signals based either on secondorder properties or on statespace representations. Moreover, a way to check the WLM condition has been given and the correlation structure of WLM signals has been devised. Finally, WL forwards and backwards Markovian representations have been presented from which some applications are illustrated in the signal estimation and simulation fields.
Appendix 1
Proof of Theorem 1
To prove the implication 1)⇒2) observe that if x_{ t } is a WLM(n) signal then for any s < t,
which implies that {\widehat{\mathbf{x}}}^{\mathit{WL}}\left(t\right\tau \le s) is of the form (5) with K(t,s) defined in (4). Moreover, the rows are of the form
for i=1,…,n. The inverse implication, 2)⇒1), is checked similarly.
Finally, the proof of 2)⇔3) is similar to the one given in Theorem 1 of [16].
Proof of Theorem 2
The proof of 2)⇔3) is similar to that of Theorem 1 by taking into account that E\left[{\mathbf{x}}_{t}^{b}{{\mathbf{x}}_{s}^{b}}^{H}\right]=\mathit{R}(t+n1,s+n1). Now, we prove 1)⇔3). Following a similar reasoning to that used in the proof of Theorem 1 in [16], we have that (7) is equivalent to the condition
and thus,
from which, applying Theorem 1, it follows that x_{ t } is a WLM(n) signal. In a similar way the implication 1)⇒3) is proven.
Proof of Proposition 1
Taking into account that {\widehat{x}}^{\mathit{WL}}(t+ji\tau \le t)={x}_{t+ji} for j ≤ i ≤ n − 1 we obtain (8) and (9). Likewise, (13) follows from (29), (8), and (9).
Now, from (6) we get
and together with (13) we demonstrate (10), (11), and (12).
On the other hand, (14) and (15) can be proven following a similar reasoning to that of Theorem 2 in [16].
Finally, by using the Hilbert projection theorem and (5) we have
where {\mathit{w}}_{t}={[{w}_{t},{w}_{t}^{\ast},0,\dots ,0]}^{T} is the innovations process which, by construction, is uncorrelated with x_{ s } for t ≥ s. Thus,
with E\left[{\mathit{w}}_{t}{\mathit{w}}_{t}^{H}\right]={\mathit{Q}}_{t} given in (16).
Proof of Theorem 3
If x_{ t } is a WLM(n) signal then, from (13) and (30), we have
where w_{ t } is the first component of w_{ t }. Hence, denoting {\mathit{k}}_{t}={\mathit{K}}_{\left[1\right]}^{T}(t+1,t)={[{k}_{1,t},\dots ,{k}_{2n,t}]}^{T} we obtain (17). On the other hand, from the Hilbert projection theorem and (7) we get
where {\mathit{w}}_{t}^{b}={[0,\dots ,0,{w}_{t}^{b},{{w}_{t}^{b}}^{\ast}]}^{T} is the backwards innovations process which, from construction, is uncorrelated with x_{ s } for t ≤ s. Hence, {x}_{t}={\mathit{K}}_{[2n1]}(t+n1,t+n){\mathbf{x}}_{t+1}^{b}+{w}_{t+1}^{b} with {w}_{t+1}^{b} the 2n − 1 component of {\mathit{w}}_{t+1}^{b}. Thus, denoting {{\mathit{k}}^{b}}_{t+1}^{T}={\mathit{K}}_{[2n1]}(t+n1,t+n), (18) is obtained.
Conversely, suppose that x_{ t } has the representation (17). Denote \mathcal{H} the closed span generated by the set \{{x}_{\tau},{x}_{\tau}^{\ast},\tau \le t\}. By using Proposition 2.3.2 of [19], to prove that {\widehat{x}}^{\mathit{WL}}\left(t\right\tau \le s)={\widehat{x}}^{\mathit{WL}}(ts,s1,\dots ,sn+1) for any s < t is equivalent to {\widehat{x}}^{\mathit{WL}}(t+1\tau \le t)={\widehat{x}}^{\mathit{WL}}(t+1t,t1,\dots ,tn+1) for all t. Thus, projecting (17) onto \mathcal{H} and taking Proposition 2.3.2 of [19] into account we have
where {\u0175}^{\mathit{WL}}\left(t\right\tau \le t) is the projection of w_{ t } onto \mathcal{H}. The hypothesis (19) guarantees that w_{ t } is uncorrelated with x_{ s }and {x}_{s}^{\ast} for t ≥ s. Hence, {\u0175}^{\mathit{WL}}\left(t\right\tau \le t)=0 and x_{ t } is a WLM(n) signal.
The proof for the backwards representation (18) is similar.
Proof of Corollary 1
Since x_{ t } is a SOS signal then the matrices R(t + h,t), h=1,2,…, are independent of t. Thus, from (4) we obtain k_{i,t}=k_{ i }for all i and t. Finally, taking (31) into account we have
which gives (20) defining g_{1,i}=k_{2i + 1}and g_{2,i}=k_{2i + 2}.
Proof of Proposition 2
From (25) and Theorem 3 it follows that {\mathbf{x}}_{t}^{b} has the representation (32). Then by using (22) we obtain
and thus the result follows.
Abbreviations
 MMSE:

Minimummean square error
 SL:

Strictly linear
 WL:

Widely linear
 WLM:

Widely linear Markov
 WSM:

Widesense Markov.
References
Kailath T, Sayed AH, Hassibi B: Linear Estimation. Prentince Hall, New Jersey; 2000.
Poor HV, Chang Ch: A reducedcomplexity quadratic structure for the detection of stochastic signals. J. Acoust. Soc. Am 1985, 78(5):16521657. 10.1121/1.392803
Jones PW, Smith P: Stochastic Processes. An Introduction. Chapman & Hall/CRC, Boca Raton; 2010.
Doob JL: Stochastic Processes. John Willey, New York; 1953.
Mandic DP, Goh VSL: Complex Valued Nonlinear Adaptative Filters. Noncircularity, Widely Linear and Neural Models. John Willey, New York; 2009.
Adali T, Haykin S: Adaptive Signal Processing: Next Generation Solutions. WileyIEEE Press, New York; 2010.
NavarroMoreno J, Estudillo MD, FernándezAlcalá RM, RuizMolina JC: Estimation of improper complexvalued random signals in colored noise by using the hilbert space theory. IEEE Trans. Inf. Theory 2009, 55(6):28592867.
Picinbono B, Bondon P: Secondorder statistics of complex signals. IEEE Trans. Signal Process 1997, 45(2):411420. 10.1109/78.554305
Goh VSL, Mandic DP: An augmented extended kalman filter algorithm for complexvalued recurrent neural networks. Neural Comput 2007, 19: 10391055. 10.1162/neco.2007.19.4.1039
NavarroMoreno J: ARMA prediction of widely linear systems by using the innovations algorithm. IEEE Trans. Signal Process 2008, 56(7):30613068.
Cheong Took C, Mandic DP: A quaternion widely linear adaptive filter. IEEE Trans. Signal Process 2010, 58(8):44274431.
Buzzi S, Lops M, Sardellitti S: Widely linear reception strategies for layered spacetime wireless communications. IEEE Trans. Signal Process 2006, 54(6):22522262.
RubinDelanchy P, Walden AT: Simulation of improper complexvalued sequences. IEEE Trans. Signal Process 2007, 55(11):55175521.
Beutler F: Multivariate widesense Markov processes and prediction theory. Ann. Math. Stat 1963, 34(2):424438. 10.1214/aoms/1177704154
Mandrekar V: On multivariate widesense Markov processes. Nagoya Math. J 1968, 3: 719.
Kasprzyk A, Szczotka W: Covariance structure of widesense Markov processes of order k ≥ 1. Appl. Math 2006, 33(2):129143.
Dini DH, Mandic DP, Julier SJ: A widely linear complex unscented Kalman filter. IEEE Signal Process. Lett 2011, 18(11):623626.
Dini DH, Mandic DP: Class of widely linear complex Kalman filters. IEEE Trans. Neural Netw. Learn. Syst 2012, 23(5):775786.
Brockwell PJ, Davis RA: Time Series: Theory and Methods. SpringerVerlag, New York; 1991.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 2.0 International License ( https://creativecommons.org/licenses/by/2.0 ), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
About this article
Cite this article
EspinosaPulido, J.A., NavarroMoreno, J., FernándezAlcalá, R.M. et al. Widely linear Markov signals. EURASIP J. Adv. Signal Process. 2012, 256 (2012). https://doi.org/10.1186/168761802012256
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/168761802012256