 Research
 Open Access
 Published:
Energy and spectral efficient Doppler diversity transmissions in highmobility systems with imperfect channel estimation
EURASIP Journal on Wireless Communications and Networking volume 2015, Article number: 140 (2015)
Abstract
This paper studies energy and spectral efficient Doppler diversity transmissions in the presence of imperfect channel state information (CSI). Fast timevarying fading in highmobility communication systems introduces Doppler diversity that can benefit system performance. On the other hand, it is more difficult to estimate and track fast timevarying channel; thus, channel estimation errors might seriously degrade system performance in highmobility systems. The tradeoffs between channel estimation errors and Doppler diversity are studied by using two precoding schemes, a simple repetition code, and a rate 1 Doppler domain multiplexing (DDM) scheme. The repetition code can achieve the maximum Doppler diversity at the cost of a lower spectral efficiency, and the DDM scheme can achieve the energy and spectral efficient Doppler diversity transmissions. Unlike many other Doppler diversity systems that assume perfect CSI, we explicitly consider the impacts of imperfect CSI on the design and performance of both precoding schemes. Optimum and suboptimum receivers for both schemes are developed by studying the statistical properties of channel estimation errors. The analytical error probabilities of the receivers are expressed as functions of a number of system parameters, such as the maximum Doppler spread, the percentage of pilot symbols for channel estimation, the energy allocation between the pilot and data symbols, etc. The analytical and simulation results indicate that both precoding schemes can achieve the maximum Doppler diversity order through the optimization of the various parameters, even in the presence of imperfect CSI.
Introduction
With the increasing demands of highspeed railways and aircraft communications, wireless communications in highmobility environment have attracted considerable attentions during the past few years. Signals in highmobility systems could experience large Doppler shifts in the order of kilohertz, while most conventional wireless communication systems are designed for Doppler shifts up to a few hundred hertz. The large Doppler shifts result in fast timevarying fading, which is one of the main challenges for the design of reliable highmobility systems. On the other hand, fast timevarying fading caused by large Doppler shifts in highmobility systems provides Doppler diversity, which can be used to benefit system performance.
Several methods are proposed to exploit the Doppler diversity gain inherent in fast timevarying fading [13], and they provide efficient countermeasures against fading. Most existing works on Doppler diversity assume that perfect channel state information (CSI) can be obtained at the receiver. However, in highmobility systems, it is a nontrivial task to estimate and track the fast timevarying fading with high precision and credibility. Channel estimation errors are usually inevitable and might have significant impacts on system performance when the Doppler frequency is high. The performance of precoded orthogonal frequency division multiplexing (OFDM) systems with channel estimation error is studied in [4]. It is shown in [5] that for a singleinput multipleoutput (SIMO) system with identically and independently distributed (i.i.d.) fading, the conventional maximal ratio combining (MRC) receiver is no longer optimum with imperfect CSI. The results of [4] and [5] cannot be applied to highmobility systems because they both assume the systems experience quasistatic fading channels. The optimum designs of highmobility systems in the presence of imperfect CSI are studied in [6,7] and [8] in terms of different design metrics, such as the bit error rate (BER), symbol error rate (SER), or spectral efficiency. However, Doppler diversity is not considered in these works.
In this paper, we study the energy and spectral efficient designs of Doppler diversity transmissions in highmobility systems with imperfect CSI. The design aims at achieving a balanced tradeoff between Doppler diversity gain and channel estimation errors by considering a large number of system parameters, such as the percentage of pilot symbols among the total transmitted symbols, the energy allocation between pilot and data symbols, and the maximum Doppler shifts. Two precoding schemes are employed to achieve Doppler diversity: a simple repetition code, and a rate 1 Doppler domain multiplexing (DDM) code [3]. The repetition precoder can achieve the full Doppler diversity at the cost of a lower spectral efficiency, and the rate 1 DDM precoder can achieve energy and spectral efficient Doppler diversity transmissions. By analyzing the statistical properties of channel estimation errors, we develop the optimum and suboptimum receivers for both precoding schemes. The main contributions of this paper are summarized as follows.

The Doppler diversity transceivers in this paper are developed by explicitly utilizing the properties of imperfect CSI, whereas most existing works on Doppler diversity assume perfect CSI. For example, the DDM scheme is originally proposed in [3] for systems with perfect CSI, and its performance degrades significantly with imperfect CSI, given the fact that conventional receivers are no longer optimum with imperfect CSI. We address this problem by developing new transceiver structures that explicitly consider the impact of channel estimation errors.

The impacts of imperfect CSI on system performance are identified by developing the theoretical error probabilities of the new transceivers. The new analytical results are expressed as functions of the secondorder statistics of the channel estimation errors, and they reveal the tradeoff between Doppler diversity and channel estimation errors.

The new transceiver structure along with the theoretical error probabilities can be used to guide the development of practical Doppler diversity systems. With the analytical and simulation results, the various system parameters that yield the optimum tradeoff between channel estimation errors and Doppler diversity are identified.
The rest of this paper is organized as follows. Section 2 introduces the system model and the two precoding schemes. Section 3 shows the channel estimation. The tradeoff between Doppler diversity and channel estimation errors is studied in Section 4 with the repetition code. Section 5 presents the receiver design and performance analysis of the rate 1 DDM scheme. Numerical and simulation results are presented in Section 6, and Section 7 concludes the paper.
System model
The system model is presented in this section. At the transmitter, the modulated data symbols are precoded to achieve Doppler diversity transmissions. Pilot symbols are inserted among the precoded data systems to track and estimate the fast timevarying fading channel.
Precoding
The data symbols to be transmitted are divided into slots. Denote the modulated data symbols in one slot as $\textbf {s} = [s_{1}, \cdots, s_{N_{s}}]^{T} \in {\mathcal S}^{N_{s}}$ , where ${\mathcal S}$ is the modulation alphabet set, A ^{T} is the matrix transpose operation, and N _{ s } is the number of data symbols in one slot. To achieve the maximum Doppler diversity, precoding is employed to spread out the data symbols in the time domain before transmission. The precoded data vector can be represented by
where ${\mathbf \Theta } \in {\mathcal C}^{N_{c} \times N_{s}}\phantom {\dot {i}\!}$ is a precoding matrix with ${\mathcal C}$ being the set of complex numbers, and N _{ c }≥N _{ s } is the length of the codeword. Define the kth precoded data symbol of $\textbf {c} = [c_{1}, \cdots, c_{N_{c}}]^{T}\phantom {\dot {i}\!}$ as c _{ k }, where k=1,⋯,N _{ c }. The coding rate is $\rho = \frac {N_{s}}{N_{c}}$ and the precoding matrix satisfies trace(Θ Θ ^{H})=N _{ s }, with A ^{H} being the matrix Hermitian operation. Two precoding schemes are considered in this paper.
The first precoding scheme is a simple rate $\frac {1}{N_{c}}$ repetition code with N _{ s }=1 and ${\mathbf \Theta } = \frac {1}{\sqrt {N_{c}}}\textbf {1}_{N_{c}}$ , where $\textbf {1}_{N_{c}}$ is a length N _{ c } allone vector. Since the repetition code spreads one data symbol over the entire slot, it is guaranteed to achieve the maximum Doppler diversity at the cost of a low spectral efficiency. The repetition code provides the best possible performance in terms of Doppler diversity gain, and its performance can serve as a lower bound for practical precoding schemes [9].
The second precoding scheme is the rate 1 DDM scheme originally proposed in [3]. It can achieve the maximum Doppler diversity without sacrificing spectral efficiency, at the cost of mutual interference among the data streams. The precoding matrix is ${\mathbf \Theta } = \mathbf {F}_{N_{c}}^{H}$ , where N _{ c }=N _{ s } and $\mathbf {F}_{N_{c}} \in {\mathcal C}^{N_{c} \times N_{c}}$ is the normalized discrete Fourier transform (DFT) matrix with the (n,k)th element being $\frac {1}{\sqrt {N_{c}}} e^{j2\pi (n1)(k1)/N_{c}}$ . With DDM, the codeword can be alternatively expressed as
where θ _{ n } is the nth column of Θ. With such a precoding scheme, the nth data symbol is spread out over the entire slot as θ _{ n } s _{ n }, such that the maximum Doppler diversity can be achieved. However, the orthogonality between the columns of the precoding matrix will be destroyed by the fast timevarying fading channel. Therefore, there will be mutual interferences among the data symbols at the receiver. The interference can be partly removed by means of Doppler domain equalization. It should be noted that perfect CSI at the receiver is assumed by [3], and this assumption is usually not true for highmobility systems.
Pilotassisted transmission
After precoding, N _{ p } equallyspaced pilot symbols are inserted in each slot to assist channel estimation at the receiver. The number of precoded data symbols N _{ c } and pilot symbols N _{ p } are chosen as N _{ c }=N _{ p }(K−1), such that there are K−1 precoded data symbols between a pair of adjacent pilot symbols. The kth pilot symbol is denoted as $x_{i_{k}}=p_{k}$ , where i _{ k }=k K is the index of the kth pilot symbol, for k=1,…,N _{ p }. Similarly, the kth data symbol is denoted as $x_{n_{k}} = c_{k}$ , where $n_{k} = k + \lfloor \frac {k1}{K1} \rfloor $ is the index of the kth coded data symbol, for k=1,⋯,N _{ c }, where ⌊a⌋ is the largest integer not larger than a. Define $\textbf {x}_{p} = [x_{i_{1}}, \cdots, x_{i_{N_{p}}}]^{T}\phantom {\dot {i}\!}$ and $\textbf {x}_{d} = [x_{n_{1}}, \cdots, x_{n_{N_{c}}}]^{T}\phantom {\dot {i}\!} $ as the pilot and coded data vectors, respectively. Pilot and data symbols can be from different modulation alphabet sets. Without loss of generality, pilot symbols are assumed to be constant amplitude symbols, i.e., p _{ k }^{2}=1, which is not necessarily the case for data symbols. With such a slot structure, the pilot percentage can be defined as $\delta = \frac {N_{p}}{N} = \frac {1}{K}$ , with N=N _{ p }+N _{ c }.
Due to the different roles of channel estimation and data transmission, different amounts of energy are allocated to the pilot and coded data symbols. Denote the energy for each pilot and coded data symbol as E _{ p } and E _{ c }, respectively. The total energy in one slot is thus E _{0}=E _{ p } N _{ p }+E _{ c } N _{ c }. The energy per uncoded information bit can be calculated as $E_{b}=\frac {E_{p} N_{p} + E_{c} N_{c}}{N_{s} \log _{2} M}$ , where $M = {\mathcal S}$ is the cardinality of the modulation constellation set. Define an energy allocation factor $\eta = \frac {E_{p} N_{p}}{E_{0}} \in (0, 1)$ as the fraction of the total energy per slot allocated to pilot symbols, E _{ p } and E _{ c } can be expressed by
where ρ is the precoding code rate. Under a fixed E _{ b }, η, and ρ, increasing the pilot percentage δ will decrease the energy for each pilot symbol; however, it will not affect the energy per coded data symbol.
The pilot and coded data symbols are transmitted over the timevarying fading channel. The pilot symbols observed at the receiver can be represented by
where ${{\mathbf y}_{p}\! = [y(i_{1}), \cdots \!, y(i_{N_{p}})]^{T} \!\in \! {\mathcal C}^{N_{p} \times 1}}$ and ${{\mathbf z}_{p} \!= [z(i_{1}), \cdots \!,} z(i_{N_{p}})]^{T} \in {\mathcal C}^{N_{p} \times 1}$ are the received pilot vector and additive white Gaussian noise (AWGN) vector, respectively, and X _{ p }=diag(x _{ p }) is a diagonal matrix with the main diagonal being x _{ p }, and $\textbf {h}_{p} = [h(i_{1}), \dots, h(i_{N_{p}})]^{T} \in \mathcal {C}^{N_{p} \times 1}\phantom {\dot {i}\!}$ is the discretetime channel fading vector. The AWGN vector is a zeromean symmetric complex Gaussian random vector (CGRV) with covariance matrix $N_{0} \textbf {I}_{N_{p}}\phantom {\dot {i}\!}$ , where I _{ N } is a size N identity matrix.
Similarly, the data symbols observed at the receiver can be denoted as,
where ${{\mathbf y}_{d}\!=\, [y(n_{1}),\! \cdots \!, y(n_{N_{c}})]^{T} \!\!\in \! {\mathcal C}^{N_{c} \times 1}}$ and ${{\mathbf z}_{d} \!= [\!z(n_{1}), \cdots \!,} z(n_{N_{c}})]^{T} \in {\mathcal C}^{N_{c} \times 1}$ are received coded data symbols and AWGN, respectively, X _{ d }=diag(x _{ d }) is a diagonal matrix with the precoded data vector x _{ d } on its main diagonal, and $\textbf {h}_{d} = [h(n_{1}), \dots, h(n_{N_{c}})]^{T} \in \mathcal {C}^{N_{c} \times 1}$ .
Channel estimation
The channel is assumed to experience widesense stationary uncorrelated scattering, and h(n) is a zeromean symmetric complex Gaussian random process with the covariance function
where a ^{∗} denotes complex conjugate, $f_{_{\text {D}}}$ is the maximum Doppler spread of the fading channel, T _{ s } is the symbol period, and J _{0}(x) is the zeroorder Bessel function of the first kind.
In fast timevarying fading, the channel coefficients vary from symbol to symbol by following the time correlation in (7). As a result, the channel coefficients of the pilot symbols, h _{ p }, are different from the channel coefficients of the data symbols, h _{ d }.
In order to track the fast time variation of the fading coefficients, the channel estimation can be performed in two steps. In the first step, the receiver obtains an estimate of the channel coefficients at pilot locations as
where ${\mathbf W}_{p} \in {\mathcal C}^{N_{p} \times N_{p}}$ is the minimum mean squared error (MMSE) matrix designed to minimize the mean squared error (MSE) $\frac {1}{N_{p}}{\mathbb E}\left (\\hat {{\mathbf h}}_{p}  \textbf {h}_{p}\^{2}\right)$ . In the second step, the channel coefficients at data locations can be estimated by performing time domain interpolation over ${\hat {\textbf {h}}}_{p}$ as
where ${\mathbf W}_{d} \in {\mathcal C}^{N_{c} \times N_{p}}$ is the MMSE matrix to minimize $\frac {1}{N_{c}}{\mathbb E}\left (\\hat {{\mathbf h}}_{d}  \textbf {h}_{d}\^{2}\right)$ .
It is shown in [10] and [11] that the above twostep MMSE is equivalent to a singlestep MMSE as
where ${\mathbf W} = \textbf {W}_{p} \textbf {W}_{d} \in {\mathcal C}^{N_{c} \times N_{p}}$ is the MMSE matrix that can minimize $\frac {1}{N_{c}}{\mathbb E}\left (\\hat {{\mathbf h}}_{d}  \textbf {h}_{d}\^{2}\right)$ . In (10), the received pilot symbols are used to estimate the channel coefficients of data symbols through time interpolation; thus, the fast timevariation of the fading coefficients can be accurately tracked.
Based on the orthogonal principal, ${\mathbb E}\left [\left (\hat {{\mathbf h}}_{d}  \textbf {h}_{d}\right)\textbf {y}_{p}^{H}\right ] = \textbf {0}$ , the MMSE estimation matrix can be solved as
where A ^{−1} is the matrix inverse operation, $\textbf {R}_{\textit {dp}} = {\mathbb E}\left [\textbf {h}_{d} \textbf {h}_{p}^{H}\right ] \in {\mathcal R}^{N_{c} \times N_{p}}$ and $\mathbf {R}_{\textit {pp}}=\mathbb {E}\left [{\mathbf h}_{p} {\mathbf h}_{p}^{H}\right ] \in \mathcal {C}^{N_{p} \times N_{p}}$ with their elements defined in (7).
Define the error vector $\textbf {e}_{d} =\textbf {h}_{d}  \hat {{\mathbf h}}_{d}$ . The correlation of the error vector, $\textbf {R}_{\textit {ee}} = {\mathbb E}\left [\textbf {e}_{d}\textbf {e}_{d}^{H}\right ] \in {\mathcal R}^{N_{c} \times N_{c}}$ , can be calculated as
where $\gamma _{p} = \frac {E_{p}}{{\sigma _{z}^{2}}}$ is the signaltonoise ratio (SNR) of the pilot symbols, $\mathbf {R}_{\textit {dd}} = {\mathbb E}\left [\textbf {h}_{d} \textbf {h}_{d}^{H}\right ] \in {\mathcal R}^{N_{c} \times N_{c}}$ , and $\textbf {X}_{p}\textbf {X}_{p}^{H} = \textbf {I}_{N_{p}}$ is used in the derivation of the above equation.
For the design and analysis of the diversity receiver in the presence of imperfect CSI, it is necessary to obtain the statistical properties of ${\hat {\textbf {h}}}_{d}$ by considering the effects of channel estimation errors because the receiver performs detection based on the knowledge of the estimated channel coefficients ${\hat {\textbf {h}}}_{d}$ .
Given the pilot symbols, the vector y _{ p } is zeromean Gaussian distributed with covariance matrix ${\mathbb E}[\textbf {y}_{p} \textbf {y}_{p}^{H}] = E_{p} \textbf {X}_{p}\textbf {R}_{\textit {pp}} \textbf {X}_{p}^{H} + {\sigma _{z}^{2}} \textbf {I}_{N_{p}}$ . From (10), the estimated channel coefficient vector is a linear transformation of a zeromean Gaussian vector. Therefore, ${\hat {\textbf {h}}}_{d}$ is zeromean Gaussian distributed with the covariance matrix
To gain further insights, define the estimation MSE as $\sigma _{e,N_{c}}^{2} = \frac {1}{N_{c}} \text {trace}(\textbf {R}_{\textit {ee}})$ . When N _{ c } goes to infinity while keeping a finite δ, the asymptotic MSE ${\sigma _{e}^{2}} = {\lim }_{N_{c} \to \infty }\sigma _{e,N_{c}}^{2}$ is [12, Proposition 1],
where $\omega _{_{\text {D}}}=2\pi {f_{_{\text {D}}}} T_{s}$ , γ _{ b }=E _{ b }/N _{0} is the equivalent SNR of the uncoded bit, and $\nu =\eta \left (\frac {1}{\delta }1\right)\rho \log _{2} M$ .
Tradeoff between channel estimation errors and Doppler diversity
The tradeoff between channel estimation errors and the maximum Doppler diversity gain is studied in this section with the help of the simple repetition code.
Optimum diversity receiver with imperfect CSI
With the repetition precoder, the received data vector in (6) can be simplified to
It should be noted that the total energy of the N _{ c } repeated data symbols is normalized to E _{ c } to ensure fair comparison with other precoding schemes and uncoded systems.
The receiver performs detection based on the received data vector y _{ d } and the knowledge of the estimated CSI vector ${\hat {\textbf {h}}}_{d}$ . Since both h _{ d } and ${\hat {\textbf {h}}}_{d}$ are zeromean Gaussian distributed, the error vector $\textbf {e}_{d} =\textbf {h}_{d}  \hat {{\mathbf h}}_{d}$ is zeromean Gaussian distributed. Conditioned on ${\hat {\textbf {h}}}_{d}$ , h _{ d } is Gaussian distributed with mean $\textbf {u}_{d{\hat h}_{d}} = {\mathbb E}[\textbf {h}_{d}{\hat {\textbf {h}}}_{d}] = {\hat {\textbf {h}}}_{d}$ and covariance matrix $\textbf {R}_{dd{\hat h}_{d}} = {\mathbb E}\left [\left (\textbf {h}_{d}\textbf {u}_{d{\hat h}_{d}}\right)\left (\textbf {h}_{d}\textbf {u}_{d{\hat h}_{d}}\right)^{H} {\hat {\textbf {h}}}_{d}\right ] = \textbf {R}_{\textit {ee}}$ .
Therefore, conditioned on the transmitted data symbol s _{ m } and the estimated CSI vector ${\hat {\textbf {h}}}_{d}$ , the received data vector y _{ d } is complex Gaussian distributed, i.e., $\textbf {y}_{d}  ({\hat {\textbf {h}}}_{d}, s_{m}) \sim {\mathcal CN}\left (\textbf {u}_{y{\hat h}_{d}, s_{m}}, \textbf {R}_{yy{\hat h}_{d}, s_{m}}\right)$ . The conditional mean and conditional covariance matrix are given as follows
For system with equiprobable symbols, the maximum likelihood (ML) detection rule can minimize the detection error probability. From (16), maximizing the likelihood function, $p(\textbf {y}_{d}{\hat {\textbf {h}}}_{d}, s_{m})$ , is equivalent to minimize the following cost function
It can be easily shown that
where ℜ denotes the real part operator, and α is the decision variable defined as
Therefore, the optimum decision rule in the presence of imperfect CSI can be written as
Performance analysis
The error performance for systems with Mary phase shift keying (MPSK) modulation and operating with imperfect CSI and the optimum decision rule is derived based on the statistical properties of the estimated CSI ${\hat {\textbf {h}}}_{d}$ .
From [5], the conditional error probability (CEP) for systems with MPSK modulation is
where $\textbf {Q} = \left (\textbf {R}_{\textit {ee}} + \frac {1}{\gamma _{c}} \textbf {I}_{N} \right)^{1}$ .
The above CEP depends on the estimated CSI through the quantity $\beta = {\hat {\textbf {h}}}_{k}^{H} \textbf {Q}{\hat {\textbf {h}}}_{k}$ . Since the estimated CSI vector ${\hat {\textbf {h}}}_{k}$ is zeromean Gaussian distributed with covariance matrix $\textbf {R}_{{\hat d}{\hat d}}$ given in (13), β is a quadratic form of a complex Gaussian random vector with the moment generating function (MGF) [13]
where t is a dumb variable.
The MGF in (22) depends on the matrices R _{ dd } and R _{ ee }. Perform eigenvalue decomposition of R _{ ee } as
where $\textbf {V} \in {\mathcal C}^{N_{s} \times N_{s}}$ contains the orthonormal eigenvectors of R _{ ee }, and ${\Lambda } = \text {diag}{[\varphi _{1}, \cdots, \varphi _{N_{c}}]}\phantom {\dot {i}\!}$ is the diagonal matrix containing the corresponding eigenvalues.
Based on the Sylvester’s determinant identity, the MGF in (22) can be rewritten as
where $\boldsymbol {\Omega } \in {\mathcal R}^{N_{c} \times N_{c}}\phantom {\dot {i}\!}$ is a diagonal matrix with the kth diagonal element being $\varphi _{k} + \frac {N_{c}}{\gamma _{c}}\phantom {\dot {i}\!}$ .
Perform eigenvalue decomposition on the Hermitian matrix $\boldsymbol {\Upsilon } = \boldsymbol {\Omega }^{\frac {1}{2}} \textbf {V}^{H} \textbf {R}_{{\hat d}{\hat d}} \textbf {V} \boldsymbol {\Omega }^{\frac {1}{2}} = \textbf {U}\textbf {D}\textbf {U}^{H}\phantom {\dot {i}\!}$ , where $\textbf {U} \in {\mathcal C}^{N_{c} \times N_{c}}\phantom {\dot {i}\!}$ contains the orthonormal eigenvectors of Υ, and $\textbf {D} = \text {diag}\{[d_{1}, \cdots, d_{N_{c}}]\}\phantom {\dot {i}\!}$ is the diagonal matrix containing the corresponding eigenvalues. Then the MGF in (24) can be simplified to
From (21) and (25), the unconditional SER, $P(E) =\mathbb {E} [ P(E\hat {{\mathbf h}}_{d}) ]$ , can then be expressed as
which can be easily evaluated by numerical calculation.
In case of perfect CSI, we have the assumption R _{ ee }=0 so that the decision variable in (19) simplifies to $\alpha = \frac {1}{{\sigma _{z}^{2}}}\sqrt {\frac {E_{s}}{N_{s}}} \textbf {h}_{d}^{H} \textbf {y}_{d}$ . Thus, the optimum decision rule in Proposition (20) degrades to the conventional maximal ratio combining (MRC) receiver. From (26), the unconditional SER simplifies to
where g _{ n } is the eigenvalue of R _{ dd }, and it is directly related to the Doppler diversity gain.
However, in case of imperfect CSI, the result in (20) shows that MRC is no longer optimal. The presence channel estimation error affects the decision process, and the new optimum decision rule has to take into consideration the statistical properties of the channel estimation error quantified in the matrix R _{ ee }.
The SER expressions in (26) reveal the tradeoff between Doppler diversity and channel estimation errors, and they provide a lower bound on the performance of systems with practical precoders.
Doppler domain multiplexing in the presence of imperfect CSI
The design and performance of a practical rate 1 DDM precoding scheme is studied in this section. A suboptimum receiver is developed by studying the statistical properties of the estimated channel coefficients, and the corresponding analytical error performance is derived.
Doppler domain equalization with imperfect CSI
From (2) and (6), the received coded data samples can be expressed as
where ${\hat {\textbf {H}}}_{d} = \text {diag}({\hat {\textbf {h}}}_{d})$ and E _{ d }=diag(e _{ d }) are diagonal matrices with ${\hat {\textbf {h}}}_{d}$ and e _{ d } on their main diagonals, respectively.
The system can be considered as an equivalent MIMO system with the equivalent channel matrix being $\hat {{\mathbf H}}_{d} {\mathbf \Theta }$ , which introduces interference among the symbols in the frequency domain. In addition, interference is introduced by the channel estimation error.
We propose to detect the transmitted symbols by using an iterative softinput softoutput (SISO) block decision feedback equalizer (BDFE) [14]. Unlike conventional BDFE that assume perfect CSI at the receiver, the BDFE in this paper is developed by considering the statistical properties of the estimated channel coefficients and the channel estimation errors.
The input to the SISOBDFE equalizer is the a priori loglikelihood ratio (LLR) of the information bits, and the output of the SISOBDFE equalizer is the a posteriori LLR of the information bits. The a priori LLR at the ith iteration is the a posteriori LLR at the (i−1)th iteration. Based on the a priori LLR of the bits, the a priori mean, ${\bar s}_{n}$ , and variance, ${\sigma _{n}^{2}}$ , of the symbol s _{ n } can be calculated as described in [14].
The structure of the SISOBDFE in one iteration is shown in Fig. 1. The BDFE contains two filters, a feedforward filter A _{ n } and a feedback filter B _{ n }, both will be calculated by using the properties of $\hat {{\mathbf H}}_{d}$ and E _{ d }.
The output of the feedforward filter is ${\textbf {r} = \textbf {A}_{n}} \left (\textbf {y}_{d}  \hat {{\mathbf H}}_{d} {\bar {\textbf {s}}}_{n} \right)$ , where ${\bar {\textbf {s}}}_{n} = [{\bar {s}_{1}}, \dots, {\bar {s}_{n1}}, 0, {\bar {s}_{n+1}}, \dots, {\bar {s}_{N_{c}}}]^{T} \in {\mathcal C}^{N_{c} \times 1}$ is the a priori mean vector. The nth element of ${\bar {\textbf {s}}}_{n}$ is set to 0 during the detection of s _{ n } to avoid instability caused by positive feedback. The output of the feedforward filter is then used to subtract the output of the feedback filter, $\textbf {B}_{n} ({\hat {\textbf {s}}}  {\bar {\textbf {s}}}_{n})$ , where B _{ n } is a strict upper triangular matrix with the main diagonal being 0, ${\hat {\textbf {s}}}$ is the hard decision at the output of the BDFE, and the soft output of the BDFE filter, ${\tilde {\textbf {s}}}_{n}$ , can then be represented by
where the soft output ${\tilde {\textbf {s}}}_{n}$ is used for the detection of s _{ n }.
Since B _{ n } is strictly upper triangular, the detection is performed in a reverse order, that is, s _{ n+1} is detected before s _{ n } and the hard decision of s _{ n+1} is fed back to facilitate the detection of s _{ k } for k<n+1.
Based on the common assumption of correct past decisions, the error vector ${\epsilon }_{n} = {\tilde {\textbf {s}}}_{n}  \textbf {s}$ can be written as
With the orthogonality principle, ${\mathbb E} \left [{\epsilon }_{n} \mathbf {y}_{d}^{H} \right ] = 0$ , the feedforward matrix A _{ n } can be derived as
where ${\mathbf G}_{\text {mmse}} = {\mathbf R}_{s y} {\mathbf R}_{\textit {yy}}^{1}$ is the linear MMSE matrix for estimating s from y _{ d }, $\textbf {R}_{\textit {sy}} = {\mathbb E}\left [({\mathbf s}{\bar {\textbf {s}}}_{n}) \left ({\mathbf y}_{d}  \hat {{\mathbf H}}_{d} {\bar {\textbf {s}}}_{n} \right)^{H}\right ]$ , and $\textbf {R}_{\textit {yy}} = {\mathbb E}\left [\left ({\mathbf y}_{d}  \hat {{\mathbf H}}_{d} {\bar {\textbf {s}}}_{n} \right) \left ({\mathbf y}_{d}  \hat {{\mathbf H}}_{d} {\bar {\textbf {s}}}_{n} \right)^{H}\right ]$ . From (27), we have
where $\phantom {\dot {i}\!}{\mathbf R}_{\textit {ss}} = \text {diag}\left [\sigma _{1}^{2}, \cdots, \sigma _{N_{c}}^{2}\right ]$ with ${\sigma _{n}^{2}}$ being the a priori variance of s _{ n }, $\textbf {T} = \textbf {E}_{d}\left ({\mathbf \Theta } {\mathbf R}_{\textit {ss}} {\mathbf \Theta }^{H}\right) \textbf {E}_{d}^{H}\phantom {\dot {i}\!}$ . Since E _{ d } is diagonal, the (m,n)th element of T is $(\textbf {T})_{m,n} = e_{m} e_{n}^{*} \bar {v}_{\textit {mn}}$ , where e _{ m } is the mth element of the channel estimation error vector e, and $\bar {v}_{\textit {mn}}$ is the (m,n)th element of the matrix $\overline {\textbf {V}} = {\mathbf \Theta } {\mathbf R}_{\textit {ss}} {\mathbf \Theta }^{H}\phantom {\dot {i}\!}$ . Thus ${\mathbb E}\left [(\textbf {T})_{m,n}\right ] = (\textbf {R}_{\textit {ee}})_{\textit {mn}} \cdot v_{\textit {mn}}$ . Therefore, the matrix ${\mathbb E}\left (\textbf {T} \right) = \textbf {R}_{\textit {ee}}\odot {\overline {\textbf {V}}}\phantom {\dot {i}\!}$ , where $(\overline {\textbf {A}} \odot \overline {\textbf {B}})_{m,n} = (\overline {\textbf {A}})_{m,n} (\overline {\textbf {B}})_{m,n}\phantom {\dot {i}\!}$ is the elementwise multiplication between two matrices.
Combining (29) and (30) yields ${\epsilon }_{n} = (\mathbf {B}_{n}+\mathbf {I}_{N_{c}}) {\xi}_{n}\phantom {\dot {i}\!}$ , where ${\xi }_{n} = {\mathbf R}_{s y} {\mathbf R}_{\textit {yy}}^{1} ({\mathbf y}_{d}  \hat {{\mathbf H}}_{d} {\bar {\textbf {s}}}_{n})  ({\mathbf s}{\bar {\textbf {s}}}_{n})\phantom {\dot {i}\!}$ is the error vector of a linear MMSE equalizer. Based on the orthogonal principle ${\mathbb E}({\xi }_{n} \textbf {y}_{d}^{H}) = 0\phantom {\dot {i}\!}$ , we have $\mathbf {\Phi }_{\xi \xi }={\mathbb E}\left ({\xi }_{n}{\xi }_{n}^{H}\right) = \textbf {R}_{\textit {ss}}^{n} \textbf {R}_{\textit {sy}}\textbf {R}_{\textit {yy}}^{1} \textbf {R}_{\textit {sy}}^{H}\phantom {\dot {i}\!}$ . Combining (31), (32), and the matrix inverse lemma, we have
The covariance matrix of the BDFE error vector ε _{ n } can then be calculated as $\mathbf {\Phi }_{\epsilon \epsilon } = (\mathbf {B}_{n}+\mathbf {I}_{N_{c}}) \mathbf {\Phi }_{\xi \xi } (\mathbf {B}_{n}+\mathbf {I}_{N_{c}})^{H}\phantom {\dot {i}\!}$ . We need to identify the feedback matrix B _{ n } such that trace(Φ _{ ξ ξ }) is minimized under the constraint that B _{ n } is strictly upper diagonal. The problem can be solved through the Cholesky decomposition of $\mathbf {\Phi }_{\xi \xi }^{1}$ as [15]
where ${\mathbf U}_{\xi } \in {\mathcal C}^{N_{c} \times N_{c}}$ is an upper triangular matrix with unit diagonal elements, ${\mathbf D}_{\xi } \in {\mathcal R}^{N_{c} \times N_{c}}$ is a diagonal matrix, and ${\mathbf L}=\sqrt {{\mathbf D}_{\xi }} {\mathbf U}_{\xi }$ . With the Cholesky decomposition described in (33), the feedback matrix B _{ n } can be calculated as $\textbf {B}_{n} = \textbf {U}_{\xi }\textbf {I}_{N_{c}}\phantom {\dot {i}\!}$ . Consequently, the error covariance matrix of the BDFE equalizer is $\mathbf {\Phi }_{\epsilon \epsilon } = \textbf {D}_{\xi }^{1}$ .
Error performance analysis
The pairwise error probability (PEP) and a BER lower bound of the rate 1 DDM system with imperfect CSI is developed in this subsection.
Based on the estimated channel coefficients, (6) can be rewritten as
where $\textbf {w}_{d} = \sqrt {{E_{c}}} {\mathbf X}_{d} {\mathbf e}_{d} + {\mathbf z}_{d}$ is the equivalent noise component associated with the estimation error and AWGN. If we treat w _{ d } as noise, then the maximum likelihood rule for deciding between two possible codewords for MPSK modulated systems, c _{ α } and c _{ β }, is
where X _{ α }=diag(c _{ α }) and X _{ β }=diag(c _{ β }).
Based on the ML decision rule in (35), the PEP of deciding in favor of X _{ β } when X _{ α } is transmitted can be calculated as
where $Q_{z} = \ {\mathbf y}_{d}  \sqrt {{E_{s}}} {\mathbf X}_{\beta } \hat {{\mathbf h}}_{d} \^{2}  \ {\mathbf y}_{d}  \sqrt {{E_{s}}} {\mathbf X}_{\alpha } \hat {{\mathbf h}}_{d} \^{2}$ .
Define ${\mathbf z}_{1} = \sqrt {E_{s}} ({\mathbf X}_{\alpha }{\mathbf X}_{\beta }) \hat {{\mathbf h}}_{d} + \textbf {w}_{\alpha }$ and z _{2}=w _{ α }, where $\textbf {w}_{\alpha } = \sqrt {{E_{s}}} {\mathbf X}_{\alpha } {\mathbf e}_{d} + {\mathbf z}_{d}$ . Conditioned on X _{ α } and X _{ β }, both z _{1} and z _{2} are complex Gaussian distributed. Then Q _{ z } can be expressed in a quadratic form of a complex Gaussian random vector as
The characteristic function (CHF) of Q _{ z } is given by [13]
where W _{ z } has the form
with ${\mathbf R}_{z_{1}} = E_{s} ({\mathbf X}_{\alpha }{\mathbf X}_{\beta }) {\mathbf R}_{\hat {d}\hat {d}} ({\mathbf X}_{\alpha }{\mathbf X}_{\beta })^{H} + \textbf {R}_{w_{\alpha }}$ , $\textbf {R}_{w_{\alpha }} = {\mathbb E}\left [\textbf {w}_{\alpha } \textbf {w}_{\alpha }^{H}\right ] = E_{s} \textbf {X}_{\alpha } {\mathbf R}_{\textit {ee}} \textbf {X}_{\alpha }^{H} + {\sigma _{z}^{2}} {\mathbf I}_{N_{c}}$ .
Performing eigenvalue decomposition of R _{ z } yields ${\mathbf R}_{z} = {\mathbf U}_{z} {\mathbf D}_{z} {\mathbf U}_{z}^{H}$ . The CHF in (37) can be alternatively represented by
where λ _{ i } is the ith nonzero eigenvalue of the rank N _{ w } matrix ${\mathbf D}_{z}^{\frac {1}{2}} {\mathbf U}_{z}^{H} {\mathbf K}{\mathbf U}_{z} {\mathbf D}_{z}^{\frac {1}{2}}$ .
When the values of λ _{ i } are all unique, which is usually the case for practical systems, we can rewrite the CHF through partial fraction expansion as
where $\mu _{i} = \prod _{n=1, n \neq i}^{N_{w}} \frac {\lambda _{i}}{\lambda _{i}  \lambda _{n}}$ .
The PDF of Q _{ z } can be derived from the CHF as $f_{Q_{z}}(x) = \frac {1}{2\pi } \int _{\infty }^{\infty } \Psi _{Q_{z}}(t) e^{j x t} dt$ , and the result is
where u(x) is the unit step function, and sgn(x)=1 if x≥0 and −1 otherwise.
Without loss of generality, assume the eigenvalues are ordered in an ascending order, and there are $N_{w}^{}$ negative eigenvalues, the PEP can be written as
where λ _{ n } for $n=1, \dots, N_{w}^{}$ are the negative eigenvalues.
For a pair of information vectors, s _{ α } and s _{ β }, their Hamming distance are defined as D _{ H }(s _{ α },s _{ β })=∥s _{ α }−s _{ β }∥_{0}, where ∥a∥_{0} is the l _{0} norm operator that returns the number of nonzero elements in the vector a. Intuitively, the BER can be reduced by assigning codeword pairs with smaller PEP to information vector pairs with larger Hamming distance, such that the probability of error events with a large number of bit errors is small. Equivalently, when D _{ H }(s _{ α },s _{ β }) is small, a good precoding scheme should yield a relatively large PEP P(X _{ α }→X _{ β }), where X _{ α } and X _{ β } are the codewords of s _{ α } and s _{ β }, respectively.
Figure 2 shows the PEP of the DDMprecoded system with BPSK modulation and different Hamming distances between information vector pairs. The normalized Doppler frequency is $f_{_{\text {D}}} T_{s} = 0.01$ and block length is 100. For a given Hamming distance D _{ H }, we randomly generate an information vector s _{ α }, then randomly change D _{ H } bits in s _{ α } to get s _{ β }. The PEP for a given Hamming distance is obtained by averaging over 10^{5} such information vector pairs. It can be seen that the PEP decreases as D _{ H }(s _{ α },s _{ β }) increases, and the system performance is dominated by error events with 1 bit information error.
Since the system performance is dominated by error events with 1 bit information error, we can get a BER lower bound of a BPSK modulated system by considering only error events with 1 bit error
In the equation above, the outer summation with respect to α is used to average over all $2^{N_{c}}\phantom {\dot {i}\!}$ possible values of s _{ α }. When N _{ c } is large, the averaging operation is timeconsuming. However, it can be evaluated by using a large number of randomly generated s _{ α } instead. It should be noted that our analysis can be easily extended to any MPSK modulated system by considering the dominant error events.
Simulation results
Analytical and simulation results are presented in this section to study the tradeoff between Doppler diversity and channel estimation errors and to validate the performance of the two precoding schemes in the presence of imperfect CSI. All systems employ a symbol rate of 0.1 Msym/s and operating at 1.9 GHz. The block length is N=100. When the movement speed is between 56.8 and 568.4 km/hr, the corresponding range of Doppler spread is between 100 Hz ( $f_{_{\text {D}}} T_{s} = 10^{3}$ ) to 1 KHz ( $f_{_{\text {D}}} T_{s} = 10^{2}$ ).
Figure 3 compares the performance between the two precoding schemes with both analytical and simulation results. For the rate $\frac {1}{N_{c}}$ repetition coded system, the analytical BER described in (26) serves as the lower bound for Doppler diversity system in the presence of channel estimation error, while the simulated BER is obtained by utilizing the optimum receiver described in (20). The rate 1 DDM system employs the suboptimum SISOBDFE receiver and the analytical results are obtained by using (44). The value of $f_{_{\text {D}}} T_{s}$ is 0.01, and the energy allocation factor is selected as η=0.1. The analytical BER of the rate 1 system matches very well with the corresponding simulation results. The analytical BER lower bound of the DDMprecoded system can reasonably predict the performance of the system with the suboptimum SISOBDFE receivers. For both systems, the pilot percentage has significant impacts on the BER performance. When δ=0.02, the BER performance is dominated by channel estimation errors, such that the Doppler diversity gain is very small. When δ=0.05, the repetition coded system can achieve the maximum Doppler diversity order as the system with perfect CSI. The BER of the DDM precoded system is 1.9 dB worse than the repetition coded system at 10^{−3} and δ=0.05, mainly due to the interference introduced by the rate 1 precoder. Similar performance is observed for highlevel modulations.
Similar results are given in Fig. 4 with higher modulation order. For the rate $\frac {1}{N_{c}}$ repetition code, the analytical BER is approximated by dividing the SER over log2(M), and a perfect match is observed between the analytical and simulation SER. For the rate 1 DDM scheme, the analytical results serve as the lower bound for the simulation results.
Figure 5 shows the analytical and simulated BER as a function of the energy allocation factor η for system with the repetition precoding scheme. The normalized Doppler spread is $f_{_{\text {D}}} T_{s} = 0.01$ and E _{ b }/N _{0}=16 dB. Simulation and analytical results match very well. For all system configurations, the BER is convex in η. The optimum η increases as δ increases. The optimum η is 0.06, 0.09, 0.1, 0.1 and 0.1 for systems with δ=0.01,0.02,0.05,0.1, and 0.5, respectively. A larger δ means more pilot symbols per slot; thus, more energy needs to be allocated to the N _{ p } pilot symbols in order to achieve a good channel estimation quality. Similar results are also observed for systems with DDM precoders. We will use the energy allocation factor 0.1 for all subsequent examples, and it provides a robust and good performance for all system configurations.
Figure 6 shows the analytical BER lower bound as a function of pilot percentage, under various values of $f_{_{\text {D}}} T_{s}$ . The energy allocation factor is η=0.1 and E _{ b }/N _{0}=16 dB. When δ is small, e.g., δ<0.05 for $f_{_{\text {D}}} T_{s} = 0.01$ , increasing δ slightly yields significant performance gains because the performance is dominated by channel estimation errors. Once δ reaches a threshold, e.g., δ≥0.1 for $f_{_{\text {D}}} T_{s} = 0.01$ , increasing δ further has very small impact on system performance, due to the fact that the impacts of channel estimation error is small once δ becomes large. It is interesting to note that increasing δ from 0.1 to 0.5 might degrade system performance because E _{ p } is a decreasing function in δ. More importantly, when δ is small (e.g. δ<0.02), a higher $f_{_{\text {D}}} T_{s}$ might result in a worse BER due to channel estimation errors. On the other hand, when δ is large enough such that the channel estimation quality is good enough, the BER performance increases with $f_{_{\text {D}}} T_{s}$ due to the higher Doppler diversity gains collected by the proposed systems.
To illustrate the impact of Doppler shift on error performance, the BER of system with repetition code is shown as a function of $f_{_{\text {D}}} T_{s}$ in Fig. 7, under various values of pilot percentage. The E _{ b }/N _{0} is fixed at 16 dB. For all systems, the BER is a quasiconvex function of $f_{_{\text {D}}} T_{s}$ . The performance first improves as $f_{_{\text {D}}} T_{s}$ increases due to the Doppler diversity and then degrades due to the channel estimation errors. For systems with δ = 0.02, 0.05, and 0.1, the minimum BER is achieved at $f_{_{\text {D}}} T_{s}$ = 0.004, 0.015, and 0.045, respectively. A higher pilot percentage may not always yield a better performance, and the performance improvement is not significant when the channel estimation is good enough.
Conclusions
The maximum Doppler diversity transmissions for highmobility systems in the presence of channel estimation errors have been studied in this paper. The tradeoff between Doppler diversity and channel estimation errors has been studied by using a repetition code and a rate 1 Doppler domain multiplexing scheme. The analytical performance of both systems have been obtained by analyzing the statistical properties of the channel estimation errors, and they quantitatively identify the impacts and interactions of a number of system parameters, such as the pilot percentage, the maximum Doppler spread, and the energy allocation factor between pilot and data symbols, etc. It has been shown that the error probability is quasiconvex in Doppler spread and monotonically decreasing in the pilot percentage. The performance of systems with a sufficiently high pilot percentage can approach that of a system with perfect CSI. On the other hand, if the pilot percentage is too low, the benefits of Doppler diversity are offset by channel estimation error such that a system with a lower Doppler spread could get a better performance.
References
 1
AM Sayeed, B Aazhang, Joint multipathDoppler diversity in mobile wireless communications. IEEE Trans. Commun. 47(1), 123–132 (1999). doi:10.1109/26.747819.
 2
X Ma, GB Giannakis, Maximumdiversity transmissions over doubly selective wireless channels. IEEE Trans. Inform. Theory. 49(7), 1832–1840 (2003). doi:10.1109/TIT.2003.813485.
 3
J Wu, in Proc. IEEE GLOBECOM. Exploring maximum Doppler diversity by Doppler domain multiplexing (San Fransico, USA, 2006), pp. 1–5. doi:10.1109/GLOCOM.2006.897.
 4
KI Ahmed, C Tepedelenlioglu, A Spanias, Performance of precoded OFDM with channel estimation error. IEEE Trans. Signal Process. 54(3), 1165–1171 (2006). doi:10.1109/TSP.2005.862898.
 5
J Wu, C Xiao, Optimal diversity combining based on linear estimation of Rician fading channels. IEEE Trans. Commun. 56(10), 1612–1615 (2008). doi:10.1109/TCOMM.2008.060598.
 6
S Savazzi, U Spagnolini, On the pilot spacing constraints for continuous timevarying fading channels. IEEE Trans. Commun. 57(11), 3209–3213 (2009). doi:10.1109/TCOMM.2009.11.080024.
 7
N Sun, J Wu, Maximizing spectral efficiency for high mobility systems with imperfect channel state information. IEEE Trans. Wireless Commun. 13(3), 1462–1470 (2013). doi:10.1109/TWC.2014.012314.130772.
 8
N Sun, J Wu, in Proc. IEEE GLOBECOM. Maximizing spectral efficiency with imperfect channel information in high mobility systems (Atlanta, USA, 2013). doi:10.1109/GLOCOM.2013.6831591.
 9
W Zhou, J Wu, P Fan, in Proc. IEEE ICC. Maximizing Doppler diversity transmisssions for high mobility systems with imperfect channel state information (Sydney, Australia, 2014). doi:10.1109/ICC.2014.6884267.
 10
J Wu, N Sun, Optimum sensor density in distortion tolerant wireless sensor networks. IEEE Trans. Wireless Commun. 11, 2056–2064 (2012). doi:10.1109/TWC.2012.041612.110127.
 11
N Sun, J Wu, Optimum sampling in spatialtemporally correlated wireless sensor networks,. EURASIP J. Wireless Commun. Networking. 2013, 5 (2013). doi:10.1186/1687149920135.
 12
N Sun, J Wu, in Proc. IEEE MILCOM. Minimum error transmissions with imperfect channel information in high mobility systems (San Diego, USA, 2013). doi:10.1109/MILCOM.2013.160.
 13
S Siwamogsatham, MP Fitz, JH Grimm, A new view of performance analysis of transmit diversity schemes in correlated Rayleigh fading. IEEE Trans. Inform. Theory. 48(4), 950–956 (2002). doi:10.1109/18.992794.
 14
J Wu, YR Zheng, Low complexity softinput softoutput block decision feedback equalization. IEEE J. Sel. Areas Commun. 26(2), 281–289 (2008). doi:10.1109/JSAC.2008.080205.
 15
A Stamoulis, GB Giannakis, A Scaglione, Block FIR decisionfeedback equalizers for filterbank precoded transmissions with blind channel estimation capabilities. IEEE Trans. Commun. 49(1), 69–83 (2001). doi:10.1109/26.898252.
Acknowledgements
The work of Weixi Zhou and Pingzhi Fan was supported by the Chinese 973 Program (No. 2012CB316100), the NSFC project (No. 61471302), the 111 Project (No. 111214), the MoE Key Grant Project (No. 311031100), and the Young Innovative Research Team of Sichuan Province (No. 2011JTD0007). The work of Jingxian Wu was supported in part by the U.S. National Science Foundation (NSF) under Grants ECCS1202075 and ECCS1405403.
Author information
Additional information
Competing interests
The authors declare that they have no competing interests.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0), which permits use, duplication, adaptation, distribution, and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Received
Accepted
Published
DOI
Keywords
 Doppler diversity
 Channel estimation error
 Highmobility communications