IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 57, NO.

3, MARCH 2010

703

Semi-Blind Most Significant Tap Detection for Sparse Channel Estimation of OFDM Systems
Feng Wan, Student Member, IEEE, W.-P. Zhu, Senior Member, IEEE, and M. N. S. Swamy, Fellow, IEEE
three classes: training-based, blind, and semi-blind methods. First, training-based methods, such as the least square (LS), employ known training signals to render an accurate channel estimation and thus require a consumption of spectral resources [2]. Blind channel-estimation algorithms, such as those proposed in[3]–[5], which exploit the second-order stationary statistics, correlative coding, and other properties normally have a better spectral efficiency. With the idea of both the training-based and blind algorithms, semi-blind channel-estimation techniques can potentially enhance the quality of channel estimation [6]–[8]. A wireless channel can often be modeled as a sparse channel in which the delay spread could be very large, but the number of significant paths is normally very small [9]–[18]. Broadly speaking, there are two kinds of approaches for the sparse channel estimation. The first one estimates the complex amplitude and the delay of each path based on a nonsampling spaced parametrical channel modeling [9]. The second kind is based on the sparsity assumption of the equivalent discrete-time channel [11]–[17], [19] in which only a few taps in the long-tapped delay line are considered most significant. By exploiting the sparse structure of the channel, some improved channel-estimation algorithms have been developed for OFDM systems [10]–[12], [17], [20], [21] and code division multiple access systems [15], [16]. In this paper, we are concerned with the development of a very efficient sparse-channel-estimation approach for OFDM and MIMO–OFDM systems that requires only a few OFDM symbols. As the nonsampling-spaced sparse channel estimation requires a large number of OFDM symbols for the estimation of the delay subspace, we focus only on the sampling-spaced approach. It should be mentioned that almost all of the sampling-spaced sparse-channel-estimation methods in the literature utilize a training sequence and follow two steps: 1) detect the positions of the most significant taps (MSTs), which are also referred to as the nonzero taps in some of the literature, and (2) obtain an improved channel estimate by exploiting the position of the MSTs. The key to these methods is the MST detection. Several techniques based on initial LS estimation have been developed for MST detection [10]–[12]. In these methods, an unstructured channel, i.e., a channel presumably having all nonzero taps, is considered in the initial LS estimation. Based on the preliminary estimate, the MST detection can simply be conducted by choosing a number of the largest amplitude channel taps [10]. To improve the detection accuracy, a generalized Akaike information criterion is utilized to estimate the MSTs in an iterative fashion [11], which is then simplified to a noniterative scheme [12]. The matching-pursuit (MP)-based MST detection method has also extensively been investigated by many researchers [13]–[16]. In the MP method, the MST is detected by sequentially selecting the column in a mixture

Abstract—In this paper, a very efficient semi-blind approach for the detection of most significant taps (MSTs) in sparse orthogonal frequency-division multiplexing (OFDM) channel estimation is developed. The least square (LS) estimation problem of sparse OFDM channels is first formulated, showing that the key to sparse channel estimation lies in the detection of the MSTs. An in-depth study of the second-order statistics of the signal received through a noise-free sparse OFDM channel reveals the sparsity and other properties of the correlation functions of the received signal. These properties lead to a direct relationship between the positions of the MSTs of the sparse channel and the most significant lags of the correlation functions, which is then used in conjunction with a pilot-assisted LS estimation to detect the MSTs in a semi-blind fashion. It os also shown that the new MST detection algorithm can be extended for the estimation of multiple-input–multiple-output (MIMO)–OFDM channels. A number of computer-simulation-based experiments for various sparse channels are carried out to confirm the effectiveness of the proposed semi-blind approach. Index Terms—Most significant tap (MST) detection, multiple-input–multiple-output (MIMO), orthogonal frequency-division multiplexing (OFDM), semi-blind, sparse-channel estimation.

NOTATIONS Throughout this paper, we adopt the following notations. Pseudoinverse. Kronecker delta. Transpose. complex-conjugate transpose; Circular convolution; Frobenius norm; Delta function; A stacking of the columns of the involved matrix into a vector. I. INTRODUCTION T IS WELL known that channel estimation is of crucial importance to orthogonal frequency-division multiplexing (OFDM) and multiple-input-multiple-output (MIMO)–OFDM systems [1]. Broadly speaking, channel-estimation techniques for OFDM or MIMO–OFDM systems can be categorized into

I

Manuscript received March 26, 2008; revised September 26, 2008. First published June 02, 2009; current version published March 10, 2010. This work was supported in part by the Natural Sciences and Engineering Research Council (NSERC) of Canada and in part by the Regroupement Stratégique en Microsystémes du Quebéc (ReSMiQ). This paper was recommended by Associate Editor W. Utschick. The authors are with the Center for Signal Processing and Communications, Department of Electrical and Computer Engineering, Concordia University, Montreal, QC H3G 1M8, Canada (e-mail: f_wan@ece.concordia.ca; weiping@ece.concordia.ca; swamy@ece.concordia.ca). Digital Object Identifier 10.1109/TCSI.2009.2023765

1549-8328/$26.00 © 2010 IEEE

by using the acquired MST positions. it is shown that the semi-blind MST detection algorithm can readily be extended to the sparse channel estimation of MIMO–OFDM systems. OFDM Signal Model For an OFDM system. which employs an active tap-detection criterion involving the correlation between the training pilots and the received signal.. is proposed. [19]. denoted as After adding a CP. Section III presents a semi-blind MST detection approach for the estimation of sparse OFDM channels. It is known that a wireless channel can often be modeled as a sparse channel that contains many zero taps in the uniform delay line [11]–[17]. very little work on blind MST detection is found in the existing literature. is referred to as the effective channel throughout this paper. there is only one cyclic-prefix (CP)-based blind method dealing with the MST detection of the sparse channel estimation for OFDM systems [17]. the analysis of the second-order statistics of the received signal passing through a sparse channel is conducted for OFDM systems. namely where denotes the number of subcarriers. In this case. Assume that the subcarriers. we first introduce the trainingbased LS channel-estimation algorithm for OFDM systems. the received frequency-domain signal after the DFT processing is given by Considering that OFDM systems are designed for broadband wireless communications. say.e. leading to a relationship between the positions of the MSTs and the most significant lags (MSLs) of the correlation functions of the received signal. the channel with respect to the th MST (namely. Section VI concludes this paper by highlighting some of the contributions presented. this detection scheme needs a large number of OFDM symbols as well as a large CP length in order to obtain precise MST positions. including the analysis of the second-order statistics of the received signal through a sparse channel and the development of a new MST detection algorithm. has been proposed in [19]. after removing the CP. the received signal can be described as (1) Then.704 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS. At the receiver. the transmitted th OFDM symbol can be written as a vector of the frequency-domain signals. the signal bandwidth is always larger than the coherence bandwidth. the sampled version of the continuous-time channel response[22]. the time-domain signal model for the frequency-selective fading channel is given by (2) where is the number of OFDM symbols within which the channel remains unchanged and is a spatiotemporally uncorrelated noise with zero mean and variance . Sparse LS Estimation Prior to the development of the sparse LS algorithm for the estimation of the effective channel. The inverse discrete Fourier transform (IDFT) processing gives the time-domain OFDM signal. Section II formulates the sparse LS channel-estimation problem. Based on this relationship. implying that the channel is frequency selective. an efficient MST detection algorithm. which requires only a small number of OFDM symbols and a very small number of pilots. from to . Unfortunately. The rest of this paper is organized as follows. Most of the existing channel-estimation methods work on the equivalent discrete-time channel. However. the nonzero tap) can be expressed as (3) are integers with . the available information of user data could be applied to both the MST detection and the channel estimation. [15] and the LMS method [19]. Thus. The transmitted and the received pilot vectors can be defined as where . the channel can be described by If the length of the CP is not less than the channel length . In this paper. First. Another MST detection. Section V comprises a number of experiments validating the proposed approach and showing the significant advantage of the sparse solution over the regular LS method. Finally. Section IV extends the new semi-blind MST detection algorithm to the MIMO–OFDM channel estimation. VOL. 57. i. MARCH 2010 matrix of the pilot signal that matches best the residual vector of the received signal. Furthermore. of each OFDM symbol carry the pilot signal. a refined channel estimate can be obtained in the second step. B. an LS estimation of the sparse channel vector with respect to the MSTs are derived for OFDM systems. Assuming that the channel is constant during a number of consecutive OFDM symbols. NO. To increase the spectral efficiency. This is usually implemented with the structured version of the training-based approaches such as the LS method [11]–[13]. 3. To distinguish from . It has been shown that the performance of these sparse-channel-estimation methods is much better than that of the unstructured channel-estimation alternatives. II. The common problem of the aforementioned sparse-channelestimation methods is that a large number of pilots is needed in order to render an accurate MST detection and channel estimation. Once the MST information has been acquired. each OFDM symbol is then sent out by the transmit antenna. our main objective is to develop an efficient semi-blind MST detection approach for sparse OFDM channel estimation. Then. To the best knowledge of the authors. the discrete-time channel can be modeled by an -tap finite-impulse response (FIR) filter. PROBLEM FORMULATION A.

. From (5). For the th OFDM symbol. .e.e. . Clearly. for example. . due to the sparse feature of most of the values of the channel. It is clear from (13) and (16) that as long as is a zero matrix.. we would like to express in terms of the effective sparse channel . channel as (17) . In the following. III. .. we obtain (13) .: SEMI-BLIND MST DETECTION FOR SPARSE CHANNEL ESTIMATION OF OFDM SYSTEMS 705 It should be noted that the pilot signal might not be located at be a the same position in each OFDM symbol. (4) can be rewritten as (8) where and is a matrix whose th column is the th column of . the nonzero elements of occur only when . (15) Here. are zero. . we first analyze the second-order statistics of the signal received through the sparse channel and then propose a novel semi-blind MST detection algorithm. the key to the structured LS estimation of the effective channel lies in the detection of the MSTs. [26]. the value of mainly depends on . Using (3) and assuming that the MSTs are correctly estimated. .. one can form a matrix. . i. . the received frequency-domain pilot signal with respect to OFDM symbols can be obtained as (5) where where section. by taking only the rows of associated pilot subcarriers. Let matrix formed by the first columns of a DFT matrix . one can obtain an LS minimization problem. . is zero. . Due to the fact that . i. . Since the channel is sparse. which can be. . . (6) whose solution is given by (7) In the aforementioned LS channel-estimation method as well as in many existing OFDM channel-estimation methods such as those in[23] and [24]. and represent the frequency-domain noise corresponding to in (2). . . . (5) can be re-expressed as (9) where . (10) where (14) . the channel estimate obtained is not efficient when the channel is sparse. [25]. In other words. defined as (12) In this section. in general. is a zero matrix for most of the values of ranging from 0 to .WAN et al. . In the following is a lower triangular matrix for . i. in the absence of noise. Thus. . Obviously. Using (2) and (3) in (12). the sparse case of the wireless channel has not been taken into consideration. . SEMI-BLIND MST DETECTION ALGORITHM A. we would like to propose a sparse LS solution for OFDM channel estimation.e. an LS estimate of the sparse channel with respect to the MSTs can be obtained as (11) Obviously. and show that has only a few MSLs. Moreover. It was shown in [2] that with the (4) . . . . From (4). Accordingly. we can rewrite (15) as .. there is only a small number of choices of which make a nonzero matrix. the index has been omitted for the sake of notational convenience. Second-Order Statistics of the Received Signal Through Sparse Channel It is well known that the correlation function of the received plays a crucial role in blind or semi-blind signal vector channel estimation[5]. (16) From (9). By substituting (14) and can be expressed in terms of the effective (16) into (13).

Case D3. which are common to all the eight cases. are not included in Table I for notational simplicity. corresponding to and .2 lies in the number of MSLs in addition to the expression of . and in (16) cannot be unity at the same time. there are only two nonzero effective channel elements and .. in terms of the effective channel matrix for the cases . the corresponding two potential MSLs are overlapped.1: if . From (17). by using a if if if otherwise. . all the potential MSLs can be written as . and the expression of and . thus. When . . As such. i. both and could be unity simultaneously. When . When some of the values of . Using (17). it is clear from (17) that has different expressions depending on the relationship between and .e. All other MSLs can be determined from the relationship among . and their differences. (21) It is seen that the main difference between the two cases D3. and. As seen in the next subsection. NO. The two common MSLs of the two cases are and . i. It will be shown through simulation studies that.2: if (20) if if if if otherwise. one can easily obtain In the case of to the first and the last MSLs as with respect (22) (23) which is similar to the cases of and . and happen to be identical. Using (17). For example. yielding the only MSL. Evidently. and . the two presumable MSLs overlap. B. As some of these values could be identical as discussed for the case of . each corresponding to . for any value of . 57. VOL. the structure of can be exploited to detect the MSTs of the sparse channel. . one can find that there are only two nonzero values of . and . However. Let us consider the simplest case when . 3. Noting that . . the corresponding two ’s are combined. the position of the MSLs are dependent on the values of . the number of the MSLs would be reduced. . In this case. i..e. Detection of MSTs We now propose an MST detection algorithm based on the structure of as analyzed in the previous subsection. . one can obtain the following results. MARCH 2010 TABLE I EXPRESSION OF CORRELATION FUNCTION r(l) FOR D = 4 In obtaining (17). one can have a total of eight possible cases as summarized in Table I.1 and D3. a table similar to Table I can easily be designed by simple computer programming.2 corresponds to a possibility that the MSLs at and are overlapped as well as the MSLs at and are overlapped. if equals . If is not identical to .706 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS.e. Note that the correlation matrices with respect to the first and the last MSLs. The idea is to determine the MST position based on the MSL position of . Other MSLs depend actually on the relationship between and . We have obtained earlier the expressions of . has been assumed without loss of generality.. (18) (19) which means that the MSL positions of are . Case D4. .e.. Case D3. In the following. i. respectively. Case D4. we disclose the relationship between the MSTs with the MSLs. where . the number of the MSLs of can be different. . Then. . leading to two different MSLs. . .8 indicates a possibility of having a maximum of seven MSLs of including the first and the last ones. If any two of ’s are identical. a unit signal variance.

First.1: When with .WAN et al. there is only one additional tap to be determined.1: .1 in Table I. If Then. The correlation function of timated version can be estimated as (24) where for . whose positions are denoted as with . which gives and . our new MST detection method has a high detection accuracy. when the test is “yes. which yields and . which gives three possible MST detection results. Note that in the noisy case. In the case of . . we have and .3. indicating a possibility of misdetection of Case W4. 1 is already the improved version of a straightforward implementation. Obviously. we should have case D3. respectively. respectively.: SEMI-BLIND MST DETECTION FOR SPARSE CHANNEL ESTIMATION OF OFDM SYSTEMS 707 small number of OFDM symbols and pilot subcarriers. Let be the potential MSTs of the sparse channel and is a matrix. In the following. the circular convolution (2) in the noisy case can be rewritten in the matrix form as (25) Substituting (25) into (24) and using (13). we obtain and . instead of computing all the three costs . a nonzero value even if . it is clear that the positions of the first and the last MSTs are simply given by (29) If .2: When with . The complete scheme is shown in Fig. which could not only save the calculation of one cost but also enhance the detection performance by excluding the false MST candidate Case W4. one can have the following three possible solutions. we present an LS criterion similar to the sparse LS channel-estimation algorithm proposed in Section II-B. (28) where is a predetermined length. denoted as .e. (29) and (30) give the only two nonzero taps if . whose th column is the th column of . Accordingly. is found to be much smaller than . while the second and the third terms are the errors introduced by the signal and noise perturbations. . (32) Evidently. i. the cost reaches its minimum when gives true MSTs. Case W4. would be dominant provided that the signal-to-noise ratio (SNR) is not extremely low.3 from comparison. Assume that a total of MSLs of have been detected. Then. . By letting (30) Obviously.1 (Case W4. Second. a test on is performed to possibly reduce the three candidates to two. in general. . the information on their positions can be utilized to determine the MST position of the sparse channel as discussed in the previous subsection. Using the idea of joint channel estimation and zero detection proposed in[27]. and Case W4.3.1. However. and and making a detection among the three values. readily have In order to make a decision among the three choices.1 (Case W4.2. is.2 from (21) or D4.2) as Case W4. . which is readily given by where for . namely: Case W4.3: When with . 1. we have (26) where (27) represents the signal perturbation and is the perturbation error introduced by the noise. . and the coefficient is used to adjust the average absolute value of . We now determine the possible ’s according to the structures of for different values of and .2: .” we have found that could be very close to if Case W4. W4. we determine from where . Once the MSLs of are acquired. Case W4. and W4. It should be stressed that the detection scheme in Fig. at a low level of SNR.3 is true. we have only the case D3. Case W4. On the other hand.2) is true. Case W4.. our MST detection only needs to calculate a few costs as given by (31) with respect to Cases W4. the first term in the right-hand side (RHS) of (26) is an ideal correlation matrix of the received signal vector with neither the signal perturbation nor the noise corruption. a cost function for the new MST detection method can be established from (9) as (31) where and . we first introduce the esof .1 according to (20). Although we have in general . and .3: . Prior to developing the new method. The second step of the new MST detection algorithm can be described as follows. the first step of our method is to detect the MSLs of by comparing its absolute value with the threshold. we and . when Case W4. the MSLs can be detected from the norm of since the first term in the RHS of (26).

4: . NO.5: . the conventional methods need a large number of pilots or OFDM symbols to conduct a training-based or CP-assisted MST detection. the MSLs might be incorrectly detected. 1. . and . we have employed a scaling coefficient to distinguish W4. For each of these combinations. in some extremely noisy cases.1: . when . In the improved detection scheme. Case W5. for a moderate SNR. and . when and . our technique uses only a small number of OFDM symbols and pilots to implement an accurate detection. . . . as shown in Fig. and . and . . For example. which normally require searching all the possible MST candidates over the entire channel length . which would lead to the cases that there are no corresponding MST detected due to the logic in the MST detection scheme. First. . for the estimation of MSTs. . SPARSE-CHANNEL ESTIMATION OF MIMO–OFDM SYSTEMS In this section. The final decision can be made based on the minimum of all the combinations. In this case. . and . Case W5. 1. More importantly. by exploiting the MSL information of . Case W5. when and . 3. Clearly. a minimum cost can then be obtained by using the aforementioned proposed scheme. for . The corresponding MST detection scheme is described in Fig. together with a small number of pilots.5: . MST detection scheme for W = 4. we can use a joint MSL and MST detection scheme. the proposed detection method reduces the search range to a very small number of MST candidates. when and . Case W5. when . Overall.2: . 2. MST detection scheme for W = 5. . the channel coefficients are estimated in an LS fashion as given by (7). 2) The estimated MSLs are utilized.1 and W4. and . 2. Fig. which gives the following five detection results. and Case W5. . 57. we have found that is a proper choice. . 3) Based on the estimated MSTs.708 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS. In contrast. and . The proposed sparse LS channel-estimation approach can be summarized as follows. This table can easily be constructed even if has a large value. since the expression of can simply be derived by computer programming as stated before. therefore. .3: . Consider a MIMO–OFDM system with the vertical Bell Laboratories layered space-time . Case W5.2: . the key to the aforementioned detection scheme is to design a table listing all the possible MST candidates corresponding to the same value of using the expression of obtained in the previous subsection. The value of should be chosen according to the level of SNR. Similarly. Case W5. a relatively small threshold is used to get more candidate lags. Case W5. VOL. leading to all possible combinations of different values of and positions of lags. and . we would like to extend the proposed semi-blind MST detection algorithm for sparse channel estimation of MIMO–OFDM systems. Case W5. one can have the following possible situations: Case W5. IV. .4: . in comparison with many of the MST detection methods available in the literature.3: .3 from W4. MARCH 2010 Fig. 1) MSLs are first estimated by comparing the absolute value of the correlation function with the threshold calculated from (28). Note that.2. It will be shown in Section V that the performance of the proposed sparse LS channel-estimation algorithm is significantly superior to that of the regular LS method.1: .

respectively. In the following. .. the nonzero tap) can be expressed as (34) are integers with . where represents the channel response between the th receive antenna and the th transmit antenna. . . .: SEMI-BLIND MST DETECTION FOR SPARSE CHANNEL ESTIMATION OF OFDM SYSTEMS 709 (V-BLAST) structure. . MSE versus SNR. . in which all elements of are assumed to have the same MSTs. 3. an LS estimate of the sparse channel can be obtained as (40) (35) where and are the transmit and received signal vector in the link between the th transmit and the th receive antennas. . which consists of transmit and receive antennas. . In this case. the second-order statistics of the received signal for the MIMO–OFDM systems is a correlation matrix. . which can be. . it can be proved that (38) where the relationship of the terms are given in (39) at the bottom of the page. . the channel matrix with respect to the th MST (namely.. . and . we first derive a sparse LS algorithm for the effective channel matrix . where .WAN et al. . Here. By defining . . Using (34) and (35). From (38). The MIMO version of (4) can be easily obtained as where Fig. (37) and using a manner similar to the derivation of (9). written as (41) where (36) (42) . we consider the extension of the proposed semi-blind MST detection algorithm for the MIMO–OFDM systems. . . Accordingly. . . in general.. . the th tap of the MIMO channel can be described by a channel matrix as . . the time-domain signal model (2) in single-input–single-output (SISO) OFDM systems can be modified for the MIMO case as (33) where is the time-domain signal with respect to the th transmit antenna and and are the time-domain signal and noise with respect to the th receive antenna. (39) . . one can obtain Now. . . In contrast to the correlation function for OFDM systems. we assume a sparse MIMO channel. Then. .

2) Instead of using (28). 3. SIMULATION RESULTS A. . The constant in (28) is set to 0. MARCH 2010 In the previous two equations. One can find that when only five OFDM symbols are used. The simulation involves 2000 Monte Carlo runs of the transmission of ten OFDM symbols with 30 pilot subcarriers. (50) . the MSE performance versus the SNR is investigated. In our simulation. Experiment A2: MSE Versus Pilot Length: In this experiment. Fig. . for MIMO–OFDM channel estimation. respectively. OFDM Channel Estimation We consider an OFDM system in which the number of subcarriers is set to 1024. . Fig. . channel matrix V. Substituting (42) into (41) and using (33) and (34). Fig. NO. the sparse LS method can achieve a gain of 6 dB over the regular LS method. the QPSK modulation is used.i. yielding a sparse LS estimate of the effective . and is given in the equation shown at the bottom of the page. in particular. When 25 OFDM symbols are used. Experiment A1: MSE Versus SNR: Here.8. the proposed method is more advantageous for a smaller number of OFDM symbols used.d. the MIMO channels can be estimated in the LS sense. . and (16). . the symbol index has been omitted. it is observed that the performance improvement from the sparse LS method becomes 4.710 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS. one can prove that (43) can be rewritten as (47) By comparing (47) with (17). Obviously. it achieves a gain of about 5. (46). in Table I is replaced with . Noting that . where is the number of Monte Carlo iterations and and are the true and the estimated channel vectors with respect to the th Monte Carlo iteration. Accordingly. in which each tap is independent identically distributed (i. 1) The symbols and in the SISO OFDM case should be replaced by and . It is seen that the sparse LS method performs significantly better than the original LS method. Therefore.5 dB over the regular LS version when the SNR is 10 dB. One can find that the sparse LS method can achieve about 6-dB gain over the regular LS method. . we obtain (43) where (44) Once the MSTs are obtained by using the proposed semi-blind algorithm. 3 shows the MSE plots of the proposed sparse and original LS methods.. the cost function in the MST detection can be established from (38) as (49) where . as shown in the following. we investigate the channel-estimation performance versus the number of pilot subcarriers. . 5 shows the MSE plots from 1000 Monte Carlo iterations for 30 pilot subcarriers with OFDM symbols varying from 5 to 25 at an SNR of 10 dB. . . as given by (40). where has the same definition as in Section III-B. (45) can be rewritten as (46) Using (44). The MSTs are . (45) with . the threshold used for MSL estimation is calculated by (48) 3) Instead of using (31). the proposed semi-blind MST detection algorithm in Section III can be applied to channel estimation of MIMO–OFDM systems after modifications. . 4 shows the MSE plots from 2000 Monte Carlo iterations for five OFDM symbols with pilot subcarriers varying from 15 to 35 at an SNR of 10 dB. . .) . and the length of the CP is 30. respectively. we investigate the channel-estimation performance versus the number of OFDM symbols. and. 57. VOL. The estimation performance is evaluated in terms of the MSE of the channel estimate as given by . and a sparse Rayleigh channel modeled by a three-nonzero tap FIR filter is assumed. and complex normal distributed.9 dB. one can find that the difference and are between the two equations is that the scalers replaced by the matrices and . . Experiment A3: MSE Versus the Number of OFDM Symbols: In this experiment.

5 dB. Specifically. The channel is modeled by a four-tap sparse Rayleigh channel with . that the MSE of the sparse method is slightly worse than that of the ideal sparse method. An ideal sparse LS method with the knowledge of the true MST information is also simulated for comparison. we consider the proposed sparse LS method with the MST detection. which has a similar sparsity as in the previous channel. 5. each tap corresponds to a random matrix whose elements are i. MIMO–OFDM Channel Estimation Based on the same condition for the OFDM system as in the previous subsection. Fig. 6. an additional gain in the channel-estimation performance can be expected. 6 shows the MSE plots from the proposed as well as the ideal sparse methods along with two regular methods. The error probability of the MST detection versus SNR is shown in Table II. 7. the error probability is . 4.d. and . The training sequence is generated using the method in [2]. and . the LS sparse solution is significantly better than the two regular methods. MSE versus the number of pilot subcarriers. Fig. of which 30 are used as pilot for training purpose. which does not considerably affect the MSE of the channel estimation for most of the SNR levels. It is also noted that the proposed sparse method significantly outperforms the two regular methods. the bit error rate (BER) performance of the MIMO–OFDM system is investigated by using the estimated channel matrix and an ordered V-BLAST decoder. we consider now a MIMO–OFDM system with two transmit and four receive antennas. i. We now show the estimation result of another channel. Fig. for this exponential channel. It is interesting to see such a phenomenon. by utilizing the proposed MST detection algorithm in conjunction with our previously developed semi-blind channel-estimation method in [8]. This is because the MSE plot is the averaged result from 20 000 Monte Carlo iterations. . TABLE II ERROR PROBABILITY OF THE MST DETECTION VERSUS DIFFERENT SNR B. It is seen that the sparse LS method is highly consistent with its ideal version.WAN et al. [8]. the error probability is very small. Experiment B2: BER Versus SNR: Now.e. complex Gaussian variables with zero mean and unit variance. meanwhile.i. It is seen that the ideal and the proposed sparse LS methods deliver almost the same estimation results. the channel is considwith ered as Rayleigh distributed. the sparse LS method is consistently superior to the original method by nearly 6. . the channel estimation performance in terms of the MSE as a func- tion of the SNR is investigated for a three-tap sparse channel . . . The simulation involves 20 000 Monte Carlo runs of the transmission of one OFDM symbol at 1024 subcarriers. The . First.: SEMI-BLIND MST DETECTION FOR SPARSE CHANNEL ESTIMATION OF OFDM SYSTEMS 711 Fig. However. In particular. and . MSE versus the number of OFDM symbols. It is clear that a high performance of the MST detection is achieved despite the fact that only one OFDM symbol is used. As for the estimation of the effective channel. For the purpose of comparison. but undergoes an exponential decay with variances of the three significant taps as . MSE versus SNR for a Rayleigh fading channel.. the channel is first estimated by the regular LS method and our previously developed semi-blind method[7]. when dB. It is of interest to mention that. Experiment B1: MSE Versus SNR: In this experiment. It is seen from Fig.

W. 12.-S. ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers. and T. vol. Rauhut. no. and M. no. The semi-blind MST detection algorithm has been extended to MIMO–OFDM channel estimation. pp. 1126–1136. 1. pp. [6] Y. Fig. I. W. For example. Tanner. pp. Commun. 2. no. 11. 53. pp. J. “An investigation into time-domain approach for OFDM channel estimation. Ng. 7. F. Bhashyam. 2007. 3239–3242. Jan. 6.” IEEE Trans. Bhargava. no. no. no. no. [1] Y.. 56. 631–635. vol. N.” IEEE Trans. S. Lee. 3. pp. 2007. NO.” IEEE Signal Process. Lin. 784–789. no. Technol. vol. Signal Process. 2006. “Linear prediction based semi-blind channel estimation for MIMO–OFDM system. [5] Y. H. [18] G. “Blind channel estimation for MIMO–OFDM systems via nonredundant linear precoding. Commun. R. R. Raghavendra. and J. Dec. Nov. 670–685. E90-B.. Lett. Cotter and B. and J. Jun. Nallanathan. Zhu. “Estimation of fading channels with large possible delay spreads. Briefs. and K. Raghavendra and K. 151. Jayalath.. S. 56. 2007. 1571–1579. K. Iltis. vol. [13] S. Borah. “A semi-blind channel estimation approach for MIMO–OFDM systems. IEEE ISCAS. vol. no. Jung. Clearly. I. vol.. Deng. VOL. pp. Exp. pp.-P. A. Jul. 7. K. no. [15] S.” IEEE Trans. pp.-A. D. K. Reg. 52–55. Papers. Hwang.. pp. “Parametric channel estimation for pseudo-random tile-allocation in uplink OFDMA. Fig. pp. Swamy. MARCH 2010 the new approach does not require estimating all the channel taps. “Estimation of frequency-selective CDMA channels with large possible delay and Doppler spreads. Signal Process. no. 2008.” IEEE Trans.” IEEE Trans. no. no. 55. 2004. it has saved a large amount of computations compared with regular channel-estimation methods. Pfander. 2. S. simulation involves 4000 Monte Carlo runs of the transmission of one OFDM symbol with 20 pilot subcarriers. F. Signal Process. “Improving channel estimation in OFDM systems for sparse multipath channels. D. vol. Circuits Syst. Gao and A. Jul. no. Giridhar. Their valuable comments and suggestions have greatly helped improve the presentation of this paper. 2006. pp. vol. Giridhar. pp. This relation has then been exploited to develop a highly efficient semi-blind MST detection algorithm that requires only a small number of OFDM symbols and pilot subcarriers for an LS-based detection. K. 51. As . L. “Design and implementation of MIMO–OFDM baseband processor for high-speed wireless LANs. R. Borah. Heath. Barhumi. Jan. no.. Jul.-P. the performance improvement provided by the proposed sparse LS method gets more prominent compared with the regular one as SNR increases. Lett. “Highly efficient sparse multipath channel estimator with Chu-sequence preamble for frequency-domain MIMO DFE receiver. M. Lior. pp. [3] C. N. “Identification of matrices having a sparse representation. Powers. Eng. Zeng. Jul. Jan. MSE versus SNR for a channel with exponential decay. 2821–2834. 8. W.” IEEE Trans.” IEEE Signal Process. 4. W. “A matching-pursuit/GSIC-based algorithm for DS-CDMA sparse-channel estimation. S.. Kim and R. Tsai. G. 46. Reg. H. 50. Leus. 2004. 240–248. Lett. 11. N. Computer simulations based on various sparse channels have confirmed that the proposed sparse-channel-estimation approach significantly outperform the regular OFDM and MIMO–OFDM channel-estimation techniques without utilizing the MST detection. Kim. [9] M. [8] F. [14] Y.. Veh. pp. Jan.” Proc. “Semi-blind channel estimation and equalization for MIMO space-time coded OFDM. Technol. and M. Papers. 463–474. 2007. Yoon. 8. II. [11] M. Commun. 2005. 2103–2110. Signal Process. 2003. no. Aug. pp. E. Nov. 1615–1624. 55. [16] D. Circuits Syst. 12–15. 11.” IEEE Trans. vol. [10] H.” Electron. vol.” in Proc. 2000. [17] C. and M. 2008. 4. the performance gain of the proposed sparse LS method is boosted to 24 from 3 dB. Dec.” IEEE Trans. Lam. BER versus SNR.” IEEE Trans. 2. 2007. 5370–5381. S. Mar. [2] I. Inst. Chung. Athaudage and A. 6. Feb. 559–566. Wan. vol. R. Broadcast. Signal Process. [4] F. Mar. 8 shows the BER performance of the proposed sparse LS method as well as its original version. E. H. 2002.712 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS. vol. Liu and D.. no. vol. and E. the Associate Editor.. pp.. 1. REFERENCES Fig. 3. pp. S. J. 2003.” IEICE Trans. Chen and C. Minn and V. 39.” IEEE Trans. 2008. 6. 2007. VI. “Optimal training design for MIMO–OFDM systems in mobile wireless channels. vol. and the Editor-in-Chief for their careful review of this paper. “Delay-spread estimation using cyclic-prefix in wireless OFDM systems. when SNR is increased to 17 from 5 dB. vol. 7. Zhu... “Blind channel estimation for MIMO–OFDM systems.. 56. H. 57. 5376–5388. Shin. A relationship between the MSTs of the sparse channel and the MSLs of the correlation function of the OFDM signal received in the absence of noise has first been revealed. “Blind-channel identification for MIMO single-carrier zero-padding block-transmission systems.” IEEE Trans. 54. vol. “Sparse channel estimation via matching pursuit with application to equalization. 130–131. Circuits Syst.” IEEE Trans. [12] J. Moonen. 1. 55. vol. Kim. 374–377. Elect. [7] F. pp. CONCLUSION A semi-blind MST detection approach has been proposed for the sparse channel estimation of OFDM and MIMO–OFDM systems. 55. Rao. and J. Veh. R. Swamy. Wan.

OFDM. He currently serves as an Associate Editor of Circuits. Feb. He was a Postdoctoral Fellow from 1991 to 1992 and a Research Associate from 1996 to 1998 with the Department of Electrical and Computer Engineering. the M.WAN et al. I. Wuhan. Papers. 53. Lett. 2003. and M. the Chinese Government Award for Outstanding Self-Financed Students Abroad from China Scholarship Council. [26] K. Hoang. During 1993–1996. Y. 1992). Calgary.” Electron. respectively. Vice-President in 1976. vol. and the M. N. “Subspace-based blind channel estimation for OFDM by exploiting virtual carriers. 3. L. Digital Communications. and was also its coordinator for Concordia University. Analysis and Design (Prentice Hall International U. 1. Wireless Commun. and the Ph. no. QC. In May 2009. 141–150.. “Sinusoidal modeling and adaptive channel prediction in mobile OFDM systems. Montreal. 1. and the Postdoctoral Fellowship from Natural Sciences and Engineering Research Council of Canada. Reg. QC. vol. Simeone. Aug. statistical estimation theory and its applications in MIMO. and the International Tuition Fee Remission Award from Concordia University.-A. he has been a Research Assistant with Concordia University. Saskatoon. 8. the Graduate Fellowship. He received the Doctor of Science (Honoris Causa) degree in engineering by Ansted University “in recognition of his exemplary contributions to the research in Electrical and Computer Engineering and to Engineering Education. Editor-in-Chief of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-I from June 1999 to December 2001. Swamy (S’59–M’62–SM’74–F’80) received the B. speech and image processing. Canada. no. Networks and Algorithms (Wiley. degree in electrical engineering from Wuhan University. I.” IEEE Trans. 2008. 1783–1791. [21] I. degree in electrical engineering from Southeast University. Montreal. President in 2004. N. In August 2001. circuits. during which time. of Posts and Telecommunications. the Institution of Engineers (India). 4. respectively.) degree in mathematics from Mysore University. Canada. Golden Jubilee Medal in 2000. China. Homer. andNeural Networks in a Softcomputing Framework (New York: Springer. Program Chair for the 1973 IEEE Circuits and Systems (CAS) Symposium. 424–435. China. Inouye. 55. Canada. “Blind identification of MIMO channels using optimal periodic precoding. “Sparse channel estimation with zero tap detection. Reg. Proakis. Concordia University instituted the M.E. He is currently a Research Professor and the Director of the Center for Signal Processing and Communications with the Department of Electrical and Computer Engineering. Jan. 2006). Canada.” Since July 2001. Vice-Chair for the 1999 IEEE CAS Symposium. Roy. Circuits Syst. He has also taught with the Electrical Engineering Department. Wei-Ping Zhu (SM’97) received the B. 901–911. Past-President in 2005.. pp. he holds the Concordia Chair (Tier I) in Signal Processing. vol.D. vol. Swamy is a Fellow of a number of professional societies including the Institute of Electrical Engineers (United Kingdom). in 1957. no. Oliver. 830–832.Sc. and C. Aravind.K. Papers. pp. He is the coauthor of two book chapters and four books: Graphs.” IEEE Trans. Chinese translation—Education Press. “Pilot-based channel estimation for OFDM systems by tracking the delay-subspace. in 2009. in 2000. Wong and B.D. Y. He was also the recipient of the Doctoral Thesis Completion Award. He has served the IEEE in various capacities such as the President-Elect in 2003. Wireless Commun. M. in 1982 and 1985. and Ph. 1995). vol. QC. including the Education Award in 2000. Jun. and M. he was an Associate Professor with the Department of Information Engineering. 44. Wan is the recipient of the Best Paper Award at the 2008 IEEE International Conference on Neural Networks and Signal Processing. He is the recipient of many IEEE-CAS Society awards. M. Associate Editor of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS during June 1985 to May 1987. Dr. degrees in electrical engineering from the University of Saskatchewan. 2. the Engineering Institute of Canada. Kohno. Circuits Syst. He has published extensively in the areas of number theory. I. [20] J.. Systems and Signal Processing. 5. (Hons.” IEEE Trans. S. S. Canada.E. vol. 1.E.-S. Nanjing. NS. Concordia University. Beijing. and with the University of Calgary. Mysore. and U.” IEEE Trans. Wireless Commun. Evans. in 2003. and the Institution of Electronic and Telecommunication Engineers (India). Switched Capacitor Filters: Theory. Since July 2001. Technical University of Nova Scotia.. and signal processing for wireless communication. 1981. systems and signal processing.D. Zhu was an Associate Editor of IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS PART I: FUNDAMENTAL THEORY AND APPLICATIONS from October 2001 to December 2003. Taiwan. He was a founding member of Micronet from its inception in 1999 as a Canadian Network of Centers of Excellence until its expiration in 2004. 1995. pp. the Diploma degree in electrical communication engineering from the Indian Institute of Science. [27] C. General Chair for the 1984 IEEE CAS Symposium. Mareels. AB. “Enhanced detection-guided NLMS estimation of sparse FIR-modeled signal channels. Nanjing. 315–325. Lin and Y. I. 4. Prabhu. and a member of the Board of Governors of the CAS Society. 3rd ed. 2007. 1601–1615. China. 13. no. degree in electrical engineering from Concordia University. Since 2004. 19. Canada. [24] O. Feng Wan (S’09) received the B.” IEEE Trans. in 1991. 2008. Chinese Academy of Sciences. Halifax. Bangalore. Vice President (Publications) during 2001–2002. [22] J. C. 2008. no. including Nortel Networks and SR Telecom Inc. no. in 1954. 6. as well as his dedication to the promotion of Signal Processing and Communications Applications. R. he worked in hi-tech companies in Ottawa. Montreal. Papers. G. Reg. Circuits Syst. Mitra. Swamy Research Chair in Electrical Engineering as a recognition of his research contributions. degree in electrical engineering from the Institute of Electronics. vol. filter banks.E. pp. Ltd. and K. His research interests include the areas of digital filters. Bar-Ness.. Apr. pp. 54. S. 1743–1763. Kawamoto. Carbonelli. India.” IEEE Trans. Concordia University. May 2007. In 2008. “Sparse channel estimation in OFDM systems by threshold-based pruning.: SEMI-BLIND MST DETECTION FOR SPARSE CHANNEL ESTIMATION OF OFDM SYSTEMS 713 [19] J. India. no. Apr. Vedantam. [25] C. Concordia University. Jan. New York: McGrawHill. Signal Process. [23] C. and is the holder of five patents. and smart antenna wireless communications. 1987). degrees in electrical engineering from Nanjing University of Posts and Telecommunications. in 1960 and 1963. His research interests are in the area of digital signal processing. and the Ph. pp. 2006. he was awarded the title of Honorary Professor by the National Chiao Tong University. 56.. Dr. 1984. Russian translation—Mir Publishers. vol. Spagnolini. where he served as the Founding Chair of the Department of Electrical Engineering from 1970 to 1977.” IEEE Trans. From 1998 to 2001. 2004. M. Graphs: Theory and Algorithms (Wiley. “A matrix pseudo-inversion lemma for positive semidefinite Hermitian matrices and its application to adaptive blind deconvolution of MIMO systems. he has been with Concordia University’s Electrical and Computer Engineering Department as an Associate Professor. and the 1986 Guillemin-Cauer Best Paper Award. Li and S. University of Saskatchewan. Nanjing University .Sc. and Dean of Engineering and Computer Science from 1977 to 1993. Chen. he developed the faculty into a research oriented faculty from what was primarily an undergraduate one. pp. as well as with the Department of Mathematics. and U. Dr. pp. no.