You are on page 1of 8

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO.

4, APRIL 2002

937

On Maximum-Likelihood Detection and Decoding for Space-Time Coding Systems


Erik G. Larsson, Member, IEEE, Petre Stoica, Fellow, IEEE, and Jian Li, Senior Member, IEEE
AbstractSpace-time coding (STC) schemes for communication systems employing multiple transmit and receive antennas have been attracting increased attention. In this paper, we address two interrelated problems: detection of space-time codes under various interference conditions and information transfer from the STC detector to an error-correcting channel decoder. By taking a systematic maximum-likelihood (ML) approach to the joint detection and decoding problem, we show how to design optimal detectors and how to integrate them with a channel decoder. We also discuss various aspects of channel modeling for STC communication receivers. In particular, while many previous works on space-time coding assume that the channel is a stochastic quantity, we find that a deterministic channel model can have some advantages for the receiver design. Finally, we illustrate our results by numerical examples. Index TermsInterference suppression, maximum-likelihood estimation, maximum-likelihood sequence detection, MIMO systems, space-time coding, soft information.

Fig. 1. MIMO communication system with concatenated space-time coding and channel coding.

I. INTRODUCTION OMMUNICATION systems utilizing multiple transmit and receive antennas, which are referred to as multiple-input multiple-output (MIMO) or space-time coding (STC) systems, have received a significant amount of attention in recent years. In this paper, we consider the communication system depicted in Fig. 1. A space-time code is concatenated with an outer error-correcting channel code, and an interleaver is employed in a standard manner between the outer channel code and the space-time code. An important topic in the design of a communication system is the optimal transfer of information from the STC detector to the channel decoder. For single-input single-output (SISO) systems, this topic is mature and well understood; see, for instance [1][3] for a treatment of systems with multiple concatenated error-correcting channel codes. The simplest way of transferring information from the detector to the decoder is to transmit so-called hard decisions on each received symbol. However, a significant performance gain can be obtained by transmitting information about the symbol decision reliability as well. This reliability information is usually referred to as soft decisions.
Manuscript received March 8, 2001; revised December 10, 2001. This work was supported in part by the National Science Foundation under grant CCR0097114 and the Swedish Foundation for Strategic Research (SSF). The associate editor coordinating the review of this paper and approving it for publication was Dr. Alle-Jan van der Veen. E. G. Larsson and J. Li are with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32611 USA (e-mail: larsson@dsp.ufl.edu). P. Stoica is with the Department of Systems and Control, Uppsala University, Uppsala, Sweden. Publisher Item Identifier S 1053-587X(02)02384-X.

In this paper, we take a unified approach to interference-resistant detection of symbols transmitted over an unknown MIMO channel and optimal (in an ML sense) information transfer from the STC detector to the channel decoder. In particular, we will see how the concept of soft information is transferred to the MIMO case. While considerable efforts have been spent on optimal code-design for space-time coding systems [4][8], optimal detector design and interference suppression for MIMO systems is only beginning to attract attention (see, e.g., [9]). Despite this fact, the topic is of fundamental importance since it is well known that all practical communication systems will inevitably suffer from both multiuser (co-channel) interference and interference from external microwave sources. We consider so-called quasistatic fading, and hence, the channel is assumed to be constant only during the block for which we formulate the ML detection problem but may vary randomly from block to block. A large part of the work on receiver structures for MIMO systems is based on the assumption that all channel gains are realizations of a Gaussian random process. Despite being thought of as reasonable from a physical modeling point of view, we find that this assumption can lead to complicated algebraic expressions for the design of interference suppressing detectors and to a somewhat ill-defined problem in the context of optimal soft information transfer. As a remedy to this problem, we propose to model the propagation channel as a deterministic unknown when designing the STC detector. However, for the purpose of performance analysis, assumptions on the channel statistics have to be made, and consequently, a stochastic model has to be used. The paper is organized as follows. In Section II, we describe the system model that will be used throughout this work. In Section III, we show that the optimal soft information entities are equal to the values of certain concentrated likelihood functions.

1053587X/02$17.00 2002 IEEE

938

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 4, APRIL 2002

Explicit expressions for these likelihood values under various assumptions are derived in Section IV, where we also discuss the MIMO channel modeling. We present some numerical examples in Section V, and Section VI concludes the paper. II. SYSTEM MODEL Referring to Fig. 1, consider the transmission of a message , where each element belongs to a finite constellation . The message is encoded by a channel encoder by adding redundancy that is used for correcting possible errors in the transmission over a noisy channel [1], [10]. In our model, the . We channel encoder operates on a symbol level, i.e., on ; we use the time denote the encoded symbols by and the time index for index for the source symbols . Note that in general, the number of the encoded symbols encoded symbols is different from the number of source symbols and that if a convolutional encoder is used, there is an implicit one-to-one correspondence between and in the sense depends only on , where is the memory that length of the code. Following the channel encoder, each codeword is passed on to the interleaver that changes the order of the symbols to ensure that the fading that affects adjacent symbols is independent. The interleavingdeinterleaving is completely transparent to the system, and hence, we consider the symbols and received data before the interleaver and after the deinterleaver, respectively, and use the time index for that purpose. Consider next the transmission of one encoded symbol over the MIMO channel. We assume that the system employs transmit antennas and receive antennas and that the channel gain is frequency independent so that the transmission channel matrix , where the element can be described by a of is the complex gain between transmit antenna and receive antenna at time . The STC encoder maps each symbol onto a space-time code matrix and transmits of across the antennas the columns during consecutive time intervals. The received signal can then be written (1) where and are defined above, and is a matrix that contains thermal noise

and interference. We assume that the receiver noise is temporally white and , i.e., the complex Gaussian with spatial covariance matrix are distributed according to columns of for and for . The assumption is temporally white is customary but, that the receiver noise clearly, an approximation. In particular, the interference may consist of modulated signals that are not perfectly white. We will make one of the following two assumptions about the spatial noise covariance. a) The noise is spatially and temporally white, i.e., , where is an unknown noise variance. b) The noise is temporally white but spatially colored, i.e., is an unknown positive definite matrix. This is a standard model for a communication channel subject to interference and jamming.

We assume that the channel as well as the noise statisor are constant during one symbol interval but may tics change independently from one symbol interval to another. The channel usually changes in a random fashion in a fast-fading environment, especially since the interleaver spreads consecutive encoded symbols over time (and possibly frequency). The contains both thermal noise and co-channel innoise term terference, and the power of the interference can vary signifior for different cantly with . Hence, the noise statistics values of are also modeled as being independent. That the interference power varies with is especially true, for instance, in a time-division multiaccess (TDMA) system, where the interferers may be unsynchronized with the signal of interest, where they also use frequency hopping [11], [12], and where the interleaver spreads adjacent symbols over a relatively long time. In this paper, we treat exact ML detection and decoding, given and the noise statistics or are unthat the channel known to the receiver for all . In particular, we consider the situation where the fading may be so fast that the channel cannot be explicitly estimated in a reliable way using training. To perform ML detection in this case, the likelihood function for the problem under study has to be minimized jointly with respect to the unknown channel, noise covariance, and symbols. In particular, this means that some assumption has to be made on the . A very common assumption on for unknown matrix the derivation of receiver structures is that its elements are independent zero-mean complex Gaussian random variables with a (see, e.g., [7], [8], and common but unknown variance [13]). This assumption, which we will refer to as the stochastic channel model, may be reasonable from a physical modeling point of view (but, of course, it does not hold in a correlated fading environment [14]). As an alternative to assuming that the is a Gaussian random variable, we can simply let it channel enter the likelihood function as a deterministic unknown quantity. We will refer to this alternative as the deterministic channel model. In this work, we will treat both the stochastic and deterministic channel models and discuss their advantages and disadvantages for receiver design. It turns out that for hard decisions , the deterministic and and spatially white noise stochastic assumptions give similar results, whereas in the case of soft decisions, they can give different results. Furthermore, the stochastic channel model may not be feasible for detector design in the case where the noise is assumed to be spatially colored. We stress that even though the source that generates the channel coefficients (that is, the fading or wave propagation process) may, in practice, be considered to be random, this does not contradict the deterministic assumption made by the , as a function of , may be detector. The channel matrix random in nature, yet for a given , it is an unknown constant and can therefore be treated as a deterministic matrix in the likelihood function (corresponding to one symbol interval). Indeed, the assumption on the channel being stochastic or deterministic is not related in any way to the nature or rapidness of the fading process. The deterministic and stochastic assumptions simply represent two different ways for the receiver to model an unknown quantity. Although both the deterministic and stochastic models can be defended from various standpoints,

LARSSON et al.: MAXIMUM-LIKELIHOOD DETECTION AND DECODING

939

our results indicate that the deterministic viewpoint tends to be more favorable for receiver design. This observation can be contrasted with results from the array processing literature [15], where a stochastic signal model is favored over a deterministic one; note, however, that the problem studied there is different in nature. Note also that for the purpose of performance analysis, a statistical model for the channel may have to be adopted. Symbol detection is only possible if the likelihood function corresponding to the received data has a unique maximum. We will see below that since the channel is unknown, a necessary condition for this maximum to be unique is that the code mahas more columns than rows, that is, . This can trix be achieved by a suitable code design or induced artificially by considering two or more consecutive symbols at the same time and assuming that the channel is constant during the transmission of those symbols. An especially eloquent embodiment of the latter idea is the so-called differential space-time modulation (DSTM) scheme [8], [13], [16]. In a DSTM system, each is mapped onto a matrix , where encoded symbol is an algebraic group of unitary matrices. Furthermore, the space-time code matrices are chosen in the following special way: (2) here instead of for the space-time code maWe write trix since we will later consider two consecutive symbols sito denote those two symbols multaneously, and we will use concatenated with each other. Note that (2) motivates the term , it is clear that differential. Given an initial matrix for all . For a DSTM system, it is straightforward to design a detection scheme that does not require knowledge of the channel. To demonstrate this, note that the received signal corresponding to two consecutive symbols can be written as (3)

into account properly when designing, for instance, an ML sequence detector. However, if the differential encoding is performed after the interleaving, this effect can be neglected. III. ML DECODING Let denote a variable that contains all information about the MIMO channel during the transmission of the symbol , viz., the channel gains and the noise statistics at time . For instance, in the case of a deterministic channel and white noise, . An ML approach to the decoding of a transmitted message amounts to jointly maximizing the conditional log-likelihood with respect to the function of the received data and the variables , that is message argmax (5) and If we assume sufficient interleaving, in other words, that are independent for , the log-likelihood function in (5) can be written

(6) Define the following concentrated log-likelihood function: (7) Then, from (6) and (7), it follows that

(8) Equations (5) and (8) show that the ML sequence detection with reamounts to maximizing . If a convolutional spect to all possible messages depends only on , channel code is used, where is the memory length of the code. Hence, in the steady , (8) reduces to state, that is, when

is included in the unknown channel where the code matrix . We write here to stress the assumption that the channel during the transmission of the two symbols is considered to be constant. Using this technique, it is shown, for instance, in [13] that argmax Re Tr (4)

(9) Note that the implicit correspondence between the time index before the encoder and that after the encoder is crucial here. The minimization of (9) can be implemented efficiently using the Viterbi algorithm (VA) to facilitate ML sequence detection at a constant computational load per symbol. See, for example, [1], [2], [10], and the references therein for discussions about the VA and details on its implementation. From the above discussion, it is clear that the optimal soft information entities to be transferred from the STC detector to the channel decoder are the concentrated likelihood functions for each received symbol given dif. Hence, ferent hypothetical transmitted messages optimal (in a ML sense) detection requires the maximization of

gives the hard ML decision under the assumption of a stochastic dechannel model and spatially white noise. Hereafter, Tr denotes the conjugate transpose, and Re notes the trace, denotes the real part. Note that in a DSTM system, the sequence of received is used for the blocks detection of the sequence of symbols or, rather, space-time . Due to the overlapping of code matrices with , the noise in the sequence is not white; in particand ular, the two consecutive noise matrices are not independent. This fact must be taken

940

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 4, APRIL 2002

the likelihood functions with respect to the channel state infor. Since the received symbol depends mation entities only through the enon the transmitted message or, equivalently, , we write incoded symbol when considering the concentrated stead of has been eliminated). Herelikelihood function (from which after, we also drop the time index to simplify the notation whenever no confusion can occur. The following discussion focuses on optimal soft information transfer. On the other hand, a detector that employs hard decisions takes a definite decision on each received symbol but provides no reliability information about this decision. The decision on each symbol is taken by maximizing the likelihood function. The most likely symbol, but not the likelihood value itself, is sent to the channel decoder. In the above framework, this can be equal to unity (1) for the most implemented by setting likely symbol and to zero (0) otherwise. The values 1 and 0 are used only to denote a decision that a certain symbol was/was not sent and are not related to the symbol itself or its amplitude in any way (in fact, the values 0 and 1 can be replaced by any other two constants). For instance, if higher order modulation is employed so that each symbol is associated with several information bits, each symbol corresponds to a certain constellation point, and hence, hard decision amounts to deciding on a particwould take on 1 ular constellation point. In this case, for the most likely constellation point and 0 for all others. As is well known, and as we will see later, soft-information transfer yields a significant performance gain, which, even in the MIMO case, comes at a modest additional computational cost. IV. ML DETECTION We now proceed to derive the ML STC detector. For the deterministic channel model, we treat both the case of spatially white and colored noise. However, for the stochastic channel model, we consider for simplicity only the case of white noise. The corresponding detectors will be referred to as deterministic white ML (DWML); deterministic colored ML (DCML); stochastic white ML (SWML). For simplicity, we consider only semi-unitary space-time code matrices . Yet, we treat both detection for one received and symbol at a time, in which case, we require detection of differential codes [see (2) and (3)]. In the former is of dimension , case, the received signal is by assumption. In the latter case, the received and according signal is . To accommodate both cases, to (3), and hence, be of dimension , where we will let corresponds to either or . Consequently, we and that . Note that assume that never has full rank (as long as ). A. Deterministic Channel and White Noise Assume that the receiver noise is white so that and . The model for the received signal is given by (1), and hence, the logarithm of the likelihood function for under the assumption of a deterministic the observed data

channel as

can be written, after dropping irrelevant constants,

Tr Maximization of (10) with respect to that

(10) amounts to requiring

Tr

(11)

which has the unique and strictly positive (with probability 1, for all valid ) solution Tr Insertion of (12) into (10) yields Tr (13) (12)

Maximizing (13) with respect to is a quadratic optimization problem, and it is easy to see that the solution is given, provided (see, e.g., [17] and [18]), by that (14) Insertion of (14) in (10) and dropping constant terms gives the sought-after DWML concentrated likelihood function

Tr Tr (15)

B. Deterministic Channel and Colored Noise We next consider the case where the receiver noise is spatially colored. Here, is a general positive definite matrix, and . The log-likelihood function for the observed data under the assumption of a deterministic channel can be written, after dropping irrelevant constants, as

Tr

(16)

denotes the determinant. The maximization of the where log-likelihood function in (16) with respect to an unstructured Hermitian matrix amounts to requiring that

Tr (17)

LARSSON et al.: MAXIMUM-LIKELIHOOD DETECTION AND DECODING

941

where Re Im and Im denotes the imaginary part. By using some standard matrix algebra results [17], [18], (17) can be shown to be equivalent to Tr Tr (18) is the th column of the identity matrix. It is easy to where see that (18) is equivalent to (19) which obviously has the unique solution (20) is positive definite with probaNote that as long as bility 1 for all possible . Insertion of (20) in (16) and dropping constant terms make (16) proportional to (21) denote the fact that the Now, let the matrix inequality is non-negative definite, and observe that matrix

corresponding to the variance of the channel gains, and is the unknown noise power. Note that and vary with , but we drop their dependence on for notational convenience. We remark here that the model used in [7], [8], and [13] implicitly , which is an assumption that we will not assumes that make. Even if the model can be normalized by the noise power and the variance of the channel gains would be so that , both and are unknown and should enter the data model and the likelihood function as unknown parameters. Note . that for the stochastic channel model, Under the assumptions above, the rows of , or, equivalently, , are independent and Gaussian with covarithe columns of ance matrix (27) Consequently, the likelihood function for the received data can be written as Tr Note first that (29) whenever where we have used the fact that the dimensions of and are conformable [17], [18] and that . Furthermore, it follows by the matrix inversion lemma [17], [18] that (30) Using these observations, the likelihood function in (28) can be written as (28)

(22) where the inequality becomes an equality if and only if (23) Hence, minimizes all eigenvalues of and, therefore, the determinant in (21) as well. Insertion of (23) in (21) gives the sought DCML concentrated likelihood function

Tr (24) Note the similar structure of (15) and (24). In addition, note (strictly) for both (15) and (24) to be that we need , the term inside the trace in meaningful; indeed, if (15) or the determinant in (24) would be identical to zero. This shows that if the channel is unknown, symbol detection is only possible if the code matrix has more columns than rows (cf. Section II). C. Stochastic Channel and White Noise In the case of a stochastic channel, we model the received signal as follows [7], [8], [13]: (25) or equivalently (26) elements of as well as the elements where the are independent and complex Gaussian distributed with of is an unknown parameter unit variance. The variable

Tr

(31)

The that maximizes (28), or, equivalently, (31), coincides with for any and . the that maximizes Tr Hence, the hard ML symbol decisions under the assumption of a stochastic channel are given by argmax Tr (32)

which in the case of the differential code in (2) readily simplifies to (4). A comparison with (15) reveals that the hard ML decisions obtained under the assumption of a deterministic channel are equal to those obtained under the assumption of a stochastic channel using (32). In particular, for the DSTM scheme discussed in Section II, this means that the deterministic and stochastic channel models give the same result [viz. (4)], as far as hard decisions are concerned. However, the case of optimal soft decision turns out to be slightly more involved. The computation of optimal soft information under the assumption of a stochastic channel model requires the concentraand . This tion of the likelihood function with respect to is complicated by the fact that the set is open. Since is open, the global maximum must occur at an interior point of , where the gradient of the likelihood function

942

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 4, APRIL 2002

with respect to fied that the only a given ) Tr Tr

is equal to zero. It can readily be verifor which the gradient is zero is (for

(see, e.g., [13]). For this system,

, and

Tr

(33)

(36) and hence, the likelihood function (34) assumes the same value and . for An alternative possible approach to the problem associated with the concentration of the stochastic likelihood func, i.e., perform the maximization subtion is to allow and belong to the compact ject to the constraint that . In this case, the that set maximize (31) must be either a point where the gradient of is zero or a point on the border of . As and with probability 1. Fur, the likelihood function thermore, on the line takes on the form Tr (37)

Furthermore, it is not difficult to show that the Hessian main (33) trix at this point is positive definite; therefore, . Clearly, in (33) is posiis a maximum of tive with probability 1, but there is a nonzero probability that , in which case, there is no point in where the gradient is zero. Hence, there is a nonzero probability of that the maximum of the likelihood function in does not exist, or, in other words, that it is impossible to concentrate the likelihood function. This problem can be overcome by relaxing as well. While the optimization problem to allow negative this is an approximation in the sense that the original estimation problem is modified, it is in some analogy to what is customarily done for the so-called stochastic ML method in sensor array processing [15]. Doing so, we insert (33) into (31) and obtain, after a few lines of algebra, the concentrated stochastic white ML (SWML) likelihood function

Tr Tr (34) The hard and soft decisions based on (34) are, in general, different from those obtained by DWML. That the hard decisions are different can be understood by considering the fact that the was implicitly used in obtaining (32) and imconstraint plicitly neglected when obtaining (34). to be negative might be thought of as a conveAllowing nient way to guarantee the existence of a solution to the optimization problem and, hence, as a way to concentrate the likelihood function. However, this fix does not circumvent the difficulty associated with the stochastic channel model. In fact, it is not difficult to find an example where the method based on (34) fails. Consider, for instance, a DSTM communication system transmit antennas and receive antennas with employing the 2 2 BPSK unitary group code (35)

is easily found to and its maximum value with respect to Tr . A point be where the gradient is zero exists only if in (33) is positive. It follows that the likelihood function concentrated over is as in (38), shown at the bottom of the page. The detector based on (38) will be referred to as the modified-SWML (M-SWML). For the counterexample (35), it can be verified that , and hence, exactly one of the in (33) is negative. For this reason, we expect that the M-SWML method (38) will not fail for this example, unlike SWML in (34). See Section V for examples of the performance of M-SWML. V. NUMERICAL EXAMPLES We illustrate our findings by considering a DSTM commutransmit antennas and nication system with receive antennas. For simplicity, we use the differential 2 2 BPSK unitary group code in (35), and to simplify the simulation work involved, we use a (7, 3) binary Hamming block code [1] and a brute-force decoding algorithm that simply scans through all possible codewords and chooses the one with the largest likelihood value. Example 1Thermal Noise Only: We consider a situation where the channel noise is spatially and temporally white are independent outcomes Gaussian, that is, the columns of distribution. Fig. 2 shows an estimate of the of a bit-error-rate (BER) after decoding, obtained by Monte Carlo simulation, as a function of the signal-to-noise ratio (SNR). are independent All elements of the channel gain matrix

Tr otherwise

(38)

LARSSON et al.: MAXIMUM-LIKELIHOOD DETECTION AND DECODING

943

Fig. 2. Results for Example 1. Thermal noise only.

Fig. 3. Results for Example 2. One strong interferer plus thermal noise.

realizations of a Gaussian random process with variance , . In this case, the DWML and the SNR is defined as detector outperforms the DCML detector by 3 dB, which was expected since DWML uses the a priori information that the noise is spatially white. Furthermore, the usage of optimal soft information transfer to the channel decoder yields a gain of 1.5 dB. Example 2Strong Interference: In this example, we consider a system with one strong interference signal. The interference signal is temporally white and has a power equal to that of the signal of interest and a channel gain that is also zero-mean Gaussian and independent of all other random quantities. White Gaussian noise is added to the received signal in the same way as for Example 1. Fig. 3 shows the BER versus the SNR for the different schemes. The DCML receiver outperforms the DWML receiver significantly. Furthermore, we observe that the soft information transfer also yields a substantial gain in this case. In our example, the ability of the DCML receiver to cancel interference comes at a relatively high cost, that is, more than a 3-dB performance loss when there is no interference. This is is too small; hence, the matrix related to the fact that in (20) is a poor estimate of the noise covariance matrix. This or by modisituation can be alleviated by choosing fying the STC detector to assume that is constant over several time intervals. As a further remark on this aspect, we note that [9] uses another approach to perform interference cancellation in a STC system at the cost of a reduced data rate. Despite the fact that the approach in [9] yields a simple receiver structure, it requires the noise statistics as well as the channel to be constant over several symbol intervals, which is a condition that may not be met in practice. For this reason, a performance comparison between our technique and that in [9] is not directly relevant. Example 3Stochastic versus Deterministic Channel Modeling: In this example, we compare the quality of the soft information provided by the DWML concentrated likelihood function (15) to that provided by M-SWML [see (38)]. We consider a scenario identical to that in Example 1. Fig. 4 shows an estimate of the BER after decoding using soft decisions based on (15) and (38) and hard decisions based on (15) [or, equivalently,

Fig. 4. Results for Example 3. Deterministic and stochastic channel modeling.

(32)], respectively. Note that we show no curve corresponding to SWML [see (34)] since we demonstrated in Section IV-C that this method will fail, at least for the code matrix (35) that we use in this example. We can see from Fig. 4 once again that the use of soft information improves the performance over hard decisions. We also note that as far as soft decisions are concerned, the performance of DWML is slightly better than that of M-SWML. VI. CONCLUDING REMARKS This paper has presented a unified treatment of detection and decoding for space-time coding systems under unknown MIMO channel conditions. In particular, interference resistant detection algorithms were derived, and it was shown how optimal (in a ML sense) soft information can be transferred from the space-time code detector to an error-correcting decoder. Furthermore, certain aspects related to deterministic and stochastic channel modeling were discussed. In particular we showed that from a receiver design point of view, deterministic channel modeling can sometimes be more convenient than stochastic channel modeling.

944

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 4, APRIL 2002

REFERENCES
[1] J. G. Proakis, Digital Communications, 3rd ed. New York: McGrawHill, 1995. [2] H. Meyr, M. Moeneclaey, and S. A. Fechtel, Digital Communication Receivers: Synchronization, Channel Estimation and Signal Processing. New York: Wiley, 1998. [3] U. Hansson and T. M. Aulin, Soft information transfer for sequence detection with concatenated receivers, IEEE Trans. Commun., vol. 44, pp. 10861095, Sept. 1996. [4] V. Tarokh, N. Seshadri, and A. R. Calderbank, Space-time codes for high data rate wireless communications: Performance criterion and code construction, IEEE Trans. Inform. Theory, vol. 44, pp. 744765, Mar. 1998. [5] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, Space-time block codes from orthogonal designs, IEEE Trans. Inform. Theory, vol. 45, pp. 14561467, July 1999. [6] G. Ganesan and P. Stoica, Space-time block codes: A maximum SNR approach, IEEE Trans. Inform. Theory, vol. 47, pp. 16501656, 2001. [7] B. M. Hochwald and T. L. Marzetta, Unitary space-time modulation for multiple-antenna communications in Rayleigh flat fading, IEEE Trans. Inform. Theory, vol. 46, pp. 543564, Mar. 2000. [8] B. M. Hochwald and W. Sweldens, Differential unitary space-time modulation, IEEE Trans. Commun., vol. 48, pp. 20412052, Dec. 2000. [9] J. Liu, J. Li, H. Li, and E. G. Larsson, Differential space-code modulation for interference suppression, IEEE Trans. Signal Processing, vol. 49, pp. 17861795, Aug. 2001. [10] S. Lin and D. J. Costello Jr., Error Control Coding: Fundamentals and Applications. Englewood Cliffs, NJ: Prentice-Hall, 1983. [11] G. L. Stber, Principles of Mobile Communication. Norwell, MA: Kluwer, 2001. [12] M. Mouly and M.-B. Pautet, The GSM System for Mobile Communications. Paris, France: Telecom, 1992. [13] B. L. Hughes, Differential space-time modulation, IEEE Trans. Inform. Theory, vol. 46, pp. 25672578, Nov. 2000. [14] H. Blcskei and A. J. Paulraj, Performance of space-time codes in the presence of spatial fading correlation, in Proc. 34th Asilomar Conf. Signals, Syst., Comput., vol. 1, Pacific Grove, CA, Nov. 2000, pp. 687693. [15] B. Ottersten, M. Viberg, P. Stoica, and A. Nehorai et al., Exact and large sample ML techniques for parameter estimation and detection in array processing, in Radar Array Processing, S. Haykin et al., Eds. New York: Springer-Verlag, 1993, ch. 4. [16] V. Tarokh and H. Jafarkhani, A differential detection scheme for transmit diversity, IEEE J. Select. Areas Commun., vol. 18, pp. 11691174, July 2000. [17] P. Stoica and R. Moses, Introduction to Spectral Analysis. Upper Saddle River, NJ: Prentice-Hall, 1997. [18] T. Sderstrm and P. Stoica, System Identification, Hemel Hempstead, U.K.: Prentice-Hall Int., 1989.

Petre Stoica (F94) received the D.Sc. degree in automatic control from the Bucharest Polytechnic Institute (BPI), Bucharest, Romania, in 1979 and an honorary doctorate degree in science from Uppsala University (UU), Uppsala, Sweden, in 1993. He is a Professor of system modeling with the Department of Systems and Control at UU. Previously, he was a Professor of System Identification and Signal Processing at BPI. He held longer visiting positions with the Eindhoven University of Technology, Eindhoven, The Netherlands; Chalmers University of Technology, Gothenburg, Sweden (where he held a Jubilee Visiting Professorship); UU; the University of Florida, Gainesville; and Stanford University, Stanford, CA. His main scientific interests are in the areas of system identification, time-series analysis and prediction, statistical signal and array processing, spectral analysis, wireless communications, and radar signal processing. He has published seven books, ten book chapters, and some 450 papers in archival journals and conference records on these topics. The most recent book he co-authored with R. Moses is entitled Introduction to Spectral Analysis (Englewood Cliffs, NJ: Prentice-Hall, 1997). Recently, he co-edited two books on signal processing advances in wireless and mobile communications (Englewood Cliffs, NJ: Prentice-Hall, 2000). He is on the Editorial Boards of five journals in the field: Journal of Forecasting; Signal Processing; Circuits, Systems, and Signal Processing; Digital Signal Processing-A Review Journal; and Multidimensional Systems and Signal Processing. He was a co-Guest Editor for several special issues on system identification signal processing, spectral analysis, and radar for some of the aforementioned journals as well as for Proceedings of the IEE. Dr. Stoica was co-recipient of the IEEE ASSP Senior Award for a paper on statistical aspects of array signal processing and recipient of the Technical Achievement Award of the IEEE Signal Processing Society for fundamental contributions to statistical signal processing with applications in time-series analysis, system identification, and array signal processing. In 1998, he received a Senior Individual Grant Award of the Swedish Foundation for Strategic Research. He was also co-recipient of the 1998 EURASIP Best Paper Award for Signal Processing for a work on parameter estimation of exponential signals with time-varying amplitude, a 1999 IEEE Signal Processing Society Best Paper Award for a paper on parameter and rank estimation of reduced-rank regression, a 2000 IEEE Third Millennium Medal, and the 2000 W.R.G. Baker Paper Prize Award for a work on maximum likelihood methods for radar. He was a member of the international program committees of many topical conferences. From 1981 to 1986, he was a Director of the International Time-Series Analysis and Forecasting Society, and he has been a member of the IFAC Committee on Modeling, Identification, and Signal Processing since 1994. He is also a member of the Romanian Academy and a Fellow of the Royal Statistical Society.

Erik G. Larsson (S99M02) received the M.Sc. degree in applied physics and electrical engineering from Linkping University, Linkping, Sweden, in 1997 and the Ph.D. degree in electrical engineering from Uppsala University, Uppsala (UU), Sweden, in 2002. From 1998 to 1999, he was a research engineer at Ericsson Radio Systems AB, Stockholm, Sweden, where he was involved in algorithm design and standardization of location services for the GSM system. From 1999 to 2000, he was a research and teaching assistant with the Department of Systems and Control, UU. Since September 2000, he has been with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, where he is currently an assistant professor. His research interests include wireless communications, statistical signal processing, spectral analysis, and radar applications. He has several publications in international journals and holds three U.S. patents.

Jian Li (S87M91SM97) received the M.Sc. and Ph.D. degrees in electrical engineering from The Ohio State University (OSU), Columbus, in 1987 and 1991, respectively. From April 1991 to June 1991, she was an Adjunct Assistant Professor with the Department of Electrical Engineering, OSU. From July 1991 to June 1993, she was an Assistant Professor with the Department of Electrical Engineering, University of Kentucky, Lexington. Since August 1993, she has been with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, where she is currently a Professor. Her current research interests include spectral estimation, array signal processing, and signal processing for wireless communications and radar. She is a Guest Editor for Multidimensional Systems and Signal Processing. Dr. Li is a member of Sigma Xi and Phi Kappa Phi. She received the 1994 National Science Foundation Young Investigator Award and the 1996 Office of Naval Research Young Investigator Award. She is currently an Associate Editor for IEEE TRANSACTIONS ON SIGNAL PROCESSING. She is presently a member of the Signal Processing Theory and Methods (SPTM) Technical Committee of the IEEE Signal Processing Society and an Executive Committee Member of the 2002 International Conference on Acoustics, Speech, and Signal Processing, which is to be held in Orlando, FL, in May 2002.

You might also like