296

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

Timing and Frequency Synchronization for the Uplink of an OFDMA System

Michele Morelli

AbstractThis paper deals with timing and frequency recovery for the uplink of an orthogonal frequency-division multiple access (OFDMA) system. The frequency estimator is derived from ad hoc reasoning, whereas the timing estimator is based on the maximum-likelihood criterion. Both schemes rely on the repetition of a fixed pilot symbol. Their main feature is that they provide feedforward estimates and allow synchronization in only two OFDM blocks. In contrast to other existing methods, they do not require that the subcarriers of a given user occupy adjacent positions in the signal bandwidth. This makes it possible to interleave subcarriers of different users so as to optimally exploit the frequency diversity of the channel. Theoretical analysis and computer simulations are used to assess the performance of the proposed synchronizers. It is found that the degradations due to residual frequency and timing errors are negligible at signal-to-noise ratios of practical interest. Index TermsFrequency estimation, maximum-likelihood (ML), orthogonal frequency-division multiple access (OFDMA), timing estimation.

I. INTRODUCTION

RTHOGONAL frequency-division multiple access (OFDMA) is a multiplexing technique in which several users simultaneously transmit their own data by modulating an exclusive set of orthogonal subcarriers. OFDMA has gained increased interest in the last few years and has been proposed for the uplink of wireless communication systems [1], [2] and cable TV (CATV) networks [3]. Its main advantage is that separating different users through frequency-division multiple access (FDMA) techniques at the subcarrier level can mitigate multiple-access interference (MAI) within a cell. Also, compared with single-carrier multiple-access systems, OFDMA offers increased robustness to narrowband interference, allows straightforward dynamic channel assignment, and does not need adaptive time-domain equalizers, since channel equalization is performed in the frequency domain through one-tap multipliers [3]. For all this to be true, however, proper frequency and timing synchronization is necessary to maintain orthogonality among the active users. Frequency offsets due to Doppler shifts and/or oscillator instabilities produce interchannel interference

Paper approved by L. Vandendorpe, the Editor for Transmission Systems of the IEEE Communications Society. Manuscript received November 7, 2001; revised July 30, 2002 and June 10, 2003. This work was supported by the Italian Ministry of Education under the FIRB project PRIMO. This paper was presented in part at the MCSS, Munich, Germany, 2001. The author is with the University of Pisa, Department of Information Engineering, 56100 Pisa, Italy (e-mail: michele.morelli@iet.unipi.it). Digital Object Identifier 10.1109/TCOMM.2003.822699

(ICI) and must be counteracted to avoid severe error-rate degradations. Timing errors result in intersymbol interference (ISI) between consecutive OFDM symbols. Using a guard interval (cyclic prefix) provides intrinsic protection against timing errors at the expense of some reduction in the data throughput due to the extra overhead. However, timing accuracy becomes a stringent requirement in practical applications where, to minimize the overhead, the cyclic prefix is made only just greater than the length of the channel impulse response (CIR). Frequency and timing recovery for single-user OFDM has received much attention in the last few years, and several solutions are available in the technical literature (see [4][8] and references therein). Unfortunately, they are only suited for a broadcast (downlink) scenario and cannot be directly used in the uplink of a multiuser system, because each user must be separated from the others at the base station (BS) before his synchronization parameters can be estimated. A possible separation method is to assign a group of adjacent subcarriers to each user and then pick them up through a filter bank [2] at the BS. However, grouping the subcarriers together prevents the possibility of optimally exploiting the channel diversity. A deep fade might hit a substantial number of subcarriers of a given user if they are close together [1]. In this paper, we propose an alternative method for estimating the frequency and timing offsets of a new user entering the system. As in [2], we assume that the other users have already been acquired and aligned to the BS references. The estimates of the users offsets are then returned on a downlink control channel and exploited to adjust the transmitter clock and carrier frequency [2]. Note that frequency and timing compensation cannot be accomplished at the BS, as the correction of one users offsets would misalign the other users. As we shall see, joint maximum-likelihood (ML) frequency and timing recovery is impractical, as it involves a bidimensional (2-D) numerical search. Therefore, we look for simpler solutions. In particular, we propose a suboptimum algorithm that involves the repetition of a given pilot symbol and gives an explicit frequency estimate. The algorithm is based on ad hoc reasoning, and its accuracy is close to the CramerRao bound. The frequency estimate is then exploited to perform timing recovery by resorting to an ML-based method. In comparison with [2], the proposed scheme has two major advantages. First, it can be applied when the subcarriers of different users are interleaved and, second, it provides estimates in just two OFDM symbols, whereas in [2], the estimates from several symbols must be filtered to achieve sufficient accuracy. However, using pilot symbols results in an extra overhead with respect to [2]. Also, as

0090-6778/04$20.00 2004 IEEE

MORELLI: TIMING AND FREQUENCY SYNCHRONIZATION FOR THE UPLINK OF AN OFDMA SYSTEM

297

Fig. 1.

Block diagram of the k th OFDMA transmitter.

Fig. 2.

Block diagram of the k th OFDMA receiver.

shown later, interleaving the subcarriers of different users produces MAI during the acquisition phase of the new user. The rest of the paper is organized as follows. The next section introduces the OFDMA system and formulates the problem. The signal model is described in Section III, whereas the joint ML frequency and timing recovery is addressed in Section IV. Sections V and VI deal with reduced-complexity alternative schemes. Simulation results are discussed in Section VII and, finally, some conclusions are offered in Section VIII. II. OFDMA UPLINK SYSTEM A. System Description We consider the uplink of an OFDMA system employing subcarriers and accommodating a maximum of simultaneously active users. Each user transmits on a set of assigned subcarriers. The block diagram of the transmitter for the th user is shown in Fig. 1. The stream is serial-to-parallel (S/P) converted and parof data titioned into adjacent blocks of length . The th block is denoted , with the superscript meaning transpose operation. Vector is extended with the insertion of zeros to produce and is fed to an -point inverse discrete Fourier transform (IDFT) unit. The components of are defined as if otherwise where (1)

given user are grouped together. Also, some components of may be set to zero to account for virtual carriers [9]. The time-domain samples from the IDFT unit are (2)

and are arranged in a vector . In order to eliminate any interference between adjacent OFDM symbols, an -point prefix is appended to such that for . The resulting vector drives a linear modulator with imand signaling interval , pulse response where is the OFDM symbol duration. The transmitted signal reads

(3) where counts the OFDM blocks and counts the time-domain samples within a block. As discussed later in Section II-B, and serve to compensate for the the parameters in Doppler shift and transmission delay incurred by passing through the channel. The receiver at the BS is sketched in Fig. 2. The incoming waveform is first filtered and then sampled at rate . After S/P conversion, the cyclic prefix is removed and the remaining samples are fed to an -point discrete Fourier transform (DFT) unit. The DFT output is finally passed over to the equalizer and the data detector. The waveform arriving at the BS is the superposition of the signals from the active users. The th signal experiences a Doppler shift due to the mobile speed, and a propagation , depending on the distance between the user and the delay

is the index of the subcarrier modulated by . The function associating the th component of to is referred to as user-specific frequency mapping. Note can be any integer within the interval that the indexes ), since we do not assume that the subcarriers of a (0,

298

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

BS. Let us consider the quantities and . For convenience, we decompose in an integer part and a fractional part with respect to the sampling period, i.e., (4) with and from the matched filter at . Then, the sample may be written as (5)

ingly, we have and are given by

and

. Note that

(7) (8) is the mobile speed, and where is the speed of light, is the distance between the user and the BS. in (5) for yields Setting (9) where the contribution of user #1 has been separated from the and . Note others. The task of the BS is to estimate need not be estimated, as long as that the fractional delay is shorter than the difference between the length of the . The reason for this is that cyclic prefix and the length of the fractional delay appears as a linear phase across the DFT outputs and can be compensated for by the channel equalizer, which does not distinguish between phase shifts introduced by the channel and those deriving from the timing offset. III. SIGNAL MODEL Denote the th block of samples from the receiver filter (after elimination of the cyclic prefix) and assume that, for any user, has a finite support , with . Then, letting and collecting (2), (6), and (9) yields

where is thermal noise and nent of the th user

is the signal compo-

(6) is the length of the extended In this equation, is the OFDM symbols (including the cyclic prefix), is the sample of overall CIR of the th user, and at . Note that the fractional part of the propagation delay has been incorporated into . B. Synchronization Policy As mentioned earlier, the correction of one users frequency and timing cannot be accomplished at the BS, since this operation would misalign the other users. Therefore, in our scheme, the BS performs only frequency and timing estimation, whereas adjustment of the synchronization parameters is made at the users side based on instructions transmitted on the control channel. This is accomplished as follows. Assume that the ticks of the time reference from the BS are transmitted at the on the carrier frequency (close instants to the uplink carrier frequency). Because of the propagation delay and the Doppler shift, the ticks are received by user at on the frequency . To keep his signal aligned at the BS, user must transmit at on the frequency . This amounts to saying that the parameters and appearing in (3) are chosen equal to and , respectively. We consider the situation where all the active users, except the first, are already synchronized and, accordingly, we take for . To get information about his own misalignment, user #1 transmits with the timing phase and frequency received through the downlink control channel. This means that he sets and . Accord-

(10) where is a diagonal matrix (11) while and are matrices with entries

(12) and (13), shown at the bottom of the page. Also, is a ( zero-mean Gaussian vector with covariance matrix is the identity matrix of order ).

(13)

MORELLI: TIMING AND FREQUENCY SYNCHRONIZATION FOR THE UPLINK OF AN OFDMA SYSTEM

299

To separate the user of interest from the others, we compute and select those outputs that correspond to the the DFT of subcarriers of user #1. This produces the -dimensional vector (14) where is an matrix with entries

, we initially keep and To find the maximum of fixed and let vary. In these circumstances, the maximum is achieved for (19) where (20)

(15) Then, substituting (10) into (14) and exploiting the identity , which holds for ( is matrix with all-zero entries), yields an (16) is Gaussian with zero mean and covariwhere ance matrix . This equation is a crucial step in our derivation. It says that exhibits no interuser interference. Clearly, this relies on the assumption that the already-present for users are perfectly synchronized, i.e., . The effect of users imperfect synchronization is addressed later by simulation. To perform frequency and timing estimation, the BS observes vectors in succession ( is a design parameter). The symbols transmitted by the user of interest are assumed known at the receiver (data-aided operation). Also, for for , i.e., user #1 simplicity, we let transmits identical and known OFDM blocks (pilot symbecomes independent bols). In these circumstances, of . Then, dropping the superscript designating user #1 for simplicity, (16) reduces to (17) with to estimate section. . Our task is to exploit vectors and . An ML approach is considered in the next

means Hermitian transpose. Next, and the superscript substituting (19) into (18) and maximizing with respect to and , produces the ML estimates of and (21) with

(22) As is seen, computing and from (21) requires a 2-D grid search over the set spanned by ( , ). Such a search may be impractical even for the BS. In the next two sections, we propose a suboptimal solution in which the estimate of is given in closed form, while is estimated through a 1-D search. V. REDUCED-COMPLEXITY FREQUENCY ESTIMATION Consider the correlation (23) where is a design parameter not greater than . To see can be used to estimate , let us write (17) in the how equivalent form (24) with produces . Then, collecting (23) and (24)

(25) where output, while is the total signal energy at the DFT is a disturbance term

IV. ML JOINT FREQUENCY AND TIMING ESTIMATION Given the unknown parameters ( , , ), from (17), we see that the vectors are Gaussian and independent, with and covariance matrix . Acmean cordingly, the likelihood function of ( , , ) takes the form

(26) (18) Next we define the angles (27) where , , and are trial values of , , and , while denotes the Euclidean norm of the enclosed vector. is arbitrarily taken equal to unity. At large signal-towhere noise ratio (SNR) values, the last term in (26) (Noise Noise

300

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

TABLE I COMPUTATIONAL LOAD

contribution) can be neglected. Then, denoting the imag, it can be shown that the following approxiinary part of mation holds [10], [11]:

(28) where is the value of reduced to the interval [ , ). If is adequately small, the quantity in brackets in the right-hand side (RHS) of (28) is (with high probability) less than and reduces to (29) is the sum of a deterministic This equation indicates that component proportional to plus a zero-mean random disturbance. Based on the model (29), the best linear unbiased estimate of is found to be [11] (30) where is the following smoothing function: (31) The variance of given by [11] (conditioned on a given channel vector ) is

are inde2) Equation (31) indicates that the weights pendent of the training sequence. This might lead to the erroneous conclusion that RCFE operates in a nondataaided (NDA) fashion. Actually, as shown by (23) and (27), the frequency estimate is obtained by comparing the . phase of the received vectors Therefore, in order to work properly, the RCFE needs that only the frequency offset (and not the modulation phases) contributes to the phase shift between successive training symbols. This requires the repetition of a fixed OFDM for block during the acquisition process (i.e., ). In this respect, the RCFE might be classified as a data-aided (DA) estimator. , the RCFE reduces to the frequency esti3) With mator proposed by Moose [7]. correlations involve approx4) Computing the complex products and imately complex additions. The overall computational load required by the RCFE is summarized in the first row of Table I. In writing these figures, we have borne in mind that a complex product amounts to four real products plus two real additions, while a complex addition is equivalent to two real additions. It should be stressed that no grid search is needed with RCFE as (30) provides an explicit expression for . 5) It is interesting to compare the performance of RCFE with the modified CramerRao bound (MCRB) [12, p. 60] (34)

(32) and depends on , , and through the coefficient . Note that, achieves a minimum for that as varies, reads (33) (35) In the sequel, (30) is referred to as the reduced-complexity frequency estimator (RCFE). The following remarks are in order. 1) The RCFE gives unbiased estimates as long as the approximation (29) holds true. This occurs at high SNR values , otherwise the moduloprovided that operation in (28) makes the relation between and highly nonlinear, and the corresponding estimates are inaccurate. In summary, the estimation range of the RCFE . is limited to where (36) is the SNR loss at the DFT output due to frequency and timing offsets (note that becomes unity for ). From (35) we see that decreases and approaches as the number of observed vectors grows large. Also, it is worth noting that the quantity This bound has been computed bearing in mind that the and the signal energy of user observation length is , as is seen from #1 per received sample is (10). The loss of RCFE with respect to the MCRB is given by the ratio of (33) to (34) and reads

MORELLI: TIMING AND FREQUENCY SYNCHRONIZATION FOR THE UPLINK OF AN OFDMA SYSTEM

301

in (35) represents the performance loss due to discarding the signal energy contained in the cyclic prefix.

is the th entry of and where is the ( , )th entry of . The overall operations required by RCTE are summarized in the second row of Table I. B. RCTE With Uniformly Spaced Subcarriers A possible drawback of the RCTE is the large memory needed matrices in (40). We show that this to store the difficulty may be overcome by resorting to a properly designed frequency mapping rule. Assume that the subcarriers assigned to user #1 are uniformly spaced over the available transmission bandwidth at a distance from each other, i.e., of (42) In addition, let the length of the cyclic prefix be equal to the (note that is a fixed number of subcarriers, i.e., design parameter, independent of the effective number of active users). Then, it can be shown that (43) where is an matrix with entries (44) and , being the symbol transmitted on the th subcarrier. To proceed further, assume that the have unit amplitude (as occurs with phase-shift keying (PSK) modulation). Then, substituting (43) into (39) and (40), and bearing and , we get in mind that (45) where denoting by -point IDFT of and , we see that is given in (20). Finally, the can be written as (46) where is the value of reduced to the interval [0, ]. The following remarks are of interest. 1) From (46), it is clear that is periodic in of period . Therefore, the corresponding timing estimates are ambiguous by multiples of . To avoid this ambiguity, the cell radius and the number of subcarriers assigned to a . given user must be chosen in such a way that As shown later, this condition is easily met in practical situations. 2) In comparison with the general case of nonuniformly spaced subcarriers, the memory requirement is greatly reduced, as no matrix needs to be precomputed and stored to evaluate the RHS of (46). while

VI. REDUCED-COMPLEXITY TIMING ESTIMATION A. Derivation of the Timing Estimator If the frequency offset were perfectly known, the ML timing estimator could be derived using the same arguments of Section IV to produce (37) where is computed setting in (22), while is the maximum expected value of . In practice, however, is unknown, and we replace it by its estimate . This means that instead timing estimation is performed by maximizing . However, from (22) we see that computing of involves a matrix inversion. This can be avoided by observing that (in a well-designed system) Doppler shifts are much smaller than the distance between adjacent subcarriers. In other words, , and the matrix in (11) can be replaced we have . Correspondingly, reduces to by (independent of ) and the timing estimator reads (38) with (39) and (40) The following remarks are of interest. matrices are independent of 1) The . Therefore, they can be precomputed and stored so that no matrix inversion is needed in the computation . This greatly reduces the complexity of the of proposed scheme. For this reason, in the sequel we call it reduced-complexity timing estimator (RCTE). 2) The computational complexity of RCTE can be assessed have been as follows. Assume that the entries of precomputed. Then, computing from the vectors requires complex multiplications and complex additions. Also, evaluating needs a total real products and real additions. of In writing these figures, we have taken into account that is Hermitian and, accordingly, can also be written as

(41)

302

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

3) Computing requires the IDFT of . The latter can be efficiently performed through fast Fourier transform (FFT) techniques. 4) The processing load involved in (46) is assessed as follows. Assuming binary phase-shift keying (BPSK) or quaternary phase-shift keying (QPSK) data symbols, the requires complex products computation of and complex additions. Also, the FFT needs complex products and a total of complex additions. Finally, each modulus extraction requires two products and one addition. This leads to the third line of Table I, where the acronym USS stands for uniformly spaced subcarriers. C. Performance Analysis The performance of RCTE is assessed in terms of probability of making a timing error. An error event is declared to occur whenever the estimate gives rise to ISI at the DFT output. is greater than zero or less This amounts to saying that [2]. Note that no single correct timing than estimate exists. Depending on the CIR duration and the cyclic prefix length, there may be several values corresponding to zero-ISI operation. Computing the exact expression of the probability of , say , is a difficult task. For this reason, we will only give upper . To begin, we denote by the set of and lower bounds to and such that integers belonging to or . Then we observe that a timing error occurs whenever for some . Formally (47)

Taking the probabilities of both sides of (47) and using the union bound yields (50) On the other hand, from (47), it is seen that This produces the lower bound if .

(51)

VII. SIMULATION RESULTS The performance of the proposed frequency and timing synchronizers has been investigated by computer simulation in a frequency-selective fading channel. The following assumptions have been made. A. System Parameters . 1) The total number of subcarriers is subcarriers uniformly spaced at a 2) Each user has distance from each other. s. 3) The sampling interval is 4) The number of observed OFDM blocks is either or . The parameter with RCFE is set to . 5) The channel responses have length . Their components vary independently and are modeled as complex-valued Gaussian random variables with zero mean and an exponential power delay profile

where is the error event in a binary context in which and are the only hypothesized values of the timing offset (pairwise error event). The probais computed in the Appendix under the assumpbility of tion that is a zero-mean Gaussian vector, the SNR is large, and . The result is (48)

where (49)

and are the nonzero eigenvalues of the madefined in (A9). It is worth noting that detrix pends on the training sequence, the channel statistics, and the . Also, from (48), it timing offset through the quantities approaches zero as the noise variance is seen that decreases, or the number of observed vectors grows large. indicate that However, simulation results obtained with is nonzero, even in the absence of thermal noise.

(52) A channel snapshot is generated at each simulation run observed blocks. The conand is kept constant over stant in (52) is chosen such that the signal energy of user #1 per received sample is normalized to unity, i.e., . Correspondingly, the SNR equals , where is the variance of the Gaussian noise. 6) The cell radius is km, so that the maximum props, this agation delay is 6.66 s. Recalling that makes the maximum of equal to 26 and the search in . (38) can be limited to 7) Application of RCTE-USS would require a cyclic prefix samples. This is nearly twice the channel of duration and may represent an excessive overhead for the system under investigation. For this reason, we have set in the simulations. . As shown in 8) The maximum frequency offset is Section II-B, is related to the system parameters by the following equation: (53) where is the carrier frequency, the mobile speed, and the GHz, it is seen that speed of light. Assuming corresponds to a mobile speed greater than 300 km/h.

MORELLI: TIMING AND FREQUENCY SYNCHRONIZATION FOR THE UPLINK OF AN OFDMA SYSTEM

303

Fig. 5. Timing metric with  = 10,  = 0:3, and SNR = 10 dB.

Fig. 3. MSE of RCFE versus SNR with M = 2, 4.

is 15 dB and is either 2 or 4. It is seen that the performance is less than 0.3. Other simuladegradation is quite limited if tions (not shown for space limitations) indicate that RCFE gives , which is about what we exunbiased estimates for pected . The performance of RCTE is evaluated in terms of the proba, as well as of the average ISI power, due to imperfect bility and a timing estimates. Note that for a fixed error CIR , the ISI power is given by [13] (54) where if if otherwise. (55)

Fig. 4. MSE of RCFE versus  with M = 2, 4.

B. System Performance Fig. 3 illustrates the mean-square error (MSE) of the frequency estimates provided by RCFE as a function of SNR. The or . Marks indicate simulation retrue offset is sults, while thin lines are drawn to ease the reading of the graphs. and The MCRB is also indicated as a reference. For , the loss with respect to the bound is approximately . As expected, the MSE 2 dB, and reduces to 1 dB with degrades as increases. This is also shown in Fig. 4, illustrating the MSE of the frequency estimates as a function of . The SNR

Fig. 5 shows a realization of the timing metric for and . The true timing offset is , and the has a maximum at , SNR is 10 dB. It is seen that with secondary peaks at a distance of multiples of from the maximum. Note that if the SNR is decreased, a peak other than the first from left might occasionally prevail. If this were to happen, the estimate would be far away from , and a very large timing error (outlier) would be incurred. Outliers may have catastrophic effects on the system performance and must be avoided. This can be achieved by choosing the cell radius small is less than . In these circumstances, no enough so that secondary peaks can occur on the interval . Note that in the system under investigation, the condition holds with a cell radius of 2 km, which is typical for outdoor cellular mobile radio systems. as a function of SNR with . Fig. 6 illustrates or , and the timing The frequency offset is either is computed using the estimate provided by metric RCFE. Marks indicate simulation results, while the thick lines labeled UB and LB represent the upper and lower bounds

304

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

Fig. 6.

P (") versus SNR with 

= 0 and

= 0:3.

Fig. 8. BER with uncoded BPSK.

Fig. 7.

Average ISI power versus SNR with M = 2, 4.

given in (50) and (51). As expected, simulations obtained with lay between the bounds, which are very tight at interme, however, exhibits a diate/high SNR values. With floor, meaning that the timing estimator is sensitive to frequency offsets. This comes from the fact that in deriving the RCTE, we in (11) by , and this approxihave replaced the matrix mation is valid only with very small values. Similar conclusions can be drawn from inspection of Fig. 7, which shows the average ISI power versus SNR with either

or . Note that doubling increases the performance of RCTE by approximately 3.5 dB. The overall system performance has been computed in terms of uncoded bit-error rate (BER). Fig. 8 illustrates the BER for a BPSK system employing the proposed frequency and timing esand . The curve labeled PCK cortimators with and reflects the responds to perfect channel knowledge degradations due to frequency and timing errors. Curve CFTE assumes channel, frequency, and timing estimates. Channel estimation is performed by replacing and with their estimates and in (19). Theoretical results with ideal synchronization are also indicated for reference. We see that the results with PCK and with ideal synchronization are very close, meaning that the performance loss due to frequency and timing errors is negligible. Imperfect channel estimation is seen to bring about a degradation of 1 dB. Fig. 9 shows the BER of the already-present users during the acquisition process of user #1. Here, the frequency offset of the desired user is modeled as a random variable uniformly dis, ], whereas is kept fixed to 20. As we can tributed over [ see, the curves exhibit a floor that increases with . The reason is that the signal of user #1 is not orthogonal to those of the other users, due to the frequency and timing offsets. This produces interference at the DFT outputs that inevitably degrades the system performance. It is worth noting that this problem only occurs during the acquisition phase of user #1, which is limited to a few OFDM blocks. So far we have assumed that all the active users, except #1, are perfectly aligned to the BS references. In practice, synchronization of the already-present users is not ideal, and the resulting interference at the DFT outputs deteriorates the performance of the proposed frequency and timing estimators. Fig. 10 shows simulations illustrating the performance loss incurred by RCFE

MORELLI: TIMING AND FREQUENCY SYNCHRONIZATION FOR THE UPLINK OF AN OFDMA SYSTEM

305

VIII. CONCLUSIONS Synchronization algorithms have been proposed for frequency and timing recovery in the uplink of an OFDMA system. They exploit knowledge of pilot symbols and require a limited computational load. In contrast to other existing methods, they have a fast acquisition and can be used in applications where the subcarriers of different users are interleaved. Their performance has been investigated by theoretical analysis and computer simulations. It has been found that a receiver endowed with these algorithms has virtually the same performance as a perfectly synchronized system. APPENDIX In this Appendix, we compute the probability of the pairwise defined in Section VI-C. To simplify timing error event and observe that the analysis, we assume , where , and is obtained by setting in (39). We begin by collecting (17), (20), and (39). This yields (A1)
Fig. 9. BER of the already-present users during the acquisition of user #1.

with (A2) and . The noise vector is given by (A3) and has a Gaussian distribution with zero mean and covariance matrix . Substituting (A2) into (A1) and bearing in mind (40) produces

(A4) where is the orthogonal complement of . Next we assume that the SNR is sufficiently high so that the last term in (A4) (Noise Noise contribution) can be neglected. In these circumstances, it is seen that conditioned on is a Gaussian random variable with mean a given , and variance , where
Fig. 10. Performance loss of RCFE and RCTE due to users imperfect synchronization.

(A5) Therefore, the probability of to be conditioned on is found

and RCTE in a scenario where the signals of the interfering users are not perfectly synchronized. In particular, frequency errors are modeled as random variables with uniform distribution over , ], whereas timing errors take integer values within [ , [ ] with uniform probability. The SNR is 15 dB and the frequency and timing offsets of user #1 are and . It is seen that both the MSE of the frequency estimates and the average ISI power increase with and . However, the loss with (perfect synchronization respect to the ideal case of the already-present users) is quite tolerable.

(A6) where (A7)

306

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 52, NO. 2, FEBRUARY 2004

Note that depends on the channel realization only through . The average probability can over , the be computed by integrating probability density function (pdf) of , i.e., (A8) where denotes the interval spanned by . , we need to specify the statistics of . To compute We assume that is a zero-mean Gaussian vector with co. Also, using the Cholesky decomposition variance matrix , (A5) becomes , where has a Gaussian distribution with zero mean and covariance matrix (A9) It can be shown [14] that has the same pdf as (A10) where are the nonzero eigenvalues of and are complex-valued independent Gaussian random variables with zero mean and unit variance. Using this can be written in the form property, (A11)

REFERENCES
[1] S. Kaiser and K. Fazel, A spread-spectrum multicarrier multiple-access system for mobile communications, in Proc. 1st Int. Workshop on Multicarrier Spread Spectrum, Apr. 1997, pp. 4956. [2] J. J. van de Beek, P. O. Borjesson, M. L. Boucheret, D. Landstram, J. M. Arenas, P. Odling, C. Ostberg, M. Wahlqvist, and S. K. Wilson, A time and frequency synchronization scheme for multiuser OFDM, IEEE J. Select. Areas Commun., vol. 17, pp. 19001914, Nov. 1999. [3] H. Sari and G. Karam, Orthogonal frequency-division multiple access and its application to CATV networks, Eur. Trans. Telecommun., vol. 9, pp. 507516, Dec. 1998. [4] T. M. Schmidl and D. C. Cox, Robust frequency and timing synchronization for OFDM, IEEE Trans. Commun., vol. 45, pp. 16131621, Dec. 1997. [5] J. J. van de Beek, M. Sandell, and P. O. Borjesson, ML estimation of timing and frequency offset in OFDM systems, IEEE Trans. Signal Processing, vol. 45, pp. 18001805, July 1997. [6] F. Daffara and A. Chouly, Maximum-likelihood frequency detectors for orthogonal multicarrier systems, in Proc. Int. Conf. Communications, June 1993, pp. 766771. [7] P. H. Moose, A technique for orthogonal frequency-division multiplexing frequency offset correction, IEEE Trans. Commun., vol. 42, pp. 29082914, Oct. 1994. [8] B. Yang, K. B. Letaief, R. S. Cheng, and Z. Cao, Timing recovery for OFDM transmission, IEEE J. Select. Areas Commun., vol. 18, pp. 22782290, Nov. 2000. [9] M. Luise, R. Reggiannini, and G. M. Vitetta, Blind equalization/detection for OFDM signals over frequency-selective channels, IEEE J. Select. Areas Commun., vol. 16, pp. 15681578, Oct. 1998. [10] U. Mengali and M. Morelli, Data-aided frequency estimation for burst digital transmission, IEEE Trans. Commun., vol. 45, pp. 2325, Jan. 1997. [11] M. Morelli and U. Mengali, An improved frequency offset estimator for OFDM applications, IEEE Commun. Lett., vol. 3, pp. 7577, Mar. 1999. [12] U. Mengali and A. N. DAndrea, Synchronization Techniques for Digital Modems. New York: Plenum, 1997. [13] M. Speth, S. Fechtel, G. Fock, and H. Meyr, Optimum receiver design for wireless broadband systems using OFDMPart I, IEEE Trans. Commun., vol. 47, pp. 16681677, Nov. 1999. [14] P. Ho and D. K. P. Fung, Error performance of interleaved trellis-coded PSK modulations in correlated Rayleigh fading channels, IEEE Trans. Commun., vol. 40, pp. 18001809, Dec. 1992. [15] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed. New York: McGraw-Hill, 1991.

where

is the characteristic function of

, given by [15] (A12)

To proceed further, we expand (A12) by partial fractions. Assuming that the eigenvalues are all different, this produces (A13) where is defined in (49). Then, (A11) becomes (A14) and substituting into (A8) leads to (48).

Michele Morelli received the Laurea (cum laude) in electrical engineering and the Premio di Laurea SIP from the University of Pisa, Pisa, Italy, in 1991 and 1992, respectively. From 1992 to 1995, he was with the Department of Information Engineering, University of Pisa, where he received the Ph.D. degree in electrical engineering. In September 1996, he joined the Centro Studi Metodi e Dispositivi per Radiotrasmissioni (CSMDR) of the Italian National Research Council (CNR) in Pisa, where he held the position of Research Assistant. Since 2001, he has been with the Department of Information Engineering, University of Pisa, where he is currently an Associate Professor of Telecommunications. His research interests are in wireless communication theory, with emphasis on synchronization algorithms and channel estimation in multiple-access communication systems.