You are on page 1of 28

Three Non-Pilot based Time- and Frequency

Estimators for OFDM
Jan-Jaap van de Beek1

Per Ola Börjesson1

Daniel Landström1
Per Ödling1

Marie-Laure Boucheret2

Julia Martinez Arenas2
Sarah Kate Wilson1

Elsevier’s Signal Processing – Special Issue COST 254 workshop.


Luleå University of Technology, Division of Signal Processing, SE–971 87 Luleå, Sweden.


Ecole National Supérieure de Télécommunications, Site de Toulouse,
BP4004 CEDEX, Toulouse, France.

Jan-Jaap van de Beek
Division of Signal Processing
Luleå University of Technology
SE–971 87 Luleå, Sweden

+46 920 72029


+46 920 72043


This paper has been presented in part at the COST 254 Workshop, July 7-9, 1997, Toulouse, France,

at the International Conference for Universal Personal Communications, October 13-16, 1997, San Diego,
CA, USA and at Gretsi, September 15-19, 1997, Grenoble, France.


Abstract – Time-domain Maximum-Likelihood (ML) estimators of time
and frequency offsets are derived for three Orthogonal Frequency Division
Multiplexing (OFDM) signal models: a pulse-shaped one-shot OFDM signal, a stream of multiple OFDM signals and an OFDM signal in a dispersive
channel environment. We then develop structures to simplify their implementation. Simulation results show the relative performance and strengths
of each of these three estimators.



In this paper we focus on the estimation of time and frequency offsets for Orthogonal Frequency-Division Multiplexing (OFDM) symbols for synchronization purposes in
wireless environments. The synchronization of an OFDM transmitter and receiver is
important because OFDM systems, are in general more sensitive to time and frequency
offsets than single carrier systems [17]. Not only may synchronization errors cause
intersymbol-interference (ISI), they also can cause the loss of orthogonality between the
subcarriers resulting in intercarrier-interference (ICI). The sensitivity of OFDM systems
to synchronization errors has been documented in, e.g., [16][27][17][7]. Most time and
frequency estimators for OFDM require pilot symbols, for example [3][26][14]. However,
too many pilot symbols can lower the overall information rate. Hence methods that do
not use pilots are desirable. Such methods have been investigated in [4][5][19][20][23][24]
and patented in [19]. The method described in [23] differs from the previous methods
in that it describes the time-domain maximum-likelihood (ML) estimator for OFDM
systems with a cyclic-pre…x [15] in an additive white Gaussian noise (AWGN) channel.
This estimator not only exploits the redundancy in the cyclic-pre…x, but also the relative power of the received signal to determine both the time and frequency offset of the
received OFDM symbol. One of the main contributions of this paper is the knowledge
that the estimator is optimal in the ML sense and as such gives an upper bound on the
performance of a time and frequency estimator.
However, the estimator in [23] was derived for the case of a static AWGN channel and

First. The paper does not target a particular application. The three cases are as follows. we present the ML estimator of a time offset for systems with channel dispersion. We show that 3 . Here. and thus under a dispersive channel [9]. Secondly. in Section 2 we discuss general properties of the OFDM system that are used in the estimation procedure. This paper describes three signal models for wireless OFDM systems and derives the ML time and frequency estimator associated with each. we propose an estimator for a stream of OFDM symbols that can help compensate for such an SNR threshold. this threshold may reduce the applicability of the estimation concept. We develop an ML estimator that is based knowledge of the channel dispersion. wireless OFDM systems often operate in a multipath environment. The purpose is also to build knowledge about the structure of ML estimators for OFDM systems so that one can use this knowledge to design for a speci…c system. The goal of the paper is to provide a toolbox of techniques that can be used to improve the performance of the estimator from [23] by adapting the signal model to more accurately model the system. pulse shaping is used in many wireless OFDM systems to reduce the out-of-band emission [8][11][21].as such it is not necessarily suitable for wireless OFDM channels. Finally. proposed for multiuser communication systems [25]. we present the ML estimator of time and frequency offsets for systems in which these parameters vary slowly. Instead we will be working with signal models for a class of estimators and show how different system properties can be incorporated in the model. In Section 3 we present three signal models and the associated ML estimators. the ML estimator’s performance suffers below a given SNR threshold [23]. Secondly. In Section 4 we illustrate the performance of the estimators with simulation results and we discuss these in Section 5. First we present the ML estimator for time and frequency offsets for systems using pulse shaping. it will suffer when applied to a pulse-shaped OFDM signal. For example. This paper is organized as follows. We focus on the role of the signal model. We present an estimator for pulse-shaped systems [10][13]. Because the ML estimator in [23] was derived for a non-pulse shaped OFDM system. rather than on a particular application. for bandlimited signals. Applying the estimator derived in [23][19] will result in a error ‡oor in the time and frequency offset estimation. Finally.

The phase offsets will in a coherent system be compensated for in a channel equalizer and the system performance depends on the performance of the channel estimator and equalizer. The time offset will then appear as phase offsets of the demodulated data symbols and a channel estimator can not distinguish these from channel phase distortions. In the receiver the cyclic pre…x is removed and the data are demodulated by means of a DFT. The sensitivity to frequency offsets for the AWGN channel has analytically been investigated in. In [17] the ICI is interpreted as a degradation of the SNR and quantitatively 4 . They describe the amount of redundancy in the signal that the estimator can incorporating system properties as pulse shaping or knowledge about the stability of the clocks and oscillators. For larger time offsets ISI and ICI occur [7]. it avoids ISI and ICI [15]. As long as the time offset  and the length of the channel impulse response together are smaller than half of the length of the cyclic pre…x L. Reference [23] gives a thorough investigation of the estimator’s performance in relation to N and L. The sensitivity to a symbol-time offset of  samples has been investigated in [7]. among others. ISI and ICI are avoided. The transmitter modulates complex data symbols on a large number of carriers Figure 1 by means of an inverse discrete Fourier transform (DFT). For the synchronization concept described in this paper the number of subcarriers N and the length of the cyclic pre…x L are important parameters. A frequency offset ε (normalized to the intercarrier spacing) results in ICI. [17]. Each block of samples is cyclically extended with a pre…x before it is transmitted over the channel [15]. The amount of ICI is proportional to ε2 and the signal-to-noise ratio (SNR). If the cyclic pre…x is longer than the length of the channel impulse response. In a differential system the system performance depends on how fast the channel is varying. 2 OFDM systems and synchronization Figure 1 shows the structure of the OFDM system on which our signal models are based. the estimator performance can signi…cantly be improved. Symbol timing requirements are relaxed by increasing the length of the cyclic pre…x.

model (2) applies. if k1 = k2 + N and 0  k2 < L. In the next section we discuss three models of the received OFDM signal. N samples ahead. Thus. n(k ) is complex AWGN with variance  2n . In a real system the transmitted signal consists of a stream of OFDM symbols. each containing this redundancy. For example. The ML estimator based on (2) yields a fast one-shot estimator 5 . By observing this correlation one may tell where the OFDM symbol is likely to start. Cs (k1 . because much of the statistical structure of the transmitted signal s(k ) is transferred to the received signal r(k ). (1) Notice that this model only re‡ects the appearance of one OFDM symbol (one cyclic pre…x) in the transmitted signal. We focus on the estimation of the unknown offset parameters  and ε from the received data r(k ). Based on the above assumptions. otherwise. It is possible to estimate  and ε from r(k ). In case the transmitter does not employ pulse-shaping and there is no channel dispersion. We will compare the signal models developed in this paper with our reference model (2).      0.2 dB. (2) where s(k ) is a sample of the Gaussian process with covariance function (1). Each transmitted OFDM symbol contains L consecutive samples which are pairwise correlated with L other consecutive samples. In each model we assume that the transmitted discrete-time signal is Gaussian [6]. if the original working SNR is 10 dB. k2 ) =   if k2 = k1 + N and 0  k1 < L. we assume that the transmitted signal s(k ) is Gaussian with covariance function    if k1 = k2 . a frequency offset which is 4% of the intercarrier spacing results in interference that decreases the SNR with 0. for a AWGN channel.investigated. As we will see in the rest of the paper.       2s . the frequency offset can also be estimated exploiting this redundancy. Because of the cyclic pre…x this signal is not white. in [23] the received signal r(k ) is modelled as r(k ) = s(k  )e j 2εk/N + n(k ) ∞ < k < ∞.

We repeat the result from [23] here for later reference b 0 = arg max {| N ()| + 0 ()} . Then. ε)ML of  and ε. the covariance function of the received signal contains the information due to (multiple) cyclic pre…xes. The term N (m) collects this correlation and the term 0 (m) compensates for high contributions due to large samples rather than a large correlation. The joint maximum likelihood estimate ([ . ε) r . ε)ML = arg max  . pulse shaping or channel dispersion and noise. 3 Estimators of time and frequency offsets In this section we generalize model (2) and estimator (3) to include pulse shaping. (5) For the signal models described in this section. ε0 = b 1 ] N (b  0 ).ε rH Cr 1 (. k=m mX +L 1  (4) |r(k )|2 + |r(k + N )|2 . and channel dispersion. the quadratic form can be put in an explicit expression which leads to implementable structures. and N (m ) = 2 0 (m ) = mX +L 1 r  (k )r (k + N ) . k=m and where = SNR . 2 (3) where b 0 is the symbol-time offset. bε0 is the frequency offset. We introduce the vector r for the received signal. SNR + 1 Figure 2 This estimator extracts the information carried by the cyclic pre…x by correlating the received signal and a delayed version. given the received Gaussian data vector r with known covariance matrix Cr becomes ([ .of  and ε [23]. 6 . with covariance matrix Cr . multiple symbols.

Pulse shaping affects the performance of an estimator in two ways.   2  pT (k ) = 1. N +L in [12]. The choice of pulse shapes is beyond the scope of this paper.1 Offset estimators for pulse shaped systems Some OFDM systems require pulse shaping in order to suppress the system’s sidelobes and out-of-band emission. (7) P  k < N + L. and thus also the performance of 7 . where p(k ) is the pulse shape. Equation (8) models adjacent symbols as white Gaussian with time-invariant average power. 2 P 1 . k = 0  k < N + L. First. we model the received signal as r (k ) = g (k )s(k  )e j 2εk/N + n(k ) ∞ < k < ∞. For some pulse shapes and parameter choices. [7][8][11][21][12]. (8) where .g. k = 0  k < N + L. one negative and one positive. e. (6) 0  k < P.        1 1 + cos k (N +L P )  . Examples of pulse shapes found in existing systems or system proposals are (see Figure 3) the Hanning pulse.  1 pH (k ) = 1 2  cos 2k N +L 1  investigated in [7][12]. The use of pulse shaping is suggested in. For the choice pR (k ) = 1.. As in the reference model (2). and   p(k ) 0  k < N + L g (k ) = . For a general pulse shape.3. s(k ). we model knowledge (its correlation properties and its shape) for one OFDM symbol only (N + L consecutive samples) in the transmitted signal.   1  1 cos P k 1  . For most practical pulses it reduces the amplitude in the parts of the signal that are cyclically repeated. the tones in the OFDM symbol Figure 3 lose their orthogonality. ε. model (8) reduces to (2). it will change the amplitude in some parts of the signal. P k <N +L P. This reduces the correlation in the signal.  1 otherwise. for instance either (6) or (7). and n(k ) as in (2). r(k ). and the Tukey pulse.

Since this term does not depend on  it 1 (thus. the net effect of the pulse shaping is positive. Secondly. To derive the ML estimator for the pulse-shaped OFDM system. ε) = | N ()| cos {] N ( ) + 2ε} + 0 (). In Appendix A it is shown that this function becomes  (. the reduction in amplitude introduces a time-varying signal power that also carries information about the symbol time offset. In order to make the …lter length …nite. This is equivalent to subtracting n it from the log-likelihood function  (. Subwill not change the maximizing argument b tracting this constant from the log-likelihood function is equivalent to rede…ning h0 (k ) 8 . 2 (11) The …lter h0 (k ) has in…nite length. For some systems. we subP∞ 2 1 tract the constant 12 SNR+1 k= ∞ |r (k )| from 0 ( ). h0 (k m) |r(k )|2 .  0 otherwise  SNRg 2 (k+N )+1   0k<L   SNR(g2 (k)+g2 (k+N ))+1 2 SNRg (k N )+1 h0 (k ) = N k <N +L . b ε1 = 1 ] N (b  1 ). where N (m ) = 0 (m ) = ∞ X k= ∞ ∞ X hN (k m )r (k )r  (k + N ) . ε). ε) = ([ rH Cr 1 (. ε) r. there is no performance loss). This information may improve the performance of an estimator. as simulations will show.the estimator. SNR(g 2 (k)+g 2 (k N ))+1    1  otherwise (10) SNRg 2 (k)+1 The SNR is the ratio of the average signal energy to the average noise energy (SNR =  2s / 2n ). (9) k= ∞ and where   hN (k ) = 2SNRg (k)g (k+N ) SNR(g 2 (k)+g 2 (k+N ))+1 0k<L . ε) that maximizes the log-likelihood function  (. we …nd the pair . The ML estimator maximizes the log-likelihood function and becomes b 1 = arg max {| N ()| + 0 ()} .

 SNRg 2 (k+N )+1 g 2 (k)    SNR(  g 2 (k)+g 2 (k+N ))+1   2 (k )  g 1   2 SNRg (k)+1 h0 (k ) = SNRg 2 (k N )+1 g 2 (k)    SNR(  g 2 (k)+g 2 (k N ))+1     0 as 0k<L Lk<N N k <N +L . Here we assume that the transmitted signal s(k ) has 9 . This …lter now has …nite length and is thus implementable. Figure 4 shows the estimator structure. Figure 4 Estimator (11) exploits two types of information. the time-varying signal power contains information about . (12) where .2 Offset estimators for systems with M consecutive OFDM symbols The signal models (2) and (8) incorporate a priori knowledge about one transmitted OFDM symbol in the transmitted signal and use only N + L samples in the estimation procedure. the signal model (8) reduces to (2) and the estimates b 1 and b ε1 coincide with b  0 and bε0. 3. A similar estimator performing one-shot estimation in a time–division multiuser systems is investigated in [10][13][22]. otherwise where  is as in estimator (3). and collected by the …lter hN (k ). and n(k ) as before. Moreover. ε. The estimator extracts this information by means of the …lter h0 (k ). It bene…ts from the statistical structure of M consecutive OFDM symbols. When p(k ) = 1. This …lter h0 (k ) functions much like a matched …lter to |r(k )|2 . We model the received signal r(k ) as r(k ) = s(k  )e j 2εk/N + n(k ) ∞ < k < ∞. First the correlation between the samples in the cyclic pre…x is used. In this section we describe a signal model that incorporates multiple OFDM symbols to improve the estimator performance. Estimators based on these models are one-shot estimators in the sense that they generate estimates of the time and frequency offset for each symbol by exploiting the information carried by only that symbol.

For instance. .covariance function    if k1 = k2      s2 if k = k + N and m (N + L)  k < m (N + L) + L and m = 0 . clock drifts are such that for moderate M. In appendix B. be extended to incorporate pulse shaping as well. (14) Thus the optimal way (in a maximum likelihood sense) to process the M consecutive symbols is to average the M functions N () and 0 () in (4) to create the log-likelihood function. 10 . . the drift from one symbol to the next may be large. ε) = | N ( )| cos {] N ( ) + 2ε} + 0 ( ). where N (i) = M X1 N (i + m (N + L)). the assumption holds. not all of the received symbols obey the perfect clock we assume. k2 ) =   if k2 = k1 + N and m (N + L)  k1 < m (N + L) + L and m = 0 . In practical applications there may be a drift in the transmitter symbol clock relative to the receiver clock. In most applications. . in systems with a multiple access scheme based on time-division. . in a straightforward way. M      0 otherwise The purpose of this model is to increase the information available to the estimator and to generate a stable symbol clock. however. The optimal estimator for model (12) now becomes b 2 = arg max {| bε2 = N ( )| 1 ] N (b 2). M 1 2 2 Cs (k1 . m=0 and where N (m) and 0(m) are as de…ned in (4). The estimator (14) can. (13) m=0 0 (i) = M X1 0 (i + m (N + L)). 2 + 0 ( )} . we derive the log-likelihood function of  and ε to get  (. In particular when M is large.

the correlation structure of s(k ) is not transferred so directly to the received signal r(k ). for instance the European Digital Audio Broadcasting (DAB) network [1]. the Digital Video Broadcasting (DVB) [2] and the Universal Mobile Telecommunications System (UMTS) [25] dispersion in the channel will affect the correlation properties of the received OFDM signal. For the purpose of data detection in an OFDM receiver. As we will see though. (15) where . the time and frequency synchronization concept targeted by this paper works on the received signal r(k ) before the DFT. h(k ). As with the previous models. We can write this matrix as Cr = HCs HH + 2n I. UMTS). is not as tractable as in the previous cases and some approximations are necessary to reach an implementable estimator. where the matrix Cs is the correlation matrix of the transmitted OFDM signal s(k ) 11 . Given a channel impulse response. DVB. optimal synchronization which is conceptually straightforward. Because of the dispersive channel.3. ε. we model the received signal r(k ) as r(k ) = (h  s)(k  )e j 2εk/N + n(k ) ∞ < k < ∞. we will show that estimation is possible. this assumption and the following derivation will give an upper bound on the performance of an estimator that does not use pilots and is therefore useful when evaluating other estimator’s performance. However. the receiver will have to track a time-varying channel in others (such as DAB. h(k ). While this assumption holds for some applications (such as copper wire channels). s(k ) and n(k ) are as de…ned above. However. We assume that we know the channel impulse response. the ML estimator depends on the covariance matrix Cr of the received data vector r.3 Offset estimators for systems with channel dispersion In many applications. this channel dispersion appears as a multiplicative distortion of the data symbols and correction is straightforward. In this section we state a signal model incorporating r(k )’s correlation due to the channel dispersion. Although the channel colors the received signal. We assume that the length H of h(k ) is smaller than the length of the cyclic pre…x L.

For this reason. so does the complexity of the ML operator. The only non-zero elements of Ccp () are concentrated in a (N + L + H )  (N + L + H ) submatrix.. The …rst term C0 is a band matrix representing the channel correlation.e. The optimal estimator for model (15) is then b 3 = arg max   rH Cr 1 () r . from previous applied work [23] we have seen that the performance of frequency estimators based on models (2) and (8) is sufficient in a dispersive channel. Secondly. . i. k2 )(1). we show a derivation of the ML estimator with a more manageable complexity.and whose (k1 . First we want a tractable estimator that gives an indication of likely performance in a dispersive case so that we can gain an understanding of what synchronization information is contained in r(k ). The third term Ccp ( ) is a matrix representing the total correlation of samples located in the repeated parts. whereas the performance of the time estimators 2. (16) As the size of Cr increases. 2 . C0 = 2s HHH . h(H 1)). Speci…cally. The …rst and second terms form a bandmatrix C0 +  2n I that is independent of  . the cyclic pre…x. and H is a matrix whose entries are of the form. is not.j =  2s if |j i| = N + . The second term 2n I is a matrix representing the noise correlation. where     2 if |j i| = N + . ignoring the estimation of ε. Hi.  + L 1]   s I()i. That is. j ∈ [. Ccp ( ) = HI()HH . Our motivation In our derivation. Cr = C0 + 2n I + Ccp ( ) . We can rewrite Cr as the sum of three separate matrices. we present only b for this is twofold. h(1). .  + L 1] . . Thus we focus on estimation of b 12 . i ∈ [. and the dependency of  appears only in the position of this submatrix. k2 )th entries are Cs (k1 .(i:i+H 1) = (h(0).     0 otherwise Note that only Ccp () depends on the unknown  . .

this does not necessarily mean that the estimator performance degradation is negligible. Studying the behavior of the matrix A () for different communication scenarios gives some indication about how information about  is contained in the received signal. Although the difference between the matrices may often be quite small. Variations in 1  shift the elements of A () along the diagonals. combining (16) and (17) we get n  b rH A () r + rH C0 +  2n I 3 = arg max   H r A () r .k  +i (0) rk+i . Using a matrix inversion lemma [18].j () rj rk A =  ( )  k = j XX i ek rk A k 13 . Substraction of rH (C0 +  2n I) r is P∞ 2 1 the equivalent of the substraction of 12 SNR+1 k= ∞ |r (k )| in the derivation of the n estimator for pulse shaping (11). this is a means of getting a more practical estimator. Because Ccp () is a block matrix with zeros outside an (N + H + L)  (N + H + L) submatrix. is approximately also limited to a (N + H + L)  (N + H + L) submatrix. Based on the observation that information about  is largely concentrated in a relatively small section of e () whose entries A (). we now de…ne the matrix A () = Cr 1 () C0 + 2n I  1 . the matrix A () . we will proceed with the approximation of A ( ) by a matrix A in positions outside the submatrix of size (N + H + L)  (N + H + L) are zero. We can further simplify our time-offset estimator by noting that: e ( ) r rH A XX e k. which is direct product of other matrices and Ccp ( ) . = arg max (17) 1 o r  since the maximizing argument does not depend on rH (C0 +  2n I) 1 r. we can show that    1  1 A ( ) = C0 +  n2 I Ccp () I + C0 + n2 I Ccp () 1 C0 +  n2 I  1 .To simplify our estimator. Rather. Then.

rl+ A l .= XX i e l.l+i (0) rl+ +i .


e e where A (0) is the -independent matrix A ().

Without channel dispersion. This result …nally 1 diagonals of A yields the time offset estimate ( b  3 = arg max (NP +L) ) i () . The estimate b (18) can. only the terms 0 (m). (18) i= ( N +L) where i (m ) = ∞ X h i (k m )r (k )r  (k + i ) . If The pulse shape appears in the covariance matrix Cr and in the diagonals of A pulse shaping is incorporated. This estimator structure is similar to the structure (3). …lters hi (k ) with i equal or close to N. First the products r(k )r (k + i) are formed. for a channel with a short impulse response. (19) k= ∞ e (0). the estimator reduces to (3) for ε = 0. Since all non-zero elements of =0 e (0) are concentrated in an (N + L + H )  (N + L + H ) submatrix there are only A 2 (N + L + H ) e (0) that contain non-zero elements. In general. A Low-complexity approximations of (18) can be obtained by disregarding …lters with a relatively small contribution to the quadratic form and just regarding the most signi…cant diagonals. but used in an environment without channel dispersion (18) reduces to (11). Filter hi (k ) has …nite support of length (N + L + H ) |i|. the complexity and the approximation error relative to the optimal estimator can be varied by varying the number of …lters. These are Figure 5 fed into the …lterbank containing …lters with impulse responses hi (k ) (the diagonals of e (0)). N (m ) and N (m) contain non-zero values and estimator (18) 3 then coincides with b 0 . Also. in a straightforward way be extended to incorporate pulse shaping as well. 14 . e (0). . Estimator (18) suggests the implementation structure shown in Figure 5. This result is important and where hi (k ) is the negative of the ith diagonal of A because one can now implement the estimator with …lters hi (k ). 0 and N contribute most. Typically.

 = 0. 7. We model the channel with the following static discrete-time exponentially decaying channel impulse response h( ) = e  /3 . The channel dispersion is a maximum of 8 samples. Figure 7 shows the estimator variance for the frequency estimators designed as described above. . . . In order to reduce ISI and ICI in a dispersive channel environment the system employs a cyclic pre…x of 16 samples. 4. We …rst investigate how the pulse shape in this system affects the estimator performance.2 Performance in a dispersive environment We now assume that the system operates in a dispersive environment. Thus. Figure 8 shows the performance of the time offset estimators (3). we assume that the transmitter shapes 8 samples of an OFDM symbol (4 samples at each side) with the Tukey pulse shape. designed with no knowledge of the channel dispersion 15 . designed for no pulse shaping.1 Performance in a system with pulse shaping We illustrate the estimator concepts described in the previous sections with the following system. the transmitter employs the Tukey pulse shape. and (11).4 Simulations 4. Both estimators are then extended with Figure 6 the averaging concept (14). designed with no knowledge of the pulse shape nor the channel dispersion. Also. As above. (11). Figure 6 shows the variance of the time-offset estimators (3). This error ‡oor can be decreased by averaging but consideration of the pulse shape in the estimator design removes this error ‡oor. one estimator does not take the pulse shape into account and one does. designed for pulse shaping. yielding another two estimators (see Figure 6). In our system scenario the inclusion Figure 7 of the pulse shape in the model does not signi…cantly improve the performance of the frequency estimator. Consider a …ctitious OFDM system with 64 subcarriers. . Ignoring the pulse shape in the design of the estimator results in a performance error ‡oor.

Simulations show that the use of this estimator in systems with pulse shaping is bene…cial in both AWGN and dispersive environments. The simulation results in this section suggest that in systems with pulse shaping. all of the estimators can improve on the performance of the ML estimator designed for an AWGN channel [23]. designed without knowledge of the pulse shape but incorporating the channel dispersion. estimator performance is increased by incorporating the pulse shape in the signal model. The estimator designed for a dispersive channel can be quite complex. but is useful since it gives a bound on the achievable performance.but incorporating the pulse shape. When applied to a wireless system. We have also found that when the clock drift is small. where time and frequency offsets vary slowly. 5 Summary We have presented three signal models and their respective ML estimators of time and frequency offsets for OFDM systems. one estimator takes Figure 8 neither pulse shaping nor dispersion into account (reference estimator). 16 10 . We presented the ML estimator based on a signal model that considers the pulse shape. Thus. Both estimators are also extended Figure 9 with the averaging concept (14). one estimator takes the pulse shape only into account. Each of the models confronts a difficulty in estimating the time and frequency in a wireless OFDM system. and one estimator recognizes the channel dispersion only. In this scenario the performance of each estimator experiences an error ‡oor because none of the estimators is perfectly matched to the true signal properties. Figures 9 and 10 show the performance of the joint estimators (3) (the reference estimator) and (11) (the pulse shaping estimator). averaging can signi…cantly improve the error ‡oor of the estimator. Using knowledge of the pulse shape property improves the performance more than using knowledge of the channel dispersion alone. In our dispersive environment the use of estimator Figure (11) combined with averaging decreases the error ‡oor. Moreover. in systems with channel dispersion. and (18) for ε = 0. averaging decreases the error ‡oor.

j 2ε  where D is the discriminant of C. and C its covariance matrix. r is a complex Gaussian vector with joint density function f (r) log f (r) = rH C 1 r+ log c2 . ε) . k∈ / [.A The ML estimator for pulse shaping For model (8) the log-likelihood function  (.+L 1]∪[+N. First. and f (r(k )) denotes the Gaussian density function for other values of k .+L 1]  f (r(k )) . By calculating the densities f (r(k )) and f (r) we will …nd the log-likelihood function  (. Now. ε) can be written as  Y Y f (r(k ). r(k + N )). + log c1   where  2s  E |s(k )|2 and 2n  E |n(k )|2 . . r(k + N ))   (. k ∈ [. +N +L 1] where f (r(k ). log f (x) =  1 SNRg 2 (k + N ) + 1 |r(k )|2 D  2 + SNRg (k )g (k + N  ) Re e D 1 SNRg 2 (k D r (k )r (k + N ) . de…ne h iT r = r(k ) r(k + N ) . Secondly.  + L 1] (samples coupled through the cyclic extension). C becomes   H SNRg 2 (k ) + 1 2  C = E r r = n SNRg (k + N )g (k  SNRg (k  )e j 2ε  )g (k + N 2 SNRg (k + N  )e j 2ε . r(k ) is a complex Gaussian variable with density f (r(k )) and log f (r(k )) = |r (k )|2 2 2s g (k )2 +n ∞<k<∞ . ) + 1 Thus. D =  n2 SNR g 2 (k  ) + g 2 (k + N Straightforward calculations now yield 17   ) + 1 |r (k + N )| 2   ) + 1 . r(k + N )) is the joint Gaussian probability density function for the pair (r(k ). ε) = log  k∈[.

ε) = k∈[. as before. ε) = | N ()| cos {] N ( ) + 2ε} + 0 (). the index of the …rst sample of the …rst complete symbol. 0 Im . m=0 The observation samples r(k ) can now be divided into the samples r(k ). the probability density function of the observation can be written as f (r |. . k ∈ / I ∪ I 0 which are independent. r(k + N )) + hN (k log f (r(k )) k∈ / [.+N +L 1]  ) Re e j 2ε  r (k )r (k + N ) + k= + L 1 X h0 (k  ) | r ( k ) |2 . Using these properties. k ∈ I ∪ I 0 . . The arrival time  is.X  (. Finally  (. k= where hN (k ) and h0 (k ) de…ned in (10). r(k + N )) k∈I Y k∈I / f (r(k )) = Y f (r(k ). . . . Consider the cyclic pre…xes Im and their copies Im for each symbol m = 0. . modelling 0 the unknown channel delay. . . f (r(k )) f (r(k + N )) k k∈I 18 . and de…ne the union of all these indexes I  I0  M [1 m=0 M [1 Im . which are pairwise dependent. where N () and 0 () as de…ned in (9). m(N + L) +  + N + L 1} . r(k + N )) Y f (r(k )) . .+L 1]∪[+N. . B The ML estimator for M consecutive symbols Assume that the observation interval contains M complete OFDM symbols. 0 Im  {m(N + L) +  + N. . m(N + L) +  + L 1} . M 1: Im  {m(N + L) + . .+L 1] = + L 1 X X log f (r(k ). and the remaining samples r(k ). ε ) = Y f (r(k ).

pages 804–809. prETS 300 744. Conf. 1995. are de…ned in (13). A new frequency detector for orthogonal multicarrier transmission techniques. July 1995. r(k + N )) . r(k + N )) Y1 Y f (r(k ). Intern. Maximum-likelihood frequency detectors for orthogonal multicarrier systems. [5] F. Conf. Frequency synchronization algorithms for OFDM systems suitable for communication over frequency-selective fading channels. the loglikelihood function given the observation of M symbols now becomes  (. portable and …xed receivers. Classen and H. Chicago. Chouly. June 1994. Adami. Sweden. ETSI – European Telecommunications Standards Institute. pages 1655–1659. [4] F. = f r k f r k N f r k f r k N ( ( )) ( ( + )) ( ( )) ( ( + )) m=0 k∈I k∈I m Using the statistical properties of r(k ) and following the lines of appendix A. 1996. European Telecommunications Standard. In Proc. Daffara and O. Geneva. Daffara and A. IEEE Vehic. [2] Digital broadcasting systems for television. IEEE Vehic. The remaining part can be rewritten because of the independence of samples from different symbols ! M Y f (r(k ). Technol.The last factor is independent of  and ε and can thus be omitted. In Proc. Valbonne. ETS 300 401. References [1] Radio broadcasting systems. May 1993. Digital Audio Broadcasting (DAB) to mobile. Meyr. IL. Technol. sound and data services. ε) = M L 1 X1  +X m=0  Re e r (k + m (N + L))r(k + m (N + L) + N ) j 2ε  k= + M L 1 X1  +X m=0 = | where N ( ) and N ( )| cos {] N ( ) 0 ( ) + 2ε} + |r(k + m (N + L))|2 k= 0 ( )). Stockholm. Feb. 19 . Commun. [3] F. pages 766–771. France. In Proc. volume 2. Conf... Sept.. volume 3. Switzerland.

Martinez Arenas. USA. 1993. Adjacent channel interference in an OFDM system. France. Boucheret. 1996. C. Technol. J. Jakes. Clipping distortion in DMT ADSL systems. Gudmundson and P. Gross and D. Andersson. Microwave mobile communications. 1996. In Proc. [8] R. Classic Reissue. 1997.-J. Grenoble. and C. Royal Institute of Technology. Le Floch. pages 918–922. In GRETSI.-L. [10] D. 83(6):982–996. M. [11] B. Sept. Ecole National Supérieure des Télécommunications. van de Beek. IEEE Vehic. Landström. 42(10):2908–2914. Ödling.-O. In English. and P. San Diego.. Oct. 29(24):2080–2081. [13] J. O. Ödling. Börjesson. D. 1974. Malmgren. IEEE Press. Berrou. Nov. J. Commun. IEEE. Apr. Conf. Martinez Arenas. Coded orthogonal frequency-division multiplexing. 1994. 1997. [14] P. J. van de Beek. Application des transmissions à porteuses multiples aux communications radio mobiles. [9] W. IEEE Trans. Conf. Haas. Veeneman. Universal Personal Commun. Sweden. Oct. Int. A technique for orthogonal frequency-division multiplexing frequency offset correction. Lett. Stockholm. Time and frequency offset estimation in OFDM systems employing pulse shaping.-L.[6] R. P. pages 279–283. volume 2. Paris. Landström. Single Frequency Broadcasting Networks. M. [12] G. Proc. Moose. Electron. and P. 20 . Jan. Atlanta. Boucheret. France. PhD thesis. P... CA. In Proc. O. GA. Apr. Performance of a joint time and frequency offset estimator for different pulse shapes in OFDM systems. June 1995. thesis.-J. Börjesson. New Jersey. 1997. Alard. [7] M.. Phd. Piscataway. M.

van de Beek. IEEE Trans. CO. Commun. Sandell. Peled and A. United States Patent. Pollet. Monnier. Feb. [17] T. Signal Proc. [24] J. van de Beek. L. Nov. Ruiz. [23] J.. Lopez. Optimal …nite duration pulses for OFDM. Boucheret. Low-complex frame synchronization in OFDM systems.. 1997. COST 254. Tourtier. Vahlin and N. Toulouse. Martinez Arenas. Börjesson. M. P. 21 . M.-J. Moeneclaey. [18] L. and P. M. 5(5–6):379–403.602. (5. and P. 1980. Y. In Proc. July 1997. Speech. Scharf. Japan.835). Ödling. 43(2/3/4):191–193. Addison-Wesley. July 1997. IEEE Int. Isaksson. D. Feb/Mar/Apr 1995. Börjesson. J.-O. pages 2054–2058. van Bladel. BER sensitivity of OFDM systems to carrier frequency offset and Wiener phase noise. Sugita. M. 1991. [20] P. Pollet and M. In Proc. and P. France. 1993. 1995. Denver. and time series analysis. Jan. [22] J. Statistical signal processing: Detection. Synchronizability of OFDM signals. Conf. Signal Processing. ML estimation of timing and frequency offset in OFDM systems. [21] A. [16] T. Holte.. In Proc. Commun. and T. Ofdm synchronization demodulation unit. M. Sandell. 1996.-J. O. Conf. Landström. pages 982–986. Nov. Globecom. Signal Proc. O. IEEE Trans. volume 3.: Image Commun.. Multicarrier modem for digital HDTV terrestrial broadcasting. Int. 44(1):10–14. Seki. An all-digital symbol synchronizer in an OFDM modem for wireless comunication. Börjesson. and M. van de Beek.. [19] T. 45(7):1800–1805. J. and P. Universal Personal Commun. estimation. Dec. Acoust. 1995. Frequency domain data transmission using reduced computational complexity algorithms. Singapore. Submitted to. L. IEEE Trans.[15] A. pages 964–967. Ishikawa. Tokyo. In Proc. R.-J.. Moeneclaey.

Commun. [26] W.. 42(3):302–313. Technol. May 1997. Feb/Mar/Apr 1995. Vehic. Bad Aibling. Germany. Synchronization requirements for multi-user OFDM on satellite mobile and two-path Rayleigh-fading channels. 22 . Leung. IEEE Trans. 1993.. OFDM/FM frame synchronization for mobile radio data communication. Warner and C. 43(2/3/4):887–895. Description of Telias OFDM based proposal (working document in the OFDM concept group). Schlegel. Aug.[25] M. Wahlqvist et al. [27] L. ETSI STC SMG2 meeting no 22. Technical Report Tdoc 180/97. D. IEEE Trans. Wei and C.

23 .Transmitter Receiver n(k) x1 s(k) xk IDFT P/S y1 r(k) S/P h(k) xN DFT yk yN Figure 1: The OFDM system model.

2 0 -10 0 10 20 30 40 50 60 70 80 90 0. |2 zN r(k) |. 24 . | (.4 1.4 0.2 1 0. Rectangular pulse (dashed) and Tukey pulse (solid).8 0.2 0.)* θ argmax γ(.9 1 time average power 10 0 -10 -20 -30 -40 -50 0 0.3 0.) Sliding sum ε -1 2π Correlation part Figure 2: Structure of the estimator designed for an AWGN channel.6 0.4 0.1 0.5 0.6 0.Energy part Φ(. ) Sliding sum | .7 frequency Figure 3: Spectrum of pulse shaped OFDM signals. average power 1. |2 | .8 0.

) -1 2π ε Correlation part Figure 4: Structure of the estimator designed for the AWGN channel and pulse shaping. ) | . z-1 r(k) z-1 z-1 (.)* h0(k) h1(k) h2(k) hM(k) argmax θ Figure 5: Structure of the estimator designed for a dispersive channel. | (.Energy part h2(k) Φ(.)* h1(k) argmax θ γ(. |2 r(k) zN |. 25 .

Time offset estimator variance (squared sample) 10 2 10 1 10 0 10 -1 10 -2 0 5 10 15 20 SNR (dB) Figure 6: Variance of the time offset estimator in a system with pulse shaping. The estimator designed for pulse shaping with (dotted) and without (dashed) averaging over 10 symbols. The estimator designed without pulse shaping with (solid) and without (dash-dotted) averaging over 10 symbols. 26 . The estimator designed without pulse shaping with (solid) and without (dash-dotted) averaging over 10 symbols. The estimator designed for pulse shaping with (dotted) and without (dashed) averaging over 10 symbols. Frequency offset estimator variance (normalized) 10 -2 10 -3 10 -4 10 -5 10 -6 0 5 10 15 20 SNR (dB) Figure 7: Variance of the frequency offset estimator in a system with pulse shaping.

and the estimator designed with only the knowledge about the pulse shape (dashed). The estimator designed for pulse shaping with (dotted) and without (dashed) averaging over 10 symbols.Time offset estimator variance (squared sample) 10 2 10 1 10 0 10 -1 0 5 10 15 20 SNR (dB) Figure 8: Variance of the time offset estimator for a system with pulse shaping and channel dispersion. the estimator designed with only the knowledge about the channel dispersion (dash-dotted).The estimator designed for the AWGN channel with (solid) and without (dash-dotted) averaging over 10 symbols. 27 . The estimator designed for the AWGN channel (solid). Time offset estimator variance (squared sample) 10 2 10 1 10 0 10 -1 10 -2 0 5 10 15 20 SNR (dB) Figure 9: Variance of the time offset estimator in a system with pulse shaping and channel dispersion.

The estimator designed for the AWGN channel with (solid) and without (dash-dotted) averaging over 10 symbols. 28 . The estimator designed for pulse shaping with (dotted) and without (dashed) averaging over 10 symbols.Frequency offset estimator variance (normalized) 10 -2 10 -3 10 -4 10 -5 0 5 10 15 20 SNR (dB) Figure 10: Variance of the frequency offset estimator for a system with pulse shaping and channel dispersion.