IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008 1 Synchronization, Channel Estimation, and Equalization in MB-OFDM Systems Y. Li, Student Member, IEEE, H. Minn, Senior Member, IEEE, and R. M. A. P. Rajatheva, Senior Member, IEEE
Abstract— This paper addresses preamble-based low com- synchronization method, while UWB systems demand a low plexity synchronization, channel estimation and equalization complexity approach. for Zero-padded (ZP) MB-OFDM based UWB systems. The proposed synchronization method consists of sync detection, The MB-OFDM systems from [2] [3] have the following coarse timing estimation, fine timing estimation, and oscillator distinctive characteristics if compared to conventional OFDM frequency offset estimation. The distinctive features of MB- systems: 1) different channel responses and channel energies OFDM systems and the interplay between the timing and carrier across different bands, 2) different carrier frequency offsets frequency hopping at the receiver are judiciously incorporated across different bands, 3) the use of zero padding (ZP) instead in the proposed synchronization method. In order to apply the low complexity one-tap frequency-domain equalizer, the required of cyclic prefix (CP), 4) the interplay between the timing circular convolution property of the received signal is obtained and the frequency hopping (a mismatched timing point at the by means of an overlap-add method after the frequency offset receiver will yield a mismatched frequency hopping and hence compensation. The proposed low complexity channel estimator a significant performance degradation). These characteristics for each band is developed by first averaging the over-lap- provide diversity components as well as additional design con- added received preamble symbols within the same band and then applying time-domain least-squares method followed by the straints, and hence, should be taken into account in the designs discrete Fourier transform. We develop an MMSE equalizer and of synchronization, channel estimation, and equalization. its approximate version with low complexity. We also derive the There exist several preamble-based synchronization and probability density functions of the UWB channel path delays, channel estimation methods for regular OFDM systems (e.g., and using them we present several optimization criteria for our see [8]–[11] and references therein), but only a few counter- proposed synchronization, channel estimation, and equalization. The effectiveness of our proposed methods and optimization parts for MB-OFDM-based UWB systems [12]–[15]. These criteria are confirmed by computer simulation results. existing works have overlooked most of the distinctive features of MB-OFDM systems. The works in [13] and [14] applied Index Terms— BLUE, channel estimation, equalization, fre- quency offset, OFDM, synchronization, timing offset, UWB. the concept of utilizing the first significant channel tap for timing improvement, which was previously proposed in [10] [11] for regular OFDM systems, in MB-OFDM systems. [13] I. INTRODUCTION considered a ZP MB-OFDM system with the same frequency NE of the promising ultra-wide band (UWB) tech- offset and channel impulse response for all bands, while O nologies is multi-band orthogonal frequency division [14] considered a CP MB-OFDM system with independent multiplexing (MB-OFDM) [1] which has been proposed for channels across different bands and no frequency offset. the IEEE 802.15.3a and ECMA standard [2] [3]. OFDM is a The work in [12] developed best linear unbiased estimation relatively mature technology and has been adopted in digital (BLUE) based oscillator frequency offset (OFO) estimators broadcasting, wireless LAN and MAN standards. OFDM has for MB-OFDM systems by exploiting some of the MB-OFDM several advantages such as low complexity equalization in characteristics, but other synchronization, channel estimation dispersive channels and the spectral scalability/adaptability. and equalization tasks were not addressed and the interplay However, OFDM is more susceptible to nonlinear distortion between the timing and the frequency hopping was not consid- at the transmit power amplifier [4] and more sensitive to ered. An MMSE-based one-tap equalizer for ZP MB-OFDM synchronization errors [5] [6]. It is shown in [7] that the MB- systems was recently proposed in [15], but no optimization OFDM based UWB receivers are quite sensitive to carrier for ZP MB-OFDM systems was attempted, and the method is, frequency estimation errors. Hence, the application of OFDM as will be shown in this paper, just an approximate MMSE. technology in UWB systems requires a high performance The overlap-add (OLA) method was proposed in [16] for the Manuscript received June 5, 2007; revised September 23, 2007; accepted low complexity equalization in ZP OFDM systems, but the November 6, 2007. The associate editor coordinating the review of this paper effects of OLA window lengths and timing offsets were not and approving it for publication was E. Serpedin. This paper was presented considered. in parts at ICC 2007. Y. Li and H. Minn are with the Department of Electrical Engineering, In this paper, we propose low-complexity synchronization, University of Texas at Dallas, P.O. Box 830688, EC 33, Richardson, TX channel estimation and equalization methods and correspond- { } 75080 USA (e-mail: yxl044000, hlaing.minn @utdallas.edu). ing parameter optimizations for ZP MB-OFDM systems based R. M. A. P. Rajatheva is with the Asian Institute of Technology, PO Box 4, Klong Luang 12120, Pathumthani, Thailand (e-mail: [email protected]). on the preamble of [2] [3]. The main contributions of this The work of H. Minn was supported by the Erik Jonsson School Research paper are as follows. Excellence Initiative, the University of Texas at Dallas, USA. Digital Object Identifier 10.1109/T-WC.2008.070601 1) We derive the probability density functions (pdf) and the 1536-1276/08$25.00 c 2008 IEEE 2 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008
statistics of the channel path delays of the IEEE 802.15 UWB channel model from [17]. Note that these pdfs are also useful in evaluating or optimizing the energy capture of other UWB systems. 2) We present distinctive characteristics of the ZP MB- OFDM systems which are worthy to be taken into account into the system design but have been overlooked by the existing works. 3) We present how to optimize the parameter settings Fig. 1. The structure of the preamble pattern 1 from [2] and [3] (Also of the proposed synchronization, channel estimation, shown are the three preamble parts used in our synchronization method. and equalization methods by utilizing the distinctive The preamble symbols of carrier 2 and 3 within Part-a are not used in our characteristics of the ZP MB-OFDM systems and the synchronization method. channel path delay pdfs we derive. Note that existing works neither exploited all of these characteristics nor addressed all synchronization, channel estimation and 24 periods with an inverted polarity for the last period in equalization tasks together. each band. The channel estimation sequence is composed of 6 4) We derive an exact MMSE equalizer for ZP MB-OFDM identical periods of the frequency-domain channel estimation systems, and its low-complexity implementation. Our sequence. The structures of the preamble pattern 1 in each of MMSE equalizer reduces to the equalizer in [15] after the three different bands are shown in Fig. 1 (see [3] [12] for a high SNR approximation. other patterns). 5) Our investigation provides new insights for improving The MB-OFDM-based systems from [2] and [3] use ZP the receiver functionalities with low complexity. For guard intervals consisting of Npre prefix and Nsuf suffix zero example, it reveals different performance sensitivities samples, instead of the conventional CP guard interval. An to different parameters, and provides guidelines in set- OFDM symbol including the guard interval has a total of M = ting these parameters for robustness and complexity- N +Ng samples where Ng = Npre+Nsuf and, N is the number reduction. of the OFDM sub-carriers. Between ith and (i +1)th OFDM The rest of this paper is organized as follows. Section II symbols in [2] and [3], there are 37 guard samples consisting describes the signal and channel models. Section III introduces of Nsuf =32suffice samples and Npre =5prefix samples. the statistics of the UWB channels. The proposed synchro- The preamble sample index 0 corresponds to the first non-zero nization, channel estimation, and equalization methods are sample of the first preamble symbol. Denote the low-pass- presented in Section IV, and the parameter optimizations of equivalent time-domain samples (sampled at N/T =1/Ts, N the proposed methods are described in Section V. Simulation times the sub-carrier spacing) transmitted in the q-th frequency results and discussions are provided in Section VI, and the band by {sq(k)} and their corresponding frequency-domain paper is concluded in Section VII. symbols by {s˜q(n)}. During the m-th symbol interval, i.e., Notations: The superscripts T and H represent the transpose k ∈ [mM − Npre, (m +1)M − Npre − 1], {sq(k)} can be all and the Hermitian transpose, respectively. The kth row, nth zeros if no signal is transmitted in the qth band. X X · column element of a matrix is denoted by [ ]k,n. E[ ] The IEEE 802.15.3a UWB RF channel model described in represents a statistical expectation. The N × N unitary DFT [17] is given by F f f f F matrix is denoted by =[ 0, 1,..., N−1], while K is the sub-matrix of F consisting of the first K columns and Lh K F (:,i : k) represents [f , f +1,...,f ]. The corresponding i i k hRF(t)=X α δ(t − T − τ ) (1) frequency-domain variable of a time domain variable x is k,l l k,l l=0 k=0 denoted by x˜. {·} represents the real part of {·}.TheN ×N identity matrix is denoted by I and e represents the ith i where T , τ , and X are random variables representing column of I. l k,l the delay of the l-th cluster, the delay (relative to the l- th cluster arrival time) of the k-th multipath component of II. SIGNAL AND CHANNEL MODELS the l-th cluster, and the log-normal shadowing, respectively. We consider the MB-OFDM-based UWB systems from The channel coefficients are defined as a product of small- [2] [3] where the carrier frequency is hopped within a pre- scale and large-scale fading coefficients, i.e., αk,l = pk,lξlβk,l ± defined set of carrier frequencies {fq} (corresponding to where pk,l takes on equiprobable 1 to account for signal { } disjoint frequency bands) from symbol to symbol, according inversion due to reflections, and ξlβk,l are log-normal to a time-frequency code. The MB-OFDM system from [2] distributed path gains. Then we have E [αk,l(Tl,τk,l)] = 0 2 − Γ − | | Tl/ τk,l/γ and [3] has 4 different preamble patterns (for 4 pico-nets), and E αk,l(Tl,τk,l) =Ω0e e . Γ and γ are the each associated with a different time-frequency code. Each cluster decay factor and ray decay factor, respectively. Details preamble pattern is constructed by a synchronization sequence of the channel models are referred to [17]. portion followed by a channel estimation sequence portion. In this paper, we consider a low-pass-equivalent system that The synchronization sequence is composed by successively absorbs the carrier-frequency hopping into the channel impulse repeating a time-domain preamble sequence (symbol) over response (CIR). The sample-spaced low-pass-equivalent CIR LI et al.: SYNCHRONIZATION, CHANNEL ESTIMATION, AND EQUALIZATION IN MB-OFDM SYSTEMS 3 for the q-th band is given by Λ and λ are the cluster arrival rate and the ray arrival rate, respectively. The m-th moment of Zk,l is obtained as Lh K − 2 ( + ) m j πfq Tl τk,l − − − m d ΦZk,l (s) hq(n)=X αk,le p(nTs Tl τk,l t0) E[z ]= | =0 k,l dsm s l=0 k=0 (2) m i −1 m −i−1 m −i −m+i where the effect of the combined transmit and receive filter = (−l − n)(−Λ) (−k − u)(−λ) . =0 i =0 =0 with the impulse response p(t) whose span is [−t0,t0] has i n u (6) been included in the CIR, and the delay t0 is inserted for the causality. Applying the residue theory, we obtain the pdf of Zk,l as { q} { q − } ∞ Define tl = tl (i): i =0,...,N 1; l =1,...,Lq −sz q f (z )= Φ (s)e k,l ds where {t (i):i =0, 1, ..., N − 1} denotes the time-domain Zk,l k,l Zk,l l −∞ sample index set corresponding to the l-th non-zero preamble l k (l+k) (R1 + R2)Λ λ (−1) k =0or l =0 symbol period in the q-th band with N ≤ N ≤ N +Nsuf . L = q δ(z ) k =0&l =0 is the number of nonzero preamble symbols in the q-th band k,l (7) and depends on the preamble pattern and the band index q.Let { } where xq(k) denote the low-pass-equivalent time-domain channel ⎧ −1 i−1 ( + ) ⎪ l −Λzk,l l m=0 k m output signal samples corresponding to the q-th band. Then ⎪ (−1) e =0 ⎨⎪ i i!(l−i−1)! the corresponding low-pass-equivalent time-domain received (−k−i) (l−i−1) q (Λ − λ) z u(zk,l) k =0& l =0 samples {r¯ (t (i))} in the q-th band can be expressed as 1 R1 = k,l q l ⎪ (−1)l (l−1) −Λ ⎪ z e zk,l u(z ) k =0&l =0 ⎩⎪ (l−1)! k,l k,l q 2 q ( ) q q j πυq tl i /N 0 l =0 r¯q(tl (i)) = e xq(tl (i)) + n(tl (i)) (3) ⎧ −1 i−1 ( + ) q ⎪ k −λzk,l k m=0 l m where {n(t (i))} are zero-mean independent and identically- ⎪ (−1) e =0 l ⎨⎪ i i!(k−i−1)! distributed circularly-symmetric complex Gaussian noise sam- (−l−i) (k−i−1) (Λ − λ) z u(zk,l) k =0& l =0 2 q 2 R2 = k,l ples with variance σ = E{|n(t (i))| }, and υq is the ⎪ (−1)k (k−1) − l ⎪ z e λzk,l u(z ) k =0& l =0 normalized (by the subcarrier spacing) carrier frequency offset ⎩⎪ (l−1)! k,l k,l (CFO)oftheq-th band which is related to the normalized OFO 0 k =0, υ by with u(·) being the unit step function. Since {Tl} and {τk,l} are independent, we have υq = bq υ, for q =1, 2, 3. (4)
fTl,τk,l (x, y)=ftl (x)fτk,l (y) (8) For the frequency synthesizer in [1], we have [b1,b2,b3]= 13 15 17 where fTl (x)=fZ0,l (x) and fτk,l (y)=fZk,0 (y). 16 , 16 , 16 (c.f. [12]). Since the receiver does not know Next, we find the covariance matrix of the CIR for each in advance when the preamble will arrive, we denote the band. Let L denote the maximum length of the sample-spaced low-pass-equivalent received signal as {rq(k)} which contains CIR vector of each band and hq denote the L × 1 CIR vector noise-only samples followed by the received preamble samples of the qth band. Depending on the actual length of each {r¯q(m)}. The number of the initial noise-only samples is a channel realization, hq may contain some zero samples. The random variable depending on the arrival time of the preamble covariance matrices of hq are the same for all bands as will and the receiver’s sync detection window timing. be apparent in the following. Hence, we simply denote the covariance matrix of hq by Ch. From (2), we can calculate 2 the element of Ch as III. STATISTICS OF THE UWB CHANNEL MODELS ∞ ∞ Lh K 2 ∆ Ch(m, n)= E |αk,l(Tl,τk,l)| We first derive the pdf of the delay Zk,l = Tl + τk,l for the 0 0 =0 =0 k-th ray of the l-th cluster of the UWB channel, from which l k × p(mTs − Tl − τk,l − t0)p(nTs − Tl − τk,l − t0) the pdfs of Tl and τk,l are obtained. × f (T ,τ ) dT dτ (9) Note that T0 =0, τ0,l =0; and Zk,l can be expressed as Tl,τk,l l k,l l k,l the sum of inter-arrival times of the first l clusters and the first where we have used the statistical independence among k rays of the lth cluster. Since inter-arrival times of the UWB {αk,l(Tl,τk,l)}. 2 channel paths are independent exponential random variables, Substituting (7)-(8) and E |αk,l(Tl,τk,l)| = − −T /Γ −τ /γ Zk,l is the sum of (l +k 2) independent exponential random Ω0e l e k,l into (9) gives variables. The moment generating function of Zk,l is given by Lh K T T −Tl/Γ −τk,l/γ Ch(m, n)= Ω0e e Λ l λ k 0 0 Φ (s)= , ∀k, l except k = l =0. l=0 k=0 Zk,l − − Λ s λ s × p(mTs − Tl − τk,l − t0) p(nTs − Tl − τk,l − t0) (5) × fZ0,l (Tl)fZk,0 (τk,l) dTl dτk,l (10)
1An arbitrary carrier phase offset can be absorbed into the CIR and hence 2We neglect the shadowing parameter X in this paper and hence realiza- it is not included explicitly in the signal model. tions of {αk,l} are not normalized. 4 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 11, NOVEMBER 2008 which can be numerically evaluated. We will use the above A. Sync Detection and Coarse Timing Estimation channel covariance matrix in optimizing our proposed methods The carrier frequency of the receiver is initially set to in Section V. Note that the pdfs of the UWB channel path the one that corresponds to the first preamble symbol, and delays we derive are also useful in calculating or optimizing denote the corresponding band index by q¯. During the sync the energy capture of other UWB systems. detection process, the receiver carrier frequency hopping is not performed since it requires timing information which is IV. PROPOSED SYNCHRONIZATION,CHANNEL unavailable yet. We use the first two preamble symbols in the ESTIMATION AND EQUALIZATION q¯th band to detect the arrival of the preamble. Then the sync In this paper, we consider sync detection, timing offset esti- detection metric is given by mation, OFO estimation, channel estimation and equalization M(k)=|R ¯(1,k,N )| . (13) for the packet-based MB-OFDM system in [2] and [3]. We a,q assume that the receiver knows what preamble the transmitter For complexity reduction4 and energy saving, the computation uses but not the OFO and the arrival time of the preamble. of M(k) is performed only at k =0, ∆, 2∆, ..., where ∆ The whole interval of the 30 preamble symbols with the (an integer) is a design parameter. { q } M sample index set tl (i) is divided into three disjoint parts Once (k) becomes larger than a pre-defined threshold – part-a, part-b, and part-c – with the corresponding sample MSD, say at k = kSD, the receiver decides that the preamble { q } { q } { q } index sets ta,l(i) , tb,l(i) , and tc,l(i) , respectively. All has been detected and starts finding the coarse timing point the distances of the preamble symbol pairs within part-x are as follows: denoted by {dx(m):m =1, 2,...} where x is a, b,or kct = arg max {M(k)} (14) c representing the corresponding part. For example, for the kSD≤k≤kSD+W1−1 preamble pattern 1, we may set the first part as the first 6 where W1 is the length of the window within which the coarse symbol interval, the second part as the next 18 symbol interval, timing point is searched and W1 can be set to M.Alower and the third part as the last 6 symbol interval (see Fig. 1). In { q } { q } { q } { q complexity version of the above coarse timing estimator can this case, ta,l(i) = tl (i):l =1, 2 , tb,l(i) = tl (i): q q be implemented by finding the point k1 corresponding to the l =3, 4, 5, 6, 7, 8}, {t (i)} = {t (i):l =9, 10}, {db(m)} c,l l maximum of the metric values at k = kSD, kSD +∆1, kSD + = {3M,6M,9M,12M,15M}. We also define dx(0) = 0 for 2∆1, ... , and then the point corresponding to the maximum convenience. of the metric values at k1 −∆1/2, k1 −∆1/2 +1, ..., Define the correlation term with a correlation distance k1 + ∆1/2. Here, ∆1 (an integer) is a design parameter. dx(m) between non-zero received preamble symbols within Based on the coarse timing point kct, the receiver determines part-x in the q-th band as at what time instant it should hop the carrier frequency to R (m, kt,N ) demodulate the rest of the preamble. x,q − ∗ = ( 1) rq (k)rq(k + dx(m)) (11) B. Fine Timing Offset Estimation {k, k + dx(m)} ∈{ q } The fine timing estimation is based on the preamble part-b kt + tx,l(i) while the coarse timing estimation is based on part-a. Hence, where the correct points for the coarse timing point kct and the fine { }∈{ q } timing point kft correspond to the beginning of part-a and part- 1, if k + dx(m) kt + tb,8(i) = (12) b, respectively. We denote the beginning of part-b by kref .The 0, else proposed fine timing estimator is described by 5 note that i =0, 1,...,N − 1. is a parameter introduced 3 by the burst cover sequence (see [2] [3] for details). During kft = −η + arg max |Rb,q(1,i,N )| ∈( + − + + ) part-a where the sync detection and coarse timing estimation i kct Da W2,kct Da W3 q=1 are performed, the receiver carrier frequency hopping is not (15) performed. During the remaining parts, the receiver carrier where Da is the length (in samples) of part-a and η is a fixed frequency is hopped according to the preamble pattern. The timing adjustment. For CP-based OFDM systems, η should be fine timing and the OFO estimations are based on the received set such that the timing point is most of the time within the preamble symbols within part-b, while channel estimation task ISI-free part of the CP [10]. We will describe how to set η is based on part-c 3. The reason for the use of only part- for the considered ZP OFDM system in Section VI. b for the fine timing estimation is to allow processing time so that timing adjustment (for frequency hopping instants) C. OFO Estimation and Compensation based on the fine timing estimation can be made at the end We adopt the BLUE-based OFO estimator from [12] with of the preamble. The receiver tasks are performed in the a few differences. In [12] the OFO estimator utilizes the first following order: sync detection, coarse timing estimation, fine 21 preamble symbols, while in this paper the OFO estimator timing estimation, OFO estimation and compensation, channel estimation, equalization and data detection. 4M2(k) can also be used in place of M(k). 5Note that although more correlation terms can be included in the fine tim- 3Note that preamble symbols from part-b can also be used to improve the ing estimation, for low complexity implementation we just use the correlation channel estimation. terms with the correlation distance db(1). LI et al.: SYNCHRONIZATION, CHANNEL ESTIMATION, AND EQUALIZATION IN MB-OFDM SYSTEMS 5 is based only on part-b allowing processing time and resource Next, CFO compensation for the qth band is performed for prior tasks of sync detection and coarse timing estimation. starting from preamble part-b as Furthermore, the correlation window length for each preamble ∆ −j2πmυb q /N symbol is fixed at N samples in [12] but is set to N in this rq(kft + m)e = r˘q(kft + m)form =0, 1,.... paper which will be optimized for better performance. (26) The OFO estimator is given by
3 1T C−11 D. Channel Estimation q=1 θ bqvq υ = q (16) 3 T −1 2 The use of zero-amplitude guard samples in MB-OFDM =1 1 C 1 b q θq q UWB systems [2] [3] instead of conventional CP guard Hq samples saves transmit power and allows a sufficient amount where υq = ωq(m)θq(m) (17) of time for the transmitter and receiver to switch between m=1 carrier frequencies. To obtain the circular convolution prop- N erty required for the frequency-domain one-tap equalization, θq(m)= angle{Rb,q(m, kft,N )} (18) 2πdb(m) we apply the OLA method [16] to the frequency-offset- compensated received signal. Specifically, for every l-th non- with ωq(m) being the m-th component of the weighting vector zero OFDM symbol in the qth band after CFO compensation, −1 −1 −1 ≤ ≤ { ω = C 1 1T C 1 . (19) the OLA method adds M0(0 M0 Nsuf ) samples r˘q(kft+ q θq θq lM +N), r˘q(kft +lM +N +1),...,r˘q(kft +lM +M0 +N −1)} C { { − } In (19), θq is the covariance matrix of θq(m):m = to r˘q(kft +lM), r˘q(kft +lM +1),...,r˘q(kft +lM +M0 1) , } 1 (l) 1,...,Hq , Hq is a design parameter (see [12]) and denotes resulting in an N × 1 vector yq with the k-th element given the all-one column vector, whose size should be apparent from by the expression. Two approaches (Methods A and B) for the ⎧ ⎨⎪r˘q(kft + lM + k) calculation of {Cθ } are given in [12]. Due to the different q (l) window length in this paper, the covariance matrices {C } y (k)= +r˘q(kft + lM + k + N), if 0 ≤ k ≤ M0 − 1 θq q ⎩⎪ become slightly different from those in [12]. For example, r˘q(kft + lM + k), else. for preamble pattern 1 or 2, denoting the number of non-zero (27) preamble symbols within the combined part-b in the qth band The OLA length M0 need not be the same for preamble sym- ¯ by Lq, and applying Method A, we obtain bol and data symbol. We use Mp and Md to denote the OLA length for preamble symbol and data symbol, respectively. C (m, n)=g(m, n) θ⎧q After applying the OLA method, the preamble symbols in ( ¯ − ) 2 ⎪ Lq m N σ ¯ ⎪ m + 2 ( ) , if m = n & m