You are on page 1of 4

SYMBOL TIMING RECOVERY ISSUES IN MULTICARRIER MODULATION Fred Daneshgaran ECE Dept., Calif. State Univ.

, Los Angeles, (USA) Email: fdanesh@calstatela.edu


ABSTRACT The timing sensitivity of the MultiCarrier Modulation (MCM) schemes such as OFDM plays an important role in the design and application of such systems for digital communications. The timing sensitivity is particularly important in MCM since timing offsets can cause large Adjacent Channel Interference (ACI) due to the spectral overlap among the adjacent frequency channels. We address several issues in connection with the design of clock recovery loops for MCM based on ML criterion exploiting the timing information available from the multiple channels. We present the results of the study of both the theoretical and practical implementation aspects of the promising techniques. INTRODUCTION MultiCarrier Modulation (MCM) is an attractive technique for data transmission over bandlimited slowly fading channels, particularly useful for audio [1] or video broadcasting and Digital Subscriber Line (DSL) over copper wires. MCM has the potential of having a performance approaching the channel capacity via proper channel coding and shaping the power spectrum of the transmitted signal to compensate for the linear ltering effects of the channel. There are now several different avors of the MCM technique among which we may list the Discrete Fourier Transform (DFT) based Orthogonal Frequency Division Multiplexing (OFDM) and Discrete MultiTone (DMT) modulation, and Wavelet based OFDM, denoted WOFDM. In the DFT based OFDM, the orthogonal subchannels are essentially obtained by modulating a rectangular window function of duration (the OFDM symbol duration) by sinusoids such that the resulting modulated signals are orthogonal. . All other MCM techniques to be an integer multiple of This is achieved by setting the subchannel carrier spacing may be viewed under a common umbrella by noting that the basic technique for generating orthogonal frequency channels with overlapping spectra (e.g., WOFDM systems) is that of starting with a Nyquist pulse and splitting the spectrum of this pulse into overlapping orthogonal frequency subchannels. The splitting sequences, which are in effect the discrete counterpart of the continuous Nyquist pulses, can be either of the discrete wavelet variety or generated using paraunitary lterbank structures. The key distinction between DFT based OFDM scheme and other MCM schemes in this light is that the former uses a time limited rectangular pulse that is orthogonal to its translated versions provided the shift period is an integer multiple of , while the latter uses a Nyquist pulse where the elementary pulse is allowed to spread over several symbol periods. In order to encompass as broad a variety of MCM schemes as possible, our focus here shall be on the class of MCM schemes obtained using lterbanks. We note however that the popular OFDM scheme can indeed be implemented using a lterbank structure, although the FFT based implementation is more efcient when the number of subchannels is large ). (e.g., The timing sensitivity in MCM is particularly important in multichannel modulation since timing offsets can cause large Adjacent Channel Interference (ACI) due to the spectral overlap among the adjacent channels. In practice, often training sequences are used in order to obtain estimates of the channel parameters and for carrier and clock synchronization that are needed for signal processing at the receiver. Hence, for all practical purposes we may assume that the receiver has acquired the carrier and focus on timing recovery issues during tracking. In this paper, we focus on ML symbol synchronizers which derive the timing information from all the subchannels without the need for any pilot symbols [2, 3, 4].

FREQUENCY SUBCHANNEL GENERATION IN MCM VIA CHANNEL SPLITTING In applications, the synthesis lterbank is used at the transmitter while the analysis lterbank is used at the receiver. The signal processing functions can be partitioned into two distinct parts, the rst is the digital signal processing functions, while the second is the analog interface to the channel represented as a multiplication by an impulse train followed by going through the shaping lter at the transmitter, and matched ltering followed by sampling at the receiver. We may generically denote these operations as digital to analog and analog to digital conversion. However, there is a subtle difference between how an ADC and DAC normally operate versus what is meant when a shaping pulse and its matched lter is used in the transmission chain. In practice, typical DAC at the transmitter effectively uses a rectangular pulse as a shaping lter followed by post ltering via an analog low-pass lter to limit the signal bandwidth. Similarly, a typical ADC is essentially a cascade of a sample and hold circuit followed by quantization. While the rectangular shaping pulse of duration may indeed be used as an elementary shaping pulse, the post ltering process needed to limit the signal bandwidth may result in a system where the zero ISI criterion is not satised. In order to ensure the satisfaction of the zero ISI condition and indeed to control the spectral properties of the transmitted signal, the shaping lter itself is often implemented as a Finite Impulse Response (FIR) lter followed by digital to analog conversion and analog post ltering at the transmitter, and the inverse signal processing at the receiver. In such a case, while the input to the shaping lter is at the symbol rate, its output is at a much higher rate depending on the number of samples used per symbol duration for the implementation of the FIR shaping lter (i.e., the DAC and ADC operate at a much higher rate than the symbol rate at the transmitter and receiver respectively). We shall refer to the shift orthogonal sequences used to split the spectrum of our functions as splitting sequences. We recall that a set of splitting sequences with periods satisfy the following orthogonality conditions: . if and As noted above, such sequences can be obtained in the design of QMF lterbanks.

ML TIMING ESTIMATOR AND ITS JITTER VARIANCE We shall make the following assumptions for the analysis that follows: 1) the spectrum of the original shaping pulse is split into orthogonal subchannels; 2) PAM modulation is used in each subchannel whereby the data sequences are equiprobable iid and independent for different subchannels; 3) the data sequence associated with -th subchannel is denoted by ; 4) the effect of the transmission channel in a given subband is an amplitude with attenuation denoted by and a group delay both of which are assumed to be estimated accurately via a training sequence, so that we may consider them to be known quantities at the receiver; and 5) the difference in the group delay for adjacent subchannels is assumed to be small enough that the adjacent subchannels essentially maintain their relative orthogonality. (inclusion Under the above assumptions, the received signal is of in the noise term is for analytical convenience) where, is the -th Nyquist pulse (with orthogonality period ) obtained as a result of splitting the spectrum of with shift orthogonality period , and is the timing parameter to be estimated. Note that here we have assumed that all the subchannels have equal rates. Projection of on the locally generated basis yields , where as usual, denotes the correlation between the noted functions, and are iid Gaussian random variables . Let us dene . The sufcient statistic for observation over -symbols per subchannel (i.e., symbols in total) is the set of values . The approximate decision directed ML estimate of is obtained from

(1)

is the -th detected symbol of the -th subchannel. Assuming that the receiver is operating at low bit error where rates, we may use instead of for analytical evaluation of the jitter variance. Note that we may absorb the factor . in by dening so that When the timing error is small, we can use the Taylor series expansion of the derivative of the likelihood function about for evaluation of the variance of the timing error. The parameter can be solved for by nding the using only the rst two terms in the Taylor series expansion of this function. Using this method, and zero of making the approximation that at moderate to high SNR levels (here denotes . expectation), the variance of the timing jitter in the tracking mode can be expressed as

Replacing

in the likelihood function and collecting terms we have since the cross product term . Now, since . It can be veried that has zero expectation and

(2)

(3)

The derivation of

is straight forward but tedious. We omit the details for brevity:




(4)

Using the following equality



(5)

the above expression simplies to

(6)

Putting the pieces together we have

(7) (8)

Clearly, the energy distribution and the channel gain in each subband has an impact on the variance of the timing jitter and indeed the jitter variance may be optimized with respect to the energy distribution in the subchannels. However, many times the energy distribution within the subbands is adjusted by the desire to maximize the data throughput of the overall scheme, and not based on considerations of the effect of this distribution on the timing jitter. It is possible to show that if the shaping pulses are obtained by ltering the original pulse with a power complementary lterbank, then equation (7) coincides with the timing jitter variance of a single channel modulation using the shaping pulse and observed symbols.

IMPLEMENTATION OF THE SYMBOL SYNCHRONIZER

The implementation of the approximate DD ML M-channel synchronizer can be considered an extension of the symbol synchronizer presented in [5] to multichannel modulation, whereby multiple arms are used in the synchronizer [2, 3, 4]. Aside from the low frequency channel, the use of higher frequency subchannel can introduce false lock points in the S-curve of the synchronizer. The following Staged Synchronization Technique (SST) can be used to avoid false locks: 1) for a number of symbol intervals say , only the lowpass channel data are transmitted, and only the lowpass channel branch is used to reconstruct the recovered clock signal at the receiver. During this phase we are using the lowpass channel S-curve, which has no false lock points; 2) after the initial symbol periods, the transmitter goes into normal

0.1

0.01

0.001

0.0001

1e-05

1e-06

1e-07

1e-08 -5

10

15

20 EBN0

25

30

35

40

_ Figure 1: The timing jitter variance of the proposed symbol synchronizer obtained via simulation as a function of SNR. The upper dashed line is obtained using the lowpass channel timing information only, the lower dashed line is obtained using both channels simultaneously for timing estimation, and the solid line is obtained using the root raised-cosine shaping pulse with roll-off 1, used at twice the rate.

operation mode and transmits data on all subchannels. Since the sampling instant is already close to the correct zerocrossing point of the overall S-curve, the timing information provided by the other highpass channels will increase the synchronizer sensitivity without generating false locks. The increased complexity of the synchronizer training phase is justied by the gain that can be obtained in the sensitivity of the overall S-curve. To verify the analytical results for general shift orthogonal Nyquist pulses with overlapping spectra, let be a root raised-cosine pulse with roll-off 1 spanning the lowpass channel, and be its sine modulated version, spanning the highpass channel. Simulation results of SST depicting the variances of the timing jitters obtained in the tracking mode are depicted in Fig. 1. Note that there is nearly two orders of magnitude reduction in the variance of the timing jitter for a wide range of SNR values, when the timing information for both channels are used.

References
[1] B. LeFloch, R. Halbert-Lassalle, and D. Castelain, Digital Sound Broadcasting to Mobile Receivers, IEEE Trans. Consum. Elec., Vol.35, No.3, pp.493-503, Aug. 1989. [2] F. Daneshgaran and M. Mondin, Symbol Synchronization Using Wavelets, Proc. IEEE MILCOM95, pp.891-895, San Diego, CA, Nov. 1995. [3] F. Daneshgaran, M. Mondin, and F. Dovis, Symbol Synchronization for Multichannel Modulation, Proc. Conf. on Inform. Sci. and Systems, CISS98, pp.465-470, Princeton, NJ, March 1998. [4] F. Daneshgaran, M. Mondin, and F. Dovis, ML Symbol Synchronization for Multichannel Modulation: Analysis and Implementation, Proc. Conf. on Inform. Sci. and Systems, CISS99, pp.850-854, The John Hopkins Univ., Baltimore, MD, March 1999. [5] M. Moeneclay, Two Maximum-Likelihood Symbol Synchronizers with Superior Tracking Performance, IEEE Trans. on Comm., Vol. Com-32, No.11, Nov. 1984

You might also like