TWO-DIMENSIONAL PILOT-SYMBOL-AIDED CHANNEL ESTIMATION BY WIENER FILTERING

Peter Hoeher, Stefan Kaiser, and Patrick Robertson Institute for Communications Technology German Aerospace Research Establishment (DLR) P.O. Box 1116, D-82230 Oberpfaffenhofen, Germany E-mail: First name. Lastname@dlr.de

ABSTRACT
The potentials of pilot-symbol-aided channel estimation in

two dimensions are explored. In order to procure this goal, the discrete shift-variant 2-D Wiener filter is derived and analyzed given an arbitrary sampling grid, an arbitrary (but possibly optimized) selection of observations, and the possibility of model mismatch. Filtering in two dimensions is revealed to outperform filtering in just one dimension with respect to overhead and mean-square error performance. However, two THE 2-D WIENER FILTER cascaded orthogonal l - D filters are simpler to implement and 2 shown to be virtually as good as true 2-D filters. The formulation of the 2-D discrete minimum meansquare error (MMSE) estimation problem and its solution is as follows [3]. Assume it is of interest t o esti1 INTRODUCTION mate a wide-sense stationary (WSS) 2-D stochastic proCCORDING to the sampling theorem, any band- cess h ( k , I) @ disturbed by a WSS additive discrete limited l-D stochastic process or deterministic sig- noise process n(k,I ) E @,where 0 5 IC 5 l - 1 and i nal is uniquely representable by samples taken at least 0 5 1 5 L - 1 are the two indices of the process (e.g., at the Nyquist rate. If the process is observed in noise, time/frequency indices for the example of channel estihowever, a perfect reconstruction is not possible. The mation) and 1f.L is the arbitrary (finite or infinite) block Wiener filter is the optimum (minimum mean-squared size. Hence, the received bisequence is error) linear filter/smoother/predictor, if the noise is additive. The sampling theorem also holds for multi- r ( k , I ) = h ( k , I ) + n ( k , I ) , V O ~ k ~ I i - 1 ,_ < l _ < L-1. O dimensional processes and signals [l]. Generalizations of (1) the Wiener filter for multiple dimensions exist as well, Assume further that r(k, I) a is observed only at samboth for continuous and discrete signals [2, 31. Applica- ple locations (pilot locations) P. The corresponding intions considered so far include image processing, sensor- dices are denoted as k and I t , respectively, where again array processing, and geophysical applications. 0 5 k 5 K - 1 and 0 5 I 5 L - 1. The number of observations (later: pilot symbols) is Ngrid llPll K . L . = Pilot-symbol-aided channel estimation of time- and/or The estimator is chosen to be discrete, linear, and genfrequency-selective channels is another potential applicaerally shift-variant: tion. The basic principle of l-D pilot-symbol-aided channel estimation is to multiplex pilot (training) symbols k(k, I) = w ( k ,I ; IC, I) ?.(IC, I t ) , (2) known to the receiver into the data stream [4, 51. Hence, {k,l)EP the receiver is able to estimate the process (the channel) at any time given the observations at the pilot loca- where &IC,/) E @ is the estimated bisequence and tions, assuming the pilot density is sufficient w.r.t. the w ( k ,I; k,1) E @ is the shift-variant impulse response of channel bandwidth. l-D pilot-symbol-aided channel es- the estimator or filter. The optimal filter order, i.e. the timation is well understood; different filters have been number of filter coefficients, is Ntap= N g r i d ; a simplifiinvestigated [4,5 , 6, 71. cation is introduced in Section 5. An example is shown When the channel is probed simultaneously in both time in Fig. 1 illustrating the three special cases of predicand frequency domains, as proposed by the first author, tion, filtering and smoothing w.r.t. the actual locationls) the overhead of pilot symbols can be reduced signifi- {k,I } given a sampling grid P . cantly [8]. This scheme is particularly attractive for The mean square error is defined as multi-carrier modulation such as Orthogonal FrequencyDivision Multiplexing (OFDM) [8, 93. Applications in-

clude the upcoming European Terrestrial Digital Video Broadcasting and Digital Audio Broadcastingstandards, mobile radio, and line modems. However so far, to our best knowledge, only filtering with two cascaded orthogonal l-D filters was studied for this particular application, referred to as 2~ l-D filtering. In this paper, we explore the potentials of pilot-symbol-aided channel estimation in two dimensions.

<

0-8186-7919-0/97 $10.00 8 1997 IEEE

1845

PERFORMANCE ANALYSIS
@, I ) = gT(k,I) E.

For a given set of filter coefficients t ,(2) can be written u in matrix form as
0 data-symbol

pilot-symbol
1 - , . 0 0 0 0 0 0 ~ 0

(10)

000.0CC00 000000000
0000 000000000 0 0000000.000.0000 0 000000000000000

K = 16 L = 13
k = 12 1=6
NtW=4
N@,j= 11

Substitution into (3) yields the MSE [3]

J(g(k,l)) = e; -gT(IC,I)g*(IC,r) -gT(k,1)8*(IC,1)

+wT(k,q9w*(k,I); . ) E R, J(

(11)

t
t

4
k

K-1

where OH is the hermitian operator (conjugate transpose). This equation is valid for any FIR filter g ,i.e., also for model mismatch. If the cross- and autc-covariance functions are known, the MMSE is obtained by inserting (9) into (11):

Distances needed for computation of:

- auto-covariance
--cross-covariance

Figure 1: Random sampling grid illustrating prediction, filtering and smoothing, and illustrating the computation of cross- and auto-covariances.

J,(k, 1) F U; - p ( k , I)2-lg*(k1 I).

(12)

GENERALIZED TWO-DIM.

where E[.] denotes the expectation and g is the Ntap SAMPLING THEOREM by 1 filter coefficient vector. The optimal filter in the sense of minimizing (3), i.e. the MMSE estimator or 2- It is interesting to recall that the multi-dimensional D Wiener filter, is obtained by applying the orthogonal Wiener theory is a generalization of the sampling theorem, its pointed out in [7] for the l-D case. The projection theorem [3]: multi-dimensional Wiener filter is the optimal predictor/filter/smoother in the presence of additive noise, for E[(h(k,I ) - k ( k , I ) ) . T*(kll l)] = 0 finite or infinite (bi)sequences when given an arbitrary V 0 5 k 5 K - 1, 0 5 I 5 L - 1, (4) sampling grid P. It provides a compact tool to analyze where ()* denotes the complex conjugate. Although (4) the MSE including model mismatch and aliasing effects. holds for all {k, l}, in Section 6.2 {k, I} are from the pilot grid P. We assume that the Wiener filter Equations (9) and (12) (respectively and (11))manifest is physically realizable, i.e., the coefficients w ( k ,1; k,1) the generalized 2 - 0 sampling theorem. The conventional exist. The optimum filter coefficients are denoted as 2-D sampling theorem is given by the special case where 1)) w,(k, I ; IC, I). After basic operations, the 2-D discrete J(w(IC, = 0 , V {k,I}. This implies that the bisequence h ( k , I) is bandlimited in both dimensions, sampling is Wiener-Hopf equation is obtained: sufficient in both dimensions (according to the known bandwidths of h ) , and the additive noise process n ( k , I ) E [ h ( k ,I) f * ( I C , 11 = 9 w o ( k ,I ; k/, 1) is zero.
{k,l)EP

.E[r(k, ?-*(IC, I)], r)

Let

V { k , P} E P. (5)

SIMPLIFICATIONS

The 2-D Wiener filter exists iff 9is invertible. Coefficient (6) sets can be pre-computed and stored. So far, the number of coefficients per set, Ntap, assumed to be equal was to the number ~f observations, Ng.-;d. The complexity @(k - k, 1 - 1) = E[r(k,/ I ) r * ( k ,I)] (7) can be significantly reduced, however, if only a subset denote the auto-covariance. Inserting (6) and (7) into 7(k,I) E P of Nlap < N g r i d observations is chosen, where Ntap = I IT(k, I)11. The performance/complexity (5) yields trade-off is adjustable. Suitable selection criteria are: OT(k,I) = &k,I)$), (8) 1.) Optimize 7(k,l) for each location { I C , ) } so that where 2 is an Ntap by Ntap auto-covariance matrix and J ( t u ( k , l ) ) is minimized. This optimum rule is rather possibilities e(k,I) is an Ntapby 1 cross-covariance vector and ()T complex for practical block sizes: denotes the transpose. Hence the optimal solution (if must be checked. 2.) Search the Ntapsample locations existent) is {k,1) nearest to the actual location {k,I } . Nearest wT(k, = p ( k , I-). I) (9)

6 ( k - k, 1 - I) = E[h(IC, T * ( I C , l)] I) denote the cross-covariance and

(2:;)

1846

could be w.r.t. A.) the sum of horizontal and vertical 6.2 Pilot-Aided Channel Estimation I~~ distances, ~ k - k ~ + ~ I -(suitable if the normalized chan- Let us denote the spacing between pilot symbols in the nel bandwidths are similar in both dimensions; using the time domain by N K and in the frequency domain by N L . Euclidean distance - P ) Z (I - 11)2 appeared to be Given the normalized channel bandwidths fD,,,T, and worse), or B.) some form of weighted distances, such r,,,AF, the sampling theorem states that as lk - kl fD,,,T, )I - 11r,,,AF for the example in Section 6.2. . and r m a x A FN L 5 1/2. (17) fD,,,T, * N K 5 1/2

d(k

We define a balanced design as fDma,TS . N K % r,,,AF. N L . A rule of thumb is 2 x oversampling 6.1 Channel Model to achieve reasonable complexity (filter length) and perA wide-sense stationary uncorrelated scattering (WSS- formance, i.e., US) mobile radio channel model can be expressed as fD,,,Ts . N K % 1/4 and r,,,AF. N L M 1/4. (18)

EXAMPLE

Example: Consider fo,,, = 320 HZ (e.g., v = 192 km/h 62 fo = 1.8 GHz) and rmaX lops (one-sided, 2 = (13) i.e., the maximum delay is 20,us) as worst-case paramewith E[lh(k,I)12]= U: = 1 [lo]. N echoes superpose ters of a typical land mobile radio system. . incoherently. Each path is characterized by a random phase &, a random Doppler shift fD, and a random de- Given these constraints, a balanced design with 2x overlay r,, where l n N . The corresponding joint prob- sampling applied to an OFDM system is as follows: ability density function p 4 , f D , T ( df, ~r ) is needed to ran, fD,,,T$NK = 320 HZ . 2 6 0 ~3~ . 1/4 = domly choose &, fD, and r,. The symbol duration, T, , T,,+AFNL = l o p s . 4.1667 kHz . 6 = 1/4 (19) is the spacing in the time domain and the carrier spacing, AF, is the spacing in the frequency domain. fD,,,Ts , and rmazAFare the normalized one-sided channel band- where N K = 3, N L = 6, T = T - A = 260 ps-20 ,us = 240 ,us is the useful symbol duration, A = 2r,,, = widths in time and frequency domains, respectively. 20 ,us is the guard interval, and 1/T = AF = 4.1667 kHz Due to the assumption of uncorrelated scattering and for is the carrier spacing. independent phases This scheme was investigated for short blocks with A = P 4 , f D , 7 ( 4 , fD, = p $ ( 4 ) p f D ( f D ) p T ( r ) , (14) 58 and L = 115. A rectangular sampling grid with N K = 3 and N L = 6 consists of N g r i d = 400 pilot symexcluding where p f D ( f D ) and p 7 ( r ) are proportional to the so bols; the overhead (pilot density) is ~ / ( N K N L ) edge effects. Other grids were examined as well. Perforcalled Doppler power spectrum and delay power specmance criterion is the MSE averaged over all symbols; trum, respectively [lo]. edge effects are hence included. Perfect knowledge of Example: Consider a uniform Doppler power spec- the covariances is assumed to avoid any spurious effects. trum (with fo,,, being the one-sided maximum Doppler Analytical results were verified by simulations. frequency) and a uniform delay power spectrum (with l-D channel estimation serves as a benchmark. We rmax being the one-sided maximum echo delay). Assume looked at filtering in the frequency domain; the overthe received signal r ( k ,I ) is observed in white Gaussian head is ~ / N L excluding edge effects. The results are noise with average SNR = E , / N o , where E, is the averreported in Fig. 2 given various numbers of coefficients, age energy per symbol (data and pilot symbols are here Ntop. About five filter coefficients are sufficient. assumed to be transmitted with the same energy) and NO is the one-sided noise spectral density. Then, the The second strategy under investigation is 2 x l-D chancross-covariance is nel estimation [SI: Filtering in the frequency domain is followed by filtering in the time domain, both using 1e A t , A f ( k- P I , I - P) = D Wiener filters. The actual order of filtering (time or frequency first) is arbitrary due to linearity. The coreAt(6 - k) = responding results given a rectangular grid are plotted - VI) = in Fig. 3. The MSE is better than for l-D filtering, although the total pilot overhead is reduced by a factor of and the auto-covariance is three.

< <

Finally, 2-D channel estimation given the rectangular grid is investigated. 10, 25 and 100 coefficients were chosen: seen from the standpoint of complexity, 10 taps

1847

corresponds t o 2 x 5 taps for 2x 1-D, 25 taps comesponds t o 5 x 5 taps, and 100 t a p s serves as a benchmark. T h e results are reported in Fig. 4. Since Ntap < Ngrid and is very large, the weighted distance 1 - Ifom,.zTS 1 - 11r7naSAF was chosen as the favored selection criteria, see Section 5. If Ntap = 10 = 2.5, the MSE is similar t o the MSE of 2x 1-D filtering for practical SNR. If Ntap = 25 = 5 . 5, the MSE of 2-D filtering is always better.

(2:;)
+

[9] T. Mueller, K. Brueninghaus, and H. Rohling, Performance of coherent OFDM-CDMA for broadband mobile communications, Wireless Personal Communications, Kluver Academic Publishers, vol. 2, pp. 295-305, 1996. [IO] P. Hoeher, A statistical discrete-time model for the WsSUS multipath channel,), IEEE Trans. Vehicular Techn., vol. 41, no. 4, pp. 461-468, Nov. 1992.
10

7 CONCLUSIONS
T h e discrete shift-variant 2-D Wiener filter was derived and analyzed given a n arbitrary sampling grid P and including model mismatch. To reduce complexity, Ntap can be m a d e smaller than Ngrld. If so, t h e set of observations I(, should be optimally chosen from P. T h e 1) MSE was expressed in closed form; the analysis is valid for any FIR filter. A generalized 2-D sampling theorem was proposed. Application was presented for pilot-symbol-aided channel estimation. 2-D Wiener filtering was compared with 2x l-D Wiener filtering and l-D Wiener filtering. Filtering in two dimensions was shown t o outperform filtering in j u s t one dimension with respect t o overhead a n d mean-square error performance. However, for a similar computational effort, t h e performance of 2x 1-D Wiener filtering is similar t o 2-D Wiener filtering.

10

L H

102

io3

10
EJN,, dB

~i~~~~ 2: A~~~~~~MSE tering.

10

VerSUS

SNR for l - wiener ~ fil-

10

References

R.M. Mersereau and T.C. Speake, The processing of periodically sampled multidimensional signals, IEEE Trans. Acoust., Speech, Szgnal Proc., vol. ASSP-31, io3 no. 2, pp. 188-194, Feb. 1983. C.W. Helstrom, Image restoration by the method of least squares, J . Opt. Soc. Amer., vol. 57, pp. 297-303, 1 0 o l d . 5 0 1 0 0 1 5 0 200250 ~ 3 0 0 - 3 5 0 ~ ~ 0 Mar. 1967. E N. dB J, M.P. Ekstrom, Realizable Wiener filtering in two dimensions, IEEE Trans. Acoust., Speech, Signal Proc., Figure 3: Average MSE versus SNR for 2x 1-D Wiener vol. ASSP-30, no. 2, pp. 31-40, Feb. 1982. filtering given a rectangular grid. M.L. Moher and J.H. Lodge, TCMP - A modulation and coding strategy for Rician fading channels, IEEE J. Sel. Areas Commun., vol. 7 , pp. 1347-1355, Dec. 1989. A. Aghamohammadi, H. Meyr, and G. Ascheid, A new 10 method for phase synchronization and automatic gain control of linearly modulated signals on frequency-flat fading channels, IEEE Trans. Commun., vol. COM-39, L 10.2 H pp. 25-29, Jan. 1991. J.K. Cavers, An analysis of pilot symbol assisted modulation for Rayleigh fading channels, IEEE Trans. Ve1o - ~ hicular Techn., vol. VT-40, pp. 686-693, Nov. 1991. S. Fechtel, Verfahren und Algorithmen der robusten 10 Synchronisation fur die Dateniibertragung iiber disper00 5 0 100 150 200 2 5 0 300 350 400 sive Schwundkande, Ph.D. thesis, RWTH Aachen, E N dB J, Germany, 1993. P. Hoeher, TCM on frequency-selective land-mobile Figure 4: Average MSE versus SNR for 2-D Wiener filfading channels, in Proc. 5th Twrenia Int. Workshop tering given a rectangular grid. on Dig. Comm., Tirrenia, Italy, pp. 317-328, Sept. 1991.

: lo2

1848