On Channel Estimation in OFDM Systems
On Channel Estimation in OFDM Systems
On Channel Estimation in OFDM Systems
Abstract - The use of multi-amplitude signaling have also addressed this property.
schemes in wireless OFDM systems requires the track- In Section 11, we describe the system model. Section I11
ing of the fading radio channel. This paper addresses discusses the minimum mean-square error (MMSE) and
channel estimation based on time-domain channel least-squares (LS) channel estimators. The MMSE esti-
statistics. Using a general model for a slowly fad- mator has good performance but high complexity. The
ing channel, we present the MMSE and LS estimators
LS estimator has low complexity, but its performance is
and a method for modifications compromising between
complexity and performance. The symbol error rate not as good as that of the MMSE estimator. We present
for a 16-QAM system is presented by means of simu- modifications to both MMSE and LS estimators that use
lation results. Depending upon estimator complexity, the assumption of a finite length impulse response. In
up to 4 dB in SNR can be gained over the LS estima- Section IV we evaluate the estimators by simulating a
tor. 16-QAM signaling scheme. The performance is presented
both in terms of mean-square error (MSE) and symbol
error rate (SER).
I . INTRODUCTION
Currently, orthogonal frequency-division multiplexing 11. SYSTEMDESCRIPTION
(OFDM) systems [l]are subject t o significant investiga- We will consider the system shown in Fig. 1, where zk:
tion. Since this technique has been adopted in the Euro- are the transmitted symbols, g ( t ) is the channel impulse
pean digital audio broadcasting (DAB) system 121, OFDM response, E ( t ) is the white complex Gaussian channel noise
signaling in fading channel environments has gained a and yk are the received symbols. The transmitted symbols
broad interest. For instance, its applicability t o digital x k are taken from a multi-amplitude signal constellation.
TV broadcasting is currently being investigated [3]. The D/A and A/D converters contain ideal low-pass filters
The use of differential phase-shzji! keying (DPSK) in with bandwidth l/Ts, where T, is the sampling interval.
OFDM systems avoids the tracking of a time varying chan- A cyclic extension of time length TG (not shown in Fig. I.
nel. However, this will limit the number of bits per symbol for reasons of simplicity) is used to eliminate inter-block
and results in a 3 dB loss in signal-to-noiseratio (SNR) interference and preserve the orthogonality of the tones.
[4]. If the receiver contains a channel estimator, multi- We treat the channel impulse response g(t) as a time-
amplitude signaling schemes can be used. limited pulse train of the form
In [5] and [6], 16-QAM modulation in an OFDM sys-
tem has been investigated. A decision-directed channel- g(t) = ams(t - T m T s ) , (1)
tracking method, which allows the use of multi-amplitude m
schemes in a slow Rayleigh-fading environment is analysed where the amplitudes a , are complex valued and 0 5
in [ 5 ] . T,T, 5 TG, i.e., the entire impulse response lies inside
In the design of wireless OFDM systems, the channel is
usually assumed to have a finite-length impulse response.
A cyclic extension, longer than this impulse response, is
put between consecutive blocks in order to avoid inter-
block interference and preserve orthogonality of the tones
[ 7 ] . Generally, the OFDM system is designed so that the
cyclic extension is a small percentage of the total sym-
bol length. This paper discusses channel estimation tech-
niques in wireless OFDM systems, that use this property
of the channel impulse response. Hoeher [6] and Cioffi [8] Fig. 1: Base-band OFDM system
io
,
20
,
30
,
40
*
50
,
. -........................... ....... ...-*--.
*
6n k.
816
where QMMSE
can be shown t o be
t
This MMSE channel estimator (9) has the form shown in
A
817
LS estimator (13) for all SNRs, since the excluded taps loo
contain only noise.
More generally, for non-T-spaced channels, any subset
of the taps in g may be taken into account when modi-
fying the MMSE and the LS estimator. The size of this
subset determines the complexity for both types of modi-
fied estimators.
In Section IV we consider the case where this subset
consists of the taps go . . . gL+K-1 and Q N - K . . . gN-1,i.e. ,
the first L taps as well as K extra taps on each side.
I V . SIMULATIONS
I
A. Simulated Channels io4' ; 10 15 20 25 30
In the simulations we consider a system operating with Average SNR [dB]
a bandwidth of 500 kHz, divided into 64 tones with a total Fig. 6: Mean-square error for three modified MMSE estimators.
symbol period of 138 p s , of which 10 ps is a cyclic prefix.
Sampling is performed with a 500 kHz rate. A symbol
thus consists of 69 samples, five of which are contained in
the cyclic prefix (i.e. L = 5 ) . 50,000 channels are ran-
domized per average SNR, each consisting of five pulses,
of which four have uniformly distributed delays over the
interval 0-10 ps, while one tap is always assumed to have
a zero delay, corresponding to a perfect time synchroni-
sation of the sampling instants. The multipath intensity
profile is assumed to be d ( ~w)e C 7 I T r m s , where 7;msis 1/4
of the cyclic extension. We have used Monte-Carlo simu-
lations t o generate the R,, for this channel model. This
covariance matrix together with the true noise variance c:
is used in the MMSE estimations t o follow. The average
SNR per symbol in Fig. 3 is defined as E{lhk12}/E{lnk/2},
since E{(xk12) is normalised to unity.
The following estimators are used: Average SNR [dB]
Fig. 7: Mean-square error for three LS estimators.
[ Estimator I Notation I Taps used I Size Q' 1
dimensions of QL will give lower mean-square error for
all SNRs.
Fig. 7 shows the mean-square error versus average SNR
for the MMSE, LS, LS-0, LS-5 and LS-10 estimators. Con-
trary to the modification of the MMSE estimator: the
modification of the LS estimator reduces the mean-square
error for a range of SNRs.However, the same approxima-
tion effect as in the modified MMSE estimators shows up
at high SNRs. An interesting observation is, that for every
B. Mean-square Error SNR there exists an optimal size of QLs, which gives the
smallest mean-square error compared t o the other modi-
Fig. 6 shows the mean-square error versus SNR for fied LS estimators.
the MMSE, LS, MMSE-0, MMSE-5 and MMSE-10 esti-
mators. The difference between the modified MMSE esti- C. Symbol-error Rate
mators and the MMSE estimator is due to the fact that
parts of the channel statistics are not taken into account The symbol-error rate (SER) curves presented in this
in the former. For low SNRs,this approximation effect section ase based on the mean-square errors of the channel
is small compared t o the channel noise, while it becomes estimations presented in the previous section. For the
dominant for large SNRs. The curves level out to a value calculation of SER, we have used the formulae presented
determined by the energy in the excluded taps. Larger in [lo]. These formulae find the symbol error rate of a
818
10 1O0
W Y
Q,
U
2 2
k
tl 10
3
0
P
E
tn
10 10
5 10 15 20 25 30 5 10 15 20 25 30
Average SNR [dB] Average SNR [dB]
Fig. 8: Symbol error rate for three modified MMSE estimators. Fig. 9: Symbol error rate for three modified LS estimators
5 10 15 20 25 30
V . CONCLUSIONS Average SNR [dB]
Fig. 10. Comparison of SER between a modified MMSE estimator
The estimators in this study can be used t o efficiently and the corresonding modified LS estimator.
estimate the channel in an OFDM system given a certain
knowledge about the channel statistics. The MMSE esti- M. Alard and R. Lassalle, Principles of modulation and chan-
mator assumes a priori knowledge of noise variance and nel coding for digital broadcasting for mobile receivers, EBU
channel covariance. Moreover, its complexity is large com- Review, no. 224, pp 3-25, August 1987.
B. Marti et al., European activities on digital television broad-
pared to the LS estimator. For high SNRs the LS estima- casting-from company to cooperative projects, EBU Techni-
tor is both simple and adequate. However, for low SNRs, cal Review, no. 256, pp. 20-29, 1993.
the presented modifications of the MMSE and LS estima- John Proakis, Digital Communications, McGraw-Hill, 1.989.
tors will allow a compromise between estimator complex- Sarah Kate Wilson, R. Ellen Khayata and John M. Cioffi,
ity and performance. For a 16-QAM signaling constella- 16-$AM modulation with orthogonal frequency-division mul-
tiplexing in a Rayleigh-fading environment, In Proc. VTC-
tion, up to 4 dB gain in SNR over the LS estimator was 1994, pp. 1660-1664, Stockholm, Sweden, June 1994.
obtained, depending on estimator complexity. Even rela- Peter Hoeher, TCM on frequency-selective land-mobile fading
tively low-complex modified estimators, however, perform channels, In Proc. of the 5th Tirrenia International Workshop
significantly better than the LS estimator for a range of o n Digital Communications, Tirrenia, Italy, September 1991.
John A.C. Bingham, Multicarrier modulation for data trans-
SNRs. mission: an idea whose time has come, IEEE Communications
Magazine, 28(5):5-14, May 1990.
John M. Cioffi, Stanford University, Private communication.
REFERENCES Louis L. Scharf, Statistical Signal Processzng, Addison-Wesley,
1991.
[I] Leonard J . Cimini, Jr., Analysis and simulation of a digital
mobile channel using orthogonal frequency-division multiplex- Sarah Kate Wilson, Digital Audio Broadcasting in a Fading and
Dispersive Channel, PhD-thesis, Stanford University, August
ing, IEEE D u n s . Comm., Vol. 33, no. 7, pp 665-675, July
1994.
1985.
819