Communication Theory

LINEAR MMSE CHANNEL ESTIMATION FOR GSM

Jens Baltersee, Gunnar Fock, Heinrich Meyr Lihbor Yiin
Aachen University of Technology Lucent Technologies
Integrated Signal Processing Systems Microelectronics Group
Templergraben 55 1247 S. Cedar Crest Blvd.
52056 Aachen, Germany Allentown, PA 18103, U.S.A.
balterse @ ert.rwth-aachen.de yiin @ 1ucent.com

Abstract ing multiple sets of estimators, each one corresponding to a

A new channel estimation algorithm capable of exploiting different sampling time. Here, we resort to a Bayesian ap-
a-priori knowledge about modulation and receive filters and proach in order to incorporate the a-priori knowledge into
the power delay profile of the physical channel is presented. the channel estimator. In other words, it is necessary to
In order to incorporate the a-priori knowledge into the linear model the CIR taps as random variables with a given prior
minimum mean square error (LMMSE) channel estimator, a PDF. With this approach it is possible to find a linear esti-
Bayesian approach is applied to an appropriate GSM system mator which is optimal in a mean square error (MSE) sense
model. The performance of the new algorithm is compared with respect to the assumed prior PDF.
to a conventional least squares (LS) channel estimator by
means of analysis and simulations. Simulations are carried 2 Signal Model and LS channel estimation
out for the standard GSM channel profiles (TU50, HT100, In a digital communication system, such as GSM, informa-
RA2.50). Information about the shape of the power delay tion symbols { a k } are sent at rate 1 / T over a known ef-
profile is assumed to be unavailable to the receiver. The fective channel h with inter-symbol interference (ISI). This
merit of the new channel estimator is confirmed by yield- is a reasonable assumption which follows the concept of
ing an MSE approximately 1 - 2 dB lower than the LS es- synchronised detection [l] for which a channel estimate h
timator. The lower MSE translates into a BER advantage must be formed and subsequently used for detection as if
of approximately 0.25 dB. Exploiting additional knowledge it were the true known channel. Furthermore, the channel
about the shape of the power delay profile results in further is assumed to be time-invariant for the duration of a burst.
BER performance improvements of up to 0.3 dB. In GSM systems, the received signal is approximately ban-
dlimited to 1/(2T) and therefore symbol rate sampling guar-
1 Introduction antees sufficient statistics for the tasks of detection and syn-
The “classical” channel estimator that is normally employed chronisation. The GSM standard provides midamble train-
in GSM receivers is based on the Least Squares (LS) ap- ing sequences of length 26 that can be exploited to calcu-
proach and does not exploit any available a-priori informa- late a symbol spaced estimate of h. Denote the training se-
tion about the channel impulse response (CIR). However, quence as a = (ao,. . . , ~ . 5 ) ~Furthermore,
. if the channel
it is highly desirable to incorporate a-priori knowledge into memory is L then write the ( L S 1) x 1 T-spaced CIR vec-
the channel estimator, because this will improve the qual- tor h = (ho,. . . ,h ~ )In~matrix
. notation, the corresponding
ity of the channel estimate and translate directly into better linear transmission model is given by
BER performance of the receiver. The effective CIR that
is “seen” by the equaliser of the GSM receiver consists,
roughly speaking, of the convolution of the transmit filter,
the physical channel p ( t ) , and the receive filters. The char-
acteristics of the transmit filter and any digital receive filter
are known to the receiver, and can thus be exploited. In
ri = ( ;y:i
aL+15+i aL+14+i
_jj
*“
“i(”)
a15+i hL
+ ni

1997, Khayrallah et al. [4] presented a constrained LS so- = Aih+ni (1)

lution in order to incorporate the a-priori information into
the channel estimator. However, their approach suffers from The received signal samples are stored in the 16 x 1 vec-
a drawback in that a symbol spaced physical channel is as- tor ri, and the subscript i determines the exact position of
sumed (which is not to be confused with the symbol spaced the samples within a received burst, usually by means of
effective channel). Consequently, the correct sampling in- a so-called coarse timing estimate. The exact procedure is
stance within the symbol period has to be estimated by us- not within the scope of this paper, and suffice to say that 16

Global Telecommunications Conference - Globecom’99 2523

0-7803-5796-5/99/$10.0001999 E
I EE

Authorized licensed use limited to: Albert Chan. Downloaded on August 27, 2009 at 17:12 from IEEE Xplore. Restrictions apply.
Communication Theory

wide sense stationary uncorrelated scattering we have

v
a
v
,
L

' I . - L-.'1
I
c
OT LT time It is now possible to build a model of the discrete-time effec-
tive channel h (see Figure 2). Basically it consists of three
Figure 1: Power delay profile of physical channel. parts. The first part is an approximation of the GMSK mod-
ulation as described in [5]. This approximation is equivalent
received signal samples are used, because this facilitates an to a PAM modulation using n/2-phase rotated BPSK sym-
easy implementationof the conventional LS channel estima- bols and transmit pulses Co(t) and CI(t). We define
tor. If, for ease of notation, the subscript i is omitted, then
the LS channel estimator for the above linear transmission
model is given by
a0,k = (5)

r=-1

The symbols { U o , k } are simply the rotated original BPSK

Note, that the linear signal model is only an approximation, symbols { a k } which are subsequently filtered with the
because the GSM system employs a non-linear modulation pulse Co(t). This in itself constitutes the optimal linear
format called Gaussian Minimum Shift Keying (GMSK) [2]. approximation to GMSK. The symbols { U ] & } are rotated
However, it is possible to find a close linear approximation and nonlinearly combined BPSK symbols which are then
to GMSK [5].Thus it becomes possible to treat GMSK as a filtered with the C1 ( t ) pulse. Both resulting signals are then.
linear modulation with transmit filter Co(t), without notice- added to form a reasonably accurate approximation to the
ably compromising the performance of the receiver. For the GMSK modulation. In theory it is possible to approximate
LMMSE channel estimator that is presented next, an even the GMSK modulation ever more closely by including more
closer approximation is feasible that includes some of the higher-order terms, but in practice this is not necessary,
non-linearities inherent in GMSK. because these higher-order terms are negligibly small
( M -60dB). The second part is the discrete-time physical
3 Linear Minimum Mean Square Error channel p. The discrete channel results from filtering
(LMMSE) Channel Estimation the noise corrupted output of the continous time channel
p ( t ) with an anti-aliasing filter (AAF) of cutoff frequency
In order to incorporate the a-priori knowledge, the CIR vec- 1/2T and then sampling with a clock period of T . Note.,
tor h is modelled as a random vector with mean E(h) = 0 that for simplicity we consider the AAF as part of the:
and known covariance matrix Chh. Each' tap in h is as- physical channel. This can be justified, because normally
sumed rayleigh faded. The noise vector is Gaussian with the following digital receive LP filter dominates the overall
PDF !7((O,C,). In that case, the optimal linear minimum frequency response, and incorporating the a-priori knowl-
mean square error (LMMSE) estimator is given by [3] edge about the AAF does not result in further performance:
improvements. Therefore, the a-priori information that we
hLMMSE = Chr ( c r r + Cn)-' r would like to exploit is contained in the first part, the GMSK.
= XLMMSE
r (3) approximation, and the third part, the digital LP filter g k .
Apart from the channel memory L, the physical channel
From the above equation it is seen that in order to find the is unknown. From now on, the symbols { Q k } and { q , k }
LMMSE estimator and assess its performance, we have to are treated as deterministic quantities, calculated from the
determine the covariance matrices Chr, Crr, and C,. How known midamble sequence. Time-continous versions ao(t),
this can be accomplished will be shown in the following. a l ( t ) ,and&) of the symbols { U O , k } , { U l , k } , and the digital
Firstly, we assume that we do not have any prior knowledge LP filter, respectively, can be created by expanding { a O , k } ,
about the power delay profile of the physical channel, apart { a l , k } and { g k } with an -fold expander. Now we make the
from the channel memory L. In other words, the delays of following definitions
the multipath rays are, on average, uniformly distributed in
an interval [O,LT],and the powers of the rays are, on aver-
age, equal. The power delay profile of the resulting time-
continous physical channel p ( t ) is shown in Figure 1. For

2524 Globol Telecornmunicotions Conference - Globecom'99

Authorized licensed use limited to: Albert Chan. Downloaded on August 27, 2009 at 17:12 from IEEE Xplore. Restrictions apply.
 Communication Theory

GMSK discrete-time digital

approximation physical channel LP filter derotation

continous-time

Figure 2: Channel model.

where * stands for'convolution. Both of these time- is unknown to the receiver and has to be replaced in practice
continous functions f r ( t ) and fh(t) contain the a-priori by either an estimate or some desired SNR operating (de-
knowledge which is available at the receiver. Time- sign) point. The latter option has the great advantage that the
continous functions are required here, because in order to estimation matrix XLJI.IMSE can be pre-computed and addi-
calculate the covariance matrices, it is necessary to convolve tional on-line computations are not necessary. Note, that the
these functions with the power delay profile R p ( t ) . Now, resulting estimator is very insensitive to any mismatch be-
with the above definitions in mina, it is possible to calculate tween the true SNR and the design SNR and a near optimal
the covariance matrices Chh, C,, Chr: Firstly, the sampled performance gain is therefore feasible over a wide range.
received midamble signal which is not subject to AWGN is
written as. 4 Performance Analysis of LMMSE channel
estimation
' J(k) =
1: p(t)f,(kT - t)dt (9)
In order to assess the performance of the channel estimation
Then the covariance matrix Cr,.is given by algorithm, it is necessary to define a performance measure
E. We define E as the mean squared error (MSE) between the
icCrrlk/
= E { ? * ( k ) ? ( l ) }= channel h and its estimate h. The MSE can be written as the
1: R p ( t ) f f ( k T-t)f,(ZT -t)dt (10) sum of a noise attenuation factor h, weighted with the noise
power P, and a self-noise factor h, weighted with the signal
power P,, i.e. we have
The covariance matrix Chh can be calculated using fh(t) as

and, similarly the cross-covarikce Ch,.is given by '

Remember that both, the LS and the LMMSE channel es-

timators, can be put into the general linear form h = Xi-,
where X can be the estimation matrix of one or the other
Following the same reasoning as before, the digital LP filter approach. In order to calculate the noise attenuation A,,,the
{&} is taken to dominate the combined frequency response estimation matrix X is simply applied to the received noise
of both, the AAF and the digital filter. In that case, the noise sample vector n, and we get
covariance matrix of C , is determined only by the impulse
response of the digital LP filter {gk}, i.e. A, = tr ( E { X n n H X H } )

Lg-I = tr (XCUX") (15)

[Cn1kl - eJ(k-')n/'. 0,'. gJ-kgj-1 (l3) The self-noise factor hs is given by the mean squared error
J=o
betwey the channel h and its estimate ho, where the esti-
where the digital LP has Lg coefficients, and 0: is the power mate ho is computed from received signal samples r which
of the lowpass-filtered AWGN process n ( t ) . Of course, 0: are not corrupted by the noise vector n (hence self-noise). If

Global TelecommunicationsConference,.- Globecom'99 2525

Authorized licensed use limited to: Albert Chan. Downloaded on August 27, 2009 at 17:12 from IEEE Xplore. Restrictions apply.
Communication Theory

we define (TU50, HT100, RA250) lie underneath thus guaranteeing

that performance does not decrease for any possible channel
fh(2) = [ f h ( - 2 ) , f h ( T - . t ) ,...,fh(LT-Z)IT (16) realisation. The simulation results confirmed an additional
BER performance improvement of 0.3dB for the TU50 and
and similarly f r ( z )using 16 samples of the function f r ( 7 ) , RA250 channels, and 0.2dB for the HTlOO channel.
then h, can be written as
m 6 Conclusion
= [ _ R p ( 7 ) Ifh(2) -Xfr(z)12dz (17) A new LMMSE channel estimation algorithm is presented
that easily outperforms the “classical” LS-type channel es-
Using E, it is possible to compare different channel estima- timator. This is accomplished by incorporating a-priori in-
tion algorithms. The most convenient way to perform the formation about known modulation and receive filters. The
comparison is by calculating the ratio of two E’S belonging merit of the new channel estimator is confirmed by yielding
to different channel estimation algorithms, i.e. an MSE approximately 1 - 2 dB higher than the LS estima-
tor. For the GSM system, the improved channel estimate
translates into BER improvements of up to 0.25dB. On one
hand, this improvement is not enormous, but on the other
hand the computational complexity required by the LMMSE
Figure 3 shows p plotted against the SNR for the LS and
channel estimator is comparable to that of the LS chan-
LMMSE channel estimation algorithms. The ratio shown
nel estimator. Much more complex channel tracking algo-
in the Figure is computed with respect to the LS channel
rithms that try to accurately estimate the time-variations of
estimator. It is seen that the LMMSE channel estimation
the channel within one burst achieve BER improvements in
algorithm easily outperforms the LS channel estimation al-
the same ballpark. If additional knowledge about the shape
gorithm. In the low SNR region (from OdB to 10dB) the
of the power delay profile is available at the receiver, fur-
mean square error of the LMMSE channel estimate is ap-
ther BER improvements of up to 0.3dB are possible. The
proximately 1dB better than that of tkz LS channel esti-
LMMSE approach is not limited to GSM systems and can
mate. In the higher SNR regions (up to 30dB) this advan-
be applied to any communication system that uses a training
tage increases up to approximately 2.5dB, indicating that
sequence assisted channel estimator.
the LMMSE channel estimator has a much lower self-noise
factor than the LS channel estimator. Note, that a trade- Acknowledgment
off between noise suppression and self noise is possible by
designing the estimator for a specific SNR via the noise co- This work was supported by Lucent Technologies Inc.
variance matrix of equation (13).
References
5 Simulation Results [l] Heinrich Meyr, Marc Moeneclaey and Stefan Fech-
The LS and LMMSE channel estimation algorithms are tel, Digital Communication Receivers: Synchroniza-
simulated with a GSM receiver under various conditions. tion, Channel Estimation and Signal Processing, John
The reference sensitivity test was performed for the TU50, Wiley and Sons, New York, 1998.
HT100, and RA250 channel profiles. All the tests were per- [2] “Digital Cellular Telecommunications System; Modu-
formed for the class I1 (uncoded) bits. The simulation re- lation (GSM 05.04 version 5.0.1)”, ETSI, May 1997.
sults are shown in Figures 4 to 6. For the TU50 and RA250
channel profiles, the LMMSE channel estimator results in a [3] Steven M. Kay, Fundamentals of Statistical Signal Pro-
BER advantage of approximately 0.25dB, whereas for the cessing - Estimation Theory, Prentice Hall, Englewood
HTlOO channel profile the advantage is 0.15dB. One pos- Cliffs, New Jersey, 1993.
sibility to improve the quality of the LMMSE channel esti- [4] A.S. Khayrallah, R. Ramesh, G. E. Bottomley, and D.
mates even further is to include more a-priori knowledge in Koilpillai “Improved Channel Estimation with Side In-
the power delay profile R p ( t ) .For example, all three channel formation”, IEEE Vehicular Tech. Con$, pp. 1049-1053,
models used in the reference sensitivity tests (TU50, HT100, Phoenix AZ, May 1997.
RA250) have the energy concentrated towards the first taps,
and this knowledge can be incorporated into the power de- [5] Pierre A. Laurent, “Exact and Approximate Construc-
lay profile R p ( t ) by shaping it accordingly. An exponen- tion of Digital Phase Modulations by Superpostion of
tial power delay profile was chosen as indicated by way Amplitude Modulated Pulses (AMP)”, IEEE Trans.
of example in Figure 7, so that the taps from all channels Comm., Vol. 34, pp. 150-160, February 1986.

2526 Global Telecommunications Conference - Globecom‘99

Authorized licensed use limited to: Albert Chan. Downloaded on August 27, 2009 at 17:12 from IEEE Xplore. Restrictions apply.
 Communication Theory

Comparison: MSE of channel estimation

I
I I
I I I
I I

-1 I
I
I
I

I I I I I
I I I I I
I I I I I
0 I I I I I

0 5 10 15 20 25 30
SNR

Figure 3: Performance comparison of LS and LMMSE channel estimation schemes.

TU50 -- EERll HT1W -- BERll

7 8 9 10 0.04'
8 9 10
SNR (dE) SNR (dE)

Figure 4:Reference sensitivity test, class I1bits, TU50 chan- Figure 6: Reference sensitivity test, class I1 bits, HTlOO
nel. channel.

RA250 -- BERll
, f"wnenti4' Pwrer Oday Proll: for MMS,E Cham: Estlmty ,
o,

7 8 9 10
SNR (dE)
Delay

Figure 5: Reference sensitivity test, class I1 bits, RA250

channel. Figure 7: Exponential Power Delay Profile.

Global Telecommunications Conference - Globecom'99 2527

Authorized licensed use limited to: Albert Chan. Downloaded on August 27, 2009 at 17:12 from IEEE Xplore. Restrictions apply.