Optimal Whitening Approach for

Improved Channel Estimation in GSM Systems

S J Thiruvengadam, Deepthi Chander, G.Venkatesh and V Praveen Kumar
Signal Processing Laboratory,
Department of Electronics and Communication Engineering,
Thiagarajar College of Engineering,
Madurai – 625 015.

Abstract In this paper, a transformation of the LS estimate of ‘h’ is

chosen such that the covariance of the error in its estimate
In GSM applications, the receiver detects the symbols (and hence of the estimate) is constrained to be white. W e
transmitted by different users, after making an estimate of apply a whitening transformation W, to hL S ,such that the
the channels through which the symbols are received. transformed vector is ‘as close as possible’ in an MMSE
Currently GSM employs the Method of Least Squares (LS) sense to hL S. We consider the synchronous case of two
to estimate the unknown channels. We propose to apply an users trying to use the same frequency band in the same
optimal whitening transformation to this LS estimate of the time-slot. It is also proved that the LS estimator followed
channels to improve the accuracy of the estimate, at
by a whitening transformation becomes an estimate that
moderate to low Signal to Noise Ratio (SNR) prevalent in yields the CRLB for biased estimators. For two users using
mobile environments. This estimate also achieves the the same frequency at the same time slot (synchronous
CRLB. Further, it enables the assignment of non-
case), this proposed estimator permits GSM to allocate non-
orthogonal midambles to GSM users accessing the same orthogonal midambles to the users as an alternate to the
frequency band in the same time slot in the synchronous existing system of assigning only uncorrelated midambles
case.
to two such users to prevent interference.
Key words: LS Estimation, Best linear Unb iased estimator Section 2 deals with signal and channel model. Section 3
(BLUE), whitening Transformation, CRLB. explores the LS channel estimation currently used in GSM
system. Section 4 describes the whitening transformation
1. Introduction employed and derives the CRLB. Section 5 concludes the
paper. `
Cellular networks today try to cater to a large number of
users simultaneously. The interference due to other users 2. Signal Model
places a constraint on this goal. In the GSM network, the
number of users in a cell is system-dependent with every
cell assigned a set of frequency bands, each of which can The structure of the physical content of a time slot, i.e. a
accommodate upto 8 users in different time-slots. burst, specified by the GSM standard, is shown in
Figure 1. A midamble of Nm = 26 symbols, defined by the
In GSM networks, the symbols are detected by employing a training sequence code (TSC), is placed in the center of
Trellis search to give a good receiver performance. each burst, and it is used for estimating radio channel
However, the knowledge of the channel impulse response is conditions. Two 58-symbol long blocks on each side of the
inevitable for the detector. The current scheme of midamble contain data. A guard period (GP) of 8.25
estimation in GSM employs the LS estimate to determine symbols at the end of each burst prevents signals in
the unknown channel coefficients. Amongst all unbiased consecutive time slots from overlapping. Hereinafter, we
estimators, the LS estimator is certainly the minimum will ignore both guard period and tail binary symbols (TB),
variance estimator. Also, if the noise in the channel is and observe only NB =142 symbols of data and midamble
additive white Gaussian noise (AWGN), it becomes the as the user's burst.
Best Linear Unbiased Estimate (BLUE). However, when
the SNR is low or moderate, the receiving samples ’y’, are T B Data symbols TSC Data symbols TB GP
not sensitive to changes in channel coefficients ‘h’, in 33 58 26 58 3 8
which case the LS estimate of channel coefficients
(hL S ) becomes a poor estimate of ‘h’. The error in 1time slot =.577ms
estimation of ‘h’ can have a large dynamic range in its
variance and spectral shape. Figure1 A Typical GSM Burst

1
While propagating through the radio channel, the signal is proportional to path loss and shadowing coefficients of
subject to different time varying distortions due to each user's signal. Thus, the signal of interest arrives at
interference, fading and noise. In this paper, we consider a MS1 with compensated attenuation factors: ws1ws1 -1 =1 and
flat fading channel that does not induce inter-symbol wp1 wp1-1 =1 while the interfering signal will be received at
interference. In this case, apart from additive noise, the MS1 attenuated by relative shadowing and path loss
channel introduces only an unknown attenuation that scales coefficients given respectively by:
the signal amplitude. The total attenuation is the result of
−α / 2
the following processes that are assumed mutually ω ωp1  r1 
independent. ωs = s1 ; ωp = = 
ωs 2 ωp 2  r 2 
1. Path loss: Signal power decreases according to
the power law of the distance between the transmitter and The baseband signal model can be expressed as:
the receiver. In the general case, the path loss attenuation
can be expressed as
y(i ) = x1 (i )h 1 + ω s ω p x 2 (i )h 2 + n(i ),
2 −α
ωp  r 
=  where y(i) is a Na×1 vector of the received signals at Na
2 r 
ωp  0 antennas of MS1 at the ith symbol interval, 1= i = NB; x1, x2
0 ? { - 1; +1} are ith transmitted symbols from BTS to MS1
and MS2 , n is a complex Gaussian noise vector (zero mean,
where a is the path loss exponent that depends on the with variance ¬2 ) and h 1 and h 2 are channel impulse
environment, wp and wp0 are amplitude weights at distance r response vectors containing i.i.d. complex Gaussian
and reference distance ro, respectively. coefficients for each receiving antenna. Thus, they can be
viewed as unique spatial signatures characterizing each
2. Large scale fading (shadowing) is due to large user's signal space.
obstacles in the propagation path that block the signal. It is
a slowly varying process, modeled stochastically with log-
normal distribution: 3. Channel Estimation
Since both the useful and the interfering signal arrive at the
3. Small scale fading is the result of multipath mobile receiver from the same BTS, bursts will completely
propagation. The amplitude of the received faded signal is overlap, as shown in Figure 2. Due to attenuation, power
modeled as a random variable with Rayleigh distribution. levels differ significantly in general. The midambles

2.1 Users in the Same Cell [x1 (58+1)… x1(58+26)] T = m1 and

Consider the case where two users served by the same base [x2 (58+1)…x2(58+26)] T =m2
transceiver station (BTS). The cell is modeled as a circular \will be perfectly aligned. Therefore, in order to enable
area of radius R[4], with the BTS in the centre and the identification and estimation of different channels for each
users uniformly, independently identically distributed user, it is necessary that they use different midambles. We
(i.i.d.) within. The position of each user is determined by assume that MS1 knows the interferer's midamble m2 .
the polar coordinates (r;ø), which are random variables with
the following probability density functions (PDFs):
2π r 2r
p( r ) = = 2 ,0 ≤ r ≤ R BS1à MS1
πR 2
R
1
p( ϕ) = ,0 ≤ ϕ ≤ 2 π Data Midamble Data
2π
BS1à MS2
Let ws1 ,wp1 be the shadowing and path loss factors
between the BTS and MS1,and ws2 ,wp2 the equivalent Figure 2 :Overlapping Bursts
coef_cients for the BTS - MS2 link. Since we assume
perfect power control performed in the downlink, the base Let us consider now a communication system in the
station will adjust the transmit power to be inversely presence of co-channel interference that is shown in Figure
2
3. Two synchronised co-channel signals have independent m 0 , n 
complex channel impulse response  
m 1 , n 
h L,n =[h 0,n , h 1,n …. h l,n] T ,n=1,2
m n = . , n = 1,2
 
and where L is the length of the channel memory. The sum . 
of the co-channel signals and noise n is sampled in the
m 
receiver. The joint demodulation problem is to detect the  p+ L −1, n 
transmitted bit streams a1 and a2 of the two users from the
received signal y. To assist that joint detection operation The circulant training sequence matrices are denoted by
the Joint channel estimator provides channel estimates h1 ^
and h 2^
m L, n .............m1, n m 0 ,n 
 
m L+1, n ...........m 2 ,n m1, n 
S1 a1 Ch1 noise
M n = ................................ 
 
Rx Joint
................................. 
S2 a2 Ch2
+ Filter detector m ..........m P ,n m P −1,n 
 L+ p −1,n 

Joint ch and they are gathered into one large matrix

estimator
~
M = [M 1 M 2 ]
Figure 3: block diagram of GSM system
With these notations the received signal y is again given by
The complex channel impulse responses of the two
synchronous co-channel signals are expressed with a vector
h~ as follows

~ h L , 1  The LS channel estimates can be found simultaneously for

h=  the both users by minimizing the squared error quantity,
h L , 2  which produces in the presence of AWGN the following
solution
containing the channel taps of the individual signals
denoted by
h 0 , n 
 
h 1 , n  3.2 Limitation in the existing GSM System in
h L, n = . , n = 1,2 Channel Estimatio n
  The resultant h is not uncorrelated, as can be proved from
.  its covariance computation:
h 
 L ,n  Covariance of ĥ = (MH M)-1 M H E(y y H) M (MH M)-1
=E (nnH ) (M H M)-1
Hence, h~ has totally 2×(L+1)elements. Both the =σ2 (M H M)-1 ,
transmitters send their unique training sequences with a Under conditions of low to moderate SNR, the covariance
reference length of P and guard period of L bits. The of the channel estimate can have a large dynamic range,i.e
sequences are denoted by the error in the estimate of ‘h’ can have a large dynamic
range. The covariance matrix is not diagonal,i.e total
decorrelation does not exist.In GSM , only four pairs of
midambles per slot are orthogonal to each other .This
implies that we are forced to allocate only these pairs to
users of the same cell, operating in the same time slot.
3
4. Whitening Approach to Channel E( a k b k ) ≤ E( a k b k ) ≤ (E(a k2 )E(bk2 )) = c 2 E(a 2k )
1/ 2

Estimation: (3)

The LS estimator seeks the estimate of h that results in a

With equality if and only if ( )
bk = c / d k a k with
data vector y that is ‘as close as possible ’ to the original probability 1.
ˆ =cD-1/2,From (1)
W
data vector y. However we wish to minimize the error
between h and it’s estimate. In conditions of low to
moderate SNR, the data vector y is not very sensitive to Ŵ = cVD -1/2 V*=c Ca -1/2 (4)
changes in h, so that large error in estimating h may Minimizing (2) with respect to c,
m
translate into small errors in estimating ‘y’. To improve
the performance of LS estimator, we make use of a c = (1/m) ∑k =1
dk
modification of the LS channel es timate, based on the
concept of MMSE whitening[1],[2]. and the optimal whitening transformation is
Ŵ = c VD-1/2 V*=cCa-1/2
Theorem 1 (MMSE Whitening Transformation) :
4.1 Whitened Channel Estimate:
m
Let a e R b e a random vector with positive definite
covariance vector C a =VDV*, where D is a diagonal matrix Since h LS =h+n, where n = (MH M)-1 M H n, and (the
and V is a unitary matrix .Let w be the optimal whitening covariance of the noise component n in h LS) the covariance
transformation that minimizes the MSE between the input a of h LS, is given by ¬2 (M HM)-1 ,.
and the output b =Wa with covariance
From Theorem 1, it follows that the optimal whitening
Cb =c2 Im where c > 0 Then transformation is proportional to w = (M*M)1/2 so that the
WTLS estimator, denoted by HWTLS, has the form
W =aVD -1/2 V* = aCa-1/2
Where HWTLS = ß (M*M)1/2 xL S = ß (M*M)1/2M*y

1. If c is specified then a = c Minimizing the above equation with respect to

ß, the optimal value of ß is given by
2. If c is chosen to minimize the MSE then
ß = Tr ((M*M)1/2 ))
m Tr(M*M)
a = (1/m) ∑ dk
k =1 4.2 Cramer-Rao Lower Bound (CRLB)
[for biased estimators]
proof: Let Ca= VDV*, ( eigen decomposition).
Define,a =V*a; b =V*b;therefore,Ca =D; i.e, the elements For biased estimators, whitening estimator attains
of a are uncorrelated. Also, since V* is unitary, Cb = c 2 I m, CRLB[2],[3].
same as Covariance of b .Hence MSE of (a ,b)=MSE of (a
∂ ∂
, b ). Cov(estimate of h) = [ Im + B(h) ] J-1 [ I m + B(h)]*
It is straightforward to show that, ∂h ∂h
Ŵ = V Wˆ V *; (1) ∂2
where J(h)= -E [ log p(h,y)];
ˆ 2
To determine W , mean square error is given by, ∂h

m m
( Fisher information matrix)
ε MSE = k∑=1 E((a k − b k ) 2 ) = k∑=1 d k + mc 2 − 2k∑=1 E( a k b k )
(2) For the whitening estimator, the bias is given by
B(hWTLS) = (ß(MH M)1/2-Im) h
where d k = E (a ) are the eigen values of Ca .
2
k

From the Cauchy-Schwartz inequality

4
∂ biased estimators. Hence it is bound to outperform all
(B(hWTLS )) = (ß(M HM)1/2 -Im) existing methods of channel estimation in GSM
∂h environments. We can also employ the whitening approach
When the noise is Gaussian J(h) = M* M in the Multi-User-Detection (MUD) of received symbols.
Therefore the CRLB on the variance of any estimator with
bias B(hWTLS);
References:
Var (hWTLS) = ß 2 ((M H M)1/2 (MH M)-1(MH M)1/2 [1] Yonina.C.Eldar, Alan. V. Oppenheim, “MMSE
= ß 2 Im. Whitening and Subspace whitening”, IEEE Transactions on
Inform. Theory, Vol.49, No.7, pp 1846-1851 ,July 2003.
However, the whitening transformation transforms the least
square estimate to an estimate with covariance= ß 2 Im. [2]Yonina.C.Eldar, Alan. V. Oppenheim, “Covariance
Hence the whitening estimator is the optimum estimator of Shaping Least-Squares Estimation”, IEEE Transactions on
the unknown channel coefficients. Signal Processing, Vol.51, No.3,pp 686-697, March 2003.

Figure(4) describes the imp roved performance of the [3]Steven M Kay “Fundamentals of Statistical Signal
whitening estimator, with respect to the LS estimator by Processing: Estimation Theory.” Upper Saddle River, NJ:
means of a plot of the mean squared error of the estimate as Prentice Hall, 1993.
a function of SNR. Here, we have considered low to
moderate SNR values ranging from -10db to +10 db.As can [4]Maja Loncar, Christoph F.Mecklenbrauker, Ralf
be seen,at an SNR of 0db,a mean-squared error of 10 2..5 of R.Muller “ Co-Channel interference mitigation in GSM
the LS estimate has been brought down to 100 by using the networks by iterative estimation of channel and data”, May,
whitening transformation. 2002.

[5]P.A.Ratna and A.Hottinen and Z-C.Honkasalo, “Co-

channel interference canceling receiver for TDMA Mobile
systems”, Proc.of IEEE Int.Conf.on Communications
(ICC),pp 17-21, 1995

[6] S.Verdu, Multiuser Detection, New York, 1998.

Figure(4): Comparison of MSE between LS and Whitening

Approach.

5. Conclusion
In this paper, we propose the application of whitening on
two synchronous, power imbalanced users in a GSM-like
system. From the results obtained mathematically and
through simulations, we can infer that the use of whitening
to the LS estimate improves the performance of GSM
channel estimation, and hence the overall performance in
detection of symbols. Among all the linear estimators for a
given bias, this estimator yields the minimum variance, as it
theoretically obeys the Cramer-Rao Lower bound for
5