CHANNEL TRACKING FOR FAST FADING LONG-CODE WCDMA

Youngchul Sung and Lang Tong

School of Electrical and Computer Engineering

Cornell University
Email: ys87,ltong
@ee.cornell.edu

ABSTRACT The paper is organized as follows. The data model of

a CDMA system is described in Section 2. A blind chan-
A new technique for tracking of fast fading channels in long
nel tracking method based on linear interpolation channel
code CDMA is proposed. Using a linear interpolation model,
model is proposed in Section 3. In Section 4, the perfor-
the proposed method tracks the time-varying channel with-
mance of the proposed method is assessed by Monte Carlo
out pilot symbols or with sparsely inserted pilots. The pro-
simulations and compared with the estimation with block
posed method can be implemented efficiently using a state-
fading model.
space inversion technique. A new identifiability condition is
established and the performance of the method is assessed
through the mean square error and bit error rate. 2. DATA MODEL

1. INTRODUCTION We consider an asynchronous CDMA system with  users

with long spreading sequences of spreading gain  and slot-
Recent code division multiple access systems such as WCDMA ted transmissions of  symbol size. As illustrated in Fig.1,
adopted coherent detection schemes with pilot symbols in user  symbol sequence 
 is scrambled with the long
the reverse link to increase the system capacity. In such spreading code   and transmitted.
systems, channel estimation with a reasonable performance
h (t) fs=1/Tc
over a wide range of fading rates plays a crucial role. h (t)
i,2

y (t) y [n]
User i d (t) i,1 h (t) i Chip i
Channel estimators based on the block fading model do s (t)
i
i
... i,L
MF

not work satisfactorily for fast fading where channel changes C (t) i MUI w(t)
rapidly within a time interval between pilot symbols; Addi-
tional pilot insertion is required with the conventional es- Fig. 1. System model
timation methods. Several channel tracking methods are
proposed based on time multiplexed pilot symbols and in- We assume that the channel of a particular user  con-
terpolation techniques e.g. [1] [8]. These methods utilize sists of  independent multipaths each of which is a ban-
only pilot symbols over multiple slots, which requires pilot dlimited deterministic waveform with bandwidth  the max-
symbols with high SNR and long observation time. Others imum Doppler frequency [3]. Due to fast fading, we let the
consider Wiener filter approaches which requires the knowl- multipath coefficients vary from symbol to symbol while
edge of the Doppler spread and the signal-to-interference remaining constant over one symbol period  . The delay
ratio [4]. profile of multipaths is assumed to be invariant within one
In this paper, we present a new channel estimation tech- slot1 . To simplify the model, we assume that all users are
nique which utilizes multipath structure and tracks fast fad- chiprate synchronized, that is, the delays between paths are
ing channels effectively without the insertion of additional multiples of chip interval  . Specifically, the continuous-
pilot symbols. Using a linear interpolation model, the pro- time time-varying channel impulse response of user  is given
posed method estimates the channel coefficients at selected as
sampling points and tracks the channel for the whole slot. $
"
The proposed method can be implemented with a state-space  # $ 
inversion technique with a comparable amount of complex- 
 !   () +*-,  */.0   (1)
ity with conventional RAKE receiver [5]. %'&

This work was supported by the Office of Naval Research Contract 1 Fast fading results mainly from the phase variation of carrier not from

N00014-00-1-0564. the delay changes

 $

where   is the path coefficient waveform, .0 the delay A2: The noise vector is circularly symmetric complex Gaus-
`~ `
of user  relative to the slot reference. We assume that .1 and sian oyxpz{ v|h
} with possibly unknown } .
 is known2 . The received signal 2   is passed through a
chip waveform matched filter and sampled synchronously Assumption (A1) implies that the receiver knows codes of
at the chiprate. Since the channel is linear and has a finite all users. This assumption is usually valid for the uplink
impulse response, the noiseless chiprate output sample 2 3 465 and full column assumption is satisfied almost surely with
corresponding to user  is expressed as the convolution sum reasonably high spreading factors.
" $
#$ 
27 3 465 .0 3 4 *-, 5  3 465  (2) 3. FAST FADING CHANNEL ESTIMATION
%8&
3.1. Linear interpolation channel model
where .0 3 465 9 3 465  3;:7=?
< > 5 ,  3 @A5 is the discrete-time sym-
bol rate data sequence, and  3 465 the chiprate spreading se- We consider a linear interpolation channel model under the
quence. The received noiseless signal vector B DC due to the deterministic parameter assumption. Among various linear
@ th symbol DC of user  is given in a matrix form as interpolations, we consider the  -sample time domain ap-
proach3 which includes a broad range of interpolation tech-
BECFHGIECKJ8ECKECL (3)
niques such as piecewise linear, polynomial, ideal low pass
where GMEC is the Toeplitz matrix whose first column is interpolations. A similar modeling of areplacements
PSfrag time-varying chan-
made of @ AN+.1 zeros followed by vector O7DCP 3  3 @ AN nel with a truncated basis was proposed in [7].Time
Q Q We assume that the channel at an arbitrary time within
5 SRTRSRU
 3 @ N  5E5WV and additional zeros. Vector J DC is
defined as a slot is a linear combination of channels$ at sample points
Z[ ] &_^ ca bb Z[  acb which are not necessarily the pilot positions. Consider the
[ [ b  ] ^
[ EC b [ & @   b , channel path of user  . The channel DC at @ th symbol
[ ]E` ^ b 
EC ` @  
J8DCYX .. X .. R (4) ’ Žc‹‰ “ŠŒ
ˆSŽk‰‹ŠŒ ˆ ŽE‰‹ ŠŒ 
_ ˆ ŽE‰c‘ ŠŒ
\ ." d \ . d
 ] ^  " €
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€
€ path
 @   ‚tƒ…„ † ‡
DC

The size of B EC is the total number of chips of the entire  - Fig. 2.  sample interpolation model
symbol slot. For user  , the total received noiseless signal is
expressed as $
interval $ as
is modeled $

# U] ^ ] ^ ] ^ Q
DC H”'C &•  & NHlSlTlN–”'C˜— • E —  @ SRSRTRUP (7)
Be f GMDCKJ'DCgTDC
C %8&
where ”8C is the interpolation coefficient for symbol @ and
GM diag J8 & TRSRSRh
J8 _i
 (5) <
sample 4 . Stacking all the multipaths corresponding to the
f same symbol and user, we have
where ih 3  & RTRSRjT 5kV and GM!X 3 GM & TlSlSlUGM 5 which
has a special structuref of sparse block Toeplitz form Z[ ] &_^ ] &^ ] &^ ca bb Z[ acb
f which [ • & • ` • D —
[ lTlSl b [ ”8C & b
is exploited for efficient implementation of matrix inversion [ ]D` ^ ]E` ^ ]E` ^ b [ b
• & • ` lTlSl • D— ” C `
8
[5]. Including all users and noise, we have the multiuser
J EC .. .. .. P™
X ;š›C 
data model as \ ." ." d \ . d
] ^ ]" ^ ] ^ ”8C˜—
# • & • ` lTlSl • D—
B m G  diag J8 & T RSRSRUJ'
M _inNpo (8)
 %'& f
GMqrisNpor (6) where ™  and š›C are defined correspondingly. The inter-
polation coefficient vector š C is determined by the selected
where the overall code matrix Gt 3 G & SRTRSRUG 5, u
q
interpolation method and ™œ is the unknown channel pa-
diag vJ && SRSRTRU
J & J ` & TRSRSRUJ ` SRTRSR
J , and
m o is ad- rameter which is invariant over a slot. We further assume
ditive Gaussian noise. f We also assume f f
them following
the following
A1: The code matrix G is known and has full column
rank. A3: The matrix ™ has full column rank.
2 Rough timing knowledge is enough since we parameterize the delay 3 Frequency domain approach can be also used with  frequency com-
cluster with sufficient number of fingers w ponents for bandlimited channels.
 &
Assumption (A3) implies that the number of multipaths is Let ®¯°Pª
X ² ± with the following row partition
larger than that of the sampling points which is usually valid
Z[ acb
due to the abundance of the multipaths in most mobile chan- [ o³¦ & b
[ b
nels and several sample points are sufficient to track the o³¦ `
channel change within a slot effectively. The design pa- ®¯' .. R (13)
\ . d
rameter  is chosen considering the fading velocity so that
the channel vectors at different sampling points are linearly o³E¦ —
independent almost surely. This suffices to the validity of
the assumption (A3). Multiplying ­ EC by ®¯ from the left gives
Z[ acb Z[ acb
[ o³¦ & b [ ”'C & b
3.2. Blind multiuser channel estimation algorithm [ b [ b
o³¦ ` ” C `
­ DCP DCL
.. ..
We propose a blind channel estimator based on the linear \ . d \ . d
interpolation channel model exploiting the multipath struc- o³D¦ — ” C˜—
ture of channel. We assume that the channel and symbols
are deterministic parameters. Q
Assume ” C ¶µ› 4 SlTlSlU
 and  DC ¶µ for given
<9´ ´
The multiuser signals are separated by decorrelating or @ , which is valid for most interpolation points and mod-
regularized least square front end. The required matrix in- ulation schemes. Notice that multiplying row · , ¸ by ”'C?¹ ,
version is efficiently implemented by an algorithm using the ”8C»º respectively gives the same value ”'C!ºT”8C˜¹7TDC . Taking
state-space technique[5][6]. The output of the decorrelator difference between two rows related to @ th symbol data, a
is given by system of equations are obtained similarly as in [2]
ž G ŸB ½¼
DC¾o¿89| (14)
diag J & & T RSRSRUJ & 
J ` & SRTRSRU
J _isN–¡s
f m f where o+»X 3 o ¦ & SlSlTlUo D¦ — 5 ¦
where _l Ÿ is pseudo inverse. Let ž DC be the subvector of
and
length  corresponding to @ th symbol of user  . Due to the ÆÇÇ É7À cÂÊ Ã'Ë É À ÂcÊ ÃÌ
7 Í Î_Î_Î Î_Î_Î Í Ô ÕÕ
diagonal structure of q , the vector ž EC us given by ÇÇ ± ÕÕ
ÇÇ É7À ÂcÊ Ã'Ë Í É7À ÂcÊ ÃÏ
.
.
.
ÕÕ
ÇÇ ± ÕÕ
ž EC J EC  EC Np¡ DC ÇÇ .
. Í Í .. .
. ÕÕ
.
Q ÇÇ . .
ÕÕ
™ ;š›C  DC N–¡ DC  @ SRTRSRUP (9) ÇÇ .
. .. Í ÕÕ
.
À ÂEÃÅ% Ä ÇÇ É7À cÂÊ Ã'Ë .
Í _Î Î_Î Î_Î_Î Í É7À ÂEÊ ÃhÐ ÕÕ
`¢ Á» ÇÇ ÕÕ
where the colored noise ¡8EC has distribution of z{ v|h
} ECg , Í É7À ÂcÊ ÃÌ É7À ÂcÊ ÃÏ Í _Î Î_Î ± Í
¢ ÇÇ ± ÕÕ
and the covariance EC is the ¤£p submatrix obtained ÇÇ Í .
Í .. .. . ÕÕ
Q ÇÇ
. . . .
ÕÕ
É7À ÂcÊ ÃÌ É7À ÂEÊ ÃhÐ
. .
from the 8* ¥N @ th diagonal block of G Ÿ vG Ÿ _¦ . Í Í _Î Î_Î Í
ÇÇÈ Í Í É7À ÂcÊ ÃÏ É7À ÂcÊ ÃUÑ Í ± Í ÕÕ
±
. . . . . .
3.2.1. Noise free case
.
.
.
.
.
.
.
.
.
.
.
. Ö
Í Í Í Í É7À ÂcÊ Ã²Ò Ð'Ó0Ë É7À ÂEÊ ÃhÐ
±
Consider the noise free case. With the deterministic as-
¼ ¼
sumption on ™œ , the column space of ™œ is obtained by and ­ EC!º 9”
X C?¹ ­ DC and ­ DC?¹ H”
X C!º ­ EC for given D·7j¸›
singular value decomposition pair. Combining all the symbol data gives
§ ¢
 ¨ ¦ P©! X 3 ž  &  ž  ` SRSRTRU ž  5R (10) ½
vo+'H| (15)
f
Then ™œ is given as ½ ½¼ ½¼
where  X 3 ¦ & SlSlTlU ¦ 5¦ .
§
™œ' «ª (11) f
Q `
Proposition
½
1 (Identifiability) The Fp v×* ØÚÙL£Û
where ª  is a invertible {£¬ unknown square matrix from ` Q
matrix  is a matrix with rank  * , i.e. the column
the assumption A(3). Projecting ž DC to the column space of rank is deficient by one. Hence, o+ is the unique null space
§ ½
 , we have the following square system of  and is blindly identifiable up to a scale factor.
­ DC § Q
X ¦ ž DC{9ªh š CKECL @ SRTRSRUjPR (12) Hence, ™œ is identifiable by eq.(11).
 1
10

4. NUMERICAL RESULTS
0
10

mean square error

The performance of the proposed method focused on the −1
10

algorithm itself is evaluated first though the Cramér-Rao

bound and Monte Carlo simulation using a channel gener- −2
10
One pilot symbol
Sinc

ated by the interpolation model in section 3.1 for a single Polynomial

Piecewise linear

user case. To exclude the channel modeling error, the chan- −3

10
0 5 10 15
SNR[dB]
20 25 30

nel is generated according to the interpolation model with

three sample points and sinc interpolation coefficients. The Fig. 4. Mean square error: low-pass channel
scrambling code is generated randomly with spreading fac-
QTÜ QTÜ
tor { and slot length e µ symbols. The number
0
10

of multipaths  is Ý with equal averageÂcà magnitude. Ì The

]váá â Ì áá ^ =
−1
10

signal-to-noise ratio is defined as ÞUßjà ã . The scale 10

−2

ambiguity is resolved using one pilot symbol placed at the

BER
−3
10

left end of slot. As shown in fig. 3, the method shows a good −4

10 One pilot symbol

mean square error performance and almost reaches the the Sinc
Polynomial
Piecewise linear
Known channel

Cramér-Rao bound at medium and high SNR.

−5
10

−6
10
0 5 10 15 20 25 30
SNR[dB]

1
10
CRB
MSE Fig. 5. Bit error rate: low-pass channel
0
10

−1
10

5. CONCLUSION
mean square error

−2
10

−3
10
−4
10
−5
10
0 5 10 15 20
SNR[dB]
Fig. 3. Mean square error:interpolation channel
The performance of the proposed method is evaluated
for a bandlimited waveform channel. Since we assume the
$Eäå
channel is deterministic, the channel waveform is gener-
ated with Jakes’s model[3] with ˜ ¶µRæ7ç and trun-
cated for one slot length. Two equal power asynchronous
users with delay of a half symbol interval are simulated.
Other parameters are the same as in the interpolation chan-
nel case. Fig. 4 shows the MSE performance of the pro-
posed method with various interpolation techniques with
three sample points. The proposed algorithm improves MSE
We propose a new blind channel estimation technique which
effectively tracks fastfading channels in long code CDMA
systems and a new blind identifiability is established. Ex-
25
30 35 40 45 50
ploiting the multipath structure and interpolation model, the
proposed method shows a significant improvement over the
channel estimation with block fading model without inser-
tion of additional pilot symbols.
6. REFERENCES
[1] H.Andoh, M.Sawahashi, and F.Adachi, “Channel Estimation Filter using Time-
Multiplexed Pilot Channel for Coherent RAKE Combining in DS/CDMA Mo-
bile Radio,” IEICE Trans. Communications, E81-B(7):1517-1526, July 1998.
[2] D. Gesbert, P. Duhamel, and S. Mayrargue, “On-line blind multichannel equal-
ization based on mutually referenced filters ,” IEEE Trans. Signal Processing ,
45(9):2307-2317, September 1997.
[3] W.Jakes Jr., Microwave Mobile Cummunications, Wiley, New Yor, 1974.
[4] B. Lindoff, C. Ostberg, and H. Eriksson, “Channel estimation for the W-CDMA
system, performance and robustness analysis from a terminal perspective ,”
In Proc. IEEE Vehicular Technology Conference, volume 2, pages 1565-1569,
1999.

performance much over the estimation using the pilot sym- [5] L.Tong,A.V.der Veen, P.Dewilde, and Y.Sung, “Blind decorrelating rake receiver
bol and the block fading model. However, the proposed for long code WCDMA,” Tech. Rep. ACSP-02-01, Cornell University, Feb. 2002.

method also shows a performance floor at high SNR due [6] P.Dewilde and A. van der Veen, Time-Varying Systems and Computations,
to the imperfect modeling of the actual channel. Figure 5 Kluwer Academic Publishers, Dordrecht, The Netherlands, 1998.

shows the average bit error rate performance of a whitened
RAKE receiver with the estimated channel. The increase of
BER with respect to SNR for the estimation with block fad-
ing model shows that the additive noise works beneficially
for detection since the estimation gives adverse values due
to lack of tracking capability.
[7] M. K. Tsatsanis and G. B. Giannakis, “Subspace Methods for Blind Estimation
of Time-Varying FIR Channels,” IEEE Trans. Signal Processing , 45(12):3084-
3093, December 1997.
[8] Hai-Wei Wang and Che-Ho Wei, “Pilot-symbol assisted channel estimation for
coherent DS-CDMA communications,” In Proc. IEEE International Symposium
on Circuits and Systems , volume 3, papes 383-386, 2000.