CORDIC based QRD-RLS Adaptive Equalizer

for CDMA Systems

Tim Zhong Mingqian, AS Madhukumar and Francois Chin

Institute for Communications Research
20 Science Park Road, Singapore 117674
Email:{rtpunq.madhu.chinfrancois}@icr.a-star.edu.sg

Abstract The conventional RAKE receiver employed in CDMA system can not satisfy
the desired demand when the propagation channel possesses a considerable
number of paths and a deep fading that causes serious inter-user and inter-
symbol interferences. This paper proposes a novel means of adaptive equalizer
based on QRD-RLS algorithm to substitute the traditional RAKE receiver.
Regarding the computational complexity of the RLS, the well-known
CORDIC algorithm has been exploited and plays a key role in the hardware
implementation of the new approach. The proposed RLS structure is simulated
extensively under different channel parameters and performance is compared
against conventional RAKE structure.

Key Words QRD-RLS, CORDIC, CDMA, Adaptive Equalizer

1. Introduction
The time-varying multipath propagation and the multiuser interference
are the two important issues that limit the capacity and performance of a
wireless communication system based on the code division multiple access
(COMA) technology. The commonly proposed schemes to deal with these
two factors are to use transmitter power control, diversity technique and/or
error control coding [1][2]. An alternative approach to combat the distortion
brought by the channel characteristic and the multiuser interference is to
replace the traditional RAKE receiver with an efficient equalizer using the
adaptive algorithms such as Least-Mean-Square (LMS) and Recursive Least-
Squares (RLS).

The original version of this chapter was revised: The copyright line was incorrect. This has been
corrected. The Erratum to this chapter is available at DOI: 10.1007/978-0-387-35618-1_37
C. G. Omidyar (ed.), Mobile and Wireless Communications
© IFIP International Federation for Information Processing 2003
250 Tim Zhong Mingqian, AS Madhukumar and Francois Chin

Regarding the time-varying Rayleigh fading channel which is typical

for a practical mobile radio system, due to the limitation of number of
RAKE fingers, it's hard to reach a satisfied system performance when the
number of channel path increases and the channel coefficient changes
quickly from time to time. In addition, the system complexity increases
dramatically as more RAKE fingers have been appended and a channel
estimation module must be included at the receiver side. Time Domain
Equalizers based on LMS or RLS can be a solution to solve the issues
related to multipath interference and errors due to channel estimation.
The RLS algorithm takes into account all the information that extend
back to the initialization and updates the estimation of the tap-weight vector
upon the arrival of new data, thus it is preferred over the LMS algorithm due
to its superior convergence properties [3]. The orthogonal triangularization
of the input matrix via QR-decomposition (QRD) has played a crucial role in
RLS filtering. Such triangularization process can be realized through a series
of Givens rotations, which is commonly employed to do the QR updating on
a sample-by-sample basis [7]. Figure 1 shows the block diagram of the
proposed RLS based adaptive equalizer structure for a CDMA receiver.

r'· ..·..······· .. ········································............................._.......................................................................................................!

i !
! Pilot !
i Oespreading RLS Weight i
! Generator !
i i
i i
I
·! d
Recel\/e !i
! i

Oesnreading FTR i
\..............................................................................................................................................................................................1

Figure]: Structure ofRLS Adaptive Equalizer in aCOMA system

Armed with the fast developing VLSI technology, systolic array has
become practically feasible when realizing the QRD-RLS in hardware
implementation. The CORDIC (Coordinate Rotation Digital Computer)
algorithm [4] has been introduced to perform the two-dimension vector
rotation instead of the conventional Givens rotations. The main idea
underlying this algorithm is to do phase shifting through a series of "micro-
rotations" using a fIxed set of elementary rotation angles. Through a proper
choice of the elementary angles all computations can be implemented
efficiently in VLSI using a sequence of shift and add/subtract operations.
Generally, a look-up-table holding the elementary rotation angles is set up in
advance to perform the phase shifting replacing the trigonometric functions
CORDIC based QRD-RLS Adaptive Equalizer for CDMA systems 251

exploited in Givens rotations, which are not applicable in hardware

realization.
This paper discusses the design and development of a CORDIC based
QRD-RLS channel equalizer in a CDMA system and compares its
performance with conventional RAKE structure.
This paper is organized as follows, section 2 briefly introduces the
mathematical principles of the QRD-RLS algorithm and discusses its
limitation due to the high computational complexity. The CORDIC
algorithm will be depicted in section 3 and its implementation in hardware
realization will also be included. Section 4 discusses the simulation studies.

2. QRD-RLS Algorithm

The basic idea underlying the RLS algorithm can be described in two
processes:
1. We calculate the output of a transversal filter produced by a set of tap
inputs and then try to obtain the error estimation by comparing the
output with the desired response.
2. We find a method to minimize this error by adjust the tap weight of
the transversal filter so that we can approach the desired response at
the output.
Let d(n) be the desired response vector at time n, u(i) the input signal
vector at time i and w(i) the tap weight we are looking for, and then the
goal is to minimize the cost function:

E= LA
n"1d(i) - w
n-, H (i)u(i)12 (1)
i=l
where A is the forgetting factor used to ensure that the effect brought
by past signals is reduced or "forgotten".
Eq. (1) can be converted to a matrix format as follows:
e= = IIA(n)d(n) - A(n)A(n)w(nf (2)
where the symbol 11-112 stands for the squared Euclidean norm and:

A =dlag
· A ,A , ... ,1) (3)

d H (n) =[d(1),d(2), ... ,d(n)] (4)

A H (n) = [u(I),u(2), ... ,u(n)] (5)
Regarding Eq. (2), since the multiplication by a unitary matrix does not
change the norm of one matrix, we apply the QRD to transform the weighted
input signal matrix A(n)A(n) into an upper triangular matrix R(n) , the first
part to the right of Eq. (2) has been transformed simultaneously into two
auxiliary matrices denoted by P(n) and V{n):
252 Tim Zhong Mingqian, AS Madhukumar and Francois Chin

Q(n)e(n) = [ p(n)] - [R(n)W(n)] (6)

V(n) 0

whereQ(n) is a unitary matrix.

It's easy to observe from Eq. (6) that the cost functione approaches its
minimum value IIV(n)11 2 when the following equation satisfies
R(n)w(n) = pen) (7)
The process can be proceeded recursively by a series of complex
Givens rotations which give rise to the unitary update transformation matrix
Q(n) :
Q(n) =J M (n)"'J 2 (n)J 1 (n) (8)
here J k (n) is the single Givens rotations transformation matrix used to
eliminate the corresponding element in the signal vector that enters the
system at time n. With all these transformation matrices, the non-zero vector
i on the bottom will be transformed into a zero vector:

JM(n) . . 1 (9)

J k (n) is given as:

1
o
cosO 0 0 sin· 0 k
Jk(n) = 0 0 0 (10)
0
0 0 0
-sinO 0 0 cosO n
k n

It can be seen from Eq. (7)-(9), as the data stream enters the equalizer
row by row, it is annihilated to zero while the matrix R(n) and pen) are
updated accordingly.
Therefore, it's easy to obtain the desired weight value through Eq. (7).
3. CORDIC ALGORITHM AND ITS IMPLEMENTATION
3.1 CORDIC Algorithm

Considering Eq.(lO), trigonometric functions have been employed here

to perform orthogonal rotations. But such functions can't be applied in
practical hardware implementation as far as fixed-point value is concerned.
CORDIC based QRD-RLS Adaptive Equalizer for CDMA systems 253

The CORDIC algorithm will be introduced to solve such problem since it

utilizes a look-up-table to do the rotation through series of sub-rotations.
There are two modes of CORDIC algorithm: the "vectoring" mode,
which is used to determine the phase and magnitude of the input vector and
the "rotation" mode, which actually performs the rotation of the input vector.
In order to execute the Givens rotations, "vectoring" mode is applied ftrstly
to determine the angle to be rotated followed by the "rotation" mode rotating
the vector through a set of sub-rotation for certain loops.
Consider a two-dimensional vector v represented by v = x + jy in the
complex plane. If the vector is rotated by an angle 0, the new vector is
presented as v = ve i8 . The angle 0 can be expanded into a set of elementary
angles OJ with pseudo-digits qj E {-1,+1} and angle expansion error Zn-l' such
that
n-l
0= Lqj ·OJ +Zn-l (11)
j=-1

here the sub-rotation angles OJ take on the following values

1! 12 (i = -1)
{ (12)
OJ = arctan(2 -j) (i =0,1,···, n -1)
The pseudo-digits qj are used to determine the direction of rotation
according to the present value of Z/. Under the vectoring mode, Zj is set to
zero during the initialization and q j is decided by the position of the current
vector Vj in the updating process, once the original vector has been rotated to
the abscissa, the phase can be obtained by the value of Zn • Similarly, under
the rotation mode, Zj is initialized by the angle desired for rotation and q j is
decided by the sign of current Zj in the updating process, once the loop ends,
the vector v has been rotated by 0 while Zj approaches zero.

3.2 Systolic array

Another fascinating feature of CORDIC is that the determination of the

rotation angle takes the same number of clock cycles as does the actual
rotation. This permits a systolic array of QRD-RLS mtering. CORDIC
algorithm has been exploited here to realize the function of every small cell
in the systolic array as shown in Figure 2. [7] There are two kinds of cells in
the ftgure differentiated by their shapes. The cell is always deftned as
CORDIC processor element (CPE), with the round one working under
''vectoring mode" and the square one working under "rotation mode".
254 Tim Zhong Mingqian, AS Madhukumar and Francois Chin

u(3) u(2) u(l) d(3)

u(2) u(l) o d(2)
u(l) o o d(l)

a w,la
Figure 2: Structure of systolic array

4. Simulation results

The proposed system is extensively simulated for different channel

parameters and the performance is compared against conventional RAKE
receiver. The CDMA based transceiver structure used in the simulation is
compliant to 3GPP standards [5]: using orthogonal variable spreading
(OVSF) as channelization codes and a segment of Gold sequence as
scrambling code. Pulse shaping filter has been included in the system to
smooth the chip signal and discard higher frequency components. Two
receiver side antennas have been utilized to improve the performance and a
two-branch structure is employed to reduce the taps, which is set as 4,
needed for the RLS equalizer. The channel model used in the downlink
simulation system is a Rayleigh fading channel with 10 paths as denoted by
L. The maximum delayspread is 20 chips. The Doppler frequency denoted
by Fd is relatively low in the simulation thus the time-varying channel
changes slowly. The processing gain has been set to 16 and the number of
user has been set as the quarter respectively the half of the system load. The
system supports a total bandwidth of 5MHz, a chip rate of 3.84Mcps and a
data rate of 144kbps. The simulation assumes perfect synchronization and no
power control matters are employed.
Figure 3 shows the simulation results of the proposed system. The solid
curve in every figure denotes the BER against SNR per bit through the
CORDIC based QRD-RLS Adaptive Equalizer for CDMA systems 255

Rayleigh fading channel when the number of user is set to single (single-user
bound) [6] , serving as a preference for comparison. It's obvious that the
RLS equalizer outperforms the conventional RAKE by nearly 4 dB at the
BER level of 10.2 in the case of quarter system load. It can be also observed
easily that the system performance deteriorates as the number of user
increases. However, the MUI caused by the increasing of the user number
doesn't affect the equalizer as heavily as for the RAKE receiver. For both
two cases, a noticeable error bottom can be observed for the RAKE receiver
case but the performance for the RLS equalizer improves with EblNo.

10'1 ::.::::;:;:;:
. -. --------- .
*
--:c:::.:::::::::.:.:::;::::
-" :: ---... ---------- -. --- -.. --.... -_...:::::E:
::::::::::t:::::::::::.:::::.:::.:: -_.... -

· ,

'0· H H
.... -.-.-.------.---- ··. --.------.----- ... ---.- . .. --- ........... -.
• • • • • • • _ • • ___ • • • • • • _.L • • • __ • • • • • __ • • • __ ____ L • • • __ • • • • • • • • • • • • •

.... . __ ....... - ... ---:-._------ .. - ----- .. ---:-, ... --- ........... -..
· .

(a) quarter system load

::: : :: :::: :: =: _:: :::. .:: :::::

: ::::: :: ::: :t.::: : :: ...:::: ::: ::::: .
...... - -- -----. -..... , .. _.............. ........
. .. ...... ........ .
"

(b) half system load

Figure 3: Performance of the proposed method in a CDMA system
256 Tim Zhong Mingqian, AS Madhukumar and Francois Chin

s. Conclusion
This paper has focused on a novel approach of signal receiving in
CDMA system: replace the RAKE with an adaptive equalizer based on
QRD-RLS algorithm due to its perfect convergence property and robustness
against multipath interference. Considering the computational complexities.
which hinder it from practical realization. the CORDIC algorithm has been
introduced to make it feasible for hardware implementation. The
complicated matrix computation has been carried out by a systolic array
consisted of CPE cells and driven by a certain clock. It can be observed from
the simulation results that the adaptive equalizer outperforms the RAKE to
an ideal extent in a multipath fading channel and the system complexity has
been reduced since the channel estimation module can be omitted.

Reference
[1] Ozan K. Tonguz, and Melanie M. Wang, "Cellular CDMA Networks Impaired by
Rayleigh Fading: System Performance with Power Control", IEEE Trans. on Veh. Tech.,
Vol. 43, No.3, August 1994.
[2] 1. Boutros and E. Viterbo, "Signal Space Diversity: A Power- and Bandwidth-Efficient
Diversity Technique for the Rayleigh Fading Channel, " IEEE Transactions on
Information Theory, vol. IT-44, pp. 1453-- 1467, July 1998.
[3] E. Eweda. "Comparison of RLS, LMS, and sign algorithms for tracking randomly time-
varying channels." IEEE Trans. on Signal Processing, vol. 42, no. 11, pp. 2937-2944,
Nov. 1994.
[4] Y. H. Hu, "CORDie-based VLSI architecture for digital signal processing." IEEE Signal
Processing Magazine, 9 (3): 16-35, 1992.
[5] 3GPP TS23.101 V3.5.0: "UE Transmission and reception (FDD)", Release 1999
[6] Proakis, J.G: "Digital Communications", Third Edition, McGraw-Hill Inc.,1995
[7] Simon Haykin: "Adaptive Filter Theory", Third Edition, Prentice-Hall Inc.,1996