LMS Algorithm Based Error Correction Technique in

In-Phase and Quadrature Channel Signal Processing

Amritakar Mandal #1, B. K. Kaushik*2, Brijesh Kumar*3,K. C. Tyagi@4, , R. P. Agarwal$5
#

Radar Laboratory, PMG, New Delhi,

New Delhi, INDIA
*
Department of Electronics and Computer Engineering, Indian Institute of Technology
Roorkee, INDIA
1amritkar2k@gmail.com
3bk228dpt@iitr.ernet.in
@

Electronics and Communication Engineering,

Dev Bhoomi Institute of Technology,
Dehradun, INDIA
$
Department of Computer Science and Engineering,
Shobhit University,Meerut, INDIA
2

bkk23fec@iitr.ernet.in
4
prajanag@gmail.com

oscillator are negligible as both the channel uses same

coherent oscillator. This paper presents a feasible design of a
LMS algorithm based adaptive filter using CORDIC for the
correction of I/Q channel mismatch to increase Signal to Noise
ratio (SNR) significantly. Reformulated trigonometric LMS
algorithm has been used for the realization of adaptive FIR
filter [3]. The sine and cosine terms generated are used in
filtering and weight updating purpose. The sine and cosine
terms can be calculated using polynomial approximation, e.g.
Taylor series. But it requires a considerable amount of
hardware space on the silicon substrate. Interpolation method
using table look-up may be the other solution. But it also
requires large number of gates and ROM memory. The
CORDIC offers the opportunity to calculate the desired errors
in a simple and efficient way due to its good convergence
property as shown in Fig. 2. This paper first describes the I/Q
channel demodulation, errors in section II; I/Q compensation
technique in section III; and finally algorithms and adaptive
filter architecture design in section IV and V respectively.

Abstract-Every communication receiver that uses in-phase and

quadrature channel signal processing technique encounters
problems related to matching of gain and phase in both the
channels. The gain and phase imbalances occur between Low
Pass Filter and Local Oscillator used in both the channels and as
a result the performance of the receivers and the quality of the
received signals are degraded. The imbalances produced may
cause insufficient attenuation in image frequency band leading to
interference. The problem needs to be compensated. This paper
represents the design and implementation of Coordinate Rotation
Digital Computer (CORDIC) algorithm based Adaptive FIR
filter using Least Mean Square (LMS) algorithm to give a
solution for correcting I/Q imbalances. The design of CORDIC in
the vector rotation mode results in high system throughput due to
its pipelined architecture where latency is reduced in each of the
pipelined stage and finds its appropriate application in real time
signal processing. The good convergence of CORDIC in the given
LMS algorithm based adaptive filter facilitates in easy
calculation of new filter weights.
Keywords I/Q imbalance, Low Pass Filter, Local Oscillator,
CORDIC, FIR filter, LMS algorithm.

I/Q CHANNEL DEMODULATION AND ERRORS

As per the Eulers theorem, vector sum of cosines
component is completely real whereas the spectrum of sine
component is totally imaginary. If the cosine and sine
components are combined, the resultant spectrum becomes
one sided with direction of rotation (positive or negative
frequency) and with known real (cosine) and imaginary (sine)
components [4].
II.

I. INTRODUCTION
A quadrate receiver consists of two distinct channels to
form in-phase and quadrate components of the received signal.
Each channel consists of a mixer, low pass filter, A/D
converter and digital signal processor unit. Coherent oscillator
(COHO) is common to both the channel with 90o phase shift.
The mismatch between low pass filters can severely limit the
performance of the adaptive canceller and the matched filters
and hence reduces the quality of the signal [1]. If noisy
oscillators are used for both channels, then I/Q error may be
generated because of gain and phase imbalances in mixer
output [2]. But the chances of generation of error due to

978-1-4577-0240-2/11/$26.00 2011 IEEE

cos c t + j sin c t = e jct

The process of recovering both real and imaginary signal
component is known as I/Q demodulation. I stand for in-phase
channel which processes cosine (real) components. Q stands
for quadrate channel which processes sine (imaginary)

226

component. The input of I/Q channel is Intermediate

Frequency (IF). If the carrier frequency of IF is fc with a time
varying amplitude a(t ) and time varying phase (t ) , then input

improvement in Mean Squared Error (MSE) with respect to

iterations, as shown in Fig. 5, and thus mismatch is neutralized.
LMS ALGORITHM AND CORDIC
The LMS algorithm [3] is one of the simplest well known
adaptive algorithms and with the help of this algorithm a
robust adaptive filter can be devised. Let the FIR filter is
linear discrete time filter. x(n) And d(n) is input and training
IV.

signal s (t ) will be:

S (t ) = a(t ) cos[(2f c t + (t )] .
In I channel, the IF signal is multiplied by reference carrier
frequency produced by crystal oscillator at zero phase
reference. The output of the I channel mixer is, I(t) , given by :
I (t ) = a(t ) cos[(2f c t + (t )]. cos(2f c t )

sequence of the FIR filter. e(n) is the error between desired

output and the obtained output. Let w be an N-tap filter

= a(t ) cos[ (t )] + a(t ) cos[4f c t + (t )]

The first term is the average value (DC) of the product and
represents cosine of the signal phase and amplitude. The
second term with high frequency component is suppressed by
Low Pass Filter (LPF). So the output of the I channel is
I (t ) = a(t ) cos[ (t )] .
Similarly, the Q channel output can be derived. The LPF
output at Q channel is:
Q(t ) = a(t ) sin[ (t )] .
Thus, I and Q channels are necessary for any angle
modulated signals because the two side bands of the RF
spectrum contain different information and there is a strong
possibility of corruption of signals if they overlap each other
to create Inter Symbol Interference (ISI).
The receiver possesses a Coherent Oscillator which is at the
same phase with the transmitter frequency. The low pass
filters in both the paths of the receiver must have identical
characteristics as any mismatch in their characteristics will
lead to I/Q errors. In practice, the errors are unavoidable and
the outcome is insufficient attenuation of the image frequency
[1,4]. Due to this phenomenon, an error compensation
technique is required.

coefficient

vector,

then,

w = [w0 ,w1,........,wN 1 ] T

and x(n) = [x(n),x(n 1 ),.......,x(n N + 1 )] T . The mean square

error is = E[e 2(n)] = E[e(n)eT (n)] . The minimum mean
square error will give the optimum filter weight. Using the
steepest descent search algorithm, an optimum filter weight in
the form of wopt = R 1 p is obtained. The i -th iteration using
steepest descent search algorithm is given by the expression:
w(i +1 ) = w(i) + [p Rw] . Replacing R and p by

x(n)x T (n) and x(n)d(n) respectively, LMS algorithm is

obtained in the form: w (n+1 ) = w (n) + x (n) e ( n) . Here
is called the step size.
The Delayed LMS (DLMS) is equivalent to the LMS
algorithm but the gradient is evaluated at some previous
iteration, say L iterations before present iteration. Then the
updated equation of DLMS algorithm can be expressed
as, w (n+1 ) =w (n) + x (n-L) e (n-L) .
Let the tap weight wk in the equation wk = Ak sin k be
such that each wk , Ak wk + Ak , maps uniquely to a

, + ] .Then using LMS algorithm,

2
2
recursively updated Trigonometric LMS (TLMS) algorithm
can be formulated as (n +1 ) = (n) + (n)x(n)e(n) where

k in the interval [

ERROR COMPENSATION TECHNIQUE IN I/Q CHANNELS

In the receiver model same coherent oscillator is used for
both in-phase and quadrate channel. So error due to mismatch
in oscillator phase or gain is negligible. The total error may be
produced by the LPFs in both channels. The basic principle is
to make the input signal through the filters identical and then
compensate for the error between the two outputs. The
compensation filter is an adaptive filter which uses LMS
algorithm to reduce the error.The calibration of the signal can
be performed with random input signal as shown in Fig. 3. To
perform the calibration in the circuit, both the mixers are
III.

N 1

e(n) = d(n) Ak sin k (n) x(n k) .

The

Trigonometric

k =0

LMS algorithm is suitable for CORDIC based realization [3],

since the two quantities, such as sin k x(n k) and

cos k x(n k) in, k = 0, 1, ., N-1, required for filtering and

updatation 0 i M 1 , denote a sequence of 1s that
determine the direction of each elementary rotation.
When M is the total number of elementary rotation angles,
i-th angle i is given by:

isolated. This means both switches SW1 and SW2 are open. A
known calibrated signal source is then incorporated to the
input of LPFs . Let the quadrate channel low pass filter is
mismatched. So it is obvious that I/Q error will persist. To
compensate the mismatch between the channels, adaptive
algorithm based filter is incorporated. Now the circuit is
needed to be trained so that the two signals at the filter output
would become identical.
After training, the weights are calculated and the transfer
function of new filter is obtained and added to the network to
neutralize the error [1]. It is observed that the errors have been
decreased significantly as the LMS filter output showed

m ,i =

1
m

tan

[ m2

s ( m ,i )

] =

2 s ( 0 ,i )
1 s ( 1 , i )
tan 2
tanh 1 2 s ( 1, i )

Where m=0, 1 and 1 correspond to the rotation operation in

linear, circular, and hyperbolic coordinate system respectively.
For a given value of , the CORDIC iteration is given

227

by: x i + 1

1
y i + 1 = 2 i
i

i x
i
y
1
i

i 2

range will be [ 2 , 2] . The difference in binary

representation between these two angles is one bit.
Overflow arises when a rotational angle crosses a positive
right angle to a negative one. To avoid overflow, an overflow
control is added. It checks for the sign of the operands
involved in addition or subtraction and the result of the
operation. If an overflow is produced, the result keeps its last
sign without affecting the final result. In the overflow control,
the sign of zi determines whether addition or subtraction is to
be performed.
of the k-th coefficient can be computed by using single
CORDIC processor in adaptive filter as shown in Fig. 4.
The theory of CORDIC computation is to decompose the
desired rotation angle into the weighted sum of a set of
predefined elementary rotation angles [5, 6]. Each of them can
be accomplished with simple shift-add operation for a desired
rotational angle . It can be represented for M iterations of an
input vector (x,y)T setting initial conditions: x0=x, y0=y, and

and z i +1 = z i i i ,

where i = tan1 2i .
In case of counter clockwise rotation of a vector, the
recursively
updated
equations
are
generated
as: xi +1 = xi cos i i yi sin i i
and yi +1 = y i cos i i + xi sin i i .
The above equations can be simplified in the following
form: xi +1 = cos i i(xi yi tan i i )

yi +1 = cos i i(yi + xi tan i i ) .

i
Here, tan i is restricted to 2 . Thus, multiplication is
transformed to an arithmetic right shift. Since cosine is an
even function, therefore cos( ) = cos( ) The iterative
equation can be reduced toi
x i +1 = K i (xi y i i 2
) and y i +1 = K i (y i + xi i 2 i ) .

where K i = cos arctan 2 i = (1 + 2 2i ) is known as

gain factor for each iteration. If M iterations are performed,

then scale factor, K, is defined as the multiplication of every
Ki. So,
M 1
M 1
K = K i = (1 + 2 2i ) .
i=0

M 1

M 1

i =0

i =0

z0= as z f = i i . If zf=0 holds, then = i i ,

i.e. the total accumulated rotation angle is equal to . i

i=0

The elementary functions sine and cosine can be computed

using the rotation mode of the CORDIC algorithm if the initial
vector starts at ( K ,0) with unit length [7].
PIPELINED ARCHITECTURE OF CORDIC
In Pipelined CORDIC architecture, a number of rotational
modules are incorporated and each module is responsible for
one elementary rotation [8-10]. The modules are cascaded
through intermediate latches (Fig. 1). Every stage within the
pipelined CORDIC architecture, only adder/subtraction is used.
The shift operations are hardwired using permanent oblique
bus connections to perform multiplications by 2-i. The precomputed values of i-th iteration angle i require at each
module are stored at a ROM memory location as shown in
Table-I. The delay is adjusted by using proper bit-length in the
shift register. Since no sign detection is needed to force zf=0,
the carry save adders are well suited in this architecture. The
use of these adders reduces the stage delay significantly. With
the pipelining architecture, the propagation delay of the
multiplier is the total delay of a single adder. So ultimately the
throughput of the architecture is increased by many folds as
the throughput is given by: 1/(delay due to a single adder).
If an iterative implementation of the CORDIC is used, the
processor would take several clock cycles to give output for a
given input. But in the pipelined architecture, each pipeline
stage takes exactly one clock cycle to pass one output [3].
The most recurrent problem for a CORDIC implementation
is overflow. Since the first tangent value is 20=1 then rotation
V.

228

Figure 1. Pipelined CORDIC Architecture

Figure 2. Convergence of CORDIC w.r.t number of iterations

TABLE I.

PRE-COMPUTED ANGLES

2 i = tan i

i = arctan( 2 i )

i in radian

0
1
2
3
4
5
..

1
0.5
0.25
0.125
0.0625
0.03125
..

45o
26.565o
14.063o
7.125o
3.576o
1.7876o
..

0.7854
0.4636
0.2450
0.1244
0.0624
0.0312
.

proposed design. The property of good convergence of

CORDIC is efficiently used in adaptation and filtering with
LMS algorithm.

Figure 5. Convergence of error at the output of adaptive filter with step size

REFERENCES
[1]

Isis Mikhael, Wasfy. B, Pravinkumar, Adaptive IQ Mismatch

Cancellation For Quadrature IF Receiver, Proceedings of 4th WSEAS
International Conference on Signal Processing, Robotics and
Automation, 2005.
[2] Fred Harris, Digital filter equalization of analog gain and phase
mismatch in I-Q receivers, 5th IEEE International Conference on
Universal Personal Communications, vol. 2, pp. 793-796, 1996.
[3] M. Chakraborty, A. S. Dhar and Moon Ho Lee, A Trigonometric
Formulation of the LMS Algorithm for Realisation of Pipelined
CORDIC, IEEE Trans. Circuits and Systems, vol. 52, no. 9, pp. 530534, Sep.2005.
[4] Mikko Valkama, M. Renfors, V. Koivunen, Advanced Methods For I/Q
Imbalance Compensation in Communication Receivers. IEEE
Transactions on Signal Processing, vol. 49, No. 10, pp. 2335-2344, Oct.
2001.
[5] J.E. Volder. "The CORDIC Trigonometric Computing Technique". IRE
Transactions on Electronic Computing, vol EC-8, pp 330-334, Sept.
1959.
[6] Y.H. Hu. "CORDIC-Based VLSI Architectures for Digital Signal
Processing" IEEE Signal Processing Magazine, Vol. 9, No. 3, pp. 1635, 1992.
[7] Andraka R.A., "Survey of CORDIC Algorithms for FPGA Based
Computers, Proceedings of the 1998 ACM/SIGDA 6th International
Symposium on FPGAs, pp 191-200, Monterey, California, Feb.22-24,
1998.
[8] A. Mandal, K.C. Tyagi, B.K. Kaushik, VLSI Architecture Design and
Implementation for Application Specific CORDIC Processor, 2nd
IEEE International Conference on Advances in Recent Technologies in
Communicaion and Computing (ARTCom), pp. 191-193, Oct 16-17,
2010.
[9] S. Wang, V. Piuri, E. E. Swartzlander. Jr., "Granularly-pipelined
CORDIC processors for sine and cosine generators, IEEE International
Conference on Acoustics, Speech, and Signal Processing ICASSP, vol.
6, pp. 3298-3301, 1996.
[10] Y.H. Hu. "The Quantization Effects of the CORDIC Algorithm" IEEE
Trans. Signal Processing, vol. 40, No. 4, pp. 834-844, Apr. 1992.

Figure 3. Calibration techniques for filter mismatch correction

Figure 4. CORDIC based adaptive filter

VI. CONCLUSION
This paper presented the error correction technique in inphase and quadrature channel with a CORDIC based adaptive
FIR filter using LMS algorithm. The use of pipelined
CORDIC computational architecture makes implementation of
this kind of adaptive filter easier. Numbers of micro-rotations
have been adjusted so as to achieve better convergence and
speed of operation while minimizing delay in error
computation. CORDIC algorithm is used to achieve high
degree of computational accuracy facilitating better
application in I/Q channel calibration. The inherent issue of
CORDIC i.e. overflow is quite appropriately resolved using

229