You are on page 1of 4

FIXED AND FLOATING POINT ERROR ANALYSIS OF

QRD-RLS AND STAR-RLS ADAPTIVE FILTERS *


Kalavai J. Raghunath and Keshab K. Parhi
Department of Electrical Engineering
University of Minnesota, Minneapolis, MN-55455

ABSTRACT can be applied with little increase in hardware complexity.


A pipelined version of the algorithm has also been devel-
The QR decomposition based recursive leastsquares oped and this is referred to as the pipelined STAR-RLS
(RLS) adaptive filtering (referred to as QRD-RLS) algo- or PSTAR-RLS [6]. The PSTAR-RLS algorithm can be
rithm is suitoble for VLSf implementation since it has good pipelined to operate a t very high speed. The STAR-RLS
numerical properties and can be mapped to a systolic array. algorithm [4] and the pipelined STAR-RLS algorithm (or
Recently, a new fine-grain pipelinable STAR-RLS algo- PSTAR-RLS) [6] have lower complexity, half the inter-cell
rithm was developed based on scaled tangent rotation. The communication and are square-root free, as compared with
pipelined STAR-RLS algorithm, referred to as PSTAR- the QRD-RLS algorithm.
RLS, is useful for high-speed applications. The stability of
The accumulation of quantization noise is an important
QRD-RLS, STAR-RLS and PSTAR-RLS has been proved
concern in the implementation of adaptive digital filters.
but the performance of these algorithms in finite-precision
The cost of implementation is strongly dependent on the
arithmetic has not yet been analyzed. The aim of this paper
number of bits used to represent the signals in the fil-
is to determine ezpressions for the degradation in the per-
ter. Hence one would like to use fewer number of bits
formance of these algorithms due to finite-precision. By for words without affecting the performance of the algo-
ezploiting the steady-state properties of these algorithms,
rithm. This motivates the finite precision error analysis
simple closed-form ezpressions are obtained which depend
of QRD-RLS and STAR-RLS algorithms. The stability
only on known parameters. Since floating-point o r fized-
of QRD-RLS, STAR-RLS and PSTAR-RLS was recently
point arithmetic representations may be used in practice,
proved in [7],[6]. The effect of finite-precision on recursive
both representations are considered in this paper. The
least-squares (RLS) adaptive filters has been considered
results show that the PSTAR-RLS and STAR-RLS algo-
before, for example in [8]. The RLS algorithms consid-
rithms perform better than the QRD-RLS especially in a
ered in these papers refer to the original algorithm which
floating-point representation. The theoretical ezpressions
does not use any decomposition techniques. A theoretical
are found to be in good agreement with the simulation re-
finite-precision analysis of the QRD-RLS algorithm has,
sults.
however, not been done before. The only paper on this
issue, to the best of our knowledge, is [9]. Here the error
I INTRODUCTION propagation of quantization noise is investigated assuming
a single error has occurred. The aim of this paper is to
Recursive least squares (RLS) based adaptive filters are make a general finite-precision analysis of the QRD-RLS,
used in applications such as channel equalization, voice- STAR-RLS and PSTAR-RLS algorithms, while still keep
band modems, digital mobile radio, beamforming, speech ing the final expressions simple. To do this we exploit the
and image processing. The QRD-RLS algorithm [1],[2],[3] steady-state properties of the algorithms, determined in
is the most promising algorithm since it is known to have [101,[4Il [GI.
very good numerical properties and can be mapped to a In a practical implementation either floating-point or
systolic array. The speed (or sample rate) of the QRD-RLS fixed-point arithmetic may be used. The errors resulting
algorithm is limited by the recursive equations in its cells. from these representations have different statistics. K e e p
A new STAR-RLS algorithm [4] was recently developed ing these things in view, we have carried out the analy-
which can be used for high-speed applications, for exam- sis separately for the fixed-point and floating-point arith-
ple, in communications, magnetic recording, image pro- metic.
cessing etc. This algorithm uses scaled tangent rotations
(STAR) instead of the Givens rotations which are normally We need to define a figure of merit for finite precision im-
used. These rotations are so designed that look-ahead [5] plementation, which can be used to decide the wordlength
and also to compare the three algorithms. The estimation
*This research was supported by the office of Naval Research error can be directly obtained as the final output from the
under contract number N00014-91-J-1008. systolic array for the QRD-RLS, STAR-RLS and PSTAR-

m-81
$3.00 0 1994 I E E
0-7803-1775-0/94
RLS algorithms. Hence, we use the deviation in the esti- due to finite-precision quantization, and find the average
mation error due to finite-precision as the figure of merit. value of the same. The errors are propagated down the
Thus, we find expressions for the squared average devia- systolic array for all three algorithms and the deviation in
tion in the estimation error for the three algorithms for the estimation error is found. The intermediate results are
fixed-point and floating-point arithmetic. not shown here. The expected values of the square of the
deviation in the estimation error for STAR-RLS, PSTAR-
RLS and QRD-RLS are shown in that order in Table B
I1 BACKGROUND below.
TABLE B
The QRD-RLS, STAR-RLS and PSTAR-RLS algo-
rithms are described in [3],[4],[6]. To obtain simplify
the finite-precision performance expressions we exploit the
steady-state properties of the algorithms. We find expres-
sions for the steady-state properties and these are shown
in Table A.
TABLE A

Algorithm Rotation Parameter r', h2


IV FLOATING-POINT
STAR t', = - &A)
aa 1
M1?(1--x) RESULTS
In the last section the deviation due to finite-precision
with fixed-point was analyzed. In this section the analysis
is repeated for a floating-point representation. A floating-
QRD :c =A;& = 1- A A-
I-x x&-j +
point representation consists of a k 1 bit mantissa and a
k, bit exponent. For the mantissa 1 bit is used as sign bit.
Rounding is used on the mantissa to quantize it down to
+
k 1 bits. Thus the number representation here requires
k, + k + 1 bits. If "#" represents a division, multiplication,
addition or subtraction operation, then quantizing the re-
sult to a floating-point format would involve an error t,
i.e.,
+
f l o a t ( ~ # b )= a#b(l c) (4.1)
where c is assumed to be a zero-mean noise process with
a variance of uc2 which is equal to .182-2k. We carry
out the same analysis as in the case of fixed-point. The
expected squares of the deviation in the estimation error
for STAR-RLS, PSTAR-RLS and QRD-RLS, in that order,
I11 FIXED-POINT RESULTS are shown in Table C.
TABLE C
In this section we analyze the performance of the three
algorithms when fixed-point arithmetic is used. All quan-
tities in the algorithm are represented with k, bits for the
integer part and k bits for the fractional part Thus, the
fixed-point representation would require k, k 1 bits, + +
with one bit being used for sign. We also assume that
rounding is used in the operations. If "*" represents a
multiply or divide operation, then
fiz(a * b) = a * b + 6 (3.1)
where t is a zero mean white noise process with variance V COMPARISON
of uc2 = 2-2k/12 181. Additions and subtractions do not
introduce any error (assuming there is no overflow). The To compare the performance of the different algorithms
approach used here is to follow the arithmetic operations in we can use the expression for the deviation in the estima-
the different cells and keep track of the errors which arise tion error. Note that the expressions are similar in form

111-82
and this greatly simplifies the comparisons. First we com- inputs are quantized to the given format (fixed-point or
pare the floating-point and fixed-point implementations for floating-point) and the wordlength. The actual operation
the same algorithm. It should be kept in mind that the is carried out and the result is then quantized. In the case
variance U,’ is different for the two implementations. For of floating-point, each number is represented as a vector of
STAR-RLS and QRD-RLS, we find that the fixed-point length 2, one for the exponent and the other for mantissa.
implementation is always better than floating point. The The handling of the mantissa and exponent is done aa in
same is true for PSTAR-RLS if hardware. The rounding scheme used is the round-to-the-
nearest scheme.
The simulations are run on a 5000 long data stream and
the average results are computed by averaging over the last
With p = 5 and A = .98, this condition is equivalent to 1250 samples. This ensures that the results are calculated
M 2 4. after steady-state has been reached. The infinite precision
Next, we compare the different algorithms with the same results are also calculated alongside, and the deviation is
arithmetic being used. In the case of fixed-point arithmetic determined. The plots show the deviation in estimation
it can be shown that PSTAR-RLS always performs better error as a function of the number of bits for the fractional
than STAR-RLS. The same would be true for floating- part k. The theoretical expressions are plotted with dashed
point implementation if line while the simulation results are shown by ”+”. Fig.1
shows the results for the fixed-point case while Fig.2 shows
the results for the floating-point case.
References
The above would be satisfied if p is not too small (p > 3).
Thus, in general PSTAR-RLS has better numerical prop- [I] W. M. Gentleman and H. T. Kung, “Matrix triangu-
larization by systolic arrays,” Proc. SPIE Real Time
erties than the STAR-RLS algorithms. This would not be Signal Processing IV, vol. 298, pp. 298-363, 1981.
expected since more computations are involved in PSTAR- [2] J. G. McWhirter, “Recursive least-squares minimiza-
RLS. But it should be noted that the computation of tan- tion using a systolic array,” Proc. SPIE Real Time
gent in PSTAR-RLS is simpler, and the effect of forgetting Signal Processing IV, vol. 431, pp. 105-li2, 1983.
factor is to multiply the delayed version of cell content by [3] S. Haykin, Adaptive Filter Theory. Englewood Cliffs,
Xp instead of A. This factor controls the finite-precision
- NJ: Prentice Hall, 1986.
deviation and thus helps PSTAR-RLS. Thus, the extra ad- [4] K. J. Raghunath and K. K. Parhi, “High-speed RLS
dition operations needed in PSTAR-RLS get compensated using scaled tangent rotations (STAR),” Proc. of
due to these reasons. IEEE Intl. Symp. on Circuits and Systems (ISCAS-
Next, we compare STAR-RLS and QRD-RLS. In the 93), pp. 1959-1962, May 1993.
case of fixed-point it can be shown that STAR-RLS per- [5] K. K. Parhi and D. G. Messeeschmitt, “Pipeline inter-
forms better than QRD-RLS if leaving and parallelism in recusive digital filters- part
I: Pipelinin usin scattered look-ahead and decom-
X ( l - 1.5XZM-’) position,” fEEE %“actions on Acoustics, S ech,
M > - (5.3) and Signal Processing, pp. 1099-1117, July 1 9 8 r
2.5(1 -X)(2XzM-’ - 1 ) ’
[6] K. J. Raghunath and K. K. Parhi, “Pipelined imple-
The right-hand-side (RHS) is positive only when .5 < mentation of hi h-speed STAR-RLS adaptive filters,”
Proc.. SPIE, Akanced Si nal Processin A Lgorithms,
xZM-1
- < .67. The condition (5.3) will hence be satis- Architectures, and Implkentatrons b, vol. 2027,
fied for most of time except when X2M-’ is close to and 1993.
greater than .5. In the case of floating-point arithmetic, [7] H. Leung and S. Haykin, “Stability of recursive QRD
STAR-RLS has a better performance than QRD-RLS if LS algorithms ;sin finite precision systolic array
.25 < Since is close to 1, this condition would implementation IBEE Transactions on Acoustics,
S ech, and Signal Processing, pp. 760-763, May
be satisfied for most applications. lK9.
[8] S. H. Ardalan and S. T. Alexander, “Fixed-point er-
ror analysis of the exponential windowed RLS algo-
VI SIMULATION RESULTS rithm for time-varying systems,” IEEE Transactions
on Acoustics, Speech, and Signal Processing, vol. 35,
pp. 770-783, June 1987.
Next, we show results of the simulations conducted to
verify the theoretical expressions developed in this paper. [9] H. Dedieu and M. Hasler, “Error propagation in re-
cursive QRD LS filter,” Proc. of IEEE Intl. Conf. on
We use the 11 tap equalizer example of [3], which hass also Acoustics, Speech and Signal Processing (ICASSP),
been used in [4] and [6]. pp. 1841-1844, 1991.
To simulate the finite-precision behavior of the algo- [lo] K. J. R. Liu, K. Yao, and C. T. Chiu, “Dynamic
rithms we carry out the computations as would be done range, stabilit and fault-tolerant capability of finite-
in an actual circuit. We have developed subroutines in C precision RLgsystolic array based on Givens rota-
language to mimic the operations of finite-precision arith- tion,” IEEE Transactions on Circuits and Systems,
pp. 625-636, June 1991.
metic. For every operation of add, multiply, etc., the

III-83
(4 STAR-RLS
0-
+
Q -20:'. t
*. +
40.
"*
'4 +.

1
w -80
+.
+-+Y
Y
Y
,g -la,. -+-t
**. t
3
d-120.
+.+.+
'
-140'

(d) PSTAR-RLS@=lo)
Or

Fig. 1 Fixed-point results (- - - theoretical, + simulated)

-140' -1401 .
10 I5 20 10 15 20
Numbcrofbiu fobadimd pltol) Numtu dbiu for fnstionl pn(k)

(c) PSTAR-RLS @-s)

Or

Fig. 2 Floating-point results (- - - theoretical, + simulated)

111-84

You might also like