Advances in Communications, Computing, Networks and Security 7

Implementation of RLS Algorithm

H.V.Ajmera and M.Z.Ahmed
Fixed and Mobile Communications, University of Plymouth, Plymouth, UK
e-mail: [email protected]

Abstract
An algorithm for recursively calculating the least square solution to adaptive filter, for its
implementation on a dsPIC is described in this paper. It shows an unified approach for
deriving the fixed-point Recursive Least Square (RLS) algorithm to estimate the optimal set of
filter coefficients. The objective is to study implementation effects on the Recursive Least
Square algorithm used in the mean-square manner for adaptive filter for a dsPIC. Using the
matrix operation as well as the round-off error model, the effects of machine implementation
on RLS algorithm has being found out theoretically confirmed by simulation results. The main
sources of errors can be thought of: division operation involved in updating filter variables.
Due to such errors, the digitally implemented algorithm shows a divergence rather than
convergence. A couple of solutions are provided to give better performance, so that the
simulated results for filter coefficients for fixed and floating point are found to be similar.

Keywords
RLS algorithm, Finite Precision Effect, Error Analysis

Introduction

In recent years, recursive least square (RLS) filters have emerge as a powerful tool
for adaptive filtering, prediction and identifying (Haykin, 1996). However, when
implemented in a finite precision environment, RLS algorithm can suddenly become
unstable and also divergence becomes a problem. A number of papers have being
published that deals with the effects that finite word length has on the RLS algorithm
(Adali and Ardalan, 1987; Ardalan, 1986; Bottomley and Alexander, 1991). Its always
being a difficult issue to analyses such algorithm due to recursive nature of this filter.
Consequently, certain simplifications have being made in order to yield useful
results. One simplifying approach is to assume that kalman gain is available with
finite precision range (Adali and Ardalan, 1987; Ardalan, 1986). A second approach
was to bias the round-off error in each filter quantity to achieve a more stable
performance (Bottomley and Alexander, 1991). It has also being shown that the
prewindowed growing memory RLS algorithm is unstable (Adali and Ardalan, 1987),
while other assume that some of the term didnt contribute to the error. All the
elements of matrices and vectors in the RLS algorithm will deviate from their correct
values due to quantization effects, giving three separate effects: 1) The inverse
autocorrelation matrix may become indefinite due to accumulation of the error, 2)
Propagation of the error due to recursive use of the algorithm and 3) With use of
exponentially forgetting factor, the errors also grows exponentially. The output
signal is computed by convolving the input sample with tap coefficients and the
weight vector are updated by taking the product of kalman gain with the predicted
246

Section 4 Computing, Communications Engineering and Signal Processing & Interactive

Intelligent Systems

error. In this paper, there are no simplifications made. Specifically, 1) Quantized

input signal and desires signal are available, 2) Neglect the second order noise term
if the magnitude is smaller than 1. The paper is organised as: In section II, the
infinite precision RLS algorithm is being discussed, while in section III, fixed-point
error are injected into the algorithm with results being reviewed in section IV.

RLS Algorithm

The basic aim of the least square algorithm is to minimize the sum of the squares of
the difference between the desired signal r(i) and the filter output y(i). An important
characteristic of RLS is the computation of the estimate of the inverse correlation
matrix from the input data u(i), which helps the minimization process. The adapted
coefficient hn(i), n=0,1,2N aim at minimizing the given objective function. In the
case of the least square method, the objective function is given by
n

(n) =

n i e(i)

(2.1)

i0

where, e(i)= r(i) y(i)

(2.2)

and is forgetting factor. To find optimum values of the tap-weight vector h(n), it
necessary that (n) achieves its minimum value. This is done by taking partial
derivative with respect to the tap coefficients hn. The optimum value for the filter
coefficient is obtained when the partial derivative is set to zero. The resulting
expression is given by:
1
hn= R u (n ) .rdu(n)

(2.3)

where, Ru(n) and rdu(n) are auto correlation of the u(n) and cross correlation matrix
between u(n) and r(n) respectively. Using the matrix inversion lemma, and defining
term kalman gain, the inverse of the autocorrelation matrix is given by
1
(n
P(n + 1) = R u

1
1) = [P(n) k(n + 1) uT(n + 1).P(n) ]

(2.4)

Using eq. (2.2), (2.3) and (2.4) we develop a recursive solution for updating the least
square estimate h(n) for the tap weight at iteration n as follows:
hn+1 = h(n) + k(n + 1) e(n + 1)

(2.5)

Thus, we can conclude that the new estimates of the tap coefficients are calculated
based on the inner product of the old estimate and the current input sample.

247

Advances in Communications, Computing, Networks and Security 7

Fixed-Point Implementation of RLS Algorithm

In this section, a fixed point analysis approach is developed to analyze RLS

algorithm. The approach develop is clear- model all the round-off errors resulting
from fixed-point implementation and depending on this model, build up exact
recursive equation for total error in the algorithm. A typical approach towards the
modeling is the assumption that addition and subtraction do not introduce any roundoff error and this holds true as long as there is no overflow. Thus, for most of cases,
the culprit left are multiplication and division, as the source of round-off error in
fixed implementation of RLS algorithm. The round-off error in the product of a
multiplication can be expressed as
f[xy] = xy + xy
where xy is the infinite precision product, xy is relative error and is independent of
x,y and also of xy. Similar result do exist in case of division.
The error term e(i) consists of the difference between a desired filter output r(i) and
the fixed-point output of a filter y(i), where e(i) is the fixed-point error term. Here
h(n) denotes fixed-point weight coefficients.
'

e(i) = r(i) y(i) = r(i) h n uT(i) (i)

(3.1)

The optimum value of weight tap coefficients is found out in a same way as done for
conventional RLS algorithm and for fixed-point RLS is given by:
rdu(n) du(n) = Ru(n).hn

(3.2)

where, du(n) is cross correlation matrix between u(i) and error (n). As from the
fixed-point model defined previously, the multiplication and division may result in
quantization error, denoting the fixed-point kalman gain by k(n + 1) while (n) as
the fixed-point error, the kalman gain is expressed as:
k(n + 1) = k(n + 1) + (n)

(3.3)

Similarly, the error introduced in the calculation of the inverse autocorrelation matrix
is (n) and P(n + 1) as the inverse auto correlation matrix using fixed-point RLS
algorithm, then

1
P(n + 1) =
(n)]

[ P(n) k(n + 1) uT(n + 1).P(n) + (n) uT(n + 1).P(n)

+
(3.4)

Once the fixed-point inverse autocorrelation is calculated, we can then use eq. (2.5),
(3.2), (3.3) and (3.4) to update the weight vector by recursive sequence.
h(n + 1) = h(n) + k(n + 1)e(n + 1) NN(n)h(n)
248

(3.5)

Section 4 Computing, Communications Engineering and Signal Processing & Interactive

Intelligent Systems

where, NN(n) = (n) uT(n + 1).P(n) + (n). This is the theoretical expression for the
fixed-point weight error vector.

Experimental Analysis

The section below provides the effect of fixed-point errors on the performance of the
RLS algorithm through simulation results. The programs were stimulated in C and
MPLAB. The algorithm is based on the equations (3.1), (3.3), (3.4) and (3.5). In the
calculation, the inverse autocorrelation matrix is initialized as P(0) = -1I, where,
= 0.005 which ensure that P-1(n) is a positive definite matrix. All the other vectors
except the input and the desired response are set to zero. The desired response to the
algorithm was a sine wave, while the input response consists of delayed version of
the sine wave added with a sine wave of different frequency. The table shown below
is calculated for different forgetting factor lambda .
Iteration

=0.1

=0.5

0
1
2
3

0.0000
0.0000
0.0371
-0.013

0.0000
0.0000
-0.0413
0.4645

4
5
6
7

-0.069
-0.2171
-0.0514
-0.0706

0.3786
0.4446
-0.328
0.2280

8
-0.7693 -0.1528
9
-0.395 0.3817
10
-1.233 -0.1596
Table 1: Comparison of (i) for different
With larger value of , the magnitude of error remains less than 0.6 and its value
remains more stable, although the filter continues to be unstable due to the higher
value of . The smaller value of makes the filter more unstable as compared to
higher values of , and also sets the maximum value of (i). This smaller value of
has benefit of offering the least value of (i). With the smaller value of forgetting
factor, the maximum value shown is almost the double of its counterpart.
One of the most important filter quantity is the kalman gain, as the value of k(n)
affects all the filter quantities directly. This can be seen from eq. (3.4) used to
calculate the inverse of auto correlation matrix and from eq. (3.5), where current tap
coefficient vector uses the value of kalman gain. Hence, quantization in kalman gain
k(n), would amplify the total error in the filter and severely affect the stability of the
algorithm. To see how the errors are manifested in practice, we show the results of
computer experiment.
Table 2 shows results for different value of at iteration 2. As observed from above,
the error vector (n) increases as the value of the is being increased from 0.1 to 0.5.
249

Advances in Communications, Computing, Networks and Security 7

The variation in the error for the case when =0.1 is very small, while the simulation
results shows a tremendous growth in the error (n) for the latter case. This means
the magnitude of could be said to be proportional to .
With the presence of inverse autocorrelation, the successive taps becomes
decorrelated in RLS algorithm, making the algorithm self-orthogonalizing. The
update value of P(n) is calculated using the previous value of P(n), the product of
kalman gain and the tap inputs. Table 3 below shows the error (n).
=0.1

=0.5
Iteration 2

1.9711
3.2636
0.8482
5.3929

0.0673
12.7649

0
Table 2: Comparison of (n) for different

=0.1
=0.5
2.4592 0.5371 2.218 0 3.0142 1.5050 0.2129 0
0.6029 0.0188 3.8376 0 1.4353 1.4719 0.8287 0

2.5396
4.6352 9.7435 0 0.3748
0.7289 3.7256 0

0
0
0
0
0
0
0
0

Table 3: Comparison of (n) for different

As the autocorrelation matrix in symmetric, so is the case with the error in it. The
values of (n) also follows this symmetric property. As seen from the table, (n)
shows maximum value for the minimum value of forgetting factor .
Having calculated all the error in each filter variables, the final step will be to
calculate the total error in the tap coefficients affected due to the errors present in the
filter variable. In this sense, the errors generated in the filter vectors and matrices are
multiplied by different filter variables to produce additional errors. In the eq. (3.5),
there is an error NN(n) which is given as:
2.5076 5.5667
4.078
10.4329 5.1587. 9.7086

27.2169 13.9655 17.2931

0
0
0
Table 4: Error NN(n)

0
0
0

The magnitude of the error produce is of order of 10. An error with such magnitude
definitely tends to make the filter highly unstable and also losses the convergence
property for which the adaptive filter are famous for. This error is then multiplied by
the current tap coefficients.
250

Section 4 Computing, Communications Engineering and Signal Processing & Interactive

Intelligent Systems

Table 5 shows the amount of the error which is added to the previous tap coefficients
to give the new coefficients. In other words, this is the error by which the new taps
coefficient will be shifted. One peculiar characteristics of this error is the dependence
on the previous tap coefficient as seen in eq. (3.5). In short, the tap coefficients are
updated as addition of previous tap coefficients with the product of kalman gain and
error e(n), with the addition of the product NN(n)h(n).
3.0379
4.6637

6.0451

0
Table 5: Total Error in Tap Coefficient

There are two problems which can be associated with the error result found out
above: Firstly, a fixed-point implementation does not guarantee to prevent the
overflow and underflow, a problem which results in nothing but divergence.
Secondly, the stability may not be guaranteed. The problem can be solved by
preventing the overflows. This can be done either by using floating-point, not
possible in our case or by modifying the code. From a designer point of view, a
careful fixed-point implementation of the filter is to be developed, so that overflow
and underflow are minimized. Secondly, the filter is to be made stable. The
following are the way to minimize the error:

Changing the value of forgetting factor

The minimum value for error in can be obtained when the value of is taken to be
minimum shown in table 2. Successively, as seen from table 3, with minimum value
of forgetting factor the error reaches a value of 10 in the first few iteration only.
The minimum value for this error can be obtained when the value of the is
maximum. Successively, as a designer there has to be a trade-off made in selecting
the value of the forgetting factor.

1
Reducing the value of k by factor of 2

k
Changing the value of k to 2 and by keeping the value =0.5, has the advantage of
decreasing the error and also compensates the error in the auto correlation matrix.
With the above solution in the mind, let us recalculate the errors in different term.
Figure 1 below shows the stimulation result for the average mean square error (i)
against the number of iteration for the original value of kalman gain k(n). This
situation sets the maximum value of the average mean square error (i), which
comes to be more than 2. However, for most of the iteration the error value remains
close to 1. Accordingly, this indicates a huge amount of error in the least square
solution, which results in divergence. As seen form the graph, there is continuously
rise and fall in the value, with a sudden rise reaching a peak value of about 2.4,
indicating the instability of the filter.
251

Advances in Communications, Computing, Networks and Security 7

Figure 1: Average Mean Square Error (i)

Now let us analyze the graph when the value of the kalman gain is reduced by 2, so
k(n)
that the input to the filter is now 2 . It can clearly seen in figure 2 that the value of
the average mean square error (i) has decreased from the maximum value of 2.4 to
a value just less than 1.77. Consequently, the graph seen is more stable and tries to
maintain a similar value of 0.75, but at certain iteration the value is doubled then its
previous iteration. There is peculiar similarity about both the graphs- at a point there
is sudden linear rise followed by liner fall and again linear rise.

Figure 2: Average Mean Square Error (i) for scaled kalman gain
252

Section 4 Computing, Communications Engineering and Signal Processing & Interactive

Intelligent Systems

As seen from Figure 2, the error in the (i) has decreased. The error in kalman gain
(n) as shown decreases due to scaling and as shown in table 3, with increase in
value of forgetting factor the error (n) decreases. With these details, the error
NN(n) is calculated as shown in table 6.
When this calculated value of NN(n) is compared with the previous value, one can
easily state that the error had reduced to almost half of the previous one. Now let us
compute the total error due to which the current estimate tap coefficients are shifted.
This is achieved in the same way as earlier by multiplying the error shown below
with the previous tap coefficients.
0.235 3.9352 1.867 0
7.8091 1.4737 3.1074 0

15.7651 10.0565 7.4675 0

0
0
0
0
Table 6: Error NN(n) for scaled kalman gain

There is tremendous decrease in the total error as seen in table 7, when matched
against the outcome of the simulation performed without the correction in the
program. This may due to the reason that the error in the previous coefficient vector
may be less as well as the other error has decreased. Performing similar scaling of
various filter quantities may help the designer to obtain convergence or convergence
may be possible with the above solution after may be 100s of iteration or the error
value will be so less that the magnitude may be less than 1.
0.4504
0.9807

1.1111

0
Table 7: Total Error in Tap Coefficient for scaled kalman gain

Conclusion

A surprising divergence is observed on implementing the algorithm on a fixed-point

processor. From the mathematical expressions, we can say that the least square
solution includes a huge amount of error, confirmed with the stimulated results.
From the theoretical expression, it is shown that in equation for optimum filter
coefficients, the only error present is the correlation matrix of (i) and tap inputs u(i).
The expression for least square solution is same as that for infinite precision with
finite precision least square solution being the result due to arithmetic operation
carried on the finite precision filter quantities plus the error multiplied by the old
estimate of tap coefficient. It was shown that the error contribution of certain fixedpoint operations increased as the forgetting factor increases while errors due to other
operations decreased. The cause of error in some of the variables has being
identified, pointing towards the division operation carried out to calculate these
253

Advances in Communications, Computing, Networks and Security 7

terms. With the above facts, it forces the designer to make trade-off in the value of
forgetting factor . The value of lambda could be taken equal to 0.5. The (n) will
still show a higher error value. Consequently, the values in kalman gain k(n) should
be then reduced by a factor of . With this solutions, the error decreases which is
shown in table 7.

References

Adali, T. and Ardalan, S. H. (1987). Fixed-point roundoff error analysis of the exponentially
windowed RLS algorithm for time-varying systems. Proceedings of the Acoustics, Speech,
and Signal Processing, 1987. ICASSP-87., 1987 International Conference on.
Ardalan, S. (1986). "Floating-point error analysis of recursive least-squares and least-meansquares adaptive filters." Circuits and Systems, IEEE Transactions on 33(12): 1192-1208.
Bottomley, G. E. and Alexander, S. T. (1991). A novel approach for stabilizing recursive
least squares filters. Signal Processing, . IEEE Transactions on 39(8): 1770-1779.
Haykin, S. S. (1996) Adaptive filter theory, Upper Saddle River, N.J., Prentice Hall.

254