GB2380367A

GB2380367A - an adaptive filter (or equaliser) which can increase or decrease its taps in response to errors in the input signal

Info

Publication number: GB2380367A
Application number: GB0110517A
Authority: GB
Inventors: James Noras; Felip Riera-Palou
Original assignee: University of Bradford
Current assignee: University of Bradford
Priority date: 2001-04-28
Filing date: 2001-04-28
Publication date: 2003-04-02
Also published as: GB0110517D0

Abstract

An adaptive filter comprising an input, a series of taps, and an output, each tap being arranged to generate an output comprising a past signal value weighted by a coefficient, the values of the tap coefficients being updated according to an algorithm which minimises a predetermined error measurement in the combined tap outputs that together provide the adaptive filter output, wherein the filter is provided with switching means operative to increase or decrease the number of taps included in the series of taps is as required in response to variations of the error included in a signal provided at the input, the increase or decrease of the number of taps being carried out in accordance with the output of comparison means which compares the incremental error reduction provided by different numbers of taps. The taps may be increased/decreased by switching over segments of taps.

Description

ADAPTIVE FILTER The present invention relates to an adaptive filter, and particularly though not exclusively to an adaptive filter suitable for use at a receiver end of a communications channel.

Any communications channel will introduce distortion to a signal when that signal is transmitted across the channel. Distortion is particularly accentuated in wireless systems, with multipath propagation and fading both causing significant degradation of a signal. A well known method of reducing signal distortion is through the use of a linear equaliser located at a receiver end of a channel. A linear equaliser comprises a finite impulse response filter which aims to invert the impulse response of a channel, thereby recovering an original transmitted signal. The filter comprises a series of taps, each of which generates an output comprising a past signal value weighted by an equaliser coefficient (tap gain). The outputs of the taps are summed to provide a filter output in which the effects of signal distortion are reduced.

If the characteristics of a communications channel are known and static, the coefficients of the taps of the linear equaliser can be computed a-priori and fixed for subsequent communication. On the other hand, if the channel is unknown or timevarying, the coefficients of the taps must be adapted in order to maintain an optimum level of performance (the linear equaliser is an adaptive filter).

The coefficients of the equaliser may be updated using one of several known adaptive algorithms. The two most commonly used adaptive algorithms are the least mean squares (LMS) and the recursive least squares (RLS) algorithms. LMS is a computationally simple and numerically robust algorithm that works well enough for most applications. Its main drawback is its poor convergence properties (time required to obtain an optimum set of tap coefficients) in some types of channels. In contrast, RLS provides extremely fast convergence, independently of channel characteristics, but it requires many more operations than LMS. A further

disadvantage of LMS is that it is very sensitive to numerical round-off errors when implemented in fixed-point processors.

The computational complexity of an adaptive algorithm is directly related to the number of taps included in an adaptive filter which implements the algorithm. The LMS requires 2N products and 2N additions per iteration, where N is the number of taps in the filter. RLS requires 2N2 products, 1. 5N2 additions and 1 division per iteration.

It has been shown that the number of taps included in a linear equaliser has a significant effect on the utility of the equaliser [Variable Length Equalisers for Broadband Mobile Systems"F. Riera-Palou, J. M. Noras, D. G. M. Cruickshank. IEEE Vehicular Technology Conference, Boston, September 2000]. The output error obtained using equalisers having different numbers of taps has been shown to be dependent upon the characteristics of a channel (channel spectrum, level of interference, Doppler spread). In particular, the symbol energy per noise power spectral density of the signal (where the noise power spectral density includes any unwanted signal, including interfering signals from other users in the system) has a significant effect upon the output error obtained.

There is a trade-off in terms of equaliser length between the output error provided by an equaliser, and the number of computations required to maintain the tap coefficients at optimum values. A long equaliser will in general provide low output error but will require many computations, whereas a short equaliser will require few computations but will provide a higher output error. The large number of computations required by a long equaliser is disadvantageous for several reasons. A time delay may be introduced by the computations (this is particularly undesirable for communications applications). The amount of power required to carry out the calculations may be large, a particular disadvantage for mobile communications applications. A long equaliser may also suffer from the further disadvantage that it over-equalises a signal. Over-equalisation occurs when the number of taps is significantly greater than the

number required to give optimal reduction of signal distortion, with excess taps tending to compound the effect of noise in the signal.

It is known to set the length of a linear equaliser prior to initiation of communication in order to attempt to provide good filter performance without wasting computation cycles. For example, PCT/IL97/00315 describes a filter in which operation is initiated with a maximum number of taps in operation, a quality criteria is measured, and if the quality criteria is above a predetermined threshold then M taps are removed from the filter. The process is repeated until the quality criteria falls below the predetermined threshold, whereupon M taps are added back to the filter. Once the filter length needed to satisfy the quality criteria has been determined, the length of the filter is fixed and communication is begun. The filter described in PCT/IL97/00315 works well for a communication channel having properties which do not vary over time (for example a land-line telephone link). However, where the properties of the communication channel vary over time, the performance of an equaliser will rapidly become sub-optimal and will remain so unless the properties of the communication happen to return to their initial values.

It is an object of the present invention to provide an adaptive filter which overcomes or mitigates the above disadvantage.

According to a first aspect of the invention there is provided an adaptive filter comprising an input, a series of taps, and an output, each tap being arranged to generate an output comprising a past signal value weighted by a coefficient, the values of the tap coefficients being updated according to an algorithm which minimises a predetermined error measurement in the combined tap outputs that together provide the adaptive filter output, wherein the filter is provided with switching means operative to increase or decrease the number of taps included in the series of taps is as required in response to variations of the error included in a signal provided at input, the increase or decrease of the number of taps being carried out in accordance with the output of comparison means which compares the incremental error reduction provided by different numbers of taps.

The comparison means may comprise a comparator. The predetermined error measurement may be the Mean Squared Error.

Suitably, the taps are grouped into segments, each segment comprising a predetermined number of taps, and the number of taps included in the adaptive filter is incremented or decremented in segments of taps.

Suitably, the comparison means is arranged to compare the error of outputs from the segments of taps.

Suitably, the compared error is an accumulation of the difference between detected symbols and received data.

Suitably, the output of the final segment of the adaptive filter and the output of the preceding segment of the adaptive filter are compared by the comparison means, and the number of segments comprising the filter is incremented or decremented in accordance with the output of the comparison means.

Suitably, if the output of the comparison means indicates that a significant error reduction has been provided by the final segment, a further segment is added to the adaptive filter.

Suitably, the accumulated square error of the output of the final segment of the adaptive filter is compared with a value comprising the accumulated square error of the output of the preceding segment multiplied by a first weighting.

Suitably, the value of the first weighting is used to adjust the predisposition of the filter to add a further segment to the adaptive filter.

Suitably, the output of the comparison means indicates that only a minor error reduction, or no error reduction, has been provided by the final segment then the final segment is removed from the adaptive filter.

Suitably, the accumulated square error of the output of the final segment of the adaptive filter is compared with a value comprising the accumulated square error of the output of the preceding segment multiplied by a second weighting.

Suitably, the value of the second weighting is used to adjust the predisposition of the filter to remove the final segment from the adaptive filter.

Suitably, the value of the second weighting is less than one.

Suitably, if the output of the comparison means indicates that a medium error reduction has been provided by the final segment then the number of segments is not changed.

Suitably, the range of medium errors for which the number of segments is not changed is determined by the difference between the first weight and the second weight.

Suitably, the number of taps per segment is selected prior to operation of the adaptive filter.

Suitably, the number of taps per segment is four.

Suitably, an adder is provided at a given segment of the adaptive filter, the adder being arranged to sum a value generated by that segment with values generated by each preceding segment.

Suitably, an adder is provided at each segment of the adaptive filter to sum the value generated by that segment with a previously determined sum obtained from the preceding segments.

Suitably, an adder is provided at the final segment of the adaptive filter to sum the values generated by all segments of the adaptive filter, and a second adder is provided at the preceding segment of the adaptive filter to sum the values generated by all but the final segment of the adaptive filter.

Suitably, the adaptive filter is a linear equaliser.

Suitably, the adaptive filter is a feedback filter.

Suitably, the linear equaliser and the feedback filter are combined to form a Decision Feedback Equaliser, with the linear equaliser functioning as a forward filter, with the number of taps included in the forward filter or the feedback filter being increased or decreased as required in response to variations of the error included in the input signal.

Suitably, taps are switched from the forward filter to the feedback filter, and from the feedback filter to the forward filter, in response to variations of the error included in the input signal.

Suitably, outputs are periodically obtained from a multitude of segments of taps, and comparison means are used to compare the outputs to determine whether a step change of the number of segments of taps included in the adaptive filter is required.

Suitably, the multitude of segments of taps includes segments which were not included in the adaptive filter during recent operation of the adaptive filter.

Suitably, the multitude of segments comprises substantially all of the segments of taps.

Suitably the accumulated difference is squared to provide an accumulated square error prior to the comparison.

Suitably, the accumulated difference is raised to a power greater than two prior to the comparison.

According to a second aspect of the invention there is provided a method of operating an adaptive filter comprising an input, a series of taps, and an output, the method comprising arranging each tap to generate an output comprising a past signal value weighted by a coefficient, updating the values of the tap coefficients according to an algorithm which minimises a predetermined error measurement in the combined tap outputs that together provide the adaptive filter output, wherein the method further comprises increasing or decreasing the number of taps included in the series of taps is as required in response to variations of the error included in a signal provided at input, the increase or decrease of the number of taps being carried out in accordance with a comparison of the incremental error reduction provided by different numbers of taps.

The second aspect of the invention may suitably incorporate features of the first aspect of the invention.

A specific embodiment of the invention will now be described by way of example only, with reference to the accompanying figures in which: Figure 1 is a schematic illustration of a communications system to which the invention is applied; Figure 2 is a schematic illustration of an equaliser which implements the invention; Figure 3 is a graph illustrating the shows the evolution of symbol energy per noise power spectral density of a signal; Figure 4 is a graph which compares the output of an implementation of the invention with fixed length equalisers; Figure 5 compares the number of products required by implementations of the invention and by fixed length equalisers; and

Figure 6 is a schematic illustration of a nonlinear filter which implements the invention.

A communications system to which the invention is applied is shown schematically in figure 1. An input signal d (n) passes through a channel c (n) which introduces distortion into the input signal. White Gaussian noise is also introduced into the input signal d (n), as indicated by v (n) in figure 1. The resulting signal u (n) is input to an equaliser w (n) which generates an output y (n). Estimated data bits d (n) of the signal are generated by passing the equaliser output y (n) through a decision device. The equaliser output y (n) is passed to an inverting input of an adder, which during training of the equaliser determines the difference between the equaliser output y (n) and the input signal d (n) and passes the difference back to the equaliser. This feedback allows initial coefficients of taps comprising the equaliser to be determined. Since the properties of the communications channel are time varying, the filter coefficients of the equaliser must be adapted over time, and this is done by passing the estimated data bits d (n) of the signal to the adder.

Figure 2 shows schematically the structure of an equaliser which implements the invention. The equaliser comprises an N-tap equaliser which, instead of comprising a conventional single series of taps, has been split into K segments of P taps per segment. This structure permits the length of the equaliser to be varied by adding or subtracting segments of P taps. The number of segments K is an integer factor of the total number of taps in the equaliser (N=KP). The parameter P allows selection of the level of"performance granularity"provided by the segments. When P is large, adding or removing a segment will probably have a much more significant effect than for a small P. If P is given a value of 1, then the equaliser expands and/or contracts on a tap by tap basis.

The taps of the equaliser each act in the conventional manner, i. e. the coefficients of each tap are modified during operation of the equaliser to minimise a least mean squares (LMS) or recursive least squares (RLS) error measurement. However, in

addition to the conventional modification of the tap coefficients, the number of segments of taps used in the equalisation process is increased or decreased as required.

The criterion used to determine the number of segments of taps to be used should include only information available to the equaliser subsystem, that is, no external information should be required (such as noise level, Doppler spread). Furthermore, the criterion should be easy to compute, to avoid incurring significant computational cost which in a worst case scenario could outweigh the benefit provided by reducing the number of taps.

The criterion used by the invention is based on an error measurement which will be referred to hereafter as the accumulated squared error (ASE). The ASE is defined by:

In digital communication systems the most common measure of performance is the bit error rate (BER), also called the probability of bit error (Pe). Although BER is a very meaningful metric of system performance, it is difficult to deal with from a mathematical point of view when trying to optimise subsystems. A more tractable index is the MSE, which is given by:

In plain words, MSE is the average square of the difference between what it is received and what should be received. MSE and BER are related through Saltzberg's bound which predicts that reductions in the MSE will probably, but not necessarily always, lead to reductions in the BER.

ASE (n) and MSE (n) are related by:

ASE is a known error measurement, but has previously only been applied to the control of adaptive filter coefficients, for example in the RLS algorithm. The control of adaptive filter coefficients is a very complicated and computationally expensive process.

Referring to figure 1, it can be seen that ASE (n) can be computed very easily from the signal e (n) which is used to determine the coefficients of the adaptive filters. No external information is required to measure ASE (n). All that is required is that e (n) is squared and is accumulated. Referring to figure 5, ASE (n) can be computed for every active segment of the equaliser, therefore we can compute ASEo (n), ASE, (n),.... ASEK-l (n).

In order to define a criterion that determines when to change the length of the equaliser the invention computes two different ASEi (n). At any given moment the

equaliser is made of L active segments (L < K). At every iteration ASEL (n) and ASEL- ] (n) are computed. The decision algorithm used is :

If ASEL (n) aupASEL-J (n) --- ? > add one segment (P extra taps) If ASEL (n) : ad. ASEL-I (n)- > remove one segment (P less taps) with 0 < a, a 1 and a a.

The comparison between ASEL (n) and ASEL-I (n) is carried out by a comparator (not shown in figure 2).

This criterion can be translated into plain words as: If the equaliser with L segments performs significantly better than the one with L-1 segments add another segment. If the equaliser with L segments performs similarly to the one with L-1 segments remove the last segment. This limits the number of segments comprising the filter to those that really make a significant contribution in the equalisation process.

The invention assesses the contribution of the last operating segment of the filter by comparing the output of the last operating segment with that of the preceding segment. The invention does not attempt to compare the output of the last operating segment with the output that would be generated by the next segment if that segment were to be added to the equaliser. The reason why this is not done is because it would require adaptive algorithm coefficients to be generated for the next segment, a computationally very intensive process. By monitoring the output of the last operating segment of the filter and the output of the preceding segment, the invention avoids incurring this extra intensive processing.

The variables aup ans ad determine the amount of improvement needed from the last segment for its contribution to be deemed significant. Values of aup very close to 1 will tend to expand the filter even if the performance benefit is very small. On the other hand, values for aup very close to zero will tend to keep the equaliser short unless there is an important benefit in terms of ASE reduction. Similarly, values of a very close to 1 will tend to keep the equaliser long while values close to 0 will increase the tendency of the equaliser to shrink. The gap between a up and a dw determines the ASE level that is considered as optimal. If the values of aup and adware very close, the equaliser will tend to expand and contract sequentially. Since ASE (n) and MSE (n) are linearly related, minimisation of one implies minimisation of the other. The constraints 0 < a, a < l and aa,, are not rigorous limits, but provide a well behaved system. If value of ad ils set to be greater than 1 then the invention will still function correctly, but its performance will be somewhat degraded.

Similarly, if aup > adw then the invention will still function correctly, but the number of segments of taps may oscillate around the optimal value.

The criterion used to control the equaliser length is advantageous because it does not use any information external to the equalising subsystem. Furthermore, the whole criterion and decision uses only 4 products and 2 additions independently of the number of taps of the filter, computationally a relatively low burden.

It is not necessary to include K-l adders in the equaliser. An alternative configuration could use two adders connected to a switching device arranged to select the outputs of the last two segments. Implementation details will depend on the hardware/software used to implement the equaliser, but will be apparent to those skilled in the art.

The invention is advantageous because it prevents the use of an excessive number of taps by detecting the conditions where they would not offer any MSE reduction. The invention is also capable of detecting whether it is favourable to exploit further the equalisation process by increasing the number of taps. As a result, the invention provides a computationally efficient method of minimising the distortion of a received signal.

The parameters aup and adw can be used as"tuners"in order to achieve a desired level of performance or to limit the amount of processing of the equalisation process. This tuning may be used, for example, to purposely degrade the equaliser performance in order to save power.

The advantages provided by the invention do not incur any significant increase in hardware/software with respect to a conventional linear equaliser.

As an example of the functioning of the invention we show in Figures 3,4 and 5 the data obtained from comparing fixed equalisers of various lengths (13 taps, 23 taps and 33 taps) with an equaliser according to the invention having aup = 0.6, ad = 0. 99 and a segment length of 4 taps (in the graphs, aup ana aware denoted by a+ and a-). The system to be equalised was an 11-tap static channel whose noise level (No) was dynamic. Figure 3 shows the evolution of the symbol energy per noise power spectral density of the signal, figure 4 shows the MSE curves for the different equalisers, and figure 5 compares the number of products. The results are an average of 30 independent runs. Figure 4 indicates that the invention offers the best performance (lowest MSE) and figure 5 shows that the computational complexity of the invention

lies between the 13-tap and the 23-tap equaliser. In particular, comparing the invention with the 23-tap equaliser it can be seen that the invention provides a better performance better and saves a significant number of products.

The values oc = 0. 6, ad = 0. 99 are used in the above example because they provide efficient optimisation of the number of segments of taps. It will be appreciated that other values could be used, in accordance with the guidelines set out above.

As noted above, selection of the number of taps per segment provides control of the level of"performance granularity"provided by the segments. When the number of taps per segment is large, adding or removing a segment is likely to have a much more significant effect than for a small number of taps per segment. It has been found that the use of segments comprising 4 taps provides very good results. It is preferred not to use segments comprising single taps as it has been found that signal distortion may include plateaus which are not traversed by the addition of a single tap. Where this to occur, the addition of a single tap would provide no reduction of distortion and would erroneously lead to the conclusion that the distortion had been minimised, whereas the addition of a segment, comprising for example 4 taps, would lead to a significant reduction of distortion.

It is common to provide a linear equaliser in which the spacing of the taps is chosen to be less than the symbol (bit/chip) period. Equalisers of this type are known as Fractionally Spaced Equalisers (FSE). Usually the separation of the taps is chosen to be half the symbol period (T/2). FSE's are advantageous because their operation is not affected by the sampling phase of an input signal. They also avoid aliasing in systems using an excess of bandwidth. A FSE covering the same time span as a conventional linear equaliser will always have more taps than the conventional equaliser (a T/2 FSE will have double number of taps of the conventional linear equaliser). It will be appreciated that, since the number of taps in a FSE is significantly greater than in a conventional linear equaliser, the ability to control the number of taps included in the equaliser so as to provide good distortion reduction without requiring excessive numbers of computations is particularly important.

It will be appreciated by those skilled in the art that modifications may be made to the

above embodiment of the invention. For example, the criterion used to determine

whether the number of segments of taps is optimal is E ld (n)-y (n) 12. A more general 1=0 t ! representation of the criterion is y ! J (n)-y (n) r, where m is any number greater 1=0

than or equal to 2. The value m=2 is used by the above described embodiment of the invention selected because it incurs the least amount of computations. The operation of the invention may be modified such that instead of continuously monitoring the output of the last operative segment and the output of the preceding segments, extra segments could periodically be added to the equaliser and the outputs from the extra segments compared to each other to determine whether a step change in the number of segments would be beneficial. This may be done for example every 10 milliseconds, and could be carried out instead of or in addition to the continuous monitoring of segments. The number of extra segments could be pre-set to a single value. Alternatively, every possible number of segments could be added, with the output from every segment compared.

Referring to figure 6, the invention could be applied to a Decision Feedback Equaliser comprising a segmented feedback filter and a segmented forward filter. The objective of the feedback filter is to cancel out completely the interference of previously detected symbols. Decision Feedback Equalisers are particularly useful for communications channels which include spectral nulls in the passband. The performance of the segments of the feedback filter is monitored in the manner described above, with the output of the last active segment and that of the preceding segment being compared using the ASE measurement. Segments are added to the feedback filter and subtracted from the feedback filter when the ASE measurement indicates that it is beneficial to do so.

The ASE measurement may be used to determine whether to use a linear equaliser on its own, or to add a feedback filter thereby providing a Decision Feedback Equaliser. Where this is done, one or more segments of taps may be switched into either the feedforward or the feedback loop as required. There may be instances in which it is preferable to use the feedback filter on its own.

Although the described embodiment of the invention relates to a communications channel it will be appreciated that the invention may be applied wherever adaptive filters are used, for example in the operation of a multi-use detector, or in the control of adaptive antennas.

Claims

CLAIMS 1. An adaptive filter comprising an input, a series of taps, and an output, each tap being arranged to generate an output comprising a past signal value weighted by a coefficient, the values of the tap coefficients being updated according to an algorithm which minimises a predetermined error measurement in the combined tap outputs that together provide the adaptive filter output, wherein the filter is provided with switching means operative to increase or decrease the number of taps included in the series of taps is as required in response to variations of the error included in a signal provided at the input, the increase or decrease of the number of taps being carried out in accordance with the output of comparison means which compares the incremental error reduction provided by different numbers of taps.
2. An adaptive filter according to claim 1, wherein the taps are grouped into segments, each segment comprising a predetermined number of taps, and the number of taps included in the adaptive filter is incremented or decremented in segments of taps.
3. An adaptive filter according to claim 2, wherein the comparison means is arranged to compare the error of outputs from the segments of taps.
4. An adaptive filter according to claim 3, wherein the compared error is an accumulation of the difference between detected symbols and received data.
5. A adaptive filter according to any of claims 2 to 4, wherein the output of the final segment of the adaptive filter and the output of the preceding segment of the adaptive filter are compared by the comparison means, and the number of segments comprising the filter is incremented or decremented in accordance with the output of the comparison means.

<Desc/Clms Page number 17>
6. An adaptive filter according to claim 5, wherein if the output of the comparison means indicates that a significant error reduction has been provided by the final segment, a further segment is added to the adaptive filter.
7. An adaptive filter according to claim 6, wherein the accumulated square error of the output of the final segment of the adaptive filter is compared with a value comprising the accumulated square error of the output of the preceding segment multiplied by a first weighting.
8. An adaptive filter according to claim 7, wherein the value of the first weighting is used to adjust the predisposition of the filter to add a further segment to the adaptive filter.
9. An adaptive filter according to any of claims 5 to 8, wherein if the output of the comparison means indicates that only a minor error reduction, or no error reduction, has been provided by the final segment then the final segment is removed from the adaptive filter.
10. An adaptive filter according to claim 9, wherein the accumulated square error of the output of the final segment of the adaptive filter is compared with a value comprising the accumulated square error of the output of the preceding segment multiplied by a second weighting.
11. An adaptive filter according to claim 10, wherein the value of the second weighting is used to adjust the predisposition of the filter to remove the final segment from the adaptive filter.
12. An adaptive filter according to claim 11, wherein the value of the second weighting is less than one.

<Desc/Clms Page number 18>
13. An adaptive filter according to any of claims 5 to 12, wherein if the output of the comparison means indicates that a medium error reduction has been provided by the final segment then the number of segments is not changed.
14. An adaptive filter according to claim 13, wherein the range of medium errors for which the number of segments is not changed is determined by the difference between the first weight and the second weight.
15. An adaptive filter according to any of claims 2 to 14, wherein the number of taps per segment is selected prior to operation of the adaptive filter.
16. An adaptive filter according to claim 15, wherein the number of taps per segment is four.
17. An adaptive filter according to any of claims 2 to 16, wherein an adder is provided at a given segment of the adaptive filter, the adder being arranged to sum a value generated by that segment with values generated by each preceding segment.
18. An adaptive filter according to claim 17, wherein an adder is provided at each segment of the adaptive filter to sum the value generated by that segment with a previously determined sum obtained from the preceding segments.
19. An adaptive filter according to claim 17, wherein an adder is provided at the final segment of the adaptive filter to sum the values generated by all segments of the adaptive filter, and a second adder is provided at the preceding segment of the adaptive filter to sum the values generated by all but the final segment of the adaptive filter.
20. An adaptive filter according to any preceding claim, wherein the adaptive filter is a linear equaliser.

<Desc/Clms Page number 19>
21. An adaptive filter according to any of claims 1 to 19, wherein the adaptive filter is a feedback filter.
22. An adaptive filter according to claim 20 and 21, wherein the linear equaliser and the feedback filter are combined to form a Decision Feedback Equaliser, with the linear equaliser functioning as a forward filter, with the number of taps included in the forward filter or the feedback filter being increased or decreased as required in response to variations of the error included in the input signal.
23. An adaptive filter according to claim 22, wherein taps are switched from the forward filter to the feedback filter, and from the feedback filter to the forward filter, in response to variations of the error included in the input signal.
24. An adaptive filter according to claim 2 or any claim dependent thereon, wherein periodically outputs are obtained from a multitude of segments of taps, and comparison means are used to compare the outputs to determine whether a step change of the number of segments of taps included in the adaptive filter is required.
25. An adaptive filter according to claim 24, wherein the multitude of segments of taps includes segments which were not included in the adaptive filter during recent operation of the adaptive filter.
26. An adaptive filter according to claim 25, wherein the multitude of segments comprises substantially all of the segments of taps.
27. An adaptive filter according to claim 4, wherein the accumulated difference is squared to provide an accumulated square error prior to the comparison.
28. An adaptive filter according to claim 4, wherein the accumulated difference is raised to a power greater than two prior to the comparison.

<Desc/Clms Page number 20>
29. A method of operating an adaptive filter comprising an input, a series of taps, and an output, the method comprising arranging each tap to generate an output comprising a past signal value weighted by a coefficient, updating the values of the tap coefficients according to an algorithm which minimises a predetermined error measurement in the combined tap outputs that together provide the adaptive filter output, wherein the method further comprises increasing or decreasing the number of taps included in the series of taps is as required in response to variations of the error included in a signal provided at the input, the increase or decrease of the number of taps being carried out in accordance with a comparison of the incremental error reduction provided by different numbers of taps.
30. An adaptive filter substantially as hereinbefore described with reference to the accompanying figures.
31. A method of operating an adaptive filter substantially as hereinbefore described with reference to the accompanying figures.