Filters coefficients in symbol spaced DFE

Hi,

I'm trying to equalize a channel with sever multipath using a DFE (12,12)
with LMS adaption algorithm.
The relative power of the replicas are quite high w.r.t the main path. (max
-4 dB). The equalizer is catastrophic.
From the learning curve analysis I can observe that the error is still high
after processing the training sequence.
Morover, the forward filter coefficients are very small compared to the
feedback filter ones (10^-3 vs 0.2).
Is there any conclusion I can draw from these info?
Thanks

Alberto

alberto.fuggetta wrote:

> Hi,
> 
> I'm trying to equalize a channel with sever multipath using a DFE (12,12)
> with LMS adaption algorithm.
> The relative power of the replicas are quite high w.r.t the main path. (max
> -4 dB). The equalizer is catastrophic.
> From the learning curve analysis I can observe that the error is still high
> after processing the training sequence.
> Morover, the forward filter coefficients are very small compared to the
> feedback filter ones (10^-3 vs 0.2).
> Is there any conclusion I can draw from these info?
> Thanks

Feedback path adaptation is nasty nonlinear problem. Your filter either 
falls into a local minimum or the adaptation is unstable.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com

>Hi,
>
>I'm trying to equalize a channel with sever multipath using a DFE (12,12)
>with LMS adaption algorithm.
>The relative power of the replicas are quite high w.r.t the main path.
(max
>-4 dB). The equalizer is catastrophic.
>From the learning curve analysis I can observe that the error is still
high
>after processing the training sequence.
>Morover, the forward filter coefficients are very small compared to the
>feedback filter ones (10^-3 vs 0.2).
>Is there any conclusion I can draw from these info?
>Thanks
>
>Alberto
>

Check the eigenvalue spread of the channel. If it is very high then LMS
will not perform well. Try to use RLS and see if you get any better
performance.

Refer to Proakis Comms or Haykin's Adaptive Filter Theory book.

Chintan

>
>
>alberto.fuggetta wrote:
>
>> Hi,
>> 
>> I'm trying to equalize a channel with sever multipath using a DFE
(12,12)
>> with LMS adaption algorithm.
>> The relative power of the replicas are quite high w.r.t the main path.
(max
>> -4 dB). The equalizer is catastrophic.
>> From the learning curve analysis I can observe that the error is still
high
>> after processing the training sequence.
>> Morover, the forward filter coefficients are very small compared to the
>> feedback filter ones (10^-3 vs 0.2).
>> Is there any conclusion I can draw from these info?
>> Thanks
>
>Feedback path adaptation is nasty nonlinear problem. Your filter either 
>falls into a local minimum or the adaptation is unstable.

Or maybe his symbol timing has not been locked down well enough for a one
sample per symbol equalizer to pull in. Trying 2 samples per symbol might
provide insight into the system's behaviour.

Steve

Hi Steve,

already tried with 2 samples per symbol but the result does not change.
:-(
I also tried computing the received samples autocorrelation matrix, just
multiplying the samples vector for its complex conj.
Is it correct?



>>
>>
>>alberto.fuggetta wrote:
>>
>>> Hi,
>>> 
>>> I'm trying to equalize a channel with sever multipath using a DFE
>(12,12)
>>> with LMS adaption algorithm.
>>> The relative power of the replicas are quite high w.r.t the main path.
>(max
>>> -4 dB). The equalizer is catastrophic.
>>> From the learning curve analysis I can observe that the error is still
>high
>>> after processing the training sequence.
>>> Morover, the forward filter coefficients are very small compared to
the
>>> feedback filter ones (10^-3 vs 0.2).
>>> Is there any conclusion I can draw from these info?
>>> Thanks
>>
>>Feedback path adaptation is nasty nonlinear problem. Your filter either 
>>falls into a local minimum or the adaptation is unstable.
>
>Or maybe his symbol timing has not been locked down well enough for a one
>sample per symbol equalizer to pull in. Trying 2 samples per symbol might
>provide insight into the system's behaviour.
>
>Steve
>
>

steveu wrote:

>>
>>alberto.fuggetta wrote:
>>
>>
>>>Hi,
>>>
>>>I'm trying to equalize a channel with sever multipath using a DFE
> 
> (12,12)
> 
>>>with LMS adaption algorithm.
>>>The relative power of the replicas are quite high w.r.t the main path.
> 
> (max
> 
>>>-4 dB). The equalizer is catastrophic.
>>>From the learning curve analysis I can observe that the error is still
> 
> high
> 
>>>after processing the training sequence.
>>>Morover, the forward filter coefficients are very small compared to the
>>>feedback filter ones (10^-3 vs 0.2).
>>>Is there any conclusion I can draw from these info?
>>>Thanks
>>
>>Feedback path adaptation is nasty nonlinear problem. Your filter either 
>>falls into a local minimum or the adaptation is unstable.
>  
> Or maybe his symbol timing has not been locked down well enough for a one
> sample per symbol equalizer to pull in. Trying 2 samples per symbol might
> provide insight into the system's behaviour.

Feedforward coeffs ~ 0 -> error is not correlated with the signal -> 
adaptation process is not working right.



> Steve
> 

I tried with RLS instead of LMS but I always get bad results.
Is there the possibility that the channel is so bad that any algorithm can
work well with the DFE?

>
>
>steveu wrote:
>
>>>
>>>alberto.fuggetta wrote:
>>>
>>>
>>>>Hi,
>>>>
>>>>I'm trying to equalize a channel with sever multipath using a DFE
>> 
>> (12,12)
>> 
>>>>with LMS adaption algorithm.
>>>>The relative power of the replicas are quite high w.r.t the main path.
>> 
>> (max
>> 
>>>>-4 dB). The equalizer is catastrophic.
>>>>From the learning curve analysis I can observe that the error is still
>> 
>> high
>> 
>>>>after processing the training sequence.
>>>>Morover, the forward filter coefficients are very small compared to
the
>>>>feedback filter ones (10^-3 vs 0.2).
>>>>Is there any conclusion I can draw from these info?
>>>>Thanks
>>>
>>>Feedback path adaptation is nasty nonlinear problem. Your filter either

>>>falls into a local minimum or the adaptation is unstable.
>>  
>> Or maybe his symbol timing has not been locked down well enough for a
one
>> sample per symbol equalizer to pull in. Trying 2 samples per symbol
might
>> provide insight into the system's behaviour.
>
>Feedforward coeffs ~ 0 -> error is not correlated with the signal -> 
>adaptation process is not working right.
>
>
>
>> Steve
>> 
>

>I tried with RLS instead of LMS but I always get bad results.
>Is there the possibility that the channel is so bad that any algorithm
can
>work well with the DFE?
>

Can you explain your channel model? or how the impulse response looks?

Chintan

alberto.fuggetta wrote:

> I tried with RLS instead of LMS but I always get bad results.

Does your equalizer work at all, with a trivial channel, no spread?

VLV

Yes, it does.
If I change the channel with a similar one, having same path delays and
lower path gains, the equalizer has good performances. (BER = 10^-4 @ 12
dB)

>
>
>alberto.fuggetta wrote:
>
>> I tried with RLS instead of LMS but I always get bad results.
>
>Does your equalizer work at all, with a trivial channel, no spread?
>
>VLV
>