TED selection and comparison

Hi,

Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
Which one of these work better if we consider

1. Non fading channels
2. Fading channels

How response i.e acquisition and tracking  vary or get affected by amount
of SNR (I think all of them would have degraded performance as output error
signal of all TED is dependent on symbol estimates  i.e decision directed
or data aided), but I want to have expert comments on it as which one would
be better on low SNR and vice versa ? 

Also I wanted to know the roll of excess bandwidth on the variance of error
(self noise) of above TEDs signal at optimum sampling time i.e in lock
state would be greatly appreciated. 
 
Waiting to listen from the expert. Thanks ahead of time.

Regards,
ML

>Hi,
>
>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>Which one of these work better if we consider
>
>1. Non fading channels
>2. Fading channels
>
>How response i.e acquisition and tracking  vary or get affected by amount
>of SNR (I think all of them would have degraded performance as output
error
>signal of all TED is dependent on symbol estimates  i.e decision directed
>or data aided), but I want to have expert comments on it as which one
would
>be better on low SNR and vice versa ? 
>
>Also I wanted to know the roll of excess bandwidth on the variance of
error
>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>state would be greatly appreciated. 
> 
>Waiting to listen from the expert. Thanks ahead of time.
>
>Regards,
>ML
>
>
>

I would refer you to the book "Synchronization Techniques for Digital
Receivers" by Mengali and D'Andrea.  They cover this subject pretty well,
although coverage of fading channels is not as good, but performance vs SNR
and excess bandwidth is in there.
-Doug

>>Hi,
>>
>>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>>Which one of these work better if we consider
>>
>>1. Non fading channels
>>2. Fading channels
>>
>>How response i.e acquisition and tracking  vary or get affected by
amount
>>of SNR (I think all of them would have degraded performance as output
>error
>>signal of all TED is dependent on symbol estimates  i.e decision
directed
>>or data aided), but I want to have expert comments on it as which one
>would
>>be better on low SNR and vice versa ? 
>>
>>Also I wanted to know the roll of excess bandwidth on the variance of
>error
>>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>>state would be greatly appreciated. 
>> 
>>Waiting to listen from the expert. Thanks ahead of time.
>>
>>Regards,
>>ML
>>
>>
>>
>
>I would refer you to the book "Synchronization Techniques for Digital
>Receivers" by Mengali and D'Andrea.  They cover this subject pretty well,
>although coverage of fading channels is not as good, but performance vs
SNR
>and excess bandwidth is in there.
>-Doug
>

Thnak you for the post. Any further comments/suggesstions please ?

Regards,
ML 

On Sun, 07 Apr 2013 12:21:43 -0500, "mite_learner" <94814@dsprelated>
wrote:

>>>Hi,
>>>
>>>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>>>Which one of these work better if we consider
>>>
>>>1. Non fading channels
>>>2. Fading channels
>>>
>>>How response i.e acquisition and tracking  vary or get affected by
>amount
>>>of SNR (I think all of them would have degraded performance as output
>>error
>>>signal of all TED is dependent on symbol estimates  i.e decision
>directed
>>>or data aided), but I want to have expert comments on it as which one
>>would
>>>be better on low SNR and vice versa ? 
>>>
>>>Also I wanted to know the roll of excess bandwidth on the variance of
>>error
>>>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>>>state would be greatly appreciated. 
>>> 
>>>Waiting to listen from the expert. Thanks ahead of time.
>>>
>>>Regards,
>>>ML


>>I would refer you to the book "Synchronization Techniques for Digital
>>Receivers" by Mengali and D'Andrea.  They cover this subject pretty well,
>>although coverage of fading channels is not as good, but performance vs
>SNR
>>and excess bandwidth is in there.
>>-Doug
>>
>
>Thnak you for the post. Any further comments/suggesstions please ?
>
>Regards,
>ML 

How low is low for the SNR?   M&M works pretty low, but if you're
trying to get down to channel capacity with a low rate code you might
need something else.

Otherwise what works best where may depend on things you've not
listed, like modulation type, etc.

The excess bandwidth in the filtering mostly affects the shape of the
transitions in the eye diagrams, and since that's what many TEDs use
for tracking it may affect the TED gain in the loop analysis,
depending on how that was done.

All of what you're asking can be quantified in simulation for a
specific system of interest without a tremendous amount of effort.
IMHO it's a good way to get relevant answers for the applications of
interest.


Eric Jacobsen
Anchor Hill Communications
http://www.anchorhill.com

On 4/5/2013 2:06 PM, mite_learner wrote:
> Hi,
>
> Can anybody provide a comparison of MLTED, M&M TED and early-late TED.

Timing detection is a property of the ambiquity function of the signal.
Different TEDs are better fitted to different types of signals and/or 
different hardware/software requirements.

> Which one of these work better if we consider

Optimal, i.e. maximum likelihood, provides best performance by definition.

> 1. Non fading channels
> 2. Fading channels

> How response i.e acquisition and tracking  vary or get affected by amount
> of SNR (I think all of them would have degraded performance as output error
> signal of all TED is dependent on symbol estimates  i.e decision directed
> or data aided), but I want to have expert comments on it as which one would
> be better on low SNR and vice versa ?

It depends.

> Also I wanted to know the roll of excess bandwidth on the variance of error
> (self noise) of above TEDs signal at optimum sampling time i.e in lock
> state would be greatly appreciated.

It depends.

> Waiting to listen from the expert. Thanks ahead of time.

Start with the basics. Implement Gardner detector over BPSK.

VLV

>Hi,
>
>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>Which one of these work better if we consider
>
>1. Non fading channels
>2. Fading channels
>
>How response i.e acquisition and tracking  vary or get affected by amount
>of SNR (I think all of them would have degraded performance as output
error
>signal of all TED is dependent on symbol estimates  i.e decision directed
>or data aided), but I want to have expert comments on it as which one
would
>be better on low SNR and vice versa ? 
>
>Also I wanted to know the roll of excess bandwidth on the variance of
error
>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>state would be greatly appreciated. 
> 
>Waiting to listen from the expert. Thanks ahead of time.
>
>Regards,
>ML
>
>
>

1. Most TEDs are approximations of ML TED (hint: ML is symbol estimate into
the derivative, early-late is an approximation of that derivative).
Similarly, Gardner, Zero Crossing (ZC) are all approximations of the ML.
M&M however is based on the balancing of autocorrelation of the pulse shape
around its mean, that's why it can work on 1 sample per symbol (as opposed
to most others, which work on 2 samples per symbol).

2. For excess bandwidth or rolloff effect, consider this. M&M works at 1
sample per symbol, so has good performnace only at low rolloffs, higher
rolloffs introduce aliasing and performance deteriorates. Gardner and ZC
work better at higher rolloffs because derivatives are easy to compute with
those sharp transitions.

3. During tracking mode, M&M has no self-noise with Nyquist pulses, ZC self
noise is small too. So at high SNRs, M&M is very good but at low SNRs,
there is not much to differentiate between them at low rolloffs. Remember
that at low SNRs, nothing but coding can pull the signal out of noise.

4. Gardner is non-data-aided, and hence can work without data and with
small frequency offsets. That is a great advantage of recovering the timing
before the frequency and phase. There are modified versions in the
literature with slight improvements at the cost of complexity.

5. References: Books by Meyr 1998 (chapter 5), Mengali 1997 (chapter 7) and
Rice 2008 (chapter 8). Last two are really wonderful for good
understanding.

I'm no expert but I think the best way should be to utilize a feedforward
algorithm first after collecting a few dozens of samples to avoid hangups
and then initalize the tracking loop. For feedforward, see citations of
Oerder and Meyr 1988 paper on Google Scholar.

>On Sun, 07 Apr 2013 12:21:43 -0500, "mite_learner" <94814@dsprelated>
>wrote:
>
>>>>Hi,
>>>>
>>>>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>>>>Which one of these work better if we consider
>>>>
>>>>1. Non fading channels
>>>>2. Fading channels
>>>>
>>>>How response i.e acquisition and tracking  vary or get affected by
>>amount
>>>>of SNR (I think all of them would have degraded performance as output
>>>error
>>>>signal of all TED is dependent on symbol estimates  i.e decision
>>directed
>>>>or data aided), but I want to have expert comments on it as which one
>>>would
>>>>be better on low SNR and vice versa ? 
>>>>
>>>>Also I wanted to know the roll of excess bandwidth on the variance of
>>>error
>>>>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>>>>state would be greatly appreciated. 
>>>> 
>>>>Waiting to listen from the expert. Thanks ahead of time.
>>>>
>>>>Regards,
>>>>ML
>
>
>>>I would refer you to the book "Synchronization Techniques for Digital
>>>Receivers" by Mengali and D'Andrea.  They cover this subject pretty
well,
>>>although coverage of fading channels is not as good, but performance vs
>>SNR
>>>and excess bandwidth is in there.
>>>-Doug
>>>
>>
>>Thnak you for the post. Any further comments/suggesstions please ?
>>
>>Regards,
>>ML 
>
>How low is low for the SNR?   M&M works pretty low, but if you're
>trying to get down to channel capacity with a low rate code you might
>need something else.
>
>Otherwise what works best where may depend on things you've not
>listed, like modulation type, etc.
>
>The excess bandwidth in the filtering mostly affects the shape of the
>transitions in the eye diagrams, and since that's what many TEDs use
>for tracking it may affect the TED gain in the loop analysis,
>depending on how that was done.
>
>All of what you're asking can be quantified in simulation for a
>specific system of interest without a tremendous amount of effort.
>IMHO it's a good way to get relevant answers for the applications of
>interest.
>
>
>Eric Jacobsen
>Anchor Hill Communications
>http://www.anchorhill.com
>

Hi Eric,

A good day to you. 
To your questions, SNR would be [5,10]dB, so minimum is 5dB. Maximum data
rate would be around 250kbps, with 4FSK. Transmission is burst type and not
continuous. So, I need a short acquisition time so that minimum number of
symbols get affected. Channel would be free from multi-paths. 

And I would be glad to have your comment on, How TED gain affect the
performance of a particular TED ?. For example, phase detector gain, Kp, of
early late TED increases as excess bandwidth increases while opposite is
true for M&M. 

Thanks for your value-able feedback.

Regards,
ML

>On 4/5/2013 2:06 PM, mite_learner wrote:
>> Hi,
>>
>> Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>
>Timing detection is a property of the ambiquity function of the signal.
>Different TEDs are better fitted to different types of signals and/or 
>different hardware/software requirements.
>

Signal received is 4-FSK (converted to qpsk), before symbol recovery.  Also
I am using Corei7-2600 CPU 3.40GHz × 8. I would be using C++/python.  

>> Which one of these work better if we consider
>
>Optimal, i.e. maximum likelihood, provides best performance by
definition.
>
>> 1. Non fading channels
>> 2. Fading channels
>
>> How response i.e acquisition and tracking  vary or get affected by
amount
>> of SNR (I think all of them would have degraded performance as output
error
>> signal of all TED is dependent on symbol estimates  i.e decision
directed
>> or data aided), but I want to have expert comments on it as which one
would
>> be better on low SNR and vice versa ?
>
>It depends.
>
>> Also I wanted to know the roll of excess bandwidth on the variance of
error
>> (self noise) of above TEDs signal at optimum sampling time i.e in lock
>> state would be greatly appreciated.
>
>It depends.
>
>> Waiting to listen from the expert. Thanks ahead of time.
>
>Start with the basics. Implement Gardner detector over BPSK.
>
>VLV
>
>

Thanks for your suggestion and comment.

Regards,
ML 

>>Hi,
>>
>>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>>Which one of these work better if we consider
>>
>>1. Non fading channels
>>2. Fading channels
>>
>>How response i.e acquisition and tracking  vary or get affected by
amount
>>of SNR (I think all of them would have degraded performance as output
>error
>>signal of all TED is dependent on symbol estimates  i.e decision
directed
>>or data aided), but I want to have expert comments on it as which one
>would
>>be better on low SNR and vice versa ? 
>>
>>Also I wanted to know the roll of excess bandwidth on the variance of
>error
>>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>>state would be greatly appreciated. 
>> 
>>Waiting to listen from the expert. Thanks ahead of time.
>>
>>Regards,
>>ML
>>
>>
>>
>
>1. Most TEDs are approximations of ML TED (hint: ML is symbol estimate
into
>the derivative, early-late is an approximation of that derivative).
>Similarly, Gardner, Zero Crossing (ZC) are all approximations of the ML.
>M&M however is based on the balancing of autocorrelation of the pulse
shape
>around its mean, that's why it can work on 1 sample per symbol (as
opposed
>to most others, which work on 2 samples per symbol).
>
>2. For excess bandwidth or rolloff effect, consider this. M&M works at 1
>sample per symbol, so has good performnace only at low rolloffs, higher
>rolloffs introduce aliasing and performance deteriorates. Gardner and ZC
>work better at higher rolloffs because derivatives are easy to compute
with
>those sharp transitions.
>
>3. During tracking mode, M&M has no self-noise with Nyquist pulses, ZC
self
>noise is small too. So at high SNRs, M&M is very good but at low SNRs,
>there is not much to differentiate between them at low rolloffs. Remember
>that at low SNRs, nothing but coding can pull the signal out of noise.
>
>4. Gardner is non-data-aided, and hence can work without data and with
>small frequency offsets. That is a great advantage of recovering the
timing
>before the frequency and phase. There are modified versions in the
>literature with slight improvements at the cost of complexity.
>
>5. References: Books by Meyr 1998 (chapter 5), Mengali 1997 (chapter 7)
and
>Rice 2008 (chapter 8). Last two are really wonderful for good
>understanding.
>
>I'm no expert but I think the best way should be to utilize a feedforward
>algorithm first after collecting a few dozens of samples to avoid hangups
>and then initalize the tracking loop. For feedforward, see citations of
>Oerder and Meyr 1988 paper on Google Scholar.
>


Bundle of thanks for taking time and for writing such a detailed answer.
One further question, which one is better in terms of acquisition time?
considering 4-FSK transmission with max date rate of 250Kbps, (4-FSK is
converted to qpsk before symbol timing ). Actually, I am more concerned
with acquisition time as transmission would be bursty.

I have those books and 'll surely go through the paper as well:)

Regards,
ML

On Sun, 07 Apr 2013 23:50:15 -0500, "mite_learner" <94814@dsprelated>
wrote:

>>On Sun, 07 Apr 2013 12:21:43 -0500, "mite_learner" <94814@dsprelated>
>>wrote:
>>
>>>>>Hi,
>>>>>
>>>>>Can anybody provide a comparison of MLTED, M&M TED and early-late TED.
>>>>>Which one of these work better if we consider
>>>>>
>>>>>1. Non fading channels
>>>>>2. Fading channels
>>>>>
>>>>>How response i.e acquisition and tracking  vary or get affected by
>>>amount
>>>>>of SNR (I think all of them would have degraded performance as output
>>>>error
>>>>>signal of all TED is dependent on symbol estimates  i.e decision
>>>directed
>>>>>or data aided), but I want to have expert comments on it as which one
>>>>would
>>>>>be better on low SNR and vice versa ? 
>>>>>
>>>>>Also I wanted to know the roll of excess bandwidth on the variance of
>>>>error
>>>>>(self noise) of above TEDs signal at optimum sampling time i.e in lock
>>>>>state would be greatly appreciated. 
>>>>> 
>>>>>Waiting to listen from the expert. Thanks ahead of time.
>>>>>
>>>>>Regards,
>>>>>ML
>>
>>
>>>>I would refer you to the book "Synchronization Techniques for Digital
>>>>Receivers" by Mengali and D'Andrea.  They cover this subject pretty
>well,
>>>>although coverage of fading channels is not as good, but performance vs
>>>SNR
>>>>and excess bandwidth is in there.
>>>>-Doug
>>>>
>>>
>>>Thnak you for the post. Any further comments/suggesstions please ?
>>>
>>>Regards,
>>>ML 
>>
>>How low is low for the SNR?   M&M works pretty low, but if you're
>>trying to get down to channel capacity with a low rate code you might
>>need something else.
>>
>>Otherwise what works best where may depend on things you've not
>>listed, like modulation type, etc.
>>
>>The excess bandwidth in the filtering mostly affects the shape of the
>>transitions in the eye diagrams, and since that's what many TEDs use
>>for tracking it may affect the TED gain in the loop analysis,
>>depending on how that was done.
>>
>>All of what you're asking can be quantified in simulation for a
>>specific system of interest without a tremendous amount of effort.
>>IMHO it's a good way to get relevant answers for the applications of
>>interest.
>>
>>
>>Eric Jacobsen
>>Anchor Hill Communications
>>http://www.anchorhill.com
>>
>
>Hi Eric,
>
>A good day to you. 
>To your questions, SNR would be [5,10]dB, so minimum is 5dB. Maximum data
>rate would be around 250kbps, with 4FSK. Transmission is burst type and not
>continuous. So, I need a short acquisition time so that minimum number of
>symbols get affected. Channel would be free from multi-paths. 
>
>And I would be glad to have your comment on, How TED gain affect the
>performance of a particular TED ?. For example, phase detector gain, Kp, of
>early late TED increases as excess bandwidth increases while opposite is
>true for M&M. 

The TED gain just has to be taken into account in the design of the
timing-recovery loop.   Designing for the lower expected gain as SNR
changes usually works well, but, as always, it's worth checking the
extremes to see if additional measures need to be taken.

>Thanks for your value-able feedback.
>
>Regards,
>ML

Eric Jacobsen
Anchor Hill Communications
http://www.anchorhill.com