Hello all,

I have a small query regarding the GSM training sequence. The
recommendation GSM 05.02 suggests that bits "61 to 86" which makes
"26=5+16+5" bits are for training, whereas in many online tutorials I
see only the middle "16" bits mentioned as training bits.

Can anyone suggest should I consider 26-bits or 16-bits for a good
channel estimation.

Thanks,

Best regards,
Vimal

v...@yahoo.com (Vimal) wrote in message
news:<b...@posting.google.com>...
> Hello all,
> 
> I have a small query regarding the GSM training sequence. The
> recommendation GSM 05.02 suggests that bits "61 to 86" which makes
> "26=5+16+5" bits are for training, whereas in many online tutorials I
> see only the middle "16" bits mentioned as training bits.
> 
> Can anyone suggest should I consider 26-bits or 16-bits for a good
> channel estimation.
> 
> Thanks,
> 
> Best regards,
> Vimal

I assume you are interested in normal burst processing where 26 is the
training sequence length as per ETSI spec.
It is often necessary to synchronize the burst.It is done according to
maximum energy position calculation -some times known as long
correlation where you need the full training sequence to estimate the
position for better accuracy.Once burst position is detected mid
training sequence can be used to estimate the  CIR. Sometimes it is
known as short correlation where it employs a sliding window and
calculates the energy of each window and select the maximum
correponding the estimated channel taps.

It is very brief explanation but hope that helps.
Santosh

Thanks Santosh for your help,

So, Can I assume the following:
26-bits are used to identify the burst start-end points and then 16
middle bits are used to estimate CIR using
channel-sounding/correlation-method/MMSE.

Pl advice.

Cheers,
Vimal

s...@ntlworld.com (santosh nath) wrote in message
news:<6...@posting.google.com>...
> v...@yahoo.com (Vimal) wrote in message
news:<b...@posting.google.com>...
> > Hello all,
> > 
> > I have a small query regarding the GSM training sequence. The
> > recommendation GSM 05.02 suggests that bits "61 to 86" which makes
> > "26=5+16+5" bits are for training, whereas in many online tutorials I
> > see only the middle "16" bits mentioned as training bits.
> > 
> > Can anyone suggest should I consider 26-bits or 16-bits for a good
> > channel estimation.
> > 
> > Thanks,
> > 
> > Best regards,
> > Vimal
> 
> I assume you are interested in normal burst processing where 26 is the
> training sequence length as per ETSI spec.
> It is often necessary to synchronize the burst.It is done according to
> maximum energy position calculation -some times known as long
> correlation where you need the full training sequence to estimate the
> position for better accuracy.Once burst position is detected mid
> training sequence can be used to estimate the  CIR. Sometimes it is
> known as short correlation where it employs a sliding window and
> calculates the energy of each window and select the maximum
> correponding the estimated channel taps.
> 
> It is very brief explanation but hope that helps.
> Santosh

v...@yahoo.com (Vimal) wrote in message
news:<b...@posting.google.com>...
> Thanks Santosh for your help,
> 
> So, Can I assume the following:
> 26-bits are used to identify the burst start-end points and then 16
> middle bits are used to estimate CIR using
> channel-sounding/correlation-method/MMSE.
> 
> Pl advice.
> 
> Cheers,
> Vimal
> 

I guess so.We do not need to bother about burst end position since the
burst length is fixed and we are interested in burst by burst
synchronization. I guess initially you can try LSE based estimate or
double correlation.You can go for LMMSE smoothing - it depends on your
requirement.
Cheers
santosh
> s...@ntlworld.com (santosh nath) wrote in message
news:<6...@posting.google.com>...
> > v...@yahoo.com (Vimal) wrote in message
news:<b...@posting.google.com>...
> > > Hello all,
> > > 
> > > I have a small query regarding the GSM training sequence. The
> > > recommendation GSM 05.02 suggests that bits "61 to 86" which makes
> > > "26=5+16+5" bits are for training, whereas in many online tutorials I
> > > see only the middle "16" bits mentioned as training bits.
> > > 
> > > Can anyone suggest should I consider 26-bits or 16-bits for a good
> > > channel estimation.
> > > 
> > > Thanks,
> > > 
> > > Best regards,
> > > Vimal
> > 
> > I assume you are interested in normal burst processing where 26 is the
> > training sequence length as per ETSI spec.
> > It is often necessary to synchronize the burst.It is done according to
> > maximum energy position calculation -some times known as long
> > correlation where you need the full training sequence to estimate the
> > position for better accuracy.Once burst position is detected mid
> > training sequence can be used to estimate the  CIR. Sometimes it is
> > known as short correlation where it employs a sliding window and
> > calculates the energy of each window and select the maximum
> > correponding the estimated channel taps.
> > 
> > It is very brief explanation but hope that helps.
> > Santosh

Hi Santosh,

Assuming that I have perfect synchronization, I tried LMS and Least
Squares to estimate the CIR. Also assuming that the channel lenght is
5.

If I consider 16-bits for channel extimation, ML ie reducing to Least
Squares for Gaussian noise gives me better estimate than LMS. However,
if I consider 22 ie 5+16+1 (assuming estimator is 5-tap FIR, the last
4-training-bits are used to fill-up the first 4-taps) LMS performances
way better than ML. Another important observation that the estimation
error in case of 16-bits is less than that for 22 or 26-bits.

I guess from observations above 16-bits could be the actual training
bits for CIR estimation.

Cheers,
Vimal

s...@ntlworld.com (santosh nath) wrote in message
news:<6...@posting.google.com>...
> v...@yahoo.com (Vimal) wrote in message
news:<b...@posting.google.com>...
> > Thanks Santosh for your help,
> > 
> > So, Can I assume the following:
> > 26-bits are used to identify the burst start-end points and then 16
> > middle bits are used to estimate CIR using
> > channel-sounding/correlation-method/MMSE.
> > 
> > Pl advice.
> > 
> > Cheers,
> > Vimal
> > 
> 
> I guess so.We do not need to bother about burst end position since the
> burst length is fixed and we are interested in burst by burst
> synchronization. I guess initially you can try LSE based estimate or
> double correlation.You can go for LMMSE smoothing - it depends on your
> requirement.
> Cheers
> santosh