Question about Accuracy of Carrier Frequency Estimation in OFDM

I was simulating an OFDM system, and used 802.11a type preamble (short and
long training symbols). I found that correlation based coarse frequency
synchronization with a delay of 16 samples is not very accurate. Then, even
using the long training for fine frequency estimation (same correlation but
with 64 samples delay) doesn't give very good results.

For example, if I plot the BER vs SNR curves, the BER curve goes flat due
to residual frequency offset, or if I give this input to a decoder, the
results are very bad due to the remaining CFO.

Any suggestions on accurately estimating the CFO in OFDM systems? If I use
a data-aided method, it will increase latency and add a lot of complexity
related to finding a complex sinusoid in noise.

Thanks in advance.	 

_____________________________		
Posted through www.DSPRelated.com

>I was simulating an OFDM system, and used 802.11a type preamble (short
and
>long training symbols). I found that correlation based coarse frequency
>synchronization with a delay of 16 samples is not very accurate. Then,
even
>using the long training for fine frequency estimation (same correlation
but
>with 64 samples delay) doesn't give very good results.
>
>For example, if I plot the BER vs SNR curves, the BER curve goes flat due
>to residual frequency offset, or if I give this input to a decoder, the
>results are very bad due to the remaining CFO.
>
>Any suggestions on accurately estimating the CFO in OFDM systems? If I
use
>a data-aided method, it will increase latency and add a lot of complexity
>related to finding a complex sinusoid in noise.
>
>Thanks in advance.	 
>
>_____________________________		
>Posted through www.DSPRelated.com
>

I just add that to have a negligible effect on the data detection specially
for later symbols in a block, the frequency offset should be reduced to a
mean square error of at least 10^-6 - 10^-8.	 

_____________________________		
Posted through www.DSPRelated.com

I thought the idea of a preamble is just to get things started, you still need to do ongoing carrier synchronization. 

They go together like horse and carriage.

Mark

On Tuesday, April 29, 2014 7:07:26 AM UTC-4, commsignal wrote:
> I was simulating an OFDM system, and used 802.11a type preamble (short and
> long training symbols). I found that correlation based coarse frequency
> synchronization with a delay of 16 samples is not very accurate. Then, even
> using the long training for fine frequency estimation (same correlation but
> with 64 samples delay) doesn't give very good results.
> 
> For example, if I plot the BER vs SNR curves, the BER curve goes flat due
> to residual frequency offset, or if I give this input to a decoder, the
> results are very bad due to the remaining CFO.
> 
> Any suggestions on accurately estimating the CFO in OFDM systems? If I use
> a data-aided method, it will increase latency and add a lot of complexity
> related to finding a complex sinusoid in noise.
> 
> Thanks in advance.	 
> 

Are you using exactly the format and preambles defined in those standards, or are you designing your own? 

I think you already know that there's no straight answer to your questions, which while they are excellent questions they are truly design questions. And one can't give a complete answer without knowing the channel and system constraints. 

I'm sure you already know to check whether the residual CFO/SCO is causing ICI, and to check how much SNR you have and whether you can differential encoding. So there's easily a textbook's worth of answers here. 

Are you at least able to get a quantitative idea of how much CFO/SCO you can tolerate? Or better yet, where you are relative to various performance bottlenecks? 

Hi,

if you know the signal (this -is- data aided) and can guarantee an error of
less than half a turn over a time T, you can do this:

- multiply received with conjugate of reference over T
- get the phase of the product
- apply weighting with magnitude
- apply the same weight to the phase of a unit frequency error (which is a
straight line)
- do a least-sqares fit
- the result is your error estimate

I've used this to capture a known signal from a free-running cellphone in
an automated test. The hard part is to get close enough, the above method
itself is then pretty fast. Increase the observation window as the
uncertainty goes down.
It's not particularly elegant, but gets the job done without spinning my
own basestation. How accurate exactly, don't know, but the EVM readings
looked as expected.

It would work in a simulation, sure, but then why not simply use the known
value? 	 

_____________________________		
Posted through www.DSPRelated.com