OFDM timing synchronization

Hi,

I've been struggling with OFDM timing synchronization using Schmidl 
algorithm for some time but I can't figure out how to find the sample at 
which the symbol starts. Authors seem to have the same problem but 
apparently they don't mention how to deal with it -- or is it just me 
not understanding it?

The timing metric shows me where to sample (I actually only use P metric 
because AGC term R gives false peaks for non-training symbols) and I 
sometimes get my symbols circularly shifted. This happens when maximum 
of the metric is slightly off by one or two samples.

For example after ifft of the first training symbol I transmit:

    first half    second half    cyclic prefix
    1 2 3 4 5      1 2 3 4 5       1  2  3
----------------------------------------------->
    0 1 2 3 4      5 6 7 8 9      10 11 12        sample number N


but the optimal timing point is for N=1. Knowing that useful part of 
OFDM symbol is 10 samples I take 10 samples starting from the second one 
which gives

   2 3 4 5   1 2 3 4 5   1

This sequence carries the same information but circularly shifted and 
performing fft on it does not give what was put into ifft at the 
transmitter.

How can I find out that a shift occured and by how many samples? I can't 
simply try all possible shifts to find the correct offset because 
subcarriers are altered in the channel and I need to know this correct 
offset within the symbol to perform channel estimation by 
(circshift(symb_received,n) ./ symb_expected)

I also tried to investigate Park algorithm that gives very sharp metric. 
Authors say that the proposed preamble is of the form Ppro = [C D C* D*] 
and it is possible to generate this pattern by transmittng a PN sequence 
of BPSK points on even frequencies and zeros on odd frequencies. So for 
example
   s = [1 0 1 0 -1 0 1 0 -1 0 -1 0 1 0 1 0] should work but after ifft 
it simply gives Ppro2 = [C D C D] -- no conjugates in the second half of 
the symbol. I wouldn't even expect it to yield something else. But then 
how to generate symbol of the form [C D C* D*]?

Alternatively, please recommend another algorithms I could try. My 
channel is dispersive.

Thanks!

>Hi,
>
>I've been struggling with OFDM timing synchronization using Schmidl 
>algorithm for some time but I can't figure out how to find the sample at 
>which the symbol starts. Authors seem to have the same problem but 
>apparently they don't mention how to deal with it -- or is it just me 
>not understanding it?
>
>The timing metric shows me where to sample (I actually only use P metric 
>because AGC term R gives false peaks for non-training symbols) and I 
>sometimes get my symbols circularly shifted. This happens when maximum 
>of the metric is slightly off by one or two samples.
>
>For example after ifft of the first training symbol I transmit:
>
>    first half    second half    cyclic prefix
>    1 2 3 4 5      1 2 3 4 5       1  2  3
>----------------------------------------------->
>    0 1 2 3 4      5 6 7 8 9      10 11 12        sample number N
>
>
>but the optimal timing point is for N=1. Knowing that useful part of 
>OFDM symbol is 10 samples I take 10 samples starting from the second one 
>which gives
>
>   2 3 4 5   1 2 3 4 5   1
>
>This sequence carries the same information but circularly shifted and 
>performing fft on it does not give what was put into ifft at the 
>transmitter.
>
>How can I find out that a shift occured and by how many samples? I can't 
>simply try all possible shifts to find the correct offset because 
>subcarriers are altered in the channel and I need to know this correct 
>offset within the symbol to perform channel estimation by 
>(circshift(symb_received,n) ./ symb_expected)
>
>I also tried to investigate Park algorithm that gives very sharp metric. 
>Authors say that the proposed preamble is of the form Ppro = [C D C* D*] 
>and it is possible to generate this pattern by transmittng a PN sequence 
>of BPSK points on even frequencies and zeros on odd frequencies. So for 
>example
>   s = [1 0 1 0 -1 0 1 0 -1 0 -1 0 1 0 1 0] should work but after ifft 
>it simply gives Ppro2 = [C D C D] -- no conjugates in the second half of 
>the symbol. I wouldn't even expect it to yield something else. But then 
>how to generate symbol of the form [C D C* D*]?
>
>Alternatively, please recommend another algorithms I could try. My 
>channel is dispersive.
>
>Thanks!
>
Both Schmidl and Park algorithms assume the AWGN channel. The  property of
two identical symbols is destroyed by the dispersive channel. A better
approach is to compute the cross correlation between the received signals
and  the preamble, but this approach increases the complexity of frame
synchronization.