Re: [Discuss-gnuradio] Spectrum distortion in wideband OFDM samples

[Per Zetterberg]

On Fri, 2011-01-28 at 02:09 -0500, Sangho Oh wrote:
> Hello, I am transmitting 20Mhz OFDM symbols using UHD code.
> When I checked the PSD of the received samples using UHD
> (rx_samples_to_file)
> I found there is a significant difference in PSD between the
> transmitted and received signal. 
> 
> 
> Reception decimation is 4 and I used re-sampled data (4/5) for OFDM
> demodulation.
> Here is the figure showing the PDS difference. 
> 
> 
> http://i1181.photobucket.com/albums/x423/notilas/untitled.jpg
> 
> 
> Someone guessed that this is a natural thing related to the low pass
> filter in RF parts of Gnuradio.
> However, I would like to ask valuable comments from radio experts.  
> 


The spectrum looks very strange so I would guess saturation or
something.

However, I have had some experiences I would like to discuss. I have
experimented with OFDM waveforms on USRP1, USRP2 and also another
home-brewed testbed. What I generally find is that an oversampling of a
factor of two is required. Thus with 25MHz sample-rate I can squeeze in
a maximum 12.5MHz signal. I can use more but at the expense of much
worse performance. I think the main problem is nonlinearities. The
nonlinearities causes intermods. If the sampling frequency is less than
twice the bandwidth, 3rd order intermods will fold into the desired
signal. 

Another observation is that, given that I have the luxury of two times
oversampling (or more), it is better to use an fft on all the samples
and null (don't use) some of the subcarriers, than do up/downsampling
after/before the fft. I don't why, but I guess that it would not be very
hard to prove it with maths.

I wish I had a reference on these issues (conference, journal-paper or
book). Most papers are ignoring all these practical issues .... 


BR/
Per