On Tue, Jun 9, 2009 at 5:12 AM, John Melton <span dir="ltr"><<a href="mailto:John.Melton@sun.com">John.Melton@sun.com</a>></span> wrote:<br><div class="gmail_quote"><div> </div><blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;">
I have modified ghpsdr to have the FFT use 32768 samples (4096 actual samples and the rest padded with zeros). This allows me to do a 16X zoom on the bandscope (i.e. I can see 3840 KHz) where I get a 1:1 sample/pixel display. You really cannot see much additional information. It is simply the original display stretched out. After all, it only had 4096 samples to start with.</blockquote>
<div><br>The way to think of what you're seeing is something like this. Increasing the number of points in the FFT increases the number of bins *between* the ones you see already. But unless you do some other, additional work, there will be leakage among the new bins. That's why I said you need to be careful about windowing. Furthermore, to keep up the accuracy of the power estimates, you need to use other techniques to reduce the variance in each bin. That involves making multiple looks at data in each frame. Simple averaging amond frames is merely one of the simplest ways to do that.<br>
<br>Polyphase and multitaper estimation is probably the most common method used currently since it's not too difficult to implement and the computational complexity is not bad. However, if you're really concerned with zooming in on arbitrary segments and getting truly high-resolution power and frequency estimates, then ARMA techniques are where you want to go. But they can be expensive.<br>
<br>73<br>Frank<br>AB2KT<br></div></div><br>-- <br>You better buy yourself a 'lectric guitar / You better play / There's no other way / To beat the blues... -- Pete Townshend, "My Baby"<br>