|Subject:||Re: [Discuss-gnuradio] Handling of IQ files|
|Date:||Sun, 20 Mar 2016 19:18:13 +0100|
|User-agent:||Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.6.0|
So correlating is digitally mixing something with a predetermined sequence?um, no. Not in the sense that I'd use "correlating".
Mixing in this context is simple point-wise multiplication.
I guess, then, if you have a PSK waveform you might multiply it by something, starting at various offsets in the signal, to try and decode it?A correlation of two digital signals is pretty much that, you take the first signal, multiply it with the second, sum that up, and then shift either signal, rinse, repeat. You get a vector of correlation coefficients which can be interpreted as "how similar is the first signal to the second being shifted by a different sample delays".
Generally, if you have a known pulse shape (i.e. your transmitter takes the transmit symbol, in case of BPSK that would be either +1 or -1, and multiplies it with the pulse shape) you can correlate against that pulse shape, and you'll get the factor with which the pulse was multiplied (which is your transmit symbol combined with whatever your channel does to the signal).
Because we were talking about the DFT/FFT:
Look up the Discrete Fourier Transform in its Matrix form: You take a vector of samples and multiply it with a matrix to get the discrete Fourier transform of , which I shall denote :
(i.e. a square matrix, representing the N-point DFT)
(i.e. column vectors).
Remember how you do matrix-vector multiplication:
To get the result's element in its -th result row and the -th result column, you take row from the left factor (i.e. a row of , which has rows, so ), and multiply it point-wise with -th column of the right factor (, which only has 1 column, so ) and calculate the sum.
Now, point-wise multiplication & summation is what is defined as dot product, so what happens here is that you take your sample vector and calculate the dot products with each row of the DFT Matrix . The rows of the DFT matrices happen to be nothing else but samples of complex sinusoids.
The FFT is different (I actually don't know how it works, other than it operates on 2^n samples), but the output is the same.Exactly, an FFT is just a specific implementation of the DFT (because you really don't want to do vector-matrix multiplication if you can avoid it, because you'd end up doing multiplications and summations), and is mathematically equivalent. In a first step, the fact that the DFT matrix is pretty regular and all entries appear twice is used to save calculations. Modern FFTs do all kind of cool optimizations that I don't understand the least, but what I learnt is that today, the Radix-2 algorithm isn't the only FFT algorithm in existence, and so it turns out that for example the FFTW library, which GNU Radio uses to do its FFTs, can deal with any DFT that has a prime factorization with "nearly no" prime factors >7 .
So, yeah, DSP often gets a lot easier if you just consider it an exercise in linear algebra :)
On 20.03.2016 17:45, Henry Barton wrote:
|[Prev in Thread]||Current Thread||[Next in Thread]|