# GPS CA Signal Acquisition

The coarse acquisition C/A GPS signal has a symbol rate or chip rate of 1.023 Mcps with code period of 1 ms. The code bandwidth is +/-1.023MHz (to 1st sinc nulls). I'd like to record this signal using a USRP.

What's the minimum complex sample rate that I'd need to use so I could acquire this signal in software? 1 sample/chip = 1.023e6? 2 samples/chip = 2.046e6?

Since the signal is coming from a satellite, there will be an unknown carrier frequency offset due to Doppler. From what I've read, for GPS the Doppler is +/- 10 KHz. What's the maximum residual frequency offset allowable when doing the despreading of the CA code? 500 Hz? Or, what's the SNR loss due as a function of frequency missmatch due to residual Doppler?

• chips/s != bits/s. That's the idea of having chips!! – Marcus Müller Oct 3 '18 at 15:58

What's the maximum residual frequency offset allowable when doing the despreading of the CA code? 500 Hz? Or, what's the SNR loss due as a function of frequency missmatch due to residual Doppler?

The maximum frequency offset is dependent on SNR required for acquisition as the roll-off of correlation versus frequency offset is due entirely to the duration of the correlation.

The magnitude of correlation vs frequency offset is a Sinc function with the first null at 1/T Hz where T is the duration of the correlation interval in seconds. So if your correlate over 1 full PRN period, which is 1 ms, the correlation will go to zero at a 1 KHz Doppler offset. Note the the PRN is repeated 20 times before a data transition (the data rate if 50 Hz), so you could actually correlate over 20 ms for 13 dB more processing gain (Processing gain is 10Log(N)), at the expense of reducing the Doppler tolerance (the first null would appear with a 50 Hz offset in this case!).

A good demonstration of this is the DFT itself for those familiar with it. Each bin in the DFT is a correlation over the time interval of the sequence to the particular frequency in that bin. The DFT can be described as a bank of filters with each filter approaching a Sinc magnitude response as N gets larger (when a rectangular window is used), with the first null appearing in the adjacent bin and then additional nulls in every other bin (hence we show that each bin is uncorrelated to the others). Given the time domain sequence is N samples long in time, and there are N bins also in frequency, then the spacing between bins is 1/N, as we would expect by the description above.

This is demonstrated in the generic graphic below showing the correlation to a 3 Hz sine wave (red waveform) buried in noise over a 2 second correlation interval.

From this we also see how the post correlation magnitude is reduced as a function of frequency offset, while the noise, if white, would not be affected- from which we can determine SNR vs frequency offset.

We also see the trade space involved between processing gain and Doppler offset tolerance. If the received SNR is strong, we can reduce the duration for the correlation and then benefit from a wider frequency acquisition window. For example with GPS we could correlate over half a PRN sequence (0.5 ms) at the expense of 3 dB in processing gain, with the benefit of having the first null out at 2 KHz instead of 1 KHz. In practice I have typically used the metric of half the main lobe as my Doppler range for the correlation (so for GPS would be 1 KHz wide or +/- 500 Hz when correlating over 1 ms), but in weaker signal conditions a smaller search range may be necessary in order to maximize the SNR required for acquisition. (Meaning how far I would step the frequency to repeat the correlation again while searching for signals; typically with a 1ms correlation, I would step 1 KHz but if I felt more SNR may be necessary I may search in finer steps, I could also increase the correlation time which would also require finer frequency steps in the search but here I was just referring to being less sensitive to the roll-off within a set correlation time by not stepping as far in frequency before repeating the correlation again).

Also see this post related to GPS acquisition and a joint delay and frequency acquisition approach that can acquire in one PRN sequence (at the expense of significant processing): GPS signal acquisition

Also to note, once within a Doppler bin, the I and Q correlated outputs can then used to determine the precise Doppler offset and do carrier tracking with implementations described here (where we see the algorithm to determine the precise frequency offset from two adjacent I Q correlation outputs is quite simple): High modulation index PSK - carrier recovery

• Nice summary of the important issues on a topic that isn't well-understood by many. – Jason R Oct 4 '18 at 11:58
• Thank you @JasonR. Also a possible interesting approach would be to window the signal in the correlator which would both widen the Doppler acquisition range (main lobe) and reduce the otherwise relatively large side-lobes reducing false acquisitions at the expense of SNR (given by the window). This is nothing I have ever tried nor am I sure it is a good idea (given the SNR loss when we need all the SNR we can get during acquisition. Your comment made me think of how universal and helpful this 1/T sinc vs freq offset is throughout many applications in signal processing so good to understand. – Dan Boschen Oct 4 '18 at 12:20
• @DanBoschen What's the typical minimum SNR required for acquisition? 13 dB? – Seth Oct 4 '18 at 15:23
• This is ultimately a probability of false alarm vs probability of detection question in addition to speed of acquisition but as a rough guideline 6-10dB should be sufficient for post correlation SNR. You can refer to the error rate vs SNR waterfall curve for BPSK (which is what the post correlation signal really is if you do symbol synchronous correlation) for further insight on performance implications. Note once acquired the signal should be rotated to just I for decisions and Q not used as that will only contribute noise. Use of I and Q is important for Doppler wipe-off (Carrier recovery) – Dan Boschen Oct 4 '18 at 15:36
• @Seth I quickly went through the paper and would prefer a more statistical appraoch based on the standard deviations of the noise at the peak and off peak to determine probability of false alarm vs probability of detection, more consistent with an SNR approach, although I see how they can use the peak levels of each as a quick metric comparison (although hardly robust give the statistics of the possible value of a peak in the normally distributed noise!). I don't see it as of any use for any detailed analysis. My (quick) two cents. However my prior comment re detection dynamic range applies. – Dan Boschen Oct 12 '18 at 21:17

The critical sampling rate is the bandwidth with complex baseband sampling, where a single sample is in fact a complex number.

So, in theory, 1.023 MS/s would be enough. You'd typically go higher to mitigate the filtering imperfection effects as well, as you've noticed, problems due to doppler. (

Also, not all devices can sample at all rates, and 1.023 MHz isn't really an easy rate to achieve.

Look at existing SDR GPS receivers like GNSS-SDR to see how they sample.

What's the maximum residual frequency offset allowable when doing the despreading of the CA code?

That's a bit of a different question.

Let's do a quick estimate here:

• If I remember correctly, the spreading sequence is 10230 chips in length
• That means a symbol takes 10230/1.023e6 = 10⁻² s
• Modulation is BPSK -> maximum allowable phase shift $$\frac \pi2$$, i.e. ¼ of a cycle
• $$\frac{\frac 14}{10^{-2}\,\text{s}}= 25\,\text{Hz}$$ is the maximum allowable residual frequency error.
• Each PRN spreading sequence is actually 1023 bits or chips not 10230. 25 Hz seems way too stringent? I would guess that SNR gradually decreases as residual frequency offset increases, but I don't know. I don't think oversampling helps mitigate frequency offset. – Seth Oct 3 '18 at 17:06
• Seth is correct, the spreading code duration is 1023 chips, or 1 msec. – Jason R Oct 3 '18 at 17:54
• ah, in that case, 250 Hz. Yes, SNR just gradually decreases, but with 500 Hz you'd then get an exact 0 as correlation coefficient, so I wouldn't call that "graceful degradation". – Marcus Müller Oct 3 '18 at 19:08