Turns out there was a significant error in the experiment I posted yesterday.
For the base station data for the post-processed RTKLIB solution I had downloaded a RINEX file from the CORS website. This contained only GPS data, and there was no option to request GLONASS measurements, so I had assumed this receiver only supports GPS measurements. It turns out though, that the real-time stream from the UNAVCO NTRIP server for the same receiver includes GLONASS measurements as well as GPS.
So, when I compared the Tersus real-time solution to the RTKLIB post-processed solution, the Tersus solution included double-differences for the GLONASS satellites while the RTKLIB solution did not. This caused me to conclude that the Tersus solution algorithm was better than RTKLIB when really it just had more measurements to work with.
The extra GLONASS measurements won’t change the number of measurements used for ambiguity resolution, at least for RTKLIB, since the inter-channel biases prevent the GLONASS measurements being used for ambiguity resolution unless the base and rover receivers are using identical hardware. They do, however, increase the number of double differences used for the float solution which will help indirectly with ambiguity resolution.
Fortunately I was able to download a RINEX file from the UNAVCO website which included the GLONASS measurements and re-run the RTKLIB solution with that data. In this case, the two solutions look much more similar.
Here is a comparison of the initial time-to-fix and the following six times-to-fix after the antenna obstructions. Yellow/green is Tersus, olive/blue is RTKLIB. The Tersus initial acquire is still noticeably faster than the RTKLIB initial acquire but five of six of the RTKLIB re-acquires are faster than the Tersus re-acquires. Of course, time-to-fix by itself is not a very good metric because it does not take into account what level of confidence is required by the two solutions before asserting a fix.
Here is a zoom-in of the higher frequency, smaller amplitude differences. Again, the two solutions look much more similar than they previously did. As I mentioned before, I believe the DC offsets are probably caused by me not paying close attention to the various offsets in the setup. The variation in the z-axis is larger than I am used to seeing and presumably comes from the long baseline.
At this point, the two solutions look similar enough that with this limited amount of data, it is difficult to say one is better than the other. That doesn’t mean they are equal, just that I don’t believe this particular experiment can differentiate between them with any confidence.
Part of the original experiment was to compare the Tersus data to the M8T data. Adding the extra GLONASS base measurements did not appreciably change the M8T solution so the conclusions from that part of the experiment don’t change. The Tersus data with 12 satellite pairs available for ambiguity resolution was still significantly better than the M8T data with only six satellite pairs.
So, maybe I was a little hard on RTKLIB a couple of posts ago when I first looked at it’s dual frequency capabilities. Intuitively it still feels to me that there should be some advantage in explicitly using knowledge of the physical relationship between the dual frequency measurements in the ambiguity resolution but the data supports the commentors who argued otherwise.