Thanks for taking your time to explain all this. It's quite cool. I spent some time with your link and also with the wikipedia page about GRACE. This is quite an ingenious experiment. First I thought the two satellites were at different altitudes, one in LEO and one in, let's say GSO. The high altitude satellite perceives the Earth more or less like a mass point, while the low altitude satellite is much more sensitive to the distribution of mass, especially at the Earth's surface. You could in principle use several satellites at different altitudes, and do some sort of "triangulations". This way, you could potentially learn about the mass distribution deep inside the Earth. Anyway, the actual way it's done is also cool, a lump of mass, like a mountain pulls faster the leading satellite, Jerry, and the second satellite, Tom, lags a little behind. Later on, Tom goes over the same mass and catches on.
That being said, the complexity of this model sounds absolutely humongous. Complex models are exciting for mathy guys to work with, but are inevitably subject to more model risk. As coders say, the number of bugs is roughly proportional with the number of lines of code.
You're going to say that this numerical calculation is bug free. A numerical calculation with 4 thousand mass element (mascons) and distance measurements probably in the billions, and lots and lots of approximations and contributions from the sun and the moon and the planets. Of course you are not going to say it is bug free. Should we trust ice loss calculations coming from GRACE? To be honest, after I read your explanation and all other things about GRACE, I'm less skeptical. I don't fully trust it, but I don't fully distrust it either.
But as you said, this is not that relevant, since the actual result published was not based on GRACE, but rather on ice flow models. Now these models I take with quite a large grain of salt. Why? Models produce estimates, and the estimates generally include an uncertainty level. The uncertainty reflects only how the uncertainty in inputs propagates to the uncertainty in the final result, not the model uncertainty per se.
Take a look at the figure 1 in [1]. It shows the precipitation in Antarctica over 2.5 decades based on several models. The average number (labeled "multi-model mean") shows very little variability from year to year, and you could infer that very little uncertainty as well. However, this is just an artifact of taking the average of several not very highly correlated time series. The model uncertainty is better reflected by the wide range of the numbers produced by the different models. If you look at this range, it's about 200mm/y, which is 10 times the net ice loss we are talking about here.
Do you see why I'm a bit skeptical?
But anyway, thanks again for taking your time with me. I learned something cool today.
[1] https://www.sciencedirect.com/science/article/pii/S187396521...