Closed mateuszbaran closed 3 years ago
I did not yet have time to investigate, but one thing that I had was that exp/log was sometimes less accurate compared to the ProjectionRetraction
(for very small steps).
It's not this I think, it's also unstable on R^n and for relatively large steps (but small enough that CG2 works well). My guess is that the sequence of operations is wrong somewhere but I don't see where.
Ok, I will also take a look when I find time.
Thanks!
The first short question, where I am not sure in CG3 is: You are implementing an explicit variant of Eq. (2.3), right? Is that algorithm using retractions or the Lie group action? I am not sure there (since exp(X)p is not that explicitly mentioned what that is).
That's a relatively complex thing but yes, Eq. (2.3) is I think the most clear presentation of explicit Runge-Kutta frozen coefficient methods I've seen and what I was trying to implement here. There are no group actions as far as I can tell, these are flows of the frozen coefficients fields (as explained above Eq (2.1)), and these circles are, I think, compositions of these flows. Each e^{h b_r F^r_p}
looks like a tangent moved from some point to p
, and I've interpreted this composition as a linear combination in the tangent space (since all of these tangent vectors are at the same point, but maybe I'm wrong).
But why is there then a p
at the end of the y_k+1
line and why was it not directly written as a sum? I think understanding exactly that might be the solution here (for CG2 there is just one e^...
so there it is clear and sound). I thought that e^(hb_1F_p^1)p
would be exp_p(hb_1F_p^1)
and one would continue from there (though that would mean that hb_2F_p^2
is from another tangent space and now that I write that it seems strange, too?).
Here https://arc.aiaa.org/doi/pdf/10.2514/1.G004578, Eq (29) it reads like it is several exp_p(Xi)
which are then combined with the Lie group action (\circ) ?
No, it doesn't look like Lie group action because above Eq. (2.3) from [^OwrenMarthinsen1999]
there is no mention of Lie groups. It's a composition of flows of integral curves (see above Eq. (2.1)) and I think p
means that we start from p
. But apparently I don't understand how these flows are supposed to work in practice.
Initially implemented in #1 , it seems to not work and was cut from that PR: