In my experience, the usual trick is to do a few iterations of "linearize and solve the new convex problem". Sometimes, you can get super clever and use LM: htt

In my experience, the usual trick is to do a few iterations of "linearize and solve the new convex problem". Sometimes, you can get super clever and use LM: https://sites.cs.ucsb.edu/~yfwang/courses/cs290i_mvg/pdf/LMA.pdf https://sites.cs.ucsb.edu/~yfwang/courses/cs290i_mvg/pdf/LMA... Look there in equation 6: That's the FIM being left multiplied when solving these types of problems. (under standard gauss-newton step, which is also common) Another tidbit: If you apply matrix inversion lemma to eq 6, you can get the (Extended)Kalman Filter update steps. Somewhat related: https://robotics.stackexchange.com/questions/1180/information-filter-instead-of-kalman-filter-approach/1185#1185 https://robotics.stackexchange.com/questions/1180/informatio...