Hi, this sounds like an interesting problem! Unfortunately, I think choosing the starting value is quite problem dependent. So, maybe the equations indicate that there is some candidate for a good initial value? If there is no obvious candidate, I could think of two possible ways: 1. I guess the easiest is to just start BFGS multiple times from random starting values. For example, you could start BFGS 100 or more times with a different starting value each. Then you can just keep the best result. This might also give you some insight in how sensitive the optimization problem is in regard to the starting value. Do all runs result in similar values or are the results highly different? 2. Alternatively if there are way too much local minima, you could try a global optimization algorithm. There are some implemented in the nlopt library (seems to be available via nloptr in R, but I don't use R). However, I expect those to have a way longer runtime than BFGS. I am also not sure how they scale with the dimensionality of the problem. In both cases you would need to have bounds on the reasonable range where the starting value could be, either for the sampling or the global optimization algorithms. Personally, I would start with the multiple random initializations if I don't have good prior knowledge about what a good starting value would be.

Thanks! I knew the meaning of the "Sci" but somehow never connected this to the prononciation :-/ (And the prononciation is literally on the first page of the documentation :D)

@@FolkerHoffmann lol, all good. Thanks for the great content!

Thank you! I am happy to hear this!

Thank you for the kind words :-)

Thank you!

Thanks! You are of course completely right with your correction!