Why UKF‘s results are "nan" in the Loreanz attractor with Noisy non-linear observations ?

[FongT8]

Dear Authors,

I have a question about it. Why UKF‘s results are "nan" in the Loreanz attractor with Noisy non-linear observations ?

Looking forward to your reply. Thank you!

Best wishes!

[XiaoyongNI]

'nan' is possibly due to the exceed of Python's largest number limit. When the trakectory length T is long + both f() and h() are highly non-linear + the noise variances Q and R are high, the setting becomes very difficult for UKF to accurately predict. In this case, if we calculate square loss accumulated over the long trajectory, it may easily become a very large number.