Author of this post. I have an undergraduate in Applied Mathematics and my training in the "definition -> proposition -> proof" style of mathematics probably comes through in the article more than I wanted it to.
That being said, I began studying Differential Geometry and Lie Groups as part of my graduate degree in Electrical Engineering. Engineers think about problems very differently than mathematicians and I've benefited a lot from taking a more geometric-based and visual approach to learning in the years following my undergraduate.
So, my prescription would be to play around with math ideas when you see them. Create a script to draw what you are trying to visualize. This was my first time using the `manim` library and I gained a deeper appreciation and intuition for the ideas presented in the article even though I've studied them dozens of times!
Overall, learning math is a slow and deliberate exercise. Don't get down on yourself if you don't understand something at first glance. Feel free to pause, verify an idea (either visually or with a formal proof) and then continue on a more firm base of understanding.
Quaternions and SLERP are absolutely a fundamental part of 3D vision (and game development too). However, I wanted to focus this post mainly on the question "why is optimizing on the unit sphere difficult?" As the post stands, it's already quite verbose.
Maybe I'll find some time to do a deep dive on common Lie Groups used in computer vision e.g. SO(3), SE(3) and Sim(3) and also the common representations used for those groups.