> for any problem there is an ideally sized model that when trained will produce the lowest error rate.

You studied ML before discovery of "double descent"?

https://youtu.be/z64a7USuGX0