1: \begin{abstract}

2: We summarize the usual implementations of the large $N$ limit of $O(N)$ models and show in detail why and how they can miss some physically important fixed points  when they become  singular in the limit $N\to\infty$. Using Wilson's renormalization group in its functional nonperturbative versions, we show how the singularities build up as $N$ increases. 

3: In the Wilson-Polchinski version of the nonperturbative renormalization group, we show that the singularities are cusps, which become boundary layers for finite but large values of $N$. The corresponding fixed points  being never close to the Gaussian, are out of reach of the usual perturbative approaches. We find four new fixed points  and study them  in all dimensions

4:  and for all $N>0$  and show that they play an important role for the tricritical physics of $O(N)$ models. Finally, we show that some of these fixed points  are bi-valued when they are considered as functions of $d$ and $N$  thus revealing  important and nontrivial homotopy structures.

5:  The  Bardeen-Moshe-Bander phenomenon that occurs at $N=\infty$ and $d=3$ is shown to play a crucial role for the internal consistency of all our results.

6: \end{abstract}

7: