Learning is PAC-possible if enough statistical data and finite 
 - {E_{out}}(g)} \right| > \varepsilon } \right] \leqslant 2 * M * \exp ( - 2{\varepsilon ^2}N)$$)
when M is infinite,we need to estabish a finite quantity that replaces M
 - {E_{out}}(g)} \right| > \varepsilon } \right] \leqslant 2 * {m_h} * \exp ( - 2{\varepsilon ^2}N)$$)
always $$)
,we called it a break point.
positive rays: = N + 1 = {\rm O}(N)$$)
break point at2
positive intervals:  = \frac{1}{2}{N^2} + \frac{1}{2}N + 1$$)
break point at 3
convex sets: no break point
2D perceptrons: break point at 4
So learning is possible again when we find a finite