The following question arise from the proof of Levy's theorem in Richard Bass - Stochastic processes (can be seen via Google books, its on page 77).
So we have $(M_t)_{t\geq 0}$ a continuous local martingale, $M_0=0$ adapted to $\{\mathcal{F}_t\}$ s.t. $<M>_t = t$.
We let $t_0>0$ and define $N_t=M_{t_0+t}-M_{t_0}$ and now its routine (Bass claims) to show $<N>_t=t$.
Had the quadratic variation been linear - then no problem, but by my objection is that we only know the mixed variation is a biliniar form so that $$<N>_t=<M_{t_0+t}-M_{t_0},M_{t_0+t}-M_{t_0}>=<M_{t_0+t}>+<M_{t_0}>-2<M_{t_0},M_{t_0+t}>$$ which I cant see is the sought.
On the other hand, for square integrable martingales we have that $E[(M_S-M_T)^2|\mathcal{F_S}]=E[M_S^2-M_T^2|\mathcal{F_S}]$ (his proposition 9.6 p 56) so in this case the quadratic variation would be linear (one can just check it works) - but how is this not in conflict with the above bilinearity?