In literature about trust region methods I found the following statement:
First they define for $t\in (0,1)$
$f(x_k+p)=f(x_k)+\nabla f(x_k)^Tp+\frac{1}{2}p^T\nabla^2f(x_k+tp)p$
and by using $B_k$ as an approximation to the Hessian (with $B_k$ symmetric)
$m_k(p)=f(x_k)+\nabla f(x_k)^Tp+\frac{1}{2}p^TB_kp$.
Then it says the difference between $m_k(p)$ and $f(x_k+p)$ is $\mathcal{O}(\Vert p\Vert^2)$.
I'm not sure, why this statement about the difference holds.
Obviously the difference between those functions is in the second-order term. The only statement I remember from previous lessons for approximations of Hessians is that the difference is $\mathcal{o}(\Vert x_{k+1}-x_{k}\Vert)$. Since $x_{k+1}$ is in this case equal to $x_k+p$ it follows $\mathcal{o}(\Vert p\Vert)$, but this doesn't lead to the statement.
I would be really happy for advices.
Edit: $f(x_k+p)$ is the Taylor-series expansion around $x_k$ from $f$ and $m_k(p)$ is the model function, which should be minimized in every trust region iteration.