BFGS

Newton’s Method

$$\begin{aligned} x_{k+1} &= x_k - [H(x_k)]^{-1}\nabla f(x_k)^\intercal \end{aligned}$$

Quasi Newton’s Method

$$\begin{aligned} x_{k+1} &= x_k - \alpha_kS_k {\nabla f(x_k)}^{T} \end{aligned}$$

If $S_k$ is inverse of Hessian, then method is Newton’s iteration; if $S_k=I$, then it is steepest descent

BFGS

BFGS is a quasi newtons method where we approximate inverse of Hessian by $B_k$. The search direction $p_k$ is determined by solving $$B_kp_k = -\nabla f(x_k)$$ A line search is performed in this search direction to find next point $x_{k+1}$ by minimising $f(x_k+\gamma p_k)$. The approximation to hessian is then updated as $$\begin{aligned} B_{k+1} &= B_k + \alpha_k u_ku_k^\intercal + \beta_k v_kv_k^\intercal \\ u_k &= \nabla f(x_{k+1})-\nabla f(x_k) \\ \alpha_k &= \frac{1}{\alpha u_k^\intercal p_k} \\ v_k &= B_kp_k \\ \beta_k &= \frac{-1}{p_k^\intercal B_kp_k} \end{aligned}$$




tags: optimisation