BFGS update

From Glossary

Jump to: navigation, search

A way to update an approximation of the hessian, used for unconstrained optimization. The update is

B' = B + \frac{(q - Bp)(q - Bp)^T}{p^T(q - Bp)},

where LaTeX: B' is the update from LaTeX: B and LaTeX: p and LaTeX: q are defined in the BFGS method for the LaTeX: k-th iteration. Taking the inverse to compare with the DFP update,

H' = H + \frac{1 + q^T H q}{(q^T p)^2} \left( pp^T \right) - 
\frac{1}{p^T q} \left( pq^T H + H qp^T \right)

where LaTeX: H' is the update from LaTeX: H.

Personal tools