Skip to content
nonconvexcard

The BFGS update satisfies the secant condition

ProofAdvanced⏱ 30 minAdvanced Optimizationbfgssecantquasi-newtonpositive definiteness

Problem

Let s=xt+1−xts = x_{t+1}-x_t, g=∇f(xt+1)−∇f(xt)g = \nabla f(x_{t+1})-\nabla f(x_t), and

B+=(I−sg⊤g⊤s)B(I−gs⊤g⊤s)+ss⊤g⊤s.B^+ = \Big(I-\frac{sg^\top}{g^\top s}\Big)B\Big(I-\frac{gs^\top}{g^\top s}\Big) + \frac{ss^\top}{g^\top s}.

Verify that B+g=sB^+g = s, and show that B≻0B\succ0 and g⊤s>0g^\top s>0 imply B+≻0B^+\succ0.

Write a complete argument. Every inequality you use should be one you could state precisely.

Solution

Opening this before you have a real attempt on paper is the fastest way to feel like you understand something you cannot reproduce.

Related problems

All problems
SolvedProblemFieldTypeDifficultyTime
Newton's method is affine invariantAdvanced OptimizationProofIntermediate30 min
Conjugate gradient and Krylov optimalityAdvanced OptimizationProofAdvanced1 hour
Local quadratic convergence of Newton's methodAdvanced OptimizationProofAdvanced1 hour