Documentation

IntroductoryLecturesOnConvexOptimization_Nesterov_2004.Chap07.Proposition_7_37

noncomputable def quasiNewtonUpdatedHessian {n : } (δ : ) (G : { G : Matrix (Fin n) (Fin n) // G.PosDef }) (g : EuclideanSpace (Fin n)) :
Matrix (Fin n) (Fin n)

The rank-one Hessian update G + (δ / (1 - δ)) * (ggᵀ / ‖g‖_{G}^{*2}) from the relative-accuracy quasi-Newton scheme, written using the canonical Chapter 7 dual norm owner.

Instances For
    theorem quasiNewtonUpdatedHessian_def {n : } (δ : ) (G : { G : Matrix (Fin n) (Fin n) // G.PosDef }) (g : EuclideanSpace (Fin n)) :
    quasiNewtonUpdatedHessian δ G g = G + (δ / (1 - δ) / (positiveDefMatrixNorm G ).dualNorm g ^ 2) Matrix.vecMulVec g.ofLp g.ofLp

    Expanding quasiNewtonUpdatedHessian δ G g gives the displayed rank-one update formula for the next Hessian matrix.

    theorem quasiNewtonUpdatedHessian_posDef {n : } {δ : } ( : δ Set.Ioo 0 1) (G : { G : Matrix (Fin n) (Fin n) // G.PosDef }) (g : EuclideanSpace (Fin n)) :

    The rank-one update preserves positive definiteness when δ ∈ (0, 1).

    theorem quasiNewtonRankOneHessianUpdate_and_inverse {n : } {δ : } ( : δ Set.Ioo 0 1) (Gk : { G : Matrix (Fin n) (Fin n) // G.PosDef }) (gk : EuclideanSpace (Fin n)) :
    quasiNewtonUpdatedHessian δ Gk gk = Gk + (δ / (1 - δ) / (positiveDefMatrixNorm Gk ).dualNorm gk ^ 2) Matrix.vecMulVec gk.ofLp gk.ofLp (quasiNewtonUpdatedHessian δ Gk gk)⁻¹ = (↑Gk)⁻¹ - (δ / (positiveDefMatrixNorm Gk ).dualNorm gk ^ 2) ((↑Gk)⁻¹ * Matrix.vecMulVec gk.ofLp gk.ofLp * (↑Gk)⁻¹)

    Proposition 7.37: if ψ_k is a quadratic estimating function with positive-definite Hessian G_k, then the next Hessian is obtained by the rank-one update G_{k+1} = G_k + (δ / (1 - δ)) * g_k g_kᵀ / ‖g_k‖_{G_k}^{*2}, and its inverse satisfies the Sherman-Morrison identity G_{k+1}^{-1} = G_k^{-1} - δ * G_k^{-1} g_k g_kᵀ G_k^{-1} / ‖g_k‖_{G_k}^{*2}.