The stored gradient difference of a PCG pair is the line-search step times the curvature step
ProvedLimitedBFGS.SQN.pcgPair_y_eqlimitedbfgslinear-algebrapcg
Let pcgPair A b H₀ x₀ k = (s_k, y_k) be the secant pair stored after the step from x_k, on the quadratic f(x) = ½ xᵀAx + bᵀx with g(x) = Ax + b.
Claim. y_k = A s_k.
Why. y_k = g_{k+1} - g_k = A(x_{k+1} - x_k) = A s_k, since the gradient of this quadratic is the affine map x ↦ Ax + b.
Preamble
import Mathlib import Definitions.Def_LimitedBFGS_SQN_pcgPairs import Definitions.Def_LimitedBFGS_SQN_sqnIter open Matrix
Formal statement
namespace LimitedBFGS.SQN
/-- The second component of `pcgPair A b H₀ x₀ k` is `y_k = grad A b x_{k+1} - grad A b x_k`,
and on the quadratic `grad A b (x_{k+1} - x_k) = A *ᵥ s_k`. -/
theorem pcgPair_y_eq {n : ℕ} (A : Matrix (Fin n) (Fin n) ℝ) (b : Fin n → ℝ)
(H₀ : Matrix (Fin n) (Fin n) ℝ) (x₀ : Fin n → ℝ) (k : ℕ) :
(pcgPair A b H₀ x₀ k).2
= A *ᵥ (pcgPair A b H₀ x₀ k).1 := by sorry
end LimitedBFGS.SQN
Source
Nocedal, Updating Quasi-Newton Matrices with Limited Storage, Math. Comp. 35 (1980), p. 775, Property (b).