Eq. (43): the self-scaled direction under the three orthogonalities
ProvedShannoCG.SCONB.sconbDirection_eq_43conjugate-gradientp2o-batch-p200bp2o-gran-per-chapterp2o-plan-paperp2o-v1quasi-newton
Let with , and let be the self-scaled two-update direction ( from (37), its BFGS update with ). Assume
- (exact search at step );
- ;
- (conjugacy).
Then
This is the algebraic core of Shanno's reduction: once the three orthogonalities hold, the self-scaled direction is an explicit combination of , and .
Formalization Note Stated for arbitrary vectors. No condition on is needed, since both sides use the same quotient.
Preamble
import Mathlib import Definitions.Def_ShannoCG_SCONB_gammaScale import Definitions.Def_ShannoCG_SCONB_sconbDirection open Matrix
Formal statement
namespace ShannoCG.SCONB
/-- Shanno, *Conjugate Gradient Methods with Inexact Searches*, Math. Oper. Res. 3(3) (1978), §IV, p. 251 (PDF 8), eq. (43): if `p_k' g_{k+1} = 0`, `p_t' g_{k+1} = 0` and
`y_k' p_t = 0`, the self-scaled two-update direction is
`d_{k+1} = −(p_t'y_t / y_t'y_t) g_{k+1} + (y_t'g_{k+1} / y_t'y_t) p_t
+ (y_k'g_{k+1} / p_k'y_k)(p_t'y_t / y_t'y_t) p_k`.
**Formalization Note.** Stated generically (`g` plays `g_{k+1}`). The three orthogonalities are
hypotheses: exact search at `k`, the fact M4 (`p_t'g_{k+1} = 0`) and the conjugacy `y_k'p_t = 0`.
`p_t ⬝ᵥ y_t ≠ 0` keeps (37) genuine; no condition on `p_k ⬝ᵥ y_k` is needed since both sides
use the same quotient. -/
theorem sconbDirection_eq_43 {n : ℕ} (pt yt pk yk g : Fin n → ℝ) (hpy : pt ⬝ᵥ yt ≠ 0)
(hexact_k : pk ⬝ᵥ g = 0) (horth : pt ⬝ᵥ g = 0) (hconj : yk ⬝ᵥ pt = 0) :
sconbDirection pt yt pk yk g =
-(gammaScale pt yt • g) + ((yt ⬝ᵥ g) / (yt ⬝ᵥ yt)) • pt
+ ((yk ⬝ᵥ g) / (pk ⬝ᵥ yk) * gammaScale pt yt) • pk := by sorry
end ShannoCG.SCONB
Source
Shanno, Conjugate Gradient Methods with Inexact Searches, Math. Oper. Res. 3(3) (1978) 244–256, DOI 10.1287/moor.3.3.244, p. 251 (PDF p. 8), §IV, eq. (43)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.