Algorithm 1.
1: M ← initial matrix (M0). |
2: δ ← initial damping scalar (δ0). |
3: repeat |
4: Compute gradient (g) and Hessian (H) from Jacobian terms (Jj). |
5: repeat |
6: δ ← δ × 10. |
7: Find ΔM from vec (ΔM) ← (H + δI)−1g. |
8: until f (M− ΔM) < f (M). |
9: M ← M − ΔM. |
10: δ ← δ × 10−2. |
11: Orthogonalize M (keep mean vector t unchanged). |
12: until convergence. |