Learning-Based Numerical Methods for the Nonlinear Matrix Equation X2 = A.
- 문서 유형
- 연구출판물 · 학술지 논문
- 등재
- KCI
- 저널 순위
- KCI Q1
- 저널
- East Asian Mathematical Journal42(3) · pp.179 · 2026
- 피인용
- KCI 0
- ISSN
- 1226-6973
저자
- 김현민제1저자부산대학교
초록
We study the problem of approximating the principal squareroot A1/2 of a nonsingular matrix A ∈ Cn×n with no eigenvalues on R≤0. Classical iterative methods—Newton, Denman–Beavers, Halley, productform,Pad´e, and the Schur method of Bj¨orck–Hammarling—each requireO(n3) work per step and may converge slowly near the branch cut. We castthe map F : A 7→ A1/2 as an empirical risk minimization problem overdeep ReLU networks, establish a Rademacher-complexity generalizationbound, and show that the principal square-root map is qualitatively approximableby ReLU networks on the compact spectral band An,α,β. Twohybrid algorithms are proposed in which a trained network warm-startsNewton or Denman–Beavers refinement; their iteration savings follow directlyfrom the quadratic convergence of Newton’s method without newconvergence proofs.
키워드
matrix square rootnonlinear matrix equationprincipal square rootNewton iterationDenman–Beavers iterationSchur decompositionlearning-based numerical methodshybrid algorithm.
식별자
- KCI
- ART003342243



로그인
연구물 목록으로