パラメータ付き線形方程式を省メモリで解くGMRES

We develop Toeplitz infinite GMRES for solving large sparse analytic parameterized systems $A(s)x(s)=z$ at many parameter values. The method exploits a block upper triangular Toeplitz structure in the companion Krylov sequence to construct the Arnoldi process and recover solution approximations without storing full Arnoldi vectors. We derive incremental recurrences requiring $\mathcal{O}(np^2+p^3)$ arithmetic and $\mathcal{O}(np+p^2)$ storage for $p$ Arnoldi steps, excluding factorization setup and assuming linear-cost coefficient actions and triangular solves. A dynamic generator-refreshing strategy addresses cancellation in the basic recurrence. For matrix functions admitting a fixed separated representation, we further develop an implicitly rebased method based on a compact, contractive nilpotent matrix. Both variants preserve the Arnoldi process in exact arithmetic, with exact matrix-function actions required for the rebased method. A residual lower bound for square-summable Taylor coefficients explains a scaling obstruction outside the normalized unit disk. Theoretical analysis and numerical experiments illustrate the computational efficiency of the proposed methods.