Some results on Krylov solvability in Banach space and connections to spectral theory

Abstract: This article contains the first steps in a general analysis of the problem of Krylov solvability of the inverse linear problem in a Banach space. In contrast to the well-studied Hilbert space setting, the Banach space setting presents particular difficulties in creating the connection between Krylov solvability and structural properties of the Krylov subspace itself. At the centre of this is the fact that the closed Krylov subspace may not always have a topological complement. We also develop spectral tools in order to attack the problem using the resolvent operator and exploiting its holomorphic properties on the resolvent set.

• The paper demonstrates that a complemented Krylov subspace in Banach spaces is neither necessary nor sufficient for Krylov solvability.
• It uses explicit constructions in sequence spaces to contrast Hilbert and Banach theories, exemplifying with weighted backward shift operators.
• Resolvent and spectral methodologies are developed to establish conditions for Krylov solvability, impacting algorithm design and stability analysis.

Krylov Solvability in Banach Spaces: Structural and Spectral Perspectives

Introduction

This paper systematically investigates the underlying mathematical structure and solvability of inverse linear problems via Krylov subspace methods in the context of Banach spaces. While the theory and practice of Krylov methods are mature in Hilbert spaces, their extension to Banach spaces is highly nontrivial due to the absence of a canonical inner product and associated geometrical tools. The paper focuses on two main axes: the interplay between the structural properties of Krylov subspaces (especially the existence of topological complements) and the solvability of associated inverse problems, and the utility of spectral methods, specifically the resolvent operator, for characterizing Krylov solvability of closed (possibly unbounded) operators.

Krylov Subspaces in Banach Spaces: Structural Obstacles

A Krylov subspace associated with an operator $A$ and vector $g$, denoted $K_A(g)$, consists of the closure of the span of $\{g, Ag, A^2g, \dots\}$. The core question is: under what conditions does Krylov solvability hold, i.e., when is the solution to $Af = g$ contained in $K_A(g)$?

The paper rigorously analyzes whether the closed Krylov subspace is complemented, i.e., whether there exists a closed subspace $G$ such that $K_A(g) \cap G = \{0\}$ and $K_A(g) + G = X$. This property is automatic in Hilbert spaces but can fail in Banach spaces, with significant implications for solvability and the structure of solution projections.

Critically, the paper shows through explicit constructions—primarily in sequence spaces like $\ell^\infty(\mathbb{N})$ and $g$0—that:

• The lack of a topological complement for $g$1 does not preclude Krylov solvability (Example: weighted backward shift with spectrum covering the origin).
• The existence of a topological complement does not guarantee Krylov solvability.
• The presence or absence of a complemented structure for $g$2 is neither necessary nor sufficient for Krylov solvability.

The notion of the "Krylov intersection" is central: in Banach spaces, whether the intersection $g$3 is trivial can yield necessary and sufficient conditions for Krylov solvability, but only when $g$4 is complemented. The existence of continuous projections and decompositions is directly linked to the complemented property, as is standard in the theory of Banach spaces.

Spectral and Resolvent-Theoretic Tools

Given the limitations of structural analysis, the paper pivots to resolvent and spectral methods for closed operators $g$5 on Banach spaces. The operator resolvent $g$6 is exploited for several purposes:

• Polynomial and Rational Krylov Connections: For bounded $g$7, the resolvent can be uniformly approximated by polynomials in $g$8 for $g$9 outside the spectrum. This property ensures that polynomial Krylov subspaces capture rational Krylov behavior in this regime, providing a tool for examining the reach of Krylov methods.
• Spectral Mapping and Connected Components: If $K_A(g)$0 lies in $K_A(g)$1 for some $K_A(g)$2 in the resolvent set, and the Krylov subspace is invariant under $K_A(g)$3, then $K_A(g)$4 remains in $K_A(g)$5 for all $K_A(g)$6 in the same connected component. This result leverages holomorphicity of the resolvent and analytic perturbation techniques.
• Solvability at Isolated Spectrum Points: The reduced resolvent formalism is used to analyze solvability at isolated points of the spectrum, drawing on complex function theory (the Riesz-Dunford functional calculus). If the spectral projection $K_A(g)$7 associated to the isolated point $K_A(g)$8 satisfies $K_A(g)$9, then the inverse problem $\{g, Ag, A^2g, \dots\}$0 is Krylov solvable, and the solution lies in the closure of polynomial expressions of $\{g, Ag, A^2g, \dots\}$1 applied to $\{g, Ag, A^2g, \dots\}$2.
• Compact Operators and Projections: For compact operators with discrete spectrum, it is shown that spectral projections and associated nilpotent operators can be uniformly approximated by polynomials in $\{g, Ag, A^2g, \dots\}$3, providing an explicit mathematical link between the spectral theory of compact operators and Krylov subspaces.

Contrasts with Hilbert Space Theory

A recurrent theme is the contrast with Hilbert space analogues. Many properties that hinge on orthogonality, direct sum decompositions, and self-adjointness fail or become more subtle in Banach spaces. The absence of a universal dual pairing, and thus a well-defined projection framework, means that structural results must be replaced with more operator-theoretic and complex-analytic arguments.

Practical and Theoretical Implications

The results have several layered implications:

• Algorithmic Selection: Krylov methods such as GMRES and conjugate gradients may or may not converge to true solutions depending on the operator and the Banach space in question. A careful spectral analysis is required even for basic assertions on solvability.
• Ill-posed Problems: The framework is relevant for inversion problems in infinite-dimensional settings where operators are not normal or self-adjoint and where solution approximability is delicate.
• Perturbation and Regularization Theory: The resolvent-based approach provides tools for analyzing the stability and approximation properties of methods in the presence of operator perturbations, non-compactness, and spectral degeneracies.
• Mathematical Physics and PDEs: Many linear inverse problems arising in applied mathematics, especially those involving unbounded operators in Banach function spaces (such as $\{g, Ag, A^2g, \dots\}$4), fall into the studied setting, so the machinery developed connects directly to numerical analysis for differential and integral equations.

Future Directions

Extension of the results to more general classes of operators (e.g., non-sectorial, non-compact, or with continuous spectra) can provide finer criteria for Krylov solvability. Further, investigation into the perturbation behavior of Krylov subspaces and polynomial approximation of operator functions in non-Hilbertian frameworks remains an open field. The connections laid out between operator theory, complex function theory, and computational methods highlight a promising theoretical foundation for advancing methods for solving operator equations in Banach spaces.

Conclusion

This study rigorously delineates the limitations of geometric, projective approaches to Krylov solvability in Banach spaces and demonstrates the robustness of spectral and resolvent-based methods. The key outcome is that, in Banach spaces, traditional Hilbert space intuitions regarding complemented subspaces and iterative solvability must be replaced by more nuanced, operator-theoretic criteria, especially when dealing with ill-posed and infinite-dimensional inverse problems. The analytic properties of the resolvent operator emerge as central to understanding when Krylov-type algorithms can yield meaningful solutions in non-Hilbertian contexts (2604.25686).