In numerical analysis, we often assume that the operators we work with—whether they represent physical collisions or the evolution of a dynamical system—are "well-behaved." Mathematically, this usually means they are bounded: applying the operator to a state doesn't cause the magnitude of that state to explode.
However, many real-world systems are governed by unbounded or degenerate operators. In nonlinear dynamics, transfer operators (like the Koopman operator) can grow without bound. In kinetic theory, collision operators in semi-infinite domains often lack the "coercivity" needed for stable computation. When we apply standard tools like Krylov subspace methods or Galerkin projections to these operators, they often fail to converge, or worse, the error increases as we add more data.
A shared strategy has emerged across these fields: instead of attacking the difficult operator directly, we transform it into a bounded, stable form using a resolvent or damping transformation, solve the problem in that "safe" space, and then map the solution back.
The Problem of Numerical Instability
Standard numerical methods are built on the assumption of continuity and boundedness. As noted by Yuka Hashimoto and Takashi Nodera in Krylov subspace methods for estimating operator-vector multiplications in Hilbert spaces, when an operator is unknown and unbounded, increasing the subspace dimension or adding more data samples can lead to stagnant or even increasing error.
This happens because the "tail" of an unbounded operator—the part that maps to very high-energy or high-frequency states—is not captured by finite approximations. In the context of quantum mechanics or the Schrödinger equation, Nicolas Boullé and Matthew J. Colbrook point out in On the Convergence of Hermitian Dynamic Mode Decomposition that traditional proofs of spectral convergence simply do not apply to these unbounded cases.
Shift-Invert Transformations for Dynamical Systems
To handle unbounded transfer operators in nonlinear dynamical systems, researchers like Yuka Hashimoto and Yoshinobu Kawahara employ a shift-invert transformation.
The intuition is simple: if an operator $A$ is unbounded, its "inverse" (specifically the resolvent $(zI - A)^{-1}$) is often bounded and smooth. In Krylov Subspace Method for Nonlinear Dynamical Systems with Random Noise, the authors apply a shift-invert Arnoldi method. By "lifting" the system into a Reproducing Kernel Hilbert Space (RKHS) and working with the shifted inverse, they map a divergent problem into a numerically stable one. This allows for high-accuracy anomaly detection and time-series prediction even when the underlying physics are mathematically "explosive."
Damping and Recovery in Kinetic Equations
A parallel challenge exists in kinetic theory, where researchers solve equations describing how particles move and collide in a "half-space" (a domain that is infinite in one direction). Here, the collision operators are often degenerate, meaning they don't provide enough "push" (coercivity) to ensure the solution decays properly at infinity.
In A convergent method for linear half-space kinetic equations, Jianfeng Lu, Qin Li, and Weiran Sun introduced a "damp-then-recover" strategy. The mechanism works in two steps:
- Add Damping: Artificially introduce a damping term to the operator. This makes the operator coercive and ensures the solution naturally decays to zero at the far-field limit.
- Recover the Solution: Because the original equations are linear, the authors can systematically subtract the artificial damping effect to recover the true physical solution.
This technique, further refined in Half-space kinetic equations with general boundary conditions, converts a mathematically ill-posed boundary layer problem into a stable system that can be solved with standard spectral methods.
The Theoretical Foundation: Strong Resolvent Convergence
Why do these transformations work so reliably? The answer lies in the theory of strong resolvent convergence.
As Juha Sarmavuori explains in Strong Resolvent Convergence of Finite Matrix Approximations in Numerical Integration, when we approximate an operator with a finite matrix, we shouldn't ask if the matrix converges to the operator itself (which might be impossible if the operator is unbounded). Instead, we ask if the resolvent of the matrix converges to the resolvent of the operator.
Because the resolvent is a bounded operator, it provides a rigorous mathematical bridge. This framework allows us to prove that numerical integration on infinite domains or spectral analysis of quantum Hamiltonians will converge to the correct physical values, even when the operators involved lack the boundedness required by classical theorems. By focusing on the resolvent, researchers can provide "quasi-optimal" numerical schemes for systems that were previously considered too unstable to simulate accurately.