Regularizing Unbounded Operators for Stable Numerical Estimation

15 concepts7 papers1 proposal

When analyzing complex dynamical systems, fluid flows, or kinetic equations, researchers often "lift" the system's state space into an infinite-dimensional space of functions. In this view, highly nonlinear dynamics become linear through operators such as the Koopman, Liouville, or transfer operators. However, a major mathematical hurdle arises: these operators are frequently unbounded.

Intuitively, an unbounded operator can take a well-behaved, smooth function and map it to a function with an arbitrarily large norm—akin to how taking the derivative of a highly oscillatory function amplifies its amplitude. When data-driven algorithms like Dynamic Mode Decomposition (DMD) or Krylov subspace methods attempt to approximate these operators from empirical data, this lack of boundedness causes severe numerical instability. As Hashimoto et al. 2021 demonstrate, standard iterative methods can fail to converge entirely in these scenarios; adding more data or increasing the subspace dimension can paradoxically cause the approximation error to stagnate or grow.

To restore numerical sanity, a powerful shared strategy has emerged across different domains: instead of approximating the unbounded operator directly, transform it into a bounded or compact operator. A bounded operator guarantees that small errors in data do not explode during computation. Once the stable, transformed operator is estimated or solved, the original physical properties can be systematically recovered.

Shift-Invert Transformations and Resolvents

One of the most elegant ways to tame an unbounded operator $A$ is the shift-invert transformation, which evaluates the operator's resolvent:

$$(A - \sigma I)^{-1}$$

where $\sigma$ is a scalar shift.

To understand why this works, imagine an operator with eigenvalues stretching out to infinity. By subtracting a shift $\sigma$ and taking the inverse, the largest eigenvalues (which cause the numerical instability) are mapped close to zero. Meanwhile, the eigenvalues closest to $\sigma$ are magnified. The entire infinite spectrum is effectively mapped into a bounded, well-behaved domain.

Unbounded Spectrum:  [λ₁] ---- [λ₂] -------- [λ₃] ------------------> ∞
                                   │ (Shift-Invert)
                                   ▼
Bounded Spectrum:    [ 0 <------------------ (λ₃-σ)⁻¹ --- (λ₁-σ)⁻¹ ]

Hashimoto et al. 2019 applied this shift-invert Arnoldi framework to the transfer operators of nonlinear dynamical systems corrupted by random noise, demonstrating that it ensures numerical convergence where standard subspace methods fail. In follow-up work, Hashimoto et al. 2021 proved that this regularization maps the potentially divergent operator into a bounded space, enabling exponential convergence of operator-vector multiplications as the subspace dimension grows.

A closely related approach is taken by Kostic et al. 2025 for learning continuous Markov semigroups. Instead of directly estimating the unbounded infinitesimal generator of the semigroup, the authors leverage the Laplace transform of transfer operators to estimate its resolvent. This provides a bounded and stable alternative that remains robust even under high-frequency sampling conditions.

Operator Scaling for Compactness

In continuous-time DMD, researchers often deploy the Liouville operator to avoid the discretization errors inherent in discrete-time Koopman frameworks. However, because the Liouville operator is a differential operator, it is inherently modally unbounded.

To resolve this, Rosenfeld et al. 2021 and Rosenfeld et al. 2022 introduced the concept of a scaled Liouville operator. By introducing a scaling parameter strictly less than one within a reproducing kernel Hilbert space (RKHS), the authors mathematically transform the unbounded differential operator into a compact operator.

Compact operators are the closest infinite-dimensional cousins to finite matrices; they map bounded sets to relatively compact sets, which guarantees strong norm convergence and uniform error bounds for data-driven estimation procedures. Because the scaling parameter can be chosen to be arbitrarily close to one, the compact operator remains numerically indistinguishable from the true physical system within machine precision, while unlocking rigorous convergence guarantees for the DMD procedure.

The Damp-then-Recover Strategy in Kinetic Equations

The philosophy of operator transformation also extends to solving partial differential equations, such as linear half-space kinetic equations. In this domain, the challenge stems from degenerate or non-coercive collision operators, which fail to provide the necessary dissipation to guarantee that solutions decay appropriately at infinity.

To bypass this limitation, Lu et al. 2014 and Lu et al. 2017 developed a "damp-then-recover" (or "damping adding-removing") procedure. The authors temporarily inject artificial damping terms into the non-coercive operator, transforming it into a well-behaved, coercive form. This modification ensures mathematical well-posedness and numerical stability for standard spectral or Galerkin solvers. Once the damped system is solved, the authors exploit the linearity of the underlying equations to recover the true physical solution via a systematic linear superposition of special solutions.

Summary

Whether through shift-invert resolvents, parameter scaling in an RKHS, or the addition of temporary damping terms, the underlying principle remains identical: numerical methods require boundedness to converge reliably. By wrapping unstable, unbounded physical operators in clever mathematical transformations, these frameworks allow us to leverage standard, highly efficient numerical linear algebra tools without sacrificing physical fidelity.

Key Foundational Work

A convergent method for linear half-space kinetic equations (2014) — contributed 2 concept elements to this direction.
Half-space kinetic equations with general boundary conditions (2017) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_shift_invert_operator_transformation_6b26b538,
  title = {shift-invert operator transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/shift-invert-operator-transformation/rdr_6b26b538f59a4c08a9c449877c625343},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_shift_invert_operator_transformation_6b26b538_v1,
  title = {shift-invert operator transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/shift-invert-operator-transformation/rdr_6b26b538f59a4c08a9c449877c625343/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake