Eliminating Spurious Eigenvalues in Numerical Operator Approximation

36 concepts23 papers1 proposal

Many core problems in science and engineering rely on understanding the "vibrational modes" or spectra of infinite-dimensional operators. Whether we are analyzing the Koopman operator to find linear patterns in fluid turbulence, computing the Laplacian to understand the geometry of a manifold, or solving the Schrödinger equation in quantum mechanics, the goal is the same: find the eigenvalues and eigenfunctions that govern the system's long-term behavior.

However, computers cannot directly manipulate infinite-dimensional operators. We must compress them into finite-dimensional matrices using data-driven approximations or numerical discretization. This compression introduces a severe mathematical pathology known as spectral pollution: the appearance of spurious eigenvalues—mathematical ghosts that exist in the finite matrix but have no relation to the true physical system.

Infinite-Dimensional Operator ──(Discretization/Data-Driven Projection)──> Finite Matrix
       │                                                                        │
       ▼                                                                        ▼
 True Spectrum (Bound states, gaps)                                      Polluted Spectrum 
                                                                    (True states + Spurious Ghosts)

Understanding why these artifacts emerge and how to systematically eliminate them is a crucial frontier in numerical analysis and operator learning.

The Geometry of Spectral Hallucination

Why do finite approximations hallucinate eigenvalues? The root cause lies in how truncation distorts the underlying space.

When we project an infinite-dimensional operator onto a finite set of basis functions, we implicitly assume that the discarded, high-frequency components do not influence the low-frequency states we care about. But operators couple these spaces together. Truncation acts like a hard boundary, forcing energy that should escape into higher dimensions to reflect back into our finite subspace.

In data-driven settings, such as Extended Dynamic Mode Decomposition (EDMD), this issue manifests as what Kostic et al. term metric distortion. When approximating an operator from finite, noisy data, the empirical geometry of the data sample deviates from the true geometry of the underlying Hilbert space. This distortion causes the regression algorithm to mistake projection errors or stochastic noise for genuine, stable dynamical features, as noted by Colbrook et al..

The consequences of these artifacts are severe:

  • False Instabilities: In fluid dynamics or partial differential equations (PDEs), spurious positive eigenvalues can falsely indicate that a stable system will blow up, as shown by Narayan et al. and Hu et al..
  • Geometric Distortion: In manifold learning and dimensionality reduction, numerical solvers that fail to isolate the true underlying eigenvectors suffer from coordinate leakage, causing the resulting low-dimensional maps to fold or compress at the boundaries, as demonstrated by Brand.
  • Spectral Invisibility: Spurious eigenvalues can fill physical spectral gaps, completely hiding true physical states like quantum bound states, a problem detailed by Colbrook et al..

Mechanisms for Mitigation

To safely use operator learning in practice, researchers have developed several distinct strategies to identify, filter, or entirely prevent spectral pollution.

1. Structural Preservation and Symmetry

If the original infinite-dimensional operator possesses a structural property—such as self-adjointness (symmetry)—the discrete matrix approximation ought to respect it.

Standard data-driven projections often yield non-symmetric matrices, allowing eigenvalues to leak into the complex plane and create unphysical instabilities. To combat this, Harlim et al. developed symmetric discrete formulations for approximating differential operators on manifolds. By algebraically forcing the discrete operator to be symmetric, the eigenvalues are mathematically locked onto the real line and guaranteed to be non-negative. This structural preservation provides rigorous mathematical stability for higher-order modes where spectral pollution typically dominates.

2. Verification via Spectral Residuals

A direct way to catch a ghost eigenvalue is to measure how well it satisfies the original operator equation. For a true operator $A$, an eigenpair $(\lambda, \psi)$ must satisfy $A\psi = \lambda\psi$. In a finite approximation, we can explicitly compute the spectral residual:

$$\text{Residual}(\lambda) = \|A\psi - \lambda\psi\|$$

If $\lambda$ is a true eigenvalue, the residual will shrink toward zero as the approximation refines. If $\lambda$ is a spurious artifact of the truncation, the residual will remain large.

Colbrook et al. leveraged this intuition to introduce Residual Dynamic Mode Decomposition (Res-DMD), providing a formal verification mechanism for data-driven dynamical models. This residual-based filtering allows practitioners to prune away unphysical components. Building on this concept, Zhang et al. utilize residual calculations to reliably compare complex dynamical systems without contamination from spectral pollution, while Xu designed neural networks (ResKoopNet) that directly minimize these spectral residuals during training to actively search for continuous spectra without introducing fixed-basis artifacts.

3. Shifting to the Resolvent Operator

Instead of discretizing the operator $A$ directly, a highly robust alternative is to compute its resolvent operator, defined as:

$$R(z) = (zI - A)^{-1}$$

where $z$ is a point in the complex plane.

Why does this prevent spectral pollution? The resolvent acts as a localized, smooth filter. While the original operator $A$ might be unbounded or have an infinitely long tail that causes truncation errors to explode, the resolvent maps the infinite-dimensional space into a well-behaved, bounded operator.

Colbrook et al. proved that by applying high-order rational kernels to the resolvent, one can compute the continuous spectral measures of self-adjoint operators to near machine precision without ever encountering spectral pollution. Similarly, when learning the dynamics of random systems, Kostic et al. shifted the learning target from the raw infinitesimal generator to its smoothed resolvent inverse. Combined with energy-based metrics, this approach completely eliminates the mathematical ghosts that haunt traditional stochastic learning models.

Direct Discretization:
[ Infinite Operator A ] ───(Truncate)───> [ Finite Matrix ] ───> Hallucinates Spurious Eigenvalues

Resolvent Mapping:
[ Infinite Operator A ] ───(Invert & Smooth)───> [ Bounded Resolvent (zI-A)⁻¹ ] ───> Immune to Pollution

4. Grid Convergence and Continuous Frameworks

Finally, practitioners can exploit the fact that spurious eigenvalues are highly sensitive to the specific choice of discretization.

As shown by Hu et al., if you vary the number of collocation points or grid resolution in a numerical solver, the true physical eigenvalues will remain stable and converge to fixed values. Spurious eigenvalues, by contrast, will shift wildly across the complex plane or explode as the grid changes, making them easy to identify and discard.

Alternatively, one can bypass matrix discretization entirely. Zvonek et al. introduced a randomized algorithm, ContHutch++, which estimates the trace and density of states of continuous integral operators using continuous stochastic processes. By avoiding the intermediate step of converting a continuous operator into a finite matrix, the method circumvents both spectral pollution and its counterpart, spectral invisibility, entirely.

Key Foundational Work

Continuous nonlinear dimensionality reduction by kernel Eigenmaps (2003) — contributed 1 concept elements to this direction.
Uniform estimation of the eigenvalues of Sturm–Liouville problems (1980) — contributed 1 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_spectral_pollution_76be9ed7,
  title = {spectral pollution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spectral-pollution/rdr_76be9ed7feab4ebba18ef80efd7c0221},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_spectral_pollution_76be9ed7_v1,
  title = {spectral pollution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spectral-pollution/rdr_76be9ed7feab4ebba18ef80efd7c0221/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake