Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment

arXiv:2606.24851 · cs.LG · Submitted 2026-06-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment".

Jane: The paper was written by Jason Sulskis and Sathya Ravi from University of Illinois at Chicago and Georgia Tech Research Institute, Electronic Systems Laboratory, Applied Embedded Systems Division and Georgia Tech Research Institute, Applied Embedded Systems Division, Electronic Systems Laboratory.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Paper discussion segment 1: Tom: So we’ve established that the authors of "Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment" provide structural guidelines. Now, let’s think about the initial implications—what does this mean for our general workflow?

Jane: Before this paper, if we were tackling a new PDE—say, one modeling heat flow versus one modeling ocean currents—our instinct was to test both Fourier and Hartley bases and see which gave us the best performance on our training data. That approach is fundamentally flawed because we are treating structural limitations as mere hyperparameters to optimize away.

Lu: The improvement here is that we don't need to guess; we can mathematically predict the optimal architecture by simply analyzing the underlying symmetry of the governing equation itself. This shifts us from empirical testing to deductive design.

Meng: From an industrial standpoint, that predictive capability is a massive time saver and cost reduction tool. Instead of running dozens of computationally expensive simulations just to compare basis performance, we can get a theoretical answer first.

Lalam: It’s about moving beyond the "black box" approach entirely. We are gaining transparency into *why* one method works better than another, based on mathematical principles rather than just observed data patterns.

Tom: Exactly. The authors emphasize that the underlying symmetry of the physics is the ultimate arbiter of computational design choice, which is a powerful shift in methodology for applied mathematics.

Jane: It formalizes a kind of deep respect between physical laws and computational tools; they must align perfectly. To really drive this home, we need to look at how different types of physics demand different mathematical symmetries. That leads us nicely into the discussion of specific processes—like diffusion versus wave propagation—which is what the next section covers.

Paper discussion segment 2: Tom: We’ve discussed how the paper's summary provides concrete mathematical conditions for basis selection. To build on that, let's focus on how those conditions apply to specific physical processes, using "Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment" as our guide.

Jane: The authors show us a clear distinction: if the physics is dominated by steady, purely diffusive processes—the kind of physics that are self-contained in space—they direct us toward utilizing the real Hartley basis first.

Lu: This is fascinating because it suggests that when energy dissipation and spatial containment are key, we can achieve near-optimal performance using a vastly simpler, lower-dimensional structure compared to forcing a complex model onto it.

Meng: For an engineer looking at materials science problems, where heat diffusion or steady chemical concentrations are modeled, this recommendation is incredibly practical. It means we can start with the simplest possible computational framework and still achieve high fidelity results.

Lalam: It suggests that mathematically, these self-contained processes don't require the full complexity of the complex plane to be accurately represented by our neural operator.

Jane: And conversely, for those dynamic, wave-like problems where energy is constantly being transported—like advection or pure wave propagation—we are directed toward the full complex Fourier space.

Tom: The implication here is that when motion and transport are dominant features, we *must* use the complex basis to capture the phase shifts and directional flow correctly. The improvement isn't just theoretical; it translates directly into computational efficiency and memory savings in production models.

Jane: It’s about telling the hardware: "For this specific type of physics, you only need to look at these specific mathematical components." This dramatically reduces the complexity that needs to be learned by the neural network itself. But what happens if we move away from smooth processes? That's our next question.

Paper discussion segment 3: Tom: We’ve seen that "Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment" gives us specific advice based on process type. Now, let's deepen our understanding by focusing on the architectural improvements this paper suggests for deployment.

Jane: To recap, the paper moves us beyond merely testing bases; it predicts them based on symmetry. This is revolutionary because it means we no longer need to guess—we can mathematically predict the optimal architecture by simply analyzing the underlying symmetry of the governing equation itself.

Lu: The diagnostic capability is so powerful for deployment. Think of it as a pre-flight check for your AI model. If your system exhibits characteristics that are highly dissipative, like those self-contained diffusive processes, the paper tells us to build the solver around a real Hartley basis first.

Meng: For industrial modeling, this means we can create specialized "hardware profiles" for different types of physics. We don't treat all PDEs equally; we tailor the computational structure to match the physics profile.

Lalam: This framework is inherently building trust into the AI model because it ensures that our computational tool isn't fighting against a fundamental physical law, which addresses a core concern about AI reliability.

Jane: And for those dynamic, wave-like problems where energy transfer is constant—advection or wave propagation—the full complex Fourier space is mandatory. The improvement is that we are using the minimum necessary mathematical complexity required by the physics.

Tom: Ultimately, this shifts our workflow from a brute-force optimization task—"Which model performs best?"—to an informed design task: "What is the fundamental structure of the physics I am simulating, and what is the simplest mathematical tool that respects that structure?" It's a paradigm shift toward mathematical harmony.

Jane: However, while this framework handles smooth operators beautifully, real-world systems often have

Conclusion: Tom: We've seen how the authors use their work in "Real vs. Complex Spectral Bases for Neural Operators: The Role of Green's Function Alignment" to give us a powerful diagnostic tool, moving away from just testing models toward understanding why they are performing optimally.

Jane: It really highlights that the physical laws governing a process—whether it’s steady diffusion or active transport—determine the best mathematical language we should be using to express that physics in our AI model.

Lu: That predictive power is incredible for the entire field, moving from just empirical results to having a structured theory of design based on self-adjoint versus phase-carrying operators.

Meng: From a practical standpoint, this means we can build highly efficient solvers tailored specifically to the industries they serve without wasting time comparing every single configuration.

Lalam: The concept aligns with my view that AI should be fundamentally respectful of natural constraints, and it’s exciting to see our tools designed to respect the symmetry of nature itself.

Tom: Exactly, Jane; we're moving toward building an architecture where the physics dictates the design, not vice versa.

Jane: It’s been such an insightful discussion, Tom; it has truly clarified a path forward for all researchers in both AI and physics who are looking at these models.

Lu: I'm feeling very optimistic about how this level of mathematical clarity can open up new avenues for exploring other complex systems that have remained elusive.

Meng: We definitely need to consider how this framework scales to real-world deployment across various industries, though, and ensure our hardware can keep up with those efficiency gains.

Tom: Well, we have a lot of ground covered today on the power of structured design in AI; I hope that next time we can explore how these operators handle discontinuities—stay tuned!

University of Illinois at Chicago · Georgia Tech Research Institute, Electronic Systems Laboratory, Applied Embedded Systems Division · Georgia Tech Research Institute, Applied Embedded Systems Division, Electronic Systems Laboratory

cs.LG

Submitted: 2026-06-23

Updated: 2026-09-03

Code: https://github.com/jaysulk/hartley-neural-operator

Importance score: 87/100

The gist: The paper investigates the performance differences between using standard complex Fourier bases (FNO) versus real Hartley bases (HNO) for spectral neural operators, particularly focusing on how the

Key concepts

Spectral Bases
These are mathematical tools used by Neural Operators to represent physical processes. The paper focuses on selecting the right 'basis'—a set of functions—to accurately model a specific type of physics without needing to rely solely on trial-and-error testing.
Real Hartley Basis
This basis is recommended for modeling steady, purely diffusive processes. These are self-contained physical systems where energy dissipation is key. Using this simpler, lower-dimensional structure allows the model to achieve high fidelity results efficiently.
Complex Fourier Space
This complex basis is necessary for dynamic, wave-like problems where energy is constantly being transported. It accurately captures phase shifts and directional flow, ensuring the neural operator can correctly represent motion and transport features.

Terminology

Summary

The paper investigates the performance differences between using standard complex Fourier bases (FNO) versus real Hartley bases (HNO) for spectral neural operators, particularly focusing on how the alignment of these bases with physical properties—such as phase or symmetry—affects approximation accuracy. This analysis is crucial because choosing the appropriate basis can determine whether a learned operator can accurately model complex physical dynamics, especially those involving strong transport or non-self-adjoint behavior.

Basis Comparison and Phase Alignment

The choice between Fourier and Hartley bases is dictated by the mathematical properties of the underlying differential operator. The paper notes that the more phase the propagator carries, the larger FNO’s advantage, establishing a gradient from heat (phaseless) to advection/Burgers/Navier-Stokes (strongly phase-carrying). This suggests a monotone ordering: the more phase the propagator carries, the larger FNO’s advantage. The basis preference is thus set by the relative strength of advection to diffusion when considering nonlinear PDEs.

Approximation Error Bounds and Truncation Equivalence

When analyzing the spectral truncation error, the study finds that both bases possess identical approximation power. Theorem E.20 proves this equivalence for real functions: f - F M [f] L squared = f - H M [f] L squared. This means that the approximation power of both bases is identical—neither basis can represent a given function more accurately with M modes. Furthermore, the Sobolev Truncation Rate (Theorem E.22) confirms this equivalence for both L squared and gradient errors, showing that the error bounds are governed by similar decay rates.

Error Decomposition: Optimization vs. Truncation

The total relative L squared error of a spectral neural operator is decomposed into three components (Theorem E.23):

  1. Truncation Error: This term is identical for both bases, confirming that the basis choice does not affect the inherent approximation limit imposed by mode cutoff M.

  2. Optimization/Realizability Error: This term reflects how well the learned weights can reach the optimal spectral weights (W*). For elliptic (phaseless) operators, HNO’s aligned real parameterization drives this term to zero exactly. However, for phase-carrying operators, the real Hartley diagonal cannot reach W* at all, providing a clear advantage to FNO.

  3. Generalization Error: This depends on sample size and model complexity, independent of the spectral basis used.

Extension to Nonlinear PDEs

The framework is extended to nonlinear equations like Burgers and Navier-Stokes via local linearization (Theorem E.25). The resulting linearized operator contains terms that are classified by their phase content:

  • The advective part is pure imaginary, phase-carrying.

  • The diffusive part of the Fréchet derivative is real, phaseless, HNO-aligned.

Because the advective (phase-carrying) part dominates at typical viscosities studied, Corollary E.26 predicts that FNO remains favored for these nonlinear PDEs, aligning them on the same phase-ordering axis as linear time-dependent operators.

Improvements for AI systems

The core improvement involves moving the basis selection from an empirical choice (FNO vs. HNO) to a theoretically guided, adaptive component within the network architecture itself, informed by the operator's spectral phase properties.

Improvement: Integrate a pre-processing or initialization module that analyzes the spectral symbol lambda(k) of the target PDE operator L. This module calculates a quantitative **Phase Content Metric ** based on the ratio of Im lambda(k) to lambda(k).

What the Improved AI System Can Do:

  • Automatic Basis Selection: Instead of requiring manual selection (FNO or HNO), the system automatically determines if indicates a phase-carrying (> epsilon), phaseless (< -epsilon), or borderline (about 0) operator.

  • Dynamic Architecture Switching: The AMSB dynamically selects the optimal spectral basis implementation (Full Fourier, Real Hartley, or a hybrid/mixed basis) and adjusts the internal weight structure accordingly. This eliminates the current limitation of having to use one fixed architecture for all PDEs.

Mathematical Formulation (Conceptual):

L total = L data + lambda 1 times R truncation + lambda 2 times R phase-alignment(W theta,)

The resulting AI system is not merely a spectral neural operator; it is a Universal Spectral Solver Engine.

  1. Automatic Diagnosis: It diagnoses the underlying mathematical nature of the PDE (elliptic, hyperbolic, diffusive, advective) purely from its symbolic representation lambda(k).

  2. Adaptive Architecture: It selects and implements the mathematically optimal spectral basis (Fourier or Hartley) and weight regularization structure on-the-fly.

  3. Guaranteed Error Control: It optimizes both the approximation error (via AMTS) and the optimization error (via PAWR), ensuring that the reported solution accuracy is rigorously bounded by theoretical convergence rates rather than empirical testing.

Sources

Related papers