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

Paper: 2606.24851 Authors: Jason Sulskis, Sathya Ravi Categories: cs.LG

The Gap

Fourier Neural Operators (FNO) have become a standard tool for learning PDE solution operators by parameterizing global convolutions in the complex Fourier domain.
But for real-valued PDE solutions, the complex FFT stores redundant information through conjugate symmetry — half the spectrum is a mirror image.
Prior work took this inefficiency as a given; no one had systematically tested whether a purely real spectral basis could be better, and under what conditions.

The paper fills this gap by constructing an exact real-valued mirror of FNO (the Hartley Neural Operator, HNO) and establishing a theoretical criterion — Green’s function symmetry — that predicts which basis wins.

[FNO complex FFT] --> [conjugate redundancy for real PDEs] --> [waste?]
     |                                                              |
     v                                                              v
[Can we keep same params with real basis?] ------> [HNO: DHT replaces FFT, iso-parametric]
     |                                                              |
     v                                                              v
[Theory: Green's function alignment] ------> [elliptic: real diag. works]  [time-dep: need complex]
     |                                                              |
     +------------------ Experimental benchmarks -------------------+
                                 |
                                 v
[Match basis to operator symmetry: no universal winner]

The Increment

One sentence: Before this paper, researchers implicitly assumed complex Fourier bases are always appropriate for neural operators; after this paper, we know the best spectral basis is a property of the operator itself — elliptic operators align with real Hartley, time-dependent operators align with complex Fourier.

Core Mechanism

The Hartley Neural Operator (HNO) is structurally identical to FNO: it takes a function input, lifts it to a latent space, applies a spectral convolution layer (transformation → multiplication → inverse transformation), and projects to the output.
The only change is replacing the Fast Fourier Transform (FFT) with the Discrete Hartley Transform (DHT).

The DHT maps real-valued signals to a real-valued spectrum of the same length (no conjugate symmetry), so each spectral mode gets exactly one real multiplier.
FNO, by contrast, discards the redundant half of the spectrum (keeping only N/2+1 complex modes) and uses two real numbers per mode (real and imaginary parts).
Because HNO keeps twice as many frequency “bins” as FNO, but uses half as many parameters per bin, the total parameter count is identical — the two models are iso-parametric.

This means the comparison is purely about the spectral basis: real (Hartley) vs. complex (Fourier).
The core mechanism is that each spectral multiplier acts as a diagonal operator on the Green’s function of the PDE in the frequency domain.
For operators with real symmetric Green’s functions (e.g., Poisson, biharmonic), a real diagonal multiplier can exactly represent the solution, so HNO is favored.
For operators with phase-shifting Green’s functions (e.g., wave, advection), a complex multiplier is required to capture the rotation, so FNO is favored.

Input function f
      |
      v
Lift to latent dim (same for HNO and FNO)
      |
      v
Spectral transform:
    FNO: FFT --> complex spectrum (N/2+1 bins)
    HNO: DHT --> real spectrum (N bins)
      |
      v
Diagonal multiplier:
    FNO: complex mult = (a+ib)*x  (2 real ops per bin)
    HNO: real mult = w*x         (1 real op per bin)
      |
      v
Inverse transform:
    FNO: IFFT
    HNO: IDHT
      |
      v
Project to output u

Think of it like a tuner’s equalizer for audio signals.
FNO is a stereo equalizer with phase knobs — it can handle both magnitude and time delay (phase).
HNO is a monophonic equalizer — it only adjusts magnitude.
If the audio signal has no phase information (like a pure amplitude envelope), the monophonic EQ works perfectly and is simpler.
But if the signal carries phase shifts (like echoes), you need the stereo phase knobs.

The Green’s function of a PDE tells us what kind of “audio signal” the operator produces.
Elliptic operators produce signals where the response is symmetric and in-phase — monophonic works.
Wave and advection operators produce signals where the response lags or leads — stereo phase knobs are necessary.

Key Concepts

• Green’s function symmetry: The Green’s function is the impulse response of a differential operator — what output you get when you apply a point source. For self-adjoint elliptic operators (Poisson, biharmonic), the Green’s function is real and symmetric: G(x, y) = G(y, x). This symmetry is diagonalized by a real spectral basis (the Hartley transform). For time-dependent operators (wave, advection), the Green’s function picks up a phase: G(x, t; y, ts) involves oscillations or transport that cannot be captured by a real diagonal multiplier alone.
  Concrete example: Poisson equation Δu = f has Green’s function G(x,y) = -1/(2π) ln|x-y| in 2D — real and symmetric. The wave equation □u = f has Green’s function involving Heaviside and D’Alembertian — contains phase information.

• Phase content of an operator: How much the operator shifts the phase of frequency components. The wave equation introduces a frequency-dependent phase shift (ejωt), while the heat equation ∂u/∂t = κΔu has no phase shift — it only decays magnitudes. Advection shifts phases linearly with wavenumber. The more phase content, the harder it is for a real-diagonal basis to approximate the operator. This paper shows that heat equation sits exactly on the borderline: both bases perform equally.

Framework Shift

Before (mainstream approach):
  FNO architecture: FFT -> complex mult -> IFFT
  Philosophy: "Complex Fourier basis is the natural choice for PDEs"
  Hidden assumption: Half the spectrum thrown away (conj. sym.) is acceptable
  Evaluation: Compare FNO variants with different parameter counts, but never change basis alone

After (this paper):
  HNO architecture: DHT -> real mult -> IDHT (exact real mirror)
  Philosophy: "Best basis is operator-dependent; match basis to Green's function symmetry"
  Key insight: Iso-parametric comparison isolates basis choice
  Evaluation: Phase content as continuous criterion predicts performance ranking

From “one universal spectral basis” to “basis alignment with operator symmetry” — the core shift is recognizing the spectral basis as a hyperparameter that must match the intrinsic structure of the PDE.

Expert Assessment

Problem choice: Real gap. The redundancy in FNO’s complex spectrum was known but never exploited theoretically or empirically to decide which basis works when. This sits at the intersection of neural operator design and spectral theory of PDEs — a timely and clean question.

Method maturity: Clever, not brute force. The iso-parametric construction is elegant: by keeping parameter count identical, the authors isolate the sole variable (spectral basis). The Green’s function alignment theory is a simple but powerful lens. One could argue that a simpler “try both and pick” approach already existed, but this paper provides a *predictive rule, which is more valuable.

Experimental integrity: Baselines are fair — same architecture, same training, same random seeds. The PDE suite is diverse: Poisson, biharmonic (elliptic); heat, wave, advection, Burgers, Navier-Stokes (time-dependent). One red flag: no comparison with other neural operator families (e.g., DeepONet, attention-based operators). But the paper’s scope is explicitly about FNO-family, so it’s defensible.

Writing quality: Clear and logical. The abstract is dense but communicates the core message. The “elliptic-vs-time-dependent split” is well motivated. The weakest section is the discussion of phase quantification — the metric “operator phase content” is intuitive but not precisely defined. A rewritten section with a concrete phase measure (e.g., imaginary part of Green’s function in frequency) would strengthen the paper.

Verdict: strong accept — delivers a predictive rule, cleanly designed experiments, and a clear theoretical lens that can guide future neural operator design.

Takeaways

1. For practitioners choosing a neural operator: test the PDE’s Green’s function symmetry. If your operator is elliptic (Poisson, elasticity, etc.), try HNO — it may outperform FNO with no additional cost. If your operator has explicit time dependence (wave, convection, etc.), stick with FNO or consider complex-valued architectures.
2. For architecture designers: iso-parametric comparisons are underused. Many neural operator papers add complexity without controlling for capacity. This paper shows how to isolate one design choice fairly.
3. For theorists: the diagonalization perspective (spectral multiplier ≈ Green’s function diagonal in frequency) offers a bridge between neural operators and classical potential theory. The concept of “phase content” as a continuous axis might unify other operator classes beyond elliptic/time-dependent.