KANO: Kolmogorov-Arnold Neural Operator

1. Problem Statement

The paper addresses a critical limitation in current Neural Operators, specifically the Fourier Neural Operator (FNO), when applied to position-dependent dynamics (governed by variable-coefficient Partial Differential Equations, PDEs).

• The Spectral Bottleneck: FNO relies on a "pure-spectral" approach, hard-coding the encoding of functions into a truncated Fourier basis. While efficient for spectrally sparse operators (where coefficients decay rapidly), FNO struggles with operators that are spectrally dense.
• Position-Dependent Dynamics: In many physical systems (e.g., fluid flow with varying viscosity, quantum Hamiltonians with position-dependent potentials), the operator involves multiplication by spatial variables (e.g., $x \cdot u(x)$). In the Fourier domain, spatial multiplication becomes a dense convolution (Toeplitz matrix), destroying the sparsity that FNO relies on.
• Consequences: To approximate these dense operators, FNO requires super-exponential growth in model size to achieve a target accuracy. Consequently, FNO models trained on specific data distributions fail to generalize to unseen inputs (out-of-distribution failure) because the learned spectral off-diagonals are tied to the training data distribution rather than the underlying physical law.
• Lack of Interpretability: Existing operator networks are often "black boxes." There is a need for models that not only approximate operators but also recover the underlying closed-form symbolic equations (scientific laws) from data.

2. Methodology: The KANO Framework

The authors propose KANO (Kolmogorov–Arnold Neural Operator), a dual-domain neural operator that jointly parameterizes the operator in both spatial and spectral bases using Kolmogorov–Arnold Networks (KANs).

Core Architecture

Unlike FNO, which uses Multi-Layer Perceptrons (MLPs) for latent mapping and hard-coded Fourier transforms, KANO utilizes:

1. Pseudo-Differential Symbol Calculus: KANO represents the operator using a learnable symbol $p(x, \xi)$ that depends on both spatial coordinates ($x$) and frequency modes ($\xi$).
2. Kohn-Nirenberg Quantization: The operator is computed via a quantization formula that integrates the symbol $p(x, \xi)$ with the input function $a(y)$ over both spatial and spectral domains:
   $$\mathcal{O}_m(p)[a](x) = \left(\frac{h}{L}\right)^d \sum_{\xi \in \Xi} \sum_{y \in Y} e^{i(x-y)\cdot\xi} p(x, \xi) a(y)$$
   This allows the model to perform convolutions in the spectral domain (handling derivatives) and multiplications in the spatial domain (handling position-dependent terms) simultaneously.
3. KAN Sub-networks:
   • The symbol $p(x, \xi)$ is parameterized by a KAN sub-network.
   • The non-linear activation function $\Phi$ is also a KAN.
   • Why KAN? KANs replace fixed node activations with trainable univariate functions on edges. This structure is inherently interpretable, allowing the learned functions to be visualized and approximated by symbolic regression to recover closed-form mathematical expressions.
4. Dual-Domain Sparsity: KANO exploits the fact that differential terms are sparse in the spectral domain, while localized/multiplicative terms are sparse in the spatial domain. By learning the symbol in both domains, KANO achieves a compact representation for variable-coefficient PDEs.

Theoretical Analysis

The paper provides a rigorous theoretical proof of FNO's limitations and KANO's advantages:

• FNO Limitation (Theorem 1): The authors prove that for position-dependent operators, FNO suffers from the curse of dimensionality. The projection error (due to truncating the Fourier tail) forces the latent network size to scale super-exponentially with respect to the desired accuracy $\epsilon$.
• KANO Advantage (Theorem 2): KANO avoids this bottleneck. By using the Kohn-Nirenberg quantization, the projection error scales polynomially with the grid width $m$. Furthermore, the latent network (KAN) approximates the smooth symbols with polynomial scaling. Thus, KANO achieves the desired accuracy with a polynomial increase in model size, making it practical for generic position-dependent dynamics.

3. Key Contributions

1. Theoretical Proof of FNO's Bottleneck: The paper formally demonstrates that FNO cannot efficiently converge to generic position-dependent dynamics due to the super-exponential scaling of model size required to handle the algebraic decay of Fourier tails induced by position operators.
2. Novel Architecture (KANO): Introduction of the first interpretable neural operator that jointly parameterizes spatial and spectral bases using KANs within a pseudo-differential operator framework.
3. Symbolic Recovery: KANO enables the extraction of closed-form symbolic formulas from the learned operator, a capability largely absent in previous operator learning methods.
4. Robust Generalization: The architecture is designed to generalize robustly to out-of-distribution inputs, overcoming the fragility of FNO on unseen data subspaces.

4. Experimental Results

The authors evaluated KANO on synthetic operators and a quantum physics benchmark.

A. Synthetic Operator Benchmarks

• Tasks: Learning three position-dependent operators ($G_1, G_2, G_3$) involving terms like $x^2 \cdot f$, $x \cdot \partial_x f$, and non-linearities.
• Setup: Models were trained on "Group A" (Periodic, Chirped Cosine, Sine Beats) and tested on "Group B" (Sinc Pulse, Gaussian-Hermite, Wave Packet) to test out-of-distribution generalization.
• Performance:
  • Generalization: FNO and its variants (U-FNO, AM-FNO) showed significant loss increases (fragility) on Group B. KANO maintained consistent low loss across both groups.
  • Efficiency: KANO achieved 10x lower relative $\ell_2$ loss than FNO while using only 0.03% of the parameters (152 parameters vs. 566k).
  • Symbolic Recovery: After training, the KAN edges were visualized and frozen. The model successfully recovered the exact ground-truth symbolic formulas (e.g., $x^2 - \partial_{xx}$) with coefficients accurate to the fourth decimal place.

B. Quantum Hamiltonian Learning Benchmark

• Tasks: Learning the ground-truth Hamiltonian for a Quartic Double-Well (DW) potential and a Nonlinear Schrödinger Equation (NLSE) from projective measurement data (Position and/or Momentum Probability Mass Functions).
• Scenarios:
  • Full-type: Training with full wave function data (ideal, physically unattainable).
  • Pos-type: Training with only position data.
  • Pos&Mom-type: Training with both position and momentum data.
• Results:
  • State Infidelity: KANO achieved a state infidelity of $\approx 6 \times 10^{-6}$, outperforming FNO (trained on full data) by four orders of magnitude ($\approx 1.5 \times 10^{-2}$).
  • Data Efficiency: KANO trained on "Pos & Mom" data achieved performance indistinguishable from the ideal "Full" data baseline.
  • Symbolic Reconstruction: KANO successfully reconstructed the Hamiltonian terms (e.g., $x^4 - x^2 + 0.5\xi^2$) with high precision. It correctly identified that global phases and certain cubic coefficients were unobservable from the available data, demonstrating physical consistency.

5. Significance and Impact

• Paradigm Shift: KANO shifts the goal of operator learning from "universal approximation" (fitting data) to "universal generalization" and "scientific discovery" (recovering laws).
• Interpretability: It bridges the gap between deep learning and symbolic mathematics, allowing AI to not just predict outcomes but to reveal the governing equations of physical systems.
• Scalability: By theoretically and empirically proving that KANO avoids the curse of dimensionality inherent in spectral methods for variable-coefficient PDEs, it opens the door to solving complex, real-world physics problems that were previously intractable for FNO.
• Scientific AI: The ability to learn from partial observations (projective measurements) and recover closed-form Hamiltonians makes KANO a powerful tool for quantum system identification and other scientific domains where full state data is unavailable.

In summary, KANO represents a significant advancement in Scientific AI by combining the efficiency of spectral methods with the spatial flexibility of KANs, resulting in a model that is highly accurate, parameter-efficient, and capable of discovering interpretable physical laws.