Programmable k-local Ising interactions and shallow optical Kolmogorov--Arnold networks through repeated data encounters

Based on the provided manuscript, here is a detailed technical summary of the work by Stroev and Berloff.

Problem Statement

Photonic processors are inherently efficient at linear transformations, but implementing independently programmable higher-order interactions (beyond pairwise couplings) typically requires nonlinear optical media or a reduction of the problem to pairwise models via quadratization. While quadratization is possible, it introduces overhead in the form of ancillary variables and penalty constraints. Existing photonic Ising machines and simulators largely target pairwise couplings, and those that realize higher-order terms often rely on material nonlinearities (e.g., $\chi^{(2)}$ or Kerr effects) which lack per-hyperedge coefficient control. There is a need for a photonic architecture that can directly evaluate sparse, structured $k$-local Ising objectives with independently programmable couplings and interaction orders, without requiring material optical nonlinearities or ancillary variables.

Methodology

The authors propose a repeated-encounter architecture that combines linear optical propagation with square-law detection to generate nonlinear responses to encoded data. The core mechanism involves routing specific groups of spins (hyperedges) to resolved optical channels and subjecting them to multiple passes ("encounters") through a spin-dependent mask.

1. Architecture:

   • Data Encoding: A spatial light modulator (SLM) applies a spin mask $D_s$ to the optical field.
   • Routing and Selection: A fan-out element directs the field to a Fourier plane where specific hyperedges are selected via windows.
   • Reciprocal Recollection: The selected field is returned through the same live spin mask (the "reciprocal" step). This creates a second data encounter.
   • Detection: The system uses a bank of taps to measure the field after each encounter. Direct intensity measurements provide even powers of the spin sum ($S_q^2, S_q^4, \dots$), while phase-referenced interference between adjacent taps provides odd powers ($S_q^1, S_q^3, \dots$).
   • Reconstruction: A calibrated linear combination of these measured signals reconstructs the specific $k$-spin Walsh term (the product of spins $\prod s_i$) with a programmed sign and strength.

2. Theoretical Framework (Tight Encounter Theorem):

   • The authors prove that for a system where the incident field is independent of the spin configuration, the mask is diagonal in optical modes, and detection is square-law, the multilinear spin degree of any detected observable is bounded by twice the number of data encounters ($R$).
   • Consequently, a $k$-spin term requires a minimum of $R = \lceil k/2 \rceil$ encounters.
   • They demonstrate that a rank-one recollection construction (returning the selected mode as a uniform beam onto the corresponding spin macropixels) attains this lower bound, allowing the generation of $S_q^p$ at the $p$-th encounter.

3. Continuous-Variable Extension:

   • By replacing binary spins with bounded continuous signed-amplitude transmittances, the same architecture generates polynomial edge functions. This enables the implementation of shallow Kolmogorov–Arnold Networks (KANs), where network edges are trainable univariate polynomials of degree up to $2R$.

Key Contributions

• Tight Encounter Theorem: A rigorous proof establishing that $R \ge \lceil k/2 \rceil$ is the minimum optical depth required to generate a $k$-local interaction in linear optics with square-law detection, and an explicit construction (rank-one recollection) that achieves this bound.
• Independent Programmability: The ability to program hyperedge identity, interaction order, and signed coupling strength independently without quadratization ancillas or material nonlinearities.
• Optical Weighting via Cardano Seeding: A method to encode the non-constant weights of a $k=4$ term directly into the return-path optics using grating depths, solvable via a closed-form algebraic solution (Cardano's formula) for the required grating parameters.
• Numerical Verification: A finite-dimensional discrete-Fourier model of an ideal folded 4f relay ($1024 \times 1024$ grid) used to validate the $R=2, k=4$ case.

Results

The paper presents numerical results from a finite-discrete Fourier relay model:

• Encounter Scaling: The "reciprocal recollection" route successfully generated a quartic response ($S_q^4$) in the second tap, with an effective exponent $\nu_{eff} \approx 2.02$. In contrast, a control "fixed-patch" route (data-independent return) remained quadratic ($\nu_{eff} \approx 1.06$), confirming the necessity of the second encounter for higher-order terms.
• Walsh Spectrum: The reciprocal recollection method achieved a normalized four-body Walsh coefficient of $0.98$, whereas the fixed-patch control yielded $4 \times 10^{-7}$ (below the noise floor), demonstrating that the four-body information is carried by the second encounter.
• Reconstruction Accuracy: For a four-spin term ($k=4$), the reconstructed energy across all 16 spin configurations showed an RMS error of $0.129$ (baseline aperture) and $0.037$ (half-radius aperture). The "class margin" (separation between target energy classes) was $1.81$ and $1.95$ respectively, compared to an ideal of $2.0$.
• Leakage Analysis: The primary deterministic error source was identified as finite-window leakage, which introduces spurious lower-order Walsh components. The error scales linearly with post-trim macropixel amplitude/phase errors.
• KAN Edge Recovery: In a fixed-projection ridge bank simulation ($R=2$), the system successfully recovered coefficients for polynomial edge functions (degrees up to 4) with a test Mean Squared Error (MSE) of $1.9 \times 10^{-7}$ using exact gradients.

Significance and Claims

The authors position this work as a fundamental advancement in the programmability of photonic hardware for optimization and machine learning:

• Efficiency: The architecture provides a "tight" bound on the optical depth required for $k$-local interactions, proving that no scheme of this class can do better per pass.
• Hardware Independence: It enables direct higher-order mapping without the need for nonlinear materials or the overhead of quadratization, making it suitable for sparse and structured objectives where hyperedges can be routed to unique channels.
• Versatility: The same physical hierarchy supports both discrete Ising objectives and continuous-variable polynomial functions for KANs.
• Modesty in Validation: The authors explicitly state that the results are numerical validations within an idealized finite Fourier model. They do not claim experimental demonstration of the hardware. They note that experimental transfer would require addressing measured aberrations, SLM-camera responses, photon counts, and phase drift. Furthermore, the paper does not claim to benchmark solver performance (time-to-solution) against other platforms or to provide a native-versus-quadratized performance comparison; it focuses on the theoretical construction and its numerical verification.

In summary, the paper introduces a theoretically grounded, linear-optical architecture capable of evaluating arbitrary-order Ising interactions and KAN edges through repeated data encounters, validated numerically to show that the proposed "reciprocal recollection" mechanism successfully generates the required higher-order nonlinearities.