DPD-KAN: Kolmogorov-Arnold Networks for Low Complexity Digital Predistortion in 5G Analog Radio-over-Fiber Systems

Technical Summary: DPD-KAN for 5G Analog RoF Systems

Problem Statement
Radio-over-fiber (RoF) is a critical technology for distributing wireless signals in dense 5G and prospective 6G networks. While Digital RoF (D-RoF) ensures high signal fidelity, it introduces significant bandwidth overhead and power consumption challenges for dense Remote Radio Unit (RRU) deployments. Analog RoF (A-RoF) offers a more efficient alternative with lower latency and complexity; however, A-RoF links are highly susceptible to signal degradation. In systems utilizing directly modulated lasers (DMLs), impairments arise from laser saturation, clipping, intensity-dependent frequency chirp, and chirp-dispersion interactions within standard single-mode fiber (SSMF).

Digital Predistortion (DPD) is the standard technique for mitigating these nonlinearities. While Machine Learning (ML) models, such as Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs), have demonstrated superior performance over classical Volterra-based methods (e.g., Generalized Memory Polynomials, GMP), their computational complexity and power consumption remain barriers to practical deployment in fronthaul links. Furthermore, Kolmogorov-Arnold Networks (KANs), which utilize learnable activation functions on edges rather than fixed functions at nodes, have shown promise in optical baseband equalization and RF power amplifier DPD but have not yet been investigated for DPD in A-RoF systems.

Methodology
The authors propose DPD-KAN, the first KAN-based DPD model designed specifically for a 5G New Radio (NR) A-RoF fronthaul link. The experimental setup involves a 100 MHz bandwidth, 64-QAM 5G NR waveform upconverted to a 3.6 GHz RF carrier. The optical transmission utilizes a 1550 nm VCSEL DML over a 1 km SSMF link, followed by photo-detection and capture via a real-time oscilloscope.

The DPD-KAN architecture employs a dual-path residual structure:

1. Linear Path: An identity mapping that forwards the uncorrected signal.
2. Non-linear KAN Path: Learns the residual correction ($\Delta I, \Delta Q$). Unlike MLPs, KANs use learnable activation functions $\phi(x)$ on the edges, defined as a combination of a base weight, a SiLU function, and a weighted sum of B-spline basis functions.

The model processes a three-dimensional input vector comprising the In-phase (I) and Quadrature (Q) components of the baseband signal and the RF input power. Given the 1 km link length, memory effects from chromatic dispersion were found to be negligible; thus, a memoryless architecture was adopted to address static impairments (VCSEL electro-optic nonlinearity and driver amplifier saturation). The model weights are uniformly quantized to 8-bit resolution during inference.

For validation, an Indirect Learning Architecture (ILA) was used to train the models. The authors compared the proposed DPD-KAN against a baseline MLP (using the same residual architecture) and a conventional GMP model. Complexity was measured in Bit Operations (BOPs) to accurately reflect the cost of quantized networks.

Key Results
The study evaluates performance based on Error Vector Magnitude (EVM), Adjacent Channel Leakage Ratio (ACLR), and Normalized Mean Square Error (NMSE) across varying RF input powers and complexity levels (BOPs).

• Performance at Low Complexity: At approximately $10^4$ BOPs, DPD-KAN significantly outperforms both MLP and GMP. Specifically, at an RF input power of 5 dBm, DPD-KAN achieves a 24.2% lower EVM than an equivalent complexity MLP and a 29.6% lower EVM than a higher complexity ($10^5$ BOPs) GMP model.
• Spectral Efficiency: At 3 dBm RF input power, the $10^4$ BOPs KAN model reduces spectral regrowth by approximately 5 dB compared to a $10^4$ BOPs MLP and a $10^5$ BOPs GMP model.
• Complexity Efficiency: A comprehensive analysis of 200 independent trials reveals that DPD-KAN converges to optimal performance at a significantly lower computational cost. To achieve an EVM threshold of less than 2%, DPD-KAN requires approximately 52% fewer BOPs than the MLP baseline (1.32 $\times$ $10^4$ BOPs for KAN vs. 2.75 $\times$ $10^4$ BOPs for MLP).
• High Complexity Performance: At higher complexity levels ($10^5$ BOPs), DPD-KAN achieves an EVM of 0.91%, an ACLR of -47.46 dBc, and an NMSE of -40.81 dB, performing comparably to the equivalent MLP but with a demonstrated advantage in convergence efficiency at lower complexities.

Significance and Claims
The paper claims to demonstrate the first application of KANs for DPD in 5G A-RoF fronthaul links, addressing a fundamentally different challenge than previous KAN applications in baseband equalization or standalone RF power amplifiers. The primary significance of this work lies in proving that KANs offer a more efficient nonlinear basis for the distortion profiles of VCSEL-based A-RoF links compared to fixed-activation MLPs and Volterra-based GMPs, particularly under stringent complexity constraints.

The authors conclude that DPD-KAN is a suitable architecture for low-complexity linearization in fronthaul A-RoF systems, offering a pathway to deploy effective ML-based DPD in resource-constrained 5G environments without the prohibitive power and computational overhead associated with traditional deep learning models.