TL;DR
This paper introduces KAN-LUT, a highly efficient soft demapper for nonlinear satellite communication links based on Kolmogorov–Arnold Networks. By representing learned functions as small 1D lookup tables, the authors achieve near-optimal accuracy (within 0.2 dB of the gold standard) at a fraction of the computational and memory cost of traditional neural networks or complex statistical models.
What Problem Does This Paper Address?
Next-generation "high-throughput" satellites are evolving from simple "bent-pipe" transponders into "regenerative" payloads that demodulate and decode signals directly in orbit. However, doing this on-board is difficult because satellites have extremely tight power and thermal budgets.
The core challenge lies in soft demapping: converting received complex signals into bit log-likelihood ratios (LLRs) that the error-correction decoder (LDPC) can understand. In satellite links, high-power amplifiers (TWTAs) are run near saturation to maximize efficiency, which introduces nonlinear distortion. Standard demappers struggle with this distortion, while accurate "impairment-aware" models are computationally expensive. As data rates push toward 1 Terabit per second, the industry needs a way to perform accurate demapping without melting the satellite's hardware.
Methods
The authors leverage a new architecture called a Kolmogorov–Arnold Network (KAN). Unlike traditional Multilayer Perceptrons (MLPs) that place fixed activation functions on nodes and learn weights on edges, KANs place learnable univariate (one-variable) functions directly on the edges.
The "secret sauce" of this paper is the KAN-LUT approach:
- Training: The network is trained using B-splines to learn the optimal nonlinear mapping for demapping.
- Implementation: Because the edge functions are univariate, they can be pre-computed and stored as 128-entry Lookup Tables (LUTs).
- Inference: During operation, the satellite doesn't perform heavy matrix multiplications. Instead, it performs simple table reads, linear interpolation, and additions.
- Inputs: The model takes three features: the In-phase ($z_I$) and Quadrature ($z_Q$) components of the received signal, and the estimated signal-to-noise ratio (SNR).
Datasets
The study evaluates the demapper using simulated data based on industry-standard satellite communication protocols:
- Modulation Schemes: 16-APSK, 32-APSK, and 64-APSK (Amplitude Phase-Shift Keying), as defined in the DVB-S2 and DVB-S2X standards.
- Channel Model: A memoryless Saleh model was used to simulate the nonlinear distortion of a Traveling-Wave-Tube Amplifier (TWTA), combined with Additive White Gaussian Noise (AWGN) for both uplink and downlink.
- Scale: Training was performed on datasets ranging from $10^7$ to $3 \times 10^8$ symbols to ensure robust convergence.
Key Insights and Ideas
- Defeating the Curse of Dimensionality: A direct multidimensional LUT for 32-APSK demapping would require roughly 40 MiB of memory. By using the KAN structure, the memory requirement drops to just 32 KiB, as the storage grows linearly with the number of edges rather than exponentially with input dimensions.
- Efficiency Gains with Complexity: As the constellation order increases (from 16 to 64-APSK), the computational cost of traditional statistical models (LogMAP-CG) nearly quadruples. In contrast, the KAN-LUT cost increases only slightly because only the output layer expands.
- Robustness to Drift: The authors found that a KAN-LUT trained at a specific operating point remains competitive even when the amplifier back-off or link conditions drift, matching the robustness of traditional model-based demappers.
Key Results
- Accuracy: KAN-LUT stays within 0.2 dB of the performance of the LogMAP-CG (the high-complexity reference) across all tested APSK orders.
- Computational Savings: KAN-LUT matches or slightly exceeds the accuracy of MLPs but at only 1/5th of the computational cost.
- Relative Cost: Compared to the LogMAP-CG reference, KAN-LUT’s cost falls from 87% (for 16-APSK) to just 27% (for 64-APSK).
- Hardware Readiness: For a 1 Tbit/s link, KAN-LUT would require roughly half the operations per second compared to the standard LogMAP-CG approach, making it a prime candidate for real-time regenerative payloads.
Limitations and Future Work
- Fixed-Point Arithmetic: The current study uses floating-point operations. The authors note that future work should evaluate the impact of fixed-point quantization on the LUT entries and interpolation logic.
- Channel Memory: The Saleh model used here is memoryless. Real-world satellite channels can exhibit memory effects (intersymbol interference), which might require KANs with recurrent structures or taps.
- Hardware Validation: While the operation counts are promising, the authors suggest that actual FPGA or ASIC measurements are needed to confirm the power savings in a physical implementation.
Summary updated