In-Context Learning Based Efficient Spectrum Sensing

The radio frequency (RF) spectrum is essential for wireless communication but is becoming increasingly limited due to the rapid growth in device usage. Real-time spectrum sensing facilitates dynamic spectrum sharing, but conventional methods face significant challenges, including the high power consumption of analog-to-digital converters (ADCs) and the computational demands of Fast Fourier Transforms (FFTs). To address these limitations, prior work introduced a frequencydomain analog signal processor. This processor includes a digitally tunable narrow-bandpass filter implemented with programmable dispersion-engineered elements and a scalable pathsharing delayed signal combiner. However, the naive spectrum sweeping method employed in this design remains highly timeand energy-intensive. In this work, we improve the spectrum sensing efficiency of the analog signal processor by co-designing a sensing matrix generation method with a decoder-based transformer for in-context spectrum recovery. Specifically, we introduce a novel algorithm for sensing matrix generation that leverages the hardware design of the analog signal processor. We show that the generated sensing matrices can be interpreted as part of the well-designed prompts for a transformer with specifically designed parameter matrices to solve the sparse spectrum sensing problem efficiently through its in-context learning capability. To characterize the efficiency of the in-context learningenabled spectrum sensing approach, we provide rigorous theoretical guarantees on the in-context spectrum sensing and evaluate the performances through empirical results. Compared to baseline approaches, our method achieves significant improvements in accuracy.