A K-Means Clustering Data Selection Method for Low-Complexity Least Squares in Digital Predistortion

Digital Predistortion (DPD) is a widely adopted technique to mitigate power amplifier (PA) nonlinearity, where the Least Squares (LS) method is commonly used for predistorter model coefficient estimation. However, the LS method faces a fundamental trade-off between estimation accuracy and computational complexity, where the input data length plays a important role. To address this, this paper proposes a data selection method combining k-means clustering and a Look-Up Table (LUT). First, the k-means algorithm adaptively extracts cluster centers as representative data points, which retain the critical statistical characteristics of the original long sequence by grouping samples with similar nonlinear response patterns. Second, the LUT-based simplification stores precomputed centroids offline and enables hardware-friendly online data selection via squared Euclidean distance discrimination, avoiding real-time clustering overhead. Simulation and experimental results verify that the proposed method only requires input data length of around 250 to achieve −41.5dB and −51 dB value in Normalized Mean Square Error (NMSE) and Adjacent Channel Power Ratio (ACPR), respectively. Moreover, the LUT simplification introduces negligible performance degradation while significantly reducing hardware latency and resource consumption.