Principal Component DPD

The core mechanism that separates the array's nonlinear distortion into orthogonal spatial components. Principal Component Analysis (PCA) is applied to the covariance matrix of the PA output signals, identifying the directions of maximum variance. The first few principal components capture the dominant nonlinear modes common across the array, while higher-order components represent uncorrelated noise. By linearizing only the top 2-4 components, the system achieves near-full-array performance with a fraction of the computational cost.

• Reduces an N-element array to K principal modes (K << N)
• Captures correlated distortion caused by beamforming-dependent nonlinearity
• Orthogonal decomposition ensures no redundant compensation

The DPD coefficients are derived from the spatial covariance matrix of the transmitted signals rather than from per-antenna feedback. During training, the system computes the covariance of the PA output signals across all array elements, then performs eigendecomposition to extract the principal eigenvectors. These eigenvectors form a reduced-rank basis that represents the collective nonlinear behavior of the array. The predistorter is then trained to invert the nonlinear response projected onto this low-dimensional subspace.

• Uses eigendecomposition or singular value decomposition (SVD)
• Training complexity scales with K² rather than N²
• Adapts dynamically as beamforming weights change the spatial distortion profile

The number of significant principal components is directly tied to the beamforming configuration. A single narrow beam concentrates nonlinear distortion into one dominant spatial mode, requiring only 1-2 components. Multi-user beamforming with spatially separated streams excites multiple independent distortion modes, increasing the required rank. Principal Component DPD adapts the number of active modes based on the instantaneous spatial multiplexing order, allocating computational resources only where needed.

• Single-beam scenario: 1-2 principal components sufficient
• MU-MIMO with 4 users: Typically 3-5 components required
• Rank adaptation prevents over-provisioning or under-compensation

Principal Component DPD is not a standalone behavioral model but a dimensionality reduction wrapper applied to existing DPD architectures. The spatial PCA projection is combined with temporal basis functions such as memory polynomials or generalized memory polynomials (GMP). The composite model applies nonlinear basis functions in the time domain, then projects the result onto the principal spatial modes. This decoupled structure allows independent optimization of temporal memory depth and spatial rank.

• Compatible with Volterra, MP, and GMP temporal models
• Spatial and temporal dimensions are independently configurable
• Enables reuse of existing single-antenna DPD IP with spatial extension

Principal Component DPD naturally integrates with over-the-air (OTA) feedback architectures. Instead of requiring per-antenna observation receivers, a small set of far-field probes captures the radiated distortion in the dominant spatial directions. The PCA framework maps these sparse OTA measurements back to the principal component coefficients, enabling array-level linearization without the hardware cost of N individual feedback paths. This is critical for mmWave and sub-THz arrays where integrated per-element couplers are impractical.

• Reduces feedback receiver count from N to K+1
• Far-field probes aligned with beam directions capture dominant modes
• Eliminates coupler insertion loss in the transmit path

The primary value proposition of Principal Component DPD is the dramatic reduction in coefficient count and multiply-accumulate operations (MACs). For a 64-element array with a memory polynomial of order 7 and memory depth 3, a full per-antenna DPD requires approximately 64 × 21 = 1,344 coefficients. With PCA reducing to 3 dominant modes, the coefficient count drops to 3 × 21 = 63—a 95% reduction. This translates directly to lower FPGA resource utilization and power consumption.

• 64-element array, full DPD: ~1,344 coefficients
• 64-element array, PCA-DPD (K=3): ~63 coefficients
• Power savings: Typically 60-80% reduction in DPD processing power