Beamforming-Aware DPD

Unlike static DPD, beamforming-aware techniques continuously adjust the predistorter coefficients to track active impedance mismatch at each power amplifier. As the beam is steered, the load impedance seen by each PA changes, altering its AM/AM and AM/PM characteristics. This approach models the nonlinearity as a function of the beamforming weight vector, not just the input signal envelope.

In a phased array, nonlinear distortion products are not radiated uniformly. The beam-squint effect causes frequency-dependent steering, meaning intermodulation products may be directed differently than the fundamental beam. Beamforming-aware DPD models this spatial behavior to suppress spectral regrowth in specific angular directions, protecting adjacent channel users in the far-field.

Advanced implementations merge DPD with zero-forcing or minimum mean square error precoding in a single optimization step. This joint processing simultaneously corrects PA nonlinearity and mitigates multi-user interference. The technique is particularly relevant for MU-MIMO downlink scenarios where per-antenna distortion and inter-user crosstalk must be addressed holistically.

For massive MIMO arrays with hundreds of elements, per-branch DPD is computationally prohibitive. Sub-array DPD groups antennas with similar nonlinear behavior—determined by their position in the array and mutual coupling environment—and applies a single predistorter per cluster. Principal component DPD further reduces dimensionality by identifying and compensating only the dominant spatial modes of distortion.

Traditional DPD uses per-branch couplers for feedback, which does not capture far-field array effects. Over-the-air DPD uses a remote observation receiver to sample the combined radiated field, enabling correction of nonlinearities as they appear at the intended receiver location. This approach inherently accounts for antenna mutual coupling and array manifold effects that per-element feedback misses.

Graph neural network DPD models the antenna array as a graph where nodes represent PAs and edges represent coupling paths. This structure naturally captures spatial dependencies and generalizes across beamforming states. Physics-informed DPD embeds known PA behavioral models—such as the Volterra series or memory polynomial—into the network architecture, improving data efficiency and extrapolation to unseen steering angles.

The root cause of why beamforming-aware DPD is necessary. As the array steers its beam by adjusting the phase and amplitude of each element, the impedance seen by each individual power amplifier changes dynamically. This active impedance variation—caused by mutual coupling between elements—modulates the PA's nonlinear behavior as a function of the beam angle. Without accounting for this, a static DPD model calibrated at boresight will fail when the beam scans off-axis.

• Key metric: Voltage Standing Wave Ratio (VSWR) variation over scan angle
• Impact: AM/AM and AM/PM curves shift with beam weight selection
• Mitigation: Load-insensitive PA design combined with weight-aware predistortion

An adaptive linearization strategy that explicitly compensates for the time-varying load impedance presented to each PA in a phased array. Unlike conventional DPD that assumes a fixed 50-ohm termination, load modulation DPD incorporates the instantaneous reflection coefficient (Γ) as an input parameter to the predistorter model.

• Architecture: Look-up tables or polynomial models indexed by both signal envelope and load state
• Training: Requires characterization across the full Smith chart or at representative beam states
• Trade-off: Higher model dimensionality vs. improved linearity under beam steering
• Application: Critical for Doherty PAs in arrays, where load modulation is inherent to the amplifier architecture itself

A signal processing method that mitigates unintended electromagnetic interaction between adjacent antenna elements. In a dense array, the signal transmitted by one element couples into neighboring elements, creating a composite distortion that appears at each PA's output. Cross-coupling cancellation models this as a linear crosstalk matrix that must be inverted jointly with the nonlinear DPD function.

• Modeling: S-parameter coupling matrix extracted via full-wave simulation or over-the-air measurement
• Joint formulation: The combined nonlinear PA + linear crosstalk system requires a MIMO Volterra or neural network model
• Complexity scaling: Full MIMO DPD complexity grows as O(N²) with array size; sub-array and sparse techniques reduce this

A linearization architecture where the combined far-field radiated signal is captured by a probe antenna and used as the feedback signal for DPD training. This approach inherently accounts for all array-level impairments—PA nonlinearity, mutual coupling, and beamforming effects—in a single closed-loop measurement.

• Advantage: No need for per-element feedback receivers; captures true radiated distortion
• Challenge: Feedback signal is a spatial superposition; requires beam-specific training sequences
• Implementation: Single or multiple probe antennas placed in the far-field or compact antenna test range
• Use case: Production calibration and periodic recalibration of deployed arrays

A complexity-reduction technique for massive MIMO systems where a single DPD engine linearizes a cluster of antenna elements that share similar nonlinear characteristics. By grouping elements with correlated distortion behavior—typically those in close physical proximity or driven by identical PA designs—the number of independent predistorters is dramatically reduced.

• Grouping criteria: Spatial proximity, PA batch matching, or measured distortion correlation
• Trade-off: Linearization performance vs. hardware/software resource savings
• Scalability: Enables DPD for arrays with 64, 128, or 256+ elements where per-element DPD is infeasible
• Synergy: Often combined with coefficient sharing across elements within a sub-array

A dimensionality reduction technique that identifies and compensates for the dominant spatial modes of nonlinear distortion in a massive MIMO array. Rather than modeling each PA individually, principal component DPD decomposes the array's distortion into a small set of orthogonal spatial patterns and applies predistortion in this reduced-dimensional space.

• Method: Singular value decomposition (SVD) or principal component analysis (PCA) on the array's distortion covariance matrix
• Result: Typically 3-8 dominant modes capture >95% of the nonlinear distortion energy
• Benefit: Linearizes the entire array with far fewer coefficients than per-element or full MIMO approaches
• Foundation: Exploits the fact that mutual coupling and beamforming create structured, low-rank distortion