Single-Feedback Receiver DPD

A single observation receiver captures PA outputs one at a time through an RF switch matrix, creating a time-division multiplexed feedback stream. The switch sequentially connects the coupler output from each antenna branch to the shared receiver chain. Key implications:

• Training time scales linearly with the number of PAs
• Coefficient update rate is inversely proportional to array size
• Requires precise time-alignment between the transmitted reference and the delayed feedback sample
• Switch settling time and isolation directly impact measurement fidelity

Eliminates the need for N parallel observation receivers in an N-element array, dramatically reducing component count, power consumption, and board area. A single high-performance ADC and downconverter chain is shared across all branches. Trade-offs:

• Reduced bill of materials (BOM) cost by up to 80% compared to per-branch feedback
• Lower total power dissipation in the feedback path
• Single point of failure in the observation chain
• ADC dynamic range must accommodate the full power variation across all PAs in the array

Because each PA is sampled infrequently, its DPD coefficients may become stale between observation windows. This is particularly problematic during:

• Rapid beam steering events that change active impedance
• Fast-varying envelope conditions in wideband signals
• Thermal transients from bursty traffic patterns

Mitigation strategies include coefficient interpolation between updates, predictive aging models, and prioritizing sampling for PAs experiencing the largest operating point changes.

The RF switch matrix is the critical component enabling single-feedback operation. Design requirements:

• High isolation (>40 dB) between channels to prevent leakage from unsampled PAs contaminating the measurement
• Low insertion loss to preserve feedback SNR
• Fast switching speed to minimize dead time between samples
• Termination management for unselected ports to maintain impedance stability

Common implementations use SPNT switches or cascaded binary switching trees with absorptive terminations on inactive ports.

Accurate DPD coefficient extraction requires sample-level alignment between the transmitted baseband reference and the feedback observation. In single-feedback systems:

• Each PA path has a different loop delay due to varying trace lengths and switch paths
• Delay must be calibrated per-branch and stored in a lookup table
• Fractional delay interpolation is often required for sub-sample alignment
• Correlation-based delay estimation using known training sequences is the standard approach

Misalignment exceeding 0.1 samples can significantly degrade linearization performance.

Single-feedback DPD is particularly well-suited to time-division duplex (TDD) massive MIMO systems where:

• The guard period between downlink and uplink slots provides a natural window for sequential PA sampling
• Channel reciprocity can be exploited to infer downlink distortion from uplink measurements
• The relatively static nature of TDD slot assignments allows predictable training schedules

In reciprocity-based DPD, the single receiver used for uplink reception can be time-shared as the observation receiver during dedicated calibration intervals.

The dominant training architecture for single-feedback systems. The inverse PA model is identified by swapping the input and output of the post-distorter during training, avoiding the need for a pre-distorted signal as a target. The observed PA output becomes the model input, and the original baseband signal becomes the target. This decoupling makes it robust to feedback noise and eliminates the assumption that the PA is exactly invertible.

A resource-efficient technique critical for massive MIMO where a single set of DPD basis function coefficients is applied across multiple antenna branches. By clustering PAs with similar nonlinear characteristics—often based on physical proximity or thermal symmetry—the single-feedback receiver trains one model and shares it across a sub-array. This dramatically reduces the coefficient storage and update computation overhead.

The workhorse batch estimation algorithm for extracting predistorter coefficients from sequentially sampled feedback. The single receiver captures time-multiplexed snapshots from each PA, constructing a composite observation matrix. A least-squares solve then computes the optimal coefficients that minimize the error between the desired linear output and the observed nonlinear output across all branches simultaneously.

A complexity-reduction method that partitions a large array into clusters of elements sharing a single DPD engine. Each sub-array is served by one feedback path, and the single-feedback receiver cycles through elements within that cluster. This balances linearization performance against hardware cost by exploiting the spatial correlation of nonlinear behavior among adjacent elements in a phased array.

Real-time adaptive methods that update DPD coefficients as the single-feedback receiver cycles through array elements. Techniques include recursive least squares (RLS) and least mean squares (LMS) variants that process each new observation sample incrementally. These algorithms track time-varying PA behavior caused by thermal drift, aging, and beamforming weight changes without requiring full batch recomputation.

A pruning technique that identifies and selects only the most significant basis functions from a large candidate set. In a single-feedback architecture, where training time scales with the number of coefficients, sparsity reduces the observation matrix dimensions. Methods like orthogonal matching pursuit or LASSO regularization eliminate redundant terms, yielding a compact predistorter that maintains linearization performance while reducing computational load.