ET-DPD Closed-Loop

A dedicated feedback path that downconverts and digitizes a coupled sample of the power amplifier output. This observation receiver captures the actual transmitted waveform, including all ET-induced distortions, providing the error signal needed for adaptation.

• Typically uses a high-dynamic-range ADC with bandwidth exceeding 3-5x the signal bandwidth
• Must maintain phase coherence with the transmit path for accurate error computation
• Enables capture of supply-dependent AM/AM and AM/PM distortion in real operating conditions

An online learning algorithm that continuously compares the observed output with the desired input to compute updated predistorter coefficients. This engine runs in the background during live transmission.

• Implements least-squares (LS) or recursive least-squares (RLS) estimation
• Updates coefficients on a frame-by-frame or sub-frame basis
• Tracks slow-varying effects like thermal memory drift and supply modulator aging
• Convergence time typically under 100 ms for stable operating conditions

Precise temporal alignment between the reference transmit signal and the observed feedback signal is critical. Even sub-sample misalignment corrupts the error calculation and causes the DPD to diverge.

• Integer-sample alignment via cross-correlation of reference and feedback
• Fractional-sample alignment using Lagrange interpolation or Farrow filters
• Must also compensate for ET delay mismatch between RF and supply paths
• Typical alignment accuracy requirement: < 1% of symbol period

The dominant closed-loop topology where the predistorter coefficients are extracted by training a postdistorter on the feedback signal, then copying those coefficients to the forward predistorter.

• Avoids the nonlinear optimization problem of direct learning
• Assumes the PA characteristic is invertible and the postinverse equals the preinverse
• Robust to measurement noise in the observation path
• Standard implementation for ET-DPD systems due to stability guarantees

The closed-loop system must capture training data that spans the full dynamic range of supply voltage variation. This ensures the DPD model characterizes PA behavior across all ET operating points.

• Data capture synchronized with shaping function traversal
• Requires excitation signals with sufficient PAPR and envelope diversity
• Captures iso-gain contour transitions where nonlinearity changes rapidly
• Typical capture duration: 10,000 to 100,000 samples per adaptation cycle

Closed-loop adaptation introduces the risk of coefficient divergence under anomalous conditions. Robust systems implement guard mechanisms to prevent spectral mask violations.

• ACLR monitoring on the feedback signal to detect degradation
• Coefficient magnitude clamping to prevent excessive gain expansion
• Freeze-on-fault logic that holds last-known-good coefficients
• Automatic recovery and re-convergence when normal conditions resume

A dedicated feedback receiver that attenuates and downconverts a coupled sample of the PA output to baseband for digitization. It must exhibit higher linearity and bandwidth than the transmitter itself to avoid corrupting the error signal. Key specifications include:

• Spurious-free dynamic range (SFDR) exceeding 70 dB
• Bandwidth typically 3-5x the signal bandwidth for adequate harmonic capture
• IQ imbalance below -40 dBc to prevent coefficient estimation bias

The digital processing block that computes new predistortion coefficients by comparing the transmitted reference signal with the observed feedback. Common algorithms include:

• Least Squares (LS) for batch estimation
• Recursive Least Squares (RLS) for continuous tracking
• Least Mean Squares (LMS) for low-complexity gradient descent
• Normalized LMS (NLMS) for improved stability with varying signal power The update rate must be fast enough to track thermal transients and supply voltage changes.

The critical process of time-aligning the reference and feedback signals to within a fraction of a sample period before coefficient estimation. Misalignment by even 0.1 samples can severely degrade linearization performance. Techniques include:

• Integer delay correction via FIFO buffers
• Fractional delay compensation using Farrow interpolators
• Correlation-based delay estimation using the transmitted waveform's autocorrelation properties
• Frequency-domain phase slope measurement for sub-sample precision

The process of scaling the feedback signal to match the reference signal's amplitude before error computation. The PA's gain varies dynamically with supply voltage, temperature, and frequency, requiring continuous normalization. Methods include:

• RMS power ratio estimation over a sliding window
• Peak correlation for rapid acquisition
• Adaptive gain tracking using a first-order IIR loop Incorrect normalization introduces gain compression artifacts into the DPD model.

The selection or generation of waveforms used to excite the PA during coefficient extraction. Effective training signals must:

• Exercise the full dynamic range of supply voltage and input power
• Match the statistical distribution (PAPR, bandwidth) of the intended modulation
• Avoid peak regrowth that saturates the observation receiver
• Support persistent excitation conditions for invertible matrix solutions Common choices include OFDM symbols, noise-like signals, and multi-tone combs.

A supervisory mechanism that prevents the DPD from diverging due to faulty feedback, model mismatch, or extreme operating conditions. Protection strategies include:

• Coefficient magnitude limiting to prevent excessive gain expansion
• ACLR watchdog that reverts to safe coefficients if spectral regrowth exceeds thresholds
• Feedback signal integrity checks for saturation, clipping, or loss of signal
• Hysteresis in update decisions to prevent oscillation between coefficient sets