Learning Rate

The learning rate μ scales the gradient vector to determine the magnitude of each coefficient update. A large μ accelerates convergence but risks overshooting the Wiener solution, causing parameter oscillation. A small μ ensures smooth convergence but may fail to track time-varying PA nonlinearities caused by thermal memory effects or bias drift. In practice, the step size must satisfy 0 < μ < 2/λ_max, where λ_max is the largest eigenvalue of the input autocorrelation matrix, to guarantee bounded-input bounded-output stability.

The learning rate establishes a fundamental trade-off in adaptive DPD systems:

• Convergence Rate: Proportional to μ — larger values drive the mean squared error (MSE) toward the minimum faster, critical for rapid cold-start linearization.
• Steady-State Misadjustment: Also proportional to μ — excess MSE above the theoretical Wiener minimum caused by gradient noise from stochastic estimation. This trade-off is quantified by the misadjustment factor M ≈ μ·tr(R)/2, where tr(R) is the trace of the input correlation matrix.

Practical DPD implementations often employ dynamic learning rate schedules to optimize both acquisition and tracking phases:

• Step-Decay: μ is reduced by a factor γ after a fixed number of iterations, enabling fast initial convergence followed by refined steady-state performance.
• Exponential Decay: μ(t) = μ₀·e^(-αt) provides smooth asymptotic reduction.
• Performance-Based Adaptation: μ adjusts based on monitored ACLR or EVM metrics, increasing when distortion rises due to PA temperature changes.
• Gear-Shifting: Discrete μ switching triggered by error threshold crossings.

In Normalized Least Mean Squares (NLMS) adaptation, the effective learning rate is normalized by the input signal power: μ_eff = μ̃ / (||x(n)||² + δ). This provides:

• Input Power Invariance: Prevents gradient amplification during high-PAPR signal peaks common in OFDM waveforms.
• Faster Convergence: For correlated inputs, NLMS converges faster than standard LMS with equivalent steady-state error.
• Regularization Parameter δ: Prevents division by zero during silent intervals and improves numerical conditioning. The normalized formulation is essential for wideband DPD where signal envelope power fluctuates dramatically.

The learning rate directly determines the tracking capability of the adaptive DPD system in non-stationary environments:

• Thermal Tracking: PA characteristics drift with temperature on millisecond-to-second timescales; μ must be sufficient to follow this slow parameter wander.
• Doherty Load Modulation: In Doherty PAs, the effective nonlinearity changes with instantaneous envelope power; higher μ enables tracking of these dynamic AM-AM/AM-PM characteristics.
• Forgetting Factor Equivalence: In Recursive Least Squares (RLS) algorithms, the forgetting factor λ plays an analogous role — λ closer to 1 provides better steady-state performance, while smaller λ enables faster tracking.

Fixed-point FPGA implementations impose constraints on learning rate selection:

• Quantization Effects: Very small μ values may suffer from stalling where coefficient updates fall below the LSB threshold, halting adaptation entirely.
• Gain Scaling: The learning rate is often implemented as a power-of-two bit shift to replace multipliers with barrel shifters, restricting μ to discrete values like 2^(-k).
• Coefficient Freeze Logic: When input power drops below a threshold, μ is effectively set to zero to prevent noise-driven coefficient drift in the absence of excitation.
• Loop Delay Compensation: The effective learning rate must account for the total loop delay through the TX chain and feedback receiver to maintain stability.

The foundational iterative optimization method that updates DPD coefficients by computing the gradient of the cost function on small batches of signal samples. In online training, SGD directly uses the instantaneous error signal to step toward the optimal predistorter inverse. The learning rate scales this gradient, determining the magnitude of each coefficient adjustment. A learning rate that is too large causes divergence or oscillation around the minimum, while a rate that is too small results in impractically slow convergence.

A lightweight adaptive filtering algorithm prized for its minimal computational footprint in FPGA-based DPD implementations. LMS approximates the true gradient using the instantaneous squared error, making it a specific case of SGD. The learning rate in LMS directly trades off convergence rate against steady-state misadjustment — the residual error floor after adaptation settles. For signals with high PAPR, a fixed learning rate struggles because high-power samples dominate the update, potentially destabilizing the loop.

An enhancement to LMS that normalizes the learning rate by the power of the input basis function vector. This makes the effective step size self-adjusting: smaller updates occur for high-power signal samples, preventing gradient overshoot, while larger updates accelerate convergence during low-power periods. NLMS is particularly valuable in DPD systems processing modern communication waveforms with extreme peak-to-average power ratios, where signal power fluctuates dramatically across the modulation envelope.

A second-order adaptive algorithm that recursively computes the exact inverse of the correlation matrix, achieving an order-of-magnitude faster convergence than LMS-family algorithms. The forgetting factor λ acts as an exponential weighting mechanism analogous to a dynamic learning rate, controlling how quickly old data is discounted. RLS excels in tracking time-varying PA nonlinearities caused by thermal memory effects, but its O(N²) computational complexity demands careful hardware resource planning for real-time FPGA implementation.

The number of algorithm iterations required for the DPD coefficients to approach their steady-state optimal values within an acceptable error margin. A higher learning rate accelerates convergence but risks ill-conditioning and instability. In closed-loop DPD, the convergence rate must be fast enough to track PA behavioral changes due to temperature drift, voltage sag, and carrier frequency shifts. The theoretical convergence bound is determined by the eigenvalue spread of the input signal's correlation matrix — a wider spread demands a smaller learning rate for stability.

A scalar parameter λ (0 < λ ≤ 1) in recursive algorithms like RLS that exponentially weights past data, enabling the estimator to track non-stationary PA characteristics. A smaller λ applies heavier discounting, effectively increasing the algorithm's responsiveness to recent measurements — analogous to increasing the learning rate in gradient methods. However, aggressive forgetting amplifies estimation noise. In DPD applications, λ is typically tuned between 0.95 and 0.999, balancing rapid adaptation to thermal memory effects against steady-state coefficient variance.