Direct Learning Architecture

DLA operates by forming a feedback loop around the power amplifier. The system compares the original baseband input signal with the attenuated, down-converted PA output. An adaptive algorithm—typically Least Squares (LS) or Recursive Least Squares (RLS) —iteratively adjusts the pre-distorter coefficients to minimize the mean squared error (MSE) between these two signals. This direct minimization ensures that the combined DPD+PA cascade converges to a linear gain.

Unlike the Indirect Learning Architecture (ILA) , DLA does not require explicit identification of a post-distorter or a separate PA model. The adaptive algorithm inherently solves for the inverse nonlinear characteristic of the PA. By treating the DPD and PA as a single combined system, DLA directly extracts the pre-inverse function, avoiding the copy error problem associated with ILA where a post-distorter model is imperfectly transferred to the pre-distorter.

A critical advantage of DLA is its ability to compensate for transmitter chain impairments that occur after the DPD block. Errors such as IQ imbalance, LO leakage, and quadrature modulator nonlinearity are included within the closed feedback loop. The DLA algorithm automatically pre-compensates for these errors because it observes the final RF output, treating the entire transmitter chain as a single nonlinear dynamic system to be linearized.

DLA supports continuous online training to track changes in PA behavior due to thermal drift, aging, or antenna load mismatch (VSWR) . The coefficient estimation block processes batches of captured time-domain samples. Advanced implementations use regularization techniques to prevent over-parameterization and ensure numerical stability of the matrix inversion, often employing QR decomposition or Cholesky factorization for efficient hardware implementation.

The performance of DLA is heavily dependent on the condition number of the data covariance matrix used in coefficient estimation. Highly correlated input signals can lead to ill-conditioned matrices, causing coefficient divergence. To mitigate this, DLA implementations often incorporate Tikhonov regularization or variable step-size algorithms. Proper time alignment between the reference and feedback signals is critical; sub-sample delay estimation is mandatory to prevent model bias.

In FPGA or ASIC implementations, the DLA feedback path requires a high-linearity observation receiver with sufficient dynamic range to capture the PA output without adding its own distortion. The loop delay through the DPD, DAC, PA, coupler, and ADC must be precisely calibrated. Crest factor reduction (CFR) is often integrated before the DPD to limit peak signal excursions, ensuring the PA operates in a region where the DLA can effectively linearize without clipping the pre-distorted signal.

The primary alternative to the direct method. ILA first identifies a post-distorter by minimizing the error between the PA output and the post-distorter output. The converged post-distorter coefficients are then copied to the pre-distorter. This avoids the need to model the PA inverse directly but assumes commutability, which can break down in the presence of significant measurement noise or strong nonlinearities.

The defining characteristic of the Direct Learning Architecture. The system continuously monitors the power amplifier output and feeds it back to the coefficient estimator. This creates a self-correcting loop that can track time-varying changes such as thermal drift, aging, and antenna impedance mismatch. The feedback path requires a high-linearity observation receiver to avoid contaminating the error signal.

The mathematical engine of the DLA. Algorithms like Least Squares (LS), Recursive Least Squares (RLS), or Least Mean Squares (LMS) iteratively solve for the pre-distorter parameters by minimizing the error between the desired input and the attenuated PA output. The choice of algorithm involves a trade-off between convergence speed and computational complexity for FPGA implementation.

A critical distinction in DPD design. Model extraction first identifies a behavioral model of the PA, then analytically derives the inverse. Direct learning bypasses explicit PA modeling entirely, directly identifying the pre-distorter coefficients that minimize the output error. DLA is generally more robust to model structural mismatch but requires a stable closed-loop convergence guarantee.

While DLA avoids explicit inverse modeling, understanding the PA's behavior is crucial for setting the pre-distorter model structure (e.g., Memory Polynomial order and depth). The pre-distorter must be complex enough to invert the PA's nonlinearity. Key PA characteristics include:

• AM-AM Distortion: Gain compression
• AM-PM Distortion: Phase shift vs. amplitude
• Memory Effects: Thermal and electrical lag

The adaptive algorithms that run continuously within the DLA feedback loop. Unlike offline batch processing, online algorithms update coefficients sample-by-sample or block-by-block without interrupting transmission. Key challenges include maintaining numerical stability with ill-conditioned data matrices and preventing coefficient divergence during signal dropouts or PA saturation events.