Volterra Series

The Volterra series extends the concept of a Taylor series to systems with memory. The output is not just a function of the instantaneous input, but a sum of multidimensional convolution integrals of increasing order.

• Taylor Series: Models static nonlinearity (no memory).
• Volterra Series: Models dynamic nonlinearity by integrating over past input values.
• Kernel Role: Each term's kernel acts as a higher-order impulse response, quantifying how combinations of past inputs influence the present output.

The Volterra series is a universal approximator for a wide class of nonlinear time-invariant systems, meaning it can theoretically model any continuous nonlinear dynamic system to arbitrary accuracy.

• Parameter Explosion: The number of coefficients grows exponentially with the nonlinear order and memory depth.
• Practical Barrier: A full Volterra model for a typical wideband power amplifier can require thousands of coefficients, making real-time implementation infeasible.
• Consequence: This complexity directly motivates the use of simplified models like the Memory Polynomial and Generalized Memory Polynomial.

When formulated in the complex baseband, the Volterra series inherently captures both amplitude and phase distortions that are critical for power amplifier modeling.

• AM-AM Distortion: The nonlinear relationship between input amplitude and output amplitude, representing gain compression or expansion.
• AM-PM Distortion: The nonlinear relationship between input amplitude and output phase shift, a primary cause of spectral regrowth.
• Complex Kernels: The kernels themselves are complex-valued, allowing them to model the simultaneous amplitude and phase perturbations introduced by the amplifier.

The series is organized into distinct nonlinear orders, each representing a different degree of nonlinearity. This structure allows engineers to truncate the series based on the dominant distortion mechanisms.

• 1st Order (Linear): A standard linear convolution, representing the amplifier's small-signal gain and filtering.
• 3rd Order: The dominant source of in-band distortion and spectral regrowth near the carrier in differential amplifiers.
• 5th Order and Above: Models higher-order intermodulation products that affect wider bandwidths and deeper compression points.
• Odd-Order Dominance: Differential circuit topologies naturally suppress even-order harmonics, making odd-order kernels the primary focus in DPD.

Volterra kernels possess inherent symmetries that can be exploited to reduce the number of unique coefficients without losing information.

• Permutation Symmetry: The kernel value is unchanged if its time arguments are swapped (e.g., h₂(τ₁, τ₂) = h₂(τ₂, τ₁)).
• Triangular Kernel: By enforcing symmetry, the 2D integration domain for a second-order kernel can be reduced to a triangular region, halving the number of required coefficients.
• Practical Impact: This is a fundamental first step in model pruning before applying more advanced techniques like LASSO regression or tensor decomposition.

The Volterra series serves as the theoretical parent for a family of more practical block-structured and pruned models used in real-time DPD systems.

• Memory Polynomial: Derived by keeping only the diagonal terms of the Volterra kernels.
• Generalized Memory Polynomial: Adds cross-terms between the signal and its envelope to capture more complex memory effects.
• Parallel Hammerstein: A block-structured equivalent that represents a subclass of the full Volterra series.
• Sparse Volterra: A direct pruning approach that uses algorithms like Orthogonal Matching Pursuit to select only the most significant kernel coefficients.