Volterra Series

The Volterra series generalizes the linear convolution integral to higher orders by applying multi-dimensional convolution kernels. The output is expressed as a sum of increasingly complex terms:

• 1st-order kernel: Standard linear impulse response
• 2nd-order kernel: Captures quadratic interactions between two time instances
• 3rd-order kernel: Models cubic interactions across three time instances Each kernel operates on products of delayed input samples, enabling the model to capture both harmonic distortion and intermodulation products simultaneously.

Unlike static polynomial models, the Volterra series explicitly incorporates temporal memory effects through its kernel structure. The integration over multiple time variables allows the current output to depend on past inputs:

• Short-term memory: Captured by the kernel's decay along the delay axis
• Thermal memory: Modeled through slowly varying kernel coefficients
• Trapping effects: Represented by asymmetric kernel shapes The kernel support region defines the memory depth, typically truncated to a finite length M for practical implementation in DPD systems.

The full Volterra series is computationally intractable due to exponential growth in coefficients. Practical DPD implementations apply strategic reductions:

• Nonlinearity order truncation: Limiting to 3rd, 5th, or 7th order terms
• Memory depth truncation: Restricting delay taps to a finite window
• Dynamic deviation reduction: Separating static nonlinearity from low-order memory effects
• Near-diagonal pruning: Retaining only kernel coefficients near the main diagonal where most signal energy concentrates These techniques reduce coefficient count from thousands to hundreds while preserving linearization accuracy.

The Generalized Memory Polynomial (GMP) is a pruned subset of the full Volterra series specifically optimized for power amplifier modeling. It retains three categories of cross-terms:

• Aligned terms: Signal and envelope powers at the same delay
• Lagging cross-terms: Signal multiplied by lagged envelope powers
• Leading cross-terms: Signal multiplied by advanced envelope powers This structure captures the nonlinear memory effects dominant in Doherty and envelope tracking amplifiers while maintaining a coefficient count suitable for real-time adaptation at gigahertz sampling rates.

The Volterra series exhibits uniform convergence for continuous nonlinearities with finite memory when the input signal amplitude remains bounded. Key mathematical properties include:

• Fading memory requirement: The system's dependence on past inputs must decay sufficiently fast
• Radius of convergence: Determined by the amplifier's saturation characteristics
• Analyticity condition: The nonlinear transfer function must be representable by a convergent power series For strongly saturating amplifiers near compression, the series may diverge, motivating alternative basis functions such as orthogonal polynomials or spline-based models.

Extracting Volterra kernels from measured input-output data requires solving a linear regression problem in the kernel coefficients. Common identification approaches include:

• Least squares estimation: Direct pseudo-inverse solution using training data matrices
• Recursive least squares (RLS): Adaptive coefficient tracking for time-varying systems
• Orthogonal basis expansion: Using Laguerre or Kautz functions to improve numerical conditioning
• Cross-correlation techniques: Exploiting statistical properties of Gaussian input signals The indirect learning architecture applies these methods to identify the predistorter coefficients by swapping the PA input and output during training.

The dependence of a power amplifier's current output on past input values, violating the assumption of memoryless non-linearity. These effects arise from multiple physical mechanisms:

• Thermal dynamics: Transistor junction temperature changes with signal envelope, altering gain characteristics over microsecond timescales
• Bias network impedance: Low-frequency envelope components modulate the transistor's operating point through non-ideal biasing circuits
• Trapping effects: Charge capture and release in semiconductor defects creates long time-constant dependencies
• Electrical memory: Energy storage in matching network reactive components introduces short-term temporal correlations

The process of determining the Volterra kernel coefficients that define the system's non-linear behavior. For DPD applications, this involves solving a linear regression problem in the expanded feature space.

• Least squares estimation is the standard approach for coefficient extraction from input-output measurements
• The indirect learning architecture swaps the PA model's input and output to directly identify the predistorter
• Regularization techniques such as ridge regression prevent overfitting when the kernel order is high
• Recursive least squares enables online adaptation to track time-varying amplifier characteristics

Techniques to reduce the exponential growth of Volterra kernel coefficients as non-linearity order and memory depth increase. A full Volterra series quickly becomes computationally intractable for practical DPD implementation.

• Pruning removes terms with negligible contribution to the output based on statistical significance tests
• Principal component analysis projects the high-dimensional kernel space onto a lower-dimensional subspace
• Orthogonal search methods select the most relevant basis functions sequentially
• Sparse regression using L1 regularization automatically identifies the minimal set of essential kernel terms

The mathematical conditions under which a Volterra series representation accurately approximates a non-linear system. Understanding these properties is critical for reliable DPD performance.

• Fading memory systems, where the influence of past inputs decays over time, are well-represented by Volterra series
• The series converges for analytic non-linearities with bounded inputs within a certain radius of convergence
• Hammerstein and Wiener models are special cases of the Volterra series with restricted kernel structures
• Orthogonal basis expansions such as Wiener G-functionals guarantee convergence for Gaussian input signals

A comparison of data-driven neural approaches with the classical Volterra-based behavioral modeling framework for power amplifier linearization.

• Volterra DPD: Provides a principled mathematical structure with interpretable kernels, linear-in-parameters estimation, and well-understood convergence properties
• Neural DPD: Offers universal approximation capability without requiring explicit kernel selection, but introduces non-convex optimization challenges
• Real-valued time-delay neural networks can be shown to implement a subset of Volterra series operations
• Hybrid approaches use Volterra-inspired feature extraction layers followed by neural network processing to combine the strengths of both methods