Volterra Kernel

A Volterra kernel is a multidimensional generalization of the linear impulse response. While a linear system is fully characterized by a 1D function h(τ), the k-th order Volterra kernel h_k(τ_1, ..., τ_k) is a k-dimensional function that quantifies how k delayed versions of the input interact nonlinearly to produce the output. For a power amplifier, the first-order kernel captures linear gain, the third-order kernel captures third-order intermodulation distortion, and the fifth-order kernel captures fifth-order compression effects.

Volterra kernels exhibit triangular symmetry to ensure a unique representation. Since the product of input samples x(t-τ_1)·x(t-τ_2) is commutative, the kernel values h_2(τ_1, τ_2) and h_2(τ_2, τ_1) are indistinguishable. To avoid redundancy, kernels are constrained such that:

• Triangular domain: h_k(τ_1, ..., τ_k) is defined only for τ_1 ≤ τ_2 ≤ ... ≤ τ_k
• Symmetric extension: Values outside this domain are obtained by permuting indices This symmetry reduces the number of independent coefficients by a factor of k! for the k-th order kernel.

Each kernel order corresponds to a specific physical nonlinear mechanism in the system:

• 1st-order kernel: Linear transfer function, representing small-signal gain and phase response
• 2nd-order kernel: Quadratic nonlinearity, capturing harmonic generation and DC offset effects
• 3rd-order kernel: Cubic nonlinearity, the dominant source of AM-AM compression and AM-PM conversion in differential amplifiers
• 5th-order kernel: Higher-order compression effects visible near saturation Odd-order kernels dominate in push-pull and differential PA architectures due to even-order cancellation.

The kernel's structure reveals the nature of memory effects:

• Diagonal elements h_k(τ, τ, ..., τ): Represent contributions where all delayed samples are at the same time instant. These capture static nonlinearity and are retained in memory polynomial models.
• Off-diagonal elements h_k(τ_1, τ_2, ..., τ_k) with τ_i ≠ τ_j: Represent cross-memory effects where input samples at different time lags interact nonlinearly. These capture complex dynamic phenomena like envelope memory and are often pruned in simplified models. The Generalized Memory Polynomial explicitly includes cross-terms between the signal and its lagging envelope to capture these off-diagonal interactions.

In discrete-time system identification, the Volterra kernels are estimated as coefficient tensors using least squares regression. The output is expressed as a linear combination of kernel coefficients multiplied by polynomial basis functions of delayed inputs:

• The regression matrix Φ contains all polynomial combinations of delayed input samples
• The coefficient vector θ contains all kernel values flattened into a single column
• The solution θ = (Φ^H Φ)^(-1) Φ^H y minimizes the squared error between modeled and measured output Regularization techniques like LASSO are often applied to enforce sparsity and prevent overfitting when the number of kernel coefficients is large.

Full Volterra kernels suffer from the curse of dimensionality: the number of coefficients grows exponentially with memory depth and nonlinear order. Tensor decomposition techniques address this by factorizing the kernel into lower-rank structures:

• Canonical Polyadic Decomposition (CPD): Expresses the k-th order kernel as a sum of R rank-one tensors, reducing O(M^k) parameters to O(kMR) where M is memory depth and R is the tensor rank
• Parallel Cascade Representation: Decomposes the Volterra system into a bank of parallel Hammerstein or Wiener branches These decompositions enable compact Volterra models suitable for real-time DPD implementation on FPGAs.

A mathematical power series with memory that models nonlinear dynamic systems by representing the output as a sum of multidimensional convolution integrals. Each term in the series is defined by a Volterra kernel that quantifies the contribution of a specific nonlinear order and memory depth combination.

• First-order kernel: linear impulse response
• Second-order kernel: quadratic nonlinear interactions
• Third-order kernel: cubic distortion dominant in differential PAs
• Higher-order kernels capture progressively weaker nonlinear effects

A simplified Volterra model that retains only the diagonal terms of the Volterra kernels, where all delayed samples share the same time index. This dramatically reduces complexity while effectively capturing nonlinear memory effects in power amplifiers.

• Eliminates cross-terms between different delays
• Parameter count scales linearly with memory depth
• Captures AM-AM and AM-PM distortion with memory
• Standard baseline for DPD implementation

An enhanced memory polynomial model that includes cross-terms between the signal and its lagging envelope values. These cross-terms capture complex memory effects in wideband power amplifiers that diagonal-only models miss.

• Adds envelope-lag cross-terms to the standard memory polynomial
• Captures interactions between current signal and past envelope values
• Significantly improves modeling accuracy for wideband signals
• Higher parameter count than standard memory polynomial

The exponent of the input signal in a Volterra series term, defining the degree of nonlinearity being modeled. Each kernel in the series corresponds to a specific nonlinear order that determines which distortion products are captured.

• Odd orders (3rd, 5th, 7th): dominate in differential PA distortion
• Even orders (2nd, 4th): typically filtered by bandpass response
• Higher orders model gain compression at saturation
• Order selection balances accuracy against parameter count

The number of past input samples considered in a Volterra or memory polynomial model, determining the temporal span over which memory effects are captured. Each kernel dimension corresponds to a specific delay tap.

• Short-term memory: bias network impedance effects (nanoseconds)
• Long-term memory: thermal dynamics and trapping (microseconds to milliseconds)
• Deeper memory increases parameter count exponentially in full Volterra
• Memory polynomial parameter count scales linearly with depth

A Volterra model simplification technique that separates static nonlinearity from low-order dynamic behavior. It assumes the system is weakly nonlinear with memory, drastically reducing the number of parameters needed.

• Decomposes output into static + dynamic deviation components
• Retains only first-order dynamics around the static nonlinearity
• Reduces full Volterra complexity from exponential to near-linear
• Effective for PAs operating below deep compression