Abstract
The Volterra-PARAFAC (VP) nonlinear system model, which consists of a FIR filterbank followed by a memoryless nonlinearity, aims at offering a good compromise between accuracy and parametric complexity. Here, for an even better compromise, we propose a generalization with IIR filters (VPI model) and evaluate both models. For the evaluation, we consider the concrete case of two audio loudspeakers and initially compute reference Volterra kernels from their known physical state-space models, using an efficient procedure. Then, VP and VPI models are derived and their accuracy is tested. As shown, the VPI models have in this case only 15 to 26% of the parametric complexity of VP models with the same accuracy, which points to a great potential for accurate and efficient nonlinear system modeling.
Introduction
In signal processing, many physical systems—ranging from audio loudspeakers to satellite communication channels—exhibit nonlinear behavior. While linear models are often preferred for their simplicity, they fail to capture the distortion and memory effects inherent in these systems. The Volterra Filter (VF) is considered the gold standard for nonlinear modeling; however, it suffers from the "curse of dimensionality." As the model order or memory length increases, the number of parameters grows exponentially, making real-time implementation computationally prohibitive.
To address this, previous research introduced the Volterra-PARAFAC (VP) model. By applying tensor decomposition (Parallel Factor or PARAFAC) to the Volterra kernels, the system is simplified into a bank of parallel Finite Impulse Response (FIR) filters. While more efficient than standard VFs, FIR filters can still be bulky when modeling systems with long impulse responses. This paper evaluates a further refinement: the Volterra-PARAFAC IIR (VPI) model, which replaces FIR filters with low-order Infinite Impulse Response (IIR) filters.
The VPI Modeling Framework
The core innovation of the VPI model lies in its structural compactness. The PARAFAC decomposition maps a complex Volterra series into a parallel structure of linear filters followed by power-order nonlinearities.
The transition from VP to VPI involves a three-step methodology:
- Reference Kernel Generation: Ground truth Volterra kernels are derived from physical state-space models of loudspeakers using Carleman linearization.
- Tensor Decomposition: The kernels are decomposed using an Alternating Least Squares (ALS) algorithm to identify the parallel FIR filterbank.
- IIR Approximation: The resulting long FIR filters are approximated by low-order IIR filters using balanced realization and Hankel singular value reduction. This process discards the least significant states of the filter, drastically reducing the parameter count while maintaining the impulse response's essential characteristics.
Experimental Evaluation
The VPI model was tested against two specific hardware components: a small 82 mm automotive loudspeaker (Loudspeaker A) and a larger 130 mm loudspeaker (Loudspeaker B). These were stimulated with zero-mean Gaussian input signals at varying power levels to trigger different degrees of nonlinear distortion.
Results and Performance Analysis
The quantitative results demonstrate a significant leap in efficiency. For Loudspeaker A, a standard VP model required 1,300 parameters to reach a specific accuracy threshold. The VPI version achieved nearly identical performance with only 264 parameters. Across all tested configurations, the VPI model required only 15% to 26% of the parameters used by the VP model.
As shown in the performance plots, the "FIT" percentage—a metric of modeling accuracy—remains remarkably stable even as input power increases and the system's nonlinearities become more pronounced. The VPI model successfully captures these complex dynamics with a fraction of the computational overhead.
Conclusion
The Volterra-PARAFAC IIR model offers a superior compromise between accuracy and parametric complexity. By leveraging the efficiency of IIR filters within a tensor-decomposed Volterra framework, researchers can model complex nonlinear systems like loudspeakers with unprecedented efficiency. Future work may explore joint design procedures to optimize IIR filters and tensor components simultaneously, potentially unlocking even greater reductions in model size.
Summary updated