Efficient Kernel Computation for Volterra Filter Structure Evaluation

José Henrique de Morais Goulart, P.m.s. Burt

Abstract

Despite their generality, conventional Volterra filters are inadequate for some applications, due to the huge number of parameters that may be needed for accurate modelling. When a state-space model of the target system is known, this can be assessed by computing its kernels, which also provides valuable information for choosing an adequate alternate Volterra filter structure, if necessary, and is useful for validating parameter estimation procedures. In this letter, we derive expressions for the kernels by using the Carleman bilinearization method, for which an efficient algorithm is given. Simulation results are presented, which confirm the usefulness of the proposed approach.

Introduction

Volterra filters (VF) are a powerful tool for modeling nonlinear systems with memory, such as audio equipment or communication channels. However, they suffer from a "curse of dimensionality": the number of parameters grows exponentially with the nonlinearity order and memory length. If a physical state-space model of a system is available, researchers can theoretically compute its kernels (the coefficients of the Volterra series) to determine the necessary filter complexity. Unfortunately, existing analytical and numerical methods are often computationally prohibitive or numerically unstable.

This paper introduces a highly efficient algorithm for computing the kernels of Volterra filters from a known state-space model using Carleman bilinearization. By eliminating redundant calculations, the proposed method achieves significant computational speedups, providing a practical way for engineers to evaluate and choose the best filter structure for nonlinear systems.

Methodology: Efficient Carleman Bilinearization

The core innovation lies in the application of Carleman bilinearization (or linearization) to transform a nonlinear state-space system into a bilinear one, which possesses closed-form solutions for its kernels.

Standard Carleman methods produce massive matrices filled with redundant terms because they do not account for the symmetry in higher-order products of state variables. The authors propose a recursive algorithm that:

  1. Computes only non-redundant components.
  2. Builds higher-order rows using previously calculated lower-order rows.
  3. Calculates the Generalized Frequency Response Functions (GFRFs) directly.

These GFRFs are then converted to time-domain kernels via an Inverse Fast Fourier Transform (IFFT). This frequency-domain approach avoids aliasing issues common in direct time-domain sampling. Managing redundancy is critical; for a system with 3 state variables and a 4th-order nonlinearity, the non-redundant matrices contain less than 9% of the elements of their redundant counterparts.

Results and Discussion

The algorithm was evaluated using a simulated physical model of a bass loudspeaker. The model's output was compared against the Volterra filter's output using a "fit index" (FIT %) across 50 realizations of Gaussian noise.

Computational Efficiency

The proposed algorithm demonstrates a massive speedup compared to direct approaches. As shown in the table below, the efficiency gains increase dramatically with the number of state variables ($m$) and the nonlinearity order ($P$).

Table I: Speedup ratios of the proposed algorithm compared to the direct method for various system orders.

For a system with 5 state variables and a 5th-order nonlinearity, the proposed method is over 611 times faster than the standard approach.

Accuracy and Filter Design

Simulation results confirm that while linear approximations fail as input variance increases, 2nd and 3rd-order Volterra filters maintain high accuracy.

Figure 2: Fit index (%) vs. variance of the input for different Volterra filter orders.

These results allow engineers to make informed design decisions. For instance, achieving a 98% fit at high power might require moving from a 2nd-order filter (4,195 coefficients) to a 3rd-order filter (129,775 coefficients). This 31x increase in complexity can now be quantified and justified before implementation.

Conclusion

The proposed recursive Carleman bilinearization algorithm provides a robust tool for deriving Volterra kernels from state-space models. By drastically reducing computational overhead, it enables the practical evaluation of filter structures and serves as a ground truth for validating other parameter estimation procedures. While the method requires an accurate analytic state-space model, it bridges a critical gap in nonlinear system modeling and digital filter design.

Summary updated

Abstract

Despite their generality, conventional Volterra filters are inadequate for some applications, due to the huge number of parameters that may be needed for accurate modelling. When a state-space model of the target system is known, this can be assessed by computing its kernels, which also provides valuable information for choosing an adequate alternate Volterra filter structure, if necessary, and is useful for validating parameter estimation procedures. In this letter, we derive expressions for the kernels by using the Carleman bilinearization method, for which an efficient algorithm is given. Simulation results are presented, which confirm the usefulness of the proposed approach.
BibTeX
@article{goulart2012efficient,
  author = {José Henrique de Morais Goulart and P.M.S. Burt},
  title = {{Efficient Kernel Computation for Volterra Filter Structure Evaluation}},
  journal = {IEEE Signal Processing Letters},
  year = {2012},
  doi = {10.1109/lsp.2011.2182342},
}
Figure