Applying the Spectral Method for Modeling Linear Filters: Butterworth, Linkwitz-Riley, and Chebyshev filters

Technical Summary: Applying the Spectral Method for Modeling Linear Filters

Problem Statement
The paper addresses the challenge of computer modeling linear filters (specifically Butterworth, Linkwitz–Riley, and Chebyshev types) in continuous time. Traditional computer modeling approaches typically rely on time discretization, converting analog systems into discrete-time equivalents. The authors note that these methods introduce specific artifacts: the impulse invariance method suffers from aliasing if the discretization frequency is insufficient, while the bilinear transform method introduces frequency warping due to the nonlinear transformation of the frequency axis. The paper proposes a technique that avoids time discretization entirely, aiming to model the output signal as a continuous-time function directly.

Methodology
The core of the proposed technique is the spectral form of mathematical description of linear systems. This approach represents input and output signals as orthogonal expansions (spectral characteristics) relative to a chosen basis, rather than as discrete time samples.

Key methodological components include:

• Spectral Representation: Signals $g(t)$ and $x(t)$ are represented as infinite column matrices of expansion coefficients ($G$ and $X$) relative to an orthonormal basis $\{q(i, t)\}$.
• Two-Dimensional Non-Stationary Transfer Function: The linear filter itself is described not by a scalar transfer function $H(s)$, but by a two-dimensional non-stationary transfer function matrix $W$. This matrix relates the input spectral characteristic to the output via the linear algebraic operation $X = WG$.
• Derivation of $W$: The matrix $W$ is derived from the standard transfer function $H(s)$ (defined by poles and zeros) by mapping the complex variable $s$ to a derivative block matrix $P$. The relationship is expressed as $W = (a_n P^n + \dots + a_0 E)^{-1}(b_m P^m + \dots + b_0 E)$, where $P$ is the spectral representation of the derivative operator and $E$ is the identity matrix.
• Basis Selection: For the numerical experiments, the authors utilize a cosine basis (Fourier series) to ensure smoothness and high approximation accuracy.
• Truncation: For computer implementation, the infinite matrices and vectors are truncated to finite sizes ($L \times L$).
• Error Metrics: The study evaluates performance by comparing the filtered output (with phase delay compensation) against the original useful signal. The error is calculated in the $L^2$ norm over the time interval. The methodology distinguishes between the inherent error of the non-ideal filter and the error introduced by the spectral method's truncation.

Key Contributions

1. New Modeling Technique: The paper introduces a method for modeling linear filters that operates entirely in the spectral domain, allowing for continuous-time output representation without the aliasing or frequency warping associated with discrete-time discretization methods.
2. Application to Specific Filter Families: The authors explicitly derive the two-dimensional non-stationary transfer functions ($W$) for:
   • Butterworth Filters: Derived from the roots of unity.
   • Linkwitz–Riley Filters: Modeled as the cascade (square) of two Butterworth filters.
   • Chebyshev Type I and Type II Filters: Derived using Chebyshev polynomials and Joukowski transformations.
3. Algebraic Formulation: The work demonstrates that filtering can be reduced to linear algebraic operations on expansion coefficients, combining the analytical rigor of continuous systems with the implementability of numerical methods.

Results
The technique was tested using both deterministic noise (specific frequency components) and random Gaussian white noise.

• Deterministic Noise: The error ($E$) between the filtered output and the useful signal was significantly lower than the a priori error (the noise energy itself), demonstrating effective noise suppression. The results showed that increasing the filter order generally improved quality, while increasing the truncation order ($L$) beyond a certain point (e.g., $L=128$) yielded diminishing returns, indicating that the spectral method error was negligible compared to the filter's inherent limitations.
• Random Noise: Statistical analysis (mean and standard deviation of error over $10^4$ realizations) confirmed that the spectral method effectively reduces noise. The error estimates remained lower than the a priori noise energy, though the filters were less effective than with deterministic noise due to noise components falling within the passband.
• Computational Complexity: The authors estimate the complexity of finding the transfer function matrix as $O(L^3)$ using standard algorithms, noting that fast algorithms (e.g., Strassen) could reduce this.

Significance and Claims
The paper claims that the spectral method offers a distinct advantage by modeling filters in continuous time, thereby avoiding the artifacts of time discretization. The authors state that, to their knowledge, this spectral form has not been previously applied to computer modeling of these specific filter families.

The significance of the work lies in its universality; the technique is presented as applicable to other linear filters (such as elliptic, Bessel, Gaussian, and Legendre filters) once their transfer functions are known. The authors emphasize that the method combines the benefits of analytical methods (smoothness, continuous time) with the simplicity of numerical implementation. The study concludes that the spectral method error is insignificant compared to the filter error, validating the approach for practical computer modeling of linear systems.