Abstract
In a very recent paper "On eigenvalues and equivalent transformation of trigonometric matrices" (D. Zhang, Z. Lin, and Y. Liu, LAA 436, 71–78 (2012)), the authors motivated and discussed a trigonometric matrix that arises in the design of finite impulse response (FIR) digital filters. The eigenvalues of this matrix shed light on the FIR filter design, so obtaining them in closed form was investigated. Zhang et al. proved that their matrix had rank-4 and they conjectured closed form expressions for its eigenvalues, leaving a rigorous proof as an open problem. This paper studies trigonometric matrices significantly more general than theirs, deduces their rank, and derives closed-forms for their eigenvalues. As a corollary, it yields a short proof of the conjectures in the aforementioned paper.
Introduction
The design of Finite Impulse Response (FIR) digital filters is a cornerstone of modern signal processing. A critical component of this design process involves the analysis of specific $2n \times 2n$ trigonometric matrices. In 2012, researchers Zhang, Lin, and Liu identified a class of such matrices where the entries are composed of sine and cosine functions scaled by linear terms. While they were able to demonstrate through numerical experiments that these matrices consistently possess a rank of 4, providing a rigorous proof for the closed-form expressions of their non-zero eigenvalues remained an open challenge.
This paper addresses this "trigonometric puzzle" by moving beyond the specific instance studied by Zhang et al. and considering a much broader family of generalized trigonometric matrices. By identifying the underlying algebraic structure, we provide a definitive proof of the 2012 conjectures and offer a general framework for calculating eigenvalues in this class of matrices.
Methodology: Generalization and Trace Identities
The core of the approach lies in a "generalize to simplify" strategy. Rather than focusing solely on the specific FIR matrix, we define a generalized trigonometric matrix $P$ where the scaling is governed by an arbitrary symmetric rank-2 matrix $L$.
Rank Analysis
The first step involves establishing the rank of the matrix. By leveraging the properties of Hadamard (entry-wise) and Kronecker products, it is proven that these matrices have a rank of at most 4. This low-rank property is essential, as it implies that the vast majority of the eigenvalues are zero, reducing the problem to finding the four non-zero values.
Trace Identities and Newton’s Identities
The primary mathematical breakthrough is the derivation of a trace identity. It is proven that the trace of the $m$-th power of the large $2n \times 2n$ matrix $P$ is exactly twice the trace of the $m$-th power of the smaller $n \times n$ scaling matrix $L$: $$\text{tr}(P^m) = 2 \text{tr}(L^m) \quad \text{for } m=1, 2, 3, 4.$$
This identity is significant because it allows us to bypass the complex trigonometric oscillations of the larger matrix. By applying Newton’s identities, which relate the traces of matrix powers to the coefficients of the characteristic polynomial, the problem of finding eigenvalues for a high-dimensional matrix is reduced to solving a relatively simple quartic equation.
Main Results
The derivation yields explicit, closed-form formulas for the eigenvalues. For any matrix in this class defined by a rank-2 scaling matrix $L = lh^T + hl^T$, the four non-zero eigenvalues are given by: $$\lambda_{1,2} = \gamma + \delta, \quad \lambda_{3,4} = \gamma - \delta$$ where the scalars are defined as $\gamma = l^T h$ and $\delta = \sqrt{(l^T l)(h^T h)}$.
Verification of the FIR Conjecture
As a direct corollary, this general formula provides a short and rigorous proof for the specific FIR filter design matrix conjectured by Zhang et al. The eigenvalues for their specific case are confirmed to be: $$\lambda_{\pm} = \frac{n}{4} \left( n - 1 \pm \sqrt{\frac{4n^2 - 6n + 2}{3}} \right)$$ Each of these eigenvalues occurs with a multiplicity of 2, consistent with the rank-4 observation.
Discussion and Implications
One of the most striking insights from this analysis is the frequency independence of the eigenvalues. The eigenvalues of the matrix $P(\omega)$ do not depend on the digital frequency variable $\omega$. In the context of filter design, this is a highly desirable property, as it suggests that the stability and performance characteristics of the filter remain uniform across the frequency spectrum.
The success of this derivation highlights the power of identifying hidden low-rank structures within seemingly complex matrices. While this work focuses on rank-2 scaling matrices, the methodology suggests that similar derivations could be performed for higher-rank matrices, although the algebraic complexity would increase significantly. By providing a formal proof for the Zhang et al. conjecture, this paper closes an open problem in the literature and provides a robust tool for future research in signal processing and linear algebra.
Summary updated