The Resonance Bias Framework: Resonance, Structure, and Arithmetic in Quadrature Error

TL;DR

The paper reframes trapezoidal-rule quadrature error for periodic functions on uniform grids as a deterministic resonance phenomenon rather than random noise. By analyzing the tractable prototype $f(x)=\sin^2(2\pi kx)$, it derives a complex resonance function $\tilde{\chi}_P(y)$ that acts as a spectral filter, yielding the exact bias $B_P[f]=\sum_{k\neq0} c_k\tilde{\chi}_P\left(\frac{k}{P}\right)$ and connecting this to the classical aliasing sum $\sum_{l\neq0} c_{lP}$ in a way that clarifies the role of grid-frequency interactions. The framework unifies Fourier error analysis with a mechanistic, number-theoretic, and geometric interpretation, extending to higher dimensions and arbitrary smooth periodic functions. This resonance-perspective explains non-monotonic error behavior, near-resonant plateaus, and the precise dependence of error on the arithmetic relation between frequencies and grid size, with potential implications for adaptive grid design and spectral-method diagnostics. Overall, the Resonance Bias Framework provides a transparent, structure-rich lens on finite-resolution quadrature errors, grounded in harmonic analysis and number theory, and promises actionable insights for numerical analysis and computational practice.

Abstract

We study the trapezoidal rule for periodic functions on uniform grids and show that the quadrature error exhibits a rich deterministic structure, beyond traditional asymptotic or statistical interpretations. Focusing on the prototype function f(x) = sin^2(2 pi k x), we derive an analytical expression for the error governed by a resonance function chi_P(y), closely related to the Dirichlet kernel, roots of unity, and discrete Fourier analysis on the group Z/PZ. This function acts as a spectral filter, connecting the integration error to arithmetic properties such as k/P and geometric phase cancellation, visualized as vector averaging on the unit circle. We introduce the Resonance Bias Framework (RBF), a generalization to arbitrary smooth periodic functions, leading to the error representation B_P[f] = sum_{k != 0} c_k chi_P(k/P). Although this is mathematically equivalent to the classical aliasing sum, it reveals a deeper mechanism: the quadrature error arises from structured resonance rather than random aliasing noise. The RBF thus provides an interpretable framework for understanding integration errors at finite resolution, grounded in number theory and geometry.

Figures (6)

• Figure 1: The Real Resonance Function $\chi_P(y)$ for $P=20$, illustrating the resonance landscape. The function $\chi_{20}(y)$ (purple line) quantifies the grid's response. Peaks of height 1 (green circles) indicate full resonance at integer relative frequencies $y=0, 1, 2$. Zeros (red crosses) indicate perfect cancellation at $y=n/P$ where $n \not\equiv 0 \pmod{P}$. The orange star marks $y=m/P=0.23$ corresponding to the prototype function $f(x)=\sin^2(2\pi(2.3)x)$. Theorem \ref{['thm:bias_sin2']} implies the bias $B_P[f] \approx -\frac{1}{2}\chi_{20}(0.23)$, indicated qualitatively by the dashed orange line.
• Figure 2: Validation: RBF vs Direct Error for $f(x) = \sin^2(2\pi(2.3)x)$. Log-log plot of the absolute error $|B_P[f]|$ versus the number of points $P$. The Direct Numerical Error $|I_P - I|$ (blue dots) is perfectly matched by the RBF Prediction $|B_P[f]|$ from Theorem \ref{['thm:bias_sin2']} (red crosses), confirming the formula. Both oscillate around the classical $O(P^{-2})$ bound (dashed line). The annotation highlights the cancellation case at $P=92$, where $m/P = 4.6/92 = 1/20$, resulting in $\chi_{92}(1/20)=0$ and the error plateauing at $B_{92}[f] \approx C(m)$.
• Figure 3: Geometric interpretation of the complex resonance function $\tilde{\chi}_P(y)$ for $P=5$. Each plot shows $P=5$ unit vectors (arrows) starting from the origin, representing $e^{2\pi i y j}$ for $j=0,...,4$. The thick black arrow represents the sum vector scaled by $1/P$, which is $\tilde{\chi}_P(y)$. (Left) Resonance ($y=1.00$): All vectors align, $\tilde{\chi}_5(1) = 1$. (Center) Cancellation ($y=0.20 = 1/5$): Vectors form roots of unity, sum is zero, $\tilde{\chi}_5(0.20) = 0$. (Right) Partial Cancellation ($y=0.30$): Vectors partially cancel, $\tilde{\chi}_5(0.30) \approx 0.20 + 0.15j$.
• Figure 4: Fine Structure of the real resonance function $\chi_P(y)$ for $P=20$ over the interval $[0, 1]$. The plot shows the function $\chi_{20}(y)$. Resonance peaks ($\chi_P=1$) occur at the endpoints $y=0$ and $y=1$ (green circles). Exact zeros ($\chi_P=0$) occur at $y=n/P$ for $n=1, \dots, 19$ (red crosses). Key rational zeros are labeled (e.g., 1/20, 1/5, 1/4, 7/20, 1/2, 3/4, 19/20), illustrating the cancellation property.
• Figure 5: The RBF Filtering View of Aliasing for $P=20$. (Top) Example Input Spectrum $|c_k|$ (log scale, for $f(x)=e^{\cos(2\pi x)}$). (Middle) Grid Filter Response magnitude $|\tilde{\chi}_P(k/P)|$. For integer $k$, this response is $1$ only when $k$ is a multiple of $P=20$ and $0$ otherwise. The filter $\tilde{\chi}_P(k/P)$ thus exhibits a sharp selectivity, isolating the resonant frequencies $k=lP$. (Bottom) Filtered Spectrum magnitude $|c_k \tilde{\chi}_P(k/P)|$ (log scale), showing only the aliased spectral components ($k=lP$, $l \neq 0$) that survive the filtering process. The sum of these filtered components (correctly accounting for phase, not just magnitude) constitutes the bias $B_P[f] = \sum_{k \neq 0} c_k \tilde{\chi}_P(k/P) = \sum_{l \neq 0} c_{lP}$.

Theorems & Definitions (6)

• Theorem 3.1: Deterministic Bias for $\mathbf{\sin^2}$
• proof
• Remark
• Definition 4.1: Resonance Functions
• Proposition 4.2: Properties of $\mathbf{\tilde{\chi}_P}$ and $\mathbf{\chi_P}$
• Theorem 5.1: General RBF Bias Formula