Fourier Ratio in Harmonic Analysis

Updated 31 January 2026

Fourier Ratio is defined as the ℓ1 to ℓ2 norm ratio of a function's Fourier coefficients, measuring its compressibility and structural complexity.
It underpins key methods in compressed sensing and signal processing, enabling polynomial approximations and recovery from minimal measurements.
A low Fourier Ratio indicates effective sparsity and efficient learning, while a high ratio reflects spectral delocalization and challenges in recovery.

The Fourier ratio is a scalar parameter quantifying the interplay between the structural, analytic, and algorithmic complexity of a function or measure, as manifested through its representation in a fixed orthonormal basis—most notably, the Fourier basis. It appears in diverse contexts, including compressed sensing, uncertainty principles, additive combinatorics, signal processing, harmonic analysis, and high-dimensional probability. The canonical definition on a discrete domain is the ratio of the $\ell^1$ to $\ell^2$ norm of the system's expansion coefficients, generalizing naturally to continuous settings via regularized $L^1/L^2$ norms of the Fourier transform. This quantity encapsulates effective dimension: in sparse or highly compressive regimes, small Fourier ratio signals admit low-dimensional polynomial approximations and recovery from minimal measurements; conversely, large ratios reflect spectral delocalization and obstructions to efficient recovery or learning.

1. Formal Definitions and Variants

Let $f$ be a function in $\mathbb{C}^D$ with respect to an orthonormal basis $\{\phi_j\}_{j\in D}$ (e.g., the standard Fourier basis on $G$ ). The Fourier ratio in basis $\mathcal{B}$ is

$FR_{\mathcal{B}}(f) = \frac{\|c_{\mathcal{B}}(f)\|_1}{\|c_{\mathcal{B}}(f)\|_2} = \frac{\sum_{j\in D}|\langle f,\phi_j\rangle|}{\Bigl(\sum_{j\in D}|\langle f,\phi_j\rangle|^2\Bigr)^{1/2}},$

where $c_{\mathcal{B}}(f)$ is the vector of expansion coefficients. For discrete Fourier analysis on $\mathbb{Z}_N$ or related groups, this reduces to

$FR(f) = \frac{\|\widehat{f}\|_1}{\|\widehat{f}\|_2},$

where $\widehat{f}$ is the normalized discrete Fourier transform of $f$ .

In the continuous setting, for compactly supported Borel measures $\mu$ on $\mathbb{R}^d$ and $f\in L^2(\mu)$ , the scale-dependent Fourier ratio is defined using regularized norms: $F_{\mu,R}(f) := \frac{X_{1,\mu,R}(f)}{X_{2,\mu,R}(f)},$ where $X_{p,\mu,R}(f)$ are scale- $R$ Fourier norms (regularized via mollification) (Iosevich et al., 18 Dec 2025).

Related quantities include the $\ell_4/\ell_2$ ratio for restricted Fourier supports, $\mu(A) = \max_{f:\supp(\hat f)\subset A}\|f\|_4^4/\|f\|_2^4$, directly controlling uncertainty lower bounds and hypercontractivity for the Boolean cube (Kirshner et al., 2018). Other functionals, such as the bi-Fourier ratio $\min\{FR(f), FR(\widehat f)\}$ , capture two-sided complexity (Aldaleh et al., 24 Nov 2025).

2. Effective Dimension, Uncertainty, and Structure

The Fourier ratio robustly tracks the "effective sparsity" or compressibility of a function in its chosen basis, governing both analytic and combinatorial properties. For a $k$ -sparse vector,

$FR(f) \leq \sqrt{k},$

so $FR(f)^2$ quantifies effective dimension. Functions with polynomially-decaying or power-law coefficient tails have $FR(f)$ growing logarithmically in dimension, distinct from the quadratic scaling induced by truly flat (random) coefficients (Burstein et al., 22 Jan 2026, Aldaleh et al., 24 Nov 2025).

Uncertainty principles tightly constrain the joint localization of a function and its spectrum: for $f$ supported on $E$ with $\widehat{f}$ concentrated on $S$ ,

$(1-a)^2\,\frac{N}{|E|} \leq FR(f)^2 \leq \frac{|S|}{(1-b)^2}$

where $a$ , $b$ measure non-concentration (Aldaleh et al., 24 Nov 2025). The classical uncertainty relation $|E|\,|S|\geq N$ follows as a corollary. Variants such as the stability of the uncertainty principle—quantifying proximity to extremizers (affine subspaces)—are governed by the maximal $\ell_4/\ell_2$ ratio over spectral support (Kirshner et al., 2018).

The Fourier ratio also encodes additive and algebraic structure. For instance, sets $A$ with large additive energy $E_2(A,A)$ can attain higher $\mu(A)$ in the discrete cube, with explicit entropy-like formulas on Hamming spheres (Kirshner et al., 2018). Conversely, generic sparse supports force large $FR(f)$ unless the signal is highly regular (Aldaleh et al., 24 Nov 2025).

3. Recovery, Approximation, and Learning Complexity

Functions with small Fourier ratio admit strong algorithmic guarantees across recovery, approximation, and learning. The scalar bound $FR(f)\le r$ (in any bounded ONB) ensures recovery from

$m \geq C \frac{r^2}{\epsilon^2}[\log(r/\epsilon)]^2\log|G|$

random samples via $\ell^1$ -minimization, with relative error $O(\epsilon)$ (Burstein et al., 22 Jan 2026, Iosevich et al., 24 Jan 2026). This is independent of sparse support assumptions and applies equally to signals with smoothness-induced spectral compressibility (e.g., $C^2$ fields on $[0,1]^2$ have $FR=O(\log N)$ ) (Iosevich et al., 24 Jan 2026).

Fourier ratio provides explicit polynomial-degree bounds for $L^2$ , $L^1$ , and $L^\infty$ approximation: if $FR(f)\le r$ , then $k=O(r^2/\varepsilon^2)$ -degree trigonometric polynomials suffice for $\varepsilon$ -accurate $L^2$ approximations (Aldaleh et al., 24 Nov 2025, Iosevich et al., 18 Dec 2025). Algorithmic rate-distortion—the minimal program length required to approximate $f$ to accuracy $\epsilon$ —is bounded above by the same function of $r$ (Aldaleh et al., 24 Nov 2025, Burstein et al., 22 Jan 2026).

In the statistical learning framework, the VC dimension and SQ (statistical query) dimension of Boolean classes $\{f:\ FR(f)\le r\}$ are at most $\exp(Cr^2\log r)$ , yielding efficient PAC and SQ learning algorithms (Aldaleh et al., 24 Nov 2025, Burstein et al., 22 Jan 2026).

4. Geometric and Harmonic Analysis Perspectives

The continuous Fourier ratio $F_{\mu,R}(f)$ systematically links spatial and frequency-domain geometry. Sharp "fractal uncertainty principles" assert that

$\sqrt{\frac{1}{R^d |E_f^{R^{-1}}|}} \leq F_{\mu,R}(f) \leq \sqrt{\frac{|X_R|}{R^d(1-\eta)^2}}$

where $E_f^{R^{-1}}$ and $X_R$ are neighborhoods of the spatial and frequency supports. These inequalities yield recovery theorems for compactly supported measures given incomplete frequency data, as well as precise bounds on the degree of polynomial approximation in $L^p$ norms (Iosevich et al., 18 Dec 2025). Crucially, the Fourier ratio, and its decay or stalling under geometric restrictions, discriminates curved surface measures (e.g., $S^1$ ) from random fractal measures (e.g., random Cantor sets), reflecting their harmonic analytic complexity.

On convex bodies, the dimension of the normal cone $N(K)$ determines the $L^2$ -approximation degree (via stationary phase bounds and the decay of $\widehat{\mu}$ ), reinforcing the geometric control that Fourier ratio provides over functional complexity (Iosevich et al., 18 Dec 2025).

5. Localization, Restriction, and Additive Combinatorics

Localization phenomena place lower bounds on the Fourier ratio of "slices" in product spaces: if $f$ is defined on $G=H\oplus K$ , the maximal slicewise Fourier ratio over $H$ satisfies

$\max_{k\in K} FR_H(f_k) \geq \frac{FR_G(f)}{\sqrt{|K|}}$

so effective global compressibility cannot be distributed locally without loss (Burstein et al., 22 Jan 2026). This manifests sharply in multi-dimensional compressed sensing or function approximation, where attempts at coordinate-wise decomposition incur unavoidable losses in sample complexity.

For Fourier supports constrained to Hamming spheres or balls, the $\ell_4/\ell_2$ ratio $\mu(S(n,k))$ , dictated by combinatorial structure, controls hypercontractive constants, stability in uncertainty principles, and the exact growth of additive energy, with formulas involving binomial coefficients and rate-type entropy functions (Kirshner et al., 2018).

6. Algorithmic and Statistical Implications

A low Fourier ratio ensures small algorithmic rate-distortion complexity—robust approximability by low-degree models and short program descriptions. In typical signal-processing or high-dimensional data applications, this translates into efficient lossy compression, denoising, and imputation from incomplete measurements with theoretical sample-complexity guarantees (Aldaleh et al., 24 Nov 2025, Burstein et al., 22 Jan 2026).

Stable recovery results hold even in the presence of stochastic subsampling, random restrictions, or additive Gaussian noise, as perturbations to FR are quantitatively controlled (Aldaleh et al., 24 Nov 2025).

Classes with bounded FR also enjoy favorable statistical properties: learning from random examples, VC/SQ dimensions, and resilience to loss are all governed by FR rather than ambient dimension. This unifies modern compressed sensing, learning theory, and information-theoretic analysis under a single complexity parameter (Aldaleh et al., 24 Nov 2025, Burstein et al., 22 Jan 2026).

7. Specialized and Computational Constructions

The terminology "Fourier ratio" also arises in specialized algorithmic contexts, notably rational approximations of the Fourier transform by ratios of two polynomials, $R(\omega) = P(\omega)/Q(\omega)$ , constructed to approximate $F\{f\}(\omega)$ to high accuracy. This approach, based on sampling and incomplete-cosine expansions, enables efficient implementation of spectral methods and signal processing algorithms, providing a rational-function form suitable for symbolic or fast numerical computation (Abrarov et al., 2020).

In cosmology and statistical physics, the ratio of two Fourier-transformed random fields, such as the quotient $\delta(\mathbf{k})/\psi(\mathbf{k})$ for density and momentum divergence fields, has a mathematically precise one-point distribution, derivable in closed-form for complex Gaussian fields and tested against N-body simulations (Li et al., 2021). This statistical Fourier ratio provides direct access to growth rates and stochastic properties of cosmic structures.