Power Spectrum Representation Methods

• Power Spectrum Representation is a method that decomposes a signal’s energy over frequency or wavenumber, linking variance with spatial and temporal correlations.
• It employs explicit stochastic integrals and tensor formulations to facilitate numerical field generation and analytical moment computations in both scalar and vector contexts.
• Applications span diverse fields—from cosmology and turbulence studies to wireless communications and power systems—with techniques like Chebyshev polynomial expansion and compressive sampling enhancing spectral recovery.

A power spectrum representation expresses the second-order statistics or energy distribution of a stochastic process, field, or deterministic signal as a function of frequency or wavenumber. In the context of random fields, time series, multidimensional signals, or physical processes, the power spectrum encapsulates how variance or energy is partitioned over the relevant spectral domain, and provides a rigorous link between spatial/temporal correlations and their harmonic decompositions. Modern formulations extend the power spectrum framework from scalar to vector and tensor-valued fields, accommodate nonuniform domains (e.g., angle or space), integrate with advanced numerical and statistical methods, and serve as the backbone of statistical inference in fields ranging from cosmology, condensed matter, and turbulence to signal processing and wireless communications.

1. Explicit Stochastic Spectral Representation for Random Fields

The canonical representation for a stationary random field $\rho(\mathbf{r})$ is via a stochastic integral over the power spectrum. For a zero-mean, homogeneous scalar field, the two-point correlation function is

$$C(\mathbf{r}) = \langle \rho(\mathbf{0})\,\rho(\mathbf{r})\rangle = \int_{\mathbb{R}^3} e^{i\,\mathbf{k}\cdot\mathbf{r}}\,F^2(\mathbf{k})\,d\mathbf{k}$$

where $F^2(\mathbf{k})$ is the (nonnegative) power spectrum (Chepurnov, 2021).

Chepurnov's explicit construction writes the field as a Lebesgue-type stochastic integral:

$$\rho(\mathbf{r}) = \int_{\mathbb{R}^3} e^{i\,\mathbf{k}\cdot\mathbf{r}}\,F(\mathbf{k})\,\xi(\mathbf{k})\,d\mathbf{k}$$

where $F(\mathbf{k}) = \sqrt{F^2(\mathbf{k})}$ and $\xi(\mathbf{k})$ is a complex Gaussian random field, satisfying

$$\langle \xi(\mathbf{k})\,\xi^*(\mathbf{k}')\rangle = \delta_{\mathbf{k}\mathbf{k}'}, \quad \xi(-\mathbf{k}) = \xi^*(\mathbf{k})$$

This explicit spectral measure element makes higher-order moment computations and numerical field generation tractable, and guarantees statistical convergence due to the random-phase cancellations, rather than requiring explicit spatial windowing for integral convergence.

For vector fields $u_i(\mathbf{r})$, the representation involves a spectral tensor (orthogonal projector) $T_{ij}(\mathbf{k})$, yielding

$$u_i(\mathbf{r}) = \int e^{i\,\mathbf{k}\cdot\mathbf{r}}\,F(\mathbf{k})\,T_{ij}(\mathbf{k})\,\xi_j(\mathbf{k})\,d\mathbf{k}$$

with $T_{ij}T_{jk} = T_{ik}$, $T_{ij} = T_{ji}$, and $$\langle \xi_i(\mathbf{k})\,\xi_j^*(\mathbf{k}')\rangle = \delta_{ij}\,\delta_{\mathbf{k}\mathbf{k}'}$$.

Physical mode decompositions (potential, solenoidal, Alfvénic) are realized by choosing the appropriate $T_{ij}(\mathbf{k})$ projector.

2. Power Spectrum in Channel Covariance and Angular Domains

In spatially distributed sensing (e.g., wireless communications), the covariance of received signals encodes information about the angular power spectrum (APS). One can formally relate the channel covariance matrix $R$ to the continuous APS $\rho(\theta)$ via

$$R = \int_{-\pi/2}^{\pi/2} \rho(\theta)\,a(\theta)a(\theta)^H\,d\theta$$

where $a(\theta)$ is the array steering vector.

The APS can be recovered from $R$ by casting the problem in a weighted Fourier domain, with

$$r_m = \int_{-1}^{1} g(x)\,e^{i\kappa_m x}\,w(x)\,dx$$

where $g(x) = \rho(\arcsin x)$, $w(x) = 1/\sqrt{1-x^2}$, and $\kappa_m$ are steering vector parameters. The feasible set of all $g$ matching the first $M$ lags is an affine subspace; the projection-onto-linear-variety (PLV) method selects the minimum-norm $g_{plv}$ in the trigonometric-polynomial subspace $N_\perp$:

$g_{plv}(x) = \sum_{k=-N}^N a_k e^{i\kappa_k x}$

with coefficients $a_k$ determined by a positive-definite linear system $Ga=y$ (Luo et al., 29 Dec 2025).

The approach achieves exact APS recovery if and only if the ground-truth APS lies in the spanned trigonometric-polynomial subspace; otherwise, PLV returns the unique minimum-energy spectrum consistent with the observed covariance.

3. Chebyshev Polynomial Frameworks for Power Spectrum Recovery

For enhanced flexibility and regularization, APS recovery can be formulated using Chebyshev polynomial expansions in a transformed variable $t = \cos\theta$. The spectrum is expanded as

$g(t) = \sum_{n=0}^\infty \alpha_n T_n(t)$

where $T_n$ are Chebyshev polynomials (orthogonal with respect to the weight $w(t) = (1-t^2)^{-1/2}$), and the covariance admits an exact series representation

$R = \sum_{n=0}^\infty \alpha_n M_n$

with $[M_n]_{p,q} = \pi\,i^n J_n(\kappa_{p-q})$, $J_n$ being Bessel functions (Luo et al., 30 Dec 2025).

Truncating to $N$ terms yields a finite-dimensional linear (and convex) regression problem, with uniform error governed by the tail of the Chebyshev expansion. A semidefinite characterization is available: $g_N(t) \geq 0$ can be enforced by representing $g_N$ as sums of squares of polynomials, producing a tractable LMI. Regularization via derivative-based $\ell_1$ penalties promotes smooth, cluster-preserving APS estimates.

This framework is particularly effective for downlink covariance prediction from uplink data in FDD systems, leveraging the invariance of the APS across frequencies.

4. Two-Dimensional and Compressive Power Spectrum Representations

In multivariate array or sensor network contexts, the power spectrum may be defined jointly over frequency and angle (or other domains). For multiple wide-sense-stationary (WSS) sources impinging on a uniform linear array, the 2D power spectrum $P(f, \theta)$ admits a reconstruction scheme under space-time compressive sampling (Ariananda et al., 2014):

• The compressed measurements are correlated in both temporal and spatial domains.
• The set of all lagged covariance entries is vectorized, and a linear system constructed with space-time mixing matrices.
• Least-squares recovery yields a stacked vector $\mathbf{p}_{vec}$ housing $P(f, \theta)$ samples, which can be reshaped to the full 2D spectrum.

This scheme enables recovery of the joint angular-frequency power spectrum with as few as $M_s \ll N_s$ active array elements, provided spatial and temporal selectors are chosen to ensure full rank. Peaks in $P(f, \theta)$ specify source directions (DOAs) at each frequency.

5. Power Spectrum Representations in Cosmology and Astrophysics

In cosmological context, the observed power spectrum embodies a mapping from theoretical 3D ensemble-averaged spectra to directly observable quantities such as angular multipoles, redshifts, and line-of-sight Fourier modes (Raccanelli et al., 2023). Statistical corrections for unequal-time correlators and projection effects produce observable 3D power spectrum forms as well as "frequency–angular power spectrum" observables, which avoid reliance on fiducial distance models and remain robust to so-called Alcock–Paczynski distortions.

In the analysis of Lyman-alpha forest data, the observed 1D flux power spectrum is a nonlocal projection of the underlying 3D matter power spectrum:

$$\Delta_{1D}(k) = k \int_k^\infty \frac{dq}{q^2} \Delta_{3D}(q)$$

with the 3D spectrum further modified by dark matter free-streaming and baryon pressure filtering, represented by exponential cutoffs in Fourier space (Ridkokasha et al., 20 May 2025).

6. Power Spectrum Representations Beyond Classical Fourier Analysis

Alternative representations extend the power spectrum to domains inaccessible to classical Fourier techniques. In nonuniformly sampled data (e.g., uneven photon arrival times), the Lomb periodogram computes a normalized periodogram for light curves reconstructed without binning, enabling full-frequency-resolution power spectrum estimation from sparse data (Tian et al., 2014).

Further, spectral representations of active and reactive power are constructed from analytic (Hilbert transform–based) signals:

• Instantaneous power $p(t)$ is decomposed into Hermitian (active) and complementary (non-Hermitian, reactive) components.
• The spectral (Fourier) representation yields power spectra $P_H(\omega)$, $P_C(\omega)$ by convolutions of the analytic voltage and current spectra (Schar et al., 2021).
• These spectra recover the classical power triangle at each frequency, generalizing physical interpretations to nonsinusoidal and aperiodic signals.

In power systems, Hilbert-transform analysis delivers instantaneous amplitude, phase, and frequency at each sample, yielding a full time-frequency "Hilbert spectrum" that supersedes the limitations of windowed Fourier transforms, especially for broadband and transient events (Derviškadić et al., 2019).

7. Power Spectrum Map Reconstruction and Learning

Beyond direct spectral estimation, the power spectrum can be reconstructed as a function over space–frequency domains from indirect, quantized, or compressed measurements. Support-vector machine (SVM)–type solvers, using vector-valued RKHS representations, enable the estimation of power spectrum density (PSD) maps from spatially distributed, quantized sensor data, incorporating physical priors and propagation models (Romero et al., 2016). Nonparametric and semiparametric formulations reduce to convex quadratic programs and admit efficient batch and online solvers, supporting practical deployment in large-scale sensor networks.

---

The developments in power spectrum representation span explicit stochastic integral forms, structural and geometric frameworks for statistical estimation, multidimensional and compressed harmonics, semidefinite and regularized recovery algorithms, and model-free signal processing strategies. These methodologies are foundational in theoretical and applied research across mathematics, physics, engineering, and data science, continually advancing the precision, interpretability, and computational feasibility of spectral analysis.