Shaped Sinusoids & Amplitude Scaling

Updated 18 January 2026

Shaped sinusoids are engineered waveforms with tunable amplitude, frequency, phase, and envelope that enable precise signal synthesis, decomposition, and analysis.
Amplitude scaling employs systematic methods like variance normalization and Parseval tightness to ensure energy consistency and fulfill geometric constraints.
Joint shaping and scaling techniques enhance applications in wireless channel modeling, sparse frequency analysis, neural networks, and geometric visualization by improving performance and accuracy.

Shaped sinusoids are mathematical entities whose parameters—amplitude, frequency, phase, and envelope—are manipulated to control their spectral and temporal properties for use in signal processing, time-frequency analysis, channel modeling, neural representations, and geometric design. Amplitude scaling refers to explicit parameterization or normalization schemes that ensure desired variance, Parseval tightness, or geometric constraints within these shaped sinusoidal constructs. Joint consideration of shaping and amplitude scaling enables the precision synthesis, decomposition, and manipulation of signals and multidimensional random fields for both theoretical analysis and practical engineering tasks.

1. Mathematical Formulations of Shaped Sinusoids

Shaped sinusoids appear in several forms:

Sum-of-Sinusoids (SoS) model: A spatially correlated Gaussian process can be represented as $k̂(x) = \sum_{n=1}^N a_n \cos(2\pi f_n x + \psi_n)$ , extended to higher-dimensional cases for spatial channel modeling. The amplitude $a_n$ , frequency $f_n$ , and phase $\psi_n\sim\mathrm{Uniform}[-\pi, \pi]$ are tuned to match desired autocorrelation functions, and for each spatial dimension, direction cosines encode angular orientation (Jaeckel et al., 2018).
Piecewise-constant Amplitude/Phase Sinusoids: In variational sparse frequency analysis, the modeled signal is $x(n) = \sum_{k=0}^{K-1} A_k(n)\cos(2\pi f_k n + \phi_k(n))$ with $A_k(n)$ and $\phi_k(n)$ approximated as piecewise-constant, regularized via total variation penalties (Ding et al., 2013).
Generalized Oscillations ("Wave-shape" Functions): The adaptive non-harmonic model writes $x(t) = \sum_{k=1}^K A_k(t) s_k(\phi_k(t))$ , where $s_k$ are 1-periodic smooth wave-shape functions, allowing for general periodic templates beyond pure sinusoids (Lin et al., 2016).
Enveloped Sinusoid Parseval Frames: Atoms are of the form $f_{\ell,k,m}[n] = e_\ell[n-m]\,\exp(2\pi j k(n-m)/N)$ , where $e_\ell[n]=\alpha_\ell g_\ell[n]$ is a scaled envelope (e.g., Gaussian, Hann, exponential) with normalization to preserve the Parseval frame property (Goehle et al., 2022).
Sinusoidal Neural Networks and Harmonic Expansions: In SIREN models, each neuron computes $h_i(x) = \sin(\sum_j a_{ij} \sin(\omega_j x + \varphi_j) + b_i)$ , yielding a harmonic sum expansion with amplitudes determined analytically via products of Bessel functions—these can be shaped and scaled to target specific spectral properties (Novello, 2022).
Geometric Sinusoids for Visualization: For VennFan diagrams, boundaries are constructed using $r_i(\theta) = 1 + \lambda(i)\,\mathrm{sgn}(\sin(2^i\theta))|\sin(2^i\theta)|^p$ for sine-based or $r_i(\theta) = 1 + \lambda(i)\,\mathrm{sgn}(\cos(2^{i-1}\theta))|\cos(2^{i-1}\theta)|^p$ for cosine-based fans, with amplitude scaling $\lambda(i)$ ensuring geometric nesting (Csanády, 11 Jan 2026).

2. Amplitude Scaling: Principles and Normalization

Shaped sinusoids almost always require systematically chosen amplitude scalings to satisfy variance, energy, or frame-tightness constraints:

Variance normalization in SoS: $a_n^2 = 2/N$ ensures unit variance in Gaussian process simulations, decoupling amplitude design from the autocorrelation function shaping. For six-dimensional (dual-mobility) processes, a global $1/\sqrt{2}$ factor preserves variance when summing independent subfields (Jaeckel et al., 2018).
Parseval tightness in ESP frames: Envelopes $e_\ell[n]=\alpha_\ell g_\ell[n]$ must satisfy $\|e_\ell\|_2 = 1/\sqrt{NL}$ for all $\ell$ to ensure the overall tight frame constant is 1. If additional “gains” $\gamma_\ell$ are used, the total frame bound becomes $\alpha = \sum_\ell \gamma_\ell^2$ (Goehle et al., 2022).
Amplitude parameters for geometric shaping: In VennFan, choices for $\lambda(i)$ (linear, exponential, or custom with modifiers $\delta$ and $\epsilon$ ) govern how fan blades taper and avoid geometric overlap, while the exponent $p\in(0,1]$ determines the wave "fatness." Proper amplitude selection ensures nestability and legibility without self-intersection (Csanády, 11 Jan 2026).
Control in neural models: In SIREN expansions, the amplitude of each frequency component is governed by products of Bessel functions $\prod_j J_{k_j}(a_{ij})$ ; amplitude decay is controlled by selecting the $a_{ij}$ weights. Tight analytic bounds enable truncation and error control in approximations, linking amplitude scaling directly to neural network expressivity and “spectral bias” effects (Novello, 2022).

3. Methods for Constructing and Optimizing Shaped Sinusoids

Effective use of shaped sinusoids hinges on practical algorithms for their construction:

Iterative frequency and amplitude optimization: In SoS models, frequencies and directions are initialized randomly and iteratively refined (by nonlinear least-squares in circular “cuts”) to minimize the average squared error (ASE) against a sampled target ACF. Uniform amplitude assignment decouples the frequency optimization from variance control (Jaeckel et al., 2018).
Convex variational strategies: For signals with jump discontinuities (abrupt amplitude or phase changes), sparse frequency analysis leverages convex optimization with total variation (TV) and $\ell_1$ spectral sparsity regularization. Alternating Direction Method of Multipliers (ADMM) enables efficient decoupled updates and convergence guarantees (Ding et al., 2013).
Envelope and gain design: ESP frames select envelope shapes $g_\ell[n]$ (choice depends on desired time-frequency resolution) and tuning factors $\alpha_\ell$ . The envelope’s width determines localization properties, while $\gamma_\ell$ adjusts energy allocation per band (Goehle et al., 2022).
Closed-form harmonic expansions: The expansion of neurons in sinusoidal neural networks leads to exact analytical control of harmonic content and amplitude, with explicit rules for sorting and truncating the expansion (Novello, 2022).
Geometric parameterization in visualization: Explicit amplitude scaling functions $\lambda(i)$ , combined with nonlinear exponents $p$ , support algorithmic generation of non-intersecting, optimally spaced VennFan diagram boundaries. This enables systematic exploration of visual “shape space” via parameter sweeps (Csanády, 11 Jan 2026).

4. Applications in Signal Processing, Modeling, and Representation

Shaped sinusoids and amplitude scaling are deployed across a variety of domains:

Wireless channel modeling: The SoS approach is the core of spatially consistent 3GPP New-Radio channel simulation; it supports extensibility up to six spatial dimensions for device-to-device and dual-mobility cases. The method yields accuracy improvements of 4–6.5 dB in ACF matching compared with prior filtering strategies, supporting fine-grained spatial correlation control (Jaeckel et al., 2018).
Sparse decomposition and filtering: Sparse frequency analysis enables “shaped” component extraction (even with amplitude or phase discontinuities) and outperforms linear band-pass filters in jump preservation, with direct applications to EEG analysis and noisy waveform extraction (Ding et al., 2013).
Time–frequency and adaptive harmonic analysis: In adaptive non-harmonic waveform contexts (ECG, musical signals, bioacoustics), de-shape SST recovers instantaneous amplitudes, periods, and separates fundamental frequencies, robust to complex wave-shapes and amplitude modulations (Lin et al., 2016).
Overcomplete frame representation: ESP frames, parametrized by shaped envelopes and properly normalized scaling, deliver energy-preserving, sparse, and morphologically interpretable signal decompositions—demonstrating superior denoising and sparse reconstruction performance over short-time Fourier or Prony methods (Goehle et al., 2022).
Neural networks with spectral control: In SIRENs, closed-form shaping and scaling of sinusoidal neurons allows the construction of neural fields with precisely prescribed harmonic envelopes and amplitude spectra, matching signal priors and supporting applications in graphics, radiance fields, and signed distance representation (Novello, 2022).
Geometric constructions: In data visualization, shaped sinusoidal boundaries (with tuned scaling) generate combinatorially complete, aesthetically flexible, and label-optimized $n$ -set Venn diagrams for large $n$ , a feat unattainable by classical constructions (Csanády, 11 Jan 2026).

5. Accuracy, Efficiency, and Trade-offs

Shaped sinusoids provide compelling accuracy and computational properties, contingent on careful amplitude scaling:

Memory and speed: SoS models scale as $O(N)$ in memory and per-sample evaluation, compared to the $O(N^D)$ prohibitive costs of grid-filtered approaches, and support arbitrary sampling locations. All coefficients can be precomputed for reuse (Jaeckel et al., 2018).
Error control: Accuracy (as measured by ASE) improves by approximately 3 dB per doubling of $N$ ; Shape and amplitude selections are analytically decoupled so one can separately target statistical fidelity and computational tractability (Jaeckel et al., 2018).
Sparse expansions: ESP frames and group-sparse decompositions yield extreme coefficient sparsity and order-of-magnitude denoising improvements, provided that normalization is strictly enforced (Goehle et al., 2022, Ding et al., 2013).
Spectral bias and truncation: Analytic amplitude bounds in SIRENs ensure that truncation beyond a prescribed $B$ yields bounded error; low-frequency components dominate initially (spectral bias), and higher harmonics are incorporated as optimization progresses or as width increases (Novello, 2022).
Time–frequency tradeoffs: Envelope width and amplitude scaling in ESP frames directly navigate the classical time-frequency localization dilemma—wider envelopes provide sharp frequency selectivity at the expense of time resolution, and vice versa (Goehle et al., 2022).
Parameter sweep for area-fidelity in geometric applications: VennFan shows that tailored amplitudes and exponents can eliminate vanishing-region and label-overlap issues, with direct quantitative improvements in normalized region area (Csanády, 11 Jan 2026).

6. Theoretical Guarantees and Identifiability

Autocorrelation control: In SoS models, frequencies and directions can be iteratively refined such that the empirical autocorrelation of the shaped process converges to the target over both test directions and distances, as measured by average squared error (Jaeckel et al., 2018).
Energy preservation: Parseval normalization in ESP frames ensures exact preservation of signal norm through the analysis–synthesis pipeline, regardless of the envelope shape (Goehle et al., 2022).
Identifiability in time–frequency extraction: The de-shape SST offers provable separation of instantaneous frequencies, amplitude scalings, and periodic templates, with explicit error bounds under “slow variation” and spectral bandwidth assumptions (Lin et al., 2016).
Amplitude bounding and error control in harmonic expansions: Explicit estimates for the decay of Bessel-function products in SIREN models ensure that amplitude shaping is mathematically controllable and that truncation yields predictable uniform errors, supporting the formal design of approximators with specified spectral content (Novello, 2022).

7. Extensions and Customizations

Marginal transformation: After generating shaped, amplitude-scaled Gaussian random fields, nonlinear transformations (e.g., $u=0.5\operatorname{erfc}(-k̂/\sqrt{2})$ ) may be used to remap to arbitrary marginal distributions for simulation tasks (Jaeckel et al., 2018).
Custom envelope design: In ESP frames, envelopes $g_\ell$ and scaling $\gamma_\ell$ can be tailored to encode domain-specific prior information or application constraints (physical resonances, damping patterns, etc.) (Goehle et al., 2022).
Frequency and shape selection in neural networks: Initialization and architecture selection in SIREN models may leverage domain priors by aligning input frequencies to expected bands and shaping amplitude profiles to match empirical spectral features (Novello, 2022).
Flexible geometric shaping for visualization: The amplitude scaling sequence $\lambda(i)$ and exponent $p$ in shaped sinusoids provide a fine-grained “shape palette” for constructing Venn diagrams adapted to label- and region-area constraints (Csanády, 11 Jan 2026).

These frameworks demonstrate that the synthesis and representation of signals and random fields as shaped sinusoids, with rigorous amplitude scaling, underpins both foundational theoretical results and leading-edge applications in statistical modeling, procedural geometry, neural representation, denoising, and feature extraction. The analytic separation of amplitude shaping from envelope selection and phase/frequency control is critical for engineering tunable, efficient, and provably accurate systems across these domains (Jaeckel et al., 2018, Ding et al., 2013, Goehle et al., 2022, Novello, 2022, Lin et al., 2016, Csanády, 11 Jan 2026).