HOSC: Hyperbolic Oscillator with Saturation Control

• HOSC is a mechanism that combines periodic oscillatory dynamics with saturation control, applicable to both neural activations and hyperbolic PDE systems.
• It employs a tunable tanh-sin structure to decouple frequency support from gradient magnitude, enabling precise control in INR tasks and robust controller design.
• Empirical studies demonstrate its effectiveness in improving image, audio, video, and NeRF performance while ensuring bounded control actions through rigorous stability analyses.

The Hyperbolic Oscillator with Saturation Control (HOSC) refers to a class of mechanisms and mathematical structures that combine periodic oscillatory dynamics and explicit saturation control, as exemplified by the neural activation HOSC$(x)$ and control-theoretic approaches to hyperbolic PDE systems under input saturation constraints. HOSC activations serve to decouple frequency support and gradient scale, enabling both high-frequency representation and precise gradient magnitude control in implicit neural representations (INRs) (Wlodarczyk et al., 10 Jan 2026). In saturated control design for hyperbolic systems, similar principles are applied to govern boundary feedback controllers subjected to actuator saturation (Shreim et al., 2022). This article covers HOSC in both neural activation (INR) and control-theoretic contexts.

1. Mathematical Definition and Intuitive Structure

In neural network applications, HOSC is defined as

$$\mathrm{HOSC}_{\beta,\,\omega_0}(x) = \tanh\bigl(\beta\,\sin(\omega_0 x)\bigr),$$

where $\omega_0$ denotes the oscillatory period (carrier frequency) and $\beta$ the saturation parameter controlling the steepness and gradient magnitude (Wlodarczyk et al., 10 Jan 2026). Low $\beta$ values yield weak saturation, approximating a scaled sine, while large $\beta$ lead the activation towards square-wave behavior with strongly localized gradients. The tanh envelope bounds the output for numerical stability and directs gradients near zero-crossings of the carrier.

In control-theoretic settings, HOSC principles are realized via controller architectures for 1D hyperbolic PDE systems subject to actuator saturation, where the control input is subjected to a symmetric saturation $\sigma(u)$, defined element-wise as

$$\sigma_i(u) = \min\{\lvert u_i\rvert, \bar u_i\}\,\mathrm{sign}(u_i),$$

with prescribed bounds $\bar u_i > 0$ (Shreim et al., 2022). Saturation functions guarantee bounded control actions and induce nonlinear dead-zone characteristics.

2. Theoretical Properties: Gradient and Lipschitz Analysis

For HOSC activations, the derivative with respect to input $x$ is given by

$$\frac{d}{dx}\,\mathrm{HOSC}(x) = \beta\,\omega_0\,\cos(\omega_0 x) \cdot \mathrm{sech}^2(\beta\,\sin(\omega_0 x)),$$

where $\mathrm{sech}^2(u)=1-\tanh^2(u)$ (Wlodarczyk et al., 10 Jan 2026). The global Lipschitz constant is exactly $\beta\omega_0$, allowing explicit tuning of the gradient scale independently from the carrier frequency. At sine zero-crossings, the gradient is maximized and matches this upper bound. Network-level gradient bounds for typical MLP layers accumulate multiplicatively, with exponential suppression enabled by decreasing $\beta$.

In control scenarios, sector and Lipschitz bounds for the saturation nonlinearity $\phi(u) = \sigma(u) - u$ are exploited, with $\phi$ globally Lipschitz (constant 1) and satisfying sector inequalities:

$$\phi(u)^\top T (\phi(u) + u) \le 0, \quad \forall u.$$

These properties facilitate stability proofs and boundedness of solutions even under saturation (Shreim et al., 2022).

3. Empirical Performance in INR Tasks

Comprehensive benchmarking of HOSC versus SIREN, FINER, Gaussian, WIRE, and PEMLP activations covers canonical INR domains: image fitting, audio synthesis, video modeling, NeRFs, and SDFs (Wlodarczyk et al., 10 Jan 2026). HOSC demonstrates the following domain-specific behaviors:

• Image fitting (DIV2K, $512\times512$): PSNR increases monotonically with $\beta$, peaking at $40.22\,\mathrm{dB}$ for $\beta\approx14$. HOSC performs on par with FINER and outperforms SIREN, WIRE, and Gaussian activations in high-fidelity texture reconstruction.
• Audio fitting ($44.1$kHz): Best MSE achieved at moderate $\beta$, with HOSC yielding lowest error ($1.32\times10^{-6}$ for Bach), approximately $4\times$ better than SIREN.
• Video ($512\times512\times300$ frames): Optimal $\beta \approx 0.5$, with HOSC halving motion-blur artifacts relative to SIREN and yielding $2\times$ lower MSE.
• NeRFs: HOSC achieves PSNR within $0.1$–$0.2\,\mathrm{dB}$ of baseline methods. The explicit gradient-scale knob enables safe tuning without architectural modifications.
• SDFs: All periodic activations perform equivalently in low-chamfer distance, but vastly outperform Gaussian activations.

A plausible implication is that HOSC’s capacity for sharp gradient localization is most beneficial for low-dimensional, high-frequency regimes (audio, images), while modest $\beta$ values are preferable in higher-dimensional tasks for stability (Wlodarczyk et al., 10 Jan 2026).

4. Hyperparameter Selection and Regime Characterization

Empirical guidance suggests distinct $\beta$ ranges for optimal performance:

• Low-dimensional, high-frequency domains: $\beta \in [4,16]$ (audio, images) yields sharp transitions and localized gradients.
• Higher-dimensional coordinate domains: $\beta \in [0.3,1.0]$ (video, NeRF, SDF) maintains smooth saturation and mitigates gradient starvation.

Standard protocol is to fix $\omega_0$ as in SIREN ($\omega_0=30$ for images/audio, $\omega_0=1$ for geometry), sweep $\beta$ logarithmically, and select per-modality optimal values based on validation loss. A plausible implication is that tuning $\beta$ for each modality suffices across datasets and initializations, streamlining practical deployment (Wlodarczyk et al., 10 Jan 2026).

5. Control Synthesis for Hyperbolic PDEs with Saturation

Design of saturated boundary control applies HOSC principles in linear hyperbolic PDE contexts. For an $n$-dimensional state $X(t,z)$ evolving under

$$\frac{\partial}{\partial t}X(t,z) + \Lambda \frac{\partial}{\partial z}X(t,z) = N d(t,z),$$

boundary feedback at $z=0$ is imposed via saturated input, resulting in nonlinear boundary conditions (Shreim et al., 2022):

$$X(t,0) = H X(t,1) + B \sigma(u(t)), \qquad u(t) = KX(t,1).$$

Lyapunov functionals and LMIs are employed for synthesis:

• Functional: $V(X)=\int_0^1 e^{-\mu z} X(z)^\top P X(z) dz$, with bounds relating to exponential decay rates.
• LMI criteria: Sufficient conditions for stability and input-to-state stability (ISS) are formulated over feasible sets of controller gains $K$, diagonal matrices $P,T$, and scalars $\mu,\alpha$.

Numerical validations show exponential decay in the state, actuator efforts within prescribed saturation levels, and improved disturbance rejection (Shreim et al., 2022). These results establish quantitative guarantees for closed-loop systems with saturated inputs in the presence of disturbances.

6. Practical Considerations, Cost, and Limitations

HOSC activations introduce negligible computational overhead (typically $<5\%$ relative to SIREN) due to a single additional tanh and sin operation per activation (Wlodarczyk et al., 10 Jan 2026). Bounded outputs enhance training stability, particularly preventing divergence in high-gradient regimes. SIREN-style initialization is compatible.

For control-theoretic scenarios, the saturation leaves the system well-posed with globally unique mild solutions under general disturbances. LMI-based controller synthesis supports convex programming-based optimization.

Limitations include gradient starvation in smooth regions for large $\beta$ in neural applications and the lack of positional-encoding bias mitigation. In control applications, the sector bounds for saturation nonlinearity restrict the set of stabilizable systems.

7. Summary and Domain Impact

HOSC achieves a structured and tunable interface between periodic representation and gradient localization, unlocking high-fidelity learning and robust control under nonlinear saturation constraints. In neural networks, it either matches or outperforms state-of-the-art periodic activations across a diverse spectrum of INR tasks, and in control-theoretic settings, supports rigorous stability and synthesis methodologies for saturated hyperbolic systems. The explicit separation of carrier and gradient scale establishes HOSC as an analytically tractable and practically robust tool in both computational learning and mathematical control (Wlodarczyk et al., 10 Jan 2026, Shreim et al., 2022).