Curvature Regularizer: Concepts & Applications

• Curvature Regularizer is a constraint that penalizes abrupt trajectory changes, promoting linear, predictable paths in mathematical models and machine learning.
• It is mathematically formulated using cosine similarity between successive vectors in discrete settings or second derivatives in continuous models, linking variational principles with neural optimization.
• Applied across geometric flows, generative models, and signal processing, curvature regularizers enhance stability, robustness, and the accuracy of predictive tasks.

A curvature regularizer is a constraint or penalty term introduced in mathematical modeling, geometric flows, machine learning, or representation learning to explicitly reduce the curvature of trajectories, curves, or data representations—typically to promote easier prediction, analytic tractability, or improved robustness. In modern applications, the curvature regularizer is closely associated with the principle of temporal or trajectory straightening: penalizing changes in the direction (second-order variation) of a trajectory through latent, geometric, or physical spaces to encourage locally straight (low-curvature) paths.

1. Mathematical Formulation of Curvature Regularizers

In the most general setting, a curvature regularizer evaluates the deviation of a curve or trajectory from straightness, usually by penalizing rapid angular changes (high curvature) between consecutive steps. For a sequence of points $\{z_t\}$ in a high-dimensional space (e.g., neural network latent space), the discrete curvature at time $t$ is typically defined by the cosine similarity between consecutive difference vectors: $$\mathcal{C}_t = \frac{(z_{t+1} - z_t) \cdot (z_{t+2} - z_{t+1})}{\|z_{t+1}-z_t\|\, \|z_{t+2}-z_{t+1}\|},$$ where $\mathcal{C}_t = 1$ for perfectly straight (collinear) steps and decreases as the trajectory curves. A curvature regularizer is then implemented as the loss

$\mathcal{L}_{\text{curv}} = 1 - \mathcal{C}_t$

summed over all valid $t$ in the sequence. Continuous-time analogs use second derivatives to define curvature, but this discrete form is common in neural sequence modeling and planning (Wang et al., 12 Mar 2026, Niu et al., 2024, Toosi et al., 2023).

In geometric flows and PDEs, curvature penalties often take the form $\int |\kappa(s)|^2 ds$, where $\kappa$ is the geometric curvature along a curve $\gamma(s)$ (Novaga et al., 2013, Miura et al., 4 Apr 2025).

2. Historical Origins and Variational Foundations

Curvature regularizers have roots in the calculus of variations and geometric analysis. In curve evolution, the classic gradient flows of length (curve shortening) and of bending energy (curve straightening) are archetypes. The length functional $L[\gamma] = \int ds$ provides a first-order (tangent) regularizer, while the bending energy $t$0 is a canonical second-order (curvature) regularizer.

The direction energy $t$1, with $t$2 the unit tangent, interpolates between length and straightness, providing a unifying variational framework. The L$t$3-gradient flow of such energies governs “shortening-straightening” phenomena in PDEs and geometric evolution equations (Miura et al., 4 Apr 2025, Novaga et al., 2013).

3. Curvature Regularization in Representation Learning

Curvature regularizers have achieved significance in high-dimensional representation learning domains. The “temporal straightening” principle posits that in biological and robust artificial neural systems, representations of sequences (e.g., video or sensorimotor streams) should follow locally straight trajectories in feature space to facilitate prediction by linear extrapolation and improve robustness (Niu et al., 2024, Toosi et al., 2023).

In self-supervised learning for temporal data, the curvature penalty is typically formulated as the negative average cosine similarity of consecutive velocity vectors in embedding space: $t$4 where $t$5. This loss term is either optimized directly or used as an auxiliary regularizer to standard learning objectives (e.g., invariance or contrastive losses) (Niu et al., 2024).

Empirically, curvature regularization in neural representations leads to enhanced predictability, more reliable planning, improved robustness to adversarial attacks, and alignment between Euclidean and geodesic distances for goal-reaching tasks (Wang et al., 12 Mar 2026, Niu et al., 2024, Toosi et al., 2023).

4. Role in Geometric Flows and Mathematical Physics

In geometric flows, curvature regularizers underpin the evolution equations of curves and surfaces. For example, the “curve shortening-straightening flow” is the L$t$6-gradient flow for the energy functional $t$7, and yields the PDE

$t$8

for planar curves, where $t$9 is the normal vector (Novaga et al., 2013). Such fourth-order flows generalize to higher-order parabolic equations for elastic curves and surfaces, where asymptotic convergence to stationary solutions (straight lines or elasticae) can be rigorously proved under appropriate boundary or asymptotic conditions (Miura et al., 4 Apr 2025).

In time-parametrization-invariant models of mechanics and quantum gravity, curvature regularization arises implicitly as “temporal straightening”: the process of eliminating reparametrization redundancy to restore a physical time parameter. Gauge-fixing procedures (e.g., fixing the lapse function in the Jacobi variational principle) mathematically “straighten” the time coordinate (Cattaneo et al., 2016).

5. Applications in Machine Learning, Planning, and Signal Processing

Curvature regularization finds diverse applications:

• Latent world models and planning: Incorporating curvature regularizers in the representation space of world models aligns Euclidean and geodesic distances, stabilizes the Hessian of the planning loss, and dramatically increases success rates in goal-reaching and control tasks (Wang et al., 12 Mar 2026).
• Self-supervised sequence learning: Imposing curvature penalties during self-supervised representation learning on temporal data yields representations supporting robust prediction, interpretable latent factors, and resistance to adversarial perturbations (Niu et al., 2024).
• Generative models: In continuous-time flow and diffusion models, straightening the probability flow with piecewise-linear or curvature-penalizing objectives reduces truncation error, enabling accurate generation at extremely low numbers of function evaluations (Yoon et al., 2024).
• Video authenticity: Temporal curvature statistics of trajectories in neural representation space distinguish natural from AI-generated video, enabling efficient, highly accurate detection frameworks (Internò et al., 1 Jul 2025).
• Signal processing: In ultrafast photonic time-stretch applications, “temporal straightening” via removal of higher-order spectral-phase distortions achieves linear group delay, eliminating aberrations and maximizing spectral resolution (Chen et al., 2019).

6. Theoretical, Empirical, and Algorithmic Properties

Curvature regularizers are theoretically well-motivated:

• Consistency and convergence: In curve registration and geometric flows, curvature-regularized algorithms exhibit provable identifiability, consistency, and long-time convergence to stationary straight or elastica solutions under mild conditions (Novaga et al., 2013, Miura et al., 4 Apr 2025, Bhaumik et al., 2015).
• Optimization and conditioning: Reducing curvature directly improves the numerical stability of gradient-based optimization—lowering the condition number of Hessians in latent planning and reducing global truncation error in ODE-solved models (Wang et al., 12 Mar 2026, Yoon et al., 2024).
• Variance reduction: In deep generative modeling, regularizing on piecewise-linear segments (straightening) reduces gradient variance during flow matching, also yielding lower Lipschitz constants in learned neural fields (Yoon et al., 2024).

Empirically, the addition of curvature regularizers as auxiliary losses leads to:

• Improved predictiveness and robustness in vision and temporal representation models (Niu et al., 2024, Toosi et al., 2023).
• Accelerated and stabilized training and inference in generative and planning models (Yoon et al., 2024, Wang et al., 12 Mar 2026).

7. Extensions and Generalizations

Curvature regularization principles are extensible:

• Multi-domain adaptation: Steering hidden representations via mean-difference vectors (a first-moment analog of curvature guidance) addresses distribution shifts, including temporal, domain, and style drifts—often in a plug-and-play and compositionally extensible fashion (Shin et al., 24 Mar 2025).
• Functional data and signal alignment: Feature-sensitive kernel-based temporal straightening aligns two noisy functional datasets by estimating monotone warps that minimize pointwise misalignment under curvature-regularizing objectives (Bhaumik et al., 2015).
• Physical metrology: In relativistic timekeeping, “temporal straightening” of coordinate time via appropriate frame selection eliminates spurious periodicities and yields tractable, physically covariant expressions for clock synchronization in multi-body systems (Liu et al., 21 Jul 2025).

This diversity underscores the unifying geometric role of curvature regularization across analysis, modeling, inference, and control.