Force-Dual Modes: Subspace Design from Stochastic Forces

Otman Benchekroun, Eitan Grinspun, Maurizio M. Chiaramonte, Philip Allen EtterarXiv 2025 · DOI · arXiv

Abstract

Simulating high-resolution soft-body physics in real-time remains a significant challenge in computer graphics. Reduced Order Modeling (ROM) typically employs Linear Modal Analysis (LMA) to construct low-dimensional motion subspaces. However, LMA often suffers from "spooky action at a distance," where local interactions cause unnatural global deformations. We present Force-Dual Modes, a framework that builds simulation subspaces by modeling the statistics of expected forces. By propagating user-defined force distributions through a physical model, our method eliminates common artifacts and enables realistic, real-time control of complex soft bodies with massive speedups over full-order simulations.

Introduction: The Problem of Spooky Action at a Distance

Simulating squishy objects like a dragon's wing or a human bicep in real-time is computationally expensive. To achieve interactive rates, researchers use ROM to restrict motion to a small subspace of likely shapes. The standard approach, Linear Modal Analysis (LMA), assumes that forces are applied randomly and uniformly across the entire object.

In practice, interactions are usually local—such as pulling a handle or poking a surface. When an LMA subspace is used for local interaction, the entire body may wiggle or deform unnaturally. This artifact, known as "spooky action at a distance," occurs because the underlying mathematical basis does not account for the spatial distribution of input forces. Unrelated parts of a mesh move in a correlated way simply because the subspace lacks the resolution to decouple them.

Methods: A Statistical Perspective on Subspace Design

Force-Dual Modes introduces a statistical approach to subspace construction. Instead of assuming uniform noise, we allow users to define a force prior—a distribution representing where and how forces are likely to be applied.

  1. Force-to-Displacement Mapping: Input forces are modeled as a Gaussian distribution with a covariance matrix $\Sigma_F$. Using a linearized version of the physics (the system's Hessian matrix $H$), we propagate this force uncertainty to find the resulting displacement uncertainty: $\Sigma_U = H^{-1}\Sigma_F H^{-1}$.
  2. Generalized Eigenvalue Problem: We solve for the optimal low-rank basis that captures this displacement distribution.
  3. LMA as a Special Case: We demonstrate that standard LMA is a specific instance of our method where the force distribution is assumed to be spatially uncorrelated "white noise."
  4. Adaptive Subspaces: For complex multi-modal interactions, we utilize Gaussian mixtures. At runtime, the system uses Bayes' Rule to detect the active force distribution (e.g., touching the head vs. the wing) and adaptively switches the subspace to match the current interaction.
Comparison of full simulation, standard LMA, and Force-Dual LMA. Note how LMA causes the trunk to bend unnaturally ("spooky action at a distance"), while Force-Dual LMA maintains the correct shape.

Key Insights

  • Explicit Priors: LMA is not "prior-free"; it implicitly assumes an unrealistic global force distribution. Making this prior explicit allows for intentional subspace tailoring.
  • Physics-Aware Weighting: Instead of manual skinning weight painting, users can paint "force variances." The system then leverages material properties (stiffness, density) to determine natural deformations.
  • Statistical Duality: Designing in "force space" is more intuitive for users than "displacement space," as it directly maps to how we intend to interact with an object.

Results and Evaluation

The method was evaluated on high-resolution meshes, including a Dragon model (25k vertices, 103k tetrahedra) and a Pegasus model for adaptive switching.

  • Performance: The framework achieves staggering speedups. The Dragon model ran at 448 FPS, a 3,166x speedup over full-order simulation.
  • Visual Fidelity: Force-Dual LMA effectively eliminates the "spooky action" artifacts. As seen in the elephant example, local forces no longer cause distant, unrelated parts of the mesh to deform.
  • Real-Time Inverse Kinematics (IK): We demonstrate the first system capable of real-time IK on high-resolution soft characters with spring-like actuators, enabling natural, puppet-like control.
Real-time inverse kinematics control of a complex dragon model. Force-Dual LMA (right) eliminates the "gap" and "locking" artifacts seen in standard LMA (center).

Limitations and Future Work

The current theory relies on linearized dynamics. While Rotation Strain Coordinates are used to handle large rotations, the subspace construction does not yet account for extreme non-linearities like buckling or complex self-contact. Additionally, subspaces must be precomputed; significant changes in material properties during simulation would require recomputation. Future work will explore non-linear subspaces and applications in biomechanical simulations and fluid control.

Summary updated

Abstract

Designing subspaces for Reduced Order Modeling (ROM) is crucial for accelerating finite element simulations in graphics and engineering. Unfortunately, it's not always clear which subspace is optimal for arbitrary dynamic simulation. We propose to construct simulation subspaces from force distributions, allowing us to tailor such subspaces to common scene interactions involving constraint penalties, handles-based control, contact and musculoskeletal actuation. To achieve this we adopt a statistical perspective on Reduced Order Modelling, which allows us to push such user-designed force distributions through a linearized simulation to obtain a dual distribution on displacements. To construct our subspace, we then fit a low-rank Gaussian model to this displacement distribution, which we show generalizes Linear Modal Analysis subspaces for uncorrelated unit variance force distributions, as well as Green's Function subspaces for low rank force distributions. We show our framework allows for the construction of subspaces that are optimal both with respect to physical material properties, as well as arbitrary force distributions as observed in handle-based, contact, and musculoskeletal scene interactions.
BibTeX
@misc{benchekroun2025force,
  author = {Otman Benchekroun and Eitan Grinspun and Maurizio M. Chiaramonte and Philip Allen Etter},
  title = {{Force-Dual Modes: Subspace Design from Stochastic Forces}},
  year = {2025},
  eprint = {2505.23969},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.23969},
}
Figure