When modeling the movement of a beating heart, the stretching of a tongue, or the transformation of high-dimensional data, a common physical reality often goes ignored: volume does not change. In biomechanics, tissues like muscle are essentially incompressible. In machine learning, transformations that "collapse" or "explode" space lead to numerical instability and lost information.
Researchers across medical imaging, robotics, and deep learning are increasingly moving away from unconstrained deformation models toward frameworks that explicitly enforce volume preservation. By integrating divergence-free fields and Jacobian regularization, these models ensure that every "squeeze" in one direction is balanced by an "expansion" in another, leading to more physically realistic and mathematically stable results.
The Failure of Displacement Constraints
The simplest way to model motion is to calculate a displacement vector for every point in a volume. To ensure the volume stays constant, one might try to constrain the divergence of this displacement field to zero. However, as Xiaofeng Liu and Jerry L Prince point out in Incompressible Cardiac Motion Estimation of the Left Ventricle Using Tagged MR Images, this is a mathematical shortcut that fails during the large deformations of a beating heart.
The intuition is that displacement is a "jump" from point A to point B, while volume preservation is a property of the path taken. To maintain physical realism, the constraint must be applied to the instantaneous velocity field. By ensuring the velocity field is divergence-free at every moment and integrating it over time, the model naturally preserves volume even during extreme contractions. Marvin Eisenberger and Daniel Cremers take this further in Divergence-Free Shape Interpolation and Correspondence by representing velocity as the curl of a potential field, a trick from fluid dynamics that mathematically guarantees zero divergence by construction.
The Muscular Hydrostat
This principle is most visible in "muscular hydrostats"—structures like the tongue or a Hydra’s body that lack a skeletal support. In these systems, the internal fluid or tissue is incompressible. As Jianwu Dang 0001 explains in A model based investigation of activation patterns of the tongue muscles for vowel production, a tongue elongates not by growing, but by contracting transverse muscles to reduce its cross-sectional area, forcing the volume to "squirt" forward.
Without volume-preserving constraints, computational models of these movements often produce "ghost" motions or unrealistic volume drift. Jonghye Woo and colleagues demonstrated in 3D Tongue Motion from Tagged and Cine MR Images that enforcing incompressibility allows for much more accurate interpolation of motion between sparse MRI slices, as the physics of the tissue fills in the gaps that the camera cannot see.
Jacobian Regularization in Neural Networks
In the realm of deep learning, volume preservation is enforced through the Jacobian determinant. The Jacobian matrix describes how a transformation scales space locally; if its determinant is exactly $1$, the transformation is volume-preserving (equiareal).
This mathematical anchor is used in several ways:
- Stability in Dimensionality Reduction: Guannan Zhang and Jacob Hinkle use Jacobian regularization in Learning nonlinear level sets for dimensionality reduction in function approximation to prevent reversible neural networks from collapsing dimensions or becoming numerically unstable when approximating complex engineering simulations.
- Unsupervised Motion Tracking: In Deep Unsupervised Phase-based 3D Incompressible Motion Estimation in Tagged-MRI, Zhangxing Bian and Fangxu Xing incorporate a Jacobian penalty directly into the loss function. This forces the network to learn "diffeomorphic" mappings—smooth, reversible transformations that respect the physical incompressibility of biological tissue.
- Quantization and Flows: For Large Language Models, Ofir Gordon uses volume-preserving regularization in LATMiX to keep activations from vanishing or exploding during 4-bit quantization. Similarly, Ming‐Tian Zhang shows in Spread Flows for Manifold Modelling that restricting flows to be volume-preserving simplifies the calculation of entropy and divergence, making it easier to model data on low-dimensional manifolds.
Why It Works
The power of volume preservation lies in constraint-induced accuracy. By telling a model that the total volume must remain constant, we effectively remove millions of "impossible" solutions from the search space. Whether tracking a heart valve or training a normalizing flow, these constraints act as a form of inductive bias that aligns the mathematics of the model with the fundamental laws of the physical world. As Tommaso Mansi and Xavier Pennec argue in LogDemons Revisited, this doesn't just make the models more realistic—it makes them more robust, requiring less data to reach a physically plausible conclusion.