Hierarchical Volume-Preserving Maps

• The paper introduces a framework where hierarchical volume-preserving maps guarantee a unit Jacobian determinant, preserving both local and global density.
• It details implementations using Laplacian pyramids and Haar wavelet transforms that ensure norm- and distance-preservation, critical for scalable generative modeling.
• It discusses geometric grid discretizations and cascaded multi-scale models, enabling exact likelihood evaluation and robust applications in computational geometry.

A hierarchical volume-preserving map is a smooth, bijective reparameterization of a structured space (such as high-dimensional data or geometric solids) into multiple scales or hierarchical components, with the property that the transformation preserves local and global volume. The defining characteristic is that the determinant of the Jacobian matrix at every point is identically one, ensuring that the original density structure (in data or space) is not locally contracted or expanded. This property enables exact likelihood evaluation and geometrically faithful discretizations in applications ranging from generative modeling to geometric grid construction (Li et al., 13 Jan 2025, Holhoş et al., 2015).

1. Formal Definition and Mathematical Properties

Let $\mathcal{X}$ denote the original space (e.g., $\mathbb{R}^d$ for imaging), and suppose $x\in\mathcal{X}$ is mapped into $S$ hierarchical scales $(z^{(1)},\ldots,z^{(S)})$, where each $z^{(s)}$ belongs to $\mathcal{Z}^{(s)}$. A hierarchical volume-preserving map is a $C^1$-diffeomorphism $$f: \mathcal{X} \to \mathcal{Z}^{(1)} \times \cdots \times \mathcal{Z}^{(S)}$$ satisfying

$$\det\big[\nabla_x f(x)\big] = 1\quad\forall\,x\in\mathcal{X} \ .$$

Invertibility is required so that reconstruction of $\mathbb{R}^d$0 from $\mathbb{R}^d$1 is always possible. The volume preservation guarantees that, in probabilistic models, the change-of-variables formula simplifies:

$\mathbb{R}^d$2

This transformation introduces no local density distortion, facilitating tractable likelihoods and robust geometric discretizations (Li et al., 13 Jan 2025).

2. Concrete Realizations: Laplacian Pyramid and Haar Wavelet

Two canonical examples illustrate the concept in data-analytic settings:

Laplacian Pyramid:

Let $\mathbb{R}^d$3 be a norm-preserving downsampler (e.g., bilinear with scaling), and $\mathbb{R}^d$4 the corresponding upsampler. Recursively decompose $\mathbb{R}^d$5 and $\mathbb{R}^d$6 with base $\mathbb{R}^d$7 and inverse $\mathbb{R}^d$8 up to $\mathbb{R}^d$9. Because $x\in\mathcal{X}$0 and $x\in\mathcal{X}$1 are norm-preserving, the overall transform is a tight frame: it satisfies Parseval’s identity and all singular values of its Jacobian equal one. Thus, $x\in\mathcal{X}$2 (Li et al., 13 Jan 2025).

Haar Wavelet Transform:

Decompose the input using orthonormal filters $x\in\mathcal{X}$3 into subbands at each scale via convolutions and striding. The transform matrix $x\in\mathcal{X}$4 is orthonormal ($x\in\mathcal{X}$5), implying $x\in\mathcal{X}$6. The multiscale (hierarchical) transform is exactly volume-preserving and distance-preserving (Li et al., 13 Jan 2025).

Map Type	Construction Principle	Volume Preservation Mechanism
Laplacian Pyramid	Norm-preserving linear tight frame	Parseval’s identity $x\in\mathcal{X}$7 $x\in\mathcal{X}$8
Haar Wavelet	Orthonormal matrix decomposition	$x\in\mathcal{X}$9 (orientation set to $S$0)

3. Geometric Instances on Polyhedral Domains

In geometric contexts, area- and volume-preserving maps have been constructed for families of convex polyhedra $S$1, defined as the union of a right prism of height $S$2 over a regular $S$3-gon, plus two congruent pyramidal caps. The surface $S$4 can be mapped bijectively and area-preservingly to a sphere $S$5 by explicit algebraic and trigonometric formulas, constructed zone-wise with area-preserving Jacobians. For a corresponding family of solids $S$6, a volume-preserving homeomorphism onto the solid ball $S$7 can be constructed if and only if $S$8 satisfy explicit algebraic constraints on the Jacobian; exact homeomorphisms exist for specific $S$9 (Holhoş et al., 2015).

4. Hierarchical Grid Construction and Refinable Partitions

Hierarchical volume-preserving maps naturally permit construction of uniform, refinable, bijective grids for both analytic and geometric applications.

• Surface Grids: Uniform partitions of each pyramid face in $(z^{(1)},\ldots,z^{(S)})$0 yield $(z^{(1)},\ldots,z^{(S)})$1 quadrilaterals of equal area, extendable to HEALPix-type iso-latitude pixelations for $(z^{(1)},\ldots,z^{(S)})$2, $(z^{(1)},\ldots,z^{(S)})$3, and appropriate $(z^{(1)},\ldots,z^{(S)})$4.
• Solid Grids: Subdivision of $(z^{(1)},\ldots,z^{(S)})$5 into $(z^{(1)},\ldots,z^{(S)})$6 tetrahedra, recursively refining each into four, yields tetrahedral partitions with uniform volume at each refinement level: $(z^{(1)},\ldots,z^{(S)})$7, each of volume $(z^{(1)},\ldots,z^{(S)})$8. Volume-preserving mapping to the ball ensures exact volumetric discretization (Holhoş et al., 2015).

Partition Type	Construction	Properties
Surface	Equiareal quadrilaterals	Uniform area, hierarchical
Solid	Recursive tetrahedral refinement	Uniform volume, refinable

5. Cascaded Multi-Scale Models and Likelihood Tractability

In probabilistic modeling, especially cascaded (multi-scale) diffusion models, hierarchical volume-preserving maps enable exact, tractable likelihood evaluation. Classical cascaded models suffer from intractable marginalization over extraneous latent scales. Under a hierarchical volume-preserving reparameterization $(z^{(1)},\ldots,z^{(S)})$9, the likelihood satisfies

$z^{(s)}$0

without the need for Jacobian correction. The model can then be trained using exact diffusion ELBOs on each scale, with the total objective decomposed as a sum of ELBOs for the unconditional base scale and conditional super-resolution scales (Li et al., 13 Jan 2025).

6. Theoretical Guarantees and Connections to Optimal Transport

Hierarchical volume-preserving maps provide strict area- or volume-preservation: the local area-distortion factor is identically one in all zones, and similarly the Jacobian for 3-volume is one under the solid domain mappings. In data-analytic cascaded models, the training objective under such maps is provably equivalent (up to a constant) to minimization of a weighted sum of Earth Mover’s Distances (EMD/Wasserstein-p metrics) between the true and learned score fields:

$z^{(s)}$1

where hierarchical transforms such as Laplacian pyramids or Haar wavelets admit linear-time upper bounds on the EMD by summing $z^{(s)}$2-norms of the hierarchical coefficients. Thus, volume preservation not only makes the likelihood tractable but also effects a perceptually meaningful score matching in generative modeling (Li et al., 13 Jan 2025).

7. Special Cases and Integration with Established Schemes

The mapping framework subsumes established equal-area discretization schemes such as HEALPix by explicit construction: for $z^{(s)}$3, $z^{(s)}$4, and $z^{(s)}$5, the quadrilateral partition of the mapped polyhedron reproduces the iso-latitude, equal-area pixelization standard in cosmological data analysis, with the mapping formulas yielding the HEALPix parametrization on each cap (Holhoş et al., 2015). This construction demonstrates the unified applicability of hierarchical volume-preserving maps in both geometric grid generation and modern generative modeling architectures.