Volume-Preserving Encoders

• Volume-preserving encoders are mapping functions that maintain a prescribed notion of volume, ensuring preservation of geometric invariants like simplex or Jacobian volumes.
• They are implemented across different paradigms including linear compressive embeddings, invertible flows for exact lossless encoding, and neural networks to stabilize gradients.
• Empirical evaluations show that enforcing volume preservation architecturally can enhance performance in dimensionality reduction, dynamics modeling, and classification tasks.

Searching arXiv for the cited papers and closely related work to ground the article. Volume-preserving encoders are representation maps constructed so that some mathematically specified notion of volume is preserved under encoding. In the literature, the preserved object is not unique. One line of work studies linear dimension reduction that approximately preserves the volumes of finite simplices in an $n$-point Euclidean dataset; another studies same-dimensional bijections with Jacobian determinant $1$, so that local phase-space or latent-space volume is preserved exactly; a third uses generating differential forms to parameterize exact volume-preserving maps in $\mathbb{R}^3$; and a fourth builds deep feedforward architectures whose hidden transformations are volume preserving in order to control gradient behavior (Zouzias, 2010, Zhang et al., 2021, Zhu et al., 2022, Verdier et al., 2014, MacDonald et al., 2019). A central distinction running through this literature is that exact preservation of ambient $N$-dimensional volume is impossible under a dimension-reducing map $N \to d < N$, so “volume preservation” in compressive settings must instead refer to lower-dimensional simplex volumes or other derived geometric quantities (Zouzias, 2010).

1. Competing notions of volume preservation

The discrete geometric formulation treats an encoder as a linear map $f:\mathbb{R}^N \to \mathbb{R}^d$ acting on a finite point cloud $P \subset \mathbb{R}^N$, and asks whether the volumes of convex hulls of small subsets can be approximately preserved after projection. For a subset $S\subset P$, the relevant quantity is the normalized relative change

$$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$

which behaves like a per-dimension multiplicative distortion. In this setting, the encoder is compressive, linear, and only approximately volume preserving on finite subsets rather than globally measure preserving (Zouzias, 2010).

The flow-based and dynamical-systems formulations instead use the classical criterion

$\det \frac{\partial F(x)}{\partial x} = 1,$

or its equivalent measure-theoretic statement $1$0 for bounded open sets $1$1. Here the encoder is same-dimensional and bijective. In normalizing flows this removes the log-determinant term from the change-of-variables formula; in source-free dynamics it encodes incompressibility; and in exact map representations it is expressed through pullbacks of differential forms and generating forms (Zhang et al., 2021, Zhu et al., 2022, Verdier et al., 2014).

A plausible implication is that the phrase “volume-preserving encoder” does not denote a single architecture class but a family of constructions tied to different invariants. In compressive geometry, preservation concerns simplex volumes of subsets. In invertible flows and dynamics, it concerns infinitesimal Jacobian volume. In lossless compression, it further concerns compatibility between real-valued invertibility and discrete machine-representable bijectivity (Zouzias, 2010, Zhang et al., 2021).

2. Low-dimensional Euclidean embeddings as compressive volume-preserving encoders

In the finite-set setting, the canonical result is the existence theorem for linear maps that preserve all simplex volumes of sufficiently small subsets. For an $1$2-point set $1$3 and $1$4, there exists a linear mapping $1$5 such that for every subset $1$6 with $1$7,

$1$8

This gives normalized distortion

$1$9

for all subsets up to size $\mathbb{R}^3$0 (Zouzias, 2010).

The underlying geometry is expressed through Gram determinants. If $\mathbb{R}^3$1 and

$\mathbb{R}^3$2

then simplex volume is controlled by $\mathbb{R}^3$3, and for a linear encoder $\mathbb{R}^3$4,

$\mathbb{R}^3$5

Volume preservation therefore reduces to approximate preservation of Gram determinants of difference matrices (Zouzias, 2010).

The proof is existential and probabilistic. The map is chosen as a Gaussian random projection, and the squared volume ratio for a fixed $\mathbb{R}^3$6-point subset has distribution

$\mathbb{R}^3$7

Gordon’s theorem compares the geometric mean of this product to a single $\mathbb{R}^3$8 variable, after which incomplete-gamma tail bounds and a union bound over all subsets yield the final guarantee (Zouzias, 2010).

This formulation strictly generalizes distance preservation, since the $\mathbb{R}^3$9 case is just the Euclidean distance case. The same paper reproves a distance-distortion theorem of Matoušek with distortion

$N$0

and notes a lower bound

$N$1

so the $N$2 dependence is essentially optimal up to logarithmic factors (Zouzias, 2010).

This line of work is best understood as a theory of compressive linear geometric encoders that preserve higher-order relations on finite data. It does not address Jacobian determinant equal to $N$3, learned neural encoders, normalizing flows, or continuous measure-preserving maps, and it does not provide reconstruction guarantees beyond the geometric embedding statement (Zouzias, 2010).

3. Invertible flows for exact lossless encoding

In flow-based compression, a volume-preserving encoder is a bijection $N$4 with

$N$5

For such a flow,

$N$6

so training reduces to maximum likelihood on the prior evaluated at the latent, without a log-determinant term (Zhang et al., 2021).

The motivation in lossless compression is not only analytic simplicity. The discrete-coding argument shows that non-volume-preserving flows cause two distinct pathologies after discretization. If $N$7, coding in latent space can incur extra bits; if $N$8, the discretized map may fail to remain bijective, so exact decoding can fail. This makes unit Jacobian magnitude a structural compatibility condition between continuous flows and discrete coding schemes (Zhang et al., 2021).

The iVPF construction addresses the additional gap between real-valued invertibility and machine-level exactness. It starts from a trained continuous volume-preserving flow and replaces ordinary floating-point computations with a numerical scheme $N$9 satisfying the exact round-trip property

$N \to d < N$0

The model assumes $N \to d < N$1-precision quantization

$N \to d < N$2

and the central coupling-layer construction is the Modular Affine Transformation, which augments the affine update with an auxiliary remainder state $N \to d < N$3. The algebraic closure condition

$N \to d < N$4

is exactly what allows the rational approximants in the coupling block to close with $N \to d < N$5, making the block bijective on quantized data (Zhang et al., 2021).

The paper proves that if the input remainder $N \to d < N$6, then the output remains $N \to d < N$7-precision, the final remainder also lies in $N \to d < N$8, and $N \to d < N$9 and $f:\mathbb{R}^N \to \mathbb{R}^d$0 establish a bijection. The approximation error relative to the original continuous affine transform is

$f:\mathbb{R}^N \to \mathbb{R}^d$1

and across $f:\mathbb{R}^N \to \mathbb{R}^d$2 layers the latent discrepancy satisfies

$f:\mathbb{R}^N \to \mathbb{R}^d$3

Thus the encoder-decoder is exact as a discrete map while remaining close to the trained continuous model (Zhang et al., 2021).

Empirically, this exactness is accompanied by competitive compression performance. At $f:\mathbb{R}^N \to \mathbb{R}^d$4, on CIFAR10 the difference between iVPF and the corresponding continuous volume-preserving flow is only $f:\mathbb{R}^N \to \mathbb{R}^d$5 bpd, and the final entropy coding adds another $f:\mathbb{R}^N \to \mathbb{R}^d$6 bpd. On CIFAR10, LBB attains 3.12 bpd while iVPF attains 3.20 bpd, but iVPF uses 6.00 auxiliary bits rather than 39.86, requires 6.6 ms rather than 97 ms coding time, and uses 2 rather than 188 coding operations. Against IDF and IDF++, iVPF reports 3.20 bpd on CIFAR10, 4.03 on ImageNet32, and 3.75 on ImageNet64 (Zhang et al., 2021).

4. Intrinsic volume-preserving networks for source-free dynamics

For learning source-free dynamical systems, the volume-preserving encoder is a same-dimensional map $f:\mathbb{R}^N \to \mathbb{R}^d$7 intended to approximate the time-$f:\mathbb{R}^N \to \mathbb{R}^d$8 flow map of an ODE

$f:\mathbb{R}^N \to \mathbb{R}^d$9

If the vector field is source-free, the phase flow $P \subset \mathbb{R}^N$0 satisfies

$P \subset \mathbb{R}^N$1

which motivates building the network itself so that $P \subset \mathbb{R}^N$2 exactly rather than penalizing deviations numerically (Zhu et al., 2022).

VPNets realize this by composing intrinsically volume-preserving modules. The residual module

$P \subset \mathbb{R}^N$3

updates one coordinate block using only the complementary coordinates, so its Jacobian is block triangular with identity on the diagonal and determinant $P \subset \mathbb{R}^N$4. The activation module

$P \subset \mathbb{R}^N$5

has the same property. The linear module

$P \subset \mathbb{R}^N$6

uses structured determinant-one matrices $P \subset \mathbb{R}^N$7, so the entire affine map remains volume preserving (Zhu et al., 2022).

Two architectures are built from these components. An R-VPNet is a composition of residual modules,

$P \subset \mathbb{R}^N$8

whereas an LA-VPNet alternates linear and activation modules,

$P \subset \mathbb{R}^N$9

Because every constituent map has determinant one, both architectures are exactly volume preserving by construction (Zhu et al., 2022).

Theoretical expressivity is supplied by two approximation theorems. If $S\subset P$0 is compact and $S\subset P$1 is a volume-preserving flow map generated by a $S\subset P$2, divergence-free vector field, then for every $S\subset P$3 there exists $S\subset P$4 with $S\subset P$5, and likewise for $S\subset P$6. This establishes that exact volume preservation does not, for the target class $S\subset P$7, preclude universal approximation on compacts (Zhu et al., 2022).

The encoder interpretation is direct: these networks are reversible, same-dimensional latent-coordinate transforms suited to incompressible dynamics. They are not compressive encoders, because global preservation of $S\subset P$8-dimensional volume requires the domain and codomain to have the same dimension. The empirical section emphasizes accurate trajectory prediction, generalization beyond training trajectories, and structure-preserving rollout behavior on the 3D Volterra system and a 4D charged-particle model; it also remarks that R-VPNet is easier to optimize, while LA-VPNet may be more parameter-efficient, with reported parameter counts of 0.2K versus 2.3K on Volterra and 0.6K versus 3.8K on the charged-particle example (Zhu et al., 2022).

5. Generating forms and exact volume-preserving map representations in $S\subset P$9

A different tradition represents exact volume-preserving maps in $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$0 through generating differential forms. If $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$1, a $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$2-map $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$3 is volume preserving when

$$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$4

equivalently $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$5. Choosing primitives $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$6 with $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$7, an exact volume-preserving map satisfies

$$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$8

where $$\left(\frac{\operatorname{Vol}(f(S))}{\operatorname{Vol}(S)}\right)^{1/(|S|-1)},$$9 is a generating form. In $\det \frac{\partial F(x)}{\partial x} = 1,$0, this becomes a generating one-form, which acts as an implicit parameterization of the map $\det \frac{\partial F(x)}{\partial x} = 1,$1 (Verdier et al., 2014).

Lomelí–Meiss had derived thirty-six generating one-form cases in $\det \frac{\partial F(x)}{\partial x} = 1,$2, but the classification result shows that, up to variable relabeling and adjunction, these reduce to five essentially distinct classes. Three correspond to previously understood Hamiltonian or Lagrangian constructions, and two are novel. The five classes are named SE+SE, DL+SE, DL+DL, $\det \frac{\partial F(x)}{\partial x} = 1,$3, and $\det \frac{\partial F(x)}{\partial x} = 1,$4 (Verdier et al., 2014).

The first three classes admit clear mechanical interpretations. The SE+SE class corresponds to two successive Symplectic Euler steps applied to Hamiltonian subsystems generated by $\det \frac{\partial F(x)}{\partial x} = 1,$5 and $\det \frac{\partial F(x)}{\partial x} = 1,$6. The DL+SE class combines a discrete Lagrangian step for one subsystem with a Symplectic Euler step for another. The DL+DL class uses discrete Lagrangian structure in both stages. These constructions yield consistent first-order volume-preserving numerical integrators for divergence-free vector fields (Verdier et al., 2014).

The $\det \frac{\partial F(x)}{\partial x} = 1,$7 and $\det \frac{\partial F(x)}{\partial x} = 1,$8 classes are genuinely new. Their generating forms do not admit an obvious Hamiltonian or Lagrangian interpretation, but the paper shows that consistent first-order volume-preserving schemes still exist, at least for linear divergence-free fields under assumptions such as $\det \frac{\partial F(x)}{\partial x} = 1,$9 or $1$00. The importance of this result is classificatory: apparently many different exact volume-preserving parameterizations collapse to a small number of canonical templates once symmetry and adjoint equivalence are taken into account (Verdier et al., 2014).

For the theory of volume-preserving encoders, this suggests a distinct design paradigm: rather than parameterizing a map directly and constraining $1$01, one can parameterize potentials $1$02 and define the map implicitly through derivative relations and twist conditions. This is not a neural-network construction, but it provides a rigorous vocabulary for exact three-dimensional volume-preserving coordinate transforms (Verdier et al., 2014).

6. Feedforward volume-preserving neural networks and general design principles

A further interpretation of volume-preserving encoders appears in deep feedforward networks whose hidden layers preserve volume in order to mitigate vanishing and exploding gradients. In this setting, a hidden layer has the form

$1$03

where $1$04 is a structured volume-preserving linear transformation, $1$05 is a bias, and $1$06 is a coupled volume-preserving activation. The final layer is not volume preserving because classification reduces dimension (MacDonald et al., 2019).

The linear transform is factorized as

$1$07

Here each $1$08 is a block-diagonal direct sum of $1$09 rotations, each $1$10 is a fixed permutation chosen randomly before training begins, and $1$11 is a positive diagonal matrix with telescoping parameterization ensuring $1$12. The activation is a pairwise coupled map $1$13, with a principal example given by the coupled Chebyshev map

$1$14

applied independently to disjoint coordinate pairs (MacDonald et al., 2019).

The paper’s central claim is not exact gradient-norm preservation but that volume preservation “forces the gradient (on average) to maintain equilibrium and not explode or vanish.” The argument is geometric: a volume-preserving hidden map cannot contract all directions simultaneously, so systematic collapse or blowup of backpropagated signals becomes less likely across depth. The empirical diagnostic

$1$15

shows markedly flatter learning profiles than standard dense-ReLU baselines in a 10-layer network (MacDonald et al., 2019).

Performance is reported on MNIST and IMDB bag-of-words. With a 4-layer network, VPNN attains 98.06% on MNIST and 86.89% on IMDB, while the standard dense-ReLU baseline attains 97.42% and 86.35%, respectively. A Mixed1 model combining a standard dense layer with the coupled Chebyshev activation achieves 98.40% on MNIST and 87.16% on IMDB. Parameter scaling is reduced from $1$16 for a standard fully connected network to $1$17 for VPNN, and each hidden layer has $1$18 trainable parameters rather than $1$19 (MacDonald et al., 2019).

Taken together, the feedforward, flow-based, dynamical-systems, geometric-embedding, and generating-form literatures establish a common structural lesson. Volume preservation is most naturally enforced by architecture or parameterization rather than by a soft penalty. The preserved object, however, depends on the task: simplex volumes for compressive Euclidean embeddings, Jacobian volume for same-dimensional invertible maps, bin-volume compatibility for lossless coding, or hidden-space volume for gradient management. A common misconception is that all of these are interchangeable. The literature instead shows that they are mathematically distinct, with different dimensionality requirements, different notions of exactness, and different admissible encoder families (Zouzias, 2010, Zhang et al., 2021, Zhu et al., 2022, Verdier et al., 2014, MacDonald et al., 2019).