Most deep generative models share a foundational assumption: to learn the underlying structure of a dataset, they map a simple, flat Euclidean latent space (typically governed by a standard Gaussian prior) to the high-dimensional data space.
However, real-world data rarely conforms to flat geometry. Whether it is the cyclic nature of a rotating object, the periodic gait of a walking human, or the hollow structure of a molecular ring, data frequently lives on manifolds with non-trivial topologies. When we force a topologically complex dataset through a flat, contractible Euclidean bottleneck, the model encounters a fundamental mathematical barrier known as topological manifold mismatch.
The Geometry of the Mismatch
To understand why this mismatch breaks generative models, we must look at the concept of a homeomorphism—a continuous, bijective mapping between spaces that can be inverted continuously. Ideally, an encoder and decoder should form a homeomorphism between the latent space and the data manifold, preserving both proximity and global structure.
As shown by Ravelonanosy et al., when the mathematical topology of the latent space disagrees with the physical symmetry of the data, a homeomorphism becomes impossible.
Imagine trying to wrap a flat, solid square of rubber around a hollow sphere or a torus without cutting it, tearing it, or overlapping the edges. It is topologically impossible because a solid square is contractible (it can be shrunk to a single point continuously), whereas a sphere or a torus contains holes and voids that prevent such a contraction.
When a generative model attempts this impossible mapping via continuous neural networks, it is forced to make one of two destructive compromises: tearing or bridging.
Tearing (Discontinuity) Bridging (Invalid Samples)
Latent Space Latent Space
[---A---*---B---] [-------*-------]
/ \ |
/ \ v
v v Data Manifold
( A ) ( B ) ( * ) <- Mass placed
\ / \ / \ / in the empty
\_/ \_/ \___/ center
Data Manifold (Torn)
1. Topological Tearing (Discontinuities)
If the model prioritizes keeping the data manifold hollow or disconnected, the learned mapping must "tear." As Pérez Rey et al. and Connor et al. observe, this tearing forces points that are semantically identical or adjacent in the real world to be mapped to opposite ends of the latent space.
For instance, if you map a 360-degree rotation onto a 1D linear latent segment, 0 degrees and 359 degrees will end up on opposite sides of the line. The representation breaks its continuity, causing severe instability when interpolating within the latent space.
2. Density Bridging (Invalid Samples)
If the model instead insists on maintaining a continuous mapping from the solid latent prior, it must satisfy topological continuity by filling in the gaps. As demonstrated by Jin et al., mapping a contractible Gaussian distribution to a non-contractible hollow shell forces the generator to assign probability mass to the empty center.
The result? The model creates artificial "density bridges" across disconnected regions, as highlighted by Ross et al.. When you sample from the center of the latent space, the decoder generates impossible, blurry, or invalid samples that do not exist in the original data distribution.
Why Disentanglement Breaks Down
This topological bottleneck directly undermines representation learning and disentanglement. A core goal of disentangled representation learning is to assign distinct semantic factors of variation (like object rotation, color hue, or spatial phase) to independent latent dimensions.
However, standard frameworks assume these factors can be represented as independent Euclidean axes. Tonnaer et al. and Rotman et al. demonstrate that when a periodic or cyclic factor is forced into a flat coordinate system, the model is forced to use multiple latent dimensions to represent a single circular property. This behavior leads to inefficient, highly entangled representations.
Similarly, Keurti et al. argue that simply ensuring statistical independence or identifiability does not guarantee topological fidelity; cyclic relationships will still be distorted by Euclidean distances. In neurobiology, Baroni et al. found a parallel effect when modeling visual sensory neurons: mapping periodic neural symmetries (like spatial phase) onto linear latent spaces prevents models from capturing the true circular nature of neural responses.
Strategies for Topological Fidelity
To overcome the limitations of pushforward mappings from flat spaces, researchers have developed several distinct paradigms to align latent topology with data topology.
Non-Euclidean Latent Spaces
Instead of changing the mapping, this approach changes the latent space itself. By replacing Gaussian priors with distributions defined on closed manifolds, the latent space naturally mirrors the data's geometry.
- Hyperspherical Spaces: Davidson et al. introduced the Hyperspherical VAE (S-VAE), which replaces the standard Gaussian latent space with a von Mises-Fisher (vMF) distribution, capturing directional and spherical data without pulling mass toward an artificial origin.
- Toroidal Spaces: Rotman et al. developed the $T^D$-VAE, utilizing a torus-structured latent space built from the product of unit circles to capture multiple independent periodic factors natively.
- Arbitrary Closed Manifolds: Pérez Rey et al. generalized this via the Diffusion VAE ($\Delta$VAE), using Brownian motion to define priors on arbitrary closed manifolds.
Multi-Chart Atlases
Borrowing from differential geometry, another strategy avoids forcing a global mapping by breaking the complex manifold down into a collection of local coordinate patches, known as an atlas.
- Alberti et al. utilize a mixture model of VAEs where each component acts as a local "chart," preventing the formation of uncovered gaps and improving performance on complex inverse problems.
- Stolberg-Larsen and Sommer and Yu et al. extend this to normalizing flows, "gluing" multiple local coordinate charts together to compute exact geodesics and densities across complex surfaces like tori and spheres without introducing artificial discontinuities.
Implicit and Bottleneck-Free Architectures
A third paradigm abandons explicit coordinate-based pushforward mappings altogether.
- Implicit Manifolds: Ross et al. introduced Energy-Based Implicit Manifolds (EBIMs). Instead of mapping out of a latent space, EBIMs define the data manifold implicitly as the zero-set of a neural network. This allows the model to represent arbitrary topologies, loops, and disjoint components with perfect stability.
- Eliminating the Bottleneck: As Jin et al. demonstrate, Denoising Diffusion Probabilistic Models (DDPMs) inherently bypass this entire problem. Because diffusion models operate directly within the full data space rather than compressing data through a low-dimensional, contractible latent bottleneck, they naturally respect complex topological constraints and avoid generating mass in forbidden voids.
Key Foundational Work
Diffusion Variational Autoencoders (2020) — contributed 1 concept elements to this direction.
Hyperspherical Variational Auto-Encoders (2018) — contributed 1 concept elements to this direction.
Representing Closed Transformation Paths in Encoded Network Latent Space (2020) — contributed 1 concept elements to this direction.