Learning Tangent Bundles and Characteristic Classes with Autoencoder Atlases

Technical Summary: Learning Tangent Bundles and Characteristic Classes with Autoencoder Atlases

Problem Statement
Manifold learning aims to recover low-dimensional representations of high-dimensional data that reflect underlying geometry. The dominant paradigm relies on global Euclidean embeddings (e.g., via autoencoders, Isomap, or diffusion maps). However, this approach encounters fundamental topological obstructions: manifolds with non-trivial topology (such as the real projective plane $\mathbb{R}P^2$ or the Klein bottle) cannot be embedded in low-dimensional Euclidean space without self-intersection or loss of structure. Furthermore, even when embeddings exist, standard Euclidean coordinates do not encode differential-topological invariants like the tangent bundle, transition functions, or characteristic classes, which are essential for distinguishing orientable from non-orientable manifolds. Existing methods for estimating intrinsic dimension or tangent spaces do not resolve the deeper issue of representing manifolds where global coordinates are topologically forbidden.

Methodology
The authors propose a framework that shifts from viewing autoencoders as producing a single global embedding to treating a collection of locally trained encoder-decoder pairs as a learned atlas on the data manifold.

1. Autoencoder Atlases: Instead of a single map, the method employs a collection of charts $\mathcal{A} = \{(U_i, E_i, D_i)\}$, where $U_i$ are open sets covering the manifold $M$, $E_i$ are encoders mapping to a latent space $\mathbb{R}^d$, and $D_i$ are decoders.
2. Transition Maps and Cocycles: For overlapping charts $U_i \cap U_j$, transition maps are defined as $T_{ji} = E_j \circ D_i$. Theoretical analysis shows that reconstruction consistency ($D_i \circ E_i \approx \text{Id}$) canonically induces these transition maps to satisfy the cocycle condition ($T_{ki} = T_{kj} \circ T_{ji}$) on triple overlaps.
3. Linearization and Vector Bundles: By linearizing the non-linear transition maps, the authors construct linear transition functions $g_{ji}(x) = d(T_{ji})_{E_i(x)}$. These functions satisfy the cocycle condition for vector bundles, allowing the construction of a learned tangent bundle $T_\mathcal{A}$ over $M$. When the latent dimension matches the intrinsic dimension and the atlas is smooth, $T_\mathcal{A}$ is isomorphic to the true tangent bundle $TM$.
4. Orientability Detection via Stiefel-Whitney Classes: The first Stiefel-Whitney class $w_1(T_\mathcal{A}) \in H^1(M; \mathbb{Z}/2)$, which detects orientability, is computed directly from the learned atlas. Specifically, the sign of the Jacobian determinant of the transition maps, $\omega_{ji}(x) = \text{sign}(\det g_{ji}(x))$, forms a Čech 1-cocycle. The manifold is orientable if and only if this cocycle is a coboundary (i.e., consistent signs can be assigned to charts such that $\omega_{ji} = \nu_j \cdot \nu_i$).
5. Stability under Approximation: Recognizing that neural networks achieve only approximate reconstruction, the paper develops stability theory. It proves that the sign cocycle remains a valid Čech cocycle under explicit bounds on reconstruction error ($\epsilon$) and differential error ($\eta$), provided the transition Jacobians remain non-degenerate ($\delta > 0$). A key refinement (Theorem 4.12) shows that stability holds locally on "nerve triangles" even if global bounds are violated, provided a "bad" chart can be re-indexed to a position where its reconstruction error does not affect the cocycle condition.
6. Chart Count Constraints: The paper establishes that the minimum number of charts required is determined by the covering type of the manifold (the minimum cardinality of a good cover), rather than the structure of the tangent bundle. Non-trivial characteristic classes obstruct single-chart representations, but a trivial tangent bundle does not guarantee a single chart suffices for compact manifolds.

Key Contributions

1. Formalization of Autoencoder Atlases: The authors formalize the concept of an autoencoder atlas, proving that reconstruction consistency implies the cocycle condition for transition maps and enabling the construction of a learned vector bundle.
2. Algorithmic Orientability Detection: They provide a direct algorithmic criterion for detecting orientability by computing the first Stiefel-Whitney class from the signs of Jacobian determinants of learned transitions. This reduces the problem to checking if a sign cocycle is a coboundary (equivalent to 2-coloring the nerve graph).
3. Stability Theory for Approximate Atlases: The paper proves that the topological invariants derived from the atlas are stable under the approximate reconstruction inherent in neural network training. This includes a global stability theorem and a stronger local "per-triple" theorem that explains successful detection even when global error bounds are violated.
4. Topological Lower Bounds: The work identifies the covering type of the manifold as the fundamental lower bound on the number of autoencoder charts required, linking the architectural complexity of the model directly to the topology of the data.
5. Loss Function Design: The authors demonstrate that explicit cocycle regularization is unnecessary; minimizing the reconstruction loss alone is sufficient to drive the cocycle error to zero, as the cocycle error is bounded by the reconstruction error of the intermediate chart in a triple.

Experimental Results
The framework was validated on four datasets with known orientability:

• 2-Sphere ($S^2$): An orientable manifold. The method correctly detected orientability (coboundary test passed) using a 4-chart atlas.
• Möbius Band: A non-orientable manifold with disconnected overlaps. The method correctly detected non-orientability (coboundary test failed) using a 2-chart atlas, identifying the sign flip across the two overlap components.
• Klein Bottle: A non-orientable manifold in $\mathbb{R}^4$. Using an 8-chart atlas, the method correctly detected non-orientability in converged trials. The experiments highlighted the utility of the local stability theorem, which allowed detection even when one chart violated the global differential error bound ($\eta < 1$).
• $\mathbb{R}P^2$ (Line Patches): A non-orientable manifold represented as 100-dimensional images. Despite high codimension and large latent-side differential errors ($\eta_{lat} \gg 1$), the method correctly detected non-orientability in 4 out of 5 trials, suggesting that the tangent-restricted error (which governs the theory) remained controlled despite the high-dimensional proxy failing.

Across all converged trials (17 out of 20 total runs), the coboundary test achieved 100% accuracy in detecting the correct orientability class.

Significance
The paper claims to bridge the gap between modern manifold learning and classical differential topology. By treating learned representations as atlases, the authors demonstrate that neural networks can be used to compute differential-topological invariants (specifically the first Stiefel-Whitney class) directly from data. This provides a rigorous theoretical foundation for understanding why multi-chart architectures are necessary for topologically non-trivial data and offers a practical, algorithmic method for detecting orientability without requiring explicit topological computations during training. The work suggests that the "obstruction" to global embeddings is not a failure of the learning algorithm but a feature of the data's topology, which can be faithfully captured and analyzed through the lens of vector bundles and characteristic classes.