Causal Geometry Encoder

• Causal Geometry Encoder is a framework that extracts underlying geometric and causal structures—such as metrics, curvature, and manifolds—directly from observational and intervention data.
• It combines neural autoencoders, variational models, and causal flows to enforce invariance under interventions, ensuring robustness in representation learning and fairness.
• The approach finds applications in high-dimensional learning, quantum gravity, and spatiotemporal systems, providing theoretical guarantees and empirical performance improvements.

A Causal Geometry Encoder is a mathematical or algorithmic construct designed to encode, infer, or reconstruct the underlying geometric or causal structure—such as metric, curvature, manifolds, causal graphs, or light-cones—directly from data or physical laws, often with the goal of separating true underlying generative or causal factors from confounders, spurious correlations, or observational artifacts. These encoders are used across disparate domains including high-dimensional representation learning, Lorentzian quantum gravity, causal inference, and dynamical systems theory, where they provide the means to recover geometric invariants or causal relationships under interventions, noise, or discrete sampling.

1. Causal Geometry Encoding in Representation Learning

In the context of geometric deep learning and fairness, a causal geometry encoder aims to learn latent representations whose induced geometry—such as local Riemannian metric and curvature in ambient space—remains invariant under causal interventions. "Causal Manifold Fairness" (CMF) (Rathore, 6 Jan 2026) introduces such an encoder for counterfactual fairness, modeling the generative process as a structural causal model (SCM) $$\mathfrak{M} = \langle \mathcal{U}, \mathcal{V}, \mathcal{F} \rangle$$, with $X = f(U, A)$ where $A$ is a sensitive attribute warping the data manifold. Under counterfactual intervention $\mathrm{do}(A = a')$, the mapping $U \mapsto X$ is warped.

A neural autoencoder pair $(E, D)$ parameterizes an embedding $z = E(x)$ and decoder $D(z)$, inducing a manifold $\mathcal{M}$ in feature space. The pullback Riemannian metric is $G(z) = J_D(z)^T J_D(z)$, where $X = f(U, A)$0; higher-order geometry is encoded in the Hessians of $X = f(U, A)$1. The encoder enforces invariance under $X = f(U, A)$2 by penalizing deviations in metrics and curvatures between factual and counterfactual encodings: $X = f(U, A)$3 Such optimization enforces that the latent−to−data mapping is a Riemannian isometry under sensitive interventions. Empirically, CMF achieves reductions in counterfactual geometric errors by two to three orders of magnitude at minimal accuracy cost (Rathore, 6 Jan 2026).

2. Causal Geometry in Disentangled Representation Learning

In unsupervised learning, a causal geometry encoder can be instantiated as a variational autoencoder (VAE) whose latent coordinates are explicitly entangled by a explicit causal graph. "Causal Flow-based Variational Auto-Encoder" (Fan et al., 2023) utilizes a two-stage model: a base encoder outputs a Gaussian latent $X = f(U, A)$4, which is then transformed via a sequence of invertible "causal flows" $X = f(U, A)$5 parameterized by the DAG adjacency, yielding the final $X = f(U, A)$6. Each flow is autoregressive, where each latent dimension $X = f(U, A)$7 is conditioned only on its DAG parents, enforcing that the Jacobian of the flow reflects the ground-truth causal geometry.

The prior on $X = f(U, A)$8 is similarly defined via a flow from a standard Gaussian through the graph-coupled flows, resulting in a "geometry-preserving" prior. The overall ELBO includes KL divergence terms computable via Jacobian determinants of the triangular flow: $X = f(U, A)$9 In experiments, such causal geometry encoders produce disentangled representations aligned with generative factors and support explicit interventions (do-calculus at the latent code) while maintaining geometric fidelity to the underlying SCM (Fan et al., 2023).

3. Causal Geometry Encoding in Quantum Gravity and Spacetime Reconstruction

A Causal Geometry Encoder also arises in the geometric encoding of spacetime causal structure. In (Skakala et al., 2010), the isomorphism between Lorentzian manifolds and families of parameterized Randers geometries is exploited. Given a causal Lorentzian metric $A$0 in ADM decomposition, the null-cone structure is encoded by a 1-parameter family of Randers–Finsler norms on each spatial slice: $A$1 where $A$2 and $A$3 are constructed from the lapse, shift, and spatial metric. Reconstruction of the full causal structure (i.e., the light cones) is possible up to a conformal factor solely from these Randers data.

Similarly, the discrete Causal Set approach (Eichhorn et al., 2018) leverages order-theoretic data plus volume element to reconstruct spatial distances. For two elements $A$4 in an antichain (representing a time slice), their spatial distance is estimated as a function of the minimal suspended causal future volume $A$5 over all common future events $A$6: $A$7 A graph of antichain elements is then constructed, and the mesoscopic cutoff $A$8 is chosen to control curvature and discreteness corrections. This order-theoretic encoder, with a single tunable parameter $A$9, provably recovers the spatial metric in the continuum limit within controlled error (see also numerical convergence in (Eichhorn et al., 2018)).

4. Geometric Causal Inference via Information Flow

Beyond explicit geometry, the notion of causal geometry encoding also applies to the geometry of information flow. (Surasinghe et al., 2020) contrasts transfer entropy $\mathrm{do}(A = a')$0 (Kullback–Leibler divergence between conditionals) with a geometric causal measure $\mathrm{do}(A = a')$1 based on conditional fractal (correlation) dimension: $\mathrm{do}(A = a')$2 Here $\mathrm{do}(A = a')$3 is estimated from the scaling of interpoint distances in embedding space. $\mathrm{do}(A = a')$4 directly quantifies the "extra manifold dimension" induced by a causal dependency, sidestepping divergence issues in deterministic settings and providing a bounded geometric interpretation of causal flow. Practical algorithms involve delay embedding, correlation sum estimation, and linear regression in log–log scale for slope (dimension) estimation (Surasinghe et al., 2020).

5. Causal Geometry Encoders for Spatiotemporal Systems

For spatiotemporal fields, causality can be encoded via equivalence classes of past lightcone configurations that yield identical predictive distributions for future lightcones. (Rupe et al., 2020) formulates an encoder $\mathrm{do}(A = a')$5 that clusters past lightcones using a decay-weighted metric and merges clusters based on similarity in future predictive distributions; this yields local causal states as a compressed, geometry-preserving latent. Decoding reconstructs future fields via sampling centroids from associated predictive distributions. This approach enables both compressed Markovian forecasting and explicit quantification of how local spacetime geometry is encoded by causal structure.

6. Causal Geometry Encoding in Quantum Models

In quantum gravity, local causal geometric structure is encoded at the level of quantum field variables. The Complete Barrett–Crane model (Jercher et al., 2022) uses fields $\mathrm{do}(A = a')$6, with $\mathrm{do}(A = a')$7 and $\mathrm{do}(A = a')$8 labelling causal type (timelike, spacelike, lightlike), encoding the causal type in field arguments and constructing amplitudes via integral geometry. The model ensures that spins labelling faces of each tetrahedron are compatible with its causal type via the simplicity projector, making the encoding of causal geometry explicit and unambiguous in the quantum amplitude construction.

Domain	Causal Geometry Encoder Paradigm	Key Features
Deep Learning	Autoencoder + metric/Hessian invariance (CMF)	Latent isometry under SCM interventions; fairness/utility trade-off (Rathore, 6 Jan 2026)
Representation Learning	VAE + Causal Flows (DCVAE)	Factor-level causal structure in flows; intervention-robust latents (Fan et al., 2023)
Gravity/Spacetime	Order + volume (Causal Set), Randers geometry (ADM)	Metric reconstruction from causal order; geodesic/finsler path encoding (Eichhorn et al., 2018, Skakala et al., 2010)
Dynamical Systems	Correlation dimension (GeoC), transfer entropy	Geometric causal flow via manifold dimensionality (Surasinghe et al., 2020)
Quantum Gravity	TGFT fields with causal labelling	Explicit quantum encoding of local causal type (Jercher et al., 2022)

7. Implementation and Theoretical Considerations

Causal geometry encoders have computational and analytical requirements driven by the domain. In manifold learning, Jacobian and Hessian computations necessitate $\mathrm{do}(A = a')$9 smooth decoders and efficient use of automatic differentiation. In causal set theory, efficient graph construction (using beam-pruning, step-into-light, and shortest-path routines) is essential for scalability. Conditional dimension estimators for GeoC require careful selection of embedding and scale parameters, with approximate nearest-neighbors to accelerate correlation sum computations.

Across all instances, the encoder leverages the causal structure—whether as explicit graphs, lightcones, volume order, or adjacency matrices—to encode, regularize, or reconstruct geometry in a manner robust under interventions, noise, or quantum uncertainty, while maintaining interpretability and theoretical guarantees afforded by the underlying physical or generative axioms.