Compositional Symmetry as Compression: Lie Pseudogroup Structure in Algorithmic Agents

AI summary

Overview

• Research area: Theoretical machine learning / algorithmic information theory, intersecting with differential geometry (Lie pseudogroups, Spencer formalism), neuroscience-inspired predictive coding, and symmetry-aware deep learning. Published as arXiv:2510.10586v1 [cs.LG] on 12 Oct 2025, associated with the workshop "Symmetry and Geometry in Neural Representations."
• Technical level: Advanced. The paper assumes familiarity with Lie groups and algebras, Lie pseudogroups, jets and the Spencer complex, Noether-style conservation arguments, and Lyapunov/ISS stability.
• Scope (one sentence): A conceptual, largely theoretical paper that proposes describing natural sensory streams via local actions of finite-parameter Lie pseudogroups, and derives the structural and dynamical constraints this imposes on a neural world-tracking agent — including a symmetry-based reformulation of hierarchical predictive coding.

What This Paper Is About

In the Kolmogorov/algorithmic view, an agent is a program that compresses sensory streams by building generative programs of its environment. This paper asks what structure such compressive models can exploit, and answers that the relevant prior is a compositional symmetry: natural streams are well described by local actions of finite-parameter Lie pseudogroups on low-dimensional configuration manifolds (latent spaces). From this premise, the author derives what an agent must satisfy — in its weights and in its dynamics — in order to track such streams, and uses that structure to motivate hierarchical, symmetry-aware architectures and a group-theoretic version of predictive coding.

Key Contributions

1. A symmetry-first definition of generative models. Generative models are defined as smooth maps f : C → R^X from an M-dimensional configuration manifold (M ≪ X) to observations, with a Lie pseudogroup G acting locally on both C and R^X. A "Lie generative model" is one where every c ∈ C can be written c = γ · c₀ and f(c) = γ · I₀ for some γ ∈ G.
2. Derivation of structural and dynamical constraints from world-tracking. Modeling the agent as a neural dynamical system ẋ = F(x; w, I_θ(t)) with readout p(x) ≈ I_θ(t), the paper shows that feasible tracking forces (i) equivariance of the agent's constitutive equations and readouts, and (ii) Noether-style conserved quantities and low-dimensional invariant manifolds in the dynamics.
3. A hierarchical reduction via a flag of sub-pseudogroups. Choosing G = H₀ ⊃ H₁ ⊃ ... ⊃ H_L yields nested reduced manifolds M_k := M_{k−1}/H_k with dim M_{k+1} < dim M_k, giving a geometric account of the "blessing of compositionality" and a link to the Spencer complex of compatibility conditions.
4. A symmetry-based, self-contained version of predictive coding. Higher layers receive only coarse-grained residual transformations (prediction-error coordinates) in the quotient directions H_{k−1}/H_k unresolved at lower layers, with an explicit canonicalize-then-coarse-grain-update loop.

Main Findings

• Equivariance is required for tracking. If the input stream is generated by a finite-parameter Lie pseudogroup acting on a reference image, then effective tracking demands that internal dynamics respect the same action: f(γ · x; w, γ · I_θ) = γ · f(x; w, I_θ) and p(γ · x) = γ · p(x). For infinitesimal γ = exp(ε T_k), this reduces to linear commutation conditions T_k f = (∂f/∂x) T_k x that must hold for every Lie-algebra generator T_k.
• Structural constraints on weights. The equivariance requirement places constraints on the parameters w — described as "weight tying, zero blocks, etc." — identical in spirit to group-equivariant CNNs.
• Noether-style conservation under static inputs. Freezing θ makes the readout invariant, so each of the Y readout channels defines a conserved quantity. Trajectories are confined to an (X − Y)-dimensional leaf of phase space.
• Adiabatic drift stays low-dimensional. When θ(t) varies slowly, the invariant leaf drifts, but its dimension remains ≤ M + X − Y — gaining at most M extra degrees of freedom.
• A Lyapunov closure for approximate tracking. With error E = p(x) − I_θ and V = ½‖E‖², the modified dynamics ẋ = F(x; w, I_θ) + K E with gain K commuting with every T_k gives V̇ ≤ 0 under static inputs, and an ISS-type bound V̇ ≤ −α‖E‖² + β‖İ_θ‖² for suitable α, β > 0 under slow drift. The gain must satisfy that D_x p(x) K is negative semi-definite.
• Nested reduced manifolds formalize compositionality. Since every γ near the identity factors as exp(Σ_k θ_k T_k), declaring certain generators negligible at coarser levels produces the flag G = H₀ ⊃ H₁ ⊃ ... ⊃ H_L and quotients M_k = M_{k−1}/H_k. The paper states that each quotient sharpens the description while retaining the ability to recompose fine structure by re-activating generators.
• Spencer machinery as a consistency checker. The Spencer complex is described as a jet-level compatibility complex whose cohomology governs formal integrability, prolongation, and obstructions — a structured checker for whether local symmetry rules can be stitched into a coherent global transformation. Exactness of the sequence guarantees compatibility of layered constraints. All constructions in the paper are explicitly local; global statements require additional compatibility (Spencer exactness).
• A residual-based predictive hierarchy. Errors are expressed as small residuals ε_k = γ̂_k^{−1} γ_{k−1} ∈ H_{k−1}/H_k with ε_k = exp(η_k^a S_k^a), estimated by regularized least squares (e.g., Tikhonov). After splitting η_k = P_k η_k + Q_k η_k with P_k the orthogonal projector onto h_k = Lie(H_k), the level updates γ̂_k ← γ̂_k exp(P_k η_k) and passes Q_k η_k upward. Messages are canonicalized as r_k := γ̂_k^{−1} · I_θ(t) − I₀ and coarse-grained as m_{k→k+1} := C_{k→k+1}(r_k) using an H_k-invariant, H_{k+1}-equivariant coarse-grainer. An equivalent algebra-space message is η̄_{k→k+1} = Ad_{γ̂_k^{−1}}(Q_k η_k).
• No experimental results are reported. The paper contains no datasets, benchmarks, or quantitative performance measurements. It is a theoretical and conceptual contribution with a worked analogical example (Appendix C).
• Acknowledged scope limitation. The author notes that though the picture is possible in some cases (e.g., tracking robotic, jointed cats), it may fail when the generative model's latent space is very complex.

Methodology in Plain English

The paper builds an argument in layers rather than running experiments:

1. Start from compression. Take the algorithmic-agent premise — agents are programs that compress sensory streams, guided by Ockham's razor (simpler, shorter programs preferred) and serving an objective function (described as telehomeostasis, or preservation of pattern, in living systems).
2. Assume structure in the world. Rather than arbitrary inputs, assume sensory streams are generated by a small number of parameters acting on a low-dimensional configuration manifold — i.e., a finite-parameter Lie pseudogroup with local transformations.
3. Write down the agent. Model it as a neural dynamical system driven by such streams, with a projector as a readout and a "Comparator" that enforces world-tracking.
4. Derive consequences. Show that if tracking must succeed, the internal equations must commute with the same symmetry, which in turn produces conservation laws (as in Noether's theorem) and restricts trajectories to low-dimensional leaves; slow parameter drift moves those leaves without blowing up their dimension.
5. Close the loop. Add a symmetry-preserving feedback term and a Lyapunov function to handle approximate (rather than exact) tracking, using standard input-to-state stability reasoning for the drifting case.
6. Build the hierarchy. Nest sub-pseudogroups by declaring some generators negligible at coarse levels, take quotients to get nested manifolds, and interpret the quotients as the residual messages that flow upward in a predictive-coding-style loop.
7. Illustrate. The appendices formalize the hierarchy (Appendix A), contrast generative versus predictive directions of flow (Appendix B), and map the construction onto a concrete example: a Blender character rig for a cat, as in the Blender-generated movie Flow (Appendix C).

Why This Matters

Impact on research. The paper offers a unified geometric language in which compositional symmetry, hierarchical abstraction, and Lie pseudogroup theory are described as three facets of one structure. It proposes a reason why deep hierarchical architectures enjoy favorable sample complexity on hierarchical tasks, and clarifies why a bare manifold prior can be insufficient without additional geometric covering structure.

Real-world applications (as motivated or illustrated in the paper):

• Computer animation and rigging: Appendix C maps the abstract levels onto a Blender rig, with levels for camera and lens, global body/root, torso/spine chain (product-of-exponentials), limbs/paws/tail, facial morphology (shape keys), fur/appearance, illumination gauge, and environment/camera jitter. A minimal state vector is described as including camera (6 + lens), root (6), roughly 20–40 joint angles, 10–30 shape-key coefficients, 3–8 groom principal components, and 9 spherical-harmonic lighting coefficients.
• Robotics and articulated tracking: the author cites tracking of robotic, jointed figures as a case where the approach can apply.
• Neuroscience and predictive coding: the formalism is presented as a tentative implementation of hierarchical predictive processing in the vein of Friston (2018), in which higher levels receive only residuals along unresolved symmetry directions.
• Architecture design for machine learning: the structural constraints derived (weight tying, zero blocks, commutation with generators) are described as being "identical in spirit" to group-equivariant CNNs, suggesting symmetry-aware designs as principled agent architectures.

Industry relevance. The most direct relevance is to practitioners building equivariant or symmetry-constrained neural networks, and to graphics/animation pipelines where parameter hierarchies are already organized in the coarse-to-fine order the paper formalizes. Notably, the paper reports no empirical validation, so industrial adoption would first require the empirical tests the author lists as future work.

Future Directions

1. Generalization to stochastic inputs. The author lists SDE analogues of the Lyapunov formulation (Equation 5) and analysis of robustness as open work.
2. Constructing valid gain operators. Developing K operators for valid Lyapunov world-tracking problems is stated as a remaining task.
3. Empirical tests. Controlled experiments with equivariant architectures under known generative symmetries are proposed but not carried out.
4. Formal links to Spencer exactness and beyond. Connecting the formalism rigorously to Spencer exactness, moduli stacks, and integrability guarantees in practical learning systems. The paper also notes that symmetry-discovery methods provide a route to learning G from data, closing the loop between structure and learning.

Target Audience

This paper is aimed at researchers comfortable with differential geometry and group theory who work on the theory of representation learning, symmetry-aware and equivariant neural networks, and geometric deep learning. It will also interest computational neuroscientists working on predictive coding and hierarchical Bayesian accounts of cortex, and mathematically inclined roboticists or graphics researchers who think about articulated motion as Lie-group composition. Readers looking for empirical benchmarks, datasets, or implementable recipes will not find them here — the paper is a conceptual framework accompanied by a worked animation-rig analogy.

Authors’ abstract

In the algorithmic (Kolmogorov) view, agents are programs that track and compress sensory streams using generative programs. We propose a framework where the relevant structural prior is simplicity (Solomonoff) understood as \emph{compositional symmetry}: natural streams are well described by (local) actions of finite-parameter Lie pseudogroups on geometrically and topologically complex low-dimensional configuration manifolds (latent spaces). Modeling the agent as a generic neural dynamical system coupled to such streams, we show that accurate world-tracking imposes (i) \emph{structural constraints} -- equivariance of the agent's constitutive equations and readouts -- and (ii) \emph{dynamical constraints}: under static inputs, symmetry induces conserved quantities (Noether-style labels) in the agent dynamics and confines trajectories to reduced invariant manifolds; under slow drift, these manifolds move but remain low-dimensional. This yields a hierarchy of reduced manifolds aligned with the compositional factorization of the pseudogroup, providing a geometric account of the ``blessing of compositionality'' in deep models. We connect these ideas to the Spencer formalism for Lie pseudogroups and formulate a symmetry-based, self-contained version of predictive coding in which higher layers receive only \emph{coarse-grained residual transformations} (prediction-error coordinates) along symmetry directions unresolved at lower layers.