Compositional Symmetry as Compression: Lie Pseudogroup Structure in Algorithmic Agents

Basic Information

• Paper ID: 2510.10586
• Title: Compositional Symmetry as Compression: Lie Pseudogroup Structure in Algorithmic Agents
• Author: Giulio Ruffini (Neuroelectrics, Starlab, BCOM, Barcelona, Spain)
• Classification: cs.LG cs.AI cs.IT math.IT q-bio.NC
• Publication Date/Venue: Under Review - Proceedings Track 2025
• Paper Link: https://arxiv.org/abs/2510.10586

Abstract

Based on the framework of algorithmic information theory (Kolmogorov theory), this paper proposes that intelligent agents are programs that track and compress sensory streams through generative programs. The authors present a framework that understands relevant structural priors as compositional symmetries: natural data streams can be well-described through the local action of finite-parameter Lie pseudogroups on geometrically and topologically complex low-dimensional configuration manifolds. Modeling agents as universal neural dynamical systems coupled to such data streams, the paper demonstrates that accurate world tracking requires: (1) structural constraints—equivariance of agent constitutive equations and readouts; (2) dynamical constraints—under static inputs, symmetries induce conserved quantities in agent dynamics and restrict trajectories to reduced-dimensional invariant manifolds. This produces a hierarchical structure of reduced manifolds aligned with pseudogroup compositional decomposition, providing a geometric explanation for the "blessing of compositionality" in deep models.

Research Background and Motivation

Core Problem

The core problem addressed by this paper is: How can one construct a symmetry-based theoretical framework for algorithmic agents that enables them to efficiently compress and track natural data streams with compositional structure?

Research Significance

1. Compression and Structure Discovery: Within the Kolmogorov theory framework, the core task of agents is to construct compressed models for understanding environments, where symmetry provides a natural structured compression mechanism
2. Theoretical Foundation for Deep Learning: Provides mathematical theoretical explanation for the superior sample complexity of deep models on hierarchical tasks
3. Geometric Foundation for Predictive Coding: Provides a symmetry-based geometric theoretical framework for predictive coding

Limitations of Existing Methods

1. Insufficient Manifold Assumptions: Manifold priors alone without additional geometric covering structures are inadequate
2. Lack of Structured Compression Theory: Existing methods lack a unified theoretical framework connecting symmetry, compression, and hierarchical learning
3. Lack of Mathematical Foundation for Predictive Coding: Traditional predictive coding lacks rigorous mathematical formalization

Core Contributions

1. Proposes a generative model framework based on Lie pseudogroups: Defines generative models as local actions of finite-parameter Lie pseudogroups on configuration manifolds
2. Establishes world tracking dynamics theory with symmetry constraints: Proves that accurate tracking requires equivariance constraints and Noether-type conserved quantities
3. Constructs geometric theory of hierarchical dimensionality reduction: Establishes hierarchical structures of nested invariant manifolds through pseudogroup compositional decomposition
4. Provides symmetry-based predictive coding implementation: Formalizes hierarchical predictive processing where higher layers receive only coarse-grained residual transformations
5. Connects Spencer formalism theory: Links the Spencer complex of Lie pseudogroups with agent hierarchical structures

Methodology Details

Task Definition

The core task studied in this paper is constructing algorithmic agents capable of tracking and compressing sensory data streams with compositional symmetries. The input is a data stream generated by Lie pseudogroups, and the output is the agent's internal state representation and world tracking performance.

Theoretical Framework

1. Generative Model Definition

Definition 1 (Generative Model): A generative model is a smooth mapping from an M-dimensional configuration manifold C to observation space R^X:

f: C → R^X, I = f(c)

Definition 2 (Lie Generative Model): If there exists a Lie pseudogroup G acting on C and R^X such that for any c ∈ C, there exists γ ∈ G satisfying:

c = γ·c₀, f(c) = γ·I₀

then f is called a Lie generative model.

2. World Tracking Dynamics

The agent's high-dimensional state x ∈ R^X follows the neural network equation:

ẋ = F(x; w, I_θ(t))  (2)

The world tracking constraint is:

p(x(t)) ≈ I_θ(t)  (3)

3. Equivariance Requirements

Effective tracking requires internal dynamics to respect the same group action:

∀γ ∈ G: f(γ·x; w, γ·I_θ) = γ·f(x; w, I_θ)
p(γ·x) = γ·p(x)  (4)

Technical Innovations

1. Recursive Structure of Compositional Symmetry

Using the exponential map of Lie pseudogroups, complex transformations can be decomposed as:

γ = exp(∑ᵣₖ₌₁ θₖTᵏ)

This provides recursive compositional parameterization enabling structured compression.

2. Noether-type Conserved Quantities

Under static inputs, equivariance leads to readout invariance: p(x) = const. Each readout channel defines a conserved quantity, restricting trajectories to (X-Y)-dimensional phase space leaves.

3. Hierarchical Coarse-graining

Through pseudogroup flags:

G = H₀ ⊃ H₁ ⊃ ... ⊃ H_L

Construct nested reduced manifolds:

M₀ ⊃ M₁ := M₀/H₁ ⊃ ... ⊃ M_L

4. Predictive Hierarchy Implementation

Each layer k predicts Îₖ = γ̂ₖ·I₀, computing residuals:

rₖ := γ̂ₖ⁻¹·I_θ(t) - I₀  (8)

Applying coarse-graining operators:

mₖ→ₖ₊₁ := Cₖ→ₖ₊₁(rₖ)  (9)

Experimental Setup

Proof of Concept: Blender Cat Model

The paper provides a concrete implementation example in the appendix, using Blender's cat character rig as a practical application of Lie pseudogroup hierarchical structure:

Hierarchical Structure Mapping

1. Level 1: Camera and lens - SE(3) × R
2. Level 2: Global body/root - SE(3)
3. Level 3: Torso/spine chain - R^n_spine
4. Level 4: Limbs/claws/tail - R^n_limb
5. Level 5: Facial morphology - R^d_face
6. Level 6: Appearance/fur/material - R^d_mat
7. Level 7: Lighting and environment - SE(3) × R^d_SH

Compositional Action Implementation

Using the Product of Exponentials (PoE) model:

T(θ) = (∏ₙ∈chain e^[Sₙ]θₙ) M

Experimental Results

Theoretical Validation

1. Equivariance Constraints: Proves that tracking constraints and invariant compatibility require equivariance
2. Conservation Laws: Under static inputs, each readout channel defines a conserved quantity
3. Dimensionality Reduction Constraints: Trajectories are restricted to low-dimensional invariant leaves
4. Hierarchical Compatibility: Spencer complex ensures integrability of hierarchical constraints

Conceptual Implementation

The Blender example demonstrates:

• Practical implementation of local group decomposition γ = γ^(7)γ^(6)...γ^(1)
• Geometric meaning of nested quotient spaces Mₖ = Mₖ₋₁/Hₖ
• Propagation mechanism of prediction residuals in quotient directions

Related Work

Symmetry and Deep Learning

• Group Equivariant Networks: The paper's equivariance constraints are spiritually similar to group equivariant CNNs
• Invariance Learning: Lie group learning for visual invariance by Miao & Rao (2007) and others
• Symmetry Discovery: Symmetry inference methods by Moskalev et al. (2022) and others

Manifold Learning and Compression

• Manifold Hypothesis: Extends traditional manifold hypothesis with geometric covering structures
• Hierarchical Representation: Related to hierarchical representation learning in deep models
• Algorithmic Information Theory: Compression theory based on Kolmogorov complexity

Predictive Coding

• Traditional Predictive Coding: Predictive processing theory by Friston (2018) and others
• Hierarchical Prediction: The paper provides mathematical formalization based on symmetry

Conclusions and Discussion

Main Conclusions

1. Symmetry as Compression: Compositional symmetry provides a natural structured compression mechanism for natural data
2. Necessity of Equivariance: Accurate world tracking requires equivariance of agent dynamics
3. Hierarchical Geometry: Compositional decomposition of Lie pseudogroups naturally leads to nested reduced manifolds
4. Geometric Foundation for Predictive Coding: Provides a rigorous mathematical framework for predictive coding based on residual transformations

Limitations

1. Local Assumptions: All constructions are local; global statements require additional compatibility conditions
2. Complex Latent Spaces: May fail when the latent space of the generative model is very complex
3. Practical Implementation Challenges: Gap exists between theory and practical neural network implementation

Future Directions

1. Stochastic Input Generalization: Extend to stochastic inputs and analyze robustness
2. Lyapunov Operator Development: Develop effective K operators for world tracking problems
3. Empirical Validation: Test equivariant architectures under controlled symmetric generation
4. Spencer Exactness: Establish formal connections with Spencer exactness, moduli spaces, and integrability guarantees for practical learning systems

In-Depth Evaluation

Strengths

1. Theoretical Innovation: Innovatively combines Lie pseudogroup theory with algorithmic agent theory
2. Mathematical Rigor: Provides rigorous mathematical formalization connecting multiple mathematical fields
3. Unification: Unifies compression, symmetry, and hierarchical learning under a single framework
4. Practical Guidance: Provides theoretical guidance for equivariant network design
5. Interdisciplinary Value: Connects mathematics, machine learning, neuroscience, and other fields

Weaknesses

1. Insufficient Experimental Validation: Primarily theoretical work lacking sufficient experimental verification
2. Complexity: Mathematical formalization is complex, potentially limiting practical application
3. Assumption Limitations: Depends on the assumption that data is indeed generated by Lie pseudogroups
4. Missing Implementation Details: Insufficient detail in translating theory to practical algorithms

Impact

1. Theoretical Contribution: Provides new perspective on mathematical foundations of deep learning
2. Methodological Value: Guides design of symmetry-aware neural architectures
3. Cross-disciplinary Impact: May influence computational neuroscience, robotics, and other fields
4. Long-term Value: Established theoretical framework has long-term research value

Applicable Scenarios

1. Domains with Clear Symmetries: Such as robotics and geometric transformations in computer vision
2. Hierarchical Data: Data types with natural hierarchical structures
3. Compression Tasks: Applications requiring structured compression
4. Predictive Coding Systems: Predictive coding implementations requiring theoretical foundation

References

The paper cites abundant related work, including:

• Cover & Thomas (2006): Foundations of algorithmic information theory
• Goldschmidt (1967), Seiler (2010): Spencer theory of Lie pseudogroups
• Poggio et al. (2016, 2020): Compositionality theory in deep learning
• Friston (2018): Predictive coding theory
• Lynch & Park (2017): Lie group methods in modern robotics

Overall Assessment: This is a highly theoretical work attempting to establish a mathematical theoretical framework based on Lie pseudogroups for algorithmic agents. While the mathematical formalization is rigorous and innovative, it requires more experimental validation to demonstrate practical value. This work provides new mathematical tools for understanding symmetry and hierarchical structure in deep learning, with significant theoretical importance.