Structured Dynamics in the Algorithmic Agent

Symmetry in the world forces structure in the mind — this paper derives why neural dynamics live on low-dimensional manifolds from first principles, not empirical observation.

The core intuition is simple: if you want to compress data, you have to exploit its regularities. And regularities, mathematically, are symmetries. A dataset of cat images is compressible because any cat image can be obtained from a reference cat by applying rotations, translations, and scale changes — a Lie group (a smooth, continuous family of transformations). The paper formalizes this by defining a Lie generative model: a generative model whose outputs are related to each other by group transformations. The compression lives in the group parameters, not the raw pixels.

The key move is then to ask: what happens to an agent that successfully tracks inputs from such a model? The agent's internal state must shadow the world — its predictions must match incoming data, driving prediction error toward zero. This world-tracking constraint is not just a loss function; it's a structural constraint on the agent's dynamics. The paper proves that when the world is generated by an r-parameter Lie group, the agent's state-space trajectories are forced onto a reduced manifold and must obey conserved quantities — one per Lie algebra generator. This is Noether's theorem applied to cognition: every symmetry in the world the agent tracks becomes a conservation law in the agent's dynamics. The manifold hypothesis — the empirical observation that neural activity lives in far lower dimensions than the number of neurons — is thus not a lucky accident but a mathematical necessity for any agent doing its job.

The hierarchical extension is elegant. If the world has compositional structure (objects made of parts, scenes made of objects), the agent's state space develops nested manifolds: each level of the hierarchy introduces additional constraints that carve out a lower-dimensional subspace inside the previous one. Recognizing "white cat with blue eyes" is a sequence of manifold refinements, each level more specific than the last. This mirrors what we see in visual cortex, from V1 to IT, though the paper is careful to note the neuroscience connections are conceptual rather than quantitatively validated.

The paper is primarily theoretical and honest about its gaps. The biggest open problem it identifies — and does not solve — is learning: the whole framework assumes the agent already possesses a world model with the right group structure. How an agent discovers that structure from raw data is left explicitly open. The approximation regime (real agents only approximately track the world) is handled via Lyapunov stability arguments but without tight quantitative bounds on how much manifold structure degrades with tracking error. Empirical validation is indirect, pointing to existing neuroscience literature on neural manifolds rather than designing experiments to test the specific conservation-law predictions the theory makes.