A system can have a huge state space and still spend almost all its life on a much smaller shape inside it.
Imagine a sheet of paper floating inside a room.
The room is three-dimensional.
The sheet is two-dimensional.
A point constrained to the sheet can move in the room, but not everywhere in the room.
Its possible motion is restricted to a lower-dimensional surface.
That intuition leads us toward the idea of a manifold.
Quick Route
- State space: all represented configurations available to the model.
- Manifold: a lower-dimensional structured subset on which the system actually tends to live or move.
- Trajectory: the path followed across that manifold or state space.
- Vector field: the local rule indicating which direction the state tends to move next.
- Attractor: a state or set toward which trajectories tend to converge.
Canonical Job
Manifold owns one reader job in Cognitive Art:
Why can a system with many possible variables occupy a much smaller, structured family of states?
In mathematics, a manifold is a space that can look locally like ordinary Euclidean space while possessing richer global geometry.
In modern neuroscience, the word is often used for low-dimensional structure embedded in high-dimensional neural population activity.
A 2025 Nature Neuroscience review, A Neural Manifold View of the Brain, describes neural manifolds as mathematical descriptions of the possible collective states of neuronal populations under intrinsic and behavioural constraints. It also emphasises open questions about how these structures arise and how directly they correspond to biological mechanism.
A 2025 Nature Reviews Neuroscience article, Neural Manifolds: More Than the Sum of Their Neurons, similarly frames the field around the question of whether manifold structure is merely useful analysis or something more fundamental about circuit function.
One-sentence answer: A manifold is a lower-dimensional structured set embedded inside a larger state space, representing the configurations a system actually occupies under its constraints and dynamics.
Manifold Is Not State Space
The Cognitive Art article What Is a State Space? owns the possibility field.
Manifold owns the structured subset that the system actually uses.
State space:
every coordinate combination the representation permits.
Manifold:
the lower-dimensional region compatible with the system’s real constraints and coordinated activity.
A hundred-dimensional state space can contain a manifold of only three or four effective dimensions.
Manifold Is Not Dimension Reduction
Dimension reduction is a method.
A manifold is a hypothesised or inferred geometric structure.
Principal component analysis, UMAP, t-SNE and nonlinear manifold-learning methods can all produce lower-dimensional representations.
But the existence of a pretty low-dimensional embedding does not prove the underlying system truly lives on that exact manifold.
Method can reveal structure.
Method can also manufacture apparent structure.
Manifold Is Not Cluster
A cluster is a region where cases group densely.
A manifold can be continuous and extended.
Imagine points distributed along a curved ribbon.
There may be no separate clusters at all.
The important structure is the ribbon itself.
Manifold Is Not Attractor
An attractor is defined by return dynamics.
A manifold is geometric structure.
An attractor can itself be a manifold.
But not every manifold is attracting.
This boundary matters because modern neuroscience often studies both manifolds and attractor dynamics in the same dataset.
Why Low-Dimensional Structure Appears
Many variables do not vary independently.
Neurons are connected.
Muscles are coordinated.
Mechanical constraints couple joints.
Tasks demand organised combinations rather than arbitrary activation.
These dependencies collapse the effective degrees of freedom.
The system may contain thousands of microscopic variables while behaviour uses only a much smaller coordinated subspace.
Constraints Carve the Manifold
The Cognitive Art article What Is Constraint? owns limits on possibility.
Manifold shows what repeated constraints can look like geometrically.
Connectivity constrains neural co-activation.
Anatomy constrains movement.
Task rules constrain legal action.
Prior learning constrains which transitions are easy.
The occupied states form structure because the system cannot move arbitrarily.
Local and Global Geometry
A manifold can look simple locally and complex globally.
A globe looks flat if you inspect a small patch.
Travel far enough and curvature becomes unavoidable.
This idea matters in cognition because locally linear approximations can work beautifully while global structure remains nonlinear.
One neighbourhood may be well approximated by a plane.
The whole representational space may curve, fold or wrap.
The Nearest-Point Trap
Two points can be close in straight-line distance through the ambient space while far apart along the manifold.
Imagine two points on opposite folds of a curled sheet.
Spatially close.
Structurally distant along the sheet.
This is why distance depends on the geometry the system actually uses.
Manifold and Trajectory
The Cognitive Art article What Is a Trajectory? owns motion through state space.
If a system is constrained to a manifold, trajectories live mostly on that manifold.
This immediately changes prediction.
Instead of asking what happens in the entire ambient state space, we ask what movement is available on the occupied surface.
Manifold and Vector Field
A manifold gives the surface.
A vector field gives local direction and speed.
At each point on the manifold, the vector field tells us how the state tends to evolve next.
The next Cognitive Art article makes that local flow field explicit.
Manifold and Perturbation
The Cognitive Art article What Is a Perturbation? owns deliberate displacement.
Push activity along the manifold.
Push it off the manifold.
Compare the response.
If off-manifold states rapidly relax back while on-manifold states persist or evolve naturally, the manifold is doing more than decorating the plot.
Neural Manifolds
Neural manifold research begins with simultaneous recordings from many neurons.
Each moment produces a high-dimensional population state.
Across behaviour, those states often occupy a structured lower-dimensional region.
The 2023 Nature Reviews Neuroscience perspective A Unifying Perspective on Neural Manifolds and Circuits for Cognition argues that manifold descriptions become especially informative when they can be linked back to underlying connectivity and circuit mechanisms.
This is a crucial standard.
A low-dimensional shape is interesting. A low-dimensional shape tied to behaviour, perturbation and circuit mechanism is much stronger science.
Manifolds and Learning
Learning can reshape neural population geometry.
Some work suggests that skills are learned more easily when new activity patterns remain within or near pre-existing manifold constraints, while learning that requires movement outside familiar neural subspaces can be slower.
This does not mean every human learning difficulty is a manifold problem.
The evidence is strongest in specific motor and neural-control paradigms.
The Educational Analogy
A learner may have many theoretically possible strategies.
In practice, only a small family is reachable with current knowledge.
We can describe that as a manifold analogy:
the learner’s effective possibility space is smaller than the full mathematical possibility space.
But unless the dimensions and dynamics are explicitly modelled, this remains a design metaphor rather than a neuroscientific claim.
Manifold in Mathematics
Mathematics gives the term precise technical definitions.
A circle is a one-dimensional manifold embedded in two-dimensional space.
A sphere is a two-dimensional manifold embedded in three-dimensional space.
The local dimension describes how many independent directions of motion exist near a point.
Cognitive Art preserves this geometric intuition without replacing formal differential geometry.
Manifold in AI
Machine-learning representations are often analysed for low-dimensional structure.
Embeddings can contain semantic or task-related geometry.
But human neural manifolds and machine latent spaces should not be treated as identical objects simply because both admit low-dimensional analysis.
Same mathematics can describe different mechanisms.
Failure 1: Embedding Equals Manifold
A dimensionality-reduction plot looks curved, so the curve is declared the system’s true manifold.
Repair: test stability across methods, samples and perturbations.
Failure 2: Low Dimension Equals Simple System
The observed manifold is low-dimensional, so the underlying mechanism is assumed simple.
Repair: remember that complex circuits can generate low-dimensional coordinated activity.
Failure 3: Manifold Equals Cluster
Continuous structure is chopped into artificial groups.
Repair: inspect continuity before imposing categories.
Failure 4: Geometry Equals Mechanism
A low-dimensional shape is treated as the causal explanation for behaviour.
Repair: connect geometry to circuit constraints, intervention and prediction.
Failure 5: Universal Manifold Story
Every cognition problem is described as movement on a manifold.
Repair: use the framework only where state variables, dimensional structure and behavioural linkage are defensible.
Repair Path
- Define the full state representation.
- Estimate whether activity occupies a lower-dimensional region.
- Test local dimensionality and geometry.
- Compare reasonable embeddings.
- Relate manifold coordinates to behaviour.
- Perturb along and away from the manifold where possible.
- Connect geometry to mechanism.
- Revise the manifold model if the world does not return the predicted dynamics.
The Manifold Audit
- Manifold inside which state space?
- What defines its local dimension?
- Which constraints keep the system near it?
- Is the geometry stable across samples?
- Could the embedding method have created the shape?
- Which trajectories lie on it?
- What happens under off-manifold perturbation?
- Does the geometry predict behaviour?
- Can the manifold be linked to circuit mechanism?
- What evidence would falsify the manifold interpretation?
Research Notes and Further Reading
For a current broad review, see Perich, Narain and Gallego, A Neural Manifold View of the Brain (Nature Neuroscience, 2025).
For the biological-meaning debate, see Gallego, Neural Manifolds: More Than the Sum of Their Neurons (Nature Reviews Neuroscience, 2025).
For linking manifold geometry to circuits and computation, see Langdon, Genkin and Engel, A Unifying Perspective on Neural Manifolds and Circuits for Cognition (Nature Reviews Neuroscience, 2023).
World Return
A manifold model earns trust when it predicts reachable states, trajectories and responses to intervention better than a simpler rival.
If the supposed manifold disappears when the task changes slightly, the geometry may be too brittle to deserve canonical status.
Final Thought: Possibility Is Larger Than Practice
The full state space may be enormous.
The system uses only a thin structured part of it.
That thin part is where its habits, constraints, skills and dynamics become visible as geometry.
A manifold is the reminder that a system does not live everywhere it could theoretically exist. It lives where its structure allows it to move.