A brain region can contain ten thousand neurons without using ten thousand independent directions of activity.
That is the dimensionality problem.
Imagine recording hundreds of neurons at once.
Every neuron gives you another coordinate.
If you record 300 neurons, the raw population activity lives in a 300-dimensional measurement space.
But those neurons do not usually vary independently.
Some rise together.
Some oppose one another.
Some combinations barely occur at all.
The activity may therefore occupy far fewer effective degrees of freedom than the number of neurons suggests.
Neural dimensionality asks how many independent collective patterns the population actually uses.
Quick Read
- Neuron count: how many neurons are measured.
- Dimension: one axis along which population activity can vary.
- Neural dimensionality: how many independent or effectively independent population activity directions are substantially explored.
- Manifold: the structured shape formed by the population states that are possible or occupied.
- Dimensionality reduction: a method for representing high-dimensional recordings with fewer coordinates.
- Participation ratio: one common effective-dimensionality measure based on how variance is distributed across population modes.
- Relative dimensionality: dimensionality considered relative to population size.
One-sentence answer: Neural dimensionality is the effective number of independent collective degrees of freedom explored by a neural population, which can be far smaller than the number of recorded or existing neurons.
Canonical Job
Neural Dimensionality owns one reader job in Cognitive Art:
How many independent directions of population activity does a neural circuit actually use, and what controls that number?
This question has become newly important because modern neuroscience can record from hundreds or thousands of neurons at once.
On 19 August 2026, Dahmen and colleagues published Strong and localized recurrence controls the dimensionality of neural activity across brain areas in Nature Neuroscience. The study analyses large-scale cortical recordings and develops a circuit-level account in which recurrent connectivity and specific network motifs help regulate the dimensionality of population activity.
The important shift is conceptual.
Dimensionality is not merely a plotting convenience.
It can be a signature of how the circuit is organised and what computations it is positioned to perform.
Neuron Count Is Not Dimensionality
Suppose you record 1,000 neurons.
The raw state vector has 1,000 coordinates.
That does not mean the neural activity uses 1,000 independent degrees of freedom.
If all 1,000 neurons rise and fall together in nearly the same proportion, the population may behave almost like one collective variable.
If they divide into ten strongly coordinated groups that vary somewhat independently, the effective dimensionality might be closer to ten.
If every neuron fluctuates independently, dimensionality can approach the population size.
The central distinction is therefore:
number of measured coordinates ≠ number of independently used coordinates.
Neural Dimensionality Is Not Dimension
The Cognitive Art article What Is a Dimension? owns the general idea of an axis along which things can differ.
A neural dimension is one weighted combination of neuronal activity.
For population vector r, one dimension v produces one coordinate:
a = vᵀr.
Dimensionality asks how many such independent directions are needed to capture the population activity at the resolution relevant to the analysis.
Dimension is one axis.
Dimensionality is the effective count of axes in use.
Neural Dimensionality Is Not Manifold
The Cognitive Art article What Is a Manifold? owns the structured shape occupied by possible neural states.
Dimensionality and manifold are closely related but not identical.
A circle is a one-dimensional manifold embedded in a two-dimensional plane.
A curved sheet is a two-dimensional manifold embedded in three dimensions.
Likewise, a neural population can live in a very high-dimensional measurement space while its states lie near a much lower-dimensional manifold.
The 2025 Nature Neuroscience review A neural manifold view of the brain describes neural manifolds as the low-dimensional collective states available to a population under intrinsic and extrinsic constraints.
The distinction is useful:
dimensionality asks how many directions; manifold asks what shape those directions form.
Neural Dimensionality Is Not Dimensionality Reduction
This is another common collapse.
Dimensionality is a property we try to estimate.
Dimensionality reduction is a family of methods used to analyse or represent data.
Principal component analysis, factor analysis, Gaussian-process factor analysis, demixed PCA and nonlinear manifold methods all reduce description complexity in different ways.
The classic Nature Neuroscience review Dimensionality reduction for large-scale neural recordings by Cunningham and Yu explains why population studies need such tools and why method choice must follow the scientific question.
A method that displays three components does not prove the brain has dimensionality three.
The analysis may have chosen three for visualisation.
The true effective dimensionality may be much larger.
The Simplest Intuition: Independent Knobs
Imagine a machine with 100 gauges but only three independent control knobs.
Turning knob 1 moves forty gauges.
Turning knob 2 moves another overlapping set.
Turning knob 3 shifts nearly all of them a little.
You can observe 100 numbers.
But only three independent latent controls are needed to generate most of the variation.
That machine has many measured variables but low effective dimensionality.
Neural populations can behave similarly.
Covariance Creates Collective Dimensions
If neurons varied independently, each would contribute a nearly separate degree of freedom.
But neurons covary.
Some pairs tend to rise together.
Some tend to move oppositely.
Many are connected directly or indirectly through recurrent networks.
These relationships concentrate variance into collective modes.
Instead of needing one independent axis per neuron, a smaller set of population directions can explain substantial activity.
Principal Components and the Variance Spectrum
Principal component analysis provides one useful way to see this.
The first principal component captures the direction of greatest variance.
The second captures the next greatest orthogonal variance.
Continue until all variance is accounted for.
If the first few components contain most of the variance, the activity looks low-dimensional.
If variance is spread across many components, the activity looks higher-dimensional.
But the phrase “most variance” requires a criterion.
90%?
95%?
99%?
Different thresholds produce different counts.
That is why neural dimensionality should not be treated as one method-free integer written directly into nature.
The Participation Ratio
One widely used measure avoids choosing a hard explained-variance cutoff.
Let λ₁, λ₂, … be the variance captured by the principal components.
The participation ratio is:
DPR = (Σλᵢ)² / Σλᵢ².
If ten components contribute equally and everything else contributes almost nothing, the participation ratio is close to ten.
If one component dominates, it approaches one.
If variance is spread broadly across many modes, the value increases.
Dahmen and colleagues used the participation ratio as a principal dimensionality measure in their 2026 analysis while also checking related metrics.
Effective Dimensionality Is Metric-Dependent
There is no single dimensionality measure that answers every scientific question.
Possible definitions include:
- participation ratio,
- number of principal components above a variance threshold,
- number of statistically significant modes,
- intrinsic manifold dimension,
- predictive latent dimension,
- task-relevant dimension.
These can disagree without one necessarily being wrong.
Each asks a slightly different question.
Dimensionality is always dimensionality under a declared definition, timescale and dataset.
Absolute Versus Relative Dimensionality
A population can be low-dimensional relative to neuron count while still being high-dimensional in absolute terms.
This is an important boundary in the 2026 Nature Neuroscience paper.
If a cortical network contains 10,000 neurons but effectively explores hundreds of dimensions, dimensionality is strongly constrained relative to population size.
But hundreds of dimensions is not “simple.”
Low relative dimensionality can still support a large computational repertoire.
High Dimensionality Is Not Automatically Better
High dimensionality can separate many different conditions.
That can help a simple readout classify complex combinations.
The classic Nature study The importance of mixed selectivity in complex cognitive tasks showed how nonlinear mixed selectivity can create high-dimensional population representations that support a rich repertoire of linear readouts.
The Cognitive Art article What Is Mixed Selectivity? owns that mechanism.
But more dimensions can also make learning noisier, require more data, reduce smoothness or harm generalisation.
More is not automatically better.
Low Dimensionality Is Not Automatically Better
Low-dimensional activity can be compact, stable and easier to generalise from.
But if too many task distinctions collapse into the same small space, the system may lose flexibility.
A two-dimensional code cannot linearly separate every possible combination that a hundred-dimensional code can.
The right dimensionality therefore depends on the computation.
Dimensionality is a resource with costs and benefits, not a score where larger always wins.
The Classification–Generalisation Trade-Off
The 2026 Nature Neuroscience study makes this trade-off explicit.
Higher-dimensional activity can make different inputs more separable.
Lower-dimensional compression can help related inputs share structure and generalise.
This means neural computation may benefit from changing dimensionality depending on the job.
Exploration may benefit from expansion.
Abstraction may benefit from compression.
Stable action may benefit from a narrow output subspace.
High-Dimensional Visual Coding
The visual cortex provides a useful counterexample to any claim that population activity is always strongly low-dimensional in an absolute sense.
In the 2019 Nature paper High-dimensional geometry of population responses in visual cortex, Stringer and colleagues analysed responses of very large neuronal populations to natural images and found high-dimensional population structure with a characteristic power-law eigenspectrum.
The work also connected dimensionality to smoothness: a representation can remain high-dimensional while still being constrained enough that small input changes do not produce arbitrarily large representational jumps.
This matters because “high-dimensional” does not mean “random.”
Structure can extend across many dimensions.
The 2026 Mechanism: Strong Recurrence
The new step in 2026 is a mechanistic bridge from observed dimensionality to circuit organisation.
Dahmen and colleagues argue that strongly recurrent cortical networks can constrain relative dimensionality.
The logic begins with covariance.
In a recurrent network, activity from one neuron can propagate through many paths and return through the network.
Different neuron pairs share different collections of those paths.
This creates a broad distribution of pairwise covariances.
Those covariance patterns concentrate neural activity into collective modes and thereby constrain dimensionality.
In the study’s theoretical account, stronger recurrency makes dimensionality especially sensitive to network structure.
Recurrence Is Not a Magic Low-Dimensionality Switch
It would be too strong to conclude:
recurrent networks are low-dimensional.
Dimensionality depends on connectivity strength, network motifs, cell types, state, inputs and timescale.
The 2026 result is more precise:
in the cortical regimes studied, recurrent connectivity provides a tractable mechanism capable of regulating the effective dimensionality of activity.
Network Motifs Can Tune Dimensionality
A network motif is a small recurring connectivity pattern.
Examples include specific arrangements among pairs or triplets of cells.
Dahmen and colleagues analysed synaptic physiology data covering more than 32,000 neuron pairs and found motifs implicated by their theory in both mouse and human datasets.
This matters because it connects a macroscopic population property to small-scale circuit statistics.
The chain becomes:
local connectivity → covariance structure → population dimensionality.
This is not yet a universal law for every brain circuit.
It is a concrete, testable bridge.
Dimensionality Can Change With Brain State
A major reason not to assign one permanent dimensionality to a brain region is that population statistics change over time.
The 2026 cortical analysis identified distinct neural states and found dimensionality changes across them.
Behaviour such as running was associated with state transitions in the recordings.
This makes dimensionality dynamic.
A population can occupy one regime with more effective degrees of freedom and another with fewer.
The correct question is often not:
what is the dimensionality of area X?
but:
what is its dimensionality during this state, task, timescale and analysis?
Mixing States Can Artificially Inflate Dimensionality
Suppose a population occupies one low-dimensional pattern while the animal rests and another low-dimensional pattern while it runs.
Combine both conditions into one covariance matrix.
The mixture can look more dimensional than either state alone.
This is why stationarity and state segmentation matter.
Dimensionality can describe the structure of the data mixture rather than the structure of one stable computation if conditions are collapsed carelessly.
Sampling Changes the Estimate
Record 20 neurons and estimate dimensionality.
Then record 2,000.
The answer may change.
A dimensionality estimate depends on how many neurons, trials and time points are available.
Finite samples can create both underestimation and unstable overfitting.
The 2026 study explicitly examined how dimensionality scales with population size and developed extrapolation procedures rather than assuming small recordings reveal the full network directly.
Dimensionality and Population Code
The Cognitive Art article What Is a Population Code? owns information represented by joint activity across many neurons.
Dimensionality describes the degrees of freedom available to that population code.
A population code can be high-dimensional or low-dimensional.
Neither adjective tells you by itself what variable is encoded.
Dimensionality measures representational capacity and structure, not semantic content.
Dimensionality and Mixed Selectivity
Mixed selectivity can increase dimensionality by making neurons respond to diverse nonlinear combinations of task variables.
This can expand the number of directions available to downstream linear readouts.
The original Rigotti study linked high-dimensional prefrontal representations to correct task performance and found dimensionality collapse on error trials.
But high dimensionality can arise from mechanisms other than mixed selectivity.
Dimensionality is the property.
Mixed selectivity is one mechanism that can generate it.
Dimensionality and Neural Readout
The Cognitive Art article What Is a Neural Readout? owns the transformation from population activity into downstream consequence.
A high-dimensional source population can feed a one-dimensional readout.
For example, hundreds of population degrees of freedom may exist while one motor command depends on one weighted projection.
Most source dimensions can therefore be output-null for that receiver.
Source dimensionality and output dimensionality are different quantities.
Dimensionality and Communication Subspace
The Cognitive Art article What Is a Communication Subspace? owns the lower-dimensional directions through which population activity in one region relates to another.
The source area may have dimensionality 100.
The communication subspace to one target may use five dimensions.
Another target may align with a different seven-dimensional subspace.
This yields a useful hierarchy:
population dimensionality describes the room; communication subspace describes one doorway through it.
Dimensionality and Temporal Code
Dimensionality also depends on how activity is represented through time.
If each neuron contributes one firing-rate value per trial, the representation may have one dimensionality.
If the same trial is represented with fine time bins, latency and temporal patterns create additional coordinates and potentially additional effective modes.
The Cognitive Art article What Is a Temporal Code? owns whether those timing distinctions carry information.
Neural Dimensionality owns how many independent modes the chosen representation actually explores.
Dimensionality and Timescale
Bin spikes at 5 ms.
Then repeat at 100 ms.
Fast patterns that were distinct at 5 ms may be merged at 100 ms.
Dimensionality can therefore change with temporal resolution.
Any serious dimensionality claim should state its timescale.
Dimensionality and Learning
Learning can expand, compress or rotate neural representations.
Early learning may recruit many degrees of freedom while the system searches for a workable solution.
Practice can sometimes produce more compact task-aligned activity.
Other tasks may require expansion because the learner must separate previously conflated conditions.
There is no universal rule that expertise must reduce dimensionality.
The direction depends on task demands, measurement scale and circuit.
Dimensionality and Robustness
A representation spread across many fragile dimensions may be expressive but sensitive to noise.
A strongly compressed representation may be robust but less flexible.
The Cognitive Art article What Is Robustness? owns performance under perturbation.
Neural Dimensionality contributes one structural variable that can affect that robustness.
Dimensionality and Generalisation
Suppose every training example gets its own almost orthogonal direction.
The system can distinguish the training set extremely well.
But it may fail to identify what novel examples share.
Compression forces examples to share structure.
That can improve generalisation when the compressed dimensions correspond to meaningful invariants.
Again, too much compression destroys distinctions.
The useful dimensionality is task-dependent.
Dimensionality and Smoothness
High-dimensional representations can still be smooth.
The 2019 visual-cortex study showed that population response variance followed a power-law spectrum compatible with a smooth mapping from natural images into a high-dimensional neural representation.
This is important because dimensionality alone does not tell us whether a representation is well behaved.
Two representations can have similar dimensionality but different geometry, smoothness and readout properties.
Dimensionality and Causality
Dimensionality is often estimated observationally.
That tells us how population activity is structured.
It does not by itself tell us why.
The 2026 recurrence paper is valuable because it proposes a mechanistic link from circuit connectivity to dimensionality and tests that link across recordings, network theory and synaptic physiology datasets.
Stronger causal tests would manipulate connectivity or effective recurrency and test whether dimensionality changes as predicted while competing variables are controlled.
The Dimensionality Expansion Test
Suppose task complexity increases.
If the circuit needs to represent more independent combinations, dimensionality may increase.
A useful analysis asks:
- Does dimensionality increase on held-out data?
- Are the new dimensions task-related rather than measurement noise?
- Do readouts gain useful separability?
- Does behaviour improve with the expansion?
Expansion becomes meaningful when it buys computation.
The Dimensionality Compression Test
Suppose training reduces dimensionality.
That could mean several things.
- the system found a more efficient task-aligned representation,
- irrelevant variability disappeared,
- different conditions collapsed together,
- the network became less expressive.
Dimensionality decrease is therefore not automatically improvement.
Inspect what information survives.
The Subsampling Test
Estimate dimensionality using 50 neurons.
Then 100.
Then 200.
If the estimate keeps rising rapidly, the recording has not revealed the saturation regime.
If it stabilises, confidence improves.
This is one reason large-scale datasets matter.
The Null-Model Test
A low-dimensional pattern can emerge from trivial shared drive.
For example, every neuron increases when the animal starts running.
That creates one strong common dimension.
Before calling the dimension computationally meaningful, compare against models that account for:
- movement,
- arousal,
- global input,
- recording drift,
- trial timing.
Dimensional structure is real as a statistic even when its interpretation is wrong.
The State-Specific Test
Compute dimensionality separately in distinct neural or behavioural states.
Then compare.
If dimensionality changes, ask whether the change is explained by:
- different cell participation,
- different covariance structure,
- different inputs,
- different effective recurrence,
- different task variables.
This turns “dimensionality changed” into a mechanistic question.
The Neural Dimensionality Audit
- Which population is being measured?
- How many neurons were recorded?
- What time bin or temporal representation was used?
- Which dimensionality metric was used?
- Was dimensionality computed within a stable state or across mixed states?
- How does the estimate scale with neuron count and trial count?
- Is the claim absolute dimensionality or relative dimensionality?
- What information is carried by the major dimensions?
- Does higher dimensionality improve useful separability?
- Does lower dimensionality preserve generalisation?
- What circuit mechanism could generate the covariance structure?
- Does perturbation support the proposed mechanism?
Failure 1: Neuron Count Equals Dimensionality
Five hundred neurons are recorded, so the representation is described as 500-dimensional.
Repair: measure the effective independent modes.
Failure 2: Three-Dimensional Plot Equals Three-Dimensional Brain State
A PCA figure shows three axes and the representation is called three-dimensional.
Repair: distinguish visualisation dimension from estimated activity dimension.
Failure 3: High Dimensionality Equals Better Coding
A higher participation ratio is automatically celebrated.
Repair: ask what computation the extra dimensions enable and what they cost.
Failure 4: Low Dimensionality Equals Simplicity
A population with hundreds of effective dimensions is called simple because it is low-dimensional relative to ten thousand neurons.
Repair: report both absolute and relative scale.
Failure 5: One Metric Becomes the Definition
Participation ratio is treated as the one true dimensionality.
Repair: state the metric and compare alternatives where conclusions depend on it.
Failure 6: Behavioural States Are Mixed Together
Rest, movement and task epochs are pooled and one dimensionality value is reported.
Repair: estimate state-specific structure before combining.
Failure 7: Low-Rank Fit Becomes Biological Mechanism
A low-dimensional latent model fits well and is treated as proof that the circuit literally contains the inferred latent variables.
Repair: separate useful description from identified mechanism.
Failure 8: Correlation Structure Is Interpreted Without Circuit Evidence
Dimensionality is low, so strong recurrence is assumed automatically.
Repair: compare alternative shared-input and connectivity explanations, then test circuit predictions.
Repair Path
- Define the population, condition and timescale.
- Inspect firing-rate and covariance structure.
- Choose a dimensionality measure that matches the question.
- Cross-check with at least one alternative measure when conclusions are sensitive.
- Subsample neurons and trials to test estimation stability.
- Separate behavioural or neural states before pooling them.
- Map task variables onto the major population modes.
- Test whether dimensionality changes improve separability, compression or readout.
- Compare circuit mechanisms that could produce the observed covariance.
- Use perturbation or independent physiological evidence to strengthen causal claims.
Neural Dimensionality in Education: Analogy Only
A student can make ten different mistakes.
But those ten errors may come from only two underlying causes:
- weak fraction sense,
- poor equation control.
Ten visible coordinates.
Two latent degrees of freedom.
This is not a claim that learning performance literally shares neural dimensionality metrics.
It is a systems analogy:
many observed variables can be generated by a much smaller set of independent underlying causes.
Neural Dimensionality in Organisations
A company tracks 400 metrics.
After analysis, most move with four underlying factors:
- demand,
- capacity,
- quality,
- cash.
The organisation has hundreds of measurements but a much smaller effective operating dimension at that moment.
Again, the analogy is not neuroscience.
The transferable principle is:
the number of things you measure can be very different from the number of independent things the system is actually doing.
Research Notes and Further Reading
For the newest mechanistic account linking cortical dimensionality to strong recurrence, covariance structure and network motifs, see Dahmen and colleagues, Strong and localized recurrence controls the dimensionality of neural activity across brain areas (Nature Neuroscience, 2026).
For the modern population-manifold framework and its relation to behaviour, see Perich, Narain and Gallego, A neural manifold view of the brain (Nature Neuroscience, 2025).
For practical foundations of population dimensionality reduction, see Cunningham and Yu, Dimensionality reduction for large-scale neural recordings (Nature Neuroscience, 2014).
For high-dimensional natural-image responses and the relationship between dimensionality and smooth representational geometry, see Stringer and colleagues, High-dimensional geometry of population responses in visual cortex (Nature, 2019).
For high-dimensional mixed selectivity and flexible linear readout, see Rigotti and colleagues, The importance of mixed selectivity in complex cognitive tasks (Nature, 2013).
World Return
A neural-dimensionality account earns trust when its estimate is stable across held-out samples, robust to reasonable metric choices, appropriate to the population size and timescale, and linked to specific information or computation rather than reported as an isolated number.
A mechanistic account becomes stronger when circuit variables such as recurrent connectivity predict dimensionality changes and independent perturbation or physiological evidence supports the proposed link.
If dimensionality changes disappear after behavioural state, sampling or nonstationarity is controlled, the stronger computational interpretation must weaken.
Final Thought: Complexity Is Not the Number of Neurons
The nervous system contains enormous numbers of neurons.
But coordinated populations can move together through a much smaller set of collective possibilities.
Sometimes the system expands those possibilities to separate many conditions.
Sometimes it compresses them to generalise, stabilise or communicate efficiently.
The important quantity is therefore not how many cells exist.
It is how many independent collective directions the circuit actually explores and exploits.
Neural dimensionality is the count of freedom inside the crowd.