Metastable Neural Activity (MNA)

• Metastable Neural Activity is characterized by transient, stable epochs of neural firing patterns that persist long enough for information coding and transition abruptly.
• It serves as a computational framework to explain sensory coding, working memory, decision making, and event segmentation across diverse brain regions.
• Statistical tools like Hidden Markov Models uncover discrete state transitions in MNA, linking observed neural patterns to underlying dynamical systems.

Metastable neural activity (MNA) denotes neural population dynamics that unfold as a sequence of relatively stable but transient regimes. In cortical ensemble recordings, a metastable state can be “a vector of firing rates across simultaneously recorded neurons that can last for several hundred milliseconds before giving way to the next state in a sequence”; in dynamical-systems terms, metastability corresponds to trajectories spending long but finite times in almost-invariant regions of state space rather than remaining indefinitely in a strict attractor (Mazzucato et al., 2017, Rossi et al., 2023). Across gustatory, frontal, parietal, motor, hippocampal, and whole-brain measurements, and across clustered spiking networks, neural fields, and interacting point-process models, MNA has been used to explain sensory coding, expectation, working memory, replay, decision making, and naturalistic event segmentation (Camera et al., 2019, Brinkman et al., 2021, Gozukara et al., 29 May 2026).

1. Definition and conceptual boundaries

MNA is characterized by regime-like population activity: within a state, firing-rate statistics, spatial patterns, or phase relations are comparatively stable; across states, transitions are abrupt and usually much faster than dwell times. This makes metastable activity distinct from smoothly drifting trajectories and from purely featureless noise. In the broad formulation of dynamical-systems theory, metastable regimes are long-lived but transient, and they correspond to almost-invariant sets: regions of state space in which trajectories spend long times but from which they eventually escape with non-zero probability (Rossi et al., 2023).

This definition places MNA between classical attractor dynamics and random wandering. A stable attractor is invariant: trajectories in its basin converge to it and remain there. Multistability is the coexistence of several such attractors, but in the strict deterministic setting there are no transitions among them. By contrast, metastability consists of long-lived but transient visits to multiple regimes, with transitions generated by noise, deterministic repelling structures such as unstable manifolds or chaotic saddles, or slow parameter drifts. If noise becomes so strong that trajectories are dominated by random diffusion, coherent regimes disappear and metastability is lost (Rossi et al., 2023).

The neuroscience literature uses related but not identical vocabularies—metastable states, HMM states, microstates, attractor hopping, synchronization/desynchronization switching, or integration–segregation dynamics—but these can be unified under the same umbrella: recognizable epochs of neural activity produced by trajectories dwelling in almost-invariant regions. This suggests that MNA is best understood as a population-level organization principle rather than a phenomenon tied to one recording modality, one brain area, or one particular analysis method (Camera et al., 2019, Brinkman et al., 2021).

2. Dynamical formalisms and canonical mechanisms

A generic dynamical description writes neural activity as a state vector $x(t)\in\mathbb{R}^N$ evolving under deterministic flow, noise, and input,

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$

where $\theta$ denotes system parameters, $\xi(t)$ is typically white Gaussian noise, and $I(t)$ is external drive. In this setting, almost-invariant sets are formalized through a natural measure $\mu$ and a persistence probability

$$\rho(A) = \mathrm{Prob}(Tx \in A \mid x \in A) = \frac{\mu(T^{-1}A \cap A)}{\mu(A)},$$

with $\rho(A)$ close to 1 but not equal to 1 for a metastable region $A$ (Rossi et al., 2023).

Different dynamical mechanisms generate such regions. Noise-induced attractor hopping occurs when weak fluctuations drive rare escapes between otherwise stable wells. Stable heteroclinic cycles and heteroclinic networks produce deterministic sequential visits to saddle neighborhoods. Attractor-merging crisis, interior crisis, ghost dynamics near saddle-node bifurcations, chaotic saddles, on–off intermittency, in–out intermittency, and mixed-mode oscillations all produce long-lived but transient regimes with distinct dwell-time statistics. Depending on mechanism, dwell-time distributions may be approximately exponential, show power-law tails, or scale with distance to a bifurcation (Rossi et al., 2023).

A canonical cortical realization is the clustered spiking-network model developed for gustatory cortex. There, metastable states are attractor-like configurations of a balanced recurrent network with potentiated intra-cluster excitation and depressed inter-cluster excitation. In a mean-field reduction, the firing rate $r$ of a focal cluster obeys

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$0

with fixed points given by $\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$1. An effective potential

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$2

makes the attractor picture explicit: minima are metastable states, maxima are unstable saddles, and the barrier

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$3

controls transition rates. Higher $\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$4 implies rarer transitions; lower $\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$5 yields faster metastable dynamics (Mazzucato et al., 2017).

At a broader theoretical level, MNA is not equivalent to criticality or chaos. Metastability can occur near bifurcations and may exhibit criticality-like sensitivity, but it is defined by long-lived transients and finite escape, not by thermodynamic critical scaling. Likewise, chaos can generate metastable subregions, but metastability can also be noise-induced or heteroclinic without chaos (Rossi et al., 2023).

3. Statistical identification and empirical signatures

The principal statistical tool for extracting metastable states from spiking ensembles has been the Hidden Markov Model (HMM). In the standard formulation, each time bin has a latent state $\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$6 with Markov transitions

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$7

and conditionally independent Poisson spike-count emissions

$\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$8

Parameters $\dot{x} = f(x,\theta) + \sigma \,\xi(t) + I(t),$9 are estimated by Expectation–Maximization via the forward–backward and Baum–Welch algorithms, and the number of states is selected by information criteria such as BIC, by cross-validation, or with nonparametric Bayesian variants (Camera et al., 2019, Mazzucato et al., 2015).

Study-specific thresholds have been used to convert posterior state probabilities into discrete state sequences. In gustatory cortex, one analysis treated a state as active when its posterior probability exceeded 80% for at least 25 consecutive 2-ms bins, whereas another required posterior $\theta$0 for at least 50 consecutive ms. These thresholds operationalize the defining empirical signature of MNA: piecewise-stationary firing patterns with abrupt boundaries (Mazzucato et al., 2017, Mazzucato et al., 2015).

Empirically, MNA is recognized by several convergent signatures. Single-trial activity is segmented into discrete, quasi-stationary states; dwell times are much longer than transition times; the same states can recur during both ongoing and evoked activity; and single neurons often exhibit multiple distinct firing rates across ensemble states. In ongoing gustatory-cortex activity, 3–7 HMM states per ensemble were reported, with mean $\theta$1, state durations approximately exponential with mean $\theta$2 ms, and at least 42% of neurons showing 3 or more distinct firing rates across states. During evoked activity, the number of states increased to 4–11 per taste with mean $\theta$3, mean durations dropped to $\theta$4 ms, and only 8% of neurons had more than 2 distinct firing rates, indicating stimulus-induced quenching of single-neuron multistability and trial-by-trial variability (Mazzucato et al., 2015).

HMM-state inference is complemented by other segmentation methods. In naturalistic fMRI, Greedy State Boundary Search (GSBS) identifies boundaries that maximize within-state similarity and between-state dissimilarity of multivariate activity patterns. EEG/MEG microstate clustering, direct clustering in latent state space, change-point methods, and switching dynamical systems all recover related state structure. This convergence suggests that MNA is not an artifact of one statistical formalism but a robust empirical regularity across scales and modalities (Gozukara et al., 29 May 2026, Rossi et al., 2023).

4. Canonical case studies in cortex

A central case study is gustatory cortex (GC), where MNA was developed as both an empirical description and a mechanistic explanation. In the clustered leaky integrate-and-fire model of ongoing and evoked GC activity, excitatory neurons are partitioned into clusters with stronger intra-cluster connectivity and weaker inter-cluster coupling, while global inhibition stabilizes the network. Mean-field analysis yields many stable configurations with different numbers of active clusters, and finite-size fluctuations make the network hop among them. This architecture reproduces ongoing sequences of metastable states and explains why individual neurons can show three or more distinct firing rates across states: the firing rate of a cluster depends on how many clusters are co-active (Mazzucato et al., 2015).

Expectation provides a particularly explicit example of metastability-based computation. In a “general expectation” paradigm, an auditory cue predicts that some tastant will soon arrive but carries no information about which tastant. In the model, the anticipatory cue is implemented as a transient modulation of external currents to a random 50% subset of excitatory neurons, with zero mean and positive spatial variance. This increases heterogeneity without changing the mean afferent current. Mean-field analysis shows that the cue flattens the effective transfer function, reduces energy barriers between metastable states, shortens cluster lifetimes and inter-activation intervals, and thereby accelerates the intrinsic pre-stimulus state sequence. When taste input arrives, transitions into coding states occur sooner: in simulations, the latency for first activation of stimulus-selective clusters falls from $\theta$5 s in unexpected trials to $\theta$6 s in expected trials, and population decoding latency shifts from $\theta$7 s to $\theta$8 s; in GC data, the latency of the first taste-coding state shifts from $\theta$9 s to $\xi(t)$0 s (Mazzucato et al., 2017).

Motor cortex provides a complementary example in which metastable states organize behavior rather than sensory coding. In secondary motor cortex (M2) during a self-initiated waiting task, HMMs uncovered sequences of long-lived neural patterns preceding the action sequence Poke In $\xi(t)$1 Poke Out $\xi(t)$2 Water Poke In. Pattern dwell times had coefficient of variation $\xi(t)$3 and skewness $\xi(t)$4, matching the right-skewed timing variability of the behavior. A recurrent attractor network reciprocally coupled to a low-dimensional feedforward network reproduced both the reliable sequence order and the variable transition timing, with correlated variability in the feedback loop generating stochastic transitions and predicting a low-dimensional structure of noise correlations that was empirically verified in M2 (Recanatesi et al., 2020).

Beyond GC and M2, metastable dynamics have been reported in prefrontal cortex, visual area V4, orbitofrontal cortex, hippocampal CA1, and human fMRI. In these settings, states have been related to task difficulty, attentional performance, option values, spatial representations during sharp-wave ripples, and working-memory conditions. This suggests that MNA is not specific to one cortical system but recurs whenever population dynamics must combine discrete representation with variable timing (Camera et al., 2019, Brinkman et al., 2021).

5. Model classes beyond clustered cortical attractors

Clustered spiking networks are only one realization of MNA. A broad dynamical taxonomy includes noise-induced attractor hopping in multistable systems, stable heteroclinic cycles and networks, chaotic attractors with metastable subregions, interior crises, ghost dynamics near destroyed periodic orbits, chaotic saddles and transient chaos, synchronization intermittency, mixed-mode oscillations, and chaotic itinerancy. Each mechanism defines metastable regions differently and leaves distinct signatures in dwell-time distributions, transition asymmetries, and sensitivity to perturbations (Rossi et al., 2023).

Spatially extended neural-field models provide another route. In a stochastic ring model of working memory with a quantized staircase nonlinearity,

$\xi(t)$5

the staircase firing-rate function

$\xi(t)$6

creates multiple stable bump amplitudes separated by unstable ones. Noise drives rare transitions between these discrete amplitude levels, while bump position remains marginally stable and diffuses. The reduced amplitude dynamics form a one-dimensional multi-well potential, and phase variance scales as $\xi(t)$7, so larger bump amplitude implies less positional wandering and, in the paper’s interpretation, greater certainty (Cihak et al., 2024).

Metastability can also be derived bottom-up from stochastic spiking networks with synaptic fatigue. In a mesoscopic description of hippocampal replay derived from Linear-Nonlinear Poisson neurons with short-term synaptic depression, metastable replay events arise from the interplay of finite-size fluctuations and local fatigue. The resulting stochastic neural-mass equations reproduce population spikes, Up–Down states, and replay events, and in the replay regime they generate greater variability in event content, direction, and timing than the deterministic Romani–Tsodyks model (Pietras et al., 2022).

Rigorous probability theory has established metastability in interacting spiking systems with leakage and in piecewise-deterministic models of excitatory populations. In one class of interacting point processes, the normalized extinction time converges to a mean-one exponential random variable as the population size diverges,

$\xi(t)$8

showing a long-lived active regime followed by rare collapse to an absorbing silent state (Laxa, 2022). In another model with saturating spike rates and exponential leak, the rescaled exit time of the mean firing rate from a neighborhood of the non-linear equilibrium also converges to an exponential law as system size increases, providing a mathematically explicit metastable regime in which finite networks hover for exponentially long times near an active mean-field equilibrium before extinction (Löcherbach et al., 2020).

6. Functional roles, controversies, and current directions

MNA has been linked to a broad set of computational roles. The central recurring claim is that metastability provides a balance between stability and flexibility: states persist long enough to encode or compute, yet remain transient enough to permit switching. This supports flexible computation, transient coordination of large-scale networks, dynamic functional connectivity, structured trial-by-trial variability, and temporal coding in which the transitions themselves can be informative (Rossi et al., 2023, Camera et al., 2019).

Working memory has motivated several explicit metastable mechanisms. In one rate-model framework, a metastable active regime is created by placing an excitatory–inhibitory circuit just beyond a saddle-node bifurcation of an active state, leaving a long-lasting transient “ghost” of persistent activity. Gamma-band oscillations or noise can then stabilize that regime and support retention, with stronger effects for common noise and in-phase gamma when intercircuit coupling is fast (Novikov et al., 2021). In a multiple-timescale network with fast and slow populations, slow dynamics retain context and act as bifurcation parameters that sequentially stabilize and destabilize fast task-related states, enabling context-dependent working memories and non-Markovian transitions among metastable states (Kurikawa, 2021). A later extension formalized this principle as “stability control of metastable states,” showing that neuronal gain, external input strength, and task difficulty modulate dwell times and transition times through their effect on a stability factor $\xi(t)$9 (Kurikawa et al., 12 Apr 2025).

Conceptually, metastability has also been connected to event segmentation in naturalistic cognition. On this view, the “neural states” of event-segmentation research and the metastable states of dynamical-systems neuroscience are two perspectives on the same phenomenon: relatively stable patterns of activity nested across timescales, with boundaries marking brief windows of connectivity reconfiguration, predictive-model updating, and memory encoding. Longer-duration states in higher-order regions constrain and are shaped by faster states in sensory and motor regions, producing a spatio-temporally nested hierarchy of states (Gozukara et al., 29 May 2026).

Several debates recur in the literature. Metastability must be distinguished from multistability: multistability denotes multiple coexisting attractors with no deterministic transitions, whereas metastability requires long-lived but transient regimes. It must also be distinguished from criticality: metastability may occur near bifurcations and can display criticality-like sensitivity, but it does not require critical scaling or a special point between order and disorder. Likewise, chaos can generate metastable subregions, but metastability can also arise from noise-driven escape, heteroclinic dynamics, or slow parameter drift. The literature also distinguishes spontaneous metastability, driven by intrinsic dynamics, from driven metastability, induced by external inputs; and repeatable from non-repeatable metastability, depending on whether the same regimes are revisited (Rossi et al., 2023, Kelso, 2023).

Current directions include identifying mechanisms from data using dwell-time statistics and state-space reconstruction, developing operator-based methods to detect almost-invariant sets directly in high-dimensional recordings, clarifying how metastable dynamics support cognition in active naturalistic behavior, and relating pathological regimes such as seizures to the same formal framework. A plausible implication is that future progress will depend on linking three levels that have often been treated separately: state-detection methods, mechanistic circuit models, and cognitive theories of event structure and predictive control (Rossi et al., 2023, Gozukara et al., 29 May 2026).