Neural Criticality in the Brain

Introduction to Brain Dynamics and Phase Transitions

The hypothesis that the mammalian brain operates near a critical phase transition - commonly referred to as neural criticality - has emerged as a central organizing framework in theoretical and computational neuroscience. In statistical physics, criticality describes a system poised precisely at the boundary between two distinct phases, typically order and disorder. When applied to biological neural networks, the critical brain hypothesis posits that local cortical circuits, and the brain as a macroscopic system, dynamically tune themselves to an operational setpoint that balances synchronized, regular firing with asynchronous, irregular firing ¹²³⁴.

The primary mechanism for observing this phenomenon in neural tissue stems from the discovery of neuronal avalanches. In 2003, Beggs and Plenz demonstrated empirically that spontaneous activity in the local field potentials (LFPs) of mature organotypic rat cortical slices propagates in scale-invariant cascades ²⁵⁶⁷. These avalanches exhibit specific statistical signatures analogous to self-organized criticality (SOC) observed in other complex physical systems, conforming to power-law distributions for both event size and duration ⁶⁷⁸⁹.

Operating at this precise boundary is theorized to provide profound functional advantages. Mathematical and computational models indicate that networks situated at the critical point maximize their dynamic range, optimize information transmission, and access the largest possible repertoire of activity patterns ^1251011. Over the past two decades, experimental evidence has expanded from in vitro slice preparations to in vivo macroscopic recordings using electroencephalography (EEG), magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI) ^211121314. However, the hypothesis remains the subject of rigorous investigation, particularly regarding state-dependent shifts, methodological artifacts in measurement, and the specific phase transitions that govern complex network behavior ³⁴⁸¹⁵.

Theoretical Foundations of Critical Dynamics

Neuronal Avalanches and Branching Processes

The standard empirical test for neural criticality relies on the mathematical framework of branching processes. In a neural branching process, the activity of the network is binned into discrete temporal windows, and the propagation of activity is measured across consecutive bins. A core metric derived from this analysis is the branching parameter, denoted as $\sigma$, which represents the average number of subsequent neural events (or node activations) triggered by a single preceding event within an avalanche ^2111617.

When $\sigma = 1$, the network is in a critical state, allowing activity to cascade through the network without immediately dissipating or growing exponentially. In this regime, the probability distributions of avalanche sizes $P(S)$ and durations $P(T)$ follow power laws, specifically $P(S) \sim S^{-3/2}$ and $P(T) \sim T^{-2}$ ⁵⁶¹⁷¹⁸. If $\sigma < 1$, the system operates in a subcritical state where activity rapidly attenuates, characterized by exponential decay in avalanche distributions. Conversely, if $\sigma > 1$, the network enters a supercritical state, marked by runaway excitation and bimodal size distributions dominated by massive, system-wide bursts ^2111619.

The Edge of Chaos in Recurrent Networks

While avalanche analysis is rooted in discrete spatial and temporal events, a parallel conceptualization of criticality originates in continuous dynamical systems theory. The "edge of chaos" refers to a transition point in parameter space where a network shifts from ordered, stable dynamics to chaotic, highly sensitive dynamics ^20212223. Sompolinsky and colleagues mathematically demonstrated this transition in randomly connected recurrent neural networks subject to "quenched noise" or frozen disorder in their synaptic weights, demonstrating that irregularity in connectivity patterns generates dynamic noise and chaotic fluctuations ²²²³²⁴.

In this continuous framework, the transition point is identified using the maximum Lyapunov exponent $\lambda$. In the ordered regime ($\lambda < 0$), perturbations to the network decay over time. In the chaotic regime ($\lambda > 0$), small perturbations amplify exponentially, indicating extreme sensitivity to initial conditions. At the exact edge of chaos ($\lambda \approx 0$), networks possess a rich repertoire of coexisting, topologically complex dynamics ²¹²³²⁵. Recent computational analyses reveal that networks initialized near this transition differentiate between "rich" structured learning strategies and "lazy" high-dimensional learning strategies, optimizing representation capabilities ²³²⁵. Furthermore, the perturbational complexity index (PCIst) - a metric used to clinically discriminate levels of consciousness - peaks precisely at this edge of chaos in recurrent models ²³²⁵.

Computational Thermodynamics and Reversibility

Researchers have also explored thermodynamic analogs for this computational edge. One paradigm reconceptualizes the edge of chaos not merely as a balance between macroscopic stability and randomness, but as the boundary between reversible and irreversible computation ²⁶. Under Landauer's Principle, irreversible computation deletes information and dissipates heat, thereby generating entropy. Operating near the critical boundary is hypothesized to minimize unnecessary energetic dissipation while maintaining sufficient adaptability to form complex internal models. This framework grounds the critical brain hypothesis in evolutionary metabolic efficiency, suggesting that biological neural networks are optimized to achieve maximum information throughput with minimal entropic cost ²⁶.

Methodological Approaches and Measurement Challenges

Classical Scaling Relations and Subsampling Artifacts

Traditional criticality analysis relies on the statistical aggregation of population-level signals and thresholding these signals to identify avalanches ^13162728. However, estimating parameters like the branching ratio from biological data is mathematically challenging due to subsampling effects. Because standard recordings capture only a tiny fraction of total neurons, conventional branching estimators often systematically underestimate the true parameter, creating an illusion of subcriticality ⁸¹⁶²⁹.

To address this, researchers apply scaling relations that relate the critical exponents of avalanche size ($\tau$), avalanche duration ($\tau_t$), and the scaling of average size with duration ($1/\sigma \nu z$). At standard criticality, these exponents must satisfy the crackling noise relation: $\frac{\tau_t - 1}{\tau - 1} = \frac{1}{\sigma \nu z}$ ¹⁶²⁹. Recent observations in awake and anesthetized animals demonstrate that while the cortex displays critical scaling, the specific exponents often deviate from the classical mean-field directed percolation (MF-DP) universality class ²⁹³⁰³². Fontenele et al. (2019) demonstrated that cortical activity traverses an intermediate level of spiking variability where these exponents follow a distinct linear relation, suggesting that the brain's critical phase transition may belong to a more complex universality class, or that the observed exponents are fundamentally warped by network topology and experimental subsampling ²⁹³⁰³².

Phenomenological Renormalization Group (pRG) Analysis

To bypass the thresholding artifacts inherent in avalanche detection, recent methodologies employ the phenomenological renormalization group (pRG) approach ⁸³¹³². The pRG methodology groups highly correlated neurons into clusters and scales these clusters across multiple spatial resolutions. If the system is critical, the distributions of cluster activities, variance, and free energy scale cleanly across observation resolutions, demonstrating true fractal self-similarity ³¹.

While pRG provides robust evidence of scale-invariant properties across multiple recording modalities, including extracellular electrophysiology and calcium imaging, methodological investigations warn of sensitivities. Temporal binning, measurement nonlinearities, and the specific algorithms used for deconvolution heavily influence the resulting scaling exponents. Thus, strict standardization is required to isolate inherent neural dynamics from measurement-induced scaling artifacts ¹⁵³¹.

The Phenomenon of Hidden Criticality

A persistent complication in the field is the frequent empirical observation of non-scale-invariant data, leading some researchers to reject the critical brain hypothesis when power laws poorly fit the data. However, recent computational and mathematical models demonstrate the existence of "hidden criticality" ⁸³¹³⁵.

In high-dimensional neural systems, activity patterns can segregate into distinct low-dimensional subspaces. Standard analyses average the activity of a large neural population into a one-dimensional time series, effectively collapsing these subspaces. If a network contains strongly anti-correlated sub-populations - a common physiological feature of biological circuits involving local inhibitory interneurons - these anti-correlations can destructively interfere during population averaging ⁸³⁵. Consequently, the averaged signal may exhibit strong oscillations or exponential decay, falsely signaling a subcritical or non-critical state, even when the network's underlying dimensional manifolds contain perfect scale-invariant, critical dynamics ⁸³⁵.

To clarify the distinctions in analytical approaches and their respective sensitivities, the dominant methodologies are summarized below:

Functional Consequences of Operating Near Criticality

Optimization of Dynamic Range

The primary evolutionary argument for neural criticality is the maximization of computational utility. A system poised at criticality yields the maximum possible dynamic range, defined as the span of stimulus intensities over which the network produces distinguishable, non-saturated responses ^10363334.

This relationship mimics classical pharmacological dose-response analyses, where the efficacy of a stimulus is mapped to a sigmoidal function (e.g., the four-parameter logistic model). In such models, the optimal sensitivity corresponds to the steepest slope of the sigmoid curve ^10394041. When transposed to neural networks, the computational dynamic range exhibits a pronounced bell-shaped tuning curve relative to the network's phase state. The dynamic range peaks precisely when the control parameter (the branching ratio) approaches 1.0.

In subcritical networks, sensory stimuli rapidly decay; the network lacks the recursive excitation necessary to amplify faint inputs, leading to a restricted dynamic range dominated by a high threshold of detection and low overall sensitivity ¹⁰³⁴. In supercritical networks, activity amplifies uncontrollably. Consequently, even minor stimuli trigger massive, system-wide bursts, meaning the network is perpetually saturated and unable to encode the specific intensity or nuance of the stimulus. Deviations toward either subcritical or supercritical states result in rapid, non-linear losses of sensory and informational capacity ¹⁰³³³⁴. Interestingly, empirical data aligns with mathematical models showing that activity-dependent synaptic depression plays an unanticipated but necessary role in preventing total saturation in supercritical regimes in vivo ¹⁰³³.

Information Transmission and Repertoire

At the critical boundary, neural networks achieve an optimal balance. Information transmission, encoding capacity, and long-range spatiotemporal correlations are maximized ^25112035. A critical network structure sustains long-range temporal correlations (LRTCs) without succumbing to global hypersynchrony. This allows the cortex to support localized, specific processing modules while retaining the capability to integrate information globally across disparate anatomical brain regions ³⁶³⁷.

State-Dependent Shifts in Cortical Dynamics

Resting State Versus Cognitive Task Performance

Although foundational models suggest that perfect criticality is universally optimal, empirical evidence reveals that the brain undergoes dynamic shifts away from the critical point depending on cognitive and behavioral states. A counterintuitive finding in modern neuroimaging is that while the resting brain operates closely to the critical point, engaging in focused, task-driven cognition reliably shifts the network into a subcritical regime ¹¹⁶³⁸.

In experiments utilizing visually presented Choice Reaction Tasks (CRT), heightened cognitive demand consistently correlated with subcritical dynamics ¹. This shift serves a crucial functional purpose. While criticality maximizes overall information integration, a highly excitable, critical network is also prone to interference from irrelevant environmental stimuli. Shifting into a subcritical state functions as a dynamic neural filter, attenuating extraneous sensory input and broad-network crosstalk, thereby protecting the continuous, focused processing of task-relevant information ¹³⁸. Interestingly, performance speed and accuracy in these tasks are optimized when the brain balances task focus with partial, residual activation of the Default Mode Network (DMN) - a state of being "in the zone" - indicating that strict subcriticality is functionally tempered by underlying resting-state critical features ¹.

Sleep, Wakefulness, and Homeostatic Restoration

Prolonged wakefulness poses a severe physiological challenge to critical dynamics. During waking hours, the brain undergoes experience-dependent, Hebbian associative plasticity. While these localized synaptic changes encode new memories and adaptations, they structurally alter the global network topology, gradually accumulating imbalances in excitation and inhibition ³¹⁸³⁸.

As wakefulness extends, the accumulating synaptic weights push the cortical network away from criticality, often leading to a supercritical or increasingly disordered state characterized by reduced information processing capacity and cognitive fatigue ³¹⁸³⁸. Sleep acts as a homeostatic restorative mechanism. Non-rapid eye movement (NREM) sleep and its associated complex switching of dynamic slow-wave states re-tune synaptic strengths globally, pruning unnecessary connections and returning the network topology to the optimal critical regime ³¹⁸³⁸. Meta-analyses of voltage imaging and spiking data confirm that sleep consistently restores scale-free dynamics in cortical networks, mapping sleep mathematically as a restorative vector pushing the brain back toward the peak of its computational fitness landscape ³.

Neuromodulatory Tuning of the Excitation-Inhibition Balance

The maintenance of criticality depends on the exquisite balance between excitatory and inhibitory (E/I) neurotransmission, mediated primarily by glutamate and $\gamma$-aminobutyric acid (GABA) ^3940414243. To navigate the narrow parameter space of the critical boundary, the brain relies on ascending neuromodulatory systems that act as dynamic control dials for local network excitability.

Dopaminergic Modulation of Neural Excitability

Dopamine plays a multifaceted role in shifting network dynamics by modulating the intrinsic excitability of neurons and the efficacy of glutamatergic pathways ¹⁹³⁹⁴⁴. Arising from midbrain structures, dopamine regulates E/I balance through receptor-specific intracellular signaling cascades:

• D1-Like Receptors: Activation of D1 receptors stimulates the Protein kinase A (PKA) pathway, increasing intracellular cyclic adenosine monophosphate (cAMP). This signaling cascade subsequently inhibits KCNQ-mediated potassium currents via the ERK signaling pathway. The net effect is an enhancement of pyramidal neuron and medium spiny neuron (MSN) excitability, making these cells more likely to participate in avalanche propagation. Moderate D1 activation is vital for maintaining the critical regime, optimizing neural gain, and facilitating plasticity ¹⁹³⁹.
• D2-Like Receptors: Conversely, D2 activation decreases cAMP levels, downregulating excitatory signaling and promoting local inhibition. D2 receptors further modulate G-protein-coupled inward rectifier potassium (GIRK) channels to regulate neuronal electrical activity. This inhibitory mechanism prevents the network from transitioning into runaway supercriticality ³⁹.

Too little or too much dopamine signaling disrupts this delicate homeostatic balance, forcing the network into subcritical degradation or supercritical instability, respectively ¹⁹.

Cholinergic Influence on Information Processing

Acetylcholine (ACh) operates concurrently to tune E/I balance, originating from basal forebrain and brainstem nuclei (e.g., LDT and PPT) ⁴¹⁴⁵. ACh serves primarily to facilitate attention, orientation, and the suppression of irrelevant noise by modulating excitation via two distinct receptor classes ³⁹⁴¹:

• Through nicotinic receptors located on presynaptic axonal terminals, ACh enhances the release of glutamate, promoting rapid, localized excitatory signaling ³⁹⁴¹⁴⁵.
• Through muscarinic M1 receptors, ACh modulates KCNQ-mediated currents by stimulating the phosphorylation of KCNQ2. In excitatory neurons, this typically facilitates channel closure, increasing neuronal firing rates. Conversely, in specific GABAergic interneurons, muscarinic activation negatively modulates excitation, tightly regulating local circuit timing and dampening broad excitation ³⁹⁴¹⁴⁵.

By interacting antagonistically with dopaminergic transmission in certain regions, such as the dorsal striatum, cholinergic tone ensures that the network does not become broadly hyper-excitable. Instead, ACh finely adjusts the effective branching ratio of the local network to suit current behavioral demands, generally driving the targeted subcriticality necessary for focused attention ⁷³⁹⁴⁵.

The comparative mechanisms of these neuromodulators are summarized in the table below:

Clinical Correlates of Deviations from Criticality

Because complex neural computation relies heavily on proximity to the critical point, persistent deviations - whether structurally, genetically, or chemically induced - correlate directly with specific neuropsychiatric and neurological disorders.

Schizophrenia and Excitation-Inhibition Imbalance

Schizophrenia is increasingly understood through the lens of profound E/I imbalance, specifically characterized by an abnormal elevation of network excitability and a failure of local inhibition ^42434647. A central pathophysiological feature of schizophrenia is the functional impairment of parvalbumin (PV)-containing GABAergic interneurons, coupled with hypofunction of NMDA receptors (involving the GRIN1, GRIN2B, and GRIN3B genes) ⁴²⁴⁶⁴⁷.

PV interneurons are critical for pacing high-frequency gamma oscillations and providing the fast-spiking inhibition necessary to constrain glutamatergic avalanches ⁴²⁴⁶. When PV function degrades, the network shifts away from the critical point. Clinical high-risk (CHR) individuals and first-episode patients demonstrate altered Glx/GABA ratios via magnetic resonance spectroscopy, correlating directly with aberrant resting-state gamma-band power ⁴²⁴³. In computational biophysical models of the prefrontal and posterior parietal cortices, simulating this specific E/I disruption leads directly to the disintegration of working memory attractor states. The mnemonic activity pattern undergoes random drift that decreases the precision of responses, mirroring the cognitive context-integration deficits observed clinically in paradigms like the dot pattern expectancy task ⁴⁷⁴⁸. Restoring this balance - potentially through novel compounds targeting muscarinic receptors on PV interneurons, or positive allosteric modulators of $\alpha$5-GABA-A receptors - represents a promising therapeutic frontier aimed directly at returning the cortex to stable criticality ⁴⁶.

Epilepsy and Supercritical Transitions

Epilepsy is traditionally conceptualized as a disorder where the brain spontaneously shifts into a supercritical state, generating massive, saturated avalanches of activity recognized as clinical seizures ^49505152. In a supercritical regime, the branching parameter continuously exceeds 1.0, meaning that external inputs or spontaneous internal noise trigger uncontrolled cascade amplification ¹¹⁵⁰.

However, high-resolution computational analyses of epileptic dynamics, utilizing tools like the Epileptor model, reveal a more complex picture. During seizure-free (interictal) intervals, the cortex of epilepsy patients often operates in a highly stable, subcritical or critical regime, indistinguishable from healthy brains ⁴⁹. The onset of a seizure is now modeled mathematically as a bifurcation in a bistable dynamical system. Due to transient fluctuations in neural excitability and connectivity, the boundary between the "healthy" critical state and the "ictal" supercritical state degrades ⁵⁰⁵¹⁵³.

As network excitability increases, the system's resilience to perturbations decreases - a nonlinear dynamical phenomenon known as critical slowing down. Optogenetic and electrical probing in murine and human models confirms that as the system approaches the transition boundary, recovery times from small perturbations lengthen significantly. Once the control parameter breaches the critical point, the system is mathematically forced to transition into the seizure regime ⁵⁰⁵³. Consequently, epilepsy is not defined by a state of permanent supercriticality, but rather by a pathological structural vulnerability to crossing the critical boundary ⁴⁹⁵⁰.

Global Perspectives: Large-Scale Computational Modeling and Genetics

The empirical verification of the critical brain hypothesis increasingly relies on extreme-scale computational simulations and large cohort studies spearheaded by global neuroscience institutes.

Digital Twin Brains and Billion-Neuron Simulations

At the Institute of Science and Technology for Brain-Inspired Intelligence at Fudan University (China), researchers have developed a "Digital Twin Brain" (DTB) ⁵⁴⁵⁵⁶³. This cortico-subcortical spiking network model encompasses up to 86 billion neurons, utilizing patient-specific multimodal MRI data (including DWI and T1-weighted imaging) to constrain its topological architecture across 16,043 distinct anatomical voxels ⁵⁵⁶³.

When the DTB is scaled up to 20 billion active neurons with an average synaptic connectivity degree of 100, the spontaneous resting-state simulations generate BOLD fMRI signals that strongly match empirical human data (achieving a high Pearson correlation coefficient). Crucially, the avalanche size and duration in the DTB naturally approach power-law distributions (with exponents of 1.13 and 1.18, respectively), confirming that realistic, data-constrained macroscopic architecture organically yields critical dynamics without artificial parameter tuning ⁵⁴⁵⁵⁶³. The similarity between the modeled and empirical data strictly increases with scale, demonstrating that large-scale network complexity is a fundamental prerequisite for generating true emergent criticality ⁵⁴⁵⁵.

Genetic Heritability of the Branching Parameter

Furthermore, recent large-scale fMRI studies from the Human Connectome Project (HCP) have established that maintaining a critical state is under significant genetic control. By analyzing resting-state fMRI data from monozygotic (MZ) and dizygotic (DZ) twins, researchers demonstrated that the branching ratio ($\sigma$) is highly heritable and consistently hovers around 1.0 (MZ: $1.00 \pm 0.10$, DZ: $1.02 \pm 0.10$) ⁵⁶. The data also revealed a profound U-shaped relationship between the branching ratio and temporal Renormalization Group metrics (such as the Hurst exponent, $d_2$). Subjects whose branching ratio approached exactly 1 exhibited the lowest $d_2$ values, confirming optimal scale-free dynamics ⁵⁶. The heritability of these critical parameters provides a direct biological mechanism linking genetic vulnerabilities to the systemic E/I imbalances seen in highly heritable psychiatric disorders like schizophrenia and depression ⁴²⁵⁶.

Simultaneously, researchers at the Brain Institute (UFRN) and the Federal University of Pernambuco in Brazil have investigated how criticality scales in vivo across discrete cortical states ^30325758. Using pRG approaches, they analyzed neural activity in the primary visual cortex of behaving mice, linking non-trivial scaling directly to successful task performance ³². Furthermore, using data from both anesthetized and freely moving animals, they demonstrated that universal dynamics and the critical phase transition occur precisely at intermediate values of spiking variability ^29303258. These insights prove that the critical point is not an abstract mathematical artifact, but a measurable biological setpoint mapped dynamically across the vigilance spectrum.

Conclusion

The question of whether the brain operates at the edge of chaos yields a highly nuanced affirmative. The brain is not permanently locked in a singular, immutable critical state; rather, neural criticality serves as an optimal dynamical setpoint. During restful wakefulness, the system hovers near this precise boundary to maximize its readiness, computational dynamic range, and information transmission efficiency. When targeted cognitive focus is required, neuromodulators actively shift the network into temporary subcriticality to filter environmental noise. When associative synaptic burdens become too high during prolonged wakefulness, the system risks tipping into supercritical dysfunction, necessitating the homeostatic, restorative intervention of sleep.

The edge of chaos is fundamentally a thermodynamic and informational compromise. By utilizing complex neuromodulatory systems - predominantly dopamine and acetylcholine - to continuously tune the excitation-inhibition balance, the brain navigates a computational landscape that avoids both the restrictive entropy of total order and the dissipative noise of complete chaos. Understanding this intricate balance not only sheds light on the mechanical underpinnings of healthy cognition but precisely traces the physiological origins of pathological dysfunction when the brain strays too far from the critical edge.