Theoretical foundations of studying criticality in the brain

Neuroscience is dawning upon revealing physics foundations of the brain (Abbott, 2008). Ever since the 1970s, the term neurophysics has been suggested as a term to indicate the essential role of physics in understanding the brain (Scott, 1977). More recently, substantial progress has been accomplished in studying brain connectivity and brain functions with statistical physics theories (Lynn & Bassett, 2019).

For brain connectivity, physics provides insights for its emergence, organization, and evolution. Random graphs (Betzel et al., 2016; Betzel & Bassett, 2017a), percolation (Breskin, Soriano, Moses, & Tlusty, 2006; Guo et al., 2021), and other physics theories of correlated systems (Haimovici, Tagliazucchi, Balenzuela, & Chialvo, 2013; Wolf, 2005) are applied to reveal the underlying mechanisms accounting for the origins of brain network properties. Complex network theories act as the foundation of characterizing brain connectivity organizational features (e.g., community; Betzel & Bassett, 2017b; Betzel, Medaglia, & Bassett, 2018; Khambhati, Sizemore, Betzel, & Bassett, 2018), hub (Deco, Tononi, Boly, & Kringelbach, 2015; Gong et al., 2009), and small-world (Bullmore & Sporns, 2012; Deco et al., 2015; structures) and embedding attributes into physical space (Bassett et al., 2010; Kaiser & Hilgetag, 2006). Network evolution driven by neural plasticity helps to explain the dynamics of brain connectivity structures during information processing (Del Pozo et al., 2021; Galván, 2010; Montague, Dayan, & Sejnowski, 1996; Robert & Vignoud, 2021; Song, Miller, & Abbott, 2000). For brain functions, physics presents possible explanations for the origin of information processing capacities from collective neural activities. From single neuron dynamics models (Gerstner, Kistler, Naud, & Paninski, 2014), stochastic network models of neural populations and circuits (Tian, Li, & Sun, 2021; Tian & Sun, 2021), mean-field neural mass models of brain regions (David & Friston, 2003; Touboul, Wendling, Chauvel, & Faugeras, 2011), eventually to models of entire brain networks (Hopfield, 1982; Schneidman, Berry, Segev, & Bialek, 2006), important efforts have been devoted to characterize information-processing-related neural dynamics across different scales. Networks with memory capacities (e.g., Hopfield networks; Tyulmankov, Fang, Vadaparty, & Yang, 2021), which are equivalent to Ising models under specific conditions (Lynn & Bassett, 2019), have been applied to study neural information storage and recall (Haldeman & Beggs, 2005; Krotov & Hopfield, 2020), adaptation to environment changes (Shew et al., 2015), information transmission optimization (Beggs & Plenz, 2003), dynamic range maximization (Kinouchi & Copelli, 2006; Shew, Yang, Petermann, Roy, & Plenz, 2009), and neural computation power (Bertschinger & Natschläger, 2004). These models are further related to maximum entropy models (e.g., specific fine-tuned Ising models) that predict long-range correlations observed among neurons (Ganmor, Segev, & Schneidman, 2011; Schneidman et al., 2006). Moreover, general theories of free-energy principle (Friston, 2009, 2010; Guevara, 2021) and information thermodynamics (Capolupo, Freeman, & Vitiello, 2013; Collell & Fauquet, 2015; Sartori, Granger, Lee, & Horowitz, 2014; Tian & Sun, 2022) are suggested as the unified foundations of perception, action, and learning in the brain.

If one needs to specify one of the most focused and controversial topics among all the works mentioned above, brain criticality may be a potential candidate (Beggs & Timme, 2012). The hypothesis of the critical brain has received increasing attention in recent decades, serving as a possible mechanism underlying various intriguing but elusive phenomena in the brain. In light of our limited understanding of the complex nature of collective neural dynamics, these phenomena include, to name a few, efficient transitions between cortical states (Fontenele et al., 2019), maximal dynamic ranges of neural responses (Antonopoulos, 2016; Gautam, Hoang, McClanahan, Grady, & Shew, 2015; Kinouchi & Copelli, 2006; Shew et al., 2009), optimized information transmission and representation (Antonopoulos, 2016; X. Li & Small, 2012; Shew, Yang, Yu, Roy, & Plenz, 2011), and numerous other issues concerning brain functions that we have mentioned above. One can see Beggs (2007), Chialvo (2010), Cocchi, Gollo, Zalesky, and Breakspear (2017), Hesse and Gross (2014), and Shew and Plenz (2013) for systematic reviews of the diverse function advantages implied by brain criticality and their experimental demonstrations. From a Darwinian perspective, one potential reason for the brain to feature criticality lay in that the most informative parts of external world principally occur at a borderline between purely ordered and purely disordered states (information would be trivial in a purely ordered world while it would be incomprehensible in a purely disordered world). Becoming critical may be a potential way for the brain to adapt to the complex world, where nontrivial information has a finite opportunity to occur (Bak, 2013; Chialvo, 2010). To date, generic features of a critical brain with the characteristics discussed above, such as divergent correlation length, neuronal avalanches with power law behaviors, and long-range correlations on the microscopic scale (e.g., neural populations), have been extensively observed in mathematical models in conjunction with experimental data (e.g., Beggs & Plenz, 2003; Dalla Porta & Copelli, 2019; Fosque, Williams-García, Beggs, & Ortiz, 2021; Gireesh & Plenz, 2008; Hardstone, Mansvelder, & Linkenkaer-Hansen, 2014; Petermann et al., 2009; Poil, Hardstone, Mansvelder, & Linkenkaer-Hansen, 2012; Poil, van Ooyen, & Linkenkaer-Hansen, 2008; Ponce-Alvarez, Jouary, Privat, Deco, & Sumbre, 2018; G. Scott et al., 2014; Shew et al., 2009; Shriki et al., 2013; Tagliazucchi, Balenzuela, Fraiman, & Chialvo, 2012; Tkačik et al., 2015).

Our work does not aim at repeatedly reviewing experimental advances concerning brain criticality and its biological significance, given that they have been comprehensively summarized by existing reviews (Beggs, 2007; Chialvo, 2010; Cocchi et al., 2017; Hesse & Gross, 2014; Muñoz, 2018; Shew & Plenz, 2013). On the contrary, our motivation is to present a systematic and accessible review of the theoretical methods applied to achieve these advances, which have not received necessary attention yet.

These theoretical foundations are initially thought to be incomprehensible and irrelevant to neuroscience. However, practice suggests that omitting these physical and mathematical backgrounds does not significantly improve the accessibility of studies on brain criticality. Instead, the lack of detailed explanations of theoretical foundations has frequently misled neuroscientists, leading to diverse confusions about the precise meaning, identification criteria, and biological corollaries of brain criticality. As a result, criticality, an analytic statistical physics theory with solid foundations, unnecessarily becomes an elusive black box for neuroscientists. To address this issue, we use the terminology of neuroscience to present a self-contained framework of brain criticality, reviewing and reformulating (1) physical theories that analytically derive brain criticality and (2) statistic techniques that computationally estimate brain criticality from empirical data. Given the frequent misunderstanding of neural avalanches, our discussions primarily focus on brain criticality analysis on the microscopic scale of the brain. The objectives guiding our review are tripartite: (1) explaining why brain criticality matters in the brain, (2) understanding what is brain criticality and what it conveys about the brain, and (3) confirming how to identify potential brain criticality and ensure the validity of analyses.

Brain criticality frequently confuses neuroscientists since too many distinct phenomena are studied under this name without being properly classified. In this review, brain criticality refers to a family of critical processes in neural dynamics where erratic fluctuations appear to reduce dynamic stability. To present a systematic classification framework, we discuss three fundamental perspectives concerning brain criticality. Table 1 provides all the necessary glossaries in comprehensible forms.

First, the brain, similar to other biological systems, generally exhibits temporal evolution from initial states that are far away from equilibrium (Gnesotto, Mura, Gladrow, & Broedersz, 2018; Lynn, Cornblath, Papadopoulos, Bertolero, & Bassett, 2021). These departures from equilibrium arise due to diverse endogenous causes (Gnesotto et al., 2018; Perl et al., 2021) to break the detailed balance to support consciousness, sensing, and adaptation (Lynn et al., 2021; Perl et al., 2021). Therefore, potential critical phenomena underlying neural dynamics, at least in most neural dynamics models and empirical datasets, are basically nonequilibrium and cannot be characterized by equilibrium statistic mechanics. In Figure 1A, we illustrate the difference between equilibrium and nonequilibrium dynamics.

Second, there exist two types of general mechanisms underlying the existence of brain criticality. One type of mechanisms either arise from the external manipulations outside the brain (e.g., researchers manipulate the tonic dopamine D1-receptor stimulation; Stewart & Plenz, 2006, 2008) or adjust network topology (Kaiser & Hilgetag, 2010; Rubinov, Sporns, Thivierge, & Breakspear, 2011; S. Wang & Zhou, 2012) or belong to the top-down biological processes that globally function on neural dynamics inside the brain (e.g., anesthesia effects; Fontenele et al., 2019; Hahn et al., 2017; Ribeiro et al., 2010) as well as sleep restoration effects (Meisel, Olbrich, Shriki, & Achermann, 2013). Neural dynamics is passively fine-tuned toward or away from ordinary criticality (OC) by these exogenous mechanisms, similar to ordinary critical phenomena that require the fine tuning of order parameters.

Another type of mechanisms includes all endogenous factors of neural dynamics (e.g., neural plasticity mechanisms such as spike-timing dependent synaptic plasticity (Effenberger, Jost, & Levina, 2015; Meisel & Gross, 2009; Shin & Kim, 2006), short-term synaptic plasticity (Levina, Herrmann, & Geisel, 2007, 2009), retro-synaptic signals (Hernandez-Urbina & Herrmann, 2017), and Hebbian rules (De Arcangelis & Herrmann, 2010; De Arcangelis, Perrone-Capano, & Herrmann, 2006), which locally function on neural dynamics as drive and dissipation components. The interactions between these components naturally form feedback control loops to support the self-organization of neural dynamics toward the critical point (Beggs, 2007; Chialvo, 2010). This spontaneously emerged brain criticality, distinct from ordinary critical phenomena, is conjectured as a kind of self-organized criticality (SOC) (Chialvo, 2010). In Figure 1B, we present conceptual illustrations of ordinary criticality and self-organized criticality in the brain.

Third, brain criticality frequently occurs in nonstandard forms due to stimulus drives or endogenous factors. On the one hand, slight and inconstant deviations from perfect brain criticality can be seen on the actual values of characteristic variables, differentiating the characterized phenomena from the standard criticality (Fosque et al., 2021; Williams-García, Moore, Beggs, & Ortiz, 2014). On the other hand, all statistical physics relations required by perfect brain criticality are still adhered by these actual characteristic variables, distinguishing the brain from being noncritical (Fosque et al., 2021; Williams-García et al., 2014).

For ordinary criticality, its nonstandard form is referred to as quasi-criticality (qC) (Fosque et al., 2021; Williams-García et al., 2014). Diverse mechanisms can force the brain to depart from perfect ordinary criticality and exhibit quasi-critical neural dynamics, among which, stimulus derive may be the most common one (Fosque et al., 2021; Williams-García et al., 2014). In general, sufficiently strong stimulus drives can capture or even govern neural dynamics. Similar to the situation where external inputs suppress irregular neural dynamics (Molgedey, Schuchhardt, & Schuster, 1992), the stimuli that are too strong may evoke intense but less changeable neural dynamics to make the brain depart from the perfect critical point (Fosque et al., 2021; Williams-García et al., 2014). Let us take the qC phenomenon introduced by Fosque et al. (2021) and Williams-García et al. (2014) as an instance. Under specific conditions, the actual brain state may be close to a Widom line in the three-dimensional space defined by the stimulus intensity υ, refractory period length τ, and branching ratio κ (i.e., the time-dependent average number of subsequent neural activities caused by a single neuron activation event (Haldeman & Beggs, 2005). The Widom line is a line of all the combinations of (υ, τ, κ) where the susceptibility of neural dynamics is maximized (Fosque et al., 2021; Williams-García et al., 2014). The susceptibility is defined by lim_x→0$∂y∂x$⁠, where y is the neural dynamics state and x denotes a factor that affects y. In general, one can understand susceptibility as the degree to which fluctuations in the state of each neuron can propagate to neighbored neurons (Williams-García et al., 2014). Being close to the Widom line suggests the existence of quasi-criticality in the brain. Moving along the Widom line as the stimulus intensity increases, the susceptibility of neural dynamics decreases, and the branching ratio at maximal susceptibility will decrease as well (Fosque et al., 2021; Williams-García et al., 2014). Significant deviations from the Widom line suggest noncriticality (i.e., the subcriticality where neural dynamics is disordered and the super-criticality where neural dynamics is ordered; Williams-García et al., 2014). In Figure 1C, we conceptually illustrate how stimuli imply qC in the brain. In Figure 2D, the qC phenomenon in Fosque et al. (2021) and Williams-García et al. (2014) is shown in detail.

As for self-organized criticality (SOC), its nonstandard form is defined according to statistical physics criteria. Perfect self-organized criticality only exists in conserved neural dynamics (e.g., see integrate-and-fire neurons analyzed by Levina et al., 2007), where system energy (i.e., neural activities) either conserves within the system and only dissipates at the system boundary, or dissipates inside the system (i.e., bulk dissipation) with a dissipation rate vanishing in the system size limit (Malcai, Shilo, & Biham, 2006). Under more general conditions where neural dynamics are not conserved (e.g., see leaky integrate-and-fire neurons analyzed by Levina et al., 2007; Millman, Mihalas, Kirkwood, & Niebur, 2010; Rubinov et al., 2011; Stepp, Plenz, & Srinivasa, 2015; where neural dynamics dissipates within the system due to voltage leak), perfect self-organized criticality can be broken by any rate of bulk dissipation (Bonachela, De Franciscis, Torres, & Muñoz, 2010; Bonachela & Muñoz, 2009; Buendía, di Santo, Villegas, Burioni, & Muñoz, 2020b; de Andrade Costa, Copelli, & Kinouchi, 2015). Stronger bulk dissipation implies larger deviations from perfect self-organized criticality (De Arcangelis et al., 2006). Consequently, the self-organization process of nonconserved neural dynamics only make the brain hover around the critical point. Any further closeness toward the critical point requires the fine tuning of order parameter by additional exogenous mechanisms, which is different from pure self-organized criticality (Bonachela et al., 2010; Bonachela & Muñoz, 2009; Buendía et al., 2020b; de Andrade Costa et al., 2015). This nonconserved self-organization process is termed as self-organized quasi-criticality (SOqC) (Bonachela & Muñoz, 2009). Similar to SOC in conserved dynamics, neural plasticity mechanisms, such as spike-timing-dependent synaptic plasticity (Rubinov et al., 2011), Hebbian rules (De Arcangelis et al., 2006), short-term synaptic depression in conjunction with spike-dependent threshold increase (Girardi-Schappo et al., 2021), and inhibitory plasticity in conjunction with network topology (Ma, Turrigiano, Wessel, & Hengen, 2019), can serve as underlying self-organization mechanisms of SOqC. Because purely conserved neural dynamics is relatively rare in empirical data (e.g., neural dynamics is conserved for integrate-and-fire neurons; Levina et al., 2007; and leaky integrate-and-fire neurons whose presynaptic inputs are exactly equal to the sum of voltage leak and potential costs during neural spiking (Bonachela et al., 2010), we suggest that SOqC may be more common in the brain than SOC. In Figure 1D, we present conceptual instances of the two-stage approaching process towards the critical point in the brain with SOqC.

The above discussion has presented a classification framework of brain criticality, that is, ordinary criticality (OC), quasi-criticality (qC), self-organized criticality (SOC), and self-organized quasi-criticality (SOqC). In Figure 1E, we compare between these four types of brain criticality in term of susceptibility. In general, susceptibility diverges at the critical point in a brain with standard criticality (e.g., OC and SOC) while it does not diverge in the quasi-critical region of a brain with nonstandard criticality (e.g., qC and SOqC). In Figure 1F, we summarize the commonness and difference between these four types of brain criticality discussed in our review. From a neuroscience perspective, a brain with critical neural dynamics is expected to be near the critical point and prepared for tremendous changes in cortical states during a short duration. This intriguing property coincides with the experimentally observed efficient transitions between cortical states (e.g., Cardin, 2019; Holcman & Tsodyks, 2006; Jercog et al., 2017; H. Lee, Wang, & Hudetz, 2020; Reimer et al., 2014) and, therefore, interests researchers for the potential existence of brain criticality. The importance of identifying brain criticality in neural dynamics is beyond brain criticality itself because it implies an opportunity to explain and predict brain function characteristics by various statistical physics theories built on nonequilibrium criticality.

To identify potential nonequilibrium criticality in the brain, researchers actually characterize neural dynamics as a physical system with absorbing states and avalanche behaviors (Hinrichsen, 2000; Larremore, Carpenter, Ott, & Restrepo, 2012; Lübeck, 2004). In general, one needs to consider the propagation of neural dynamics where neurons are either activated (“on” state) or silent (”off” state) (Dalla Porta & Copelli, 2019). A silent neuron may be activated with a probability defined by the number of activated presynaptic neurons and the coupling strength θ among neurons (e.g., neural correlation; Franke et al., 2016). An activated neuron spontaneously becomes silent at a constant rate (e.g., after the refractory period; Kinouchi & Copelli, 2006; Squire et al., 2012). These definitions naturally support to distinguish between different phases of neural dynamics. Here we review two kinds of phase partition that are active in neuroscience.

The second group of phases are synchronous and asynchronous phases (di Santo, Villegas, Burioni, & Muñoz, 2018; Fontenele et al., 2019; Girardi-Schappo et al., 2021). As their names suggest, these two phases correspond to the situations where synchronization emerges or disappears in neural activities, respectively. Synchronization refers to the cases where “on” states appear in an oscillatory, although not strictly periodic, manner. To quantify its potential existence, we can measure the variability of neural dynamics using the coefficient of variation (CV) (di Santo et al., 2018; Fontenele et al., 2019; Girardi-Schappo et al., 2021) or the Kuramoto order parameter (Acebrón, Bonilla, Pérez Vicente, Ritort, & Spigler, 2005; Arenas, Díaz-Guilera, Kurths, Moreno, & Zhou, 2008). CV can be defined from diverse perspectives, yet the most common definition is the ratio between the standard deviation and the mean of the interspike interval length (di Santo et al., 2018; Fontenele et al., 2019; Girardi-Schappo et al., 2021). A higher value of CV implies the reduction of synchronization. For most neural dynamics data, an empirical choice of the CV threshold that separates between synchronous and asynchronous phases may be ≃1 (Fontenele et al., 2019) or ≃$32$ (Fontenele et al., 2019). The Kuramoto order parameter ω ∈ [0, 1] measures the coherent degree of neural dynamics based on the Kuramoto model of oscillators (for detailed definitions see Acebrón et al., 2005; Arenas et al., 2008). Perfect synchronization emerges when ω = 1 and vanishes when ω = 0 (Acebrón et al., 2005; Arenas et al., 2008).

The boundary between these two phases is the critical point, at which the brain is on the edge of exhibiting self-sustained (for absorbing and active phases) or synchronous (for synchronous and asynchronous phases) neural dynamics. Perturbations (e.g., the propagation of “on” state among neurons) to the absorbing or asynchronous phase do not have characteristic lifetime and size. These perturbations, referred to as neural avalanches, are expected to exhibit power law properties in their lifetime (time difference between the first and last activation of neurons in between complete quiescent epochs) and size (number of active neurons along with the excursion) distributions (Hesse & Gross, 2014; Hinrichsen, 2000; Larremore et al., 2012; Lübeck, 2004). In general, the emergence of neural avalanches implies the slowing down of neural dynamics, that is, the brain state recovery process toward the baseline state after fluctuations changes from fast (exponential) to slow (power law) (Cocchi et al., 2017; Hesse & Gross, 2014). The dynamic stability of neural dynamics is limited by the slow recovery and, therefore, cannot robustly counteract perturbations. Consequently, small perturbations initiated on the microscopic scale may still make the brain change sharply on the macroscopic scale (Cocchi et al., 2017; Hesse & Gross, 2014). In Figure 2B, we conceptually illustrate how the recovery process slows down when the brain is close to the critical point or the quasi-critical region.

The relation between neural avalanches and brain criticality is frequently neglected or misunderstood. Neural avalanche data alone is not sufficient to determine the concrete type of brain criticality (i.e., OC, qC, SOC, and SOqC) unless additional information about the mechanisms underlying neural avalanche emergence is provided (e.g., if neural dynamics is conserved or self-organizing). To explore a concrete type of brain criticality, researchers need to explicitly present its definition depending on different control parameters (e.g., the balance between excitatory and inhibitory neurons in CROS models; Hardstone et al., 2014; Poil et al., 2012) and order parameters (e.g., active neuron density and synchronous degree; Dalla Porta & Copelli, 2019). A brain criticality hypothesis without strict definitions of control and order parameters is not informative (Cocchi et al., 2017; Girardi-Schappo, 2021). To present conceptual instances, we illustrate four possible critical phenomena in Figure 2, each of which corresponds to a concrete brain criticality type.

To produce ordinary criticality (OC), we can control neural dynamics and manipulate 〈θ〉, the expectation of coupling strength θ among all neurons (e.g., averaged neural correlation), by some top-down and global biological effects. These effects, for instance, may be anesthesia effects (e.g., by ketamine-xylazine; Ribeiro et al., 2010; and isoflurane; Hahn et al., 2017) or sleep restoration effects; Meisel et al., 2013). We use the Kuramoto order parameter ω (Acebrón et al., 2005; Arenas et al., 2008) as the order parameter to define synchronous and asynchronous phases (di Santo et al., 2018; Fontenele et al., 2019). As 〈θ〉 increases, we may see transitions from asynchronous to synchronous phase in some situations (see a similar instance in Villegas, Moretti, & Muñoz, 2014). One can see Figure 2C for conceptual illustrations.

To produce quasi-criticality (qC), we can manipulate refractory period length τ, branching ratio κ, and stimulus intensity υ as control parameters (e.g., control τ and κ by pharmacological perfusion or ionic concentration adjustment; Chiappalone et al., 2003; Shew et al., 2011). There exist a disordered phase (subcritical), an ordered but asynchronous phase (supercritical), and a synchronous (quasi-periodic) phase in the space of (υ, τ, κ) (Fosque et al., 2021; Williams-García et al., 2014). These phases can be characterized by specific order parameters related to synchronization. As υ increases, a qC phenomenon emerges in the space, where the quasi-critical region is defined by all combinations of (υ, τ, κ) whose susceptibility values are at least half-maximum. Cross-over behaviors (i.e., a generalization of phase transition with finite susceptibility) emerge when the quasi-critical region has overlaps with at least two phases (Fosque et al., 2021; Williams-García et al., 2014). In Figure 2D, we show this qC phenomenon in details.

To study self-organized criticality (SOC), we consider the conserved neural dynamics generated by integrate-and-fire neurons (Levina et al., 2007). The order parameter is active neuron density ρ, whose dynamics is controlled by parameter 〈θ〉_A, the averaged coupling strength θ between activated neurons and their postsynaptic neurons (here A denotes the set of activated neurons). In specific cases, the considered neural dynamics may self-organize to the critical point under the joint effects of excitatory and inhibitory neurons, neural spiking processes (activation and silence), as well as neural plasticity. In Figure 2E, we conceptually illustrate a case where these endogenous mechanisms enable the brain to self-organize to the criticality between absorbing and active phases.

To analyze self-organized quasi-criticality (SOqC), we consider the nonconserved neural dynamics affected by two homeostatic adaptation processes, that is, the short-term depression of inhibition and the spike-dependent threshold increase. These processes are controlled by $yˆ$⁠, the maximum inhibitory coupling strength, as well as τ_x and τ_y, the decay timescales of neural activation threshold increase and synaptic depression. These control parameters affect neural activation threshold x and inhibition strength y to shape neural dynamics states (e.g., the active neuron density ρ). With appropriate x, y, and ρ, neural avalanches with power law behaviors will occur to indicate the criticality between an asynchronous phase (stochastic oscillations) and a synchronous phase (periodic oscillations). According to Girardi-Schappo et al. (2021), x and ρ self-organize to their appropriate values through quasi-critical fluctuations under biologically reasonable conditions (i.e., τ_x ≫ 1) while y hovers around the expected value. Additional fine tuning of y based on exogenous mechanisms are necessary to place neural dynamics at the perfect criticality. Meanwhile, synaptic homeostasis is discovered as constantly canceled by the variation of the activation threshold, impeding neural dynamics from self-organizing to perfect criticality. In Figure 2F, we conceptually illustrate the defined SOqC phenomenon in a similar manner of Figure 2D and Figure 2E. As for the precise description of quasi-critical fluctuations, one can see Girardi-Schappo et al. (2021) for details.

To this point, we have conceptually introduced the phenomenological properties of brain criticality. To verify the hypothetical brain criticality, one needs to learn about analytic brain criticality theories and the properties of neural avalanche predicted by them. Below, we present accessible expositions of these theoretical foundations.

One of the main challenges faced by neuroscientists in studying ordinary criticality (OC), quasi-criticality (qC), self-organized criticality (SOC), and self-organized quasi-criticality (SOqC) is how to understand their theoretical relations (Girardi-Schappo, 2021). Overcoming this challenge is crucial for understanding why we can verify the existence of different types of brain criticality with certain theoretical tools. To present a concise and thorough review, we first focus on brain criticality between absorbing and active phases, where we generalize the ideas in Bonachela and Muñoz (2009) and Buendía, di Santo, Bonachela, and Muñoz (2020a) to present a possible framework for unification.

Because the critical point ν_c = −$ab$ is not necessarily a steady state, it can be disturbed by diverse factors (e.g., by stimuli). Unless there exist certain ideal exogenous mechanisms that persistently enlarge g(⁠$x→$⁠, t) whenever f(⁠$x→$⁠, t) increases, the fine tuning of neural dynamics cannot cancel the effects of f(⁠$x→$⁠, t). Consequently, the fine tuning process may only enable the brain to reach a quasi-critical region where the susceptibility of neural dynamics is relatively large. The initial OC vanishes and is replaced by qC.

Taken together, neuroscientists can approximately verify the existence of brain criticality in the space of absorbing and active phases with specific tools coming from directed percolation theory. This is because OC, qC, SOC, and SOqC exhibit or approximately exhibit directed percolation behaviors under certain conditions. The verification may be inaccurate since the approximation holds conditionally. As for the brain criticality between asynchronous and synchronous phases, however, the universality class properties become rather elusive because an analytic and complete theory of synchronous phase transitions in the brain remains absent yet (see Buendía, Villegas, Burioni, & Muñoz, 2021; di Santo et al., 2018; for early attempts). Although some behaviors of absorbing phase transitions can be observed in synchronous phase transitions (e.g., see Buendía et al., 2021; di Santo et al., 2018; Fontenele et al., 2019; Girardi-Schappo et al., 2021), there also exist numerous differences between them (e.g., see Buendía et al., 2021; Fontenele et al., 2019; Girardi-Schappo et al., 2021). As suggested by Dalla Porta and Copelli (2019), it remains elusive if directed percolation properties are applicable, at least conditionally applicable, to analyzing synchronous phase transitions. More explorations are necessary in the future.

There are numerous properties of brain criticality predicted by directed percolation theory, among which, neural avalanche exponents (the power law exponents of lifetime and size distributions), scaling relation, universal collapse shape, and slow decay of autocorrelation are applicable in both analytic derivations and statistical estimations from empirical data. These properties are our main focus. For convenience, we summarize important glossaries and symbol conventions before we discuss theoretical details (Table 2).

As we have mentioned above, neural avalanches are expected to exhibit power law properties in their lifetime and size distributions when the brain is at the critical point (Hinrichsen, 2000; Larremore et al., 2012; Lübeck, 2004). Therefore, it is pivotal to confirm the detailed values of neural avalanche exponents. To analytically derive these exponents, one can consider critical branching process (di Santo, Villegas, Burioni, & Muñoz, 2017; García-Pelayo, Salazar, & Schieve, 1993; Gros, 2010; Harris, 1963; Otter, 1949), neural field theory (Robinson, 2021), mean-field Abelian sandpile models (Janowsky & Laberge, 1993; D. S. Lee, Goh, Kahng, & Kim, 2004), and avalanches in networks (Larremore et al., 2012). The key idea to derive neural avalanche exponents shared by these existing theories is to confirm the explicit forms of 𝒫_T(t) and 𝒫_S(s), the probability distributions of the lifetime and size of neural avalanches, under ideal conditions (e.g., when the maximum lifetime and size are unlimited and can be infinitely large). In real cases where lifetime and size are restricted because the brain is a finite system, slight deviations from idea values may be observed but theoretical derivations of neural avalanche exponents principally hold.

The derivations of avalanche exponents α = 2 and β = $32$ are nontrivial. However, few neuroscience studies elaborate on these details, impeding researchers from understanding the theoretical foundations of brain criticality in the brain. The importance of these derivations is beyond the detailed values of avalanche exponents since they reveal the fundamental properties of neural dynamics (Cocchi et al., 2017; di Santo et al., 2017; Girardi-Schappo, 2021). In Box 1, we sketch an original idea to derive these avalanche exponents in the terminology of neuroscience. In Figure 3A, we present graphical illustrations of our idea in Box 1.

There are three important things to remember. First, the lifetime exponent α = 2 and size exponent β = $32$ can only be treated as ideal exponents under mean-field assumptions of directed percolation. There are numerous factors, such as granularity, network topology, and neural dynamics variability, that can be considered in derivations to affect the detailed values of avalanche exponents (Bonachela & Muñoz, 2009; Girardi-Schappo, 2021). In Table 3, we summarize the possible intervals of α and β in empirical neural data. Second, α and β alone are not sufficient to verify the existence of brain criticality. Even when the actual values of α and β in empirical data are exactly equal to theoretical predictions, they may still not satisfy the scaling relation and universal collapse. Meanwhile, as we shall discuss later, estimating α and β in practice is statistically error prone. Third, one cannot confirm or disprove a detail type of brain criticality based on α and β unless additional information is provided. Although four types of brain criticality exhibit (e.g., OC) or approximately exhibit (e.g., qC, SOC, and SOqC) directed percolation behaviors under certain conditions, these preconditions are difficult to verify in practice.

In Lübeck (2004), one can further learn about how brain criticality is mapped to a directed percolation transition characterized by ordinary critical exponents. Meanwhile, one can see how to connect these neural avalanche exponents with second-order phase transition exponents (Lübeck & Heger, 2003).

In Figure 3D, we illustrate examples of autocorrelation slow decay in critical cases and compare them with noncritical cases. Compared with other properties previously discussed, a slow autocorrelation decay can be readily verified by conventional data fitting. However, we need to note that one should not confirm or reject the possibility of brain criticality only based on the decay characteristic of autocorrelation in Equations 53–55. This is because Equations 53–55 only serve as the approximate descriptions of long-range correlations at criticality. The strict criterion χ, ξ → 0 is rarely seen in empirical data while the determination of whether χ and ξ are sufficiently small in the looser criterion is relatively subjective.

In summary, we have reviewed the physical foundations of identifying and characterizing criticality in the brain. Based on these analytic derivations, we attempt to present systematic explanations of what is brain criticality and how to identify potential criticality in neural dynamics. Nevertheless, physical theories alone are not sufficient to support neuroscience studies because the implementation of these theories on empirical data is even more challenging than the theories themselves. To overcome these challenges, one needs to learn about statistic techniques to computationally estimate brain criticality from empirical data.

While most properties of neural avalanches analytically predicted by the physical theories of brain criticality can be estimated by conventional statistic techniques, there exist several properties that frequently imply serious validity issues and deserve special attention. Below, we discuss them in detail.

Another error-prone step is the calculation and evaluation of the universal collapse shape, which is closely related to scaling relation analysis. Deriving the collapse shape from empirical data may be problematic because the goodness evaluation of collapse shape is rather subjective (e.g., depends on personal opinions about whether all data follows function 𝓗(·) in Equation 48) in most cases (Marshall et al., 2016). Although important efforts have been devoted to quantify if a given dataset exhibits shape collapse (Bhattacharjee & Seno, 2001; Shaukat & Thivierge, 2016), common approaches in practice still depend on specific shape collapse algorithms that search potential scaling parameters (e.g., γ in Equation 48) in a data-driven manner (Marshall et al., 2016). In these algorithms, thresholding on neural avalanches before analyzing the shape collapse is a standard preprocessing scheme to control noises (e.g., set an avalanche size threshold and remove all data below the threshold) (Marshall et al., 2016; Papanikolaou et al., 2011). While experimental noises are partly limited, unexpected excursions of scaling parameters away from theoretical predictions may occur after thresholding as well (Laurson, Illa, & Alava, 2009). To our best knowledge, the effects of thresholding on brain criticality analysis are nonnegligible. Although being highly practical, thresholding may lead to significant transient effects to cloud the true scaling property (Villegas, di Santo, Burioni, & Muñoz, 2019). Therefore, any qualitative evaluation of collapse shape after thresholding is questionable regardless of its practicability. Although an ideal approach requires further explorations, we suggest researchers to consider the following methods: (1) estimate γ by area fitting (e.g., follow Equation 47 in scaling relation analysis) and collapse shape fitting (e.g., follow Equation 48 in collapse shape analysis), respectively; (2) compare between γ derived by these two kinds of fitting and measure the difference. Search for a threshold that minimizes the difference (e.g., makes variation amplitude < 1%) and maintains a reasonable sample size (e.g., maintains > 80% samples); (3) given the chosen threshold and corresponding γ, measure the difference (e.g., the dynamic time warping; Keogh & Pazzani, 2001) between 〈S(t|T)〉T^1−γ derived on neural avalanches with different lifetime T in the plot of 〈S(t|T)〉T^1−γ vs. $tT$⁠. Denote the shape collapse error as the averaged difference. Combining these three steps, researchers may partly avoid the errors implied by subjective judgment. Similar ideas can be seen in Marshall et al. (2016).

Finally, the analysis of slow decay of autocorrelation is also error-prone in practice. Although this approach is practical and has been extensively applied (e.g., Pausch et al., 2020; Wilting & Priesemann, 2019), the criterion to determine if the decay is truly slow (e.g., χ > 0 in Equation 54 and ξ > 0 in Equation 54 are sufficiently small) remains ambiguous. A fixed criterion (e.g., χ, ξ < 0.5) may serve as an explicit condition of a slow decay. However, this presupposed criterion may deviate from real situations. For instance, the baseline of decay rate in a noncritical brain may be essentially high (e.g., χ, ξ > 10). Even though the decay rate drops significantly when the brain becomes critical (e.g., χ, ξ ≃ 1), the presupposed criterion is still unsatisfied and leads to unnecessary controversies on criticality hypothesis. Given that ξ is principally independent from spatial subsampling on neurons or brain regions at criticality (Pausch et al., 2020), we suggest researchers to consider the following approaches: (1) do spatial subsampling in both critical and non-ritical brains to derive two groups of χ or ξ (one group for criticality and another group for noncriticality); (2) use appropriate statistic tests (e.g., choose t test; Kanji, 2006; Kolmogorov–Smirnov test; Berger & Zhou, 2014; or Wilcoxon-Mann-Whitney test; Fay & Proschan, 2010; according to sample distribution properties) to verify if two groups of χ or ξ belong to different distributions. Test if the expectation and variance of χ or ξ drops significantly from the noncritical group to the critical group according to certain effect sizes.

In summary, statistic techniques bridge between brain criticality theory and empirical data. However, misconception and misuse of statistical analyses of neural avalanche properties still occasionally appear in practice. Although existing techniques remain imperfect in brain criticality analysis, we wish that our discussion may inspire future studies.

Ever since brain criticality was introduced into neuroscience, it is frequently speculated as contradictory with other traditional neuroscience hypotheses, such as the conjectured hierarchical processing characteristic of neural information (Felleman & Van Essen, 1991) and the asynchronous-irregular characteristic of neural dynamics (e.g., neurons spike independently in Poisson manners; Burns & Webb, 1976; Softky & Koch, 1993; Stein, Gossen, & Jones, 2005). Meanwhile, the differences between brain criticality and scale-free neural dynamics (Chialvo, 2010; He, 2014; Martinello et al., 2017) are frequently neglected. Before we put an end to our review, we discuss the relations between brain criticality and these neuroscience theories.

The hierarchical processing of neural information (Felleman & Van Essen, 1991) is initially speculated to contradict critical neural dynamics since hierarchical topology has not been used as an explicit condition to analytically derive criticality (e.g., see derivations in di Santo et al., 2017; García-Pelayo et al., 1993; Gros, 2010; Harris, 1963; Janowsky & Laberge, 1993; Larremore et al., 2012; D. S. Lee et al., 2004; Otter, 1949; Robinson, 2021). On the contrary, random graphs without strict hierarchical structures seem to be more widespread in criticality derivations. Recently, this speculation has been challenged by numerous discoveries of the facilitation effects of hierarchical modular structures on critical dynamics (E. J. Friedman & Landsberg, 2013; Kaiser & Hilgetag, 2010; Rubinov et al., 2011; S. Wang & Zhou, 2012). Meanwhile, computational analysis suggests that information transmission in standard feed-forward networks is maximized by critical neural dynamics (Beggs & Plenz, 2003). Parallel to neuroscience, a recent machine-learning study empirically observes and analytically demonstrates that artificial neural networks, a kind of hierarchical structure, self-organize to criticality during learning (Katsnelson, Vanchurin, & Westerhout, 2021). Therefore, brain criticality is not necessarily contradictory with hierarchical information processing, yet more analyses are required to understand how brain criticality affects hierarchical processing schemes.

Brain criticality and the asynchronous-irregular (AI) characteristic may correspond to distinct encoding schemes in the brain (Girardi-Schappo et al., 2021; Wilting & Priesemann, 2019). While AI characteristic can minimize redundancy (Atick, 1992; Barlow et al., 1961; Bell & Sejnowski, 1997; Van Hateren & van der Schaaf, 1998) to improve neural encoding (Van Vreeswijk & Sompolinsky, 1996), brain criticality may optimize encoding performance utilizing a series of reverberations of neural activities (Bertschinger & Natschläger, 2004; Boedecker, Obst, Lizier, Mayer, & Asada, 2012; Del Papa, Priesemann, & Triesch, 2017; Haldeman & Beggs, 2005; Kinouchi & Copelli, 2006; Shew & Plenz, 2013; X. R. Wang, Lizier, & Prokopenko, 2011). The coexistence of empirical evidence of AI and brain criticality characteristics initially confuses researchers since these characteristics are hypothesized as contradictory to each other (Girardi-Schappo et al., 2021; Wilting & Priesemann, 2019). In experiments, AI characteristic is supported by small correlations between the spike rates of different neurons in cortical microcircuits (Cohen & Kohn, 2011; Ecker et al., 2010) and exponential distributions of interspike intervals (Carandini & Stevens, 2004; Kara, Reinagel, & Reid, 2000) while brain criticality characteristic is observed in neural dynamics recorded from multiple species (e.g., awake monkeys; Petermann et al., 2009, anesthetized rats; Gireesh & Plenz, 2008, slices of rat cortices; Beggs & Plenz, 2003; Shew et al., 2009, and humans; Poil et al., 2008). A recent study demonstrates that cortical spikes may propagate at somewhere between perfect criticality (e.g., OC or SOC depending on whether underlying mechanisms are exogenous or endogenous) and full irregularity (Wilting & Priesemann, 2019), similar to the cases of qC and SOqC. Meanwhile, it is known that stimulus drives suppress irregularity in neural activities (Molgedey et al., 1992). These results imply that brain criticality may not necessarily contradict AI characteristic. On the contrary, they may coexist when stimulus drives are too weak to disturb brain criticality (e.g., OC or SOC) and suppress AI characteristics. In our previous discussions, we have analytically proven that neural avalanche exponents, the fundamental properties of brain criticality, can still be derived under the condition of independent neuron activation, a key feature of AI characteristics (Wilting & Priesemann, 2019). This result suggests that brain criticality and AI characteristics do not contradict each other. As for the case where stimulus drives are nonnegligible, a recent study presents an elegant theory to prove that two homeostatic adaptation mechanisms (i.e., the short-term depression of inhibition and the spike-dependent threshold increase) enable synaptic excitation/inhibition balance, AI characteristics, and SOqC to appear simultaneously in the same neural dynamics (Girardi-Schappo et al., 2021). Similarly, it is suggested that neural dynamics with criticality or with AI characteristics can be generated by the same neural populations if the synaptic excitation/inhibition balance is fine tuned appropriately (J. Li & Shew, 2020).

Neural dynamics with power law behaviors is a necessary but insufficient condition of brain criticality. This property is frequently neglected in practice. Power law behaviors are widespread in nature because they can be generated by diverse mechanisms, such as exponential curve summation and preferential attachment (Mitzenmacher, 2004; Reed & Hughes, 2002). It has been reported that the aggregate behaviors of noncritical stochastic systems may also create scale-free dynamics within a limited range (Touboul & Destexhe, 2010, 2017). In the brain, the generic scale-free properties can be implied by neutral dynamics, a kind of dynamics where the population size of neutral individuals (or dynamically homogeneous individuals) does not tend to increase or decrease after adding a new individual that is neutral to existing ones (see neutral theories for further explanations; Blythe & McKane, 2007; Liggett, 2006). This generic property can generate power law neural avalanches without criticality (Martinello et al., 2017). Meanwhile, bistability phenomena, a kind of fine-tuned or self-organized discontinuous phase transitions with limit cycles rather than critical points, can also create neural dynamics with power law properties (Buendía et al., 2020a; Cocchi et al., 2017; di Santo, Burioni, Vezzani, & Muñoz, 2016). Consequently, we emphasize that neural avalanche exponents alone are insufficient to prove or disprove any brain criticality hypothesis. These power law exponents are meaningless for brain criticality hypothesis unless they satisfy the scaling relation.

Given what have been reviewed above, we arrive at a point to conclude on the current progresses and limitations in establishing theoretical foundations of different types of brain criticality, that is, ordinary criticality (OC), quasi-criticality (qC), self-organized criticality (SOC), and self-organized quasi-criticality (SOqC). As we have suggested, an inescapable cause of various controversies is the nontriviality of physical theories that analytically derive brain criticality and statistical techniques that estimate brain criticality from empirical data. Immoderate omitting of these theoretical foundations, especially their imperfection, in practice may lead to confusions on the precise meaning, identification criteria, and biological corollaries of brain criticality. To address these problems, we have introduced the mainstream theoretical foundations of brain criticality, reformulated them in the terminology of neuroscience, and discussed their error-prone details.

Thanks to the increasing efforts devoted to improving theoretical frameworks of criticality in the brain, researchers have seen substantial progresses in explaining various important neuroscience problems, including but not limited to efficient cortical state transitions (Fontenele et al., 2019), dynamic range maximization in neural responses (Kinouchi & Copelli, 2006; Shew et al., 2009), and optimization of information transmission and representation (Shew et al., 2011). These advances have been comprehensively reviewed by existing works (Beggs, 2007; Chialvo, 2010; Cocchi et al., 2017; Hesse & Gross, 2014; Muñoz, 2018; Shew & Plenz, 2013) and are not discussed in detail in our review. The benefits of studying brain criticality, as we have suggested, lay in the possibility to analyze brain function characteristics with numerous statistical physics theories relevant to brain criticality, such as directed percolation (Hinrichsen, 2000; Lübeck, 2004), conserved directed percolation (Bonachela et al., 2010; Bonachela & Muñoz, 2008), and dynamical percolation theories (Bonachela et al., 2010; Steif, 2009). These theories characterize the brain as a physical system with avalanche behaviors, enabling researchers to analyze various propagation, synchronization, and correlation properties of neural dynamics (e.g., continuous phase transitions). These properties intrinsically shape neural information processing (e.g., encoding; Bertschinger & Natschläger, 2004; Boedecker et al., 2012; Del Papa et al., 2017; Haldeman & Beggs, 2005; Kinouchi & Copelli, 2006; Shew & Plenz, 2013; X. R. Wang et al., 2011; transmission; Shew et al., 2011, and memory; Haldeman & Beggs, 2005; Krotov & Hopfield, 2020) and can be readily recorded in neuroscience experiments. Therefore, the nonequilibrium dynamic processes and potential criticality defined by statistical physics theories are highly applicable to characterizing brain functions. As we have discussed in Figure 2, researchers can consider diverse brain criticality phenomena in neural dynamics by defining different control (e.g., the balance between excitatory and inhibitory neurons; Hardstone et al., 2014; Poil et al., 2012) and order (e.g., active neuron density; Dalla Porta & Copelli, 2019) parameters, corresponding to multifarious biological mechanisms underlying neural dynamics (e.g., synaptic depression; Levina et al., 2007). Meanwhile, the definition of neural avalanches can flexibly change from neural spikes and local field potentials to global cortical oscillations. The flexibility of brain criticality and neural avalanche definitions enables researchers to analyze different functional properties on distinct organizational levels in the brain.

The limited theoretical foundations of brain criticality in the brain, however, have become irreconcilable with their increasingly widespread applications. Although the analytic theories of brain criticality have solid physics backgrounds, they needlessly become black boxes for neuroscientists in practice. On the one hand, the details of brain criticality theory frequently experience immoderate neglecting in neuroscience studies. On the other hand, to our best knowledge, there is no accessible and systematic introduction of the statistical physics foundations of brain criticality in the terminology of neuroscience yet. These obstacles severely impede neuroscientists from comprehensively understanding brain criticality, eventually motivating us to present this review. When we turn to bridging between brain criticality theories and experiments, one can find nonnegligible gaps separating between theories and experiments. Although numerous biological factors (e.g., neural plasticity; De Arcangelis et al., 2006; Levina et al., 2007, 2009; membrane potential leakage; Levina et al., 2007; Millman et al., 2010; Rubinov et al., 2011; Stepp et al., 2015; retro-synaptic signals; Hernandez-Urbina & Herrmann, 2017; spatial heterogeneity; Girardi-Schappo, Bortolotto, Gonsalves, Pinto, & Tragtenberg, 2016; Moretti & Muñoz, 2013; and refractory period; Fosque et al., 2021; Williams-García et al., 2014) have been considered in brain criticality characterization, existing theories more or less suffer from deviations from actual neural system properties. For instance, the requirements of conserved neural dynamics and an infinite timescale separation between the dissipation and drive processes required by SOC may not be biologically realistic (Muñoz, 2018). The implicit requirement of a sufficiently large system size by the mean-field theories of brain criticality may not always be satisfied during neural avalanche recording, implying nonnegligible finite size effects (Girardi-Schappo, 2021). Meanwhile, to precisely verify the existence of a detailed type of brain criticality (e.g., confirm the actual universality class) in empirical neural data is principally infeasible. As we have explained, the common criteria used for brain criticality hypothesis verification, such as neural avalanche exponents (Bauke, 2007; Clauset et al., 2009; Deluca & Corral, 2013; Marshall et al., 2016; Yu et al., 2014; scaling relation; Lübeck, 2004; Lübeck & Heger, 2003; universal collapse shape; Bhattacharjee & Seno, 2001; Laurson et al., 2009; Marshall et al., 2016; Papanikolaou et al., 2011; and slow decay of auto-correlation; Pausch et al., 2020; Wilting & Priesemann, 2019), are derived according to directed percolation theory under mean-field assumptions. Among four types of brain criticality in absorbing phase transitions, only OC originally belongs to directed the percolation universality class, while qC, SOC and SOqC conditionally exhibit directed percolation behaviors. In most cases, one can only verify if the brain is plausibly at criticality (e.g., whether neural avalanches obey universal collapse and have the power law exponents that satisfy the scaling relation). When observed neural avalanche exponents depart from their mean-field approximation results but still satisfy the scaling relation, there may exist an OC phenomenon affected by nonmean-field factors (e.g., network topology; Girardi-Schappo, 2021) or exist a certain qC, SOC, or SOqC phenomenon caused by diverse mechanisms. Additional information of neural dynamics properties is inevitably required to determine the category belonging to the hypothesized brain criticality, which poses daunting challenges to neuroscience experiment designs. Moreover, the potential validity issues of applying the theoretical tools derived from directed percolation theory to verify brain criticality in synchronous phase transitions deserve special attention (for similar opinions see Dalla Porta & Copelli, 2019). It remains controversial if absorbing and synchronous phase transitions robustly share specific features (see reported similarities; Buendía et al., 2021; di Santo et al., 2018; Fontenele et al., 2019; Girardi-Schappo et al., 2021; and differences; Buendía et al., 2021; Fontenele et al., 2019; Girardi-Schappo et al., 2021). Any speculated relations between these two kinds of critical phenomena should be tested with caution. Furthermore, statistic techniques to estimate and verify brain criticality from empirical data are yet imperfect. The estimation of some properties of neural avalanches is error prone in practice and may lead to serious validity issues. Although we suggest compromised solutions to these issues, more optimal approaches are required in future studies.

We submit that this review not only summarizes the latest developments in the field of studying criticality in the brain, but also serves as a blueprint for further explorations. Below, we offer concrete recommendations of future directions.

First, we suggest researchers carefully rethink the theoretical foundations of criticality in the brain. Immoderately omitting these foundations in neuroscience needlessly muddies an already complex scientific field and leads to potential validity issues. While we have presented a self-contained framework of brain criticality to characterize neural dynamics as a physical system with avalanches, plentiful details are uncovered in this article (e.g., the Landau–Ginzburg theory; di Santo et al., 2018) because the statistical physics theories of brain criticality are essentially grand. We recommend researchers to further improve our work and explore a more accessible and systematic reformulation of related physics theories, such as directed percolation, conserved directed percolation, dynamic percolation, and nonequilibrium dynamics, in the context of neuroscience. Moreover, we note that these theories are not initially proposed for brain analysis. It is normal to see gaps between these theories and real situations of the brain. We urge researchers to develop new variants of criticality formalism that is more applicable to the brain or even explore new universality classes of continuous phase transitions.

Second, neuroscience is in urgent need of new physical theories and statistical techniques to bridge between brain criticality hypotheses and experiments. Although existing theories and techniques have become increasingly widespread and cover most of the pivotal details of brain criticality, there remain various limitations, as we have suggested. Specifically, we suggest five potential directions to resolve these problems: (1) combine brain criticality theories with large-scale neural dynamics recording or computation to include more realistic biological details into brain criticality theories and establish a closer connection with experimental observations; (2) try to summarize, standardize, and subdivide these theories according to the concrete biological meanings of brain criticality phenomena, prerequisites of model definitions, and scopes of application—try to avoid abusing or misusing of different brain criticality theories; (3) develop open-source toolboxes of theoretical models and statistical techniques to routinize brain criticality analysis in neuroscience studies (one can see existing efforts to achieve this objective; Marshall et al., 2016); (4) establish open-source, multispecies, and large-scale datasets of neural dynamics recorded from both critical and noncritical brains—validate different statistic techniques of brain criticality estimation and testing on these datasets and, more importantly, confirm appropriate baselines to define the criteria of brain criticality identification (see notable contributions in Girardi-Schappo, 2021); (5) explore new nonequilibrium statistical physics theories for synchronous phase transitions or analytically verify the theoretical validity of directed percolation formulation of synchronous phase transitions.

Third, parallel to neuroscience, the discoveries of critical phenomena in other learning and computation systems also merit attention. Learning or computing at the edge of chaos has been proven as a mechanism to optimize the performance of learners (e.g., recurrent neural networks; Bertschinger & Natschläger, 2004). The well-known residual connections can control the performance degradation of artificial neural networks because they enable networks to self-organize to criticality between stability and chaos to preserve gradient information flows (Yang & Schoenholz, 2017). It is recently demonstrated that any artificial neural network generally self-organizes to criticality during the learning process (Katsnelson et al., 2021). In the future, it would be interesting to explore whether information processing processes in brains and artificial neural networks can be universally characterized by a unified criticality theory.

Overall, we anticipate the potential of well-validated studies of criticality in the brain to greatly deepen our understanding of neural dynamics characteristics and their roles in neural information processing. Laying solid theoretical foundations of studies is the most effective and indispensable path to contributing to this booming research area.

Authors are grateful for discussions and assistance of Drs. Yaoyuan Wang and Ziyang Zhang from the Laboratory of Advanced Computing and Storage, Central Research Institute, 2012 Laboratories, Huawei Technologies Co. Ltd., Beijing, 100084, China.

Yang Tian: Conceptualization; Formal analysis; Investigation; Methodology; Visualization; Writing – original draft; Writing – review & editing. Zeren Tan: Formal analysis; Methodology. Hedong Hou: Formal analysis; Methodology. Guoqi Li: Validation; Writing – review & editing. Aohua Cheng: Formal analysis; Writing – review & editing. Yike Qiu: Validation; Writing – review & editing. Kangyu Weng: Validation; Writing – review & editing. Chun Chen: Validation; Writing – review & editing. Pei Sun: Conceptualization; Project administration; Supervision; Validation; Writing – original draft; Writing – review & editing.

Pei Sun, The Artificial and General Intelligence Research Program of Guo Qiang Research Institute at Tsinghua University, Award ID: 2020GQG1017.

Neurophysics:

A branch of biophysics that develops and uses physics theories to study the neural system.

Brain connectivity:

Refers to the anatomical connectivity formed by synaptic connections and the functional connectivity formed by dynamic interactions among neurons.

Statistical physics:

A branch of theoretical physics that develops mathematical theories of the characterization and approximations of large populations with inherently stochastic natures.

Percolation theory:

A physics theory that characterizes critical phenomena and phase transitions from a probabilistic and geometric perspective.

Hopfield network:

An Ising model of a neural network, which serves as a content-addressable memory system with binary nodes or continuous variables.

Information thermodynamics:

A branch of statistical physics that develops mathematical theories of the exchange between information quantities and thermodynamic quantities.

Dynamic range:

The intensity or amplitude range of stimulus inputs that is encoded by the given neural dynamics.

Fine tuning:

A concept in theoretical physics, which refers to the case where system parameters must be precisely manipulated in order to fit with certain observations.

Discontinuous phase transition:

A kind of phase transition where the first-order derivative of the order parameter diverges when transition happens.