March 28, 2026 by Tejasri Gururaj, Phys.org
Collected at: https://phys.org/news/2026-03-human-brain-critical.html
A recent study published in Physical Review Letters reveals that many widely used signatures of criticality in brain data may be statistical artifacts. They propose a more robust framework that, when applied to whole-brain fMRI data, confirms the brain operates near, but not exactly at, a critical point.
Neuroscientists have long found the idea fascinating—that the brain operates near a “critical point,” a phase transition between stable and chaotic dynamics. Theory suggests this sweet spot enhances computational flexibility, dynamic range, and sensitivity to inputs. Evidence has mounted over the years from neural recordings showing approximate scale invariance and power-law behavior across spatiotemporal scales.
The concept has even influenced AI, particularly reservoir computing, where networks near the “edge of chaos” tend to perform best. However, the field faces a persistent concern: are these criticality signatures intrinsic to the brain’s recurrent dynamics, or do external inputs and data limitations shape them?
Two common features of neural recordings—temporally autocorrelated signals and limited data sampling—can mimic the statistical fingerprints of criticality, even in systems with no genuine collective dynamics whatsoever.
Phys.org spoke to Rubén Calvo Ibáñez, a Ph.D. student at Universidad de Granada and one of the co-authors of the study. “I’ve always been drawn to fundamental questions—how complicated behavior emerges from simple rules. What excited me about complex systems and non-equilibrium physics is that you can bring those tools to messy, real biological data, like brain activity, and still ask principled questions.”
Spurious signatures
To detect criticality in neural data, researchers typically look for power-law scaling patterns, where activity looks statistically similar across many different scales. This is done in two main ways: by analyzing the eigenvalue spectrum of the covariance matrix of brain activity (a PCA-based approach), and through phenomenological renormalization group (PRG) analysis, which tracks how activity statistics change as neurons are progressively grouped into larger clusters.
Both methods, however, share a blind spot. When neural signals are autocorrelated, meaning they vary slowly and smoothly over time rather than fluctuating independently, these scaling signatures can appear even in systems with no genuine collective dynamics. Pair that with limited sampling, where the number of recorded time points is small relative to the number of brain regions being studied, and the problem compounds.
“The main artifact we explore is a combination of temporal correlations and subsampling. To test this, we built a simple model of brain activity with no connectivity between regions. In such a disconnected system, there is no mechanism for collective dynamics—yet if the inputs each region receives introduce long autocorrelation times, the apparent scaling exponents can vary continuously with that correlation time,” said Calvo.
This is particularly problematic for fMRI, where the bold signal is inherently slow and recording sessions are short, making it an ideal breeding ground for spurious criticality signatures.
A framework to test criticality
To separate real criticality from artifacts, the team built and extended two theoretical models, each designed to isolate a different piece of the puzzle.
The first is a linear recurrent firing-rate model, a standard tool in computational neuroscience, where each brain region influences others through interconnected feedback. A key parameter, g, controls the overall coupling strength, tuning the network from stable, quickly-damped activity toward the edge of instability. Their key insight was that time coarse-graining is mathematically equivalent to driving the network with autocorrelated “colored” noise, making scaling signatures sensitive to preprocessing choices.
As a deliberate counterexample, they also studied a system with g set to zero, meaning no interactions between regions whatsoever. Driving each independent region with slow noise under limited recording conditions was enough to produce covariance statistics with fake power-law tails, statistically indistinguishable from those of a genuinely critical network.
To cut through these artifacts, the framework relies on three practical tools. First, time-shift randomization, where each region’s timeline is shuffled independently, preserving slow fluctuations while destroying any genuine coordination between regions. Second, data pooling across participants, which increases the effective number of time points and reduces sampling error. Third, exponent matching, checking whether fMRI scaling signatures align with the recurrent model’s predictions rather than the artifact baseline.
“What has been scrutinized much less systematically is whether other commonly used scaling signatures can also be produced by noncritical mechanisms. Our contribution is to provide that missing critique and, importantly, a practical way to tell apart genuine collective dynamics from artifacts,” said Calvo.
The framework was then applied to the LEMON dataset, comprising resting-state fMRI scans from 136 healthy participants, 183 brain regions, and roughly ten-minute recording sessions per subject.
What the data revealed
When applied to the pooled fMRI data, the framework delivered a clear verdict. Genuine near-critical signatures emerged at the population level, with an effective coupling strength of approximately 0.88, where 1.0 marks the critical point. In other words, when brain activity is analyzed collectively across participants rather than individually, the group-level dynamics sit close to, but safely below, the critical threshold.
After time-shift randomization, those signatures collapsed almost entirely, confirming that what remained in the original data reflected true collective dynamics rather than artifacts. The extracted scaling exponents closely matched predictions from the recurrent firing-rate model, suggesting the near-critical behavior stems from reverberant network activity rather than structured inputs. Notably, the brain sits slightly below the critical point rather than exactly at it.
“Operating near a critical point can retain many of the proposed computational benefits, such as rich multiscale collective modes and strong but controllable amplification, while avoiding the drawbacks of sitting exactly at criticality, where small perturbations can lead to instability, runaway activity, or reduced robustness,” said Calvo.
Looking ahead, the team hopes to build connectome-informed models linking criticality signatures directly to the brain’s structural architecture. They also want to test how the distance to criticality shifts with age, disease, or cognitive state. The framework itself, they note, is broadly applicable well beyond neuroscience, anywhere claims of near-critical dynamics are made.
Publication details
Rubén Calvo et al, Robust Scaling in Human Brain Dynamics Despite Correlated Inputs and Limited Sampling Distortions, Physical Review Letters (2026). DOI: 10.1103/36v9-wtm8.
Journal information: Physical Review Letters
Leave a Reply