Power-Law Scaling in the Brain Surface Electric Potential

Neuroscience has spent decades mapping the brain's rhythms, such as alpha waves, beta waves, and theta waves, as if the brain were a radio tower broadcasting on specific frequencies. Those rhythms are real, and they relate to behavior: the occipital alpha rhythm, which runs at 8 to 12 hertz, suppresses when you open your eyes. Beta rhythms in the 18 to 25 hertz range quiet down when you move your hand. Researchers have built entire careers reading these bands. However, they may have been studying the melody while missing the underlying instrument entirely. A team at the University of Washington increased the sampling rate high enough to see what was hiding beneath, and what they found changes how we think about what a brain signal actually is. The dominant framework treats the power spectral density, or PSD, which is a measure of how much signal exists at each frequency, as a collection of distinct peaks, each one representing a named rhythm. Miller and colleagues wanted to test a different hypothesis: that above the region of those rhythms, the spectrum follows a smooth mathematical form called a power law. Written out, the hypothesis states that P of f equals A times f to the minus x. This means the power at any frequency equals a constant A, scaled by the frequency raised to a negative exponent. On a log-log plot, that's a straight line with a slope of negative x. The question was whether the brain's high-frequency spectrum actually looks like that, and if so, what is x? The problem was practical before it was scientific. Earlier electrocorticography studies, or ECoG, meaning recordings from electrode arrays placed directly on the cortical surface in epilepsy patients awaiting surgery, were sampled at 1 kilohertz. This effectively truncated the spectrum above around 250 hertz. You can't characterize a high-frequency power law if you can't see the high frequencies. So, Miller and colleagues recorded from twenty patients with subdural platinum electrode arrays covering the lateral frontal, temporal, and parietal cortex. For four of those subjects, they increased the sampling rate to 10 kilohertz. The electrodes were small, with a diameter of 2.3 millimeters and spaced 1 centimeter apart, embedded in silastic. Channels overlying blood vessels or seizure foci were excluded. To reduce common-mode noise, the team applied bipolar rereferencing, expressing each channel as the voltage difference between nearest-neighbor electrode pairs, resulting in ninety-one usable bipolar channels. What they found in the high-frequency regime was strikingly clean. Above approximately 80 hertz, the power spectral density follows a power law with an exponent of exactly 4.0. The four subjects recorded at 10 kilohertz produced fitted exponents of 3.97, 3.94, 3.97, and 4.02. Across 151 individual electrode-pair fits, the mean exponent was 4.01, with a standard deviation of just 0.13. Miller and colleagues report the scaling index as chi equals 4.0 plus or minus 0.1. That's a remarkably tight number, and it held across different subjects, different regions of cortex, and different levels of neural activity. There's also a feature at the bottom of this high-frequency range that deserves attention. The spectrum doesn't smoothly extend down into the low frequencies. Instead, there's a bend, a "knee," at a corner frequency, referred to as f-zero, of approximately 75 hertz, estimated more precisely across the 1 kilohertz dataset as 77 hertz, with a standard deviation of 14 hertz. Below that knee, the power law changes slope. Above it, the exponent is 4. That knee implies a characteristic timescale. If you take one divided by two pi times f-zero, you get something on the order of 2 to 4 milliseconds. That number turns out to be important and will reappear in the model. Now comes the finding that makes this genuinely elegant. When Miller and colleagues used a simple finger-movement task to drive cortical activity in five of their subjects, and then compared the spectrum during movement to the spectrum at rest, the power law didn't change shape. The exponent remained at 4. What changed was A, the amplitude — the entire broadband spectrum shifted upward uniformly, like turning up the volume on a song without touching the equalizer. The measured shift in exponent during movement was 0.03 plus or minus 0.09, which is statistically indistinguishable from zero, with a p-value of about 0.104. However, the active-to-rest amplitude ratio across channels was 1.76 on average, with a standard deviation of 0.31, and the effect was unmistakably real, with a p-value below ten to the negative fourteen. This stands in sharp contrast to what the named rhythms are doing at the same time. The alpha and beta oscillations actually suppress during movement; they decrease while the broadband signal increases. Two different processes are moving in opposite directions, coexisting in the same recording. The team had to remove the alpha and beta peaks with a principal-component method before they could cleanly quantify the broadband change. Once those narrow-band rhythms were stripped away, the broadband amplitude increase was unambiguous and consistent across every channel tested. So, what biological process could produce this? Miller and colleagues built a minimal simulation to find out. They modeled a single pyramidal neuron receiving inputs from 6,000 synapses. Each synapse generated a Poisson-distributed train of presynaptic spikes, and each incoming spike produced a postsynaptic current with a fast rise and an exponential decay. All 6,000 of those currents were summed and integrated across the dendritic membrane, with a leakage time constant of around 100 milliseconds. That summed signal, the dendritic current dipole, which an ECoG electrode is actually sensitive to, was then converted to a power spectral density. The result was a spectrum with a knee near 70 hertz and a high-frequency tail that follows one over frequency to the fourth power. P of f is proportional to f to the minus four, just like the data. The reason this works is mathematical. When you superpose many independent random events — Poisson arrivals — each convolved with the same exponential decay kernel, you naturally get a characteristic corner frequency and a power-law tail above it. The corner frequency corresponds to the decay timescale of the synaptic currents, which in this simulation was about 2.3 milliseconds. That's the same 2 to 4 millisecond timescale implied by the knee in the data. The model also reproduces the volume-knob behavior. Simulating Poisson input rates of 15, 30, and 60 spikes per synapse per second, the fitted high-frequency exponent remained at 4.0 at all three rates. However, the amplitude scaled with firing rate: A at 30 spikes per second was about 1.96 times A at 15 spikes per second, and A at 60 spikes per second was 4.03 times the 15-spike baseline. That's close to linear — double the firing rate results in roughly double the amplitude. The observed experimental amplitude ratio of about 1.76 between active and rest cortex is consistent, the authors suggest, with roughly a doubling of mean population input rate beneath each electrode. Miller and colleagues also address a tempting alternative explanation: self-organized criticality, the idea that the brain sits at a phase transition and generates power laws as a consequence of critical dynamics. They argue the data don't require it. Because the measured exponent is close to 4, which is an integer multiple of 2, a simpler story works. Filtered noise from randomly arriving synaptic events, convolved with an exponential kernel, provides a clean spectrum that follows the fourth power without invoking complex critical-state machinery. The conclusion drawn by the paper is conceptual as much as empirical. Alongside the synchronous, narrowband oscillations that neuroscience has long focused on, there exists an asynchronous, broadband, scale-free signal that tracks the mean population firing rate. These two processes coexist. The rhythms and the broadband power law ride together in the same recording, but they respond differently to the same behavioral event and likely reflect different aspects of cortical computation. Broadband power shifts, Miller and colleagues note, have been shown to capture the timing of finger movements with higher fidelity than narrow-band approaches and to correlate more tightly with mean firing rate than any single frequency band. The open question they leave behind is the right one to carry with you: what determines the amplitude A? What controls those population-level changes in broadband power? The exponent, it turns out, is fixed by biology. The volume knob is where the action is. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.