Generating Deep Hierarchical Amplitude Modulation in Cortical Circuits

The brain's own wiring can manufacture the statistical signature of rogue waves — not by accident, but through a specific architectural trick in how cortical columns are connected.

The paper starts from a framework called hierarchical amplitude modulation (HAM): the idea that the brain encodes information in nested layers, where slow rhythms (delta, theta) multiplicatively modulate the amplitude of faster ones (gamma), and those products compound across layers. A deep enough product of this kind produces a heavy-tailed amplitude distribution — meaning rare, outsized surges are not exponentially suppressed but follow a power law. These are the neural equivalent of rogue waves. The puzzle is that a standard neural mass model — the workhorse computational unit for a cortical column — contains no explicit multiplier. It has linear synaptic filtering, membrane summation, and a single sigmoid (S-shaped) firing-rate function. Where does the multiplication come from?

The answer is elegant and sits in plain sight: the dendrite adds, and the sigmoid multiplies. On the sub-threshold, approximately exponential flank of the sigmoid, summing several slow inputs at the membrane is mathematically equivalent to multiplying their envelope contributions together. The paper writes this out explicitly: σ(v₀ + Σᵢ sᵢ) ≈ σ₀ · Πᵢ(1 + κsᵢ). The sum at the membrane becomes a product at the output. No extra circuitry needed — the curvature of the transfer function does it.

This immediately dictates the right architecture. To get a deep product on a gamma carrier, you need many slower-band inputs arriving directly at that carrier's population — not passed through a serial chain of columns. A serial chain (delta → theta → alpha → gamma, each feeding the next) nests additively and saturates: the amplitude is provably bounded at 1/(1−m) regardless of depth, and the tail stays thin. The network architecture, by contrast, lets delta, theta, alpha, and beta all land simultaneously on the gamma column's membrane, where the sigmoid multiplies them into a genuine product. Peak amplitudes grow as (1+m)ⁿ with depth n, and the envelope becomes log-normal with a heavy upper tail (generalized-Pareto index ξ > 0). The paper proves this analytically and gives a closed-form expression for how envelope kurtosis grows with the number of converging bands.

The numerical confirmation uses a network of laminar neural-mass columns wired according to the human connectome (80 cortical regions), with each region assigned a frequency band based on its cortical type — agranular limbic cortex carries delta, granular primary sensory cortex carries gamma. Orienting edges slow-to-fast and routing them to the fast population's drive realizes the architecture. Both a minimal five-band convergent star and the full whole-brain network cross into ξ > 0 on the gamma envelope at a moderate coupling strength. The decisive control: phase-randomizing the signals (same power spectrum, scrambled phases) collapses the heavy tail entirely — confirming the tail is genuine phase-coupled multiplicative depth, not a spectral artifact. Importantly, the summed field across all regions stays bounded throughout; the heavy tail is a local, carrier-level property, not a global blowup.