CALLIOPE · BCOM

Hierarchical Amplitude Modulation (HAM) is proposed as a mathematically grounded, circuit-level mechanism for hierarchical predictive coding in the brain, offering a unified explanation for three major empirical regularities in neural recordings. The framework formalizes a multiplicative cascade in which slower brain rhythms modulate the amplitude of faster rhythms, creating nested oscillatory envelopes that encode information at multiple hierarchical levels simultaneously. A central theoretical result demonstrates that avoiding spectral overlap among sideband clusters in such a cascade requires geometric frequency spacing with a common ratio r ≳ 2–3, directly accounting for the empirically observed logarithmic arrangement of canonical oscillation bands from delta through gamma. The same cascade structure naturally generates 1/f^α aperiodic power spectra, with the exponent governed by the closed-form relation α = 2 ln(2/m) / ln(r), where m is modulation depth and r is the inter-band frequency ratio, thereby explaining state-dependent variability in aperiodic slopes across brain regions and pharmacological conditions. Proof-of-concept implementation within the Laminar Neural Mass Model (LaNMM) confirms that biophysically realistic cortical circuits can perform both modulation and demodulation through intrinsic nonlinearities and cross-frequency coupling. The framework generates testable predictions linking spectral slope changes to neuromodulatory state, cognitive load, and consciousness level, with implications for brain stimulation design and the architecture of artificial neural systems.

The brain's nested rhythms aren't just noise — they're a multiplexing scheme, and HAM is the first framework to derive why the specific frequency ratios we observe are mathematically necessary.

The core idea is borrowed from radio engineering. In AM radio, a slow information signal modulates the amplitude of a fast carrier wave. HAM extends this to multiple nested layers: slow delta waves modulate the amplitude of theta, which modulates alpha, which modulates beta, which modulates gamma. Information gets encoded at every level of this envelope hierarchy simultaneously. The brain isn't just oscillating — it's broadcasting on multiple channels at once, stacked inside each other like Russian dolls.

Here's where the math gets interesting. When you stack these modulation layers, each faster oscillation generates "sideband" frequencies — spectral echoes at (f_carrier ± f_modulator). If you pack your frequency bands too close together, these sidebands bleed into neighboring bands and the whole encoding scheme collapses — you can't decode anything cleanly. The paper derives the exact condition for avoiding this: each frequency band must be more than twice the sum of all lower frequencies. The simplest way to satisfy this is geometric spacing, where each band is a constant ratio r apart. Crucially, r must be at least 2. The empirically observed ratio between delta, theta, alpha, beta, and gamma is approximately 2–3. HAM doesn't just accommodate this fact — it predicts it as a mathematical necessity.

The same cascade structure also explains the brain's ubiquitous 1/f^α "pink noise" background spectrum. The paper derives a closed-form expression: α = 2 ln(2/m) / ln(r), where m is modulation depth and r is the frequency ratio. This is a genuinely useful formula. It says the spectral slope isn't a fixed property of the brain — it's a readout of how strongly the hierarchy is modulating. Stronger modulation (higher m, as under attention or psychedelics) flattens the slope; weaker modulation (anesthesia, sleep) steepens it. Emerging empirical data on aperiodic slopes across brain states is consistent with this, though the paper is careful to frame this as prediction rather than established fact.

The paper also grounds HAM in a biophysically realistic cortical circuit model (LaNMM), showing that the sigmoid-shaped input-output nonlinearity of neurons naturally produces amplitude modulation when slow and fast inputs combine — no special machinery required. Demodulation (reading out the encoded signal) follows from simple envelope extraction, which neural circuits can implement via rectification and low-pass filtering. This closes the loop from abstract communication theory to actual cortical implementation.

The main open questions the paper acknowledges honestly: modulation depth m isn't directly measurable from recordings, so the key formula's predictions are currently tested only indirectly; the analysis assumes relatively weak modulation and fixed band spacing, while real brains are noisy and dynamic; and causality is unresolved — do brains use log-spacing because of HAM constraints, or do both emerge from something else? These are fair caveats for a theoretical framework paper. What HAM offers is a single organizing principle that unifies three previously separate empirical puzzles — band spacing, 1/f spectra, and cross-frequency coupling — under one mathematical roof.