The Rosetta Stone of Neural Mass Models

The motivation

Why We Needed a Common Language

If you work in computational neuroscience, you've almost certainly faced this situation: you read a paper using a Kuramoto model, another using Wilson–Cowan equations, a third using Jansen–Rit, and you find yourself wondering—how do these formalisms actually relate to each other? Which assumptions did each team make, and what would it take to translate results from one framework into another?

We found ourselves asking exactly these questions. Brain dynamics unfold across every level of neural organization—from single-neuron spiking to the macroscopic rhythms we measure with fMRI, MEG, and EEG—and neural mass models (NMMs) are our primary tool for making sense of this hierarchy. But the NMM zoo has grown wild. Each formalism carries its own notation, its own assumptions, and its own conventions for how external inputs and inter-regional coupling enter the equations. Choosing a model often feels less like a principled design decision and more like picking a dialect at random.

"Understanding is the ability to see one thing in many ways." — R. P. Feynman

We set out to build a bridge—what we call a Rosetta Stone—that arranges the major neural mass formalisms along a single, navigable ladder of increasing biological detail. The guiding principle turned out to be surprisingly simple: every one of these models, no matter how complex, rests on the same dynamical engine.

The core idea

One Motif to Rule Them All: Push–Pull

The central insight of our work is that every neural mass model can be decomposed into a push–pull interaction between two variables. One variable (call it x, loosely "excitation") drives the system forward; the other (y, loosely "inhibition") drags it back. This antagonistic dance is what generates oscillations—from the simplest harmonic oscillator all the way up to biologically detailed laminar neural mass models.

We start at the very bottom of the ladder: an undamped harmonic oscillator, the purest archetype of phase dynamics. Its equations are about as simple as it gets:

Rung 1 — Undamped Harmonic Oscillator

$$\dot{x} = -\omega\, y, \qquad \dot{y} = \omega\, x$$

When $y$ is positive it pushes $x$ down; when $x$ is positive it pushes $y$ up. This reciprocal chase—never quite catching up—is the push–pull motif in its most elemental form. In the complex plane: $\dot{z} = i\omega\, z$.

From this starting point, we systematically add ingredients—damping, nonlinearity, synaptic filtering, transfer functions, coupling—and at each step we show that the push–pull core is preserved. Every new rung makes explicit exactly which assumptions are introduced.

Key Insight

Every neural oscillator—whether described in phase coordinates, complex amplitudes, or excitatory–inhibitory firing rates—can be decomposed into a push variable that drives the system away from equilibrium and a pull variable that restores it. This single motif is the dynamical thread connecting all the models on our ladder.

The framework

Climbing the Ladder: Six Rungs of Complexity

We organized the major families of neural mass models into a ladder. Each rung adds a specific layer of biological realism while preserving the push–pull core beneath:

Phase-Only Oscillator

The undamped harmonic oscillator. Phase advances uniformly, amplitude is fixed. The Kuramoto model for synchronization lives here.

Damped Harmonic Oscillator

Adds a damping coefficient for decay or growth. A linear resonator at the heart of Ornstein–Uhlenbeck models for resting-state fMRI.

Stuart–Landau Oscillator

A cubic nonlinearity self-regulates amplitude, creating robust limit cycles via Hopf bifurcation. The canonical model for brain rhythm onset.

Wilson–Cowan (WILCO)

Amplitude and phase become excitatory and inhibitory firing rates with sigmoidal transfer functions. The E–I push–pull is now explicit.

NMM1 (Second-Order Synapses)

Realistic synaptic filters with distinct rise and decay times. Post-synaptic potentials become the dynamical variables; reproduces alpha–gamma interactions.

NMM2 (Next-Generation)

Derived exactly from QIF neurons. Both the transfer function and synaptic couplings are dynamical. Macroscopic equations emerge rigorously from spiking.

The mathematics

The Equations at Each Rung

One of the things we wanted to make very clear in this paper is how the equations transform as you climb the ladder. Let me walk you through the key ones.

Damped Harmonic Oscillator

Adding a real damping coefficient $\alpha$ (negative for decay, positive for growth) to the harmonic oscillator gives us:

Rung 2 — Damped Harmonic Oscillator

$$\dot{z} = (\alpha + i\omega)\, z + F(t)$$

Here $z = x + iy$ is the complex state, $\alpha$ sets the decay time, $\omega$ is the natural frequency, and $F(t)$ is an external (complex-valued) forcing term—which can represent stimulation, sensory drive, or coupling from other regions. This is already the linearized Stuart–Landau model.

For a whole-brain network with $N$ coupled nodes, each placed on a weighted connectome with optional delays:

Coupled Damped Oscillator Network

$$\dot{z}_i = (\alpha_i + i\omega_i)\, z_i + G\sum_{j=1}^{N} C_{ij}\big[z_j(t - \tau_{ij}) - z_i(t)\big] + \hat{F}_{e,i}(t), \quad i = 1,\dots,N$$

$C_{ij}$ are connectome weights, $G$ is a global coupling gain, $\tau_{ij}$ are propagation delays, and $\hat{F}_{e,i}(t)$ is external forcing on node $i$.

Stuart–Landau Oscillator

The critical step from linear to nonlinear dynamics. We add an amplitude-dependent cubic term that saturates growth and stabilizes a limit cycle—mathematically, the normal form of a Hopf bifurcation:

Rung 3 — Stuart–Landau (Complex Form)

$$\dot{z} = (\alpha + i\omega)\, z \;-\; (\gamma + i\beta)\,|z|^2\, z \;+\; F(t)$$

The new parameters: $\gamma > 0$ governs nonlinear amplitude saturation (prevents runaway excitation), and $\beta$ introduces amplitude–phase coupling ("shear"). The limit-cycle radius is $r^* = \sqrt{\alpha/\gamma}$ when $\alpha > 0$.

In polar form, writing $z = r\,e^{i\theta}$, the amplitude and phase dynamics separate cleanly:

Stuart–Landau in Polar Coordinates

$$\dot{r} = \alpha\, r - \gamma\, r^3, \qquad \dot{\theta} = \omega - \beta\, r^2$$

The amplitude equation shows the balance between linear growth ($\alpha\,r$) and nonlinear saturation ($\gamma\,r^3$). Phase velocity depends on amplitude through the shear parameter $\beta$.

Wilson–Cowan (WILCO)

Here we reinterpret the abstract oscillator variables as biologically meaningful excitatory ($x$) and inhibitory ($y$) firing rates, coupled through a weight matrix and passed through sigmoidal transfer functions $\sigma$:

Rung 4 — Wilson–Cowan E–I Rate Model

$$\tau_x\,\dot{x}_i + x_i = \sigma_x\!\Big(w_{xx}\,x_i - w_{xy}\,y_i + P_{x,i} + \hat{F}_{e,i}(t) + \sum_{j \neq i} C_{ij}\,x_j\Big)$$ $$\tau_y\,\dot{y}_i + y_i = \sigma_y\!\Big(w_{yx}\,x_i - w_{yy}\,y_i\Big), \quad i = 1,\dots,N$$

The push–pull motif is now explicitly an E–I loop: excitation ($x$) drives inhibition ($y$) through $w_{yx}$, while inhibition restrains excitation through $w_{xy}$. The sigmoidal $\sigma$ ensures bounded firing rates.

NMM1: Second-Order Synaptic Dynamics

The Jansen–Rit family replaces instantaneous synaptic filtering with second-order operators $L_\alpha$ that shape post-synaptic potentials with realistic rise and decay time constants:

Rung 5 — NMM1 (Second-Order Synapses)

$$L_x[x_i] = \sigma_x\!\Big(-w_{xy}\,y_i + \hat{F}_{e,i}(t) + \sum_{j\neq i} C_{ij}\,x_j(t - \tau_{ij})\Big)$$ $$L_y[y_i] = \sigma_y\!\Big(w_{yx}\,x_i\Big), \quad i = 1,\dots,N$$

where $\;L_\alpha = \frac{1}{\gamma_\alpha}\!\left(\tau_\alpha^2\,\frac{d^2}{dt^2} + 2\tau_\alpha\,\frac{d}{dt} + 1\right)$ is a second-order differential operator with gain $\gamma_\alpha$ and time constant $\tau_\alpha$. This produces the characteristic PSP waveforms seen in real EEG data.

NMM2: Next-Generation (QIF-Based)

The final rung is qualitatively different: instead of postulating the equations, we derive them exactly from an infinite population of quadratic integrate-and-fire neurons. Both the transfer function and synaptic couplings become dynamical:

Rung 6 — NMM2 (Next-Generation Mean-Field)

$$r_x^{(i)} = \Phi_x\!\Big[C_{xx}\,s_x^{(i)} - C_{xy}\,s_y^{(i)} + \hat{F}_e^{(i)}(t) + \sum_{j\neq i} C_{ij}\,s_x^{(j)}\Big], \quad s_x^{(i)} = \hat{K}_x[r_x^{(i)}]$$ $$r_y^{(i)} = \Phi_y\!\Big[-C_{yy}\,s_y^{(i)} + C_{yx}\,s_x^{(i)}\Big], \quad s_y^{(i)} = \hat{K}_y[r_y^{(i)}], \quad i=1,\dots,N$$

Here $r$ are firing rates, $s$ are synaptic variables, $\Phi$ are dynamic (not static) transfer functions derived from the Lorentzian ansatz, and $\hat{K}$ are synaptic kernels. This is the first neural mass model with a rigorous microscopic foundation.

The thread

What Connects All These Models

We don't just list these models side by side—we show the transformations between them. Each rung is a specialization or generalization of its neighbors. The Stuart–Landau is the nonlinear extension of the damped oscillator. Wilson–Cowan is a biological reinterpretation of the Stuart–Landau's Hopf bifurcation, where abstract amplitude and phase become concrete E and I firing rates. NMM1 upgrades Wilson–Cowan with dynamic (rather than instantaneous) synaptic filtering. And NMM2 derives the whole architecture rigorously from spiking neurons.

At every level, the push–pull motif persists. In the harmonic oscillator, it's the $-\omega y$ and $+\omega x$ cross-coupling. In Stuart–Landau, the cubic term adds amplitude regulation but the rotational core remains. In Wilson–Cowan, it's the explicit E$\to$I and I$\to$E synaptic weights. In NMM1 and NMM2, the same antagonism is filtered through increasingly realistic synaptic dynamics.

Quick Reference: Models at a Glance

Model	Core Equation	Best For
Phase-Only / Kuramoto	$\dot{\theta}_i = \omega_i + G\sum C_{ij}\sin(\theta_j - \theta_i)$	Synchronization, large networks, analytic results
Damped Harmonic	$\dot{z}_i = (\alpha_i + i\omega_i)\,z_i + \text{coupling}$	Resting-state FC/PSD, closed-form statistics
Stuart–Landau	$\dot{z} = (\alpha+i\omega)\,z - (\gamma+i\beta)\|z\|^2 z$	Rhythm onset, metastability, envelopes
Wilson–Cowan	$\tau\dot{x} + x = \sigma(w_{xx}x - w_{xy}y + \cdots)$	Bistability, pharmacological modeling, E–I balance
NMM1 (Jansen–Rit)	$L_\alpha[x] = \sigma(\text{synaptic input})$	EEG/MEG spectra, alpha–gamma coupling
NMM2 (QIF-based)	$r = \Phi[C\,s + F],\; s = \hat{K}[r]$	Rigorous spike-to-mass bridge

Stimulation

How External Forcing Enters: From tES to DBS

One aspect we were particularly careful about is standardizing how external inputs—stimulation, sensory drive, pharmacological modulation—enter each model. In every formalism, external forcing $\hat{F}_e(t)$ splits into a deterministic physiological drive $f(t)$, an electric-field contribution $\Lambda[\vec{E}(t)]$, and additive noise $\hat{\eta}(t)$:

External Forcing Decomposition

$$\hat{F}_e(t) = f(t) + \Lambda[\vec{E}(t)] + \hat{\eta}(t)$$

For weak electric fields (tES), the coupling is linear: $\Lambda[\vec{E}] = \vec{\lambda} \cdot \vec{E}(t)$. This standardized treatment means we can design stimulation protocols at one level of the ladder and predict their effects at another.

Why it matters

What This Means in Practice

For Experimentalists

The ladder provides a clear decision tree for model selection. Need synchronization and phase-locking? Start at the Kuramoto level. Interested in amplitude regulation and metastability? Stuart–Landau. Modeling pharmacological interventions or E–I balance? Wilson–Cowan or NMM1. Want a rigorous bridge from spikes to population dynamics? NMM2. No more guesswork.

For Theorists

By exposing the shared dynamical core, we enable systematic parameter transformations between formalisms. Results derived in one model can be mapped to another along the ladder. This also clarifies what is gained and what is lost at each level of abstraction—crucial for model selection and comparison.

For Brain Stimulation & Clinical Work

The standardized treatment of external forcing across all models makes it possible to design stimulation protocols (tES, TMS, DBS) at one level of the ladder and translate their predicted effects to another—bridging the gap between abstract models and clinical interventions.

Looking ahead

Open Frontiers

We see several exciting directions ahead. One is extending exact mean-field reductions beyond QIF neurons to more realistic single-cell models—such as Hodgkin–Huxley-type neurons—possibly via moment closures or renormalization-group methods. Another is enriching laminar models with cell-type specificity (PV, SOM, VIP interneurons), layer-specific projections, and short-term synaptic plasticity.

At the whole-brain scale, incorporating subcortical structures, neuromodulatory gradients informed by gene-expression atlases, and directed laminar connectomes would bring these models closer to the heterogeneous reality of the brain. We're also excited about coupling this framework with modern inference methods—variational data assimilation, Bayesian model comparison, active inference—to turn these models into forward engines for multimodal neuroimaging data and in silico perturbation experiments.

The Bottom Line

We've transformed a fragmented collection of neural mass models into a coherent, navigable ladder. Each rung distills a distinct set of assumptions about synapses, transfer functions, and coupling—while sharing the same push–pull dynamical core. Our hope is that this common language will make it easier to compare models, design experiments, and ultimately bridge microcircuit physiology, whole-brain activity, and cognition within a single coherent framework.

Reference

Castaldo, F., de Palma Aristides, R., Clusella, P., Garcia-Ojalvo, J., & Ruffini, G. (2025). Rosetta Stone of Neural Mass Models. arXiv:2512.10982. https://arxiv.org/abs/2512.10982