Mean Field Adaptation Oscillations

# Mean-Field Theory of Rich Oscillatory Dynamics in Low-Rank Recurrent Networks with Activity-Dependent Adaptation

## Paper Reference
**arXiv**: [2606.30366v1](https://arxiv.org/abs/2606.30366v1)  
**Authors**: Bowen W. Zheng, Earl K. Miller, Ila R. Fiete (MIT)  
**Date**: June 29, 2026  
**Keywords**: mean-field theory, oscillatory dynamics, adaptation, low-rank networks, chaotic dynamics

## Core Contribution

Develops a complete dynamical mean-field theory (DMFT) for random recurrent networks with low-rank connectivity structure and firing-rate-driven adaptation. Reveals how a single parameter (adaptation strength β) drives networks through four distinct dynamical regimes while preserving single-neuron irregularity.

## Key Theoretical Framework

### Model Architecture
- **Network**: N rate neurons with membrane potential xi and adaptation current ai
- **Connectivity**: J = g/√N * W + 1/N * m*n^T (random bulk + rank-one structure)
- **Adaptation**: τa * ai_dot = -ai + β * tanh(xi)
- **Timescales**: τm = 1 (fast), τa ≫ 1 (slow adaptation)

### Four Dynamical Regimes
Increasing adaptation strength β drives progression:

1. **Regime I: Static Coherent State**
   - Overlap κ settles near symmetric fixed points ±κ*
   - Small fluctuations, stable nodes

2. **Regime II: Noise-Sustained Oscillation** (novel discovery)
   - Coherent fixed points become stable foci (damped oscillators)
   - Chaotic background acts as broadband noise driving sustained oscillations
   - Oscillations concentrate at adaptation frequency
   - Transitions from regular → irregular as β increases

3. **Regime III: Irregular Switching**
   - Chaotic fluctuations drive transitions between symmetric wells
   - Switching mediated by slow adaptation variable κa
   - Frequency increases with β

4. **Regime IV: Global Oscillation**
   - Hopf bifurcation of coherent mode
   - Stable limit cycle carries network between wells
   - Coexists with chaotic single-neuron fluctuations

### Two Instability Mechanisms

1. **Chaos Onset** (driven by random connectivity g)
   - Spectral boundary: gc(β) depends on population-averaged squared gain χ²,x
   - Adaptation compresses operating range → concentrates neurons in high-gain region → lowers chaos threshold
   - Counterintuitive: adaptation (stabilizing feedback) facilitates chaos onset

2. **Hopf Bifurcation** (driven by adaptation β)
   - Coherent mode loses stability through frequency-dependent transfer function
   - Ĝ(s) = (sτa + 1) / [(s + 1)(sτa + 1) + βχ̄x]
   - Resonant frequency ω* ≈ [b(2+b)]^(1/4) / √τa where b = βχ̄x

### Mathematical Innovation

**Fixed-Point Diffeomorphism**: 
- F(x*) = x* + β*tanh(x*) = h (total input)
- Local gain: c(h) = sech²(F⁻¹(h))
- Adaptation compresses gain by factor 1/(1+β) near population center
- Population susceptibilities: χ̄x = E[c(h)], χ̄eff = E[c(h)/(1+βc(h))]

**Spectral Self-Consistency** (Hermite expansion):
- Expand tanh(x* + σx*z) in probabilist Hermite polynomials
- Per-neuron firing-rate spectrum: Sδφ,i(f) = Σp (bp)²/ri(τ)^p / p!
- Captures nonlinear transformation of Gaussian fluctuations

**Reduced 3D Model**:
- Variables: (κ, κa, Q) - coherent overlap, adaptation overlap, chaos intensity
- Captures full bifurcation structure
- Explains noise-sustained oscillation mechanism

## Physiological Relevance

### Brain State Transitions
- **Wakefulness**: Low β (cholinergic suppression of adaptation currents) → irregular activity
- **Sleep/Anesthesia**: High β (reduced cholinergic drive) → slow Up-Down alternations
- **Intermediate**: Waxing-waning rhythmic episodes (sleep spindles)

### Single Architecture, Multiple Phenomena
- One parameter sweep produces dynamics spanning multiple brain states
- Explains coexistence of population oscillations with single-neuron irregularity
- Provides mechanistic account of state transitions

## Experimental Validation

- **Parameters**: N=4000, τa=30, σm=σn=2.0, γ=0.7
- **Regime transitions**: Confirmed across g = {1.2, 1.8, 2.0}
- **Phase diagrams**: (β, g) plane mapped with both instability boundaries
- **Code**: https://github.com/Bowen-Zheng-99/rnn_adapt

## Key Insights

1. **Adaptation as Tuning Knob**: Complementary to connectivity structure for controlling oscillations
2. **Noise-Sustained Oscillation**: Linearly stable focus + broadband noise = sustained oscillations at resonant frequency
3. **Adaptation Paradox**: Stabilizing feedback at single-neuron level but destabilizing at population level
4. **Minimal Mechanism**: Low-rank structure + adaptation sufficient for rich oscillatory repertoire
5. **Theoretical Completeness**: Full phase diagram with analytical boundaries and reduced model

## Methodological Patterns

### When to Use
- Analyzing oscillatory dynamics in recurrent networks
- Understanding adaptation effects on population activity
- Modeling brain state transitions
- Designing networks with controllable oscillations

### Implementation Checklist
- [ ] Define rank-one connectivity structure (m, n vectors)
- [ ] Set adaptation dynamics (τa, β)
- [ ] Compute population susceptibilities (χ̄x, χ̄eff)
- [ ] Determine chaos threshold gc(β) from spectral boundary
- [ ] Determine Hopf boundary from coherent mode stability
- [ ] Map phase diagram in (β, g) plane
- [ ] Validate with network simulations (N ≥ 4000)

## Extensions Discussed

- Higher-rank connectivity can layer connectivity-driven oscillations on top
- Multiple adaptation channels for richer dynamics
- Frequency-dependent transfer function generalizes to other slow variables

## Related Work Context

- **Low-rank networks**: Mastrogiuseppe & Ostojic (2018), Landau & Sompolinsky (1998)
- **Adaptation in random networks**: Muscinelli et al. (resonant chaos)
- **Slow variables**: Clark & Abbott (rich phase diagrams without spatial structure)
- **Bridging gap**: This work shows low-rank + adaptation → spatially coherent oscillations

## Activation Triggers
**Keywords**: mean-field theory, oscillatory dynamics, adaptation, low-rank networks, chaotic dynamics, brain state transitions, Up-Down states, noise-sustained oscillation, Hopf bifurcation

**Use Cases**:
- Theoretical neuroscience research on oscillatory mechanisms
- Computational modeling of sleep/anesthesia dynamics
- Network design with controllable oscillatory properties
- Understanding adaptation effects on population activity