Second Order Synaptic Motifs Nonlinear Dynamics

---
name: second-order-synaptic-motifs-nonlinear-dynamics
description: Dynamic mean-field theory for how four second-order synaptic motifs (chain, reciprocal, convergent, divergent) shape nonlinear random-network dynamics. Covers outlier/bulk eigenvalue reshaping, ferromagnetic/limit-cycle/glassy regimes, quenched-heterogeneity chaos suppression, and Lyapunov geometry. Activation: synaptic motifs, DMFT, path-integral, network motifs, glassy dynamics, limit cycles, chaos dimensionality.
license: MIT
metadata:
  arxiv_id: "2609.14251"
  published: "2026-09-13"
  authors: "Jun Yang, Hannah Choi (Georgia Institute of Technology)"
  tags: [synaptic-motifs, dynamic-mean-field-theory, random-neural-networks, chaos, limit-cycles, glassy-dynamics, connectomics]
---

# Nonlinear Dynamics of Random Neural Networks with Second-Order Synaptic Motifs

Methodology from arXiv:2609.14251 — "Nonlinear dynamics of random neural networks with second-order synaptic motifs" (Jun Yang & Hannah Choi, Georgia Tech, Sep 2026).

## Overview and Key Innovation

Classical random-network theory (Sompolinsky–Crisanti–Sommers) assumes i.i.d. connectivity, but biological circuits over/under-represent local **motifs** across species (C. elegans, Drosophila, zebrafish, mouse, rat, human). This work derives a **path-integral DMFT** for the SCS firing-rate model with four second-order motif correlations, showing each motif class has a *distinct* dynamical signature — going beyond prior linear-network and spiking-network-correlation analyses.

**Four motifs** (Pearson correlations between connectivity entries):
- **Chain** ρ(J_ij, J_jk) — generates outlier eigenvalues AND renormalizes effective mean coupling
- **Reciprocal** ρ(J_ij, J_ji) — classic symmetric feedback, slows dynamics
- **Convergent** ρ(J_ij, J_ik) — converts mean activity into quenched heterogeneity
- **Divergent** ρ(J_ik, J_jk) — rescales temporal noise only

**Positivity constraint**: |τ_chn| ≤ √(τ_con·τ_div) — convergent and divergent correlations cannot both vanish; chain studies use minimal symmetric choice τ_con = τ_div = |τ_chn|. Scaling: three-neuron motifs must be O(N⁻¹) with N·τ finite to have O(1) aggregate effect (each edge belongs to O(N) such motifs).

## Core Theory

### 1. Model

```
ẋ_i = −x_i + Σ_j J_ij φ(x_j),   φ = tanh (odd) or (1+tanh)/2 (nonnegative)
⟨J_ij⟩ = J₀/N,  ⟨⟨J_ij²⟩⟩ = g²/N
```

Gaussian ensemble with motif correlations = dense approximation to sparse SONET networks (τ_chn = (p_chn − p²)/(p(1−p))).

### 2. Eigenspectrum: Outliers + Bulk

Chain/reciprocal generate **outliers**:
```
λ± = [J₀ ± √(J₀² + 4g²(N·τ_chn + τ_rec))]/2
```
Positive τ_chn → two real outliers; negative τ_chn → complex-conjugate pair (oscillatory instability).

Motifs also reshape the **bulk** (not just outliers — key novel point). Bulk stays elliptical with semi-axes g(A±B)/√A:
```
A = 1 − τ_con − τ_div        B = τ_rec − 2τ_chn
g_eff = g·(A+B)/√A           (effective gain; zero state loses stability at Re λ = 1)
```

### 3. Single-Site DMFT Equation (path-integral saddle point)

```
ẋ(t) = −x(t) + J₀⟨φ(t)⟩ + g²B ∫χ_φ(t,t′)φ(t′)dt′      [retarded self-interaction]
       + g²N·τ_chn ∫χ_φ(t,t′)⟨φ(t′)⟩dt′               [chain mean feedback]
       + η(t)
⟨η(t)η(t′)⟩ = A·g²·C_φ(t,t′) + g²N·τ_con·⟨φ(t)⟩⟨φ(t′)⟩
```
with response χ_φ = δφ/δη and autocorrelation C_φ.

### 4. Regime-by-Regime Effects

**Positive chain correlations → ferromagnetic states.** In stationarity the chain term reduces to a *static renormalization of mean coupling* (using ∫χ_φ = ⟨φ′⟩):
```
J₀ → J₀ + g²·N·τ_chn·⟨φ′⟩·⟨φ⟩
```
Phase diagram (g, N·τ_chn) has 4 regimes: paramagnetic, ferromagnetic (2 symmetric fixed points), homogeneous chaos (HC), structured chaos (SC, chaotic with nonzero mean). Boundaries: PM–HC at g=1 with λ+<1; PM–FM at λ+=1 with g<1; HC–SC at ⟨φ′⟩λ+=1; SC–FM from bulk edge at fixed point. **Interpretation**: E-I balance can partly arise from local connectivity structure, not only population-mean cancellation.

**Negative chain correlations → limit cycles (J₀>0).** Complex-conjugate outliers destabilize at J₀⟨φ′⟩ = 2 (oscillatory HC instability). Five regimes including **LC+Chaos bistability** — small-mean initial conditions converge to HC, large-mean to LC. A new *structural mechanism for neural oscillations*: motif under-representation, not just coupling sign, drives rhythms.

**Strong negative chain → glassy multistable regime.** Both negative τ_chn and positive τ_rec increase B, producing slow dynamics and fixed-point proliferation — glassiness in **fully asymmetric** networks (τ_rec=0), previously seen only with partial symmetry. New route to glassy dynamics.

**Convergent vs divergent motifs — the sharpest distinction:**
- Divergent: enters ONLY through prefactor A of temporal noise — invisible in zero-mean HC phase
- Convergent: adds static term g²N·τ_con·⟨φ⟩² to noise — **converts nonzero mean activity into quenched heterogeneity**
- Temporal-variance fraction: q_T = [C_x(0)−C_x(∞)] / [C_x(0)−C_x(∞) + g²(AC_φ(∞)+N·τ_con·⟨φ⟩²)]
- With nonnegative φ (inhibitory population, J₀<0): active states have ⟨φ⟩>0, so convergent motifs suppress temporal chaos (q_T < 1) while divergent motifs do not

### 5. Chaos Geometry Beyond Single-Site DMFT

- **Lyapunov spectrum** (numerical): negative chain + positive reciprocal suppress chaos (lower LLE, KS entropy, KY dimension); convergent/divergent lower KS entropy and KY dimension **even at fixed g_eff** — dimensionality drops without changing spectral edge
- **Participation-ratio dimension** (analytical, two-cavity method): three-neuron motifs increase off-diagonal activity covariance, reducing PR dimension. In linear noise-driven networks D_PR = 1−g_eff² (motifs act only via g_eff); in autonomous chaotic nonlinear networks motifs reduce dimensionality directly — a circuit-level mechanism constraining neural-dynamics dimensionality
- **Transient dynamics**: chain feedback leaves imprint on relaxation even when stationary statistics match i.i.d. networks

## Implementation Guidance

1. **Network generation**: sample jointly Gaussian J with prescribed Pearson correlations under positivity constraints (paper Appendix G); verify eigenvalue bulk matches g(A±B)/√A prediction
2. **DMFT solve**: stationary states via fixed-point equations for (µ_x, Δ_x); chaotic states via frequency-domain solve of Eqs. (15)-(16) with autocorrelation and response kernels
3. **Phase diagrams**: simulate grids over (g, N·τ_chn) with fixed J₀ sign; compare against analytic boundaries (Eqs. 18–22, 24)
4. **Lyapunov analysis**: compute spectrum from long trajectories; extract LLE, KS entropy (Σ positive exponents), KY dimension (interpolated zero of cumulative sum)
5. **Quenched-vs-temporal decomposition**: measure q_T from time-averaged per-neuron means vs temporal variance to detect convergent-motif effects

## Pitfalls

- **Scaling matters**: O(1) three-neuron correlations make bulk corrections leading-order but networks non-self-averaging; O(N⁻¹) scaling (N·τ = O(1)) keeps DMFT tractable — pick scaling deliberately
- **Gaussian approximation breaks down in extremely sparse networks** — use sparse-network DMFT instead
- **Rate ≠ spiking**: this is a firing-rate theory; synchrony effects of motifs in spiking networks are cell-type dependent (only chain and convergent affect synchrony significantly in prior numerics) — the quenched-heterogeneity result provides intuition for why
- **Fixed-point complexity**: Kac–Rice topological-complexity extension to motif ensembles is open; complexity trends from GLV partial-symmetry models are qualitative analogies only
- **Reciprocal ≠ chain-in-bulk**: B = τ_rec − 2τ_chn means negative chain *and* positive reciprocal both slow dynamics — do not attribute slowdown to symmetry alone

## Applications

- **Cortical circuit theory**: connectomics-informed DMFT — plug measured motif statistics (mouse CA3 development, human cortical data) into dynamical-regime prediction
- **Reservoir computing design**: motif tuning as a structural knob for memory capacity vs. chaos dimensionality trade-offs
- **E-I balance theory extension**: chain-motif contribution to balanced-state self-consistency
- **Oscillation mechanism taxonomy**: distinguish structure-driven (motif) vs. parameter-driven rhythms in ECoG/MEG
- **Disease models**: motif dysregulation → glassy multistability as candidate dynamical phenotype

## Related Skills

- `synaptic-motifs-mean-field-theory` (arXiv:2606.27946) — complementary low-rank multi-population theory for heterogeneous motifs; this skill covers the single-population nonlinear DMFT regimes
- `cavity-method-rnn-analysis` — two-site cavity technique for large nonlinear RNNs
- `predictable-mean-field-chaos-rnn` — Krylov mean-field chaos predictability
- `effective-plasticity` — network-level plasticity quantification

## Source

- arXiv:2609.14251v1 [q-bio.NC, cond-mat.dis-nn, nlin.CD], submitted 2026-09-13
- Jun Yang, Hannah Choi — School of Mathematics & Quantitative Biosciences, Georgia Tech
- Acknowledges David Clark (source code of Ref. 38); Sloan Research Fellowship in Neuroscience