| name | topological-effective-connectivity-hodge |
| description | Information-theoretic framework coupling Hodge decomposition with lead-lag mutual information for directed brain network analysis - separates feed-forward drive, feedback loops, and cyclic flow around topological holes. |
| tags | ["neuroscience","brain-networks","topology","effective-connectivity","hodge-decomposition","information-theory","directed-graphs"] |
| version | 1 |
| arxiv | 2606.08407v1 |
| date | 2026-06-07T00:00:00.000Z |
Topological Effective Connectivity Modeling in Brain Networks
Overview
Nonparametric, information-theoretic framework for characterizing directed information flow in brain networks with recurrent feedback loops, using discrete Hodge decomposition coupled with lead-lag mutual information.
arXiv: 2606.08407v1
Published: 2026-06-07
Keywords: Topological Data Analysis, Hodge Decomposition, Effective Connectivity, Brain Networks, Directed Information Flow
Core Problem
Challenge: Neural circuits have recurrent feedback loops, but most directed dependence tools assume DAG structure to resolve directional ambiguity.
Gap: DAG assumption cannot represent:
- Recurrent excitation/inhibition loops
- Cortico-cortical feedback
- Thalamocortical loops
- Hippocampal circuits
The Hodge Decomposition Framework
Three Orthogonal Components
The edge flow decomposes into:
Edge Flow = Gradient + Curl + Harmonic
┌─────────────┬──────────────┬─────────────────┐
│ Component │ Interpretation │ Brain Meaning │
├─────────────┼──────────────┼─────────────────┤
│ Gradient │ Hierarchical │ Feed-forward │
│ │ feed-forward │ drive │
│ │ relationships │ │
├─────────────┼──────────────┼─────────────────┤
│ Curl │ Triangle-level │ Local feedback │
│ │ circulation │ loops │
│ │ │ (E-I circuits) │
├─────────────┼──────────────┼─────────────────┤
│ Harmonic │ Cyclic flow │ Global loops │
│ │ around holes │ (large-scale │
│ │ │ networks) │
└─────────────┴────────────────┴─────────────────┘
Mathematical Formulation
Hodge Decomposition on Simplicial Complex:
Given edge flow f: E → ℝ:
f = f_grad + f_curl + f_harm
where:
- f_grad = dφ (gradient of potential φ)
- f_curl = δβ (co-gradient of 2-form β)
- f_harm ∈ ker(Δ) (harmonic)
Properties:
- Orthogonal decomposition: ⟨f_grad, f_curl⟩ = 0
- Unique decomposition for given flow
- Topology-dependent harmonic component
Lead-Lag Mutual Information
Time-Delayed Information Flow
I_lead_lag(X → Y) = I(X_{t-τ}; Y_t) - I(X_t; Y_{t-τ})
Advantages over Granger Causality:
- Nonparametric (no linear assumption)
- Captures nonlinear dependencies
- Information-theoretic foundation
Implementation
def lead_lag_mi(X, Y, delay_bins):
"""
Compute lead-lag mutual information.
Args:
X, Y: Time series (neural activity)
delay_bins: List of time delays to test
Returns:
direction: 'X→Y' or 'Y→X'
magnitude: Information flow strength
"""
forward_mi = mutual_info(X[:-delay], Y[delay:])
backward_mi = mutual_info(Y[:-delay], X[delay:])
net_flow = forward_mi - backward_mi
if net_flow > 0:
return 'X→Y', net_flow
else:
return 'Y→X', -net_flow
Combining Hodge + Lead-Lag
Step-by-Step Procedure
Step 1: Build Network
for region_i, region_j in region_pairs:
direction, strength = lead_lag_mi(activity_i, activity_j)
edges.append((region_i, region_j, direction, strength))
Step 2: Create Edge Flow
edge_flow = {}
for (source, target), direction, strength in edges:
if direction == 'source→target':
edge_flow[(source, target)] = strength
else:
edge_flow[(target, source)] = -strength
Step 3: Hodge Decomposition
gradient_flow = compute_gradient_component(edge_flow)
curl_flow = compute_curl_component(edge_flow)
harmonic_flow = compute_harmonic_component(edge_flow)
Step 4: Interpretation
- Gradient: Identify hierarchical processing streams
- Curl: Find local recurrent circuits
- Harmonic: Detect global oscillatory loops
Brain Network Interpretations
Gradient Component (Feed-Forward)
Examples:
- Sensory → Association hierarchy
- Visual V1 → V2 → V4 → IT
- Motor M1 → Spinal cord output
Interpretation: Unidirectional information propagation following anatomical hierarchy.
Curl Component (Feedback Loops)
Examples:
- Excitatory-Inhibitory microcircuits
- Cortico-thalamic loops
- Local cortical columns
Interpretation: Bidirectional, local feedback maintaining stability, gating, or gain control.
Harmonic Component (Global Cycles)
Examples:
- Hippocampal-Prefrontal-Striatal loop
- Default mode network cycles
- Whole-brain oscillations
Interpretation: Sustained reverberations, memory maintenance, state transitions.
Key Advantages
1. DAG-Free
- Traditional: Assume acyclic structure → miss feedback
- Hodge: Explicitly models cycles → captures full dynamics
2. Disentangled
- Traditional: Mixed forward/backward signals
- Hodge: Separate components → clear interpretation
3. Topology-Aware
- Traditional: Graph-level metrics only
- Hodge: Hole detection → identifies global loops
4. Nonparametric
- Traditional: Linear Granger causality
- Hodge: Information-theoretic → nonlinear capture
Applications
1. Cortical Processing Streams
Identify feed-forward sensory processing vs. feedback attentional modulation.
2. Disease Diagnosis
- Schizophrenia: Abnormal harmonic flow (disrupted global integration)
- Alzheimer's: Reduced gradient (impaired hierarchical processing)
- Parkinson's: Enhanced curl (overactive basal ganglia loops)
3. BCI Optimization
Optimize electrode placement based on gradient/curl balance for stable decoding.
4. Network Control
Identify controllable nodes (gradient) vs. stabilizing loops (curl).
Comparison with Existing Methods
| Method | Handles Loops | Separates Components | Nonlinear | Topology |
|---|
| Granger Causality | ❌ (assumes DAG) | ❌ | ❌ (linear) | ❌ |
| Transfer Entropy | ❌ (no loop handling) | ❌ | ✓ | ❌ |
| Dynamic Causal Modeling | ✓ (explicit) | ❌ | ❌ (linear) | ❌ |
| Hodge + MI | ✓ | ✓ | ✓ | ✓ |
Implementation Notes
Required Data
- Multi-region neural activity time series
- Sufficient length for MI estimation (> 1000 samples)
- Known anatomical connections (optional, for validation)
Computational Steps
- Preprocessing: Normalize, detrend, remove artifacts
- Delay Selection: Test multiple τ, select peak MI
- MI Estimation: Use binning or KDE methods
- Simplicial Complex: Build from regions + connections
- Hodge: Linear algebra decomposition (eigenvectors)
Tools
- Python:
gudhi for simplicial complexes
- Python:
sklearn.metrics.mutual_info_score
- MATLAB: Custom Hodge decomposition scripts
Key Insights
- Topology Essential: Brain dynamics inherently cyclic → DAG methods insufficient
- Three Types: Feed-forward (gradient), local feedback (curl), global loops (harmonic)
- Disentanglement: Same network has multiple overlapping flow types
- Clinical Relevance: Component balance differs in neurological disease
- Method Integration: Combines topological + information-theoretic approaches
Activation
Use when:
- Analyzing directed brain connectivity with feedback
- Disentangling feed-forward from feedback flows
- Detecting global oscillatory loops
- Comparing connectivity in healthy vs. diseased brains
- Building interpretable network models
Trigger words: Hodge decomposition, effective connectivity, directed brain network, feedback loops, topological analysis, curl, gradient, harmonic, recurrent circuits, lead-lag mutual information
References
- Original paper: arXiv:2606.08407v1
- Hodge theory: Lim, 2020 (Hodge Laplacian on graphs)
- TDA: Edelsbrunner & Harer, 2010
- Brain loops: Felleman & Van Essen, 1991 (cortical hierarchy)