Stochastic Graph Heat Connectivity

---
name: stochastic-graph-heat-connectivity
description: "Stochastic Graph Heat Modelling methodology for brain connectivity estimation. Uses noise-driven heat diffusion on graphs to estimate directed, multivariate, dynamic, model-based connectivity from neurophysiological data. Extends traditional coherence methods with graph-based PDE formulation and regularization. Activation: brain connectivity, graph heat modelling, neurophysiological data, directed connectivity, coherence, graph PDE, effective connectivity"
metadata:
  arxiv_id: "2606.29098"
  published: "2026-06-27"
  authors: "Stephan Goerttler, Min Wu, Fei He"
  tags: ["brain-connectivity", "graph-heat-modelling", "neurophysiological-data", "directed-connectivity", "graph-pde"]
---

# Connectivity Estimation using Stochastic Graph Heat Modelling

## Core Methodology

**Stochastic Graph Heat Modelling** uses noise-driven heat diffusion on graphs to estimate brain connectivity from neurophysiological data (EEG/MEG/fMRI).

### Mathematical Framework

1. **Heat Equation on Graphs**: Model neural signal propagation as heat diffusion
   ```
   ∂u/∂t = Δ_G u + noise
   ```
   where Δ_G is the graph Laplacian

2. **Noise-Driven Approach**: Instead of deterministic diffusion, incorporate stochastic noise terms that represent neural activity fluctuations

3. **Directed Connectivity**: The heat flow direction reveals causal/directional relationships between brain regions

### Key Innovations

1. **Relaxed Noise Assumptions**: Previous work assumed specific noise distributions; this extension allows more general noise models

2. **Regularization**: Added regularization terms improve robustness and prevent overfitting

3. **Model-Based**: Unlike correlation/coherence methods, this is an explicit generative model with interpretable parameters

4. **Dynamic & Multivariate**: Captures time-varying, multivariate interactions simultaneously

## Advantages Over Traditional Methods

### Compared to Correlation/Coherence
- **Directed**: Identifies causal direction, not just correlation
- **Model-based**: Explicit generative model vs. statistical association
- **Multivariate**: Handles multiple regions simultaneously
- **Dynamic**: Captures time-varying connectivity

### Compared to Granger Causality
- **Graph-based**: Explicit spatial structure via graph Laplacian
- **Physics-informed**: Based on heat diffusion physics
- **Regularized**: Better stability in high-dimensional settings

## Validation

### Simulation Studies
- Controlled experiments with known ground-truth connectivity
- Varying signal-to-noise ratios
- Comparison against:
  - Standard coherence
  - Granger causality
  - Other graph-based methods

### Real Data Applications
- EEG/MEG datasets
- fMRI resting-state data
- Task-based paradigms

## Implementation Notes

### Graph Construction
1. Start with anatomical/functional regions as nodes
2. Initialize with spatial proximity or structural connectivity
3. Iteratively update edge weights based on heat flow fitting

### Parameter Estimation
- Edge weights represent connection strengths
- Optimized via maximum likelihood or least-squares fitting
- Regularization prevents overfitting (L1/L2 penalties)

### Computational Complexity
- Graph Laplacian computation: O(N²) for N regions
- Heat equation solving: O(N³) per time step
- Parallelizable across time windows

## Applications

### Neuroscience
- **Effective connectivity**: Directional information flow
- **Network dynamics**: Time-varying connectivity patterns
- **Clinical biomarkers**: Connectivity changes in disease

### Other Domains
- Social network influence propagation
- Financial market correlations
- Climate system interactions

## Limitations

1. **Computational Cost**: More expensive than simple correlation
2. **Model Assumptions**: Assumes heat diffusion is appropriate generative model
3. **Spatial Resolution**: Limited by sensor/source localization accuracy
4. **Temporal Resolution**: Requires sufficient sampling rate

## Extensions & Future Work

1. **Non-linear heat models**: Incorporate non-linear diffusion terms
2. **Multi-scale graphs**: Hierarchical graph structures
3. **Integration with fMRI**: Combine with hemodynamic models
4. **Real-time estimation**: Online algorithms for BCI applications

## Key References

- arXiv:2606.29098 [stat.ML]
- Graph heat equation literature
- Neurophysiological connectivity methods
- Stochastic PDE theory