Fmri Mahalanobis Bures Whitening

---
name: fmri-mahalanobis-bures-whitening
description: "De-individualizing fMRI signals via Mahalanobis whitening and Bures geometry — methodology for distilling meaningful information from fMRI by treating data whitening as quantum-inspired state de-individualization using Bures distance."
activation_keywords:
  - fMRI de-individualization
  - Mahalanobis whitening brain
  - Bures distance fMRI
  - fMRI functional connectivity
  - quantum-inspired brain analysis
  - 功能连接去个体化
  - fMRI白化
  - Bures几何脑成像
  - Alzheimer fMRI biomarker
  - functional connectivity preprocessing
categories:
  - neuroscience
  - medical-imaging
  - quantum-physics
arxiv_id: "2511.07313"
arxiv_url: "https://arxiv.org/abs/2511.07313"
authors: "Aaron Jacobson, Tingting Dan, Martin Styner, Guorong Wu, Shahar Kovalsky, Caroline Moosmueller"
created: "2026-06-08"
---

# De-Individualizing fMRI Signals via Mahalanobis Whitening and Bures Geometry

## Description

Methodology for processing fMRI signals through Mahalanobis data whitening to extract meaningful information about subjects and experimental stimuli. Provides a quantum-inspired interpretation of whitening as a two-stage de-individualization process motivated by the Bures distance from quantum mechanics. Applications include improving Alzheimer's diagnosis accuracy, especially in preclinical stages.

## Core Concepts

### Mahalanobis Whitening for fMRI
- Standard fMRI analysis suffers from subject-specific confounds
- Mahalanobis whitening: x_whitened = Σ^(-1/2) · (x - μ)
- Removes individual covariance structure while preserving stimulus-relevant information
- Two-stage process: (1) remove subject identity, (2) preserve experimental signal

### Bures Distance Connection
- Bures distance: D_B(ρ₁, ρ₂) = √[2(1 - Tr√(√ρ₁·ρ₂·√ρ₁))]
- Quantum metric for distinguishing quantum states
- Applied to fMRI covariance matrices as "quantum states" of brain activity
- Provides geometrically meaningful distance between brain states

## Methodology Steps

### Step 1: Compute Subject Covariance
```
For each subject s:
  Σ_s = Cov(fMRI_time_series_s)
  μ_s = Mean(fMRI_time_series_s)
```

### Step 2: Apply Mahalanobis Whitening
```
Global covariance: Σ_global = Average(Σ_s) across subjects
Whitening matrix: W = Σ_global^(-1/2)
Whitened data: x' = W · (x - μ_global)
```

### Step 3: Bures Distance Analysis
```
Treat each subject's covariance as density matrix: ρ_s = Σ_s / Tr(Σ_s)
Compute Bures distance between subject states:
  D_B(ρ_s, ρ_t) = √[2(1 - Tr√(√ρ_s · ρ_t · √ρ_s))]
```

### Step 4: De-Individualization Validation
- Verify that whitened data no longer encodes subject identity
- Confirm that stimulus/experimental effects are preserved
- Use cross-validation on downstream tasks (classification, prediction)

## Applications

1. **Alzheimer's diagnosis**: Improved accuracy in preclinical stage detection
2. **Cross-subject fMRI analysis**: Remove individual confounds for group studies
3. **Biomarker discovery**: Isolate disease-relevant signals from individual variation
4. **Brain-computer interfaces**: Standardized features across subjects
5. **Longitudinal studies**: Track changes within subjects over time

## Key Advantages

- **Non-parametric**: No assumptions about data distribution
- **Geometry-preserving**: Bures distance respects quantum-information geometry
- **Computationally efficient**: Standard linear algebra operations
- **Interpretable**: Clear separation of individual vs. stimulus factors

## Error Handling

### Issue: Singular Covariance Matrix
- **Solution**: Add small regularization (Σ + ε·I) before inversion
- **Typical ε**: 1e-6 to 1e-4 times max eigenvalue

### Issue: Over-whitening (loss of signal)
- **Solution**: Use partial whitening: W_α = (α·I + (1-α)·Σ)^(-1/2)
- **Tuning**: Cross-validate α on downstream task performance

## Resources

- Paper: [arXiv:2511.07313](https://arxiv.org/abs/2511.07313)
- Related: Metabolic quantum limit to MEG information capacity (arXiv:2511.06401)