# Projected Bures Covariance Pooling

- ID: 3150
- Canonical URL: https://synthcore.org/idea/3150/projected-bures-covariance-pooling
- API JSON: https://synthcore.org/api/idea/3150.json
- API Markdown: https://synthcore.org/api/idea/3150.md
- Verification status: unverified
- Source: [arXiv:2609.03762](https://arxiv.org/abs/2609.03762)
- Category: geometry
- Solves: stability, accuracy, speedup
- ML areas: embedding, cnn, graph-nn, optimizer
- Math tags: geometry, optimal-transport, linear-algebra, spectral-theory
- Ratings: usefulness 7/10; difficulty 5/10; novelty 6/10

## Idea description

Replace Euclidean or unprojected covariance averaging with a projected Bures-Wasserstein barycenter layer. Each unit-step barycenter update is followed by eigenvalue clipping into \([\alpha,\beta]\), preserving positive definiteness and preventing ill-conditioning without an additional eigendecomposition.

## Mathematical statement

For positive semidefinite matrices \(P,Q\in\mathrm{PSD}(d)\), the squared Bures-Wasserstein distance is \(\mathrm{B}(P,Q)=\operatorname{Tr}(P+Q)-2\mathrm{F}(P,Q)\), where \(\mathrm{F}(P,Q)=\operatorname{Tr}[(P^{1/2}QP^{1/2})^{1/2}]=\|P^{1/2}Q^{1/2}\|_1\). Given SPD inputs \(A_1,\ldots,A_m\) and weights \(w_i\ge0\) with \(\sum_iw_i=1\), the BW barycenter minimizes \(\Phi(S)=\frac12\sum_iw_i\mathrm{B}(S,A_i)\). The paper's projection lemma states that for \(X=U\operatorname{diag}(\lambda_1,\ldots,\lambda_d)U^\top\), \(\Pi_{[\alpha,\beta]}(X)=U\operatorname{diag}(\operatorname{clip}(\lambda_j,\alpha,\beta))U^\top\) is the closed-form BW projection onto \(\mathcal C_{\alpha,\beta}=\{S:\alpha I\preceq S\preceq\beta I\}\), and it is non-expansive: \(d_{\mathrm{BW}}(\Pi(X),\Pi(Y))\le d_{\mathrm{BW}}(X,Y)\). Applying this projection to the unit-step BW fixed-point map \(T\), \(S_{t+1}=\Pi_{[\alpha,\beta]}(T(S_t))\), gives the claimed rate \((1-\kappa^{-3/2})\) when the ensemble condition number is \(\kappa\) and inputs lie in the spectral interval.

## Key formulas

- $$\mathrm{B}(P,Q)=\mathrm{d}^{2}_{\mathrm{BW}}(P,Q):=\operatorname{Tr}[P+Q]-2\mathrm{F}(P,Q),\qquad \mathrm{F}(P,Q):=\operatorname{Tr}\!\left[\sqrt{P^{1/2}QP^{1/2}}\right]=\left\|P^{1/2}Q^{1/2}\right\|_{1}.$$
- $$\Phi(S)=\frac12\sum_{i=1}^{m}w_i\,\mathrm{B}(S,A_i),\qquad w_i\ge0,\quad \sum_{i=1}^{m}w_i=1.$$
- $$\Pi_{[\alpha,\beta]}(X)=U\operatorname{diag}\!\big(\operatorname{clip}(\lambda_1,\alpha,\beta),\ldots,\operatorname{clip}(\lambda_d,\alpha,\beta)\big)U^{\top},\quad X=U\operatorname{diag}(\lambda_1,\ldots,\lambda_d)U^{\top}.$$
- $$S_{t+1}=\Pi_{[\alpha,\beta]}(T(S_t)),\qquad d_{\mathrm{BW}}(S_t,S_\star)\lesssim(1-\kappa^{-3/2})^t d_{\mathrm{BW}}(S_0,S_\star).$$

## Implementation notes

(1) Integration point: add this as a differentiable covariance-aggregation module in a CNN, graph neural network, or set encoder. For each sample, form SPD feature covariances \(A_i=H_i^\top H_i/n+\epsilon I\), where \(H_i\) is a patch, token-neighborhood, or augmentation feature matrix. Replace arithmetic pooling with \(K\) projected BW barycenter iterations. (2) Pseudocode: initialize \(S\leftarrow\sum_iw_iA_i\). For \(t=1,\ldots,K\), compute \(R_i=(S^{1/2}A_iS^{1/2})^{1/2}\), form the unit-step proposal \(X\leftarrow(\sum_iw_iR_i)^2\), eigendecompose \(X=U\operatorname{diag}(\lambda)U^\top\), and update \(S\leftarrow U\operatorname{diag}(\operatorname{clip}(\lambda,\alpha,\beta))U^\top\). Feed \(S\), or \(\log S\), to the next layer and backpropagate through the matrix operations with stabilized eigendecomposition gradients. (3) Compute the BW update and clipping exactly; choose \(\alpha\) and \(\beta\) from running lower and upper eigenvalue quantiles of the inputs, with gradients stopped through these bounds. Estimate the empirical contraction by measuring successive \(d_{\mathrm{BW}}(S_t,S_{t-1})\); the theorem supplies the non-expansiveness and claimed rate, while task improvement is empirical. (4) First experiment: compare arithmetic pooling, log-Euclidean pooling, unprojected BW iterations, and projected BW iterations on CIFAR-100 using a small ResNet and covariance features, with \(K\in\{3,5,10\}\). Measure validation accuracy, barycenter objective, condition numbers, NaN or gradient-failure frequency, and loss versus eigendecomposition FLOPs. Success is lower instability and faster objective decrease at equal compute, with no extra eigendecomposition relative to the unprojected BW baseline.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.