# Dual-Gibbs Gaussian latent layer

- ID: 1548
- Canonical URL: https://synthcore.org/idea/1548/dual-gibbs-gaussian-latent-layer
- API JSON: https://synthcore.org/api/idea/1548.json
- API Markdown: https://synthcore.org/api/idea/1548.md
- Verification status: unverified
- Source: [arXiv:2607.28706](https://arxiv.org/abs/2607.28706)
- Category: sampling
- Solves: speedup, stability, sample-efficiency
- ML areas: graph-nn, vae, sampling
- Math tags: graph-theory, linear-algebra, probability, spectral-theory
- Ratings: usefulness 6/10; difficulty 7/10; novelty 7/10

## Idea description

Replace slow primal-coordinate Gibbs sampling of a graph-structured Gaussian latent feature field with Gibbs sampling in the Fourier-dual factor graph. Use the primal convergence formula to determine when dual sampling is worthwhile, and map dual samples back to neural features through the graph incidence operator.

## Mathematical statement

The model has density p(x) proportional to exp(-x transpose Q x divided by 2), with precision Q = D_s inverse + B D_sigma inverse B transpose. Here B is an oriented graph incidence matrix, D_s contains vertex variances, and D_sigma contains edge variances. For random-sweep Gibbs, define A = I - diag(Q_11 inverse, ..., Q_nn inverse) Q, where n is the number of vertices. The paper gives the contraction rate r_p = (n inverse times (n - 1 + lambda_max(A))) raised to n. For a homogeneous k-regular graph, this becomes r_p = (1 - (sigma squared divided by (sigma squared + k s squared)) times (1 divided by n)) raised to n. The dual variables satisfy tilde_X = B tilde_Y, and their covariance obeys the covariance-complement identity D_s to the power minus one-half times Cov(X) times D_s to the power minus one-half, plus D_s to the power one-half times Cov(tilde_X) times D_s to the power one-half, equals I. The implementation exploits the paper's claim that dual convergence is topology-independent for cyclic homogeneous models and that effective dual convergence improves with graph algebraic connectivity, while measuring actual integrated autocorrelation time empirically.

## Key formulas

- p(x) proportional to exp(- x transpose Q x / 2), and Q = D_s inverse + B D_sigma inverse B transpose.
- A = I - diag(Q_11 inverse, Q_22 inverse, ..., Q_nn inverse) Q, and r_p = (n inverse times (n - 1 + lambda_max(A))) raised to n.
- r_p = (1 - (sigma squared / (sigma squared + k s squared)) times (1 / n)) raised to n for a homogeneous k-regular graph.
- D_s^(-1/2) Cov(X) D_s^(-1/2) + D_s^(1/2) Cov(tilde_X) D_s^(1/2) = I, with tilde_X = B tilde_Y.

## Implementation notes

Integrate this as an exact sampling subroutine inside a graph VAE, Bayesian GNN, or neural world model whose latent state is a scalar or low-dimensional feature field indexed by graph vertices. The neural network produces likelihood parameters, while the conditional latent prior uses Q = D_s inverse + B D_sigma inverse B transpose. First construct and store the oriented sparse incidence matrix B, and parameterize positive vertex and edge variances with softplus. As a baseline, perform random-sweep primal Gibbs: when vertex i is selected, update x_i from a Gaussian with variance Q_ii inverse and mean m_i = minus Q_ii inverse times the sum over j not equal to i of Q_ij x_j; perform n random updates per sweep. For the dual version, build the normal-factor representation with edge variables y_e = x_u - x_v, Fourier-transform the local Gaussian factors, maintain dual edge variables tilde_Y, and update one dual factor at a time using its scalar Gaussian conditional. Recover graph features with tilde_X = B tilde_Y, adding any separate vertex contribution required by the chosen factorization. Estimate algebraic connectivity lambda_2(L), with L = B B transpose, by a few Lanczos iterations and use it only for mode selection. Validate chains with the covariance identity and compare effective sample size per second, integrated autocorrelation time, ELBO variance, and reconstruction error. Start with 32-node and 64-node cycle, grid, and expander graphs, a two-layer GCN decoder, and a Gaussian reconstruction likelihood. Success means at least 2x higher effective samples per wall-clock second on low-connectivity graphs without worse likelihood.

## Disclaimer

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