# Cluster MCMC for rare neural trajectories

- ID: 2459
- Canonical URL: https://synthcore.org/idea/2459/cluster-mcmc-for-rare-neural-trajectories
- API JSON: https://synthcore.org/api/idea/2459.json
- API Markdown: https://synthcore.org/api/idea/2459.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.17511](https://arxiv.org/abs/2608.17511)
- Category: sampling
- Solves: sample-efficiency, stability, accuracy
- ML areas: world-model, ssm, diffusion-sampling
- Math tags: stochastic-processes, statistical-mechanics, probability, dynamical-systems
- Ratings: usefulness 8/10; difficulty 7/10; novelty 7/10

## Idea description

Train or sample a neural state-space model in trajectory space rather than drawing complete rollouts independently. Construct a space-time path graph whose vertices are latent states and local transition events, then update connected clusters of the entire trajectory using conditional Gibbs or Swendsen-Wang-like moves while preserving fixed initial, terminal, or event-count constraints. This should replace exponentially small forward-rollout success probabilities with local conditional updates that change long portions of a valid trajectory at once.

## Mathematical statement

For a continuous-time Markov model with state trajectory $z(t)$, transition intensity $r_\theta(z\to z')$, and escape rate $R_\theta(z)=\sum_{z'\neq z}r_\theta(z\to z')$, the unnormalized path weight is $W_\theta[\omega]=\rho(z_0)\exp[-\int_0^T R_\theta(z(t))\,dt]\prod_k r_\theta(z_{k^-}\to z_{k^+})$, where $\omega$ contains event times and states. Uniformization chooses $\Gamma(z)\ge R_\theta(z)$ and inserts virtual events, giving a fixed opportunity rate. The resulting kernel is $K_\theta(z,z')=r_\theta(z\to z')/\Gamma(z)$ for $z'\neq z$ and $K_\theta(z,z)=1-R_\theta(z)/\Gamma(z)$. A cluster proposal must satisfy detailed balance $W(\omega)q(\omega\to\omega')=W(\omega')q(\omega'\to\omega)$; conditional Gibbs cluster resampling gives acceptance probability one. In the paper's edge construction, nonnegative rates require $\Gamma_x\ge\alpha_{i,j}+\alpha_{j,i}-\psi+|\Delta\phi|$ and $0\le\psi\le\min(\alpha_{i,j}-\phi^{\uparrow},\alpha_{j,i}-\phi^{\downarrow})$.

## Key formulas

- $$W_\theta[\omega]=\rho(z_0)\exp\!\left[-\int_0^T R_\theta(z(t))\,dt\right]\prod_{k=1}^{K}r_\theta(z_{k^-}\!\to z_{k^+}),\qquad R_\theta(z)=\sum_{z'\ne z}r_\theta(z\!\to z').$$
- $$\Gamma(z)\ge R_\theta(z),\qquad K_\theta(z,z')=\begin{cases}r_\theta(z\to z')/\Gamma(z),&z'\ne z,\\1-R_\theta(z)/\Gamma(z),&z'=z.\end{cases}$$
- $$\Gamma_x\ge\alpha_{i,j}+\alpha_{j,i}-\psi+|\Delta\phi|,\qquad 0\le\psi\le\min\big(\alpha_{i,j}-\phi^{\uparrow},\alpha_{j,i}-\phi^{\downarrow}\big).$$
- $$\tau_{\mathrm{int}}(A)=\frac12+\sum_{\ell=1}^{\infty}\mathrm{Corr}\big(A(\omega_0),A(\omega_\ell)\big),\qquad \mathrm{ESS}\approx\frac{M}{2\tau_{\mathrm{int}}(A)}.$$

## Implementation notes

1. Exact integration point: use this as a sampler or trainer for a neural continuous-time state-space or world model. Let the latent state be a binary or discretized vector $z_t$, and let an MLP produce positive local rates $r_\theta(z\to z')=\operatorname{softplus}(f_\theta(z,z'))+\epsilon$. Condition on a fixed start state and a rare terminal predicate, such as reaching a target class or exceeding a threshold during a long rollout. 2. Pseudocode: initialize one valid trajectory; compute each escape rate $R_\theta$; choose $\Gamma=(1+\delta)\max R_\theta$ with $\delta=0.05$; insert virtual self-events; build a space-time graph from adjacent time slices and local event factors; seed random graph vertices; grow clusters using conditional factor weights while freezing boundary vertices; resample each cluster from its conditional distribution; remove virtual events; repeat. 3. Computed from the paper: uniformization, local factorization, detailed-balance cluster construction, and rejection-free conditional updates. Estimated empirically: neural rates, cluster bond probabilities, the safe upper bound for $\Gamma$, and integrated autocorrelation time. 4. First cheap experiment: use a 20-node SIS-like binary latent process or a two-dimensional neural jump process with horizon $T=100$, conditioning on an outbreak or target event whose forward probability is below $10^{-4}$. Compare independent forward rollouts, single-site Metropolis path updates, and cluster CPMC. Measure valid-trajectory rate, effective samples per network evaluation, and autocorrelation of event size. The prediction is that cluster updates retain nonzero valid-sample production as event probability decreases, while single-site updates exhibit rapidly increasing $\tau_{\mathrm{int}}$; at the smallest event probability, cluster $\tau_{\mathrm{int}}$ should be at least five times smaller and cluster acceptance should exceed 0.5.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a readable fixed-endpoint trajectory-space cluster MCMC experiment with uniformization and reversible-kernel checks. Uniformization became nonnegative at Gamma/Rmax >= 1, detailed-balance error was 1.7e-18, and measured bond probabilities matched 1-exp(-2J) within 0.011. Cluster updates improved autocorrelation at strong coupling, but were worse at weak coupling, so the observed win is conditional rather than universal.

### Mechanism check

- Verdict: Built a readable fixed-endpoint trajectory-space cluster MCMC experiment with uniformization and reversible-kernel checks. Uniformization became nonnegative at Gamma/Rmax >= 1, detailed-balance error was 1.7e-18, and measured bond probabilities matched 1-exp(-2J) within 0.011. Cluster updates improved autocorrelation at strong coupling, but were worse at weak coupling, so the observed win is conditional rather than universal.
- Confidence: 8/10
- Limitations: The test used a 1D Ising-like fixed-endpoint toy rather than a neural continuous-time state-space model or a terminal event with probability below 1e-4. It did not test neural rates, event-time resampling, a full space-time factor graph, FLOP-normalized comparisons, or high-dimensional trajectories.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_cluster.py](https://synthcore.org/code/846/bench_cluster.py)
- [bench_report.json](https://synthcore.org/code/846/bench_report.json)
- [cluster_mcmc_experiment.py](https://synthcore.org/code/846/cluster_mcmc_experiment.py)
- [report.md](https://synthcore.org/code/846/report.md)
- [report_bench_2026-09-01T123429.md](https://synthcore.org/code/846/report_bench_2026-09-01T123429.md)
- [results.json](https://synthcore.org/code/846/results.json)
- [Download all files as ZIP](https://synthcore.org/download/846)

## Disclaimer

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