# Unbiased Path-Rejection Langevin Corrector

- ID: 1822
- Canonical URL: https://synthcore.org/idea/1822/unbiased-path-rejection-langevin-corrector
- API JSON: https://synthcore.org/api/idea/1822.json
- API Markdown: https://synthcore.org/api/idea/1822.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.05022](https://arxiv.org/abs/2608.05022)
- Category: sampling
- Solves: accuracy, sample-efficiency, stability
- ML areas: diffusion-sampling, diffusion
- Math tags: stochastic-processes, probability, measure-theory, numerical-analysis
- Ratings: usefulness 7/10; difficulty 8/10; novelty 8/10

## Idea description

Replace a discretized Langevin sampler used with a neural energy model by a short underdamped diffusion proposal followed by exact path-space rejection correction. The correction uses a Girsanov likelihood ratio and an unbiased randomized estimator, so accepted samples target the continuous-time diffusion rather than a biased Euler chain.

## Mathematical statement

The paper's underdamped Langevin diffusion is the phase-space SDE dX_t=P_t dt and dP_t=(-gamma P_t-nabla V(X_t))dt+sqrt(2 gamma)dB_t, where X_t is position, P_t is momentum, V is the target potential, gamma>0 is friction, and B_t is standard Brownian motion. For a proposal drift with momentum component     ilde b_p(x,p), define the drift discrepancy Delta(x,p)=(-gamma p-nabla V(x))-    ilde b_p(x,p). Since the diffusion coefficient in momentum is sigma=sqrt(2 gamma)I, Girsanov's path-density ratio over [0,T] is R_T=exp{int_0^T (Delta(X_t,P_t)/sqrt(2 gamma))^T dB_t - (1/2)int_0^T ||Delta(X_t,P_t)||^2/(2 gamma)dt}. The paper's key computational claim is that, for a suitably short T and an appropriate rejection construction, R_T can be replaced by an unbiased estimator using O(1) queries to nabla V. A generic unbiased estimator for an exponential integral used inside this construction is: if 0<=g(t)<=lambda, U_i iid Uniform[0,T], N~Poisson(lambda T), then E[product_{i=1}^N(1-g(U_i)/lambda)]=exp(-int_0^T g(t)dt). Here g(t)=||Delta(X_t,P_t)||^2/(4 gamma), with lambda an upper bound. The accept/reject step requires a nonnegative unbiased ratio estimator and a valid envelope M satisfying R_T<=M almost surely or via the paper's corresponding randomized rejection bound.

## Key formulas

- $$\mathrm{d}X_t=P_t\,\mathrm{d}t,\qquad \mathrm{d}P_t=\bigl(-\gamma P_t-\nabla V(X_t)\bigr)\,\mathrm{d}t+\sqrt{2\gamma}\,\mathrm{d}B_t.$$
- $$R_T=\exp\left\{\int_0^T\frac{\Delta(X_t,P_t)^\top}{\sqrt{2\gamma}}\,\mathrm{d}B_t-\frac12\int_0^T\frac{\|\Delta(X_t,P_t)\|^2}{2\gamma}\,\mathrm{d}t\right\},\qquad \Delta(x,p)=(-\gamma p-\nabla V(x))-\widetilde b_p(x,p).$$
- $$N\sim\operatorname{Poisson}(\lambda T),\quad U_i\stackrel{\mathrm{iid}}\sim\operatorname{Uniform}[0,T],\quad \widehat E=\prod_{i=1}^{N}\left(1-\frac{g(U_i)}{\lambda}\right),\quad \mathbb{E}[\widehat E]=\exp\left(-\int_0^T g(t)\,\mathrm{d}t\right),\quad 0\le g(t)\le\lambda.$$
- $$\text{Accept the proposed path with probability }\alpha=\frac{\widehat R_T}{M};\qquad \mathbb{E}[\widehat R_T\mid\text{path}]=R_T.$$

## Implementation notes

(1) Integration point: modify the inference-time sampler for a neural energy model E_theta(x), where V(x)=E_theta(x) or V is a known data-consistency energy plus a neural prior. Maintain position x and momentum p. Use a cheap proposal drift     ilde b_p, for example -gamma p minus a stale, cached, low-rank, or smaller-network gradient, while the target drift uses the current full gradient nabla E_theta(x). Do not alter network weights during the first experiment. (2) Pseudocode: initialize x,p; for each proposal draw Brownian increments and integrate dX=p dt, dP=    ilde b_p(x,p)dt+sqrt(2 gamma)dB with a fine internal step; query the full gradient at randomized Poisson times U_i and at the endpoints; compute Delta=(-gamma p-nabla E_theta(x))-    ilde b_p; accumulate the stochastic integral Delta^T dB/sqrt(2 gamma) and the quadratic term ||Delta||^2/(4 gamma); form the nonnegative unbiased estimator \hat R_T using the paper's randomized path-ratio estimator, or initially use the Poisson product estimator for the quadratic exponential and a discretized stochastic-integral factor; accept with probability min(1,\hat R_T/M), otherwise restart from the previous state. (3) Compute from the mathematics: the underdamped proposal, Girsanov correction, Poisson randomization, and rejection decision. Estimate empirically: the envelope M, the maximal discrepancy bound lambda, the acceptance rate, gradient queries per accepted sample, and bias from time discretization. Start with bounded-gradient synthetic potentials so that lambda and M are known; then test a neural energy with gradient clipping to enforce a practical bound. (4) Cheap experiment: use 2D and 10D strongly convex quartic potentials and a small MLP energy, comparing corrected proposals against MALA, ULA, and ordinary underdamped Euler using the same number of full energy-gradient calls. Measure target mean and covariance error, KL or Renyi-divergence estimates, acceptance rate, and effective samples per gradient query. Success is lower discretization bias at equal compute, or at least a 2x larger stable integration step than ULA while preserving the same distributional error. First validate unbiasedness by fixing a path and averaging thousands of ratio-estimator repetitions.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built an MVP with underdamped proposal integration, discretized Girsanov diagnostics, and the nonnegative Poisson quadratic-factor estimator. The core mechanism manifested: Poisson estimates were unbiased across lambda, variance matched the predicted scaling, and zero discrepancy produced an exactly-one estimator. The constant-drift Girsanov check also matched E[R]=1; the corrected MH analogue removed Gaussian ULA variance bias, but this is supporting evidence rather than a full exact path-rejection implementation.

### Mechanism check

- Verdict: Built an MVP with underdamped proposal integration, discretized Girsanov diagnostics, and the nonnegative Poisson quadratic-factor estimator. The core mechanism manifested: Poisson estimates were unbiased across lambda, variance matched the predicted scaling, and zero discrepancy produced an exactly-one estimator. The constant-drift Girsanov check also matched E[R]=1; the corrected MH analogue removed Gaussian ULA variance bias, but this is supporting evidence rather than a full exact path-rejection implementation.
- Confidence: 9/10
- Limitations: The full exact path-rejection sampler was not implemented: the stochastic-integral factor is only a discretized diagnostic, and no almost-sure finite rejection envelope M was constructed. The sampling comparison uses exact random-walk MH as a correction analogue rather than the complete unbiased underdamped rejection algorithm. No neural energy, high-dimensional target, FLOP-normalized benchmark, or GPU experiment was run.

### 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_report.json](https://synthcore.org/code/678/bench_report.json)
- [bench_run.py](https://synthcore.org/code/678/bench_run.py)
- [energy_track.py](https://synthcore.org/code/678/energy_track.py)
- [report.md](https://synthcore.org/code/678/report.md)
- [report_bench_2026-08-31T213902.md](https://synthcore.org/code/678/report_bench_2026-08-31T213902.md)
- [results.json](https://synthcore.org/code/678/results.json)
- [sampler.py](https://synthcore.org/code/678/sampler.py)
- [verify.py](https://synthcore.org/code/678/verify.py)
- [Download all files as ZIP](https://synthcore.org/download/678)

## Disclaimer

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