# Noise-aware randomized Milstein sampler

- ID: 1610
- Canonical URL: https://synthcore.org/idea/1610/noise-aware-randomized-milstein-sampler
- API JSON: https://synthcore.org/api/idea/1610.json
- API Markdown: https://synthcore.org/api/idea/1610.md
- Verification status: unverified
- Source: [arXiv:2607.29578](https://arxiv.org/abs/2607.29578)
- Category: sampling
- Solves: accuracy, speedup, stability
- ML areas: diffusion-sampling, diffusion, inference-speedup
- Math tags: stochastic-processes, numerical-analysis, probability, dynamical-systems
- Ratings: usefulness 5/10; difficulty 5/10; novelty 6/10

## Idea description

Replace an Euler-Maruyama reverse-diffusion sampler with a scalar or coordinatewise randomized Milstein step that uses an autodifferentiated score or drift derivative and explicitly tolerates noisy coefficient and Brownian evaluations. Use the paper's additive error law to stop refining the time grid when discretization error falls below the neural-oracle noise floor.

## Mathematical statement

The paper studies the scalar SDE $\mathrm{d}X(t)=a(t,X(t))\,\mathrm{d}t+b(t,X(t))\,\mathrm{d}W(t)$ with terminal approximation error measured in $L^r$, where $a$ is the drift, $b$ is the diffusion coefficient, $W$ is standard Brownian motion, and $r\geq2$. Its main bound is $\|X(T)-\widehat X_n(T)\|_{L^r}\leq C\left(n^{-\min\{\gamma_1+1/2,\gamma_2\}}+\delta_1+\delta_2+\delta_3\right)$, where $n$ is the number of time steps, $\gamma_1$ and $\gamma_2$ are temporal Holder exponents of the drift and diffusion, respectively, $\delta_1$ bounds noisy drift observations, $\delta_2$ bounds noisy diffusion and diffusion-derivative observations, and $\delta_3$ bounds noisy Wiener-path observations. The Milstein correction is $\frac12 b\,\partial_xb\,((\Delta W)^2-h)$, where $h$ is the step size and $\Delta W$ is a Brownian increment. This correction captures leading stochastic variation omitted by Euler. In a diffusion sampler, $a$ is the learned reverse drift, $b$ is the prescribed diffusion amplitude, and $\partial_xb$ is the scalar derivative or a diagonal/Jacobian-vector analogue.

## Key formulas

- $$\mathrm{d}X(t)=a(t,X(t))\,\mathrm{d}t+b(t,X(t))\,\mathrm{d}W(t),\qquad X(0)=\eta.$$
- $$\|X(T)-\widehat X_n(T)\|_{L^r}\leq C\left(n^{-\min\{\gamma_1+\frac12,\gamma_2\}}+\delta_1+\delta_2+\delta_3\right),\qquad r\geq2.$$
- $$X_{k+1}=X_k+h\,\widetilde a(\tau_k,X_k)+\widetilde b(\tau_k,X_k)\,\widetilde{\Delta W}_k+\frac12\widetilde b(\tau_k,X_k)\,\widetilde{b_x}(\tau_k,X_k)\left(\widetilde{\Delta W}_k^2-h\right).$$
- $$h=T/n,\qquad \tau_k\sim\mathrm{Uniform}[t_k,t_{k+1}],\qquad t_k=kh,\qquad \widetilde a=a+e_a,\ \widetilde b=b+e_b,\ \widetilde{b_x}=\partial_xb+e_{b_x},\ \widetilde{\Delta W}_k=\Delta W_k+e_{W,k}.$$

## Implementation notes

(1) Exact integration point: modify the inference-time sampler in a score-based diffusion model, initially in one dimension or with diagonal diffusion. At each step, evaluate the reverse drift network $a_\theta(t,x)$, the diffusion amplitude $b(t,x)$, and $b_x(t,x)$ or the appropriate diagonal derivative using automatic differentiation. (2) Pseudocode: set $h=T/n$ and $t_k=kh$; for each $k$, draw $\tau_k\sim\mathrm{Uniform}[t_k,t_{k+1}]$, draw $\Delta W_k\sim\mathcal N(0,h)$, optionally add measured or synthetic noise $e_{W,k}$, compute $\tilde a=a_\theta(\tau_k,x_k)$, $\tilde b=b(\tau_k,x_k)$, and $\widetilde{b_x}=\partial_xb(\tau_k,x_k)$, then update $x_{k+1}$ with the displayed Milstein formula. Use common random numbers for Euler and Milstein comparisons. (3) The mathematical quantity computed exactly is the Milstein correction and the exponent $p=\min(\gamma_1+1/2,\gamma_2)$. Estimate $\delta_1$ and $\delta_2$ by comparing the deployed network and derivative against a larger teacher network or repeated noisy evaluations; estimate $\delta_3$ from Brownian quantization or injected increment noise. Choose the smallest $n$ satisfying $n^{-p}\leq\widehat\delta_1+\widehat\delta_2+\widehat\delta_3$. (4) First experiment: use a one-dimensional Ornstein-Uhlenbeck reverse process and a nonlinear score model trained on a two-component Gaussian mixture. Compare Euler-Maruyama, deterministic-time Milstein, and randomized Milstein at equal network evaluations under controlled score and Brownian noise. Measure Wasserstein distance to the exact target, terminal strong error under shared Brownian paths, and runtime. Success means lower error than Euler at equal steps and saturation at the predicted oracle-noise floor; the adaptive rule should achieve the same error with fewer steps than a fixed high-resolution sampler.

## Disclaimer

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