Variance-Reduced ASG Scheme

• Variance-Reduced ASG Scheme is a class of algorithms designed to significantly decrease variance in simulating averaged dynamics of slow-fast stochastic systems.
• The core method uses control variates by exploiting the strong temporal correlation between consecutive estimator values when using the same random seed.
• This approach enables efficient and accurate simulation by dramatically reducing estimator variance compared to standard methods, provided proper initialization and reinitialization are applied.

A variance-reduced ASG (Averaged Stochastic Gradient) scheme is a class of algorithms designed to significantly decrease variance in the estimation of macroscopic or effective dynamics when simulating systems governed by multiscale stochastic differential equations (SDEs), particularly in slow-fast settings. The central aim is to enable accurate and efficient numerical simulation of the "averaged" behavior of the slow variables, where the driving drift or force arises from averages over the rapidly evolving (fast) stochastic subsystem. This approach builds upon the use of control variables—specifically, leveraging strong correlations along the temporal axis of the slow variable evolution—to construct estimators with much lower variance compared to naive Monte Carlo methods.

1. Variance Reduction via Control Variables

A core innovation of the variance-reduced ASG scheme is the use of control variates derived from the temporal sequence of estimators. Consider a slow-fast system where, at each macroscopic time step $n$, the drift term for the slow variable is estimated by averaging microscopic simulations of the fast variable via a Markov chain Monte Carlo (MCMC) method. The variance-reduced drift estimator at step $n$, denoted as $\overline{F}(X^n)$, is constructed using both the current and previous values of the estimator, but crucially, both are computed with the same random seed or path, yielding strongly correlated noise terms:

$$\overline{F}(X^n) = \widehat{F}(X^n, \omega_n) - \left[\widehat{F}(X^{n-1}, \omega_n) - \overline{F}(X^{n-1})\right]$$

Here:

• $\widehat{F}(X^n, \omega_n)$ is the standard MCMC-based estimator for the drift at $X^n$ using random seed $\omega_n$,
• $\overline{F}(X^{n-1})$ is the variance-reduced estimator from the previous step.

This formula exploits the fact that, when using the same random seed $\omega_n$ to sample both $X^n$ and $X^{n-1}$, the stochastic fluctuations in the two estimates are highly correlated and tend to cancel out, dramatically reducing the estimator variance.

The slow variable $X$ is then updated via a forward Euler method:

$X^{n+1} = X^n + \Delta t\, \overline{F}(X^n)$

Initialization involves an accurate estimate $\overline{F}(X^0)$, obtained from either exact calculation or from a standard estimator using many samples. To maintain low variance as the dynamics evolve, periodic reinitialization with a more accurate estimator is recommended.

2. Slow-Fast Stochastic System Structure

The variance-reduced ASG scheme is tailored to systems modeled by singularly perturbed SDEs exhibiting clear time-scale separation:

$$\begin{cases} dx(t) = f(x, y)\, dt \ dy(t) = \frac{1}{\varepsilon}g(x, y)\, dt + \frac{1}{\sqrt{\varepsilon}}\beta(x, y)\, dW(t) \end{cases}$$

• $x(t)$ is the slow variable,
• $y(t)$ is the fast variable,
• $\varepsilon \ll 1$ enforces the time-scale gap,
• $dW(t)$ is the standard Brownian motion.

The slow variable’s evolution is influenced by the fast variable, which equilibrates rapidly towards a stationary distribution for a fixed value of $x$.

3. Ergodicity and Averaged Dynamics

Averaged dynamics are justified under the assumption that the fast process $y(t)$ is ergodic with a unique invariant measure $\mu_X(dy)$ for each fixed $x = X$. In the limit $\varepsilon \to 0$, the slow variable obeys an averaged ODE:

$$\frac{dX(t)}{dt} = F(X(t)), \qquad F(X) = \int f(X, y)\ \mu_X(dy)$$

Estimation of $F(X)$ is performed through time averaging over trajectories of the fast process, or via MCMC sampling, for fixed slow variables.

4. Monte Carlo and Markov Chain Sampling

Sampling the invariant measure $\mu_X(dy)$ is ordinarily done by running the fast SDE for a number of steps $M$, starting from various initial conditions $y_0$. For each slow variable configuration, the estimator is

$$\widehat{F}(X, \omega) = \frac{1}{M} \sum_{m=1}^{M} f(X, y^m)$$

where the sequence $y^m$ is generated by discretizing the fast SDE, and $\omega$ denotes the seed that determines the random sequence of Brownian increments.

When $X$ changes incrementally over macroscopic time steps, using the same seed for simulating the fast SDEs at both $X^n$ and $X^{n-1}$ ensures strong correlation in the sample paths, which is pivotal for variance reduction.

5. Analysis of Variance and Bias

The variance-reduced ASG scheme exhibits dramatically lowered estimator variance, especially in linear systems:

• For linear problems and exact initialization (where $\overline{F}(X^0)$ is unbiased and accurate), the variance of the drift estimator decays to zero immediately and stays zero at subsequent steps:

$$\operatorname{Var}\left( \overline{F}(X^n) \right) = 0$$

• For nonlinear systems, the variance is not completely eliminated but is typically reduced by one to two orders of magnitude compared to the standard estimator, as confirmed in comprehensive numerical experiments (see Figs. 1–7 in the original paper).

The bias in the estimator is controlled by the initial estimate. Unbiased initialization ensures overall unbiasedness; otherwise, the bias is primarily dictated by the initialization error.

Periodic reinitialization (e.g., every $R$ steps) is empirically shown to be sufficient to keep variance low over long integrations.

6. Broader Implications for Multiscale and ASG-Type Methods

The control variate strategy outlined in the variance-reduced ASG scheme is broadly applicable to any multiscale simulation framework where consecutive estimates share strong statistical correlation, such as heterogenous multiscale methods (HMM), projective integration, or other averaged stochastic gradient schemes in numerical SDEs.

Key guiding principles for effective implementation include:

• Ensuring correlated estimation between consecutive steps via shared random seeds or sample paths.
• Using accurate (ideally unbiased) initialization and periodic reinitialization to manage bias and maintain low variance.
• Tuning the frequency of reinitialization based on the observed rate of drift in the slow variables and the onset of variance increase.

This approach enables the attainment of statistical accuracy comparable to that of much larger sample runs (in the standard MCMC estimator) but with far lower computational cost. For ASG methods and related multiscale schemes, it enables efficient and accurate simulation of macroscopic dynamics even when the fast subsystem is highly stochastic and computationally expensive to sample.

---

Summary Table

Aspect	Standard HMM	Variance-Reduced Scheme
Variance	$O(1/M)$	O(1/M), further reduced by control variate
Bias control	Depends on estimator	Controlled via unbiased initialization
Error correlation	Independent, uncorrelated	Correlated, manageable via reinitialization
Linear problems	Variance persists	Variance fully eliminated with exact init
Nonlinear	Significant variance	Strongly reduced; reinit. manages increases

---

The variance-reduced ASG scheme using control variates is a substantial methodological advance for the efficient numerical simulation of slow-fast stochastic systems, enabling accurate, low-variance estimation of averaged behavior with scalable computational costs, provided proper initialization and correlation structures are leveraged. Attention to ergodic properties, estimator correlation, and periodic bias correction is essential to maximize effectiveness in practical applications.