Ratio-of-Means Control-Variate Estimator

• The ratio-of-means control-variate estimator is a variance reduction method that uses auxiliary variables to improve Monte Carlo estimation of ratios of expected values.
• It jointly regresses predictable components from both the numerator and denominator to minimize variance while preserving unbiasedness under mild moment conditions.
• The method achieves significant efficiency gains in simulation and experimental design, with scalable adaptations even in high-dimensional settings.

The ratio-of-means control-variate estimator is an advanced variance reduction methodology for Monte Carlo estimation of ratios of expected values, applicable when the estimand is $R = \mathbb{E}[X]/\mathbb{E}[Y]$ and high-variance baseline estimators hamper practical efficiency. Systematically incorporating auxiliary variables (control variates) correlated with both the numerator and denominator, the estimator jointly regresses out predictable components to optimize estimator variance, yielding substantial efficiency gains without introducing bias under mild moment conditions. Recent work provides rigorous theoretical guarantees, optimality criteria, and scalable estimation procedures for both classical and high-dimensional regimes, positioning this estimator as a default tool for ratio-metric estimation in simulation and experimental design contexts (Bocquet-Nouaille et al., 15 Oct 2025, Jin et al., 2021).

1. Problem Formulation and Classical Estimator

The quantity of interest is a ratio of expectations: for i.i.d. samples $(X_i, Y_i) \sim (X, Y)$ for $i=1,\ldots,n$, estimate

$$R = \frac{\mathbb{E}[X]}{\mathbb{E}[Y]}, \quad \mathbb{E}[Y] \neq 0.$$

The standard Monte Carlo (MC) “ratio-of-means” estimator is

$$\widehat{R}_{\mathrm{MC/MC}} = \frac{\bar{X}}{\bar{Y}}, \qquad \bar{X} = \frac{1}{n} \sum_{i=1}^n X_i, \; \bar{Y} = \frac{1}{n} \sum_{i=1}^n Y_i.$$

By a first-order Taylor expansion (delta method), its variance is approximately

$$\operatorname{Var}(\widehat{R}_{\mathrm{MC/MC}}) \approx \frac{1}{n} \left[ \tfrac{1}{\mathbb{E}[Y]^2}\operatorname{Var}(X) + \tfrac{\mathbb{E}[X]^2}{\mathbb{E}[Y]^4}\operatorname{Var}(Y) - 2\frac{\mathbb{E}[X]}{\mathbb{E}[Y]^3}\operatorname{Cov}(X,Y) \right].$$

This estimator is unbiased up to order $\mathcal{O}(n^{-2})$ but can suffer prohibitively large variance, especially when $n$ is moderate or when $X$ and $Y$ are weakly correlated (Bocquet-Nouaille et al., 15 Oct 2025).

2. Control Variates: Joint Optimal Adjustment

Suppose auxiliary random vectors $U \in \mathbb{R}^p$, $V \in \mathbb{R}^q$ are available, selected for their correlation with $X$ and $Y$, respectively, and their means $\mathbb{E}[U], \mathbb{E}[V]$ are either known analytically or cheaply estimated. The estimator incorporates control correction vectors $c_X \in \mathbb{R}^p$ and $c_Y \in \mathbb{R}^q$: $$\widehat{R}_{c} = \frac{ \bar{X} - c_X^\top(\bar{U} - \mathbb{E}[U]) }{ \bar{Y} - c_Y^\top(\bar{V} - \mathbb{E}[V]) }.$$ This formulation generalizes the classical control-variates method, leveraging joint optimization across both the numerator and denominator. The strategy eliminates variance components in $X$ and $Y$ that are predictable from $U$ and $V$, respectively, while preserving unbiasedness at order $\mathcal{O}(n^{-2})$ (Bocquet-Nouaille et al., 15 Oct 2025).

3. Optimal Coefficient Determination and Variance Properties

For efficiency, the coefficient vectors $c_X$ and $c_Y$ are optimized to minimize the estimator’s leading-order variance. Defining

$$\tilde{X} = \bar{X} - c_X^\top(\bar{U} - \mathbb{E}[U]), \quad \tilde{Y} = \bar{Y} - c_Y^\top(\bar{V} - \mathbb{E}[V]),$$

and denoting $\mu_X = \mathbb{E}[X]$, $\mu_Y = \mathbb{E}[Y]$, the delta-method approximation yields

$$\operatorname{Var}(\widehat{R}_c) \approx \frac{1}{\mu_Y^2} \operatorname{Var}(\tilde{X}) + \frac{\mu_X^2}{\mu_Y^4} \operatorname{Var}(\tilde{Y}) - 2\frac{\mu_X}{\mu_Y^3} \operatorname{Cov}(\tilde{X}, \tilde{Y}).$$

Defining

$$Z = \begin{pmatrix} \bar{X} - \mu_X \ \bar{U} - \mathbb{E}[U] \ \bar{Y} - \mu_Y \ \bar{V} - \mathbb{E}[V] \end{pmatrix} \sim N\left(0, \tfrac{1}{n}\Sigma\right),$$

where $\Sigma$ is the joint covariance, let $c = \begin{pmatrix}c_X\c_Y\end{pmatrix}$ and

$$b = \begin{pmatrix} \tfrac{1}{\mu_Y} \operatorname{Cov}(X,U) - \tfrac{\mu_X}{\mu_Y^2} \operatorname{Cov}(Y,U) \ \tfrac{1}{\mu_Y} \operatorname{Cov}(X,V) - \tfrac{\mu_X}{\mu_Y^2} \operatorname{Cov}(Y,V) \end{pmatrix},$$

with covariance block

$$\Sigma = \operatorname{Var}((\bar{U} - \mathbb{E}[U], \bar{V} - \mathbb{E}[V])).$$

The unique minimizer is

$c^* = \Sigma^{-1}b,$

guaranteeing

$$\operatorname{Var}(\widehat{R}_c) \leq \operatorname{Var}(\widehat{R}_{\mathrm{MC/MC}})$$

with non-negative variance reduction, and the reduction term vanishes (i.e., estimator is fully efficient) if $b = 0$ (Bocquet-Nouaille et al., 15 Oct 2025, Jin et al., 2021).

4. Implementation Procedures and Algorithmic Details

Estimation of $c^*$ is conducted using empirical covariance estimators: $$\widehat{\Sigma} = \begin{bmatrix} \widehat{\Sigma}_{UU} & \widehat{\Sigma}_{UV} \ \widehat{\Sigma}_{VU} & \widehat{\Sigma}_{VV} \end{bmatrix}, \quad \widehat{b} = \begin{pmatrix} \tfrac{1}{\bar{Y}} \widehat{\operatorname{Cov}(X,U)} - \tfrac{\bar{X}}{\bar{Y}^2} \widehat{\operatorname{Cov}(Y,U)} \ \tfrac{1}{\bar{Y}} \widehat{\operatorname{Cov}(X,V)} - \tfrac{\bar{X}}{\bar{Y}^2} \widehat{\operatorname{Cov}(Y,V)} \end{pmatrix}.$$ The empirical optimal coefficients are $\widehat{c} = \widehat{\Sigma}^{-1} \widehat{b}$. The adjusted data are

$$\tilde{X}_i = X_i - \widehat{c}_X^\top(U_i - \bar{U}),\quad \tilde{Y}_i = Y_i - \widehat{c}_Y^\top(V_i - \bar{V}),$$

with sample means used to form the final estimator: $$\widehat{R}_c = \frac{\bar{\tilde{X}}}{\bar{\tilde{Y}}}.$$ Numerical regularization by ridge stabilization, e.g., $\bar{\tilde{Y}}+\varepsilon$ with $\varepsilon=10^{-8}$, prevents instability if the denominator becomes small. All steps are justified by delta-method expansions and finite-sample Taylor analysis (Bocquet-Nouaille et al., 15 Oct 2025).

5. Extensions and Practical Adaptations

The method accommodates multiple control variates simply by increasing the dimensions of $U$ and $V$, with the same joint optimization ($\Sigma^{-1}b$) framework. When $\mathbb{E}[U]$ or $\mathbb{E}[V]$ are themselves unknown but can be estimated from additional unlabeled samples, the same coefficient formula applies, and variance reduction holds modulo scaling by the fraction of control-only samples. In high-dimensional scenarios, regularization (ridge on $\widehat{\Sigma}$ or $\ell_1$ penalties on $c$) is recommended to prevent overfitting when $n < p+q$ (Bocquet-Nouaille et al., 15 Oct 2025).

For experimental design and randomized trials, analogous strategies regress out predicted components from both numerator and denominator using observed covariates $X_i$. This reduces to jointly residualizing treatment and control group means before forming the difference-of-ratios, as expounded in the analysis of ratio metrics in controlled experiments (Jin et al., 2021).

6. Application Domains and Empirical Performance

In multi-fidelity modeling, such as aircraft design, high-fidelity simulations yield expensive strut-mass and total-mass estimates, while low-fidelity models provide correlated, inexpensive controls. For $n=200$ or $500$ high/low-fidelity runs plus $m\approx1000$ low-fidelity controls for mean estimation, observed correlations between output and controls (0.5–0.8) lead to empirical relative variance reductions of approximately 20% for the optimal control-variates estimator, reducing high-fidelity sample requirements by the same factor (Bocquet-Nouaille et al., 15 Oct 2025). In large-scale online experiments, optimized ratio-of-means control-variates estimators show up to 80% variance reduction compared to naive estimators, with further gains over baseline approaches like CUPED in the presence of high-dimensional covariates and cross-fitting (Jin et al., 2021).

---

References

• Louison Bocquet-Nouaille, Jérôme Morio & Benjamin Bobbia. “Control variates for variance-reduced ratio of means estimators” (Bocquet-Nouaille et al., 15 Oct 2025).
• Alexander Volfovsky, et al. "Towards Optimal Variance Reduction in Online Controlled Experiments" (Jin et al., 2021).