Control Variates and Optimal Coefficients

Why Use a Control Variate?

A Monte Carlo estimate can be noisy even when the target expectation is straightforward to define. If each simulated target output is paired with another output that is strongly correlated with it and has a known expectation, that second output can help correct the estimate’s random fluctuation. This technique is called control variates.

For example, a hypothetical insurer might simulate a complicated annual loss $X$ while also recording a related quantity $C$, such as a simplified loss measure whose expected value is known analytically. When simulated values of $C$ are above their known mean, values of $X$ may also tend to be high. Adjusting for that deviation can reduce simulation error. The adjustment works only if the relationship is useful and the control’s expectation is known correctly.

Definition: Let $X$ be the simulation output whose expectation $\theta=E[X]$ is to be estimated. A control variate $C$ is a random variable simulated alongside $X$, with known expectation $\mu_C=E[C]$, that is correlated with $X$. For a fixed coefficient $b$, the adjusted output is $X-b(C-\mu_C)$. Its expectation remains $\theta$, provided $E[X]$ and $E[C]$ exist and the stated value of $\mu_C$ is correct.

The Estimator and Its Variance

Suppose the simulation produces $n$ independent, identically distributed pairs $(X_i,C_i)$, with the two outputs in each pair generated together. Write $\bar X=n^{-1}\sum_{i=1}^n X_i$ and $\bar C=n^{-1}\sum_{i=1}^n C_i$. For a fixed coefficient $b$, the control-variate estimator is

\hat{\theta}_b=\bar X-b(\bar C-\mu_C).

The minus sign is a convention: a positive $b$ subtracts the control’s deviation from its known mean. A negative optimal coefficient is possible. Since $E[\bar C-\mu_C]=0$, the adjustment has expectation zero, so $E[\hat{\theta}_b]=E[\bar X]=\theta$. Thus the estimator is unbiased when the inputs satisfy the stated assumptions.

To use the variance formula and its minimizing coefficient below, assume $E[X^2]<\infty$, $E[C^2]<\infty$, and $0<\operatorname{Var}(C)<\infty$. These conditions ensure the relevant variances and covariance are finite. The pairs must be independent across simulation replications; $X_i$ and $C_i$ within a pair need not be independent.

Theorem: Under these assumptions, for a fixed $b$, $$ \operatorname{Var}(\hat{\theta}_b)=\frac{1}{n}\left[\operatorname{Var}(X)+b^2\operatorname{Var}(C)-2b\operatorname{Cov}(X,C)\right]. $$ The variance-minimizing coefficient and the resulting variance are $$ b^*=\frac{\operatorname{Cov}(X,C)}{\operatorname{Var}(C)},\qquad \operatorname{Var}(\hat{\theta}_{b^*})=\frac{1}{n}\left[\operatorname{Var}(X)-\frac{\operatorname{Cov}(X,C)^2}{\operatorname{Var}(C)}\right]. $$

Why the Optimal Coefficient Has This Form

For one replication, the adjusted output is $Z_b=X-b(C-\mu_C)$. Subtracting a constant does not change variance, so its variance is the variance of $X-bC$. Expanding the variance gives

\operatorname{Var}(Z_b)=\operatorname{Var}(X)+b^2\operatorname{Var}(C)-2b\operatorname{Cov}(X,C).

For $n$ independent adjusted outputs, the variance of their average is this one-replication variance divided by $n$. To minimize the expression in $b$, complete the square:

\begin{aligned} \operatorname{Var}(X)+b^2\operatorname{Var}(C)-2b\operatorname{Cov}(X,C) &=\operatorname{Var}(C)\left(b-\frac{\operatorname{Cov}(X,C)}{\operatorname{Var}(C)}\right)^2\\ &\quad+\operatorname{Var}(X)-\frac{\operatorname{Cov}(X,C)^2}{\operatorname{Var}(C)}. \end{aligned}

Because $\operatorname{Var}(C)>0$, the squared term is smallest, and equals zero, at $b=b^*$. Dividing the minimized one-replication variance by $n$ gives the theorem’s variance for the sample estimator. This derivation also shows why the finite-second-moment assumptions matter: without finite variances and covariance, the quadratic expression and its minimization need not be valid.

Interpretation: If $\rho=\operatorname{Corr}(X,C)$, then the optimal one-replication variance is $\operatorname{Var}(X)(1-\rho^2)$, and the optimal estimator variance is $\operatorname{Var}(X)(1-\rho^2)/n$. A control with zero correlation gives no variance reduction; a control with strong positive or negative correlation can be effective. The sign of $b^*$ follows the sign of the covariance.

Calculating an Adjusted Estimate and Standard Error

When the population covariance and variances are provided, use them to calculate $b^*$ and the theoretical standard error. In an actual simulation, they are often unknown. A coefficient can then be estimated from paired simulation outputs. For a fixed chosen coefficient $b$, form the adjusted replication values $Z_i=X_i-b(C_i-\mu_C)$; an estimated standard error is the sample standard deviation of the $Z_i$ values divided by $\sqrt n$. If $b$ itself is estimated from the same data, treating it as fixed is generally an approximation; a separate pilot run or an appropriate uncertainty assessment may be needed.

Worked Example: A Negatively Correlated Control

A hypothetical simulation estimates an annual cost in dollars. It produces $n=100$ paired outputs with $\bar X=\$120.40$ and $\bar C=3.20$; the known control mean is $\mu_C=3.00$. The population moments are $\operatorname{Var}(X)=49$, $\operatorname{Var}(C)=16$, and $\operatorname{Cov}(X,C)=-21$, in compatible units. Find the optimal coefficient, adjusted estimate, and standard error.

Method. Use the covariance-to-variance ratio for the coefficient, then substitute it in the adjusted estimator and minimized variance formula. The negative covariance means the optimal coefficient will be negative.

Coefficient. $b^*=-21/16=-1.3125$.

Adjusted estimate. The control deviation is $\bar C-\mu_C=3.20-3.00=0.20$. Therefore, $\hat{\theta}_{b^*}=120.40-(-1.3125)(0.20)=120.40+0.2625=\$120.6625$, or $\$120.66$ to the nearest cent.

Standard error. The minimized one-replication variance is $49-(-21)^2/16=49-441/16=49-27.5625=21.4375$. The estimator variance is $21.4375/100=0.214375$, so its standard error is $\sqrt{0.214375}=\$0.463006\ldots$, or $\$0.46301$ to five decimal places.

Check and mistake avoided. The variance reduction is positive, and the standard error is less than the unadjusted value $7/\sqrt{100}=\$0.70$. A common sign error is to change the estimator to $\bar X+b(\bar C-\mu_C)$ while retaining the coefficient formula for the minus-sign convention. Here the negative coefficient under the stated convention correctly raises the estimate because the observed control is above its mean.

Worked Example: Positive Covariance and a Smaller Standard Error

A hypothetical simulation reports an output in thousands of dollars. From $n=64$ paired replications, $\bar X=52.50$ and $\bar C=10.75$, while $E[C]=11.00$. The supplied moments are $\operatorname{Var}(X)=36$, $\operatorname{Var}(C)=9$, and $\operatorname{Cov}(X,C)=12$. Calculate the optimal coefficient, adjusted estimate, and minimized standard error.

Coefficient. $b^*=12/9=4/3$.

Adjusted estimate. The control deviation is $10.75-11.00=-0.25$. Thus $\hat{\theta}_{b^*}=52.50-(4/3)(-0.25)=52.50+1/3=52.833333\ldots$ thousand dollars, or $52.833$ thousand dollars to three decimal places.

Standard error. The minimized one-replication variance is $36-12^2/9=36-144/9=36-16=20$. The variance of the estimate is $20/64=0.3125$, so the standard error is $\sqrt{0.3125}=0.559016\ldots$ thousand dollars, or $0.559$ thousand dollars to three decimal places.

Check and mistake avoided. The unadjusted standard error is $\sqrt{36/64}=0.75$ thousand dollars, so the adjusted standard error is smaller. Do not stop at $20/64=0.3125$: that is the variance of the estimator, not its standard error. The standard error is its square root.

Worked Example: Applying the Coefficient to a Hypothetical Loss Estimate

A hypothetical insurer simulates a loss $X$ in dollars and a correlated auxiliary quantity $C$, measured in control units. For $n=25$ paired replications, the simulation summary is $\bar X=\$1{,}040$, $\bar C=12.4$, and the known control expectation is $E[C]=12.0$. The supplied moments are $\operatorname{Var}(X)=25$ dollars squared, $\operatorname{Var}(C)=4$ control units squared, and $\operatorname{Cov}(X,C)=8$ dollars times control units. Find the optimal estimate and its standard error.

Coefficient. $b^*=8/4=2$ dollars per control unit.

Adjusted estimate. The observed control deviation is $12.4-12.0=0.4$ control units. The adjusted estimate is $\$1{,}040-2(0.4)=\$1{,}039.20$.

Standard error. The minimized one-replication variance is $25-8^2/4=25-64/4=25-16=9$ dollars squared. The variance of the estimate is $9/25=0.36$ dollars squared, giving a standard error of $\sqrt{0.36}=\$0.60$.

Check and mistake avoided. The unadjusted standard error is $\sqrt{25/25}=\$1.00$; the control reduces it to $\$0.60$. The coefficient has units of dollars per control unit, so multiplying it by the $0.4$-unit deviation produces a dollar adjustment. Ignoring units can conceal an incorrect scaling of the control.

Common Exam Traps

Using the wrong sign convention. For $\bar X-b(\bar C-\mu_C)$, use $b^*=\operatorname{Cov}(X,C)/\operatorname{Var}(C)$. If the estimator is instead written with a plus sign, the minimizing coefficient has the opposite sign.
Confusing variance with standard error. First divide the adjusted one-replication variance by $n$; then take the square root to obtain the standard error.
Assuming every correlated control helps equally. The benefit depends on squared correlation. A weakly correlated control may give little improvement, and an uncorrelated control gives none at the optimum.
Forgetting the known mean. The adjustment uses $\bar C-\mu_C$, not $\bar C$ by itself. Centering at the known expectation preserves the target expectation.
Omitting moment assumptions. A positive finite $\operatorname{Var}(C)$ alone is not enough for the variance result. The target and control must have finite second moments for the stated formulas.

Key takeaway: A control variate preserves the target expectation while adjusting for a correlated simulation fluctuation. With independent paired replications, finite second moments, and a control of positive finite variance, the optimal coefficient is covariance divided by control variance. The adjusted estimator’s standard error is the square root of its minimized variance divided by the replication count.

Check Your Understanding

For each problem, use the minus-sign estimator convention and report the requested result.

A simulation uses $n=50$, with $\operatorname{Var}(X)=18$, $\operatorname{Var}(C)=2$, and $\operatorname{Cov}(X,C)=3$. Find the optimal coefficient and minimized standard error. (Answer: $b^*=1.5$; standard error $=\sqrt{0.27}\approx0.5196$.)
For $n=100$, suppose $\operatorname{Var}(X)=64$, $\operatorname{Var}(C)=16$, and $\operatorname{Cov}(X,C)=-24$. Find $b^*$ and the minimized standard error. (Answer: $b^*=-1.5$; standard error $=\sqrt{0.28}\approx0.5292$.)
A run has $n=25$, $\bar X=80$, $\bar C=5.5$, and known $E[C]=6$. Given $\operatorname{Var}(X)=20$, $\operatorname{Var}(C)=4$, and $\operatorname{Cov}(X,C)=6$, find the adjusted estimate and its standard error. (Answer: $b^*=1.5$; estimate $=80.75$; standard error $=\sqrt{11/25}\approx0.6633$.)
A target has variance $100$, and a control has variance $25$ and covariance $40$ with the target. For $n=100$, find the optimal coefficient and standard error. (Answer: $b^*=1.6$; minimized one-replication variance $=36$; standard error $=\sqrt{36/100}=0.6$.)
Suppose $X$ and $C$ are uncorrelated, $\operatorname{Var}(X)=9$, $\operatorname{Var}(C)=4$, and $n=36$. What is the optimal coefficient, and does the control reduce the standard error? (Answer: $b^*=0$; no reduction; standard error remains $\sqrt{9/36}=0.5$.)

What You'll Learn