Monte Carlo Precision Settings in Stochastic Estimation

Why Monte Carlo Precision Matters

A stochastic estimation procedure may evaluate a likelihood or objective using simulated draws. At fixed model parameters and fixed data, changing the random-number stream can then change the reported value. This variation is Monte Carlo error: numerical variation caused by using a finite simulation rather than evaluating the target exactly. It is separate from sampling uncertainty in the data and from uncertainty in the estimated model parameters.

A simulation setting is not adequate merely because an algorithm completes or returns many digits. The relevant question is whether random-number variation is small enough for the intended use. A noisy objective can obscure whether two parameter values differ, make a reported objective value hard to reproduce, or lead a user to mistake simulation noise for a change in model fit. Conversely, greatly increasing simulation effort when the existing error is already negligible wastes computation.

Two counts must be kept distinct. Let \(M\) be the number of simulation draws used in one objective evaluation, and let \(R\) be the number of repeated evaluations at the same parameter value, using independent random-number streams. Increasing \(M\) can improve the precision of each evaluation. Repeating the evaluation \(R\) times helps estimate how much random-number variation remains at that setting.

Estimating Variation From Repeated Evaluations

Suppose \(Q_r(\theta)\) is the objective returned by replicate \(r\), for \(r=1,\ldots,R\). For these calculations to estimate ordinary run-to-run Monte Carlo variation, the evaluations should use the same data, model, parameter value, and simulation setting, and their random-number streams should be independent. Define their mean and sample standard deviation as follows.

The sample standard deviation \(s_{\mathrm{MC}}\) estimates the spread of an individual objective evaluation. The estimated standard error \(\widehat{SE}(\bar{Q})\) applies to the average of the \(R\) replicate values. If the practical task uses one evaluation, the run-to-run spread is usually more directly relevant than the precision of the replicate average. If the task averages multiple independent evaluations, the standard error describes that average.

Repeated evaluations do not reveal whether the expected stochastic objective is itself an accurate approximation to the desired target. They quantify variation around the expected value of the procedure being run. A procedure can be repeatable but systematically biased. For example, taking the logarithm of an unbiased likelihood estimate does not in general produce an unbiased log-likelihood estimate.

Why More Simulation Draws Usually Reduce Error

A basic Monte Carlo average illustrates the usual rate. Let \(Z_1,\ldots,Z_M\) be independent draws from the same distribution, and suppose the quantity of interest is \(\mu_Z=E[Z_i]\), with finite variance \(\sigma_Z^2\). The simulation estimate is \(\bar{Z}=M^{-1}\sum_{i=1}^{M}Z_i\). Independence gives the variance of the average:

The Monte Carlo standard deviation is therefore \(\sigma_Z/\sqrt{M}\). In practice, \(\sigma_Z\) is not known, so replicate evaluations or within-run diagnostics are used to estimate precision. The square-root result explains why multiplying the simulation size by four is expected to halve the Monte Carlo standard deviation, not reduce it by a factor of four. It relies on independent draws with finite variance and on the estimator behaving like an average. Correlated draws or highly variable contributions can make the practical improvement smaller or less predictable.

The rate concerns random variation, not every source of numerical error. It does not imply that a finite-\(M\) estimate is unbiased, that all stochastic likelihood procedures use a simple sample mean, or that a reported log objective inherits the unbiasedness of an underlying likelihood estimate. Repeated evaluation at each setting is a practical way to check the actual behavior of the objective being used.

Relating Monte Carlo Error to a Practical Decision

An error estimate is useful only in relation to the scale of the decision. If two parameter values produce objective values that differ by less than the run-to-run noise, a single evaluation may not reliably distinguish them. If the difference is large relative to the estimated noise, it is less likely to be explained by that noise alone. This comparison is a diagnostic, not automatically a formal hypothesis test or a guarantee that one parameter value is preferable.

When comparing two parameter values, using the same random-number draws for both can make their difference more precise. This approach is often called common random numbers. The comparison must then be based on paired differences, not on treating the two sets of evaluations as independent. The advantage occurs when the paired evaluations tend to move together across random-number streams; it is not guaranteed for every implementation.

Choosing and Diagnosing a Precision Setting

A practical workflow sets a tolerance before increasing simulation effort. The tolerance should be expressed in objective units and chosen for the decision being made. For example, a tolerance for comparing candidate parameter values may need to be tighter than one used merely to report a broad objective trend. Evaluate the objective repeatedly at the same fixed parameter value, estimate its spread, and compare that spread with the tolerance. Increase the simulation size if the observed variation is too large, then repeat the check. A small standard error for the replicate mean alone is not enough if the procedure will use a single noisy evaluation.

For a stochastic likelihood estimate \(\widehat{L}\), precision and bias require additional care. Even when \(E[\widehat{L}]=L\), the logarithm is nonlinear, and generally \(E[\log\widehat{L}]\ne\log L\). For estimates concentrated near a positive \(L\), a first-order Taylor expansion gives the approximate relation \(\operatorname{Var}(\log\widehat{L})\approx\operatorname{Var}(\widehat{L})/L^2\). A second-order expansion also indicates a downward log-scale bias of approximately \(-\operatorname{Var}(\widehat{L})/(2L^2)\). These approximations require relatively small variation around a positive target; they are not a correction that should be applied blindly. Repeated log-likelihood evaluations can show their variability but cannot, by themselves, determine their bias.

Check that replicate variability is not being confused with a changing computation. If parameters, data, model settings, or code differ between runs, the observed spread is not a clean estimate of Monte Carlo error at fixed conditions. Record the simulation size and random-number handling so the evaluation can be reproduced. Very few replicates give a noisy estimate of the spread; unusually large outliers or a spread that changes erratically as \(M\) increases can signal unstable contributions, insufficient simulation, or a failure of the independence and finite-variance conditions behind the simple square-root argument.

Some stochastic calculations use correlated draws, such as draws from a simulation chain. In that case, treating every draw as independent and dividing an estimated standard deviation by \(\sqrt{M}\) can understate Monte Carlo error. Use a method that accounts for dependence, such as batch means, or compare independent runs with suitable diagnostics. The repeat-evaluation formulas above also rely on independent replicates; shared random numbers are useful for paired comparisons, but they require analysis of the paired differences.

The earlier material on importance sampling for likelihood evaluation provides context for simulation-based likelihood calculations. Here the central task is narrower: quantify the random variation in the objective actually returned and decide whether that variation is acceptable. Monte Carlo precision settings do not replace model checking, data-based uncertainty assessment, or evaluation of approximation bias.

Common Mistakes

• Using \(s_{\mathrm{MC}}/\sqrt{R}\) as the noise of one run. That is the estimated standard error of the mean of \(R\) independent runs. The estimated run-to-run spread of an individual run is \(s_{\mathrm{MC}}\).
• Increasing \(M\) by a factor of four and expecting fourfold precision improvement. Under the independent-draw average assumptions, the standard deviation falls by a factor of about two.
• Calling repeatability accuracy. Small variation across runs shows precision, not that the expected value equals the desired likelihood or objective.
• Comparing paired evaluations as if they were independent. With common random numbers, estimate the mean and standard error from the replicate differences.
• Assuming the square-root formula applies to correlated draws without adjustment. Dependence reduces the information in a fixed number of draws, so use dependence-aware error estimates.