The Method of Common Random Numbers: An Example
Requires a Wolfram Notebook System
Interact on desktop, mobile and cloud with the free Wolfram CDF Player or other Wolfram Language products.
Requires a Wolfram Notebook System
Edit on desktop, mobile and cloud with any Wolfram Language product.
Variance reduction is of great interest to the creators of Monte Carlo experiments. For example, investment banks use very complicated Monte Carlo simulations to price esoteric mortgage-backed securities. These simulations often run overnight because many Monte Carlo trials are necessary to obtain (by the central limit theorem) a point estimate of some true population parameter, bounded by a relatively small confidence interval. One way to reduce the number of required Monte Carlo trials is to use a variance reduction technique.[more]
The method of common random numbers is one such technique. It is useful in Monte Carlo experiments generally, including Monte Carlo integration. We illustrate its use with a simple example.
Let and , and and . Mathematica's numerical integration techniques tell us that , but this fact is not necessarily clear from the graphs of these two functions over the unit interval.
We propose two different techniques for estimating .
1. Suppose that we estimate and by and , where and are sequences of independent random numbers from the unit interval. Then .
2. Now consider the following alternative. Let and be positively correlated, but identically distributed uniform random variables. Estimate according to the rule .
The Demonstration output shows that is a modestly better way to estimate than . The histogram on the left features a simulated collection of the quantity . The histogram on the right features a simulated collection of the quantity .
Run the simulation repeatedly to see that in this example, the method of common random numbers always results in a reduction of variance.[less]
This method of common numbers produces good, but not overwhelming, variance reduction. The method of common random numbers (also known as the method of correlated sampling, the method of matched pairs, or the method of matched sampling) does not always work. It can backfire if the the engineer of the Monte Carlo simulation creates a negative, rather than positive, correlation between the two random variables and . Often, it is useful to choose , which we do in this example.
Why does this method work?
Recall that, for example, is a sequence of independent, identically distributed random variables. The variance of the first Monte Carlo method, when is independent of , is
Now consider the second Monte Carlo method, the method of common random numbers. The variance of this Monte Carlo method, when is positively correlated to , is
If we make the additional assumption that and are either: (a) both monotonically nondecreasing; or (b) both monotonically nonincreasing, then
and we see that .
Notice that in both cases, and have the identical marginals—the method of common numbers only permits us to manipulate the joint distribution of and . Also, notice that in our example, both and are strictly monotonically increasing.
In this particular example, the variance reduction is always successful. However, notice that other measures of dispersion—like the range or the interquartile range—are not always reduced by the technique.
There are more powerful variance reduction techniques available, including antithetic variates, control variates, importance sampling, and stratified sampling. For more information on the method of common random numbers, see Sheldon Ross's textbook Stochastic Processes or Paul Glasserman's book on Monte Carlo methods.