CENTRAL LIMIT THEOREM ΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩ Central Limit Theorem If X1, X2, X3, … is a sequence of independent random variables, each taken as an observation from the same population, and if the population mean is and the standard deviation is (finite), then the random variable sequence Z1, Z2, Z3, … Xn where Z n n converges in distribution to N(0, 1). Some notes: Ω d d 1. This can be restated as { Zn } N(0, 1). The symbol should be read as “converges in distribution.” 2. Convergence in distribution is a statement about the cumulative distribution functions F1 (for Z1), F2 (for Z2), F3 (for Z3), and so on, and also F (for the limiting random variable). 3. Consider the limiting cumulative distribution function F, and let u be a continuity point of F. If F is continuous, then every u is a continuity point. If F is not continuous, then u cannot be one of the jump points of F. 4. Consider the sequence of numbers F1(u), F2(u), F3(u), … . Check that this sequence of numbers converges to F(u). 5. If every u that satisfies (3) also satisfies (4), then the sequence { Fn } converges in distribution to F. Even though convergence in distribution is a property of cumulative distribution functions, as noted in (2), it is common to say also that the corresponding sequence of random variables converges in distribution. 6. In this discussion, there is no mention of independence. 7. Here is an example that justifies the concern about continuity points in (3). 1 Suppose that the random variable Zn has the property P Z n = 1. Its n 1 cumulative distribution function is Fn(u) = I u . The sequence of n random variables clearly converges to the constant zero, corresponding to the cumulative distribution function F(u) = I(0 u). The function F has a jump at 0, so that 0 is not a continuity point. For every nonzero u, it 1 gs2011 CENTRAL LIMIT THEOREM ΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩΩ happens that { Fn (u) } F(u). This convergence does not happen at 0, as Fn(0) = 0 for every n. Then { Fn (0) } 0, while F(0) = 1. Ω 8. The previous example may be unsatisfying in that all the random variables are degenerate (non-random constants). Thus, it seems to be a statement about numbers rather than about random variables. Try replacing Zn by a 1 uniform random variable on the interval 0, ; the same demonstration n will work. 9. There are many, many variants on the Central Limit theorem. This result can be proved under much weaker assumptions than those we have used here. 10. The result can be proved, using the assumptions we have stated here, by characteristic functions. If you are willing to assume that all moments are finite, then moment generating functions can provide a proof. 11. For the assumptions used here, most people will say that the approximation to normality is adequate if n 30. That is, a sample size of 30 or more will be enough to invoke the Central Limit theorem. This is, of course, a practical statement which cannot possibly be theoretically rigorous. The Central Limit theorem works very well for nearly all real data problems if n 10. This seems to be a secret. 12. The Central Limit theorem should be used with suspicion if the result probability is extremely large or extremely small. That is, you might not want to trust the Central Limit theorem in cases in which it provides an answer like 0.001 or 0.999. 13. The Central Limit theorem is most often applied to X , the average of a sample X1, X2, …, Xn from a population with mean and standard deviation . It will usually be noted that E X = and SD( X ) = . n These statements are correct, but they are not the Central Limit theorem. The Central Limit theorem is about the limiting distribution of the X standardized random variable n . 14. Sample averages do not always have limiting normal distributions. For example, the average of a sample of Cauchy random variables will not have a limiting normal distribution. 2 gs2011