Law of Large Numbers: Weak and Strong Forms for Actuaries
Learn both forms of the Law of Large Numbers and their actuarial significance.
Weak Law of Large Numbers
The Weak Law of Large Numbers (WLLN) states that for iid random variables X1, X2, ..., Xn with mean mu and finite variance sigma^2, the sample mean X_bar converges to mu in probability: for every epsilon > 0, P(|X_bar - mu| > epsilon) approaches 0 as n approaches infinity.
The proof using Chebyshev's inequality is elegant and concise: P(|X_bar - mu| > epsilon) <= Var(X_bar) / epsilon^2 = sigma^2 / (n * epsilon^2). As n grows, the right side approaches 0, proving the result. This proof requires only finite variance, and the Chebyshev bound gives a concrete (though often loose) rate of convergence.
In actuarial terms, the WLLN guarantees that the average claim cost per policy converges to the true expected claim cost as the portfolio grows. If an insurer writes n policies with iid losses having mean mu, then the average loss per policy X_bar will be close to mu with high probability for large n. This is the mathematical foundation of insurance pricing.
Strong Law of Large Numbers
The Strong Law of Large Numbers (SLLN) provides a stronger guarantee: X_bar converges to mu almost surely, meaning P(lim X_bar = mu) = 1. The SLLN in its strongest form (the Kolmogorov version) requires only that E[|X|] < infinity, without requiring finite variance.
The distinction between convergence in probability (weak law) and almost sure convergence (strong law) is subtle. Convergence in probability means that for any fixed epsilon, the probability of X_bar being far from mu goes to zero. Almost sure convergence means that with probability 1, the sequence X_bar eventually stays close to mu forever. Almost sure convergence implies convergence in probability, but not vice versa.
While Exam P primarily tests the WLLN and its applications, understanding the SLLN provides deeper insight. The SLLN tells us that for a single insurance portfolio that grows over time, the running average of losses will converge to the true mean with certainty (probability 1). The WLLN only guarantees this convergence in a weaker probabilistic sense.
Chebyshev's Inequality and the LLN
Chebyshev's inequality is the key tool for proving the WLLN and for computing explicit bounds on the accuracy of the sample mean. The inequality states P(|X - mu| >= k*sigma) <= 1/k^2 for any random variable X with mean mu and standard deviation sigma.
Applied to the sample mean X_bar with Var(X_bar) = sigma^2/n, the bound becomes P(|X_bar - mu| >= epsilon) <= sigma^2/(n * epsilon^2). This tells you exactly how large n must be to achieve a desired accuracy with a desired confidence level.
Example: Claim amounts have mean 500 and variance 10000. How many claims are needed so that P(|X_bar - 500| < 10) >= 0.95? By Chebyshev: P(|X_bar - 500| >= 10) <= 10000/(n * 100) = 100/n. For this to be at most 0.05, we need n >= 2000. Note that Chebyshev gives a conservative (large) estimate of n; the actual required sample size using the CLT would be smaller.
Actuarial Significance of the LLN
The LLN justifies the insurance business model. By pooling a large number of independent, identically distributed risks, an insurer can predict its average loss per policy with high accuracy. The standard deviation of the average loss X_bar is sigma/sqrt(n), which decreases as the portfolio grows. With n = 10000 policies, the standard deviation of the average is sigma/100, making the average very predictable.
However, the total risk (absolute variability of total losses) still increases with n. The standard deviation of total losses S_n is sigma*sqrt(n), which grows proportionally to sqrt(n). This means that while the average loss per policy becomes more predictable, the total dollar amount of losses becomes more variable as the portfolio grows.
This distinction between risk per policy and total risk is conceptually important for actuaries. Insurance pricing relies on the per-policy risk decreasing (so premiums can be set accurately), while capital requirements and reinsurance decisions must account for the total risk increasing.
The LLN and the CLT Compared
The LLN and the CLT are complementary results about the sample mean. The LLN tells you where X_bar converges (to mu), while the CLT tells you the shape of the distribution of X_bar around mu for large n (approximately normal with variance sigma^2/n). Together, they provide a complete picture of the behavior of averages.
On Exam P, the LLN appears primarily in conceptual questions: "What happens to the sample mean as n increases?" or "Why can insurers price policies accurately for large portfolios?" The CLT appears in computational problems: "Find the probability that the total claims exceed a threshold." Know when each theorem is being tested.
A useful way to remember the relationship: the LLN says X_bar converges to a point (mu), while the CLT says it fluctuates around mu according to a normal distribution with a shrinking spread. The CLT gives a more precise description of the fluctuations that the LLN says are vanishing.
Extensions and Limitations
The LLN extends beyond iid variables. Versions exist for independent but not identically distributed variables, and even for certain types of dependent variables (under mixing conditions). These extensions are not tested on Exam P but are relevant in actuarial practice where risks may not be perfectly independent.
The LLN does not apply when the mean does not exist. For distributions like the Cauchy, which has no mean, the sample average does not converge to any value. Instead, the sample average has the same Cauchy distribution regardless of n. This is a striking failure of the LLN that highlights the importance of the finite mean assumption.
For heavy-tailed distributions with finite mean but infinite variance (such as Pareto with 1 < alpha <= 2), the SLLN still applies (the average converges to the mean), but the CLT does not (the fluctuations around the mean are not normal). The rate of convergence under the LLN is also slower for heavy-tailed distributions.
Common Pitfalls
A frequent misunderstanding is the "gambler's fallacy," which is related to a misapplication of the LLN. The LLN says that the average converges to the mean, not that future outcomes compensate for past deviations. If a fair coin has produced 10 heads in a row, the probability of tails on the next flip is still 0.5. The average converges to 0.5 not by compensating outcomes but by the past 10 flips becoming an increasingly small fraction of the total.
Another common error is confusing the LLN with the CLT. The LLN is about convergence of the average to a point; the CLT is about the distribution of the average around that point. A problem asking "what value does X_bar approach?" is an LLN question. A problem asking "what is P(X_bar > c) for large n?" is a CLT question.
Study Tips
For Exam P, understand the conceptual meaning of both the weak and strong laws, but focus your problem-solving practice on the WLLN and Chebyshev's inequality. Know how to use the Chebyshev bound to determine sample sizes. Understand the actuarial interpretation: per-policy risk decreases as 1/sqrt(n), while total risk increases as sqrt(n). Be able to articulate why the LLN justifies the insurance business model. These concepts typically appear in one to two problems per exam, often as conceptual questions rather than heavy calculations.