Checking Conditions Before Applying the Central Limit Theorem

Why Conditions Matter

The Central Limit Theorem (CLT) explains why a sum or average can be approximately normal even when the individual observations are not normal. That conclusion is conditional: the observations must satisfy assumptions that support the normal limit, and a large sample does not repair every violation. Before standardizing a sum, check how the observations are related, whether their variances exist, and whether the sample is large enough relative to the distribution’s skewness.

For exam problems aligned with the SOA Exam P (Probability) syllabus, the goal is not to apply a rule such as “use the CLT when $n\geq 30$.” That cutoff is only a rough classroom heuristic, not a theorem. A defensible decision identifies the model, checks its assumptions, and describes any remaining finite-sample concern.

Theorem: If $X_1,\ldots,X_n$ are independent and identically distributed with mean $\mu$ and finite, positive variance $\sigma^2$, then, as $n$ increases, $$ \frac{\sum_{i=1}^n X_i-n\mu}{\sigma\sqrt{n}} \ \xrightarrow{d}\ N(0,1). $$ Equivalently, the standardized sample average $(\bar X-\mu)/(\sigma/\sqrt n)$ converges in distribution to $N(0,1)$, where $\bar X=(1/n)\sum_{i=1}^n X_i$. The theorem is an asymptotic result: it does not assert that a particular finite-$n$ distribution is exactly normal.

A Practical Conditions Check

The classical theorem above gives a clean starting point for independent, identically distributed (iid) observations. For a particular application, ask the following questions in order.

Are the observations independent?
Independence is part of the classical theorem. Shared shocks, repeated measurements, sampling without replacement from a small population, or other dependence can change the variance of the sum. Do not replace a dependence check with the claim that the sample is large.

Do the observations have the same distribution?
Identical distribution makes the common mean and variance in the classical theorem available. Independent nonidentical terms can also have a normal limit, but require conditions suited to their changing distributions; the iid theorem alone does not justify that conclusion.

Is the variance finite and positive?
The classical iid theorem uses a finite variance $\sigma^2>0$. If the variance is infinite, the usual centering by $n\mu$ and scaling by $\sigma\sqrt n$ are not available, even if a finite mean exists.

Is the sample large enough for this distribution?
The theorem describes what happens as $n$ grows, not a universal minimum sample size. Symmetric, light-tailed distributions often approach normality more quickly than strongly skewed or heavy-tailed distributions.

These checks answer different questions. Independence and finite variance concern whether a theorem applies; sample size and skewness concern how convincing the approximation may be for the actual sample. A large $n$ is useful only after the model has been checked.

What Skewness Says About the Approximation

When the third absolute moment is finite, the population skewness of $X_i$ is $\gamma_1=E[(X_i-\mu)^3]/\sigma^3$. Positive skewness indicates a longer right tail; negative skewness indicates a longer left tail. Skewness is a useful warning about convergence speed, but it is not a complete diagnostic: tail behavior and unusual extreme values also matter.

For iid observations with finite third absolute moment, the skewness of the sample average is $\gamma_1/\sqrt n$. This gives a reason that averaging often reduces asymmetry, while also showing why a very skewed individual distribution may need a larger sample.

\begin{aligned} E\left[\left(\sum_{i=1}^n(X_i-\mu)\right)^3\right] &=nE[(X_1-\mu)^3],\\ \operatorname{Var}\left(\sum_{i=1}^nX_i\right)&=n\sigma^2,\\ \text{skewness of }\bar X &=\frac{nE[(X_1-\mu)^3]}{(n\sigma^2)^{3/2}} =\frac{\gamma_1}{\sqrt n}. \end{aligned}

For the first equality, expand the cube of the sum. Independence and $E[X_i-\mu]=0$ make the mixed terms have expectation zero, leaving the $n$ individual third central moments. The variance adds because the observations are independent. Dividing the third central moment by the variance to the power $3/2$ gives the stated standardized skewness. This calculation is not available if the required moments do not exist.

A symmetric distribution has zero skewness, but zero skewness alone does not prove that a finite-sample distribution is close to normal. Likewise, a finite variance is enough for the classical iid CLT, but it does not guarantee rapid convergence. Treat skewness as one part of the judgment rather than a pass-fail threshold.

Worked Example: A Skewed Distribution With a Finite Variance

A hypothetical insurer models the severity of each claim as an exponential random variable with mean 2 thousand dollars. Suppose 64 claim severities are independent and identically distributed. Assess whether a CLT approximation for their average is plausible.

Method. Check the iid and finite-variance conditions, then use skewness and sample size to discuss finite-sample plausibility. This avoids treating the theorem’s asymptotic conclusion as an exact finite-sample result.

For an exponential variable with mean 2, the variance is $2^2=4$, so the standard deviation is 2 thousand dollars. Its skewness is 2, and all its moments are finite. The stated model supplies both independence and identical distribution.

The standard error of the average of 64 severities is $$ \frac{\sigma}{\sqrt n}=\frac{2}{\sqrt{64}}=\frac{2}{8}=0.25 \quad\text{thousand dollars}. $$ The skewness of that average is $$ \frac{\gamma_1}{\sqrt n}=\frac{2}{8}=0.25. $$ For comparison, with only 4 severities the corresponding skewness would be $2/\sqrt4=1$.

The conditions for the classical iid CLT hold, and 64 observations substantially reduce the average’s skewness compared with one observation. A normal approximation is therefore plausible as an approximation to the average, though it is not exact and the checks do not guarantee a specified accuracy. The positive skewness remains a reason to be cautious about unusually far-tail conclusions.

Check. The standard error is in the same units as the severity, and it is smaller than the individual standard deviation by a factor of $\sqrt{64}=8$. The skewness calculation also decreases from 2 to 0.25, as expected when averaging iid observations.

Dependence Can Defeat the Usual Scaling

For independent variables, variances add. For dependent variables, covariance terms also enter: $$ \operatorname{Var}\left(\sum_{i=1}^nX_i\right) =\sum_{i=1}^n\operatorname{Var}(X_i) +2\sum_{i<j}\operatorname{Cov}(X_i,X_j). $$ Thus, using $\sigma/\sqrt n$ as the standard error of an average is not justified merely because each observation has variance $\sigma^2$. It also requires that the dependence structure support that calculation.

Worked Example: A Shared Shock Creates Dependent Observations

In a hypothetical portfolio, let $Z$ equal 1 or 0 with probability $1/2$ each. Let $\epsilon_1,\ldots,\epsilon_n$ be independent random variables, each equal to $1$ or $-1$ with probability $1/2$, and independent of $Z$. Define each observed amount, in thousands of dollars, by $X_i=10Z+\epsilon_i$. Assess whether the iid CLT applies to the average of 100 observations.

Method. Calculate the marginal moments and pairwise covariance, then compare the correct variance of the average with the variance obtained by incorrectly treating the observations as independent. This exposes the shared-shock dependence.

Since $E[Z]=1/2$ and $E[\epsilon_i]=0$, each observation has mean $E[X_i]=10(1/2)=5$. Also, $\operatorname{Var}(Z)=1/4$ and $\operatorname{Var}(\epsilon_i)=1$. Independence of $Z$ and $\epsilon_i$ gives $$ \operatorname{Var}(X_i)=10^2(1/4)+1=26. $$ For distinct $i,j$, the only shared random component is $10Z$, so $$ \operatorname{Cov}(X_i,X_j)=10^2\operatorname{Var}(Z)=25. $$

For $n=100$, the correct variance of the average is $$ \operatorname{Var}(\bar X) =\frac{100(26)+100(99)(25)}{100^2} =\frac{2{,}600+247{,}500}{10{,}000} =25.01. $$ If the covariance terms were wrongly omitted, the calculation would be $26/100=0.26$, a severe understatement. As $n$ grows, the correct variance is $25+1/n$, which does not approach zero because the shared $Z$ remains in every observation.

The observations are identically distributed and have finite variance, but they are not independent: every pair has covariance 25. Therefore the classical iid CLT and its usual standard error do not apply. The shared shock prevents the average from concentrating around 5 in the way independent averaging would.

Check. Setting $n=100$ in $25+1/n$ gives $25.01$, matching the direct covariance calculation. This confirms that the dependence contribution was retained.

Finite Mean Does Not Replace Finite Variance

A distribution can have a finite mean but infinite variance. In that situation, the classical iid CLT cannot be applied using the usual $\sqrt n$ scaling and a finite standard deviation. Independence and a large sample do not supply a missing variance.

Worked Example: A Pareto Model With Infinite Variance

A hypothetical loss severity $X$, measured in thousands of dollars, has density $f_X(x)=\frac32x^{-5/2}$ for $x\geq1$, and density 0 otherwise. Suppose repeated severities are iid. Decide whether the classical iid CLT applies to their average.

Method. Check the mean and second moment from the density. This avoids assuming that a finite expected loss automatically gives a finite variance.

The density is a Pareto density with scale 1 and shape $3/2$. Direct integration gives $$ E[X]=\int_1^\infty x\left(\frac32x^{-5/2}\right)\,dx =\frac32\int_1^\infty x^{-3/2}\,dx =\frac32(2)=3. $$ The second moment is $$ E[X^2]=\int_1^\infty x^2\left(\frac32x^{-5/2}\right)\,dx =\frac32\int_1^\infty x^{-1/2}\,dx =\infty. $$ Therefore $\operatorname{Var}(X)=\infty$, despite the finite mean $E[X]=3$.

The observations meet the independence and identical-distribution assumptions, but they fail the finite-variance condition. The classical iid CLT does not justify a normal approximation using $\sigma/\sqrt n$, because no finite $\sigma$ exists. This conclusion rules out that theorem’s standard application; it does not assert that every possible normalization or limit theorem must fail.

Check. The first-moment integral converges because its integrand is proportional to $x^{-3/2}$, while the second-moment integral diverges because its integrand is proportional to $x^{-1/2}$. The two moment checks are consistent with a finite mean and infinite variance.

Independent but Nonidentical Terms

The CLT also has versions for independent nonidentically distributed variables, but “independent” by itself is not enough. Let $X_{n,1},\ldots,X_{n,k_n}$ be independent variables in row $n$, with means $\mu_{n,j}$, and let $s_n^2=\sum_j\operatorname{Var}(X_{n,j})$. One standard sufficient condition is the Lindeberg condition: for every $\varepsilon>0$, $$ \frac{1}{s_n^2}\sum_j E\left[(X_{n,j}-\mu_{n,j})^2 \,\mathbf{1}\{|X_{n,j}-\mu_{n,j}|>\varepsilon s_n\}\right]\longrightarrow 0. $$ Here $\mathbf{1}\{A\}$ is 1 when event $A$ occurs and 0 otherwise. Informally, the condition prevents a small number of unusually large terms from carrying a nonnegligible share of the total variance. It is not a condition to replace with a fixed sample-size rule.

Worked Example: One Term Remains Too Influential

For each $n$, consider $n+1$ independent, mean-zero variables. Let $X_{n,1}$ equal $\sqrt n$ or $-\sqrt n$, each with probability $1/2$. Let each of $X_{n,2},\ldots,X_{n,n+1}$ equal 1 or $-1$, each with probability $1/2$. Assess whether the usual normal limit for the standardized sum is supported by the Lindeberg condition.

Method. Compare the variance of the largest term with the total variance and test the Lindeberg condition at one fixed positive $\varepsilon$. This avoids assuming that many independent terms automatically make every term negligible.

The first term has variance $n$, and each of the other $n$ terms has variance 1. Thus $$ s_n^2=n+n=2n, \qquad \frac{\operatorname{Var}(X_{n,1})}{s_n^2}=\frac{n}{2n}=\frac12. $$ For $\varepsilon=1/2$, the Lindeberg threshold is $\varepsilon s_n=\frac12\sqrt{2n}$. Since $\sqrt n>\frac12\sqrt{2n}$, the first term always exceeds the threshold in absolute value. Its contribution to the Lindeberg expression is $$ \frac{1}{2n}E[X_{n,1}^2]=\frac{n}{2n}=\frac12. $$ The remaining terms can only add a nonnegative amount. Hence the expression cannot tend to zero, and the Lindeberg condition fails.

The first term retains half the total variance at every $n$; it does not become negligible under standardization. Independence alone therefore does not support the proposed CLT approximation for this array.

Check. The total variance is $2n$, and the influential term contributes $n$, exactly half. The nonzero lower bound in the Lindeberg expression confirms the same diagnosis.

Common Exam Traps

Using a sample-size slogan as a theorem. “$n\geq30$” does not establish independence, finite variance, or adequate accuracy for a strongly skewed distribution.
Using the independent-sum variance under dependence. The incorrect calculation $26/100=0.26$ in the shared-shock example ignores covariance; the correct variance of the average is $25.01$.
Confusing a finite mean with a finite variance. The Pareto example has mean 3 but infinite variance, so the classical standardization is unavailable.
Assuming nonidentical independent terms are automatically covered by the iid theorem. For such terms, check an appropriate condition or explain why the sum is dominated by no single contribution.
Calling an approximation exact. The CLT gives convergence in distribution as sample size grows, not an exact normal model for each finite sample.

Key takeaway: Before using the CLT, specify the random variables and check independence, distributional assumptions, and finite positive variance. Then judge finite-sample plausibility using sample size relative to skewness and tail behavior. For nonidentical independent terms, also check that no term carries a persistent share of the total variation.

Check Your Understanding

For each situation, decide what the stated facts do and do not justify.

Independent iid observations are uniform on $(0,1)$, and the sample size is 8. State whether the classical iid CLT applies asymptotically and whether the facts guarantee an accurate approximation at this sample size. (Answer: The theorem applies asymptotically because the observations are iid with finite positive variance; $n=8$ does not guarantee a particular finite-sample accuracy, although symmetry and boundedness are favorable features.)
Independent iid severities have a Pareto distribution with shape parameter $1.8$ and scale 1. The mean is finite but the variance is infinite. Can the usual classical CLT standardization be used? (Answer: No. The finite-variance condition fails, so the usual $\sqrt n$ standardization by a finite standard deviation is unavailable.)
In a collection of observations, each $X_i=4Z+\epsilon_i$, where $Z$ is shared, has variance $1/4$, and is independent of mutually independent errors $\epsilon_i$ with variance 1. Find the covariance of two distinct observations and identify the dependence issue. (Answer: $\operatorname{Cov}(X_i,X_j)=4^2(1/4)=4$; the shared $Z$ makes distinct observations dependent, so the iid CLT cannot be invoked.)
For each $n$, independent centered terms have total variance $99$, with one term contributing variance 90 and the other nine terms contributing variance 1 each. What warning does the variance share give? (Answer: The largest term contributes $90/99\approx0.9091$ of the total variance, so the sum is strongly dominated by one term; independence alone does not support a CLT approximation.)
An iid distribution has finite positive variance but no finite third absolute moment. Which conclusion is justified, and what cannot be assessed by the skewness calculation in this tutorial? (Answer: The classical iid CLT still applies asymptotically; the third-moment skewness calculation is unavailable, so it does not provide a finite-sample convergence diagnostic.)

What You'll Learn