Why Sample Means Look Normal

Pose the driving question with a concrete image: a uniform probability mass function for a single die roll, contrasted with a smooth bell curve that the sample mean's distribution seems to approach.

• A single die roll is perfectly uniform — no shape, no peak.
• Yet the average of many rolls looks bell-shaped.
• How can the mean inherit a shape the source does not have?

Ask the learner to commit to one prediction before seeing the simulation, forcing an explicit hypothesis about the sampling distribution of the mean from a uniform source.

• Make one explicit prediction about the shape of sample means
• Force commitment before evidence is shown

Simulation: draw N independent samples of size n from a uniform source, plot the histogram of their means, and let the learner change n and the number of replications to see the empirical distribution emerge.

• Adjust sample size n from 1 to 50
• Adjust the number of replications from 100 to 10,000
• Watch the histogram morph from flat to bell-shaped
• Observe the spread shrink as n grows

Repeat the simulation with a strongly skewed (exponential) source to show that the same convergence happens — the source shape is not the determining factor.

• Switch the source to a skewed exponential
• Increase n and watch the bell emerge anyway
• Confirm: the result is not specific to the uniform

Explain that the distribution of a sum of independent variables is the convolution of their distributions, and convolution is a smoothing operation that erases sharp features. The width of the resulting distribution scales as 1/√n, which is the standard error.

• Sum of independent variables = convolution of their distributions
• Convolution smooths sharp peaks, edges, and corners
• Repeated smoothing leaves a Gaussian profile
• Spread of the mean = σ / √n, the standard error

Visualization: animate the convolution of two uniform distributions, then chain several together, showing how the shape progressively rounds toward a bell. Optionally overlay the matching Gaussian with width σ/√n.

• Observe one convolution step: flat + flat → triangular
• Chain convolutions and watch edges round off
• Compare each result to a Gaussian overlay

Mark the limits: extremely heavy-tailed sources (e.g., Cauchy), or tiny n on a wildly skewed source, can produce sampling distributions that are slow to look Gaussian. Independence and finite variance are the assumptions that matter.

• Sources with infinite variance defeat the standard CLT
• Very small n on highly skewed sources may not yet look normal
• Independence is required; correlated draws break the argument

Transfer test: let the learner upload or type in a small dataset (e.g., exam scores, reaction times, counts) and draw bootstrap sample means to see whether their empirical sampling distribution matches the predicted bell of width s/√n.

• Plug in a new dataset
• Draw many bootstrap sample means
• Compare the empirical histogram to the predicted Gaussian

Close by directly answering the driving question: the sample mean is a sum of independent values; summing smooths the source distribution, and the resulting shape approaches a Gaussian centered at the true mean with spread σ/√n — the Central Limit Theorem.

• Sample means become normal because averaging smooths via convolution
• The Gaussian width is the standard error σ/√n
• The source distribution's shape does not matter — only independence and finite variance
• This is why n ≈ 30 is a common rule of thumb