When the Central Limit Theorem Wobbles

State the classical central limit theorem as most learners first meet it, then surface the one word it leans on: independent. Pose the driving question visually as a clean bell curve on one side and a fuzzy, lumpy histogram on the other.

• Sum of many small wiggles → bell curve
• Independence is the engine, not an afterthought
• Driving question: what happens when the wiggles talk to each other?

Ask the learner to commit to one prediction before any simulation is shown, anchoring the comparison that follows.

• Lock in an initial hypothesis
• The answer choices encode the three canonical failure modes: width, shape, and skew

A simulation that lets the learner draw thousands of sums of N wiggles under three regimes — independent, short-range correlated, and long-range correlated — and overlays each empirical histogram with a fitted Gaussian of matched variance. A slider for N and a slider for correlation length make the breakdown visible.

• Independent: histogram and matched Gaussian overlap tightly
• Short-range correlation: same shape, wider spread than the i.i.d. case
• Long-range correlation: visible deviation in tails and sometimes skew
• Increasing N does not, by itself, restore the bell

Summarize the evidence in one composite figure: variance growth vs. N on a log-log plot, with three lines fanning apart. The independent case falls fastest; the correlated cases flatten, signaling a slower — or zero — rate of concentration.

• Variance of the sample mean shrinks like 1/N only under independence
• Dependence adds a 'stickiness' timescale that slows the collapse
• On a log-log plot, the slopes are the signature

Show the calculation behind the scenes: variance of a sum is the sum of all pairwise covariances. Independence zeroes them out; positive correlation keeps them alive. Rewrite this as Var(mean) = (σ²/N)·(1 + 2·Σρₖ) — a single number, the sum of autocovariances, controls everything.

• Var of a sum = sum of all covariances, not just the diagonal
• Define the correlation time τ as 1 + 2·Σρₖ
• Effective sample size becomes N/τ, not N
• When τ diverges, the CLT silently fails

Hand the learner a new scenario: daily temperature anomalies with known AR(1) structure. They must predict, then test, whether the annual average of 365 days looks Gaussian, and estimate the effective sample size. A second scenario contrasts it with internet traffic bursts (long-range dependence) where the answer is no.

• Same math, new domain: real time series
• Estimate τ from data, then read off the predicted width
• Long-memory series resist Gaussian limits even at large N

Mark the boundary explicitly: weak dependence with a finite correlation time preserves a Gaussian limit (a generalized CLT), while strong or long-range dependence breaks the theorem's conclusions — and no amount of averaging repairs a divergent τ.

• Weak dependence + finite τ: bell curve survives, just wider
• Divergent τ: no Gaussian limit, tails dominate
• Independence is a sufficient condition, not a necessary one

Close the loop with a single, direct sentence answering the driving question, supported by the three pieces of evidence collected along the way. Restate the role of the correlation-time sum ρ = 1 + 2·Σρₖ as the single dial that turns the bell into something else.

• Direct answer to the driving question
• Independence is what makes 1/N the right scaling
• Dependence replaces N with N/ρ; if ρ diverges, the limit is not Gaussian
• Take-home: averaging is not a universal detergent