Visual walkthrough — Central Limit Theorem — statement, proof sketch, significance

---

Step 0 — What is a "random variable", in pictures?

Before anything else: a random variable is just a number you don't know yet, produced by some experiment — the face of a die, the height of a stranger, the time until a bus arrives. We write it $X$. Its distribution is the picture of "how likely is each value" — a bar chart (discrete) or a smooth hill (continuous).

Two numbers summarise that picture:

Look at the three shapes above — a die (flat bars), an exponential (a right-leaning slide), a coin (two spikes). They look nothing alike. The whole miracle ahead is that averaging any of them lands on the same curve. Keep all three in your head.

---

Step 1 — The thing we actually study is the average, not the data

We do NOT claim the data becomes a bell. We take $n$ independent copies $X_1,X_2,\dots ,X_n$ (independent = one outcome tells you nothing about another) and form their sample mean

${\bar{X}}_n=\underset{\text{share equally}}{\underbrace{\frac{1}{n}}}\underset{\text{add all }n\text{ draws}}{\underbrace{\sum _{i=1}^n{X}_i}}.$

• ${\sum }_{i=1}^n{X}_i$ means "add up all $n$ draws" — the big $\Sigma$ is Greek capital S, for Sum.
• Multiplying by $\frac{1}{n}$ turns that total into an average.

WHY the average? One draw is hopeless to predict. But an average lets the too-high and too-low draws cancel. The picture below builds one such average from scattered dots.

WHAT IT LOOKS LIKE: scattered blue dots (individual draws, wildly spread) collapse to a single yellow tick (their average) that sits very near the true mean $\mu$. Do this experiment again and the yellow tick lands slightly elsewhere — that jitter of the yellow tick is the star of this whole page.

---

Step 2 — Repeat the whole experiment many times → a new distribution appears

Here is the mental move that trips everyone up. Freeze it:

WHY: each average is itself a random number (jitters every time). A random number has its own distribution — and that is the object the CLT describes.

WHAT IT LOOKS LIKE: many yellow ticks from Step 1, stacked into a histogram. Even though the source (green exponential, skewed) is lopsided, the pile of averages is already leaning toward symmetric and bell-shaped.

---

Step 3 — Where does that pile sit, and how wide is it?

Before proving the shape is a bell, we pin down its center and width — because a bell needs both.

$E[{\bar{X}}_n]=\mu ,\mathrm{Var}({\bar{X}}_n)=\frac{{\sigma }^2}{n},\mathrm{SD}({\bar{X}}_n)=\frac{\sigma }{\sqrt{n}}.$

• $E[{\bar{X}}_n]=\mu$: the average of averages is still the true mean — the pile is centered on the truth.
• $\mathrm{Var}({\bar{X}}_n)={\sigma }^2/n$: dividing by $n$ shrinks the spread. WHY $/n$ and not $/\sqrt{n}$ for the variance? Because variances of independent things add, so $\mathrm{Var}(\sum X_i)=n{\sigma }^2$, and multiplying a variable by $\frac{1}{n}$ multiplies its variance by $\frac{1}{n^2}$: $\frac{1}{n^2}\cdot n{\sigma }^2=\frac{{\sigma }^2}{n}$.
• $\mathrm{SD}({\bar{X}}_n)=\sigma /\sqrt{n}$: take the square root of the variance. This width has its own name, the standard error.

WHAT IT LOOKS LIKE: three piles for $n=1,10,100$. All centered on $\mu$; each is $\sqrt{10}\approx 3.16\times$ narrower than the last. The narrowing-but-never-moving is the visual signature of $\sigma /\sqrt{n}$.

---

Step 4 — Rescale so the width is frozen at 1 (standardize)

Problem: every pile in Step 4's picture has a different width, so we can't compare shapes. Fix it by standardizing — subtract the center, divide by the width:

$Z_n=\frac{\stackrel{\text{distance from truth}}{\overbrace{{\bar{X}}_n-\mu }}}{\underset{\text{one standard error}}{\underbrace{\sigma /\sqrt{n}}}}.$

• Numerator ${\bar{X}}_n-\mu$: slides the pile so its center sits at $0$.
• Denominator $\sigma /\sqrt{n}$: rescales so the width is exactly $1$.

So $Z_n$ always has mean $0$ and standard deviation $1$, no matter what $n$ or what the original $X$ was. Now every pile is drawn on the same ruler and we can ask: does the shape settle down?

WHY: to compare shapes you must remove size differences — like resizing photos to the same width before comparing what's in them.

WHAT IT LOOKS LIKE: the three different-width piles from Step 4, all stretched/shifted to sit on top of each other at center 0, width 1. Now they nearly coincide — and the curve they hug is the target of Step 5.

---

Step 5 — The "fingerprint" trick: characteristic functions

To prove the shape becomes a specific curve, comparing messy histograms directly is hopeless. We switch to a tool that turns "adding independent randoms" into plain multiplication.

WHY this tool and not densities? Two independent things summed have a convolution of densities — a messy smearing integral. But their fingerprints simply multiply: ${\phi }_{X+Y}(t)={\phi }_X(t){\phi }_Y(t)\text{(when independent).}$ Multiplication is far easier to take a limit of than repeated smearing.

WHAT IT LOOKS LIKE: the yellow dot $e^{itX}$ riding the unit circle; the red vector is the average of many such dots — that shrunken averaged vector is ${\phi }_X(t)$. As $t\to 0$ the dots barely spin, so their average vector has length $\approx 1$.

---

Step 6 — Zoom in near $t=0$: only mean and variance survive

For our standardized building block $Y=\frac{X-\mu }{\sigma }$ (mean $0$, variance $1$), the fingerprint near $t=0$ is well approximated by its first curvature — a Taylor expansion:

${\phi }_Y(s)=\underset{\text{starts at }1}{\underbrace{1}}+\underset{=0\text{ since }E[Y]=0}{\underbrace{isE[Y]}}-\underset{=-s^2/2\text{ since }E[Y^2]=1}{\underbrace{\frac{s^2}{2}E[Y^2]}}+o(s^2).$

• ${\phi }_Y(0)=1$: averaging non-spinning dots gives length 1.
• The linear term vanishes because we centered ($E[Y]=0$).
• The quadratic term $-\frac{s^2}{2}$ is the only surviving information: it is the variance, which is $1$.
• $o(s^2)$ means "junk that shrinks faster than $s^2$" — negligible near 0.

WHY Taylor here? Because in the next step every input $s=t/\sqrt{n}$ is tiny, so only the behaviour of $\phi$ right next to $t=0$ matters. Taylor is precisely the microscope for "near a point".

WHAT IT LOOKS LIKE: the true fingerprint curve and the parabola $1-\frac{s^2}{2}$ kissing at $s=0$ — indistinguishable in the shaded zoom window. That kiss is why only the variance matters.

---

Step 7 — Sum, factor, and take the limit → the bell's fingerprint

Now assemble. Standardizing the sum gives $Z_n=\frac{1}{\sqrt{n}}{\sum }_{i=1}^n{Y}_i$. Apply the multiply-fingerprints rule:

${\phi }_{Z_n}(t)=\left[{\phi }_Y\right(\frac{t}{\sqrt{n}}){]}^n={\left(1-\frac{t^2}{2n}+o\left(\frac{1}{n}\right)\right)}^n\underset{n\to \infty }{\to }{e}^{-t^2/2}.$

• We plugged $s=\frac{t}{\sqrt{n}}$ into Step 6, so $\frac{s^2}{2}=\frac{t^2}{2n}$.
• The power $n$ comes from $n$ identical independent factors multiplied.
• The limit uses ${\left(1+\frac{a}{n}\right)}^n\to e^a$ with $a=-\frac{t^2}{2}$ — the same limit that defines compound interest.

And $e^{-t^2/2}$ is exactly the fingerprint of the standard normal $N(0,1)$. Matching fingerprints ⟹ matching distributions (Lévy's continuity theorem). Done. $■$

WHAT IT LOOKS LIKE: the curves ${\left(1-\frac{t^2}{2n}\right)}^n$ for $n=1,4,20,\infty$ climbing onto the smooth red target $e^{-t^2/2}$. As $n$ grows they lock onto the bell's fingerprint.

---

Step 8 — The edge case that BREAKS the machine

Every step above needed finite variance ${\sigma }^2<\infty$. What if a distribution is so heavy-tailed that ${\sigma }^2=\infty$?

WHAT IT LOOKS LIKE: side by side, the exponential's averages tighten into a bell (green) while the Cauchy's averages stay wild and spread out (red) even at large $n$. The single occasional monster draw drags the whole average.

---

The one-picture summary

Left to right: any starting shape → take many size-$n$ averages → center and rescale by $\sigma /\sqrt{n}$ → the piled, standardized averages converge onto the one universal bell $N(0,1)$, whose fingerprint is $e^{-t^2/2}$ — unless the variance is infinite, in which case (Cauchy) the machine jams.

Recall Feynman retelling of the whole walkthrough

Pick any messy random thing — a die, a lopsided waiting time, a coin flip. One draw is unpredictable (Step 0–1). So don't predict one draw; average a bunch (Step 1). Do that averaging over and over and pile up the answers — that pile is a brand-new picture (Step 2). The pile always sits on the true mean, and it gets narrower like one-over-root-n (Step 3). Shrink and shift every pile to the same size so we can compare shapes (Step 4). To prove the shape, switch to the "fingerprint" that turns adding randoms into multiplying numbers (Step 5), zoom in so only the spread matters (Step 6), multiply the $n$ identical fingerprints and take the limit — out pops $e^{-t^2/2}$, the bell's fingerprint (Step 7). It works for anything with a finite width, and dramatically fails for the fat-tailed Cauchy (Step 8). That's the whole magic: mess in, bell out.

Recall Quick self-checks

Why divide the deviation by $\sqrt{n}$ and not $n$? ::: Dividing by $n$ over-squashes to $0$ (Law of Large Numbers); by $\sqrt{n}$ freezes the width at the standard error so a non-degenerate bell survives. Does the raw data become normal? ::: No — only the sampling distribution of the mean/sum does; raw data keeps its shape. Why do characteristic functions make the proof easy? ::: Independent sums multiply their fingerprints, and pointwise CF convergence implies convergence in distribution. Which two facts about $X$ does the limit depend on? ::: Only its finite mean $\mu$ and finite variance ${\sigma }^2$. Why does CLT fail for Cauchy? ::: Its variance is infinite, so there is no finite width to rescale by and the Taylor $-t^2/2$ term never appears.

---

See also

Confidence Intervals · Hypothesis Testing — z and t tests · Binomial Distribution · Standard Error · Law of Large Numbers