Visual walkthrough — Law of large numbers

Before line one: the only thing we assume is that you can add numbers and divide. Everything else — "mean", "variance", "spread", "converge" — we build here.

---

Step 1 — The raw ingredient: repeated random draws

WHAT: we lay the draws out as dots on a number line. WHY: before averaging anything we must literally see the scatter we are trying to tame. PICTURE: each pale-blue dot below is one $X_i$. They land all over the place — that scatter is the enemy the Law will defeat.

---

Step 2 — The one number we are chasing: the expected value $\mu$

WHAT: we mark a single yellow line, the true centre of the scatter. WHY: the whole Law is a statement about arrows pointing to this line — so it needs its own picture first. PICTURE: the yellow vertical line is $\mu$. Notice the blue dots from Step 1 are sprinkled around it, not on it. That gap is what we must close.

---

Step 3 — Measuring the scatter: variance ${\sigma }^2$

WHAT: we draw the spread as a shaded band of typical width $\sigma$ around $\mu$. WHY: $\sigma$ is the "radius of the fog"; shrinking this fog is the Law. PICTURE: the pink band spans roughly one $\sigma$ each side of the yellow $\mu$ line. Most dots live inside it.

---

Step 4 — The estimator: the sample mean ${\bar{X}}_n$

WHAT: we collapse the whole cloud of dots into a single green marker — their average. WHY: this single number, not the individual draws, is the thing the Law is about. PICTURE: watch the green marker sit near $\mu$ but not exactly on it — the small remaining gap is the error ${\bar{X}}_n-\mu$ we now attack.

---

Step 5 — The green marker is centred on the truth

WHAT: the green marker's own long-run home is exactly the yellow line. WHY: this says our estimator is unbiased — it doesn't systematically aim high or low, so any error is pure randomness, not a lean. PICTURE: imagine repeating the whole experiment many times; each run gives a green marker. They pile up symmetrically around $\mu$ — no drift.

---

Step 6 — The fog shrinks: variance of the sample mean is ${\sigma }^2/n$

This is the heart of the whole Law.

WHAT: we draw the green marker's fog for growing $n$ and watch it clamp toward the yellow line. WHY: a shrinking variance is exactly "getting closer to $\mu$" — this picture is the Law made visible. PICTURE: three stacked fogs for $n=1,10,1000$: wide, then narrow, then a razor's edge on $\mu$.

---

Step 7 — Turning "small fog" into a guarantee: Chebyshev

A shrinking fog feels like convergence, but we want a hard number. The bridge is Chebyshev's inequality.

WHAT: we shade the two "miss" tails beyond $\pm ϵ$ and cap their combined probability. WHY: Chebyshev converts the geometric fact "small variance" into the probabilistic fact "small chance of a big miss" — the exact currency the Law is written in. PICTURE: the fog with two vertical fences at $\mu \pm ϵ$; the pink tail area outside the fences is what the fraction ${\sigma }^2/(n{ϵ}^2)$ bounds from above.

---

Step 8 — Let $n$ grow: the bound (and the fog) vanish

WHAT: we plot the ceiling ${\sigma }^2/(n{ϵ}^2)$ as a curve that dives to zero. WHY: this is the finish line — "the chance of a noticeable miss vanishes" is the Law of Large Numbers itself. PICTURE: the pale-yellow $1/n$ curve sinking to the axis; underneath it the true miss-probability (blue) is trapped and dragged to zero.

---

Step 9 — The edge cases: where the Law bends or breaks

PICTURE: four little panels — finite-variance fog shrinking (good), Cauchy fog stuck wide, correlated fog barely moving, and the degenerate spike sitting on $\mu$.

---

The one-picture summary

Everything above, compressed: draws scatter (Step 1) around a true centre $\mu$ (Step 2) with spread $\sigma$ (Step 3); their average ${\bar{X}}_n$ (Step 4) is centred on $\mu$ (Step 5) with spread $\sigma /\sqrt{n}$ (Step 6); Chebyshev fences that spread (Step 7) and $n\to \infty$ collapses it (Step 8) — unless the finiteness/independence conditions fail (Step 9).

Recall Feynman retelling — say it in plain words

Imagine a paint sprayer that mostly hits a target dot but sprays a fuzzy cloud around it. One spray is messy. But if you take a hundred sprays and mark their average position, that average lands much closer to the dot — because the strays cancel out. Take a million sprays and the average is practically nailed onto the dot. That "average gets nailed as you take more sprays" is the Law of Large Numbers. The size of the cloud is the variance; averaging $n$ sprays shrinks the cloud by a factor $n$ (so the width by $\sqrt{n}$). Chebyshev is just the referee who says "a cloud this thin cannot have much paint far from the dot," and letting the number of sprays run to infinity leaves the cloud with zero width — as long as the sprayer never goes berserk (finite variance) and each spray doesn't copy the last one (independence).

Recall Quick self-test

Why does variance carry a square when we pull out the $1/n$? ::: Because a constant $c$ scales variance by $c^2$; here $c=1/n$, giving $1/n^2$. What single condition on the draws makes the variance of a sum split into a sum of variances? ::: Independence (no cross-covariance terms). Four times more data improves the typical error by what factor? ::: A factor of 2, because error scales like $\sigma /\sqrt{n}$. What breaks the Law even with independent draws? ::: Infinite variance — the fog ${\sigma }^2/n$ never shrinks.

---

Connections: this visual derivation underlies Monte Carlo Methods, the mini-batch averaging inside Stochastic Gradient Descent, the resampling in Bootstrap Sampling, and the width of Confidence Intervals. The rate of the fog's collapse is sharpened by the Central Limit Theorem, and the averaging-shrinks-variance idea reappears in the Bias-Variance Tradeoff.