Visual walkthrough — Confidence intervals

---

Step 1 — One sample gives one dot

WHAT. We have a big pile of things (people, test images, whatever). Each thing has a number. The whole pile has a true average we cannot see — call it $\mu$ (the Greek letter "mu", just a name for "the real average of everything"). We grab $n$ of them at random and average those:

$$\bar{x}=\underset{\text{divide by how many}}{\underbrace{\frac{1}{n}}}\underset{\text{add them all up}}{\underbrace{\sum _{i=1}^n{x}_i}}$$

• $x_i$ — the $i$-th measured number (the $i$ runs from $1$ up to $n$).
• $n$ — how many things we grabbed (the sample size).
• $\bar{x}$ — read "x-bar", the average of this one handful.

WHY. $\bar{x}$ is our single best guess for the invisible $\mu$. But it is only one guess from one handful.

PICTURE. Look at the figure: the faint cloud is the whole population. The four coloured dots are four different handfuls we could have drawn — each gives a different $\bar{x}$. The true $\mu$ (yellow line) sits fixed while our guesses scatter around it.

---

Step 2 — Repeat the grab many times: a distribution of averages

WHAT. Imagine grabbing a handful, writing down $\bar{x}$, and repeating thousands of times. Those $\bar{x}$ values form their own bell-shaped histogram, called the sampling distribution of the mean.

WHY. We cannot control one $\bar{x}$, but we can describe how the whole family of $\bar{x}$'s behaves. Predicting the family lets us reason about any single member.

The Central Limit Theorem promises that for a decent-sized $n$ this histogram is a normal (bell) curve:

$$\bar{X}\sim N\left(\underset{\text{centred on truth}}{\underbrace{\mu }},\underset{\text{spread shrinks with }n}{\underbrace{\frac{{\sigma }^2}{n}}}\right)$$

• $N(\text{centre},\text{spread})$ — the bell curve with that centre and that spread.
• ${\sigma }^2$ — the variance (spread${}^2$) of one raw measurement.
• ${\sigma }^2/n$ — the variance of the average; bigger $n$ ⇒ narrower bell.

PICTURE. The bell is centred exactly on $\mu$ (not shifted!) and its width is the Standard Error $\sigma /\sqrt{n}$. Watch how doubling $n$ pinches the bell narrower.

---

Step 3 — Slide and stretch onto a universal ruler (standardize)

WHAT. Every normal bell has a different centre $\mu$ and width $SE$. We convert to one master bell — centred at $0$, width $1$ — by subtracting the centre and dividing by the width:

$$Z=\frac{\stackrel{\text{how far from truth}}{\overbrace{\bar{X}-\mu }}}{\underset{\text{measured in SE units}}{\underbrace{\sigma /\sqrt{n}}}}\sim N(0,1)$$

• $Z$ — the same quantity $\bar{X}$, but relabelled in "number of standard errors away from $\mu$."
• The bottom $\sigma /\sqrt{n}$ is the $SE$ from Step 2.

WHY. On the master bell we know exactly which points cut off any tail area — those numbers are fixed constants, not something we recompute every problem. Standardizing turns "every problem is different" into "one ruler fits all."

PICTURE. Left: three different real bells. Right: they all collapse onto one standard bell after the shift-and-squeeze. The $x$-axis is now labelled in $SE$ units.

---

Step 4 — Carve out the middle $(1-\alpha )$ of the master bell

WHAT. Pick a confidence level, say $95\%$. Call the leftover $\alpha =0.05$ (the $5\%$ we are willing to miss). Split it into two equal tails of $\alpha /2=0.025$ each. The point on the master bell that leaves $2.5\%$ in the right tail is called $z_{\alpha /2}$.

$$P\left(-z_{\alpha /2}\le Z\le z_{\alpha /2}\right)=1-\alpha$$

• $\alpha$ — the total tail probability we sacrifice (miss rate).
• $\alpha /2$ — half of it, one for each side.
• $z_{\alpha /2}$ — the ruler mark cutting off that tail; for $95\%$ it equals $1.96$.

WHY split evenly? We have no reason to fear over-guessing more than under-guessing, so we treat both tails symmetrically. This is the same two-tailed logic used in Hypothesis Testing.

PICTURE. The shaded green middle holds $95\%$ of all $Z$ values; the two red tails hold $2.5\%$ each. The vertical marks at $\pm 1.96$ are $\pm z_{\alpha /2}$.

---

Step 5 — Un-standardize: flip the statement inside-out

WHAT. We now undo Step 3 algebraically to free $\mu$. Start from the inequality inside the probability and multiply every part by $SE=\sigma /\sqrt{n}$:

$$-z_{\alpha /2}\frac{\sigma }{\sqrt{n}}\le \bar{X}-\mu \le z_{\alpha /2}\frac{\sigma }{\sqrt{n}}$$

Subtract $\bar{X}$, then multiply by $-1$ (which flips both $\le$ signs):

$$\boxed{\bar{X}-z_{\alpha /2}\frac{\sigma }{\sqrt{n}}\le \mu \le \bar{X}+z_{\alpha /2}\frac{\sigma }{\sqrt{n}}}$$

• Left/right edges — random, because they carry $\bar{X}$.
• $\mu$ in the middle — fixed, just trapped between two wobbling walls.

WHY. Rearranging puts $\mu$ (what we care about) alone in the middle, so the picture becomes "a moving bracket that either catches or misses a nailed-down target."

PICTURE. The yellow $\mu$ is a fixed vertical line. Blue brackets are ten repeated experiments; each bracket is $\bar{x}\pm z_{\alpha /2}SE$ wide. About nine of ten brackets straddle the yellow line — the tenth (red) misses. The brackets move; the line does not.

---

Step 6 — Edge case: we never actually know $\sigma$

WHAT. In real life the population spread $\sigma$ is unknown. We estimate it from the same sample:

$$s=\sqrt{\frac{1}{n-1}\sum _{i=1}^n(x_i-\bar{x}{)}^2}$$

• $n-1$ (not $n$) — Bessel's correction; one degree of freedom was spent computing $\bar{x}$, so we divide by one fewer.

Swapping $s$ for $\sigma$ injects extra randomness, so the master bell is no longer good enough. We use a slightly fatter cousin, the t-distribution with $df=n-1$:

$$T=\frac{\bar{X}-\mu }{s/\sqrt{n}}\sim t_{n-1}$$

WHY a fatter curve? Because $s$ can under-guess the true spread, extreme $T$ values happen more often. Fatter tails ⇒ wider critical marks ⇒ honestly wider intervals. As $n\to \infty$, the t-curve melts back into the normal.

PICTURE. Overlay: the normal (blue) and $t$ with $df=3$ (red). Same peak, but red has thicker tails and its $2.5\%$ mark sits further out — that extra distance is the price of not knowing $\sigma$.

---

Step 7 — Degenerate limits: what breaks the picture

WHAT & WHY. Push the knobs to extremes and watch the interval respond — this is how you sanity-check any use.

• $n=1$: $SE=\sigma /\sqrt{1}=\sigma$ and $s$ is undefined (dividing by $n-1=0$). One data point tells you nothing about spread → no interval.
• $n\to \infty$: $SE=\sigma /\sqrt{n}\to 0$, the bracket collapses onto $\bar{x}=\mu$. Infinite data ⇒ zero uncertainty.
• Confidence $\to 100\%$ ($\alpha \to 0$): $z_{\alpha /2}\to \infty$, bracket widens to $(-\infty ,\infty )$. "Certain" only if you claim everything.
• Skewed data / heavy outliers: the CLT bell in Step 2 is a lie; the normal shortcut fails and you should reach for Bootstrap Methods or Bayesian Credible Intervals.

PICTURE. Margin of error $E$ plotted against $n$ — a $1/\sqrt{n}$ curve. Steep drop early, then flat: the "diminishing returns" of more data. The dashed asymptote sits at $E=0$.

---

The one-picture summary

Everything at once: raw scatter → bell of averages → standard ruler → carved middle → un-flipped bracket around $\mu$, with $t$ widening it when $\sigma$ is guessed.

Recall Feynman retelling — say it to a 12-year-old

I want the real average $\mu$ of a giant pile, but I only get one small handful, so I take that handful's average $\bar{x}$. If I imagine grabbing handfuls over and over, their averages pile into a bell centred on the true $\mu$, and the bell's width is $\sigma /\sqrt{n}$ — more data, skinnier bell. I slide-and-squeeze that bell onto a universal one centred at zero, where I know the two marks (at $\pm 1.96$) that fence in the middle $95\%$. I stretch those marks back to real units and rearrange until $\mu$ sits alone in the middle, trapped by a bracket $\bar{x}\pm 1.96\sigma /\sqrt{n}$. The bracket wobbles from sample to sample; $\mu$ stays put; about $95$ of every $100$ brackets catch it. Since I never truly know $\sigma$, I estimate it as $s$ (dividing by $n-1$), and to pay for that extra guessing I use the slightly fatter $t$-marks — which quietly become the normal marks again once I have lots of data.

Recall Quick self-test

Why divide $SE$ by $\sqrt{n}$ and not $n$? ::: Averaging cuts variance by $n$; the spread ($SE$) is the square root of variance, so it drops by $\sqrt{n}$. After computing $[167,173]$, what is "random"? ::: The two endpoints (they carry $\bar{x}$); the true $\mu$ is fixed. When must you switch $z\to t$? ::: When $\sigma$ is unknown and $n$ is small ($<30$); $t$'s fat tails pay for estimating $\sigma$ with $s$. What happens to the interval as $\alpha \to 0$? ::: $z_{\alpha /2}\to \infty$, so the interval blows up to $(-\infty ,\infty )$.

Related tools that reuse this machinery: A/B Testing, Cross-validation, Hypothesis Testing.