The Normal Distribution
If a statistician were allowed to keep just one curve, this would be it. The
normal distribution — the famous bell curve — is a smooth,
symmetric hump that turns up almost everywhere you look: the heights of people in a crowd, the
scores on a big exam, the tiny errors a machine makes when it cuts a plank to length, the spread of
blood-pressure readings in a hospital. Line any of these up as a histogram and the same gentle bell
keeps appearing.
What makes it magical is how little it takes to pin one down. A normal distribution is
specified completely by just two numbers: its mean
\mu (where the peak sits) and its
standard deviation
\sigma (how wide the bell is). Tell me those two numbers and I can draw
the entire curve — every height, every tail, forever. We write
X \sim N(\mu, \sigma^2)
to say X is normal with mean \mu and variance
\sigma^2 (so standard deviation \sigma).
Because nearly every classical statistical test leans on this curve, it is fair to call it the most
important distribution in all of statistics.
The shape, feature by feature
Four things define the bell, and each one is worth naming:
-
Symmetric. The left half is a mirror of the right half, folded along the centre
line at \mu. Values equally far above and below the mean are equally
common.
-
Single-peaked. There is exactly one hump, right at
\mu. Because of the symmetry, the mean, median, and
mode all coincide — they sit together at that peak. (In a skewed distribution they pull
apart; in a normal one they lock together.)
-
Tails that never quite touch zero. The curve swoops down toward the axis on both
sides but never actually reaches it. In principle a normal variable could take any value
at all — it's just wildly unlikely to be far out. The tails get thin astonishingly fast.
-
Set by two numbers. \mu and
\sigma, and nothing else. Everything about the shape follows from them.
Centre and width: what the two knobs do
The two numbers play two very different roles, and it helps to feel them separately:
-
\mu sets the centre — slide it and the whole bell
glides left or right without changing shape at all, like sliding a paper cut-out along
the table.
-
\sigma sets the width — a small
\sigma gives a tall, narrow, tightly-focused bell; a large
\sigma gives a low, broad, spread-out one. Since the total area is
always 1, squeezing the bell narrower forces it to grow
taller to keep its area — like squeezing a lump of clay.
The height of the bell at each value x is given by its density formula:
f(x) = \frac{1}{\sigma\sqrt{2\pi}}\,\exp\!\left(-\frac{(x-\mu)^2}{2\sigma^2}\right).
You don't need to memorise this to use it — just notice that x only ever
appears as (x-\mu)^2, the squared distance from the centre. That
squared distance is exactly why the curve falls away smoothly and symmetrically on both sides of
\mu, and why it plunges so fast far from the mean.
(A deeper, measure-theoretic treatment of the same curve — standardisation, the Gaussian integral,
and more — lives at
the normal distribution in finance.)
Shape the bell yourself
Slide \mu to move the curve, and
\sigma to make it wider or narrower. The faint curve is
the fixed standard normal N(0,1) for comparison. Watch how a narrower
bell grows taller — the area underneath stays 1 no matter what you do.
- The normal N(\mu,\sigma^2) is the symmetric bell curve set by its mean \mu and standard deviation \sigma.
- \mu moves the centre; \sigma sets the width.
- It is symmetric about \mu, so mean = median = mode, all at the peak.
- Density: f(x) = \dfrac{1}{\sigma\sqrt{2\pi}}\,e^{-(x-\mu)^2/(2\sigma^2)}; total area 1, so a narrower bell is taller.
Worked example 1 — where does the bulk of the data sit?
Adult male heights in a country are roughly N(178, 7^2) — mean
178 cm, standard deviation 7 cm. Where does
"most" of the data live?
For any normal distribution, the bulk clusters within a couple of standard deviations of the mean.
Roughly two-thirds of values fall within one \sigma of
\mu — here 178 \pm 7, i.e. between
171 and 185 cm. Almost everybody (about 19 in
20) falls within two \sigma —
178 \pm 14, i.e. 164 to
192 cm. A man of 206 cm sits four standard
deviations out, which is why you almost never meet one. (The exact percentages are the
empirical rule,
coming up next.)
Worked example 2 — comparing two bells
Two classes sit the same test. Class A scores N(70, 5^2); Class B scores
N(70, 12^2). Same mean — so what's different?
Because \mu is identical, both bells are centred at 70; neither
is shifted. But Class A's smaller \sigma = 5 makes a tall, narrow bell:
almost everyone scored close to 70, a very consistent class. Class B's
\sigma = 12 makes a low, broad bell: plenty of scores in the 50s and in
the 80s, a much more mixed group. Now change instead the mean — give Class C
N(80, 5^2) — and its bell is Class A's exact shape, just slid 10 marks to
the right. Different \mu shifts; different
\sigma re-shapes.
Worked example 3 — the standard normal
One special bell gets its own name. The standard normal is the normal distribution
with mean 0 and standard deviation 1 —
N(0, 1). It's the "reference bell," centred at zero and one unit wide.
Why single it out? Because any normal distribution can be slid and squeezed onto it. If
X \sim N(\mu, \sigma^2), then subtracting the mean and dividing by the
standard deviation,
z = \frac{x - \mu}{\sigma},
rescales X to a standard normal. The number z
answers "how many standard deviations from the mean is this value?" — so our
206 cm man from example 1 has
z = (206 - 178)/7 = 4. This single trick means we only ever need
one table (or one function) for the entire infinite family of normal curves. Those rescaled
values are called
z-scores.
The normal distribution is so useful that people reach for it reflexively — and that is a genuine
trap. Assuming normality when your data is actually skewed or heavy-tailed leads to badly
wrong conclusions. Several famous quantities are emphatically not normal:
- Income and wealth — strongly right-skewed, with a long tail of the very rich that a bell simply cannot reproduce.
- City sizes and word frequencies — follow power laws, where a few giants dominate.
- Financial returns — have heavy tails: extreme crashes and spikes happen far more often than a normal curve predicts.
That last one has bitten hard. Many of the risk models blamed for the 2008 financial
crisis assumed market moves were normally distributed, which made catastrophic swings look
essentially impossible — "a once-in-a-billion-years event." Then those "impossible" moves happened
several days running. The lesson: before you assume a bell, look at the tails. If rare
extremes matter, the normal distribution can be dangerously optimistic.
The curve first arose from errors, not people. Around 1809, Carl Friedrich
Gauss was trying to pin down the orbits of planets
and asteroids from telescope readings that never quite agreed. He asked: if my measurements scatter
around the true value, what's the most reasonable shape for that scatter? Out popped this exact
curve — which is why to this day it is often called the Gaussian distribution.
Decades later, Francis Galton built a gorgeous physical demonstration: the quincunx,
or "bean machine." Drop balls in at the top, let each one rattle down through a triangle of pegs —
bouncing left or right at random at every peg — and watch where they land in the bins at the bottom.
Every ball's final position is the sum of many little random left/right nudges… and the piles that
build up form a bell, right before your eyes. It is the
central limit theorem
made out of wood and marbles: sum up enough small independent randomness and you always get the same
curve. Once you've seen the beans pile into a bell, "why is everything normal?" stops feeling like a
coincidence.
See it explained