Normal Distribution

The bell curve — why so much data naturally clusters around the middle.

Key Points

Defined by mean (μ) and standard deviation (σ). Mean sets the center, SD sets the spread.
Perfectly symmetric — mean = median = mode. The bell is identical on both sides.
The 68-95-99.7 rule: 68% of data falls within 1σ, 95% within 2σ, 99.7% within 3σ of the mean.
The total area under the curve = 1 (all probabilities sum to 1).
Many statistical tests assume normality — violating this assumption can invalidate results.

What it is

The normal distribution is a symmetric, bell-shaped probability distribution defined entirely by two parameters: the mean (center of the bell) and the standard deviation (width of the bell). It is the most important distribution in statistics because a huge range of real-world measurements — heights, IQ scores, measurement errors, blood pressure — naturally follow it. It is also the foundation of most classical statistical tests and the central limit theorem.

Why it exists

The normal distribution emerges naturally from the central limit theorem: when you average many independent random variables, the result tends toward a normal distribution regardless of the original distribution. This is why errors in measurement, biological traits, and aggregated effects all cluster into bell shapes. Its mathematical properties make it tractable — many calculations that are impossible for other distributions have clean closed-form solutions for the normal distribution.

What it measures

The normal distribution describes how probable different values are for a continuous variable. The peak (at the mean) is the most probable value. Values further from the mean are progressively less probable. The standard deviation controls how quickly probability drops off — a narrow bell (small σ) means most values are close to the mean, a wide bell (large σ) means values are spread far from it.

Intuition

Imagine measuring the heights of 10,000 adults. A few people are very short, a few are very tall, and the vast majority cluster around 170cm. If you plotted this as a histogram, the bars would form a bell shape — high in the middle, tapering symmetrically on both sides. This is the normal distribution. The mean is 170cm, the SD might be 10cm, which means 95% of people fall between 150cm and 190cm (within 2 SDs). The distribution tells you not just where the center is, but how confident you can be that any given person is near it.

Example

Exam scores: mean = 70, standard deviation = 10. Distribution is approximately normal.

68% of students scored between 60 and 80. 95% scored between 50 and 90. A score of 90 is 2 standard deviations above the mean — in the top 2.5% of the class. A score of 50 is 2 SDs below the mean — in the bottom 2.5%.

Formulas

Probability Density Function

f(x) = \frac{1}{\sigma\sqrt{2\pi}} e^{-\frac{1}{2}\left(\frac{x-\mu}{\sigma}\right)^2}

f(x) — probability density at value x — height of the curve at that point
μ — mean — the center of the bell
σ — standard deviation — controls width of the bell
e — Euler's number ≈ 2.718 — base of natural logarithm
1/σ√2π — normalizing constant — ensures total area under curve equals 1

Z-Score (Standardization)

z = \frac{x - \mu}{\sigma}

z — z-score — how many standard deviations x is from the mean
x — the observed value
μ — mean of the distribution
σ — standard deviation of the distribution

When to use

Use when data is continuous and you expect it to cluster symmetrically around a central value — heights, weights, measurement errors
Use the normal distribution assumption when applying t-tests, ANOVA, linear regression, and most classical statistical tests
Use z-scores to compare values from different normal distributions on a common scale
Use it to detect outliers — values beyond 3 standard deviations from the mean are flagged as anomalies

When not to use

Do not assume normality for income, prices, or any quantity that cannot be negative — these are typically right-skewed
Do not use normal distribution for count data (number of events) — use Poisson distribution instead
Do not apply tests that assume normality to small samples without checking — the central limit theorem requires sufficiently large n
Do not use it for binary outcomes — use Bernoulli or binomial distributions instead

Common mistakes

Assuming all continuous data is normally distributed — checking with a histogram or Q-Q plot before applying parametric tests is essential
Confusing the probability density f(x) with probability — the density can exceed 1, only areas under the curve are probabilities
Ignoring skewness — a histogram that looks roughly bell-shaped but has a long tail is NOT normal and will violate test assumptions
Using z-scores without checking normality — z-scores are meaningful summaries of position only within a normal distribution

In practice

Normal distribution is assumed by linear regression (residuals should be normally distributed), t-tests, ANOVA, and many other classical methods. In feature engineering, checking whether a feature is normally distributed before applying normalization matters — log transformation is commonly applied to right-skewed features to make them more normal. In scikit-learn, StandardScaler standardizes features to have mean=0 and SD=1 (a standard normal), which is prerequisite for many algorithms like SVM and PCA. scipy.stats.norm provides pdf, cdf, and sampling functions.

Was this concept helpful?

Spot an error or gap? Suggest an edit →

Related concepts

variance-standard-deviationmean-median-modehypothesis-testingprobability-basicscorrelation-vs-causation