The normal distribution, also called the Gaussian distribution or the bell curve, is a continuous probability distribution that is symmetric around its mean and fully described by two parameters: the mean (μ), which sets its center, and the standard deviation (σ), which sets its spread. In statistics, it is the most widely used model for continuous measurements and the foundation of many methods of estimation and hypothesis testing.
Its importance comes from the central limit theorem: sums and averages of many independent random effects tend toward a normal shape, whatever the shape of the individual effects. That is why measurement errors, averages of samples and many biological and physical quantities are approximately normal.
Shape and parameters of the normal distribution
The normal distribution has a single peak at the mean, and its density falls off symmetrically and ever more steeply on both sides, approaching zero but never reaching it. In a normal distribution, the mean, median and mode coincide. Its density is proportional to the exponential of minus the squared distance from the mean divided by twice the variance.
Changing the parameters only shifts or stretches the same shape. Increasing μ moves the curve right; increasing σ makes it lower and wider, because the total area under the curve always equals one. The variance is σ squared.
| Property | Value | Meaning |
|---|---|---|
| Center | Mean μ, equal to median and mode | Location of the peak |
| Spread | Standard deviation σ | Distance to the inflection points |
| Skewness | 0 | Perfectly symmetric |
| Excess kurtosis | 0 | Reference level for tail heaviness |
| Support | All real numbers | Any value is possible in theory |
| Standard form | μ = 0, σ = 1 | Basis for z-scores and tables |
The empirical rule and z-scores
The empirical rule, also called the 68-95-99.7 rule, states how much probability lies within whole numbers of standard deviations from the mean. It is a quick way to judge whether a value is ordinary or unusual.
| Interval | Share inside | Share outside |
|---|---|---|
| μ ± 1σ | About 68.3 percent | About 31.7 percent |
| μ ± 1.96σ | 95 percent | 5 percent |
| μ ± 2σ | About 95.4 percent | About 4.6 percent |
| μ ± 2.576σ | 99 percent | 1 percent |
| μ ± 3σ | About 99.7 percent | About 0.3 percent |
Any normal variable can be converted to the standard normal distribution by computing a z-score: subtract the mean and divide by the standard deviation. A z-score of 2 means the value lies two standard deviations above the mean, regardless of units. The values 1.96 and 2.576 are the familiar critical values for two-sided 95 and 99 percent confidence intervals.
Why the normal distribution is so common
The central limit theorem explains most appearances of the normal distribution. When a quantity is the sum of many small, independent influences with finite variance, its distribution approaches normal as the number of influences grows. The same applies to the sampling distribution of the mean: averages of samples are approximately normal even when individual observations are not.
How large a sample must be depends on the underlying data. For roughly symmetric data, samples of a few dozen are usually enough; for strongly skewed data, such as incomes or waiting times, hundreds of observations may be needed. This is why the standard error, confidence intervals and many tests assume normality of the estimate, not necessarily of the raw data.
The theorem has limits. It requires independent observations and finite variance; with strongly dependent data, such as consecutive readings of a slowly drifting sensor, or with extremely heavy-tailed data, averages can stay far from normal. Products of many factors tend toward a log-normal shape instead, because their logarithms form a sum.
Normal distribution compared with related distributions
Many real datasets resemble a bell curve only loosely, and choosing a better-fitting distribution can change conclusions. The normal model should be compared with plausible alternatives.
| Distribution | Shape | Typical use |
|---|---|---|
| Normal | Symmetric, light tails | Measurement errors, averages |
| Student t | Symmetric, heavier tails | Means from small samples |
| Log-normal | Right-skewed, positive values | Incomes, file sizes, response times |
| Binomial | Discrete counts of successes | Conversions out of a fixed number of trials |
| Poisson | Discrete counts per interval | Arrivals, defects, events |
| Uniform | Flat over a range | Random number generation |
Binomial and Poisson counts are well approximated by a normal distribution when expected counts are large, which is why many tests on proportions use normal approximations. The Student t distribution converges to the normal as degrees of freedom grow, and the two are practically identical beyond about 30 to 50 degrees of freedom for most purposes.
Checking normality and common mistakes
Before relying on a normal model, analysts should check the data visually and numerically. A Q-Q plot, which compares sample quantiles with theoretical normal quantiles, reveals skewness and heavy tails more clearly than a histogram. Formal tests such as Shapiro-Wilk or Anderson-Darling exist, but with large samples they flag tiny, harmless deviations, and with small samples they miss large ones.
- Assuming normality by default for skewed, bounded or count data.
- Underestimating tail risk: real financial and operational data often has far more extreme events than a normal model predicts.
- Reading 3σ as impossible: in a million observations, a few thousand are expected beyond three standard deviations.
- Applying the empirical rule to multimodal data, where mean and standard deviation describe the data poorly.
- Confusing the data with the estimate: tests often need a normal sampling distribution, not normal raw values.
In plain words
Measure the height of thousands of adults and draw how many people have each height. Most are close to the average, fewer are very short or very tall, and the picture looks like a bell. That shape is the normal distribution.
It shows up whenever many small, independent influences add together. Two numbers describe it completely: the average, which sets the middle, and the spread, which says how wide the bell is. Once those are known, it is easy to say how unusual any single value is.
A real-world example
A beverage plant fills bottles labeled 500 ml. Measurements over several weeks show that fill volumes follow an approximately normal distribution with a mean of 503 ml and a standard deviation of 1.5 ml. Using z-scores, the quality engineer calculates that a bottle under 500 ml lies two standard deviations below the mean, so about 2.3 percent of bottles would be underfilled.
To reduce underfills to roughly one in a thousand, the target must sit about 3.1 standard deviations above 500 ml. The plant can either raise the mean to about 504.7 ml, which costs extra product, or reduce variation. After servicing the filling heads, the standard deviation drops to 1.0 ml, and a mean of 503.1 ml reaches the same goal with less wasted liquid.