Is a distribution always a bell curve?

No. It can be skewed, multimodal, discrete or take many other forms.

Is a density value a probability?

For a continuous model, probability is area over a range; density height alone is not probability at a point.

Can two distributions share a mean and variance?

Yes. Even matching those summaries does not determine every feature of shape or tail behavior.

AI Glossary / Statistics and probability / Beginner

Distribution

Definition

A distribution describes how values of a variable are spread across possible outcomes or ranges. An empirical distribution records the frequencies in observed data; a probability distribution assigns probabilities to possible events under a model or population description. Distributions can be discrete, continuous or mixed. Their shape, location, spread and tail behavior often matter more than any single summary statistic. The term here concerns statistical data, not distributing software or products.

What it means

Knowing an average does not tell you how often extreme outcomes occur or whether values form one group or several. Two processes can have the same mean but very different variability and operational risks. A distribution preserves more of that pattern. For a single variable, it might describe waiting times or categories of request. A joint distribution describes several variables together, and a conditional distribution describes one variable under specified values or conditions of others. These distinctions matter when a machine-learning system encounters a different population from its training data.

How it works

Define the variable, observation unit, reference population and time period. Summarize observed values using counts, proportions, an empirical cumulative distribution or an appropriate graph. For discrete outcomes, probability masses specify probabilities at individual values. For a continuous model, probabilities over intervals are obtained from density areas; a density height alone is not an event probability. A cumulative distribution gives the probability of being at or below a threshold. If fitting a theoretical family, estimate parameters and assess whether its assumptions and tails match the data. Keep uncertainty about the fitted distribution separate from the variability it describes.

Three practical examples

Service operations: inspect the distribution of response times to understand both typical waits and slow-tail experiences.
AI development: compare feature and label distributions across training, validation and deployment datasets.
Demand planning: evaluate probabilities of low, ordinary and unusually high demand instead of relying solely on average sales.

Think of it this way

Think of an average as one landmark on a map and a distribution as the map's terrain. The landmark is useful, but it cannot tell you where the steep slopes, gaps or distant edges lie. A decision about navigating the terrain needs more than the location of that single point.

Worked example

Consider two hypothetical collections of four waiting times. Process A has values [4, 4, 6, 6] minutes, while process B has [0, 0, 10, 10]. Both means equal five minutes. Treating each observation as one quarter of its empirical distribution, A has variance [(4-5)^2 + (4-5)^2 + (6-5)^2 + (6-5)^2]/4 = one squared minute. B has empirical variance 25 squared minutes. No A observation exceeds eight minutes, whereas half of B's observations do. Thus equal means conceal very different observed spread and threshold exceedance. These are empirical summaries using divisor four, not the usual unbiased sample-variance formula using divisor three. Four observations cannot establish the future tail probability with confidence, but the comparison clearly demonstrates why the distributions contain information absent from the means.

Illustrative case study

Fictional case: Silver Support reviews an AI-assisted service process. Its average response time remains unchanged, yet a distribution plot shows more very fast responses and more unusually slow ones. Analysts investigate request mix and handoff patterns before deciding whether the change benefits customers. They compare similar operating periods and report the number of observations behind the graphs. This invented case illustrates distributional review; it does not establish a causal explanation or claim that a particular average conceals problems in every setting.

Common misconceptions and mistakes

Do not assume all data follow a normal distribution. A histogram's appearance depends on bin widths and boundaries, so it is one representation rather than the distribution itself. A dataset's finite empirical distribution differs from the unknown population distribution. Similar individual-variable distributions do not guarantee a similar joint distribution: relationships between variables can change. Zero observed occurrences do not prove zero population probability. A named family is a modeling assumption, and a good fit in the center may coexist with poor tail predictions where operational consequences are greatest.

Business use

Businesses use distributions to make uncertainty and variation visible in forecasting, quality management and model evaluation. State whether a figure describes observed frequencies, a fitted population model or a predictive distribution for future outcomes. Compare relevant conditions rather than pooling unlike populations without explanation. When a model's environment changes, examine both individual features and their relationships to outcomes. Distribution monitoring can reveal changes worth investigating, but a detected shift does not automatically prove declining model performance or identify the action required.

Check your understanding

Why should a manager who cares about waits above eight minutes distinguish the two example processes?

Answer: Their equal averages do not capture the relevant event. In the observed data, A has no waits above eight and B has two out of four. The distribution directly addresses the threshold question, while the mean alone does not.

Related terms

Probability distribution
Empirical distribution
Histogram
Cumulative distribution function
Mean
Variance
Quantile
probability density
probability mass
joint distribution
distribution shift

Learn next

compare histograms, empirical cumulative curves and summary statistics for the same dataset. Construct examples with equal means but different tails. Then distinguish marginal, joint and conditional distributions and consider which distribution a forecasting or classification decision actually needs.

Sources

Google, Machine Learning Glossary, distribution: https://developers.google.com/machine-learning/glossary NIST/SEMATECH, Distribution: Location, Spread and Shape: https://www.itl.nist.gov/div898/handbook/ppc/section1/ppc131.htm NIST/SEMATECH, Histogram: https://www.itl.nist.gov/div898/handbook/eda/section3/histogra.htm NIST/SEMATECH, What is a Probability Distribution?: https://www.itl.nist.gov/div898/handbook/eda/section3/eda361.htm

Previous: Distributed tree search Next: Distributional - relational database