Is a realization always a number?

No. It may be a vector, category, function or other value supported by the probability model.

Can two realizations be identical?

Yes. Repeated experiments can produce the same value, particularly for discrete variables.

Is a sample path a realization?

Yes. It is a realization of a stochastic process across its index set, such as time.

AI Glossary / Mathematics and optimization / Beginner

Realization

Definition

A realization is a particular value taken by a random variable when an outcome of a random experiment is specified or observed. For a stochastic process, a realization can be an entire sequence or path of values rather than a single number.

What it means

A random variable describes how possible outcomes are translated into values. A realization is one result of that mapping. If X represents tomorrow's number of support requests, then X is the random variable and an observed count such as 47 is a realization.

The distribution describes the possible values and their probabilities. A realization does not replace that distribution or show what must happen on the next occasion. Several different underlying outcomes may produce the same realized value.

In simulation, the value generated during one run is also called a realization. It is a result under the simulation's assumptions, not necessarily an observation of the real world. Distinguishing observed and simulated realizations helps prevent a model output from being presented as measured evidence.

How it works

Formally, a random variable X maps an outcome ω from a sample space to a value X(ω). Once ω is fixed, the corresponding value is a realization. Uppercase X and lowercase x are often used to distinguish the random variable from a particular realized value, although notation varies.

For a pair of random variables, a realization is a pair. For a time-indexed stochastic process, fixing the outcome determines a path across time. One process realization may therefore contain hundreds of measurements, and those measurements need not be independent.

A sample commonly refers to a collection of observations or realizations obtained under a sampling scheme. Independence is an additional assumption, not a consequence of calling values a sample.

For a continuously distributed variable, a particular point can have probability zero while still being a possible realized value. Probabilities for intervals remain meaningful. Instrument rounding adds another distinction: a recorded display value may represent a range of underlying values rather than the exact mathematical quantity.

Three practical examples

A demand simulator generates one possible month of daily orders; the complete generated trajectory is one realization of the demand process.
A sensor records a particular temperature, giving a measured realization of a modeled temperature variable subject to the sensor's precision.
A data scientist repeats a randomized training procedure and treats each run's final validation score as one realized outcome of that experimental setup.

Think of it this way

Think of a recipe for drawing a ticket from a box and reading the printed number. The drawing rule describes the experiment, while the number on the selected ticket is the realization of the number-valued variable.

The analogy is limited because real probability models can have infinitely many possible values, and a process can produce a whole path. It nevertheless separates the generating description from one produced result.

Worked example

One outcome, several random variables

Toss a fair coin twice independently. The four equally likely outcomes are HH, HT, TH and TT. Define X as the number of heads and Y as an indicator that the first toss is heads.

For the outcome HT, the realization of X is 1, the realization of Y is 1, and the joint realization is (1,1). For TH, X is also 1 but Y is 0. Observing X=1 therefore does not reveal which of those two underlying outcomes occurred.

Across all four outcomes, X takes values 2,1,1,0 with probabilities 1/4,1/4,1/4,1/4 at the outcome level. Its value distribution is P(X=0)=1/4, P(X=1)=1/2 and P(X=2)=1/4. Its expectation is 1.

A realized value of zero does not contradict that expectation. Expectation summarizes the distribution; it is not a promise that each realization equals the average. Repeating the two-toss experiment produces further realizations, which must be interpreted under the stated experimental conditions.

Illustrative case study

A fictional parcel-delivery company simulates demand for the next seven days. One generated path contains a sharp Wednesday peak, and an early presentation labels that chart as the expected forecast.

The analyst corrects the caption to identify it as one simulated realization. The team then presents a collection of paths and appropriate summaries to explain the range of outcomes implied by the model.

When actual orders arrive, they form an observed path that can be compared with the model's predictions. A mismatch in one week does not by itself establish the entire model's adequacy or inadequacy; the team examines the pattern, assumptions and repeated evidence.

Common misconceptions and mistakes

A realization is the probability distribution. It is a particular value or path generated under that distributional description.
The expected value must occur in each run. Individual results can differ, and some expectations are not possible individual outcomes.
Different outcomes always have different realized values. Random variables can map several outcomes to the same value.
A long path supplies independent samples automatically. Time dependence may connect its values.
Simulated realizations are observed facts about the world. They are generated results under model assumptions.

Business use

The term helps teams communicate uncertainty without confusing scenarios, averages and observations. Label charts and reports according to what they show: one observed path, one simulated path, an average across runs or an interval summary.

For reproducible simulation work, record the model version, input assumptions and random-generation setup. Reproducing a path helps audit the calculation, but it does not validate the assumptions that generated it. When evaluating a system, also distinguish repeated runs on the same data from independent new observations of the underlying business process.

Check your understanding

In the coin example, observing X=1 identifies either HT or TH. Does it identify Y uniquely?

Answer: No. Y is one for HT and zero for TH. A realization of one random variable need not determine the underlying outcome or every other variable.

Related terms

Learn next

Read random variable and stochastic process. Define two different measurements of the same experiment, then list outcomes that give the same realization for one measurement but different realizations for the other.

Sources

Aalto Dictionary of Machine Learning, realization: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. Source for the mapping from an underlying outcome to a realized value. The coin experiment, path interpretation and fictional parcel scenario are original explanations and illustrations.

Previous: Real-time search Next: Realized variance