AI Glossary / Statistics and probability / Advanced

Multivariate random variable

Definition

A multivariate random variable, also called a random vector, is an ordered collection of random variables defined on the same probability space. One outcome of the underlying experiment determines a value for every component.

What it means

The vector describes several quantities together. Its joint distribution specifies probabilities for combinations of values, while marginal distributions describe individual components. The components need not be independent, identically distributed, or measured in the same units.

How it works

Define the components and their order, then specify their joint probability law. A discrete vector can have a joint probability mass function. An absolutely continuous vector has a joint density. Other possibilities include mixed distributions and distributions concentrated on a lower-dimensional set.
A joint cumulative distribution function describes the probabilities that all components fall below specified thresholds. Summing or integrating out components yields marginal distributions where the relevant mass or density representation exists.

Three practical examples

A fictional AI evaluation records a failure indicator and a delay indicator for each request, preserving which outcomes occurred together.
A fictional sensor experiment records temperature, pressure, and vibration from the same sampling interval as a three-component vector.
A fictional service simulation generates a request count and a waiting time jointly, allowing different kinds of variables within one model.

Think of it this way

A single photograph can capture several people's positions at the same instant. Keeping the positions together preserves relationships that separate collections of individual positions could lose. A random vector similarly keeps several quantities attached to the same underlying outcome.

Worked example

Consider an original experiment that selects one of six recorded requests with equal probability. Let X indicate whether the selected request failed, and let Y indicate whether it exceeded a delay threshold. Both indicators take values zero or one.
Requests one and two have vector (0, 0). Request three has (0, 1), request four has (1, 0), and requests five and six have (1, 1).
The four possible vector values therefore have probabilities one third, one sixth, one sixth, and one third respectively. The six underlying request outcomes map to four distinct vector values. A realized vector does not necessarily identify the original request uniquely.
The probability X=1 is one sixth plus one third, or one half. The probability Y=1 is also one half. However, the probability that both equal one is one third, not the one quarter obtained by multiplying those marginal probabilities.
Thus the two indicators are dependent in this invented population. Among the three failed requests, two also exceeded the delay threshold, so P(Y=1 given X=1)=2/3. Among the three nonfailed requests, only one exceeded it, giving P(Y=1 given X=0)=1/3.
Both expected values are one half. Because the product XY equals one only when both indicators equal one, E(XY)=1/3. Their covariance is therefore one third minus one quarter, or one twelfth.
Each indicator has variance one quarter. Their correlation is (1/12)/(1/4)=1/3. These numerical summaries describe the specified six-request population; they do not establish a causal relationship between failure and delay.
Now transform the vector into S=X+Y. The possible values zero, one, and two each have probability one third. Its mean is one, and its second moment is (zero plus one plus four)/3=5/3. Its variance is consequently 5/3 minus one, or 2/3.
The same result follows from adding the component variances and twice their covariance: one quarter plus one quarter plus twice one twelfth equals two thirds. This agreement checks the joint calculation.
Consider D=X-Y as another transformation. It equals minus one with probability one sixth, zero with probability two thirds, and one with probability one sixth. Its mean is zero and its variance is one third.
Although S and D have covariance zero in this example, they are not independent. Whenever S=1, D must be either minus one or one, so it cannot equal zero. Yet D=0 has unconditional probability two thirds. A transformation does not make dependence disappear merely because a covariance happens to vanish.
For a separate continuous example, draw U uniformly from zero to one and set V=U. Both individual variables are uniform. Their joint probability P(U no greater than 0.6 and V no greater than 0.4) is 0.4, because both conditions reduce to U no greater than 0.4.
If these uniform components were independent, that probability would be 0.6 times 0.4=0.24 instead. The marginal distributions alone therefore do not determine the joint law.
In the actual V=U construction, every vector lies on the diagonal line segment. The probability U=V is one, but the probability of any specified point such as (0.5, 0.5) is zero. There is no ordinary joint density with respect to two-dimensional area for this line-supported distribution.
Finally, let B be an independent fair zero-or-one indicator and T a uniform waiting time between zero and one. The vector (B, T) combines a discrete and a continuous component. The probability B=1 and T no greater than 0.3 is one half times 0.3=0.15. Random vectors are not restricted to collections of variables of one type.

Illustrative case study

Fictional company Cedar Evaluation initially stores failure and delay summaries separately. Analysts retain the original request-level pairs so they can calculate joint events and examine how the two outcomes occur together.
They specify the indicator definitions, observation period, and sampling rule before generalizing beyond the recorded requests. The fictional case illustrates the value of preserving pairing rather than inferring relationships from separate totals.

Common misconceptions and mistakes

An observed vector is a realization, not the full random object.
Marginal distributions do not uniquely determine a joint distribution.
Zero covariance does not generally imply independence.
Continuous marginal distributions do not guarantee a joint area density.
The mean vector need not be one of the possible observed values.

Business use

Define the joint event relevant to the decision. In the fictional evaluation, the fraction of requests that are both failed and delayed cannot be recovered from the two marginal fractions alone. Keep the pairing and the population definition intact.

Check your understanding

What is the probability of both indicators being one?

Answer: One third.

What is the variance of their sum?

Answer: Two thirds.

Why does the uniform diagonal example lack an area density?

Answer: All its probability lies on a line segment.

Related terms

Marginal distribution
Covariance matrix
Realization
Random vector
joint distribution
statistical independence
mixed distribution

Learn next

Compare joint and marginal distributions, then examine conditional distributions and transformations of random vectors.

Sources

Marco Taboga, StatLect, Random vectors: https://www.statlect.com/fundamentals-of-probability/random-vectors . Request population, transformations, continuous constructions, and company scenario are original illustrations.

Previous: Multivariate procedures Next: Multivariate regression