Does full mean all variables in reality?

No. It means all variables selected for the probability model.

Can marginals be recovered from the joint?

Yes, by summing or integrating over unwanted variables.

Can every joint be written as a product of marginals?

No. That factorization requires independence; the general chain rule uses conditional probabilities.

AI Glossary / Statistics and probability / Intermediate

Full joint probability distribution

Definition

A full joint probability distribution specifies the probabilities of all combinations of values for all random variables included in a probability model. For finitely many discrete variables, it can be represented as a table with one probability for each complete assignment. Entries must be nonnegative and sum to one. Full refers to the model's chosen variable set, not every possible feature of the real world. Continuous variables require an appropriate joint distribution representation, such as a density when one exists, rather than a finite table of point probabilities.

What it means

Separate marginal distributions describe individual variables but generally do not determine how those variables occur together. The full joint distribution includes their dependence structure as well as their individual behavior. From it, one can obtain probabilities for subsets of variables and condition on observed evidence, subject to the usual conditions. This makes the full joint a conceptual foundation for probabilistic reasoning. In practice, explicitly storing every combination can be infeasible, which motivates structured representations such as Bayesian networks and other graphical models.

How it works

Choose variables and define their possible values. For a finite discrete model, assign a probability to each complete combination and verify normalization. To obtain a marginal probability, sum over variables not needed in the query. To obtain a conditional probability given evidence with positive probability, restrict to compatible assignments and normalize by the evidence probability. The chain rule factors a joint distribution into successive conditional distributions without assuming independence. Additional conditional-independence assumptions can simplify those factors, but must be justified by the model. For continuous quantities, corresponding marginalization generally uses integration and conditional distributions require suitable density or measure-theoretic treatment.

Three practical examples

Diagnostic reasoning: combine uncertain indicators and possible system states in a coherent probability model.
Scenario analysis: represent how demand and supply conditions occur jointly instead of assuming their separate probabilities can always be multiplied.
AI modeling: use a structured factorization to answer probabilistic queries without necessarily enumerating a complete table during every calculation.

Think of it this way

Think of a map showing every combination of two switches, with a weight attached to each configuration. Totals across rows or columns describe one switch alone. Keeping the full map also reveals whether the switches tend to be on together more often than separate summaries would suggest.

Worked example

Suppose a hypothetical model has two binary variables: A indicates a delayed delivery and B indicates an adverse operating condition. Define the full joint probabilities as P(A=1,B=1)=0.12, P(A=1,B=0)=0.18, P(A=0,B=1)=0.08 and P(A=0,B=0)=0.62. The four entries sum to one. Summing the two A=1 entries gives P(A=1)=0.30; summing the two B=1 entries gives P(B=1)=0.20. Therefore P(A=1 given B=1)=0.12/0.20=0.60. Multiplying the marginals instead gives 0.30 times 0.20 = 0.06, which differs from the actual joint probability 0.12, so A and B are not independent in this model. The conditional association does not establish that changing B would cause a corresponding change in A. The values are constructed for illustration and are not operational estimates.

Illustrative case study

Fictional case: Laurel Logistics Lab initially combines separate delay and disruption rates as if the events were independent. After reviewing paired records, analysts build a joint model that preserves their association and check it against later data. They use the model for scenario calculations while clearly documenting omitted variables and uncertain estimates. This invented case illustrates why coherent dependence modeling matters, without claiming that the resulting model proves a causal mechanism or predicts every disruption accurately.

Common misconceptions and mistakes

Do not assume marginals uniquely determine the full joint distribution. Multiplying marginal probabilities requires the relevant independence assumption. Conditional independence differs from unconditional independence and can change after evidence is observed. A density value is not a point probability for a continuous variable. A zero estimated cell count need not imply that the underlying event is impossible. Full does not mean perfectly known or free of modeling error. Even a valid normalized table may describe the wrong population or omit important states, and probabilistic conditioning is not the same operation as a causal intervention.

Business use

Businesses can use joint models when decisions depend on combinations of uncertain events rather than one event at a time. Start with a manageable variable set and document how probabilities were estimated or elicited. Check normalization, impossible combinations and consistency with available evidence. Explicit tables grow quickly: 20 binary variables already require 1,048,576 cells, with one fewer free probability after normalization. Structured models can reduce representation costs under suitable assumptions, but efficient inference is not guaranteed for every structure. Communicate uncertainty in estimated probabilities as well as the probabilities produced by the model.

Check your understanding

P(A)=0.3 and P(B)=0.2. Is P(A and B) necessarily 0.06?

Answer: No. That product is justified only with independence. A joint model or other information about dependence is needed to determine the intersection probability.

Related terms

Learn next

Learn marginalization and conditioning on a small joint table, then study the chain rule. Explore how conditional independence supports compact graphical representations and why observational probabilities alone do not determine intervention effects.

Sources

Russell and Norvig, AIMA chapter 14 teaching slides, representing full joint distributions with Bayesian networks: https://aima.eecs.berkeley.edu/2nd-ed/slides-pdf/chapter14a.pdf Russell and Norvig, AIMA chapter 14 teaching slides, probabilistic inference: https://aima.eecs.berkeley.edu/~russell/aima/2nd-ed/slides-pdf/chapter14b.pdf AIMA code project, probability-model implementations: https://github.com/aimacode/aima-java/blob/AIMA3e/README.md

Previous: Full functional dependency Next: Full-rank