How does an MRF differ from a Bayesian network?

An MRF uses an undirected graph, while a Bayesian network uses a directed acyclic graph and conditional probability factorization.

What is the partition function?

It is the normalizing quantity that turns unnormalized configuration scores into probabilities.

Does an edge indicate causation?

Not by itself. In an MRF it represents a statistical interaction within the model.

AI Glossary / AI and machine learning / Advanced

Markov random field

Definition

A Markov random field, or MRF, is an undirected probabilistic graphical model that represents conditional-dependence structure among variables. Its local Markov property says that a variable is conditionally independent of the other non-neighboring variables once its neighbors are known.

What it means

Some relationships are easier to express without choosing a direction. Neighboring image labels, for example, may tend to agree, while each also depends on local evidence. An MRF represents such interactions with an undirected graph.

The graph is not necessarily a physical grid, and the word “Markov” does not require a sequence over time. Nodes can represent pixels, sites or other variables connected by relevant relationships. The model concerns conditional dependence, not an automatic claim of causation.

A missing edge has a conditional-independence meaning under the graph's Markov interpretation. It does not necessarily mean that the two variables are independent without conditioning. Indirect paths can still connect them. Understanding that distinction is important when reading a network diagram.

How it works

An MRF can be specified through potential functions over cliques, which are fully connected subsets of nodes. Potentials assign nonnegative compatibility scores to local configurations. Multiplying the potentials and dividing by a normalizing constant gives a probability distribution when the construction is valid.

The normalizing constant, called the partition function, sums or integrates the unnormalized scores over all configurations. Potentials do not each need to be probabilities. For strictly positive distributions, the Hammersley - Clifford theorem connects the Markov properties with factorization over graph cliques under its conditions.

Inference can seek marginal probabilities or a most probable configuration. These are different tasks and can be computationally difficult on general graphs. Special structures and approximate algorithms can make particular problems manageable. Model design must balance local evidence and interaction strengths so that smoothing or agreement does not overwhelm real distinctions.

Three practical examples

1. An image-analysis model combines evidence for each pixel's label with preferences for compatible labels at neighboring pixels.

2. A spatial-analysis model represents dependence among nearby locations while distinguishing local measurements from neighborhood interactions.

3. A research system models jointly interacting discrete variables using clique potentials and compares exact and approximate inference on a small benchmark.

Think of it this way

Think of a set of neighboring tiles whose arrangements receive compatibility scores. A good overall arrangement must balance all local preferences together. The limit is that MRF probabilities require global normalization, and local compatibility does not by itself give the probability of the complete arrangement.

Worked example

Consider two binary variables A and B with one potential: score 2 when they agree and score 1 when they differ. The four configurations have scores 2, 1, 1 and 2, so the partition function is 6.

The probability of A = 0 and B = 0 is 2 / 6 = 1/3. The total probability of agreement is (2 + 2) / 6 = 2/3.

The potential value 2 is not itself a probability. It becomes part of a probability only after normalization over all configurations. This tiny example is easy to enumerate; larger networks can make the same normalization task expensive.

Illustrative case study

A fictional inspection team uses an MRF to make neighboring region labels more consistent in product images. The initial model removes isolated noise, but it also erases small genuine defects because the agreement preference is too strong.

The team evaluates defect detection separately from visual smoothness and adjusts the balance between image evidence and neighborhood interaction. It checks several defect sizes and lighting conditions before selecting the model.

The case illustrates that a plausible structural prior can help and harm. A smoother output is not automatically a more accurate or more useful interpretation.

Common misconceptions and mistakes

An MRF is necessarily a time-series model. It can represent spatial or other undirected dependencies.
Every potential is a probability table. Potentials are compatibility scores that require appropriate normalization.
A missing edge implies unconditional independence. The relevant implication is conditional on the graph's separation structure.

Business use

MRFs can support structured prediction where outputs interact, including image and spatial analysis. They provide an explicit way to combine local evidence with relationships among neighboring variables.

Ask what the nodes, edges and potentials represent, and which inference task the system performs. Evaluate whether interaction terms preserve important small or unusual patterns. Include inference cost and approximation quality in the assessment. A useful graph should encode justified relationships rather than merely produce a visually coherent result.

Check your understanding

A clique potential assigns score 4 to one configuration. Why can that number be greater than one without violating probability rules?

Answer: A potential is an unnormalized compatibility score. Probabilities are obtained after combining potentials and dividing by the partition function.

Related terms

Probabilistic Graphical Model
Markov network
Conditional independence
clique
potential function
partition function
Gibbs distribution

Learn next

Read conditional independence and factor graphs. Enumerate a two-variable model by hand, calculate its partition function and distinguish a marginal probability from a most probable joint configuration.

Sources

Carnegie Mellon University, Undirected Graphical Models, authored course notes: https://www.cs.cmu.edu/~epxing/Class/10708-19/notes/lecture-03/. Shalizi, Markov Random Fields: https://stat.cmu.edu/~cshalizi/dst/20/lectures/22/lecture-22.html. Oxford research notes on MRF theory: https://www.fmrib.ox.ac.uk/datasets/techrep/tr00yz1/tr00yz1/node4.html. The binary model and inspection-team scenario are fictional.

Previous: Markov property Next: Markov reward process