Can a probability measure be a fuzzy measure?

In an appropriately defined setting, an additive probability measure satisfies the relevant monotonicity and normalization conditions. Fuzzy measure theory permits more general cases.

Does fuzzy measure mean uncertain measurement?

Not in this mathematical usage. It refers to a set function with stated properties.

Must the measure be nonadditive?

It need not be additive, but additive instances are allowed. The distinction is the relaxation of the requirement.

AI Glossary / Graphs and symbolic AI / Advanced

Fuzzy measure theory

Definition

Fuzzy measure theory studies monotone set functions that need not be additive, together with ways to combine or integrate information using them. In a common finite setting, a normalized fuzzy measure assigns each subset a value between zero and one, increasing or staying equal when the subset grows.

What it means

Ordinary additive measures combine disjoint parts by summing their values. A fuzzy measure relaxes that requirement. This makes it possible to represent situations in which a combination of information sources or decision criteria contributes differently from the sum of its individual parts.

The term can be misleading. A fuzzy measure need not assign fuzzy numbers, and the sets it evaluates can be ordinary crisp subsets. It is also distinct from a membership function: membership grades describe individual elements' fit to a fuzzy set, whereas a fuzzy measure assigns values to subsets of a universe.

Capacities and monotone measures are closely related terminology. Definitions differ in normalization and additional continuity conditions, so this entry uses the normalized finite case for its example.

How it works

Choose a finite universe N, such as a set of evaluation criteria. Define a set function v with v(empty set) = 0 and v(N) = 1. Require monotonicity: if A is contained in B, then v(A) must not exceed v(B).

Values for combinations can express interaction. If two criteria overlap in what they contribute, their joint value can be lower than the sum of their individual values. If they complement each other, the joint value can exceed that sum. Additivity remains a special case rather than a forbidden possibility.

An aggregation procedure such as a Choquet or Sugeno integral uses the set function together with input scores. These are different procedures, not interchangeable names. A general finite measure requires values for many subsets, so practical models often impose structure or learn a constrained representation.

Three practical examples

A decision-support model accounts for overlap between two product-quality indicators rather than counting their shared contribution twice.
An information-fusion study represents the joint usefulness of several sources when combinations matter beyond individual source weights.
A service-evaluation prototype models complementary criteria, such as capability and availability, whose combined contribution differs from separate contributions.

Think of it this way

Think of evaluating a toolkit. A screwdriver and a nearly identical screwdriver may add less value together than a screwdriver and the matching set of bits. The worth of a subset depends on the combination.

The analogy illustrates nonadditivity, not a universal pricing formula. Monotonicity also assumes that adding an available item cannot reduce the modeled value; real costs or harmful additions may require a different representation.

Worked example

Checking a capacity and aggregating two scores

Let N contain criteria A and B. Define v(empty set) = 0, v({A}) = 0.6, v({B}) = 0.5 and v({A, B}) = 1. These values satisfy monotonicity because each singleton value lies between zero and the full-set value.

They are not additive: the singleton values sum to 1.1, while their joint value is 1. In this invented decision model, that difference represents overlap in contribution; it is not an arithmetic error.

Now assign criterion scores A = 0.8 and B = 0.3. For the discrete Choquet integral with these nonnegative scores, the common bottom level contributes 0.3 × v({A, B}) = 0.3. The remaining height, 0.8 - 0.3 = 0.5, belongs only to A and contributes 0.5 × 0.6 = 0.3. The aggregate is 0.6.

This is not obtained by directly computing 0.6 × 0.8 + 0.5 × 0.3, which gives 0.63 and treats singleton values as ordinary additive weights. The example's capacity and scores describe a preference model, not event probabilities.

Illustrative case study

The fictional Willow Evaluation team combines two quality indicators that partly measure the same underlying service feature. A simple weighted score makes that feature dominate its comparison.

The team explores a capacity-based model and elicits values for combinations, including cases where both indicators are strong. Reviewers then inspect whether the values satisfy monotonicity and whether the resulting preferences match the intended decision process.

Willow does not declare the new model objectively correct merely because it captures interaction. It tests sensitivity to the elicited values and documents the shared meaning of the input scales. The case illustrates an alternative aggregation design without claiming that the model improved an actual procurement decision.

Common misconceptions and mistakes

A fuzzy measure is just a membership function. They assign values to different kinds of objects and serve different roles.
Nonadditive means monotonicity can be ignored. Monotonicity is central to the capacity convention used here.
All subset values must add to one. Normalization applies to the full universe, not the sum of arbitrary subset values.
Choquet and Sugeno integrals always produce the same answer. They combine scores differently and suit different assumptions.

Business use

The theory is useful when a decision model has meaningful interaction between criteria or information sources. It can make assumptions about redundancy and complementarity explicit, but it also increases the burden of eliciting and validating parameters.

For n criteria, a fully general normalized capacity has 2^n - 2 nontrivial subset values before further constraints reduce freedom. Ask whether the additional complexity is supported by data or expert judgment. Ensure that the input scales and aggregation rule have a defensible interpretation, and show how rankings change when uncertain capacity values vary.

Check your understanding

Answer: A proposed normalized capacity has v({A}) = 0.8 and v({A, B}) = 0.7.

Is it valid under the stated definition?

Answer: No. Since {A} is contained in {A, B}, monotonicity requires the joint value to be at least 0.8. The displayed pair violates that requirement.

Related terms

Sugeno integral
Membership function (mathematics)
Capacity
monotone measure
Choquet integral
nonadditive aggregation
criterion interaction
fuzzy set
membership function

Learn next

Keep the worked capacity but swap the two criterion scores. Calculate the Choquet aggregate by separating the common score level from the additional level, then explain why the result changes when the stronger score belongs to B.

Sources

A primary research paper defines capacities and explains criterion interactions: https://www.sciencedirect.com/science/article/abs/pii/S0165011420304887

Research on monotone set functions and Choquet integration: https://www.sciencedirect.com/science/article/pii/0165011495001816

Choquet's foundational work is available in its journal archive: https://archive.numdam.org/item/AIF_1954__5__131_0/

The finite capacity, arithmetic and Willow case are original illustrations. Normalization conventions are explicitly limited to the finite example.

Previous: Fuzzy logic Next: Fuzzy number