Does mutual utility independence require an additive utility function?

No. Structured interaction terms can be compatible with it, as the worked example demonstrates.

Is it the same as mutual preferential independence?

No. Preferential independence compares certain outcomes; utility independence compares lotteries over attributes with the complement fixed.

Can statistically dependent attributes be utility independent?

Yes. Statistical dependence describes outcomes, while utility independence describes preferences over specified uncertain prospects.

AI Glossary / Graphs and symbolic AI / Intermediate

Mutually utility independent

Definition

Attributes are mutually utility independent when every subset is utility independent of its complement: preferences over lotteries on that subset do not depend on the fixed levels assigned to the remaining attributes. This is a condition about preferences under uncertainty. It is stronger in scope than merely ranking certain outcomes and is different from statistical independence between random attributes.

What it means

A lottery is an uncertain prospect with specified possible outcomes and probabilities. To assess utility independence, compare two lotteries that vary one attribute or a subset, while keeping the other attributes fixed. Then change those fixed attributes equally for both lotteries and check whether the preference ordering changes.
The word “mutually” requires the corresponding condition across subsets and complements, not just one direction. Utility independence can simplify the structure of a multiattribute utility function, but it does not mean that all attributes contribute through a simple weighted sum. Interaction terms may remain compatible with it.

How it works

Within expected-utility theory, utility independence means that conditional utility functions for the varying subset preserve its lottery preferences as the complement changes. Positive affine transformations preserve those preferences: multiplying all relevant utilities by a positive constant and adding a common constant does not change expected-utility rankings.
Assessment therefore asks more than whether a person always prefers a higher value of one attribute. It asks whether their ordering of uncertain prospects over that attribute changes with the fixed context. Representation results for mutually utility-independent attributes can lead to structured multiplicative or additive special-case models under the required assumptions. A formal assessment must specify attribute domains and preference conditions rather than relying on an unexplained formula.

Three practical examples

A service-design analyst can test whether preferences between uncertain delivery-speed prospects change when both prospects offer the same fixed support level.
A planning assistant can assess whether a user's willingness to accept variability in one benefit depends on a fixed level of another benefit.
A decision-model reviewer can investigate whether interaction terms are compatible with the elicited lottery preferences instead of assuming all interactions violate independence.

Think of it this way

Think of comparing two uncertain game rewards while receiving the same guaranteed bonus with either choice. If changing that common bonus changes which gamble you prefer, the gamble preference depends on the bonus context. The analogy is limited because utility independence requires systematic comparisons across a defined domain, and several attributes may vary jointly.

Worked example

Consider an invented expected-utility function U(x,y) = x + y + xy for attributes x and y between 0 and 1. The scale is unnormalized; dividing every value by 3 would normalize its best outcome without changing preferences.
For fixed y, rewrite it as U(x,y) = (1 + y)x + y. A lottery over x has expected utility (1 + y)E[x] + y. Since 1 + y is positive throughout the domain, its ranking against another x-lottery depends only on their expected x values, not the fixed y. The same argument applies symmetrically to lotteries over y with x fixed. Thus this two-attribute example satisfies mutual utility independence.
Nevertheless, the interaction matters when both attributes are uncertain. Lottery A gives (0,0) or (1,1), each with probability 0.5. Its expected utility is 0.5 × 0 + 0.5 × 3 = 1.5. Lottery B gives (0,1) or (1,0), each with probability 0.5, so its expected utility is 1. Both lotteries have the same marginal distributions for x and y, yet they are not indifferent. Mutual utility independence therefore does not justify discarding the joint outcome structure or assuming additive expected utility.

Illustrative case study

In a fictional training-service design, an analyst asks staff to compare uncertain depth-of-content outcomes while delivery support is held fixed. Staff then repeat the comparisons with a different common support level. Their risk preferences change: a variable-depth prospect is acceptable with intensive support but not with minimal support.
The team records that the tested utility-independence assumption fails for those outcomes. It gathers more targeted preference information instead of forcing the responses into a convenient separable model. The case illustrates how a failed assumption guides further elicitation; it does not report actual training outcomes or a universal human preference pattern.

Common misconceptions and mistakes

Mutual utility independence is not probabilistic independence. It concerns which lotteries a decision maker prefers when other attributes are fixed, not whether attributes occur independently in a dataset. The worked lotteries deliberately show why their joint dependence can still affect expected utility.
Do not infer the condition from a few certain-outcome rankings. Utility independence includes attitudes to uncertainty, so a value function that ranks certain outcomes may lack the information needed. Also, utility independence in one direction need not establish the mutual condition. State exactly which subset-complement comparisons have been assessed.

Business use

For organizations, the concept helps balance model simplicity against faithful representation of trade-offs under uncertainty. If appropriate independence conditions hold, preference assessment can be more manageable. If they fail, a simplified model may recommend actions that do not reflect the decision owner's actual priorities.
Document whose preferences are represented, the attribute ranges, the assessed lotteries, and remaining assumptions. Separate uncertainty estimates from utility assessments. When recommendations are sensitive to an untested interaction, highlight that dependency and test it directly. An elegant mathematical representation is useful only when its meaning and scope are clear.

Check your understanding

For the worked utility, compare an x-lottery with mean 0.6 against one with mean 0.4 when y is fixed at 0.5.

Answer: Their expected utilities are 1.5 × 0.6 + 0.5 = 1.4 and 1.5 × 0.4 + 0.5 = 1.1. The first is preferred, and changing fixed y within the domain does not reverse that ordering.

Related terms

Utility independence
Multiattribute utility theory
Expected utility
lottery
preferential independence
interaction term
positive affine transformation
additive independence

Learn next

Next, compare mutual utility independence with additive independence. Use the two worked joint lotteries, which share the same marginals, and explain why their different expected utilities rule out a purely additive representation for those lottery preferences. Keep deterministic outcome rankings separate from this stronger uncertainty question.

Utility independence

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, Chapter 16 slides, stochastic preference structure and utility independence: https://aima.cs.berkeley.edu/slides-pdf/chapter16.pdf . Relevant original-author sections were read. The two-attribute utility function and lottery comparisons are original, algebraically checked illustrations; no human preference data were collected.

Previous: Mutually Exclusive Events Next: n-best list