Mutual Utility Independence (MUI)

• MUI is defined as the condition where every subset of attributes is utility independent of its complement, leading to simplified utility representations.
• Under MUI, utility functions can collapse into additive or multiplicative forms, effectively compressing the multiattribute utility structure.
• MUI serves as a benchmark in MAUT for comparing alternative independence notions, influencing computational methods and graphical utility models.

Searching arXiv for the specified papers and closely related work on Mutual Utility Independence. Mutual Utility Independence (MUI) is a classical condition in multiattribute utility theory (MAUT) under which every subset of attributes is utility independent of its complement. In the literature surveyed here, MUI is the strongest standard form of utility independence and is associated with highly compressed utility representations, notably additive or multiplicative single-attribute forms. At the same time, later work uses MUI as a point of contrast: additive and conditional additive independence are presented as structurally different notions with distinct computational consequences, and Shoham argues that classical utility independence, including MUI, is not the direct analogue of probabilistic independence that one would want for Bayesian-network-like utility reasoning (Bacchus et al., 2013, Shoham, 2013).

1. Classical definition in multiattribute utility theory

Let $V=\{v_1,\ldots,v_n\}$ be a finite attribute set, with state space

$S=\prod_{i=1}^n d_{v_i}.$

A utility function $u(V)$ induces a preference ordering on lotteries over $S$ by expected utility: $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$ For $X\subseteq V$, letting $Y=V-X$, and fixing an assignment $y$ to $Y$, the conditional preference over $X$ given $S=\prod_{i=1}^n d_{v_i}.$0, denoted $S=\prod_{i=1}^n d_{v_i}.$1, is defined by comparing lotteries on $S=\prod_{i=1}^n d_{v_i}.$2 after extending them to the full state space with probability $S=\prod_{i=1}^n d_{v_i}.$3 on $S=\prod_{i=1}^n d_{v_i}.$4 (Bacchus et al., 2013).

Utility independence is then the invariance of these conditional preferences across assignments to the complement. In the standard formulation, $S=\prod_{i=1}^n d_{v_i}.$5 is utility independent of $S=\prod_{i=1}^n d_{v_i}.$6 when conditional preferences for lotteries on $S=\prod_{i=1}^n d_{v_i}.$7 do not depend on the particular value assigned to $S=\prod_{i=1}^n d_{v_i}.$8. Shoham reviews the same classical notion in set-based notation: if $S=\prod_{i=1}^n d_{v_i}.$9 and $u(V)$0 are attribute sets, then $u(V)$1 is utility independent of $u(V)$2 when conditional preferences on lotteries on $u(V)$3 given $u(V)$4 do not depend on the particular value of $u(V)$5 (Shoham, 2013).

MUI is the global strengthening of this condition. The formulation given in the survey literature is that every subset of variables is independent of its complement. In equivalent notation,

$u(V)$6

Shoham gives the same condition as: attributes $u(V)$7 are mutually utility independent if every subset of $u(V)$8 is utility independent of its complement. He also distinguishes a weaker condition, singulary utility independence, under which every individual attribute $u(V)$9 is utility independent of its complement $S$0 (Bacchus et al., 2013, Shoham, 2013).

2. Representation theorems and hierarchy of strength

The immediate significance of utility independence lies in the functional restrictions it imposes on $S$1. For a single subset $S$2, utility independence of $S$3 from its complement is equivalent to the existence of a representation

$S$4

with $S$5. This is the basic structural theorem for ordinary utility independence (Bacchus et al., 2013).

For MUI, the representation collapses much further. The survey result states that every subset of variables is independent of its complement in $S$6 if and only if there exist single-variable functions $S$7 such that either

$S$8

for some constant $S$9, or

$$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$0

The literature described in the survey treats this as an extremely strong conclusion, because it reduces a general utility function on the product space to a representation built only from single-attribute functions (Bacchus et al., 2013).

The same section distinguishes full MUI from the weaker requirement that each individual variable be utility independent of the rest. Under that weaker condition, $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$1 is only a multilinear combination of single-variable functions $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$2, not necessarily additive or multiplicative. Shoham summarizes the representational hierarchy in compatible terms: singular utility independence leads to a multilinear form; MUI leads to a special form with $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$3 simple utility functions and $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$4 constants, though still possibly exponentially many operations; and additive independence leads to an additive form. He further remarks that the MUI form “always collapses to one of the following two special cases,” namely multiplicative or additive (Bacchus et al., 2013, Shoham, 2013).

3. Distinction from additive and conditional additive independence

MUI is not identical to additive independence, and the distinction is central to the comparative literature. Additive independence is defined over a partition $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$5 of $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$6: if two lotteries have the same marginals on each block $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$7, then they are indifferent. Its representation theorem is

$$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$8

Shoham states the same idea informally as preferences over lotteries depending only on marginal probability distributions and not on the joint probability distribution (Bacchus et al., 2013, Shoham, 2013).

The standard counterexample uses binary attributes $$P_1 \succeq_u P_2 \iff \sum_{s\in S} P_1(s)u(s)\ge \sum_{s\in S} P_2(s)u(s).$$9 for healthy and $X\subseteq V$0 for wealthy, with

$X\subseteq V$1

In this example, $X\subseteq V$2 is utility independent of $X\subseteq V$3, and $X\subseteq V$4 is utility independent of $X\subseteq V$5, in the classical sense. But the attributes are not additively independent. Shoham demonstrates this with two lotteries having the same marginals: $X\subseteq V$6 and

$X\subseteq V$7

for which

$X\subseteq V$8

After changing only $X\subseteq V$9 from $Y=V-X$0 to $Y=V-X$1, the function becomes additively independent and can be written as

$Y=V-X$2

with $Y=V-X$3. The example shows that classical utility independence can hold while additive independence fails (Shoham, 2013).

The survey literature then introduces conditional additive independence (CAI) for disjoint $Y=V-X$4 with $Y=V-X$5. The defining representation is

$Y=V-X$6

CAI is positioned as neither as strong as additive independence nor as generally applicable as utility independence, but it yields decompositions that are often more computationally useful than the product utility functions associated with MUI and the multilinear forms associated with weaker utility-independence assumptions (Bacchus et al., 2013).

4. Shoham’s reinterpretation of utility independence

Shoham’s contribution is not a refinement of classical MUI inside MAUT, but a reinterpretation of conditional utility and utility independence. He argues that the standard MAUT notion is an “objective” kind of conditioning, because one conditions on facts about the world, whereas the desired probabilistic analogue should be “subjective,” conditioning on utility-bearing factors or sources of utility (Shoham, 2013).

To formalize that reinterpretation, he introduces utility distributions as a special case of additive MAU functions. An additive MAU function has the form

$Y=V-X$7

Shoham then restricts attention to Boolean attributes $Y=V-X$8 and defines a TIOLI (“take it or leave it”) utility function by

$Y=V-X$9

A utility distribution is the normalized case in which

$y$0

and still

$y$1

Shoham explicitly states that utility distributions, whose structure is identical to that of probability distributions, can be viewed as a special case of additive multiattribute utility functions (Shoham, 2013).

Within this framework, subjective conditional utility is defined by

$y$2

and utility independence becomes

$y$3

Shoham’s explicit claim is that classical MAUT utility independence, including MUI, is not structurally analogous to probabilistic independence, whereas this factor-based notion is intended to be the direct analogue. A plausible implication is that the classical role often assigned to MUI in compact utility reasoning is being displaced rather than generalized in this framework (Shoham, 2013).

5. Graphical representation and computational role

The graphical consequences of these different independence notions are not uniform. In Shoham’s proposal, utility distributions and the new definition of utility independence motivate utility networks, intended to do for utilities what Bayesian networks do for probabilities. Their semantics are teleological rather than causal: a parent factor explains why a child factor or object is desired. Shoham presents this as the conceptual basis for compact graphical representation, but the paper does not provide a full $y$4-separation theorem or an explicit local factorization equation analogous to

$y$5

for utility networks (Shoham, 2013).

By contrast, the theory of conditional additive independence is supplied with a precise undirected-graph semantics. The survey paper proves that CAI always has a perfect representation as separation in an undirected graph $y$6: for $y$7,

$y$8

It also gives the associated clique decomposition theorem: $y$9 is a CA-independence map for $Y$0 if and only if $Y$1 has an additive decomposition over the maximal cliques of $Y$2. The computational motivation is explicit: additive decompositions permit expected utility calculations to exploit linearity of expectation, whereas MUI-induced product forms and multilinear forms need not yield the same simplifications (Bacchus et al., 2013).

6. Scope, limitations, and adjacent frameworks

The contemporary relevance of MUI is sharpened by comparison with frameworks that do not assume scalar MAUT forms at all. Multi-objective influence diagrams (MOIDs) use utility values that are vectors in $Y$3, ordered only partially, typically by Pareto dominance. The paper introducing MOIDs does not mention MUI, utility independence, preferential independence, or multilinear utility. Instead, it assumes an additively decomposable structure over local utility components, where each local utility function is vector-valued and the overall utility is their sum in $Y$4; user tradeoffs are incorporated as pairwise comparisons that induce a stronger dominance relation via convex cones (Marinescu et al., 2012).

This contrast is methodologically important. The MOID framework is best understood as avoiding the need for a single scalar utility aggregation rather than deriving one from MUI-like assumptions. In that setting, the central issues are partial orders, Pareto sets, $Y$5-coverings, and tradeoff-induced dominance, not the representation theorems associated with mutual utility independence (Marinescu et al., 2012).

Within the narrower domain of scalar MAUT, however, MUI remains a canonical benchmark. It marks the point at which ordinary utility-independence assertions become strong enough to force additive or multiplicative single-attribute structure. The later literature does not discard that fact, but it substantially reframes its significance: MUI is a powerful classical regularity condition, yet not necessarily the most useful notion for graphical representation, for efficient expected-utility computation, or for utility reasoning designed to parallel probabilistic conditional independence (Bacchus et al., 2013, Shoham, 2013).