Choquet Rank-Dependent Utility Model

• Choquet Rank-Dependent Utility is a decision model that fuses rank-based risk weighting with Choquet ambiguity handling to assess risky and ambiguous prospects.
• It separates pure risk, evaluated via rank-dependent utility on non-ambiguous events, from ambiguity captured by a nonadditive capacity aligned with a reference probability.
• The model is axiomatized through regularity, risk, and ambiguity conditions, offering robust applications in finance, insurance, and decision analysis.

The Choquet rank-dependent utility (CRDU) model is a decision model under ambiguity that combines a rank-dependent utility-type treatment of risk with a Choquet integral-type treatment of ambiguity. In a Savage setting with a distinguished sub-$\sigma$-algebra $\mathcal G$ of non-ambiguous events and a reference probability $\mathbb P$ on $\mathcal G$, it represents preferences through a utility function $u$, a probability distortion $g$, and generalized probabilistic beliefs $\nu$, according to

$$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$

Its defining novelty is that, in the absence of ambiguity, it reduces to rank-dependent utility (RDU) rather than expected utility (EU) (Oosten et al., 13 Sep 2025).

1. Formal representation and primitive components

The model is built on a separation between a risk component and an ambiguity component. The utility function $u:\mathbb R\to\mathbb R$ is a strictly increasing vNM utility function, unique up to positive affine transformations. The distortion function $g:[0,1]\to[0,1]$ is strictly increasing and continuous; it transforms probabilities before they enter the Choquet integral and captures the risk weighting part of preferences. The generalized probabilistic belief $\mathcal G$0 is a continuous capacity, not necessarily additive, and is required to be risk conforming: $\mathcal G$1 Thus $\mathcal G$2 agrees with the trusted reference probability on non-ambiguous events, while allowing nonadditivity outside that domain (Oosten et al., 13 Sep 2025).

The representation uses the composed capacity $\mathcal G$3, meaning the set function $\mathcal G$4. This gives the model a two-layer structure. The function $\mathcal G$5 values consequences, $\mathcal G$6 encodes rank-dependent probability weighting for pure risk, and $\mathcal G$7 encodes ambiguity through a nonadditive belief object. The paper identifies $\mathcal G$8 with the decision maker’s matching probability: for an event $\mathcal G$9, $\mathbb P$0 is the probability level of a non-ambiguous event that makes betting on $\mathbb P$1 indifferent to betting on that non-ambiguous event (Oosten et al., 13 Sep 2025).

A central implication of this architecture is that pure risk and ambiguity are not collapsed into a single nonlinear expectation. Acts measurable with respect to $\mathbb P$2 are treated as “pure risks” with no distributional ambiguity, while ambiguous acts are evaluated through the nonadditive capacity $\mathbb P$3. This decomposition is the model’s basic conceptual move.

2. Relation to CEU, RDU, and expected utility

CRDU extends Choquet expected utility (CEU) by inserting the distortion $\mathbb P$4 into the Choquet evaluation. CEU has the form

$\mathbb P$5

with a capacity $\mathbb P$6. CRDU instead evaluates acts via

$\mathbb P$7

so the ambiguity-belief capacity is first transformed by $\mathbb P$8, and only then integrated through the Choquet integral (Oosten et al., 13 Sep 2025).

On the pure-risk domain $\mathbb P$9, the model becomes

$\mathcal G$0

This is exactly the RDU form. By contrast, EU is recovered only in the special case $\mathcal G$1 and $\mathcal G$2. The distinction is therefore sharp: CEU typically reduces to EU on non-ambiguous acts, whereas CRDU is constructed so that the non-ambiguous benchmark is RDU, not EU (Oosten et al., 13 Sep 2025).

The Choquet component of CRDU is closely aligned with the CEU literature on comonotonicity. In a Savage-style framework with psychological gambles, $\mathcal G$3-coherence—coherence restricted to gambles whose supports consist of pairwise comonotonic acts—is equivalent to a convex capacity and a Choquet expected utility representation

$\mathcal G$4

This situates the Choquet part of CRDU within the established logic that rank-sensitive aggregation emerges from comonotonic structure rather than from additive probability alone (Cassese, 2023).

A common misconception is to treat CRDU as merely CEU with an extra parameter. The 2025 model rejects that interpretation. Its central contribution is precisely the separation between ambiguity, carried by $\mathcal G$5, and rank-dependent risk weighting, carried by $\mathcal G$6, so that pure-risk behavior is RDU behavior.

3. Axiomatization, uniqueness, and matching probability

The axiomatization is developed in a Savage setting with pure risk. The main regularity and risk axioms are (RC) Risk conformity, (M) Monotonicity, (SRM) Strict risk monotonicity on $\mathcal G$7, and (C) Continuity. The additional axioms are (RS) Risk symmetry, a weak independence-type axiom over coin-flip events, and (SCI) Subjective comonotonic independence. The key representation theorem states: $\mathcal G$8 Moreover, $\mathcal G$9 is unique up to positive affine transformations, $u$0 is unique, and $u$1 is unique and equals the $u$2-matching probability (Oosten et al., 13 Sep 2025).

The matching probability provides an event-level representation of ambiguity. For any event $u$3, there exists a non-ambiguous event $u$4 such that

$u$5

The paper also proves

$u$6

Thus $u$7 orders events directly and continuously, while remaining anchored to $u$8 on the trusted non-ambiguous source (Oosten et al., 13 Sep 2025).

Under the additional null set consistency axiom (NSC), the model admits an alternative representation

$u$9

Here $g$0 is an act-dependent distortion, $g$1 coincides with the common distortion $g$2 on non-ambiguous acts, and deviation of $g$3 from $g$4 measures ambiguity in act $g$5. If $g$6 for all $g$7, the model collapses to pure RDU under $g$8, which the paper identifies with ambiguity neutrality (Oosten et al., 13 Sep 2025).

The paper also axiomatizes the dual-utility subclass obtained by setting the utility linear, yielding

$g$9

This places Yaari-type dual utility inside the CRDU family as a special case (Oosten et al., 13 Sep 2025).

4. Ambiguity attitudes, neutrality, and robust representation

CRDU distinguishes several ambiguity attitudes through the generalized probabilistic belief $\nu$0. Ambiguity neutrality is defined through first-order stochastic dominance relative to some probability measure $\nu$1, and for CRDU it holds if and only if the matching probability $\nu$2 is actually a probability measure. In that case $\nu$3 is additive, and because $\nu$4, the model behaves like pure RDU on all acts (Oosten et al., 13 Sep 2025).

Comparative ambiguity aversion is represented directly by the matching probabilities. If $\nu$5 is more ambiguity averse than $\nu$6, then

$\nu$7

and the paper proves the characterization

$\nu$8

The model also discusses reference ambiguity aversion, equivalent to

$\nu$9

that is, the reference probability lies in the core of the capacity (Oosten et al., 13 Sep 2025).

The diversification property

$$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$0

holds for CRDU if and only if $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$1 is concave and $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$2 is supermodular. On pure risks, strong risk aversion corresponds to concavity of $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$3 and convexity of $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$4. The paper explicitly notes that, unlike in CEU, strong risk aversion plus ambiguity aversion does not collapse neatly into diversification seeking; the interaction between risk and ambiguity remains richer (Oosten et al., 13 Sep 2025).

The model also admits a robust interpretation analogous to maxmin expected utility: $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$5 This holds if and only if $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$6 is supermodular. The crucial difference from standard MEU is that the inner evaluation still uses the rank-dependent distortion $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$7, not plain expected utility (Oosten et al., 13 Sep 2025).

The broader Choquet ambiguity literature shows how these capacity-based ambiguity attitudes carry into strategic settings. In finite strategic-form games with Choquet expected utility, believed events are exactly unambiguous events of capacity $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$8, and ambiguity love or aversion leads to smaller or larger Choquet rationalizable sets of action profiles (Dominiak et al., 2019). This suggests a direct route for importing CRDU-style ambiguity primitives into epistemic game theory.

5. Foundational issues: comonotonicity, heterogeneity, and generalized rank dependence

The CRDU model belongs to a family of Choquet and rank-dependent representations whose technical foundations depend on how ranking and comparability are obtained. Classical Choquet and RDU axiomatizations typically rely on comonotonicity and constant acts. In a two-dimensional heterogeneous product set $$X \succsim Y \iff \int_{\Omega} u(X)\, d(g\circ \nu)\;\ge\;\int_{\Omega} u(Y)\, d(g\circ \nu).$$9, where the coordinates are not a priori commensurate, those tools are unavailable. An axiomatization on such a domain shows that a Choquet-integral representation can still be derived without assuming commensurateness of criteria a priori; the ranking step is recovered endogenously through cone-based structure and trade-off consistency (Timonin, 2015).

This issue matters because Choquet and rank-dependent models require a meaningful comparison of levels across dimensions. In the heterogeneous-product framework, the representation

$u:\mathbb R\to\mathbb R$0

is obtained even though comonotonicity cannot be defined in the usual way. The result shows that the ranking stage needed by Choquet/RDU models need not be imposed ex ante; it can be constructed from preference structure itself (Timonin, 2015).

A related lesson appears in decision analysis. The Choquet integral has been analyzed as a unifying aggregation operator whose special cases include MIN, MAX, order statistics, lattice polynomials, and convex-capacity maximin models. In that framework, the usual comonotonicity-based construction implies state-independence: if equal outcomes across states satisfy the comonotonic ordering structure assumed in the axioms, then all state utilities collapse to a single utility function. The same literature also emphasizes that learning only the capacity typically requires assuming that all dimensions are evaluated on a common scale, an assumption described as rarely justified in practice (Timonin, 2016).

The generalized rank-dependent function literature extends these issues further. For mappings of the form

$u:\mathbb R\to\mathbb R$1

which include expected utility, dual utility, rank-dependent utility, and signed Choquet functionals, probabilistic risk aversion—defined as quasi-convexity in probabilistic mixtures—is determined by the distortion function. The only quasi-convex cases are those with either a convex distortion or the class of scaled quantile-spread mixtures (Wang et al., 2022). This places CRDU within a broader distortion-based landscape in which the geometry of the weighting function remains decisive.

Another nearby model is the two-stage rank-dependent theory for decision under risk and ambiguity, represented by

$u:\mathbb R\to\mathbb R$2

That theory extends Quiggin’s RDU from risk only to risk plus ambiguity, reduces to variational preferences when $u:\mathbb R\to\mathbb R$3 is the identity, and is dual to variational preferences when $u:\mathbb R\to\mathbb R$4 is affine (Laeven et al., 2023). Relative to CRDU, it locates rank dependence inside a robust outer minimization over priors, rather than in a single matching-probability capacity.

6. Dynamic extensions and applications

The broader RDU and Choquet literature shows that probability weighting creates substantial dynamic and computational consequences. In a complete continuous-time market, RDU preferences are time inconsistent because the weighting of probabilities changes as conditioning information changes. For sophisticated consistent planners, equilibrium terminal wealth can still be derived explicitly and has the same form as in the classical Merton model, except that the market price of risk is scaled by a deterministic function determined by a highly nonlinear singular ordinary differential equation (Hu et al., 2020).

In incomplete markets, deterministic strict equilibrium strategies for CRRA investors with rank-dependent utility are characterized through singular ODEs for remaining variance exposure. With time-invariant weighting, there is at most one nonzero deterministic strict equilibrium strategy; with time-variant weighting, there may be infinitely many nonzero deterministic strict equilibrium strategies, parameterized by positive solutions of a nonlinear singular ODE (Wei et al., 2024). A related forward-performance development constructs rank-dependent predictable forward performance processes under conditionally complete markets with exogenous distortion functions updated periodically, reducing the update step to an integral equation solved via Volterra theory (Angoshtari et al., 2024).

Insurance provides another operational setting. In a Pareto optimal insurance problem with a rank-dependent utility insured, a risk-neutral insurer using the mean-variance premium principle, and incentive-compatible moral-hazard-free contracts, the original non-concave maximization problem involving Choquet expectation is transformed into a concave quantile optimization problem. The optimal contract is characterized by a second-order ordinary integro-differential equation with a nonlocal operator (Xu, 2021).

Robust optimization has also absorbed Choquet/RDU structure. For risk measures represented by Choquet integrals possibly induced by a probability weighting function and ambiguity sets defined by $u:\mathbb R\to\mathbb R$5-divergences, the robust optimization problem can be reformulated into a rank-independent problem for concave, convex, and inverse $u:\mathbb R\to\mathbb R$6-shaped weighting functions. In the concave case, the reformulation can be further turned into a convex optimization problem with explicit conic representability for canonical examples, and cutting-plane and piecewise-linear approximation algorithms yield tight upper and lower bounds and converge asymptotically (Jin et al., 17 Feb 2025).

Outside finance and insurance, Choquet-based rank-dependent aggregation has been used in path planning under uncertainty. With a capacity $u:\mathbb R\to\mathbb R$7 over scenarios and an increasing disutility $u:\mathbb R\to\mathbb R$8, the criterion

$u:\mathbb R\to\mathbb R$9

defines Choquet expected disutility over path-cost vectors. The resulting path problem is NP-hard, but exact heuristic search methods based on lower bounds from the core of the dual capacity were developed and shown to be practically efficient on tested instances (Galand et al., 2012).

Taken together, these developments position CRDU as a model at the intersection of Choquet ambiguity, rank-dependent probability distortion, and robust decision analysis. Its distinctive contribution is not merely a new representation formula, but a framework in which pure-risk behavior is rank-dependent, ambiguous events are represented by generalized probabilistic beliefs, and the interaction between the two remains explicit rather than absorbed into a single additive benchmark.