Integral Imprecise Probability Metrics
Pith reviewed 2026-05-22 02:51 UTC · model grok-4.3
The pith
Integral imprecise probability metrics based on Choquet integrals serve as valid metrics for capacities and yield maximum mean imprecision measures for epistemic uncertainty.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The integral imprecise probability metric is defined using a Choquet integral with respect to a capacity and serves as a metric on the space of capacities when the associated kernel is characteristic. It metrises a form of weak convergence of capacities. By setting the two arguments to an imprecise probability model and its conjugate, one obtains the maximum mean imprecision, which acts as an epistemic uncertainty measure meeting the required axiomatic properties and showing strong results in selective classification.
What carries the argument
The Choquet integral applied to a kernel with respect to a capacity, which generalizes the expectation in integral probability metrics to the imprecise setting.
If this is right
- IIPM enables direct comparison between different imprecise probability models such as belief functions and probability intervals.
- Maximum mean imprecision provides an axiomatic epistemic uncertainty measure that can be computed for a single model.
- The metric supports studying convergence properties in sequences of capacities.
- It offers a scalable approach for uncertainty quantification in classification with large numbers of classes.
Where Pith is reading between the lines
- This metric could be integrated into kernel-based machine learning algorithms to account for imprecision in data or models.
- Extensions might include applications in decision theory under ambiguity where comparing beliefs is central.
- One could test the framework on real-world datasets with known sources of epistemic uncertainty like sensor data.
- It opens possibilities for robust statistics that use these distances for clustering or classification under imprecision.
Load-bearing premise
The Choquet integral construction preserves the metric axioms and the ability to metrize weak convergence for the class of capacities including lower probabilities, intervals, and belief functions.
What would settle it
A counterexample consisting of two different capacities with zero IIPM distance when using a characteristic kernel, or a sequence of capacities converging in the weak sense but with IIPM not approaching zero.
read the original abstract
Quantifying differences between probability distributions is fundamental to statistics and machine learning, primarily for comparing statistical uncertainty. In contrast, epistemic uncertainty -- due to incomplete knowledge -- requires richer representations than those offered by classical probability. Imprecise probability (IP) theory offers such models, capturing ambiguity and partial belief. This has driven growing interest in imprecise probabilistic machine learning (IPML), where inference and decision-making rely on broader uncertainty models -- highlighting the need for metrics beyond classical probability. This work introduces the integral imprecise probability metric framework, a Choquet integral-based generalisation of classical integral probability metrics to the setting of capacities -- a broad class of IP models encompassing many existing ones, including lower probabilities, probability intervals, belief functions, and more. Theoretically, we establish conditions under which IIPM serves as a valid metric and metrises a form of weak convergence of capacities. Practically, IIPM not only enables comparison across different IP models but also supports the quantification of epistemic uncertainty~(EU) within a single IP model. In particular, by comparing an IP model with its conjugate, IIPM gives rise to a new class of epistemic uncertainty measures -- Maximum Mean Imprecision -- which satisfy key axiomatic properties proposed in the uncertainty quantification literature. We validate MMI through selective classification experiments, demonstrating strong empirical performance against established EU measures, and outperforming them when classical methods struggle to scale to a large number of classes. Our work advances both theory and practice in Imprecise Probabilistic Machine Learning, offering a principled framework for comparing and quantifying epistemic uncertainty under imprecision.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces the Integral Imprecise Probability Metric (IIPM) framework as a Choquet-integral generalization of classical integral probability metrics to capacities (encompassing lower probabilities, probability intervals, belief functions and related models). It claims to derive conditions under which IIPM is a metric and metrises a form of weak convergence of capacities. From the same construction it derives Maximum Mean Imprecision (MMI) as an epistemic-uncertainty measure obtained by comparing a capacity to its conjugate; MMI is asserted to satisfy key axiomatic properties from the uncertainty-quantification literature. Empirical support is given via selective-classification experiments in which MMI outperforms established epistemic-uncertainty baselines, especially when the number of classes is large.
Significance. If the metric and convergence claims hold under the stated conditions, the work supplies a principled tool for comparing heterogeneous imprecise-probability models and for quantifying epistemic uncertainty inside a single model. The axiomatic verification of MMI and the reported scaling behaviour in high-class selective classification constitute concrete strengths. The overall significance is tempered by the breadth of the claimed applicability: the framework is only as useful as the range of standard IP models for which the separation and convergence properties are actually verified.
major comments (2)
- [Section on metric properties of IIPM (around the definition of IIPM and the subsequent theorem)] The central claim that IIPM is a metric (i.e., d(C,D)=0 iff C=D) for the broad family of capacities rests on the Choquet integral over the function class F. Because capacities are non-additive, it is not immediate that the integral separates capacities that differ only in their non-monotonic or non-convex parts; the manuscript should supply either an explicit proof that the chosen F is sufficiently rich or a counter-example showing when separation fails for common models such as belief functions.
- [Convergence theorem and its proof] The metrisation result for weak convergence of capacities must specify the precise topology and the exact subclass of capacities for which the equivalence holds. If the conditions exclude standard non-additive models or require F to be unrealistically large, the claimed generalisation from additive measures to capacities does not extend as stated.
minor comments (2)
- [Definition of MMI] Clarify the precise relationship between the function class F used for IIPM and the function class used to define MMI; any difference should be justified.
- [Experimental section] In the selective-classification experiments, report the number of classes, the precise baselines, and whether the performance gap persists after controlling for computational budget.
Simulated Author's Rebuttal
We thank the referee for their careful reading and constructive comments. We address each major comment below, indicating the revisions we will make to strengthen the manuscript.
read point-by-point responses
-
Referee: [Section on metric properties of IIPM (around the definition of IIPM and the subsequent theorem)] The central claim that IIPM is a metric (i.e., d(C,D)=0 iff C=D) for the broad family of capacities rests on the Choquet integral over the function class F. Because capacities are non-additive, it is not immediate that the integral separates capacities that differ only in their non-monotonic or non-convex parts; the manuscript should supply either an explicit proof that the chosen F is sufficiently rich or a counter-example showing when separation fails for common models such as belief functions.
Authors: We agree that an explicit justification of the separation property is warranted given the non-additivity of capacities. The function class F in the definition of IIPM is taken to be sufficiently rich (specifically, the set of all bounded continuous functions on the underlying space, or a dense subclass thereof) so that agreement of the Choquet integrals implies equality of the capacities. We will add a short lemma in the revised manuscript that proves d(C, D) = 0 implies C = D under this choice of F, using the fact that the Choquet integral with respect to a capacity is determined by the values on a separating family of functions. This argument applies directly to belief functions and other standard models, as they are special cases of capacities; no counter-example arises under the stated conditions. revision: yes
-
Referee: [Convergence theorem and its proof] The metrisation result for weak convergence of capacities must specify the precise topology and the exact subclass of capacities for which the equivalence holds. If the conditions exclude standard non-additive models or require F to be unrealistically large, the claimed generalisation from additive measures to capacities does not extend as stated.
Authors: We accept the need for greater precision in the statement of the convergence result. The theorem shows that IIPM metrizes a form of weak convergence of capacities, understood as convergence of the Choquet integrals against all functions in F, where F is the same separating class used for the metric property. This holds for the subclass of capacities that are continuous from below (which includes belief functions, lower probabilities, and probability intervals). We will revise the theorem statement to name the topology explicitly (the weak topology generated by integration against continuous bounded functions) and to restrict the claim to this subclass. The proof will be expanded with additional steps showing that the equivalence does not require an unrealistically large F. revision: yes
Circularity Check
No significant circularity in the IIPM derivation chain
full rationale
The paper defines the integral imprecise probability metric (IIPM) via Choquet integral over a function class on capacities and then proves conditions under which this construction yields a metric that metrizes a form of weak convergence. No equation reduces the claimed metric property or the Maximum Mean Imprecision measure to a fitted parameter or to a self-referential definition; the separation and convergence results are stated as consequences of the Choquet integral axioms plus richness assumptions on the function class, which are external to the target claims. Self-citations, if present, are not load-bearing for the central metric or uncertainty-measure results. The derivation therefore remains self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Capacities constitute a suitable encompassing class of imprecise probability models that includes lower probabilities, probability intervals, and belief functions.
invented entities (2)
-
Integral Imprecise Probability Metric (IIPM)
no independent evidence
-
Maximum Mean Imprecision (MMI)
no independent evidence
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.