Conformal Prediction Regions are Imprecise Highest Density Regions

Eyke H\"ullermeier; Michele Caprio; Yusuf Sale

REVIEW 3 major objections 4 minor 1 cited by

Conformal Prediction Regions are Imprecise Highest Density Regions

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Conformal prediction regions coincide with highest-density regions

desk verdict Intended equivalence is correct, but Definition 4's exact-equality IHDR doesn't exist for most α, and Proposition 5's proof is wrong as written. read the letter →

arxiv 2502.06331 v2 pith:TQDP4GQC submitted 2025-02-10 stat.ML cs.LGmath.PR

Michele Caprio , Yusuf Sale , Eyke Hüllermeier This is my paper · ORCID

classification stat.MLcs.LGmath.PR MSC 68T3762M2060G2520M3215A80

keywords conformalpredictionimpreciseprobabilitiescredalsetshighestdensityregionsconsonanceplausibilityfunctionsclouds

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to establish that conformal prediction regions, the standard model-free prediction sets with uniform coverage guarantees, are exactly the Imprecise Highest Density Regions of a credal set built from the conformal transducer. This gives conformal prediction a complete imprecise-probability interpretation: the same set can be obtained by asking what the smallest region is that every distribution in the induced credal set regards as at least $1-\alpha$ probable. The equivalence matters because it transfers the robust, set-of-distributions semantics of imprecise probabilities onto conformal prediction, and it opens a new bridge between two previously separate literatures. The paper also records a new algebraic property of consonant plausibility functions and connects conformal prediction to the imprecise-probability concept of a cloud.

What carries the argument

The load-bearing object is the pair consisting of a consonant conformal transducer and the upper probability it induces. Consonance, $\sup_{\tilde y\in\mathbb{Y}}\pi(\tilde y,y_n)=1$, makes $\Pi_{y_n}(A)=\sup_{\tilde y\in A}\pi(\tilde y,y_n)$ a consonant plausibility function, and its credal set is $M(\Pi_{y_n})=\{P\in\Delta\mathbb{Y}: P(A)\le \Pi_{y_n}(A)\ \forall A\in\Sigma_{\mathbb{Y}}\}$. An Imprecise Highest Density Region is the smallest set $A$ with $P[Y_{n+1}\in A]\ge 1-\alpha$ for every $P\in M(\Pi_{y_n})$. The proof uses the cloud $[\gamma,\pi]$ with $\gamma(y)=\pi(y)$ when $\pi(y)\le 1/2$ and $1-\pi(y)$ otherwise, together with Neumaier's probabilistic constraint on clouds, to show that any IHDR must lie inside $R_\alpha(y_n)$; minimality then forces equality. The paper also records that a consonant plausibility function is a monoid homomorphism from $(\Sigma_{\mathbb{Y}},\cup)$ to $([0,1],\max)$, an algebraic fact it flags for future use.

What would settle it

For a finite grid of candidate values, a fixed dataset, and a fixed nonconformity measure satisfying consonance, compute $\pi(\cdot,y_n)$ and compare, for every $\alpha$, the set $\{y:\pi(y,y_n)>\alpha\}$ with the smallest set $A$ such that $\sup_{y\in A}\pi(y,y_n)\ge 1-\alpha$; any disagreement at any $\alpha$ refutes Proposition 5, as would a Monte Carlo violation of the coverage bound under an exchangeable process.

Watch

Extended reading notes

Core claim

The central claim is Proposition 5: once a conformal transducer $\pi$ is consonant, meaning $\sup_{\tilde y\in\mathbb{Y}}\pi(\tilde y,y_n)=1$ for every dataset, the $(1-\alpha)$-Imprecise Highest Density Region of the credal set $M(\Pi_{y_n})$ equals the classical conformal prediction region $R_\alpha(y_n)=\{y\in\mathbb{Y}:\pi(y,y_n)>\alpha\}$ for every $\alpha\in[0,1]$ and every $n\in\mathbb{N}$. Corollary 6.1 then attaches the standard conformal guarantee, $P[Y_{n+1}\in \mathrm{IR}^M_\alpha]\ge 1-\alpha$, uniformly in $n$ and in every exchangeable distribution $P$ on $\mathbb{Y}$. In other words, the region conformal prediction draws can be obtained by a fully imprecise-probability procedure: build the consonant plausibility function $\Pi_{y_n}(A)=\sup_{y\in A}\pi(y,y_n)$, form the closed convex set of probabilities it dominates, and extract the smallest set to which every distribution in that set assigns at least $1-\alpha$ probability.

Load-bearing premise

The equality rests on consonance, meaning every data set yields a conformal transducer with supremum 1 over the output space; the coverage guarantee also assumes exchangeability and, as stated in Section 6, that the true distribution's support is the whole output space.

Editorial extensions

If this is right

For a fixed nonconformity measure satisfying consonance, the conformal prediction region and the imprecise highest density region are the same set for every significance level and sample size, so interpretations and guarantees transfer in both directions.
The IHDR inherits conformal validity: $P[Y_{n+1}\in \mathrm{IR}^M_\alpha]\ge 1-\alpha$ holds uniformly over $n$ and over all exchangeable data-generating distributions, meaning the robust-Bayesian route loses no coverage.
Because the equality holds for every $\alpha$, the entire conformal transducer contour $\pi(\cdot,y_n)$ acts as a plausibility contour, so conformal prediction can be presented as a model-free possibilistic inference method rather than only a frequentist one.
Tuning the non-conformity measure is still essential: Proposition 7 shows that for any $\Psi$ there is another measure $\Psi'$ whose conformal region is strictly smaller while preserving the same uniform guarantee, so the IHDR and CPR are never minimal across all scores.
The monoid-homomorphism property suggests that consonant plausibility functions, and hence consonant conformal transducers, can be combined by tropical addition or maximum while staying in the same class, an algebraic handle the paper points to as future work.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If consonance fails, the equality as stated no longer holds; comparing $\{y:\pi(y,y_n)>\alpha\}$ with the IHDR of the raw credal set gives a quantitative measure of how much the consonance distortion shifts the region, and the adjusted transducers $\pi'$ and $\pi''$ show the mismatch is resolvable only by changing the region.
The paper's closed-world assumption, that the true distribution's support is the whole space $\mathbb{Y}$, is logically separate from the equality; a testable extension is to re-derive the coverage guarantee under a stated strict-support alternative and see exactly where the uniform-in-$P$ claim needs qualification.
The cloud proof suggests a direct split-conformal analogue: replace the transductive transducer by an inductive one and, as long as consonance is enforced, the same equivalence should reproduce the split conformal region; this is not shown in the paper but is implied by the structure of the argument.
The monoid homomorphism opens a composition rule not explored here: combining two consonant plausibility functions by pointwise maximum yields another consonant plausibility function, so one could pool conformal scores from different nonconformity measures and ask whether the resulting region remains valid.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper is an honest, small contribution: it shows that if you build a credal set from a consonant conformal transducer, the imprecise highest density region derived from that credal set is the same as the conformal prediction region, with the usual coverage guarantee. The authors openly say the key proposition is an immediate consequence of Couso et al. (2001, Theorem 2), so the novelty is the cloud-based proof route, the monoid-homomorphism observation (which is maxitivity in a new hat), and the framing. That is fine, but it is translation, not a new theorem.

The intended theorem is true, but the paper states it incorrectly. Definition 4 requires the lower probability of the IHDR to be exactly 1−α. The conformal transducer from Algorithm 1 takes values in {k/(n+1)}, so for any α that is not a grid value, no set has lower probability exactly 1−α. The object IR^M_α then does not exist, and Proposition 5's equality is undefined. The paper's own equation (8) and the alternative proof in (11) use the ≥ version, so the fix is clear: replace equality with ≥ in Definition 4 (or in Proposition 5, restrict α to the grid). The stress-test note is correct, and the paper needs this correction before publication.

The written proof of Proposition 5 is also wrong: from P(R_α) ≥ 1−α = P(IR^M_α) they claim monotonicity gives IR^M_α ⊆ R_α. Monotonicity goes the other way, and the inference doesn't follow. The alternative proof in (11) works only if the intersection is over sets with lower probability ≥ 1−α, i.e. Π(A^c) ≤ α, not Π(A) ≥ 1−α as printed. There is a notation ambiguity between Π and Π that should be cleared up.

Proposition 6 as stated is false: the "if and only if" claims P ∈ M(Π) is equivalent to P({π > α}) ≥ 1−α. The reverse direction fails for probabilities outside the credal set that still put mass on that set. The corollary's conclusion is correct, but the proof via Proposition 6 is not.

The paper is theoretically clean, clearly structured, and honest about limitations (Section 5.2 discussion, closed-world assumption in Section 6). No code, data, or experiments. It is for readers working at the CP-IP interface; practical CP users will not find much new.

It deserves a serious referee because the core claim is correct and supported by external published work, but the main proposition's statement and proof need revision. I'd bring it to group discussion; I wouldn't cite it in the next year.

Best.

Referee Report

3 major / 4 minor

Summary. The paper claims that, under the consonance assumption on the conformal transducer, the Imprecise Highest Density Region (IHDR) derived from the credal set M(Π_yn) induced by the transducer equals the classical conformal prediction region R_α(y_n) = {y : π(y,y_n) > α}, and that this region retains the standard uniform coverage guarantee. The main result is Proposition 5, with a coverage corollary (Corollary 6.1). The paper also introduces a cloud-based proof strategy, discusses a monoid-homomorphism property of consonant plausibility functions, and considers the effect of the nonconformity measure and the closed-world assumption.

Significance. The conceptual bridge between conformal prediction and imprecise probability is interesting and potentially useful: if stated with the correct '≥' formulation of an IHDR, the result gives a clean interpretation of conformal prediction regions as the smallest sets that every distribution in the consonant credal set deems at least (1−α) probable. The cloud-based argument and the uniform coverage statement are valuable. However, the current formulation has definitional and proof issues that must be corrected before the result is reliable.

major comments (3)

[Section 4, Definition 4 and Section 5] The exact equality in Definition 4 makes the object IR^M_α ill-posed for most α in the conformal setting. The conformal transducer returned by Algorithm 1 takes values in {k/(n+1) : k = 1, …, n+1}; for any α not equal to one of these values, no subset A satisfies P_yn(A) = 1−α (equivalently, no A has sup_{y∉A} π(y) = α), so IR^M_α does not exist and the equality in Proposition 5 is undefined. The paper's own consequence (8) and equation (11) rely on the '≥' version. The theorem should be restated with the IHDR defined as the minimal set A with P(A) ≥ 1−α for all P ∈ M(Π_yn); under that corrected definition, R_α(y_n) = {y : π(y) > α} is indeed the unique minimal set, and the proposition is true.
[Section 5, Proposition 5, equations (10)–(11)] The proof of Proposition 5 does not establish the claimed inclusion. Equation (10) derives Π_yn(R_α) ≥ Π_yn(IR^M_α); from monotonicity of upper probabilities one can only conclude R_α ⊆ IR^M_α if the inequality goes the other way, which is not the case. The sentence 'By the monotonicity of lower probabilities ... IR^M_α ⊆ R_α' is logically backwards. Equation (11) is also stated with the upper probability Π_yn; the intersection of all A with Π_yn(A) ≥ 1−α is generally much smaller than R_α (for three points with distinct π values it is empty). The correct identity uses the lower probability: R_α = ⋂{A : P_yn(A) ≥ 1−α} = {π > α}. The proof should be rewritten around this identity or via Couso et al. (2001) under the corrected definition.
[Section 5, Proposition 6] The statement of Proposition 6 is ambiguous because of quantifier placement. If it is read as a pointwise equivalence 'for each α, P ∈ M(Π_yn) iff P(R_α) ≥ 1−α', it is false: a probability measure can satisfy the inequality for a single α while violating set-wise dominance by Π_yn. If the intended statement is 'P ∈ M(Π_yn) iff for all α ∈ (0,1], P(R_α) ≥ 1−α', then it is correct, and the proof should make this quantifier explicit. As written, the proposition is not a valid characterization of M(Π_yn).

minor comments (4)

[Section 5] The notation is inconsistent: Definition 4 defines an IHDR via the lower probability P, but Section 5 writes Π_yn(IR^M_α) = 1−α for the upper probability. This should be reconciled.
[Section 2.2, Lemma 1] Lemma 1 is immediate from the defining property of a consonant plausibility function; the claim that it is a 'new algebraic property' and the suggested connection to Algebraic Statistics are not developed further in the paper.
[Section 6] The closed-world assumption (support of the true distribution equals Y) should be reconciled with the 'uniformly in P' statement of Corollary 6.1; the coverage hold for all exchangeable P, but the equality in Proposition 5 requires consonance for each data set, and the text does not clarify how the support condition interacts with this requirement.
[Section 5.2, Table 1] The table appears garbled in the provided text; the columns for upper and lower probability should be clearly labeled and separated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main equivalence is a direct mathematical identity sharing the conformal transducer, with external load-bearing references; the flagged Definition 4/Proposition 5 tension is a well-posedness gap, not a circular reduction.

full rationale

The derivation chain for Proposition 5 is not circular. The paper takes the conformal transducer pi(., y_n) from Algorithm 1 and Vovk et al. (2005) as input, defines the consonant upper probability Pi_{y_n}(A) = sup_{y~ in A} pi(y~, y_n) in Section 3, Eq. (7), forms the credal set M(Pi_{y_n}), and then observes that the strong alpha-cut {y : pi(y, y_n) > alpha} is the smallest set whose lower probability is at least 1 - alpha. This identity is unpacked in Eq. (11): R_alpha(y_n) = cap{A : Pi_{y_n}(A) >= 1 - alpha}. Each load-bearing ingredient, namely the conformal coverage theorem (Vovk et al. 2005, Theorem 11.1), the cloud constraint (Neumaier 2002; Augustin et al. 2014), and the consonant-plausibility representation (Couso et al. 2001; Augustin et al. 2014), is an external, independently published result. Cella and Martin (2022, 2021) are also external to the present author set, and consonance is stated explicitly as an assumption rather than smuggled in. The authors do cite their own prior work, but only for background, examples, and future directions, not for the central equality. Separately, the proof has a well-posedness gap that is not circular: Definition 4 requires exact equality Pi_{y_n}(IR^M_alpha) = 1 - alpha, while the proof and Eq. (11) use the >= formulation; for the grid-valued transducer produced by Algorithm 1, exact equality generally fails for alpha not equal to k/(n+1), so Proposition 5 as literally stated is ill-posed. This is a correctness concern, not a circular reduction of the conclusion to the premises.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on exchangeability for its coverage guarantee, on consonance to define the credal set, and on a closed-world support condition. Consonance is the most fragile premise: it is an assumption on the transducer, not on the data, and enforcing it may change the prediction region. No free parameters are fitted; the only user input is the significance level alpha. No new entities are introduced.

assumptions (4)

domain assumption Exchangeability of the data process Y1, Y2, ...
Section 2.1 assumes the process is exchangeable; this is necessary for the conformal coverage guarantee in Vovk et al. (2005, Theorem 11.1) and hence for Corollary 6.1.
ad hoc to paper Consonance of the conformal transducer: sup_{y~ in Y} pi(y~, y_n) = 1
Equation (6) in Section 3. The entire credal set construction and the equality IHDR = CPR depend on it. The paper notes it can be enforced by modifying the transducer, which changes the prediction region.
domain assumption Closed world assumption: support of the true data generating process equals Y
Stated in Section 6 as a tacit assumption. If the true support is smaller, the uniform-over-all-exchangeable-P coverage statement may need qualification.
domain assumption Existence and attainability of the minimum-measure (1-alpha)-IHDR
Definition 4 assumes an IHDR exists. For a consonant plausibility the strong alpha-cut attains the minimum, so the assumption is satisfied by the cited result (Couso et al. 2001), but it is a premise of the definition.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Conformal Prediction Regions are Imprecise Highest Density Regions." pith.science (2026). https://pith.science/paper/TQDP4GQC

@misc{pith2026250206331,
  author       = {Pith},
  title        = {Pith review of: Conformal Prediction Regions are Imprecise Highest Density Regions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQDP4GQC}},
  note         = {Machine review of arXiv:2502.06331}
}

read the original abstract

Recently, Cella and Martin proved how, under an assumption called consonance, a credal set (i.e. a closed and convex set of probabilities) can be derived from the conformal transducer associated with transductive conformal prediction. We show that the Imprecise Highest Density Region (IHDR) associated with such a credal set corresponds to the classical Conformal Prediction Region. In proving this result, we establish a new relationship between Conformal Prediction and Imprecise Probability (IP) theories, via the IP concept of a cloud. A byproduct of our presentation is the discovery that consonant plausibility functions are monoid homomorphisms, a new algebraic property of an IP tool.

Figures

Figures reproduced from arXiv: 2502.06331 by the authors.

**Figure 1.** Top: Our proposed, “indirect” methodology to derive a prediction region. We first use the consonant conformal transducer π to derive credal set M(Π), and then extract from the latter the IHDR IRα. Bottom: Classical CP methodology, in which the Conformal Prediction Region is obtained as in (1). We conclude with a discussion on the open problems and unanswered questions on the nature of CP, and on the relationship bet… view at source ↗

**Figure 2.** Visual representation of M(Πyn ) in our example. As we can see, it is “pushed” towards the boundary of the unit simplex. We also depicted p emp = (0.2, 0.3, 0.5)⊤ as a black dot. 6. Conclusion In the present work we study the conformal construction of credal sets. We show that the IHDR generated by the credal set induced by the conformal transducer π is equivalent to the classical CPR, and it retains the same unifor… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Quantification of Credal Uncertainty: A Distance-Based Approach
cs.AI 2026-03 accept novelty 6.0 of 10

IPM distances yield total, aleatoric (set-valued then endpoint-summarized), and epistemic (half-diameter) uncertainty measures for multiclass credal sets; TV recovers the binary Hüllermeier et al. decomposition with l...

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

J., and Moral, S

Abellán, J., Klir, G. J., and Moral, S. (2006). Disaggregated total uncertainty measure for credal sets. International Journal of General Systems, 1(35):29–44. Acharya, J., Daskalakis, C., and Kamath, G. (2015). Optimal testing for properties of distributions. Advances in Neural Information Processing Systems,

work page 2006
[2]

Caprio, M., Sale, Y., Hüllermeier, E., and Lee, I

World Scientific. Caprio, M., Sale, Y., Hüllermeier, E., and Lee, I. (2024b). A Novel Bayes’ Theorem for Upper Probabilities. In Cuzzolin, F. and Sultana, M., editors,Epistemic Uncertainty in Artificial Intelligence, pages 1–12, Cham. Springer Nature Switzerland. Caprio, M. and Seidenfeld, T. (2023). Constriction for sets of probabilities. InInternational...

work page 2024
[28]

N., Barber, R

Angelopoulos, A. N., Barber, R. F., and Bates, S. (2024). Theoretical foundations of confor- mal prediction. Augustin, T., Coolen, F. P., De Cooman, G., and Troffaes, M. C. (2014).Introduction to imprecise probabilities, volume

work page 2024
[29]

Walley, P.(1991).Statistical Reasoning with Imprecise Probabilities, volume42of Monographs on Statistics and Applied Probability

Springer. Walley, P.(1991).Statistical Reasoning with Imprecise Probabilities, volume42of Monographs on Statistics and Applied Probability. Chapman and Hall, London. Wimmer, L., Sale, Y., Hofman, P., Bischl, B., and Hüllermeier, E. (2023). Quantifying aleatoric and epistemic uncertainty in machine learning: Are conditional entropy and mutual information a...

work page 1991
[31]

J., and Candès, E

Gibbs, I., Cherian, J. J., and Candès, E. J. (2023). Conformal prediction with conditional guarantees. arXiv preprint arXiv:2305.12616. Gilboa, I. and Marinacci, M. (2016). Ambiguity and the bayesian paradigm. InReadings in formal epistemology: Sourcebook, pages 385–439. Springer. Gong, R. and Meng, X.-L. (2021). Judicious judgment meets unsettling updati...

arXiv 2023
[42]

Shafer, G

Princeton university press, Priceton, NJ. Shafer, G. and Vovk, V. (2008). A tutorial on conformal prediction.Journal of Machine Learning Research, 9(3). Sullivant, S. (2018).Algebraic Statistics. American Mathematical Society. Vasile, M., editor (2021). Optimization Under Uncertainty with Applications to Aerospace Engineering, volume 6 ofPhysics and Astro...

work page 2008
[219]

Zaffalon, M

Springer. Zaffalon, M. (2002). The naive credal classifier.Journal of Statistical Planning and Inference, 105(1):5–21. Imprecise Probability Models and their Applications. The University of Manchester, Oxford Road, Manchester, UK M13 9PL Email address: michele.caprio@manchester.ac.uk Ludwig-Maximilian University, Akademiestraße 7, Munich, Germany 80799 Em...

work page 2002
[244]

Destercke, S., Dubois, D., and Chojnacki, E. (2008). Unifying practical uncertainty rep- resentations – i: Generalized p-boxes. International Journal of Approximate Reasoning, 49(3):649–663. Dubois, D., Nguyen, H. T., and Prade, H. (2000).Possibility Theory, Probability and Fuzzy Sets Misunderstandings, Bridges and Gaps, pages 343–438. Springer US, Boston...

work page 2008

Show all 13 references

[540]

Denk, R., Kupper, M., and Nendel, M. (2020). A semigroup approach to nonlinear Lévy processes. Stochastic Processes and their Applications, 130(3):1616–1642. Denoeux, T. (2000). A neural network classifier based on Dempster-Shafer theory. IEEE Transactions on Systems, Man, and...

2020
[591]

Barber, R

John Wiley & Sons. Barber, R. F., Candes, E. J., Ramdas, A., and Tibshirani, R. J. (2023). Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845. Conformal Prediction Regions are Imprecise Highest Density Regions 17 Berger, J. (1984). The robust b...

2023
[2024]

and Martin, R

Cella, L. and Martin, R. (2021). Valid inferential models for prediction in supervised learning problems. In International Symposium on Imprecise Probability: Theories and Applications, pages 72–82. PMLR. Cella, L. and Martin, R. (2022). Validity, consonant plausibility measur...

2021
[2025]

Ellsberg, D. (1961). Risk, ambiguity, and the savage axioms. The quarterly journal of economics, 75(4):643–669. Gao, R., Xie, L., Xie, Y., and Xu, H. (2018). Robust hypothesis testing using wasserstein uncertainty sets.Advances in Neural Information Processing Systems,

1961
[9254]

18 Michele Caprio, Yusuf Sale, and Eyke Hüllermeier Couso, I., Montes, S., and Gil, P. (2001). The necessity of the strong α-cuts of a fuzzy set. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , 09(02):249–262. Cuzzolin, F. (2020). The geometry of ...

2001

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

J., and Moral, S

Abellán, J., Klir, G. J., and Moral, S. (2006). Disaggregated total uncertainty measure for credal sets. International Journal of General Systems, 1(35):29–44. Acharya, J., Daskalakis, C., and Kamath, G. (2015). Optimal testing for properties of distributions. Advances in Neural Information Processing Systems,

work page 2006

[2] [2]

Caprio, M., Sale, Y., Hüllermeier, E., and Lee, I

World Scientific. Caprio, M., Sale, Y., Hüllermeier, E., and Lee, I. (2024b). A Novel Bayes’ Theorem for Upper Probabilities. In Cuzzolin, F. and Sultana, M., editors,Epistemic Uncertainty in Artificial Intelligence, pages 1–12, Cham. Springer Nature Switzerland. Caprio, M. and Seidenfeld, T. (2023). Constriction for sets of probabilities. InInternational...

work page 2024

[3] [28]

N., Barber, R

Angelopoulos, A. N., Barber, R. F., and Bates, S. (2024). Theoretical foundations of confor- mal prediction. Augustin, T., Coolen, F. P., De Cooman, G., and Troffaes, M. C. (2014).Introduction to imprecise probabilities, volume

work page 2024

[4] [29]

Walley, P.(1991).Statistical Reasoning with Imprecise Probabilities, volume42of Monographs on Statistics and Applied Probability

Springer. Walley, P.(1991).Statistical Reasoning with Imprecise Probabilities, volume42of Monographs on Statistics and Applied Probability. Chapman and Hall, London. Wimmer, L., Sale, Y., Hofman, P., Bischl, B., and Hüllermeier, E. (2023). Quantifying aleatoric and epistemic uncertainty in machine learning: Are conditional entropy and mutual information a...

work page 1991

[5] [31]

J., and Candès, E

Gibbs, I., Cherian, J. J., and Candès, E. J. (2023). Conformal prediction with conditional guarantees. arXiv preprint arXiv:2305.12616. Gilboa, I. and Marinacci, M. (2016). Ambiguity and the bayesian paradigm. InReadings in formal epistemology: Sourcebook, pages 385–439. Springer. Gong, R. and Meng, X.-L. (2021). Judicious judgment meets unsettling updati...

arXiv 2023

[6] [42]

Shafer, G

Princeton university press, Priceton, NJ. Shafer, G. and Vovk, V. (2008). A tutorial on conformal prediction.Journal of Machine Learning Research, 9(3). Sullivant, S. (2018).Algebraic Statistics. American Mathematical Society. Vasile, M., editor (2021). Optimization Under Uncertainty with Applications to Aerospace Engineering, volume 6 ofPhysics and Astro...

work page 2008

[7] [219]

Zaffalon, M

Springer. Zaffalon, M. (2002). The naive credal classifier.Journal of Statistical Planning and Inference, 105(1):5–21. Imprecise Probability Models and their Applications. The University of Manchester, Oxford Road, Manchester, UK M13 9PL Email address: michele.caprio@manchester.ac.uk Ludwig-Maximilian University, Akademiestraße 7, Munich, Germany 80799 Em...

work page 2002

[8] [244]

Destercke, S., Dubois, D., and Chojnacki, E. (2008). Unifying practical uncertainty rep- resentations – i: Generalized p-boxes. International Journal of Approximate Reasoning, 49(3):649–663. Dubois, D., Nguyen, H. T., and Prade, H. (2000).Possibility Theory, Probability and Fuzzy Sets Misunderstandings, Bridges and Gaps, pages 343–438. Springer US, Boston...

work page 2008