acceptodds
Under review as a conference paper at ICLR 2027

Structural Limits of the Information-Theoretic Uncertainty Decomposition

Abstract

Uncertainty estimation in machine learning typically decomposes uncertainty into aleatoric uncertainty (AU) and epistemic uncertainty (EU) using the standard information-theoretic framework. However, in practice, two critical issues arise: entanglement (AU and EU are highly correlated) and epistemic collapse (EU magnitude shrinks with increasing model capacity). We analyze this framework on a functional level and discover that significant portions of the assumed AU, EU range are infeasible in finite settings, and cannot be attained with any class probabilities. We characterize how this infeasible region scales with the number of classes and Monte Carlo samples (e.g., from ensembles with members), revealing it is bounded by . Crucially, the infeasible region's boundary helps explain epistemic collapse: when model confidence is high, is guaranteed by this fundamental structural limitation. Our findings show that increasing ensemble size mitigates epistemic collapse by reducing the infeasible area. Lastly, we caution against interpreting AU and EU as independent quantities in low AU regimes, since we show they are coupled when .

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty, Aleatoric, Epistemic, Epistemic Collapse, Ensemble

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uuKGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work looks to critically examine the information-theory breakdown of total uncertainty into its aleatoric and epistemic components. The work reasons that the significant portions of the possible range for AU and EU are actually unreachable. The stated main contribution of the work is the derived bound that AU is ≤log⁡2/N\le \log{2}/N.

Critical strengths
  1. The greater topic is valuable and of interest to the community. It is an attempt to map out the more detailed nuances of the standard decomposition of uncertainty and an attempt to impede epistemic collapse (model overconfidence), which is important for OOD detection.
  2. The conjecture presented discusses the ways to attain minimum AU for a given TU, and is plausible.
Critical weaknesses
  1. While plausible, the paper does not describe the conjecture clearly, D is essentially the set of plausible ensemble votes where one member of the ensemble is uncertain, and D is also used to denote data earlier in the paper. The empirical backing with relatively simple datasets such as MNIST and CIFAR-10 are not sufficient in themselves to strongly back a claim.

  2. The figures and the author's message can be strengthened by explicitly displaying the areas which are infeasible/unreachable.

  3. Some observations presented as "theorems" are trivial. The contributions of this work then cannot be said to contain these "theorems".

    Theorem 1 is about the traditional bounds of TU and EU, which are apparent from the definition and well-known. This makes Figure 2 also mostly trivial.

    Theorem 2 is about the fact that at AU=0, the possible values of TU becomes discrete, because they require the ensemble members to have zero entropy, meaning every ensemble is a pure vote and their disagreement is then the only source of uncertainty.

    Theorem 3 is about the uncertain ensemble member among the items that D contains. Since D is the set of ensemble votes where one member is uncertain, the AU must be the entropy of that one member divided by N, which must be less than the maximum possible binary entropy divided by N. Again, this is trivial.

    Theorem 4 is the least trivial of the group though it is still quite simple. It describes the curves seen in earlier figures, specifically the lowest curve. It's claim about the first "bubble" drives the argument in L377 about ensuring AU > EU, but this is extremely vague and not discussed in the theorem at all.

I would recommend that the narrative is fully restructured in terms of presentation of these ideas. However, even if this is done, the contrubution for the work remains too minor at this point in time, with insufficient backing.

Questions for authors
  1. Are properly calibrated models still a concern when it comes to epistemic collapse?
Reviewer_KYgROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors study the canonical information-theoretic uncertainty decomposition of (e.g.) Houlsby et al., and show theoretically that not all AU, EU tuples are actually attainable. This fact is used to explain several known shortcomings of said decomposition: that AU and EU are heavily correlated, that EU vanishes, etc. The paper includes proofs of the binary case as well as simulations of the multi-class one.

Critical strengths

The paper is quite clearly written and easy to follow. More importantly, it introduces yet another reason why the bare information-theoretic uncertainty decomposition, which has been rapidly falling out of favor recently, is flawed. I don't think I've seen this particular geometric argument before.

Critical weaknesses

The fundamental issue with the paper in my eyes is that it does not do enough work to justify why the infeasibility result is even necessarily a problem. Since the information-theoretic decomposition derives both estimates from the same underlying model prediction, AU and EU are clearly going to be correlated in some way; you've laid out one specific way in which this is the case. But without a competing vision of what these two quantities should be or how they should behave, I as a reader am left wondering what to believe about this. Other critiques of the naked uncertainty distribution either try to propose different decompositions entirely ([1]) or imbue said decomposition with additional semantics ([2], [3]). The latter two depend directly on the decomposition, so your infeasibility results must also apply there, but it's very unclear to me that they represent a defect of the decomposition; in that context, they merely show the representational limits of a specific (very small) ensemble.

Some other issues:

  1. Several claims in the abstract far overstep the actual theoretical results. "Our findings show that increasing ensemble size mitigates epistemic collapse by reducing the infeasible area. -> the paper does not jointly vary model capacity and ensemble size, so it can't really justify any of its claims like this about epistemic collapse, which is specifically an observation about larger and larger models. "Crucially, the infeasible region’s boundary helps explain epistemic collapse: when model confidence is high, AU > EU is guaranteed by this fundamental structural limitation." -> similarly, an account of epistemic collapse must also justify why model confidence even increases (or that it does at all) as model size increases. This paper does not have the tools to do that.
  2. The proof only covers the binary case, and is not even about the entire unfeasible region.
  3. "For this, we track performance for out-of-distribution (OOD) detection as it allows us to assess the entanglement between AU and EU in practice. For OOD detection, we expect EU to be more informative than AU, and we therefore say that there is entanglement when AU performance exceeds EU performance." -> this entire section doesn't make much sense to me. In a perfect model with zero epistemic uncertainty, AU is also capable of distinguishing between different domains: some domains have much higher AU than others. In an imperfect model, maybe OOD inputs cause the model to veer into a part of its output space (e.g. an abstention or something) where there genuinely is just higher AU.

[1] @InProceedings{pmlr-v267-kirchhof25b, title = {Position: Uncertainty Quantification Needs Reassessment for Large Language Model Agents}, author = {Kirchhof, Michael and Kasneci, Gjergji and Kasneci, Enkelejda}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {81665--81677}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/kirchhof25b/kirchhof25b.pdf}, url = {https://proceedings.mlr.press/v267/kirchhof25b.html}, }

[2] @inproceedings{3692070.3692971, author = {Johnson, Daniel D. and Tarlow, Daniel and Duvenaud, David and Maddison, Chris J.}, title = {Experts don't cheat: learning what you don't know by predicting pairs}, year = {2024}, publisher = {JMLR.org}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, articleno = {901}, numpages = {59}, location = {Vienna, Austria}, series = {ICML'24} }

[3] @inproceedings{ ahdritz2025provable, title={Provable Uncertainty Decomposition via Higher-Order Calibration}, author={Gustaf Ahdritz and Aravind Gollakota and Parikshit Gopalan and Charlotte Peale and Udi Wieder}, booktitle={The Thirteenth International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=TId1SHe8JG} }

Questions for authors

See above.

Reviewer_BxpPOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the information theoretic split split of total uncertainty (TU) into aleatoric uncertainty (AU, mean entropy of N posterior samples) and epistemic uncertainty (EU=TU-AU, or the mutual information between uniformly sampled posterior sample index and the predicted classe). The paper starts from the observation that not every (AU,TU) is attainable. The papers proves simple elementary bounds on the quantities. Conjecture 1 states that the lower envelope of the (AU, TU) image of configurations with N - 1 deterministic samples and one sample on a simplex edge (the set D) is the infeasible boundary; the conjecture is proved when there are two classes. Theorem 3 shows that the boundary curves satisfy AU <= log(2)/N, and Theorem 4 identifies the "first bubble" as the lowest-TU boundary curve. The analysis is complemented by experiments which the EU fraction EU/TU increases with N, and that it decreases with model capacity.

Critical strengths
  1. If you are a practitioner using this information theoretic decomposition then this paper warns you to use it with caution.

  2. The conjecture for 2 classes is proved and seems correct.

  3. Theorem 3 and 4 gives useful information on the feasability region.

Critical weaknesses
  1. All results seems fairly easy and the proof of the conjecture for C=2 seems very special and not likely to extend (discussion of this is lacking).

  2. The experiments lack error bars and accuracy analysis.

  3. Maybe this is a personal bias, but I find the whole premise of the problem somewhat unconvincing. While I can understand the definition of AU (maybe), the identification of EU does not seem well grounded. And the whole paper makes exactly this point. From a conceptual point of view aleatoric and epistemic uncertainty should be rather independent.

Questions for authors
  1. Can you explain what are the obstacles in proving the conjecture for higher value C?

  2. Why does Figure 1 use 5 classes for CIFAR 10? How many classes are used in section 4.1 (figure 7) for CIFAR 10?

  3. can you provide error bars and accuracy for your experiment to separate the geometric effects from estimator bias?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.