Assembling the evidence…
Estimating the dimensionality of the latent representation needed for prediction -- the task-relevant dimension -- is a difficult, largely unsolved problem with broad scientific applications. We cast it as an Information Bottleneck question: what embedding bottleneck dimension is sufficient to compress predictor and predicted views while preserving their mutual information (MI). This repurposes neural MI estimators for dimensionality estimation. We show that standard neural estimators with separable/bilinear critics systematically inflate the inferred dimension, and we address this by introducing a hybrid critic that retains an explicit dimensional bottleneck while allowing flexible nonlinear cross-view interactions, thereby preserving the latent geometry. We further propose a one-shot protocol that reads off the effective dimension from a single over-parameterized hybrid model, without sweeping over bottleneck sizes. We validate the approach on synthetic problems with known task-relevant dimension. We extend the approach to intrinsic dimensionality by constructing paired views of a single dataset, enabling comparison with classical geometric dimension estimators. In noisy regimes where those estimators degrade, our approach remains reliable. Finally, we demonstrate the utility of the method on multiple physics datasets.
standard neural estimators with separable/bilinear critics systematically inflate the inferred dimension
analytic Gaussian treatment and consistent controlled comparisons across multiple latent settings
In noisy regimes where those estimators degrade, our approach remains reliable
direct noise comparisons support the direction but cover constructed paired-view regimes
the recovered effective dimensionality deff tracks the growth of correlated domains and exhibits a clear scaling collapse across system sizes
finite-size collapse is shown in one simulated physical system with theory-guided scaling
deff recovers the expected phase-space dimensions for the single (2) and double (4) pendulum
known dimensions are recovered with threshold sensitivity shown, though stopping differs from the general protocol
Derived from the full evaluation — not a separate score.
Strengths
Appendix A.2 derives why the Gaussian optimal critic contains terms a bilinear bottleneck cannot represent directly, while Figures 2 and 7 connect that mechanism to observed dimensionality inflation and critic behavior.
↳ Appendix A.2; Figures 2 and 7
The experiments vary latent distributions, linear and nonlinear observation maps, noise, batch size, sample size, and critic architecture, while reporting identifiable failure regimes rather than only successful cases.
↳ Figures 2–5 and 10–14
The Discussion explicitly states that estimates depend on MI estimation, view construction, architecture, and sufficient samples, and calls for formal success and failure conditions beyond the tested settings.
↳ Discussion, paragraph 3
Limitations
The Ising and pendulum studies test whether the estimator recovers theoretically expected scaling or known phase-space dimensions, not whether it resolves a previously unknown scientific quantity.
↳ Section 3.5; Figure 6
The general max-test rule is replaced in the pendulum analysis by the earliest epoch reaching a fraction of maximum test MI. Figures 16–17 bound sensitivity to that choice, but the change reduces procedural uniformity.
↳ Section 3.4; Appendix F.2; Figures 16–17
Architectures, optimization settings, and compute are reported in detail, but the supplied manuscript contains no explicit code or trained-model availability statement.
↳ Appendices E.2–E.3; Acknowledgments
The score is driven by the combination of an analytic Gaussian treatment in Appendix A.2 and controlled experiments across critic architectures, latent distributions, observation maps, noise levels, and sample sizes. Figures 2–5 provide strong support for the central mechanism and estimator under benchmark conditions, while Figures 10–14 extend the robustness analysis. The physics studies in Figure 6 add realistic data types but remain known-answer validations rather than demonstrations on an unresolved scientific problem. Appendix F.2 also changes the stopping rule for pendulum videos, with Figures 16–17 showing useful but not complete sensitivity control.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Strong3.9
Confidence highThe score reflects a clear methodological advance: identifying critic-induced dimensionality inflation, introducing a hybrid alternative, and adding a one-shot participation-ratio estimator. The advance is meaningful but extends an existing MI and symmetric-information-bottleneck research line rather than establishing a wholly new framework.
“separable/bilinear critics systematically inflate the inferred dimension”
Analytic derivation, known-dimension teacher problems, critic comparisons, noise and finite-sample tests, and detailed architecture and training specifications provide strong design-execution fit. The pendulum-specific stopping rule and absence of a stated code release reduce procedural uniformity and independent reproducibility.
“We therefore choose t* as the earliest epoch at which the test MI reaches a fixed fraction of its maximum.”
The argument progresses coherently from the critic-architecture problem through synthetic validation, finite-data procedures, and physics applications. Claims are generally scoped to tested settings, and the Discussion explicitly identifies dependencies on MI estimation, view construction, architecture, and sample size.
“It will therefore be important to formalize conditions under which it succeeds or fails beyond the settings explored here.”
The paper engages classical intrinsic-dimension estimation, neural MI estimation, CCA, and joint-embedding approaches while comparing several of them experimentally. Contextual breadth is tempered by limited applied benchmarking and relatively brief engagement with some recent MI-based dimensionality work.
“the estimate depends on the success of the underlying MI estimation, on view construction, on the choice of embedding-network architecture”
Lower confidence on Methodological Rigour — domain match limited.
Caveats3 of 4 checks
The pendulum application uses a different stopping rule from the general finite-data protocol. The departure is transparent and accompanied by threshold-robustness analyses, so it is a bounded consistency issue rather than evidence that the reported result is unreliable.
The computational and simulation-based study presents no applicable human- or animal-subjects issue, and no contradictory availability or conflict statement was identified. Code availability is not stated, but that absence is treated as a reproducibility limitation rather than a conduct concern.
Flags: 0 declared / 5 total
Reference integrity check could not run: references could not be reliably parsed from the manuscript; verification skipped rather than risk false verdicts.
0 references in manuscript
No retraction notice found in Retraction Watch.
Sources: Retraction Watch ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The method includes a finite-data procedure and is demonstrated on Ising configurations and pendulum videos. Both applications primarily validate recovery of known or theoretically constrained structure rather than deployment in an open scientific-analysis workflow.
AI-generated, human-governed. Something look off? Contact us to request a review.