- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes a procedure to certify whether a given mixture-of-products distribution approximates a reference distribution better than every product distribution, using Wasserstein distance based on normalised Hamming distance. The reference distribution is the average of observation-dependent posteriors. Although each posterior treats the variables as independent, dependencies can arise when the posteriors are averaged. The method compares a lower bound on the smallest error achievable by any product distribution with an upper bound on the proposed mixture’s error. Improvement is certified if the upper bound falls below the lower bound.
Critical strengths
- The paper develops lower and upper bounds that support certification against the entire family of product distributions, rather than comparison with one fitted baseline. The lower bound includes an approximation guarantee, while the transport hierarchy provides progressively tighter upper bounds.
- The authors account for the availability of only finite data. Separate fitting and independent scoring, together with statistical corrections, turn the mathematical bounds into a certification procedure with a stated confidence guarantee.
Critical weaknesses
- The manuscript does not appear fully finished. The title presents two title-like statements which are both different to the one on OpenReview. Moreover, the appendices repeatedly refer to earlier versions: “v10” and “V0” in Appendix C, withdrawn quantities and an old bias bound in Appendix I.2, and an incompletely reproduced historical summary in Appendix J.4. I’m concerned that this submission has not been fully finished, or an incorrect version has been submitted.
- The paper is not sufficiently accessible. It is technically very dense and requires significant reconstruction using the appendix. The main paper should be self-sufficient, even if detailed proofs remain in the appendices. It would benefit from clearer motivation, definitions explained in words, connections between the constructions, and explanations of the main proof ideas.
- The motivating application is underexplored. While the authors acknowledge that the method has not been tried on RSSMs, it is unclear how it would be applied in practice. Access to categorical posterior probabilities appears plausible, but obtaining independent future observations conditional on a fixed recurrent state, stochastic state and action seems less straightforward. This leaves a substantial gap between the mathematical assumptions and practical applicability.
- The experiments leave an important gap between improvements in bounds and improvements in certification. They demonstrate that the new constructions, including prefix coupling and stopping trees, can tighten upper bounds. However, they do not establish whether these constructions produce successful certificates where the baseline fails. Computational costs also need further evaluation.
- The experiments are insufficiently explained in the main text. Table 1 uses “Strong,” “Soft” and “Independent” conditions whose definitions are deferred to the appendix. The basic definitions needed to interpret the results should accompany them.
- I was left unsure of the broader implications of the work. The RSSM motivation is given, but the practical benefit of the guarantee is not adequately explored. Other potential applications also receive little discussion. In addition, a more direct comparison with related proof constructions would help clarify the methodological advances over prior work.
Questions for authors
- How would the required independent scoring data be collected for the RSSM application?
- Do the new coupling constructions provide additional successful certifications that the baseline does not?
- What are the runtime and memory costs of the certification methods, particularly the new coupling constructions?
- Which proof constructions constitute the main advances over the closest prior work?
- What practical benefits are expected in the RSSM setting, and are there other applications where the framework would be useful?
3 comments
Thanks for your response.
I’m afraid it doesn’t resolve my main concerns. For example, your Q1 answer still leaves me unsure how you’d apply this to real RSSMs. And on readability, the response leaves it ambiguous how you’d actually change the current manuscript.
Thank you for clarifying what remained unclear. We have uploaded the revised manuscript. Applying the audit to a trained RSSM requires a way to draw independent next observations under the condition being audited. Section 8.2 (p. 9) now explains how we do this in our four-device simulator. We fix the observed history, the announced device availability, and the next inspection action. Using the known environment dynamics, we average over the hidden physical states consistent with that information to obtain the distribution of the next sensor reading. We draw 65,536 independent readings from this distribution and use the frozen RSSM to compute their posterior probabilities at the same predictive state. Appendix L.2.2 (p. 76) gives the full protocol. For a practitioner with only a trained model and ordinary replay data, this conditional sampling access remains an additional requirement. We have also revised the opening of Section 1 (p. 1) to explain how blocks that are independent for each individual observation can become dependent when we average over observations. Before introducing the formal constructions, we now explain in words what L and U bound and why U < L establishes that the specified candidate improves on every product distribution.
Thank you for your constructive comments. We acknowledge the presentation and version-history issues and aim to upload a clearer revision as soon as possible. Below we briefly address the main questions. The new experimental results have not yet been incorporated into the current manuscript.
Q1. How are independent scoring observations obtained for the RSSM application?
The required conditional samples are not supplied automatically by standard RSSM training. In our controlled experiments, the known environmental transition and observation laws allow us to compute a predictive observation distribution conditional on a fixed public history, action, and declared availability information. We draw independent scoring observations from this distribution and evaluate their posterior probabilities using the frozen RSSM. Thus, the environment supplies the conditional observations, while the RSSM supplies the posterior probabilities. We will explicitly state this additional sampling-access assumption and the refined conditioning used in the experiments.
Q2. Do the new couplings provide additional successes?
Yes. For G = 16 and n = 65,536, the strong baseline passes 0/20 independent scoring replicates, while short-prefix and stopping-tree couplings each pass 20/20. The comparisons use the same reference distribution, candidate, lower-bound witnesses, scoring data, and statistical budget. We also obtained additional numerical passes in trained-RSSM development experiments.
Q3. What are the runtime and memory costs?
For the G = 16 synthetic setting, the median computation times for prefix and stopping-tree couplings are 0.00976 and 0.01951 seconds, respectively. Their peak process memory usage is 72.3 and 73.1 MiB, respectively, including Python and library overhead. These measurements concern coupling computation on compressed finite-support inputs and exclude training and posterior extraction. We will provide the full cost tables and clarify their scope.
Q4. Which proof constructions constitute the main advances?
Our contributions include four components: the 2e + tau approximation guarantee for the local-compression lower bound; a coupling hierarchy combining prefix aggregation and state lifting, which preserves the full marginal distributions, guarantees monotonic tightening of the upper bound, and reaches the exact transport distance at full expansion; a stopping-tree construction that retains parent nodes and jointly optimizes transport and stopping masses, together with its integration into the hierarchy; and a common statistical event permitting selection among valid couplings for a fixed audit problem. We will explain these constructions more explicitly in the revision.
Q5. What are the practical benefits and other potential applications?
Where suitable conditional sampling access is available, a potential use during training is to screen prior structures at frozen checkpoints. When U < L, even the best product distribution cannot match the specified mixture candidate's approximation quality under the audited representation, reference distribution, and condition. The certificate therefore identifies an approximation limitation that further parameter optimization within the product family cannot eliminate. This can inform when to explore a richer prior and potentially avoid additional training effort within a family that cannot attain the target approximation quality under those fixed conditions.
The additional numerical passes show that tighter couplings make this structural diagnosis possible in cases where the original strong baseline remains inconclusive. Beyond RSSMs, the same comparison could support choices between factorized and mixture approximations in other discrete latent-variable models satisfying the same probability-access and independent-scoring assumptions.