- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper studies an AVI approach for the inference of the Fréchet variance. They utilize available ingredients in the literature to implement the approach (Algorithms 1 and 2), and later study their theoretical guarantees (Section 3). The article also includes an experimental section where they outperform the current fixed-time approximation of the Fréchet variance.
Critical strengths
- The models seems to outperform the current fixed-time approximation of the Fréchet variance in their experiments. If generalizable (as it appears, by their theoretical guarantees) it can be useful in some use-cases.
- The proofs seem mathematically correct upon a first screening, although they could be benefited by a light rewrite for clearness.
Critical weaknesses
- The paper prose is very unstructured, going to apparent proof's sketches after some results are introduced with too much detail, difficulting the reading experience.
- Over-complicated wording and complex mathematical formulations often replaces mathematical rigor and simple explanations/motivations. For example, (i) what boundaries are, why do they care about them and a simple visual representation of these boundaries are replaced by two unmotivated formulas in (1), and (ii) the following definition states a confidence band "on a possibly extended probability space, [if] positive reference radii satisfy (...)" without explaining what any of these two terms mean in their context. Note that these are two examples of a common writing habit in this article, making it very difficult to digest.
- There is no cohesive story in the paper: the motivation for any steps taken is unclear, and so are the challenges needed to overcome for the writing of this article. For example, in Section 2, the reason to study any of the terms listed as subsections is not properly addressed, making it seem too out-of-the-blue. This, in turns, makes it hard to understand the actual contributions this manuscript makes (not in terms of new theorems, but which challenges it overcame) and whether its contributions are incremental or substantial.
Questions for authors
Q1. What are these representative geometries in 2.4 and why do we care?
Q2. Could you consider adding more context to the methodology of this article?
Q3. What are the discussions after Theorem 1 and 2 about? Wouldn't they fit better on a remark, with a proper title stating what they are?
3 comments
I thank the author's for their response.
I've checked the revised manuscript---it still has the same prose and over-complicated wording mentioned in the review. Further from adding more clarity and a simpler tone to the wording, the revised version seems to further accentuate the weaknesses regarding its writing (W2), and do not fix the story. It is very hard to properly gauge the contribution of a paper, if it is unclear in its delivery.
Due to this, and after checking the other reviews and responses, I've decided to maintain my score.
Thank you for returning to the manuscript and giving us your candid assessment. The question behind the paper is how to track uncertainty in the spread of geometric data as new observations arrive. Because the centre of these data is also estimated, its error must be accounted for when the intervals are checked repeatedly. The paper studies this effect and develops both an offline procedure and an online procedure that processes the observations in one pass. We hope this statistical problem and the analysis addressing it retain value as we work to communicate them more clearly.
We respect your decision and appreciate your perspective on the narrative and readability. Your comments give us useful guidance for improving the exposition. Thank you for your time and the care you have taken to explain your concerns.
Thank you for your time and valuable comments.
W1. The revision separates method definitions, theorem statements, and interpretation. The paragraphs following Theorems 1 and 2 now have explicit remark titles. We retain the compact explanation of why offline profile is second order and remove the duplicate score/Hessian derivation from the main text; its complete appendix proof remains. The online discussion separately identifies predictable losses, centre tracking, and studentization, with references to the corresponding proofs.
W2. We now explain the boundary before introducing its formula. For a standard Gaussian partial sum Gn, a boundary bα(n) satisfies
Pr∣Gn∣≤bα(n) for every n≥1≥1−α.
Thus ±bα(n) are simultaneous envelopes for cumulative fluctuations. Dividing by n gives an average scale radius, and multiplying by the estimated log-scale standard deviation gives the reported radius snbα(n)/n.
The new deterministic table illustrates the distinction. For α=0.05, Gaussian mixture parameter ρ=1, and illustrative scale s=1:
These deterministic values evaluate equation (1).
An extended probability space supplies auxiliary Gaussian coupling variables used in the proof while preserving the observed data distribution. Writing
Tn−logV=τGn/n+en,
the reference radius is
hn∗=τbα(n)/n+∣en∣.
It absorbs the coupling and estimator errors and is generally unobservable. The proof establishes simultaneous reference coverage and hn/hn∗→1 almost surely. The main text explains the reference construction and the computed radius separately, with the growing start scope of the coverage guarantee.
W3. The methodology now begins with its organizing question: how can scalar sequential inference be transferred to a dispersion estimator whose centre is itself estimated? For Offline, we fit the centre creates a profiling remainder. For Online, we evaluate losses at the preceding centre creates an accumulated population excess and a martingale remainder. The analysis controls these terms, verifies scale consistency, and then transfers the scalar boundaries to the log-variance estimates. Exponentiation ensures positive variance scale intervals. This sequence makes the purpose of each component explicit and locates the contribution in the estimator specific transfer.
Q1. The representative geometries connect the general construction to concrete targets and updates. Euclidean space gives total coordinate variance and isolates the fitted centre versus predictable centre distinction without curvature. Affine invariant SPD matrices provide the nonlinear covariance example used in the application. The target remains scalar dispersion, and the online procedure retains the same two accumulated loss summaries. Additional geometric formulas remain in the appendix.
Q2. The added explanations in W2–W3 provide the requested context before the formulas: what is being estimated, why a simultaneous boundary is needed, and which additional errors arise from offline fitting or online tracking.
Q3. The discussions following Theorems 1 and 2 now appear under the titles “Why profiling is second order” and “Predictability, tracking and studentization.” The monitoring and Euclidean comparisons also have explicit remark headings. Formal proofs remain in the appendix.
Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.