- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper introduces FIDE (Fiber-Informed Differentiable Experimental Design), which is a framework to evaluate and rank multiple designs that are previously chosen to meet a prescribed scientific accuracy requirement. For each candidate design, the method uses a minimum relative entropy (I-projection) update to complete the design-induced partial information into a full probability law relative to a trusted reference law. The resulting law path is then evaluated using the proposed Full Action objective.
I understand such problems broadly from my work in Uncertainty Quantification (UQ). In particular, I am intimately familiar with I-projections and the work by Csiszar (cited by the authors).
FIDE evaluates a time-dependent law by computing the minimum additional probability-transport velocity needed to realize that law relative to the reference dynamics. The selected design is thus the scientifically adequate design that requires the least dynamical revision of the baseline model. The individual tools are largely familiar to the UQ community (e.g., I-projections and optimal transport). The primary novelty is the synthesis into a baseline-preserving secondary criterion for experimental design.
The paper develops several mathematical characterizations of the proposed objective (variational, transport, and dual formulations) and derives decomposition and differentiability results. Numerical experiments on multiple sensor-design problems illustrate how the proposed baseline-preserving criterion behaves in practice and how its rankings differ from those obtained using more conventual design objectives.
Critical strengths
- The mathematical development is thorough and technically sophisticated. The extensive appendices appear to provide a complete mathematical framework that is internally consistent and supports the proposed framework.
Critical weaknesses
- While the mathematical development is thorough and appears to be technically sound, the primary conceptual contribution appears relatively modest by comparison. From a UQ perspective, many of the core ingredients are well established. While the novelty appears to lie within how all these ingredients are utilized, and it seems entirely reasonable to do this, I am not convinced that the underlying design principle is sufficiently novel or impactful to justify the breadth of mathematical machinery that was brought to bear on this problem.
Questions for authors
No questions. I believe I understand the proposed framework and its motivation. My concerns primarily focus on the significance and impact of the underlying design principle rather than points requiring clarification.
Thank you for clarifying your assessment. We would like to sharpen the significance of the focused design problem.
Aggregate observations can leave an evolving population law underdetermined even when they suffice for a specified scientific task. Several measurement designs can satisfy the task while inducing reconstructed paths with different dynamical implications. Primary-task adequacy alone does not determine how to resolve this remaining choice.
FIDE supplies a reference-relative criterion for that choice when preserving an existing dynamical mechanism is an explicit objective. Its minimum-correction and dual characterizations explain what the criterion measures, while the Full–Tangent–Hidden decomposition identifies corrections omitted by measured moment rates.
The two downstream examples address complementary comparisons. The forecasting example separates Full from static W₂ and forward KL; Full and Tangent select the same design in that example. The occupancy study provides the comparison between Full and Tangent. Within each study, primary risk is identical across candidates, and the downstream loss is excluded from selection. These results show that the distinctions between the criteria can have scientific consequences within the stated settings.
We recognize that reference preservation is a conditional preference and that its usefulness depends on the reference. Our contribution does not resolve non-identifiability or establish reference correctness: instead it provides a mathematically characterized design principle for operating under those limitations.
We respectfully ask that the contribution’s significance be assessed through this residual design problem and its demonstrated consequences. Its focused scope limits the application claim, but we believe the connection between aggregate measurement design, law completion, and reference-relative dynamics constitutes useful methodological knowledge within the scope of OED.