FIDE: Fiber-Informed Differentiable Experimental Design for Reference-Relative Law Dynamics

Title

FIDE: Fiber-Informed Differentiable Experimental Design — Optimizing Dynamical Laws Under a Baseline-Preserving Setting

Meta Description

FIDE (Fiber-Informed Differentiable Experimental Design) introduces a conditional minimal-correction principle to resolve residual design degrees of freedom in law dynamics under aggregate measurements, optimizing sensor selection and experimental design in baseline-preserving scenarios.

Article Content

In scientific experiments, aggregate measurements often cannot uniquely determine the evolving population law. Different measurement systems may be equally suitable for a given scientific task, yet they constrain different aspects of the law. How should we choose under such residual design degrees of freedom? The paper "FIDE: Fiber-Informed Differentiable Experimental Design for Reference-Relative Law Dynamics" proposes a new method — Fiber-Informed Differentiable Experimental Design (FIDE) — offering a conditional minimal-correction principle for experimental design in baseline-preserving settings.

Background: Insufficiency of Aggregate Measurements and Design Degrees of Freedom

In many scientific fields, we can only obtain aggregate measurement data (such as population averages) and cannot directly observe individual behavior. These aggregate measurements are usually insufficient to uniquely determine the underlying dynamical law, so that multiple laws can fit the data equally well. This phenomenon is known as "residual design degrees of freedom."

The paper points out that different measurement systems may be equally suitable for a given scientific task, but they constrain different aspects of the law. For example, in sensor selection, different sensor combinations may all provide enough information to complete a prediction task, but they constrain system behavior in different ways.

In the baseline-preserving setting, we have an existing dynamical model as a working baseline that remains valid beyond the quantity of explicit interest to the task. The goal is to design experiments (such as choosing measurement modalities) to minimize corrections to the baseline model while completing the scientific task.

Information Projection is a core concept in information geometry: finding the distribution "closest" to a reference distribution under given constraints, typically using KL divergence as the distance measure. In the context of FIDE, information projection is used to "complete" the aggregate constraints induced by each candidate design into a full dynamical path: taking the frozen baseline model as reference, among all possible paths satisfying the observational constraints, select the one with minimal correction relative to the baseline. This operation is essentially an orthogonal projection onto the constraint manifold, giving comparisons across designs a unified baseline. The differentiability of information projection relies on implicit differentiation of the first-order optimality conditions of the constrained optimization problem, thereby backpropagating the gradient of the projection result with respect to design parameters — this is precisely the mathematical foundation that enables FIDE to perform gradient-based optimization.

The FIDE Method: Information Projection and the Minimal-Correction Principle

The core idea of FIDE is to complete the aggregate constraints of each candidate design via information projection. Specifically, for each candidate design, FIDE performs information projection relative to a common frozen reference (i.e., the baseline model) to obtain a full path. Then, scientifically adequate designs are ranked by the amount of dynamical correction required to realize these full paths.

The paper defines a measure called "Full action," namely the minimal mean-squared velocity correction required to realize the full path. Equivalently, its square root is the sharp reference-relative expected-rate discrepancy over the gradient-normalized Sobolev class. This definition forms a conditional minimal-correction principle for baseline-preserving design, rather than a generic truth-recovery objective.

FIDE's computation is differentiable, allowing gradient-based optimization. Experiments show that the method remains computationally tractable for 32-dimensional configuration laws.

The Sobolev class and gradient normalization are key background for understanding the Full action measure. A Sobolev space is a function space requiring both the function itself and its (weak) derivatives to belong to some Lebesgue space, with a norm that penalizes both the magnitude of function values and their rate of change. In FIDE, the square root of Full action is interpreted as the "sharp reference-relative expected-rate discrepancy over the gradient-normalized Sobolev class," meaning the measured object is not the Euclidean distance of the path itself but the correction cost measured by the mean-squared magnitude of the velocity field (i.e., the time derivative). Gradient normalization weights the velocity field by the gradient scale, making state components of different dimensions or sensitivities comparable in the correction cost. This construction makes Full action sensitive to the "severity of correction" rather than merely endpoint differences, better matching the physical intuition of dynamical-law correction.

Experimental Findings: Preference Shifts and Predictive Impact

The paper presents several key findings from experiments:

• Changes in sensor selection: FIDE's preferences alter sensor selection. Unlike conventional methods, FIDE tends to select designs that require minimal correction to the baseline model.
• Impact on downstream prediction: Depending on baseline reliability, FIDE may help or harm withheld downstream predictions. When the baseline model is reliable, FIDE improves prediction performance; when the baseline is unreliable, it may have negative effects.
• Prospective use: FIDE can be used prospectively from aggregate predictions, i.e., making design choices without real data.
• Computational feasibility: For 32-dimensional configuration laws, FIDE remains computationally feasible.

These findings highlight the difference between the conditional minimal-correction principle in baseline-preserving design and generic truth-recovery objectives.

Withheld downstream prediction refers to prediction tasks or data intentionally not used during the experimental design phase, used afterward to assess the impact of design choices on unseen prediction targets. In FIDE's experiments, the design phase performs sensor selection based only on aggregate measurements and the baseline model, while the withheld downstream prediction tests how well the selected sensor combination generalizes to new prediction tasks. This evaluation paradigm distinguishes "information visible at design time" from "targets tested after design," avoiding optimistic bias from directly optimizing evaluation metrics during design. FIDE's findings show that when the baseline model is reliable, designs selected by the minimal-correction principle perform better on withheld predictions; when the baseline is unreliable, the principle may systematically favor erroneous directions consistent with the baseline, degrading withheld prediction performance.

Significance and Outlook

FIDE offers a new perspective on experimental design: in baseline-preserving settings, we should not seek to recover the "true" law, but rather to make minimal corrections on top of the existing baseline. This principle better matches real scientific workflows, where baseline models are often reliable starting points.

The method has potential applications in sensor selection, experimental optimization, and related areas. Future work can further explore FIDE's applicability across domains and how to handle cases where the baseline model is unreliable.

The paper's code and further details have not yet been released, but this method opens a new direction for differentiable experimental design.

One-line Summary

FIDE selects experimental designs under baseline preservation using information projection and a minimal-correction principle.

Paragraph Summary

The paper addresses the problem that aggregate measurements cannot uniquely determine dynamical laws, leaving residual design degrees of freedom, and proposes Fiber-Informed Differentiable Experimental Design (FIDE). FIDE performs information projection for each candidate design relative to a frozen baseline model, completing aggregate constraints into full paths, then ranks scientifically adequate designs by Full action (minimal mean-squared velocity correction), forming a conditional minimal-correction principle for baseline-preserving settings. The method is differentiable, supports gradient-based optimization, and remains computationally feasible for 32-dimensional configuration laws. Experiments show that FIDE shifts sensor-selection preferences, its impact on withheld downstream predictions depends on baseline reliability, and it supports prospective design without real data.