Technical Summary: Measurement Geometry and Design for Trustworthy Generative Inverse Problems

1. Problem Statement

Generative models, particularly diffusion models, have become standard priors for solving inverse problems (e.g., $y = Ax^* + e$) in fields like medical imaging. While these models enable reconstruction from highly underdetermined measurements, they introduce a fundamental "trust problem": a reconstructed image may be visually plausible and consistent with the measurements, yet be a "hallucination" generated by the prior along directions unobserved by the measurement operator $A$.

The central challenge is distinguishing between a reconstruction supported by actual data and one filled in by the prior. In medical imaging, where acquisition is constrained by scan time, radiation dose, and hardware, this distinction is critical. Existing compressed sensing theory often relies on global properties of random measurement operators (e.g., Restricted Isometry Property) or assumes fixed operators. This paper addresses the specific question of scan-specific acquisition: given a fixed generative prior and an initial set of measurements, how can one determine if the operator can distinguish the target from nearby plausible alternatives, and how can one design additional structured measurements to resolve remaining ambiguities?

2. Methodology

2.1 Measurement-Geometry Perspective

The authors propose viewing the problem through the lens of measurement geometry. They model the generative prior as defining a local data manifold $\mathcal{M}$ around the true signal $x^*$. The core hypothesis is that a measurement operator is trustworthy only to the extent that it observes the tangent directions of this local manifold. If the operator is insensitive to a tangent direction, two visually distinct images on the manifold may produce nearly identical measurements, leading to ambiguity.

2.2 Local Stability and Compatibility Indices

The paper formalizes this intuition through a local stability theorem. Let $U_*$ be a basis for the tangent space $T_{x^*}\mathcal{M}$. The authors define two key quantities:

1. Local Measurement Quality ($\alpha$):
   $$\alpha(A; x^*) := \sigma_{\min}(AU_*)$$
   This represents the smallest singular value of the operator restricted to the tangent space. It serves as a "worst-direction certificate." Theorem 1 proves that under local regularity assumptions, $\alpha$ controls the tangent component of the reconstruction error. If $\alpha$ is small, the prior cannot distinguish between the true signal and a nearby alternative, leading to potential hallucinations.

2. Log-Determinant Score ($S_\tau$):
   $$S_\tau(A; x^*) := \log \det(U_*^\top A^\top A U_* + \tau I_r)$$
   This quantity rewards the preservation of volume across the entire local tangent space. While $\alpha$ provides a theoretical bound for the worst-case direction, $S_\tau$ is proposed as a practical, operational objective for greedy measurement selection.

2.3 Posterior-Cloud Measurement Design

To address the design question (how to choose better measurements), the authors introduce a two-stage sequential acquisition framework that does not require training a new sampling policy:

1. Stage 1: Acquire an initial set of structured measurements (e.g., a fixed mask or projection angles) to obtain a coarse reconstruction.
2. Stage 2: Use the pretrained generative prior to sample a posterior cloud conditioned on the Stage 1 measurements. From this cloud, estimate a local subspace (tangent basis).
3. Selection: Greedily select additional feasible measurements (e.g., specific k-space lines or projection angles) that maximize the incremental log-determinant score $S_\tau$ on the estimated subspace.

This approach adapts the measurement operator to the specific geometry of the current scan's posterior without ground-truth data or retraining.

3. Key Contributions

1. Measurement-Geometry Framework: The paper establishes a geometric perspective for trustworthy generative inverse problems, identifying "tangent ambiguity" as the root cause of measurement-induced hallucinations.
2. Local Compatibility Index: It introduces the index $\alpha(A; x^*)$ and the score $S_\tau(A; x^*)$ to quantify the compatibility between a measurement operator and the local data manifold, providing a theoretical link between operator design and reconstruction stability.
3. Adaptive Design without Learned Policies: The authors propose a "posterior-cloud" design rule that selects structured measurements at test time based on the current posterior geometry. This avoids the need for training a separate mask network or sampling policy, distinguishing it from fully learned acquisition methods.
4. Empirical Validation: The framework is validated across three distinct structured acquisition settings:
   • Row Sampling (MNIST): Demonstrating that specific row patterns can induce hallucinations (e.g., reconstructing a '9' as a '7') and that geometry scores correlate with reconstruction error.
   • Tomographic Angle Selection (MNIST-CT): Showing that adaptive angle selection based on posterior clouds outperforms fixed and random angle strategies.
   • FastMRI Cartesian Sampling: Demonstrating that posterior-cloud design improves upon strong, non-learned baselines (e.g., variable-density and Poisson-like masks) in k-space line selection, even when preserving calibration regions (ACS).

4. Results

• Correlation with Error: In MNIST row sampling experiments, the proposed geometry scores ($\hat{\alpha}$ and $\hat{S}_\tau$) showed strong negative correlations with reconstruction error (Pearson/Spearman correlations around -0.5 to -0.6), confirming that better manifold observability leads to lower error.
• Hallucination Mitigation: The study visualized how insufficient masks lead to measurement-consistent hallucinations. The proposed adaptive designs successfully reduced these failure modes by selecting measurements that cross "stroke regions" separating plausible alternatives.
• Performance Gains:
  • In MNIST-CT, the adaptive posterior-cloud method achieved the lowest relative error across all tested angle budgets (e.g., 0.3145 vs. 0.3231 for uniform spacing in a 7+7 setting).
  • In FastMRI (R=6), the adaptive method using a VDS prefix followed by posterior-cloud selection achieved a NRMSE of 0.0620 and SSIM of 0.9574. This outperformed the strong non-learned baseline of ACS + VD-Poisson-like (NRMSE 0.0662, SSIM 0.9528), demonstrating that test-time adaptation can refine even well-tuned static masks.

5. Significance and Claims

The paper claims that trustworthiness in generative inverse problems is not solely a property of the prior or the solver, but fundamentally depends on whether the measurement operator resolves the locally plausible alternatives induced by that prior.

• Explanatory Power: The framework provides a geometric explanation for hallucinations: they occur when the operator fails to observe tangent directions where the prior allows significant variation.
• Design Principle: It offers a practical, conservative design principle for structured acquisition. By optimizing for local volume preservation (via the log-determinant score) on the posterior cloud, one can select measurements that reduce ambiguity without requiring ground-truth training data.
• Modesty of Claims: The authors explicitly state that their claims are local and directional rather than global. The log-determinant score is a practical surrogate for the worst-case singular value, not a guarantee that every greedy step improves the worst-case bound. Furthermore, the comparison is against non-learned baselines; the paper acknowledges that fully learned acquisition policies (trained on large datasets) may achieve higher absolute performance but answers a different question by relying on training distributions rather than test-time posterior geometry.

In conclusion, the work bridges the gap between generative priors and measurement design, arguing that for a fixed solver, the measurement operator remains a "first-class determinant" of reconstruction reliability.