Technical Summary: The Geometry of Memorization

Problem Statement

Continuous-time generative frameworks, such as Flow Matching, construct probability paths between base and target domains by optimizing time-dependent velocity fields. While optimal transport theory suggests straight-line trajectories at global optimality, empirically trained neural networks often develop complex, non-linear path deformations influenced by architecture and dataset characteristics.

A critical challenge in this domain is diagnosing generative memorization, where high-capacity models overfit by collapsing continuous latent spaces onto discrete training samples. Existing diagnostic methods are predominantly extrinsic and data-dependent, relying on post-hoc nearest-neighbor searches against training pools or membership inference attacks. These approaches require access to the training data and external baselines, which limits their utility in auditing scenarios where data privacy or availability is constrained. Furthermore, tracking the internal geometry of these models via exact spatial derivatives (e.g., constructing the full spatial Jacobian) is computationally prohibitive for high-dimensional image models, as it requires $O(d)$ backward passes for an ambient dimension $d$.

Methodology: Finite-Time Spectral Sensitivity (FTSS)

To address these limitations, the paper introduces Finite-Time Spectral Sensitivity (FTSS), denoted as $g(t)$. This is a gradient-free, forward-pass metric designed to expose the geometry of flow matching models by tracking the root-mean-square (RMS) singular value of the state-transition matrix.

Mathematical Foundations

The FTSS is defined based on the propagation of infinitesimal perturbations $\delta_t$ along a trajectory governed by the ordinary differential equation (ODE) $d x_t / dt = v_\theta(x_t, t)$.

1. Definition: $g(t)$ measures the expected directional norm of the flow map $\phi_{t \to 1}$ acting on isotropic perturbations on the unit sphere $S^{d-1}$.
2. Spectral Interpretation: Proposition 2.2 establishes that $g(t)$ is directly equivalent to the RMS singular value of the state-transition matrix $\Phi(1, t; x_t)$. Specifically, $g(t)^2$ represents the expected mean-square singular value:
   $$g(t)^2 = \frac{1}{d} \mathbb{E}_{x_t} \left[ \sum_{i=1}^d \sigma_i^2(t; x_t) \right]$$
3. Advantage over Determinants: Unlike the Jacobian determinant, which can mask dimensional collapse if one axis expands while another contracts (e.g., $\sigma_1 \to 0, \sigma_2 \to \infty$), FTSS aggregates the spectrum additively. This ensures that local dimensional collapse is not obscured by outlier expansions.
4. Entropy Bound: The paper derives an upper bound on the differential entropy of the generated target distribution, showing that a drop in global FTSS $g(0)$ fundamentally restricts the information capacity of the generated distribution.

Computational Implementation

To circumvent the $O(d)$ cost of dense Jacobian construction, the authors propose Algorithm 1, a parallelized, gradient-free Monte Carlo estimator.

• Mechanism: The algorithm injects isotropic perturbations $\epsilon u$ into cached intermediate states, propagates them to the terminal time via forward-pass trajectory simulations, and computes the RMS gain.
• Complexity: The cost is $O(K \cdot T \cdot \text{Cost}(\phi))$, where $K$ is the number of trajectories and $T$ is the time grid size. This requires only forward passes, making it scalable for high-dimensional models.

Key Contributions

1. Formalization of FTSS: The paper formalizes $g(t)$ as an intrinsic geometric observable, proving its equivalence to the RMS singular value of the state-transition matrix and establishing it as a continuous proxy for stable rank.
2. Theoretical Bounds: Using the arithmetic-geometric mean inequality, the authors bound the absolute Jacobian determinant pointwise by the local RMS singular value. This leads to a proof of an explicit upper bound on the differential entropy of the generated distribution (Proposition 2.5).
3. Spectral Collapse Ratio ($M$): The authors define a metric $M := g_{\min} / g_{\min}^{\text{full}}$, where $g_{\min}$ is the minimum FTSS value during trajectory generation. They demonstrate that this ratio characterizes the transition from generalization to memorization.
4. Internal Geometric Audit: The paper proposes a framework (Algorithm 2) that detects generative memorization using purely internal trajectory dynamics. This removes the necessity for external membership queries or baseline data comparisons.

Empirical Results

The paper presents empirical characterizations of spectral collapse across different training data sizes, model architectures, and datasets.

• Spectral Collapse Phenomenon: In low-data regimes where models memorize, the FTSS curve exhibits a distinct "spectral collapse" during the late stages of trajectory generation ($t \to 1$). The velocity field "pinches" the high-dimensional flow of paths into a lower dimension to reach sparse, discrete training targets.
• Diagnostic Accuracy: The Spectral Collapse Ratio $M$ accurately identifies overfitted continuous-time models. Generalizing models maintain stable effective dimensions (stable $g(t)$), whereas overfitting causes a significant drop in $g_{\min}$, leading to a low $M$ value.
• Scalability: The method successfully audits models where exact Jacobian-based diagnostics would be computationally infeasible.

Significance and Claims

The paper claims that FTSS provides a scalable, intrinsic diagnostic for generative memorization. By shifting the focus from extrinsic data comparisons to internal geometric properties, the framework allows for the detection of overfitting without access to the training pool. The authors posit that memorization is not merely a statistical artifact but a structural pathology of the transport map, characterized by the collapse of the flow's spectral properties. This approach offers a new perspective on auditing generative models, linking the geometry of the velocity field directly to the information capacity and stability of the generated distribution.