Fréchet Mean Flow in TDA

• Fréchet Mean Flow is a statistical construct that defines a continuous path of the Fréchet mean for evolving geometric and topological data.
• It employs a probabilistic framework to overcome non-uniqueness and discontinuity issues inherent in classical mean computations on persistence diagrams.
• The method integrates seamlessly into modern pipelines, enhancing applications in geometric deep learning, bootstrap estimation, and visualization of vineyards.

The Fréchet Mean Flow is a statistical construct designed to provide a robust, well-behaved notion of averaging in geometric and topological data analysis, particularly within the context of Riemannian manifolds and the space of persistence diagrams equipped with the Wasserstein metric. It is defined as the continuous evolution of the Fréchet mean along time-varying families of data, providing a stable statistical path even in settings where the classical mean is non-unique or discontinuous. The construct has critical implications for topological data analysis, especially in the context of vineyards—families of persistence diagrams indexed by time or another parameter. The Fréchet Mean Flow also plays a foundational role in recent developments in learning on non-Euclidean spaces, such as hyperbolic neural networks, where the mean must respect underlying geometry rather than being computed in the Euclidean tangent space (Munch et al., 2013, Lou et al., 2020).

1. Classical Fréchet Mean: Definition and Limitations

Let $(M, g)$ be a Riemannian manifold with distance function $d(\cdot,\cdot)$. For data points $x^{(1)},\ldots,x^{(t)} \in M$ with weights $w_1,\ldots,w_t > 0$, the (weighted) Fréchet mean $\mu_{\mathrm{fr}}$ is defined by

$$\mu_{\mathrm{fr}} = \arg\min_{y \in M} f(y),\quad f(y) = \sum_{l=1}^t w_l\, d(x^{(l)}, y)^2.$$

This variational characterization implies optimality via the condition

$$\sum_{l=1}^t w_l\, \log_{\mu_{\mathrm{fr}}}(x^{(l)}) = 0,$$

where $\log_y(x)$ denotes the Riemannian logarithm at $y$ (Lou et al., 2020).

In the context of persistence diagrams, the Fréchet mean is similarly defined as the minimizer of the average squared $p$-Wasserstein distance to a collection of diagrams $X_1, \ldots, X_n$:

$F(X) = \frac{1}{n} \sum_{i=1}^n W_p(X, X_i)^2.$

However, minimizers are not necessarily unique, and for continuously varying input (as in vineyards), the pointwise Fréchet mean may fail to vary continuously, causing severe challenges for statistical analysis and visualization of time-varying topological features (Munch et al., 2013).

2. Probabilistic Fréchet Mean and Construction of the Flow

The probabilistic Fréchet mean (PFM) replaces the ambiguous selection of a single minimizer with a probability distribution over all possible mean diagrams, effectively "spreading" mass over the space of diagrams according to their likelihood under infinitesimal data perturbations. For diagrams $X_1, ..., X_n$, independent random perturbations are introduced at each off-diagonal point (using a specific "tremble" kernel). For each sampled perturbation, a unique optimal grouping is computed, and these groupings are aggregated to form the atomic mean diagrams $Y_G$. The PFM is then given as

$$\mu_X = \sum_{G \in \mathcal{G}} P(H_X = G) \cdot \delta_{Y_G},$$

defining a probability measure on the space of persistence diagrams $D_p$. Each atomic diagram is weighted by the probability of its grouping arising under the perturbation process (Munch et al., 2013).

This construction ensures that the mean flow $(t \mapsto \mu_{X_1(t), ..., X_n(t)})$ is continuous, even when classical means would be discontinuous. This probabilistic approach resolves both the non-uniqueness and discontinuity issues inherent to the classical Fréchet mean.

3. Main Theoretical Results: Existence, Uniqueness, and Regularity

• Probabilistic Uniqueness: For every finite collection of diagrams, the probabilistic Fréchet mean is uniquely defined as a probability measure depending only on the input and the perturbation parameter $\alpha$, independent of algorithmic tie-breaking.
• Hölder Continuity: The map assigning to each $n$-tuple of diagrams its PFM is Hölder continuous (with exponent $1/2$) with respect to the 2-Wasserstein distance on finite diagrams:

$$W_2^{\mathbb{P}}(\mu_X, \mu_Y) \leq C \sqrt{\overrightarrow{d_2}(X, Y)},$$

where $\overrightarrow{d_2}$ is the sum-of-squares metric over diagrams.

• Continuous Mean Flow: For time-varying diagrams (vineyards) $\gamma_1(t), ..., \gamma_n(t)$, the probabilistic Fréchet mean defines a continuous path $t \mapsto \mu_{\gamma_1(t), ..., \gamma_n(t)}$, constituting the Fréchet Mean Flow (Munch et al., 2013).

4. Algorithms and Practical Approximation

Monte Carlo techniques provide efficient practical approximations to the PFM:

• For a fixed perturbation parameter $\alpha$, repeatedly sample perturbed diagrams, compute the optimal matching via bipartite or multi-diagram matching (e.g., Hungarian algorithm), and aggregate frequencies of atomic mean diagrams.
• Sampling from $\mu_X$ is achieved by repeatedly applying the tremble and mean-of-grouping procedure.
• Computational complexity per sample is $O(NK)$ for perturbation and $O((NK)^3)$ in the worst case for matching, but in practice, small $\alpha$ and pruning near-diagonal points reduce the computational burden (Munch et al., 2013).

5. Fréchet Mean Flow in Riemannian Geometry and Machine Learning

A generalization of the Fréchet Mean Flow applies to arbitrary Riemannian manifolds. The variational definition extends directly. Differentiating the Fréchet mean with respect to inputs is possible by extending the argmin-differentiation lemma to manifolds, allowing the computation of Jacobians required for backpropagation in geometric deep learning (Lou et al., 2020).

• The Fréchet Mean Flow is the solution path $y(t)$ to the gradient flow ODE:

$$\frac{d}{dt} y(t) = -\nabla_y f(\{x^{(l)}\}, y(t)),$$

with discrete update:

$$y_{k+1} = \exp_{y_k}(-\eta \nabla_y f(\{x^{(l)}\}, y_k)).$$

In hyperbolic space, explicit closed-form updates without tuning parameters are available, enabling efficient integration into neural network layers. The resulting algorithms converge globally on Hadamard manifolds (Lou et al., 2020).

6. Applications and Integration in Modern Pipelines

• Time-varying Persistence Diagrams (Vineyards): The Fréchet Mean Flow provides a continuous, robust mean path for visualizing and statistically summarizing vineyards.
• Moments and Inference: One may compute variance flows and confidence intervals for evolving diagrams.
• Geometric Deep Learning: The Fréchet mean replaces tangent-space aggregation in hyperbolic graph convolutional networks (GCNs) and underpins the construction of hyperbolic batch-normalization layers, leading to improved performance on highly non-Euclidean datasets. Algorithmic details for integration are provided and are fully differentiable with explicit convergence guarantees (Lou et al., 2020).
• Bootstrap and Statistical Estimation: The Fréchet Mean Flow allows for bootstrapped estimation of the "mean vineyard" from subsampled trajectories in evolving point clouds (Munch et al., 2013).

7. Significance and Resolution of Classical Pathologies

The Fréchet Mean Flow, particularly via the probabilistic construction, overcomes two fundamental barriers in topological and geometric statistics:

• Non-uniqueness of minimizers in the space of persistence diagrams is resolved by shifting from individual point estimates to probability distributions over means.
• Discontinuity of the mean along time-varying data is mitigated, as the mean flow is H\"older continuous even when the underlying diagrams vary continuously but their classical means jump discontinuously.

This approach supplies a well-founded statistical framework for mean and variance in evolving topological and geometric datasets, seamlessly integrating into both theoretical and applied pipelines in topological data analysis and geometric learning (Munch et al., 2013, Lou et al., 2020).