Optimal Transport-Based Metrics

• Optimal transport-based metrics are mathematical tools that compare distributions by quantifying the minimal cost required to transform one into another while incorporating geometric and statistical features.
• They extend classical formulations like the p-Wasserstein distance with variants such as unbalanced, energy, and graph-based metrics to suit diverse data modalities.
• These metrics are widely applied in machine learning, generative modeling, and spatial analysis, offering robust solutions with computational challenges in high-dimensional settings.

Optimal transport-based metrics are a class of mathematical tools for comparing probability measures, data distributions, or other structured objects by quantifying the “cost” of transforming one distribution (or object) into another. Their defining characteristic is that they seek to solve a global optimization—typically via the minimization of a cost functional over couplings or transport plans—subject to prescribed marginal (mass-conservation and/or unbalanced) constraints, often encoding geometric, topological, or statistical information not accessible to local or pointwise metrics. These metrics have seen widespread adoption due to their statistical, geometric, and computational properties, as well as their ability to unify the analysis, learning, and comparison of complex data modalities.

1. Classical and Extended Formulations of Optimal Transport Metrics

The p-Wasserstein distance $W_p(\mu, \nu)$ is the canonical OT-based metric for probability measures $\mu$ and $\nu$ on a measurable space $X$, defined by the Kantorovich formulation: $$W_p(\mu, \nu) = \left( \inf_{\pi \in \Pi(\mu, \nu)} \int_{X \times X} c(x, y)^p\, d\pi(x, y) \right)^{1/p}$$ where $\Pi(\mu, \nu)$ is the set of couplings with marginals $\mu$ and $\nu$, and $c(x, y)$ is a ground cost (often $\|x-y\|$ for $\mu$0) (Canas et al., 2012).

Extensions include:

• Energy distances: Which replace the ground cost with kernel-induced similarities, connecting to MMD and Cramer distances.
• Unbalanced Optimal Transport: Metrics such as the Hellinger–Kantorovich and interpolating WFR distances generalize Wasserstein to arbitrary nonnegative measures, introducing Kullback–Leibler–type divergences in their minimization (Rana et al., 23 Sep 2025, Chizat et al., 2015, Bauer et al., 26 May 2026).
• Sobolev and graph-based metrics: These admit closed-form or scalable computation on graphs or trees, and exhibit negative definiteness, facilitating kernel construction (Le et al., 2022).
• Partition, structure-aware, and chain-rule metrics: Metrics are defined over objects with combinatorial, semantical, or compositional structure, such as graph partitions, C-sets, or statistical mixtures, by designing appropriate hierarchical, nested, or conditional ground metrics (Nielsen et al., 2018, Patterson, 2019, Abrishami et al., 2019).

2. Metric Properties and Statistical Guarantees

All OT-based metrics formulated via cost minimization over couplings or transport maps are bona fide metrics on suitable domains:

• Nonnegativity, symmetry, triangle inequality: Directly inherited from properties of the cost and the coupling polytope (Canas et al., 2012, Nielsen et al., 2018).
• Zero iff equality: $\mu$1 if and only if $\mu$2 under mild assumptions.
• Metric on mixtures and complexes: Conditional or multi-level metrics such as Chain Rule OT (CROT) inherit metricity if their ground distance is a metric (Nielsen et al., 2018).

Statistical theory for OT metrics provides:

• Quantization error rates: $\mu$3 (empirical law of large numbers) satisfies a minimax lower bound $\mu$4 and (under regularity) high-probability upper bounds $\mu$5, quantifying the curse of dimensionality (Canas et al., 2012).
• Sample efficiency and optimal quantization: k-means or similar quantizers nearly attain these rates in expectation.
• Unbiasedness and convergence in learning contexts: Mini-batch energy distances yield unbiased stochastic gradients and almost sure convergence as batch size grows, with additional convergence claims in adversarial settings (Salimans et al., 2018).

3. Computation: Exact, Regularized, and Scalable Algorithms

Computation of OT-based metrics hinges on the structure of the cost and marginal/mass constraints:

• Exact LP or network flow: For discrete distributions, classical OT reduces to a linear program or network-flow problem, with cubic or worse complexity in the general case (Li et al., 2014, Canas et al., 2012).
• Entropic Regularization: The Sinkhorn–Knopp algorithm computes an entropic-regularized OT metric, solving a strictly convex program via matrix scaling, with per-iteration cost quadratic in support sizes and rapid empirical convergence (Salimans et al., 2018, Rana et al., 23 Sep 2025).
• Closed-form and analytic solutions: Special cases (1D, graphs, trees, functional transforms like CDT/SW2) enable exact solutions via sorting or simple matrix operations (Le et al., 2022, Martín et al., 2024).
• Greedy and hierarchical constructions: Metrics defined on indices or partitions (e.g., nested OT for $\mu$6 in fixed-point analysis, assignment on partition graphs) admit specialized algorithms exploiting inside-out, nested, or block-structured plans (Bravo et al., 2021, Abrishami et al., 2019).
• Approximate and scalable methods: Sliced Wasserstein and LOT-type transforms trade higher-dimensional coupling for computation along projections or via fixed reference measures (Martín et al., 2024).

4. Adaptivity: Learned and Structured Ground Costs

Optimal transport metrics are highly sensitive to the choice of ground cost. Recent advances propose:

• Learned Mahalanobis/Riemannian metrics: Data-adaptive approaches learn a symmetric positive definite matrix (or even a spatially-varying tensor field) parameterizing $\mu$7, optimized via Riemannian geometry and alternating Sinkhorn updates (Jawanpuria et al., 2024, Scarvelis et al., 2022, Jeong, 7 May 2026).
• Adversarially learned features: In generative modeling (GANs), the feature space itself, over which the OT cost is measured, is also learned adversarially, yielding a more discriminative and task-adaptive metric (Salimans et al., 2018).
• Residual metric adaptation: Fine-grained adaptation by learning a residual transformation of embedding space, correcting for the deficiencies or invariances of a fixed embedding (Jeong, 7 May 2026).

Empirical results consistently show improved alignment, discrimination, or generative fidelity from such adaptive OT metrics compared to static ground costs.

5. Extension to Complex Domains: Unbalanced, Graph, and Structured Data

OT-based metrics extend to numerous contexts:

• Unbalanced optimal transport: The Hellinger–Kantorovich and Wasserstein–Fisher–Rao distances interpolate between OT and entropic/growth metrics, allowing comparison of measures with different total mass and modeling transport with creation/destruction (Rana et al., 23 Sep 2025, Chizat et al., 2015, Bauer et al., 26 May 2026). The Wasserstein–Ebin metric further lifts these concepts to the space of Riemannian metrics, establishing a geometric hierarchy via Riemannian submersions (Bauer et al., 26 May 2026).
• Graph and partition-valued metrics: Variants for measures on graphs, graph partitions, or more abstract C-sets lead to fast, structure-aware comparisons and yield convex relaxations of intractable combinatorial metrics (Le et al., 2022, Abrishami et al., 2019, Patterson, 2019).
• Statistical mixtures and hierarchical models: The chain-rule OT metric generalizes Wasserstein distances to spaces of statistical mixtures or jointly convex divergences, facilitating upper-bounds and efficient approximate inference (Nielsen et al., 2018).

Applications span generative modeling, biostatistics, spatio-temporal prediction (e.g., demand relocation evaluated by OT loss (Wiedemann et al., 2024)), signal processing, metric learning, and data morphometry.

6. Practical Considerations, Strengths, and Limitations

OT-based metrics offer several unique advantages:

• Global, geometry-sensitive comparison: They respect both the geometry of supports and mass arrangement, penalizing not just magnitudes but the spatial or structural arrangement of errors or differences (Wiedemann et al., 2024, Li et al., 2014).
• Scale and invariance properties: Proper choice of cost function and regularization yields metrics invariant under translation, scaling, or other group actions, critical for robust statistical learning (Martín et al., 2024).
• Embedding and discrimination: OT-based transforms linearize complex mass-preserving deformations, facilitating discrimination, regression, and embedding of non-Euclidean data (Martín et al., 2024).
• Interpretable diagnostics: Discrete OT plans provide per-element transport information, enabling precise attribution of mismatch or error (e.g., in audio (Jeong, 7 May 2026)).

However, limitations persist:

• Computational cost: Full OT remains expensive for large-scale or high-dimensional data, motivating entropic regularization or projection methods (Martín et al., 2024, Rana et al., 23 Sep 2025).
• Curse of dimensionality in statistical rates: OT distances require more samples to reliably estimate convergence, with rates degrading exponentially in ambient dimension (Canas et al., 2012).
• Sensitivity to ground cost and support: Comparison depends crucially on cost specification and support overlap; metric learning addresses this at additional optimization cost (Jawanpuria et al., 2024, Scarvelis et al., 2022).
• Algorithmic hyperparameter selection: Regularization strength (entropic, unbalanced), mini-batch sizes, and tuning of learned components often require validation (Salimans et al., 2018, Rana et al., 23 Sep 2025).

7. Empirical Impact and Application Domains

OT-based metrics are deployed across domains:

• Machine learning and generative modeling: Energy distances, OT-based GANs, neighbor embeddings with unbalanced OT deliver improved generation quality, class separability, and clustering performance (Salimans et al., 2018, Rana et al., 23 Sep 2025).
• Data representation and morphometry: LOT, CDT, RCDT, and related transforms linearize class structure, enabling robust, interpretable feature extraction and downstream analysis in signal, image, and biostatistical data (Martín et al., 2024).
• Spatial, spatiotemporal, and structural evaluation: OT-based metrics provide interpretable, scale-appropriate error measures in transport systems, resource allocation, and spatiotemporal prediction, outperforming pointwise losses and quantifying real-world costs (Li et al., 2014, Wiedemann et al., 2024).
• Optimal metric learning and trajectory inference: Riemannian metric learning via OT regularizes nonparametric geodesic inference, improving prediction of population or biological flows (Scarvelis et al., 2022).

The development of OT-based metrics continues to drive innovation in statistical methodology, scalable computation, and data-adaptive learning frameworks, with ongoing research addressing computational bottlenecks, theoretical extensions, and practical deployments across scientific disciplines.