Lévy–Prokhorov Metric

• Lévy–Prokhorov Metric is a probability metric that quantifies the distance between Borel measures on a metric space and metrizes weak convergence.
• Its formulations via coupling, optimal transport, and test-function dualities provide versatile frameworks for analyzing convergence and stability in probability theory.
• Applications in statistical inference, machine learning, and functional data analysis highlight its role in robustly characterizing distributional shifts and geometric rigidity.

The Lévy–Prokhorov metric ($d_{LP}$) is a probability metric characterizing the distance between Borel probability measures on a metric space. It metrizes weak convergence of measures, underlies the topology of weak convergence (often called Prokhorov topology), and admits several optimal transport and duality formulations. It appears equivalently as the Prokhorov or Ky Fan metric, and supports both coupling-based and test-function dualities. Its rigidity properties distill isometries of measure spaces to push-forwards by affine isometries, making it fundamental in probability theory, stochastic processes, statistical inference, functional data analysis, and machine learning.

1. Definition and Formal Structure

Let $(X, d)$ be a complete separable metric space, and $\mathcal{P}(X)$ the set of Borel probability measures on $X$. For $\mu, \nu \in \mathcal{P}(X)$ and $\epsilon > 0$, the $\epsilon$–neighborhood of a Borel set $A$ is $A^\epsilon = \{ x \in X : d(x, A) < \epsilon \}$.

The Lévy–Prokhorov distance is

$$d_{LP}(\mu, \nu) = \inf\left\{ \epsilon > 0 : \mu(A) \leq \nu(A^\epsilon) + \epsilon, \, \nu(A) \leq \mu(A^\epsilon) + \epsilon \ \forall A \subseteq X \text{ Borel} \right\}.$$

Key properties:

• $(X, d)$0 iff $(X, d)$1.
• $(X, d)$2 for all $(X, d)$3.
• If $(X, d)$4 is complete and separable, so is $(X, d)$5.
• $(X, d)$6 induces the topology of weak convergence on probability measures (Gehér et al., 2017, Pakshirajan et al., 2021, Aolaritei et al., 19 Feb 2025, Zhou et al., 2020, Kutta et al., 26 Jun 2025).

Several equivalent formulations exist:

• Coupling/transport plan: $(X, d)$7 where $(X, d)$8 is a coupling of $(X, d)$9 and $\mathcal{P}(X)$0 (Abraham et al., 2012, Aolaritei et al., 19 Feb 2025).
• Random variable interpretation: $\mathcal{P}(X)$1 iff there exist random variables $\mathcal{P}(X)$2 with $\mathcal{P}(X)$3, $\mathcal{P}(X)$4 bounded by $\mathcal{P}(X)$5, $\mathcal{P}(X)$6, $\mathcal{P}(X)$7 arbitrary (Aolaritei et al., 19 Feb 2025).
• Predicate lifting: For discrete distributions, $\mathcal{P}(X)$8, known also as the Ky Fan metric (Wild et al., 27 Oct 2025).

2. Metrization of Weak Convergence and Topological Properties

On a Polish space (complete, separable metric), $\mathcal{P}(X)$9 metrizes weak convergence: $X$0 in $X$1 (weakly) (Gehér et al., 2017, Pakshirajan et al., 2021, Zhou et al., 2020, Aolaritei et al., 19 Feb 2025, Kutta et al., 26 Jun 2025). The induced topology on $X$2 is exactly weak convergence:

• On $X$3, $X$4 (Lévy distance between cdfs), and both metrize weak convergence of distribution functions (Pakshirajan et al., 2021).
• On locally finite measures, extensions integrate Prokhorov distances over expanding balls, yielding Polish metric spaces for rooted, locally compact spaces (Abraham et al., 2012).
• Quantitative relationships: For any $X$5, $X$6 (Wasserstein distance) (Kutta et al., 26 Jun 2025).

Invariance and rigidity:

• The isometry group of $X$7 is isomorphic to the affine‐isometry group of $X$8 (Gehér et al., 2017).
• Nontrivial measure-preserving isometries arise only from affine isometries of $X$9; this is central for statistical and geometric characterizations in stochastic processes.

3. Optimal-Transport, Duality, and Variants

The Lévy–Prokhorov distance admits multiple dual and transport formulations:

• Wasserstein-style (coupling): $\mu, \nu \in \mathcal{P}(X)$0 iff there exists a coupling $\mu, \nu \in \mathcal{P}(X)$1 such that $\mu, \nu \in \mathcal{P}(X)$2; equivalently, $\mu, \nu \in \mathcal{P}(X)$3 where $\mu, \nu \in \mathcal{P}(X)$4 is the cost for mass exceeding distance $\mu, \nu \in \mathcal{P}(X)$5 (Aolaritei et al., 19 Feb 2025, Abraham et al., 2012, Wild et al., 27 Oct 2025).
• Price-function/Kantorovich lifting: LP distance can be cast as a supremum over non-expansive test functions, via a single “generally” modality $\mu, \nu \in \mathcal{P}(X)$6 (Wild et al., 27 Oct 2025).
• Generalized Kantorovich–Rubinstein duality: For the LP metric, coupling-based and Kantorovich (test-function) forms coincide for all pseudometrics; this does not hold for $\mu, \nu \in \mathcal{P}(X)$7-Wasserstein with $\mu, \nu \in \mathcal{P}(X)$8 (Wild et al., 27 Oct 2025).

Comparison with other metrics:

• LP balls decompose as $\mu, \nu \in \mathcal{P}(X)$9 capturing both local (Wasserstein-$\epsilon > 0$0) and global (Total Variation) perturbations (Aolaritei et al., 19 Feb 2025).
• For $\epsilon > 0$1: $\epsilon > 0$2 for appropriately chosen $\epsilon > 0$3.

4. Isometries of Measure Spaces and Geometric Rigidity

A characterization of surjective $\epsilon > 0$4–isometries for separable Banach spaces $\epsilon > 0$5 holds: a bijection $\epsilon > 0$6 is a $\epsilon > 0$7–isometry iff there exists a surjective affine isometry $\epsilon > 0$8 and $\epsilon > 0$9 for all Borel $\epsilon$0 (Gehér et al., 2017). Key technical tools in the proof include:

• Inductive analysis of finitely-supported measures using witness functions $\epsilon$1.
• The introduction of $\epsilon$2–Lévy–Prokhorov metrics to separate atoms of measures.
• Convex-geometric techniques analyzing supports in Banach spaces.

Implications:

• The only $\epsilon$3–isometries are push-forwards under surjective affine isometries.
• This extends Banach–Stone–type results previously established for scalar-valued (one-dimensional) measures (Gehér et al., 2017).
• Measure spaces $\epsilon$4 are metrically rigid, essential for isometric classification in random element theory and probabilistic invariance principles.

5. Applications in Probability, Statistics, and Machine Learning

Probability Limit Theory and Functional Data Analysis

• The LP metric is critical in expressing quantitative central limit theorems, including rates of convergence for Gaussian approximation in both univariate and functional (infinite-dimensional) settings (Zhou et al., 2020).
• Explicit convergence rates and constants are given in both classic and sublinear expectation frameworks: For the partial sum process $\epsilon$5 to Brownian motion $\epsilon$6, $\epsilon$7 with explicit $\epsilon$8 (Zhou et al., 2020).
• In functional data analysis, LP bounds allow for strong-invariance and coupling results, essential for change-point detection and distributional approximation under weak dependence (Kutta et al., 26 Jun 2025).

Robust Conformal Prediction

• In conformal prediction, LP ambiguity sets naturally model both local (bounded) and global (outlier) distribution shifts (Aolaritei et al., 19 Feb 2025).
• Propagation through Lipschitz scoring functions yields tractable univariate LP balls, facilitating exact worst-case quantiles and coverage.
• Relation to TV and Wasserstein balls allows flexible robustness modeling.

Behavioral Metrics and Coalgebraic Distances

• In Markov process theory, the LP lifting defines $\epsilon$9-distance, coinciding with the maximal fixpoint of a behavioral distance functional and capturing approximate bisimulation (Desharnais et al., 14 Jul 2025).
• Unlike the Kantorovich lifting, LP lifting is locally non-expansive and enables efficient computation of behaviorally defined distances.
• Coalgebraic characterizations of $A$0-couplings/bisimulations arise naturally from the LP structure.

Metric Geometry and Measured Trees

• The Gromov–Hausdorff–Prokhorov metric combines LP with Hausdorff distance to metrize spaces of compact or locally compact measured metric spaces (e.g., real trees) (Abraham et al., 2012).
• The resulting space is Polish (complete, separable), with precompactness determined by bounded diameters, net sizes, and total mass.
• LP metric ensures continuous dependence of tree laws on coding functions—crucial in random geometry and continuum random tree theory.

6. Computational, Structural, and Theoretical Properties

• The set-expansion definition is computationally intractable in high dimensions, but the coupling formulation reduces to mixed-integer or flow-type programs, and the two-step W$A$1+TV decomposition is efficient for empirical tasks (Aolaritei et al., 19 Feb 2025).
• LP balls encode local-global perturbations, with the ability to decompose distributional shifts precisely between bounded transport and TV mass movement (Aolaritei et al., 19 Feb 2025).
• The LP metric underlies stability results, e.g., for conformal predictors and in machine learning settings requiring robustness to outliers or non-local rearrangement (Wild et al., 27 Oct 2025).
• The Ky Fan metric, identical to LP on discrete supports, is key in several logic and bisimulation settings (Wild et al., 27 Oct 2025).

7. Summary Table: Comparative Features with Leading Probability Metrics

Metric	Weak Convergence	Coupling Duality	Local+Global Shift Decomposition	Isometry Rigidity
Lévy–Prokhorov	Yes	Yes	Yes (W$A$2+TV)	Yes (Banach Rx)
Wasserstein $A$3	Yes for $A$4	Yes	No	No
Total Variation	No	Yes	Only global	No
Kolmogorov	No	No	No	No

• (Gehér et al., 2017) A characterisation of isometries with respect to the Lévy-Prokhorov metric
• (Pakshirajan et al., 2021) An expository note on Prohorov metric and Prohorov Theorem
• (Abraham et al., 2012) A note on Gromov-Hausdorff-Prokhorov distance between (locally) compact measure spaces
• (Zhou et al., 2020) Prokhorov distance with rates of convergence under sublinear expectations
• (Kutta et al., 26 Jun 2025) Prokhorov Metric Convergence of the Partial Sum Process for Reconstructed Functional Data
• (Aolaritei et al., 19 Feb 2025) Conformal Prediction under Levy-Prokhorov Distribution Shifts: Robustness to Local and Global Perturbations
• (Desharnais et al., 14 Jul 2025) $A$5-Distance via Lévy-Prokhorov Lifting
• (Wild et al., 27 Oct 2025) Generalized Kantorovich-Rubinstein Duality beyond Hausdorff and Kantorovich