Prokhorov Metric Overview

• Prokhorov metric is a clear definition that metrizes weak convergence for Borel probability measures on complete separable metric spaces.
• It employs coupling-based and set-enlargement formulations to establish links with total variation and Wasserstein distances.
• Generalizations such as the Gromov–Hausdorff–Prokhorov and fuzzy metrics broaden its applications in stochastic processes and robust statistical methods.

The Prokhorov metric is a fundamental tool in probability theory, optimal transport, stochastic processes, and geometric analysis for comparing Borel probability measures on metric spaces. It metrizes the topology of weak convergence of probability measures on complete separable metric spaces (Polish spaces), underpins compactness results such as Prokhorov's theorem, and facilitates generalizations to measured metric spaces and their convergence topologies. Its coupling-based and set-enlargement formulations yield crucial links to other divergences, notably total variation and Wasserstein distances, with deep implications for both classical and modern probabilistic analysis.

1. Formal Definitions and Core Properties

Given a complete separable metric space $(X, d)$, write %%%%1%%%% for the space of Borel probability measures on $X$. For $A \subset X$ and $\varepsilon > 0$, define the $\varepsilon$-neighborhood

$$A^{\varepsilon} = \{ x \in X : d(x, A) < \varepsilon \}.$$

The Prokhorov metric $\pi(\mu, \nu)$ on $\mathcal P(X)$ is defined by

$$\pi(\mu, \nu) = \inf\left\{ \varepsilon > 0 : \forall A~\text{Borel},~ \mu(A) \leq \nu(A^\varepsilon) + \varepsilon ~\text{and}~ \nu(A) \leq \mu(A^\varepsilon) + \varepsilon \right\}.$$

Equivalently, a coupling formulation holds: $\pi(\mu, \nu)$ coincides with the infimum of $\varepsilon > 0$ such that there exists a coupling $\gamma$ of $\mu, \nu$ with $$\gamma\{(x, y): d(x, y) > \varepsilon\} \leq \varepsilon$$ (Aolaritei et al., 19 Feb 2025, Löhr, 2011).

Core properties include:

• $\pi$ is a metric on $\mathcal P(X)$. $(\mathcal P(X), \pi)$ is complete and separable if $(X, d)$ is Polish (Gehér et al., 2017, Abraham et al., 2012).
• $\pi$ metrizes weak convergence: for $(\mu_n) \subset \mathcal P(X)$ and $\mu \in \mathcal P(X)$, $\pi(\mu_n, \mu) \to 0$ if and only if $\mu_n \Rightarrow \mu$ (i.e., $\int f \, d\mu_n \to \int f \, d\mu$ for every bounded continuous $f$).
• For probability measures, $\lVert\mu-\nu\rVert_{TV} \geq \pi(\mu, \nu)$; and the diameter of $(\mathcal P(X), \pi)$ is $1$ (Abraham et al., 2012).

2. Coupling Characterizations and Parametrizations

The Prokhorov metric admits a coupling ("transport") characterization directly analogous to Strassen's theorem. Let $\Gamma(\mu, \nu)$ denote the set of couplings of $\mu$ and $\nu$ (i.e., probability measures on $X \times X$ with marginals $\mu, \nu$). Then

$$\pi(\mu, \nu) = \inf \left\{ \varepsilon > 0 : \exists \gamma \in \Gamma(\mu, \nu),~ \gamma\{d(x, y) > \varepsilon\} \leq \varepsilon \right\}.$$

This extends naturally to finite (not necessarily probability) measures by allowing couplings with marginal discrepancies controlled in total variation, as in the generalized Strassen theorem (Khezeli, 2019):

$$d_P(\mu, \nu) = \min\{\epsilon \geq 0:~\exists~\alpha,~ D(\alpha;\mu,\nu)+\alpha(d > \epsilon)\leq \epsilon\},$$

where $D(\alpha;\mu,\nu)$ bounds the total variation distance between marginals and $\alpha$ is a Borel measure on $X \times X$.

Parametrized variants introduce a parameter $\lambda > 0$ scaling the neighborhood radius in the definition:

$$d_{P,\lambda}(\mu, \nu) = \inf \left\{ \alpha > 0 : \forall A,~ \mu(A) \leq \nu(A^{\lambda\alpha}) + \alpha \right\},$$

with the classical case recovering $\lambda=1$ (Berckmoes, 2016).

3. Connection to Weak Convergence, Tightness, and Compactness

Prokhorov's theorem asserts: A family $T \subset \mathcal P(X)$ is relatively compact in the weak topology if and only if it is uniformly tight, i.e., for every $\varepsilon > 0$ there exists a compact $K \subset X$ with $\sup_{\mu \in T} \mu(X \setminus K) < \varepsilon$ (Berckmoes, 2016). The topology induced by $\pi$ corresponds exactly to weak convergence, and relative compactness criteria are characterized quantitatively via the Hausdorff measure of non-compactness as follows:

$$\sup_{\lambda > 0} H_{d_{P,\lambda}}(T) = H^{ut}(T),$$

where $H^{ut}(T)$ expresses uniform tightness (Berckmoes, 2016).

4. Generalizations: Gromov–Hausdorff–Prokhorov and Related Metrics

The Prokhorov metric serves as the measure component of several extended metrics for measured metric spaces:

• Gromov–Hausdorff–Prokhorov (GHP) Distance: For two compact measured metric spaces $(X,\rho_X,\mu_X), (Y, \rho_Y, \mu_Y)$,

$$d_{GHP}((X,\rho_X,\mu_X),(Y,\rho_Y,\mu_Y)) = \inf_{\varphi, \psi}\max\{d_Z(\varphi(\rho_X),\psi(\rho_Y)),~ d_H^Z(\varphi(X),\psi(Y)),~ d_P^Z(\varphi_*\mu_X, \psi_*\mu_Y)\},$$

where the infimum is over isometric embeddings into a common Polish space $Z$ (Abraham et al., 2012).

• Gromov–Prohorov (GP) Metric: For metric measure spaces $(X_1, d_1, \mu_1), (X_2, d_2, \mu_2)$,

$$d_{GPW}(X_1, X_2) = \inf_{f_1, f_2}\, \pi(f_{1*}\mu_1, f_{2*}\mu_2),$$

the infimum taken over isometric embeddings into a common metric space (Löhr, 2011).

These constructions yield measured metric spaces as Polish spaces, fundamental as state spaces for random geometric models (e.g., continuum random trees, random maps) (Abraham et al., 2012, Löhr, 2011, Khezeli, 2019).

5. Comparisons with Other Probability Metrics

The Prokhorov metric can be tightly related to other notions of distance between probability measures:

• For total variation, $\mathrm{TV}(\mu, \nu)\leq d_{LP}(\mu, \nu)$ (where $d_{LP}$ denotes the Prokhorov or Lévy–Prokhorov metric) (Aolaritei et al., 19 Feb 2025);
• For Wasserstein $W_\infty$, $d_{LP}(\mu, \nu) \leq W_1(\mu, \nu)$ and, in the coupling-based formulation, $d_{LP}$ interpolates between TV and $W_\infty$ (Aolaritei et al., 19 Feb 2025).

A two-parameter decomposition shows that an LP ball of radius $(\varepsilon,\rho)$ corresponds to first a $W_\infty$ perturbation of up to $\varepsilon$, then a TV change of up to $\rho$ (Aolaritei et al., 19 Feb 2025).

6. Generalizations: Discrete and Fuzzy Prokhorov Metrics

a. Discrete Prokhorov Metrics in Topological Data Analysis

In contexts such as persistence diagrams, discrete analogues of the Prokhorov distance quantify similarity not in terms of measure, but cardinality:

$\pi_f(X, Y) = \inf\{ t>0 : D_{X,Y}(t) < f(t) \},$

with $D_{X,Y}(t)$ measuring, over optimal matchings, the number of unmatched points exceeding $t$ in displacement and $f$ a nondecreasing admissible function. The bottleneck and Wasserstein distances are special cases (Dłotko et al., 2021).

b. Fuzzy Prokhorov Metric

Given a compact fuzzy metric space $(X, M, *_l)$ with Łukasiewicz t-norm, the fuzzy Prokhorov metric $F_P$ compares probability measures via

$$F_P(\mu, \nu, t) = 1 - \inf\{ r \in (0,1) : \forall~A~\text{Borel},~ \mu(A) \leq \nu(A^{r,t}) + r ~\text{and}~ \nu(A) \leq \mu(A^{r,t}) + r \},$$

where $A^{r, t}$ is the union of fuzzy balls $B(x,r,t)$ about $x \in A$. This induces a fuzzy metric and metrizes weak* convergence (Repovš et al., 2011).

7. Isometry Characterization, Applications, and Implications

Surjective isometries for the Lévy–Prokhorov metric on $\mathcal P(X)$, with $X$ a separable Banach space, are characterized: for any surjective $\pi$-isometry $T:\mathcal P(X)\to\mathcal P(X)$, there exists a surjective affine isometry $\psi:X\to X$ such that $T\mu = \psi_*\mu$. This generalizes Molnár's result for real line distributions (Gehér et al., 2017). The class of measure-preserving isometries thereby encodes the intrinsic geometric structure of $\mathcal P(X)$ under the Prokhorov metric, with significant consequences for both functional analysis and probability.

Further, Prokhorov metric balls define Prokhorov-tight families and underlie effective control of both random process convergence and construction of ambiguity sets for robust statistical procedures, such as conformal prediction under distribution shift, where the metric's interpolation between TV and Wasserstein admits precise control of local and global perturbations (Aolaritei et al., 19 Feb 2025). This enables robustification in both theory and algorithmic implementations for nuanced distributional changes beyond classical settings.

---

References:

• (Gehér et al., 2017) Gehér–Titkos, "A characterisation of isometries with respect to the Lévy-Prokhorov metric"
• (Abraham et al., 2012) "A note on Gromov-Hausdorff-Prokhorov distance between (locally) compact measure spaces"
• (Berckmoes, 2016) "On the Hausdorff measure of non-compactness for the parametrized Prokhorov metric"
• (Khezeli, 2019) "Metrization of the Gromov-Hausdorff (-Prokhorov) Topology for Boundedly-Compact Metric Spaces"
• (Aolaritei et al., 19 Feb 2025) "Conformal Prediction under Levy-Prokhorov Distribution Shifts: Robustness to Local and Global Perturbations"
• (Dłotko et al., 2021) "Bottleneck Profiles and Discrete Prokhorov Metrics for Persistence Diagrams"
• (Repovš et al., 2011) "Fuzzy Prokhorov metric on the set of probability measures"
• (Löhr, 2011) "Equivalence of Gromov-Prohorov- and Gromov's Box-Metric on the Space of Metric Measure Spaces"