Probability Vector Field (PVF)

• Probability vector field (PVF) is a measure-valued generalization of classical vector fields, assigning a distribution of velocities to mass at each point.
• PVFs model dynamics such as deterministic transport, splitting, and diffusion by replacing a single velocity with a probability measure over admissible velocities.
• Key methodologies include generalized Wasserstein metrics, Euler-type discretizations, and quantile-based constructions to ensure stability and well-posedness.

Searching arXiv for the cited PVF-related papers to ground the article in current metadata. arxiv_search.query({"search_query":"id:(Cavagnari et al., 2021) OR id:(Gong et al., 2022) OR id:(Piccoli et al., 2018) OR id:(Kauffmann, 2013) OR id:(Zhang et al., 2023)","start":0,"max_results":10}) Probability vector field (PVF) denotes, in measure-dynamical and optimal-transport settings, a measure-valued generalization of a vector field: instead of assigning a single velocity to each point, it assigns to a measure $\mu$ a measure on the tangent bundle whose spatial marginal is exactly $\mu$. In this way, each portion of mass may carry a distribution of admissible velocities, and the resulting dynamics can describe deterministic transport, splitting, concentration, finite-speed diffusion, and, with an additional source term, creation or annihilation of mass. This framework underlies measure differential equations in Euclidean spaces, coupled ODE–MDE systems for structured epidemic models, and dissipative evolution problems in Wasserstein spaces (Piccoli et al., 2018, Gong et al., 2022, Cavagnari et al., 2021).

1. Definition and geometric meaning

Let $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$ be the tangent bundle, with projection $\pi(x,v)=x$ or $\pi_1(x,v)=x$. In the Euclidean measure-theoretic formulations, a PVF is a map assigning to each measure $\mu$ on $\mathbb{R}^n$ a measure on $T\mathbb{R}^n$ such that the base marginal reproduces $\mu$: $$\pi\#V[\mu]=\mu \qquad\text{or}\qquad \pi_1\#V[\mu]=\mu.$$ One formulation uses

$\mu$0

with $\mu$1 the space of positive Borel measures with finite mass and compact support; another uses

$\mu$2

with finite mass and bounded support. In both cases, the defining structural condition is the same: the PVF is a measure-valued section of the tangent bundle (Gong et al., 2022, Piccoli et al., 2018).

Informally, one may disintegrate

$\mu$3

where $\mu$4 is a probability measure on velocities at $\mu$5. The interpretation is that the mass located at $\mu$6 does not necessarily move with a single deterministic velocity; it may split among several velocity channels. This is the core distinction between a PVF and an ordinary vector field.

Classical vector fields embed as a special case. If $\mu$7 is a deterministic vector field, then

$\mu$8

is a PVF concentrated on one velocity at each point. Deterministic transport is therefore the degenerate case in which the fiber measure is Dirac-valued (Piccoli et al., 2018).

2. Measure differential equations and source terms

The basic dynamical object associated with a PVF is the measure differential equation

$\mu$9

Its weak form replaces the classical flux term by integration over the tangent bundle: a curve $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$0 is a solution if $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$1 is constant and, for every $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$2,

$T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$3

This formulation makes explicit that the PVF replaces the classical vector field in the continuity equation by a probability measure on velocities, possibly depending on the full measure $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$4 (Gong et al., 2022).

A larger class of equations incorporates sources and sinks: $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$5 In weak form,

$T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$6

The symbolic $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$7 emphasizes that the right-hand side has two distinct contributions: transport through the PVF and mass variation through the source term. For $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$8, the source controls the variation of total mass, whereas the pure PVF part is conservative (Piccoli et al., 2018).

This formalism unifies several mechanisms that are not simultaneously represented by the classical continuity equation. Because a PVF can attach multiple velocities to mass at a point, it can encode branching, concentration, or finite-speed diffusion-like spreading. Because $T\mathbb{R}^n=\mathbb{R}^n\times\mathbb{R}^n$9 is a separate measure-valued term, the same framework also accommodates creation and annihilation of mass (Piccoli et al., 2018).

3. Wasserstein structure, semigroups, and well-posedness

Analysis of PVF-driven dynamics relies on Wasserstein-type metrics adapted to possibly unequal masses. A central object is the generalized Wasserstein distance

$\pi(x,v)=x$0

which combines a transport cost with the cost of removing or adding unmatched mass. It satisfies

$\pi(x,v)=x$1

and admits the dual characterization

$\pi(x,v)=x$2

To control PVFs themselves, the MDES theory introduces a Wasserstein-like functional on tangent-bundle measures,

$\pi(x,v)=x$3

while the coupled ODE–MDE theory uses an analogous operator $\pi(x,v)=x$4 on PVF values. Both are designed to measure velocity discrepancies along base-space optimal couplings (Piccoli et al., 2018, Gong et al., 2022).

The standard hypotheses are growth and Lipschitz conditions. In the Euclidean MDES setting, the support sublinearity condition is

$\pi(x,v)=x$5

and the local Lipschitz condition on the PVF reads

$\pi(x,v)=x$6

when the supports of $\pi(x,v)=x$7 and $\pi(x,v)=x$8 lie in a common ball. The source is assumed Lipschitz in $\pi(x,v)=x$9 and uniformly support-bounded. Under these assumptions there exists a Lipschitz semigroup $\pi_1(x,v)=x$0 of solutions, with estimates of the form

$\pi_1(x,v)=x$1

together with explicit support growth bounds (Piccoli et al., 2018).

In the coupled ODE–MDE theory, the analogous conclusion is a Lipschitz semigroup $\pi_1(x,v)=x$2 for

$\pi_1(x,v)=x$3

with stability estimate

$\pi_1(x,v)=x$4

The constructive schemes are explicit-Euler-type discretizations: the MDES paper uses Lattice Approximate Solutions, while the coupled ODE–MDE paper uses an Euler–LAS scheme (Piccoli et al., 2018, Gong et al., 2022).

Uniqueness is more delicate than existence. In the MDES theory it is not derived from Lipschitz continuity alone; an additional compatibility with dynamics of atomic measures is imposed through a Dirac germ. A semigroup is required to track the prescribed short-time evolution of finite sums of Dirac masses up to order $\pi_1(x,v)=x$5 in $\pi_1(x,v)=x$6, and under this compatibility there exists at most one Lipschitz semigroup (Piccoli et al., 2018). This distinguishes the PVF setting from classical ODE theory, where local Lipschitz continuity is typically sufficient for uniqueness.

4. Canonical constructions and explicit examples

A standard example is a barycenter-based splitting PVF on $\pi_1(x,v)=x$7. For $\pi_1(x,v)=x$8, define

$\pi_1(x,v)=x$9

The associated PVF assigns velocity $\mu$0 to mass left of the median, velocity $\mu$1 to mass right of the median, and an appropriate split at the median. For the Dirac initial condition $\mu$2, the solution of

$\mu$3

is

$\mu$4

which exhibits finite-speed splitting of a point mass (Gong et al., 2022).

A broader family is quantile-based. Let $\mu$5 be increasing, and for a nonzero measure $\mu$6 define the normalized cumulative distribution

$\mu$7

Then

$\mu$8

where

$\mu$9

At continuity points of the cumulative distribution, the velocity is deterministic; at atoms, the velocity is distributed over a range determined by the quantile interval. This construction is the measure-theoretic mechanism behind finite-speed diffusion in trait space (Gong et al., 2022).

For smooth monotone $\mathbb{R}^n$0 and Dirac initial data, the induced density solution is

$\mathbb{R}^n$1

solving

$\mathbb{R}^n$2

This example shows that PVF dynamics can generate explicit density evolutions while remaining fundamentally measure-valued (Gong et al., 2022).

5. Structured epidemic dynamics and dissipative generalizations

In one application, the infected population is represented by a measure $\mathbb{R}^n$3 on a one-dimensional trait or variant space $\mathbb{R}^n$4, while the susceptible and removed populations remain scalar variables. The coupled system is

$\mathbb{R}^n$5

$\mathbb{R}^n$6

$\mathbb{R}^n$7

Here the PVF term governs transport or mutation in trait space, while the source term describes infection and removal. The total infected mass is

$\mathbb{R}^n$8

When $\mathbb{R}^n$9 and $T\mathbb{R}^n$0 are constant, the PVF preserves mass and the system reduces exactly to the classical SIR equations for $T\mathbb{R}^n$1. With symmetric mutation and $T\mathbb{R}^n$2, $T\mathbb{R}^n$3, the same framework yields a time-dependent SIR system whose coefficients are generated dynamically by the PVF-induced motion in trait space (Gong et al., 2022).

A different generalization replaces single-valued PVFs by multivalued $T\mathbb{R}^n$4-dissipative probability vector fields in the Wasserstein space $T\mathbb{R}^n$5 of Borel probability measures on a Hilbert space $T\mathbb{R}^n$6. In this setting, the evolution equations are studied through a measure-theoretic Explicit Euler scheme, and convergence is proved with optimal error estimates under an abstract CFL stability condition. The analysis does not rely on compactness arguments and also holds when $T\mathbb{R}^n$7 has infinite dimension. Limit solutions are characterized by an Evolution Variational Inequality, and existence, uniqueness, and stability of EVI solutions lead to the generation of a semigroup of nonlinear contractions (Cavagnari et al., 2021).

Taken together, these developments show that PVF-based dynamics range from finite-dimensional Euclidean measure transport with sources to nonlinear dissipative semigroup theory on Wasserstein spaces. A plausible implication is that the PVF formalism is best understood not as a single equation class, but as a family of measure-evolution frameworks sharing the same geometric idea: velocities are encoded at the level of measures rather than points.

6. Related usages, distinctions, and misconceptions

A common misunderstanding is to equate a PVF with a random vector field or an SDE on trajectories. In the measure-dynamical literature, that identification is explicitly rejected: the PVF is deterministic once $T\mathbb{R}^n$8 is given, and the probability is located in the space of velocities at each point, while the evolution is described as an ODE in the space of measures, not as stochastic trajectories. This separates PVFs from random path models, even though both involve probabilistic structure (Piccoli et al., 2018).

A distinct usage appears in canonical quantum mechanics under the name probability vector current density. There the relevant object is a vector field $T\mathbb{R}^n$9 linked to the probability scalar density $\mu$0 by the continuity equation

$\mu$1

Its physical interpretation is local average particle flux. For the nonrelativistic Hamiltonian

$\mu$2

the familiar expression is

$\mu$3

which reduces to

$\mu$4

when the vector potential vanishes. In that literature, the objection that local average particle flux would breach the uncertainty principle is explicitly rebutted (Kauffmann, 2013). This usage is conceptually different from the measure-theoretic PVF on $\mu$5.

The acronym is also unrelated across fields. In federated learning, “PVF” can denote Partial Vector Freezing, a portable module for compressing computation costs in secure aggregation. That meaning is explicitly not “probability vector field” (Zhang et al., 2023). For cross-disciplinary reading, this terminological divergence matters: the acronym alone does not identify the underlying concept.