Path-Space Action Formulation

• Path-space action formulation is a framework where the entire trajectory is the key object, optimized using an action functional across classical, stochastic, and quantum domains.
• It explicitly encodes global trajectory geometry, enabling unified analysis for prediction, planning, transport, and control in complex dynamical systems.
• Practical implementations like PDPO transform infinite-dimensional PDE problems into finite-dimensional optimizations with provable error decay and improved computational efficiency.

Path-space action formulation denotes a family of representations in which the primary dynamical object is not an instantaneous state, control, or density, but an entire trajectory, density path, or path measure equipped with an action functional. In the literature considered here, this idea appears in several technically distinct forms: as an action-minimization problem over probability densities and velocity fields in optimal transport and mean-field models; as an Onsager–Machlup functional defining probabilities of stochastic trajectories; as entropy- or KL-based objectives on action-state paths in reinforcement learning; and as the oscillatory action governing quantum path integrals on conventional, modular, noncommutative, and discrete phase spaces (Hernandez et al., 24 May 2025, Kim, 27 Jun 2026, Milton, 2015). The common structural feature is that global trajectory geometry is made explicit, so prediction, planning, transport, or propagation is formulated directly on path space rather than recovered indirectly from local updates.

1. General formal structure

In its classical and quantum-mechanical form, the action of a path $x(t)$ on $[t_i,t_f]$ is

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$

and the propagator is formally represented as

$$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$

with the functional measure defined heuristically by time-slicing and insertion of position resolutions of the identity (Milton, 2015). In Schwinger’s formulation, infinitesimal variations of dynamics or endpoints generate corresponding variations of transition amplitudes, and iterating over slices yields the path-integral representation; in the $\hbar \to 0$ limit, stationary paths satisfying the Euler–Lagrange equation dominate (Milton, 2015).

In stochastic dynamics and world models, the same path-space viewpoint takes a probabilistic rather than oscillatory form. For an effective Markovian latent SDE

$dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$

the path law is written as

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$

with Onsager–Machlup action

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$

so that the action is the negative log-density of a path up to normalization (Kim, 27 Jun 2026).

A further generalization appears in curved-space time-slicing. For second-order generators, a valid short-time representation is one reproducing normalization, mean drift, and covariance to $O(\Delta t)$, and there are infinitely many such equivalent discretizations. These discrete actions are asymptotically Gaussian, organized into a one-parameter $\alpha\in[0,1]$ family, with $[t_i,t_f]$0 singled out as Gaussian and invariant provided the increment transforms according to Itô’s formula under nonlinear changes of variables (Ding et al., 2021). This establishes that “path-space action” is not a unique formula but a structural principle whose concrete realization depends on whether paths are weighted by variational cost, probability, or quantum phase.

2. Dynamic action over density paths

A particularly explicit path-space action formulation is given by "PDPO: Parametric Density Path Optimization" (Hernandez et al., 24 May 2025). The starting point is a dynamic action over density paths connecting endpoint measures $[t_i,t_f]$1:

$[t_i,t_f]$2

subject to the continuity equation

$[t_i,t_f]$3

In the Benamou–Brenier case, $[t_i,t_f]$4 and $[t_i,t_f]$5, so the action reduces to kinetic energy (Hernandez et al., 24 May 2025).

The potential functional $[t_i,t_f]$6 can incorporate several classes of effects:

$[t_i,t_f]$7

These terms model, respectively, an external obstacle potential, internal energy such as entropy or Fisher information, and mean-field interaction (Hernandez et al., 24 May 2025).

The central reduction in PDPO is to represent the density path as a parametric pushforward of a reference density $[t_i,t_f]$8 through a smooth map $[t_i,t_f]$9. For a $$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$0 parameter curve $$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$1,

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$2

This transforms the original infinite-dimensional optimization over $$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$3 into a finite-dimensional optimization over the parameter path $$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$4 (Hernandez et al., 24 May 2025).

The induced action becomes

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$5

Using

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$6

PDPO obtains a finite-dimensional Lagrangian

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$7

with

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$8

The kinetic geometry is encoded by the pullback matrix

$$S[x] \equiv \int_{t_i}^{t_f} dt\,L(x(t),\dot x(t),t),$$9

so that

$$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$0

In this formulation, the path-space action is literally transferred from density space to parameter space (Hernandez et al., 24 May 2025).

3. Static parameter-path approximation and error control

PDPO solves the parameter-path problem through cubic-Hermite spline interpolation in parameter space. Choosing $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$1 control points $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$2 at uniform times $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$3 and the standard cubic-Hermite basis $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$4, the parameter path is approximated by

$$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$5

The action then becomes a differentiable static objective in finitely many control variables:

$$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$6

In practice, the time integral is approximated by quadrature, for example trapezoidal quadrature at $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$7 points, and the expectation over $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$8 by Monte Carlo with $$\langle x_f,t_f \mid x_i,t_i\rangle = \int D[x(t)]\,\exp\!\Bigl(\frac{i}{\hbar}S[x]\Bigr),$$9 samples (Hernandez et al., 24 May 2025).

The approximation error is controlled theoretically. If the true minimizing curve $\hbar \to 0$0 is $\hbar \to 0$1, with $\hbar \to 0$2 and $\hbar \to 0$3, then under boundedness and Lipschitz assumptions on $\hbar \to 0$4 and $\hbar \to 0$5,

$\hbar \to 0$6

where $\hbar \to 0$7 is the cubic-Hermite interpolant. In particular, if $\hbar \to 0$8, the action error decays as $\hbar \to 0$9 (Hernandez et al., 24 May 2025).

The empirical regime reported for PDPO is notably low-dimensional in control variables: using $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$0–$dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$1 control points of the spline interpolation suffices to accurately resolve both multimodal and high-dimensional problems (Hernandez et al., 24 May 2025). The same framework accommodates obstacle potentials, mean-field interactions, Fisher-information terms for stochastic control, and higher-order dynamics. For instance, a Schrödinger bridge-type contribution can be written as

$dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$2

and higher-order dynamics arise by replacing $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$3 with $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$4 (Hernandez et al., 24 May 2025).

Relative to classical Benamou–Brenier OT and Schrödinger bridge solvers, PDPO replaces an infinite-dimensional PDE-constrained problem by a static finite-dimensional nonlinear program over spline control points. The paper states that cubic splines require only $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$5 control variables, often $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$6–$dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$7, eliminate time-marching PDE solvers, and combine Monte Carlo with ODE or Neural ODE pushforwards to scale to high dimension; empirically, PDPO outperforms existing state-of-the-art approaches in benchmark tasks in both computational efficiency and solution quality (Hernandez et al., 24 May 2025).

4. Stochastic path measures, prediction, and irreversibility

In stochastic world models, the path-space action is used to unify prediction, planning, and uncertainty. "A Path-Space Formulation of Prediction in World Models" models latent dynamics by an Itô–Stratonovich SDE and assigns to each smooth path an Onsager–Machlup action. The most probable trajectory is obtained by minimizing $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$8 with a free terminal point; planning imposes a fixed endpoint; and uncertainty is captured by the second variation operator around the instanton path (Kim, 27 Jun 2026).

For constant diffusion $dz_t = f(z_t)\,dt + \sqrt{2D}\,dW_t,$9, the Euler–Lagrange equation of the Onsager–Machlup action yields

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$0

where

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$1

The paper emphasizes that the most-probable path generally differs from the deterministic mean-field rollout $P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$2, except in the special divergence-harmonic case $P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$3 (Kim, 27 Jun 2026). This directly counters the common simplification that “trajectory prediction” in stochastic latent models reduces to a deterministic rollout.

A further structural decomposition writes the drift as

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$4

separating reversible and irreversible components. The time-antisymmetric contribution to the action is

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$5

and path-wise entropy production is

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$6

The steady-state entropy-production rate can be estimated from rollouts through empirical drift, diffusion, stationary density, and current reconstruction (Kim, 27 Jun 2026).

The same paper reports that in controlled small-scale attention-based models, attention asymmetry is acquired during training in proportion to the irreversibility of the data. Symmetrizing the learned attention suppresses entropy production and selectively degrades long-horizon prediction of irreversible dynamics while preserving relaxational prediction (Kim, 27 Jun 2026). This suggests that, in this setting, path-space action is not merely descriptive; it also identifies an operational link between architecture, irreversibility, and predictive competence.

5. Reinforcement learning and control on path space

Path-space formulations in reinforcement learning appear in at least two distinct ways in the surveyed literature: as an intrinsic objective over action-state trajectories and as a proximal regularization principle for generative policies.

"Complex behavior from intrinsic motivation to occupy action-state path space" defines a trajectory of length $P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$7 as

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$8

with path distribution

$P[x(\cdot)] = \frac{1}{Z}e^{-A[x(\cdot)]},$9

Imposing smoothness, monotonicity in transition probability, and additivity over two-step paths yields the unique occupancy gain

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$0

and therefore the discounted path-space entropy

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$1

The associated value function satisfies a Bellman equation with intrinsic reward

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$2

and the optimal value obeys

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$3

The paper proves convergence of its value-iteration scheme and illustrates behaviors such as four-room exploration, hide-and-seek, cartpole “dancing,” altruism via path entropy, and robust locomotion in a high-dimensional quadruped under zero external reward (Ramírez-Ruiz et al., 2022).

A different path-space control formulation is given by "Proximal Policy Optimization in Path Space: A Schrödinger Bridge Perspective" (Gong et al., 23 Mar 2026). There the policy is a distribution over denoising trajectories

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$4

with path law

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$5

The path-space likelihood ratio is

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$6

This yields a path surrogate objective, from which the paper develops two concrete variants: GSB-PPO-Clip and GSB-PPO-Penalty. The penalty formulation uses a quadratic path-space regularizer based on drift mismatch,

$$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$7

and the experimental summary states that the penalty formulation consistently delivers better stability and performance than the clipping counterpart (Gong et al., 23 Mar 2026).

Taken together, these works indicate that path-space objectives in RL need not be tied to external reward maximization or one-step action densities. This suggests a broader view in which entropy, occupancy, and proximal regularization can all be defined on full trajectory laws rather than on marginal action distributions.

6. Quantum and generalized geometric realizations

Quantum-mechanical path-space actions remain the canonical archetype, but the surveyed literature shows substantial variation in the underlying path space.

In modular polarization, the path space is a torus $$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$8, a compact phase space of volume $$A[x(\cdot)] = \int_0^T dt\,L(x(t),\dot x(t)), \qquad L(z,\dot z)=\frac{1}{4D}\|\dot z-f(z)\|^2+\frac12\nabla\!\cdot f(z),$$9. Repeating Feynman’s time-slicing derivation for the harmonic oscillator produces a modular propagator with winding-number sum and Aharonov–Bohm-type phase,

$O(\Delta t)$0

where the modular Lagrangian is

$O(\Delta t)$1

The action exhibits explicit phase-space translations, time translations, and hidden symplectic rotations, and a modular Legendre-transform prescription is proposed for more general Hamiltonians (Yargic, 2020).

On the noncommutative plane, the path-space action of a charged particle in a magnetic field is derived by time-slicing in a Hilbert–Schmidt operator formalism. The resulting action contains an explicitly nonlocal time-derivative operator,

$O(\Delta t)$2

and supports explicit derivations of equations of motion, ground-state energies, and the Aharonov–Bohm phase (Gangopadhyay et al., 2013). In noncommutative phase-space, a related construction yields an action with magnetic-field-like terms and second-class constraints whose Dirac brackets recover the noncommutative Heisenberg algebra (Gangopadhyay et al., 2016).

A discrete analog appears in finite-dimensional quantum mechanics on $O(\Delta t)$3. There the exact discrete phase-space propagator is expressed as a sum over paths weighted by a discrete action

$O(\Delta t)$4

For affine Hamiltonians and strictly commensurate times, the fluctuation sum collapses to a deterministic shift realizing a discrete analog of classical Hamiltonian flow; by contrast, in the interacting two-qutrit example, the $O(\Delta t)$5 sector alone is non-real at finite time step and becomes trivial in the continuum limit, so coherent summation over all fluctuation sectors is necessary to reproduce entanglement dynamics (Pachon et al., 22 Apr 2026).

Further extensions include genuinely complex actions, where non-Hermitian operators $O(\Delta t)$6 and $O(\Delta t)$7 with complex spectra are used to define

$O(\Delta t)$8

on complexified phase-space contours (Nagao et al., 2011), and quasi-Hermitian position-deformed Heisenberg algebras, where a Dyson map restores Hermiticity and leads to a deformed Lagrangian

$O(\Delta t)$9

with corresponding Euclidean action and free-particle propagator (Katsekpor et al., 2024).

7. Comparative perspective and recurrent misconceptions

The surveyed literature suggests that path-space action formulations play at least three mathematically distinct roles. First, the action may be a variational cost to be minimized, as in density-path optimization and planning (Hernandez et al., 24 May 2025). Second, it may be the negative log-density of a stochastic path measure, as in Onsager–Machlup theory and entropy production (Kim, 27 Jun 2026). Third, it may appear as the phase in an oscillatory integral, as in quantum propagation (Milton, 2015, Yargic, 2020). Treating these uses as interchangeable is a category error.

A second recurrent misconception is that discretization is merely technical. In curved space, the literature explicitly states that there is no consensus on covariant time-slicing, and the $\alpha\in[0,1]$0-family of equivalent discrete actions shows that different slice actions can be continuum-equivalent while differing by spurious drift terms; only the $\alpha\in[0,1]$1 representation is manifestly scalar under nonlinear variable changes with Itô-transformed increments (Ding et al., 2021). In discrete phase-space quantum mechanics, truncating to a single fluctuation sector fails to reproduce the exact dynamics (Pachon et al., 22 Apr 2026). In generative RL, path-ratio clipping can be unstable because the ratio is a product over many denoising steps, whereas a quadratic path penalty yields smoother updates (Gong et al., 23 Mar 2026).

A third misconception is that path-space formulations necessarily increase dimensionality without improving structure. PDPO provides the opposite example: by lifting density dynamics to a parametric manifold and interpolating the parameter path, it trades an infinite-dimensional PDE-constrained problem for a finite-dimensional differentiable optimization with provable spline error decay (Hernandez et al., 24 May 2025). Conversely, in stochastic world models, the path-space formulation exposes information that is invisible in one-step conditionals, notably the decomposition into reversible and irreversible drift and the associated entropy production (Kim, 27 Jun 2026).

Overall, path-space action formulation is best understood not as a single theory but as a unifying methodology. It recasts dynamics, transport, control, and propagation at the level of whole paths, making global constraints, symmetries, irreversibility, and fluctuation structure analytically explicit. The technical content varies sharply across optimal transport, stochastic processes, reinforcement learning, and quantum mechanics, but the underlying commitment is the same: the fundamental object is a trajectory-level functional rather than a local update rule.