McKean-Pontryagin: Mean-Field Control & Transport

• McKean-Pontryagin is a mean-field framework that combines state dynamics with Pontryagin adjoint calculus for optimal control and transport.
• It reformulates stochastic sampling into a deterministic variational principle yielding a constrained Hamiltonian system that bridges classical optimal transport and Schrödinger bridge problems.
• The approach extends to diverse applications, including McKean–Vlasov SPDE control and digital twin algorithms, demonstrating versatility in handling law-dependent dynamics.

The McKean-Pontryagin approach denotes a family of optimal-control and dynamic-transport formulations in which a McKean or mean-field description of the state is combined with Pontryagin-style adjoint calculus. In the formulation developed for entropic-regularized dynamic optimal transport, the classical problem of optimizing over stochastic sample paths is reformulated as a fully variational mean-field system for a representative trajectory $X_t(a)$ and an adjoint $P_t(a)$, indexed by a label $a$, with stationarity generating constrained Hamiltonian equations, density evolution, and a Hamilton–Jacobi–Bellman structure (Reich, 31 Mar 2026). In adjacent literature, the same combination of mean-field dependence and Pontryagin duality appears in controlled McKean–Vlasov SPDEs, weak-form conditional McKean–Vlasov control with killing, and forward-in-time particle formulations for partially observed digital twins (Spille et al., 22 Jul 2025, Carmona et al., 8 Oct 2025, Opper et al., 1 Oct 2025).

1. Conceptual scope

At its most specific, the approach is a mean-field optimality principle in which probability transport or control is represented through a continuum of deterministic labels, empirical laws, or law-dependent coefficients, while optimality is encoded by adjoint variables, Hamiltonians, and pointwise minimization or maximization conditions. In the dynamic optimal transport setting, the law $\rho_t$ of $X_t$ is part of the state description, and the adjoint relation $P_t=\nabla_x\psi_t(X_t)$ ties the representative trajectory to a scalar potential (Reich, 31 Mar 2026). In McKean–Vlasov reaction-diffusion SPDEs, the state equation depends on $\mathcal L(X_t)$, the adjoint is a backward SPDE, and the first variation of the cost is represented by the Hamiltonian derivative in the control variable (Spille et al., 22 Jul 2025). In conditional McKean–Vlasov control, the mean-field state becomes the pair

$$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$

and the adjoint is a generalized McKean–Vlasov BSDE in weak form (Carmona et al., 8 Oct 2025).

Setting	State/adjoint objects	Characteristic optimality object
Dynamic optimal transport	$X_t(a)$, $P_t(a)$, $P_t(a)$0	Constrained Hamiltonian system
McKean–Vlasov SPDE control	$P_t(a)$1, $P_t(a)$2, $P_t(a)$3	Hamiltonian minimization
Conditional McKean–Vlasov control	$P_t(a)$4, $P_t(a)$5, conditional law	Generalized McKean–Vlasov BSDE
Digital twins	Particle states $P_t(a)$6, co-states $P_t(a)$7	Forward online control law

The term is not used with a single invariant meaning across all cited works. Some papers use “McKean–Pontryagin” for genuinely law-dependent dynamics, while others use “McKean-Pontryagin style” for a finite-dimensional reduction or an adjoint-based reformulation. This suggests that the expression functions both as a description of mean-field structure and as a methodological label for Pontryagin arguments driven by representative trajectories, particle systems, or auxiliary finite-dimensional parametrizations.

2. Mean-field variational reformulation of dynamic optimal transport

In the dynamic optimal transport note, the starting point is the controlled diffusion

$P_t(a)$8

with cost

$P_t(a)$9

subject to $a$0 and $a$1. The paper identifies classical optimal transport as the case $a$2 and $a$3, and the Schrödinger bridge problem as the case $a$4 (Reich, 31 Mar 2026).

The McKean-Pontryagin reformulation lifts this controlled diffusion to a continuum of labeled trajectories $a$5 with labels $a$6 distributed according to $a$7, together with adjoint variables $a$8. The law of $a$9 is $\rho_t$0. Initial and terminal marginals are imposed weakly: $\rho_t$1 A central structural point is relabeling symmetry: the labels are passive, and only the law matters.

The action functional is

$\rho_t$2

Here $\rho_t$3 is a scalar potential, and $\rho_t$4 is a Lagrange-multiplier-like field encoding the freedom associated with the stochastic component or relabeling gauge. The corresponding Hamiltonian is

$\rho_t$5

This formulation replaces explicit optimization over stochastic paths by a deterministic variational principle on a mean-field space. The paper presents this as attractive because it avoids explicit Monte Carlo optimization over stochastic paths, yields a constrained Hamiltonian system, unifies deterministic and stochastic transport, and connects naturally to HJB and FBSDE formulations (Reich, 31 Mar 2026).

3. Constrained Hamiltonian dynamics and transport equations

Stationarity of the action with respect to $\rho_t$6 yields the constrained mean-field Hamiltonian system

$\rho_t$7

$\rho_t$8

The constraints are

$\rho_t$9

$X_t$0

$X_t$1

Hence

$X_t$2

The $X_t$3-constraint is obtained by integration by parts in the mean-field variable, using

$X_t$4

and

$X_t$5

The boundary conditions are

$X_t$6

The same framework unifies deterministic and stochastic transport by tuning $X_t$7. In the deterministic regime $X_t$8, $X_t$9, the diffusive term disappears and the formulation reduces to the Benamou–Brenier dynamic optimal transport picture. In the Schrödinger bridge regime $P_t=\nabla_x\psi_t(X_t)$0, the same variational structure yields a mean-field deterministic representation of dynamic Schrödinger bridge transport (Reich, 31 Mar 2026).

The law $P_t=\nabla_x\psi_t(X_t)$1 of $P_t=\nabla_x\psi_t(X_t)$2 satisfies

$P_t=\nabla_x\psi_t(X_t)$3

while $P_t=\nabla_x\psi_t(X_t)$4 satisfies

$P_t=\nabla_x\psi_t(X_t)$5

The paper also notes the relation to Schrödinger potentials: $P_t=\nabla_x\psi_t(X_t)$6

The Hamiltonian is conserved along solutions, with conserved quantity

$P_t=\nabla_x\psi_t(X_t)$7

4. Relation to FBSDEs and extensions beyond classical entropic regularization

A major conceptual point of the transport formulation is that the freedom in $P_t=\nabla_x\psi_t(X_t)$8 can be used to recover a stochastic forward-backward system. Setting formally

$P_t=\nabla_x\psi_t(X_t)$9

and interpreting stochastic integrals in the Stratonovich sense, one obtains

$\mathcal L(X_t)$0

$\mathcal L(X_t)$1

with constraints

$\mathcal L(X_t)$2

Rewritten in Itô form, the system becomes formally equivalent to standard FBSDEs of stochastic optimal control. The important distinction emphasized in the paper is that $\mathcal L(X_t)$3 itself remains deterministic, even though $\mathcal L(X_t)$4 and $\mathcal L(X_t)$5 depend on the Brownian path (Reich, 31 Mar 2026).

The same paper makes the approach explicitly nonlocal with respect to classical entropic transport. It extends to

$\mathcal L(X_t)$6

including underdamped Langevin dynamics and more general state-dependent drifts. It can handle multiplicative noise $\mathcal L(X_t)$7 and mean-field choices such as

$\mathcal L(X_t)$8

where $\mathcal L(X_t)$9 is the covariance of $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$0, connecting the framework to Kalman–Wasserstein gradient flows.

The scalar potential $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$1 may be replaced by a vector field $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$2, giving the modified Hamiltonian

$$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$3

A quadratic penalty

$$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$4

can be added without changing the resulting equations of motion. If one drops the requirement that $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$5 be a gradient, the same machinery can treat ideal barotropic fluid dynamics by setting $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$6 and adding internal energy $$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$7 to the Hamiltonian.

The paper also sketches optimization of the diffusion itself by adding reward

$$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$8

Formal optimality then gives

$$(\mu_s,p_s)=\left(\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau),\,\mathbb{P}^\alpha[s<\tau]\right),$$9

and the conserved energy becomes

$X_t(a)$0

Accordingly, the approach is presented not as a device tied only to entropic regularization, but as a variational language for a broader class of mean-field control and transport models (Reich, 31 Mar 2026).

5. McKean–Vlasov control and SPDE generalizations

In the McKean–Vlasov SPDE literature, the McKean-Pontryagin approach takes the form of law-dependent state equations coupled to adjoint backward equations. For reaction-diffusion SPDEs of McKean–Vlasov type, the controlled state equation is

$X_t(a)$1

with cost

$X_t(a)$2

The control-to-state map is shown to be Gâteaux differentiable, with derivative $X_t(a)$3 solving a linearized McKean–Vlasov SPDE. The adjoint $X_t(a)$4 solves a linear McKean–Vlasov backward SPDE, and adjoint calculus yields

$X_t(a)$5

Under convexity of $X_t(a)$6, optimality gives the pointwise Hamiltonian inequality in expectation (Spille et al., 22 Jul 2025).

For McKean–Vlasov SPDEs with nonconvex control sets, a separate formulation uses spike variation, second-order Taylor expansion, Lions derivatives in infinite dimension, and transposition or relaxed transposition solutions for the first- and second-order adjoint equations. The state evolves as

$X_t(a)$7

and the Hamiltonian is

$X_t(a)$8

The resulting necessary condition contains the operator-valued second adjoint $X_t(a)$9 and the control-dependent diffusion increment, giving a Pontryagin-type inequality that generalizes the classical stochastic maximum principle to an infinite-dimensional mean-field setting (Chen et al., 6 Mar 2026).

Conditional McKean–Vlasov control introduces a distinct weak-form variant. The controlled process is stopped at

$P_t(a)$0

and the interaction is through the conditional law of the surviving population: $P_t(a)$1 The Hamiltonian is

$P_t(a)$2

For a control $P_t(a)$3, the adjoint $P_t(a)$4 solves a generalized BSDE in weak form, and optimality implies

$P_t(a)$5

for every competitor $P_t(a)$6. In particular,

$P_t(a)$7

$P_t(a)$8-a.e. on $P_t(a)$9, and if $P_t(a)$00 is convex in $P_t(a)$01, this pointwise condition is also sufficient. The same paper states the verification-type inequality

$P_t(a)$02

linking suboptimality to relative entropy (Carmona et al., 8 Oct 2025).

Across these variants, the defining feature is not a single canonical equation but the recurrent triad of law-dependent state dynamics, an adjoint object defined by backward or dual evolution, and a Hamiltonian optimality condition.

6. Forward-time particle formulations, partial observation, and broader usage

For partially observed control in digital twins, the approach is cast as a forward-in-time mean-field reformulation of stochastic optimal control combined with data assimilation. The physical twin follows

$P_t(a)$03

while observations are either discrete,

$P_t(a)$04

or continuous,

$P_t(a)$05

The cost is the infinite-horizon discounted functional

$P_t(a)$06

The paper introduces state particles $P_t(a)$07, co-state particles $P_t(a)$08, ensemble Kalman filtering for assimilation, and a mean-field control law

$P_t(a)$09

approximated in particles by

$P_t(a)$10

The method is presented as avoiding direct solution of a high-dimensional HJB equation or a classical stochastic two-point boundary-value problem, and as allowing online assimilation of data together with online computation of control laws (Opper et al., 1 Oct 2025).

In this setting the co-state is related to a value-gradient representation through

$P_t(a)$11

under sufficient regularity. The digital twin uses forward particle evolution, empirical covariances, and Nadaraya–Watson kernel regression,

$P_t(a)$12

together with a Schrödinger bridge construction for the mean-field generator terms (Opper et al., 1 Oct 2025).

The literature in the data also shows looser uses of the label. In a proof of the classical Pontryagin maximum principle by packages of needle variations, the “key McKean-Pontryagin-style move” is the finite-dimensional reduction from controls to needle widths $P_t(a)$13 and the extraction of the maximum principle from finite-dimensional stationarity conditions (Dmitruk et al., 2014). For control in elliptic coefficients, the data describe a “McKean-Pontryagin-style approach” based on perturbing the coefficient on a shrinking set, rescaling, deriving a topological derivative, and obtaining a pointwise Pontryagin inequality without continuity assumptions on coefficients or gradients (Wachsmuth, 2024). For controllability metrics maximization under sparsity constraints, a “matrix-valued Pontryagin maximum principle” is applied to the controllability Lyapunov differential equation, and the PMP is used to prove that a relaxed sparse scheduling problem is exact (Ohtsuka et al., 2022).

A common misconception is therefore to treat “McKean-Pontryagin” as synonymous with one specific forward-backward SDE formalism. The cited literature does not support such a restriction. In some works the primary variational object is a deterministic mean-field Hamiltonian system; in others it is a generalized McKean–Vlasov BSDE in weak form; in others still it is a forward-only particle algorithm or an analogical finite-dimensional reduction. What remains stable is the use of mean-field or representative-state structure together with Pontryagin-type adjoint optimality conditions.