---
title: 'McKean-Pontryagin: Mean-Field Control & Transport'
url: https://www.emergentmind.com/topics/mckean-pontryagin-approach
type: topic
---

# McKean-Pontryagin: Mean-Field Control & Transport

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 [2603.30019]. 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 [2507.16288] [2510.06543] [2510.00937].

## 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 [2603.30019]. 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 [2507.16288]. 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 [2510.06543].

| Setting | State/adjoint objects | Characteristic optimality object |
|---|---|---|
| Dynamic optimal transport | \(X_t(a)\), \(P_t(a)\), \(\rho_t\) | Constrained Hamiltonian system |
| McKean–Vlasov SPDE control | \(X^\alpha\), \((P,Q)\), \(\mathcal L(X^\alpha_t)\) | Hamiltonian minimization |
| Conditional McKean–Vlasov control | \(\mathbb P^\alpha\), \((Y^\alpha,Z^\alpha)\), conditional law | Generalized McKean–Vlasov BSDE |
| Digital twins | Particle states \(X_t^{(i)}\), co-states \(P_t^{(i)}\) | 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
\[
{\rm d}\tilde X_t = U_t\,{\rm d}t + \sqrt{2}\,\Sigma^{1/2}\,{\rm d}\tilde B_t,
\]
with cost
\[
\mathcal J(U)=\frac12\int_0^T \mathbb E_{\tilde X_t}\|U_t\|_R^2\,{\rm d}t,
\qquad
\|u\|_R^2=u^\top R^{-1}u,
\]
subject to \(\tilde X_0\sim \pi_0\) and \(\tilde X_T\sim \pi_T\). The paper identifies classical optimal transport as the case \(\Sigma=0\) and \(R=I\), and the Schrödinger bridge problem as the case \(\Sigma=R=I\) [2603.30019].

The McKean-Pontryagin reformulation lifts this controlled diffusion to a continuum of labeled trajectories \(X_t(a)\) with labels \(a\) distributed according to \(\rho_0(a)\), together with adjoint variables \(P_t(a)\). The law of \(X_t\) is \(\rho_t\). Initial and terminal marginals are imposed weakly:
\[
\int f(X_0(a))\,\rho_0(a)\,da = \pi_0[f],\qquad
\int f(X_T(a))\,\rho_0(a)\,da = \pi_T[f].
\]
A central structural point is relabeling symmetry: the labels are passive, and only the law matters.

The action functional is
\[
\begin{aligned}
\mathcal S =& \int_0^T \int \Big( \langle P_t,\dot X_t-U_t\rangle -\Sigma : D_x^2\psi_t(X_t) +\frac12\|U_t\|_R^2 \Big)\rho_0(a)\,da\,dt \\
&\quad -\int_0^T \int \langle P_t-\nabla_x\psi_t(X_t),\beta_t\rangle\,\rho_0(a)\,da\,dt \\
&\quad -\int \psi_T(X_T)\rho_0(a)\,da + \pi_T[\psi_T] +\int \psi_0(X_0)\rho_0(a)\,da - \pi_0[\psi_0].
\end{aligned}
\]
Here \(\psi_t\) is a scalar potential, and \(\beta_t\) is a Lagrange-multiplier-like field encoding the freedom associated with the stochastic component or relabeling gauge. The corresponding Hamiltonian is
\[
\mathcal H = \int \Big( \langle P,U\rangle +\Sigma : D_x^2\psi(X) -\frac12\|U\|_R^2 +\langle \beta,\,P-\nabla_x\psi(X)\rangle \Big)\rho_0(a)\,da.
\]

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 [2603.30019].

## 3. Constrained Hamiltonian dynamics and transport equations

Stationarity of the action with respect to \(P,X,\beta,\psi,U\) yields the constrained mean-field Hamiltonian system
\[
\dot X_t = \frac{\delta \mathcal H}{\delta P_t}=U_t+\beta_t,
\]
\[
-\dot P_t = \frac{\delta \mathcal H}{\delta X_t} = \nabla_x(\Sigma : D_x^2\psi_t(X_t)) - D_x^2\psi_t(X_t)\,\beta_t.
\]
The constraints are
\[
0=\frac{\delta\mathcal H}{\delta \beta_t}=P_t-\nabla_x\psi_t(X_t),
\]
\[
0=\frac{\delta\mathcal H}{\delta \psi_t}
= \nabla_x\cdot\Big(\rho_t(X_t)\{\beta_t+\Sigma\nabla_x\log\rho_t(X_t)\}\Big),
\]
\[
0=\frac{\delta\mathcal H}{\delta U_t}=P_t-R^{-1}U_t.
\]
Hence
\[
U_t = R P_t = R\nabla_x\psi_t(X_t).
\]

The \(\psi\)-constraint is obtained by integration by parts in the mean-field variable, using
\[
\int \Sigma : D_x^2\psi_t(X_t(a))\,\rho_0(a)\,da
=
\int \nabla_x\cdot\big(\rho_t(x)\Sigma\nabla_x\log\rho_t(x)\big)\,\psi_t(x)\,dx
\]
and
\[
\int \langle \beta_t(X_t(a)),\nabla_x\psi_t(X_t(a))\rangle\,\rho_0(a)\,da
=
-\int \nabla_x\cdot(\rho_t(x)\beta_t(x))\,\psi_t(x)\,dx.
\]
The boundary conditions are
\[
\rho_0=\pi_0,\qquad \rho_T=\pi_T,\qquad P_0=\nabla_x\psi_0,\qquad P_T=\nabla_x\psi_T.
\]

The same framework unifies deterministic and stochastic transport by tuning \(\Sigma\). In the deterministic regime \(\Sigma\to 0\), \(R=I\), the diffusive term disappears and the formulation reduces to the Benamou–Brenier dynamic optimal transport picture. In the Schrödinger bridge regime \(\Sigma=R=I\), the same variational structure yields a mean-field deterministic representation of dynamic Schrödinger bridge transport [2603.30019].

The law \(\rho_t\) of \(X_t\) satisfies
\[
\partial_t \rho_t = \mathcal L_{u_t}^\dagger \rho_t,
\qquad
\mathcal L_{u_t}f = \langle \nabla_x f, u_t\rangle + \Sigma : D_x^2 f,
\qquad
u_t(x)=R\nabla_x\psi_t(x),
\]
while \(\psi_t\) satisfies
\[
-\partial_t\psi_t = \mathcal L_{u_t}\psi_t - \frac12\|u_t\|_R^2.
\]
The paper also notes the relation to Schrödinger potentials:
\[
\hat\phi_t = e^{\psi_t},\qquad \rho_t=\phi_t\hat\phi_t.
\]

The Hamiltonian is conserved along solutions, with conserved quantity
\[
\mathcal E = \int \Big(\frac12\|U_t\|_R^2+\Sigma:D_x^2\psi_t(X_t)\Big)\rho_0(a)\,da.
\]

## 4. Relation to FBSDEs and extensions beyond classical entropic regularization

A major conceptual point of the transport formulation is that the freedom in \(\beta_t\) can be used to recover a stochastic forward-backward system. Setting formally
\[
\beta_t = \sqrt{2}\,\Sigma^{1/2}\dot B_t,
\]
and interpreting stochastic integrals in the Stratonovich sense, one obtains
\[
{\rm d}X_t = U_t\,dt + d\beta_t,
\]
\[
-\,dP_t = \nabla_x(\Sigma:D_x^2\psi_t(X_t))\,dt - D_x^2\psi_t(X_t)\circ d\beta_t,
\]
with constraints
\[
P_t=\nabla_x\psi_t(X_t),\qquad P_t=R^{-1}U_t.
\]
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 \(\psi_t(x)\) itself remains deterministic, even though \(X_t\) and \(P_t\) depend on the Brownian path [2603.30019].

The same paper makes the approach explicitly nonlocal with respect to classical entropic transport. It extends to
\[
d\tilde X_t = b(\tilde X_t)\,dt + G U_t\,dt + \sqrt2\,\Sigma^{1/2}\,d\tilde B_t,
\]
including underdamped Langevin dynamics and more general state-dependent drifts. It can handle multiplicative noise \(\Sigma(x)\) and mean-field choices such as
\[
R=\Sigma=C(\rho_t),
\]
where \(C(\rho_t)\) is the covariance of \(X_t\), connecting the framework to Kalman–Wasserstein gradient flows.

The scalar potential \(\psi_t\) may be replaced by a vector field \(\Psi_t\), giving the modified Hamiltonian
\[
\mathcal H = \int \Big( \langle P,U\rangle +\Sigma : D_x\Psi(X) -\frac12\|U\|_R^2 +\langle\beta,P-\Psi(X)\rangle \Big)\rho_0(a)\,da.
\]
A quadratic penalty
\[
\mathcal J(\Psi)=\frac{\gamma}{2}\int \|P-\Psi(X)\|^2\,\rho_0(a)\,da
\]
can be added without changing the resulting equations of motion. If one drops the requirement that \(\Psi_t\) be a gradient, the same machinery can treat ideal barotropic fluid dynamics by setting \(\Sigma\equiv 0\) and adding internal energy \(e(\rho_t)\) to the Hamiltonian.

The paper also sketches optimization of the diffusion itself by adding reward
\[
-\frac{\gamma}{2}\|\Sigma\|_F^2.
\]
Formal optimality then gives
\[
\Sigma(x)=\frac1\gamma D_x^2\psi(x),
\]
and the conserved energy becomes
\[
\mathcal E = \int \left( \frac12\|U_t\|_R^2 + \frac{1}{2\gamma}D_x^2\psi_t(X_t):D_x^2\psi_t(X_t) \right)\rho_0(a)\,da.
\]
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 [2603.30019].

## 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
\[
\begin{aligned}
dX^\alpha_t &= L X^\alpha_t\,dt +F\!\left(t,X^\alpha_t,\mathcal L(X^\alpha_t),\alpha_t\right)\,dt \\
&\quad +B\!\left(t,X^\alpha_t,\mathcal L(X^\alpha_t),\alpha_t\right)\,dW_t,
\end{aligned}
\]
with cost
\[
J(\alpha) = \mathbb E\left[\int_0^T f\!\left(t,X^\alpha_t,\mathcal L(X^\alpha_t),\alpha_t\right)\,dt + g\!\left(X^\alpha_T,\mathcal L(X^\alpha_T)\right)\right].
\]
The control-to-state map is shown to be Gâteaux differentiable, with derivative \(Z^{\alpha,\beta}\) solving a linearized McKean–Vlasov SPDE. The adjoint \((P,Q)\) solves a linear McKean–Vlasov backward SPDE, and adjoint calculus yields
\[
dJ(\alpha)\cdot\beta = \mathbb E\left[\int_0^T H_\alpha\!\left(t,X^\alpha_t,\mathcal L(X^\alpha_t),\alpha_t,P_t,Q_t\right)(\beta_t)\,dt\right].
\]
Under convexity of \(\alpha\mapsto(F,B,f)(t,x,\mu,\alpha)\), optimality gives the pointwise Hamiltonian inequality in expectation [2507.16288].

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
\[
dX(t)=AX(t)\,dt+a\bigl(t,X(t),\mathcal L(X(t)),u(t)\bigr)\,dt +b\bigl(t,X(t),\mathcal L(X(t)),u(t)\bigr)\,dW(t),
\]
and the Hamiltonian is
\[
\mathbb H(t,x,\mu,u,y,z) = \langle y,a(t,x,\mu,u)\rangle +\langle z,b(t,x,\mu,u)\rangle_{\mathcal L_2^0} -f(t,x,\mu,u).
\]
The resulting necessary condition contains the operator-valued second adjoint \(P(t)\) 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 [2603.06245].

Conditional McKean–Vlasov control introduces a distinct weak-form variant. The controlled process is stopped at
\[
\tau := \inf\{t\ge 0:\, X_t\notin D\},
\]
and the interaction is through the conditional law of the surviving population:
\[
\mathcal{L}_{\mathbb{P}^\alpha}(X_{\cdot\wedge s}\mid s<\tau).
\]
The Hamiltonian is
\[
h(t,x,a,\mu,p,z) := f(t,x,a,\mu,p)+\beta(t,x,a,\mu,p)^\top z.
\]
For a control \(\alpha\), the adjoint \((Y^\alpha,Z^\alpha)\) solves a generalized BSDE in weak form, and optimality implies
\[
\mathbb{E}^\alpha\!\left[\int_0^{T\wedge\tau} h_a(\Theta_s^\alpha,Z_s^\alpha)^\top(\alpha'_s-\alpha_s)\,ds\right]\ge 0
\]
for every competitor \(\alpha'\). In particular,
\[
\alpha_t\in \arg\min_{a\in A} h(\Theta_t^\alpha,Z_t^\alpha)
\]
\(dt\times\mathbb P\)-a.e. on \(\{t<\tau\}\), and if \(h\) is convex in \(a\), this pointwise condition is also sufficient. The same paper states the verification-type inequality
\[
J(\alpha')-J(\alpha) \ge \frac{m}{2}\, \mathbb{E}^{\alpha'}\!\left[\int_0^{T\wedge\tau}\|\alpha_s-\alpha'_s\|^2ds\right] \ge \frac{m}{L^2}\mathcal{H}(\mathbb{P}^{\alpha'}\|\mathbb{P}^\alpha),
\]
linking suboptimality to relative entropy [2510.06543].

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
\[
\dot{X}_t = b(X_t) + G(X_t)U_t + \Sigma^{1/2}\dot{B}_t,
\]
while observations are either discrete,
\[
Y_{t_n}^\dagger = h(X_{t_n}^\dagger) + \Xi_{t_n}^\dagger,
\qquad \Xi_{t_n}^\dagger \sim \mathcal N(0,R),
\]
or continuous,
\[
\dot{Y}_t^\dagger = h(X_t^\dagger) + R^{1/2}\dot{W}_t.
\]
The cost is the infinite-horizon discounted functional
\[
J(U)=\mathbb E\left[\int_0^\infty e^{-\gamma t} \left(c(X_t)+\frac12\|U_t\|^2\right)\,dt\right].
\]
The paper introduces state particles \(X_t^{(i)}\), co-state particles \(P_t^{(i)}\), ensemble Kalman filtering for assimilation, and a mean-field control law
\[
U_t^\ast = -\int_{\mathbb R^{d_x}} G(x)^\top \psi_t(x)\,\pi_t(x)\,dx,
\]
approximated in particles by
\[
U_t^\ast = \frac{1}{M}\sum_{i=1}^M G(X_t^{(i)})^\top P_t^{(i)}.
\]
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 [2510.00937].

In this setting the co-state is related to a value-gradient representation through
\[
P_t(a)=\psi_t(X_t(a)),
\qquad
\psi_t(x)=\nabla_x v_t(x)
\]
under sufficient regularity. The digital twin uses forward particle evolution, empirical covariances, and Nadaraya–Watson kernel regression,
\[
\psi_t(x)=
\frac{\sum_{i=1}^M
\exp\!\left(-\frac{1}{2\delta}\|x-X_t^{(i)}\|^2\right)P_t^{(i)}}
{\sum_{i=1}^M
\exp\!\left(-\frac{1}{2\delta}\|x-X_t^{(i)}\|^2\right)},
\]
together with a Schrödinger bridge construction for the mean-field generator terms [2510.00937].

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 \(\varepsilon\) and the extraction of the maximum principle from finite-dimensional stationarity conditions [1412.2363]. 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 [2405.04204]. 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 [2203.12828].

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.

Source: https://www.emergentmind.com/topics/mckean-pontryagin-approach