---
title: Optimal Transport Potential Networks
url: https://www.emergentmind.com/topics/optimal-transport-potential-networks
type: topic
---

# Optimal Transport Potential Networks

Optimal Transport Potential Networks denote a family of transport formulations in which the primary unknowns are potentials—nodal pressures, edge and vertex dual variables, Kantorovich or Brenier potentials, or spatio-temporal Hamilton–Jacobi value functions—and the transport object of interest is induced from them as a flow, map, plan, or weighted network. Across the literature, this perspective appears in potential-driven conductance optimization on graphs, dynamic optimal transport on networks, statistical-mechanical transport ensembles, and neural methods that parameterize scalar OT potentials directly [0906.0006]. The unifying structure is that transport is not represented only by a coupling or by a map \(T\), but by a potential field whose gradients, differences, traces, or dual transforms determine admissible movement, dissipation, or assignment.

## 1. Potential-based transport as a common formalism

In graph and network optimization, potentials arise as physical state variables. In the conductance model for optimal transport networks inspired by leaf venation, each node \(k\) carries a scalar potential \(V_k\), and the edge current is
\[
I_{kj}=C_{kj}(V_k-V_j),
\]
with Kirchhoff conservation
\[
\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.
\]
The dissipative objective is the total power loss
\[
P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,
\]
under the material constraint
\[
\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.
\]
This is explicitly a potential network: conductances are optimized, but flows are induced by nodal potential differences [0906.0006].

In dynamic optimal transport on a network, the potential variables are dual edge functions \(\phi_t^j\) and vertex functions \(\psi_t^i\). They satisfy Hamilton–Jacobi-type inequalities,
\[
\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,
\]
\[
\partial_t\psi_t^i+\frac{1}{2\kappa^2} \left| \psi_t^i-\frac1{|Z(V^i)|}\sum_{j\in Z(V^i)}\phi_t^j \right|^2\le 0,
\]
and induce optimal fluxes through
\[
F_t^j=\rho_t^j\,\partial_x\phi_t^j,\qquad
f_t^i=\kappa^{-2}\gamma_t^i \left( \psi_t^i-\frac1{|Z(V^i)|}\sum_{j\in Z(V^i)}\phi_t^j \right).
\]
Here the network topology enters through incidence relations and potential traces at the vertices [2101.03415].

In the mesoscopic measure-based model of transportation networks, the potential is the pressure \(p\) solving
\[
-\operatorname{div}\bigl((P[\mu]+rI)\nabla p\bigr)=S,
\]
with
\[
P[\mu](x)=\int_{\mathbb{S}_+^d} C\,\mu(x,dC).
\]
The network state is a probability measure \(\mu\in\mathcal{P}(\Omega\times\mathbb{S}_+^d)\), but the actual flux is still potential-driven,
\[
J = -(P[\mu]+rI)\nabla p.
\]
The associated energy
\[
E[\mu] = \int_X \left(\nabla p[\mu]\cdot C\nabla p[\mu] + \frac{\nu}{\gamma}|C|^\gamma \right)\,d\mu \;+\; \int_\Omega r\,|\nabla p[\mu]|^2\,dx
\]
combines pumping and metabolic terms under local mass conservation [2401.07922].

A common misconception is that OT potentials must always mean Brenier potentials. The cited literature shows a broader usage: potentials can be physical pressures \(V_k\) or \(p\), graph-dual Hamilton–Jacobi variables \(\phi,\psi\), generalized source–sink multipliers \(t_i,\theta_\alpha\), or neural scalar functions \(f_\omega,u_\theta\) [2504.10444].

## 2. Network optimization, dissipation, and loop formation

In the classical conductance-budget problem on a graph, the geometry of optimal transport networks depends on the exponent \(\gamma\). For the standard fixed-load problem reviewed in the leaf-venation study, the optimum is a tree for \(\gamma<1\) and a spatially distributed sheet for \(\gamma>1\), with a transition at \(\gamma=1\). The paper modifies the objective to average either over single-edge failures or over sparse fluctuating loads, while keeping the same conductance-budget constraint [0906.0006].

The damage model removes a single edge \((a,b)\) by
\[
C^{ab}_{kj}=C_{kj}\bigl(1-\delta_{ak}\delta_{bj}-\delta_{aj}\delta_{bk}\bigr),
\]
recomputes the potentials \(V_k^{ab}\), and sums the dissipations:
\[
R=\sum_{(ab)} P^{ab}.
\]
If deleting an edge disconnects the graph, \(P^{ab}\) diverges and \(R=\infty\). Consequently, any finite optimum must be one-edge-fault-tolerant and cannot be a tree. The paper emphasizes that the optimum contains more loops than the minimum needed for connectivity, because minimizing \(R\) requires transport redundancy, not only topological redundancy [0906.0006].

The fluctuating-load model activates one sink at a time,
\[
\mathcal{I}_k^a=\delta_{0k}-\delta_{ak},
\]
and minimizes the averaged dissipation
\[
F=\sum_a P^a.
\]
This again favors loops. The paper attributes the failure of the standard tree-optimality arguments to the nonlinearity of the conductance budget
\[
\sum C_{kj}^{\gamma}=\text{const},
\]
which makes optimization over averaged load states inequivalent to optimization for the average load [0906.0006].

The main computational findings are topological as well as energetic. In the damage model, the \(\gamma<1\) phase remains hierarchical but is no longer a tree; it contains extensive anastomosis and nested loops. In the fluctuating-sink model, the sharp phase transition of the standard formulation is replaced by a crossover. The bond-damage case still exhibits a cusp at \(\gamma=1\) in normalized dissipation plots, whereas the moving-sink case smooths this into a crossover. The paper also reports that damage-optimized networks retain hierarchical or fractal characteristics up to around \(\gamma<1\), while in the sink-fluctuation case such features disappear already for \(\gamma \gtrsim 0.5\) [0906.0006].

This establishes an important point for potential-network theory: once the objective is extended from single-state dissipation minimization to a multi-state average over damage or demand, loopy optimal transport structures arise naturally within the same conductance–potential formalism [0906.0006].

## 3. Statistical and dynamic OT on networks

A distinct potential-based formulation appears in the bipartite maximum-entropy model that interpolates between an entropic transport ensemble and classical cost-minimizing OT. The state is a weighted bipartite graph with row and column constraints
\[
\sum_{\alpha=1}^M w_{i\alpha}=s_i,\qquad \sum_{i=1}^N w_{i\alpha}=\sigma_\alpha,
\]
enforced in expectation. The free-energy-like variational problem maximizes
\[
\mathrm{F}[P]=\mathrm{S}[P]-\beta \mathrm{U}[P],
\]
with transport energy
\[
U(G)=\sum_{i,\alpha} w_{i\alpha}(G)\,C_{i\alpha}.
\]
Introducing multipliers \(t_i\) and \(\theta_\alpha\) yields a factorized ensemble in which edge weights are independent exponential variables with rates
\[
r_{i\alpha}=\beta C_{i\alpha}+t_i+\theta_\alpha,
\]
and expected transport intensity
\[
\langle w_{i\alpha}\rangle = \frac{1}{\beta C_{i\alpha}+t_i+\theta_\alpha}.
\]
The self-consistency conditions are
\[
s_i^*=\sum_\alpha \frac{1}{\beta C_{i\alpha}+t_i^*+\theta_\alpha^*}, \qquad \sigma_\alpha^*=\sum_i \frac{1}{\beta C_{i\alpha}+t_i^*+\theta_\alpha^*}.
\]
The paper interprets \(t_i,\theta_\alpha\) as generalized transport potentials or generalized network chemical potentials [2504.10444].

This formulation is explicitly contrasted with standard entropically regularized OT. Sinkhorn-type couplings have multiplicative Gibbs form,
\[
\pi_{i\alpha}=u_i e^{-C_{i\alpha}/\varepsilon}v_\alpha,
\]
whereas the present model has a rational additive form,
\[
\langle w_{i\alpha}\rangle=\frac{1}{\beta C_{i\alpha}+t_i+\theta_\alpha}.
\]
The paper therefore presents a potential network not through \(c\)-transforms or convex gradients, but through additive effective energies controlled jointly by node potentials and edge costs [2504.10444].

The same paper quantifies a dense-to-sparse transition using the mass share of the maximum spanning tree of the average network,
\[
m_{MST}(\beta) = \frac{\sum_{(i,\alpha)\in MST}\langle w_{i\alpha}\rangle} {\sum_{i,\alpha}\langle w_{i\alpha}\rangle}.
\]
As \(\beta\) increases, mass condenses onto low-cost links and the support approaches a sparse OT-like backbone. Finite-size analysis gives an estimated thermodynamic transition value approximately
\[
\ln \beta_{\mathrm t}^* \simeq 3.15(9)
\]
for the main case studied, while the transition is described as apparently non-critical [2504.10444].

Dynamic network OT provides yet another potential structure. On a finite connected planar network \(\mathcal G=(\mathcal V,\mathcal E)\), mass lives both on edges and at vertices, and the action is
\[
\sum_{j=1}^m \iint_{\overline{E^j}\times[0,1]} \frac{|F_t^j|^2}{2\rho_t^j}\,dx\,dt + \kappa^2\sum_{i=1}^n \int_0^1 \frac{|f_t^i|^2}{2\gamma_t^i}\,dt.
\]
The paper proves a duality formula
\[
\mathcal{W}_\kappa^2(\boldsymbol\rho_0,\boldsymbol\rho_1,\boldsymbol\gamma_0,\boldsymbol\gamma_1) = \sup_{\mathcal C}\mathcal J^\kappa(\boldsymbol\phi_t,\boldsymbol\psi_t),
\]
attainment of minimizers, and that \(\mathcal W_\kappa\) defines a distance on \(\mathcal P(\mathcal G)\). In the limit \(\kappa\to\infty\), if the vertex masses are unchanged, the model converges to an edges-only Wasserstein metric with Kirchhoff law; otherwise the cost diverges [2101.03415].

A plausible implication is that “potential network” is not a single model class but a recurrent structural device: scalar variables encode admissibility and optimality, while maps, plans, or fluxes are secondary objects induced from those scalars.

## 4. Neural dual-potential and assignment-based architectures

Neural OT methods frequently learn dual scalar functions rather than only transport maps. In the penalization framework for optimal transport, multi-marginal OT, martingale OT, and constrained OT, the primary trainable objects are dual functions in a superhedging representation,
\[
\phi(f):=\inf\Big\{\int h\,d\mu_0: h\ge f\ \text{for some }h\in\mathcal{H}\Big\}.
\]
For classical multi-marginal OT with
\[
\mathcal{Q}=\Pi(\mu_1,\dots,\mu_d),
\]
the dual class is
\[
\mathcal{H} = \Big\{ h(x_1,\dots,x_d)=h_1(x_1)+\cdots+h_d(x_d) \text{ for some }h_i\in C_b(\mathbb{R}) \Big\}.
\]
These \(h_i\) are represented by separate feedforward neural networks, and the hard constraint \(h\ge f\) is relaxed via a penalty
\[
\phi_{\theta,\gamma}(f) := \inf_{h\in\mathcal{H}} \Big\{ \int h\,d\mu_0+\int \beta_\gamma(f-h)\,d\theta \Big\},
\]
with neural approximation
\[
\phi^m_{\theta,\gamma}(f) = \inf_{h\in\mathcal{H}^m} \Big\{ \int h\,d\mu_0+\int \beta_\gamma(f-h)\,d\theta \Big\}.
\]
The paper proves value convergence and gives a primal recovery formula from the learned dual minimizer,
\[
\frac{d\hat\mu}{d\theta} = \beta_\gamma'(f-\hat h),
\]
but does not learn Monge maps directly [1802.08539].

A different neural potential construction appears in the algorithm for strong and weak OT based on a scalar network \(f_\omega\) and a stochastic transport map \(T_\theta\). The central maximin problem is
\[
\text{Cost}(\mathbb P,\mathbb Q)=\sup_f\inf_T \mathcal L(f,T),
\]
where \(f_\omega\) parameterizes the dual scalar function and \(T_\theta(x,z)\) parameterizes a deterministic or stochastic transport map. The paper explicitly calls \(f_\omega\) a potential network. In the strong-cost case, \(f\) is the Kantorovich dual variable via the standard \(c\)-transform,
\[
f^c(x)=\inf_{y\in\mathcal Y}\{c(x,y)-f(y)\},
\]
while in weak OT it is the corresponding weak-dual function. The method does not impose convexity, \(c\)-convexity, or 1-Lipschitzness on \(f_\omega\); instead, it jointly learns \(f_\omega\) and \(T_\theta\) through alternating optimization [2201.12220].

The paper’s main approximation theorem concerns stochastic transport maps:
\[
\|T_\theta-T\|_{L^2}^2\le \epsilon,\qquad
\mathbb W_2^2\big((T_\theta)_\#(\mathbb P\times\mathbb S),\mathbb Q\big)\le \epsilon.
\]
Thus the neural representation theorem is for map-based plan samplers, not for potentials themselves. In a special weak quadratic setting with \(\gamma=1\), however, the barycentric projection
\[
\overline T(x)=\int_{\mathcal Z}\widehat T(x,z)\,d\mathbb S(z)
\]
is compared with a convex-potential gradient \(\nabla\psi(x)\), giving an indirect connection to Brenier-type objects [2201.12220].

Assignment-based generative OT uses another dual-potential architecture. A scalar neural network \(\psi_w(y)\), defined on target samples, represents a Kantorovich target-side potential for a chosen transport cost \(c\). It induces an assignment
\[
y(x,w)=\arg\inf_{y \in \operatorname{supp}(\nu)} \{c(x,y)+\psi_w(y)\},
\]
and the paper shows that the gradient of the full dual objective can be replaced by the gradient of the assignment-count loss
\[
L_{\text{assign}}(w;w_0) = \sum_{j=1}^M \left( \frac{\#\{x_i:\, y(x_i,w_0)=y_j\}}{N} - \frac{1}{M} \right) \psi_w(y_j).
\]
If the resulting map \(\widetilde T(x)=\arg\min_y\{c(x,y)+\psi(y)\}\) is unique almost everywhere and satisfies \(\widetilde T_\sharp\mu=\nu\), then \((I\times \widetilde T)_\sharp\mu\) is optimal [1910.00535].

These three neural families illustrate a central distinction. Some potential networks approximate generalized dual test functions \(h_i\), some learn an unconstrained scalar dual function \(f_\omega\) jointly with a map network, and some use a target-side dual potential \(\psi_w\) to induce discrete transport assignments. None of these is, by construction, a Brenier-potential network [1802.08539].

## 5. Brenier-potential networks and PDE-constrained learning

A more restrictive meaning of optimal transport potential network emerges when the neural architecture is designed to represent the Brenier potential itself. In the Monge–Ampère PINN method, the source and target measures are
\[
\mu(dx)=f(x)\,dx,\qquad \nu(dy)=g(y)\,dy,
\]
and under the quadratic cost
\[
c(x,y)=\frac12|x-y|^2
\]
the optimal map is represented as
\[
\mathbf m(x)=\nabla u(x),
\]
with \(u\) convex. The pushforward identity yields the generalized Monge–Ampère equation
\[
\det(D^2 u(x))=\frac{f(x)}{g(\nabla u(x))},
\]
together with the transport boundary condition
\[
\nabla u(\mathcal X)=\mathcal Y,\qquad \nabla u(\partial\mathcal X)=\partial\mathcal Y.
\]
The network \(u_\theta\) is therefore a direct neural parameterization of the OT potential, and the transport map is recovered by differentiation:
\[
T_\theta(x)=\nabla u_\theta(x).
\]
Convexity is enforced architecturally with an input-convex neural network and softplus activation [2501.10162].

The ICNN layers are
\[
x^0=x,
\]
\[
x^1=\sigma\!\left(L^{(0)}x^0+b^{(0)}\right),
\]
\[
x^l=\sigma\!\left(W^{(l-1)}x^{l-1}+L^{(l-1)}x^0+b^{(l-1)}\right), \qquad 2\le l\le L,
\]
\[
u_\theta(x)=W^{(L)}x^L+L^{(L)}x^0+b^{(L)},
\]
with \(\sigma\) convex and nondecreasing and \(W^{(l)}\ge 0\) elementwise. The loss combines an interior Monge–Ampère residual,
\[
E_{PDE} = \frac1{N_c}\sum_{i=1}^{N_c} \left| \det(D^2u_\theta(p_i)) -\frac{f(p_i)}{g(\nabla u_\theta(p_i))} \right|^2,
\]
with a Hausdorff-inspired transport boundary loss \(E_{OT}\), and the total objective
\[
u_\theta = \arg\min_\theta \left\{E_{PDE}+C\,E_{OT}\right\}.
\]
The paper reports accurate low-dimensional maps and gives concrete implementation details: PyTorch, 4 hidden layers, 10 neurons per layer, softplus activation, Adam warm-start followed by L-BFGS, default \(N_c=N_b=800\), and default \(C=1\) [2501.10162].

A broader PDE-based potential network is developed through the Hamilton–Jacobi equation. Here the learned object is a scalar spatio-temporal potential
\[
u:\Omega\times[0,t_f]\to\mathbb{R},
\]
which solves
\[
\frac{\partial u}{\partial t}+h(\nabla u)=0
\]
for costs \(c(x,y)=\ell(x-y)\) with Legendre transform
\[
h(z)=\sup_{y\in\mathbb{R}^d}\{z^\top y-\ell(y)\}.
\]
The optimal velocity is
\[
v^*(x,t)=\nabla h(\nabla u(x,t)),
\]
and the forward and backward OT maps are
\[
T_{\mu}^{\nu *}(x)=x+t_f\nabla h\bigl(\nabla u^*(x,0)\bigr),
\]
\[
T_{\nu}^{\mu *}(y)=y-t_f\nabla h\bigl(\nabla u^*(y,t_f)\bigr).
\]
The characteristic equations
\[
\dot x=\nabla h(p),\qquad \dot u=-h(p)+p^\top \nabla h(p),\qquad \dot p=0
\]
imply straight-line characteristics and eliminate numerical integration at inference [2510.01153].

Training does not use a standard PINN residual. Instead it penalizes the characteristic-consistent implicit formula through
\[
L_{\mathrm{HJ}}(u_\theta)
\]
and combines it with an MMD term,
\[
L_{\mathrm{MMD}}(u_\theta)=\mathrm{MMD}\bigl((T_\mu^\nu[u_\theta])_\sharp\mu,\nu\bigr)^2,
\]
giving a single-network pure minimization framework. For quadratic cost,
\[
\ell(z)=\frac12\|z\|^2,\qquad h(z)=\frac12\|z\|^2,
\]
the maps reduce to
\[
T_\mu^\nu(x)=x+t_f\nabla u(x,0),\qquad T_\nu^\mu(y)=y-t_f\nabla u(y,t_f).
\]
The paper states that the viscosity solution uniquely characterizes the OT map and presents consistency and Gaussian stability results [2510.01153].

These two lines of work clarify a second misconception: a neural OT potential network need not be only a dual critic. In some formulations the network is the Brenier potential \(u\) itself, and in others it is a time-dependent Hamilton–Jacobi potential whose gradients at endpoint times induce both transport directions [2501.10162].

## 6. Applications, topological consequences, and open limitations

The application domains represented in the cited work are heterogeneous but structurally related. Leaf venation motivates loop-rich transport networks that remain functional after edge cuts and under spatially patchy demand [0906.0006]. Mesoscopic conductivity models are aimed at biological transportation systems such as leaf venation, blood vessel networks, and neuronal transport-like structures, with network adaptation written as Wasserstein or Fisher–Rao gradient flows in the phase space of position and conductivity tensor [2401.07922]. Dynamic OT on networks models mass stored both on edges and at vertices, with the parameter \(\kappa\) controlling the cost of edge–vertex interchange [2101.03415].

In statistical and data-driven settings, the same potential viewpoint supports different tasks. The maximum-entropy transport ensemble proposes applications in link prediction, network reconstruction, and statistical validation, especially when transport is only partially optimized and affected by noise-like factors [2504.10444]. The assignment-based generative framework allows any transport cost compatible with the \(c\)-transform, including squared Euclidean cost and image-centered costs such as SSIM-based dissimilarities, and reports quantitative results on MNIST and Fashion-MNIST [1910.00535]. The stochastic-map method addresses both strong and weak OT and is evaluated on toy examples and unpaired image-to-image translation [2201.12220]. The Brenier-PINN and Hamilton–Jacobi methods address low- to moderate-dimensional continuous OT, including Gaussian-to-uniform transport, Gaussian-to-Gaussian transport, 3D cube transport, color transfer, class-conditional transport, and latent-space image transport [2501.10162].

The following table organizes the major potential types that recur across the literature.

| Setting | Potential variable | Induced transport object |
|---|---|---|
| Conductance network | \(V_k\) | Edge currents \(I_{kj}=C_{kj}(V_k-V_j)\) |
| Dynamic OT on graphs | \(\phi_t^j,\psi_t^i\) | Edge fluxes \(F_t^j\), vertex exchange \(f_t^i\) |
| Mesoscopic network PDE | Pressure \(p\) | Flux \(J=-(P[\mu]+rI)\nabla p\) |
| Statistical bipartite OT | \(t_i,\theta_\alpha\) | Expected weights \(\langle w_{i\alpha}\rangle\) |
| Neural dual OT | \(h_i,f_\omega,\psi_w\) | Dual values, assignments, regularized couplings |
| Brenier/HJ neural OT | \(u_\theta\) | Maps from \(\nabla u_\theta\) or \(\nabla h(\nabla u_\theta)\) |

Several limitations recur. The leaf-venation conductance model assumes linear transport laws, static conductances, a fixed underlying graph, single-edge deletion for damage, one active sink at a time for fluctuations, and numerical annealing that may find local minima rather than global ones [0906.0006]. The penalized dual-neural method proves value convergence but does not focus on optimizer convergence or Monge-map recovery [1802.08539]. The strong/weak neural OT method uses an unconstrained scalar dual network rather than a convex or \(c\)-convex architecture [2201.12220]. The Monge–Ampère PINN requires Hessians and determinants and does not provide a convergence theorem for the neural approximation [2501.10162]. The Hamilton–Jacobi method relies on convex displacement costs \(c(x,y)=\ell(x-y)\) and on characteristic structure specific to that setting [2510.01153].

Taken together, these works show that optimal transport potential networks are best understood not as one algorithmic family but as a common structural principle. The potential may be a pressure, a dual scalar function, a convex Brenier potential, a Hamilton–Jacobi viscosity solution, or a set of node-specific balancing multipliers. What remains invariant is the organizing role of scalar potential variables in defining admissible transport and in selecting optimal or near-optimal maps, plans, flows, or network architectures [2101.03415].

Source: https://www.emergentmind.com/topics/optimal-transport-potential-networks