Papers
Topics
Authors
Recent
Search
2000 character limit reached

Optimal Transport Potential Networks

Updated 15 July 2026
  • Optimal Transport Potential Networks are formulations where scalar potentials (e.g., nodal pressures and dual variables) define admissible flows in graphs.
  • They unify diverse methods—from conductance optimization and dynamic network models to PDE-constrained and neural approaches—ensuring resilient, loop-rich transport structures.
  • Neural implementations approximate dual functions or Brenier potentials, bridging statistical transport ensembles with practical, parameterized network solutions.

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 (Katifori et al., 2009). The unifying structure is that transport is not represented only by a coupling or by a map TT, 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 kk carries a scalar potential VkV_k, and the edge current is

Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),

with Kirchhoff conservation

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.

The dissipative objective is the total power loss

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,

under the material constraint

12∑k∑⟨j,k⟩Ckjγ=1.\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 (Katifori et al., 2009).

In dynamic optimal transport on a network, the potential variables are dual edge functions ϕtj\phi_t^j and vertex functions ψti\psi_t^i. They satisfy Hamilton–Jacobi-type inequalities,

∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,

kk0

and induce optimal fluxes through

kk1

Here the network topology enters through incidence relations and potential traces at the vertices (Burger et al., 2021).

In the mesoscopic measure-based model of transportation networks, the potential is the pressure kk2 solving

kk3

with

kk4

The network state is a probability measure kk5, but the actual flux is still potential-driven,

kk6

The associated energy

kk7

combines pumping and metabolic terms under local mass conservation (Haskovec et al., 2024).

A common misconception is that OT potentials must always mean Brenier potentials. The cited literature shows a broader usage: potentials can be physical pressures kk8 or kk9, graph-dual Hamilton–Jacobi variables VkV_k0, generalized source–sink multipliers VkV_k1, or neural scalar functions VkV_k2 (Buffa et al., 14 Apr 2025).

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 VkV_k3. For the standard fixed-load problem reviewed in the leaf-venation study, the optimum is a tree for VkV_k4 and a spatially distributed sheet for VkV_k5, with a transition at VkV_k6. The paper modifies the objective to average either over single-edge failures or over sparse fluctuating loads, while keeping the same conductance-budget constraint (Katifori et al., 2009).

The damage model removes a single edge VkV_k7 by

VkV_k8

recomputes the potentials VkV_k9, and sums the dissipations: Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),0 If deleting an edge disconnects the graph, Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),1 diverges and Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),2. 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 Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),3 requires transport redundancy, not only topological redundancy (Katifori et al., 2009).

The fluctuating-load model activates one sink at a time,

Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),4

and minimizes the averaged dissipation

Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),5

This again favors loops. The paper attributes the failure of the standard tree-optimality arguments to the nonlinearity of the conductance budget

Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),6

which makes optimization over averaged load states inequivalent to optimization for the average load (Katifori et al., 2009).

The main computational findings are topological as well as energetic. In the damage model, the Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),7 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 Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),8 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 Ikj=Ckj(Vk−Vj),I_{kj}=C_{kj}(V_k-V_j),9, while in the sink-fluctuation case such features disappear already for ∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.0 (Katifori et al., 2009).

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 (Katifori et al., 2009).

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

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.1

enforced in expectation. The free-energy-like variational problem maximizes

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.2

with transport energy

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.3

Introducing multipliers ∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.4 and ∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.5 yields a factorized ensemble in which edge weights are independent exponential variables with rates

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.6

and expected transport intensity

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.7

The self-consistency conditions are

∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.8

The paper interprets ∑⟨j,k⟩Ckj(Vk−Vj)=Ik.\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)=\mathcal{I}_k.9 as generalized transport potentials or generalized network chemical potentials (Buffa et al., 14 Apr 2025).

This formulation is explicitly contrasted with standard entropically regularized OT. Sinkhorn-type couplings have multiplicative Gibbs form,

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,0

whereas the present model has a rational additive form,

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,1

The paper therefore presents a potential network not through P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,2-transforms or convex gradients, but through additive effective energies controlled jointly by node potentials and edge costs (Buffa et al., 14 Apr 2025).

The same paper quantifies a dense-to-sparse transition using the mass share of the maximum spanning tree of the average network,

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,3

As P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,4 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

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,5

for the main case studied, while the transition is described as apparently non-critical (Buffa et al., 14 Apr 2025).

Dynamic network OT provides yet another potential structure. On a finite connected planar network P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,6, mass lives both on edges and at vertices, and the action is

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,7

The paper proves a duality formula

P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,8

attainment of minimizers, and that P=12∑k∑⟨j,k⟩Ckj(Vk−Vj)2,P=\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}(V_k-V_j)^2,9 defines a distance on 12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.0. In the limit 12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.1, if the vertex masses are unchanged, the model converges to an edges-only Wasserstein metric with Kirchhoff law; otherwise the cost diverges (Burger et al., 2021).

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,

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.2

For classical multi-marginal OT with

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.3

the dual class is

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.4

These 12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.5 are represented by separate feedforward neural networks, and the hard constraint 12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.6 is relaxed via a penalty

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.7

with neural approximation

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.8

The paper proves value convergence and gives a primal recovery formula from the learned dual minimizer,

12∑k∑⟨j,k⟩Ckjγ=1.\frac12\sum_k\sum_{\langle j,k\rangle} C_{kj}^{\gamma}=1.9

but does not learn Monge maps directly (Eckstein et al., 2018).

A different neural potential construction appears in the algorithm for strong and weak OT based on a scalar network ϕtj\phi_t^j0 and a stochastic transport map ϕtj\phi_t^j1. The central maximin problem is

ϕtj\phi_t^j2

where ϕtj\phi_t^j3 parameterizes the dual scalar function and ϕtj\phi_t^j4 parameterizes a deterministic or stochastic transport map. The paper explicitly calls ϕtj\phi_t^j5 a potential network. In the strong-cost case, ϕtj\phi_t^j6 is the Kantorovich dual variable via the standard ϕtj\phi_t^j7-transform,

ϕtj\phi_t^j8

while in weak OT it is the corresponding weak-dual function. The method does not impose convexity, ϕtj\phi_t^j9-convexity, or 1-Lipschitzness on ψti\psi_t^i0; instead, it jointly learns ψti\psi_t^i1 and ψti\psi_t^i2 through alternating optimization (Korotin et al., 2022).

The paper’s main approximation theorem concerns stochastic transport maps: ψti\psi_t^i3 Thus the neural representation theorem is for map-based plan samplers, not for potentials themselves. In a special weak quadratic setting with ψti\psi_t^i4, however, the barycentric projection

ψti\psi_t^i5

is compared with a convex-potential gradient ψti\psi_t^i6, giving an indirect connection to Brenier-type objects (Korotin et al., 2022).

Assignment-based generative OT uses another dual-potential architecture. A scalar neural network ψti\psi_t^i7, defined on target samples, represents a Kantorovich target-side potential for a chosen transport cost ψti\psi_t^i8. It induces an assignment

ψti\psi_t^i9

and the paper shows that the gradient of the full dual objective can be replaced by the gradient of the assignment-count loss

∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,0

If the resulting map ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,1 is unique almost everywhere and satisfies ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,2, then ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,3 is optimal (Laschos et al., 2019).

These three neural families illustrate a central distinction. Some potential networks approximate generalized dual test functions ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,4, some learn an unconstrained scalar dual function ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,5 jointly with a map network, and some use a target-side dual potential ∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,6 to induce discrete transport assignments. None of these is, by construction, a Brenier-potential network (Eckstein et al., 2018).

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

∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,7

and under the quadratic cost

∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,8

the optimal map is represented as

∂tϕtj+12∣∂xϕtj∣2≤0,\partial_t \phi_t^j+\frac12 |\partial_x\phi_t^j|^2\le 0,9

with kk00 convex. The pushforward identity yields the generalized Monge–Ampère equation

kk01

together with the transport boundary condition

kk02

The network kk03 is therefore a direct neural parameterization of the OT potential, and the transport map is recovered by differentiation: kk04 Convexity is enforced architecturally with an input-convex neural network and softplus activation (Caboussat et al., 17 Jan 2025).

The ICNN layers are

kk05

kk06

kk07

kk08

with kk09 convex and nondecreasing and kk10 elementwise. The loss combines an interior Monge–Ampère residual,

kk11

with a Hausdorff-inspired transport boundary loss kk12, and the total objective

kk13

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 kk14, and default kk15 (Caboussat et al., 17 Jan 2025).

A broader PDE-based potential network is developed through the Hamilton–Jacobi equation. Here the learned object is a scalar spatio-temporal potential

kk16

which solves

kk17

for costs kk18 with Legendre transform

kk19

The optimal velocity is

kk20

and the forward and backward OT maps are

kk21

kk22

The characteristic equations

kk23

imply straight-line characteristics and eliminate numerical integration at inference (Park et al., 30 Sep 2025).

Training does not use a standard PINN residual. Instead it penalizes the characteristic-consistent implicit formula through

kk24

and combines it with an MMD term,

kk25

giving a single-network pure minimization framework. For quadratic cost,

kk26

the maps reduce to

kk27

The paper states that the viscosity solution uniquely characterizes the OT map and presents consistency and Gaussian stability results (Park et al., 30 Sep 2025).

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 kk28 itself, and in others it is a time-dependent Hamilton–Jacobi potential whose gradients at endpoint times induce both transport directions (Caboussat et al., 17 Jan 2025).

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 (Katifori et al., 2009). 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 (Haskovec et al., 2024). Dynamic OT on networks models mass stored both on edges and at vertices, with the parameter kk29 controlling the cost of edge–vertex interchange (Burger et al., 2021).

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 (Buffa et al., 14 Apr 2025). The assignment-based generative framework allows any transport cost compatible with the kk30-transform, including squared Euclidean cost and image-centered costs such as SSIM-based dissimilarities, and reports quantitative results on MNIST and Fashion-MNIST (Laschos et al., 2019). The stochastic-map method addresses both strong and weak OT and is evaluated on toy examples and unpaired image-to-image translation (Korotin et al., 2022). 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 (Caboussat et al., 17 Jan 2025).

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

Setting Potential variable Induced transport object
Conductance network kk31 Edge currents kk32
Dynamic OT on graphs kk33 Edge fluxes kk34, vertex exchange kk35
Mesoscopic network PDE Pressure kk36 Flux kk37
Statistical bipartite OT kk38 Expected weights kk39
Neural dual OT kk40 Dual values, assignments, regularized couplings
Brenier/HJ neural OT kk41 Maps from kk42 or kk43

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 (Katifori et al., 2009). The penalized dual-neural method proves value convergence but does not focus on optimizer convergence or Monge-map recovery (Eckstein et al., 2018). The strong/weak neural OT method uses an unconstrained scalar dual network rather than a convex or kk44-convex architecture (Korotin et al., 2022). The Monge–Ampère PINN requires Hessians and determinants and does not provide a convergence theorem for the neural approximation (Caboussat et al., 17 Jan 2025). The Hamilton–Jacobi method relies on convex displacement costs kk45 and on characteristic structure specific to that setting (Park et al., 30 Sep 2025).

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 (Burger et al., 2021).

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Optimal Transport Potential Networks.