Papers
Topics
Authors
Recent
Search
2000 character limit reached

Gromov-Monge Flow Matching (GMFM): Theory and Applications

Updated 1 September 2026
  • Gromov-Monge Flow Matching (GMFM) is a technique for determining source–target correspondences using Gromov-Wasserstein (GW) geometry principles and structured maps, particularly valuable for domains with different metrics or representations such as graph data
  • GMFM can be used for deterministic maps, stochastic conditional generators, or flows on quotient spaces, with applications in understanding and generating complex data structures
  • A key application of GMFM includes determining intrinsic dimension reduction during data matching, aligning graph structures via permutation symmetry, or studying dynamic optimality in data flow matching

Gromov-Monge Flow Matching (GMFM) denotes a class of flow-matching constructions in which source–target correspondences are determined by Gromov-type structural objectives and are then represented by deterministic maps, stochastic conditional generators, or flows on quotient spaces. The term combines Gromov–Wasserstein (GW) geometry, which compares intra-domain relations rather than pointwise coordinates, with Monge transport, which restricts couplings to maps, and flow matching (FM), which learns time-dependent vector fields from interpolated endpoint pairs. No single formulation encompasses all uses of the term. Existing work ranges from closed-form Gaussian Gromov–Monge maps, through reversible and entropic stochastic transport, to quotient-aware graph flow matching. A central distinction is whether the learned correspondence is deterministic, a general transport plan, or a conditional probability law.

1. Geometric and variational foundations

Let μ\mu and ν\nu be probability measures on spaces X\mathcal X and Y\mathcal Y, equipped with intra-space costs cXc_X and cYc_Y. Classical optimal transport uses a cross-space cost c(x,y)c(x,y) and minimizes a linear functional over couplings π∈Π(μ,ν)\pi\in\Pi(\mu,\nu). By contrast, GW transport compares pairwise relations:

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').

For squared Euclidean costs and p=2p=2,

ν\nu0

The associated Gromov–Monge problem restricts the coupling to a deterministic map ν\nu1 satisfying ν\nu2:

ν\nu3

The induced coupling is ν\nu4, whose conditional distribution is a Dirac mass, ν\nu5. GMFM uses such a correspondence as endpoint supervision for a dynamic interpolation.

The Gromov formulation is particularly relevant when the domains have different dimensions, coordinate systems, modalities, or representations. It does not require a meaningful pointwise comparison between ν\nu6 and ν\nu7; instead, it seeks a coupling under which relational structure is preserved. For graph data, the relational structure is induced by edges and node-pair features. For heterogeneous data, it can be represented by kernels, affinities, inner products, or geodesic costs.

The relationship between GW, Gromov–Monge, and reversible Gromov–Monge (RGM) transport is governed by nested admissible classes:

ν\nu8

which yields

ν\nu9

RGM requires forward and reverse maps X\mathcal X0 and X\mathcal X1 satisfying equality of their induced couplings,

X\mathcal X2

Under bounded costs and non-atomic measures, RGM, GM, and GW coincide at the population infimum, although this does not imply that finite neural optimization recovers an exact GW minimizer (Hur et al., 2021).

2. Gaussian Gromov–Monge maps

For Gaussian measures

X\mathcal X3

with squared-Euclidean intra-space costs, the Gaussian-restricted GW problem reduces to an optimization over the cross-covariance X\mathcal X4. After centering and diagonalizing the covariances,

X\mathcal X5

the relevant Gaussian optimization is

X\mathcal X6

Assume X\mathcal X7 and write X\mathcal X8 for the leading X\mathcal X9 block of Y\mathcal Y0. The optimal cross-covariance matches the largest covariance eigenvalues:

Y\mathcal Y1

where Y\mathcal Y2. Its objective value is

Y\mathcal Y3

The Schur complement is saturated:

Y\mathcal Y4

Consequently, Y\mathcal Y5, so the optimal Gaussian coupling is deterministic and is induced by an affine map,

Y\mathcal Y6

with

Y\mathcal Y7

In PCA coordinates, the first Y\mathcal Y8 source principal coordinates satisfy

Y\mathcal Y9

and the remaining cXc_X0 source coordinates are discarded. The map is deterministic, pushes cXc_X1 to cXc_X2, minimizes GW among Gaussian couplings, and solves the linear Gromov–Monge problem. It does not generally solve the unrestricted GW or unrestricted Gromov–Monge problem (Salmona et al., 2021).

The Gaussian-restricted distance decomposes into total-variance mismatch, matched-principal-variance mismatch, and discarded-source-dimensional energy:

cXc_X3

Thus, dimension reduction is intrinsic to the construction when cXc_X4. The source variance outside its leading cXc_X5-dimensional PCA subspace contributes necessarily to the cost.

The construction differs from the ordinary Gaussian cXc_X6 map. GW matches sorted principal variances and is invariant to independent orthogonal transformations of the source and target spaces. Ordinary cXc_X7 transport instead depends on the full matrix interaction

cXc_X8

The two maps agree under particular covariance alignments, such as shared eigenvectors with eigenvalues paired in the same order.

For unrestricted Gaussian GW, explicit lower and upper bounds satisfy

cXc_X9

The bounds coincide when cYc_Y0 and cYc_Y1, and in one dimension:

cYc_Y2

with optimizer

cYc_Y3

Singular covariance matrices are handled by removing zero-variance coordinates, solving on the effective support, and lifting the result back to the ambient space.

3. Deterministic existence and nondeterminism

The existence of a deterministic Gromov–Monge map depends strongly on the structural cost. For compactly supported Euclidean measures with an absolutely continuous source, optimal maps exist for inner-product GW costs. The construction uses a singular-value decomposition of a cross-moment matrix and decomposes the problem into a lower-dimensional base transport and fiberwise transports. Brenier-type maps solve the base problem, while measurable selection glues the fiberwise maps into a global map.

For squared-distance GW costs, the general result is weaker. Under cYc_Y4, compact supports, and cYc_Y5, an optimal 2-map exists. A 2-map has the form

cYc_Y6

where cYc_Y7 almost everywhere. It may assign two target points to the same source point with different conditional weights.

When the relevant cross-moment matrix has full rank, subtwist and 2-twist properties imply 2-map structure and, in this case, a map/anti-map decomposition. When the rank is cYc_Y8, the admissible target set is the intersection of a sphere and an affine line and contains at most two points. When the rank is at most cYc_Y9, fiber-invariant structure permits a genuine map through disintegration and spherical fiber transport.

In one dimension, numerical constructions indicate that an optimal squared-distance GW plan can be a 2-map but not a map. This provides evidence for the conjecture that there exist absolutely continuous measures for which no optimal squared-distance GW plan is deterministic. Absolute continuity of the source, which guarantees a Monge map for classical quadratic OT, does not provide the same guarantee for squared-distance GW (Dumont et al., 2022).

This distinction is fundamental for GMFM. If a Gromov optimizer is a map, a deterministic endpoint displacement c(x,y)c(x,y)0 can supervise a conventional vector field. If only a 2-map exists, the natural object is a velocity kernel,

c(x,y)c(x,y)1

A barycentric velocity,

c(x,y)c(x,y)2

generally loses the original coupling structure. Such cases motivate stochastic, branching, multivalued, or path-space formulations rather than a single deterministic ODE.

4. Static stochastic transport and conditional flow matching

Entropic GW and fused GW provide a different route to flow matching. Instead of restricting the coupling to a graph, they compute a diffuse entropic transport plan and learn its conditional rows. GENOT—Generative Entropic (Gromov) Wasserstein Flow Matching—solves a sample-based entropic OT, GW, or fused GW problem, then applies conditional flow matching to learn a neural sampler for each conditional distribution (Klein et al., 2023).

For an entropic GW objective,

c(x,y)c(x,y)3

the conditional disintegration is

c(x,y)c(x,y)4

GENOT samples target-space noise c(x,y)c(x,y)5, usually c(x,y)c(x,y)6, and learns a conditional flow

c(x,y)c(x,y)7

The terminal generator c(x,y)c(x,y)8 is trained to reproduce c(x,y)c(x,y)9. With linear interpolation

π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)0

the conditional flow-matching objective is

π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)1

At zero population loss, the generated pair has the target entropic coupling:

π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)2

The learned object is therefore generally a stochastic conditional map,

π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)3

not a deterministic Monge map. A deterministic surrogate may be defined by the conditional mean,

π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)4

but it is generally not an optimal Monge map. A Monge map is recovered only when the conditional degenerates to π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)5. The limit of vanishing entropy or noise is not shown to converge to a classical Monge map.

GENOT supports different ambient dimensions, arbitrary sample-evaluable costs, non-Euclidean or graph-geodesic costs, cross-modality translation, uncertainty quantification, and out-of-sample conditional generation. Its practical procedure repeatedly samples source and target minibatches, computes a discrete entropic OT/GW/FGW coupling, samples index pairs from that coupling, and trains the conditional velocity field on the resulting endpoint pairs.

The discrete structural solver remains the computational bottleneck. Entropic GW may require π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)6 or π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)7 time depending on the cost structure and π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)8 memory. Minibatching enables larger datasets but introduces statistical bias that is not fully characterized. GENOT therefore constitutes a stochastic, generative analogue of Gromov–Monge transport rather than a deterministic Gromov–Monge method.

5. Flow construction, vector fields, and dynamic optimality

For a coupling π∈Π(μ,ν)\pi\in\Pi(\mu,\nu)9 of endpoint variables GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').0, standard FM uses

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').1

and conditional velocity

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').2

The population vector field is

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').3

A Gromov–Monge interpolation can use a deterministic map GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').4:

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').5

If GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').6, then

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').7

and, whenever GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').8 is invertible on the relevant support, the induced affine velocity is

GWpp(cX,cY,μ,ν)=inf⁡π∈Π(μ,ν)∬∣cX(x,x′)−cY(y,y′)∣p dπ(x,y)dπ(x′,y′).GW_p^p(c_X,c_Y,\mu,\nu) = \inf_{\pi\in\Pi(\mu,\nu)} \iint \left|c_X(x,x')-c_Y(y,y')\right|^p \,d\pi(x,y)d\pi(x',y').9

For the Gaussian map, this yields an explicit affine flow, although invertibility of intermediate maps is an additional regularity condition.

A coupling-based construction does not require determinism. If p=2p=20 is jointly Gaussian, then the flow-matching field for the linear interpolation is affine:

p=2p=21

The pseudoinverse is used when the interpolation covariance is singular.

A third option interpolates Gaussian marginals directly,

p=2p=22

with either linear covariance interpolation, map-induced covariance

p=2p=23

or a Wasserstein/Bures covariance geodesic after PCA alignment. An affine field

p=2p=24

satisfies the continuity equation when

p=2p=25

These constructions provide dynamic extensions but are not selected canonically by the Gaussian GW theory.

The relation between FM, action matching, and quadratic OT provides a general design principle. Restricting vector fields to those generated by convex potentials can make FM or action matching equivalent to quadratic OT. For a convex potential p=2p=26, the endpoint is p=2p=27 and the trajectory is

p=2p=28

The associated field is generated by a family of potentials satisfying a Hamilton–Jacobi cancellation. This makes the matching loss equal to the quadratic OT dual objective up to constants, independently of the chosen coupling or intermediate marginal path (Kornilov et al., 31 Oct 2025).

That result does not establish an analogous equivalence for GW or Gromov–Monge transport. Its proof relies on Euclidean convex duality, gradients, straight-line interpolation, and the quadratic pointwise cost. A GMFM theory would require a replacement based on intrinsic metric structure, pairwise distortion, quotient geometry, or geodesic interpolation.

6. Quotient-space flow matching for graphs

For graphs, the Gromov–Monge problem arises naturally from permutation symmetry. A graph with p=2p=29 nodes and ν\nu00 feature channels can be represented as

ν\nu01

with edge features off the diagonal and node features on the diagonal. The permutation group acts simultaneously on both node indices:

ν\nu02

The unordered graph is the orbit

ν\nu03

and the graph representation space is the quotient by the permutation group.

For a compact group ν\nu04, the quotient metric is

ν\nu05

For graph tensors, this becomes

ν\nu06

The optimal permutation is a deterministic node correspondence, hence a Gromov–Monge alignment. Exact computation is a quadratic assignment problem and is NP-hard in general (Piening et al., 27 Aug 2026).

The quotient-space theory establishes that a quotient coupling can be lifted to aligned Euclidean representatives without increasing cost. If ν\nu07 couples graph orbits, there exist a representative law ν\nu08 and Euclidean coupling ν\nu09 satisfying

ν\nu10

and

ν\nu11

An optimal quotient coupling therefore induces an ordinary Euclidean coupling between optimally aligned representatives. Its linear interpolation

ν\nu12

projects to a constant-speed Wasserstein geodesic in the quotient space, with kinetic energy equal to the squared quotient Wasserstein distance.

The practical FM modification changes only the training coupling. Given source graph ν\nu13 and target graph ν\nu14, an approximate alignment ν\nu15 produces

ν\nu16

The training pair is ν\nu17, and the interpolant is

ν\nu18

The minibatch velocity loss is

ν\nu19

Permutation-equivariant architectures can be used without explicitly averaging over all permutations. Haar symmetrization of the coupling preserves the quotient coupling, and equivariant minimizers have the same flow-matching objective on the original and symmetrized couplings. If the equivariant ODE has a unique flow ν\nu20, then

ν\nu21

so the quotient flow

ν\nu22

is well-defined and independent of arbitrary node ordering.

Exact graph alignment is approximated in practice by a GW relaxation or a first lower bound (FLB). GW uses a soft node coupling and ten Frank–Wolfe iterations, followed by Hungarian rounding to a hard permutation. FLB aligns nodes by sorted eccentricities,

ν\nu23

and has cost ν\nu24 per graph pair. It is substantially cheaper than GW but weakly correlated with the exact objective. An optional outer assignment pairs source and target graphs within a minibatch; inner alignment makes intermediate structures meaningful, while outer assignment makes trajectories shorter and straighter.

Experiments on cycle graphs, continuous stochastic block model graphs, QM9, and ZINC250k show that structure-aware couplings are especially beneficial at small integration budgets. In the continuous SBM experiment, five-step GW alignment with outer assignment reduced FGW nearest-neighbor accuracy from ν\nu25 to ν\nu26, degree MMD from ν\nu27 to ν\nu28, and clustering-coefficient MMD from ν\nu29 to ν\nu30. On QM9 at five Euler steps, GW with outer assignment achieved validity ν\nu31 and FCD ν\nu32; on ZINC250k it achieved validity ν\nu33, while inner-only GW achieved FCD ν\nu34.

The graph construction demonstrates a direct GMFM principle: permutation equivariance of the architecture is insufficient if the training coupling ignores quotient geometry. Aligning representatives before interpolation prevents arbitrary node-order mismatches from producing meaningless intermediate graphs. The exact theory applies to quotient-optimal couplings, whereas practical GW and FLB alignments are approximations without universal optimality guarantees.

7. Scope, limitations, and research directions

GMFM currently denotes several related but non-equivalent paradigms.

A deterministic Gaussian construction supplies an affine endpoint map that matches PCA directions, scales by ν\nu35, and can be interpolated by an explicit affine flow. Its rigorous optimality is restricted to Gaussian couplings and the linear Gromov–Monge problem. The unrestricted Gaussian optimizer is conjectured to be Gaussian more generally, but this is not established.

Existence theory shows that inner-product GW admits an optimal map under compactness and source absolute continuity, whereas squared-distance GW may admit only a 2-map. This creates a structural obstruction to deterministic flow matching: a single vector field may be unable to represent all optimal branches. Stochastic or branching flows, multivalued networks, barycentric projections, cost modification, or restricted data classes provide possible responses.

Entropic methods such as GENOT preserve the full conditional transport law and support uncertainty, different dimensions, arbitrary sample-evaluable costs, and out-of-sample generation. They are not Monge methods, because entropic regularization generally produces diffuse conditional distributions. Their computational complexity remains tied to discrete GW or FGW solvers, and convergence to an unregularized Monge map is not established.

Quotient-aware graph GMFM provides an exact geometric interpretation for permutation symmetries. It lifts quotient-optimal couplings to aligned Euclidean representatives and leaves the FM architecture and sampler unchanged. In practice, however, exact node alignment is intractable, and approximate GW, lower-bound heuristics, rounding, minibatching, and finite-step integration introduce errors.

Several issues remain open:

  • a dynamic Benamou–Brenier-type theory for GW or Gromov–Monge transport;
  • path-independent variational principles for metric-structure-preserving flows;
  • deterministic-map existence and uniqueness under general squared-distance GW costs;
  • principled treatment of 2-maps and branching correspondences;
  • geodesic and vector-field formulations on general metric spaces and manifolds;
  • computationally efficient exact or certified approximate quotient alignment;
  • convergence of entropic conditional flows toward Monge maps;
  • dynamic reversibility constraints for forward and reverse flows;
  • stability under sampling, minibatching, discretization, and approximate alignments;
  • coupling objectives that combine metric interpretation with deterministic existence.

The central conceptual distinction is between endpoint geometry and dynamic realization. GW or Gromov–Monge optimization determines a structurally meaningful correspondence; flow matching turns selected endpoint correspondences or conditional laws into time-dependent vector fields. A static Gromov–Monge map does not automatically minimize a dynamic action, and a GW-optimal coupling does not automatically define a globally well-posed deterministic flow. Consequently, GMFM is best understood as a family of geometry-aware flow-matching constructions whose precise mathematical character depends on whether the underlying correspondence is a map, a finite-valued plan, a stochastic conditional law, or a quotient-space alignment.

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 Gromov-Monge Flow Matching.