---
title: 'Gromov-Monge Flow Matching (GMFM): Theory and Applications'
url: https://www.emergentmind.com/topics/gromov-monge-flow-matching
type: topic
---

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

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

$$
GW_p^p(c_X,c_Y,\mu,\nu)
=
\inf_{\pi\in\Pi(\mu,\nu)}
\iint
\left|c_X(x,x')-c_Y(y,y')\right|^p
\,d\pi(x,y)d\pi(x',y').
$$

For squared Euclidean costs and $p=2$,

$$
GW_2^2(\mu,\nu)
=
\inf_{\pi\in\Pi(\mu,\nu)}
\mathbb E_\pi
\left[
\left(
\|X-X'\|^2-\|Y-Y'\|^2
\right)^2
\right].
$$

The associated Gromov–Monge problem restricts the coupling to a deterministic map $T:\mathcal X\to\mathcal Y$ satisfying $T_\#\mu=\nu$:

$$
\inf_{T:T_\#\mu=\nu}
\iint
\left(
c_X(x,x')-c_Y(T(x),T(x'))
\right)^2
\,d\mu(x)d\mu(x').
$$

The induced coupling is $\pi_T=(\operatorname{id},T)_\#\mu$, whose conditional distribution is a Dirac mass, $\pi_T(\cdot\mid x)=\delta_{T(x)}$. 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 $x$ and $y$; 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:

$$
\Pi'\subseteq\Pi_{\mathcal T}\subseteq\Pi(\mu,\nu),
$$

which yields

$$
\operatorname{GW}(\mu,\nu)
\leq
\operatorname{GM}(\mu,\nu)
\leq
\operatorname{RGM}(\mu,\nu).
$$

RGM requires forward and reverse maps $F:\mathcal X\to\mathcal Y$ and $B:\mathcal Y\to\mathcal X$ satisfying equality of their induced couplings,

$$
(\operatorname{Id},F)_\#\mu
=
(B,\operatorname{Id})_\#\nu.
$$

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 [2109.14090].

## 2. Gaussian Gromov–Monge maps

For Gaussian measures

$$
\mu=\mathcal N(m_0,\Sigma_0)\subset\mathbb R^m,
\qquad
\nu=\mathcal N(m_1,\Sigma_1)\subset\mathbb R^n,
$$

with squared-Euclidean intra-space costs, the Gaussian-restricted GW problem reduces to an optimization over the cross-covariance $K=\operatorname{Cov}(X,Y)$. After centering and diagonalizing the covariances,

$$
\Sigma_0=P_0D_0P_0^\top,
\qquad
\Sigma_1=P_1D_1P_1^\top,
$$

the relevant Gaussian optimization is

$$
\max_K
\left\{
\|K\|_F^2:
\Sigma_1-K^\top\Sigma_0^{-1}K\succeq0
\right\}.
$$

Assume $n\leq m$ and write $D_0^{(n)}$ for the leading $n\times n$ block of $D_0$. The optimal cross-covariance matches the largest covariance eigenvalues:

$$
K^\star
=
P_0^\top
\begin{pmatrix}
\widetilde I_n(D_0^{(n)})^{1/2}D_1^{1/2}\\
0_{m-n,n}
\end{pmatrix}
P_1,
$$

where $\widetilde I_n=\operatorname{diag}(\pm1,\ldots,\pm1)$. Its objective value is

$$
\max\|K\|_F^2
=
\operatorname{tr}(D_0^{(n)}D_1)
=
\sum_{i=1}^n\alpha_i\beta_i.
$$

The Schur complement is saturated:

$$
K^{\star\top}\Sigma_0^{-1}K^\star=\Sigma_1.
$$

Consequently, $\operatorname{Cov}(Y\mid X)=0$, so the optimal Gaussian coupling is deterministic and is induced by an affine map,

$$
T(x)=m_1+P_1AP_0^\top(x-m_0),
$$

with

$$
A=
\begin{pmatrix}
\widetilde I_nD_1^{1/2}(D_0^{(n)})^{-1/2}
&
0_{n,m-n}
\end{pmatrix}.
$$

In PCA coordinates, the first $n$ source principal coordinates satisfy

$$
\widetilde y_i
=
\pm
\sqrt{\frac{\beta_i}{\alpha_i}}\,
\widetilde x_i,
\qquad i=1,\ldots,n,
$$

and the remaining $m-n$ source coordinates are discarded. The map is deterministic, pushes $\mu$ to $\nu$, 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 [2104.07970].

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

$$
GGW_2^2(\mu,\nu)
=
4\bigl(\operatorname{tr}(D_0)-\operatorname{tr}(D_1)\bigr)^2
+
8\|D_0^{(n)}-D_1\|_F^2
+
8\left(
\|D_0\|_F^2-\|D_0^{(n)}\|_F^2
\right).
$$

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

The construction differs from the ordinary Gaussian $W_2$ map. GW matches sorted principal variances and is invariant to independent orthogonal transformations of the source and target spaces. Ordinary $W_2$ transport instead depends on the full matrix interaction

$$
(\Sigma_0^{1/2}\Sigma_1\Sigma_0^{1/2})^{1/2}.
$$

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

$$
LGW_2^2(\mu,\nu)
\leq
GW_2^2(\mu,\nu)
\leq
GGW_2^2(\mu,\nu).
$$

The bounds coincide when $m=n$ and $D_1=\lambda D_0$, and in one dimension:

$$
GW_2^2(\mu,\nu)
=
12(\sigma_0^2-\sigma_1^2)^2,
$$

with optimizer

$$
T(x)=m_1\pm\frac{\sigma_1}{\sigma_0}(x-m_0).
$$

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 $n\geq d$, compact supports, and $\mu\ll\operatorname{Leb}_n$, an optimal 2-map exists. A 2-map has the form

$$
\pi
=
\alpha_1(\operatorname{id},T_1)_\#\mu
+
\alpha_2(\operatorname{id},T_2)_\#\mu,
$$

where $\alpha_1+\alpha_2=1$ 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 $d-1$, 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 $d-2$, 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 [2210.11945].

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

$$
\eta_x
=
\alpha_1(x)\delta_{T_1(x)-x}
+
\alpha_2(x)\delta_{T_2(x)-x}.
$$

A barycentric velocity,

$$
\bar v(x)
=
\alpha_1(x)(T_1(x)-x)
+
\alpha_2(x)(T_2(x)-x),
$$

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 [2310.09254].

For an entropic GW objective,

$$
\min_{\pi\in\Pi(\mu,\nu)}
\int
\left|c_X(x,x')-c_Y(y,y')\right|^2
\,d\pi(x,y)d\pi(x',y')
+
\varepsilon\,\mathrm{KL}(\pi\mid\mu\otimes\nu),
$$

the conditional disintegration is

$$
d\pi_\varepsilon^\star(x,y)
=
d\mu(x)\,d\pi_\varepsilon^\star(y\mid x).
$$

GENOT samples target-space noise $Z\sim\rho$, usually $\rho=\mathcal N(0,I_q)$, and learns a conditional flow

$$
\frac{d}{dt}\phi_t(z\mid x)
=
v_{t,\theta}(\phi_t(z\mid x)\mid x),
\qquad
\phi_0(z\mid x)=z.
$$

The terminal generator $\phi_1(\cdot\mid x)$ is trained to reproduce $\pi_\varepsilon^\star(\cdot\mid x)$. With linear interpolation

$$
[Z,Y]_t=(1-t)Z+tY,
$$

the conditional flow-matching objective is

$$
\mathcal L_{\mathrm{GENOT}}(\theta)
=
\mathbb E_{t,Z,(X,Y)\sim\pi_\varepsilon^\star}
\left[
\left\|
v_{t,\theta}([Z,Y]_t\mid X)-(Y-Z)
\right\|_2^2
\right].
$$

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

$$
Y=\phi_1(Z\mid X)
\quad\Longrightarrow\quad
(X,Y)\sim\pi_\varepsilon^\star.
$$

The learned object is therefore generally a stochastic conditional map,

$$
x\longmapsto\pi_\theta(\cdot\mid x),
$$

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

$$
T_\varepsilon(x)
=
\mathbb E_{Y\sim\pi_\varepsilon^\star(\cdot\mid x)}[Y],
$$

but it is generally not an optimal Monge map. A Monge map is recovered only when the conditional degenerates to $\delta_{T(x)}$. 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 $\mathcal O(n^2)$ or $\mathcal O(n^3)$ time depending on the cost structure and $\mathcal O(n^2)$ 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 $\pi$ of endpoint variables $(X_0,X_1)$, standard FM uses

$$
X_t=(1-t)X_0+tX_1
$$

and conditional velocity

$$
\dot X_t=X_1-X_0.
$$

The population vector field is

$$
u_t(x)
=
\mathbb E[X_1-X_0\mid X_t=x].
$$

A Gromov–Monge interpolation can use a deterministic map $T$:

$$
X_0\sim\mu,
\qquad
X_1=T(X_0),
\qquad
X_t=(1-t)X_0+tT(X_0).
$$

If $T(x)=Bx+b$, then

$$
X_t=M_tX_0+c_t,
\qquad
M_t=(1-t)I+tB,
$$

and, whenever $M_t$ is invertible on the relevant support, the induced affine velocity is

$$
v_t(x)
=
\dot M_tM_t^{-1}(x-c_t)+\dot c_t.
$$

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 $(X_0,X_1)$ is jointly Gaussian, then the flow-matching field for the linear interpolation is affine:

$$
u_t(x)
=
\mathbb E[X_1-X_0]
+
\operatorname{Cov}(X_1-X_0,X_t)
\operatorname{Cov}(X_t)^\dagger
\bigl(x-\mathbb E[X_t]\bigr).
$$

The pseudoinverse is used when the interpolation covariance is singular.

A third option interpolates Gaussian marginals directly,

$$
\rho_t=\mathcal N(m_t,\Sigma_t),
$$

with either linear covariance interpolation, map-induced covariance

$$
\Sigma_t=M_t\Sigma_0M_t^\top,
$$

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

$$
v_t(x)=\dot m_t+A_t(x-m_t)
$$

satisfies the continuity equation when

$$
\dot\Sigma_t=A_t\Sigma_t+\Sigma_tA_t^\top.
$$

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 $\Psi$, the endpoint is $\nabla\Psi(z_0)$ and the trajectory is

$$
x_t=(1-t)z_0+t\nabla\Psi(z_0).
$$

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 [2510.27385].

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 $N$ nodes and $C$ feature channels can be represented as

$$
E\in\mathbb R^{N\times N\times C},
$$

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

$$
(\sigma\cdot E)_{ij}
=
E_{\sigma^{-1}(i)\sigma^{-1}(j)}.
$$

The unordered graph is the orbit

$$
[E]=\{\sigma\cdot E:\sigma\in\mathfrak S_N\},
$$

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

For a compact group $G\subset O(D)$, the quotient metric is

$$
d_Q([x],[y])
=
\min_{g\in G}\|x-gy\|.
$$

For graph tensors, this becomes

$$
GM_{2,Z}^2([E],[F])
=
\min_{\sigma\in\mathfrak S_N}
\sum_{i,j=1}^N
\|E_{ij}-F_{\sigma(i)\sigma(j)}\|^2.
$$

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 [2608.26961].

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

$$
(\mathfrak q,\mathfrak q)_\#\gamma_\Gamma=\Gamma
$$

and

$$
\int\|x-y\|^2\,d\gamma_\Gamma(x,y)
=
\int d_Q(x,y)^2\,d\Gamma(x,y).
$$

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

$$
\mu_t
=
\operatorname{proj}^t_\#\gamma_\Gamma,
\qquad
\operatorname{proj}^t(x,y)=(1-t)x+ty,
$$

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 $E$ and target graph $F$, an approximate alignment $\widehat\sigma$ produces

$$
\widetilde F
=
\widehat\sigma^{-1}\cdot F.
$$

The training pair is $(E,\widetilde F)$, and the interpolant is

$$
E_t=(1-t)E+t\widetilde F.
$$

The minibatch velocity loss is

$$
J_B(v^\theta)
=
\frac1B\sum_{a=1}^B
\left\|
v^\theta(t^a,E_{t^a}^a)
-
(\widetilde F^a-E^a)
\right\|^2.
$$

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 $\gamma_t$, then

$$
\gamma_t(gx)=g\gamma_t(x),
$$

so the quotient flow

$$
\bar\gamma_t([x])=[\gamma_t(x)]
$$

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,

$$
\operatorname{ecc}_E(i)
=
\left(
\frac1N\sum_k\|E_{ik}\|^2
\right)^{1/2},
$$

and has cost $\mathcal O(N^2C+N\log N)$ 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 $0.796$ to $0.568$, degree MMD from $0.114$ to $0.018$, and clustering-coefficient MMD from $0.179$ to $0.034$. On QM9 at five Euler steps, GW with outer assignment achieved validity $0.9421$ and FCD $1.232$; on ZINC250k it achieved validity $0.7280$, while inner-only GW achieved FCD $15.024$.

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 $\sqrt{\beta_i/\alpha_i}$, 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.

Source: https://www.emergentmind.com/topics/gromov-monge-flow-matching