Metric-Flow Formulation Explained

• Metric-flow formulation is a framework where the metric actively governs dynamics, discrepancies, and interpolation paths in systems ranging from data transport to neural network training.
• It applies across diverse domains such as Wasserstein gradient flows, Fisher-information dynamics, and operator-based discrepancy measures, enhancing modeling versatility.
• Its significance lies in recasting traditional dynamics by making the metric an intrinsic variable, thereby offering novel insights and improved variational methods.

Metric-flow formulation designates a class of mathematical constructions in which a flow is expressed through a metric, a metric-induced discrepancy, or an explicitly evolving metric tensor. In the cited literature, the expression is used in several non-equivalent senses: to define transport paths that respect a data-induced geometry, to formulate gradient flows in abstract metric or Wasserstein spaces, to replace parameter flow by Fisher-metric flow, to evolve Riemannian metrics through neural-network training, and to recast field theories in terms of generalized metric variables or metric-compatible characteristic hierarchies (Kapuśniak et al., 2024, Matthes et al., 2017, Strandkvist et al., 2020, Halverson et al., 2023, Hohm et al., 2010, Mongwane, 2017). The common structural feature is that the metric is not merely background notation: it determines either the admissible motion, the discrepancy functional, the variational principle, or the dynamical variable itself.

1. Scope and principal meanings

Across the literature, metric-flow formulation appears in at least three precise roles. First, a fixed metric can define how a distributional or geometric flow is measured; this is the case in metric flow matching, Wasserstein gradient-flow theory, and inverse-problem formulations based on observation-space norms (Kapuśniak et al., 2024, Erbar et al., 2021, Li et al., 2022). Second, the metric itself can be the evolving object, as in Fisher-information metric flow and neural-network-induced flows on the space of Riemannian metrics (Strandkvist et al., 2020, Halverson et al., 2023). Third, some works are metric-based without introducing a genuine flow, but provide generalized scalar curvatures, Ricci-type tensors, or projected variational objects that are natural candidates for future flow equations (Hohm et al., 2010, Blumenhagen et al., 2015, Geiller et al., 2019).

Sense of the formulation	Dynamical object	Representative papers
Metric-constrained transport or matching	Paths, vector fields, empirical distributions	(Kapuśniak et al., 2024, Sun et al., 24 Apr 2025)
Variational gradient flow on metric spaces	Probability measures or discrete trajectories	(Matthes et al., 2017, Erbar et al., 2021)
Metric as evolving geometry	Fisher metric or Riemannian metric	(Strandkvist et al., 2020, Halverson et al., 2023)
Metric as structural field variable	Generalized metric, macroscopic metric, characteristic metric data	(Hohm et al., 2010, Blumenhagen et al., 2015, Wetterich, 2016, Mongwane, 2017)

This multiplicity matters because a common misconception is to treat all “metric-flow” papers as instances of metric evolution. The cited works show that the term often refers instead to metric-aware interpolation, metric-induced discrepancies, or metric-compatible hierarchy constructions rather than to an equation of the form $\partial_t g = \cdots$ (Kapuśniak et al., 2024, Li et al., 2022, Mädler, 2018).

2. Metric-aware transport and flow matching

In generative modeling, the metric-flow formulation is explicit in "Metric Flow Matching for Smooth Interpolations on the Data Manifold" (Kapuśniak et al., 2024). The setup is conditional flow matching between $p_0$ and $p_1$ on $\mathbb R^d$, with flow map $\psi_t$ generated by $u_t$ through

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$

Standard conditional flow matching typically uses straight Euclidean interpolants,

$x_t=(1-t)x_0+tx_1,$

and trains a vector field by

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$

The paper argues that under the manifold hypothesis, straight interpolants may leave the curved support $\mathcal M$, placing mass in off-manifold, high-uncertainty regions. The proposed replacement is a data-dependent Riemannian metric

$p_0$0

with geodesics obtained by minimizing the metric energy

$p_0$1

The practical interpolant is amortized by

$p_0$2

and learned through the geodesic objective

$p_0$3

Flow matching is then performed with the metric-weighted loss

$p_0$4

The paper emphasizes that the essential change is not only the norm in the loss, but the interpolant itself: straight Euclidean segments are replaced by approximate geodesics that depend implicitly on the full dataset. It also presents task-agnostic diagonal constructions such as the LAND metric

$p_0$5

and an RBF metric

$p_0$6

together with a simulation-free two-stage algorithm: first learn metric-aware interpolants, then train the vector field (Kapuśniak et al., 2024).

A related but distinct use appears in "Flow Matching Ergodic Coverage" (Sun et al., 24 Apr 2025). There the robot trajectory induces the empirical distribution

$p_0$7

and the paper introduces a second time variable $p_0$8 governing a flow of control sequences,

$p_0$9

This induces a path $p_1$0 satisfying

$p_1$1

where $p_1$2 is the empirical distribution flow. The metric-flow objective is then

$p_1$3

with $p_1$4 a reference flow in probability space. The paper states that this formal flow-matching problem is equivalent to a linear quadratic regulator problem with a closed-form solution, and uses that equivalence to support Stein variational gradient flow and Sinkhorn divergence flow as alternative ergodic metrics (Sun et al., 24 Apr 2025).

In both papers, the metric-flow formulation replaces a geometry-agnostic objective by a flow rule derived from a metric or measure-space geometry. What changes is not merely the loss value but the entire admissible path structure.

3. Gradient flows in abstract metric and Wasserstein spaces

A classical variational meaning of metric-flow formulation appears in "A Variational Formulation of the BDF2 Method for Metric Gradient Flows" (Matthes et al., 2017). The paper studies gradient flows of an energy $p_1$5 on a complete metric space $p_1$6 and extends the minimizing-movement philosophy beyond the implicit Euler/JKO scheme. Its central object is the BDF2 penalization

$p_1$7

with recursion

$p_1$8

In a smooth Hilbert setting this reproduces the classical BDF2 discretization

$p_1$9

but the paper formulates it without assuming linear structure or differentiability. Under lower semicontinuity, coercivity, and semi-convexity of the augmented functional, it proves well-posedness, a discrete EVI, and convergence of the piecewise-constant interpolants to a curve of steepest descent. The abstract convergence estimate is

$\mathbb R^d$0

even though the numerical experiments display behavior close to order $\mathbb R^d$1 in smooth examples (Matthes et al., 2017).

An analogous metric-flow formulation on a singular space is developed in "Gradient flow formulation of diffusion equations in the Wasserstein space over a metric graph" (Erbar et al., 2021). The state space is the Wasserstein space $\mathbb R^d$2 over a compact metric graph $\mathbb R^d$3. The paper first proves a Benamou–Brenier formula,

$\mathbb R^d$4

with the minimum taken over continuity-equation solutions on the graph. It then studies the free energy

$\mathbb R^d$5

where $\mathbb R^d$6 is relative entropy with respect to $\mathbb R^d$7 and $\mathbb R^d$8 is the interaction term

$\mathbb R^d$9

The corresponding diffusion / McKean–Vlasov equation

$\psi_t$0

is characterized by the energy-dissipation identity

$\psi_t$1

The paper stresses that metric graphs are geodesic but branching, and that entropy is not $\psi_t$2-convex along $\psi_t$3-geodesics in general. The resulting theory is therefore a direct metric-flow construction rather than an application of standard displacement-convex AGS theory (Erbar et al., 2021).

These works exhibit the variational core of the term: a flow is defined not by coordinates or vector fields in Euclidean space, but by an energy and a metric structure on the state space.

4. Information-geometric and neural metric flows

In "Beyond RG: from parameter flow to metric flow" (Strandkvist et al., 2020), the metric-flow formulation is information-geometric. A theory is a map

$\psi_t$4

and the model manifold carries the Fisher Information Metric

$\psi_t$5

A deformation of the predictive distribution by a question or observational scale,

$\psi_t$6

induces a family $\psi_t$7. The paper’s decisive claim is that ordinary RG parameter flow is only the special case in which

$\psi_t$8

with $\psi_t$9 the beta-function vector field. In general, the points $u_t$0 do not move; what flows is the metric of distinguishability itself. The mixed tensor

$u_t$1

encodes local contraction rates, and the paper argues that a generic metric flow cannot be reduced to a point flow on the original parameter manifold. It further observes that by augmenting the manifold with the extra coordinate $u_t$2, metric flow can always be represented as geometry in dimension $u_t$3, with RG corresponding to a degenerate case (Strandkvist et al., 2020).

"Metric Flows with Neural Networks" (Halverson et al., 2023) makes the metric tensor itself dynamical. A neural network parameterizes a Riemannian metric

$u_t$4

and gradient descent

$u_t$5

induces the exact metric evolution

$u_t$6

where

$u_t$7

is the metric neural tangent kernel. The paper emphasizes three generic properties of the finite-width flow: the kernel evolves in time, the flow is nonlocal, and different metric components can mix. In the infinite-width limit, the kernel freezes to $u_t$8 and the dynamics simplify to a fixed-kernel flow. Under additional architectural assumptions,

$u_t$9

so the dynamics become local:

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$0

The paper then shows that with a suitable loss this local regime realizes Perelman’s gradient formulation of Ricci flow. It also argues, through Calabi–Yau experiments, that frozen-kernel regimes perform poorly relative to finite-width networks because the latter possess an evolving metric-NTK and hence feature learning (Halverson et al., 2023).

Here the phrase metric-flow formulation is literal: the metric is the state variable, and the learning dynamics define a flow on the space of metrics.

5. Operator-theoretic, categorical, and combinatorial formulations

A different meaning of metric-flow formulation is operator-induced discrepancy. In "Flow Measurement: An Inverse Problem Formulation" (Li et al., 2022), the physical flow is encoded by the particle map

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$1

with particle density transported by

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$2

Measurement is not defined by a geometric metric on flows, but by the observation operator

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$3

where $$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$4 solves a wave equation with source $$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$5. The least-squares objective

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$6

induces the relevant discrepancy geometry. The paper explicitly identifies

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$7

as the closest object in the paper to a metric on source states, and derives the adjoint-state gradient

$$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$8

This is therefore an operator-based metric-flow formulation of measurement rather than a Riemannian or Wasserstein one (Li et al., 2022).

In "Metric Limits in Categories with a Flow" (Cruz, 2019), the word flow refers to a monoidal action of $$\frac{d}{dt}\psi_t(x)=u_t(\psi_t(x)),\qquad \psi_0(x)=x.$$9 on a category by endofunctors $x_t=(1-t)x_0+tx_1,$0, together with coherence morphisms. This structure induces weak $x_t=(1-t)x_0+tx_1,$1-interleavings and the interleaving distance

$x_t=(1-t)x_0+tx_1,$2

The paper’s main theorems show that categorical inverse limits or direct limits of shifted diagrams produce metric limits, and that the Yoneda embedding plus Kan extension yields a general completion mechanism. Here the metric-flow formulation is categorical: flow data generate a pseudometric, and convergence becomes a problem of categorical completeness (Cruz, 2019).

"A Flow Formulation for Horizontal Coordinate Assignment with Prescribed Width" (Jünger et al., 2018) uses flow in a combinatorial sense. In layered graph drawing, auxiliary network arcs carry horizontal spacing, and node coordinates are recovered by cumulative flow,

$x_t=(1-t)x_0+tx_1,$3

The cost of flow exactly matches total horizontal edge length, and prescribed width is imposed by bounding source flow. The paper states the principle succinctly: flow encodes horizontal distance, and cost encodes horizontal edge length (Jünger et al., 2018). This is a metric-flow formulation because a geometric layout metric is represented by a min-cost flow on an auxiliary graph.

These papers demonstrate that metric-flow formulation need not mean geometric flow on a manifold. It can also mean that a metric or discrepancy emerges from an operator, a categorical action, or a network-flow encoding.

6. Generalized metric variables, one-metric RG flow, and characteristic null evolution

Several papers provide metric formulations that are structurally close to flow theory without always defining a metric flow in the strict sense. In "Generalized metric formulation of double field theory" (Hohm et al., 2010), the basic field is the generalized metric $x_t=(1-t)x_0+tx_1,$4 on doubled spacetime, satisfying

$x_t=(1-t)x_0+tx_1,$5

with action

$x_t=(1-t)x_0+tx_1,$6

The theory defines a generalized scalar curvature $x_t=(1-t)x_0+tx_1,$7 and a projected generalized Ricci tensor $x_t=(1-t)x_0+tx_1,$8, but the paper explicitly states that it does not formulate any Ricci-flow analogue or other auxiliary-time flow equation. "Generalized Metric Formulation of Double Field Theory on Group Manifolds" (Blumenhagen et al., 2015) extends this structure to $x_t=(1-t)x_0+tx_1,$9, replacing partial derivatives by background-dependent covariant derivatives and adding an explicit flux term

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$0

again yielding $$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$1 and $$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$2 but no explicit flow equation.

By contrast, "Gauge invariant flow equation" (Wetterich, 2016) does define a genuine one-metric functional flow for gravity and gauge theory. Its central equation is

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$3

with

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$4

and the projected propagator determined by

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$5

The formulation depends on projection to physical fluctuations and on a gauge-invariant effective average action depending only on one macroscopic metric $$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$6. In this sense it is a one-metric RG flow rather than a geometric flow of the spacetime metric itself (Wetterich, 2016).

Two characteristic formulations recast Einstein-type equations as hierarchical metric evolutions along null directions. "Affine-null metric formulation of General Relativity at two intersecting null hypersurfaces" (Mädler, 2018) uses coordinates $$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$7 adapted to a null boundary and an intersecting null hypersurface, with affine parameters in both null directions. The metric

$$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$8

is evolved by first solving boundary transport equations in $$\mathcal{L}_{\rm CFM}(\theta)=\mathbb{E}\left\|v_{t,\theta}(x_t)-\dot x_t\right\|^2.$$9, then hypersurface equations in $\mathcal M$0, then an evolution equation for $\mathcal M$1. "Characteristic Formulation for Metric $\mathcal M$2 Gravity" (Mongwane, 2017) does the same in Bondi–Sachs form for vacuum metric $\mathcal M$3 gravity. There the Bondi–Sachs metric variables $\mathcal M$4 and the scalar curvature $\mathcal M$5 satisfy a nested radial hierarchy, with $\mathcal M$6 evolved by the trace equation

$\mathcal M$7

Both papers are aptly described as metric-flow formulations because metric data are propagated hierarchically along null directions.

Finally, "Metric formulation of the simple theory of 3d massive gravity" (Geiller et al., 2019) derives closed nonlinear metric field equations from a first-order Lorentz-breaking theory. The paper explicitly states that there is no purely metric local nonlinear action, and it does not introduce a flow. Its relevance is therefore structural: it isolates the metric sector and yields modified Einstein equations with higher-derivative terms, but not a metric-flow equation (Geiller et al., 2019).

7. Limitations, ambiguities, and recurring themes

The most persistent limitation is semantic rather than technical: the literature does not use metric-flow formulation in a single universal sense. In some works the metric determines the geometry of interpolants or discrepancies; in others it is the evolving state variable; in still others it is a structural field variable without an auxiliary-time flow. Treating these as interchangeable obscures the mathematical content (Kapuśniak et al., 2024, Strandkvist et al., 2020, Hohm et al., 2010).

Several papers also record concrete ambiguities or restrictions. The MFM paper notes a minor ambiguity in the RBF metric notation: $\mathcal M$8 appears with an $\mathcal M$9-dependence in one place, while later the weights are described as learned positive coefficients $p_0$00; the intended construction appears to be standard learned positive RBF weights (Kapuśniak et al., 2024). The flow-measurement paper reports relative errors of $p_0$01 for a virtual ADCP and $p_0$02 for the proposed method in one experiment, while the prose says the new method is more accurate; the paper itself therefore contains an apparent typo or reversal in that comparison (Li et al., 2022).

On the analytical side, the abstract metric BDF2 theory proves only order $p_0$03 convergence under weak assumptions, even though the formal smooth-space method is second order and the numerical experiments behave accordingly (Matthes et al., 2017). On metric graphs, the entropy is not $p_0$04-convex along $p_0$05-geodesics in general because branching destroys the standard semiconvex structure, so the gradient-flow formulation requires direct regularization arguments rather than a straightforward AGS displacement-convexity theory (Erbar et al., 2021). In neural metric flows, frozen-kernel regimes correspond to fixed-kernel methods and exhibit poor learning of numerical Calabi–Yau metrics relative to finite-width networks with evolving metric-NTK (Halverson et al., 2023). In the one-metric FRG formulation, the paper argues for the existence of an optimal macroscopic field yielding a closed flow, but does not give a rigorous global proof of that existence (Wetterich, 2016).

What recurs across these otherwise disparate settings is the relocation of dynamics from naive coordinate evolution to a geometry-sensitive object. That object may be a data-induced Riemannian metric, a Wasserstein distance, an interleaving pseudometric, a Fisher metric, a generalized metric on doubled spacetime, or an observation-space norm. The phrase metric-flow formulation therefore names a structural strategy rather than a single theorem: formulate motion, matching, optimization, or field equations so that the metric is constitutive of the dynamics rather than an external afterthought.