Edge Flow: Models and Applications

• Edge Flow is a multifaceted concept involving edge-centric dynamics, from gradient descent at the stability threshold to edge signal propagation on graphs and optical flow in computer vision.
• It encompasses mathematical models where ‘edge’ denotes a critical locus, seen in systems like dynamical optimization, spectral and harmonic flows, and segmentation architectures.
• Practical applications include enhancing deep-learning optimization, performing efficient graph signal processing, and analyzing boundary phenomena in fluid mechanics and plasma physics.

Searching arXiv for recent and relevant papers on “Edge Flow” and closely related uses of the term. Edge Flow is a polysemous technical term rather than a single standardized object. In recent literature it denotes, among other things, a continuous-time model for gradient descent at the edge of stability, an edge-signal formalism on graphs and cell complexes, edge-aware optical-flow and interactive-segmentation architectures, and several boundary-centered flow phenomena in fluid mechanics, plasma physics, topology, causality, and geometric analysis (Marion, 16 Jun 2026, Spreuer et al., 29 Aug 2025, Revaud et al., 2015, Beneitez et al., 2019). A common motif is the treatment of “edge” as a privileged locus of dynamics: a stability threshold, a graph 1-cell, a motion boundary, a physical interface, or a singular geometric stratum. This suggests that the expression is best understood as a family of domain-specific constructs rather than a unified theory.

1. Terminological scope and principal usages

The term appears in several distinct technical settings.

Domain	Mathematical object	Representative paper
Deep-learning optimization	Three coupled ODEs at the edge of stability	(Marion, 16 Jun 2026)
Graph signal processing	Edge-flow signals $f \in C_1=\mathbb{R}^m$ on oriented edges	(Spreuer et al., 29 Aug 2025)
Optical flow	Edge-preserving interpolation or regularization	(Revaud et al., 2015)
Interactive segmentation	Edge-guided flow for click-conditioned masks	(Hao et al., 2021)
Topological insulators	Spectral flow of APS-boundary Dirac operators (“edge flow”)	(Yu et al., 2016)
Causal fairness	Belief flowing along a causal edge with indirect paths attenuated	(Ravishankar et al., 2020)

In graph settings, an edge flow is literally a quantity attached to oriented edges. In vision, the term is usually shorthand for flow estimation or feature propagation that explicitly respects image edges. In optimization, Edge Flow is the title of a dynamical model intended to interpolate gradient descent when the largest Hessian eigenvalue hovers near $2/\eta$. In topological and causal settings, the term acquires an analytical meaning: spectral crossing counts in one case, and direct causal influence along an edge in the other.

The breadth of usage is sufficiently large that any encyclopedia treatment must be disambiguating. The same phrase can refer either to a state variable defined on edges, a dynamical law indexed by edges, or a boundary-localized transport phenomenon.

2. Edge Flow at the edge of stability in optimization

In the optimization literature, Edge Flow is a continuous-time model for full-batch gradient descent in the edge-of-stability regime, where the sharpness

$S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$

hovers near $2/\eta$ and classical gradient-flow arguments cease to be reliable (Marion, 16 Jun 2026). The model decomposes the iterate dynamics into a slowly drifting center $\theta$, a unit oscillation direction $v$, and an oscillation magnitude $\alpha$. Its governing ODEs are

$$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$

$$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$

$$\frac{d\alpha}{dt} = \bigl(v^T H(\theta)v-\tfrac{2}{\eta}\bigr)\alpha,$$

with $2/\eta$0 and symmetrized loss

$2/\eta$1

The construction is explicitly mechanistic. The center follows a modified gradient flow on the symmetrized loss, the direction tracks a top eigenvector of the Hessian through Rayleigh-quotient dynamics, and the magnitude grows or decays exponentially depending on whether the instantaneous sharpness exceeds or falls below $2/\eta$2 (Marion, 16 Jun 2026). The salient interpretive point is the self-stabilization feedback loop: when $2/\eta$3, the oscillation amplitude increases, which alters the center dynamics through the two-point gradient average and pushes the center toward lower curvature; when $2/\eta$4, the oscillation decays and progressive sharpening resumes.

The corresponding discretization, Edge Gradient Descent, requires exactly two gradient evaluations and one Hessian–vector product per substep: $2/\eta$5

$2/\eta$6

Empirically, the model was tested across MLP, CNN, ResNet, and ViT architectures, with both MSE and cross-entropy losses on small CIFAR-10 subsets, and was reported to track gradient-descent dynamics at least as faithfully as previously proposed continuous-time edge-of-stability models while also resolving the oscillation of sharpness at onset (Marion, 16 Jun 2026). Within this literature, “Edge Flow” therefore names a specific reduced-order dynamical system rather than a generic edge-aware algorithm.

3. Edge flows on graphs, cell complexes, and edge-centric network dynamics

In graph signal processing, an edge flow is a signal defined on oriented edges. For a graph $2/\eta$7 with fixed edge orientation, any flow on $2/\eta$8 is represented as

$2/\eta$9

and a collection of $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$0 samples becomes

$S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$1

The recent cell-complex inference problem asks for a lift of the graph to a 2-dimensional cell complex $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$2 by choosing 2-cells whose boundary operator $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$3 permits a sparse Hodge decomposition of the observed edge flows (Spreuer et al., 29 Aug 2025). The 1-Hodge Laplacian is

$S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$4

and any flow decomposes uniquely into gradient, curl, and harmonic components: $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$5

The optimization target is the harmonic residual,

$S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$6

under the constraint $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$7. Since the general problem is NP-hard, the matrix-factorization heuristic MFCI relaxes the discrete search over $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$8, factorizes the harmonic component, and then discretizes retained factors into simple cycles by a deterministic or random-walk heuristic (Spreuer et al., 29 Aug 2025). Computational experiments reported that MFCI is significantly less computationally expensive than prior heuristics, with only marginally worse performance in most settings, and that in specifically noisy settings it outperforms the previous state of the art in both solution quality and computational speed.

A distinct but related edge-centric program studies how flows redistribute after an edge failure. For an undirected weighted network with incidence matrix $S(w_t)=\lambda_{\max}\!\bigl(\nabla^2L(w_t)\bigr)$9, conductance matrix $2/\eta$0, Laplacian $2/\eta$1, and pseudoinverse $2/\eta$2, Schaub et al. define the edge-to-edge transfer matrix

$2/\eta$3

and the flow-redistribution matrix

$2/\eta$4

where

$2/\eta$5

is the edge embeddedness (Schaub et al., 2013). Here the “flow” is not merely a signal on edges but the failure-induced redistribution of load from one edge to all others, encoded by the Laplacian pseudoinverse. The framework is explicitly global and was illustrated on the Iberian Power grid, road networks, and the C. elegans neuronal network (Schaub et al., 2013).

A third graph-theoretic use concerns evolutionary flows on edge weights. The local entropy flow and cohesion flow are ODEs

$2/\eta$6

designed for community detection and node classification (Zhao et al., 25 Jun 2026). The local entropy $2/\eta$7 is a symmetric KL-divergence of two local $2/\eta$8-walks, while $2/\eta$9 is an exponential of a reciprocal-sum cohesion score. The theory establishes global existence and uniqueness for both flows, and for sparse graphs their per-iteration complexity is $\theta$0 when the average degree is $\theta$1 (Zhao et al., 25 Jun 2026). In this branch of the literature, “edge flow” means the evolution of weights attached to graph edges rather than a transport field over them.

4. Edge-aware flow in computer vision

Computer-vision uses of the term are dominated by optical-flow estimation and closely related tasks in which motion boundaries should align with image edges. EpicFlow is an early canonical formulation: it starts from approximately $\theta$2 sparse correspondences and produces a dense optical flow by edge-preserving interpolation based on an edge-aware geodesic distance

$\theta$3

with contour cost $\theta$4 derived from SED (Revaud et al., 2015). Dense interpolation is performed by either Nadaraya–Watson weighting or a locally weighted affine model, followed by a single-scale variational refinement. The method was reported to achieve MPI-Sintel test AEE $\theta$5, KITTI test AEE-noc $\theta$6, and to run in about $\theta$7 s per MPI-Sintel pair on one $\theta$8 GHz CPU core (Revaud et al., 2015). Its significance lies in replacing coarse-to-fine warping by sparse-to-dense interpolation that explicitly resists smoothing across motion boundaries and occlusions.

The variational tradition remains active. The edge-preserving $\theta$9/$v$0 model for optical flow minimizes

$v$1

with $v$2 or $v$3, and is solved numerically by the Chambolle–Pock primal-dual algorithm, iterated median filtering, and a weighted median post-processing step (Doshi et al., 2022). On Middlebury, it was reported to achieve the best average angular and end-point errors among several Horn–Schunck-based variational methods, with overall averages AAE/EPE $v$4 (Doshi et al., 2022).

Deep architectures introduce edge awareness more explicitly. EA-Net is a two-stage flow-based video frame interpolator consisting of edge-guided flow estimation and edge-protected frame synthesis (Zhao et al., 2021). It injects Canny-derived edge maps through three fusion mechanisms, refines the coarse bidirectional flows in a second U-Net, and synthesizes the intermediate frame through

$v$5

Two discriminators, one on RGB frames and one on edge maps, support adversarial training. Reported results include $v$6 on Vimeo90K and $v$7 on Adobe240-fps (Zhao et al., 2021).

A different deployment axis is edge hardware. NeuFlow v2 is a lightweight optical-flow model built for edge devices, with a shallow multi-scale backbone, cross-attention and global matching at $v$8 resolution, and CNN-based recurrent refinement (Zhang et al., 2024). It was reported to run at over $v$9 FPS on $\alpha$0 images on a Jetson Orin Nano, with Sintel clean EPE $\alpha$1, Sintel final EPE $\alpha$2, and KITTI 2015 EPE $\alpha$3 (Zhang et al., 2024). Here “edge” refers to deployment constraints rather than motion boundaries.

The name EdgeFlow also appears in interactive segmentation. This architecture couples an HRNet-18 backbone, OCR segmentation head, interaction branch, and edge-guided flow modules so that user clicks and previous-mask edges are fused at multiple stages (Hao et al., 2021). The flow module predicts a feature-alignment field

$\alpha$4

which warps image and interaction features into better correspondence. The method was reported to achieve state-of-the-art performance without post-processing or iterative optimization, with end-to-end inference per click of about $\alpha$5 ms on a V100 (Hao et al., 2021).

5. Edge-centered flow phenomena in fluid mechanics and plasma physics

In wall-bounded shear flows, the dominant technical use of “edge” is the edge manifold separating qualitatively different long-time outcomes. For a spatially developing Blasius boundary layer, the perturbation field $\alpha$6 obeys

$\alpha$7

and the edge manifold $\alpha$8 is the codimension-1 hypersurface separating trajectories that return to the laminar boundary layer from those that become turbulent (Beneitez et al., 2019). The observable

$\alpha$9

is used to discriminate outcomes, and a moving-box technique allows long-time tracking in the parameter range relevant to bypass transition. The dynamics on the tracked edge trajectory features a localized packet of streamwise streaks and vortices, recurrent streak switching, and, at long enough times, coexistence of Tollmien–Schlichting waves with the coherent structure characteristic of edge trajectories (Beneitez et al., 2019). The paper proposes a reinterpretation of the edge as a manifold dividing bypass transition from classical TS transition once the laminar Blasius profile itself is linearly unstable.

Zammert and Eckhardt identified closely related edge-state dynamics in plane Poiseuille flow. In short and narrow domains they found a periodically bursting edge state with period of order $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$0 time units, localized to one side of the channel in the wall-normal direction (Zammert et al., 2013). In large domains they found an attracting state on the laminar–turbulent boundary that is localized in both streamwise and spanwise directions and chaotic, a “spotlike edge state” (Zammert et al., 2014). In square duct flow, Scherer, Uhlmann, and Kawahara reported that the edge state is a chaotic attractor in the full space but becomes time-periodic in mirror-symmetric subspaces; the dynamics alternates quiescent phases and bursting episodes, with wall switching mediated by streak–vortex interactions (Scherer et al., 28 Mar 2025). These results establish that “edge” in hydrodynamic transition theory is a state-space separator populated by coherent structures, not a transport field defined on physical boundaries.

Plasma and environmental flow literatures use “edge flow” more literally. In tokamak pedestal stability, two-fluid MHD calculations showed that the stabilizing effect of toroidal rotation on edge localized modes can be significantly enhanced by increasing the edge plasma density; for the fastest-growing high-$$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$1 mode, at $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$2 the normalized growth rate reduction changes from $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$3 at $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$4 to $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$5 at $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$6 (Cheng et al., 2016). Over flexible aquatic vegetation, direct numerical simulations of a half-channel partially obstructed by stems revealed a canopy-edge vortex driven by the drag discontinuity at the canopy tip, with mean streamwise vorticity magnitude $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$7–$$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$8, drawing high-momentum fluid into the canopy and ejecting low-momentum fluid upward (Rota et al., 21 Jul 2025). In these cases, the edge is a spatial interface where transport and stability are reorganized.

6. Spectral, causal, and geometric extensions

In topological-insulator theory, “edge flow” refers to the spectral flow of a one-parameter family of $$\frac{d\theta}{dt} = -\tfrac12\Bigl[\nabla L(\theta+\alpha v)+\nabla L(\theta-\alpha v)\Bigr],$$9-dimensional Dirac operators with Atiyah–Patodi–Singer boundary conditions induced by Kramers degeneracy (Yu et al., 2016). For the family $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$0, $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$1, the spectral flow is

$$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$2

The parity of this spectral flow equals the Kane–Mele $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$3 invariant, bulk states do not contribute because they remain gapped, and in the Fu–Kane spin-pump model a single upward crossing yields $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$4 (Yu et al., 2016). The phrase “edge flow” therefore names a boundary-spectrum counting invariant rather than a transport process.

In causal fairness, Edge Flow is the belief flowing along a causal edge after indirect path influences are attenuated (Ravishankar et al., 2020). For a parent set $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$5, the paper defines

$$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$6

where the scaling factor $$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$7 averages path-specific effects along the direct edge set. Under an error-free linear model of the conditional probability table, edge-unfairness equals the corresponding edge weight, and the paper proves that cumulative unfairness in any downstream decision is absent when edge unfairness is absent (Ravishankar et al., 2020). Here the term acquires a causal-semantic meaning: the “flow” is a direct contribution to belief propagation in a causal Bayesian network.

Geometric analysis contributes yet another adjacent usage through the edge Yamabe flow. Bahuaud and Vertman study the normalized Yamabe flow on compact Riemannian manifolds with incomplete edge singularities and negative Yamabe invariant, proving uniqueness, long-time existence, and convergence from initial metrics with everywhere negative scalar curvature (Bahuaud et al., 2016). In conformal form, with

$$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$8

the scalar equation becomes

$$\frac{dv}{dt} = 2\Bigl(H(\theta)-\bigl(v^T H(\theta)\,v\bigr)I\Bigr)v,$$9

with $$\frac{d\alpha}{dt} = \bigl(v^T H(\theta)v-\tfrac{2}{\eta}\bigr)\alpha,$$0 and $$\frac{d\alpha}{dt} = \bigl(v^T H(\theta)v-\tfrac{2}{\eta}\bigr)\alpha,$$1 (Bahuaud et al., 2016). Although this is not usually called “Edge Flow” without qualification, it illustrates a broader analytic convention in which “edge” modifies a flow evolving on spaces with edge singularities.

Across these spectral, causal, and geometric usages, “edge flow” ceases to mean either optical flow or graph-signal transport. It instead denotes a counting invariant, a direct causal contribution, or a PDE evolution constrained by edge geometry. This breadth is not accidental: the term systematically migrates to settings where edges are the primary carriers of structure, whether those edges are physical, combinatorial, spectral, or singular.