Zero-Flow Criterion in Graphs, Fluids, and Learning

• Zero-Flow Criterion is a cross-domain concept that uses the vanishing of flow-like quantities to certify hidden structural, topological, and analytical properties.
• In graph theory, it underpins nowhere-zero flow existence in signed graphs and informs reconfiguration and coloring methodologies with concrete flow bounds.
• In fluid mechanics and machine learning, zero-flow conditions validate convergence, dissipation limits, and representational sufficiency, revealing critical system behaviors.

“Zero-Flow Criterion” does not designate a single universally standardized theorem across contemporary research. In current usage, it denotes a family of class-specific criteria in which the vanishing of a flow-like quantity, or the existence of a nowhere-zero flow, certifies a structural, topological, analytic, or statistical property. In signed-graph theory, the term is naturally attached to criteria for existence of nowhere-zero flows and bounded $k$-flows; in fluid mechanics, it appears in Kato-type vanishing-dissipation criteria for inviscid limits; in dynamical systems, it arises from boundary tangency where the normal component of the flow vanishes; and in representation learning, it is an explicit midpoint criterion for rectified flows trained with independent coupling (Kaiser et al., 2016, Sueur, 2017, Sánchez-Gabites, 2024, Wang et al., 31 Jan 2026). This suggests that the expression is best understood as a cross-domain label for vanishing or nonvanishing conditions that detect hidden structure.

1. Graph-theoretic meaning: nowhere-zero flow existence

In graph theory, the most classical zero-flow interpretation is not “flow equals zero,” but the criterion for avoiding zero values on edges. For a signed graph, a nowhere-zero $k$-flow is an assignment of an orientation and values from $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$ to the edges so that Kirchhoff’s law holds at every vertex, and a signed graph is flow-admissible when it admits at least one nowhere-zero flow (Parsaei-Majd, 2 Oct 2025). The survey literature gives the fundamental criterion: a signed graph is flow-admissible if and only if every edge belongs to a signed circuit, where a signed circuit is either a balanced circuit or a barbell (Kaiser et al., 2016).

Several recent class-specific sharpenings turn flow-admissibility itself into an effective zero-flow criterion. For signed series-parallel graphs, flow-admissibility is equivalent to the condition that every edge lies in a signed circuit, and every flow-admissible signed series-parallel graph has a nowhere-zero $6$-flow (Kaiser et al., 2014). For signed circular ladders and signed Möbius ladders, flow-admissibility implies a nowhere-zero $5$-flow, with one explicit exception: the unique signed $CL_4$ with three negative edges up to switching equivalence has a nowhere-zero $6$-flow but no nowhere-zero $k$-flow for any positive integer $k<6$ (Parsaei-Majd, 2 Oct 2025). For 3-edge-connected signed graphs, the criterion becomes coarser but still bounded: every flow-admissible, 3-edge-connected signed graph has a nowhere-zero $8$-flow (Esperet et al., 19 Dec 2025).

In ordinary graph flow theory, the same criterion often appears in orientation language. A graph admits a nowhere-zero $k$0-flow if and only if it admits a modulo $k$1-orientation, and for graphs with independence number at most $k$2 there is an exact obstruction characterization: $k$3 admits a nowhere-zero $k$4-flow if and only if $k$5 and the $k$6-reduction of $k$7 is not in $k$8 (Li et al., 2017). A plausible implication is that, in combinatorial usage, “Zero-Flow Criterion” most often means a criterion for the nonexistence of zero values on edges rather than a criterion for vanishing of a transport field.

2. Structural, constructive, and reconfiguration criteria in combinatorics

Beyond existence, the literature develops local structural criteria that force low-order nowhere-zero flows. In signed cubic bipartite graphs with exactly two negative edges, flow-admissibility and bipartiteness imply a nowhere-zero $k$9-flow with $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$0 (Parsaei-Majd, 2 Oct 2025). A second criterion in the same setting states that if a signed cubic bipartite graph has a $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$1-factorization $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$2 such that two of the induced $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$3-factors are balanced, then the graph admits a nowhere-zero $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$4-flow; the construction is explicit, obtained by sending value $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$5 along each circuit of $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$6 and value $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$7 along $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$8 (Parsaei-Majd, 2 Oct 2025). For signed ladders, positive-square deletion provides the main reduction lemma: if a signed ladder has a positive square and the reduced graph has a nowhere-zero $\{\pm 1,\pm 2,\ldots,\pm(k-1)\}$9-flow, then the original graph also has a nowhere-zero $6$0-flow (Parsaei-Majd, 2 Oct 2025).

A different constructive criterion comes from the module-theoretic viewpoint on flows. If $6$1, both $6$2 and $6$3 admit nowhere-zero $6$4-flows, and the overlap has at most $6$5 common edges, then $6$6 admits a nowhere-zero $6$7-flow; for nowhere-zero $6$8-flows, the overlap bound improves to $6$9 edges when the common edges induce a connected subgraph (Zhang et al., 2021). This is a genuine zero-exclusion criterion: linear combinations of subgraph flows are used to avoid cancellations on the overlap.

The reconfiguration viewpoint shifts the question from existence to connectivity of the space of nowhere-zero flows. The reconfiguration graph $5$0 has as vertices all nowhere-zero $5$1-flows, and two flows are adjacent when the support of their difference is a cycle (Esperet et al., 19 Dec 2025). The central conjecture states that for every 2-edge-connected graph $5$2, the reconfiguration graph $5$3 is connected (Esperet et al., 19 Dec 2025). The paper shows that group structure affects the answer: $5$4 is disconnected, while $5$5 is connected (Esperet et al., 19 Dec 2025). This sharply distinguishes reconfiguration from classical existence theory, where only the group order matters.

A related criterion appears in cubic graph coloring. If a bridgeless cubic graph $5$6 has a perfect matching $5$7 such that $5$8 admits a non-conflicting nowhere-zero $5$9-flow with respect to $CL_4$0, then $CL_4$1 admits a normal $CL_4$2-edge-coloring (Mkrtchyan, 2024). Here the zero-flow condition is not on $CL_4$3 itself but on a contraction by the complementary $CL_4$4-factor, and the “non-conflicting” restriction prevents a local $CL_4$5 obstruction along $CL_4$6 (Mkrtchyan, 2024).

3. Boundary-layer and interfacial criteria in fluid mechanics

In viscous–inviscid limits, “zero-flow criterion” typically refers not to zero velocity, but to vanishing dissipation or vanishing interfacial defect. For vortex sheets, the main result is a Kato-type criterion for convergence of Navier–Stokes Leray solutions $CL_4$7 to a vortex-sheet Euler solution $CL_4$8. If, in an $CL_4$9-neighborhood $6$0 of the sheet, the dissipation satisfies

$6$1

and the gradient mismatch across the sheet satisfies

$6$2

then

$6$3

(Sueur, 2017). The paper explicitly interprets this as a zero-dissipation or zero-defect criterion near a moving internal interface rather than a criterion that the flow itself vanishes (Sueur, 2017).

The stochastic analogue preserves the same conceptual structure. For stochastic Navier–Stokes flows in a bounded domain with no-slip boundary, the analogue of Kato’s criterion is the vanishing of viscous dissipation in an $6$4-thick boundary strip: $6$5 Under the theorem’s assumptions, this is equivalent to strong mean convergence to the deterministic Euler solution: $6$6 (Goodair et al., 2023). The paper further shows that the standard noise scaling $6$7 is optimal in the sense that it is the smallest noise-amplitude scale for which the full Kato equivalence survives without strengthening the boundary-layer hypothesis (Goodair et al., 2023).

A recurring misconception is that these are “zero-flow” criteria in the literal sense of zero velocity. The cited results say something narrower and more technical: the relevant vanishing quantity is boundary-layer or interfacial dissipation, and in the vortex-sheet case also a cross-sheet gradient defect (Sueur, 2017, Goodair et al., 2023).

4. Boundary tangency, topology, and zero-normal-flow sets

In topological dynamics, the term is closest to a zero-normal-flow or tangency criterion. For a flow in $6$8 and an isolating block $6$9, the relevant input is the boundary decomposition

$k$0

where $k$1 is the tangency set, i.e. in the smooth case the subset of the boundary where the normal component of the flow vanishes (Sánchez-Gabites, 2024). If $k$2 is connected and $k$3 for the maximal invariant set $k$4, then $k$5 must be a handlebody and there must exist a complete system of cutting disks $k$6 such that each $k$7 intersects the set of tangency curves transversally at two points (Sánchez-Gabites, 2024).

The contrapositive is the operative criterion: if $k$8 is not a handlebody, or if no complete cut system exists with each boundary circle meeting the tangency set exactly twice, then $k$9 (Sánchez-Gabites, 2024). This is a boundary-only criterion: it extracts cohomological information about $k<6$0 from the geometry of the zero-normal-component locus on $k<6$1, together with the adjacent entry/exit decomposition. The paper also proves the criterion is “as sharp as possible with the information it uses as an input,” because whenever the boundary coloring satisfies the necessary conditions, there exists a realizing flow with $k<6$2 a single rest point (Sánchez-Gabites, 2024).

A plausible synthesis is that this literature uses “zero-flow” most literally: the input is precisely the boundary subset where the normal component vanishes. The output, however, is not a statement about transport intensity but about invariant-set topology.

5. Energetic and network interpretations

Some literatures explicitly state that they do not introduce a standalone zero-flow criterion, yet they identify closely related vanishing conditions. In nonequilibrium thermodynamics of viscous flow, the relevant object is the extended general evolution criterion

$k<6$3

The paper states that it does not provide a true “zero-flow criterion”; the closest rigorous statement is the zero-surface-term criterion

$k<6$4

which is ensured by fixed boundary conditions, in particular $k<6$5, and recovers the classical Glansdorff–Prigogine general evolution criterion if and only if the surface term is zero (Hochberg et al., 2024). The physical point is that this is not a criterion for $k<6$6; steady nonequilibrium flow with nonzero entropy production is allowed (Hochberg et al., 2024).

In power systems, the branch-flow model with line shunts also alters the meaning of “zero flow.” The paper explicitly notes that it does not define a named “Zero-Flow Criterion,” but its equations imply an exact terminal criterion. With shunts, $k<6$7, $k<6$8, and $k<6$9 are terminal quantities at each end of a $8$0-model line, and for a line end $8$1,

$8$2

where $8$3 (Zhou et al., 2020). The paper emphasizes the main interpretive consequence: with nonzero line shunts, zero series transfer $8$4 does not imply zero terminal branch flow, because the shunt current can remain nonzero (Zhou et al., 2020).

These two usages are methodologically similar. In both, the vanishing condition applies not to the bulk field simpliciter but to a boundary contribution or terminal quantity. A plausible implication is that “Zero-Flow Criterion” often functions as shorthand for a carefully localized zero condition.

6. Representation learning and statistical sufficiency

In machine learning, “zero-flow criterion” appears as an explicit formal theorem rather than an interpretive label. For rectified flow trained with independent coupling, the midpoint velocity field satisfies

$8$5

if and only if

$8$6

This is Theorem 3.1 of “Zero-Flow Encoders” (Wang et al., 31 Jan 2026). The midpoint formula is

$8$7

and the converse direction is proved under the assumption that the endpoint variables are independent and have non-vanishing characteristic functions (Wang et al., 31 Jan 2026).

The conditional version is the paper’s main statistical application. For a representation $8$8, the conditional zero-flow condition

$8$9

is equivalent to

$k$00

hence to

$k$01

(Wang et al., 31 Jan 2026). The paper uses this equivalence as a certificate of sufficiency, and derives a simulation-free loss that jointly trains the velocity field and the representation in amortized Markov blanket learning and self-supervised representation learning (Wang et al., 31 Jan 2026).

This usage is unusually sharp terminologically. The paper explicitly names the midpoint equivalence the “zero-flow criterion,” whereas in several other literatures the same phrase is a retrospective interpretation. It also highlights a limitation: the theorem is pointwise in $k$02, but the practical loss only enforces vanishing at sampled $k$03, and the authors note that a more rigorous justification is needed (Wang et al., 31 Jan 2026).

Across these domains, the unifying theme is precise but abstract: a vanishing condition on a flow, dissipation, boundary flux, tangency set, or transport field is used to certify a hidden property that is otherwise difficult to observe directly. In signed graphs, that property is nowhere-zero flow existence or bounded flow number; in fluid mechanics, inviscid convergence; in dynamical systems, nontrivial cohomology of an invariant set; in thermodynamics, recovery of the classical evolution inequality; in power systems, terminal no-flow conditions under shunt-aware modeling; and in representation learning, equality of endpoint distributions or conditional sufficiency (Kaiser et al., 2016, Sueur, 2017, Sánchez-Gabites, 2024, Hochberg et al., 2024, Zhou et al., 2020, Wang et al., 31 Jan 2026).