Group of Flows in RG, Graphs, and Networks

• Group of flows is defined as organized collections of flows that encapsulate fixed points, phase boundaries, or algebraic constraints across RG, graph theory, and network analyses.
• The topic highlights practical methodologies like coarse-graining in RG, finite-element discretization in functional RG, and hypothesis-driven aggregation in forensic network studies.
• Different applications use distinct frameworks—from computational coupling to entropic ordering—to extract actionable insights from complex systems.

The expression “group of flows” does not denote a single universal construction. In the literature considered here, it names several distinct objects: the family of renormalisation-group trajectories generated by iterating a coarse-graining map; truncated subsystems of coupled functional-renormalisation equations; abelian-group-valued flows and group connectivity in graphs; and higher-level aggregates of network flow records called superflows (Watson et al., 2021, Sattler et al., 2024, DeVos et al., 2020, Collins et al., 2024). What these uses share is a passage from an isolated flow to an organized collection whose structure carries physical, algebraic, computational, or forensic meaning.

1. Terminological scope

In renormalisation-group language, a flow is the sequence

$H \mapsto R(H) \mapsto R^{(2)}(H) \mapsto \cdots$

obtained by repeated coarse-graining. In that setting, the “group of flows” is not a mathematical group in the algebraic sense, but the family of RG trajectories generated on a parameter space of initial Hamiltonians. In functional RG numerics, a “group” of flows can mean a truncated coupled subsystem extracted from the infinite hierarchy

$$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$

In graph theory, the term shifts meaning entirely: a $\Gamma$-flow is a flow with values in an additive abelian group $\Gamma$, and $\Gamma$-connectedness requires avoiding any prescribed edge-labeling. In network forensics, a superflow is an aggregation of one or more ordinary flows based on an analyst-specific hypothesis (Watson et al., 2021, Sattler et al., 2024, DeVos et al., 2020, Collins et al., 2024).

Domain	Meaning of the “group” structure	Representative formalism
RG theory	Family of trajectories on Hamiltonian/coupling space	$H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$
FRG numerics	Truncated coupled subsystem of the flow hierarchy	$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$
Graph theory	Flows valued in an abelian group $\Gamma$	$\phi:E\to\Gamma$
Network forensics	Hypothesis-driven aggregation of many flow records	$h:2^F\to\{\mathrm{true},\mathrm{false}\}$

This suggests that the phrase is best understood relationally rather than ontologically. In each case, the central issue is not merely the existence of one flow, but the organization of many flows by fixed points, algebraic constraints, computational couplings, or semantic hypotheses.

2. Renormalisation-group families of trajectories

The RG meaning is the most explicit use of the phrase. A renormalisation-group map acts on local interactions as

$$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$0

with the desired properties that $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$1 is computable, preserves gappedness and gaplessness, preserves phase indicators such as order-parameter non-analyticity, and keeps the renormalised terms within the same operator family with recursively transformed couplings. Fixed points are Hamiltonians $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$2 left invariant by the map, while deviations are tracked by couplings $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$3 (Watson et al., 2021).

In the standard RG picture, Hamiltonians that flow to the same fixed point are in the same phase. Fixed points represent scale-invariant long-distance physics, and universality arises because many microscopically different systems flow to the same macroscopic fixed point. The “group of flows” is therefore organized into basins of attraction, with the basin structure encoding phase structure rather than algebraic closure (Watson et al., 2021).

A rigorous realization of this viewpoint is given by the block-renormalisation-group construction for the undecidable 2D Hamiltonian of Cubitt, Pérez-García, and Wolf. The map is built from $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$4 blocking and truncation, implemented by isometries, for example

$$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$5

The construction separates a classical tiling layer, a quantum computation layer, and a dense-spectrum auxiliary layer, permitting exact coarse-graining of the different sectors before recombination (Watson et al., 2021).

The importance of this formulation is that it treats the family of RG trajectories itself as a physically meaningful object. Phase boundaries, fixed-point selection, and the geometry of basins become properties of the entire family, not merely of individual trajectories.

3. Complexity, chaos, and uncomputable RG behaviour

The same RG framework also supports a sharp distinction between ordinary, chaotic, and uncomputable flows. Ordinary flows move toward a finite number of fixed points and are simple enough to analyze in terms of relevant, irrelevant, or marginal couplings. Chaotic flows exhibit exponential sensitivity to initial conditions, but exact initial data still determine the long-time behaviour in principle. Uncomputable flows are stronger: trajectories can remain arbitrarily close for an uncomputably large number of iterations and then diverge abruptly to different fixed points in different phases (Watson et al., 2021).

The central theorem of the undecidable construction is that each RG step is computable, the RG map is explicitly defined, and the map converges to the correct fixed points, yet the total RG trajectory is uncomputable. In particular, for every finite $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$6, the renormalised couplings are computable, but the overall flow, its basins of attraction, and the iteration at which divergence occurs are uncomputable. The flow thereby lifts undecidability from the spectral-gap problem into the coarse-graining dynamics itself (Watson et al., 2021).

This behaviour is made quantitative through the parameter $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$7 in the renormalised local terms. If the universal Turing machine does not halt, then for all sufficiently large $$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$8,

$$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\Gamma_k^{(2)},\dots,\Gamma_k^{(n+2)},\dot\Phi).$$9

If it does halt, then there is an uncomputable iteration $\Gamma$0 after which

$\Gamma$1

The fixed-point structure is nevertheless simple: the flow ends in either a gapped Ising-like fixed point or a gapless XY-like fixed point (Watson et al., 2021).

This establishes a notion of unpredictability stronger than chaos. In chaos, the obstacle is sensitivity to data; here the obstacle is the halting problem. A plausible implication is that RG theory can encode limits of predictability that are not exhausted by dynamical instability.

4. Monotonicity and irreversibility in quantum-field-theoretic flows

A different sense in which families of flows are organized appears in information-theoretic and defect-theoretic monotones. Quantum Rényi relative entropies define a one-parameter family of distances between density matrices along an RG trajectory. They are monotone in $\Gamma$2, monotone under inclusion of regions, and satisfy

$\Gamma$3

For boundary RG flows, the infrared limit obeys

$\Gamma$4

so every consistent trajectory between the two boundary fixed points must satisfy an endpoint-controlled bound (Casini et al., 2018).

For line defects in arbitrary spacetime dimension $\Gamma$5, perturbing a defect CFT by relevant defect operators produces an irreversible defect RG flow with a canonical decreasing entropy function

$\Gamma$6

The corresponding gradient formula,

$\Gamma$7

implies $\Gamma$8 and generalizes the Affleck–Ludwig $\Gamma$9-theorem to line defects in arbitrary dimension (Cuomo et al., 2021).

These results organize a group of flows by monotone quantities rather than by mere reachability of fixed points. The left-hand side is trajectory-sensitive, while the right-hand side is controlled by endpoint data. This gives RG families a thermodynamic or entropic order structure: not every interpolation between fixed points is admissible, and those that are admissible satisfy quantitative inequalities.

5. Holographic flow networks and geometric flow structures

Holographic constructions make the family structure of flows especially explicit. For Wilson loops in $\Gamma$0 SYM, the locally BPS loop and the pure gauge Wilson loop are the endpoints of an RG flow induced by worldsheet boundary perturbations. The interpolating family is

$\Gamma$1

with $\Gamma$2 the ordinary Wilson loop and $\Gamma$3 the BPS loop. At weak coupling,

$\Gamma$4

so $\Gamma$5 is the UV attractor and $\Gamma$6 are IR attractors. At strong coupling, general boundary potentials $\Gamma$7 generate a richer hierarchy involving Neumann, Dirichlet, mixed, partially Dirichlet, and kinked loop configurations (Polchinski et al., 2011).

A more geometric holographic flow appears for M5-branes wrapping a calibrated Kähler four-cycle. There, the Kähler metric itself is allowed to evolve with the holographic coordinate $\Gamma$8, with arbitrary Kähler UV data imposed by the topological twist. The flow equations derived in seven-dimensional gauged supergravity wash out much of the initial metric data, and at the IR fixed point, where $\Gamma$9, the metric becomes Kähler–Einstein: $\Gamma$0 This is interpreted as a Kähler uniformization flow rather than an ordinary Ricci flow (Fluder, 2017).

In gravity-oriented FRG, area-metric gravity supplies another example of a structured family of flows. The theory contains ordinary length-metric degrees of freedom and additional shape-mismatching degrees of freedom in left-handed and right-handed sectors. The mass beta functions show that these masses are in general even more relevant than dictated by canonical scaling, which generically drives them large compared to the Planck mass and supports decoupling of the non-metric sector. The same analysis finds that parity symmetry does not emerge under the RG flow, and that the beta function of the Immirzi parameter has zeros at vanishing and at infinite Immirzi parameter (Borissova et al., 2 Jul 2025).

Geometric analysis supplies a further, distinct relation between groups and flows. For non-collapsed ancient Ricci flows with complete time-slices and bounded curvature on compact time-intervals, the manifold has finite fundamental group, and $\Gamma$1 is a quotient of the fundamental group of the regular part of any tangent flow at infinity. Here the flow does not form a group; rather, the existence and asymptotics of the flow constrain an associated topological group (Bamler, 2021).

6. Computational grouping of functional RG equations

In functional RG practice, a “group of flows” often means a numerically tractable subsystem cut from the infinite hierarchy of flow equations. DiFfRG formalizes this viewpoint through the generalized flow

$\Gamma$2

from which one derives the tower

$\Gamma$3

The framework organizes these into field-dependent FE functions, momentum-dependent variables, and EoM-based extractors, permitting mixed systems in which field-space PDEs and momentum-space ODE-like vertex flows coexist (Sattler et al., 2024).

The discretisation strategy is explicitly PDE/ODE based. Field-dependent flows are written in weak form and solved with finite-element methods, with continuous Galerkin, discontinuous Galerkin, direct DG, and local DG variants. Time stepping is likewise stratified: implicit solvers such as SUNDIALS IDA for stiff PDE-like flows, explicit solvers such as Boost ABM, RK45, and RK78 for large momentum-vertex systems, and ImEx schemes for mixed problems. Automatic Jacobians are generated by forward automatic differentiation via autodiff (Sattler et al., 2024).

Stability of the resulting flows depends strongly on regulator choice. The Principle of Strongest Singularity orders comparable regulators by the strength of the regulator-induced singularity in the threshold integrals near the infrared singular point. In the local potential approximation, after normalizing $\Gamma$4, the order relation reduces to

$\Gamma$5

Within the standard class $\Gamma$6, the Litim regulator is the unique optimal regulator in LPA, and regulators with an infinite contact set can reach scales roughly $\Gamma$7–$\Gamma$8 times smaller than regulators with only one contact point (Zorbach et al., 2024).

The computational significance is that the “group” of flows is not merely a bookkeeping device. It determines what can be discretized together, how stiffness propagates between sectors, and how infrared singularities are stabilized in practice.

7. Algebraic and forensic groupings of flows

Outside RG theory, the phrase acquires two sharply different meanings. In graph theory, let $\Gamma$9 be an oriented graph and $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$0 an additive abelian group. A function $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$1 is a $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$2-flow if it satisfies Kirchhoff’s law at every vertex,

$H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$3

The graph is $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$4-connected if for every $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$5 there exists a $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$6-flow $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$7 such that $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$8 for every edge. For oriented 3-edge-connected graphs with $H\mapsto R(H)\mapsto R^{(2)}(H)\mapsto\cdots$9, the number of such avoiding flows is at least

$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$0

for abelian groups $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$1 of size $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$2; in particular, for $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$3 this is exponential in $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$4. For $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$5 and $k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$6, there are at least

$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$7

avoiding flows (DeVos et al., 2020).

In forensic network analysis, by contrast, a superflow is an aggregation of one or more flows based on an analyst-specific hypothesis about traffic behavior. A hypothesis is formalized as

$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$8

and a superflow decomposition partitions

$k\partial_k \Gamma_k^{(n)}=\mathrm{Flow}(\cdots)$9

with each $\Gamma$0 satisfying $\Gamma$1. Efficiently monitorable and subset-closed hypotheses admit greedy maximal decompositions in linear time with memory proportional to the number of reported superflows. The case studies are explicitly operational: a CNN homepage fetch in private mode contacted 36 different sites, consisted of 228 flows, and flowed to 36 IP addresses; the proposed webpage superflow representation reduced the footprint from $\Gamma$2 bytes to $\Gamma$3 bytes. A scan-256 superflow substitutes for 256 flows and compresses the footprint from roughly 8 KB to 32 bytes, while allotted scan-256 reduced darkspace footprint by about $\Gamma$4 to $\Gamma$5 (Collins et al., 2024).

The contrast is instructive. In group-connectivity theory, the group is algebraic and the problem is existence and multiplicity of avoiding assignments. In superflow analysis, the grouping is semantic and hypothesis-driven, and the problem is compact, maximally informative decomposition. Both cases replace a single flow object by a structured family, but the governing structures are entirely different: one is abelian-group algebra, the other analyst-defined relational logic.

Across these literatures, “group of flows” is therefore best regarded as a family resemblance term. It can denote trajectories organized by fixed points, flows valued in a group, coupled computational subsystems of FRG equations, or higher-level aggregates of ordinary network flows. The unifying pattern is that the scientifically meaningful object is not an isolated flow, but the architecture relating many flows to one another.