Fixed-Point Flows Overview

• Fixed-point flows are dynamic constructions where system behavior is anchored by fixed points, fixed-point equations, or trajectories connecting distinct configurations.
• They are applied in fields such as classical dynamics, network calculus, and separation logic to analyze entropy, compute worst-case delays, and establish rigorous flow invariance.
• Fixed-point flow frameworks leverage fixed-point equations to solve reverse-time dynamics in generative models, trace RG trajectories, and underpin algorithmic fixed-point iterations.

In the papers considered here, “fixed-point flows” denotes several related constructions rather than a single standardized object. The expression is used for continuous flows whose dynamics are constrained by the presence or absence of fixed points, for reverse-time procedures that solve a fixed-point equation at each discrete step, for RG or kinetic trajectories joining distinguished fixed points, and for abstract frameworks in which a flow or a propagated state is itself defined as a fixed point of an operator. Across these uses, the common structure is a flow, action, or iterative dynamics whose behavior is organized by fixed points, fixed-point equations, or trajectories connecting distinguished fixed configurations.

1. Principal meanings and common structure

A first major meaning comes from classical dynamics, where a flow is a continuous action of $\mathbb{R}$ on a compact space and “fixed-point free” means that no orbit is a single point. In this setting, fixed points, or their absence, govern entropy, centralizers, and lower bounds for periodic flows (Liang et al., 2021, Han et al., 2023).

A second meaning is algorithmic. In time-sensitive networks with cyclic dependencies, Fixed-Point TFA formulates worst-case flow analysis as a fixed-point problem on burstiness and delay variables, while cut-free variants iterate the same underlying monotone map directly (Plassart et al., 2021). In separation logic, a flow graph is assigned its least flow by a least fixed point equation, and this least-fixed-point semantics is then used to define transfer functions, composition, and footprints (Meyer et al., 2023). In rough analysis, a Banach fixed point construction for rough differential equations is compared directly with Bailleul’s flow approach, and the relevant bridge is an algebraic cocycle condition on the underlying Hopf algebra (Bruned et al., 19 Feb 2026).

A third meaning appears in generative modeling. For rectified flows, inversion can be written as solving, at each time step, a fixed-point equation induced by the forward solver; the resulting reverse dynamics is therefore a fixed-point flow in the sense that reverse-time evolution is implemented by fixed-point solves rather than by directly integrating a reverse ODE (Kim et al., 16 Jun 2026). In flow-based language modeling, self-conditioning is interpreted as a fixed-point iteration of the denoiser, and this leads to a two-dimensional class of self-conditioned flows called fixed-point flows, where one dimension is the flow process and the other is the fixed-point iteration (Yoo et al., 1 Jul 2026).

A fourth meaning is structural rather than algorithmic: fixed points organize nonequilibrium and RG dynamics. In Bjorken flow, the attractor is identified as the solution joining an early-time collisionless fixed point to a late-time hydrodynamic fixed point (Blaizot et al., 2021). In the boson Boltzmann equation, thermal and nonthermal power-law fixed points organize RG-like flow diagrams in a low-dimensional parameter space (Fukushima et al., 2017). In large-$N$ Chern-Simons–matter theories, a marginal fixed line is lifted at order $1/N$ to discrete fixed points connected by RG flows (Aharony et al., 2018). In scalar theories with kinetic terms $\phi \Box^k \phi$, non-perturbative RG trajectories connect different Gaussian fixed points (Buccio et al., 2022).

2. Fixed-point-free continuous flows and centralizer structure

For a compact metric flow $(M,\phi)$, the fixed-point-free condition excludes points $x$ with $\phi_t(x)=x$ for all $t\in\mathbb{R}$. In the literature considered here, this condition is not merely technical. For packing entropy of flows, it is what makes time reparametrization meaningful without degenerating into arbitrary time stretching at a stationary orbit (Liang et al., 2021). For centralizers of separating flows, it is what allows orbitwise time changes to be uniquely controlled on small time intervals (Han et al., 2023).

A separating flow is one for which there exists $\delta>0$ such that whenever $d(\phi_t(x),\phi_t(y))<\delta$ for all $N$0, the point $N$1 must lie on the $N$2-orbit of $N$3. Under this hypothesis, together with the absence of fixed points, the centralizer becomes quasi-trivial: if a continuous flow $N$4 commutes with $N$5, then there exists a continuous function $N$6, invariant along $N$7-orbits, such that

$N$8

Thus every commuting flow is an orbitwise time change of the original one (Han et al., 2023). The same pattern extends to separating $N$9 $1/N$0-actions on compact Riemannian manifolds without boundary: any commuting $1/N$1-action is given by an orbit-invariant matrix field $1/N$2 through

$1/N$3

This identifies fixed-point-free separating flows as a regime of strong centralizer rigidity rather than one of symmetry proliferation (Han et al., 2023).

3. Entropy and orbit complexity in fixed-point-free flows

For a compact metric flow without fixed points, Liang and Lei define packing topological entropy on subsets using reparametrization balls

$1/N$4

These are continuous-time analogues of Bowen balls, but with a time change $1/N$5 built in, so the notion is adapted to orbit equivalence rather than only to the time-$1/N$6 map (Liang et al., 2021).

The resulting entropy theory is a packing version of subset entropy for flows. For a non-empty compact subset $1/N$7, the main theorem is the variational principle

$1/N$8

where $1/N$9 is the upper local entropy defined through the decay of reparametrization-ball measures (Liang et al., 2021). This is the continuous-time analogue of the Feng–Huang packing entropy formula for maps, but with the fixed-point-free assumption playing an essential role in the control of time reparametrizations.

The absence of fixed points enters explicitly through a lemma stating that if two orbits remain $\phi \Box^k \phi$0-close under a reparametrization $\phi \Box^k \phi$1, then $\phi \Box^k \phi$2 must remain close to the identity: $\phi \Box^k \phi$3 This estimate underlies the adapted $\phi \Box^k \phi$4-covering lemma for reparametrization balls and hence both directions of the packing-entropy variational principle (Liang et al., 2021). A plausible implication is that, for fixed-point-free flows, local entropy and orbit geometry remain tightly coupled even when orbit comparison is allowed to vary speed continuously.

4. Fixed-point theorems and lower bounds for actions

In symplectic and almost complex geometry, fixed points of periodic flows appear as equilibrium points of circle actions. For symplectic periodic flows generated by symplectic $\phi \Box^k \phi$5-actions, if the Chern class map is somewhere injective, then the action has at least

$\phi \Box^k \phi$6

fixed points (Pelayo et al., 2010). Without additional assumptions, there is no symplectic periodic flow with exactly one equilibrium point on a compact symplectic manifold, and there is no such flow with exactly two equilibrium points when $\phi \Box^k \phi$7 (Pelayo et al., 2010).

A complementary lower-bound theory is developed for almost complex manifolds under the vanishing condition $\phi \Box^k \phi$8. There, the number of isolated fixed points of a periodic flow is bounded below by an explicitly computed arithmetic function $\phi \Box^k \phi$9, and the proof combines equivariant $(M,\phi)$0-theory with number-theoretic results on sums of squares and triangular numbers (Godinho et al., 2014). These bounds confirm many cases of Kosniowski’s conjecture and apply in particular to manifolds with torsion first Chern class, including symplectic Calabi–Yau manifolds (Godinho et al., 2014). The same framework yields divisibility properties for the number of fixed points, improving related statements of Hirzebruch (Godinho et al., 2014).

A different fixed-point theorem arises in topological dynamics of automorphism groups. If $(M,\phi)$1 for a Fraïssé structure $(M,\phi)$2, and $(M,\phi)$3 has the free joint embedding property, then every null $(M,\phi)$4-flow has a fixed point. If $(M,\phi)$5 has the free amalgamation property, then every tame $(M,\phi)$6-flow has a fixed point (Codenotti, 16 Jul 2025). Here the mechanism is a generalized Kechris–Pestov–Todorčević correspondence due to Nguyen Van Thé, specialized to the algebras of null and tame functions.

In operator algebras, fixed-point constructions also generate flows. For certain fixed point algebras arising from product type actions on ITPFI factors, the associated flow of weights can be computed explicitly and shown to be approximately transitive; under those conditions, the corresponding fixed point factors are themselves ITPFI (Munteanu, 2018). In this sense, the fixed-point algebra is controlled not only by the action but by the measurable flow attached to its weights.

5. Fixed-point equations as flow computation

In network calculus, cyclic dependencies in time-sensitive networks obstruct the one-pass Total Flow Analysis available in acyclic graphs. Fixed-Point TFA resolves this by cutting the graph to obtain a cycle-free network, analyzing the cut network, and iterating a map

$(M,\phi)$7

on cut-link burstiness bounds (Plassart et al., 2021). The central result is that the choice of cut does not affect convergence or the final worst-case delay and backlog bounds: the cut-based formulation and a cut-free formulation based on the global map

$(M,\phi)$8

either both have a finite fixed point or both do not, and when they do, they yield the same unique bounds (Plassart et al., 2021). SyncTFA, AsyncTFA, and AltTFA are therefore different realizations of the same fixed-point flow computation (Plassart et al., 2021).

In separation logic for graphs, the term “flow” is defined directly by least fixed point theory. A flow graph $(M,\phi)$9 has a least flow $x$0 satisfying

$x$1

and this least-fixed-point semantics is the basis for transfer functions, graph composition, and footprints (Meyer et al., 2023). The framework is explicitly “based on standard fixed point theory, guarantees least flows, rules out vanishing flows, and has an easy to understand notion of footprint as needed for soundness of the frame rule” (Meyer et al., 2023). The same paper defines transfer functions via least fixed points and then uses a further fixed-point computation to infer footprints automatically (Meyer et al., 2023).

In rough analysis, the relation between fixed points and flows is algebraic rather than order-theoretic. Under a Hopf-algebraic assumption requiring degree-$x$2 cocycles and a tree-like basis, the algebraic assumptions needed for a Banach fixed point argument imply the algebraic assumptions needed for Bailleul’s flow approach (Bruned et al., 19 Feb 2026). The obstruction is sharp: the Hopf algebra of multi-indices does not satisfy the cocycle condition, which gives “a rigorous result on the impossibility, observed in practice, of performing a fixed point argument for multi-indices rough paths and multi-indices in Regularity Structures” (Bruned et al., 19 Feb 2026). This sharply separates settings where fixed-point and flow methods coexist from settings where only flow methods are available.

6. Generative models and fixed-point flows

In rectified flows, generation is defined by the probability-flow ODE

$x$3

but inversion is posed discretely: given $x$4, one seeks $x$5 such that the forward solver reproduces $x$6. This leads to the residual equation

$x$7

and the associated fixed-point map

$x$8

A fixed-point inversion method therefore constructs the reverse trajectory by solving a fixed-point problem at each step (Kim et al., 16 Jun 2026).

The paper “Root-Selecting Fixed-Point Inversion for Rectified Flows via Trajectory Straightness” observes that this per-step equation can have multiple fixed-point solutions in practice, and that different selections lead to substantially different inversion trajectories, reconstruction quality, and editing quality (Kim et al., 16 Jun 2026). For rectified flows, trajectory straightness is proposed as a principled root-selection criterion. The resulting method, SelFix, uses Halpern iteration and a straightness anchor to converge, under local nonexpansiveness assumptions, to the straightness-minimizing exact inverse root (Kim et al., 16 Jun 2026). In this setting, a fixed-point flow is reverse-time evolution implemented by a sequence of local fixed-point solves, with a geometric selection principle over the fixed-point set.

A related but distinct construction appears in flow-based language modeling. Self-conditioning augments the denoiser by feeding it its own current estimate, and this is interpreted as solving a fixed-point iteration that bootstraps denoising performance (Yoo et al., 1 Jul 2026). The paper formulates fixed-point flows as “a two-dimensional class of self-conditioned flows, where the first dimension represents the flow process and the second represents the fixed-point iteration,” proves that they define valid flow maps, and distills them into a flow map LLM, FMLM$x$9, by compressing both the fixed-point iteration and the flow process (Yoo et al., 1 Jul 2026). The resulting model outperforms state-of-the-art self-conditioned models and few-step models in one- and few-step generation on OpenWebText (Yoo et al., 1 Jul 2026). Here fixed-point flow is neither a static fixed point nor a continuous-time attractor, but a denoising flow explicitly indexed by an inner fixed-point iteration.

7. Fixed points as organizing structures in kinetic and RG flows

In nonequilibrium kinetic theory, fixed points often organize the entire trajectory rather than merely serving as endpoints. For Bjorken flow in relaxation time approximation, a two-mode truncation of the kinetic hierarchy can be reduced to a single nonlinear equation for a dimensionless variable $\phi_t(x)=x$0, and the attractor is identified as “the particular solution of this non linear equation that joins two fixed points: one corresponding to the collisionless, early time regime, the other corresponding to late time hydrodynamics” (Blaizot et al., 2021). The paper further argues that extending hydrodynamics to early time amounts essentially to improving the accuracy of the location of the collisionless fixed point (Blaizot et al., 2021).

In the boson Boltzmann equation, fixed points include the Bose–Einstein thermal distribution, the Rayleigh–Jeans and Maxwell–Boltzmann limits in dense and dilute truncations, Kolmogorov–Zakharov power-law spectra, and self-similar solutions (Fukushima et al., 2017). A two-parameter ansatz $\phi_t(x)=x$1 induces RG-like flow diagrams in the $\phi_t(x)=x$2 plane, with fixed points, critical lines, attractive and repulsive directions, and basins of attraction (Fukushima et al., 2017). The KZ-I and KZ-II solutions organize particle and energy cascades, while the thermal line is globally attractive in the full theory (Fukushima et al., 2017).

In large-$\phi_t(x)=x$3 Chern-Simons–matter theories, the quasi-bosonic theories form fixed lines at strict $\phi_t(x)=x$4, parameterized by a sextic coupling $\phi_t(x)=x$5, but at first subleading order in $\phi_t(x)=x$6 the beta function becomes a cubic polynomial in $\phi_t(x)=x$7 (Aharony et al., 2018). The paper conjectures that this produces three fixed points for $\phi_t(x)=x$8 at every non-zero value of the ’t Hooft coupling, leading to three distinct regular bosonic and three dual critical fermionic conformal fixed points (Aharony et al., 2018). The resulting RG structure combines slow flows along the $\phi_t(x)=x$9 direction with fast flows descending from supersymmetric $t\in\mathbb{R}$0 theories (Aharony et al., 2018).

A still more literal version of “flows between fixed points” appears in scalar theories with kinetic terms $t\in\mathbb{R}$1. Using the non-perturbative RG, one finds trajectories between different Gaussian fixed points, with the anomalous dimension changing continuously so that the field has the correct dimensions at the endpoints of the respective free theories (Buccio et al., 2022). These theories are asymptotically free both in the infrared and in the ultraviolet, and they illustrate “the fact that a diverging coupling can actually correspond to a free theory” (Buccio et al., 2022). In that sense, fixed-point flows are RG trajectories on theory space connecting distinct free theories that are perturbatively unrelated in a single local chart.

Taken together, these lines of work show that fixed-point flows are best understood as a family of structurally related ideas. In some settings, the flow is constrained by the existence or nonexistence of fixed points; in others, the flow is computed by a fixed-point iteration; in still others, the flow is the trajectory joining distinguished fixed configurations. What unifies them is not a single definition but a recurrent architecture: fixed points serve as algebraic anchors, geometric constraints, variational descriptors, or asymptotic regimes that determine the global organization of the dynamics.