Parameter Flows in Dynamics and Inference

• Parameter flows are parameter-dependent families of transformations that index dynamics, spectral geometry, or posterior transport via external parameters.
• They span diverse applications from smooth dynamical systems and bifurcation analysis to flow-based inference and fluid simulation, underlining practical computational strategies.
• This unified framework highlights coherent parameter dependency and structural invariants, bridging disparate mathematical constructions and simulation methods.

Searching arXiv for papers that use “parameter flows” or closely related terminology across dynamics, matrix flows, and flow-based inference. Parameter flows are parameter-dependent families of flows or flow-based transports in which either a dynamical system is indexed by one or more external parameters, or an invertible transformation is used to represent parameter dependence in inference, density estimation, or simulation. Current arXiv usage suggests that the term spans several non-equivalent constructions: cohomological rigidity of smooth flows, one-parameter matrix families and their eigencurves, bifurcations under parameter variation, conditional and continuous normalizing flows for posterior transport, factorized density morphing under nuisance parameters, and parameter-conditioned simulators for fluid or population dynamics (Matsumoto, 2010, Lagunowich et al., 11 Jan 2026, Valsecchi et al., 29 Jun 2026, Schwerdtner et al., 25 May 2026).

1. Scope and principal meanings

The common structure is a map that depends coherently on an external parameter, but the mathematical object that “flows” differs substantially by field. In smooth dynamics, the parameter acts on time reparametrizations or on a family of vector fields. In matrix analysis, the parameter indexes a family $A(t)$ whose spectrum and invariant subspaces evolve. In simulation-based inference, the flow is an invertible transport from a simple base distribution to a posterior or likelihood-related density over parameters. In density-morphing and population-dynamics settings, the parameter may index nuisance configurations, physical time, or both (2002.0188, Lagunowich et al., 11 Jan 2026, Schwerdtner et al., 25 May 2026, Fresca et al., 2021).

Domain	Flow object	Representative formulation
Smooth dynamics	Parameterized vector-field flow	$f = X(g)+c$
Matrix theory	One-parameter matrix family	$A(t)\in \mathbb{C}^{n\times n}$
Flow-based inference	Conditional density transport	$p(\theta\mid x)$ or $p(\bm{x},\theta\mid \text{context})$
Density morphing	Parameter-dependent transport map	$T_\nu$ or $\Phi(a,t,s)$

This breadth matters because many papers use the same phrase for objects with very different invariants. In one tradition, the decisive questions are ergodicity, invariant forms, and conjugacy classes. In another, they are Jacobians, tractable likelihoods, and posterior sampling speed. A technically accurate account therefore has to treat “parameter flows” as a family of related usages rather than a single settled construction.

2. Smooth-dynamical and bifurcation-theoretic meanings

In smooth dynamics, the strongest usage is parameter rigidity. For a nonsingular smooth flow generated by a vector field $X$ on a closed manifold $M$, parameter rigidity means that for every smooth real-valued function $f$ there exist a smooth function $f = X(g)+c$0 and a constant $f = X(g)+c$1 such that

$f = X(g)+c$2

Matsumoto proved that a parameter rigid flow on a closed orientable $f = X(g)+c$3-manifold is smoothly conjugate to a linear flow on the $f = X(g)+c$4-torus with badly approximable slope (Matsumoto, 2010). The same work derives several structural consequences: unique ergodicity, existence of an invariant smooth volume form, minimality, and an isomorphism between closed normal $f = X(g)+c$5-forms and $f = X(g)+c$6. The classification is sharp in the stated category but explicitly does not address nonorientable $f = X(g)+c$7-manifolds, and it does not prove higher-dimensional analogues (Matsumoto, 2010).

A different one-parameter usage appears in homogeneous dynamics. For one-parameter unipotent flows on

$f = X(g)+c$8

Kelmer and Mohammadi proved the logarithm law

$f = X(g)+c$9

for $A(t)\in \mathbb{C}^{n\times n}$0-almost every $A(t)\in \mathbb{C}^{n\times n}$1, after normalizing the metric so that cusp neighborhoods satisfy $A(t)\in \mathbb{C}^{n\times n}$2 (Kelmer et al., 2011). In the arithmetic case they further obtain a full Borel–Cantelli statement for the family of cusp neighborhoods. Here the “parameter” is the time variable of the unipotent subgroup, and the central observable is the rate of cusp excursion rather than transport of densities.

Bifurcation theory supplies a third meaning. For typical one-parameter families of gradient flows on a $A(t)\in \mathbb{C}^{n\times n}$3-sphere with holes, with at most six singular points, the only codimension-one bifurcations are saddle-node and saddle connection, together with their boundary analogues, and the relevant topological type is encoded by separatrix diagrams (Bilun et al., 2023). The same article proves that bifurcation points are isolated in typical one-parametric families of flows without closed trajectories and with fixed points on the boundary. Parameter variation is therefore represented combinatorially as adjacency between Morse flow types.

The rimming-flow literature extends this perspective to multi-parameter PDE bifurcation. For the thin-film model

$A(t)\in \mathbb{C}^{n\times n}$4

the steady state $A(t)\in \mathbb{C}^{n\times n}$5 depends real analytically on the gravity and hydrostatic-pressure parameters, and the corresponding Fréchet derivative has a conjugate pair of eigenvalues whose real part satisfies

$A(t)\in \mathbb{C}^{n\times n}$6

This yields a Hopf threshold curve

$A(t)\in \mathbb{C}^{n\times n}$7

separating exponentially stable from unstable steady states and generating periodic solutions along transverse parameter paths (Karabash et al., 2024).

3. Spectral and matrix-flow formulations

In matrix analysis, a parameter flow is a family

$A(t)\in \mathbb{C}^{n\times n}$8

together with the geometry of its eigenvalue curves and the question of whether all matrices in the family admit one common block decomposition (2002.0188). The relevant structural property is uniform decomposability: $A(t)\in \mathbb{C}^{n\times n}$9 with one fixed similarity $p(\theta\mid x)$0. In the Hermitian case the similarity is unitary. Uhlig’s main corrective point is that the classical Hund–von Neumann–Wigner no-crossing rule is valid only for indecomposable Hermitian flows; decomposable Hermitian flows can exhibit genuine eigencurve crossings because their spectra are superpositions of independent block spectra (2002.0188).

This distinction between true crossings and avoided crossings is operationalized geometrically. For indecomposable Hermitian families, eigencurves do not intersect and near-collisions appear as hyperbolic veering. For decomposable families, crossings arise because branches belong to different invariant blocks. The non-Hermitian case is much harder because eigencurves live in $p(\theta\mid x)$1-space, where intersections are rare even for decomposable flows. The paper therefore formulates, but does not prove, a generalized Hund–von Neumann–Wigner theorem for general complex or real flows (2002.0188).

A complementary line of work studies how to compute the coarsest simultaneous block structure directly from sampled matrices. The central idea is to diagonalize one flow matrix $p(\theta\mid x)$2, transform another matrix $p(\theta\mid x)$3 into that eigenbasis, threshold numerically small entries, and recover common invariant subspaces from the resulting logical $p(\theta\mid x)$4-$p(\theta\mid x)$5 pattern matrix. In the Hermitian case this uses a unitary eigenbasis; in the general case it uses a nonsingular eigenvector matrix when the sampled matrix is diagonalizable (Uhlig, 2020). The same paper extends the method to static matrices by introducing the auxiliary Hermitian flow

$p(\theta\mid x)$6

It also introduces $p(\theta\mid x)$7-normal classes, meaning matrices that are unitarily reducible to block-diagonal form with maximal repeated block size $p(\theta\mid x)$8 (Uhlig, 2020).

Across both papers, matrix-flow analysis treats the parameter as indexing global spectral geometry rather than local perturbation only. The decisive objects are invariant subspaces, eigencurve topology, and parameter-independent similarities.

4. Flow-based parameter inference and posterior transport

A major contemporary meaning of parameter flows is normalizing-flow-based parameter estimation. In this usage, the flow is an invertible map from a simple base distribution to a parameter posterior, a likelihood-ratio-related density, or a joint state–parameter law. The attraction is exact change-of-variables evaluation, flexible non-Gaussian modeling, and very fast amortized sampling.

For nonlinear dynamical systems, conditional normalizing flows have been used for joint state and parameter inference. The central formulation is conditional density estimation of either latent states alone or joint state–parameter variables, with explicit attention to forward estimation, backward estimation, time inversion, chained predictions, and rollout under non-Gaussian and multimodal uncertainty (Lagunowich et al., 11 Jan 2026). The key move is to extend the target variable so that the learned conditional density covers unknown system parameters together with latent states, rather than treating parameters as fixed nuisance quantities.

In simulation-based particle-physics inference, Contrastive Normalizing Flows are used as a front end for uncertainty-aware parameter estimation under nuisance-induced distribution shift. The specific target is a physical parameter proportional to the signal fraction in a signal-plus-background mixture,

$p(\theta\mid x)$9

with the Higgs signal strength $p(\bm{x},\theta\mid \text{context})$0 as the parameter of interest. The method combines classifier-based likelihood-ratio estimation, a contrastive normalizing flow, and downstream frequentist inference with a profile/binned likelihood and a Neyman construction (Elsharkawy et al., 13 May 2025).

Gravitational-wave astronomy has adopted several closely related flow formulations. Autoregressive conditional flows were first used to learn $p(\bm{x},\theta\mid \text{context})$1 directly from strain data, with a change-of-variables model

$p(\bm{x},\theta\mid \text{context})$2

and a stronger latent-variable variant, CVAE+, used inverse autoregressive and masked autoregressive flows in encoder, prior, and decoder; the paper reports that sampling requires less than two seconds to draw $p(\bm{x},\theta\mid \text{context})$3 posterior samples (2002.01274). Continuous normalizing flows trained by flow matching were then applied to $p(\bm{x},\theta\mid \text{context})$4-dimensional massive-black-hole-binary inference, including a detector-response symmetry transformation and a reported speed of roughly $p(\bm{x},\theta\mid \text{context})$5 per sample relative to the nested-sampling table shown in the paper (Liang et al., 2024). An image-based estimator, GP12, combines a ResNet-18 with a conditional flow acting on spectrogram embeddings and produces $p(\bm{x},\theta\mid \text{context})$6 posterior samples in $p(\bm{x},\theta\mid \text{context})$7 seconds on a V100 GPU, although its weakest parameters are luminosity distance and, for many events, chirp mass (Lanchares et al., 12 May 2025). Neural spline flows have also been used for $p(\bm{x},\theta\mid \text{context})$8-dimensional microlensed gravitational-wave inference, reducing the average reported inference time from $p(\bm{x},\theta\mid \text{context})$9 s for Bilby/dynesty to $T_\nu$0 s while preserving good recovery of the main lens and source parameters (Qin et al., 27 May 2025). In a separate Bayesian pipeline, normalizing flows are used twice—inside the Jim sampler and inside the learned harmonic mean estimator—to obtain both posterior samples and evidence, with reported speedups of $T_\nu$1 in $T_\nu$2D and $T_\nu$3 in $T_\nu$4D relative to Bilby + dynesty in the experiments (Polanska et al., 2024).

Across these works, the flow is not merely a generic density estimator. It is a parameter-transport mechanism whose learned Jacobian structure is exploited for posterior evaluation, evidence estimation, nuisance robustness, or exact sampling from an amortized approximation.

5. Parameter-dependent density morphing and two-parameter transports

Another current usage concerns parameter-dependent families of densities, where the goal is not a posterior over parameters but a density that deforms smoothly as parameters vary. Factorizable Normalizing Flows address this by separating a high-fidelity nominal density from a parameter-dependent morphing transformation. The core formula is

$T_\nu$5

The nuisance dependence is expanded as low-order polynomials in each component of $T_\nu$6, with additive per-parameter terms and optional pairwise interactions. In the factorized case, each parameter’s effect is learned from samples in which only that parameter is varied, and the joint response is obtained by summation at inference time; without cross-terms the scaling is linear in the number of nuisance parameters (Valsecchi et al., 29 Jun 2026).

A related but conceptually distinct development is the two-parameter flow for learning population dynamics from unlabeled time marginals. Here one distinguishes physical time $T_\nu$7 from an auxiliary sampling time $T_\nu$8. A conditional flow model first learns, for each fixed $T_\nu$9, a transport from a base law $\Phi(a,t,s)$0 to the marginal $\Phi(a,t,s)$1,

$\Phi(a,t,s)$2

with associated map

$\Phi(a,t,s)$3

Differentiation of the same map in the $\Phi(a,t,s)$4-direction then defines a unique induced physics-time velocity

$\Phi(a,t,s)$5

Under the paper’s $\Phi(a,t,s)$6 assumptions, $\Phi(a,t,s)$7 satisfies the continuity equation for the observed marginals, and the compatibility of the $\Phi(a,t,s)$8- and $\Phi(a,t,s)$9-directions yields the PDE

$X$0

The construction avoids couplings between adjacent physical marginals, supports high-dimensional conditional flow matching, and, unlike minimal-kinetic-energy optimal-transport formulations, can produce non-gradient dynamics with rotational components (Schwerdtner et al., 25 May 2026).

These two directions share an emphasis on coherent parameter dependence, but they differ in what the parameter indexes. In FNF, $X$1 indexes nuisance or systematic variation of an observation density. In two-parameter flows, $X$2 indexes physical evolution while $X$3 indexes an auxiliary generative transport.

6. Parameterized fluid simulation and reduced models

Parameter-conditioned generative modeling of fluid flows treats the entire trajectory distribution as the parameter-dependent object. One approach learns

$X$4

where $X$5 is the full flow field and $X$6 is a vector of operating parameters. In the two-dimensional cylinder problem,

$X$7

and in the three-dimensional half-cylinder problem $X$8. The model uses latent variables, a parameter-conditioned latent prior and transition law, and an auxiliary mutual-information term

$X$9

to force latent states to retain information about $M$0. In the reported experiments, generated simulations are about $M$1 faster than CFD, and a sweep over nearly $M$2 conditions took under $M$3 hours versus nearly two months for CFD (Morton et al., 2019).

A different non-intrusive route is the POD-enhanced deep-learning reduced-order model. Here the learned object is a direct space–time–parameter map

$M$4

implemented as a POD compression, a convolutional autoencoder defining a nonlinear trial manifold, and a feedforward network that maps $M$5 to latent coordinates. The method is applied to flow around a cylinder, an FSI benchmark, and blood flow in a cerebral aneurysm. Reported online speedups are $M$6, $M$7, and $M$8, respectively, with single-time-query speedup in the FSI case reaching $M$9 (Fresca et al., 2021).

Parameter Extension Simulation represents yet another parameter-flow strategy. PES defines the target as a flow solution at desired parameter values computed from a reference solution through an asymptotic expansion,

$f$0

with the target DNS solution written as

$f$1

The required reference family is supplied by Controlled Eddy Simulation, in which the filtered momentum equation contains an artificial force term scaled by a weight coefficient $f$2. The paper reports that the $f$3-order PES solution, equivalently the reference CES solution, has similar accuracy to a traditional LES solution while its computational cost is much lower, and that higher-order PES improves model accuracy further (Jin, 2019).

These fluid-dynamical works do not use “flow” in the invertible-normalizing-flow sense. Their commonality lies instead in learning or extrapolating a parameterized family of evolving states or fields.

7. Limitations and open problems

Several open problems recur across the literature. In rigidity theory, the orientable closed $f$4-manifold classification leaves open the nonorientable case because it is unclear whether lifting a parameter rigid flow to the orientable double cover preserves parameter rigidity; higher-dimensional analogues are also not proved (Matsumoto, 2010). In matrix theory, the generalized no-crossing picture for general complex or real matrix flows remains conjectural, and the relationship between eigencurves in $f$5 and fixed-similarity block decomposability is described as difficult (2002.0188). The invariant-subspace algorithm for block decomposition is effective in many examples, but it depends on thresholding numerically small entries, can be sensitive to repeated eigenvalues and Jordan structure, and does not furnish a perturbation theory for approximate decomposability (Uhlig, 2020).

Flow-based inference faces a different set of limitations. Two-parameter population-dynamics flows recover the unique velocity induced by the selected base-to-marginal transport $f$6, not “the” microscopic stochastic law, and different learned transports to the same marginals can induce different physics-time velocities (Schwerdtner et al., 25 May 2026). Factorizable density morphing assumes low-order polynomial dependence and approximate additivity across nuisance parameters; if pairwise interactions are needed, model and data complexity grow as $f$7 (Valsecchi et al., 29 Jun 2026). In gravitational-wave Bayesian pipelines, normalizing flows can struggle with multimodality because invertible maps preserve topology and forward-KL training is mode-covering; the multimodal-base remedy used for evidence estimation is effective in the reported examples but explicitly left as incomplete for harder posterior geometries (Polanska et al., 2024). The image-based GP12 model likewise shows that flow expressivity is not enough when the conditioning representation discards information or lacks noise conditioning; the paper identifies luminosity distance, lower chirp masses, and missing PSD conditioning as its clearest current weaknesses (Lanchares et al., 12 May 2025).

A plausible synthesis is that “parameter flows” now denotes a broad methodological family unified less by a single formal definition than by a shared ambition: to represent coherent dependence on external parameters at the level of trajectories, invariant structures, spectra, densities, or posteriors. The mathematical and computational tools differ sharply across fields, but the recurring technical themes are the same—regularity under parameter change, tractable transport, structural invariants, and the boundary between exact and approximate continuation.