Parametrisation Method in Computational Models

• Parametrisation method is a design principle that represents high-dimensional objects using reduced coordinates while preserving essential constraints like invariance, analyticity, and orthogonality.
• It applies structured coordinate systems to linearise and simplify complex systems across dynamical analysis, CFD, CAD, and neural network operator models.
• Its implementations include solving conjugacy equations, leveraging gauge freedom, and ensuring reconstruction fidelity in reduced-order and phenomenological models.

Searching arXiv for recent and foundational papers on the parametrisation method and closely related parametrisation frameworks. Parametrisation method denotes a family of constructions in which a high-dimensional object is represented by a smaller or more structured set of coordinates while preserving the constraints that make the object useful: analyticity for form factors, invariance for dynamical manifolds, orthogonality for operators, or feasibility for geometry and meshes. Across the literature, this may mean solving a conjugacy equation $F(K(s)) = K(R(s))$ for invariant manifolds, embedding native CAD or FFD variables into a reduced latent basis, representing affine maps by translation–rotation–symmetric factors, or imposing asymptotic, threshold, or conservation constraints on phenomenological models (Berg et al., 2019, Serani et al., 2022, Kaji et al., 2015, Kirk et al., 2024).

1. Scope and defining structures

In the cited work, parametrisation is not a single algorithm but a recurring design principle. The common pattern is to select coordinates that make the target object easy to compute with while retaining the structure of the original problem. In dynamical systems, the target object is an invariant manifold together with its internal dynamics; in geometry and design optimisation, it is a shape or transformation; in phenomenological modelling, it is a constrained functional form for an observable or constitutive law (Berg et al., 2019, Łaniewski-Wołłk, 2013, Kirk et al., 2024).

Domain	Parametrised object	Structural constraint
Discrete dynamical systems	Center manifold embedding $K$ and internal map $R$	Invariance $F\circ K = K\circ R$
CFD shape optimisation	Surface and volume-mesh deformation field	Smooth mesh motion, optional fixed regions
CAD/FFD reduction	Native design parameters in latent variables	Reconstruction to original parameter space
Affine geometry	$Aff(3)$ with $\det(\hat A)>0$	Surjective Euclidean coordinates
Weakly coupled oscillators	Perturbed invariant torus and phase dynamics	Normal-form phase reduction
Hadronic form factors	$F(q^2)$ on spacelike/timelike domains	Analyticity, threshold, asymptotics

A plausible implication is that parametrisation is most useful when the coordinates are chosen to linearise, factorise, or otherwise expose the governing constraints rather than merely compress the number of variables. This is explicit in the use of gauge choices for center manifolds, generalized inner products in reduced geometry models, and outer functions or weight factors in dispersive form-factor constructions (Berg et al., 2019, Serani et al., 2022, Kirk et al., 2024).

2. Invariant manifolds and conjugacy equations

In the dynamical-systems sense, the parameterization method seeks an embedding $K$ of an invariant manifold and a reduced dynamics $R$ such that

$F(K(s)) = K(R(s)).$

For center manifolds of discrete maps $K$0, the construction is carried out on a spectral splitting $K$1, with $K$2 and $K$3. The method fixes the reparametrization freedom by prescribing the gauge $K$4, solves a fixed-point problem for $K$5, and simultaneously returns the center manifold and a conjugate center dynamics. The cited theorem gives explicit derivative bounds for $K$6, $K$7, $K$8, and $K$9, proves that $R$0 is a global diffeomorphism, and provides an almost-solution estimate that converts an invariance defect $R$1 into $R$2-bounds on $R$3 (Berg et al., 2019).

The same conjugacy viewpoint is extended in geometric singular perturbation theory. There, one computes a slow manifold $R$4, reduced dynamics $R$5, and fast fibre bundle $R$6 from a coordinate-independent conjugacy equation $R$7 together with its variational counterpart. The resulting homological equations provide order-by-order formulas for slow manifolds, fast fibres, and nested reduced flows. A notable consequence is the systematic detection of hidden timescales: terms such as $R$8 and $R$9 can generate an infra-slow dynamics on a lower-dimensional critical manifold even when the original problem appears only two-scale (Lizarraga et al., 2020).

For weakly coupled oscillator networks, the target manifold is an invariant torus. The embedding $F\circ K = K\circ R$0 and reduced phase field $F\circ K = K\circ R$1 satisfy $F\circ K = K\circ R$2, and reducibility yields a tangent–normal decomposition $F\circ K = K\circ R$3. Fourier-modewise solution of

$F\circ K = K\circ R$4

makes it possible to place the reduced phase dynamics directly in normal form by eliminating nonresonant tangential modes. In the three-node Stuart–Landau chain analysed in the paper, first-order coupling vanishes in normal form and remote synchronisation appears at order $F\circ K = K\circ R$5 through an explicit phase equation $F\circ K = K\circ R$6 (Gracht et al., 2023).

3. Geometric, CAD, and affine parametrisations

In CFD optimisation, the parametrisation method is implemented as morphing-based deformation of both the surface and the volume mesh. A sparse set of morphing nodes $F\circ K = K\circ R$7 is assigned displacements $F\circ K = K\circ R$8, and the displacement field is interpolated with a Gaussian kernel

$F\circ K = K\circ R$9

Precomputing

$Aff(3)$0

turns geometry-to-mesh propagation into a fixed linear map. Automatic parametrisation is obtained by maximizing the posterior variance $Aff(3)$1 over the surface mesh, thereby adding control points where the geometry is least represented. The same framework accommodates fixed regions through the tapered kernel $Aff(3)$2, where $Aff(3)$3 and $Aff(3)$4 (Łaniewski-Wołłk, 2013).

Parametric Model Embedding addresses a different problem: native parameters already exist, but one wants a reduced latent space without abandoning the original CAD, Bezier, or FFD representation. The method augments geometric samples $Aff(3)$5 with original parameter samples $Aff(3)$6 into

$Aff(3)$7

and solves a generalized weighted PCA eigenproblem

$Aff(3)$8

with zero weight on the parameter block so that the retained geometric eigenvalues and geometric modes are identical to standard KLE/PCA. The embedding of native parameters into the same latent coordinates is then

$Aff(3)$9

leading to the reconstruction

$\det(\hat A)>0$0

This preserves direct access to the original parameterization while reducing the design-space dimension. In the reported examples, $\det(\hat A)>0$1 for the airfoil and $\det(\hat A)>0$2 for the hull retain $\det(\hat A)>0$3 of geometric variance, with geometric NMSE of approximately $\det(\hat A)>0$4 and $\det(\hat A)>0$5, respectively (Serani et al., 2022).

A concise affine parametrisation uses the Euclidean space

$\det(\hat A)>0$6

and the map

$\det(\hat A)>0$7

Here $\det(\hat A)>0$8 parametrises rotation, $\det(\hat A)>0$9 parametrises the SPD scale–shear factor, and $F(q^2)$0 is the translation. The inverse uses the polar decomposition $F(q^2)$1, with $F(q^2)$2 and $F(q^2)$3, giving $F(q^2)$4. The map is surjective on orientation-preserving affine transformations and restricts linearly to subclasses such as $F(q^2)$5, $F(q^2)$6, $F(q^2)$7, and $F(q^2)$8 (Kaji et al., 2015).

4. Reduced-order models, forced responses, and operator parametrisations

For geometrically nonlinear structures discretised by finite elements, the parametrisation method constructs an invariant manifold directly in physical coordinates by introducing separate displacement and velocity mappings,

$F(q^2)$9

together with reduced dynamics $K$0. A central identity is the arbitrary-order relation $K$1, which allows the homological equations for each monomial to be reduced to a bordered linear system of half the size one would obtain from a generic first-order formulation. Three parametrisation styles are distinguished. Graph style retains all reduced-dynamics monomials and preserves a graph over the master coordinates; complex normal form retains only resonant monomials; real normal form keeps conjugate resonances and yields purely real reduced equations. The cantilever-beam example shows that graph style fails beyond a folding point of the invariant manifold, whereas normal-form styles pass over the fold (Vizzaccaro et al., 2021).

The non-autonomous extension introduces dummy coordinates for forcing, such as $K$2 and $K$3, so that forcing becomes autonomous in the enlarged phase space. The homological quantity

$K$4

then includes multiple occurrences of the forcing frequency. This is the mechanism by which superharmonic resonances appear automatically in the reduced equations: monomials with $K$5 or $K$6 generate $K$7 or $K$8 forcing content, enabling direct ROMs for $K$9 and $R$0 superharmonic resonances. In the physical-basis formulation, the displacement-level homological equation becomes

$R$1

with a bordered augmentation that enforces the chosen normal-form style (Vizzaccaro et al., 2023).

The validity of these expansions is local. Three practical criteria for their validity range are assessed in nonlinear vibrations: an invariance-residual lower bound using

$R$2

a singularity criterion based on the homological operator $R$3, and series tests based on Cauchy and d’Alembert rules. Across Duffing, 2-DOF, and finite-element beam examples, the invariance-residual criterion with $R$4 acts as a robust lower bound, while homological singularity and Cauchy estimates provide upper bounds for the radius of validity (Stabile et al., 18 Mar 2026).

Orthogonal operator parametrisation furnishes a different reduced-structure example. In recurrent neural networks, a hidden-to-hidden matrix $R$5 can be parametrised as a product of Householder reflections,

$R$6

or, equivalently, in compact WY form,

$R$7

This guarantees $R$8, stabilises gradient magnitudes, and gives $R$9 forward and backward cost per time step. With $F(K(s)) = K(R(s)).$0, the parametrisation spans the full orthogonal group $F(K(s)) = K(R(s)).$1 (Mhammedi et al., 2016).

5. Analytic and phenomenological parametrisations in physics

In hadronic physics, parametrisation is often constrained by analyticity and asymptotics. A recent pion-vector-form-factor construction maps the cut $F(K(s)) = K(R(s)).$2-plane to the unit disk with

$F(K(s)) = K(R(s)).$3

and writes

$F(K(s)) = K(R(s)).$4

The weight $F(K(s)) = K(R(s)).$5 removes zeros of the outer function at $F(K(s)) = K(R(s)).$6 and enforces $F(K(s)) = K(R(s)).$7, while the denominator encodes the $F(K(s)) = K(R(s)).$8 pole on the second Riemann sheet without introducing first-sheet zeros. In the reported fits, $F(K(s)) = K(R(s)).$9 yields five free parameters and $K$00-value $K$01, while the nominal $K$02 fit yields six free parameters and $K$03-value $K$04 (Kirk et al., 2024).

Cosmological parametrisations show the same principle in a different guise. A kinematic luminosity-distance expansion uses

$K$05

with $K$06, $K$07, and $K$08 estimated globally and per HEALPix sky pixel. This parametrisation is independent of the cosmological equation of state and was used to derive backreaction terms in $K$09 and $K$10; the measured backreaction remained below detectability, of order $K$11 for the $K$12 correction and $K$13 for $K$14 relative to the linear term (Carvalho et al., 2016). A distinct dark-energy construction fixes the scalar field directly as $K$15, reconstructs a sum-of-exponentials potential

$K$16

and treats $K$17 as a conformal coupling to CDM, with $K$18. In that model, $K$19 is independent of $K$20 (Fonseca et al., 2021).

Modified-gravity and turbulence parametrisations emphasise scale transitions and stochastic closure. In nonlinear cosmological structure formation, a screened effective coupling is written as

$K$21

where $K$22 matches the unscreened linear PPF amplitude, $K$23 is a time-, mass-, and environment-dependent screening scale, and $K$24 encode screened radial scaling and transition sharpness. The same paper relates the screened power-law exponent to the scalar-field equation through

$K$25

with examples such as $K$26 for Vainshtein screening and $K$27 for Yukawa suppression (Lombriser, 2016). In tropical-convection parametrisation, the convective area fraction $K$28 is sampled either instantaneously from $K$29 or from a Markov chain with transition probabilities $K$30, where $K$31 is the large-scale mid-level vertical velocity. The first method is memoryless; the second captures short-lag temporal correlations but suffers from transition sparsity (Gottwald et al., 2015).

A common misconception is that phenomenological parametrisations are inherently unconstrained. The surveyed examples show the opposite: the parametrisations are explicitly engineered to enforce normalization, threshold behaviour, asymptotic scaling, conservation laws, or screening limits (Kirk et al., 2024, Lombriser, 2016).

6. Structural properties, computation, and limitations

The main numerical challenge is that a useful parametrisation must be both expressive and structurally stable. Dense Gaussian kernels in mesh morphing can be ill-conditioned when $K$32 is too small or morphing nodes cluster; the remedies listed are moderate $K$33, minimum node separation, and a nugget term $K$34 (Łaniewski-Wołłk, 2013). In invariant-manifold computations, small divisors are avoided in the normal directions by hyperbolicity of the fast or stable/unstable blocks, but tangential resonances must be handled explicitly through gauge freedom or normal-form choices (Berg et al., 2019, Gracht et al., 2023).

Validity is also limited by topology and locality. The direct vibration literature shows that local asymptotic expansions may cease to be reliable at a finite amplitude and that graph-style parametrisations fail at manifold folds (Stabile et al., 18 Mar 2026, Vizzaccaro et al., 2021). In manifold learning, large-$K$35 diffusion-map limits recover PCA on linear manifolds through the eigenstructure of a distance-derived matrix $K$36, but nontrivial nonlinear manifolds require distance modifications; the Swiss roll can be unfolded by replacing long edges with graph shortest paths, whereas a partial sphere shows that even near-neighbour distances may need modification to obtain a $K$37-$K$38, non-singular parametrisation (Gear, 2012).

Several papers emphasise that preserving original coordinates or constraints can be as important as reducing dimension. PME reconstructs native CAD or FFD parameters but can yield latent points whose reconstructed parameters violate box bounds, so the reported optimisation uses a penalty $K$39 with $K$40 (Serani et al., 2022). The affine-transform parametrisation is surjective on the connected component with $K$41 but excludes reflections and can lose robustness near very small $K$42, where the authors recommend a high-quality polar decomposition before applying their closed-form logarithms and exponentials (Kaji et al., 2015).

The literature therefore suggests a precise working criterion for the term. A parametrisation method is not merely a coordinate choice; it is a coordinate choice together with a structural contract. That contract may be invariance, orthogonality, analyticity, confocality, screening behaviour, or design feasibility. When the contract is respected, the parametrisation becomes a computational instrument rather than a descriptive convenience.