Parametrization Methods in Applied Mathematics

• Parametrization method is a strategy that replaces implicit, high-dimensional objects with explicit maps, ensuring properties like injectivity, stability, and smoothness.
• It is widely applied in fields such as control theory, dynamical systems, and geometric design, where it transforms complex constraints into tractable representations.
• The approach also serves as a powerful tool for statistical inference and modeling in physical systems, balancing structural constraints with computational efficiency.

The parametrization method is a broad mathematical and computational strategy for replacing an implicit, high-dimensional, or structurally constrained object by an explicit family of variables, coordinates, or maps whose properties are easier to analyze, optimize, or compute. In the literature surveyed here, parametrization appears as a low-dimensional latent map for geological fields, $\mathbf y = G(\mathbf z)$, as a conjugating immersion $K:X_c\to X$ for invariant manifolds, as an exact description of all systems consistent with noisy data, and as a device for turning constrained inference on non-Euclidean domains into unconstrained optimization on Euclidean space (Chan et al., 2017, Berg et al., 2020, Brändle et al., 2024, Leger, 2023). Across these settings, the central issue is not merely the existence of parameters, but the structural quality of the parametrization: injectivity or almost injectivity, stability, boundedness, smoothness, analytic continuation, preservation of constraints, and computational reuse.

1. Formal roles of parametrization

A parametrization may be understood as an explicit map from a parameter space into a target class of objects, or as an explicit representation of all admissible members of a constrained set. In the geological setting, the target is a high-dimensional permeability or porosity field $\mathbf y\in\mathbb R^M$, while the parameter space is a latent vector $\mathbf z\in\mathbb R^m$ with $m\ll M$, and the parametrization is the nonlinear generator $\mathbf y=G(\mathbf z)$ (Chan et al., 2017). In the copositive-polynomial setting, the target is the cone $\mathcal C_d$ of monic copositive univariate polynomials of degree $d$, and the map $\Phi_d:(\mathbb R_{\ge 0})^d\to\mathcal C_d$ is surjective and generically injective (Hong et al., 2024). In direct data-driven control, the target is the set $\Sigma_\Theta$ of all models consistent with errors-in-variables data, and the parametrization is exact: $K:X_c\to X$0 This converts an implicit data-consistency relation into an explicit admissible family (Brändle et al., 2024).

The formal status of a parametrization can vary. Some constructions are explicitly bijective or diffeomorphic. The statistical-inference cookbook builds maps between constrained parameter spaces and Euclidean spaces and insists on bijectivity in order to preserve identifiability, asymptotic arguments, and compositional use with automatic differentiation (Leger, 2023). Other constructions are only “almost injective” or “generically injective.” For copositive polynomials, non-uniqueness occurs when the same polynomial has multiple nonnegative double roots, so the map is surjective and almost injective rather than globally one-to-one (Hong et al., 2024). At the categorical level, parametrization is treated as a universal construction: a non-pure operation $K:X_c\to X$1 is transformed into $K:X_c\to X$2, and the paper proves that this parameterization process is provided by a free functor, while parameter passing is a natural transformation (0908.3634).

These examples indicate that “parametrization method” does not denote a single algorithm. It denotes a family of constructions whose common purpose is to replace a difficult object by a structured coordinate description, while making explicit which properties of the original problem are preserved.

2. Invariant objects and conjugacy-based parametrization

In dynamical systems, the parametrization method is centered on invariance equations. For discrete systems $K:X_c\to X$3, the center manifold is written as the image of a map $K:X_c\to X$4, together with a reduced dynamics $K:X_c\to X$5, satisfying

$K:X_c\to X$6

For ODEs, the corresponding equation is

$K:X_c\to X$7

This formulation does more than prove existence of a manifold: it constructs both the embedding and the reduced dynamics, with the reduced map $K:X_c\to X$8 treated as part of the unknown (Berg et al., 2020). A distinctive consequence is the freedom to choose a convenient normal form for the conjugate dynamics, which the paper exploits in a reaction-diffusion application to derive explicit information on a period-doubling bifurcation and associated heteroclinic structure (Berg et al., 2020).

For stable and unstable manifolds of low-dimensional maps, the same idea becomes a series-based analytic representation. In generalized Hénon-type maps, the unstable manifold is written as

$K:X_c\to X$9

and substitution into the invariance equation $\mathbf y\in\mathbb R^M$0 yields recursive equations for the coefficients (Anastassiou et al., 2016). In four-dimensional coupled maps, the parametrization becomes a two-variable series

$\mathbf y\in\mathbb R^M$1

which provides a local analytic description of two-dimensional stable and unstable manifolds and supports high-precision computation of homoclinic intersections (Anastassiou et al., 2016).

A closely related but methodologically distinct use appears for state-dependent delay perturbations of a planar ODE with a limit cycle. Rather than constructing the full infinite-dimensional SDDE flow, the paper seeks special solutions of the form

$\mathbf y\in\mathbb R^M$2

derives a functional invariance equation for $\mathbf y\in\mathbb R^M$3, and proves a posteriori existence of a periodic solution and a two-parameter family of solutions asymptotic to it (Yang et al., 2020). This suggests a characteristic strength of parametrization methods in singularly perturbed settings: they can target a finite-dimensional invariant family without requiring a global phase-space theory.

A more data-driven variant appears in coarse-grained analysis of microscopic simulators. There, stable and unstable manifolds near a coarse saddle are represented as polynomial graphs over stable or unstable eigenspaces, and the coefficients are identified by enforcing invariance through a coarse timestepper built from short bursts of microscopic simulation (Siettos et al., 2019). The parametrization method thus survives even when no explicit macroscopic ODE or PDE is available.

3. Parametrization as an information and inference device

In several applications, parametrization is judged by how much information it extracts from fixed data while respecting structural constraints. For the dark-energy equation of state $\mathbf y\in\mathbb R^M$4, the problem is explicitly posed as theoretical optimization of the parametrization using the Fisher information matrix. Near the maximum-likelihood point, the parameter uncertainty ellipsoid satisfies

$\mathbf y\in\mathbb R^M$5

and its volume scales as

$\mathbf y\in\mathbb R^M$6

The proposed criteria for a “good” parametrization are high-redshift stability and a large determinant $\mathbf y\in\mathbb R^M$7, because maximizing $\mathbf y\in\mathbb R^M$8 minimizes the allowed parameter volume (Lee, 2011). Within the tested two-parameter family $\mathbf y\in\mathbb R^M$9, linear and logarithmic choices can yield larger determinants than CPL but diverge as $\mathbf z\in\mathbb R^m$0. The proposed alternative,

$\mathbf z\in\mathbb R^m$1

remains finite at high redshift and yields a determinant larger than CPL on the same $\mathbf z\in\mathbb R^m$2 data (Lee, 2011).

A complementary cosmological development is the divergence-free class

$\mathbf z\in\mathbb R^m$3

constructed so that the equation of state remains finite on the whole physical redshift domain $\mathbf z\in\mathbb R^m$4, unlike CPL, which diverges as $\mathbf z\in\mathbb R^m$5 (Feng et al., 2012). The two simplest members,

$\mathbf z\in\mathbb R^m$6

were constrained with MCMC using WMAP seven-year full CMB power spectrum, BAO from SDSS DR7, and Union2 supernovae, with dark-energy perturbations included through the standard $\mathbf z\in\mathbb R^m$7 crossing prescription (Feng et al., 2012).

In neural spectral analysis, parametrization again appears as model selection under structural constraints. The Multi-Modal Spectral Parametrization Method models the aperiodic EEG background as a continuous piecewise power law,

$\mathbf z\in\mathbb R^m$8

and supplements it with Gaussian oscillatory peaks (Racz et al., 23 May 2025). A test statistic

$\mathbf z\in\mathbb R^m$9

decides whether the bimodal model is warranted or whether the spectrum should be reduced to a unimodal power law (Racz et al., 23 May 2025). The paper reports accurate recovery of slopes and breakpoint frequency in simulation and finds that broadband resting-state EEG spectra often exhibit steeper low-frequency and flatter high-frequency scaling regimes (Racz et al., 23 May 2025).

Across these examples, parametrization is not simply a re-expression of variables. It functions as a figure-of-merit design problem, where boundedness, continuity, and identifiability directly affect the inferential power of the model.

4. Geometry, shape deformation, and embedded design spaces

In geometric design and computational mechanics, parametrization often serves to reduce dimensionality while preserving manufacturable or solver-compatible structure. In CFD optimization, an automatic morphing-based parametrization is built from a small set of morphing nodes and a Kriging/RBF displacement field,

$m\ll M$0

A greedy maximum-variance criterion selects the morphing nodes, and a precomputed interpolation matrix

$m\ll M$1

makes later deformations efficient through $m\ll M$2 (Łaniewski-Wołłk, 2013). The same framework extends to fixed regions by modifying the kernel so that it vanishes on a prescribed set $m\ll M$3 (Łaniewski-Wołłk, 2013).

In reduced-order shape optimization, parametric model embedding augments geometry data with the original design variables,

$m\ll M$4

and assigns zero weight to the appended parameter block. The resulting generalized PCA problem preserves the same nonzero eigenvalues and geometric modes as the original KLE/PCA reduction, while also yielding embedded components that recover the original parameterization (Serani et al., 2022). This construction addresses the pre-image problem for industrial CAD-like models by keeping reduced coordinates tied to the original design variables (Serani et al., 2022).

In stereo-SLAM and VIO, co-planar parametrization replaces multiple independent landmark variables by shared plane parameters. A planar point with pixel coordinates $m\ll M$5 satisfies

$m\ll M$6

so co-planar points and lines can be represented through their 2D image coordinates and the parameters of the supporting plane (Li et al., 2020). The paper reports that this reduces the number of updated items and shrinks the Hessian, with the second synthetic sequence dropping from around $m\ll M$7–$m\ll M$8 items in traditional formulations to $m\ll M$9 items in the co-planar formulation (Li et al., 2020).

A more algebraic geometric example is the parametrization of translational surfaces. Given an irreducible implicit surface $\mathbf y=G(\mathbf z)$0, the goal is to determine whether it admits a standard translational parametrization

$\mathbf y=G(\mathbf z)$1

The construction is based on first finding a rational curve on the intersection

$\mathbf y=G(\mathbf z)$2

and then recovering the second generating curve from a factorization of $\mathbf y=G(\mathbf z)$3 (Perez-Diaz et al., 2014). The paper gives necessary and sufficient conditions and an explicit constructive procedure (Perez-Diaz et al., 2014).

5. Physical and learned parametrizations

In subsurface modeling, Wasserstein GANs provide a nonlinear parametrization of geological models by learning a generator $\mathbf y=G(\mathbf z)$4 such that

$\mathbf y=G(\mathbf z)$5

with $\mathbf y=G(\mathbf z)$6 (Chan et al., 2017). The contrast with PCA is explicit: PCA captures covariance structure through a truncated linear basis, while WGAN parametrization is intended to learn non-Gaussian and multipoint structures characteristic of channelized geology (Chan et al., 2017). The paper reports that GANs reproduce both geological structures and induced flow statistics more faithfully than PCA in the tested uncertainty-propagation problems (Chan et al., 2017).

In small-$\mathbf y=G(\mathbf z)$7 QCD phenomenology, the nuclear structure function is parametrized by reusing a Froissart-bounded proton fit and rescaling the saturation variable according to the ASW prescription. The nuclear saturation scale is written as

$\mathbf y=G(\mathbf z)$8

which leads to a nuclear structure function $\mathbf y=G(\mathbf z)$9 with the same functional form as the proton one but evaluated at the rescaled scale (Boroun et al., 2022). The method is intended for the low-$\mathcal C_d$0 regime and is reported to agree well with E665, EMC, and NMC data for $\mathcal C_d$1 (Boroun et al., 2022).

For resonance analysis, the extended Lee–Friedrichs model is proposed as a practical parametrization of line shapes that respects threshold behavior, unitarity, and analyticity. In the $\mathcal C_d$2 relativistic case, the propagator is

$\mathcal C_d$3

with spectral density chosen to encode relativistic phase space, orbital threshold behavior, and ultraviolet suppression (Cui et al., 21 Feb 2025). The poles are zeros of the continued inverse resolvent on appropriate Riemann sheets, and the method is presented as a natural framework for two-pole structures in which one pole is shifted from a discrete state and another is dynamically generated (Cui et al., 21 Feb 2025).

In direct data-driven robust control with error-in-variables, parametrization becomes a bridge from noisy regression data to robust-analysis machinery. The Sherman–Morrison–Woodbury formula converts the explicit data-consistent family into a linear fractional transformation with uncertainty blocks $\mathcal C_d$4 and $\mathcal C_d$5, enabling guaranteed $\mathcal C_d$6-norm analysis through a single semidefinite program whose size does not grow with the number of samples (Brändle et al., 2024). This is a notably strong form of parametrization: it is exact, modular, and immediately compatible with robust-control tools.

6. Recurring criteria, limitations, and points of contention

The surveyed literature repeatedly evaluates parametrizations by a small set of structural criteria. Injectivity or identifiability is one. The statistical-inference cookbook explicitly rejects many-to-one maps because they destroy identifiability, while several other papers settle for almost injectivity or generic injectivity when exact bijection is topologically or algebraically unavailable (Leger, 2023, Hong et al., 2024). Stability and boundedness are another recurring criterion. In dark-energy phenomenology, high-redshift divergence is treated as a serious defect; in SDDEs and Richards equation, parametrization is designed to regularize a singular or degenerate problem so that Newton or fixed-point arguments regain uniform control (Lee, 2011, Brenner et al., 2016).

Several papers also challenge common assumptions about ostensibly “model-independent” methods. In dark-energy studies, principal component analysis is described as not optimal for maximal constraints because its binned structure yields fewer contributing terms in the Fisher determinant and greater sensitivity to bin placement and data distribution (Lee, 2011). In manifold learning, standard diffusion maps can fail in two opposite regimes: for small $\mathcal C_d$7, the Jacobian becomes nearly singular near boundaries, while for large $\mathcal C_d$8, distant points on a folded manifold become falsely close, producing a non-injective embedding (Gear, 2012). The large-$\mathcal C_d$9 asymptotics of the left eigenvectors reduce to PCA, and the paper therefore advocates modifying the distance matrix by graph shortest paths when intrinsic geometry is badly represented by ambient Euclidean distance (Gear, 2012).

Parametrizations are likewise sensitive to model class and implementation choices. In CFD mesh morphing, the smoothness and quality of deformation depend on the kernel $d$0, and the method still requires a user-specified variance threshold or node budget (Łaniewski-Wołłk, 2013). In WGAN-based geological parametrization, some artifacts remain, training depends on architecture and tuning, and the learned generator cannot represent structures absent from the training set (Chan et al., 2017). In the Lee–Friedrichs resonance framework, the exponential form factor is acknowledged as a practical approximation that may introduce far-away singularities, although the authors argue that these are irrelevant to the physical region (Cui et al., 21 Feb 2025).

A plausible implication is that parametrization should be treated less as a neutral change of variables than as a structural modeling choice. The map itself determines what kinds of singularities are excluded, which symmetries are preserved, what uncertainty set becomes tractable, and whether the resulting coordinates support exact recovery, stable numerics, or meaningful reduced dynamics.