Nonlinear Coordinate Transformation

• Nonlinear coordinate transformation is an invertible, non-affine mapping that recasts variables to expose hidden structures and simplify complex governing equations.
• It is applied in fields like transformation optics, dynamical systems, and numerical analysis to preserve invariants and linearize nonlinear models.
• This approach enables effective PDE simplification, optimized coordinate representations, and improved stability in algorithms for high-dimensional data analysis.

Searching arXiv for the specified paper and closely related nonlinear coordinate transformation work. Nonlinear coordinate transformation denotes a change of variables whose dependence on the original coordinates is not affine linear and may, depending on context, involve field dependence, higher derivatives, nonlocal terms, manifold constraints, or learned nonlinear maps. Across mathematical physics, dynamical systems, numerical analysis, geometry processing, and data analysis, such transformations are used either to preserve the form of governing equations under nontrivial reparameterization, to simplify nonlinear models by passing to adapted coordinates, or to expose latent structure that is obscured in a fixed coordinate frame. The topic is technically heterogeneous: in some settings the transformation law remains local and Jacobian-based, while in others it acquires second derivatives, quadratures, harmonicity constraints, or optimization-defined representations (Gratus et al., 2018, Paul et al., 2011, He et al., 2010).

1. Conceptual scope and formal character

A nonlinear coordinate transformation is, at minimum, an invertible map between coordinate systems that is not restricted to linear changes of basis. In transformation optics, the formalism is stated for arbitrary space-time coordinate transformations $x^\alpha \rightarrow x^{\alpha'}(x^\alpha)$, including nonlinear ones, and the transformed constitutive tensors are determined by the Jacobian and its determinant (Paul et al., 2011). In nonlinear systems theory, a state transformation may take the form $\bar{x} = T x$ with $T$ invertible, after which the transformed dynamics become $\dot{\bar{x}} = T f(T^{-1}\bar{x})$; in the cited ROA framework this is used systematically to generate multiple Takagi-Sugeno representations of the same nonlinear system (Sel et al., 29 Mar 2026). In diffusion theory, the transformation may be defined by an integral,

$\alpha(x) := \int_0^x \frac{dz}{g(z)},$

with $y=\alpha(x)$, thereby converting a problem with degenerate mobility on a bounded interval into one with linear mobility on $\mathbb{R}$ under an Osgood integrability condition (Ansini et al., 2019).

The formal role of nonlinearity varies by domain. It may alter how physical objects transform, as in electromagnetic quadrupoles whose components do not obey naive tensorial rules under general coordinate changes (Gratus et al., 2018). It may also be used constructively to simplify dynamics: the Kong-Zhang transformation for relativistic strings yields harmonic coordinates under which the nonlinear string equations reduce to linear wave equations (He et al., 2010). In data analysis and visualization, the map is itself the learned object: a deep CNN implements a nonlinear mapping $f(\cdot)$ that sends high-dimensional data to a low-dimensional visualizable space while preserving angular similarity structure (Zheng et al., 2016).

A plausible implication is that “nonlinear coordinate transformation” is best treated as a family of techniques rather than a single theory. What unifies the family is the use of non-affine reparameterization to preserve, reveal, or simplify structure.

2. Transformation laws beyond tensors

One of the most explicit departures from standard tensorial behavior appears in the transformation theory of electromagnetic quadrupoles. The quadrupole current source is represented by

$$J^a(x) = \frac{1}{2} \int \gamma^{abc}(\tau) \frac{\partial^2}{\partial x^b \partial x^c} \delta(x - C(\tau)) d\tau,$$

with $\gamma^{abc}=\gamma^{acb}$ and $\bar{x} = T x$0 (Gratus et al., 2018). Under a general coordinate transformation $\bar{x} = T x$1, the quadrupole components do not transform by a triple Jacobian alone. The full rule contains second derivatives of the coordinate map and a nonlocal worldline integral,

$\bar{x} = T x$2

where $\bar{x} = T x$3 is the Jacobian and $\bar{x} = T x$4 its second derivative evaluated on the worldline (Gratus et al., 2018).

This transformation law is unusual in two distinct senses. First, the chain rule acting on second derivatives of a distribution produces second-derivative terms of the coordinate map. Second, the paper states that the integral term has not been seen before in coordinate transformations (Gratus et al., 2018). The consequence is that quadrupoles are not tensorial objects in the naive sense, unlike dipoles, whose components transform using only first derivatives. The same work gives a concrete example in which a quadrupole that is free of dipole terms in polar coordinates acquires a dipole term $\bar{x} = T x$5 in Cartesian coordinates after the full nonlinear transformation is applied (Gratus et al., 2018).

This establishes a broader point: nonlinear coordinate transformations can change the apparent multipole content of a source. The paper states this directly as “There is no such thing as pure quadrupole” under general coordinate changes (Gratus et al., 2018). In the same spirit, the covariant formulation of transformation optics treats arbitrary nonlinear space-time transformations through tensorial constitutive laws, but there the transformation remains local in the Jacobian-based sense because the fields and susceptibilities are encoded in four-dimensional tensors (Paul et al., 2011). The contrast is instructive: not every nonlinear coordinate transformation leads beyond tensor calculus, but some distributional source models do.

3. Geometric and covariant frameworks in field theory

In transformation optics, nonlinear coordinate transformation is part of a geometric design principle: physical media are engineered so that electromagnetic propagation reproduces the behavior of fields in a transformed coordinate system. The covariant formalism expresses Maxwell’s equations through the field strength tensor $\bar{x} = T x$6, the displacement tensor $\bar{x} = T x$7, and a hierarchy of susceptibility tensors $\bar{x} = T x$8, with a general transformation law that applies to linear, nonlinear, bianisotropic, magneto-optical, and moving media (Paul et al., 2011). Because the formalism is covariant, arbitrary space-time coordinate transformations are admissible, and magneto-electric coupling terms appear automatically (Paul et al., 2011).

A distinct extension incorporates nonlinearity into the transformation itself. In the unified approach to nonlinear transformation materials, the constitutive relations are generalized so that both the background medium and the effective metric may depend on the field: $\bar{x} = T x$9 This permits nonlinear backgrounds, nonlinear coordinate transformations, or both (Sklan et al., 2017). The paper uses this framework to discuss cloaking an optical soliton in a Kerr medium, modeling nonlinear gravitational solutions through field-dependent metrics, and controlling transport in a Debye solid via a temperature-dependent transformation (Sklan et al., 2017).

The formal connection between geometry and material response is thus two-tiered. In the covariant treatment, arbitrary coordinate transformations preserve the form of the field equations and dictate how susceptibility tensors transform (Paul et al., 2011). In the nonlinear-material treatment, the transformation may itself depend on the fields, so the geometry–material relation becomes field-responsive (Sklan et al., 2017). This suggests that nonlinear coordinate transformation is not merely a passive relabeling of points; in transformation materials it can become an active constitutive design mechanism.

A related but different geometric use arises in nonlinear relativity. The Fock transformation is a rational nonlinear map between inertial coordinates with an invariant length scale $T$0, and the associated momentum transformation is chosen so that $T$1 remains invariant, allowing plane waves for free particles (Bouda et al., 2012). A later treatment of inertial transformations goes further by deriving the most general inertial frame transformation as projective linear rather than affine linear,

$T$2

and relates the resulting conditions to Schwarzian differential equations (Agia, 31 Dec 2025). In both cases, nonlinearity is tied to kinematics rather than media, but the common feature is that physically meaningful invariants survive only after the transformation law is reformulated with care.

4. Linearization, regularization, and adapted coordinates in differential equations

A major use of nonlinear coordinate transformation is to recast nonlinear equations into forms better suited to analysis. For relativistic strings, the transformed coordinates introduced in the cited note are harmonic coordinates, and under this transformation the nonlinear relativistic string equations simplify into linear wave equations (He et al., 2010). The transformation is built from the initial data through integral formulas for $T$3, and the paper shows that solutions in the original and transformed formulations are diffeomorphic (He et al., 2010). Here the nonlinearity is transferred from the PDE to the coordinate map.

A comparable strategy appears in ordinary differential equations through the generalized Cole-Hopf-Darboux transformation

$T$4

which relates a nonlinear second-order ODE to a linear second-order ODE under compatibility conditions (Humi, 2012). For the case $T$5, a sufficient condition is given by $T$6 together with an intrinsic formula for $T$7, and the framework connects nonlinear equations to special functions, including Airy-function representations of a Painlevé II case (Humi, 2012). Although this is not a coordinate change in the geometric sense, it is a nonlinear transformation of dependent variables with the same conceptual aim: transfer complexity to a transformation and expose a simpler partner equation.

In nonlinear diffusion with degenerate fast-decay mobility, the coordinate change

$T$8

induces a mass-preserving rescaling of the density and turns the original equation on $T$9 into a nonlinear diffusion equation with linear mobility on $\dot{\bar{x}} = T f(T^{-1}\bar{x})$0 (Ansini et al., 2019). The Osgood condition

$\dot{\bar{x}} = T f(T^{-1}\bar{x})$1

is decisive because it maps the bounded interval to the entire real line and removes the need to prescribe boundary conditions in the transformed problem (Ansini et al., 2019). The paper then proves existence and uniqueness for the rescaled density $\dot{\bar{x}} = T f(T^{-1}\bar{x})$2 as a weak solution (Ansini et al., 2019).

In classical mechanics, the invertible linearization map for the quartic oscillator combines a nonlinear algebraic deformation of space with a nonlinear time deformation involving a quadrature, preserving energy while establishing a one-to-one correspondence with the harmonic oscillator (Anderson, 2012). For the quartic case, the time differential transforms as

$\dot{\bar{x}} = T f(T^{-1}\bar{x})$3

and the construction extends to all attractive even-power potentials (Anderson, 2012). This suggests a recurring pattern: nonlinear transformations are often most effective when they are paired with an adapted time reparameterization rather than a spatial map alone.

5. Dynamical systems, control, and numerical approximation

In nonlinear systems analysis, coordinate transformations are used to reduce conservatism in certified stability estimates. The TS-based ROA method constructs multiple Takagi-Sugeno models, each obtained from the original nonlinear system under a distinct linear coordinate transformation $\dot{\bar{x}} = T f(T^{-1}\bar{x})$4, computes a local ROA for each transformed system, maps the ellipsoidal region back, and takes the union

$\dot{\bar{x}} = T f(T^{-1}\bar{x})$5

as the overall estimate (Sel et al., 29 Mar 2026). The baseline Lyapunov analysis uses a common quadratic form $\dot{\bar{x}} = T f(T^{-1}\bar{x})$6 and LMIs $\dot{\bar{x}} = T f(T^{-1}\bar{x})$7, while the transformed-system procedure repeats the same tractable subproblem in different coordinates (Sel et al., 29 Mar 2026). A related 2025 extension replaces common quadratic Lyapunov functions with piecewise quadratic ones, again under multiple transformed representations, to enlarge the certified ROA (Sel et al., 17 Jul 2025).

For high-dimensional tensor approximation, rank reduction via coordinate flows formulates the search for a useful transformation as an optimization problem on $\dot{\bar{x}} = T f(T^{-1}\bar{x})$8. The transformed function is $\dot{\bar{x}} = T f(T^{-1}\bar{x})$9, and the cost functional is a surrogate based on FTT singular values (Dektor et al., 2022). The resulting Riemannian gradient descent on the matrix manifold produces quasi-optimal linear coordinate transformations that lower tensor rank and improve efficiency for linear and nonlinear PDEs (Dektor et al., 2022). The paper explicitly notes that the results for linear transformations open the possibility for generalizations to larger classes of nonlinear transformations (Dektor et al., 2022).

In discontinuous Galerkin computation, mapped DG interpolation handles analytic coordinate transformations, including mappings of the form $\alpha(x) := \int_0^x \frac{dz}{g(z)},$0 with $\alpha(x) := \int_0^x \frac{dz}{g(z)},$1 arbitrary and even nonlinear (Francisquez et al., 2021). The mapped field is constructed by Galerkin projection, the procedure is quadrature-free, and the reported accuracy is $\alpha(x) := \int_0^x \frac{dz}{g(z)},$2-order in the DG representation and $\alpha(x) := \int_0^x \frac{dz}{g(z)},$3-order in the cell averages (Francisquez et al., 2021). The method preserves certain moments exactly and is applied in settings such as sheared-shift and twist-shift boundary conditions in 3D and 5D simulations (Francisquez et al., 2021).

These examples share a methodological feature: the transformation is not merely descriptive but algorithmic. It is chosen because a transformed representation yields larger ROAs, lower tensor ranks, tractable projections, or exact preservation of moments.

6. Coordinate construction from data and geometry

In data-driven settings, nonlinear coordinate transformation often means constructing coordinates adapted to latent topology or discrimination objectives. Persistent cohomology and circular coordinates addresses the limitation of real-valued nonlinear dimensionality reduction on data with intrinsic circle structure. The method uses persistent cohomology to identify significant $\alpha(x) := \int_0^x \frac{dz}{g(z)},$4 classes, then harmonic smoothing and integration to construct circle-valued coordinate functions on the data set (0905.4887). The conceptual basis is the correspondence

$\alpha(x) := \int_0^x \frac{dz}{g(z)},$5

so a nontrivial cohomology class yields a meaningful $\alpha(x) := \int_0^x \frac{dz}{g(z)},$6-valued coordinate (0905.4887). This broadens the notion of coordinate beyond $\alpha(x) := \int_0^x \frac{dz}{g(z)},$7-valued embeddings.

A neural version appears in end-to-end data visualization by metric learning and coordinate transformation. There, a deep CNN learns a nonlinear mapping $\alpha(x) := \int_0^x \frac{dz}{g(z)},$8 from high-dimensional input to a low-dimensional metric space, guided by Triangular Similarity,

$\alpha(x) := \int_0^x \frac{dz}{g(z)},$9

which is shown to be equivalent to Cosine Similarity (Zheng et al., 2016). The learned outputs are normalized onto a sphere or hypersphere and then “unfolded” through angular coordinate transformation to produce complementary visualizations (Zheng et al., 2016). The second view yields better classification results than state-of-the-art methods in the visualizable spaces, with reported MNIST accuracies of 94.76% in 2D, 98.05% in 3D, and 98.66% in 4D for the unfolded view (Zheng et al., 2016).

In geometry processing, Nonlinear Rotation-Invariant Coordinates represent a triangular mesh by stacking edge lengths and dihedral angles,

$y=\alpha(x)$0

which determines the immersion up to rigid motion once triangle inequalities and integrability conditions are satisfied (Sassen et al., 2019). The admissible set of such coordinates is described as a nonlinear submanifold $y=\alpha(x)$1, with integrability reformulated through quaternions to facilitate first and second derivative calculations (Sassen et al., 2019). Optimization is then carried out directly in this coordinate space, and the paper reports that this is particularly effective for problems involving near-isometric deformations (Sassen et al., 2019).

A different data-driven use occurs in geospatial correction. Residual-based neural correction first applies a parametric geometric transformation $y=\alpha(x)$2, computes structured residuals, and then trains a neural network $y=\alpha(x)$3 to predict only the residual distortion, leading to corrected coordinates

$y=\alpha(x)$4

The method is evaluated on simulated and real-world image georeferencing tasks and is reported to be more accurate and stable than direct neural conversion or classical transformation models under sparse or structured control point configurations (Rofatto et al., 19 Apr 2025).

7. Misconceptions, limitations, and recurrent themes

A common misconception is that coordinate changes preserve the “type” of an object in a straightforward tensorial sense. The quadrupole example demonstrates that this fails for distributional multipole sources under general nonlinear transformations: second derivatives and a worldline integral appear, and dipole content can emerge where none was present in the original coordinates (Gratus et al., 2018). A second misconception is that inertial-frame transformations must be affine linear. The cited derivation of nonlinear inertial transformations shows instead that preserving the Law of Inertia yields projective-linear transformations, with affine linearity recovered only after imposing invariance of the speed of light in all directions (Agia, 31 Dec 2025).

Another recurrent limitation is that nonlinear transformations can simplify one part of a problem while complicating another. In mapped DG interpolation, exactness for analytic mappings is accompanied by diffusion and aliasing effects under repeated shearing, especially when unresolved scales are generated (Francisquez et al., 2021). In tensor rank reduction, linear coordinate flows already improve approximation efficiency, but the paper notes that nonlinear transformations would introduce metric-tensor complications for differential operators (Dektor et al., 2022). In non-smooth dynamics, discontinuous substitutions regularize impact equations and encode restitution directly into transformed variables or time, but they also require careful treatment of generalized functions and may replace explicit discontinuities with nonlinear switching terms (Pilipchuk, 2011).

Across the literature, three themes recur. First, nonlinear coordinate transformation is often motivated by invariance: preserving Maxwell’s equations, multipole currents, or the law of inertia under nontrivial reparameterization (Paul et al., 2011, Gratus et al., 2018, Agia, 31 Dec 2025). Second, it is frequently a device for simplification: transforming nonlinear strings to linear waves, degenerate mobility to linear mobility, or quartic oscillators to harmonic ones (He et al., 2010, Ansini et al., 2019, Anderson, 2012). Third, it increasingly serves as an optimization variable in its own right, whether in TS-based ROA enlargement, tensor-rank reduction, learned visualization, or residual correction (Sel et al., 29 Mar 2026, Dektor et al., 2022, Zheng et al., 2016, Rofatto et al., 19 Apr 2025).

Taken together, these works show that nonlinear coordinate transformation is not a peripheral technicality but a central mechanism for reconciling representation, invariance, and computability across a wide range of nonlinear problems.