Transformed Gaussian Random Fields

• Transformed Gaussian Random Fields are generated by applying deterministic or stochastic mappings to Gaussian fields to alter marginal distributions and spatial or geometric properties.
• Various constructions—including pointwise transforms, copula-based adjustments, coordinate changes, and Lévy subordination—enable tailored modeling for different applications.
• These fields retain an underlying Gaussian dependence structure, facilitating efficient simulation, inference, and applications in geostatistics, PDEs, meteorology, and cosmology.

Transformed Gaussian random fields are random fields obtained from Gaussian random fields by applying deterministic or stochastic transformations that alter marginals, support, geometry, or indexing while retaining a Gaussian-derived dependence structure in some form. Across the literature, the term covers several distinct but related constructions: pointwise maps $X(x)=T(G(x))$, copula-based marginal replacement $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$, thresholded or truncated latent Gaussian fields, coordinate-transformed fields $X(t)=Z(f(t))$, subordinated constructions on manifolds, and Lévy fields driven by positive transforms of Gaussian random fields (Deiml et al., 19 Aug 2025, Prates et al., 2012, Cheng et al., 2019, Merkle et al., 2022).

1. Main construction paradigms

A common starting point is a Gaussian random field $G$ or Gaussian Markov random field $\varepsilon$, followed by a transformation that changes the observable field while preserving enough structure to permit analysis. In the pointwise formulation emphasized in quantum sampling and in latent geostatistical models, one writes

$X(x)=T(G(x)),$

with $T:\mathbb{R}\to\mathbb{R}$ applied independently at each spatial point. This produces a generally non-Gaussian field whose spatial correlation is inherited from the covariance of $G$ (Deiml et al., 19 Aug 2025, Allard et al., 2020).

A second formulation replaces Gaussian margins while keeping Gaussian-copula dependence. If $\varepsilon\sim N_n(0,\Psi)$, then

$$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$

defines a transformed Gaussian random field with marginal laws $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$0 and dependence parameterized by $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$1; when the underlying Gaussian field is Markov, the result is a transformed Gaussian Markov random field (TGMRF) (Prates et al., 2012).

A third formulation acts on the index set rather than the field values. If $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$2 is a $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$3 diffeomorphism and $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$4, then $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$5 remains a Gaussian random field, but anisotropy and geometry are altered by the pullback through $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$6 (Cheng et al., 2019). On the sphere, time-dependent transformed fields are obtained by coordinate change through Brownian motion or subordinate Brownian motion, for example

$Z_i=F_i^{-1}(\Phi(\varepsilon_i))$7

or by subordinated semigroup action on the spherical harmonic expansion (D'Ovidio, 2012).

A fourth formulation composes a Gaussian field with a Lévy process. Given a Gaussian random field $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$8, a positive transformation $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$9, and a Lévy process $X(t)=Z(f(t))$0, the Gaussian subordinated Lévy field is

$X(t)=Z(f(t))$1

This produces non-Gaussian fields with generally discontinuous sample paths and spatial dependence inherited from $X(t)=Z(f(t))$2 (Merkle et al., 2022).

Construction	Representative form	Representative source
Pointwise transform	$X(t)=Z(f(t))$3	(Deiml et al., 19 Aug 2025)
Truncated latent field	$X(t)=Z(f(t))$4	(Allard et al., 2020)
Gaussian-copula transform	$X(t)=Z(f(t))$5	(Prates et al., 2012)
Diffeomorphic transform	$X(t)=Z(f(t))$6	(Cheng et al., 2019)
Spherical coordinate change	$X(t)=Z(f(t))$7	(D'Ovidio, 2012)
Lévy subordination	$X(t)=Z(f(t))$8	(Merkle et al., 2022)

2. Marginal structure, support constraints, and dependence preservation

Pointwise transforms are chiefly used to impose admissible ranges or non-Gaussian marginal behavior. The quantum-field paper highlights clipping and truncation, exponentiation, logistic-type maps, and thresholding/indicator transforms as typical choices. These enforce boundedness, positivity, or binary phase structure while preserving correlation inherited from the Gaussian precursor (Deiml et al., 19 Aug 2025). In depositional-sequence modeling, the latent field $X(t)=Z(f(t))$9 is transformed into a zero-inflated thickness field through

$G$0

with $G$1 a key parametric choice. This produces non-negative, spatially correlated, zero-inflated thickness fields and cumulative stratigraphic surfaces $G$2 (Allard et al., 2020).

In the copula-based literature, transformed Gaussian fields are designed to decouple marginals from dependence. The Gaussian copula

$G$3

preserves the conditional-independence structure of the underlying Gaussian field under coordinate-wise monotone transforms. This permits gamma margins for positive intensities, beta margins for probabilities in $G$4, and other continuous marginals while retaining sparse Markov structure through the precision matrix $G$5 (Prates et al., 2012). The same principle underlies non-separable spatio-temporal TGMRFs, where the marginals of $G$6 can be chosen independently of the dependence matrix $G$7 (Azevedo et al., 2020).

A useful correction to a common simplification is that transformed Gaussian random fields are not uniformly non-Gaussian. Pointwise nonlinear maps and Lévy subordination usually destroy Gaussianity, but coordinate changes of the form $G$8 preserve Gaussianity because they act on the index set rather than on the Gaussian values themselves (Cheng et al., 2019). The literature therefore suggests that “transformed Gaussian random field” is an umbrella term rather than a single distributional class.

The meteorological power-transformation literature gives a scalar analogue of this marginal-calibration perspective. For nearly Gaussian variables, the transformation

$G$9

with $\varepsilon$0 close to $\varepsilon$1, is used because it is bijective on $\varepsilon$2 and preserves symmetry. The kurtosis of $\varepsilon$3 when $\varepsilon$4 is

$\varepsilon$5

which provides a direct mechanism for matching empirical kurtosis to a transformed-Gaussian model (Gonçalves, 2013).

3. Geometric, temporal, and operator-based transformations

Geometric transformations act on the parameter space and can preserve subtle structural invariants. For smooth Gaussian random fields on Riemannian manifolds related by a diffeomorphism $\varepsilon$6, the gradient and Hessian satisfy

$\varepsilon$7

where $\varepsilon$8 represents $\varepsilon$9. Consequently, critical points correspond under $X(x)=T(G(x)),$0, Hessian index is preserved, expected numbers of critical points match after domain mapping, and the height distribution at critical points is invariant under the transformation. In the linear anisotropic case $X(x)=T(G(x)),$1, expected numbers of critical points become proportional to those of the isotropic field by the factor $X(x)=T(G(x)),$2, while the height distribution remains the same (Cheng et al., 2019).

Anisotropic metric transformations provide another geometric layer. If the canonical metric of an $X(x)=T(G(x)),$3-Gaussian random field satisfies

$X(x)=T(G(x)),$4

and the covariance matrices are uniformly nondegenerate, then the effective index dimension is

$X(x)=T(G(x)),$5

Under $X(x)=T(G(x)),$6, the paper on polar sets proves

$X(x)=T(G(x)),$7

for any independent $X(x)=T(G(x)),$8 with $X(x)=T(G(x)),$9-Lipschitz sample paths, so sets with sufficiently small Hausdorff dimension are polar for the transformed field $T:\mathbb{R}\to\mathbb{R}$0 (Söhl, 2012).

On the sphere, transformed Gaussian fields are built through coordinate change and subordination. Starting from an isotropic spherical Gaussian field

$T:\mathbb{R}\to\mathbb{R}$1

the field

$T:\mathbb{R}\to\mathbb{R}$2

solves

$T:\mathbb{R}\to\mathbb{R}$3

where $T:\mathbb{R}\to\mathbb{R}$4. The transformed angular power spectrum is

$T:\mathbb{R}\to\mathbb{R}$5

so subordination acts as a spectral damping mechanism, with polynomial or exponential high-frequency decay depending on $T:\mathbb{R}\to\mathbb{R}$6 (D'Ovidio, 2012).

In Wiener-space geometry, transformed Gaussian random fields take the form $T:\mathbb{R}\to\mathbb{R}$7, where $T:\mathbb{R}\to\mathbb{R}$8 is a Gaussian path indexed by $T:\mathbb{R}\to\mathbb{R}$9 and $G$0 is a sufficiently smooth Wiener functional. The resulting excursion geometry is described through infinite-dimensional Gaussian Minkowski functionals and an infinite-dimensional Gaussian kinematic formula for $G$1 (Taylor et al., 2011).

4. Representation, simulation, and computational frameworks

Classical simulation of Gaussian precursors remains foundational. FFT-based methods generate stationary Gaussian random fields on rectangular grids with complexity $G$2, using the spectral density $G$3 and the representation

$G$4

This framework directly supports transformed Gaussian fields by postprocessing the Gaussian sample through a pointwise map $G$5, for example lognormal or bounded transforms, with the transformation step costing only $G$6 (Lang et al., 2011).

A different representation strategy rotates the underlying Gaussian Hilbert space. If $G$7 is a vector of independent standard normals and $G$8 is an $G$9-dependent orthogonal map, then

$\varepsilon\sim N_n(0,\Psi)$0

is a Gaussian process whose covariance kernel is

$\varepsilon\sim N_n(0,\Psi)$1

for the $\varepsilon\sim N_n(0,\Psi)$2-th row $\varepsilon\sim N_n(0,\Psi)$3 of $\varepsilon\sim N_n(0,\Psi)$4. This basis-adaptation viewpoint yields reduced Wiener chaos representations in which measure concentration is shifted into a lower-dimensional subspace and a mesoscale Gaussian process captures intermediate structure of the quantity of interest (Tsilifis et al., 2016).

For Gaussian subordinated Lévy fields, approximation combines truncation or interpolation of the Gaussian field with time discretization of the Lévy process. If $\varepsilon\sim N_n(0,\Psi)$5 approximates $\varepsilon\sim N_n(0,\Psi)$6 and $\varepsilon\sim N_n(0,\Psi)$7 approximates $\varepsilon\sim N_n(0,\Psi)$8, then under the paper’s assumptions

$\varepsilon\sim N_n(0,\Psi)$9

which quantifies the joint approximation error for Lipschitz observables $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$0 (Merkle et al., 2022).

The recent quantum contribution takes transformed Gaussian random fields as the primary computational object. It constructs a quantum state approximating the target Gaussian or transformed field with accuracy $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$1 in time

$$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$2

and combines state preparation with amplitude estimation and a quantum pseudorandom number generator to estimate linear and nonlinear observables, including mixed and higher-order moments, with total complexity

$$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$3

(Deiml et al., 19 Aug 2025).

In copula-based TGMRF models, inference is typically Bayesian. The methodology uses CAR-type precision matrices, MCMC, and model comparison by LPML, DIC, and in the spatio-temporal setting WAIC; the papers report that LPML is more stable than DIC in the non-Gaussian TGMRF settings they study (Prates et al., 2012, Azevedo et al., 2020).

5. Scientific applications

In hydrogeology and petroleum geostatistics, transformed latent Gaussian random fields model depositional sequences conditionally on borehole data. Each layer thickness is a thresholded and power-transformed latent Gaussian field allowing null thickness, and stacked cumulative thicknesses define continuous or smooth stratigraphic surfaces $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$4 (Allard et al., 2020).

In spatial generalized linear mixed models, TGMRFs provide non-Gaussian latent fields for conditional means. Gamma Markov fields model Poisson intensities, beta Markov fields model Bernoulli rates, and Gaussian copulas preserve spatial dependence while allowing gamma, beta, or log-normal margins. The same idea extends to non-separable spatio-temporal TGMRFs with interpretable parameters $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$5, including spatio-temporal Gamma random fields for abundance data (Prates et al., 2012, Azevedo et al., 2020).

In random media and PDEs, transformed Gaussian fields enforce physical admissibility of coefficient fields. The quantum sampling paper treats microstructure and PDE coefficient fields of the form $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$6, where $$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$7 imposes boundedness or positivity, and emphasizes direct on-device generation from a small set of statistical parameters to avoid a classical input bottleneck (Deiml et al., 19 Aug 2025). The Gaussian subordinated Lévy field paper places transformed Gaussian fields directly into elliptic diffusion coefficients,

$$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$8

and studies both standard and adaptive finite elements for the resulting random elliptic PDE (Merkle et al., 2022).

On the sphere, time-dependent coordinate-changed and subordinated Gaussian fields modify the angular power spectrum in ways relevant to cosmological modeling. The paper explicitly connects polynomial and exponential spectral damping to phenomena such as the Sachs–Wolfe effect and Silk damping (D'Ovidio, 2012).

In excursion theory, the exponential transform of a Gaussian field,

$$Z_i=F_i^{-1}\bigl(\Phi(\varepsilon_i)\bigr), \qquad i=1,\dots,n,$$9

defines a non-Gaussian field functional whose rare-event structure is asymptotically equivalent to the excursion of an auxiliary Gaussian field $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$00. The same work constructs an efficient Monte Carlo estimator with polynomial-time complexity in $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$01 for computing $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$02 to prescribed relative accuracy (Liu et al., 2012).

In meteorology, nearly Gaussian forecast errors are modeled through the power transform $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$03 with odd-over-odd rational $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$04, selected through the kurtosis formula

$Z_i=F_i^{-1}(\Phi(\varepsilon_i))$05

For daily maximum-temperature forecast errors, the transformed data $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$06 passed Lilliefors and Shapiro–Wilk normality tests and outperformed Laplace and Pearson type IV fits in the reported comparison (Gonçalves, 2013).

6. Conceptual boundaries and current directions

The literature suggests that transformed Gaussian random fields are best understood as a family of constructions rather than a single canonical model. Some transformations primarily alter marginals and support while retaining Gaussian-copula or Gaussian-Markov dependence (Prates et al., 2012); some impose thresholding, truncation, or positivity constraints for physical modeling (Allard et al., 2020, Deiml et al., 19 Aug 2025); some change geometry or coordinates without changing Gaussianity itself (Cheng et al., 2019, D'Ovidio, 2012); and some add new stochastic layers, such as Lévy subordination or Wiener-function transforms, producing discontinuous or highly non-Gaussian fields (Merkle et al., 2022, Taylor et al., 2011).

Several recurring conditions delimit tractability. Low-complexity covariance representation, efficient implementability of the transformation, and sufficient regularity of the Gaussian precursor are central in simulation and inference (Deiml et al., 19 Aug 2025, Lang et al., 2011). In latent truncation models, high-dimensional truncated normal probabilities and identifiability between presence probability and scale are explicit computational constraints (Allard et al., 2020). In spatio-temporal TGMRFs, positive definiteness is enforced through diagonal dominance of the precision structure, which restricts the parameter space for $Z_i=F_i^{-1}(\Phi(\varepsilon_i))$07 (Azevedo et al., 2020). In Lévy-subordinated models, discontinuous sample paths increase realism but reduce PDE solution regularity and favor adaptive over uniform discretizations (Merkle et al., 2022).

A plausible synthesis is that the enduring appeal of transformed Gaussian random fields lies in the separation they offer between Gaussian structure and application-specific departures from Gaussianity. Gaussian fields provide covariance, Markov, spectral, or geometric machinery; transformations impose boundedness, positivity, zero inflation, heavy tails, binary phase structure, anisotropy, non-separability, or rare-event emphasis. The result is a broad analytic framework in which Gaussian methods remain central even when the observable field is no longer Gaussian.