Prior-Aligned Anisotropic Gaussian Distribution

• Prior-Aligned Anisotropic Gaussian Reference Distribution is defined by aligning its covariance with domain-specific structures, making it more effective than isotropic models in interpreting geometric information.
• The approach models scene elements as 2D Gaussian disks on local tangent planes, which improves registration accuracy and depth interpretation in SLAM via geometry-aware optimization.
• It unifies various applications—from SLAM and SPDEs to GP kernels and Bayesian anisotropy estimation—demonstrating empirical improvements in performance and numerical conditioning.

A prior-aligned anisotropic Gaussian reference distribution is a Gaussian construction in which covariance is explicitly aligned with domain structure rather than treated as isotropic or merely axis-aligned. In "G2S-ICP SLAM" (Pak et al., 24 Jul 2025), the term denotes a geometry-aware RGB-D Gaussian Splatting SLAM representation in which each scene element is modeled as a Gaussian “disk” constrained to the local tangent plane, and the resulting anisotropic covariance prior is injected into Generalized ICP and rendering. In adjacent literatures, closely related constructions align anisotropy with physics-informed simulations, base models that penalize deviation from isotropy, explicit $SO(3)$ rotations, or reference images, indicating that the phrase names a family of covariance-alignment strategies rather than a single universal probabilistic object (Berild et al., 2023, Llamazares-Elias et al., 2024, Warrior et al., 11 May 2026, Lv et al., 2019).

1. General definition and covariance geometry

In its most explicit form, the construction separates orientation from principal scales. In G2S-ICP SLAM, each primitive has covariance

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$

so the Gaussian support is finite along tangent directions and zero, or near-zero for numerical stability, along the normal (Pak et al., 24 Jul 2025). In a rotationally anisotropic GP kernel, the same idea appears as

$$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$

with $R=\exp([\omega]_\times)\in SO(3)$ from an axis-angle parameterization, while an aligned Gaussian reference covariance is written

$$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$

(Warrior et al., 11 May 2026). In non-stationary anisotropic SPDE models, anisotropy is encoded through

$$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$

which determines local principal dependence directions and strengths (Berild et al., 2023).

This suggests a common abstraction: the reference distribution is not defined by Gaussianity alone, but by how its covariance or precision is oriented with respect to geometry, transport, correlation directions, or structural priors. The “prior-aligned” qualifier is therefore substantive. In the oceanographic SPDE setting, the GRF acts as a prior distribution whose anisotropy $H(s)$ and $\kappa(s)$ is estimated from external physics-informed simulations, so the reference distribution is aligned with dominant transport directions and spatial variability implied by the ocean model (Berild et al., 2023). In penalized-complexity formulations, the base model is the spatially constant Gaussian field $M_0=N(0,1\otimes1)$, and the prior shrinks correlation range toward infinity and anisotropy toward zero unless data support added complexity (Llamazares-Elias et al., 2024).

2. Tangent-plane Gaussian disks in G2S-ICP SLAM

In G2S-ICP SLAM, the scene is represented by Gaussian disks constrained to local tangent planes rather than volumetric isotropic 3D Gaussians. The reference distribution for element $i$ is

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$0

where $$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$1 is the Gaussian center and $$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$2 is aligned with the local tangent plane. The disk-to-world mapping is

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$3

equivalently expressed with

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$4

and the Gaussian kernel on the disk is

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$5

This formulation models the local surface as a 2D Gaussian disk aligned with the underlying geometry, leading to more consistent depth interpretation across multiple viewpoints compared to conventional 3D ellipsoid-based representations with isotropic uncertainty (Pak et al., 24 Jul 2025).

The map is estimated from RGB-D frames. Pixels are back-projected to 3D points $$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$6, a local neighborhood $$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$7 is formed, and the sample covariance

$$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$8

is analyzed by PCA. The eigenvectors $$\Sigma_i = R_i S_i S_i^\top R_i^\top,\qquad R_i = [t_i^1,t_i^2,n_i],\qquad S_i = \operatorname{diag}(s_i^1,s_i^2,0),$$9 and eigenvalues $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$0 determine $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$1, $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$2, $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$3, and $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$4 or the point itself depending on construction policy. During optimization, normals are additionally supervised from depth gradients via

$$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$5

which refines the local orientation estimate (Pak et al., 24 Jul 2025).

The anisotropic covariance prior is then written as

$$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$6

with $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$7, $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$8, and $$M = R \Lambda_\ell R^\top,\qquad \Lambda_\ell=\operatorname{diag}(\ell_1^{-2},\ell_2^{-2},\ell_3^{-2}),$$9. In the implementation, the scale regularization sets the normal scale to zero, $R=\exp([\omega]_\times)\in SO(3)$0, so $R=\exp([\omega]_\times)\in SO(3)$1; for numerical stability and invertibility in GICP, a tiny $R=\exp([\omega]_\times)\in SO(3)$2 may be used in practice, but the intended geometric prior is $R=\exp([\omega]_\times)\in SO(3)$3 and $R=\exp([\omega]_\times)\in SO(3)$4 (Pak et al., 24 Jul 2025).

3. Embedding the anisotropic prior in Generalized ICP

The core SLAM role of the prior-aligned anisotropic Gaussian reference distribution is registration. G2S-ICP SLAM performs frame-to-model registration using GICP, with each point represented by a Gaussian reference distribution $R=\exp([\omega]_\times)\in SO(3)$5 aligned to the surface. For source set $R=\exp([\omega]_\times)\in SO(3)$6 and target set $R=\exp([\omega]_\times)\in SO(3)$7, the residual for a correspondence $R=\exp([\omega]_\times)\in SO(3)$8 is

$R=\exp([\omega]_\times)\in SO(3)$9

with pose $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$0. Under the standard GICP assumption, the residual covariance is

$$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$1

and the maximum-likelihood objective is

$$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$2

Because $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$3, $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$4 preferentially weights the normal component of residuals; as $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$5, the objective approximates point-to-plane ICP with

$$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$6

Compared to isotropic ellipsoids, the anisotropic, plane-aligned prior makes $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$7 higher for normal-aligned residuals and lower for tangential components, biasing the optimizer to correct depth misalignments first (Pak et al., 24 Jul 2025).

The same prior also changes numerical conditioning. Optimization is performed via Gauss-Newton or Levenberg-Marquardt on $$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$8, with anisotropic weighting altering the normal equations

$$\Sigma = R\operatorname{diag}(s_1^2,s_2^2,s_3^2)R^\top$$9

The paper states that this improves conditioning by suppressing degenerate directions typical for planar correspondences when isotropic covariances are used (Pak et al., 24 Jul 2025).

The tangential scales are depth-dependent:

$$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$0

where $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$1 is the depth of $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$2 along the camera ray. This yields improved visual uniformity and supports stable ICP weighting. Tracking uses GICP with Mahalanobis objective; correspondences are based on nearest neighbor search in the map; tracking keyframes are selected by geometric correspondence ratio; and mapping keyframes are inserted at fixed intervals (Pak et al., 24 Jul 2025).

4. Rendering, normal supervision, and geometry-aware optimization

The same anisotropic reference distribution is used for rendering. In image space, the contribution of Gaussian $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$3 at pixel $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$4 is weighted by

$$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$5

where $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$6 is obtained by linearizing the camera projection of the 3D covariance:

$$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$7

Because $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$8 has negligible support along $$H(s)=\gamma(s)I_3+v(s)v(s)^\top+\omega(s)\omega(s)^\top,$$9, the image-space ellipse aligns with the projected tangent directions, preventing depth bleeding characteristic of volumetric, isotropic ellipsoids. Front-to-back compositing yields

$H(s)$0

$H(s)$1

and surface alignment improves multi-view depth interpretability because the disk intersects view rays consistently, unlike volumetric ellipsoids whose intersections change with viewpoint (Pak et al., 24 Jul 2025).

Mapping refines Gaussian attributes with a geometry-aware loss that enforces consistency of appearance, depth, and normals. The photometric, depth, and normal terms are

$H(s)$2

$H(s)$3

$H(s)$4

The Geometry-Aware Normal loss is

$H(s)$5

and the total loss is

$H(s)$6

with $H(s)$7, $H(s)$8, and $H(s)$9 (Pak et al., 24 Jul 2025).

A recurrent misconception is that anisotropy in this setting is only a rendering heuristic. In G2S-ICP SLAM, the same covariance prior is simultaneously a rendering primitive, a registration weight model, and a geometric regularizer. The paper’s formulation explicitly states that the anisotropic covariance prior is used both for rendering and for scan registration in GICP, and that a scale regularization explicitly enforces $\kappa(s)$0 so that $\kappa(s)$1 in $\kappa(s)$2 (Pak et al., 24 Jul 2025).

5. Empirical behavior, assumptions, and limitations

The reported empirical evidence is specific. On Replica, G2S-ICP SLAM achieves “average Depth L1 = 0.74 cm versus GS-SLAM’s 1.16 cm and GS-ICP SLAM’s 20.54 cm (reproduced), with the highest Precision, Recall, and F1-score across scenes.” For tracking, “ATE RMSE on Replica averages 0.15 cm, outperforming or matching GS-ICP SLAM and other baselines, and competitive results on TUM-RGBD.” In ablation, “replacing 3D ellipsoids with 2D disks alone reduces rendered depth L1 from 4.18 cm to 2.08 cm and improves ATE; adding geometry-aware optimization further reduces depth L1 to 0.44 cm.” The system also reports “real-time speed (≈30 FPS) with competitive PSNR/SSIM/LPIPS compared to GS-ICP SLAM, with improved depth error maps and sharper geometry” (Pak et al., 24 Jul 2025).

These results are tied to a set of explicit assumptions. The Gaussian disk prior assumes locally planar patches. Highly curved or thin structures may violate this assumption; however, the small support $\kappa(s)$3 and normal supervision mitigate misalignment. Errors in normals from noisy depth, such as TUM-RGBD, can misalign disks, while geometry-aware normal loss and robust mapping across keyframes reduce drift. Textureless or repetitive regions remain challenging for ICP, although the anisotropic prior improves conditioning by downweighting tangential ambiguity. Real-world motion blur and missing depth also introduce instability, but the paper states that depth and normal losses plus distance-aware scaling help stabilize optimization while preserving real-time speed (Pak et al., 24 Jul 2025).

The comparison with isotropic or unaligned covariances clarifies the scope of the claimed benefit. Conventional ICP/GICP often uses isotropic covariances or unaligned ellipsoids, which in multi-view SLAM with 3DGS lead to ambiguity along surface normals and inconsistent slicing. By contrast, the prior-aligned anisotropic $\kappa(s)$4 reduces ambiguity along $\kappa(s)$5 by imposing $\kappa(s)$6 and ensures consistent multi-view interpretation by constraining support to the tangent plane (Pak et al., 24 Jul 2025).

6. Broader statistical, kernel, and Bayesian interpretations

Outside SLAM, the same conceptual template recurs with different semantics for “reference distribution.” In non-stationary anisotropic GRFs, the prior over a 3D field is $\kappa(s)$7, $\kappa(s)$8, where $\kappa(s)$9 comes from an SPDE with spatially varying $M_0=N(0,1\otimes1)$0 and $M_0=N(0,1\otimes1)$1 estimated from dense SINMOD simulations; this yields a Gaussian reference distribution whose anisotropy and variance are aligned with physical transport (Berild et al., 2023). In the PC-prior framework for stationary anisotropic Gaussian fields, the reference model is the spatially constant Gaussian field $M_0=N(0,1\otimes1)$2, and the prior penalizes the distance $M_0=N(0,1\otimes1)$3 so that $M_0=N(0,1\otimes1)$4 is shrunk toward $M_0=N(0,1\otimes1)$5 and $M_0=N(0,1\otimes1)$6 is shrunk toward $M_0=N(0,1\otimes1)$7, favoring infinite correlation range and isotropy unless evidence supports otherwise (Llamazares-Elias et al., 2024). In approximate reference-prior methodology for Gaussian random fields, “reference prior” refers instead to a spectrally approximated default prior over covariance parameters, and the marginal approximate reference prior of the correlation parameter is always proper (Oliveira et al., 2022).

A second recurring theme is explicit parameterization of orientation. The rotationally anisotropic GP kernel parameterizes a three-dimensional SPD covariance metric using three principal length-scales and an explicit $M_0=N(0,1\otimes1)$8 rotation represented by an axis-angle vector, so that orientation and principal scales are directly available for prior specification and posterior summaries (Warrior et al., 11 May 2026). Bhattacharya, Pati, and Dunson formulate an anisotropic function-estimation prior through a Gaussian process with dimension-specific scalings and show that using a homogeneous Gaussian process with a single bandwidth leads to a sub-optimal rate in anisotropic cases (Bhattacharya et al., 2011). This suggests that “prior alignment” is not only geometric; it is also asymptotic, because the prior mass must be concentrated on scale configurations compatible with unknown anisotropic smoothness.

A third theme is reference-driven anisotropy from data or auxiliary structure. The NLTG prior defines a probability measure absolutely continuous with respect to a Gaussian reference measure $M_0=N(0,1\otimes1)$9, where the Gaussian reference measure provides a flexibility of incorporating structure information from a reference image, and the covariance $i$0 induces a nonlocal, reference-driven anisotropy (Lv et al., 2019). In non-parametric anisotropy estimation for 2D differentiable Gaussian random fields, the approximate joint density $i$1 serves as a model-free “reference” distribution for anisotropy parameters and provides conservative probability and confidence regions, informed initial values for maximum likelihood estimation, and a useful prior for Bayesian anisotropy inference (Petrakis et al., 2012).

Taken together, these formulations support a narrow but robust interpretation. A prior-aligned anisotropic Gaussian reference distribution is a Gaussian object whose covariance, precision, or parameter prior is deliberately aligned with an external structural hypothesis: local tangent planes in SLAM, transport fields in SPDE priors, explicit rotations in GP kernels, similarity graphs in imaging, or anisotropy statistics in reference distributions. What changes across domains is the role of the Gaussian object—scene primitive, latent field prior, kernel metric, reference measure, or parameter prior—while the governing principle remains covariance alignment under anisotropy [(Pak et al., 24 Jul 2025); (Berild et al., 2023); (Llamazares-Elias et al., 2024); (Warrior et al., 11 May 2026); (Lv et al., 2019); (Petrakis et al., 2012)].