Gaussian Path Models Overview

• Gaussian path models are a framework employing Gaussian structures to characterize, condition, and optimize paths in continuous and discrete settings.
• They integrate covariance kernels and precision matrices to address sample path regularity, inference, and graph-theoretic constructs in high-dimensional spaces.
• Applications range from GP trajectory modeling in robotics to belief-space planning and fuzzy shortest-path formulations, enabling robust uncertainty quantification.

Gaussian path models comprise several distinct but mathematically related uses of Gaussian structure in path-centric problems. In one usage, a path is a sample path of a Gaussian process, and the central questions concern regularity, conditioning, and function-space inference. In another, a path is a graph-theoretic object inside a Gaussian graphical model, where covariances and marginals are expressed through simple paths and cycles. In robotics and planning, the phrase can refer either to trajectories represented directly by Gaussian processes or beliefs, or to deterministic paths optimized against Gaussian environment models. The term is therefore not uniform, and in the graphical-model setting it is explicitly distinct from path analysis in the SEM or causal sense (Costa et al., 2023, Giscard et al., 2014, Nguyen et al., 2021).

1. Terminological scope and common structure

The common core of Gaussian path models is that Gaussianity is attached to an object that determines, constrains, or evaluates paths. For Gaussian processes, the primitive object is a random function $f \sim \operatorname{GP}(\mu,k)$, with mean $\mu(x)$ and covariance kernel $k(x,y)$. For Gaussian graphical models, the primitive object is a Gaussian vector $X \sim \mathcal N(\mu,\Sigma)$ with precision matrix $J=\Sigma^{-1}$, whose sparsity pattern defines an undirected graph. For Gaussian belief-space planning, a path is a sequence in $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$, where each state is a Gaussian belief $b=(x,P)$. In fuzzy shortest-path formulations, edge costs are generalized Gaussian fuzzy numbers $\langle(c,\sigma);h\rangle$, where $h$ is interpreted as information reliability (Pedram et al., 2021, Moran et al., 26 May 2026).

This diversity creates a recurrent source of confusion. A Gaussian path model does not always mean a Gaussian distribution over trajectories. In some planning papers, the path itself is deterministic and the Gaussian object is instead an environment cost field, a communication model, or a belief state. In the graphical-model literature, “path” may denote simple paths and cycles of a precision graph rather than physical motion. A plausible implication is that the phrase is best read relationally: the path is the object of inference, optimization, or representation, while Gaussianity specifies the statistical or algebraic mechanism through which that object is handled.

2. Sample paths of Gaussian processes

For Gaussian processes, the foundational question is how the covariance kernel determines the regularity of sample paths. A centered process $f\sim \operatorname{GP}(0,k)$ has increment variance

$\mu(x)$0

which induces the canonical pseudometric $\mu(x)$1. The key result is a necessary-and-sufficient characterization of almost sure local almost-Hölder regularity: $\mu(x)$2 has sample paths in $\mu(x)$3 if and only if $\mu(x)$4 and the diagonal four-point increment condition for $\mu(x)$5 is $\mu(x)$6 for every $\mu(x)$7 and $\mu(x)$8. In the stationary case $\mu(x)$9, this reduces to regularity of the $k(x,y)$0-th derivatives of $k(x,y)$1 at $k(x,y)$2; in the isotropic case $k(x,y)$3, it reduces further to the one-dimensional behavior of $k(x,y)$4 near $k(x,y)$5 (Costa et al., 2023).

This perspective sharply separates sample-path regularity from RKHS regularity. The paper emphasizes that GP sample paths typically do not lie in the RKHS, and that local behavior near the diagonal, not global smoothness away from it, governs almost sure path smoothness. It also yields exact classifications for widely used kernels. A centered Matérn GP with parameter $k(x,y)$6 has sample paths in $k(x,y)$7 and no more; Wiener and Ornstein–Uhlenbeck paths are $k(x,y)$8; Wendland kernels yield $k(x,y)$9 and no more; and squared exponential, rational quadratic, periodic, linear, and polynomial constructions produce $X \sim \mathcal N(\mu,\Sigma)$0 sample paths under the stated conditions (Costa et al., 2023).

For trajectory modeling, this regularity theory gives direct control over the geometry of Gaussian paths. A coordinate-wise GP trajectory model writes

$X \sim \mathcal N(\mu,\Sigma)$1

with one GP per coordinate and Gaussian predictive marginals at each timestamp. Kernel choice then determines whether the path prior encodes rough motion, moderate smoothness, multi-scale variation, or linear drift. The practical point is that the kernel is not merely a covariance device; it is the path prior in a literal geometric sense (Nguyen et al., 2021).

3. Pathwise conditioning and kernel-imposed path structure

A second major development is the shift from finite-dimensional Gaussian conditioning to pathwise conditioning. For a GP prior $X \sim \mathcal N(\mu,\Sigma)$2 with observations $X \sim \mathcal N(\mu,\Sigma)$3 at inputs $X \sim \mathcal N(\mu,\Sigma)$4, the posterior can be represented exactly as

$X \sim \mathcal N(\mu,\Sigma)$5

which is the GP version of Matheron’s update rule. The posterior sample path is thus obtained by drawing a prior path and applying a deterministic update to the realized function values at the observation locations. This pathwise view separates posterior sampling into two parts: generation of a prior path and enforcement of the data through a kernel-based correction. It underlies scalable posterior-path approximations based on random Fourier features, truncated Karhunen–Loève expansions, SPDE constructions, sparse inducing variables, and iterative linear solvers (Wilson et al., 2020).

The same pathwise viewpoint supports exact structural constraints on Gaussian sample paths. For a centered Gaussian field $X \sim \mathcal N(\mu,\Sigma)$6 with covariance kernel $X \sim \mathcal N(\mu,\Sigma)$7, many almost-sure path properties are equivalent to covariance identities. In the composition-operator setting, $X \sim \mathcal N(\mu,\Sigma)$8 up to a modification if and only if

$X \sim \mathcal N(\mu,\Sigma)$9

For more general linear operators, the Gaussian case is handled through the RKHS and the Loève isometry $J=\Sigma^{-1}$0, which induces an operator $J=\Sigma^{-1}$1 on the RKHS and yields the equivalence

$J=\Sigma^{-1}$2

This framework makes path constraints into kernel-design constraints (Ginsbourger et al., 2013).

Concrete examples follow directly. Mean-centered paths satisfy

$J=\Sigma^{-1}$3

and can be constructed by the centered kernel transform

$J=\Sigma^{-1}$4

Additive paths are characterized by

$J=\Sigma^{-1}$5

which allows correlated univariate components, not only sums of independent one-dimensional kernels. ODE- and PDE-constrained paths are enforced by choosing kernels whose sections lie in the solution space of the homogeneous operator, as in

$J=\Sigma^{-1}$6

and

$J=\Sigma^{-1}$7

Because Gaussian process regression preserves these invariances in both posterior mean and posterior covariance, structural prior knowledge can be enforced at the kernel level and retained after conditioning (Ginsbourger et al., 2013).

4. Graph-theoretic path models in Gaussian graphical models

In Gaussian graphical models, path language takes a different meaning. The model is a Gaussian vector

$J=\Sigma^{-1}$8

with canonical density

$J=\Sigma^{-1}$9

The undirected graph $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$0 is tied to the sparsity of the precision matrix by

$\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$1

so exact marginal inference reduces to computing entries of $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$2. The path-sum formulation begins from the walk expansion with $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$3: $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$4 whenever the series converges, so covariance entries can be written as sums over walks on the graph (Giscard et al., 2014).

The central result is that infinite walk families can be resummed exactly into finite expressions over simple paths and simple cycles. Every covariance entry $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$5 has an exact path-sum representation involving only simple paths $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$6 and recursively dressed diagonal terms built from simple cycles $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$7. The resulting object is a branched continued fraction of finite depth and breadth. Crucially, this path-sum exists for every positive-definite $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$8, not only for walk-summable models satisfying $\mathbb{B}=\mathbb{R}^d\times\mathbb{S}_{++}^d$9. The method is therefore exact on arbitrary graph topologies, including loopy and non-walk-summable Gaussian graphical models (Giscard et al., 2014).

Two connections are especially important. On trees, path-sums reduce exactly to Gaussian belief propagation, recovering the Gaussian BP message-passing equations associated with Malioutov et al. On arbitrary partitions of $b=(x,P)$0, the same formulas remain valid for matrix-valued edge weights, so the method extends naturally to block and noncommutative settings where determinant-based formulas do not. The paper also recovers determinant formulas as a corollary and shows that path-sums can be evaluated recursively by enumerating simple paths and simple cycles on reduced subgraphs (Giscard et al., 2014).

The 5-cycle example makes the distinction from walk-summability concrete. For

$b=(x,P)$1

the model is positive definite for approximately $b=(x,P)$2, whereas walk-summability holds only for $b=(x,P)$3. Path-sums still return the exact inverse for values such as $b=(x,P)$4, where the Gaussian model is valid but the walk-sum is not absolutely convergent. In this literature, “Gaussian path model” therefore means a graph-theoretic path/cycle representation of Gaussian inference, not a causal path model (Giscard et al., 2014).

5. Trajectories, beliefs, and planning

When the path itself is a trajectory through physical space, Gaussian path models split into several subfamilies. In the most direct one, a trajectory is modeled coordinate-wise by Gaussian processes over time. For one coordinate $b=(x,P)$5, training data are $b=(x,P)$6, and the latent function satisfies

$b=(x,P)$7

At query timestamps $b=(x,P)$8, the posterior predictive distribution is Gaussian: $b=(x,P)$9 with the standard GP mean and covariance formulas. The method in the trajectory chapter uses one GP for longitude and one for latitude, assumes independence between them, and returns a timestamp-wise Gaussian distribution over location components. In this setting the path is genuinely stochastic, and the kernel determines whether the trajectory prior is rough, Matérn-like, RBF-smooth, multi-scale, or trend-driven (Nguyen et al., 2021).

A different construction treats the path as a curve in Gaussian belief space. A belief state is

$\langle(c,\sigma);h\rangle$0

and the path lies in $\langle(c,\sigma);h\rangle$1. The information-geometric distance is

$\langle(c,\sigma);h\rangle$2

where for lossless transitions

$\langle(c,\sigma);h\rangle$3

This $\langle(c,\sigma);h\rangle$4 is an asymmetric quasi-pseudometric and is interpreted as the minimum information gain required to steer the Gaussian belief. Collision avoidance is enforced through covariance-aware ellipsoidal constraints, and an RRT*-based algorithm searches for short, low-sensing-burden paths in the belief manifold (Pedram et al., 2021).

In informative path planning with Gaussian environmental models, the path is again not itself Gaussian; instead it is scored through a GP posterior over the environment. OLAh-GP optimizes a receding-horizon waypoint sequence

$\langle(c,\sigma);h\rangle$5

against the GP posterior that would result after sampling along all waypoints in the horizon. The propagated future variance

$\langle(c,\sigma);h\rangle$6

contains the full covariance block $\langle(c,\sigma);h\rangle$7, so correlations between future samples are modeled explicitly. For algal-bloom monitoring, the planning objective is based on the Gaussian threshold misclassification probability

$\langle(c,\sigma);h\rangle$8

summed over a grid and regularized by path length. This is a nonmyopic Gaussian-information design problem, not a GP prior over trajectories (Beaudin et al., 21 Apr 2026).

A further distinction arises in GP-based path optimization. In smooth robot navigation with a GP regression map, the Gaussian object is a continuous differentiable field of terrain traversability, obstacle distance, and posterior variance, while the path is a deterministic Bézier curve optimized by gradient descent. The loss combines path length, traversability penalty, variance penalty, obstacle penalty, and curvature penalty, and the differentiability of the GP map allows backpropagation through Bézier control points. Similarly, in planetary exploration and Gaussian sensor networks, GP or Gaussian communication models induce uncertainty or power–distortion fields that drive path selection. These formulations are path-relevant and Gaussian, but they are not probabilistic trajectory models in the strict sense (Serdel et al., 2024, Akemoto et al., 20 Mar 2025, Akyol et al., 2016).

6. Other path-space formulations

Several additional literatures use Gaussian path language in still broader senses. In normalized Gaussian path integrals, the central object is the transition density

$\langle(c,\sigma);h\rangle$9

where $h$0 is the conditional Euclidean or Wiener path integral and $h$1 is the unconditional normalization integral. The paper shows that both numerator and denominator can be computed from linear Hamiltonian Jacobi systems, yielding exact Gaussian kernels for quadratic actions such as the Ornstein–Uhlenbeck process and charged-string examples, and semiclassical approximations for nonlinear systems such as the Van der Pol oscillator driven by white noise (Corazza et al., 2020).

In large-deviation theory, the path may be a slow trajectory driven by a quadratic observable of a fast Gaussian process. For

$h$2

with $h$3 a centered stationary Gaussian process, the resulting slow path is not itself Gaussian. Nevertheless, the sample-path rate functional can be written explicitly through a Fredholm determinant and Widom asymptotics, leading to the effective action

$h$4

This is a path-space theory induced by Gaussian drivers after a nonlinear transformation (Bouchet et al., 2021).

Other uses are more indirect but still structurally important. Minimum-energy-path calculations can be accelerated by placing a GP surrogate on the energy surface and relaxing a deterministic path on the posterior mean surface, with derivative observations supplying local force information (Koistinen et al., 2017). In 3D Gaussian fields for inverse rendering, “path” means light-transport paths traced through an equivalent interaction model over overlapping Gaussian primitives, not physical robot or stochastic-process trajectories (Zhu et al., 8 Jun 2026). In fuzzy shortest-path models, edge costs are generalized Gaussian fuzzy numbers with membership

$h$5

and path ranking combines center, spread, and reliability through

$h$6

These examples show that Gaussian path models are best understood as a family of path-space constructions unified by Gaussian algebra, rather than a single canonical model class (Moran et al., 26 May 2026).

Gaussian path models therefore form a heterogeneous but coherent research area. Their unifying theme is that Gaussian structure—through kernels, precisions, beliefs, channel models, pathwise updates, or Gaussian-shaped uncertainty sets—makes path objects analytically tractable, optimizable, or exactly characterizable. The major distinctions are semantic and methodological: whether the path is a sample function, a graph path, a physical trajectory, a belief evolution, a light path, or a path integral; and whether Gaussianity belongs to the path itself, the field that generates it, or the uncertainty used to score it.