Discrete Intrinsic Gaussian Processes

• Discrete intrinsic Gaussian processes are stochastic models defined on discrete domains, incorporating state-dependent noise and intrinsic geometric structures for enhanced modeling of non-Euclidean data.
• Markov-chain constructions, Gaussian approximations, and discrete operator methods underpin these processes, providing robust frameworks for capturing dynamics in population models, graphs, and manifolds.
• Applications span from finite-population dynamics and time-series analysis to manifold learning and vector field inference, demonstrating improved precision and uncertainty quantification compared to extrinsic methods.

A discrete intrinsic Gaussian process is a stochastic process defined on a discrete domain whose statistical properties are determined by internal mechanisms—such as local Markovian transition, difference operators, or intrinsic geometric structures—rather than externally imposed noise. These models encompass state-dependent, multiplicative noise and exploit discrete analogs of intrinsic geometry or system dynamics. They provide a principled framework for modeling stochasticity and dependence in discrete, non-Euclidean, or structured spaces, including population models, graphs, point clouds near manifolds, and meshes on surfaces.

1. Markov-Chain Constructions and Finite-Population Dynamics

The archetype of a discrete intrinsic Gaussian process in population dynamics arises from embedding deterministic iterative maps into stochastic Markov chains whose large-population ($N\to\infty$) limit recovers the original map. For a mono-type population of size $N$ with $n$ individuals of type $A$ at generation $t$, the state probability vector $\bm P_t$ evolves via multinomial sampling:

$$\bm P_{t+1} = \bm Q \bm P_t, \quad Q_{nm} = \binom{N}{n} [f(m/N)]^n [1-f(m/N)]^{N-n}$$

where $f:[0,1]\to[0,1]$ is a deterministic map encoding selection or interaction. The process is constructed so that the first moment satisfies:

$$\langle n_{t+1}\rangle = N \langle f(n_t/N) \rangle$$

In the deterministic ($N\to\infty$) limit, rescaled means $N$0 obey the one-dimensional recursion $N$1, with all randomness vanishing. This construction yields a classically deterministic map perturbed by intrinsic, state-dependent noise for finite $N$2 (Challenger et al., 2013).

2. Gaussian Approximations and Stochastic Difference Equations

For large but finite $N$3, the binomial structure justifies a Gaussian approximation of transition probabilities. In rescaled variables $N$4, evolution becomes:

$N$5

yielding

$N$6

with

$N$7

Noise $N$8 is intrinsic, multiplicative, and temporally uncorrelated. The construction is valid for $N$9 and in regimes away from chaos or bifurcation, where a single-peaked Gaussian remains a good approximation (Challenger et al., 2013). This discrete-time, state-dependent difference equation exemplifies the defining features of a discrete intrinsic Gaussian process.

3. Discrete Intrinsic Gaussian Processes on Graphs and Manifolds

Discrete intrinsic Gaussian processes extend beyond population dynamics to structured spaces such as graphs or manifolds, where the underlying geometry is encoded intrinsically. For the vertex set $n$0 of a finite graph $n$1 with combinatorial Laplacian $n$2, one defines a zero-mean Gaussian process with covariance

$n$3

where $n$4 is the eigenbasis of $n$5 and $n$6 is a spectral filter (e.g., heat kernel $n$7 or Matérn-type $n$8) (Terenin, 2022). Such kernels intrinsically respect the graph's geometry and admit interpretation as the solution to a discrete SPDE:

$n$9

with $A$0 white noise, resulting in $A$1 and covariance $A$2.

In point cloud data sampled from unknown manifolds, the intrinsic approach involves learning a probabilistic embedding and metric tensor via Bayesian GPLVM, then simulating Riemannian Brownian motion and heat kernels to define the Gaussian process prior directly on the discrete point set. This procedure yields a kernel that captures manifold geometry and boundary without reliance on extrinsic distances (Niu et al., 2023).

4. Spline-Based, Difference-Operator, and Increment Models

Another major formulation arises from the discrete-time theory of Gaussian CARMA (Continuous-time AutoRegressive Moving Average) processes sampled onto discrete lattices. Consider a continuous-time process $A$3 governed by

$A$4

with $A$5 white Gaussian noise, where $A$6 and $A$7 are differential operators. After discretization, the process $A$8 satisfies

$A$9

where $t$0 is the discrete AR polynomial and $t$1 is discrete white noise. Taking $t$2-th order generalized increments

$t$3

with

$t$4

yields a stationary intrinsic Gaussian increment process whose covariance is completely characterized by a minimally-supported exponential B-spline $t$5:

$t$6

and

$t$7

The intrinsic process $t$8 is uniquely defined (up to a null space; e.g., polynomial drift for differences), and its increments $t$9 are strictly stationary (Unser et al., 2011).

5. Geometric and Topological Extensions: Meshes and Vector Fields

On triangulated meshes representing 2D discrete manifolds, intrinsic Gaussian processes leverage the discrete Laplace-Beltrami operator, discrete exterior calculus (DEC), and vector field decompositions. The core elements are:

• Construction of the cotangent-weight Laplacian $\bm P_t$0 and mass matrix $\bm P_t$1
• Spectral decomposition $\bm P_t$2 to obtain eigenmodes
• Bases for gradient (curl-free), rotated gradient (divergence-free), and harmonic (Hodge Laplacian nullspace) vector fields
• Matérn-type spectral filtering and assembly of block covariances for vector fields:

$\bm P_t$3

where the $\bm P_t$4-terms correspond to the contributions from curl-free, divergence-free, and harmonic modes, with weights set by the local geometry and the explicit spectrum of the mesh Laplacian (Gillan et al., 26 Jul 2025).

Table: Key Constructions of Discrete Intrinsic Gaussian Processes

Context	Core Operator/Structure	Innovation/Intrinsic Feature
Population dynamics	Markov chain (multinomial)	Finite-$\bm P_t$5 intrinsic binomial noise
Temporal processes	Difference operator, B-spline	Nth-order increments, stationary ARMA
Graphs/manifolds	Laplacian (graph/Riemannian/mesh)	Spectral filter (heat/Matérn kernel)
Point clouds	Probabilistic metric/BGPLVM	Brownian motion/heat kernel on learned manifold
Vector fields on meshes	Discrete exterior calculus	Matrix- and DEC-structured covariances

6. Regimes of Validity, Extensions, and Empirical Performance

The Gaussian approximation—central to most discrete intrinsic constructions—requires large populations or sample sizes, non-chaotic dynamical regimes (negative Lyapunov exponent or $\bm P_t$6), and parameters well-separated from bifurcations or critical points. In vector and graph-based models, validity is determined by mesh or graph resolution and faithful representation of geometry via intrinsic Laplacians and metric tensors.

Empirical studies demonstrate that intrinsic GP constructions, particularly those using manifold or mesh geometry, outperform naive Euclidean and graph-Laplacian Gaussian processes in sparse-label, non-Euclidean, or high-dimensional settings (e.g., manifold regression, WiFi localization, and vector field downscaling), with lower RMSE and robust uncertainty quantification in regions reflecting intrinsic geometry (Niu et al., 2023, Gillan et al., 26 Jul 2025).

7. Relation to Broader Gaussian Process Theory

Discrete intrinsic Gaussian processes are distinct from generic discrete-domain GPs by virtue of their state-dependent, geometry-driven, or system-encoded covariance structure. In contrast to extrinsic noise models, the stochasticity arises from internal combinatorial (e.g., multinomial, binomial), difference, or geometric mechanisms—respecting the underlying discrete space or process structure. The intrinsic approach is unifying in connecting population genetics, time-series, manifold learning, and vector field inference within a principled stochastic framework, often through explicit linkage to continuous analogues and discrete operator theory (Challenger et al., 2013, Unser et al., 2011, Terenin, 2022, Niu et al., 2023, Gillan et al., 26 Jul 2025).