Gaussian Random Bridges

• Gaussian random bridges are Gaussian processes modified by endpoint or linear constraints, resulting in explicitly adjusted covariance structures.
• They incorporate spectral methods and canonical representations to reveal eigenvalue shifts and geometric covariance metrics.
• These bridges extend to models with random endpoints or lengths and offer applications in stochastic control, field theory, and generative modeling.

Searching arXiv for the cited bridge papers to comply with the requirement for fresh arXiv grounding. Searching arXiv for exact papers on Gaussian bridges, generalized Gaussian bridges, and random-length bridges. Gaussian random bridges are Gaussian processes obtained by imposing endpoint or more general linear path constraints on an underlying Gaussian process, or, in related constructions, by randomizing the pinning time or pinning value while preserving conditional bridge structure. In the classical case, conditioning Brownian motion on its terminal value produces the Brownian bridge; in the broader Gaussian setting, one obtains bridge laws by conditioning on $X_1=0$, on finitely many linear functionals, or on random terminal data, and the resulting processes remain Gaussian conditional on the imposed constraints (Chigansky et al., 2017, Sottinen et al., 2012, Erraoui et al., 2017). The topic spans operator-theoretic, geometric, and stochastic-control viewpoints, ranging from covariance rank-one perturbations and canonical representations to random-endpoint Brownian bridges, constrained Ornstein–Uhlenbeck bridges, and endpoint-pinned Gaussian fields on domains (Chigansky et al., 2017, Mazzolo, 2017, Sandrić et al., 23 Jun 2026, Görgens et al., 2014).

1. Definition and basic structure

A centered Gaussian bridge is obtained from a centered Gaussian process $X=(X_t,\ t\in[0,1])$ by conditioning on an endpoint value. In the zero-to-zero case with $X_0=0$, the bridge is

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$

where $K(s,t)=\mathbb E[X_sX_t]$ is the covariance kernel (Chigansky et al., 2017). Its covariance is

$$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$

so the bridge covariance operator is a rank-one perturbation of the base covariance operator (Chigansky et al., 2017).

In the classical Brownian case, Brownian motion on $[0,1]$ has covariance $K(s,t)=s\wedge t$, while the Brownian bridge has covariance

$\widetilde K(s,t)=s\wedge t-st.$

Its eigenpairs are explicit: $$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$ in contrast with Brownian motion,

$X=(X_t,\ t\in[0,1])$0

(Chigansky et al., 2017).

A broader Hilbert-space formulation defines a generalized Gaussian bridge as the regular conditional law of a continuous Gaussian process $X=(X_t,\ t\in[0,1])$1 under finitely many linear constraints

$X=(X_t,\ t\in[0,1])$2

This subsumes ordinary endpoint pinning, multibridges pinned at several times, and constraints such as time averages (Sottinen et al., 2012). In that framework, the bridge is defined as a conditional Gaussian law on path space rather than first by a pathwise formula (Sottinen et al., 2012).

2. Covariance geometry and canonical pseudo-metrics

For Brownian bridge $X=(X_t,\ t\in[0,1])$3, the covariance is

$X=(X_t,\ t\in[0,1])$4

and the canonical pseudo-metric is

$X=(X_t,\ t\in[0,1])$5

This identifies the Brownian bridge geometry encoded by second moments (Jin, 17 Mar 2025).

That metric contrasts with Brownian motion, whose canonical metric is $X=(X_t,\ t\in[0,1])$6. The extra factor $X=(X_t,\ t\in[0,1])$7 is the signature of the bridge constraint (Jin, 17 Mar 2025). The same geometry appears in high-dimensional random bridge limits: in the square integrable case, the limit as a metric space is deterministic and equals $X=(X_t,\ t\in[0,1])$8 equipped with the pseudo-metric

$X=(X_t,\ t\in[0,1])$9

in the Gromov–Hausdorff sense (Jin, 17 Mar 2025). A plausible implication is that bridge conditioning can be identified geometrically even when the underlying convergence is formulated at the level of metric spaces rather than finite-dimensional Gaussian laws.

This covariance geometry also underlies measure-indexed constructions. In the generalized random field approach, Brownian motion is extracted from

$X_0=0$0

using

$X_0=0$1

while the Brownian bridge on $X_0=0$2 is obtained by replacing $X_0=0$3 with

$X_0=0$4

so that $X_0=0$5 is built into the indexing measure (Görgens et al., 2014). In one dimension, the membrane construction $X_0=0$6 reduces to exactly this bridge measure family because harmonic measure on $X_0=0$7 is $X_0=0$8 (Görgens et al., 2014).

3. Generalized conditioning and representation theory

The orthogonal representation of a generalized Gaussian bridge is

$X_0=0$9

with bridge covariance

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$0

(Sottinen et al., 2012). This is the direct Gaussian projection formula, but it is anticipative because it depends on the full future path through $$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$1 (Sottinen et al., 2012).

The canonical representation aims to avoid future information. For Brownian bridge, the noncanonical orthogonal form

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$2

is replaced by the canonical form

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$3

(Sottinen et al., 2012). More generally, for continuous Gaussian martingales and then for prediction-invertible Gaussian processes, one obtains filtration-preserving bridge representations via Volterra-type transforms and resolvent kernels (Sottinen et al., 2012).

This distinction between orthogonal and canonical constructions is central. The orthogonal representation is universal for continuous Gaussian processes; the canonical representation requires additional structure but preserves the information generated by the process up to time $$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$4 (Sottinen et al., 2012). In the semimartingale case, this connects directly to enlargement of filtration and information drift (Sottinen et al., 2012).

A concrete Gaussian example is the constrained Ornstein–Uhlenbeck bridge, where the process is conditioned on both endpoint and area: $$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$5 Because the underlying process is Gaussian and the constraints are linear functionals, the conditioned process remains Gaussian (Mazzolo, 2017). The anticipative representation has the explicit regression form

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$6

with explicit $$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$7, $$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$8, mean, and variance formulas (Mazzolo, 2017). A non-anticipative representation is obtained by augmenting the state with accumulated area and applying stochastic control, yielding an adapted SDE and an effective Langevin equation (Mazzolo, 2017). This is a prototypical generalized Gaussian bridge conditioned on two linear observables.

4. Spectral theory and bridge eigenproblems

The covariance operator of a Gaussian bridge differs from that of its base process by a rank-one perturbation, and bridge eigenvalues satisfy the exact transcendental equation

$$\widetilde X_t=X_t-\frac{K(t,1)}{K(1,1)}X_1,\qquad t\in[0,1],$$9

while the corresponding eigenfunctions are

$K(s,t)=\mathbb E[X_sX_t]$0

(Chigansky et al., 2017). This shows that the bridge spectrum is not determined by the base eigenvalues alone; one also needs endpoint information $K(s,t)=\mathbb E[X_sX_t]$1 (Chigansky et al., 2017).

For fractional Brownian motion with Hurst parameter $K(s,t)=\mathbb E[X_sX_t]$2, the bridge eigenvalues satisfy

$K(s,t)=\mathbb E[X_sX_t]$3

with second-order frequency shift

$K(s,t)=\mathbb E[X_sX_t]$4

(Chigansky et al., 2017). In the Brownian case $K(s,t)=\mathbb E[X_sX_t]$5, this reduces to the exact shift $K(s,t)=\mathbb E[X_sX_t]$6 (Chigansky et al., 2017).

The bridge eigenfunctions retain the oscillatory bulk plus boundary-layer structure of the base eigenfunctions, but the endpoint-conditioning modifies the boundary-layer coefficient so that the functions vanish at both ends (Chigansky et al., 2017). A plausible implication is that endpoint pinning is spectrally small in rank but not small in second-order asymptotics.

The spectral viewpoint also appears in nonclassical bridge constructions. The fractional Wiener–Weierstrass bridge is

$K(s,t)=\mathbb E[X_sX_t]$7

where $K(s,t)=\mathbb E[X_sX_t]$8 is a fractional Brownian bridge and

$K(s,t)=\mathbb E[X_sX_t]$9

competes with the Hurst parameter $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$0 in determining roughness (Schied et al., 2024). This process is centered Gaussian, pinned at $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$1 and $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$2, but has highly non-stationary increments and fractal covariance structure (Schied et al., 2024). Its sample path properties split into regimes $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$3, $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$4, and $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$5, with exact local and uniform modulus statements, $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$6-variation, Hausdorff dimension, and atomlessness of the maximum location depending on that trichotomy (Schied et al., 2024). In the regime $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$7, the exact local and uniform modulus constants become random rather than deterministic (Schied et al., 2024).

5. Random endpoints and random lengths

A Gaussian random bridge need not be pinned at a deterministic terminal value. A particularly explicit example is

$$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$8

where $$\widetilde K(s,t)=K(s,t)-\frac{K(s,1)K(t,1)}{K(1,1)},$$9 is a standard planar Brownian bridge from $[0,1]$0 to $[0,1]$1, and $[0,1]$2 is independent (Sandrić et al., 23 Jun 2026). Then

$[0,1]$3

so the terminal value is Gaussian with law $[0,1]$4, and conditionally on $[0,1]$5,

$[0,1]$6

is a planar Brownian bridge from $[0,1]$7 to $[0,1]$8 (Sandrić et al., 23 Jun 2026). The covariance is

$[0,1]$9

which interpolates between Brownian bridge at $K(s,t)=s\wedge t$0 and Brownian motion at $K(s,t)=s\wedge t$1 (Sandrić et al., 23 Jun 2026).

This random-endpoint model remains centered and isotropic, and the paper computes exact geometric functionals of its convex hull, including

$K(s,t)=s\wedge t$2

with $K(s,t)=s\wedge t$3, and

$K(s,t)=s\wedge t$4

(Sandrić et al., 23 Jun 2026). The one-dimensional projected maximum and argmax law are also explicit (Sandrić et al., 23 Jun 2026).

A different randomization concerns the pinning time. For a centered continuous Gaussian process $K(s,t)=s\wedge t$5 and a strictly positive random time $K(s,t)=s\wedge t$6, independent of $K(s,t)=s\wedge t$7, the Gaussian bridge with random length is

$K(s,t)=s\wedge t$8

so

$K(s,t)=s\wedge t$9

(Erraoui et al., 2017). If the starting process is Gaussian Markov, then the random-length bridge is again Markov with respect to its natural filtration and, under an additional mild condition, with respect to the usual right-continuous augmentation (Erraoui et al., 2017). This yields right-continuity and completeness of the completed natural filtration (Erraoui et al., 2017). In the Brownian specialization, this recovers the Brownian bridge on a random interval (Erraoui et al., 2017).

A Lévy-process formulation proves analogous results for bridges with random length in the symmetric Lévy setting, with Brownian motion as the Gaussian special case (Erraoui et al., 2019). A plausible implication is that randomization of terminal time preserves bridge-type conditional structure more robustly than Gaussianity itself, since the unconditional law becomes a mixture when one averages over $\widetilde K(s,t)=s\wedge t-st.$0.

6. Geometric, field-theoretic, and applied directions

Gaussian bridges also arise as objects extracted from selfsimilar Gaussian generalized random fields. In that setting, one starts with a measure-indexed field $\widetilde K(s,t)=s\wedge t-st.$1 and recovers pointwise bridges by choosing signed measures designed to enforce endpoint pinning. Brownian bridge corresponds to

$\widetilde K(s,t)=s\wedge t-st.$2

fractional Brownian bridge corresponds to

$\widetilde K(s,t)=s\wedge t-st.$3

and generalized bridges conditioned on $\widetilde K(s,t)=s\wedge t-st.$4 are represented by replacing $\widetilde K(s,t)=s\wedge t-st.$5 with

$\widetilde K(s,t)=s\wedge t-st.$6

(Görgens et al., 2014). In higher dimensions, the same logic yields Gaussian membranes vanishing on the boundary of a domain via $\widetilde K(s,t)=s\wedge t-st.$7, where $\widetilde K(s,t)=s\wedge t-st.$8 is harmonic measure (Görgens et al., 2014).

Bridge-like Gaussian structures also appear in random matrix asymptotics. Deterministic truncations of Haar unitary or orthogonal matrices yield a bivariate tied-down Brownian bridge in the limit, while Bernoulli random truncations produce the Gaussian field

$\widetilde K(s,t)=s\wedge t-st.$9

built from two independent one-parameter Brownian bridges (Beffara et al., 2013, Donati-Martin et al., 2013). The deterministic limit has covariance

$$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$0

whereas the random-truncation limit has covariance

$$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$1

(Beffara et al., 2013). This suggests that bridge structure can persist under subordination but change its form when the dominant randomness comes from the indexing mechanism rather than the underlying field.

A more recent applied direction treats random bridges as stochastic transports for generative models. In the Gaussian case, if $$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$2 is a Gaussian driver and $$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$3 is a random terminal variable, the initialized Gaussian random bridge has anticipative representation

$$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$4

with conditional Gaussian law

$$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$5

(Goria et al., 16 Dec 2025). In the Brownian case, the bridge SDE becomes

$$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$6

so the drift is a state-dependent reversion toward the posterior estimate of the target (Goria et al., 16 Dec 2025). The paper reports low-step sample quality competitive with diffusion baselines on MNIST and CIFAR-10 (Goria et al., 16 Dec 2025). This suggests that Gaussian random bridge conditioning can be recast as a controlled information-flow mechanism on path space.

In Schrödinger bridge theory, exact Gaussian-to-Gaussian Brownian Schrödinger bridges are likewise Gaussian random bridges with explicit time marginals, affine drift, and cost formulas (Ganguly et al., 24 May 2026). For Gaussian-mixture endpoints, the lifted formulation decomposes the problem into pairwise Gaussian bridges $$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$7 and a finite-dimensional entropic coupling $$\widetilde\lambda_n=\frac{1}{(\pi n)^2},\qquad \widetilde\varphi_n(t)=\sqrt2\sin(\pi n t),$$8, after which the projected marginal flow is a Gaussian-mixture interpolation with posterior-averaged drift (Ganguly et al., 24 May 2026). The paper identifies a nonnegative conditional label-information gap under projection, showing that the lifted optimizer generally does not coincide with the direct unlabeled Schrödinger bridge (Ganguly et al., 24 May 2026). A plausible implication is that Gaussian bridge tractability survives at the component level but not automatically at the mixture level after marginalization.

Gaussian random bridges therefore form a broad class rather than a single construction. At one end lie classical Brownian and fractional Brownian bridges, with exact covariance and spectral descriptions; at the other lie generalized bridges conditioned on multiple linear observables, random-endpoint and random-length bridges, selfsimilar field extractions, and componentwise Gaussian bridge decompositions in modern transport and generative settings (Chigansky et al., 2017, Sottinen et al., 2012, Mazzolo, 2017, Görgens et al., 2014, Erraoui et al., 2017, Sandrić et al., 23 Jun 2026, Goria et al., 16 Dec 2025, Ganguly et al., 24 May 2026). The common structural feature is that Gaussian conditioning yields explicit or semi-explicit control of covariance, filtration, and geometry, even when the resulting path law is embedded in a substantially richer stochastic or geometric framework.