Random Bridges in Stochastic Processes

• Random bridges are conditioned stochastic paths defined by imposing endpoint constraints, uniting discrete and continuous models.
• They play key roles in strong approximation theory, scaling limits, and fluctuation analysis across Brownian, Lévy, and Gaussian processes.
• Beyond probability, bridge concepts appear in graph theory and topology, exemplified by cut-edges and link bridge positions.

Random bridges are stochastic paths specified by endpoint information rather than by unconstrained dynamics alone. In the most common probabilistic usage, a bridge is obtained by conditioning a random walk, Brownian motion, Lévy process, or Gaussian process on terminal data such as $S_n=z$, $X_T=y$, positivity constraints between endpoints, or a random terminal time; conditioned path laws of this kind appear in strong approximation theory, fluctuation theory, line ensembles, convex geometry, and information-based stochastic modeling (Dimitrov et al., 2019). The term also has distinct meanings in graph theory and low-dimensional topology, where a bridge may denote a cut-edge or a bridge position of a link; these usages are related by language rather than by a common conditioning formalism (Elçi et al., 2015).

1. Foundational definitions and principal classes

A basic discrete model starts from an i.i.d. random walk

$S_n=X_1+\cdots+X_n,$

and defines the bridge of length $n$ from $0$ to $z$ by conditioning on the endpoint: $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$ In the continuous-jump setting $X$ has a density on a single interval $(\alpha,\beta)$; in the integer-valued setting the support is a single integer interval, and in both cases the bridge is extended to all $t\in[0,n]$ by linear interpolation (Dimitrov et al., 2019). A Brownian bridge of variance $X_T=y$0 from $X_T=y$1 to $X_T=y$2 on $X_T=y$3 is

$X_T=y$4

with covariance $X_T=y$5; for comparison with a length-$X_T=y$6 bridge ending at $X_T=y$7, the relevant reference process is $X_T=y$8 (Dimitrov et al., 2019).

A second standard class consists of bridges with additional pathwise constraints. For a random walk started at $X_T=y$9 and ending at $S_n=X_1+\cdots+X_n,$0, the non-negative and strictly positive bridges are

$S_n=X_1+\cdots+X_n,$1

$S_n=X_1+\cdots+X_n,$2

and excursions arise as the special case of bridges starting and ending at $S_n=X_1+\cdots+X_n,$3 while staying positive in between (Caravenna et al., 2012).

A third class randomizes the terminal time. For Brownian motion $S_n=X_1+\cdots+X_n,$4 and a strictly positive random time $S_n=X_1+\cdots+X_n,$5, independent of $S_n=X_1+\cdots+X_n,$6, the Brownian bridge on the random interval $S_n=X_1+\cdots+X_n,$7 is

$S_n=X_1+\cdots+X_n,$8

so that conditional on $S_n=X_1+\cdots+X_n,$9, the law is that of a Brownian bridge of length $n$0 (Bedini et al., 2016). For a centered Gaussian process $n$1 with covariance $n$2, the Gaussian bridge of random length $n$3 is

$n$4

again understood as a mixture of deterministic-length bridges conditional on $n$5 (Erraoui et al., 2017). For Lévy processes with transition densities $n$6, a bridge of deterministic length $n$7 from $n$8 to $n$9 is defined via the Markov bridge transition density

$0$0

and a random-length Lévy bridge is obtained by mixing these laws over a random horizon $0$1 (Erraoui et al., 2019).

Class	Defining condition	Representative source
Random walk bridge	$0$2	(Dimitrov et al., 2019)
Positive bridge / excursion	Endpoint conditioning plus $0$3 or $0$4	(Caravenna et al., 2012)
Random-length bridge	Conditional on $0$5, law equals a fixed-length bridge	(Bedini et al., 2016)

2. Strong approximation, scaling limits, and conditioned fluctuation theory

A central result for endpoint-conditioned walks is a KMT-type strong coupling for bridges. Under assumptions C1–C6 in the continuous case and D1–D5 in the integer-valued case, and for a fixed reference slope $0$6, there exists a coupling of a Brownian bridge $0$7 with variance $0$8 and all bridges $0$9 such that

$z$0

satisfies

$z$1

In particular, at the typical endpoint $z$2,

$z$3

so bridge paths admit a uniform $z$4 sup-norm approximation by Brownian bridges (Dimitrov et al., 2019). The same work shows that classical KMT for unconditioned walks does not imply the bridge result: conditioning on rare endpoints can radically alter midpoint laws, and a concrete spike-distribution counterexample shows that the analogue of the exponential-moment bound can fail without additional assumptions (Dimitrov et al., 2019).

Conditioned scaling limits appear in a complementary form for bridges constrained to stay positive. If the increments are in the domain of attraction of a strictly $z$5-stable law, and $z$6, $z$7, then the rescaled positive bridges converge in $z$8 to the bridge of the limiting stable Lévy process from $z$9 to $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$0, conditioned to stay non-negative or positive. In the Brownian case $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$1, with mean $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$2, variance $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$3, and endpoints $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$4, the rescaled bridge conditioned to stay positive converges in $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$5 to the normalized Brownian excursion (Caravenna et al., 2012). The proof uses Doob $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$6-transforms, explicit Radon–Nikodym densities, and local asymptotics for kernels such as

$$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$7

and their absolutely continuous analogues (Caravenna et al., 2012).

Endpoint conditioning also changes first-passage behavior. For a random walk bridge conditioned on $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$8, first-passage tails over a moving boundary typically retain a regularly varying exponent $$\mathcal L\big(\{S_m^{(n,z)}\}_{m=0}^n\big)=\mathcal L\big(\{S_m\}_{m=0}^n\,\big|\,S_n=z\big).$$9, but when the observation time $X$0 approaches $X$1, a phase transition appears whose form depends on the size of the boundary $X$2 relative to $X$3 (Sloothaak et al., 2017). This establishes that bridge conditioning affects not only global invariance principles but also barrier-crossing asymptotics.

3. Random terminal times, filtrations, and Markov structure

For Brownian bridges on random intervals, the random time becomes observable through the bridge itself. In the model

$X$4

one has $X$5 almost surely, so $X$6 is a stopping time for the completed natural filtration of $X$7 (Bedini et al., 2016). The posterior law of $X$8 given $X$9 is explicit: for $(\alpha,\beta)$0,

$(\alpha,\beta)$1

where $(\alpha,\beta)$2 is the Gaussian bridge density kernel. The process is Markov with respect to its natural filtration and also with respect to its right-continuous completed filtration (Bedini et al., 2016).

The Gaussian-Markov generalization preserves this structure. If $(\alpha,\beta)$3 is a centered Gaussian process with covariance $(\alpha,\beta)$4 and is Markov with respect to its natural filtration, then the Gaussian bridge of random length

$(\alpha,\beta)$5

is Markov with respect to its natural filtration; under Assumption 5.1, it is also Markov with respect to the completed natural filtration, which then satisfies the usual conditions of right-continuity and completeness (Erraoui et al., 2017). The deterministic bridge of length $(\alpha,\beta)$6 remains Gaussian-Markov, and its covariance has the explicit bridge form

$(\alpha,\beta)$7

The same framework yields explicit Bayesian formulas for the posterior law of $(\alpha,\beta)$8 given $(\alpha,\beta)$9, separating the cases $t\in[0,n]$0 and $t\in[0,n]$1 (Erraoui et al., 2017).

The Lévy case extends endpoint conditioning beyond Gaussian continuity. For a symmetric Lévy process with densities $t\in[0,n]$2 satisfying Sharpe’s integrability condition, a random-length bridge $t\in[0,n]$3 from $t\in[0,n]$4 to $t\in[0,n]$5 is defined by requiring that, conditionally on $t\in[0,n]$6, the law of $t\in[0,n]$7 is the law $t\in[0,n]$8 of the deterministic Lévy bridge (Erraoui et al., 2019). The random horizon $t\in[0,n]$9 is a stopping time for the completed natural filtration of $X_T=y$00, explicit conditional laws of $X_T=y$01 given discrete observations $X_T=y$02 are derived, and $X_T=y$03 is shown to be a non-homogeneous Markov process with explicit transition kernels. Under Assumption 3.1, the completed natural filtration is right-continuous and complete (Erraoui et al., 2019). A plausible implication is that random-length bridges form a stable bridge calculus across Gaussian and jump settings, with filtration regularity depending on density control rather than on continuity of paths.

4. Constructive, Gibbsian, and algorithmic viewpoints

Endpoint-conditioned paths can be generated exactly in discrete time. For a random walk

$X_T=y$04

with arbitrary jump density or mass function $X_T=y$05, the bridge of length $X_T=y$06 is generated by replacing the original increment law with the effective jump distribution

$X_T=y$07

where $X_T=y$08 is the backward propagator (Bruyne et al., 2021). This is an exact Doob-type construction and extends to generalized bridges with $X_T=y$09, excursions with positivity constraints, and meanders; the paper gives explicit formulas in the simple symmetric, Gaussian, and Cauchy cases, and emphasizes that Lévy flights are included (Bruyne et al., 2021).

Random bridges also appear as local resampling laws in integrable probability. For geometric random walk bridges, the set

$X_T=y$10

carries the uniform bridge measure $X_T=y$11, and interlacing ensembles of such bridges satisfy an interlacing Gibbs property (Dimitrov, 2024). Under one-point tightness, these line ensembles are tight, and any subsequential limit satisfies the Brownian Gibbs property; as an application, sequences of spiked Schur processes converge uniformly on compact sets to the Airy wanderer line ensembles (Dimitrov, 2024). This is precisely the type of application anticipated by KMT bridge couplings, where line-ensemble resampling laws require pathwise comparison of discrete bridges with Brownian bridges (Dimitrov et al., 2019).

A recent machine-learning reinterpretation treats random-bridges as stochastic transports between distributions. In that framework, random-bridges are stochastic processes conditioned to take target distributions at fixed timepoints; they can be Markovian or non-Markovian, and continuous, discontinuous, or hybrid depending on the driving process. Empirical results built on Gaussian random bridges produce high-quality samples in significantly fewer steps compared to traditional approaches, while achieving competitive Frechet inception distance scores (Goria et al., 16 Dec 2025). This suggests that the bridge formalism can be used not only to analyze conditioned paths but also to parameterize one-way generative transports.

5. Convex geometry, random matrices, and growing dimension

Finite-dimensional exchangeable bridges support exact convex-geometric formulas. If $X_T=y$12 are exchangeable and satisfy the bridge property

$X_T=y$13

then the partial sums $X_T=y$14 form a $X_T=y$15-dimensional random bridge, and the positive hull

$X_T=y$16

has distribution-free expected face numbers, conic intrinsic volumes, and tangent-cone functionals expressed through Stirling numbers of the first and second kinds (Godland et al., 2021). For example,

$X_T=y$17

and

$X_T=y$18

Here the bridge property is not a small perturbation of a walk but the defining type-$X_T=y$19 symmetry behind the cone formulas (Godland et al., 2021).

A related finite-dimensional phenomenon appears in convex position probabilities. For a random bridge $X_T=y$20 in $X_T=y$21 of length $X_T=y$22, with increments satisfying the bridge property, exchangeability, and general position, the probability that $X_T=y$23 are in convex position is

$X_T=y$24

the same universal value as for the first $X_T=y$25 points of a suitable random walk (Panzo, 2024). The proof reduces the event “not in convex position” to a vertex-count identity and then to a Stirling-number summation formula (Panzo, 2024).

High-dimensional asymptotics introduce a different geometry. In the square-integrable case, the path of random bridges viewed as random sets in $X_T=y$26 with $X_T=y$27 converges in the Gromov–Hausdorff sense to the deterministic space $X_T=y$28 equipped with the pseudo-metric

$X_T=y$29

while in the heavy-tailed case with summands regularly varying of order $X_T=y$30, the limiting metric space has a random metric derived from the bridge variant of a subordinator (Jin, 17 Mar 2025). This provides a metric-space analogue of the classical dichotomy between diffusive and stable bridge scaling.

Bridge limits also arise in random matrix theory. For the squared Frobenius norm of a deterministic $X_T=y$31 truncation of a Haar unitary or orthogonal matrix, the centered two-parameter process converges to a tied-down bivariate Brownian bridge; if rows and columns are instead selected by independent Bernoulli choices, then after centering and rescaling by $X_T=y$32 the process converges to another Gaussian field (Beffara et al., 2013). This suggests that bridge structures can emerge as universal fluctuation fields even when there is no direct endpoint conditioning of an underlying path.

6. Applications, neighboring domains, and terminological boundaries

In integrable probability and random geometry, bridge laws are local building blocks of Gibbsian path ensembles. The motivation emphasized in bridge coupling theory is that lozenge tilings, last-passage models, log-gamma polymers, Hall–Littlewood and Macdonald-process ensembles, and related discrete systems resample single curves according to random walk bridge laws between fixed endpoints, often additionally conditioned by barriers; Brownian Gibbs line ensembles provide the continuous counterpart (Dimitrov et al., 2019). The bridge perspective is therefore central to transferring regularity and scaling arguments between discrete and continuous models.

The term “bridge” also has established meanings outside conditioned stochastic processes. In the random-cluster model on a finite graph $X_T=y$33, an open edge $X_T=y$34 is a bridge if removing it increases the number of connected components: $X_T=y$35 For the random-cluster measure, exact bridge–edge identities relate the expected bridge density $X_T=y$36 to the expected open-edge density $X_T=y$37, and in two dimensions the variance of the number of bridges and non-bridges diverges below $X_T=y$38 of the cluster coupling $X_T=y$39 (Elçi et al., 2015). In complex networks, a bridge is likewise an edge whose removal disconnects the graph, and the associated bridgeness quantifies its damage to the giant connected component; analytic formulas for bridge fraction and bridgeness are available for uncorrelated random networks with arbitrary degree distributions (Wu et al., 2016). These are graph-theoretic bridges, not conditioned path measures.

Low-dimensional topology supplies another distinct usage. A random link defined by a random $X_T=y$40-bridge splitting is obtained by gluing two trivial $X_T=y$41-string tangles by a random walk in the mapping class group $X_T=y$42, and such a random link is hyperbolic with asymptotic probability $X_T=y$43 (Ichihara et al., 2016). Here “bridge” refers to bridge position of a link rather than to a stochastic bridge process.

Taken together, these literatures show that “random bridges” is not a single theory but a family of mathematically sharp constructions organized around endpoint, connectivity, or gluing constraints. In stochastic-process theory the unifying mechanism is conditioning on terminal data; in graph theory and topology the unifying mechanism is structural indispensability. The overlap in terminology is accidental, but the recurrence of bridge objects across probability, geometry, and combinatorics indicates that constrained intermediacy is itself a robust organizing principle.