---
title: Pairwise Independent Brownian Motions
url: https://www.emergentmind.com/topics/essentially-pairwise-independent-brownian-motions
type: topic
---

# Pairwise Independent Brownian Motions

Essentially pairwise independent Brownian motions are families of Brownian motions indexed by a set \(I\) such that, for almost every index \(u\), the process indexed by \(u\) is a standard Brownian motion, and for almost every pair \((u,v)\), the two indexed processes are independent. In the recent literature, the term appears in two closely related but distinct senses. In one sense, it is a measure-theoretic notion for continua of Brownian motions on a Fubini extension space, where full joint measurability is compatible with “almost everywhere” pairwise independence [2509.12096]. In another sense, it describes asymptotic phenomena in which strongly dependent pre-limit processes converge, after scaling, to Brownian motions with independent coordinates; the dependence persists at finite scale but disappears in the limit [1508.02840], [1410.1108]. These usages are connected by a common theme: independence at the Brownian level can emerge either as an exact structural property on an extended probability space or as a scaling-limit property of nonlinearly coupled stochastic systems.

## 1. Formal definition and probabilistic setting

The precise measure-theoretic notion is formulated on two probability spaces: an index space \((I,\mathcal I,\lambda)\) and a sample space \((\Omega,\mathcal F,\mathbb Q)\). A random variable \(\Theta:I\times\Omega\to\mathbb R\) is called **essentially pairwise independent** if, for \(\lambda\)-almost every \(u\in I\), the random variables \(\Theta(u,\cdot)\) and \(\Theta(v,\cdot)\) are independent for \(\lambda\)-almost every \(v\in I\) [2509.12096]. The qualifier “essentially” means that independence may fail on index sets of \(\lambda\)-measure zero.

For Brownian motions, the corresponding notion is an indexed family \(\{(B_t^u)_{t\ge 0}:u\in I\}\) such that, for every \(u\in I\), \((B_t^u)_{t\ge 0}\) is a standard Brownian motion on \((\Omega,\mathcal F,\mathbb Q)\), the map \((u,\omega)\mapsto B_t^u(\omega)\) is jointly measurable for each \(t\), and the family is essentially pairwise independent [2509.12096]. This formulation is designed for continuum-agent models, where one wants idiosyncratic Brownian noise for almost every agent while retaining joint measurability.

A distinct but compatible interpretation appears in scaling-limit results. In "Bootstrap Random Walks" [1508.02840], a collection of random walks is generated from a single i.i.d. increment sequence by repeated nonlinear “bootstrapping”; at the discrete level the coordinates are deterministic functionals of the same underlying randomness, but under diffusive scaling the limit is a multidimensional Brownian motion with independent components. In that setting, “essentially pairwise independent Brownian motions” means that the Brownian objects are independent in the limit even though the pre-limit coordinates are far from independent [1508.02840].

This suggests a useful conceptual distinction. One may speak of **exact e.p.i. Brownian families** when independence is formulated directly on an extended probability space, and of **asymptotically independent Brownian coordinates** when independent Brownian motions arise as limits of dependent systems. The literature supports both interpretations, but keeps them technically separate.

## 2. Fubini extension spaces and continuum Brownian families

The need for Fubini extensions arises because the classical product space
\[
(I\times\Omega,\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)
\]
does not support a jointly measurable continuum of independent random variables [2509.12096]. Sun’s Fubini extension enlarges this product to
\[
(I\times\Omega,\mathcal V,\mathcal Q),
\]
with the property that Fubini’s theorem continues to hold for every integrable \(\mathcal V\)-measurable function \(g\):
\[
\int_{I\times\Omega} g\,d\mathcal Q
=
\int_I\left(\int_\Omega g_u\,d\mathbb Q\right)\,d\lambda
=
\int_\Omega\left(\int_I g_\omega\,d\lambda\right)\,d\mathbb Q.
\]
This framework permits jointly measurable continuum families with prescribed marginal laws [2509.12096].

Within this setting, a process \((\mathbf B_t)_{t\ge 0}\) on \((I\times\Omega,\mathcal V,\mathcal Q)\) is a standard Brownian motion if \(\mathbf B_0=0\) \(\mathcal Q\)-a.s., its increments are Gaussian with variance \(t-s\), its increments over disjoint intervals are independent, and for \(\mathcal Q\)-almost every \((u,\omega)\) the path \(t\mapsto \mathbf B_t(u,\omega)\) is continuous [2509.12096]. The associated coordinate family is given by
\[
B_t^u(\omega):=\mathbf B_t(u,\omega).
\]

The main structural equivalence is explicit. If \(\{B^u:u\in I\}\) is an e.p.i. collection of Brownian motions, then
\[
\mathbf B_t(u,\omega):=B_t^u(\omega)
\]
defines a standard Brownian motion on the Fubini extension space [2509.12096]. Conversely, if \((\mathbf B_t)_{t\ge 0}\) is a standard Brownian motion on the Fubini extension whose coordinates are pairwise independent and identically distributed, then the coordinate processes \(B_t^u(\omega):=\mathbf B_t(u,\omega)\) form an e.p.i. collection of Brownian motions on the marginal sample space [2509.12096]. The identically distributed condition is essential: the paper notes that one can construct families whose aggregate process is Gaussian under \(\mathcal Q\) while individual coordinates are not Brownian under \(\mathbb Q\) [2509.12096].

A further consequence is the Exact Law of Large Numbers. If \(\Theta\) is integrable and e.p.i. on the Fubini extension, then for every \(\tilde\omega\in\Omega\),
\[
\int_I \Theta(u,\tilde\omega)\,\lambda(du)
=
\int_{I\times\Omega}\Theta(u,\omega)\,\mathcal Q(du,d\omega).
\]
Applied to an e.p.i. Brownian family, this yields
\[
\int_I B_t^u(\omega)\,\lambda(du)=0
\]
for every \((t,\omega)\) [2509.12096]. In continuum-agent models, this pathwise averaging property plays the role usually associated with laws of large numbers for independent sequences.

## 3. Preservation under measure change and graphon SDEs

A notable feature of the Fubini-extension framework is that it supports a Girsanov theorem preserving pairwise independence at the level of coordinates. Starting from an e.p.i. Brownian family \(\mathbf B=\{B^u:u\in I\}\) on \((I\times\Omega,\mathcal V,\mathcal Q)\), and a coordinate-wise adapted process \(\Theta=\{(\theta_{t,u})_{t\ge 0}:u\in I\}\), one defines the exponential martingale
\[
\mathcal E_t(u,\omega)
=
\exp\Bigg\{
\int_0^t \theta_{s,u}(\omega)\,dB_s^u(\omega)
-\frac12\int_0^t \theta_{s,u}(\omega)^2\,ds
\Bigg\},
\]
and the new measure
\[
\frac{d\mathcal P}{d\mathcal Q}=\mathcal E_T.
\]
The drifted global process
\[
\mathbf W_t:=\mathbf B_t-\int_0^t \Theta_s\,ds,
\qquad
W_t^u=B_t^u-\int_0^t \theta_{s,u}\,ds,
\]
is then a standard Brownian motion on \((I\times\Omega,\mathcal V,\mathcal P)\) [2509.12096].

The coordinate-wise implication is subtler. Under \(\mathcal P\), the family \(\{W^u\}\) remains essentially pairwise independent, but the individual coordinate processes need not remain Brownian on \((\Omega,\mathcal F,\mathcal P)\); for instance,
\[
\mathbb E_{\mathcal P}[W_t^u]
=
-\int_0^t \mathbb E_{\mathbb Q}[\theta_{s,u}]\,ds
\]
unless the drift has zero mean [2509.12096]. Thus the global Brownian structure is preserved on the extended space, while coordinate marginals may acquire drift. The Exact Law of Large Numbers is crucial in the independence argument after measure change [2509.12096].

The same paper uses this noise structure to reformulate graphon stochastic differential equations. For a symmetric Borel measurable graphon
\[
w:I\times I\to I,
\]
the continuum system
\[
d\theta_{t,u}
=
b\left(\theta_{t,u}, \int_I w(u,v)\,\phi_{t,v}\,\lambda(dv)\right)\,dt
+
\sigma(\theta_{t,u})\,dB_t^u,
\qquad
\phi_{t,u}:=\mathbb E_{\mathbb Q}[\theta_{t,u}],
\]
can be rewritten as a single McKean–Vlasov type equation on the Fubini extension:
\[
d\Theta_t
=
b\big(\Theta_t,\mathbb W[\Phi_t]\big)\,dt
+
\sigma(\Theta_t)\,d\mathbf B_t,
\]
where
\[
\mathbb W[\eta](u)=\int_I w(u,v)\eta(v)\,\lambda(dv).
\]
Under the stated Lipschitz and linear-growth assumptions on \(b\) and \(\sigma\), the equation admits a unique solution in the corresponding square-integrable path space [2509.12096]. In this application, essentially pairwise independent Brownian motions provide the idiosyncratic noise for a continuum of agents while retaining a jointly measurable, globally analyzable structure.

## 4. Asymptotic independence from a single source of randomness

A different route to pairwise independent Brownian motions begins from a single i.i.d. increment sequence. In "Bootstrap Random Walks" [1508.02840], one starts from the simple symmetric random walk
\[
X_n=\sum_{k=1}^n \xi_k,
\qquad
\mathbb P(\xi_n=1)=\mathbb P(\xi_n=-1)=\tfrac12,
\]
and defines new increments by partial products,
\[
\eta_n=\prod_{k=1}^n \xi_k,
\qquad
Y_n=\sum_{k=1}^n \eta_k.
\]
Although \((\eta_n)\) is again i.i.d. symmetric \(\pm1\), the dependence between \(X\) and \(Y\) is exact and invertible:
\[
\eta_n=\prod_{k=1}^n \xi_k,
\qquad
\xi_n=\eta_{n-1}\eta_n,
\]
and the filtrations coincide:
\[
\mathcal F_n=\sigma(\xi_1,\dots,\xi_n)=\sigma(\eta_1,\dots,\eta_n).
\]
Thus \(X\) and \(Y\) share the same underlying randomness and are not independent at finite time [1508.02840].

Nevertheless, under the usual diffusive scaling,
\[
\mathfrak W_n(t)=\frac{1}{\sqrt n}W_{\lfloor nt\rfloor},
\qquad
W_n=(X_n,Y_n),
\]
the process converges weakly to a two-dimensional Brownian motion with independent components [1508.02840]. More generally, after iterating the bootstrap operation and constructing
\[
W_n=(Y_{0,n},Y_{1,n},\dots,Y_{K,n}),
\]
the normalized process converges in \(D[0,1]^{K+1}\) to a \((K+1)\)-dimensional Brownian motion with independent components [1508.02840].

The independence mechanism is encoded in the covariance structure. For fixed bootstrap levels \(K\) and \(J\), the increment correlations satisfy:
\[
\mathbb E[\eta_{K,m}\eta_{J,n}]=0 \quad \text{if } m\neq n,
\]
\[
\mathbb E[\eta_{K,m}^2]=\sigma^2,
\]
and, for equal times, there is a threshold \(\omega_{|K-J|}\) such that
\[
\mathbb E[\eta_{K,m}\eta_{J,m}]
=
\begin{cases}
\sigma^2, & m<\omega_{|K-J|},\\
0, & m\ge \omega_{|K-J|}.
\end{cases}
\]
Consequently,
\[
\mathbb E[Y_{K,m}Y_{J,n}]
=
\min(m,\omega_{|K-J|}-1)\sigma^2,
\]
so the covariance between distinct coordinates stabilizes at constant order while each variance grows linearly in \(n\) [1508.02840]. After normalization by \(\sigma\sqrt n\), the cross-covariances vanish:
\[
\operatorname{Cov}\Big(\frac{Y_{K,n}}{\sigma\sqrt n},\frac{Y_{J,n}}{\sigma\sqrt n}\Big)\to 0.
\]
Since the limiting field is Gaussian with diagonal covariance matrix \(\Sigma=I_{K+1}\), the limiting coordinates are independent Brownian motions [1508.02840].

The construction extends beyond \(\{-1,+1\}\). If \(\xi_n\) takes values in a finite alphabet \(\mathcal U=\{u_0,\dots,u_{p-1}\}\subset\mathbb R\), where \(p\) is prime and \((\mathcal U,\otimes)\) is an Abelian cyclic group of order \(p\), one defines the bootstrap operator
\[
\Delta:(x_n)_n\mapsto \Big(\bigotimes_{\ell=1}^n x_\ell\Big)_n.
\]
Under the mean-zero condition
\[
\frac1p\sum_{i=0}^{p-1}u_i=0,
\]
each bootstrap increment process has the same i.i.d. law as the original increments, and the corresponding vector of walks again converges to a multidimensional Brownian motion with independent components [1508.02840].

This line of work provides a concrete realization of asymptotic independence: many independent Brownian motions arise from a single underlying source of randomness via nonlinear recycling of increments. The pre-limit dependence is strong, but it is localized in a way that becomes negligible under diffusive scaling [1508.02840].

## 5. Interacting systems and asymptotically independent Brownian limits

A related asymptotic-independence phenomenon appears in "Stirring two grains of sand" [1410.1108]. The model consists of a Brownian particle \(B_t\) moving on a flat torus \(T_r^d\), together with two unit balls whose centers \(X_t\) and \(Y_t\) move only when pushed by the Brownian particle, and reflect when they touch so that
\[
|X_t-Y_t|\ge 2.
\]
At finite \(r\), the two center processes are dependent: they share the same Brownian driver and interact through reflection [1410.1108].

The analysis is carried out in the local-time clock. If \(L_t^X\) and \(L_t^Y\) are the local times of \(B_t\) on the moving spheres surrounding the balls, and
\[
L_t=L_t^X+L_t^Y,
\qquad
\sigma_t=\inf\{s\ge 0:L_s\ge t\},
\]
then the rescaled center process observed at \(\sigma_{nt}\) satisfies an invariance principle. For \(d\ge 2\), with
\[
C_d=\sqrt{\frac{2}{(d-1)d}},
\]
the paper proves that, as \(n\to\infty\) and \(r\to\infty\),
\[
\Big\{C_d\,n^{-1/2}\big(X_{\sigma_{nt}}^r-X_0^r,\;Y_{\sigma_{nt}}^r-Y_0^r\big),\ t\ge 0\Big\}
\Rightarrow
\text{standard Brownian motion in }\mathbb R^{2d},
\]
that is, a pair of independent \(d\)-dimensional Brownian motions [1410.1108]. The diffusion coefficient depends on dimension through \(C_d\).

The same asymptotic decorrelation appears at equilibrium. The process \((B_t^r,X_t^r,Y_t^r)\) has a unique stationary distribution \(L_r^d\), and the rescaled stationary law of the centers satisfies
\[
(X_0^r/r,\;Y_0^r/r)\xrightarrow{d}\nu_1^d\times \nu_1^d,
\]
the product of uniform laws on the unit torus [1410.1108]. Thus the time-evolving processes and their stationary positions both become asymptotically independent.

This interacting-particle example differs from the bootstrap-random-walk construction in mechanism but not in outcome. In the discrete bootstrap model, asymptotic independence arises from covariance saturation and Gaussian limits [1508.02840]; in the stirred-balls model, it arises from a large-domain limit, a local-time clock, and excursion-theoretic homogenization [1410.1108]. In both settings, strongly coupled finite-scale dynamics converge to independent Brownian motions.

A useful contrast is provided by "Common Decomposition of Correlated Brownian Motions and its Financial Applications" [1907.03295]. There, two correlated Brownian motions \(B\) and \(W\) are decomposed as
\[
B_t=X_{T_t}+Y_{S_t},
\qquad
W_t=X_{T_t}-Y_{S_t},
\]
where \(X\) and \(Y\) are independent Brownian motions and
\[
T_t=\frac{t+[B,W]_t}{2},
\qquad
S_t=\frac{t-[B,W]_t}{2}.
\]
The paper shows equivalent conditions for the triplet \((X,Y,T)\) to be mutually independent [1907.03295]. This is not an e.p.i. construction over a continuum of indices, but it exhibits the same structural theme: correlated Brownian objects can be represented in terms of independent Brownian building blocks plus an auxiliary clock.

## 6. Identification, testing, and related interpretations

For Gaussian processes, pairwise independence is equivalent to vanishing cross-covariance. In "Tests of independence for pairs of paths of non-stationary Gaussian processes" [2510.23563], two standard Brownian motions \(\mathcal W_1,\mathcal W_2\) are taken jointly Gaussian with
\[
\operatorname{Cov}(\mathcal W_1(t),\mathcal W_2(s))=r\,(t\wedge s).
\]
In this framework, the two Brownian motions are independent if and only if \(r=0\) [2510.23563]. The same equivalence holds for fractional Brownian motions with \(H>1/2\), where the cross-covariance is the same covariance kernel multiplied by \(r\) [2510.23563].

The paper emphasizes that pathwise empirical correlation does not directly test independence for non-stationary Gaussian processes. For a path \(f\), define
\[
D(f)=\int_0^1 f(t)^2\,dt-\Big(\int_0^1 f(t)\,dt\Big)^2,
\]
and for two paths \(f,g\),
\[
A(f,g)=\int_0^1 f(t)g(t)\,dt-\Big(\int_0^1 f(t)\,dt\Big)\Big(\int_0^1 g(t)\,dt\Big).
\]
The associated empirical correlation is
\[
\rho=\frac{A(f,g)}{\sqrt{D(f)D(g)}}.
\]
For independent Brownian motions, \(\rho\) has a non-degenerate distribution that does not converge as the observation horizon grows; this is the paper’s formulation of Yule’s “nonsense correlation” [2510.23563]. Instead, the informative object is the discretization error \(\rho_n-\rho\), where \(\rho_n\) is computed from samples at \(t_k=k/n\).

For Brownian motion, the paper proves that
\[
n(\rho_n-\rho)=\mu(\mathcal W_1,\mathcal W_2)+\mathcal Y_n,
\]
and
\[
\frac{n(\rho_n-\rho)-\mu(\mathcal W_1,\mathcal W_2)}
{\sigma^r(\mathcal W_1,\mathcal W_2)}
\xrightarrow{\mathcal D}\mathcal N(0,1),
\]
where \(\mu\) and \(\sigma^r\) are explicit functionals of the observed paths [2510.23563]. Under the null hypothesis \(r=0\), this yields a path-based asymptotic test for independence of Brownian paths. The paper also provides a fully discrete version based on \(\rho_n-\rho_{2n}\) [2510.23563].

These testing results belong to a different layer of the subject than the construction theorems. The Fubini-extension literature studies how to realize e.p.i. Brownian families as measurable objects [2509.12096]. The bootstrap and stirring papers study how independent Brownian components emerge from dependent dynamics in the limit [1508.02840], [1410.1108]. The testing paper studies how to decide, from observed Gaussian paths, whether the pairwise independence condition \(r=0\) holds [2510.23563]. Taken together, these strands show that “essentially pairwise independent Brownian motions” is not a single theorem but a cluster of related ideas: exact measure-theoretic realizability, asymptotic decorrelation under scaling, structural decompositions into independent drivers, and statistical verification from path data.

Source: https://www.emergentmind.com/topics/essentially-pairwise-independent-brownian-motions