Fubini Extension Space in Probability

• Fubini extension space is an enlarged product probability space that preserves the exact iterated integration property of Fubini’s theorem for continuum-indexed random fields.
• It facilitates the realization of a continuum of essentially pairwise independent random variables, including standard Brownian motions, via an extended measurable structure.
• The framework supports advanced probabilistic results such as a global Girsanov theorem and graphon SDE formulations, uniting continuum-agent systems with stochastic calculus.

A Fubini extension space is an enlarged product probability space designed to retain the full iterated-integration property of Fubini’s theorem while accommodating jointly measurable continuum-indexed random fields that the ordinary product space cannot support. In the recent probabilistic formulation, the underlying set remains $I\times\Omega$, but the measurable structure and probability measure are extended from $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$ to $(\mathcal V,\mathcal Q)$, or equivalently $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$, so that a continuum of essentially pairwise independent random variables or Brownian motions can be realized as a single jointly measurable process without losing exact iterated integration (Amini et al., 15 Sep 2025).

1. Definition and measure-theoretic structure

The basic setting consists of an index space and a sample space. In the probabilistic applications emphasized in the literature, the intended index set is $I=[0,1]$, first equipped with its Borel $\sigma$-algebra $\mathcal B_I$ and Lebesgue measure $\lambda_I$. One would like to work on a product probability space of the form

$$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$

where $(I,\mathcal I,\lambda)$ is an atomless index space and $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$0 is the sample space. The standard obstruction is that the usual product probability space cannot accommodate a jointly measurable continuum of independent random variables. The Fubini extension enlarges the measurable structure and measure while preserving the product-like integral calculus (Amini et al., 15 Sep 2025).

Formally, if $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$1 and $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$2 are probability spaces, then a triple

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$3

is a Fubini extension of

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$4

if for any $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$5-integrable real-valued function $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$6, the sections

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$7

are integrable for $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$8-a.e. $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$9 and $(\mathcal V,\mathcal Q)$0-a.e. $(\mathcal V,\mathcal Q)$1, and

$(\mathcal V,\mathcal Q)$2

This exact iterated-integration identity is the defining property.

Structure	Measurable space	Integral property
Ordinary product space	$(\mathcal V,\mathcal Q)$3	Insufficient for a jointly measurable continuum of independent random variables
Fubini extension space	$(\mathcal V,\mathcal Q)$4 or $(\mathcal V,\mathcal Q)$5	Preserves the Fubini theorem for all $(\mathcal V,\mathcal Q)$6-integrable functions

The space is not merely existentially larger. The paper quotes Sun’s theorem: for the Lebesgue index space $(\mathcal V,\mathcal Q)$7, there exists an extension $(\mathcal V,\mathcal Q)$8, a probability space $(\mathcal V,\mathcal Q)$9, and a Fubini extension $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$0 such that for any measurable map

$$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$1

there exists a $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$2-measurable random variable

$$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$3

whose sections are essentially pairwise independent and satisfy

$$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$4

This identifies the Fubini extension space as a realization theorem for continuum-indexed laws, not only as an abstract extension (Amini et al., 15 Sep 2025).

2. Continuum-indexed randomness and Brownian motion

The central probabilistic notion on a Fubini extension space is essential pairwise independence. A random variable $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$5 is essentially pairwise independent if for $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$6-almost every $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$7, the random variables $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$8 and $$(\mathcal I\boxtimes\mathcal F,\lambda\boxtimes\mathbb Q)$$9 are independent for $I=[0,1]$0-almost every $I=[0,1]$1. This is weaker than full mutual independence, but it is the notion used in continuum-agent models and exact law of large numbers results (Amini et al., 15 Sep 2025).

A standard Brownian motion on the Fubini extension $I=[0,1]$2 is a process $I=[0,1]$3, with $I=[0,1]$4, such that:

1. $I=[0,1]$5, $I=[0,1]$6-a.s.;
2. for all $I=[0,1]$7,

$I=[0,1]$8

1. for any $I=[0,1]$9, the increments

$\sigma$0

are independent under $\sigma$1;

2. $\sigma$2 is continuous for $\sigma$3-a.e. $\sigma$4.

By contrast, an e.p.i. collection of Brownian motions is a family

$\sigma$5

such that $\sigma$6 is jointly measurable for all $\sigma$7, the collection is essentially pairwise independent, and for every $\sigma$8, the process $\sigma$9 is a standard Brownian motion on $\mathcal B_I$0.

The key equivalence result is Theorem 3.1. If $\mathcal B_I$1 is a collection of e.p.i. Brownian motions in the Fubini extension space, then the aggregate process

$\mathcal B_I$2

is a standard Brownian motion on $\mathcal B_I$3. The converse, stated as Proposition 3.2, requires an additional hypothesis: if $\mathcal B_I$4 is a standard Brownian motion on the Fubini extension with pairwise independent and identically distributed elements, then

$\mathcal B_I$5

forms an e.p.i. collection of Brownian motions. The paper remarks that identity in law appears necessary; without it, one may construct a family whose aggregate process is Brownian under $\mathcal B_I$6, although each component is not Brownian under $\mathcal B_I$7 (Amini et al., 15 Sep 2025).

This equivalence is closely tied to the exact law of large numbers. For an integrable e.p.i. random variable $\mathcal B_I$8,

$\mathcal B_I$9

In particular, for an e.p.i. Brownian family $\lambda_I$0,

$\lambda_I$1

A plausible implication is that the Fubini extension space merges continuum-agent averaging and ordinary stochastic-process structure into a single probability-space formalism.

3. Change of measure and preserved pairwise independence

A major application of the Brownian-motion equivalence is a Girsanov theorem on the Fubini extension space. Let

$\lambda_I$2

be a collection of e.p.i. Brownian motions, and let

$\lambda_I$3

be a real-valued process such that for all $\lambda_I$4,

$\lambda_I$5

Define the stochastic exponential

$\lambda_I$6

with sectionwise form

$\lambda_I$7

If for all $\lambda_I$8, the process $\lambda_I$9 is a $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$0-martingale, then under the measure $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$1 defined by

$$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$2

the shifted process

$$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$3

is a standard Brownian motion on

$$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$4

(Amini et al., 15 Sep 2025).

The theorem is global rather than merely sectionwise. Under the sectionwise changed measures

$$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$5

classical Girsanov yields Brownianity of each $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$6. The Fubini extension then allows one to integrate over $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$7 and recover Brownianity of $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$8 under $$(I\times \Omega,\mathcal I\otimes \mathcal F,\lambda\otimes \mathbb Q),$$9. The paper also states that pairwise independence is preserved under the changed measure, using the exact law of large numbers inside the Radon–Nikodym weighting.

The important subtlety is that global Brownianity on $(I,\mathcal I,\lambda)$0 does not imply that each fixed-$(I,\mathcal I,\lambda)$1 section is Brownian under $(I,\mathcal I,\lambda)$2. The paper computes, for bounded measurable $(I,\mathcal I,\lambda)$3,

$(I,\mathcal I,\lambda)$4

and concludes that $(I,\mathcal I,\lambda)$5 need not be Brownian under $(I,\mathcal I,\lambda)$6. This directly separates sectionwise stochastic analysis from stochastic analysis on the Fubini extension space itself.

A sufficient condition for the exponential martingales is a Novikov-type integrability assumption. If $(I,\mathcal I,\lambda)$7 is e.p.i., $(I,\mathcal I,\lambda)$8, each $(I,\mathcal I,\lambda)$9 is $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$00-measurable, and

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$01

then

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$02

forms a collection of e.p.i. $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$03-martingales (Amini et al., 15 Sep 2025).

4. Graphon SDEs and the single-equation formulation

One of the main motivations for Fubini extension spaces is the analysis of continuum systems of interacting stochastic agents. The paper considers graphon stochastic differential equations driven by a family of e.p.i. Brownian motions. A graphon is a symmetric Borel measurable function

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$04

The continuum system is

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$05

Using the Brownian-motion equivalence, the full family can be rewritten as a single process

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$06

on the Fubini extension space. The compact reformulation is

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$07

where

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$08

and the graphon interaction operator is

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$09

The paper describes this as a McKean–Vlasov type equation driven by a standard Brownian motion (Amini et al., 15 Sep 2025).

The associated function spaces are

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$10

and the path spaces

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$11

Under the hypotheses

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$12

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$13

the paper proves that if $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$14, then there exists a unique solution $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$15 to (2), and

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$16

This is one of the clearest analytic payoffs of the framework: a continuum family of SDEs can be treated as a single SDE on one probability space.

The paper also stresses that this is not exactly the standard McKean–Vlasov formulation, because the interaction occurs through $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$17 and the expectation $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$18 is taken under $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$19, over the sample-space coordinate. A plausible implication is that the Fubini extension space functions as a bridge between continuum-agent systems and standard single-space stochastic calculus, without collapsing the directional structure carried by the index variable.

5. Scope, interpretation, and common misconceptions

The most persistent misconception is that a Fubini extension space is just a larger $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$20-algebra on a product set. The defining point is stronger: it is an enlargement that preserves the exact Fubini identity for all integrable functions. The extended space is therefore simultaneously a joint measurability device and an iterated-integration device (Amini et al., 15 Sep 2025).

A second misconception is that the framework is only a technical substitute for independent families on the ordinary product space. The paper uses a weaker, but structurally appropriate, notion: essential pairwise independence. This is the notion that supports continuum-agent modeling and exact law of large numbers statements, while remaining compatible with the existence of jointly measurable fields on the extension.

A third misconception is that global statements on the Fubini extension automatically translate into sectionwise statements on the marginal sample space. The Girsanov theorem shows this is false. A process can be Brownian on $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$21 while its fixed-$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$22 sections are not Brownian under $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$23. The paper presents this as a specifically global phenomenon, tied to the distinction between $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$24- or $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$25-laws on the extension and $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$26-laws on the sample space.

The framework is therefore best understood as a product-extension formalism adapted to continuum-indexed stochastic systems. Its purpose is not to imitate ordinary product probability spaces as closely as possible, but to realize jointly measurable continuum families and preserve

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$27

on the enlarged measurable structure. This explains why the paper positions Fubini extension spaces as basic objects for exact law of large numbers, preserved pairwise independence under change of measure, and graphon stochastic dynamics (Amini et al., 15 Sep 2025).

6. Related and analogical usages in adjacent literatures

The exact term “Fubini extension space” is specific. Several papers use the phrase only analogically, or identify nearby constructions that play an extension role for some Fubini principle.

In model theory, the closest analogue is the full subcategory

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$28

whose objects are the definable sets that are fiberable over $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$29. Under stable embeddedness of $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$30, an $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$31-valued Fubini measure on $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$32 extends uniquely to $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$33; in important settings this coincides with co-analyzable sets relative to $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$34 (Dries, 27 Dec 2025). This is an extension domain for Fubini measures, not a probability-space extension.

In stochastic analysis of semimartingales, the closest “extension-space-like” object is the weak$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$35 predictable measure-valued integrand space

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$36

together with the charge-/measure-valued stochastic integral $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$37. This enlarges classical scalar integrands and fixed parameter measures to measure-valued predictable integrands and stochastic kernels, yielding a new stochastic Fubini theorem (Choulli et al., 2024).

In convergence-space theory, the relevant object is the natural distribution monad

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$38

on arbitrary convergence spaces. Commutativity of this monad yields a Fubini theorem for continuous linear functionals on spaces of scalar functions, generalizing the classical theorem for Radon measures on compact Hausdorff spaces (Lucyshyn-Wright, 2012). This is a generalized-space extension of Fubini theory, but not a Fubini extension space in the probabilistic sense.

In Banach-space tensor analysis, the nearest construction is the pair

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$39

together with canonical partial-integration operators $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$40 and $$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$41. The paper describes this as the natural candidate for a Fubini extension mechanism on the dual side of the projective tensor product (He et al., 7 Feb 2026).

These parallel usages show that “Fubini extension” has become a broader heuristic label for structures that enlarge the domain on which order-of-integration or fiberwise summation remains valid. In the strict probabilistic sense, however, a Fubini extension space is the extended product probability space

$$(\mathcal I\otimes\mathcal F,\lambda\otimes\mathbb Q)$$42

that preserves the Fubini theorem while supporting jointly measurable continuum-indexed randomness (Amini et al., 15 Sep 2025).