Infinite-Dimensional SDEs

• Infinite-dimensional stochastic differential equations are evolution equations on spaces with infinitely many degrees of freedom, extending classical finite-dimensional theory.
• They address challenges such as cylindrical noise and non-Lipschitz or singular coefficients by employing diverse solution frameworks including mild, martingale, and strong formulations.
• Their applications span Hilbert spaces, countable-site systems, and particle models, underpinning advances in SPDEs, random matrix theory, and scaling limits.

Infinite-dimensional stochastic differential equations (ISDEs) are stochastic evolution equations with infinitely many degrees of freedom. In the literature, this includes Hilbert- or Banach-space-valued equations driven by cylindrical Brownian motion, countable-site systems such as reaction-diffusion dynamics on $\mathcal E \subset [0,\infty)^{\mathcal V}$, and labeled infinite particle systems whose coordinates solve coupled SDEs with singular long-range interaction. Their analysis departs sharply from finite-dimensional Itô theory because the noise may be cylindrical, coefficients may be merely measurable, locally monotone, or non-Lipschitz, and the drift may involve conditionally convergent infinite sums or logarithmic derivatives of random point fields. Correspondingly, solution concepts include mild, martingale, weak, strong, and path-by-path formulations, together with Dirichlet-form and tail-$\sigma$-field methods (Wresch, 2017, Osada et al., 2014, Goodair, 2022).

1. Scope, state spaces, and representative formulations

The term ISDE is used for several closely related classes of stochastic equations. One class consists of evolution equations on infinite-dimensional linear spaces. Wresch studies

$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$

in a separable Hilbert space, with $B$ a cylindrical Brownian motion and $A^{-1}$ trace-class (Wresch, 2017). Bao, Ren, and coauthors formulate stochastic functional differential equations with delay on a Gelfand triple

$$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$

where $x_t(\cdot)=x(t+\cdot)$ is the segment process (Rockner et al., 2014). Chueshov and Millet’s notes develop stochastic integration for Hilbert-space-valued processes driven by cylindrical Brownian motion,

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$

with Hilbert–Schmidt integrands (Goodair, 2022).

A second class consists of coordinatewise equations indexed by a countable set. In reaction-diffusion models on $\mathcal V$, the state is $\zeta_t:\mathcal V\to[0,\infty)$ in the weighted space

$\sigma$0

and the dynamics are

$\sigma$1

Here the transition kernel $\sigma$2 induces the discrete Laplacian $\sigma$3 (Costa et al., 2020).

A third class consists of infinite particle systems. Osada, Tanemura, and collaborators study labeled configurations $\sigma$4, typically with

$\sigma$5

or singular logarithmic/Coulomb specializations thereof (Osada et al., 2014). For jump-type systems, the Brownian drivers are replaced by Poisson random fields and the dynamics become nonlocal: $\sigma$6 The jump rate depends on the configuration through Palm-measure data of an equilibrium measure $\sigma$7 on configuration space $\sigma$8 (Esaki et al., 2018).

Regime	State space	Representative form
Hilbert/Banach evolution	$\sigma$9, or $$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$0	$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$1
Countable-site systems	$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$2	$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$3
Infinite particle diffusions	$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$4 or configuration space	$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$5

These formulations already indicate that “infinite-dimensional” may refer either to a linear state space, an infinite coordinate set, or an unlabeled configuration space encoded through a labeled process.

2. Solution concepts and well-posedness mechanisms

Because coefficients in ISDEs are frequently singular or non-Lipschitz, well-posedness is organized through several distinct frameworks rather than a single canonical notion. In Hilbert spaces, a standard notion is the mild solution

$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$6

and Wresch proves path-by-path uniqueness for almost every noise realization when the drift is merely measurable and bounded but its coordinates satisfy

$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$7

This extends Davie’s finite-dimensional result to separable Hilbert spaces (Wresch, 2017).

For non-Markovian or non-globally monotone equations, martingale and strong-solution methods based on Gelfand triples are central. The delay framework of (Rockner et al., 2014) first constructs martingale solutions by Galerkin approximation, tightness, and weak convergence, then obtains strong existence and uniqueness under the local monotonicity condition (H4) and a pathwise growth assumption (H5). The same framework applies to stochastic fractional Navier-Stokes equations with delay, stochastic reaction-diffusion equations with delay, and stochastic porous media equations with delay (Rockner et al., 2014). The reflected SPDE theory of (Li et al., 4 Mar 2026) uses fully local monotonicity, penalization by the projection $$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$8 onto the unit ball, and a variational inequality

$$\mathrm dX_t = -A X_t \,\mathrm dt + f(t, X_t) \,\mathrm dt + \mathrm dB_t$$9

to prove existence and uniqueness for reflected equations in an infinite-dimensional ball. The examples explicitly include stochastic Allen-Cahn, stochastic $B$0-Laplacian, stochastic Cahn-Hilliard, and stochastic 3D tamed Navier-Stokes equations (Li et al., 4 Mar 2026).

For interacting particle ISDEs, a different mechanism dominates. Osada and coauthors formulate strong existence and pathwise uniqueness through an infinite system of finite-dimensional SDEs with consistency, the IFC condition, coupled to tail-$B$1-field arguments (Osada et al., 2014). In this approach, the finite block of labeled particles evolves in the random environment generated by the tail, and pathwise uniqueness for each conditioned finite system lifts to the full ISDE. The follow-up paper (Kawamoto et al., 2020) shows that IFC can be proved without assuming quasi-regularity or symmetry of the associated Dirichlet forms, by explicit local regularity, non-explosion, and local Lipschitz-type conditions on the finite-dimensional SDEs in random environment.

A common misconception is that an equilibrium random point field alone determines a strong labeled dynamics. The results above point in the opposite direction: equilibrium or reversibility can support unlabeled processes, but strong labeled well-posedness requires additional structures such as tail triviality, IFC, local monotonicity, or explicit control of logarithmic derivatives (Osada et al., 2014, Kawamoto et al., 2020).

3. Interacting particle ISDEs, Coulomb gases, and random matrix limits

A major branch of ISDE theory concerns infinite interacting Brownian particles with logarithmic or Coulomb interaction. Osada’s 2010 work studies

$B$2

whose equilibrium unlabeled dynamics are the Ginibre random point field in $B$3 and Dyson’s measures in $B$4 (Osada, 2010). Because the two-dimensional Coulomb interaction is long-range, the drift does not converge absolutely, and this leads to a distinctive phenomenon: the same Ginibre diffusion satisfies plural ISDEs, including one with an explicit confining $B$5 term and one without it (Osada, 2010). This is a structural difference from Ruelle-class interactions.

For Dyson’s bulk model, Tsai proves strong existence and pathwise uniqueness for all $B$6 for the conditionally convergent system

$B$7

starting from an explicit configuration class that includes the lattice $B$8 and the sine process (Tsai, 2014). The finite-$B$9 Dyson Brownian motions converge to this ISDE, and for $A^{-1}$0 the resulting spacetime process coincides with the determinantal construction of Katori and Tanemura (Tsai, 2014).

At the soft edge, Osada and Tanemura construct Airy$A^{-1}$1 interacting Brownian motions for $A^{-1}$2 with renormalized drift

$A^{-1}$3

The unlabeled process is reversible with respect to the Airy$A^{-1}$4 random point field, and the solution is the soft-edge limit of the finite-$A^{-1}$5 Dyson SDEs (Osada et al., 2014). At the hard edge, Osada and Tanemura treat Bessel random point fields and calculate the logarithmic derivative

$A^{-1}$6

thereby identifying the labeled ISDE

$A^{-1}$7

and proving quasi-Gibbsianness of the Bessel field (Honda et al., 2014).

The Coulomb program is extended in arbitrary dimension by the construction of strong solutions and pathwise uniqueness for Coulomb random point fields in all $A^{-1}$8 and all $A^{-1}$9 (Osada et al., 29 Aug 2025). A central distinction there is that the labeled $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$0-valued diffusion has no invariant probability measure, while the associated unlabeled process is reversible with respect to the Coulomb random point field (Osada et al., 29 Aug 2025). The 2024 work on logarithmic interaction and characteristic polynomials shows path-space convergence from every initial condition, including colliding initial data, to an infinite-dimensional Feller diffusion on the enhanced state space

$$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$1

and proves convergence to equilibrium given by inverse points of the Bessel determinantal point process (Assiotis et al., 2024).

These results make random matrix theory one of the principal sources of explicit ISDEs. They also show that singular drifts, non-absolutely convergent sums, and the difference between labeled and unlabeled dynamics are not peripheral complications but defining features of the subject.

4. Forward-backward systems, SPDE interfaces, and delay or reflection

ISDEs also arise as analytical devices for stochastic PDEs and related path-dependent systems. In the KPZ setting, the mollified equation

$$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$2

is represented through an infinite-dimensional forward-backward stochastic differential equation,

$$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$3

with cylindrical Brownian motion in $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$4 and quadratic nonlinearity in $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$5 (Monter et al., 2012). Existence and uniqueness are proved by the exponential transformation $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$6, $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$7, time reversal, martingale representation, a specialized Itô formula for backward-forward integrals, and local-time arguments. The terminal value $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$8 yields a probabilistic representation of the Bertini–Cancrini–Giacomin solution of KPZ as $$V \subset H \simeq H^* \subset V^*, \qquad \mathrm{d}x(t) = [A_1(t, x(t)) + A_2(t, x_t)]\,dt + B(t, x_t)\,dW(t),$$9 (Monter et al., 2012).

A broader SPDE interface appears in the stochastic calculus framework for cylindrical Brownian motion and Stratonovich equations. Chueshov and Millet emphasize that for transport-type noise the Stratonovich-to-Itô correction

$x_t(\cdot)=x(t+\cdot)$0

may be unbounded and lose derivatives, which forces the analysis into larger ambient spaces and motivates existence theories based on Galerkin approximation, tightness, Skorokhod representation, or pathwise Cauchy schemes (Goodair, 2022).

Delay and memory produce another ISDE/SPDE hybrid. The abstract delay equation on a Gelfand triple in (Rockner et al., 2014) treats coefficients depending on the full recent history $x_t(\cdot)=x(t+\cdot)$1, and under local monotonicity yields strong solutions for classes including stochastic fractional Navier-Stokes, stochastic reaction-diffusion, and stochastic porous medium equations with delay. Reflection adds a geometric constraint: (Li et al., 4 Mar 2026) studies

$x_t(\cdot)=x(t+\cdot)$2

in the unit ball of a Hilbert space, where $x_t(\cdot)=x(t+\cdot)$3 is a bounded-variation reflection process characterized by the variational inequality above.

Forward-backward doubly stochastic equations with jumps form a further extension. In a separable Hilbert space, (Al-Hussein, 2024) proves existence and uniqueness for fully coupled systems driven by two independent cylindrical Wiener processes and a compensated Poisson random measure, using monotonicity, Lipschitz conditions, and the method of time continuation. This suggests a continuous spectrum from classical Hilbert-space SDEs to coupled, path-dependent, reflected, or doubly stochastic infinite-dimensional systems.

5. Approximation, scaling limits, and asymptotic analysis

Approximation is not merely a numerical convenience in ISDE theory; in several foundational papers it is the construction principle. The reaction-diffusion ISDE of (Costa et al., 2020) is the scaling limit of interacting particle systems $x_t(\cdot)=x(t+\cdot)$4, with rescaling $x_t(\cdot)=x(t+\cdot)$5. Tightness of the rescaled processes and identification of limit points through the martingale problem yield existence, and for finite total mass initial data the limiting weak solution is unique (Costa et al., 2020).

Random-matrix-derived ISDEs are repeatedly obtained as thermodynamic or spectral scaling limits. Dyson’s bulk ISDE is the limit of finite-$x_t(\cdot)=x(t+\cdot)$6 Dyson Brownian motions (Tsai, 2014). Airy$x_t(\cdot)=x(t+\cdot)$7 ISDEs arise from the soft-edge scaling of finite-$x_t(\cdot)=x(t+\cdot)$8 Dyson SDEs (Osada et al., 2014). The 2025 Coulomb work constructs finite-particle systems in balls with reflection and frozen exterior configuration, then passes first $x_t(\cdot)=x(t+\cdot)$9 at fixed radius and then $$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$0 to obtain the full-space Coulomb ISDE (Osada et al., 29 Aug 2025). The Rider–Valko-type model in (Assiotis et al., 2024) goes further by proving path-space convergence in $$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$1 from every initial condition and by coupling the finite-$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$2 and infinite systems so that

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$3

for each fixed $$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$4 (Assiotis et al., 2024).

The asymptotic theory also includes large deviations. Kurtz and Xiong establish large deviation principles for stochastic integrals and SDEs driven by infinite-dimensional semimartingales under a uniform exponential tightness condition, with rate functions of the form

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$5

and the corresponding SDE version

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$6

The class of drivers includes Banach-space-valued semimartingales and martingale random measures (Ganguly, 2011).

At the numerical level, Buckwar, Tambue, and colleagues analyze mean-square stability for approximations of linear stochastic evolution equations. For the fully discrete recursion

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$7

asymptotic mean-square stability is characterized by

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$8

They apply this to spectral Galerkin and finite element discretizations combined with Euler–Maruyama, Milstein, Crank–Nicolson, forward Euler, and backward Euler methods (Lang et al., 2017).

6. Control, inference, and broader methodological directions

ISDEs increasingly interact with optimization, control, and statistical inference. In Bayesian nonparametric learning of SDE drift, (Ganguly et al., 2022) studies optimization problems on a Hilbert space,

$$\int_0^t B(s)\, d\mathcal W_s = \sum_{i=1}^\infty \int_0^t B_s(e_i)\, dW_s^i,$$9

and derives generalized representer-type results that recover finite expansions as special cases. Applied to drift learning for

$\mathcal V$0

this yields

$\mathcal V$1

after which Bayesian hierarchical priors such as multivariate $\mathcal V$2 and Horseshoe-type priors are used for sparse learning and uncertainty quantification (Ganguly et al., 2022).

In control theory, (Qiu et al., 2023) considers infinite-dimensional path-dependent systems on a Gelfand triple,

$\mathcal V$3

with random and path-dependent coefficients. The associated value function is a random field on path space and is characterized as the unique viscosity solution of the stochastic path-dependent Hamilton-Jacobi equation

$\mathcal V$4

where

$\mathcal V$5

The viscosity solution theory uses semimartingale decompositions of path-dependent functionals, stopping-time-based extremality, and compactness through the Gelfand triple (Qiu et al., 2023).

A final methodological theme is that the main obstacles in ISDEs are often structural rather than purely technical. The IFC condition remains difficult to verify (Kawamoto et al., 2020); logarithmic and Coulomb drifts may fail to converge absolutely and can admit plural SDE representations (Osada, 2010); labeled and unlabeled processes need not share the same invariant-measure behavior (Osada et al., 29 Aug 2025). This suggests that the subject is best understood not as a straightforward extension of finite-dimensional SDE theory, but as a collection of interacting frameworks—stochastic calculus in infinite-dimensional spaces, random point fields, SPDE methods, and asymptotic or variational constructions—organized around the problem of making infinitely many stochastic degrees of freedom analytically tractable.