Common Noise McKean–Vlasov Problems

• Common-noise McKean–Vlasov problems are weak formulations of mean-field dynamics where coefficients depend on the conditional law given shared randomness, typically a common Brownian motion.
• They are analyzed via generator-level approaches, martingale characterizations, and stochastic Fokker–Planck equations, which capture the evolution of random measure flows.
• The framework influences stability and control, leading to particle approximations and finite-dimensional reductions that are critical for numerical schemes and dynamic programming.

Common-noise McKean–Vlasov martingale problems are weak formulations of mean-field stochastic dynamics in which the coefficients depend on a conditional law given a shared source of randomness, typically a common Brownian motion. In contrast with the no-common-noise case, the law process is itself random, and the state equation, the associated measure-valued evolution, and the martingale characterization must all be formulated relative to the common-noise filtration. Across stochastic control, mean field games, stochastic differential games, and singular interacting systems, this framework appears through conditional McKean–Vlasov SDEs, stochastic Fokker–Planck equations, generators on spaces of probability measures, and verification principles expressed as martingale or submartingale conditions (Hammersley et al., 2019).

1. Conditional-law formulation and basic objects

A prototypical common-noise McKean–Vlasov SDE has the form

$$dX_t = b(X_t,\mu_t)\,dt + \sigma(X_t)\,dB_t + \sigma_0\,dW_t, \qquad \mu_t := \mathcal{L}(X_t\mid\mathcal{F}^W_t),$$

where $B$ is idiosyncratic noise and $W$ is common noise. In this formulation, the interaction is through the conditional law rather than the unconditional distribution, so $\mu_t$ is a random probability measure flow adapted to the common-noise filtration (Jianhai et al., 2024). A more path-dependent version allows

$$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$

which makes explicit that the mean-field argument may be a conditional law on path space (Hammersley et al., 2019).

The common-noise setting is equivalently described at the particle level. A standard $N$-particle system is

$$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$

and the large-$N$ limit yields a random measure-valued process $\mu_t$ rather than a deterministic flow, precisely because the empirical measure remains adapted to the common noise (Kolokoltsov et al., 2015). The same structural feature appears in controlled settings, where the coefficients may depend on the conditional law of the state and even on the conditional law of the state-control pair given the common-noise filtration (Djete et al., 2019).

A common misconception is that common noise merely replaces deterministic measure flows by random ones without altering the analytic structure. The literature shows that this is too narrow. Common noise changes the filtration, the notion of law consistency, the form of the generator, and the correct state space for dynamic programming and martingale problems; in several settings it also changes the long-time qualitative behavior, sometimes stabilizing the dynamics and sometimes inducing singular events such as blow-ups (Maillet, 2023).

2. Generator, SPDE, and martingale characterizations on measure space

For a fixed measure $\mu$, the single-particle generator is typically

$B$0

or, in the multidimensional diffusion setting with common and idiosyncratic noises,

$B$1

The corresponding martingale formulation requires that for each smooth test function,

$B$2

is a martingale, with $B$3 or its path-space analogue (Jianhai et al., 2024, Maillet, 2023).

At the law level, the conditional measure flow solves a stochastic nonlinear Fokker–Planck equation. In weak form, one representative formulation is

$B$4

with

$B$5

so the law process is itself a measure-valued diffusion driven by the common Brownian motion (Kolokoltsov et al., 2015). In density form, the same object becomes a stochastic Fokker–Planck SPDE for $B$6, and this equivalence is one of the standard bridges between SDE and martingale formulations (Hammersley et al., 2019).

A more intrinsic formulation acts directly on functionals $B$7 of measures. For smooth enough $B$8, the limiting generator in the presence of common noise is

$B$9

and the martingale condition is

$W$0

is a martingale (Kolokoltsov et al., 2015). The second-order term in measure space is the signature of common noise: it has no analogue in deterministic law flows.

This generator-level viewpoint is central in long-time analysis as well. For nonlinear stochastic Fokker–Planck equations with common noise, invariant measures on $W$1 are characterized by

$W$2

where $W$3 is the measure-space generator containing both drift terms and the common-noise-induced second-order term in the measure variable (Maillet, 2023).

3. Weak solutions, compatibility, and controlled martingale problems

In weak formulations with common noise, the dependence structure between $W$4, the common noise, and the conditional law process cannot be left implicit. A central notion is compatibility: $W$5 must be compatible with $W$6, and $W$7 must be compatible with $W$8, so that the conditional law constraint

$W$9

remains meaningful and stable under weak convergence (Hammersley et al., 2019). This avoids pathological weak limits in which the process labelled as “conditional law” fails to be the actual conditional distribution of the state.

Under boundedness and joint continuity of coefficients, weak existence can be proved by compactness of Euler approximations while keeping track of the full dependence structure of $\mu_t$0; weak uniqueness can then be obtained under non-degeneracy of the idiosyncratic diffusion and drift regularity in total variation via a Monge–Kantorovich cost function and Girsanov representation (Hammersley et al., 2019). In this sense, the common-noise martingale problem is well posed as a genuinely nonlinear weak problem with random coefficients.

Controlled problems admit an analogous reformulation on canonical spaces with two filtrations. In the weak formulation of McKean–Vlasov optimal control with common noise, one works with a global filtration $\mu_t$1 and a common-noise filtration $\mu_t$2, and sets

$\mu_t$3

The controlled state satisfies

$\mu_t$4

and the canonical martingale problem is expressed in terms of a lifted generator acting on $\mu_t$5 together with an explicit conditional-law process $\mu_t$6 on path space (Djete et al., 2019).

This two-filtration framework is technically significant because conditioning, concatenation, and measurable selection must all be performed relative to the common-noise filtration. The dynamic programming principle then takes the form

$\mu_t$7

with the state variable at the stopping time given by the updated conditional law $\mu_t$8 (Djete et al., 2019). A plausible implication is that, in common-noise settings, the measure-valued component is not merely a parameter in the coefficients but an essential part of the Markov state.

4. Weak martingale optimality in linear-quadratic and game-theoretic models

Linear-quadratic common-noise McKean–Vlasov models provide a particularly explicit realization of the martingale-problem viewpoint. In the stochastic control setting, the state dynamics are linear in the state, its mean, the control, and its mean, and the cost depends quadratically on both deviations from the mean and the mean itself. In the common-noise extension, unconditional expectations are replaced by conditional expectations given the common-noise filtration, for example

$\mu_t$9

and the coefficients may be $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$0-adapted (Miller et al., 2018).

The defining method is a weak martingale optimality principle. For a candidate control profile, one constructs value-like processes $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$1 and then

$$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$2

with the property that $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$3 has nonnegative drift for arbitrary controls and zero drift at equilibrium (Miller et al., 2018). In the LQ case one uses a quadratic ansatz in $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$4, $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$5, and linear terms, and drift cancellation yields Riccati equations together with mean-field BSDEs.

In the common-noise case, the Riccati objects become backward stochastic Riccati equations adapted to the common-noise filtration,

$$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$6

and the linear terms are governed by a mean-field BSDE driven by both idiosyncratic and common noises (Miller et al., 2018). The martingale property is therefore imposed not on a deterministic value functional but on a random functional adapted to the common noise.

The same logic appears in LQ McKean–Vlasov control. There, the weak martingale approach leads to verification via suitably chosen processes whose drift vanishes under the optimal control, with the common-noise extension replacing unconditional means by conditional means with respect to $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$7 and adding a common-noise diffusion term (Basei et al., 2018). This suggests a general pattern: in tractable common-noise models, equilibrium or optimality is often encoded by the vanishing of the drift of a carefully designed semimartingale functional of $$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$8.

5. Qualitative phenomena: blow-ups, contractivity, ergodicity, and uniqueness recovery

Common noise does not have a single qualitative effect. In singular models with positive feedback it can trigger or prevent discontinuities. In the conditional McKean–Vlasov contagion model

$$X_t = \xi +\int_0^t b(s,X_{\cdot\wedge s},\mu_s)\, ds +\int_0^t \sigma(s,X_{\cdot\wedge s},\mu_s)\, dW_s +\int_0^t \rho(s,X_{\cdot\wedge s},\mu_s)\,dB_s, \quad \mu_s = \mathscr L\big(X_{\cdot\wedge s}\mid \mathcal{F}_s^{B,\mu}\big),$$9

the loss process $N$0 may jump even though the driving noises are continuous, and the common noise can either provoke or avert such blow-ups depending on the path realization (Ledger et al., 2018). The limiting relaxed formulation uses a random probability measure $N$1 satisfying

$N$2

together with a minimal jump condition

$N$3

where $N$4 is the surviving-mass flow (Ledger et al., 2018). This is a martingale-problem framework with both continuous and jump components at the measure level.

At the opposite end of the spectrum, common noise may improve long-time stability. For one-dimensional McKean–Vlasov SDEs with common noise,

$N$5

a new asymptotic coupling by reflection yields exponential contractivity of the random measure flow: $N$6 under dissipativity and non-degeneracy assumptions, with the paper emphasizing that both the common noise and the idiosyncratic noise facilitate the exponential contractivity of the associated measure-valued processes (Jianhai et al., 2024).

A related phenomenon appears for nonlinear stochastic Fokker–Planck equations with common noise and linear dependence on the measure. In uniformly convex settings, one has uniform-in-time conditional propagation of chaos and uniqueness of the invariant measure $N$7, together with exponential convergence

$N$8

where $N$9 (Maillet, 2023). More strikingly, in certain non-convex regimes, sufficiently strong common noise restores uniqueness of the long-time behavior; the paper describes this as a uniqueness recovery phenomenon induced by common noise (Maillet, 2023).

These results contradict the idea that common noise is only a source of additional randomness. Depending on the structure, it can generate singular mass cascades, enforce contractivity through coupling, or restore uniqueness in regimes where deterministic McKean–Vlasov dynamics admit multiple invariant measures (Ledger et al., 2018, Maillet, 2023).

6. Particle limits, numerical schemes, and finite-dimensional reductions

The particle approximation of common-noise McKean–Vlasov martingale problems is by now standard, but the common-noise setting changes both the limiting object and the convergence mode. For symmetric $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$0-particle systems, the generator on functionals of the empirical measure $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$1 takes the form

$$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$2

where $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$3 is the generator of the limiting measure-valued diffusion and $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$4 is an order-$$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$5 correction (Kolokoltsov et al., 2015). This leads to explicit $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$6-propagator estimates and $$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$7-Nash equilibrium results in mean field games with common noise when the master equation is regular enough (Kolokoltsov et al., 2015).

For direct numerical approximation of common-noise McKean–Vlasov equations, one may combine an Euler time discretization with a particle discretization of the conditional law. Under Hölder continuity in time and Lipschitz continuity in state and measure arguments, the Euler approximation satisfies

$$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$8

and the full particle approximation satisfies

$$dX_t^i = b(t, X_t^i,\mu_t^N,u_t^i)\,dt + \sigma_{\text{ind}}(X_t^i)\,dB_t^i + \sigma_{\text{com}}(X_t^i)\,dW_t, \qquad \mu_t^N = \frac{1}{N}\sum_{j=1}^{N}\delta_{X_t^j},$$9

with $N$0 given by empirical-measure Wasserstein rates (Gall, 2024). In super-linear settings, explicit tamed Euler and tamed Milstein schemes for interacting particle systems with common noise achieve strong convergence of order $N$1 and $N$2, respectively (Kumar et al., 2020).

A different approximation strategy applies when the law dependence closes on finitely many conditional moments. In polynomial conditional McKean–Vlasov control with common noise, the flow of conditional laws can be reduced to a finite-dimensional Markov state consisting of conditional moments such as

$N$3

leading to a finite-dimensional stochastic control problem driven only by the common noise (Balata et al., 2018). This reduction turns an infinite-dimensional conditional-law problem into a standard Markov control problem amenable to quantization, regression by control randomization, and regress-later schemes (Balata et al., 2018).

A plausible implication is that numerical treatment of common-noise McKean–Vlasov martingale problems bifurcates into two broad regimes. In general models, one works with interacting particle systems, propagation of chaos, and SPDE approximations; in structurally closed models, one may bypass measure-valued numerics altogether by projecting onto a finite-dimensional common-noise state (Gall, 2024, Balata et al., 2018).