Stackelberg-Mean Field Equilibrium

• Stackelberg-Mean Field Equilibrium is a hierarchical framework where a leader commits to a policy and followers respond via a mean field Nash equilibrium.
• It encompasses diverse mathematical formulations, including continuous-time and discrete-time models, relaxed control, and information asymmetry approaches.
• Applications span finance, public policy, and systemic risk, with computational techniques ranging from Riccati equations to neural solvers for approximating ε-equilibria.

Stackelberg-Mean Field Equilibrium (S-MFE) denotes a hierarchical equilibrium regime in which a leader commits to a policy, the followers respond through a mean field Nash equilibrium, and the leader optimizes against the induced mean field response. In the literature, the label covers several closely related objects: an optimal leader policy together with the followers’ induced mean field equilibrium in continuum-population models; decentralized limits whose finite-$N$ implementations are only $\varepsilon$-Stackelberg equilibria; and, in some papers, mean-field-type Stackelberg differential games in which the “mean field” is an expectation term rather than a population fixed point. The unifying structure is therefore not a single canonical equation, but a leader–follower optimization architecture coupled with a mean-field consistency condition (Dayanikli et al., 2023, Moon et al., 2019).

1. Equilibrium architecture

The most standard S-MFE object in continuous time is the pair formed by a leader policy and the followers’ induced mean field Nash equilibrium. In the formulation of “A Machine Learning Method for Stackelberg Mean Field Games,” the principal solves

$$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$

where $\mathcal N(\boldsymbol\lambda)$ is the followers’ mean field Nash equilibrium set under policy $\boldsymbol\lambda$ (Dayanikli et al., 2023). In discrete time, “Master Equation for Discrete-Time Stackelberg Mean Field Games with single leader” defines a Stackelberg mean field equilibrium as a triple $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$ satisfying

$$\tilde{\sigma}^f\in BR^f(z,\tilde{\sigma}^l),\qquad z=\Lambda(\tilde{\sigma}^l,\tilde{\sigma}^f),\qquad \tilde{\sigma}^l\in BR^l(z),$$

so follower optimality, mean-field consistency, and leader optimality are imposed simultaneously (Vasal et al., 2022).

This equilibrium is hierarchical rather than symmetric. The followers play a Nash game conditional on the leader’s action or policy, while the leader solves an upper-level control or optimization problem through the followers’ equilibrium map. In that sense, S-MFE differs from ordinary mean field equilibrium, where all agents are strategically symmetric, and from ordinary Stackelberg differential games, where the lower level usually involves finitely many followers rather than a continuum or large population (Dayanikli et al., 2023, Vasal et al., 2022).

A recurrent source of confusion is that some papers use “mean field” in a weaker sense. “Linear-Quadratic Time-Inconsistent Mean-Field Type Stackelberg Differential Games: Time-Consistent Open-Loop Solutions” studies one leader and one follower, with conditional expectations of state and control in the costs, and explicitly notes that this is a “mean-field type Stackelberg differential game” rather than a mean field game with a continuum-population fixed point (Moon et al., 2019). Accordingly, not every Stackelberg problem with expectation terms is an S-MFE in the Lasry–Lions or Huang–Caines–Malhamé sense; some belong instead to the Stackelberg mean-field-type control literature.

2. Mathematical formulations

The baseline continuum-population formulation is stochastic, finite-horizon, and continuous-time. In (Dayanikli et al., 2023), a representative follower with state $X_t\in\mathbb R^m$ evolves by

$$dX_t=b(t,X_t,\alpha_t,\mu_t;\lambda(t,X_t))\,dt+\sigma\,dW_t, \qquad X_0=\zeta,$$

with follower objective

$$J^{\boldsymbol\lambda}(\boldsymbol\alpha,\boldsymbol\mu) = \mathbb E\!\left[\int_0^T f(t,X_t,\alpha_t,\mu_t;\lambda(t,X_t))\,dt +g(X_T,\mu_T;\lambda(T,X_T))\right].$$

Here the mean field is the population law $\varepsilon$0, and the fixed-point condition is $\varepsilon$1 under the optimal follower response (Dayanikli et al., 2023). This is the canonical principal–continuum-agents Stackelberg MFG model.

Several papers enrich this structure by changing the information flow. In “Mean Field Games in a Stackelberg problem with an informed major player,” the leader privately observes a type $\varepsilon$2, chooses a randomized control law depending on $\varepsilon$3, and the followers update the posterior

$\varepsilon$4

which becomes an endogenous common-information process. Conditional on that posterior process, followers solve a mean field game with random common environment, and the leader optimizes through the induced conditional population law $\varepsilon$5 (Bergault et al., 2023). This yields an S-MFE with asymmetric information and signaling.

Other formulations replace fully observed states by filtering structures. In (Si et al., 2024), neither leader nor followers observe states directly; they observe $\varepsilon$6 and $\varepsilon$7, and the decentralized controls are adapted to $\varepsilon$8 and $\varepsilon$9, even though the state dynamics are driven also by common noise $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$0. The mean field is then conditional: $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$1 In (Si et al., 20 Mar 2025), the mean field coupling contains both follower average terms $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$2 and expectation terms $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$3, and the limiting consistency variables are expressed through filtered estimates under partial observation.

A different branch of the literature retains the leader–follower hierarchy but departs from genuine continuum-population mean field games. In (Yang et al., 13 May 2026), the mean field enters through $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$4 and $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$5 inside linear-quadratic state equations and costs with random coefficients. The equilibrium is still hierarchical and mean-field consistent, but the interaction is through endogenous expectations rather than a distributional fixed point. Similarly, (Moon et al., 2019) uses conditional expectations $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$6, $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$7, and $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$8 in a two-player leader–follower model; the paper explicitly separates this mean-field-type setting from genuine MFG fixed-point models.

3. Solution concepts and analytical characterizations

S-MFE has no single universal solution method because the equilibrium notion itself varies with time consistency, equilibrium multiplicity, and information structure. In standard Stackelberg MFG models with unique follower equilibrium, a common route is to characterize the followers first and then optimize the leader on top of that response map. In (Dayanikli et al., 2023), the followers’ equilibrium under a fixed leader policy is represented by an MKV-FBSDE with consistency $$\boldsymbol\lambda^* \in \arg\min_{\boldsymbol\lambda\in\boldsymbol\Lambda} J^0(\boldsymbol\lambda), \qquad (\hat{\boldsymbol\alpha}^{\boldsymbol\lambda^*},\hat{\boldsymbol\mu}^{\boldsymbol\lambda^*})\in \mathcal N(\boldsymbol\lambda^*),$$9, after which the bilevel Stackelberg problem is reformulated as a penalized single-level McKean–Vlasov control problem.

In discrete time, (Vasal et al., 2022) derives a master-equation-type recursive characterization on the common-information state $\mathcal N(\boldsymbol\lambda)$0, where $\mathcal N(\boldsymbol\lambda)$1 is the public belief about the leader’s private state and $\mathcal N(\boldsymbol\lambda)$2 is the mean field. The backward recursion computes equilibrium prescription mappings $\mathcal N(\boldsymbol\lambda)$3, while the forward recursion updates $\mathcal N(\boldsymbol\lambda)$4 and $\mathcal N(\boldsymbol\lambda)$5. The paper proves that solving this stagewise fixed-point system is equivalent to computing all discrete-time SMFE (Vasal et al., 2022).

When follower equilibrium is non-unique, the upper-level problem can no longer treat the lower-level map as single-valued. “Optimization frameworks and sensitivity analysis of Stackelberg mean-field games” formulates a pessimistic leader objective

$\mathcal N(\boldsymbol\lambda)$6

and shows that the Stackelberg-MFG can be recast as a minimax constrained optimization problem over occupation measures, leader state occupancies, and LP/KKT variables for the followers’ MDP (Guo et al., 2022). This is a materially different S-MFE regime: the leader optimizes against the worst follower equilibrium selection rather than a unique equilibrium map.

Relaxation becomes necessary in signaling models. In (Bergault et al., 2023), the leader’s strict control problem is compactified into an optimization over laws of posterior martingale processes $\mathcal N(\boldsymbol\lambda)$7, and the relaxed objective is

$\mathcal N(\boldsymbol\lambda)$8

The corresponding follower equilibrium is encoded by a common-noise HJB–Fokker–Planck system. The existence theorem is therefore for a relaxed Stackelberg solution, not directly for strict controls (Bergault et al., 2023).

Time inconsistency produces yet another equilibrium notion. In (Moon et al., 2019), the paper explicitly rejects standard precommitted Stackelberg optimality and instead defines a time-consistent adapted open-loop equilibrium by local spike-deviation inequalities for both follower and leader. This yields equilibrium controls rather than global optima, with feedback representations obtained only after Riccati decoupling. A plausible implication is that “S-MFE” cannot be treated as a single concept across the literature without specifying whether one means precommitment, relaxed control, local time-consistent equilibrium, or exact bilevel optimality.

4. Structural variants and extensions

A large part of the S-MFE literature concerns structural departures from the simplest one-leader/continuum-followers model. One direction is partial observation and common noise. In (Si et al., 2024), both leader and followers act under partial information, the state equations contain control-dependent diffusions, and the common noise is not directly observed. In (Si et al., 20 Mar 2025), both leader and followers are partially observed, the drift terms contain both average-state and expectation terms, and the decentralized controls are derived via state decomposition and a backward separation principle. In both papers, the finite-$\mathcal N(\boldsymbol\lambda)$9 outcome is an $\boldsymbol\lambda$0-Stackelberg-Nash equilibrium rather than an exact equilibrium for the original game.

A second direction is large-population Stackelberg games with nonstandard state architectures. “Direct Approach of Linear-Quadratic Stackelberg Mean Field Games of Backward-Forward Stochastic Systems” studies a backward leader and a large number of forward followers. The limiting equilibrium is obtained by decoupling a six-dimensional FBSDE, and the resulting decentralized finite-$\boldsymbol\lambda$1 strategy profile is an $\boldsymbol\lambda$2-Stackelberg equilibrium with

$\boldsymbol\lambda$3

The paper frames this as a leader commitment problem whose lower level is a mean field game and whose upper level is an induced FBSDE control problem for the leader (Cong et al., 2024).

A third direction introduces extra hierarchy. “Dynamic Data Pricing: A Mean Field Stackelberg Game Approach” builds a three-tier game with a buyer, a broker, and a competitive seller population. Sellers play an MFG under the broker’s price, the broker solves a Stackelberg problem against that seller response, and the buyer solves a top-level Stackelberg problem against both lower layers. The final object is an $\boldsymbol\lambda$4-Stackelberg equilibrium with

$\boldsymbol\lambda$5

so the paper generalizes S-MFE to a multi-level hierarchy rather than a single leader–single population structure (Bo et al., 25 Dec 2025).

A fourth direction is random-coefficient and nonlinear mean-field-type games. In (Yang et al., 13 May 2026), random adapted coefficients obstruct deterministic Riccati decoupling, so the follower response is represented as an affine operator

$\boldsymbol\lambda$6

and the leader then solves a generalized stochastic LQ problem with operator-valued coefficients. In (Huang et al., 11 Feb 2025), the leader’s state equation after follower substitution is a fully coupled conditional mean-field FBSDE, and the limiting decentralized solution is proved to generate an approximate Stackelberg equilibrium of the nonlinear finite-$\boldsymbol\lambda$7 game. These papers suggest that the S-MFE idea persists beyond classical deterministic-coefficient LQ settings, but only after substantial modification of the analytical machinery.

5. Computation, learning, and applications

Computation of S-MFE ranges from Riccati integration to neural methods. In the time-inconsistent mean-field-type model of (Moon et al., 2019), the equilibrium controls are represented through nonsymmetric coupled Riccati differential equations and BSDEs, and the numerical section uses Euler’s method in scalar resource-allocation examples. The computations indicate no finite escape time and numerically support solvability of the coupled Riccati systems.

In true continuum Stackelberg MFGs, (Dayanikli et al., 2023) converts the bilevel problem into a single-level penalized McKean–Vlasov optimal control problem. The numerical solver combines particle approximation, Euler–Maruyama time discretization, and neural parameterizations $\boldsymbol\lambda$8, $\boldsymbol\lambda$9, and $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$0, with an additional RNN/LSTM for path-dependent contracts. The paper proves convergence of penalized near-minimizers to near-minimizers of the original Stackelberg problem and gives error controls for finite population, time discretization, and neural approximation layers (Dayanikli et al., 2023).

The most explicit recent computational framework is “Stochastic Mean-Field LQ Stackelberg Differential Games with Random Coefficients: Theory and a Deep FBSDE Picard Solver” (Yang et al., 13 May 2026). Its Deep FBSDE Picard Solver preserves Stackelberg order through four stages: follower-response learning, response-sensitivity extraction, leader optimization, and neural augmented Lagrangian enforcement of mean-field consistency. The reported diagnostics drive the follower BSDE residual to $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$1, the leader residual to $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$2, and all mean-field consistency violations below $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$3. The ablation study shows that removing the follower response sensitivity $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$4 worsens leader cost by $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$5, and the self-convergence slope under constant coefficients is about $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$6, consistent with first-order Euler–Maruyama discretization (Yang et al., 13 May 2026).

Applications span contract theory, systemic risk, public policy, epidemic control, technology adoption, finance, and data markets. The contract and systemic-risk examples in (Dayanikli et al., 2023) illustrate principal regulation of a population mean field Nash equilibrium. The discrete-time master-equation paper studies vaccine subsidy design and technology pricing (Vasal et al., 2022). The signaling paper treats an informed major player whose control path reveals information to a mean field population (Bergault et al., 2023). The three-tier data-pricing model in (Bo et al., 25 Dec 2025) analyzes buyer–broker–seller interaction under common noise. The random-coefficient LQ paper compares Stackelberg and Nash-type outcomes in a mean-variance portfolio game (Yang et al., 13 May 2026). This suggests that S-MFE is best viewed not as a narrow equilibrium definition, but as a reusable hierarchical modeling template for strategic control of large populations.

6. Limitations, misconceptions, and open issues

The main conceptual limitation is terminological nonuniformity. Some papers use “Stackelberg mean field game,” some “Stackelberg-Nash equilibrium,” some “decentralized Stackelberg equilibrium,” and some study only mean-field-type expectation couplings rather than genuine continuum fixed points. The two-player time-inconsistent model of (Moon et al., 2019) is the clearest warning: it is directly relevant to Stackelberg mean-field-type equilibrium, but it is not a mean field game with a continuum of strategic agents. Treating all such objects as interchangeable S-MFE is therefore inaccurate.

A second limitation is dependence on strong solvability assumptions. Many results are conditional on nonsymmetric or asymmetric Riccati equations, matrix invertibility conditions, or uniqueness of the followers’ equilibrium. This is explicit in (Moon et al., 2019, Cong et al., 2024, Si et al., 2024), and (Bo et al., 25 Dec 2025). In (Dayanikli et al., 2023), the admissible leader policies are restricted to those for which the follower equilibrium is unique, and the paper explicitly notes that equilibrium selection under non-uniqueness is left open. In (Guo et al., 2022), non-uniqueness is handled by worst-case equilibrium selection, but the price is a pessimistic minimax formulation with substantial sensitivity to perturbations.

A third limitation concerns robustness and finite-$(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$7 justification. Some papers prove explicit approximate-equilibrium theorems for the original finite-player game, while others do not. Approximation theorems of order $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$8 or $(\tilde{\sigma}^f,\tilde{\sigma}^l,z)$9 are available in several LQ or hierarchical models (Cong et al., 2024, Bo et al., 25 Dec 2025), but (He et al., 19 Sep 2025) explicitly motivates mean field as a large-population approximation without proving a finite-$$\tilde{\sigma}^f\in BR^f(z,\tilde{\sigma}^l),\qquad z=\Lambda(\tilde{\sigma}^l,\tilde{\sigma}^f),\qquad \tilde{\sigma}^l\in BR^l(z),$$0 convergence theorem. This suggests a methodological split between equilibrium-characterization papers and finite-population-implementation papers.

A fourth limitation is computational rather than analytical. Even when a rigorous equilibrium characterization exists, the actual numerical method may remain heuristic. In (Dayanikli et al., 2023), SGD/Adam optimization is empirical rather than globally guaranteed. In (Yang et al., 13 May 2026), the discrete learned equilibrium is validated numerically by residuals, consistency violations, and unilateral-deviation tests, not by a full verification theorem for the neural scheme. Inference from these results should therefore be calibrated: they establish practical computability in substantial model classes, but not a universal numerical theory of S-MFE.

Taken together, the literature supports a precise but plural view. S-MFE is a family of hierarchical equilibrium constructions built around leader commitment, follower mean field equilibrium, and consistency of the induced aggregate state. The exact mathematical object may be a triple $$\tilde{\sigma}^f\in BR^f(z,\tilde{\sigma}^l),\qquad z=\Lambda(\tilde{\sigma}^l,\tilde{\sigma}^f),\qquad \tilde{\sigma}^l\in BR^l(z),$$1, a policy–distribution pair, a relaxed law of posteriors, a time-consistent open-loop equilibrium, or an $$\tilde{\sigma}^f\in BR^f(z,\tilde{\sigma}^l),\qquad z=\Lambda(\tilde{\sigma}^l,\tilde{\sigma}^f),\qquad \tilde{\sigma}^l\in BR^l(z),$$2-equilibrium justified by a mean field limit. What unifies these formulations is the combination of Stackelberg hierarchy and mean field consistency; what differentiates them is the treatment of information, equilibrium selection, and implementability (Dayanikli et al., 2023, Guo et al., 2022).