Leader-Follower Multi-Agent Systems

• Leader-Follower (LeFo) paradigms are hierarchical multi-agent systems where designated leaders influence followers through dynamics, incentives, and information asymmetry.
• They are implemented across diverse models including consensus networks, Stackelberg games, density control, and robotics, enabling tailored control strategies and robust performance analyses.
• Advanced methodologies employ analytical tools like Riccati equations, forward-backward systems, and graphon aggregation to quantify system performance under varying conditions.

Searching arXiv for recent and foundational leader-follower papers relevant to the requested encyclopedia entry. arXiv.search(query="leader-follower multi-agent systems Stackelberg graphon density control navigation", max_results=10) Leader–Follower (LeFo) denotes a broad family of hierarchical multi-agent formulations in which a designated leader, a leader set, or a controllable leader population influences followers through dynamics, incentives, information, or geometry. Across the literature, LeFo appears in several technically distinct forms: leaders may be nodes with exogenous inputs in consensus networks, first movers in Stackelberg games, controllable densities in continuum models, robots defining a moving boundary for a swarm, or agents whose status is reassigned online by state-dependent rules (Sato, 2018, Maffettone et al., 2024, Chen et al., 30 Jan 2026, Vahedifar et al., 10 Jul 2025, Park et al., 2020).

1. Conceptual scope and recurring abstractions

The common structure is hierarchical but not uniform. In graph-based control, leaders are agents receiving additive inputs while followers obey interaction laws. In Stackelberg formulations, the leader commits first and followers respond through best responses or Nash equilibria. In density-control and mean-field models, leaders form a controllable population that shapes the motion or distribution of followers. In robotics, leaders can be physical boundary agents, a human operator, or a designated robot carrying privileged information. In synchronization and opinion dynamics, “leader” may denote either a fixed target-bearing subgroup or a status assigned adaptively from instantaneous state variables (Sato, 2018, Maffettone et al., 2024, Chen et al., 30 Jan 2026, Li, 2021, Park et al., 2020).

Setting	Leader role	Representative formulation
Consensus networks	Nodes with external inputs	$\dot{x}=-Lx+Mu$
Density control	Controllable density guiding followers	$\partial_t \rho^L+\partial_x(\rho^L u)=0$
Stackelberg games	First mover anticipating follower response	$S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$
RL	Sequential action within each stage	Leader acts first, follower observes and acts
Synchronization	Status-dependent asymmetric coupling	Leaders influenced only by followers

The state spaces also vary sharply. Some formulations use scalar agent states on weighted graphs, some use mixed strategies on simplices, some use densities on periodic domains, and some use stochastic processes indexed by a graphon or a population label. A plausible implication is that LeFo is better understood as a relational template—hierarchical influence with asymmetric information or actuation—than as a single model class.

2. Networked control, consensus, and formation

In leader–follower consensus networks, the canonical linear model is

$$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$

on a connected, weighted, undirected graph, where $M$ selects leader nodes and the output measures edge disagreement (Sato, 2018). In that setting, leader demotion and reselection were posed as minimization of the $H^2$ norm between original and modified transfer functions. The main theorem is unusually sharp: for a fixed demotion set, the globally optimal new leader set is exactly the original leader set minus the demoted leaders, and for a fixed number $r$ of demotions, any subset of $r$ leaders is optimal. The relative error is

$$\frac{\|G-\tilde G\|_{H^2}}{\|G\|_{H^2}}=\sqrt{r/m},$$

so the relative $H^2$ error depends only on the number of original leaders $\partial_t \rho^L+\partial_x(\rho^L u)=0$0 and the number demoted $\partial_t \rho^L+\partial_x(\rho^L u)=0$1, not on the number of agents $\partial_t \rho^L+\partial_x(\rho^L u)=0$2 (Sato, 2018).

Other LeFo performance analyses emphasize topology rather than leader reselection. In directed lattices with a virtual leader and stochastic disturbances, the steady-state variance of follower deviation grows as a square-root function in $\partial_t \rho^L+\partial_x(\rho^L u)=0$3D and a logarithmic function in $\partial_t \rho^L+\partial_x(\rho^L u)=0$4D, while remaining bounded in $\partial_t \rho^L+\partial_x(\rho^L u)=0$5D (Lin, 2016). That result links LeFo robustness to spatial dimension and directed information flow, and it gives an explicit asymptotic distinction between fragile low-dimensional leader guidance and bounded high-dimensional behavior.

Formation control with transient guarantees leads to a different topological question: when can leader inputs enforce prescribed performance bounds on all relative errors? For first-order formation dynamics on static undirected graphs, prescribed performance control is enabled not by standard controllability tests, but by cycle structure, edge neighborhoods, follower–leader–follower paths, and the “maximum follower-end subgraph” $\partial_t \rho^L+\partial_x(\rho^L u)=0$6 (Chen et al., 2023). The paper derives necessary conditions for trees and general graphs and then necessary and sufficient conditions for leader-follower formation control under transient constraints. This is a direct correction to a common oversimplification: the mere presence of leaders does not guarantee transient controllability in the prescribed-performance sense.

3. Density, mean-field, and continuum LeFo

A major line of work replaces individual trajectories by densities. In one-dimensional leader-follower density control, the leader density $\partial_t \rho^L+\partial_x(\rho^L u)=0$7 and follower density $\partial_t \rho^L+\partial_x(\rho^L u)=0$8 evolve on the periodic domain $\partial_t \rho^L+\partial_x(\rho^L u)=0$9 according to

$S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$0

with $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$1 (Maffettone et al., 2024). Feasibility is characterized through a steady-state deconvolution from a desired follower profile $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$2 to a leader reference density $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$3, and two architectures are given: feed-forward on followers with feedback on leaders, and a dual-feedback reference-governor architecture. In both cases, exponential convergence follows under explicit bounds such as $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$4 in one dimension, or $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$5 in higher dimensions (Maffettone et al., 2024).

When follower–follower interactions are included explicitly, feasibility becomes sharper. In the interacting-follower model, the steady-state leader field must offset both diffusion and follower–follower velocity,

$S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$6

and the existence of a nonnegative leader density induces necessary and sufficient bounds

$S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$7

together with a compatibility condition at points where $S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$8 (Lorenzo et al., 13 Apr 2026). The same paper gives a local stability condition

$S_F^*(S_L)=\arg\max_{s_F}U_F(S_F,S_L)$9

which tightens the non-interacting threshold and makes the dependence on interaction strength explicit (Lorenzo et al., 13 Apr 2026).

Robust multi-scale density control adds unknown bounded perturbations at the microscopic level and derives a coupled continuum model with $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$0 and $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$1 drift uncertainty. There the followers’ density can still be regulated asymptotically by a feedback law using a sign term and a feed-forward velocity, provided feasibility and timescale separation hold. An explicit lower bound on leader mass, and hence on the leader-to-follower mass ratio, is derived, giving a continuum analogue of a minimum actuation budget (Salzano et al., 17 Mar 2026).

At a different level of abstraction, controlled particle systems with leaders and followers admit rigorous micro–macro and macro–macro limits. A short-horizon feedback control

$$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$2

applied to leaders propagates coherently through both limiting procedures, with quantitative convergence estimates in Wasserstein and bounded–Lipschitz distances obtained by modulated energy methods (Albi et al., 6 Aug 2025). Relatedly, mean-field birth–death models allow direct mass transfer between follower and leader populations through nonlinear rates $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$3 and $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$4, yielding a PDE system for $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$5 and, under a structural assumption on the initial datum, an equivalent PDE–ODE system for the total spatial measure $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$6 and the label distribution $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$7 (Albi et al., 2018).

Taken together, these results show that LeFo at continuum scale is not limited to “leaders drag followers.” It includes density shaping, phase-transition-like feasibility thresholds, hierarchical mean-field reduction, and even endogenous formation or extinction of the leader population.

4. Stackelberg, stochastic control, and reinforcement learning

In game-theoretic LeFo, hierarchy is encoded as sequential optimization. A generic Stackelberg form is

$$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$8

and many recent formulations specialize this template to stochastic control or learning (Vahedifar et al., 10 Jul 2025, Chen et al., 30 Jan 2026).

One mathematically rich example is the leader-follower linear-quadratic stochastic graphon game, with one leader and a continuum of followers indexed by $$\dot{x}(t)=-Lx(t)+Mu(t), \qquad y(t)=W^{1/2}R^Tx(t),$$9. Followers’ states obey SDEs with graphon aggregation $M$0, the leader depends on the population average $M$1, and the equilibrium concept is Stackelberg–Nash: for any leader strategy, followers first attain a Nash equilibrium, and the leader then optimizes anticipating that response (Chen et al., 30 Jan 2026). The theory proceeds through graphon-aggregated forward-backward stochastic differential equations, Riccati equations, and a continuation method proving existence, uniqueness, and stability.

A related but cooperative construction appears in linear-quadratic mean-field control with one leader and many weakly coupled followers. There the leader announces a strategy first, followers choose socially optimal responses rather than noncooperative ones, and decentralized social optimality is obtained through forward-backward consistency systems. Under suitable conditions, the resulting strategies form a Stackelberg equilibrium for the social team problem (Huang et al., 2020). In closed-loop stochastic differential games, the follower’s optimal strategy is characterized by a Riccati equation and a backward stochastic differential equation, while the leader solves an LQ problem on a forward-backward system, with necessary conditions for closed-loop solvability also expressed through a Riccati equation (Li et al., 2021).

With multiple followers playing a Nash equilibrium after the leader commits, the computational picture becomes severe. For normal-form and polymatrix games, computing both optimistic and pessimistic leader-follower equilibria with multiple followers is $M$2-hard and not in Poly-$M$3 unless $M$4 (Basilico et al., 2017). That complexity result motivates the paper’s nonconvex mathematical programming formulations, global optimization procedures, and heuristic black-box search over leader commitments.

Model-free RL brings the same hierarchy into episodic Markov games. In the linear-MDP setting, the leader acts first at each step, the follower observes $M$5 and then acts, and both are non-myopic. A model-free adaptation of LSVI-UCB replaces greedy best responses by soft-max policies for both agents, which is the key step enabling uniform concentration of value functions and regret guarantees. The resulting bounds are

$M$6

for both leader and follower under bandit feedback (Ghosh, 2022). In that literature, LeFo is not just a control hierarchy; it is an online-learning problem with provable sample-efficiency guarantees.

5. Robotics, teleoperation, navigation, and embodied interaction

Robotic LeFo systems frequently reinterpret “leader” as privileged information carrier, geometric boundary, or perceptual anchor. In Tactile Internet teleoperation, the human acts as the leader and the robot as the follower in a cooperative Stackelberg game, and the same relation is used in both directions for prediction: the robot predicts the human’s next command and the human predicts the robot’s next feedback (Vahedifar et al., 10 Jul 2025). The predictors use an ARMA-style neural update,

$M$7

and optimize a MiniMax objective combining follower-side Kullback–Leibler divergence and leader-side mutual information. Reported prediction accuracy ranges from $M$8 to $M$9 for robot signals at the human side and from $H^2$0 to $H^2$1 for human operation signals at the robot side, with a Taylor-expansion-based upper bound used to certify one-sample loss mitigation (Vahedifar et al., 10 Jul 2025).

In multi-robot navigation, LeFo can be adversarially concealed rather than emphasized. A MARL-based approach uses a GNN controller and a Scalable-LSTM adversary to hide the leader’s identity by making the leader’s motion similar to followers’ motion (Deka et al., 2021). The adversary attains accuracy $H^2$2 with $H^2$3 parameters on Stage-1 trajectories of $H^2$4-agent teams, and the learned policy achieves 0-shot scaling to larger teams while lowering human accuracy and confidence in identifying the leader (Deka et al., 2021). Here the leader-follower structure is preserved functionally but obscured behaviorally.

Embodied navigation systems use still different mechanisms. In “Auto-Platoon,” the leader carries a visual marker and the follower tracks it with a pipeline built from YOLOv7, MobileNet V2, MiDaS, a Kalman filter, and a “software latching” state machine. The spacing policy uses a fixed target distance $H^2$5m, and obstacle handling is implemented through cooperative STOP signaling and a STOP-FAILSAFE mode (Puthanveettil et al., 2024). In “Swarm Herding,” the leaders are not a single robot but a two-dimensional deformable boundary: a mass–spring–damper mesh whose time-varying domain corrals followers, while the followers perform coverage over the domain through a perspective transformation from each leader-defined quadrilateral to a virtual rectangle (Xu et al., 2021).

Human experiments show that LeFo is also an interaction pattern rather than merely a controller architecture. In collaborative obstacle avoidance with asymmetric information, the overall leader was the higher-performing individual, but roles switched locally: along the obstacle-avoidance axis the obstacle-aware partner temporarily led before and around the obstacle, whereas along the goal-directed axis the higher-performing partner maintained dominance (Leskovar et al., 2022). This empirical result parallels robotic systems in which leadership is decomposed by task dimension or phase.

6. Adaptive roles, reinterpretations, and recurrent limitations

A persistent misconception is that LeFo always means a fixed set of privileged agents. Several lines of work contradict that view directly. In extended opinion dynamics, the agent set is partitioned into a follower group $H^2$6 and leader groups $H^2$7 with targets $H^2$8, but the analysis focuses on when a few leaders can dominate the entire population and drive followers toward weighted combinations of leader-group targets (Li, 2021). In synchronization, leader and follower status are reassigned at every step according to instantaneous angular velocity ranks, with a fixed follower fraction $H^2$9, producing adaptive turn-taking and hybrid synchronization transitions (Park et al., 2020). In mean-field birth–death models, followers can become leaders and leaders can become followers through the nonlinear transfer rates $r$0 and $r$1 (Albi et al., 2018). These formulations make “leader” a dynamic role, not an immutable identity.

Another misconception is that all LeFo phenomena reduce to consensus tracking. The literature includes at least three distinct mechanisms of asymmetry: exogenous actuation asymmetry, as in consensus networks and density control; sequential decision asymmetry, as in Stackelberg games and leader-follower MDPs; and informational asymmetry, as in teleoperation, obstacle-aware human dyads, and leader-identity camouflage (Sato, 2018, Ghosh, 2022, Vahedifar et al., 10 Jul 2025, Deka et al., 2021). A plausible implication is that results seldom transfer verbatim across subfields even when the label “leader–follower” is shared.

Limitations also recur, though in domain-specific form. In graph-based leader selection, directed graphs, time-varying topologies, heterogeneous agent dynamics, higher-order agents, or nonlinear dynamics are not addressed (Sato, 2018). In Tactile Internet prediction, local smoothness, near-stationarity over short horizons, and recovery for up to one-sample outages are assumed, and current inference times often exceed a strict $r$2ms TI target (Vahedifar et al., 10 Jul 2025). In identity-hiding navigation, concealment guarantees are empirical rather than formal, and the adversary model observes positions only (Deka et al., 2021). In continuum density control with interacting followers, convergence is local rather than global, and the basin estimates are conservative (Lorenzo et al., 13 Apr 2026).

Across these variations, LeFo remains a technically productive but highly heterogeneous concept. Its unifying feature is not a single equation or theorem, but the repeated use of structured asymmetry—of actuation, timing, information, or role assignment—to organize collective behavior.