Distributed Internal Model Principle

• Distributed Internal Model Principle is a framework requiring each agent to embed a dynamic internal model of the exosystem to achieve robust output regulation.
• It employs distributed consensus protocols that synchronize local estimates of the exosystem’s state and spectral parameters across time-varying, directed networks.
• The approach guarantees uniform exponential convergence and resilience to uncertainties, with applications spanning power systems, robotic formations, and more.

The Distributed Internal Model Principle (DIMP) is a foundational concept in networked and multi-agent control, extending the classical internal model principle of robust output regulation to systems composed of multiple, dynamically coupled, and often heterogeneous agents. DIMP states that, to achieve robust perfect output regulation (tracking and disturbance rejection) of reference signals or disturbances generated by some exogenous system (the exosystem), it is necessary and sufficient that each agent—operating with only partial and local knowledge—embeds a dynamic compensator whose structure incorporates a local internal representation (“internal model”) of the exosystem’s minimal polynomial. Distributed protocols enable agents to asymptotically synchronize their internal models and estimates of the exosystem’s state and dynamics across time-varying, directed, and possibly uncertain communication graphs. This framework unifies the synthesis and analysis of distributed controllers for cooperative output regulation, disturbance rejection, synchronization, and resilient operation in the presence of both plant/model uncertainty and communication constraints (Cai, 2016, Kawamura et al., 2018).

1. Core Formulation and Principle

The DIMP generalizes the centralized internal model principle by requiring that each agent in a networked multi-agent system embeds (locally) an internal model that is asymptotically congruent to the (potentially unknown) exosystem driving the outputs to be regulated. Consider a network $\mathcal{G}(t) = (V, E(t))$ of $N$ agents, each with possibly uncertain, heterogeneous LTI dynamics: $$\dot{x}_i = A_i x_i + B_i u_i + P_i w_0, \quad z_i = C_i x_i + D_i u_i + Q_i w_0$$ with the exosystem (“leader”)

$\dot{w}_0 = S_0 w_0, \quad y_0 = Q_0 w_0$

and only a subset of agents observing $w_0$ directly. The objective is to ensure $z_i(t) \to 0$ for all $i$, all admissible uncertainties, and arbitrary initial states (Cai, 2016, Kawamura et al., 2018).

DIMP mandates that each agent dynamically reconstructs both the exosystem dynamics $(S_0, w_0)$ and its minimal polynomial roots, and embeds a $q_i$-copy (for $q_i$-output agents) of the exosystem’s internal model in its local compensator, driving the local regulation error to zero. Critical requirements include stabilizability, detectability, and graph-theoretic conditions (e.g., existence of a globally reachable exosystem node over time in the communication graph, or a spanning tree in the static case).

2. Distributed Controller Architecture and Mechanisms

A canonical DIMP controller is constructed from two interacting subsystems:

1. Distributed Exosystem Generator: Each agent tracks local estimates $S_i(t)$, $w_i(t)$, updated via dynamic consensus algorithms over the communication topology:

$$\dot{S}_i = \sum_{j \in \mathcal{N}_i(t)} a_{ij}(t)(S_j - S_i), \quad \dot{w}_i = S_i w_i + \sum_{j \in \mathcal{N}_i(t)} a_{ij}(t)(w_j - w_i)$$

As long as the exosystem node is globally reachable, $S_i(t) \to S_0$ and $w_i(t) \to w_0(t)$ exponentially for all $i$ (Cai, 2016, Kawamura et al., 2018).

2. Dynamic Compensator with Internal Model Consensus: Each agent embeds a (potentially time-varying) internal model whose parameters $\lambda_i(t)$ (roots of the minimal polynomial of $S_0$) are also synchronized by consensus:

$$\dot{\lambda}_i = \sum_{j = 0}^N a_{ij}(t)(\lambda_j - \lambda_i)$$

The compensator is designed as:

$$\begin{aligned} \dot{\xi}_i &= E_i(\lambda_i) \xi_i + F_i e_i, \ u_i &= K_i(\lambda_i) \xi_i, \ e_i &= C_i x_i + D_i u_i + Q_i w_i \end{aligned}$$

where $E_i, F_i, K_i$ explicitly incorporate the $q_i$-copy internal model structure parameterized by $\lambda_i$ (Cai, 2016).

These mechanisms ensure that all agents asymptotically track both the state and the spectral parameters of the exosystem, thereby embedding a congruent internal model locally.

3. Theoretical Guarantees and Stability Analysis

The main theorems underlying DIMP establish that, under network connectivity and agent stabilizability/detectability assumptions, the distributed controller achieves uniform exponential convergence of $z_i(t) \to 0$ for all $i$:

• Consensus on Exosystem and Internal Model: Lemmas demonstrate convergence of all local copies $S_i, w_i, \lambda_i$ to their true exosystem counterparts, provided the time-varying graph uniformly contains a spanning tree rooted at the exosystem node (Cai, 2016, Kawamura et al., 2018).
• Embedding and Regulation Equations: Sylvester-type regulator equations linking the agent’s closed-loop dynamics and the exosystem must be solvable for the internal model–embedding condition.
• Robustness to Uncertainty: The controller guarantees output regulation for any admissible plant parameter uncertainties within specified (open) neighborhoods, via input-to-state stability arguments and Lyapunov techniques.
• Proof Techniques: Composite Lyapunov functions, separation principles, and input-output stability concepts are used to demonstrate uniform exponential stability and robustness (Cai, 2016, Kawamura et al., 2018).

DIMP thus yields not only robust output regulation but also resilience to plant/model uncertainty and dynamical heterogeneity across agents.

4. Methodological Variants and Extensions

DIMP has been adapted and extended to a wide range of system classes and methodologies:

• Time-varying, Directed, and Dynamic Networks: Both static and switching topologies with time-varying digraphs are accommodated. Only a subset of agents need initial access to the exosystem, and agents achieve synchronization through consensus even when the communication topology evolves (Cai, 2016, Kawamura et al., 2018).
• Infinite-dimensional Systems: The DIMP extends to distributed parameter systems with unbounded control and observation using $p$-copy internal models or the more general $\mathcal{G}$-conditions, yielding necessary and sufficient conditions for robust output regulation in Banach and Hilbert space settings (Paunonen et al., 2014).
• Data-Driven Realizations: Integral reinforcement learning and adaptive dynamic programming yield model-free, distributed control protocols. These use trajectory data to solve regularized Riccati equations, decoupling the identification into lower-dimensional algebraic steps, and drastically relaxing rank conditions. This enables DIMP for unknown, high-dimensional, or MIMO plants (Lin et al., 20 Feb 2025).
• Adversarial Settings: The same structural principle that ensures robust regulation can be exploited by attackers. An “internal model principle for the attacker” states that by embedding exosystem modes into attack signals at root nodes, an adversary can drive the consensus network to instability while remaining undetectable by local error residuals (Moghadam et al., 2017).

5. Applications and Case Studies

DIMP underlies robust distributed control and output regulation in a variety of applications:

• Cooperative Output Regulation and Synchronization: DIMP is applied in leader–follower and output synchronization problems for general LTI and uncertain agent arrays, including agents with different models, unknown parameters, and time-varying interconnections (Cai, 2016, Kawamura et al., 2018).
• Power Systems: DIMP enables distributed frequency regulation and cost-optimal generation sharing under unknown and time-varying loads in power networks. Incremental passivity is leveraged, with distributed internal models uniquely embedded at each node for disturbance rejection (Trip et al., 2014).
• Robotic Formations: In manipulator formation tasks, DIMP-based controllers combine internal model compensators for non-vanishing disturbances (e.g., step/sinusoidal torques/forces) with graph-based potential shaping and local damping, yielding distributed, robust formation-keeping even under significant local perturbations (Wu et al., 2021).
• Infinite-dimensional and Boundary-Control Systems: Robust output tracking for PDEs (such as the heat equation with boundary control) exploits DIMP via infinite-dimensional internal models or operator-theoretic embedding principles (Paunonen et al., 2014).
• Data-Driven Output Regulation: Distributed controller synthesis in high-dimensional, MIMO, and even partially unknown agent nets is enabled through PI/VI-based reinforcement learning, which identifies minimal internal model realizations directly from data in a distributed manner (Lin et al., 20 Feb 2025).

6. Adversarial and Resilient Networked Control

The distributed internal model principle not only enables robust cooperative regulation, but also determines structural vulnerabilities:

• IMP for the Attacker: If an adversary controls root nodes in a consensus network, attack signals generated by an exosystem sharing modes with the plant's internal model can destabilize or mislead the entire network, exploiting the same zero-mode dynamics DIMP leverages for regulation (Moghadam et al., 2017).
• Stealth and Resilience: Detection schemes based only on neighborhood errors are ineffective against these “zero-mode” attacks. Necessary and sufficient conditions for attack stealthiness and destabilizability have been derived, highlighting the need for protocol designs that disrupt the attacker’s access to the internal model channel, for example, through network topology reconfiguration or dynamic authentication of critical nodes.
• Future Directions: A plausible implication is that DIMP-inspired defense methods should monitor for mode embeddings, randomize communication structures, or implement cross-layer detection schemes targeting the internal model's spectral footprint (Moghadam et al., 2017).

7. Relationship to Centralized IMP and Implications

The DIMP fundamentally generalizes the classical (centralized) internal model principle:

• In centralized control, the exosystem’s model is embedded in a single controller; in DIMP, distributed consensus protocols are required to collectively reconstruct and track the exosystem’s dynamic model across all agents (Cai, 2016, Kawamura et al., 2018).
• DIMP reconciles heterogeneity, dynamic network topology, and partial information, achieving tight performance guarantees under broad connectivity, stabilizability, and regulator equation solvability.
• The theoretical structure is retained in infinite-dimensional settings, under unbounded operators and general exosystems with Jordan structures (Paunonen et al., 2014).

DIMP thus provides a unifying theoretical and algorithmic framework for robust, resilient, and scalable output regulation in complex networks, with far-reaching implications for both benign and adversarial multi-agent systems.