⚡ electrical engineering

Joint Communication-Control Strategy Optimization with Partially Nested Information Structures: The Linear-Quadratic Case

This paper formalizes a joint communication-control optimization problem for multi-agent linear-quadratic systems under partially nested information structures, establishing conditions for preserving nestedness and developing a dynamic-programming approach that yields closed-form Riccati equations for both open-loop and closed-loop communication strategies.

Original authors: Haoyi You, Kaiqing Zhang

Published 2026-08-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Haoyi You, Kaiqing Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a group of friends is trying to solve a puzzle together, but they can't see the whole picture. Each friend only sees a tiny, blurry piece of the puzzle, and they have to decide what move to make next based on just that fragment. This is the heart of decentralized control, a branch of science where many independent agents (like robots, self-driving cars, or even cells in your body) must work together without a single boss telling them what to do. The tricky part is that they don't all know the same things. One friend might know the puzzle piece is blue, while another knows it's round, but neither knows the other's secret. To solve the puzzle efficiently, they need to communicate. But here's the catch: talking takes time and energy. If they shout every little detail to everyone, they might get overwhelmed or run out of battery. If they say too little, they might make a mistake. The big question scientists are asking is: How do these agents decide exactly what to say, and when to say it, to solve the puzzle perfectly while wasting the least amount of energy?

This paper tackles that question for a specific, very common type of puzzle: one where the rules are straight lines and the "cost" of making a mistake grows like a curve (mathematicians call this a Linear-Quadratic problem). The authors, Haoyi You and Kaiqing Zhang, wanted to find the perfect recipe for these agents to jointly optimize their talking and their moving. They asked: "Can we figure out a strategy where the agents share just enough information to stay on track, without getting bogged down in complex, messy math that computers can't handle?"

The Team of Robots and the "Secret Handshake"

Imagine a team of robots trying to herd a flock of sheep. Each robot has a camera (its eyes) and a motor (its legs). They need to move the sheep to a pen, but they can't see the whole flock at once. Robot A sees the sheep on the left; Robot B sees the sheep on the right. If Robot A moves without telling Robot B, Robot B might push the sheep the wrong way, and the whole team fails.

In the past, scientists have tried to solve this by having robots share everything they see. But that's like shouting every single thought you have to your teammates while running a marathon—it's exhausting and slow. Other scientists tried to have robots share nothing, but then they often made silly mistakes because they were guessing.

The authors of this paper realized that for these specific "straight-line" problems, there's a sweet spot. They discovered that if the robots follow a specific set of rules about who knows what and when, they can find the perfect balance. They call this a Partially Nested Information Structure. Think of it like a relay race where the baton (information) is passed in a very specific order. If Robot A's move affects Robot B's view, then Robot B must know what Robot A did. But if Robot A's move doesn't change anything for Robot B, Robot B doesn't need to know. It's a "need-to-know" basis that keeps the team efficient.

The Magic of "Open-Loop" vs. "Closed-Loop"

The paper explores two ways the robots can decide what to say:

Open-Loop (The Pre-Planned Script): Imagine the robots agree on a script before the race starts. "At 1:00, I will shout 'Left!'; at 1:05, you shout 'Right!'" They don't change their minds based on what happens during the race. The authors found that if the robots stick to this pre-planned script, and if the "need-to-know" rules are followed, they can use a very neat mathematical tool called Riccati Equations to calculate the perfect moves. It's like solving a giant, complex puzzle where the pieces fit together perfectly into a smooth, predictable pattern. The computer can solve this quickly and easily.
Closed-Loop (The Live Chat): Now, imagine the robots can change their script while the race is happening. "Oh, the sheep are running left! I'll shout 'Stop!' instead of 'Left!'" This is much harder. The authors show that if the robots try to be too clever and change their minds on the fly, the math can get messy and break the "smooth pattern" they found earlier. However, they didn't give up! They developed a new way to handle this "live chat" scenario. They created a special "expanded" version of the problem where they pretend the robots know a few extra things they don't actually know yet, just to make the math work. Then, they use a step-by-step method (Dynamic Programming) to find the best moves. It's like having a GPS that recalculates the route every second, but the authors figured out how to make that GPS fast enough to be useful.

What They Actually Found

The paper proves that for these specific types of robot teams:

If they follow the "need-to-know" rules (Partially Nested), they can find a perfect, linear strategy. This means their moves are simple, straight-line calculations based on what they see. No crazy, wiggly, unpredictable math needed.
If they break those rules, the perfect strategy might not even exist, or it might be so complicated that no computer could ever solve it. The authors showed examples where breaking the rules leads to a team that just can't find a good solution.
They built a calculator for the "Open-Loop" case. They wrote down a set of equations (Riccati Equations) that anyone can use to find the perfect pre-planned script for the robots.
They extended this to the "Closed-Loop" case. They showed how to handle the "live chat" scenario by expanding the problem, making it solvable with a dynamic program that is much easier to compute than previous methods.

Why This Matters

You might wonder, "Why do I care about robots herding sheep?" Well, this isn't just about sheep. This math applies to self-driving cars coordinating on a highway, drones delivering packages in a city, or even power grids balancing electricity across a country. In all these cases, machines need to talk to each other to avoid crashes and save energy.

The authors showed that there is a "golden rule" for this communication. If the machines follow this rule, we can calculate the perfect way for them to work together. If they don't, the system might break down or become too expensive to run. By providing a clear, step-by-step method to find these perfect strategies, this paper gives engineers a powerful new tool to build smarter, more efficient, and safer autonomous systems. It turns a chaotic, impossible-sounding problem into a solvable puzzle, proving that sometimes, the best way to work together is to know exactly what to say, and exactly when to say it.

Technical Summary: Joint Communication-Control Strategy Optimization with Partially Nested Information Structures

Problem Formulation

This paper addresses the Joint Communication-Control Strategy Optimization (JCCO) problem in multi-agent linear systems with quadratic costs. The setting involves $n$ agents operating over a finite time horizon $H$ under a common-information-based (CIB) framework. The system dynamics are linear, and observations are noisy (Gaussian), forming a decentralized Linear-Quadratic-Gaussian (LQG) control problem where agents must simultaneously optimize:

Control Actions: To minimize state and control costs.
Communication Actions: To determine what private information to share with other agents, incurring a communication cost proportional to the amount of shared data.

The core challenge lies in the coupling between communication and control: the choice of communication strategy alters the Information Structure (IS) available to agents, which in turn dictates the optimal control policy. Unlike standard decentralized control where the IS is fixed, here the IS is a decision variable. The authors focus specifically on scenarios where the baseline information sharing (prior to additional communication) forms a Partially Nested (PN) structure, a condition known to ensure linearity and tractability in fixed-IS decentralized LQG problems.

Methodology

The authors develop a dynamic programming (DP) approach to solve the JCCO problem, distinguishing between open-loop and closed-loop communication strategies.

1. Structural Conditions for Linearity

The paper first establishes that for the optimal control strategy to remain linear (and thus computationally tractable) under additional communication, specific structural conditions must be met. Even if the baseline IS is PN, arbitrary additional sharing can break this property, leading to nonlinear or non-existent optimal strategies. The authors identify three critical assumptions:

Assumption II.1: Standard information evolution rules (baseline sharing, private information updates).
Assumption III.2: Agents do not share or utilize control actions that have no effect on the system state (i.e., if $B_{i,h} = 0$ , the action $U_{i,h}$ is not shared).
Assumption III.4: If an agent's control action affects the state, at least one other agent must be able to observe the effect (i.e., $E_{-i, h+1}B_{i,h} \neq 0$ ).

The paper proves via counterexamples (Lemmas III.1, III.3, III.5) that violating these assumptions can result in nonlinearity or the non-existence of team-optimal strategies.

2. Open-Loop Communication Strategies

For fixed open-loop communication strategies (where communication actions are pre-selected), the problem reduces to a decentralized LQG control problem with a fixed IS.

Strict Expansion: The authors construct a "strictly" expanded problem $\tilde{D}$ by augmenting the common information to include all past control actions that influence other agents.
SI-CIB Condition: They demonstrate that under the identified structural assumptions, this expanded problem satisfies the Strategy-Independent Common-Information-Based (SI-CIB) belief condition. This implies that the conditional belief (posterior distribution of the state and private information given common information) is a Gaussian distribution whose parameters (mean and covariance) evolve independently of the control strategies.
Riccati Equations: Leveraging the SI-CIB condition, the authors derive a finite-dimensional dynamic program. They show that the optimal control strategy is linear in the conditional mean of the state and private information. The optimal gain matrices are computed via a set of closed-form Riccati Equations, avoiding the need for infinite-dimensional belief spaces.

3. Closed-Loop Communication Strategies

The approach is extended to closed-loop strategies, where communication actions depend on current private information and past communication history.

Assumption V.1: Communication strategies must take common information and past communication actions as input.
Expansion and DP: Similar to the open-loop case, the problem is expanded to satisfy the SI-CIB condition. The authors propose a dynamic programming algorithm (Algorithm 1) that operates over the space of communication histories and the finite-dimensional conditional mean/covariance states. This yields a more tractable solution compared to solving directly over infinite-dimensional belief spaces.

Key Contributions

Formalization: The paper formalizes JCCO as an optimal control problem under the CIB framework for multi-agent linear systems, explicitly modeling communication actions as decision variables.
Structural Characterization: It identifies necessary and sufficient structural conditions (Assumptions III.2 and III.4) under which partial nestedness is preserved under additional information sharing. The paper demonstrates that violating these conditions generally leads to nonlinear optimal strategies or non-existence of solutions.
Algorithmic Solution for Open-Loop: It develops a closed-form dynamic programming approach using Riccati equations to solve for optimal control strategies given fixed open-loop communication protocols. This advances existing results on decentralized LQG control with output feedback and PN ISs (extending [17]).
Extension to Closed-Loop: It extends the methodology to closed-loop communication strategies, providing a reduced-dimensional dynamic program that is computationally more efficient than standard CIB-belief-based approaches.
Novel Connection: The work establishes a novel connection between strictly partially nested information structures and the SI-CIB condition, enabling the reduction of the decentralized problem to a centralized Markov Decision Process (MDP) with finite-dimensional sufficient statistics.

Results and Experimental Validation

Numerical Example: A concrete 2-agent, 2-step example illustrates the step-by-step implementation of the proposed method, showing how to compute filter matrices, covariance updates, and the optimal linear control gains.
Experimental Analysis: The authors evaluate the approach on four different baseline information structures (Single-Controller, Turn-based Controllers, One-Step-Delay, One-Direction-One-Step-Delay) with varying communication costs ( $\alpha$ $α$ ).
- Trade-off: Results confirm the expected trade-off: lower communication costs encourage agents to share more information, reducing control costs and total objective values. Conversely, high communication costs penalize sharing, leading to strategies with less information exchange.
- Performance: The proposed method successfully computes optimal strategies across these varying topologies and cost parameters.

Significance and Claims

The paper claims to provide a computationally tractable framework for solving joint communication-control problems in linear-quadratic settings, a domain often plagued by computational intractability due to the "curse of dimensionality" in decentralized control.

Theoretical Advancement: By proving that specific structural conditions preserve partial nestedness and enable the use of Riccati equations, the paper offers a rigorous alternative to heuristic learning-to-communicate algorithms (e.g., [8, 9]) or norm-minimization approaches for co-design.
Independent Interest: The derived dynamic programming approach for decentralized LQG control with PN ISs and output feedback is presented as a contribution of independent interest, advancing the state of the art beyond previous works that were limited to state-feedback or specific graph structures.
Modesty: The authors acknowledge limitations, noting that the results rely on Gaussian noise and linear dynamics. They explicitly state that extensions to non-Gaussian noise, non-cooperative settings, and systems with factorized states are directions for future research, rather than current capabilities. The work focuses on the theoretical derivation of optimal strategies rather than real-time deployment in complex, non-linear environments.