PaperlayerView exact revision on arXiv

Source-linked AI summary

Duality and Network Theory in Passivity-based Cooperative Control

Mathias Bürger, Daniel Zelazo, Frank Allgöwer

arXiv:1301.3676v2math.OCcs.SYeess.SY

TL;DR

The paper asks whether passivity-based cooperative control inherits the duality structure of network optimization. It introduces maximal equilibrium independent passivity and derives inverse-optimality and duality results for networks; the analysis also covers asymptotic clustering in a nonlinear traffic model.

  • Problem

    An explicit connection between passivity-based cooperative control and network optimization duality had not been established, despite modeling similarities through incidence matrices.

    Source: §1 · p. 2

  • Method

    The paper introduces maximal equilibrium independent passivity and analyzes the resulting networks through inverse-optimality and duality relations with network optimization.

    Source: §1 · p. 2

  • Results

    Networks of maximal equilibrium independent passive systems asymptotically approach network optimization solutions, and the traffic model exhibits asymptotic clustering.

    Source: §1 · p. 2, §7 · p. 18

  • Takeaways & Limitations

    Steady-state inputs and outputs can be interpreted through dual optimal flow and potential problems, connecting cooperative control variables to network optimization.

    Source: §1 · p. 2

    The output-agreement convergence result requires strongly monotone coupling nonlinearities when convergence must be ensured without knowing the required steady-state input in advance.

    Source: §4.2 · p. 12

Abstract

from arXiv · show

This paper presents a class of passivity-based cooperative control problems that have an explicit connection to convex network optimization problems. The new notion of maximal equilibrium independent passivity is introduced and it is shown that networks of systems possessing this property asymptotically approach the solutions of a dual pair of network optimization problems, namely an optimal potential and an optimal flow problem. This connection leads to an interpretation of the dynamic variables, such as system inputs and outputs, to variables in a network optimization framework, such as divergences and potentials, and reveals that several duality relations known in convex network optimization theory translate directly to passivity-based cooperative control problems. The presented results establish a strong and explicit connection between passivity-based cooperative control theory on the one side and network optimization theory on the other, and they provide a unifying framework for network analysis and optimal design. The results are illustrated on a nonlinear traffic dynamics model that is shown to be asymptotically clustering.

1 Introduction

The introduction motivates an explicit connection between passivity-based cooperative control and network optimization, then presents maximal equilibrium independent passivity as the basis for inverse-optimality and duality results. It also previews a broader network analysis and a nonlinear traffic example with asymptotic clustering.

1 Introduction

  • Passivity-based cooperative control separates agent input-output behavior from network topology, enabling analysis of interconnected systems through those properties.Equilibrium independent passivity supports convergence analysis without computing the convergence point in advance.Source: §1 · pp. 1–2
  • The paper addresses the open question of whether cooperative control inherits duality results from network optimization, despite modeling similarities through incidence matrices.The authors state that an explicit connection had not yet been established.Source: §1 · p. 2
  • The paper introduces maximal equilibrium independent passivity and uses it to derive inverse-optimality, existence and uniqueness, and coupling conditions for output agreement.The steady-state input and output correspond to an optimal flow and its dual optimal potential; the framework also treats dynamic state and coupling output as inverse optimal.Source: §1 · p. 2
  • The results extend to a broader class of networks and show asymptotic clustering in a nonlinear traffic dynamics model.The traffic example illustrates the generalized inverse-optimality and duality results.Source: §1 · p. 2

Preliminaries

  • The preliminaries introduce convex-analysis tools used in the paper, including convexity, convex conjugates, subgradients, and indicator functions.The text also defines strong monotonicity and co-coercivity.Source: Preliminaries · pp. 2–3

2 Network Optimization Theory

The section describes a graph-based optimization framework in which incidence matrices relate network flows and divergences, and potentials and tensions. Its dual formulation yields an optimal potential problem from an optimal flow problem.

2 Network Optimization Theory

  • A graph's incidence matrix encodes node-edge orientation and defines the circulation, differential, and agreement spaces used to describe network structure.The edge orientation is arbitrary, and the incidence matrix satisfies 1^⊤E = 0.Source: §2 · p. 3
  • A flow assigns an edge flux, while its divergence records the associated net flux at each node under a conservation relationship.The optimal flow problem optimizes flow and divergence subject to the conservation constraint u + Eµ = 0.Source: §2 · pp. 3–4
  • A potential assigns values to nodes, and each edge's tension is the potential difference across that edge.Flows and tensions satisfy the conversion formula µ^⊤ζ = −y^⊤u.Source: §2 · p. 3
  • The dual of the flow problem replaces divergence and flow variables with artificial variables, then dualizes their equality constraints using Lagrange multipliers.The dual objective uses convex conjugates, and its constraint ζ = E^⊤y gives the optimal potential problem.Source: §2 · p. 4
  • The section sets up a later interpretation of cooperative control for a class of passive systems using this network optimization framework.Source: §2 · p. 4

3 Passivity-based Cooperative Control

The section models cooperative control as passive dynamical systems connected through an undirected graph, then establishes convergence conditions and refines equilibrium-independent passivity. The refinement admits maximal-monotone equilibrium input-output relations, including the integrator, and supports the later connection to network optimization.

3.1 A Canonical Dynamic Network Model

  • Networks place SISO systems at graph nodes and edge controllers driven by relative outputs; controller outputs influence the two incident systems.The model includes diffusively coupled networks when the coupling functions are odd and symmetric.Source: §3.1 · pp. 4–5

3.2 Passivity as a Sufficient Condition for Convergence

  • Under the stated passive-network assumptions, node outputs converge to constant steady-state values.The convergence argument separates the system and network layers: passivity of their input-output behavior suffices, without requiring a local model of each layer's internal dynamics.Source: §3.2 · pp. 5–6

3.3 Equilibrium Independent Passivity

  • Equilibrium-independent passivity avoids computing the convergence point beforehand by requiring passivity across equilibrium points generated by constant inputs.For convergence analysis, it suffices to verify that an equilibrium trajectory exists in the relevant sets.Source: §3.3 · p. 6

3.4 Maximal Equilibrium Independent Passivity

  • Maximal equilibrium-independent passivity replaces a single-valued equilibrium map with a maximal-monotone relation, admitting integrators excluded by the original definition.The paper limits this definition to SISO systems; networks of these systems are shown to converge to solutions of dual network-optimization problems.Source: §3.4 · pp. 6–7

4 Output Agreement Analysis

The section connects output-agreement steady states to dual network optimization problems at both the plant and controller levels. Under stated passivity and monotonicity conditions, these optimization relations establish steady-state properties and support convergence to output agreement.

4.1 The Plant Level

  • At plant steady state, outputs agree, and the input and output vectors solve dual optimal flow and potential problems with opposite optimal values.The result is an inverse optimality characterization of output-agreement steady states.Source: §4.1 · pp. 8–9
  • An output-agreement steady state exists when each node is maximal equilibrium independent passive and its input and output sets are both R.Under the additional assumptions of a nonempty input set and a strongly monotone equilibrium input-output function with the stated boundary behavior, at most one steady-state pair exists.Source: §4.1 · p. 10
  • The framework applies to homogeneous and heterogeneous systems, with its new insights especially relevant when heterogeneous equilibrium input-output maps do not share an intersection point.The paper associates each node's equilibrium relation with a convex cost function and its convex conjugate potential.Source: §4.1 · pp. 8, 10

4.2 The Control Level

  • For a known feasible plant input, strongly monotone controller nonlinearities ensure controller equilibrium feasibility and make controller state and output optimal tension and flow variables in a dual problem pair.At steady state, the controller state is a tension and its output is the corresponding dual flow.Source: §4.2 · pp. 11–12
  • Strong monotonicity is a stronger assumption than sector nonlinearity; without it, the controller is not guaranteed passive at every required steady-state input.The paper motivates strong monotonicity because the required steady-state input is generally unknown at controller-design time.Source: §4.2 · pp. 11–12

4.3 The Closed-Loop Perspective

  • With output-strictly maximal passive nodes having input and output sets R and strongly monotone couplings, the closed-loop trajectories converge to optimal solutions of all four network problems and to output agreement.The plant outputs and inputs converge to their optimal potential and divergence counterparts, while controller state and output converge to tension and flow counterparts.Source: §4.3 · p. 13
  • Dynamic outputs and inputs correspond to optimal potentials and divergences, while controller state and output correspond to tensions and flows.The dynamic signals interact in a closed loop, unlike their static network counterparts; the static tensions ζ and η are not connected, whereas their dynamic counterparts are linked by an integrator.Source: Table 1 · p. 12, §4.3 · p. 13

5 A General Dynamic Network Analysis Framework

The paper generalizes its dual network optimization framework to networks whose controllers are arbitrary maximal equilibrium independent passive systems. When the generalized flow and potential problems are feasible and the nodes are output-strictly maximal passive, the network trajectories converge to their corresponding optimal variables.

5 A General Dynamic Network Analysis Framework

  • The generalized flow problem assigns costs to node divergences and controller flows, while the dual potential problem penalizes tensions with general controller cost functions.Unlike the original potential problem, the generalized problem does not necessarily force potential differences to zero.Source: §5 · p. 14
  • For output-strictly maximal passive nodes and maximal equilibrium independent passive controllers, feasibility of the generalized dual problems guarantees convergence of the network trajectories to their corresponding optimal variables.The limits solve the generalized optimal flow problem for u and µ and the generalized optimal potential problem for y and ζ.Source: §5 · pp. 14–15
  • The generalized optimization problems characterize network equilibrium configurations in both directions: feasible optimal solutions correspond to equilibria, and possible equilibria solve the problems.Source: §5 · pp. 14–15
  • The earlier output-agreement formulation is recovered when each controller's equilibrium input-output relation is the vertical line through the origin.In that case, the controller cost is the indicator function for the origin, and the generalized problems reduce to the original flow and potential problems.Source: §5 · p. 15

6 Application: Analysis of a Traffic Dynamics Model

The traffic example applies the network-optimization interpretation to heterogeneous optimal-velocity dynamics with saturating tanh couplings. Its simulations show increasingly refined clustering as the spread of preferred velocities increases.

6 Application: Analysis of a Traffic Dynamics Model

  • The model assumes heterogeneous preferred velocities, bidirectional vehicle influence, and overtaking; the neighbor set is fixed over time in the example.Source: §6 · p. 15
  • The traffic dynamics interconnect output-strictly maximal equilibrium independent passive vehicle systems with maximal equilibrium independent passive controllers.The coupling dynamics use tanh, whose integral is ln cosh and is not strongly convex; Proposition 6.1 establishes maximal equilibrium independent passivity for each coupling.Source: §6 · p. 16
  • In the traffic interpretation, velocities are potentials, vehicle influences are divergences, relative positions are tensions, and coupling-mapped positions are flows.Source: §6 · p. 16
  • For σ0=1 traffic shares a common velocity; at σ0=2.5 it clusters, with finer clustering at σ0=4 (σ1=1).The asymptotic velocities are predicted by the network optimization problems.Source: Figure 4 · p. 17, §6 · p. 18
  • GOFP constrains flows to −1 ≤ µk ≤ 1, while GOPP combines a quadratic potential cost with an ℓ1 penalty on potential differences.Source: §6 · p. 17
  • A 100-vehicle line-graph simulation uses κ=0.6 and varies the standard deviation σ0 of random preferred-velocity components.The displayed velocity trajectories use σ0=2.5 and σ1=1.Source: Figure 4 · p. 17, §6 · p. 18

7 Conclusions

The paper connects passivity-based cooperative control to network optimization through maximal equilibrium independent passivity. It shows that such networks approach optimization solutions and interprets their inputs and outputs as network variables.

7 Conclusions

  • Networks of maximal equilibrium independent passive systems asymptotically approach solutions of several network optimization problems.Source: §7 · p. 18
  • For output agreement, the steady state must be optimal for both an optimal flow problem and an optimal potential problem.Source: §7 · p. 18
  • The framework also establishes inverse optimality and strong duality for general networks that need not converge to output agreement.Source: §7 · p. 18
  • The traffic model illustrates the theory with asymptotic clustering behavior.Source: §7 · p. 18
Loading 1301.3676v2…