PaperlayerView exact revision on arXiv

Source-linked AI summary

Control Principles of Complex Networks

Yang-Yu Liu, Albert-Laszló Barabási

arXiv:1508.05384v2eess.SYmath.OCphysics.soc-ph

TL;DR

Understanding complex-network behavior requires measuring states, modeling component dynamics, and influencing selected components. This review synthesizes controllability, observability, and control methods while examining how network topology and dynamics interact, and identifies open questions for future research.

  • Problem

    Complex-systems research has focused mainly on measuring states and modeling dynamics, while the control principles of arbitrary self-organized systems remain challenging to understand.

  • Method

    The review synthesizes control-theoretic notions and results concerning controllability, observability, steering networked systems, and collective behavior across network structures and dynamics.

  • Results

    The review organizes current knowledge on controlling complex networks and shows that topology and dynamical laws jointly shape controllability and control.

  • Takeaways & Limitations

    Uncovering control principles can support exploration and understanding of the fundamental laws governing complex-system behavior.

  • Takeaways & Limitations

    Linearization can incorrectly indicate that a nonlinear system is uncontrollable, even when the original system is globally controllable.

Abstract

from arXiv · show

A reflection of our ultimate understanding of a complex system is our ability to control its behavior. Typically, control has multiple prerequisites: It requires an accurate map of the network that governs the interactions between the system's components, a quantitative description of the dynamical laws that govern the temporal behavior of each component, and an ability to influence the state and temporal behavior of a selected subset of the components. With deep roots in nonlinear dynamics and control theory, notions of control and controllability have taken a new life recently in the study of complex networks, inspiring several fundamental questions: What are the control principles of complex systems? How do networks organize themselves to balance control with functionality? To address these here we review recent advances on the controllability and the control of complex networks, exploring the intricate interplay between a system's structure, captured by its network topology, and the dynamical laws that govern the interactions between the components. We match the pertinent mathematical results with empirical findings and applications. We show that uncovering the control principles of complex systems can help us explore and ultimately understand the fundamental laws that govern their behavior.

I. INTRODUCTION

Control theory frames complex-system understanding around measuring states, modeling dynamics, and influencing system behavior. This review connects those foundations to network topology and surveys control principles across biological, social, and technological systems.

  • Complex-system analysis requires measuring state variables and mathematically modeling the dynamics of each component.
  • Control theory defines control as using inputs to make a dynamical system’s output follow a desired trajectory or reach a desired final state.Feedback applies the difference between actual and desired output to the system input.
  • State-space representations connect inputs, outputs, and internal state variables through differential or difference equations.The state mediates between inputs and outputs while emphasizing causality and internal structure.
  • Many nonlinear systems can be linearized around equilibrium points, enabling linear controllability analysis; the stick-balancing system is controllable in this formulation.
  • Complex-network control remains difficult because wiring diagrams, dynamical laws, and parameter values are often incomplete or unavailable.Human-cell protein-interaction maps cover less than 20% of potential interactions, while other systems lack analytical dynamics or precise parameters.
  • The review surveys control advances with emphasis on how network structure affects control across physical, technological, biological, and social systems.

II. CONTROLLABILITY OF LINEAR SYSTEMS

Linear controllability asks whether inputs can steer every network state to an arbitrary target, with topology determining how control signals propagate. Structural controllability extends this analysis when exact edge weights are unknown, requiring only the network’s zero–nonzero wiring pattern.

  • A system is controllable when suitable inputs can drive it from any initial state to any desired final state in finite time.
  • In an LTI network, A encodes weighted interactions among nodes, while B identifies nodes directly controlled by external signals.Actuators are nodes receiving control inputs; drivers are actuators that do not share input signals.
  • Kalman’s rank condition tests controllability through C = [B, AB, A^2B, . . . , A^(N−1)B], which must have full rank N.If the rank is deficient, reachable states remain confined to a subspace; full rank permits steering throughout the state space.
  • C. Structural Controllability: Controlling only a star network’s central node can leave peripheral states correlated and the system uncontrollable, whereas adding an input to one peripheral node restores controllability.
  • C. Structural Controllability: Structural controllability determines controllability for almost all parameter realizations from the zero–nonzero connectivity pattern, excluding pathological weight combinations.
  • C. Structural Controllability: Structural control therefore supports controllability decisions without precise edge weights, provided the network wiring diagram is accurate.

2. Graphical interpretation

Structural controllability can be determined from network topology by checking accessibility and dilation conditions, or equivalently by finding spanning cactus structures. Maximum matching then identifies minimum driver-node sets and reveals how network structure affects controllability.

  • Graphical interpretation: Structural controllability is guaranteed when the controlled digraph has neither inaccessible nodes nor dilations.An inaccessible node cannot be reached from inputs, while a dilation has fewer neighbors than nodes requiring independent control.
  • Graphical interpretation: Lin’s theorem states that an LTI system is structurally controllable if and only if its digraph is spanned by cacti.A cactus is a minimal structure without inaccessible nodes or dilations; removing an edge destroys controllability.
  • Minimum input theorem: The minimum input theorem maps driver-node identification to maximum matching rather than searching all node combinations.The unmatched nodes in a maximum matching are the driver nodes, and maximum matching can be solved in polynomial time through bipartite representation.
  • Network structure and controllability: For directed scale-free networks, γ = 2 is the critical controllability value because nD approaches 1 as γ approaches 2.For γ > 2, full controllability can be obtained by controlling only a subset of nodes; as γ approaches 2, superhubs emerge.
  • Network structure and controllability: Clustering and modularity have no discernible effect on nD, whereas degree correlations can produce linear, quadratic, or no dependence.For uncorrelated directed networks, the density of nodes with kin, kout = 1 or 2 determines maximum-matching size.

2. Solution based on PBH controllability test

The PBH controllability test provides an algebraic route to minimum driver-node identification for LTI systems with known parameters. It links controllability to eigenvalues and eigenvectors, while topology-based algorithms reveal how network structure and link weights affect the required driver set.

  • PBH controllability test: For exactly known LTI parameters, the PBH test determines controllability by checking the system matrix at each eigenvalue of A.The test need only evaluate s ∈ λ(A), because [sI − A, B] already has full rank when s is not an eigenvalue.
  • PBH controllability test: The PBH condition is equivalent to Kalman’s rank condition and requires that no left eigenvector of A be orthogonal to every column of B.This connects the minimum input problem directly to the eigenstructure of the state matrix.
  • Eigenvalue multiplicity: For undirected networks, geometric and algebraic eigenvalue multiplicities are equal because the state matrix is symmetric.The geometric multiplicity is the eigenspace dimension, μ(λi) = N − rank(λiIN − A).
  • Network examples: For undirected unweighted ER networks, nD decreases with connectivity probability p at small p but increases toward (N −1)/N for sufficiently large p.The limiting value is exact at p = 1, corresponding to the complete graph.
  • Network examples: Figure 12 identifies minimum driver nodes by comparing A − λMI, its column canonical form, all eigenvalues, and the eigenvalue with largest geometric multiplicity.Rows linearly dependent on others in the column canonical form correspond to driver nodes.

E. Minimal Controllability Problems

Minimal controllability problems differ according to whether they minimize independent signals, dedicated actuators, or actuators sharing one signal. Their computational difficulty and network-level behavior vary, while driver and link classifications expose alternative control configurations and robustness.

  • Minimal controllability problems: MCP0 minimizes independent control signals or driver nodes, whereas MCP1 and MCP2 minimize directly actuated nodes under different input-sharing constraints.MCP1 uses dedicated inputs, while MCP2 applies one scalar input through a sparse vector b.
  • Minimal controllability problems: MCP0 is easy for a general LTI system, but MCP1 and MCP2 are NP-hard.For structural controllability, MCP1 can nevertheless be solved efficiently using graph-based conditions involving root strongly connected components and assignability.
  • Driver-node configurations: Different maximum matchings can produce multiple minimum driver-node sets with the same size ND.In the example, two maximum matchings yield the driver sets {x1, x2} and {x1, x4}, both with ND = 2.
  • Link classification: Most real networks have few or no critical links, while most links are ordinary and can be removed without eliminating controllability.Critical links occur in every maximum matching; ordinary links occur in some control configurations but are not individually indispensable.
  • Link classification: In ER networks, redundant-link density peaks at ⟨k⟩c = 2e ≈5.436564, coinciding with the core-percolation threshold.The non-monotonic behavior follows the emergence of a compact core as leaf nodes are progressively removed.
  • Node classification: Beyond a critical mean degree, symmetric scale-free networks exhibit bimodal control modes: centralized control has high nr, whereas distributed control can involve more than 90% of nodes.The two modes correspond to the coexistence of high- and low-redundancy solutions after bifurcation.

3. Driver node classification

The paper classifies driver nodes by the structural features that necessitate direct control, then connects controllability to topology, self-dynamics, and control energy. Real networks cluster into three control profiles, while network structure and driver-node allocation shape controllable subspaces and energetic demands.

  • Driver node classification: Driver nodes comprise sources, external dilations, and internal dilations, yielding the control profile (ηs, ηe, ηi).Sources lack incoming links; external dilations arise from surplus sink nodes; internal dilations arise when paths branch to reach all nodes.
  • Driver node classification: Real networks cluster into three classes dominated by external dilations, sources, or internal dilations, unlike random network models.The three clusters correspond to the control-profile components (ηs, ηe, ηi).
  • Driver node classification: Neural and social networks are source dominated, whereas food webs and airport networks are internal-dilation dominated.The supplied passage connects source dominance with relatively uncorrelated agent behavior and internal-dilation dominance with mostly closed systems and conservation laws.
  • Controllable subspace and control centrality: A single node’s control centrality equals the generic dimension of the controllable subspace it can reach.For node i, rank(C(i)) gives the controllable-subspace dimension; structural control uses generic rank when parameter values are unknown.
  • Controllable subspace and control centrality: For networks with high permeability, a reasonably small driver-node set can produce a large controllable subspace.Permeability μ ranges from 0 for disconnected nodes to 1 for networks completely controllable by one driver node.
  • Self-dynamics and control energy: Identical self-dynamics can preserve the minimum driver-node count found without self-dynamics, while heterogeneous self-loop densities can minimize it at a symmetry point.With three self-loop types, equal density 1/3 yields the minimum nD for any network topology and various individual dynamics.
  • Self-dynamics and control energy: For scale-free networks, controlling every node gives Emax ∼N 1/(γ−1), whereas controlling one node gives Emax ∼eN.With a finite driver-node fraction, the maximum energy scales as Emax ∼eN/ND and decays exponentially as driver nodes increase.

K. Control Trajectories

Control trajectories need not remain local even when nearby final states are targeted. In underactuated LTI systems, trajectory geometry and numerical conditioning impose practical limits beyond algebraic controllability.

  • K. Control Trajectories: A state is strictly locally controllable when sufficiently nearby final states can be reached while the entire trajectory remains in a prescribed neighborhood.This definition uses nested balls around the initial state with radii ε and δ.
  • K. Control Trajectories: In a two-dimensional LTI example, minimum-energy trajectories from x1 > 0 to nearby states with smaller x2 necessarily cross into x1 < 0.Thus, nearby endpoints do not guarantee locally confined control paths.
  • K. Control Trajectories: When ND < N, almost all states are not strictly locally controllable, so minimum-energy trajectories are generally nonlocal.Their length generally increases with the controllability Gramian’s condition number.
  • K. Control Trajectories: Ill-conditioned controllability Gramians can make the minimum-energy input fail in practice even when the controllability matrix is well conditioned.A sharp controllability transition occurs as the number of control inputs changes.
  • K. Control Trajectories: Open-loop minimum-energy control is sensitive to noise, motivating linear feedback strategies that approach target states asymptotically while minimizing energy.The paper presents feedback as a more practical and robust alternative.

III. CONTROLLABILITY OF NONLINEAR SYSTEMS

Nonlinear controllability is difficult to characterize directly, so the review emphasizes weaker notions and linearization-based tests. Linearization can certify local controllability, but it may miss global nonlinear controllability, as illustrated by a car model.

  • Weaker notions: Local accessibility and local strong accessibility provide weaker alternatives when controllability of an arbitrary nonlinear system cannot be proved or tested.Strong accessibility requires reaching an open neighborhood at a sufficiently small exact time, whereas accessibility allows times up to a bound.
  • Nonlinear controllability: Nonlinear controllability has been studied extensively, but complete algebraic characterizations of global controllability remain elusive.For complex networked systems, the review expects that mainly weaker notions of controllability can be characterized.
  • Linearization: If a nonlinear system’s linearization at an equilibrium is controllable, the original system is locally controllable near that equilibrium for arbitrarily small input deviations.This connects nonlinear local controllability to linear time-invariant control tests.
  • Linearization: If the linearization along a trajectory is controllable as a linear time-varying system, the original nonlinear system is locally controllable along that trajectory.The trajectory-based result extends the linearization approach beyond equilibria.
  • Limitations of linearization: 4 state variables and 2 control inputs describe a front-wheel-drive car whose linearization at the origin is uncontrollable, although the nonlinear system is globally controllable.The linearized system leaves x2 and φ time-invariant, while the full nonlinear model remains globally controllable.

1. Lie brackets

Lie brackets identify state-space directions that nonlinear systems can generate through sequences of control actions, including directions unavailable from the original control fields. Their generated distributions support algebraic tests for accessibility and clarify why accessibility is weaker than controllability.

  • Lie brackets: Lie brackets generate vector fields describing directions in which a nonlinear system can move from an initial state.The relevant directions belong to the Lie algebra generated by the system’s vector fields as admissible controls vary.
  • Lie brackets: A four-step sequence of piecewise-constant inputs produces, up to order τ^2, a state displacement along [g1, g2](x0).The sequence applies g1, g2, reverse-g1, and reverse-g2 over intervals of length τ.
  • Lie brackets: Higher-order Lie brackets represent additional directions obtained through more elaborate control-input sequences, including steering maneuvers unavailable to direct inputs.For the car, [steer, drive] produces a wriggle and [wriggle, drive] produces a slide.
  • Distributions: A distribution assigns each state the span of the generating vector fields; if they are independent, its dimension is constant and the distribution is nonsingular.In the car and Brockett systems, the relevant vector fields are independent everywhere, analogous to a full-rank matrix.
  • Accessibility: The accessibility rank condition tests local accessibility by checking whether the accessibility distribution has full dimension at the initial state.The distribution is built iteratively from the drift and control fields using Lie brackets.
  • Accessibility: Accessibility does not imply controllability: a system can access all local directions while its drift continually moves it rightward, preventing local controllability.The example has full-dimensional accessibility but x1 increases whenever x2 ≠ 0.

2. Strong accessibility

Strong accessibility focuses on reaching nearby states at exact small times and coincides with controllability for linear systems. Nonlinear controllability tests remain conditional, while network-scale applications face computational and structural limitations.

  • Strong accessibility: For linear systems, strong accessibility is equivalent to Kalman’s controllability rank condition and therefore to controllability.The strong accessibility algebra excludes the drift field but incorporates its brackets with control-generated fields.
  • Limitations: General nonlinear systems still lack conditions that are both sufficient and necessary for controllability.Available sufficient conditions are described as almost necessary rather than complete characterizations.
  • Nonlinear controllability tests: When the drift lies in the span of the control vector fields, full-dimensional accessibility implies local controllability, and globally full-dimensional accessibility implies global controllability.Driftless systems, including the front-wheel-drive car, belong to this class.
  • Nonlinear controllability tests: For other control-affine systems, Sussmann’s sufficient conditions require full-dimensional accessibility and restrictions on every bad Lie bracket.Bad brackets contain an odd number of drift factors and an even number of each control-vector-field factor.
  • Network applications: Continuous-time nonlinear controllability studies have largely been limited to small neuronal network motifs, and extending them to larger symmetric networks remains challenging.Group representation theory may help analyze symmetry effects, while large real networks typically have fewer symmetries.
  • Network applications: Finding a control strategy for a general Boolean network is NP-hard, becoming polynomial-time solvable only for tree structures or networks with at most one directed cycle.Boolean networks provide a discrete-time nonlinear setting where structural restrictions affect computational tractability.

2. Nonlinear systems

For nonlinear networked systems, observability can be analyzed by translating dynamical interdependence into an inference diagram and identifying root strongly connected components. This graphical approach makes sensor selection tractable for large systems, while sufficiency is generally established only conditionally.

  • Nonlinear observability: Algebraic observability holds when state variables can be related to successive derivatives of inputs and outputs through algebraic relations.For rational systems, observability is tested through the rank of a Jacobian formed from gradients of Lie derivatives.
  • Nonlinear observability: A family of state symmetries can make internal variables indistinguishable from the measured input-output behavior, preventing full state reconstruction.In the example, monitoring x1 cannot distinguish states transformed by σλ.
  • Graphical approach: For large systems, the graphical approach reduces nonlinear observability to a property of the static inference diagram.The inference diagram links xi to xj when xj appears in xi’s differential equation, and SCC decomposition identifies structural information flow.
  • Graphical approach: Monitoring every root SCC is necessary, so the number of root SCCs is a strict lower bound on the minimum sensor-set size.A state observer fails if it does not monitor these root SCCs.
  • Sensor sufficiency: For linear systems, graph-based minimum sensors are generally insufficient, whereas for large nonlinear systems they are often sufficient because state symmetries are rare.The sufficiency result has exceptions, and rigorous proof and systematic identification of exceptional cases remain open.
  • Sensor selection: A minimum sensor set includes all pure products and one node from each larger root SCC, with multiple valid choices when SCCs contain several nodes.For root SCC sizes 1, 2, and 3, the example has 1×2×3 = 6 sensor combinations.

2. Power grid

Power-grid observability can be studied through PMU placement and graph structure, from macroscopic observability transitions to minimum dominating sets. Target observability further focuses sensing on variables of interest, while network reconstruction remains constrained by model uncertainty and data informativeness.

  • Power-grid observability: PMUs measure node voltages and adjacent line currents, allowing power-grid observability to be mapped onto a graph problem.Random PMU placement produces an observability transition as macroscopic observable components emerge.
  • Power-grid observability: The observability percolation threshold decreases as average degree or degree heterogeneity increases.The expected largest observable component can be calculated analytically for prescribed degree distributions.
  • Minimum sensor placement: Minimum PMU placement is formulated as a minimum dominating set problem, requiring every unselected node to be adjacent to a selected sensor.The general minimum dominating set problem is NP-hard.
  • Minimum sensor placement: When the underlying network has no core, generalized leaf removal solves the minimum dominating set problem exactly in polynomial time.For general graphs, available polynomial algorithms may return sets up to log N times the minimum size.
  • Target observability: Target observability seeks sensors that infer selected variables rather than the full system, such as disease-altered metabolite concentrations.The graphical approach selects sensors based on directed reachability and the amount of subsystem information that must be reconstructed.
  • State reconstruction: Observability tests and graphical sensor selection do not themselves reconstruct states; an observer must run a system replica and adjust it using measured outputs.For LTI systems, a Luenberger observer can be designed so its state asymptotically tracks the original system.
  • Network reconstruction: Network reconstruction faces a trade-off between the desired interaction-matrix property and certainty about coupling functions, while uninformative temporal data limits every reconstruction target.Prior knowledge such as edge-weight bounds can help when additional experiments are infeasible or expensive.

A. Controlling Chaos

Control methods can steer chaotic systems toward desired trajectories or unstable periodic orbits using parameter perturbations, feedback, or time-delayed signals. The OGY method is especially notable because it uses small, locally applied perturbations that can be learned from observations.

  • Controlling Chaos: Chaotic systems are highly sensitive to initial conditions, so small measurement errors can destroy long-term predictability.This sensitivity is the butterfly effect.
  • Controlling Chaos: Open-loop control designs inputs so the state converges to a desired trajectory, but it requires detailed dynamics and may consume substantial control energy.Convergence also depends on the system’s functional form and initial condition.
  • Controlling Chaos: The OGY method stabilizes a chosen unstable periodic orbit embedded in a chaotic attractor through small, time-dependent parameter perturbations.The procedure identifies low-period unstable periodic orbits, selects one, and designs perturbations near the desired trajectory.
  • Controlling Chaos: OGY has been effective in numerical and experimental systems, including magnetoelastic ribbons, diode circuits, lasers, thermal convection, and chemical reactions.Slow convergence was often reported as a cost of the method.
  • Controlling Chaos: Unlike open-loop control, OGY does not require prior dynamical knowledge because its linearization can be extracted from observations.Chaotic attractors also contain many unstable periodic orbits, providing a diverse set of candidate behaviors.
  • Controlling Chaos: Pyragas control synchronizes a system with a delayed version of itself to stabilize a desired unstable periodic orbit.The supplied passages note that sufficient conditions guaranteeing its applicability remain unknown.

B. Compensatory Perturbations of State Variables

Compensatory perturbations steer high-dimensional networked systems toward target states by modifying accessible state variables while respecting intervention constraints. The approach has practical applications but depends on an accurate detailed dynamical model.

  • B. Compensatory Perturbations of State Variables: High-dimensional networked systems require control tools that can reach target states even when those states are not directly accessible.Compensatory perturbations modify an accessible subset of state variables.
  • B. Compensatory Perturbations of State Variables: The method identifies the closest approach to the target, computes a variational matrix, and optimizes an incremental initial-state perturbation under constraints.Sequential quadratic programming solves the nonlinear optimization, followed by a basin-of-attraction test.
  • B. Compensatory Perturbations of State Variables: Control constraints can limit perturbations to selected nodes and restrict the direction or magnitude of each state-variable change.The example permits perturbing three nodes, with each variable only reducible.
  • B. Compensatory Perturbations of State Variables: A successful perturbation places the updated state inside the target attractor’s basin, after which the uncontrolled dynamics evolve toward the target.The procedure repeats when the new orbit does not reach the target basin.
  • B. Compensatory Perturbations of State Variables: The approach has been applied to mitigating cascading failures in power grids and identifying drug targets in cancer-signaling networks.These applications concern nonlinear networked systems with constrained interventions.
  • B. Compensatory Perturbations of State Variables: A major limitation is that compensatory perturbations require prior knowledge of the detailed system dynamics, and imperfect models may steer the system into a different outcome.Parameter perturbations and attractor networks provide complementary approaches for network control.
  • B. Compensatory Perturbations of State Variables: Feedback vertex set control forces the remaining regulatory network to follow desired trajectories by controlling nodes whose removal eliminates directed cycles.For broad classes of nonlinearities, controlling a set is sufficient and necessary when it is a feedback vertex set.

1. Master stability formalism and beyond

The master stability formalism separates oscillator dynamics from network structure by analyzing perturbations transverse to the synchronization manifold. It yields spectral synchronizability criteria, while pinning control extends synchronization to networks that do not synchronize spontaneously.

  • 1. Master stability formalism and beyond: The master stability formalism projects perturbations onto coupling-matrix eigenmodes, producing decoupled variational blocks for transverse stability analysis.The first eigenmode lies along the synchronization manifold; the remaining modes are transverse.
  • 1. Master stability formalism and beyond: The master stability function determines whether each eigenmode decays by evaluating the sign of its maximum Lyapunov characteristic exponent.The synchronized manifold is stable when all eigenmodes have negative master stability values.
  • 1. Master stability formalism and beyond: For bounded master stability functions, synchronization requires α1 < σλ2 ≤ ··· ≤ σλN < α2.This condition depends on oscillator dynamics through α1, α2 and on network structure through the eigenvalues.
  • 1. Master stability formalism and beyond: If R > α2/α1, synchronization is impossible for any coupling strength; if R < α2/α1, it is stable within a finite coupling interval.The interval is σmin = α1/λ2 < σ < σmax = α2/λN.
  • 1. Master stability formalism and beyond: Larger λ2 lowers the synchronization threshold σmin, making the network more synchronizable.The eigenratio R or λ2 provides a spectral criterion without referring to specific oscillators or output functions.
  • 1. Master stability formalism and beyond: The master stability function assesses only local linear stability, which is necessary but not sufficient for synchronization.Global stability methods such as Lyapunov analysis or contraction theory are needed for sufficient conditions.
  • 2. Pinning synchronizability: Pinning synchronization applies feedback to a subset of nodes so designated leaders drive the remaining followers toward a chosen trajectory.Unlike spontaneous synchronization, the target trajectory is explicitly selected in controller design.
  • 2. Pinning synchronizability: In connected networks, global pinning synchronizability can be achieved with limited pinned nodes by appropriately selecting coupling and feedback gains.For BA scale-free networks, local synchronizability is maximized near a σ-dependent control gain, while node placement also affects outcomes.

3. Adaptive pinning control

Adaptive pinning strategies adjust gains, couplings, node selection, or topology to achieve synchronization when static control is difficult. Related collective-behavior models show how changing interaction networks and noise shape alignment and phase transitions.

  • 3. Adaptive pinning control: Static pinning may require global topology knowledge and prior node selection, motivating adaptive control strategies.Adaptive methods also address cases where static pinning is initially infeasible or network links deteriorate.
  • 3. Adaptive pinning control: Adaptive control gains vary with each node’s synchronization error, and global stability can be assured when individual dynamics satisfy a Lipschitz condition.The gain responds to the deviation between the node state and the reference trajectory.
  • 3. Adaptive pinning control: Adaptive coupling gains modify mutual interaction strengths and are effective for networks of quadratic dynamical systems satisfying a specified incremental stability inequality.The condition uses a diagonal matrix and a positive scalar to bound the dynamics.
  • 3. Adaptive pinning control: Adaptive pinning selects nodes dynamically through an edge-snapping mechanism whose stable states represent unpinned and pinned status.A double-well potential provides stable equilibria at 0 and 1.
  • 3. Adaptive pinning control: Topology adaptation can activate links when local trajectory mismatches act as forcing on edge dynamics.This produces decentralized evolution of the target network topology.
  • 3. Adaptive pinning control: The Vicsek model represents collective alignment as decentralized feedback with a time-varying interaction network.Each agent updates its direction from the average direction of nearby neighbors.
  • 3. Adaptive pinning control: At small speed, decreasing perturbation magnitude produces a continuous transition from disordered motion to an ordered aligned phase.The model is governed by agent density, speed, and perturbation magnitude; noise and connectivity jointly influence alignment.
  • 3. Adaptive pinning control: For a large class of switching signals, connected unions of interaction graphs lead all agent directions to converge asymptotically to a common steady direction.The neighbor set changes over time because agents move.

2. Alignment via pinning

Pinning and navigational feedback extend alignment mechanisms to induce collective flocking and track a virtual leader. The distributed protocol addresses fragmentation and can achieve asymptotic alignment or exponential tracking under stated connectivity and feedback conditions.

  • Alignment via pinning: A single pinned leader can asymptotically align all followers when the union of their interaction graphs remains connected over each time interval.The leader moves at constant speed with fixed direction, and connectivity of the time-varying interaction graph is sufficient for eventual alignment.
  • Alignment via pinning: The flocking protocol combines cohesion, separation, and velocity alignment, but generic large systems can fragment into groups moving in different directions.For large agent populations such as N > 100, the protocol may produce fragmentation rather than flocking from generic initial states.
  • Alignment via pinning: Navigational feedback adds a group objective represented by a virtual leader, enabling agents to track a leader moving at constant velocity and thereby achieve flocking.The feedback term complements the proximity-based interaction rules that regulate inter-agent distances and velocities.
  • Alignment via pinning: Only a fraction of agents need to be informed or pinned; numerical simulations suggest that larger informed groups produce larger fractions moving at the desired velocity.The protocol therefore does not require every agent to know the virtual leader’s current state.
  • Alignment via pinning: With acceleration feedback, the protocol asymptotically tracks a virtual leader with varying velocity, and the flock’s center-of-mass position and velocity converge exponentially to the leader’s.Without this additional feedback, agents can reach a common velocity that is not guaranteed to match the leader’s varying velocity.
  • Alignment via pinning: Control theory combined with network science provides tools for understanding and externally inducing order in multi-agent systems.The review presents flocking as an example of controlling collective behavior through network interactions and control inputs.

VII. OUTLOOK

The outlook identifies unresolved challenges in stability, adaptive and multilayer network control, and quantum networks. It concludes that revealing complex-network control principles remains difficult and will require sustained interdisciplinary work.

  • A. Stability of Complex Systems: Stability remains a prerequisite for control, but general methods for constructing suitable Lyapunov functions and characterizing large diagonally stable matrices are lacking.The review notes both the practical importance of stability and the limited theoretical tools for large network-structured systems.
  • A. Stability of Complex Systems: Structural stability concerns whether qualitative trajectory behavior survives perturbations to the system model, and its theory remains underexplored in complex networked systems.For two-dimensional systems, the Andronov-Pontryagin criterion gives necessary and sufficient conditions involving hyperbolic equilibria and limit cycles and forbidden saddle connections.
  • B. Adaptive, Temporal, and Coevolutionary Networks: Adaptive, temporal, and coevolutionary networks require a framework that models network structure itself as dynamical and captures feedback between structure and nodal or edge dynamics.The review identifies controllability of such systems as a natural starting point because limitations in network structure or dynamical rules can constrain control.
  • C. Networks of Networks: Controlling networks of networks is presented as necessary for understanding complex-system control principles, with early work addressing linear controllability and observability.Relevant conditions involve overall topology, node dynamics, external control inputs, and internal interconnections.
  • D. Quantum Networks: Quantum networks can exhibit properties unavailable to classical counterparts, implying that their control will require new methodologies.The review cites quantum subgraphs of arbitrary complexity arising even at connection probabilities sufficient only for simple classical connections.
  • F. Conclusion: Revealing control principles remains a challenging problem likely to engage multiple research communities for the next decade.The review frames interdisciplinary collaboration as necessary for addressing its many outstanding questions.
Loading 1508.05384v2…