Classical control theory is built on the premise of exact state targeting. Under traditional frameworks like Kalman controllability, a system is considered controllable only if an external input can guide its state vector to any arbitrary point in the state space. However, when dealing with complex networks—such as gene regulatory networks, multi-agent systems, or stochastic environments—this strict requirement is both computationally intractable and practically unnecessary.
In a biological network, for example, a clinician does not need to force every single gene expression level to an exact decimal value; they simply want to steer the cell away from a cancerous state and into a healthy one. As noted by Gao et al. (2017) and Chen et al. (2020), determining how external inputs can override internal network logic remains a fundamental challenge. To bridge this gap, a new paradigm has emerged that shifts the focus from exact point controllability to relaxed, set-based, and distributional control.
Relaxing the Target: Herdability and Existential Control
Instead of demanding that a system reach an exact point, what if we only care that it enters a safe region or exceeds a viability threshold? This relaxation is known as herdability.
Shamma (2018) formalized herdability as a set-based reachability condition where the goal is to drive all system states above a specific non-negative threshold. This framing is highly valuable for ecological or economic systems where maintaining states above a critical level is the primary objective to avoid catastrophic collapse or extinction.
Traditional Controllability: State ───> [ Exact Point X ]
Herdability: State ───> [ Any Point > Threshold ]
Building on this, De Pasquale et al. (2021) and De Pasquale et al. (2023) explored how network topology dictates this property. They demonstrated that by analyzing the algebraic and graph-theoretic properties of signed, directed graphs, one can determine whether a system can be "herded" into a positive region without needing full controllability. For instance, in leader-follower networks or tree-like structures, the system's internal propagation pathways allow a minimal set of inputs to push the entire network past a threshold, bypassing the need to micro-manage individual node states.
A parallel concept in discrete systems is existential control. In asynchronous Boolean networks, guaranteeing that an intervention always forces a transition to a target state (absolute control) is often impossible due to non-deterministic timing and complex feedback loops (Su et al., 2023). To resolve this, Paul et al. (2018) proposed existential control: finding a minimal set of intervention points such that at least one valid sequence of transitions can lead the network to a desired stable attractor. By relaxing the requirement from a guarantee to a possibility, the authors showed that large networks can be decomposed into manageable blocks, drastically lowering the computational barrier to finding viable interventions.
Distributional Control: Steering the Long Run
When networks are inherently stochastic—such as Probabilistic Boolean Networks (PBNs) or Markov Decision Processes (MDPs)—individual state trajectories become unpredictable due to environmental noise. Here, the paradigm shifts entirely: instead of controlling the state, we control the probability distribution over states.
Trairatphisan et al. (2013) highlight how biological phenotypes can be treated as attractors in a Markov chain. In this view, a disease state is simply an undesirable attractor that traps the system's probability mass. The goal of an intervention (such as a drug combination) is to alter the network's transition probabilities to shift the long-term behavior toward a healthy attractor.
Why does this work? Instead of optimizing for the immediate path or transition speed, Qian et al. (2009) showed that it is far more effective to directly manipulate the network's stationary probability distribution. By leveraging rank-one perturbation theory, the authors designed greedy control policies that maximize the long-run probability mass residing in desirable states (e.g., forcing a cancer cell into programmed cell death) while minimizing time spent in dysfunctional states.
Crucially, this approach eliminates the need for subjective cost functions that penalize every step of a trajectory. Furthermore, as Qian et al. (2013) demonstrated, even if our structural model of the network's wiring diagram is imperfect, these distribution-shifting interventions remain remarkably robust and effective in practice.
Spectral Properties and Mixing Times
To make distributional control practical, a policy must not only identify the correct target distribution but also ensure the system converges to it rapidly. This introduces the concept of distribution regulation via spectral optimization.
Corso et al. (2026) framed steady-state control in MDPs as a problem of minimizing the time required to reach a target equilibrium. By treating the probability distribution itself as the state to be regulated, the authors developed a tractable convex optimization approach. This method synthesizes control policies by directly optimizing the spectral properties (the eigenvalue gap) of the system's transition matrix, ensuring the fastest possible mixing rate to the target distribution.
[ Current Distribution ] ───(Optimize Spectral Gap)───> [ Target Equilibrium ]
│ Fast Convergence
▼
Stable System Behavior
When transitioning between different control strategies in online settings, managing this distribution shift is critical to preventing performance loss. Chandrasekaran et al. (2023) proved that a dedicated synchronization procedure can be used to smoothly shift an agent's current state distribution toward a target policy's stationary distribution. This mechanism ensures that the agent's behavior aligns with the intended long-term characteristics within a timeframe strictly bounded by the environment's connectivity, achieving optimal regret bounds even in complex, communicating environments.
The Duality of Conditioning and Control
Perhaps the deepest mathematical intuition behind this research direction lies in the duality between probabilistic conditioning and optimal control.
Consider a stochastic process that risks exiting a safe domain (such as a biological cell facing apoptosis or an Ising network losing its activation structure, as explored by Lynn et al. (2018)). The long-term behavior of this process, conditioned on its survival, is described by a quasistationary distribution.
Budhiraja et al. (2021) established that the dynamics of a stochastic process conditioned on never reaching an exit boundary are mathematically identical to the dynamics of a process under optimal ergodic control.
Specifically, the density of a quasistationary distribution acts as the cost potential of an adjoint ergodic control problem. In this framework, an agent applies a control to minimize the average cost per unit time required to prevent the process from exiting the defined domain. This connection provides a powerful computational bridge: instead of performing difficult eigenvalue decompositions to find long-term survival paths, researchers can solve the problem through deterministic optimization methods, treating survival as the ultimate control objective.
Summary
By moving away from the rigid constraints of classical point-controllability, this unified paradigm provides a robust toolkit for managing complex, noisy networks. Whether by driving deterministic states above a vital threshold (herdability), finding existential paths to attractors, or reshaping the stationary landscape of a Markov process, these methods recognize that controlling the global, long-term statistical behavior of a system is both more achievable and more resilient than trying to micro-manage its individual components.
Key Foundational Work
Intervention in gene regulatory networks via greedy control policies based on long-run behavior (2009) — contributed 2 concept elements to this direction.
Validation of gene regulatory network inference based on controllability (2013) — contributed 1 concept elements to this direction.
Recent development and biomedical applications of probabilistic Boolean networks (2013) — contributed 1 concept elements to this direction.