In traditional optimization, the goal is to reach a solution that is both optimal and feasible. Most algorithms—especially those used in distributed systems—treat feasibility as an asymptotic property: the solution becomes "legal" only after the algorithm has fully converged.
However, in safety-critical infrastructure like power grids, robotics, and water networks, an "illegal" intermediate state is more than a mathematical nuisance; it is a potential catastrophe. If a power grid controller issues an intermediate command that violates voltage limits or supply-demand balance, the result can be immediate hardware failure or a blackout.
The research direction of Anytime Feasibility (or "per-iteration feasibility") addresses this by redesigning optimization dynamics so that every single intermediate step satisfies the system's constraints. This allows for safe early termination and robust real-time control.
The Problem with Asymptotic Feasibility
Most standard optimization frameworks, particularly those based on dual decomposition or penalty functions, approach the feasible region from the "outside." As Alireza Aghasi and Mohammadreza Doostmohammadian note in their survey of distributed algorithms, dual-based methods only satisfy global constraints once they reach a steady state.
In a real-time environment, an algorithm might be interrupted by a communication failure or a strict computation deadline. If the solver is stopped at iteration $t$, and the iterate $x_t$ is infeasible, the system has no safe command to execute. Jiarui Wang argues in Anytime-Feasible First-Order Optimization via Safe Sequential QCQP that this lack of intermediate feasibility is a primary barrier to using advanced optimization in robotics and autonomous systems.
Primal Dynamics and Invariant Manifolds
The core intuition behind anytime feasibility is to ensure the optimization trajectory stays within the feasible set $\mathcal{C}$ at all times. This is often achieved by shifting the optimization from the dual space to the primal space.
In distributed resource allocation (DRA), a common constraint is the "sum-preserving" requirement: the total supply must exactly match the total demand ($\sum x_i = D$). Hamid R. Rabiee demonstrates in Single-Bit Distributed Equality-Constraint Optimal Scheduling and Allocation that by initializing the system at a feasible point and using Laplacian-gradient dynamics, the total resource sum remains constant throughout the optimization. Because the updates are structured as balanced peer-to-peer exchanges, the global equality constraint becomes a system invariant.
Similarly, Xuyang Wu and Sindri Magnússon introduce the Distributed Feasible Method (DFM), which uses barrier functions to ensure that inequality constraints (like generator ramp rates or battery capacities) are never breached. By treating constraints as "hard" boundaries that the gradient flow cannot cross, the algorithm ensures that any intermediate state is physically realizable.
Safety Buffers and Polyak Steps
In online settings where constraints might be adversarial or time-varying, maintaining feasibility requires proactive safety margins. Spencer Hutchinson and Mahnoosh Alizadeh propose using Polyak feasibility steps to achieve anytime satisfaction. Instead of expensive projections, they use simple, computationally efficient updates that nudge the system back into the safe zone at every round, maintaining optimal regret bounds while guaranteeing safety.
This "safety-first" approach is mirrored in power systems research. José Rivera shows that primal decomposition is superior to dual methods for electric vehicle charging because it ensures the local transformer is never overloaded during the calculation process. If the charging coordination is truncated, the vehicles still have a valid, albeit suboptimal, charging schedule that won't trip the circuit breakers.
Why It Works: Intuition from Control Theory
The success of anytime feasibility often stems from blending optimization with Control Barrier Functions (CBFs). In Optimal Power Flow Pursuit via Feedback-based Safe Gradient Flow, Antonin Colot and Jorge Cortés explain that while standard gradients point toward the optimum, "safe" gradient flows are filtered to ensure they never point "out" of the feasible set.
By viewing optimization as a dynamical system, researchers like Ahmed Allibhoy have developed Safe Monotone Flows. These algorithms ensure that the state $x(t)$ satisfies $g(x(t)) \leq 0$ for all $t \geq 0$, effectively turning the feasible region into a "forward invariant" set.
Summary of Benefits
The shift toward anytime feasibility represents a move from "optimization in a vacuum" to "optimization in the loop." The core advantages identified across this research direction include:
- Safe Early Termination: Systems can stop computing and act immediately if conditions change.
- Hardware Protection: Physical limits (voltage, torque, pressure) are respected during the transient phase, not just at equilibrium.
- Robustness to Noise: As shown by Alireza Aghasi in Robust-to-Noise Algorithms, sign-based nonlinear dynamics can maintain feasibility even when communication links are unreliable.
- Scalability: By avoiding global projections and using local primal updates, these methods scale to massive networks like the Internet of Things or smart grids.