In classical constrained optimization, constraints are typically treated as a destination rather than a journey. Standard algorithms wind through "illegal" territory—the infeasible space—with the mathematical promise that they will land safely within the boundaries by the final iteration.
However, in many real-world applications, this asymptotic guarantee is a luxury we cannot afford. If an algorithm is interrupted early, or if the intermediate steps represent physical states in a real-time system, an infeasible iterate is not just imperfect—it is potentially catastrophic. This reality has driven a robust research direction focused on feasible path maintenance: designing optimization frameworks where every single intermediate iterate remains strictly feasible.
The Problem with Asymptotic Feasibility
Why do standard optimization algorithms allow intermediate iterates to violate constraints? The short answer is computational efficiency. Enforcing strict feasibility at every step requires solving expensive subproblems or performing complex projections.
As noted by Chen et al., many traditional optimization techniques only satisfy feasibility requirements at the final output. Common approaches handle constraints by penalizing violations, which inherently allows intermediate iterates to drift outside the feasible region (De Marchi et al.).
This drift becomes particularly severe under two conditions:
- Inexact Updates: When algorithms use approximate solutions for intermediate steps to save time, errors accumulate. Chowdhury et al. demonstrate that this causes the algorithm to drift away from the feasible region, rendering intermediate results practically unusable.
- Stochastic Environments: In stochastic optimization, algorithms often rely on asymptotic convergence or "shadow sequences" to find a valid solution, meaning the actual intermediate iterates do not strictly reside within the feasible set (Boţ et al.). Lin et al. point out that this lack of strict feasibility introduces severe numerical instability during stochastic updates.
Why We Need an Unbroken Feasible Path
The demand for iterative feasibility is anchored in practical deployment and safety-critical systems.
Consider real-time robotics and Nonlinear Model Predictive Control (NMPC). As Kiessling et al. explain, standard algorithms often generate intermediate solutions that violate physical constraints like motor limits or safety boundaries. In time-critical robotics, an operator might need to stop the optimizer early and execute the best available path. If that intermediate path is infeasible, the robot cannot safely execute it.
Similarly, Martinsen et al. emphasize that maintaining feasibility with respect to nonlinear inequality constraints throughout the optimization process is paramount for controller reliability in sensitive applications. If the optimizer enters a prohibited region of the state space during its search, it risks system failure.
Standard Optimization (Infeasible Path):
Start ───► [Infeasible Space] ───► [Infeasible Space] ───► Feasible Target (End)
Feasible-Path Optimization:
Start ───► [Feasible Iterate] ───► [Feasible Iterate] ───► Feasible Target (End)
│ │
▼ ▼
(Safe to deploy) (Safe to deploy)
Core Mechanisms for Maintaining Feasibility
To maintain a strictly feasible path without destroying computational efficiency, researchers utilize several key algorithmic mechanisms:
1. Search Direction Tilting and Lightweight Projections
Instead of moving blindly and projecting back to the feasible set afterward, algorithms can proactively alter their trajectory. Martinsen et al. introduce a search direction tilting mechanism that modifies the optimization path specifically to ensure intermediate iterations remain within legal bounds. When dealing with highly complex constraints, such as mixed-integer nonlinear programming, the authors of Tang et al. leverage neural networks with specialized integer correction layers and lightweight projection steps to guide solutions toward the feasible region rapidly without relying on heavy external solvers.
2. Level-Set Methods and Proximal Decompositions
Another strategy involves transforming complex, constrained problems into a sequence of simpler, manageable subproblems. Lin et al. utilize affine-minorized level-set variants to maintain a feasible solution path throughout stochastic finite-sum optimization, eliminating the instability associated with infeasible updates.
For nonconvex and nonsmooth optimization, Boob et al. (2020) and Boob et al. (2025) develop level-constrained proximal gradient methods. These frameworks break down challenging functional constraints into convex subproblems, ensuring that every iteration remains strictly inside the feasible region. Similarly, Scutari et al. show that iteratively solving strongly convex approximations allows parallel and distributed systems to naturally maintain feasibility at every iteration.
3. Progressive Relaxation in Decentralized Spaces
In decentralized environments where local information is fragmented, maintaining strict feasibility is even harder. To prevent divergence, Hong et al. introduce a multi-stage strategy where constraints are progressively relaxed, keeping intermediate solutions bounded within manageable, compact sets throughout execution.
The Acceleration Dilemma
A major frontier in this research area is combining feasible-path optimization with acceleration techniques. Numerical tricks like Anderson Acceleration speed up convergence by taking a linear combination of historical iterates. However, because the feasible region is often nonlinear or non-convex, mixing past iterates can easily throw the new accelerated iterate outside the feasible boundaries.
Jin et al. identify this exact tension as a critical challenge when applying acceleration to derivative-free projection methods. To resolve this, frameworks like the one proposed by Kiessling et al. carefully integrate Anderson Acceleration directly into Feasible Sequential Linear Programming (FSLP), proving that algorithms can achieve both high-speed convergence and strict, real-time safety guarantees.
The Theoretical Payoff
While the primary motivation for feasible path maintenance is practical safety and early-stopping capability, the paradigm offers profound theoretical advantages.
As Boob et al. demonstrate, maintaining a strictly feasible path significantly simplifies the theoretical analysis of optimization algorithms. In mathematical optimization, bounding Lagrange multipliers is notoriously difficult when iterates wander in and out of constraints. By forcing the algorithm to stay within the feasible domain, the analytical tractability and robustness of the method are substantially enhanced. Feasible-path optimization thus bridges the gap between rigorous mathematical guarantees and the uncompromising safety requirements of physical deployment.
Key Foundational Work
A new optimization algotithm with application to nonlinear MPC (2005) — contributed 4 concept elements to this direction.
Efficient Low-Rank Stochastic Gradient Descent Methods for Solving Semidefinite Programs (2014) — contributed 1 concept elements to this direction.
Parallel and distributed methods for nonconvex optimization (2014) — contributed 1 concept elements to this direction.