Feasible-Set Reshaping Technique

• Feasible-set reshaping is a collection of analytic, algorithmic, and data-driven techniques that replace complex feasible sets with surrogates having desirable properties such as convexity and constraint qualification.
• Techniques like inner convex restriction, projection-based methods, and iterative convex hull computation are used to ensure feasibility and regularity in optimization, control, and learning applications.
• Applications in biomechanics, power systems, and deep metric learning demonstrate improved convergence, reduced computation times, and enhanced safety margins.

Feasible-set reshaping is a collection of analytic, algorithmic, and data-driven techniques for systematically modifying or approximating the feasible set of a mathematical program or control system. The aim is to obtain sets with desirable computational or physical properties—such as convexity, constraint qualification (e.g., LICQ), explicit feasibility guarantees, or more accurate reflection of physical or operational constraints—while retaining practical utility for optimization, learning, or control. Techniques span convex inner restriction, geometric tightening, high-probability feasible surrogates, and direct regularization or projection of constraints.

1. Geometric and Analytic Foundations

Feasible-set reshaping leverages the geometry of the original feasible set, typically defined by nonlinear equalities and inequalities, and replaces it with an analytic or algorithmic surrogate possessing strictly desirable properties (convexity, independence, tractable membership test). Core geometric ideas include:

• Inner (convex) restriction: Construct a convex region strictly contained in the original feasible set to ensure any candidate solution is guaranteed feasible. This approach is seen in convex restriction methodologies for nonlinear algebraic constraints, e.g., power flow equations in power networks, where polynomial or trigonometric nonlinearities define feasibility boundaries that are tightly inner-approximated using quadratic envelopes or supporting hyperplanes (Lee et al., 2018, Zhang et al., 2023).
• Constraint qualification via projection: Project the feasible set onto a subspace or polyhedral structure (e.g., through a constant matrix with full-rank and positive spanning rows), guaranteeing constraint gradients are linearly independent at all active faces. This transformation enables the satisfaction of the linear independence constraint qualification (LICQ) throughout the set and yields controllers or optimizers with desirable continuity and stability properties (Wu et al., 14 Dec 2025).
• Data-driven and level-set characterization: Empirical sampling and statistical learning produce boundary classifiers (e.g., SVM) representing the feasible set as a superlevel set of a smooth surrogate function. By adjusting margins, practitioners obtain inner and outer approximations with quantifiable conservatism or coverage of the true feasible boundary (Zhou et al., 2020).

2. Algorithmic Frameworks for Reshaping

State-of-the-art feasible-set reshaping methods utilize algorithmic routines designed for tractable approximation and iterative improvement:

• Iterative Convex Hull (ICHM) with Constraints: For muscle-wrench capacity in biomechanics, the wrench feasible set is defined as a convex polytope parameterized by muscle activation bounds, with vertices found via LP-based extremal search and convex hull recomputation. Augmenting this routine with physical (joint stability) constraints reshapes the feasible polytope, reducing unsafe regions and enforcing realism in ensuing solutions. The method is iterative and relies on repeated LPs and geometric updates (Rezzoug et al., 2023).
• Convex restriction via fixed-point and envelope bounds: For nonlinear equality-constrained sets, convex inner approximations are built using fixed-point linearization (around a working point), convex and concave quadratic envelopes of nonlinearity, and polyhedral bounding tricks. Sufficient feasibility is ensured by verifying that all possible realizations within the restriction comply with original constraints (Lee et al., 2018).
• Product space and constraint reduction in projection algorithms: In multi-set feasibility problems (e.g., wavelet filter design), constraints are reshaped by merging commutative pairs (whose projections commute) into their intersection, effectively lowering product space dimension and improving convergence rates and computational cost in algorithms like Douglas-Rachford and method of alternating projections (Dao et al., 2020).
• Gradient-based alternation and batch regularization: In deep metric learning, the intersection of relaxed feasible sets is explored using alternating projection steps, each enforced via a surrogate margin-based loss plus quadratic regularization. The effect is an incremental “tightening” or reshaping of the solution toward the true feasible set defined by the totality of constraints (Can et al., 2019).

3. Application Domains

Feasible-set reshaping is a unifying principle across diverse research areas:

Area	Goal of Reshaping	Canonical Reference
Musculoskeletal simulation	Enforce joint stability and obtain realistic wrench sets	(Rezzoug et al., 2023)
Networked power systems	Guarantee feasible power flows via convex restriction	(Lee et al., 2018, Zhang et al., 2023)
Optimization-based control	Ensure LICQ for parametric QPs in control law synthesis	(Wu et al., 14 Dec 2025)
Nonlinear MPC	Efficiently approximate feasible and invariant sets	(Zhou et al., 2020)
Projection algorithms	Reduce complexity by merging constraints	(Dao et al., 2020)
Metric learning	Incrementally satisfy proximity constraints	(Can et al., 2019)

In each case, the primary objective is to secure computational tractability, physical plausibility, or regularity properties that would be absent or unreliable with the original feasible set.

4. Mathematical Guarantees and Convergence Properties

Theoretical results underpinning feasible-set reshaping techniques vary by formulation and target:

• Convex restriction and feasibility: If the reshaped (typically convex) set is non-empty and strictly contained in the original, then every point satisfies the original constraints. Tightness is controlled by the choice of surrogate envelopes, polyhedral approximations, or sample resolution (Lee et al., 2018, Zhang et al., 2023, Zhou et al., 2020).
• Constraint qualification (LICQ) restoration: For QP-based control, projection-based reshaping ensures any active set at the boundary of the reshaped polytope yields linearly independent gradients. As a result, the control law exhibits Lipschitz continuity and boundedness—crucial for applications requiring robust parametric response (Wu et al., 14 Dec 2025).
• Convergence of iterative algorithms: For projection-based or LP-search algorithms, strictly decreasing cost or coverage and convergence to a local (sometimes global) optimum are established under convexity, regularity or superregularity assumptions. For stochastic or data-driven settings, high-probability guarantees for feasibility and optimality are provided, with explicit iteration and data-complexity bounds (Lin et al., 2019, Zhou et al., 2020).

5. Quantitative Impact and Empirical Results

Empirical studies demonstrate the quantitative effects of feasible-set reshaping:

• Biomechanical feasible sets: Imposing glenohumeral stability constraints reduced unsafe (dislocating) wrenches among polytope vertices from 55–76% to 2–5%, and decreased maximal force capacity by 45–71 N, highlighting significant reshaping of achievable actions (Rezzoug et al., 2023).
• Power systems optimization: Convex-restriction-based OPF returns solutions within ~0.4% of analytic global optima while never violating physical limits, outperforming both standard convex relaxations and linearized approaches in feasibility and cost metrics (Zhang et al., 2023).
• Control law regularity: For Lipschitz continuity, the reshaped feasible set allows controller outputs to vary continuously even across problematic state transitions where original feasible sets cause discontinuity due to failure of LICQ (Wu et al., 14 Dec 2025).
• Optimizer and learning algorithms: Data-driven and iterative projection approaches consistently exhibit higher feasibility rates and reduced computation time in nonlinear MPC, stochastic programs, and deep metric learning compared to naïve, unreshaped methods or single-pass subgradient solvers (Zhou et al., 2020, Lin et al., 2019, Can et al., 2019).

6. Methodological Trade-offs and Limitations

A recurring property across feasible-set reshaping methods is conservatism: reshaped sets are stricter than originals, potentially shrinking the set of admissible solutions and sacrificing some optimality margin for guaranteed feasibility or regularity. The choice of surrogate, envelope, or projection method sets the balance between tightness and computational tractability. Pseudopolynomial or O(|E|+|N|) scaling is typical for power systems, while online complexity depends on the specific geometric or learning-based surrogate in controller design or nonlinear programming.

7. Extensions and Future Directions

Active research seeks to minimize the conservatism inherent to feasible-set reshaping, adapt set geometry online, combine data-driven boundary learning with analytic surrogates, and extend methodologies to accommodate higher relative degree safety constraints, structured uncertainty, and hybrid constraint systems. There is interest in developing adaptive or hierarchical reshaping—where the surrogate set evolves jointly with system state or empirical data—to more tightly track operational envelopes with minimal loss of admissible behaviors while maintaining computational assurance (Wu et al., 14 Dec 2025, Zhou et al., 2020, Zhang et al., 2023).