Extended Supporting Hyperplane Algorithm

• Extended Supporting Hyperplane Algorithm is a geometric and optimization-based framework that constructs polyhedral outer approximations of convex sets using supporting halfspaces.
• The algorithm iteratively projects current iterates via quadratic programming subproblems, achieving superlinear or quadratic convergence under suitable regularity conditions.
• ESHA extends to mixed-integer, nonconvex, and infeasibility detection contexts, enabling efficient constraint management and practical scalability in diverse applications.

The Extended Supporting Hyperplane Algorithm (ESHA) is a geometric and optimization-driven framework designed to efficiently solve set intersection, convex feasibility, and related optimization problems by iteratively constructing polyhedral outer approximations to convex sets via supporting halfspaces. Each iteration projects the current iterate onto a polyhedron defined by recent supporting hyperplanes, typically solving a quadratic program (QP). Under suitable regularity and curvature properties, the ESHA achieves global convergence with multiple-term superlinear or even quadratic asymptotic rates, and extends naturally to infeasibility detection. Implementation is enabled by dual active-set QP algorithms, which allow partial solves while retaining fast convergence properties. The ESHA also admits numerous variations, equivalences to classical cutting-plane approaches, and extensions to nonconvex and hybrid computational settings.

1. Problem Setting and Algorithmic Structure

The ESHA addresses two primary problems: the convex set intersection problem (SIP) and the convex inequality problem (CIP). In the SIP, the goal is to find a point in the intersection $K = \bigcap_{l=1}^r K_l$ of closed convex sets $K_l \subset \mathbb R^n$. In the CIP, the objective is to find $x$ such that $f(x) \leq 0$ for a convex $f : \mathbb R^n \to \mathbb R$, which can be viewed as intersecting the sublevel sets $\{ x : f(x) \leq 0 \}$ (Pang, 2014).

The general ESHA iteration proceeds as follows:

1. Generate supporting halfspaces: For each constraint set or function (or its subgradient at the current iterate), compute a supporting halfspace $H_i^{(l)}$ containing $K_l$, typically via projection or subgradient evaluation.
2. Aggregate cuts: Retain a window $p$ of recent (or otherwise selected) supporting halfspaces, defining a polyhedral feasible region $F_i = \bigcap_{(j,l) \in S_i} H_j^{(l)}$.
3. Project onto the polyhedron: Update the next iterate $K_l \subset \mathbb R^n$0 by (approximately) projecting $K_l \subset \mathbb R^n$1 onto $K_l \subset \mathbb R^n$2 via a strictly convex QP,

$K_l \subset \mathbb R^n$3

1. Over-relaxation (optional): Advance beyond the projection along the $K_l \subset \mathbb R^n$4 ray while preserving monotonicity (Fejér property).

This structure is shared by a variety of ESHA instantiations, including those for convex feasibility, mixed-integer programs, and constraint approximation problems (Serrano et al., 2019, Pang, 2014, Pang, 2012).

2. Convergence Analysis and Regularity Conditions

Critical to the ESHA's performance are geometric and variational properties at solution points:

• Local metric regularity: Existence of $K_l \subset \mathbb R^n$5 with $K_l \subset \mathbb R^n$6 for $K_l \subset \mathbb R^n$7 near $K_l \subset \mathbb R^n$8, often induced by a normal-cone condition (no nontrivial conic dependencies among set normals).
• Angle condition: For window size $K_l \subset \mathbb R^n$9, that among any $x$0 retained unit normals, some pair has acute angle $x$1 with $x$2, guaranteeing "coverage" of the relevant cone.
• Second Order Supporting Hyperplane (SOSH) property: $x$3 has SOSH at $x$4 if $x$5 for all $x$6 near $x$7, $x$8 a unit normal, implying that supporting halfspaces approach tangency in a quadratic fashion.

Given these, the ESHA achieves:

• Multiple-term superlinear convergence: $x$9 with sufficient window $f(x) \leq 0$0 and proper angle control.
• Multiple-term quadratic convergence (SOSH): $f(x) \leq 0$1 when SOSH holds (Pang, 2014).
• Finite termination (cones/normal alignment): For polyhedral cones or under pointed-cone alignment, the algorithm terminates in finitely many iterations (Pang, 2012).

These results generalize classical alternating projections, which converge only linearly except in the simplest cases, and connect to machinery from variational analysis such as metric regularity and normal cone calculus.

3. Quadratic Programming Subproblems and Partial Solutions

At each iteration, ESHA requires solving a strictly convex QP. Rather than requiring full resolution, global progress and fast asymptotic rates are maintained even if only partial active-set progress is made:

• The Goldfarb–Idnani dual active-set method incrementally adds or drops constraints, generating a sequence of intermediate projections $f(x) \leq 0$2 onto ever-larger subsets of the defining halfspaces. Each such partial projection satisfies Fejér monotonicity:

$f(x) \leq 0$3

so any nontrivial progress yields strict improvement towards feasibility (Pang, 2014).

• Warm-starts by maintaining the active set across outer iterations, and early termination after a small number of steps, significantly reduce computational expense without sacrificing convergence order when close to solution.

Pseudocode and operational description of the dual active-set (Goldfarb–Idnani) subroutine are given in (Pang, 2014), including cutting-plane management, constraint-dropping strategies, and over-relaxation techniques.

4. Extensions, Equivalences, and Applications

The ESHA framework subsumes, extends, and connects to several classical and modern algorithms:

• Kelley's cutting-plane algorithm: By a gauge reformulation, ESHA can be shown equivalent to Kelley's approach to nonsmooth convex minimization, with cuts arising from linearizations of the gauge function at boundary points (Serrano et al., 2019).
• Alternating projections acceleration: In Hilbert or Euclidean spaces, projecting onto the intersection of retained supporting halfspaces (rather than single sets) directly accelerates convergence, especially for families of affine subspaces (Pang, 2014).
• Mixed-integer and general nonlinear programs: By constructing outer polyhedral approximations via supporting hyperplanes derived from differentiable (possibly nonconvex) constraints with nonvanishing gradients, the ESHA achieves certified outer approximation and convergence to global optima for programs with continuous and integer variables (under KKT sufficiency on the boundary) (Serrano et al., 2019).
• Integer hull computation: In 2D polyhedral geometry, an ESHA-style approach translates facets via their supporting hyperplanes, anchored at opposite vertices, to efficiently construct the integer hull by reducing the volume of regions requiring enumeration (Mukherjee, 11 Sep 2025).
• Convex hull approximation and constraint learning: Dantzig-Wolfe-based column generation ESHA variants construct polytopic approximations with a fixed hyperplane budget, supporting practical convex hull approximation in high ambient dimensions (Barbato et al., 2024).

Table: Core ESHA Variants

Problem Domain	ESHA Mechanism	Reference
Convex Set Intersection	Project onto outer-approx. polyhedra	(Pang, 2014)
Mixed-Integer Convex Programs	Gauge-based cut generation, LP/MILP	(Serrano et al., 2019)
Affine Subspace Acceleration	Memory-based orthogonal projections	(Pang, 2014)
Integer Hull in 2D	Opposite-vertex anchored hyperplanes	(Mukherjee, 11 Sep 2025)
Hyperplane Budget Polytope Approx	Column generation, Dantzig-Wolfe	(Barbato et al., 2024)

5. Infeasibility Detection and Termination

The ESHA is equipped to recognize infeasibility in a finite number of iterations under mild conditions:

• As outer approximations shrink (nested $f(x) \leq 0$4), exact emptiness of $f(x) \leq 0$5 is detected during QP or active-set progress. If the true intersection is empty, infeasibility is certified in finitely many steps, provided supporting hyperplanes continue to cut off current iterates and there are no common recession directions among the constraint sets (Pang, 2014, Pang, 2012).
• In infeasible settings, the iterates diverge or "drift to infinity" along a direction from the intersection of recession cones, precluding the existence of strong cluster points.

This finitely certifying behavior is particularly relevant in hybrid or polyhedral geometry settings, such as integer hull computation or convex hull approximation with budgeted support.

6. Algorithm Implementation and Practical Considerations

Effective application of the ESHA involves:

• Cut selection and window management: Retaining most recent or most "active" cuts (by violation, angle, or age) keeps the QP dimension moderate, with window size $f(x) \leq 0$6 chosen to ensure normal coverage and rapid convergence when feasible (Pang, 2012, Pang, 2014).
• Redundant constraint management: Periodic removal or aggregation of nearly-parallel or inactive normals prevents QP degeneracy.
• Warm starting and partial solves: Active-set memory and partial projections economize computational resource allocation.
• Parameter tuning: Window size $f(x) \leq 0$7, over-relaxation parameter $f(x) \leq 0$8, and QP tolerance should be selected according to the alignment and regularity assumptions to maximize global and local convergence performance.

Typical per-iteration costs are dominated by QP solves with up to $f(x) \leq 0$9 linear constraints. For moderate $f : \mathbb R^n \to \mathbb R$0 and $f : \mathbb R^n \to \mathbb R$1, or in the presence of structure (e.g., sparsity, block separability), practical performance is tractable and scalable to medium-dimensional problems (Pang, 2014, Pang, 2012).

7. Theoretical and Practical Impact

The ESHA provides a unifying framework for accelerating projection-based algorithms for convex feasibility, optimization, and combinatorial geometry. Its ability to guarantee global convergence and multiple-term superlinear or quadratic local rates under regularity conditions broadens the applicability of classical methods, offering:

• Rigorous theoretical guarantees based on variational analysis, normal cone conditions, and polyhedral outer approximations.
• Broad generalizability across convex, nonsmooth, and some nonconvex settings, as well as mixed-integer and polyhedral combinatorial contexts.
• Practical relevance in optimization solvers, constraint separation, hybrid set estimation, and geometric algorithms.

The ESHA is the subject of ongoing research in algorithmic convex geometry, large-scale optimization, and computational integer programming, with extensions to stochastic, distributed, and approximation settings under active development.

References: (Pang, 2014, Serrano et al., 2019, Pang, 2014, Pang, 2012, Mukherjee, 11 Sep 2025, Barbato et al., 2024)