Bounded Optimization

• Bounded optimization is a framework that enforces variable limits, constraint structures, and curvature conditions to ensure solution existence, computability, and algorithmic efficiency.
• It underpins techniques in linear, convex, nonlinear, and integer programming, influencing areas from dynamic optimal control to polynomial and Bayesian optimization.
• Recent research leverages geometric, duality, and SOS methods to shrink computational gaps while providing robust performance guarantees in complex optimization tasks.

Bounded optimization refers to the design, analysis, and implementation of optimization tasks where additional structure—such as variable bounds, constraint-satisfaction over finite regions, properties of the objective function, or spectral/geometric restrictions—guarantees existence, computability, or algorithmic efficiency of solutions. This concept permeates a broad spectrum of mathematical programming, encompassing linear, nonlinear, convex, combinatorial, and global optimization, and appears in areas ranging from optimal control to polynomial optimization and integer programming. Specific research advances address cases where bounds on the structure (e.g., domain bounds, minor size, curvature, degree) yield theoretical guarantees or computational improvements. The following sections synthesize key developments and methodologies in bounded optimization, as established or exemplified in recent literature.

1. Boundedness in Linear Programming for Optimal Control

Boundedness in LP-based dynamic programming is critical for ensuring finite and unique solutions to control problems over infinite-dimensional state spaces. For the policy-evaluation and optimal-value LPs in infinite-horizon discounted control, the feasible set forms a convex, salient, open cone with boundedness determined by the interplay between the objective direction and the constraint cone geometry.

A pivotal result is a Farkas-type criterion: for a finite-dimensional parameterization, the LP is bounded if and only if one can express the objective's direction as a conical combination of constraint differences:

$\int \Phi\, c = \sum_{i=1}^N \lambda_i\,M_i$

where $M_i$ are constraint row differences and $\lambda_i \geq 0$ (Falconi et al., 2023). In model-free scenarios, boundedness can be enforced constructively by either selecting the objective (e.g., covariance) to match the span of sampled constraint directions, or choosing data to fit a given objective, so the Farkas condition holds exactly.

In LQR settings, boundedness of the LP is equivalent to contractiveness of the closed-loop or reachability of the system—a concrete algebraic condition. These geometric and duality-based approaches provide necessary and sufficient data-driven conditions for boundedness, extending to polynomial and SOS function classes.

2. Bounded Convex Vector Optimization and Outer/Inner Approximation

Boundedness in convex vector optimization is operationalized via the concept of a bounded upper image. Given a convex feasible set and $C$-convex objectives, the upper image

$P = \mathrm{cl}(F[S]+C)$

is required to be contained within a shifted cone, i.e., $P \subset y+C$ for some $y$ (Dörfler et al., 2020). In this regime, Benson-type outer/inner polyhedral approximation algorithms efficiently construct sandwiching representations (in both $H$- and $V$- forms) of $P$, successively cutting off vertices of the outer approximation and enriching the inner. Boundedness ensures algorithmic termination and enables practical $M_i$0-approximation guarantees, as measured by the Hausdorff distance between outer, inner, and actual images.

The vertex-selection enhancement exploits boundedness to focus refinement on the largest discrepancy between approximations, halving the number of expensive scalar subproblems required per instance and reducing the computational cost without relaxing the approximation certificate. This is empirically validated on multi-objective benchmarks and engineering design problems.

3. Bounded Global Optimization for Polynomial Programs via Binary Reformulation

For global optimization of polynomial programs over box-bounded domains, boundedness is exploited to reformulate the problem into mixed-integer linear programs with guaranteed interval bounds. Each real variable is represented as a binary expansion plus remainder, truncated per a user-specified tolerance to ensure the solution space is confined:

$M_i$1

where $M_i$2 for variable $M_i$3 (Norman, 2012). Product terms are linearized using McCormick envelopes. Solving two bounding MILPs gives provable lower and upper bounds on the original optimum, with the gap proportional to the discretization tolerances.

As variable tolerances decrease, the MILP size increases but the bounds converge exactly to the continuous global optima, allowing explicit trade-off between computation and optimality. This method leverages boundedness both in the domain and in the controlled approximation error, applicable to high-degree polynomial programs.

4. Integer Optimization on Simplices with Bounded Constraint Minors

When integer optimization is performed over simplices defined by integer matrices, bounding the largest minors of the constraint matrix has profound algorithmic consequences. Specifically, if the maximum absolute determinant of any $M_i$4 or $M_i$5 minor of $M_i$6 is bounded, the width of the simplex and associated integer optimization problems can be computed in polynomial time (Gribanov et al., 2017).

The core technical insight is that with bounded minors, all relevant normal or edge cones admit unimodular decompositions of polynomial depth:

$M_i$7

with $M_i$8, enabling reduction of the integer program to a finite set of easy feasibility checks. This boundedness property contrasts sharply with NP-hardness of the general case.

5. Bounded Curvature in Submodular and Supermodular Optimization

For set function optimization under matroid constraints, the property of bounded total curvature $M_i$9 quantifies how much the function deviates from modularity. Approximation ratios for maximizing monotone submodular functions or minimizing supermodular functions scale with this curvature:

• Submodular maximization: a $\lambda_i \geq 0$0-approximation (Sviridenko et al., 2013)
• Arbitrary monotone functions: greedy achieves $\lambda_i \geq 0$1 (increasing) or $\lambda_i \geq 0$2 (decreasing) These bounds are tight in the value-oracle model. The continuous greedy algorithm and non-oblivious local search exploit curvature-based boundedness for improved solutions, and the curvature emerges in several real-world problems, such as entropy sampling and column-subset selection for PCA.

6. Bounded Local Subgradient Variation in Nonsmooth Optimization

A function class with bounded local subgradient variation—formally, bounded maximum local variation ("Grad-BMV") or bounded mean oscillation ("Grad-BMO")—captures nuanced structure beyond Lipschitz or Hölder continuity:

$\lambda_i \geq 0$3

This property allows sharper oracle complexity bounds for subgradient methods in nonsmooth convex and nonconvex optimization. With such boundedness, both deterministic and randomized accelerated gradient methods achieve rates dependent on the local variation constant, yielding better performance for non-worst-case functions (Diakonikolas et al., 2024).

Furthermore, in parallel complexity, the mean Gaussian width of the subdifferential around the optima fundamentally limits sequential depth but may enable substantial speedups in parallel by exploiting local boundedness not captured by global Lipschitz constants.

7. Bounded-Domain Bayesian Optimization with Boundary-Aware Kernels

In Bayesian global optimization for bounded domains $\lambda_i \geq 0$4, explicit boundary-awareness in the choice of Gaussian Process kernels is crucial when optima may lie near the domain's boundary or vertices. The Beta product kernel, constructed as a product of Beta densities parameterized by the input, ensures non-stationarity tuned to the bounded domain:

$\lambda_i \geq 0$5

with each $\lambda_i \geq 0$6 defined via Beta function integrals (Nguyen et al., 19 Jun 2025). The eigenvalue spectrum of this kernel exhibits exponential decay (statistically supported across high-dimensional regimes), facilitating effective learning and optimization in Bayesian frameworks.

Empirically, the Beta kernel consistently outperforms stationary kernels (e.g., Matérn, RBF) on synthetic benchmarks and neural model compression tasks, especially when the solution lies near the hypercube's faces or vertices. Hyperparameter tuning and integration into standard GP pipelines require no modification of acquisition strategies, and the computational overhead is equivalent to leading stationary kernels.

8. Bounded Degree SOS Hierarchies for Structured Polynomial Optimization

For a structured polynomial class—separable plus lower-degree (SPLD) polynomials—bounded-degree SOS (sum of squares) hierarchies (BSOS-SPLD) introduce significant computational savings by decomposing a multivariate problem into small univariate and lower-degree blocks:

$\lambda_i \geq 0$7

with $\lambda_i \geq 0$8 (Jiao et al., 17 Feb 2025). The BSOS-SPLD approach fixes degree bounds independently of relaxation order, ensuring all SDP blocks remain small regardless of the number of hierarchy steps.

The hierarchy exhibits strong duality, asymptotic convergence, and under rank-one conditions, finite convergence with moment extraction of global minimizers. Empirical results confirm that for convex SPLD problems, this framework yields faster solves, early convergence, and smaller absolute and maximum regression errors than standard BSOS hierarchies.

---

Bounded optimization thus forms a unifying theme for advances across optimization theory and algorithms, with domain-, curvature-, subgradient-, minor-, or degree-boundedness directly impacting computability, complexity, and empirical performance across diverse applications.