Nonlinear Binary Mathematical Program

• Nonlinear binary mathematical programs are optimization problems with nonlinear objectives and/or constraints, where decision variables are binary (0 or 1).
• They utilize convexification, penalty, and reformulation techniques (e.g., PQCR, MPEC, MILP encoding) to transform nonconvex models into tractable formulations.
• These programs have practical applications in fields like image restoration, revenue management, control systems, and computational biology.

A nonlinear binary mathematical program is an optimization problem where the objective function and/or the constraints are nonlinear in the decision variables, and all or some of the variables are restricted to take binary values, typically 0 or 1. Formally, for variables $x\in\{0,1\}^n$, the program is

$$\min_{x\in\{0,1\}^n} f(x) \quad\text{subject to}\quad g_j(x)\leq 0,\; j=1,\dots,m,$$

where $f$ and $g_j$ can be nonlinear and possibly nonconvex. This class includes, but is not limited to, unconstrained 0–1 polynomial optimization, Boolean quadratic programming, and discrete nonlinear models with arbitrary polynomial, transcendental, or combinatorial structure. These programs are central in combinatorial optimization, operations research, electrical and communication systems, computational biology, and machine learning.

1. Problem Structure and Modeling Paradigms

Nonlinear binary programs admit diverse forms in both objective and constraints, e.g., polynomials (quadratic, cubic, or higher order), general nonlinearities (e.g., involving $\sin$, $\log$, $\exp$), or submodular/fractional constructs over discrete grids. Even in unconstrained settings, the presence of nonlinear terms endows the feasible region with combinatorial and geometric complexity. A typical example is the unconstrained 0–1 polynomial program: $$\min_{x\in\{0,1\}^n} f(x) = \sum_{p=1}^m c_p \prod_{i\in\mathcal M_p} x_i,$$ with each monomial possibly spanning different variable subsets and degrees up to $d\geq 2$ (Elloumi et al., 2019). Related variants include the box-constrained nonconvex quadratic integer program (Xia et al., 2014), nonlinear binary control optimization in dynamic systems (Galindro et al., 2018), and nonlinear discrete functions on simplotope domains in revenue management (He et al., 2024).

Key binary modeling paradigms include:

• Explicit logical or combinatorial structure, typically encoded via monomials or indicator variables.
• Nonlinear functions mapping discrete (binary or integer-valued) variables to nonlinear cost/convenience or physical process terms.
• Constraints coupling binary and continuous variables (MIP, MINLP, MPEC formulations).

2. Convexification and Reformulation Techniques

Due to the inherent nonconvexity induced by both nonlinearity and integrality, most solution methods focus on reformulating the problem into a form amenable to tractable optimization via convex relaxations, surrogate penalties, or exact equivalences. Major approaches include:

• Polynomial Quadratic Convex Reformulation (PQCR): PQCR recursively quadratizes any multivariate polynomial in binary variables by introducing new binary variables for high-degree monomials, then seeks a quadratic convex (PSD) equivalent by adding null quadratic forms parametrized by Lagrange multipliers. The best parameters are derived from the dual solution of a semidefinite program (SDP), yielding a convex quadratic formulation whose continuous relaxation matches the SDP bound and whose feasible solutions are equivalent to the original (Elloumi et al., 2019).
• Mixed-Binary Convex Quadratic Reformulation: For box-constrained quadratic integer programs, variables are centered/transformed and represented via binary expansions, enabling the addition of diagonal lifting terms to convexify the Hessian. The optimal convexification parameters are computed by solving an SDP, and the final model is a convex MBQP efficiently solvable by standard branch-and-bound solvers (Xia et al., 2014).
• MILP Encoding via Complementarity: Nonconvex quadratic programs are reformulated as mixed-integer linear programs by encoding the KKT complementarity conditions using big-M constraints and binary variables. This enables the use of state-of-the-art MILP solvers for globally optimal solutions (Xia et al., 2015).
• Simplotope and Staircase LP/MIP Formulations: For classes of discrete nonlinear functions (submodular, bilinear, fractional, and $L^\natural$-convex), the simplotope structure induced by unary binarization variables is leveraged to construct ideal MIP or exact LP formulations via staircase (simplicial) convex hull inequalities. This extends Lovász-extension-based methods and yields LP relaxations that are tight in many retail and inventory management applications (He et al., 2024).

3. Exact Penalty and Continuous Relaxation Methods

Several frameworks have been developed to relax binary constraints into continuous domains, subsequently enforcing integrality via penalty approaches:

• Piecewise Cubic Penalty: Binary constraints $x_i\in\{0,1\}$ are replaced by a nonlinear equality $g(x_i)=0$ where $g(x_i)$ is a piecewise cubic that vanishes only at 0 and 1. Embedding this as a penalty in the objective yields a continuous optimization problem with an explicit exact-penalty parameter threshold, independent of the global minimizer. The adaptive proximal-point algorithm (APPA) leads to guaranteed convergence to (binary) stationary points in finitely many steps, provided the original objective is strongly smooth, and extends naturally to general nonlinear binary programs (Li et al., 27 Oct 2025).
• Mathematical Programming with Equilibrium Constraints (MPEC): Binary enforcement is equivalently posed as a complementarity (e.g., $x_i(1-x_i)=0$ or $x_i u_i=0$). Penalty and augmented Lagrangian methods with exactness thresholds in the penalty parameter enable global and local minima to correspond across the original binary and relaxed continuous formulations. Block coordinate and ADMM approaches are efficient for a variety of practical instances (Yuan et al., 2016).
• Difference-of-Convexity (DC) and Majorization-Minimization: For $\{-1,1\}^n$-constrained problems, the rank-one constraint in lifted SDP relaxations is represented via a difference of convex spectral functions. Penalties on the rank gap (nuclear norm minus spectral norm) are combined with Moreau-envelope smoothing of the loss function and optimized via efficient majorization-minimization (MM) in low-rank factorization space. The DCRA method achieves explicit complexity, feasibility, and optimality-gap bounds for general nonsmooth nonlinear losses (Xiao et al., 6 Jan 2026).

4. Algorithmic Strategies and Cutting-Plane Approaches

Several algorithmic paradigms address the combinatorial explosion inherent in nonlinear binary programs:

• Cutting-Plane Algorithms: General nonlinear, possibly nonconvex, binary programs can be addressed by iteratively building a master MILP with gradient-based feasibility and optimality cuts derived from subgradients of the objective and constraints. Quantitative complexity bounds are available when the cuts are sufficiently steep, and dual (KKT-type) optimality conditions are established for global certification. This approach outperforms standard MIQP linearizations in large quadratic knapsack instances (Bui et al., 2022).
• Polynomial-Time Algebraic Methods: For unconstrained binary quadratic (and, by Rosenberg, certain multilinear) optimization, explicit reductions to (large-scale) linear programs are possible via convex-hull lifts with consistency constraints and auxiliary variables. An $O(n^{15/2})$ algorithm has been implemented and experimentally validated, providing existence results in the complexity class $P$ for these special, but important, cases (Mulero-Martínez, 2020).
• Neural Approaches and Hypergraph Reformulation: Arbitrary nonlinear binary programs (including those with transcendental elements and dense nonlinear terms) can be polynomialized on the Boolean cube and mapped to weighted hypergraph structures. Hypergraph neural network (HyperGNN) models, trained with annealed objective penalties in a fully parallel GPU pipeline, can approximately solve large-scale nonlinear BIPs. This method is competitive or superior to state-of-the-art solvers in terms of scalability and quality for synthetic polynomial and real-world combinatorial optimization problems (Bai et al., 27 May 2025).

5. Applications and Problem-Specific Structural Exploitation

Nonlinear binary mathematical programs appear extensively in diverse fields. Distinctive application examples include:

• Image Restoration and Sequence Design: High-degree unconstrained 0–1 polynomial problems arise and are addressed effectively by PQCR, yielding substantial reductions in root-node relaxation gaps and enabling the solution of previously intractable benchmark instances (Elloumi et al., 2019).
• Vineyard Replacement and Age-Structured Control: In agricultural management, nonlinear binary programs model equipment replacement or crop renewal decisions, where the objective is a nonconvex polynomial over state and binary control variables subject to age-dynamics constraints. Standard MINLP solvers can address modest instances directly; decomposition or dynamic programming is needed for large planning horizons (Galindro et al., 2018).
• Retail Revenue Management and Discrete Choice Models: The convex-hull formulation and ideal MIP/LP relaxations for nonlinear discrete functions over simplotope domains enable efficient exact solution of price-promotion and inventory optimization problems under submodular, bilinear, and fractional demand models, frequently in polynomial time (He et al., 2024).
• Sparse Recovery, Communications, and Hashing: Penalty and DC relaxation algorithms efficiently solve large-scale, nonsmooth nonlinear binary problems arising in statistical estimation, signal processing, and data mining (Xiao et al., 6 Jan 2026, Li et al., 27 Oct 2025).

6. Computational Aspects and Performance Benchmarks

Methodologies are evaluated extensively using benchmark problem classes and real-world data. For example, PQCR demonstrates root-node gap reductions of up to two orders of magnitude and consistently outperforms general MINLP solvers on instances of up to hundreds of variables (Elloumi et al., 2019). The mixed-binary convex quadratic reformulation achieves complete resolution of all tested box-constrained instances up to dimension 50, outperforming SDP- and ellipsoidal-relaxation-based solvers (Xia et al., 2014).

Neural and hypergraph-based methods are shown to scale to problems with thousands of variables where classical solvers are computationally prohibitive (Bai et al., 27 May 2025). Iterative exact-penalty, MPEC, and DC-relaxation approaches exhibit finite or explicitly bounded convergence and superior runtime or solution accuracy compared to established convex relaxations and LP/SDP rounds (Li et al., 27 Oct 2025, Yuan et al., 2016, Xiao et al., 6 Jan 2026).

7. Theoretical Foundations and Emerging Directions

Advances in the theory of nonlinear binary programming include:

• Establishing equivalence between binary programs and their relaxed or penalized surrogates under explicit penalty or convexification parameter thresholds, which depend only on problem data, not unknown solutions [(Li et al., 27 Oct 2025); (Yuan et al., 2016); (Xia et al., 2014)].
• Demonstrating polynomial-time algorithms for unconstrained binary quadratic programming and certain multilinear extensions through explicit convex-hull and consistency constraints (Mulero-Martínez, 2020).
• Providing finite, often explicit, complexity and optimality-gap guarantees for continuous relaxation algorithms, including DC and block coordinate methods (Li et al., 27 Oct 2025, Xiao et al., 6 Jan 2026).
• Expanding the set of nonlinear discrete optimization problems that admit exact LP/MIP formulations through exploitation of simplotope and submodular structure (He et al., 2024).
• Algorithmic frameworks that treat nonconvexity, nonsmoothness, and high-order coupling using a combination of exact penalty, spectral difference-of-convexity, and neural optimization methodologies (Li et al., 27 Oct 2025, Xiao et al., 6 Jan 2026, Bai et al., 27 May 2025).

Collectively, these results significantly broaden the tractable frontier for nonlinear binary mathematical programming, bridging classical polyhedral and convexification techniques, algebraic and spectral penalty methods, and large-scale modern computational and neural paradigms.