Binary Quadratic Problems (BQP)

• Binary Quadratic Problems (BQP) are discrete optimization challenges involving binary variables and quadratic terms, leading to NP-hard combinatorial structures.
• Linearization techniques like Glover–Woolsey and compact schemes convert nonconvex bilinear forms into tractable constraints, enhancing solver performance.
• Continuous relaxations and metaheuristic methods, including SDP relaxations and quantum annealing, offer scalable strategies for solving large-scale BQP instances.

A Binary Quadratic Problem (BQP) is a discrete optimization problem whose variables are binary and whose objective includes quadratic (i.e., bilinear) terms. BQPs provide a powerful modeling backbone for a wide array of NP-hard combinatorial optimization problems, including assignment, partitioning, graph-based tasks, and emerging applications in machine learning and quantum computing. The core challenge is the intersection of binary combinatorics with nonconvex quadratic structure, which induces both theoretical hardness and distinctive algorithmic phenomena.

1. Formal Definition and General Modeling Structure

A general BQP can be expressed as:

$$\min_{x \in \{0,1\}^n} \; x^\top Q x + c^\top x, \quad \text{subject to} \quad A x = b, \quad C x + D y \geq e, \quad y_{ij} = x_i x_j \;\forall (i,j) \in P,$$

where $Q\in\mathbb{R}^{n\times n}$ is symmetric, $c\in\mathbb{R}^n$, constraints can include linear equations and inequalities, and $y_{ij}$ represent bilinear products introduced through linearization. This framework subsumes classical unconstrained BQPs ($A, C, D$ absent), as well as linearly constrained (mixed-integer) variants. Nonlinearity is entirely driven by the $x_i x_j$ terms, with binary $x_i$ enforcing combinatorial structure (Mallach, 2017).

Specific instances such as the unconstrained binary quadratic problem (UBQP):

$\max_{x \in \{0,1\}^n} x^\top Q x$

or the bipartite and multi-quadratic generalizations, are recurring formulations in combinatorial optimization and machine learning (Lewis et al., 2017, Karapetyan et al., 2012).

2. Classical and Modern Linearization Schemes

Quadratic terms in BQP prevent direct usage of linear integer programming techniques. Linearization, the process of replacing products $x_i x_j$ with continuous variables $y_{ij}$ along with linking constraints, is essential.

2.1 Glover–Woolsey (McCormick) Linearization

For each $Q\in\mathbb{R}^{n\times n}$0:

$Q\in\mathbb{R}^{n\times n}$1

Applied naively, this requires $Q\in\mathbb{R}^{n\times n}$2 constraints, leading to substantial growth in model size and often weak relaxations (Mallach, 2017).

2.2 Compact Linearization

Compact linearization leverages existing linear equations with positive coefficients to eliminate redundant constraints. For an equation $Q\in\mathbb{R}^{n\times n}$3 and multipliers $Q\in\mathbb{R}^{n\times n}$4, multiplying both sides by $Q\in\mathbb{R}^{n\times n}$5 gives tailored equality constraints on $Q\in\mathbb{R}^{n\times n}$6:

$Q\in\mathbb{R}^{n\times n}$7

collectively ensuring $Q\in\mathbb{R}^{n\times n}$8 under appropriate covering conditions (C1–C2) (Mallach, 2017). In assignment-structured problems and certain combinatorial designs, this scheme yields provably at-least-as-strong relaxations as the standard McCormick approach, with dramatically fewer constraints:

Linearization	# Variables	# Constraints	LP Strength (Special Cases)
Glover–Woolsey	$Q\in\mathbb{R}^{n\times n}$9	$c\in\mathbb{R}^n$0	Baseline
Compact	$c\in\mathbb{R}^n$1 ($c\in\mathbb{R}^n$2)	$c\in\mathbb{R}^n$3	Provably as strong for $c\in\mathbb{R}^n$4 or $c\in\mathbb{R}^n$5

(Mallach, 2017)

3. Continuous Relaxations and Relaxation Hierarchies

Classical continuous relaxations substitute binary constraints $c\in\mathbb{R}^n$6 with $c\in\mathbb{R}^n$7, and either encode bilinear terms directly or via linearization. Multiple relaxation regimes exist:

• Linear Programming (LP) Relaxation: Linearizes all quadratic terms, relaxes binary constraints.
• Spectral Relaxation: Replaces $c\in\mathbb{R}^n$8 with $c\in\mathbb{R}^n$9, reduces to an eigenvalue problem. Computationally simple but yields loose bounds (Wang et al., 2013).
• Semidefinite Programming (SDP) Relaxation: Lifts to $y_{ij}$0, enforces $y_{ij}$1, $y_{ij}$2, and (optionally) $y_{ij}$3. SDP gives tight bounds (e.g., $y_{ij}$4-approximation for Max-Cut) but is expensive for large $y_{ij}$5.

Modern developments include "SDCut", a scalable SDP variant that attains comparable bounds to interior-point SDP, but at spectral method-like cost per iteration (Wang et al., 2013). Equivalence results show that doubly nonnegative relaxations and tight Shor-type SDP relaxations are fundamentally the same for BQP, including in the presence of constraints (Guo et al., 2012).

For special matrices (e.g., low completely positive rank), PTAS and QPTAS schemes exist for constrained BQPs (Elbassioni et al., 2014).

4. Algorithmic Frameworks and Metaheuristics

Due to NP-hardness, practical BQP solvers incorporate a portfolio of exact and heuristic techniques:

• Branch-and-Bound with SDP and Hypermetric Cuts: BiqBin transforms linearly-constrained BQP into Max-Cut through exact penalization, applies strengthened SDP relaxations with triangle, pentagonal, and higher-order inequalities, and uses bundle methods for bounding (Gusmeroli et al., 2020). Advanced branching (most-fractional variable selection, best-first search) and MPI-based parallel tree exploration are standard at scale.
• Column Generation with Quantum or Simulated Annealing: Pricing subproblems (i.e., finding entering binary columns) are handled by quantum or classical annealers, dramatically accelerating certain classes of constrained BQPs. Empirical studies show 2.7–1000$y_{ij}$6 speedup versus standard solvers for mid-sized problems (Hirama et al., 2023).
• Metaheuristics for Specialized Structures: Path Relinking combined with Tabu Search for unconstrained and smooth penalty-embedded BQPs (e.g., for minimum sum coloring) achieves best-known results on challenging benchmarks (Wang et al., 2013). Markov chain-based operator mixing, as in Conditional Markov Chain Search (CMCS), unifies and outperforms traditional neighborhood heuristics on bipartite BQP extensions (Karapetyan et al., 2016).
• Cover–Relax–Search Primal Heuristics: Vertex cover-based reduction methods, coupled with fractional-relaxation-guided variable fixing, yield rapid high-quality feasible solutions for large-scale BQP instances and outperform BnB and local-search baselines in primal integral metrics (Huang et al., 9 Jan 2025).

5. Robustness, Applications, and Extensions

5.1 Robust BQP and Uncertainty

For UBQP where $y_{ij}$7 is perturbed within a box, robust solutions are computed via scenario generation (fractional factorial design) and aggregation. Robustness is quantified both at the solution-level (frequency and average performance across $y_{ij}$8 instances) and variable-level (dominance criteria, diagonal dominance) (Lewis et al., 2017). Fast regression-based upper bound approximations enable near-instant performance bounds for new $y_{ij}$9 realizations.

5.2 Application Domains

BQPs model core combinatorial optimization instances:

• Quadratic Assignment Problem (QAP): Assignments, solved with compact linearizations exploiting problem structure. The Frieze–Yadegar scheme is an instance of this compact model (Mallach, 2017).
• Quadratic Traveling Salesman Problem (QTSP): Edge-incident quadratic terms handled with compact linearization, matching the Fischer–Helmberg approach (Mallach, 2017).
• Data mining tasks: Matrix factorization, discrete pattern mining, maximum biclique, cut-norm calculations are reducible to bipartite/unconstrained BQPs (Karapetyan et al., 2012, Karapetyan et al., 2016).
• Machine learning: Supervised/unsupervised hashing, binary classifier code learning, classification model compression—here, BQP subproblems are attacked by SDP relaxations, augmented Lagrangian methods, and specialized local search (e.g., sequential bit-flip with local-optimality guarantees) (Do et al., 2016, Shen et al., 2016).

5.3 BQP and Quantum Computing

Quantum annealing directly implements unconstrained BQP solutions by mapping to Ising Hamiltonians, and emerging hybrid methods integrate quantum subroutines into classical column generation or BnB frameworks. Quantum approaches are competitive for certain classes, especially when the bottleneck is unconstrained QUBO extraction (Hirama et al., 2023, Ronagh et al., 2015).

6. Theoretical Challenges and Approximation Results

• Complexity: General BQP remains NP-hard; even quadratic cover/packing with a single constraint is intractable. Current theoretical pathways exploit properties of the constraint matrices: low completely positive rank (cp-rank), positive semidefiniteness, or problem sparsity (Elbassioni et al., 2014).
• Approximation Schemes:
  • Submodular $A, C, D$0 and cp-rank-bounded $A, C, D$1: $A, C, D$2-approximation, PTAS for linear $A, C, D$3 (Elbassioni et al., 2014).
  • SDP relaxation yields $A, C, D$4-approximation for Max-Cut, $A, C, D$5-approximation for binary code inference per randomized rounding analysis (Do et al., 2016).
• Relaxation Tightness: Compact linearizations outperform or match classical McCormick bounds for several structured instances (assignment, right-hand-side-2 constraints). Hypermetric strengthened SDP relaxations further close integrality gaps for Max-Cut derived BQPs (Mallach, 2017, Gusmeroli et al., 2020).
• Scalability: Novel formulations such as SDCut and bundle-SDP with adaptive strengthening yield scalable relaxations without material bound sacrifice, enabling exact or near-exact solution of instances with $A, C, D$6 up to 10,000 in specific settings (Wang et al., 2013, Gusmeroli et al., 2020).

7. Outlook and Open Directions

• Automated Linearization: Optimization-driven selection of multiplier sets $A, C, D$7 (cover sets) in compact linearization can be cast as a MIP, yielding provably most-compact and tight LP relaxations, suitable for automatic integration into MIP solvers (Mallach, 2018).
• Algorithmic Integration of Quantum Hardware: Hybrid quantum-classical workflows (branch-and-price, Lagrangian dual subproblem solvers) are becoming practical with current quantum devices; ongoing research benchmarks these against combinatorial and branch-and-cut alternatives (Hirama et al., 2023, Ronagh et al., 2015).
• Primal-Dual Integration and Bounding: Recent primal heuristics, such as the Cover–Relax–Search algorithm, show drastic improvements in the primal integral relative to classical BnB and local search on large randomized and structured BQP benchmarks (Huang et al., 9 Jan 2025).

Direct equivalence between stronger SDP and doubly nonnegative relaxations shows relaxation strength can be preserved across convexification schemes, allowing flexibility in solver choice and model encoding (Guo et al., 2012).

BQP remains both a central challenge and a proving ground for advances in optimization theory, high-performance algorithmics, and quantum annealing systems. Continuing research focuses on scalable relaxations, automated problem reformulation, robust/uncertain BQP, tight approximation schemes, and the integration of emerging hardware for problem classes at the intersection of combinatorics and nonconvexity.