Optimization-Based Bound Tightening (OBBT)

• Optimization-Based Bound Tightening (OBBT) is a domain reduction method that computes tighter variable bounds by solving auxiliary optimization problems.
• It is applied in mixed-integer nonlinear, nonconvex, and networked optimization scenarios to significantly reduce feasible regions and enhance relaxation strength.
• Variants such as strong, weak, topology-based, and rolling-horizon OBBT balance computational cost with bound quality, enabling scalable applications in neural network verification, power systems, and polynomial optimization.

Optimization-Based Bound Tightening (OBBT) is a preprocessing and in-branching domain reduction paradigm for mathematical programs with bounded variables, especially mixed-integer nonlinear and nonconvex problems. OBBT computes tighter variable bounds via a sequence of auxiliary optimization problems, systematically reducing feasible regions and strengthening relaxations without sacrificing optimality. The method is broadly applicable and has yielded significant advances in neural network verification, power systems optimization, nonconvex flows, and combinatorial network design.

1. Core Principles and Algorithms

OBBT operates by, for each bounded variable or selected aggregate, solving two auxiliary optimization problems—one minimizing and one maximizing the variable—subject to the model’s constraints and, optionally, objective bounds. Let $\mathcal{R}$ be a (convex) relaxation of the original model and $z(x)$ its objective. For each variable $v_k$ with current bounds $[\ell_k, u_k]$, OBBT solves:

• $$\ell_k' = \min_{x \in \mathcal{R},\, z(x) \le Z^{\rm U}} v_k(x)$$
• $$u_k' = \max_{x \in \mathcal{R},\, z(x) \le Z^{\rm U}} v_k(x)$$

Upon improvement, the bounds are updated and the process repeats (potentially propagating tightenings to other variables) until a fixed-point is reached or a computational budget is exhausted. The approach is exact for convex relaxations but can be applied more generally by choosing relaxations suited to the nonconvexities and combinatorial structure of the problem. Parallelization is straightforward since all bounding subproblems are decoupled.

The technique admits multiple modes:

• Strong OBBT: Solves the problem with all integrality enforced, guaranteeing maximal bound contraction but suffering high computational cost.
• Weak OBBT: Relaxes binaries or nonlinearities, solving cheaper subproblems at the expense of weaker bounds.
• Topology- or structure-based OBBT: Maintains integrality or exact nonlinearity on local subgraphs or neighborhoods for select variables (particularly effective in network optimization or graph-structured ML) (Pineda et al., 22 Jul 2025, Hojny et al., 2024).
• Rolling-horizon/Layerwise OBBT: Applies OBBT to subgraphs, moving a window through the problem (e.g., neural net layers) to achieve near-strong bounds at scalable cost (Zhao et al., 2024).

2. OBBT in Mixed-Integer Programming and Topological Optimization

In MILPs with big-M constraints coupling binaries and continuous variables, naive OBBT often relaxes all binaries for speed, admitting fractional solutions and producing weak bounds. In networked problems such as optimal transmission switching (OTS), this breaks physical relationships (e.g., Kirchhoff’s laws), resulting in subpar relaxation strength.

Topology-aware OBBT preserves integrality for a selected local neighborhood of the target variable (e.g., a $k$-hop neighborhood around a line in the power grid), while relaxing the rest. In OTS on the IEEE 118-bus system (Pineda et al., 22 Jul 2025), a hybrid setting with $k=2$ achieves a 52% reduction in solution time (from 327s to 160s) and nearly doubles reduction in hard instances, compared to MIP or full relaxation OBBT. This local-integrality strategy generalizes to structured MILPs in energy, water, and facility design.

OBBT Variant	Bound Quality	Subproblem Cost	Suitability
All binaries LP	Weak	Fast	Unstructured MILPs
Local integrality	Strong (local/global)	Moderate to high	Networked/physical systems
Full integrality	Strongest	Highest	Small-scale or last-stage BT

3. OBBT in Neural Network Verification and Control

OBBT is indispensable in the MILP-based formal verification and certification of deep neural nets with ReLU activations. Each neuron’s preactivation bound is tightened by solving MILP subproblems up to its layer, with integrality on the ReLU indicators:

• Strong (full MILP) OBBT per neuron yields the tightest possible big-M coefficients but may be intractable for large networks (Badilla et al., 2023).
• LP-relaxation OBBT (dropping integrality) yields bounds within 10–20% of the strong OBBT, at 5–20× lower cost.
• Layerwise or rolling-horizon OBBT, in which only subsequences of layers are included per subproblem, provides nearly full-strength bounds at scalable cost and can be massively parallelized (Zhao et al., 2024, Sosnin et al., 2024).
• In mixed-integer reachability over multiple timesteps (controller verification), OBBT on the unrolled network reduces relaxation gaps by 100–1000× relative to interval propagation, and strong OBBT achieves nearly exact bounds but at high cumulative solve times (Sosnin et al., 2024).

Numerical benchmarks (Zhao et al., 2024) show rolling-horizon OBBT–RH stabilizes (fixes) almost as many ReLU neurons as full OBBT (654.2 vs. 653.9), with bound ranges an order of magnitude smaller than leading propagation methods (mean input-bound range: 12.4 for OBBT-RH vs. 53.5 for α-CROWN, 2028 for IBP) and verifies 80/90 instances under 20 min budget, outperforming prior relaxations by $2.2\times$.

4. OBBT for Nonlinear Programming, Polynomial Optimization, and Nonconvex Flows

OBBT, originally developed for general nonconvex NLP and MINLP, is central to shrinking relaxation gaps for polynomial optimization and global solvers. In RLT-based (reformulation–linearization technique) methods for polynomial problems, OBBT using SOCP or SDP relaxations (for the bounding subproblems) yields measurable improvements in bounds at moderate overhead (SOCP-based OBBT reduces average gap from 0.132 to 0.128; SDP relaxations provide no practical gain over SOCP due to cost) (Gómez-Casares et al., 2024).

In the pooling problem, OBBT is applied to arc flows and aggregate constraints, with each subproblem solved as an LP over a strong rank-one-based relaxation. On real mining instances, OBBT halves the dual gap (7.2%→3.0%) and reduces solve time by 40–50%, with diminishing returns for further polyhedral cutting (Jalilian et al., 2023). The methodology generalizes to any network–flow or rank–constrained QCQP via selection of problem-specific relaxations.

Domain	Relaxation in OBBT	Quantitative Impact
Polynomial Opt.	SOCP, SDP, RLT	SOCP gap: 0.128 vs. 0.132
Pooling	Convex hull, SOCP	Gap: 3.0% vs. 7.2% (no BT)
AC-OPF	Strengthened QC, SDP	Gap: <1% vs. 45% (no BT)

5. OBBT in Semidefinite and Conic Relaxations for Power Systems

In power system OPF (both AC-OPF and ACOTS), OBBT is applied to reduce the feasible domains of continuous variables (voltages, flows, angles), substantially strengthening relaxations:

• OBBT is executed via a sequence of convex programs (SOCP, SDP, or strengthened QC spectral relaxations) per variable or per cycle, repeatedly until convergence (Sundar et al., 2018, Guo et al., 2022).
• Cost-aware OBBT incorporates a tight upper bound on the original cost function, excluding uneconomic (but feasible) points from the auxiliary problems (Awadalla et al., 2023).
• In AC-OPF, OBBT on strengthened QC-TLM relaxations drives the optimality gap below 1% on 90% of 57 benchmark networks (vs. up to 45% gap with no tightening) (Sundar et al., 2018).

Key metrics from (Guo et al., 2022) show OBBT+cycles+QC-relaxation ("ECB") yields the tightest known QC-based relaxations, closing >90% of computational gaps on large-scale power grids within minutes.

6. Algorithmic Variants and Practical Implementation Guidelines

The success and tractability of OBBT depend on the relaxation strength, degree of integrality retained in subproblems, and exploitation of problem structure.

• Initialization: OBBT typically begins from naive interval or propagation-based bounds (e.g., IBP or loose relaxations).
• Termination: Subproblems may be cut off via bounds on total resource usage, time per subproblem, or lack of improvement.
• Fallback: If a bounding subproblem runs into numeric or timeout errors, revert to weaker or naive bounds.
• Aggressive methods: In branch-and-bound, dynamic OBBT (e.g., topology-based abt for GNN verification) can be incorporated as local cut generation at B&B nodes (Hojny et al., 2024).
• Parallelization: All OBBT subproblems per round are independent (per variable or neuron) and highly amenable to massive parallel or distributed implementation.

Best practices (Badilla et al., 2023, Zhao et al., 2024):

• Use weak (LP-relaxation) OBBT as default for large-scale or when preprocessing speed is important.
• Hybrid schemes (naive in shallow layers, weak/strong OBBT in deeper layers) yield an efficient trade-off for deep or wide networks.
• In topology-rich MILPs, retaining local integrality is nearly always superior to full relaxation for bound tightening versus computational effort.
• Always set practical per-node or global time limits and build in robust fallback mechanisms.

7. Impact, Benchmarks, and Future Directions

OBBT is now integral in the state-of-the-art for certifying robustness of neural network controllers, screening redundant power system constraints, strengthening process flows, and accelerating MINLP solvers. Benchmarks on VNN-COMP, IEEE, MINLPLib, and PGLib networks universally report tightening by OBBT halves or better the relaxation gap, cuts total solve times by up to 50%, and increases problem instances closed to optimality.

Open research directions include:

• Automatic selection of OBBT window/horizon in rolling-horizon schemes (Zhao et al., 2024).
• Learning-based adaptive selection of subproblem granularity and relaxation level (Gómez-Casares et al., 2024).
• Integration with advanced solver APIs for seamless in-branch bound tightening and cut separation (Pineda et al., 22 Jul 2025).
• Extension of topology- and symmetry-based OBBT to heterogeneous and dynamic-graph models.

OBBT constitutes a unified, broadly applicable paradigm for domain reduction in complex optimization, synthesizing convex relaxation, network topology, and discrete structure to produce the tightest provable relaxations with algorithmic efficiency and practical scalability.