Sequential Bound Tightening (SBT)

• Sequential Bound Tightening (SBT) is an iterative procedure that refines variable bounds by solving auxiliary optimization subproblems and updating convex relaxations until convergence.
• The technique leverages methods like McCormick envelopes, QC and LP relaxations to systematically narrow the feasible region, thereby improving solution quality in complex models.
• SBT is applied in domains such as AC optimal power flow, neural network verification, and global bilinear optimization, balancing computational cost with the benefits of tighter relaxations.

Searching arXiv for papers on Sequential Bound Tightening and closely related optimization-based bound tightening methods. arxiv_search(query="Sequential Bound Tightening optimization-based bound tightening arXiv", max_results=10) arxiv_search(query="Sequential Bound Tightening optimization-based bound tightening", max_results=10) Sequential Bound Tightening (SBT) is an iterative domain-reduction procedure in which bounds are tightened by solving auxiliary optimization problems, the relaxation is reconstructed using the updated bounds, and the process repeats until the bounds stop changing or the change falls below a tolerance. In recent literature, the term appears both explicitly and implicitly: some papers name the procedure SBT, while others describe essentially the same pattern as optimization-based bound tightening (OBBT), rolling-horizon OBBT, or bounds propagation. The common structure is a repeated “tighten-rebuild-tighten” loop, typically driven by convex relaxations such as McCormick envelopes, QC relaxations, LP relaxations, or mixed-integer formulations, and used to strengthen spatial branch-and-bound, neural network verification models, and AC power flow relaxations (Panthee et al., 21 Aug 2025, Panthee et al., 12 Feb 2026, Sundar et al., 2018).

1. Terminology and conceptual scope

The literature does not use a single uniform name for SBT. In AC optimal power flow, the strengthened-QC paper presents an OBBT procedure rather than a distinct “new” SBT method, but explicitly notes that it is “very much in the spirit of SBT because bounds are updated iteratively and propagated through repeated re-solving of relaxations” (Sundar et al., 2018). In ReLU-network MILP modeling, the paper on activation bounds does not use the term SBT explicitly, yet its Algorithm 1, “Bounder,” is described as exactly a sequential, layer-by-layer bound-tightening procedure (Badilla et al., 2023). In neural-network verification, the rolling-horizon method is framed as OBBT-RH rather than SBT, but it is described as a “sequential, horizon-shifting bound-tightening scheme” that reuses improved bounds from earlier windows to tighten later ones (Zhao et al., 2024).

This variation in terminology is substantive rather than merely stylistic. Some papers reserve SBT for repeated optimization-based tightening over convexified bilinear models, especially as a presolve routine before spatial branch-and-bound (Panthee et al., 21 Aug 2025, Panthee et al., 12 Feb 2026). Others use OBBT for the same iterative logic in nonconvex continuous optimization (Sundar et al., 2018). Still others treat bound tightening as a broader family that includes LP relaxations, MILP subproblems, layer-wise propagation, and problem-specific analytic derivations (Badilla et al., 2023, Sosnin et al., 2024, Ganzfried, 24 Jun 2026).

A recurring misconception is that any addition of bounds is SBT. The multiplayer imperfect-information games paper explicitly states that its method is not sequential bound tightening in the standard optimization sense, but rather “manual/analytic bound derivation” for slack and multiplier variables, supplied to the solver as finite input bounds (Ganzfried, 24 Jun 2026). Likewise, the RLT lifted-formulation paper studies explicit propagation of original-variable bounds to auxiliary variables, but emphasizes that this propagation is “not iterative inference inside the lifted model”; it is direct explicit bound transfer from $x$ to $X$ (González-Díaz et al., 23 Sep 2025). This suggests that SBT is best understood as a family of repeated solve–contract–update schemes, not as a synonym for all presolve bounding.

2. Canonical algorithmic pattern

The canonical SBT pattern is a sequence of min/max subproblems over a current relaxation $\Omega$. In AC-OPF, the OBBT step tightens voltage magnitudes and angle differences by solving

$$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$

$$\theta_{ij}^l \leftarrow \min\{\theta_{ij}:\Omega\},\qquad \theta_{ij}^u \leftarrow \max\{\theta_{ij}:\Omega\},$$

then rebuilding $\Omega$ with the new bounds and repeating until the updated bounds are identical to the previous iteration’s bounds (Sundar et al., 2018). The same logic appears in bilinear three-phase infeasibility analysis, where SBT iteratively solves, for each filtered deviation variable,

$$\Delta x_{f,i}^{l*}:=\min(\Delta x_{f,i}), \qquad \Delta x_{f,i}^{u*}:=\max(\Delta x_{f,i}),$$

subject to a convexified relaxation, and stops when the bounds stop changing by more than tolerance $\epsilon$ (Panthee et al., 21 Aug 2025).

A key implementation refinement is variable filtering. In BL-TPIA, SBT is applied only to a filtered subset of independent variables,

$\mathbf{x}_f=\{V^r,V^i\},$

while dependent variables such as $V^{\text{sq}}, G^{\text{load}}, B^{\text{load}}$ are updated analytically from the tightened voltage-component bounds (Panthee et al., 21 Aug 2025). Grid-ECO uses the same idea for a mixed-integer bilinear EV charging station placement model: only filtered variables $X$0 are tightened directly, and bounds for dependent variables $X$1 are computed in a postprocessing step (Panthee et al., 12 Feb 2026). This reduces the number of expensive bounding subproblems.

A second refinement is variable decomposition. In both BL-TPIA and Grid-ECO, filtered variables are written as nominal plus deviation terms,

$X$2

with voltage components expressed as

$X$3

The stated purpose is to give the deviation variables small, well-defined bounds that are easier for SBT to tighten, while the nominal part captures the phase-specific baseline (Panthee et al., 21 Aug 2025, Panthee et al., 12 Feb 2026).

A third refinement is objective bounding. For AC-OPF global optimization, GO-OBBT adds an objective upper bound from a feasible AC-OPF solution (Sundar et al., 2018). In Grid-ECO, which is a maximization problem, the local NLP objective value from IPOPT is used as a lower bound inside the tightening subproblems: $X$4 The paper identifies this as an explicit modification of the standard SBT scheme, which typically uses an upper objective bound in minimization settings (Panthee et al., 12 Feb 2026).

3. Relaxations, envelopes, and bound propagation mechanisms

SBT is effective only relative to a relaxation or exact subproblem whose strength depends on variable domains. In bilinear and mixed-integer bilinear models, the dominant mechanism is the McCormick envelope. For a generic bilinear term $X$5, the TPIA paper uses

$X$6

and Grid-ECO gives the corresponding McCormick system $X$7 in MIQCP form (Panthee et al., 21 Aug 2025, Panthee et al., 12 Feb 2026). The explicit rationale is that loose bounds enlarge the feasible region explored by the tree and make the relaxation weak.

In AC power systems, SBT interacts closely with bound-dependent convex envelopes. The QC relaxation of AC-OPF uses envelopes for $X$8, $X$9, $\Omega$0, and $\Omega$1, and the SDP/QC hybrid paper states that “the QC convex envelopes are determined by variable bounds” and that “as the bounds become smaller the strength of the relaxation increases” (Coffrin et al., 2015). The strengthened-QC AC-OPF paper reinforces the same point: tighter voltage-magnitude and phase-angle bounds produce stronger convex envelopes, and stronger relaxations in turn yield better bound-tightening subproblems (Sundar et al., 2018).

In ReLU-network optimization, the analogous mechanism is the dependence of big-$\Omega$2 or partitioned formulations on pre-activation bounds. The ReLU-MILP paper formulates per-neuron OBBT as maximization or minimization of a pre-activation $\Omega$3 over the network up to layer $\Omega$4, using previously computed bounds on earlier layers (Badilla et al., 2023). The neural-controller certification paper similarly emphasizes that the validity and quality of the big-$\Omega$5 formulation depend directly on valid lower and upper pre-activation bounds $\Omega$6 and $\Omega$7, and compares IBP, CROWN, $\Omega$8-CROWN, LP-OBBT, layer-wise OBBT, and MIP-OBBT as competing bound-tightening strategies (Sosnin et al., 2024).

Not all bound propagation strengthens the feasible region in the same way. In lifted RLT formulations, the paper on bound tightening in lifted formulations proves that the linear relaxation already implies the “obvious” product bounds

$\Omega$9

Its main theorem therefore identifies the explicit auxiliary-variable bounds as redundant in the feasible region of $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$0, even though their presence can still materially alter solver behavior (González-Díaz et al., 23 Sep 2025). This is an important qualification: tighter or more explicit bounds can be mathematically redundant while still changing numerics and branching decisions.

4. Major application domains

Power systems are a principal application area. In AC-OPF, OBBT tightens voltage magnitudes and angle differences, improves QC and SDP relaxations, and can materially reduce optimality gaps (Sundar et al., 2018, Coffrin et al., 2015). In constraint screening for AC-OPF, bound-tightening logic is used line by line rather than as a full fixed-point loop: each line’s flow is minimized and maximized over a convex relaxation, and a line is screened if the thermal limit cannot be reached. The cost-aware variant adds the quadratic generation-cost inequality

$$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$1

which improves redundancy removal for all tested relaxations (Awadalla et al., 2023). In topology optimization for DC-OTS, topology-aware bound tightening keeps only a $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$2-hop neighborhood of switching variables binary in each bounding subproblem, thereby preserving local structure near the target line (Pineda et al., 22 Jul 2025).

Another major domain is global optimization of bilinear electric-grid models. For three-phase infeasibility analysis, the original nonlinear problem is reformulated exactly as a bilinear program, and SBT is used as a presolve routine before Gurobi’s spatial branch-and-bound (Panthee et al., 21 Aug 2025). Grid-ECO uses the same architecture—exact reformulation into MIBLP, SBT on a convexified relaxation, then global optimization by sBnB—for EV charging station placement with binary siting and integer charger-count variables (Panthee et al., 12 Feb 2026). In both cases, SBT is not the global solver itself; it is the preprocessing stage that makes the subsequent global search tractable.

Neural networks provide a second large application cluster. In ReLU MILP modeling, sequential layer-by-layer tightening computes lower and upper pre-activation bounds needed by the big-$$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$3 encoding, with “naive,” “weak,” and “strong” versions corresponding to interval propagation, LP relaxations, and MILP solves (Badilla et al., 2023). In rolling-horizon verification, OBBT-RH applies optimization-based tightening on smaller subgraphs of the original network in a rolling-horizon sequence $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$4, reusing improved bounds from earlier windows and parallelizing neuron-wise sub-MIPs (Zhao et al., 2024). In neural-network controller certification, bounds must also be propagated through the unrolled closed-loop dynamics $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$5, so SBT-like procedures become sequential in time as well as in depth (Sosnin et al., 2024).

Additional domains broaden the methodological scope. In the pooling problem, the paper on rank-one-based relaxations uses OBBT for generalized pooling instances and a cheap, structure-exploiting bound-tightening procedure for mining-to-pooling reformulations; the tightened arc and node capacity bounds strengthen LP, MIP, and SOCP relaxations of the rank-one substructure (Jalilian et al., 2023). In multiplayer imperfect-information games, analytic finite bounds on slack and multiplier variables strengthen the McCormick relaxations inside spatial branch-and-bound, even though the paper explicitly distinguishes this from standard SBT (Ganzfried, 24 Jun 2026).

5. Computational tradeoffs and empirical behavior

A defining feature of SBT is the tradeoff between bound quality and bound-computation cost. The ReLU-network paper states that solving the per-neuron MILP OBBT problem exactly is NP-hard, and concludes that “computing the bounds with LP relaxations provides a good compromise of quality in favor of computability” (Badilla et al., 2023). Its reported observations are that weak and strong bounds are “remarkably similar,” that the gap tends to grow with depth, and that a hybrid strategy—naive in the first layers, weak in the middle layers, strong in deeper layers—is a promising practical rule of thumb (Badilla et al., 2023).

Rolling-horizon and layer-wise decompositions are one response to this tradeoff. In neural-network verification on the 90 instances of the mnist_fc benchmark from VNN-COMP, OBBT-RH+MIP verified 80 instances with 5 timeouts and total time 154.85 s, whereas OBBT+MIP verified 79 instances with 6 timeouts and total time 202.27 s; OBBT-RH also produced the tightest average bounds range, 12.38 (Zhao et al., 2024). In neural-network controller certification over $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$6 steps, the deep $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$7 network showed a particularly strong effect: MIP-OBBT reduced the relaxation gap to 40.20% and solve time to 5.72 s, whereas IBP yielded 2,217,423.69% and 996.27 s; LP-OBBT with stronger partitioned formulations gave intermediate tradeoffs (Sosnin et al., 2024).

In power and energy systems, the strongest reported gains arise when SBT is combined with exact bilinear reformulation and spatial branch-and-bound. For BL-TPIA, S-BLP reduced solution time by an average of 69.86%, reduced explored nodes by an average of 72.24%, and achieved a maximum runtime reduction of up to 97.23%; it also solved a 5194-bus, 47,568-variable, 26,816-constraint case while exploring only one sBnB node (Panthee et al., 21 Aug 2025). In Grid-ECO, the SBT-enhanced formulation reduced runtime by up to 73% and node exploration by up to 97%, while achieving a 0% optimality gap and guaranteed AC feasibility; in the downtown feeder case, the baseline failed to find any feasible solution within about 167 hours, while the SBT-enhanced formulation found a solution in about 57 minutes (Panthee et al., 12 Feb 2026).

The same tradeoff appears in topology optimization. In fully switchable DC-OTS, relaxing all switching variables in the bounding subproblems is too weak, but keeping too many binaries is too expensive. The paper identifies a “sweet spot” around $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$8: TBT-2 reduced average total time by about 45% versus the baseline MIP and cut the number of unsolved instances roughly in half, whereas higher $$v_i^l \leftarrow \min\{v_i:\Omega\},\qquad v_i^u \leftarrow \max\{v_i:\Omega\},$$9 gave stronger bounds but larger tightening overhead (Pineda et al., 22 Jul 2025). This result is structurally similar to the neural-network finding that partial or localized exactness can outperform both crude propagation and fully global tightening.

6. Limitations, variants, and recurrent misconceptions

SBT is often presented as a safe strengthening step, but its guarantees and effects depend on context. In AC-OPF relaxations, safety follows from the fact that a convex relaxation contains the original feasible set, so if a line thermal limit cannot be reached in the relaxation it can be screened safely in the original problem (Awadalla et al., 2023). In BL-TPIA, the paper states that SBT is safe relative to the incumbent NLP solution because it uses a known feasible objective $$\theta_{ij}^l \leftarrow \min\{\theta_{ij}:\Omega\},\qquad \theta_{ij}^u \leftarrow \max\{\theta_{ij}:\Omega\},$$0 as an upper bound, so the tightening does not remove feasible points that could improve the incumbent (Panthee et al., 21 Aug 2025). In Grid-ECO, the corresponding safeguard is the lower objective bound from the relaxed NLP solve in a maximization setting (Panthee et al., 12 Feb 2026).

Another misconception is that tighter explicit bounds must always help. The lifted-formulation study shows that auxiliary-variable bounds may already be implied by the RLT relaxation, so explicit propagation does not systematically shrink the branch-and-bound tree; the observed impact is “solver-dependent and highly instance-dependent,” and can help or hurt depending on the LP solver, whether OBBT/FBBT are enabled, and whether the instances are continuous or mixed-integer (González-Díaz et al., 23 Sep 2025). Similarly, the Nash-equilibrium paper reports that bounding only multipliers can be slower than bounding only slacks, even when the multiplier bounds are finite and valid, because the slack variables appear directly in the bilinear constraints and therefore have a larger effect on McCormick relaxations (Ganzfried, 24 Jun 2026).

A further limitation is that not every use of bound tightening is a full sequential fixed-point procedure. Cost-aware AC-OPF screening solves line-specific optimization problems and terminates after screening, rather than iterating all variable bounds to convergence (Awadalla et al., 2023). The pooling paper uses OBBT for generic instances but a manual stagewise update rule for mining instances; it explicitly does not introduce a classical sequential bound tightening algorithm (Jalilian et al., 2023). The multiplayer-games paper is even more explicit that its contribution is presolve-style analytic tightening, not generic SBT (Ganzfried, 24 Jun 2026).

Taken together, these results support a precise characterization. SBT is a repeated optimization-driven contraction of variable domains, usually over a convexified or mixed-integer relaxation, often combined with filtering, decomposition, or horizon restriction to keep the subproblems tractable. Its value lies in the two-way dependence emphasized in AC-OPF: tighter bounds strengthen the relaxation, and stronger relaxations yield better bound-tightening subproblems (Sundar et al., 2018). Across power systems, neural networks, and related nonconvex optimization models, that feedback loop is the defining mechanism of the method.