Dual Active-Set QP Solver

• Dual active-set QP solver is an iterative optimization method that targets the dual formulation of quadratic programming by dynamically adjusting a working set of constraints.
• It employs efficient numerical linear algebra techniques, such as Cholesky and LDLᵀ factorizations, to update the active set, ensuring fast convergence and numerical stability.
• This method is widely applied in real-time model predictive control, adversarial optimization, and polytope projection, demonstrating robust performance in varied practical scenarios.

A dual active-set quadratic programming (QP) solver is an iterative optimization technique designed to solve QP problems by operating directly on the dual formulation of the problem. These solvers are characterized by their strategy of modifying a working set of dual variables (corresponding to the active inequality constraints) and exploiting the structure of the dual QP to enable efficient updates using linear algebraic factorizations. This approach has been generalized to broad problem classes, including minimax QPs, general convex programs with mixed equality and inequality constraints, and practical scenarios such as model predictive control (MPC), adversarial optimization, and polytope projection (Ren et al., 9 Nov 2025, Schmidtobreick et al., 17 Nov 2025, Fält et al., 2019, Arnström et al., 2021, Forsgren et al., 2015).

1. Dual Formulation of QP and Minimax QP Problems

Quadratic programming problems can be written in standard form as

$$\min_{x \in \mathbb{R}^{n}} \; \frac{1}{2} x^T Q x + c^T x, \quad \text{s.t.}\;\; A x \le b,$$

where $Q \succ 0$ is symmetric positive definite, $A$ has full row rank, and the feasible set is nonempty. The Lagrangian dual introduces nonnegative multipliers $\lambda \geq 0$, leading to a dual problem of the form

$$\min_{\lambda \ge 0} \; \frac{1}{2}\lambda^T M \lambda + d^T \lambda,$$

where $M = A Q^{-1} A^T,\, d = b + A Q^{-1}c$, and the primal optimum is recovered as $x^* = -Q^{-1}(c + A^T \lambda^*)$ (Schmidtobreick et al., 17 Nov 2025, Arnström et al., 2021).

For minimax QPs with bilinear structure and coupled inequality constraints, the primal takes the form

$$\min_{x} \max_{y} \frac{1}{2} x^T G_{11}x + x^T G_{12}y + \frac{1}{2} y^T G_{22}y + c_x^T x + c_y^T y, \quad A x + B y + h \le 0,$$

where $G_{11}$ is usually positive definite, $G_{22}$ strictly negative definite, and $G_{12}$ arbitrary. The dual active-set framework is extended to this class by deriving the Lagrangian dual in the dual variables $\mu \le 0$, leading again to a convex QP structure in the dual (Ren et al., 9 Nov 2025).

2. Active-Set Strategy and Algorithmic Workflow

The essence of dual active-set methods is the iterative adjustment of the working set—i.e., the subset of constraints hypothesized to be active (binding) at the solution. At each iteration:

• The current dual working set $\mathcal{W}$ (or $\alpha$ in minimax) specifies the constraints treated as equalities.
• Solve the equality-constrained dual QP subproblem restricted to $\mathcal{W}$ via system solve.
• Test for optimality using KKT conditions: if dual feasibility ($\lambda^* \geq 0$) and primal feasibility hold for all constraints, terminate.
• Otherwise, add a violated constraint or drop an over-enforced one by a one-at-a-time update, using a step direction and length determined by the solution structure.
• Repeat until all KKT conditions are satisfied or infeasibility is detected (Ren et al., 9 Nov 2025, Arnström et al., 2021, Fält et al., 2019, Forsgren et al., 2015).

For each working set change, the underlying subproblem reduces to equality-constrained QP, with search direction and blocking index identified via specific linear algebraic updates. The process guarantees finite termination if each working set configuration appears at most once, with strict decrease in the dual or primal objective function at every iteration.

3. Numerical Linear Algebra: Factorizations and Updates

The central computational routine is the maintenance and update of matrix factorizations corresponding to the reduced dual Hessian. Predominant strategies:

• Cholesky Factorization: For minimax QPs, maintain and update a Cholesky factor $R$ for $-N_\alpha^T G^{-1} N_\alpha$; updates occur via rank-one growth (addition of a constraint) or Givens rotations (removal) (Ren et al., 9 Nov 2025).
• LDL$^{\top}$ Factorization: For general convex QPs, the system $M_{\mathcal{W},\mathcal{W}}$ is maintained via recursive LDL$^{\top}$ updates, allowing additions or removals of active constraints in $\mathcal{O}(|\mathcal{W}|^2)$ per operation (Arnström et al., 2021, Fält et al., 2019).
• Iterative Refinement: For degenerate (semi-definite) subproblems, solve regularized systems $(\bar{G}+\epsilon I)x = c$ using the current Cholesky factor. Convergence is monitored, and zero-curvature directions are used to escape unbounded subspaces (Fält et al., 2019).

All dual active-set solvers avoid explicit computation of full inverses or null-space factorizations, operating instead with efficient and numerically stable low-rank updates.

4. Convergence, Complexity, and Theoretical Guarantees

Termination is secured by the strict descent property and non-repetition of working sets (S-pairs), which follows directly from the updating rules and strict objective decrease for feasible subproblems (Ren et al., 9 Nov 2025, Arnström et al., 2021). Under generic assumptions:

• Each working set appears at most once; the number of possible sets is finite.
• Each iteration costs $\mathcal{O}(n^2)$ in the number of variables and constraints, dominated by rank-one matrix updates and triangular solves (Fält et al., 2019, Arnström et al., 2021).
• Overall complexity for $m$ constraints and $n$ variables is $\mathcal{O}(m n^2)$ under nondegeneracy, and worst-case cubic scaling if $m \sim n$.
• Infeasibility can be detected during a failed feasibility restoration step (e.g., step length infeasibility).

Specific finite-step termination theorems and exact complexity results are provided, with detailed proofs in (Ren et al., 9 Nov 2025, Arnström et al., 2021).

5. Extensions: Warm-Starting and Learning-Accelerated Dual Solvers

Warm-starting is a central feature for real-time control applications and parametric QPs. If problem data change only modestly, previous solutions—specifically, the working set, Cholesky/LDL factors, and dual/primal iterates—are recycled to achieve substantial reductions in iteration count and total computational cost (Arnström et al., 2021, Fält et al., 2019). Implementation recommendations include preserving numerical tolerances and memory-efficient structure of factorizations.

Recent advances integrate learning to optimize, notably via graph neural networks (GNNs), for working set prediction. By encoding the QP as a bipartite graph, message-passing GNNs predict probable active constraints, thus reducing cold-start iterations by ∼50% at large problem scales, with speedups of 3–5x reported in large-scale and real-time MPC settings (Schmidtobreick et al., 17 Nov 2025). Generalization to unseen problem sizes and more complex constraint structures is demonstrated.

Warm-Start Mechanism	Empirical Iteration Reduction	Notes
Smart initial set	5–10×	Often based on sign of dual gradient
GNN warm-start	1.6–2× (hard MPC instances)	Highly effective for large QPs

6. Practical Implementation and Numerical Performance

Implementation principles for dual active-set solvers emphasize:

• Initialization at an unconstrained minimizer, with careful precomputation of constraint eligibility ($\delta_i < 0$).
• Cholesky or LDL factorizations updated in-place; usage of Givens rotations for numerical robustness.
• Explicit tolerance checks for primal/dual feasibility to avoid ill-conditioning.
• Monitoring of strict objective decrease and iteration count; maximum-iteration safeguards to rule out cycling.
• Data structures and memory layouts arranged for efficient small matrix updates, with sparse adaptions as required.
• For mixed-constraint QPs, proximal-point regularization incorporates robustness to ill-conditioning (Fält et al., 2019, Arnström et al., 2021).

Benchmarks show high accuracy (residuals < $10^{-9}$), scalability to thousands of variables and constraints, and computational performance scaling between $\mathcal{O}(n^3)$ and $\mathcal{O}(n^4)$ in the most taxing minimax cases (Ren et al., 9 Nov 2025). In practical optimization, DAQP and QPDAS demonstrate robust performance for large MPC and geometric projection workloads (Fält et al., 2019, Arnström et al., 2021).

7. Representative Applications and Generalizations

Dual active-set QP solvers have been effectively applied in:

• Adversarial Mean–Covariance Optimization: Minimax QP solvers outperform baselines in stress-testing portfolio models, with strict reductions in optimal return of the target investor (Ren et al., 9 Nov 2025).
• Real-Time Model Predictive Control: DAQP and learning-accelerated variants are embedded for fast constraint handling in MPC problems (Schmidtobreick et al., 17 Nov 2025, Arnström et al., 2021).
• Polytope Projections and Feasibility Restoration: QPDAS delivers fast projections in high-dimensional geometry, demonstrating superior scaling (Fält et al., 2019).
• Mixed-Constraint and Shifted Penalty QPs: Dual active-set can be coupled with primal methods using shift-penalty approaches, supporting rapid Phase-I feasibility recovery (Forsgren et al., 2015).

The framework applies without modification to standard convex QPs by simply treating the entire Hessian as positive definite and employing the same active-set and update machinery (Ren et al., 9 Nov 2025).

---

These features make dual active-set quadratic programming solvers a core computational tool in mathematical optimization, numerical linear algebra, and modern control, with ongoing methodological advances in algorithmic efficiency, theoretical guarantees, and hybrid learning-optimization systems (Ren et al., 9 Nov 2025, Schmidtobreick et al., 17 Nov 2025, Fält et al., 2019, Arnström et al., 2021, Forsgren et al., 2015).