Conic Programming: Theory and Applications

• Conic programming is a convex optimization framework focused on minimizing functions over affine constraints intersecting closed convex cones, generalizing LP, QP, and SDP.
• It employs specialized algorithms like first-order, augmented Lagrangian, and ADMM methods to achieve provable convergence and computational efficiency for large-scale problems.
• Its applications span control, signal processing, machine learning, quantum information, and robust optimization, underscoring its versatility in addressing complex challenges.

Conic programming is a branch of convex optimization concerned with the minimization of (typically linear or convex composite) objective functions over the intersection of an affine space with a closed convex cone. This paradigm strictly generalizes linear programming (LP), quadratic programming (QP), second-order cone programming (SOCP), and semidefinite programming (SDP), while also admits rich extensions involving nonsmooth and nonlinear settings. The theoretical and algorithmic developments in conic programming underpin a wide range of methodologies in control, signal processing, robust optimization, machine learning, quantum information, and beyond.

1. Mathematical Foundations and Problem Formulation

A canonical conic convex program is expressed as: $$\min_{x \in \chi} \; p(x) = \rho(x) + \gamma(x) \quad \text{s.t.} \quad Ax - b \in \mathcal{K}$$ where

• $\rho(x)$: closed, convex, and possibly nonsmooth (e.g., indicator of a constraint or a norm),
• $\gamma(x)$: convex with a Lipschitz continuous gradient,
• $$A \in \mathbb{R}^{m \times n},\; b \in \mathbb{R}^m$$,
• $\mathcal{K}$: a closed convex cone, often non-polyhedral (e.g., positive semidefinite matrices),
• $\chi$: a compact, "simple" convex set for which Euclidean projections (and proximal steps) are efficient.

The conic constraint $Ax - b \in \mathcal{K}$ generalizes inequality and matrix inequality constraints, and the splitting between $\rho$ and $\gamma$ allows composite optimization strategies leveraging both smoothness and nonsmoothness.

Variants include:

• Linear Conic Programming: $\rho$ and $\gamma$ are linear.
• Quadratic/Conic Quadratic Programming: $\gamma$ is quadratic, $\rho$ possibly an indicator.
• Semidefinite Programming: $\mathcal{K}$ is the cone of symmetric PSD matrices.
• Conic Geometric Programming: involves exponential/affine constraints paired with a cone (Chandrasekaran et al., 2013).
• Nonconvex Conic Programs: geometric reformulations and convexifications; see (Kim et al., 2019).

2. Algorithmic Methodologies

2.1 First-Order and Augmented Lagrangian Methods

First-order methods gain computational efficiency for large-scale conic problems. Notably, the ALCC algorithm (Aybat et al., 2013) leverages an inexact augmented Lagrangian approach:

• Augmented Lagrangian:

$$\mathcal{L}_\mu(x, y) = p(x) + \frac{\mu}{2} d_\mathcal{K} (Ax - b - y/\mu)^2 - \frac{1}{2\mu} \|y\|^2$$

where $d_\mathcal{K}$ denotes distance to $\mathcal{K}$.

• Iterative steps alternate between approximately minimizing $P_k(x, y_k)$ over $x$, via accelerated proximal gradient (APG) methods, and explicit dual updates:

$$y_{k+1} = \mu_k \left[ \Pi_\mathcal{K}(A x_k - b - y_k/\mu_k) - (A x_k - b - y_k/\mu_k) \right]$$

with growing penalty parameter $\mu_k$ and decreasing subproblem tolerance.

Convergence guarantees:

• Any accumulation point of the primal iterates solves the conic program.
• Dual sequence converges uniquely to a KKT multiplier.
• $\mathcal{O}(\log(1/\varepsilon))$ outer iterations for $\varepsilon$-feasibility and optimality; total APG calls bounded accordingly.

2.2 Multi-Block and Proximal Splitting ADMM

For conic programs with multiple constraint types (linear equalities/inequalities, nonpolyhedral/polyhedral cones), specialized multi-block ADMM variants are necessary for convergence:

• sPADMM3c (Sun et al., 2014): cyclic block order $1 \to 3 \to 2 \to 3$ exploits the linearity of one block for efficient closed-form updates, ensuring convergence and empirical $20\%$ speedup (in DNN-SDPs) over naïve extensions.
• Schur complement-based ADMMs (Li et al., 2014) for quadratic conic programs incorporate blockwise Schur complement corrections into the proximal terms, decoupling blocks more effectively and yielding improved numerical performance in QSDPs.

2.3 Conic Descent and First-Order Dual-Driven Methods

Geometric reinterpretation of conic programming as searching over rays enables the design of conic descent (CD) and momentum-conic descent (MOCO) methods (Li et al., 2023):

• CD decomposes the iterate update into a 1D ray minimization and a direction search constrained to the cone.
• MOCO incorporates heavy-ball momentum (weighted history of gradients), leading to improved direction selection and convergence.
• Duality analysis provides explicit stopping criteria ($$\langle g_k, v_k \rangle \geq -O(\sqrt{\varepsilon})$$) and allows for preconditioning to enhance dual convergence.

Memory-efficient MOCO variants for SDPs maintain low-rank solutions via random sketching and incremental Burer-Monteiro-type factorizations.

2.4 Adaptive Cone Approximation and Inexact Projections

When $\mathcal{K}$ is intractable (e.g., the copositive cone), algorithms may employ an adaptive sequence of tractable outer polyhedral approximations $\mathcal{K}^k$ (Fukuda et al., 2 Jun 2024). Iteratively refining $\mathcal{K}^k$ balances per-iteration cost against overall solution accuracy. The algorithm ensures primal-dual iterates satisfy a strong sequential optimality condition (relaxed AGP), with convergence to KKT points under Robinson’s condition.

3. Duality, Sensitivity, and Strong Duality

Conic programming’s duality theory extends LP duality to arbitrary cones:

• The dual problem involves the adjoint map $A^*$ and the dual cone $\mathcal{K}^*$.
• Strong duality holds under Slater-type qualifications, closedness, or boundedness CQs (Ajayi et al., 2020). Three levels of CQ are related: closedness $\implies$ Slater $\implies$ boundedness.
• The minimax theorem and conic duality are tightly linked—even in infinite-dimensional settings, duality theories yield minimax equalities (Dimou, 2023).

Sensitivity analysis and differentiability of the optimal value function under right-hand side or objective perturbations are established via variational analysis (Luan et al., 2020): $$\phi'(b; d) = \max_{y \in S(D)} y^T d,\quad \psi'(c; h) = \inf_{x \in S(P)} x^T h$$ where $S(D)$ and $S(P)$ denote the optimal solution sets of the dual and primal, respectively.

4. Hierarchical and Specialized Problem Classes

Conic programming’s flexibility allows for embedding various program classes:

• Conic Geometric Programming (CGP) (Chandrasekaran et al., 2013): Unifies geometric and conic programming, supporting constraints of the form $\exp\{Q^T x - u\} + R^T x - v \leq 0$, $x \in \mathcal{K}$. Duals involve negative joint relative entropy.
• Higher-Order Cone Programming (Ding et al., 2018): Embedding sequences of “base cones” within higher-dimensional symmetric matrix cones leads to a hierarchy interpolating between LP, SOCP, and SDP, and—via sums of squares interpretation—between DSOS, SDSOS, and SOS constraints in polynomial optimization.
• Nonconvex Conic Reformulation (Kim et al., 2019): Nonconvex conic constraint sets are convexified (e.g., via replacing $K \cap J \cap L$ with $J \cap L$ where $J$ is a face of $\mathrm{conv}(K)$), allowing equivalence to convex conic programs under moderate assumptions.

Applications include min-max games, matrix games (simplex constraints), robust optimization (uncertainty sets as cones), entropic and kernel learning, quantum channel capacities, and more.

5. Mixed-Integer Conic Programming and Cutting Planes

• Mixed-Integer Conic Programming (MICP): Conic formulations for mixed-integer, nonlinear, and disjunctive problems are constructed using conic hull and hull (perspective) reformulations (Neira et al., 2021). They maintain the original conic structure and yield strong relaxations.
• Conic Mixed-Integer Approaches: Assortment problems under discrete choices are efficiently reformulated as conic quadratic mixed-integer programs, avoiding big-M weaknesses and exploiting McCormick inequalities for tighter relaxations (Sen et al., 2017).
• Cutting Plane Theory: For mixed-integer conic programming, minimal valid inequalities for disjunctive sets over $\mathbb{R}_+^n$ are characterized by sublinear (cut-generating) functions; the convex hull is exactly described by such inequalities together with nonnegativity (Lodi et al., 2019).

6. Certificates, Pathologies, and Robustness

Infeasibility detection, weak infeasibility, and facial reduction are central in the practical deployment of conic programming:

• Projective Geometry and Homogenization: Homogenization lifts feasibility problems into projective space, eliminating the phenomenon of weak infeasibility and enabling the design of stable, rational infeasibility certificates, even for SDPs (Naldi et al., 2018). Iterative facial reduction provides polynomial-time verifiable certificates and links to computational complexity (NP $\cap$ co-NP).
• Strong Calmness and Error Bounds: Stability/sensitivity of the KKT solution mapping is established via equivalence between strong calmness and local error bounds, extending to nonpolyhedral cones under suitable noncriticality and relative interior conditions (Liu et al., 2018).
• Robust Algorithmic Frameworks: Adaptive and memory-efficient first-order solvers such as PDCS (Lin et al., 1 May 2025) use matrix-free primal-dual hybrid gradients (PDHG), bijection-based projection algorithms for rescaled cones, and exploit modern GPU architectures for scalability and efficiency, with strong performance in large-scale, lower-accuracy regimes.

7. Applications Across Domains

Conic programming underpins advances in:

• Non-Smooth Mechanics: Recasting variational inequalities and nonsmooth mechanics (plasticity, nonlinear membranes, viscoplastic flow, minimal crack surfaces) as conic programs allows robust solution of large discretizations, leveraging epigraph and Schur complement SDP reformulations (Bleyer, 2022).
• Quantum Information: Quantification of quantum resources (entanglement, measurement incompatibility, etc.) via conic programs connects robustness measures to operational advantages in quantum discrimination tasks (Uola et al., 2018), with dual interpretations directly translating into experimentally measurable witness operators.
• Quantum-Classical Hybrid Optimization: The Quantum Conic Programming (QCP) framework utilizes parameterized linear combination-of-unitaries ansätze, shifting classical parameter optimization onto small-dimensional generalized eigenvalue problems that efficiently encode hard combinatorial constraints (Binkowski et al., 1 Nov 2024). This avoids NP-hard and barren plateau issues in variational quantum algorithms, provides robust optimality guarantees even under noisy state preparation, and is amenable to quantum hardware via LCU protocols without the need for problem-specific Hamiltonian implementation.

---

Conic programming thereby provides a unified, robust, and scalable mathematical and computational framework for convex and certain nonconvex optimization problems across mathematics, engineering, data science, and quantum information. Its generality, combined with the ongoing development of specialized algorithms and solvers, continues to expand its reach and effectiveness across diverse application domains.