Regular Facial Reduction Sequences

• Regular facial reduction sequences are iterative procedures that expose the minimal face of a convex cone by systematically removing redundancies in conic optimization problems.
• They restore strong duality and enhance algorithmic performance by replacing the original cone with its minimal face, ensuring numerical stability even without Slater’s condition.
• These sequences underpin extended dual formulations and are applied in semidefinite programming, combinatorial optimization, and polynomial systems to improve computational tractability.

Regular facial reduction sequences are systematic, iterative procedures used to regularize conic optimization problems—especially semidefinite or conic linear programs—by revealing and removing hidden redundancies in the feasible set, ultimately exposing the minimal face of the underlying convex cone necessary to contain all feasible points or slacks. These sequences play a central role in restoring strong duality, enabling algorithmic preprocessing, improving numerical stability, and interpreting dual pathologies in the absence of standard constraint qualifications, such as Slater’s condition. The theory unites geometric and analytic perspectives, linking face chains in the cone, tangent spaces, certificates of reduction, and dual solution attainability. Regular facial reduction sequences underpin extended duality theorems and are now widely applied in semidefinite programming (SDP), combinatorial optimization relaxations, polynomial systems, and beyond.

1. Facial Reduction Algorithm: Structure and Principles

Facial reduction algorithms (FRAs) start with the full feasible cone (e.g., the positive semidefinite cone in SDPs) and iteratively intersect it with successively chosen hyperplanes to peel away inactive or redundant directions, generating a sequence of nested faces: $$K = F_0 \supseteq F_1 \supseteq \ldots \supseteq F_t = F_\text{min}.$$ At each step $i$, a “reducing certificate” $y_i$ is found, lying in the intersection of the orthogonal (annihilator) of the constraint data and the dual face $F_{i-1}^*$. The new face is updated via: $F_i = F_{i-1} \cap y_i^\perp.$ This process iterates until the smallest face $F_\text{min}$ covering all feasible slacks is reached. For conic linear programs of the form $\max\{\langle c, x\rangle \mid Ax \leq_K b\}$, this regularization is equivalent to replacing the cone constraint $b - Ax \in K$ by $b - Ax \in F_\text{min} \subseteq K$, such that Slater's condition is restored.

The number of iterations is tightly controlled: there exists an integer $\ell$ (at most the minimum of the dimension of the data subspace and the length of the maximal face chain of the cone minus one) such that after $\ell$ reducing steps, the sequence reveals $F_\text{min}$ (Pataki, 2013). In the case of semidefinite cones $\mathbb{S}_+^n$, this is bounded by $n$.

2. Extended Duals and Ramana-type Constructions

Once $F_\text{min}$ is determined, the dual problem is robustly reformulated by substituting the original dual cone $K^*$ with $F_\text{min}^*$, ensuring strong duality even when constraint qualifications fail. The construction of extended duals leverages the facial reduction certificates $y_i$, which can be decomposed as $y_i = u_i + v_i$ with $u_i \in K^*$ and $v_i \in \text{tan}(u_0 + \ldots + u_{i-1}, K^*)$, subject to additional constraints. The resulting dual cone,

$$F_\text{min}^* = \{ u_{\ell+1} + v_{\ell+1} : (u_i, v_i)_{i=0}^{\ell+1} \text{ feasible in EXT} \},$$

is described via an explicit conic linear system in an extended (lifted) variable space. When the underlying cone $K$ is “nice” (e.g., the semidefinite cone or polyhedral cones), the tangent space and closure properties ensure this construction is conic representable.

In semidefinite programming, the tangent constraints can be further encoded as explicit semidefiniteness conditions on auxiliary variables (e.g., via block semidefinite matrices incorporating tangent components), recovering Ramana's dual (Pataki, 2013).

3. Applications, Examples, and Generalizations

The regular facial reduction process is illustrated through both polyhedral and semidefinite examples. In linear programming, the minimal cone often reduces to a product of an active subspace and zeros, identifying inactive inequalities (Pataki, 2013). For SDPs with block structure, the reduction process can physically identify slacks or blocks forced to be zero and reveals a lower-dimensional face supporting all feasible solutions.

Applications span:

• Polynomial system solving, where facial reduction regularizes the SDP arising from moment matrices, often reducing a large $k\times k$ matrix to smaller principal submatrices, simplifying the extraction of generators for real radical ideals (Reid et al., 2015, Wang et al., 2016).
• Matrix completion and nuclear norm minimization, where hidden degeneracy in the solution set is exposed. Facial reduction reduces the rank and the problem size, even when strict feasibility nominally holds in the original problem (Huang et al., 2016).
• Control systems, where lack of strong feasibility in an LMI can be traced to system-theoretic pathologies such as invariant zeros. Facial reduction locates and removes these via reduction to smaller subsystems (Waki et al., 2016).
• Combinatorial optimization SDPs, where affine facial reduction automatically generates exposing vectors from polyhedral problem structure, substantially reducing the size and improving regularity (Hu et al., 19 Feb 2024).

4. Theoretical Properties: Correctness, Optimality, and Bounds

The correctness of regular facial reduction sequences is established via convex geometric arguments:

• Any face containing $F_\text{min}$ can always be reduced by intersecting with the orthogonal of a reducing certificate.
• The process is guaranteed to terminate after at most $\ell$ steps, with $$\ell \leq \min\{\text{face length} - 1, \dim \text{ data subspace}\}$$ (Pataki, 2013).
• There are constructive algorithms to maximize the “cutting power” of each certificate (e.g., maximizing the rank in SDP reductions) (Permenter et al., 2014).
• In the presence of product cones with polyhedral components, polyhedral-aware variants like FRA-Poly drastically reduce the number of required steps relative to classical FRA, with cases such as the doubly nonnegative cone requiring only linear (in $n$) steps (Lourenço et al., 2015).

The representation theorems generalize the dual cones of minimal faces as projections of higher-dimensional conic systems parameterized by the full certificate decomposition, establishing a fundamental link between regularity of the sequence and dual attainability.

5. Algorithmic and Computational Aspects

Regular facial reduction sequences have a direct impact on algorithmic tractability:

• Partial facial reduction and affine FR algorithms automate preprocessing for SDPs, reducing variable dimensions, restoring strict feasibility, and preserving sparsity when structural matrix approximations (e.g., diagonally dominant or block-diagonal cones) are judiciously chosen (Permenter et al., 2014, Hu et al., 19 Feb 2024).
• Post-processing routines for dual solution lifting are provided to reconstruct dual solutions to the original problem from solutions on reduced faces, with detailed conditions for correctness (Permenter et al., 2014).
• Combinations of facial reduction with other preprocessing methods, such as chordal decomposition, offer compounded computational benefits in large-scale SDP applications (Kungurtsev et al., 2018).

Software implementations exist for several of these techniques, allowing users to preprocess SDPs (especially in SeDuMi format) automatically and to control degree and quality of reduction (Permenter et al., 2014).

6. Role in Duality, Singularity Degree, and Degeneracy Analysis

Facial reduction sequences precisely quantify and resolve forms of nonregularity in conic programs:

• The singularity degree of a problem is defined as the minimal number of facial reduction steps required to attain strict feasibility. High singularity degree is intimately linked with numerical ill-conditioning and dual nonattainment (Hu et al., 2019, Im et al., 8 Jul 2024).
• Symmetry reduction and facial reduction, when combined, can block-diagonalize and regularize very large-scale SDPs (e.g., QAPs) without increasing the singularity degree (Hu et al., 2019).
• In projection problems onto spectrahedra (e.g., nearest correlation matrix, combinatorial relaxations), ill-conditioning in semi-smooth Newton methods is directly explained by the problem's singularity degree and associated degeneracy, which facial reduction systematically eliminates (Im et al., 8 Jul 2024).
• Regular facial reduction sequences restore strong duality and numerical stability, even for instances suffering from slack inactivity, highly redundant inequalities, or boundary feasibility.

7. Broader Impact and Future Directions

The conceptual framework, algorithmic techniques, and extended duality furnished by regular facial reduction sequences have become foundational in contemporary conic and semidefinite optimization. Their scope includes:

• Regularization of pathological SDP relaxations from combinatorial optimization, polynomial equations, and systems control, often revealing structural properties such as polyhedrality and block-diagonality.
• Unification of theory and practice through explicit, implementable algorithms, extended duality, and rigorous guarantees on iteration counts and correctness.
• Inspiration for newer developments in extended formulations, preconditioned optimization, and robust symbolic/numeric hybrid methods for real algebraic and optimization problems.

Open areas include development of more predictive criteria for necessity and depth of facial reduction, automated recognition of problem structure to trigger affine or partial FR, further integration with symmetry and sparsity-exploiting methods, and analysis of the interplay between singularity degree and algorithmic performance in large-scale, real-world optimization.

---

Tables, explicit reduction formulas, and full algorithmic frameworks can be found in (Pataki, 2013, Permenter et al., 2014, Lourenço et al., 2015, Hu et al., 2019, Hu et al., 19 Feb 2024), and (Im et al., 8 Jul 2024). These resources formalize and exemplify the practical deployment and theoretical scope of regular facial reduction sequences across modern optimization settings.