Barvinok–Pataki Bound in SDP Optimization

• Barvinok–Pataki bound is a foundational result in semidefinite programming that quantifies the rank constraint via the inequality r(r+1)/2 ≥ m.
• It underpins practical approaches in combinatorial and manifold optimization, such as the Burer–Monteiro method and Stiefel manifold programs.
• Extensions of the bound reduce computational complexity but also reveal limitations, necessitating genericity conditions to avoid spurious local minima.

The Barvinok–Pataki bound is a central result in the theory of semidefinite programming (SDP), providing a quantitative description of the rank structure of solutions to SDPs with linear constraints. It establishes a precise threshold for the existence of low-rank extreme points in spectrahedra and underpins the analysis of rank-constrained nonconvex relaxations such as the Burer–Monteiro approach. The bound, its extensions, and failure modes are of critical importance for the tractability and exactness of SDP relaxations in combinatorial optimization, manifold optimization, and nonconvex optimization.

1. Formulation and Core Theorem

Let $X\in\mathbb{S}^n_+$ be the matrix variable in an SDP subject to $m$ independent affine constraints: $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$ where each $A_i \in \mathbb{S}^n$ and $\mathbb{S}^n_+$ denotes the set of $n\times n$ real symmetric positive semidefinite matrices.

The Barvinok–Pataki bound asserts that for any extreme-point solution $X^*$ of the feasible set (spectrahedron), the rank $r := \operatorname{rank}(X^*)$ satisfies: $r(r+1)/2 \geq m.$ Equivalently, for any $r$ such that $m$0, the SDP admits an optimal solution of rank at most $m$1 (Song et al., 2023, O'Carroll et al., 2022, Cifuentes et al., 2019). The proof exploits Carathéodory-type arguments about faces of the positive semidefinite cone intersected by linear spaces.

2. Applications to Stiefel Manifold Programs

When applied to linear programs over the Stiefel manifold (LPS), which are problems of the form

$m$2

the rank bound furnishes an exactness criterion for the associated SDP relaxation. The classical condition, derived via the Barvinok–Pataki theorem, is: $m$3 where $m$4 is the number of orthonormal columns, $m$5 the ambient dimension, and $m$6 the number of additional linear constraints.

The improved result shown in "Linear programming on the Stiefel manifold" relaxes this to the linear condition $m$7, which strictly enlarges the domain of exactness for the SDP relaxation, subsuming the unconstrained case $m$8 and $m$9 (Song et al., 2023).

3. Specialization to Combinatorial SDPs: Max-Cut Example

For the Goemans–Williamson SDP for Max-Cut on an $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$0-vertex graph, the only $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$1 constraints are $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$2 for $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$3 and $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$4. Applying the Barvinok–Pataki bound yields: $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$5 Thus, there always exists an optimal factorization $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$6 with $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$7 for $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$8. This bound delineates the minimum rank for which rank-constrained parametrizations can, in principle, reach all optimal solutions (O'Carroll et al., 2022).

4. Burer–Monteiro Approach and Rank Thresholds

The Burer–Monteiro method factorizes the primal variable as $A_1 \cdot X = b_1, \ldots, A_m \cdot X = b_m,$9 with $A_i \in \mathbb{S}^n$0 and directly optimizes $A_i \in \mathbb{S}^n$1 subject to the constraints (possibly nonconvex). When $A_i \in \mathbb{S}^n$2 meets or exceeds the Barvinok–Pataki bound, there is no loss of global optimality due to the rank restriction.

Recent work establishes that, for "generic" (full measure) data and in smoothed analysis, if $A_i \in \mathbb{S}^n$3, the nonconvex approach has no spurious local minima and any second-order critical point is globally optimal. Explicitly, if $A_i \in \mathbb{S}^n$4 for some fixed $A_i \in \mathbb{S}^n$5, then Burer–Monteiro can solve SDPs to arbitrary precision in polynomial time in the smoothed model (Cifuentes et al., 2019). For $A_i \in \mathbb{S}^n$6, this threshold converges to the Barvinok–Pataki value $A_i \in \mathbb{S}^n$7.

5. Limitations and Necessity of Genericity

Despite the existential guarantee of low-rank optimal solutions, the Barvinok–Pataki threshold does not preclude the appearance of spurious local minima in the nonconvex Burer–Monteiro factorization for certain concrete (non-generic) SDPs. "The Burer-Monteiro SDP method can fail even above the Barvinok-Pataki bound" demonstrates for Max-Cut that there exist deterministic instances (of zero measure) for all even $A_i \in \mathbb{S}^n$8 and $A_i \in \mathbb{S}^n$9, with spurious local minima at factorization ranks $\mathbb{S}^n_+$0 (O'Carroll et al., 2022). The construction relies on cost matrices with block structure leading to "axial" local minima that are provably non-global.

This shows that the Barvinok–Pataki bound is necessary but not sufficient for the absence of all non-global second-order critical points in the Burer–Monteiro method: the "no-spurious-minima" property requires additional genericity or smoothing of the problem data.

6. Extensions, Corollaries, and Complexity Implications

Strengthening of the rank bound in structured cases (notably Stiefel-manifold SDP relaxations) demonstrates that for problems with block-identity constraints, linear conditions such as $\mathbb{S}^n_+$1 can guarantee SDP exactness, expanding the tractability window (Song et al., 2023).

In general, for SDPs with $\mathbb{S}^n_+$2 constraints, the Barvinok–Pataki bound enables solution of the original problem via low-rank factorizations when $\mathbb{S}^n_+$3, dramatically reducing memory/storage costs (from $\mathbb{S}^n_+$4 to $\mathbb{S}^n_+$5) compared to interior point methods (Cifuentes et al., 2019). For SDPs arising in combinatorial optimization, control, and estimation, this is a foundational result underpinning both direct and approximation schemes.

$\mathbb{S}^n_+$6

7. Historical Context and Further Developments

The rank bound was established by Barvinok (1995) and Pataki (1998), initially as Carathéodory-type statements about convex combinations in spectrahedra. Extensions and refinements, particularly for structured constraint sets (Stiefel blocks), have expanded the areas of guaranteed exact relaxation. The critical realization, substantiated by recent constructions, that spurious local minima can exist even above the Barvinok–Pataki threshold in measure-zero pathological cases, motivates beyond-worst-case analysis (genericity, smoothed analysis) for practical guarantees in large-scale SDP optimization (Song et al., 2023, O'Carroll et al., 2022, Cifuentes et al., 2019).