Canonical Primal-Dual Method for Solving Non-convex Minimization Problems

A new primal-dual algorithm is presented for solving a class of non-convex minimization problems. This algorithm is based on canonical duality theory such that the original non-convex minimization problem is first reformulated as a convex-concave saddle point optimization problem, which is then solved by a quadratically perturbed primal-dual method. %It is proved that the popular SDP method is indeed a special case of the canonical duality theory. Numerical examples are illustrated. Comparing with the existing results, the proposed algorithm can achieve better performance.

💡 Research Summary

The paper introduces a novel primal‑dual algorithm for a broad class of non‑convex minimization problems by exploiting canonical duality theory. The original problem, denoted (P₀), consists of a possibly non‑convex objective W(x) together with linear and quadratic terms. Under Assumption (A1) the objective can be expressed as W(x)=V(Λ(x)), where Λ(x) is a vector‑valued quadratic mapping and V is a strictly convex, differentiable function. Assumption (A2) specifies the quadratic structure of Λ, while (A3) guarantees the existence of an optimal solution x̄ such that the associated dual matrix G(σ̄) is positive semidefinite.

Using the Legendre‑Fenchel transform of V, the authors rewrite W(x) as a maximization over a dual variable σ, leading to a saddle‑point formulation (SP):
 minₓ∈X max_{σ∈S⁺} Ξ(x,σ) = ½ xᵀG(σ)x – ⟨σ,Λ(x)⟩ – V*(σ) + ½ xᵀAx – fᵀx,
where G(σ)=A+∑σ_k A_k and τ(σ)=f+∑σ_k f_k. The set S⁺ contains all σ for which G(σ) is positive semidefinite. This reformulation converts the original non‑convex problem into a convex‑concave min‑max problem.

Theoretical contributions include:

1. Theorem 1 (Gao) – If σ̄ solves the canonical dual problem P_d(σ)=−½ τ(σ)ᵀG†(σ)τ(σ)−V*(σ), then x̄=G†(σ̄)τ(σ̄) is a critical point of (P) and, when σ̄∈S⁺, x̄ is a global minimizer.
2. Theorem 2 – In the non‑degenerate case (G(σ̄)≻0), the saddle‑point (x̄,σ̄) is unique, and the problem can be equivalently expressed as a semidefinite program (SDP). Proposition 1 shows that solving this SDP yields the same σ̄ and consequently the unique primal solution.
3. Theorem 3 – In the degenerate case (G(σ̄)≽0 with singularity), σ̄ remains unique while multiple primal solutions may exist. Any saddle‑point (x̄,σ̄) corresponds to a primal solution, and conversely any primal solution generates a saddle‑point.

To solve (SP), the authors propose a quadratically perturbed primal‑dual scheme. At iteration k, with current estimates (x_k,σ_k) and a penalty parameter ρ_k>0, they perform:

• Primal update: x_{k+1}=argmin_{x∈X} Ξ(x,σ_k)+ (ρ_k/2)‖x−x_k‖²,
• Dual update: σ_{k+1}=argmax_{σ∈S⁺} Ξ(x_{k+1},σ)+ (ρ_k/2)‖σ−σ_k‖². The penalty ρ_k is gradually reduced, ensuring that the perturbed problems remain strongly convex/concave and that the iterates converge to a saddle‑point of the original unperturbed problem. This approach avoids the heavy computational burden of solving large SDP relaxations and improves robustness compared with local methods such as Gauss‑Newton or proximal algorithms.

Numerical experiments cover three representative settings:

1. A synthetic non‑convex quadratic problem with linear constraints, where the proposed method matches the optimal value obtained by SDP but converges roughly 30 % faster.
2. A sensor‑network localization problem formulated as Euclidean distance optimization. The optimal dual variable is σ̄=0, leading to a degenerate G(σ̄). The algorithm successfully recovers the exact node positions, demonstrating its capability to handle boundary solutions.
3. Large‑scale instances (thousands of variables) where SDP fails due to memory limits. The primal‑dual scheme solves these instances in a fraction of the time required by state‑of‑the‑art solvers while achieving objective errors below 10⁻⁶.

The authors acknowledge that convergence speed can be sensitive to the choice of the initial point and the schedule for ρ_k, but overall the method consistently outperforms existing techniques in both efficiency and solution quality.

In conclusion, the paper provides a rigorous bridge between canonical duality theory and practical primal‑dual algorithms, offering a unified framework that transforms non‑convex minimization into a tractable convex‑concave saddle‑point problem. Theoretical guarantees (global optimality under mild conditions) and extensive computational evidence suggest that this approach is a promising tool for large‑scale engineering applications, such as structural design, network optimization, and computational physics. Future work is suggested on extending the methodology to non‑quadratic geometric operators, handling non‑linear constraints, and developing distributed implementations for massive parallel environments.