[Paper Review] Canonical Duality Theory for Solving Minimization Problem of Rosenbrock Function

David Yang Gao, Jiapu Zhang|arXiv (Cornell University)|Aug 5, 2011

Advanced Optimization Algorithms Research16 references3 citations

TL;DR

This paper applies canonical duality theory to solve the nonconvex Rosenbrock function minimization problem, reformulating it as a canonical dual problem with zero duality gap. The method analytically identifies global and local solutions via the triality theorem, enabling efficient and precise computation of the global minimum across dimensions, including n=4000, where standard gradient methods fail.

ABSTRACT

This paper presents a canonical duality theory for solving nonconvex minimization problem of Rosenbrock function. Extensive numerical results show that this benchmark test problem can be solved precisely and efficiently to obtain global optimal solutions.

Motivation & Objective

• To address the challenge of solving the nonconvex Rosenbrock function, a well-known benchmark in global optimization that resists standard gradient-based methods.
• To demonstrate that the canonical duality theory can reformulate the Rosenbrock minimization problem into a canonical dual problem with zero duality gap.
• To enable exact identification of both global and local minimizers using the triality theorem.
• To show that the discrete gradient (DG) method applied to the dual problem consistently converges to the global solution, even in high-dimensional cases (n=4000).
• To enable scalable, parallel computation of the dual problem using MPI for large-scale optimization.

Proposed method

• Reformulate the primal Rosenbrock problem into a canonical dual form using Legendre-Fenchel transformation and geometric admissible measures.
• Define the canonical dual function $ P^d(oldsymbol{ ho}) $ via the total complementary function $ \Xi(\mathbf{x}, \boldsymbol{\rho}) $, which links primal and dual variables.
• Introduce the shifting operators $ \delta^\flat $ and $ \delta^\sharp $ to express the dual solution in terms of dual variables $ \boldsymbol{\rho} $.
• Establish the canonical dual feasibility space $ \mathcal{S}_a^+ $ where $ \rho_i + 1 > 0 $, ensuring the existence of global solutions.
• Apply the triality theorem to classify critical points: global minima when $ \boldsymbol{\rho} \in \mathcal{S}_a^+ $, and local minima otherwise.
• Use the discrete gradient (DG) method to solve the dual problem, which yields the global solution with zero duality gap.

Experimental results

Research questions

• RQ1 Can canonical duality theory be used to solve the nonconvex Rosenbrock function minimization problem with guaranteed global optimality?
• RQ2 How does the canonical dual problem compare to the primal problem in terms of convergence and solution quality for high-dimensional cases?
• RQ3 What role does the triality theorem play in distinguishing global from local minimizers in this context?
• RQ4 Can the dual formulation be efficiently parallelized for large-scale problems, such as n=4000?
• RQ5 Why do standard gradient methods fail to locate the global minimum when initialized at (3,3,...,3) or (100,100,...,100)?

Key findings

• For all tested cases (n=2, 6, 7, 4000), the canonical dual problem solved via the DG method produced the exact global minimum $ P(\mathbf{x}) = 0 $ at $ \mathbf{x} = (1,1,\dots,1) $.
• The dual solution $ \bar{\boldsymbol{\rho}} $ was numerically zero (within machine precision) for all global minima, confirming $ P^d_+(\bar{\boldsymbol{\rho}}) = 0 $.
• When initialized at (3,3,...,3) or (100,100,...,100), the DG method on the primal problem converged to local minima with objective values >47.23824896 for n≥4000.
• For n=6 and n=7, the DG method on the primal problem yielded solutions with $ P(\mathbf{x}) \approx 4.006 $ and $ 4.094 $, respectively, while the dual method correctly identified the global minimum at 0.
• The dual problem was successfully parallelized using MPI, with each of $ n-1 $ processes computing one term of the sum, enabling scalable computation on large-scale systems.
• The canonical dual formulation ensures zero duality gap, and the triality theorem guarantees that critical points of the dual correspond to critical points of the primal with exact objective value matching.