Dantzig–Wolfe decomposition

Dantzig–Wolfe decomposition
Name	Dantzig–Wolfe decomposition
Invented	1960s
Inventor	George Dantzig; Philip Wolfe
Field	Mathematical optimization
Related	Linear programming; Column generation; Duality

Dantzig–Wolfe decomposition Dantzig–Wolfe decomposition is a mathematical programming technique for solving large-scale Linear programming problems by reformulating a problem into a master problem and subproblems, enabling efficient use of Simplex algorithm, Column generation, and duality concepts; it is widely used in industrial-scale models originating from transportation and scheduling applications. The method leverages structure in block-angular constraint matrices to decompose problems into smaller subproblems and a coordinating master problem, integrating ideas related to the Primal-dual method, Lagrangian relaxation, and decomposition approaches developed in operations research by researchers affiliated with institutions such as Stanford University, RAND Corporation, and IBM.

Introduction

Dantzig–Wolfe decomposition reformulates a structured Linear programming model with a block-angular constraint matrix into a convex combination of extreme points and rays of the subproblems' feasible regions, producing a restricted master problem whose columns are generated dynamically via pricing subproblems solved by techniques related to the Simplex algorithm and the Revised simplex method; the approach exploits properties studied by pioneers including George Dantzig, Philip Wolfe, John von Neumann, and contemporary contributors from Princeton University and Massachusetts Institute of Technology. Typical model contexts include multicommodity flow in Port of Rotterdam-scale networks, crew scheduling akin to problems studied by American Airlines, and vehicle routing variants reminiscent of challenges addressed by UPS and FedEx. The decomposition connects to theoretical frameworks advanced in seminars at INFORMS conferences and workshops at Courant Institute.

Formulation and theoretical foundations

The canonical setup assumes a master coupling constraint linking block-separable subproblems; each block corresponds to a polyhedron whose extreme points and rays are associated with columns in the master problem, a construction reflecting polyhedral theory developed by Hermann Minkowski and furthered by work at Bell Labs and AT&T. The reformulation relies on convex hull representations and on linear inequalities studied in the context of Fulkerson-type network polyhedra and Total unimodularity results proven in collaborations involving researchers at University of California, Berkeley and Cornell University. Dual variables of the master correspond to prices in subproblem objective functions, echoing equilibrium concepts found in models by Paul Samuelson and algorithmic ideas propagated through seminars at MIT Sloan School of Management and Columbia Business School.

Algorithmic implementation

Implementation alternates between solving a restricted master LP and one or more pricing subproblems; the master is typically handled with a Revised simplex method or interior-point solvers developed by software vendors such as IBM's optimization group and Gurobi Technologies, while pricing subproblems are often solved with combinatorial optimization methods influenced by research at École Polytechnique Fédérale de Lausanne and INRIA. Column generation iteratively adds columns corresponding to improving extreme points found by the subproblem, a mechanism applied in large-scale instances solved by teams at Microsoft Research and Google; stabilization techniques like Penalty methods and Trust-region methods introduced by authors affiliated with University of Oxford and Princeton University can be incorporated to mitigate tailing-off phenomena.

Convergence and computational properties

Under standard assumptions on boundedness and feasibility the algorithm converges to an optimal solution of the reformulated LP; proofs build on duality theory advanced by John von Neumann and L. V. Kantorovich, and convergence rates relate to pivoting behavior of the Simplex algorithm studied by researchers at Stanford University and University of Chicago. Practical performance depends on subproblem complexity, degeneracy patterns examined in studies at Carnegie Mellon University and Georgia Institute of Technology, and stabilization schemes proposed at INFORMS meetings; numerical stability and memory requirements motivate hybrid implementations combining column generation with Branch-and-bound and Branch-and-price frameworks developed by teams at University of Illinois at Urbana-Champaign and Imperial College London.

Applications and extensions

Dantzig–Wolfe decomposition underpins many applications: large-scale Multicommodity flow problems in transport networks studied by researchers at MIT and University of British Columbia, crew scheduling in airline operations as addressed by American Airlines and academic groups at University of Amsterdam, vehicle routing problems tackled by teams at ETH Zurich and INSEAD, and production planning models explored at Toyota and Siemens. Extensions include integer programming adaptations such as Branch-and-price introduced in collaborations involving IBM and Bellcore, stochastic variants connected to research at Columbia University and University of California, Los Angeles, and decomposition hybrids combining ideas from Benders decomposition and Lagrangian relaxation as developed by scholars at University of Toronto and University of Washington.

Historical context and development

The method originated from foundational work in the 1960s by practitioners and theorists including George Dantzig and Philip Wolfe, emerging from environments such as RAND Corporation and academic collaborations at Stanford University; subsequent maturation occurred through dissemination at SIAM conferences and in textbooks produced by authors associated with Princeton University and Wiley. Over decades, contributions from industrial research labs at Bell Labs, IBM, and university groups at INSEAD, ETH Zurich, and Cornell University expanded theoretical understanding and computational practice, integrating the technique into industrial solvers and decision-support systems used by organizations like American Airlines and logistics firms exemplified by DHL. The provenance of the method reflects intersections among optimization theory, applied operations research, and computational advances fostered at institutions such as MIT, Columbia University, and University of California, Berkeley.

Category:Mathematical optimization