1: \begin{abstract}

2: We develop a general framework for MAP estimation in discrete and

3: Gaussian graphical models using Lagrangian relaxation techniques.  The

4: key idea is to reformulate an intractable estimation problem as one

5: defined on a more tractable graph, but subject to additional

6: constraints. Relaxing these constraints gives a tractable dual

7: problem, one defined by a thin graph, which is then optimized by an

8: iterative procedure.  When this iterative optimization leads to a

9: consistent estimate, one which also satisfies the constraints, then it

10: corresponds to an optimal MAP estimate of the original model.

11: Otherwise there is a ``duality gap'', and we obtain a bound on the

12: optimal solution.  Thus, our approach combines convex optimization

13: with dynamic programming techniques applicable for thin graphs. The

14: popular tree-reweighted max-product (TRMP) method may be seen as

15: solving a particular class of such relaxations, where the intractable

16: graph is relaxed to a set of spanning trees.  We also consider

17: relaxations to a set of small induced subgraphs, thin subgraphs

18: (e.g. loops), and a connected tree obtained by ``unwinding'' cycles.

19: In addition, we propose a new class of multiscale relaxations that

20: introduce ``summary'' variables.  The potential benefits of such

21: generalizations include: reducing or eliminating the ``duality gap''

22: in hard problems, reducing the number or Lagrange multipliers in the

23: dual problem, and accelerating convergence of the iterative

24: optimization procedure.

25: \end{abstract}

26: