In this paper, we investigate the geometry of a general class of gradient flows with multiple local maxima. we decompose the underlying space into disjoint regions of attraction and establish the adjacency criterion. The criterion states a necessary and sufficient condition for two regions of attraction of stable equilibria to be adjacent. We then apply this criterion on a specific type of gradient flow which has as many as n! local maxima. In particular, we characterize the set of equilibria, compute the index of each critical manifold and moreover, find all pairs of adjacent neighbors. As an application of the adjacency criterion, we introduce a stochastic version of the double bracket flow and set up a Markov model to approximate the sample path behavior. The study of this specific prototype with its special structure provides insight into many other difficult problems involving simulated annealing.
[1]
Roger W. Brockett.
Modeling the transient behavior of stochastic gradient algorithms
,
2011,
IEEE Conference on Decision and Control and European Control Conference.
[2]
R. Brockett,et al.
Dynamical systems that sort lists, diagonalize matrices and solve linear programming problems
,
1988,
Proceedings of the 27th IEEE Conference on Decision and Control.
[3]
A. Banyaga,et al.
Lectures on Morse Homology
,
2005
.
[4]
J. Palis.
On Morse-Smale dynamical systems
,
1969
.
[5]
A. Horn.
Doubly Stochastic Matrices and the Diagonal of a Rotation Matrix
,
1954
.
[6]
R. Brockett,et al.
A Gradient Flow for the Assignment Problem
,
1991
.
[7]
A. Banyaga,et al.
Morse-Bott homology
,
2006,
math/0612316.
[8]
Roger W. Brockett.
Notes on Stochastic Processes on Manifolds
,
1997
.
[9]
J. Meza,et al.
Steepest descent
,
2010
.