H. Wolkowicz et Q. Zhao, AN ALL-INCLUSIVE EFFICIENT REGION OF UPDATES FOR LEAST CHANGE SECANT METHODS, SIAM journal on optimization, 5(1), 1995, pp. 172-191
Least change secant methods, for function minimization, depend on find
ing a ''good'' symmetric positive definite update to approximate the H
essian. This update contains new curvature information while simultane
ously preserving, as much as possible, the built-up information from t
he previous update. Updates are generally derived using measures of le
ast change based on some function of the eigenvalues of the (scaled) H
essian. A new approach for finding good least change updates is the mu
lticriteria problem of Byrd, which uses the deviation from unity, of t
he n eigenvalues of the scaled update, as measures of least change. Th
e efficient (multicriteria optimal) class for this problem is the Broy
den class on the ''good'' side of the symmetric rank one (SR1) update
called the Broyden efficient class. This paper uses the framework of m
ulticriteria optimization and the eigenvalues of the scaled (sized) an
d inverse scaled updates to study the question of what is a good updat
e. In particular, it is shown that the basic multicriteria notions of
efficiency and proper efficiency yield a region of updates that contai
ns the well-known updates studied to date. This provides a unified fra
mework for deriving updates. First, the inverse efficient class is fou
nd. It is then shown that the Broyden efficient class and inverse effi
cient class are in fact also proper efficient classes. Then, allowing
sizing and an additional function in the multicriteria problem, result
s in a two parameter efficient region of updates that includes many of
the updates studied to date, e.g., it includes the Oren-Luenberger se
lf-scaling updates, as well as the Broyden efficient class. This effic
ient region, called the self-scaling efficient region, is proper effic
ient and lies between two curves, where the first curve is determined
by the sized SR1 updates while the second curve consists of the optima
l conditioned updates. Numerical tests are included that compare updat
es inside and 'outside the efficient region.