Iterative alternating direction rnethods 725
Remark 1 Frmn such reasoning it seems clear that ADM with variable
parameters is equivalent to the two-layer scheme (26) with iteration parain-
eter T = Tk = Tk (1 ) + T~ 2) and factoriied operator
\^1 Ve will not pursue analysis of this: the ideas needed to do so have been
covered.
Remark 2 If A is a sum of p > 2 pairwise c01n1nutative operators such
that
p
A = ~ Aa, A~ = Aa > 0, baE < Aa < l::laE, b0: > 0,
<>=1
CY= 1, 2, ... ,p, AaAp = AμAa, o:,(3 = 1, 2, ... ,p,
direct applications of ADJ\!I described by ( 17) is i1npossible in principle,
so there is a real need for constructing sche111e ( 26) with the factorized
operator
p
Bk= II (E+r~"')Aa),
<>=1
In this regard, we are unaware of any exact solution to minimax proble1n
and the so-called cyclic set of the ensuing para1neters may be of help in
achieving the final aims.
The 1nain idea behind this approach is connected with the equation
related to a new iteration Yk+l:
vvhich reduces to successive solution of p equations
( E + T k (I) A 1 ) yl I ) -- F k J (E., + T(u)A k u. ) 1/(u) = .,/(o-1) < , a:=,^2 ... ,]!,
with Yk+l = ylP) incorporated.
The difference Dirichlet problem for Poisson's equation in a p-di1nen-
sional unit cube such as
AuY = Yx• • Cl:'' .,. l" '