Finding Multiple Roots of Nonlinear Algebraic Equations Using S-System Methodology
Finding Multiple Roots of Nonlinear Algebraic Equations Using S-System Methodology
Finding Multiple Roots of Nonlinear Algebraic Equations Using S-System Methodology
Michael A. Savageau
ABSTRACT
1. INTRODUCTION
suggests that methods developed to deal with S-systems will apply to a broad
class of nonlinear problems. For example, a method based on this canonical
form has been developed to solve initial value problems, and a diverse set of
benchmark tests have shown that it is fast, accurate, predictable and robust
[4]. The structural regularity of this canonical form suggests that it might also
yield methods for finding the roots of rather general nonlinear algebraic
equations.
Furthermore, the steady-state solution of an S-system reduces to a linear
algebraic problem in a logarithmic coordinate system [ll]. The result is given
bY
Y] = [Al-‘bl (3)
where
yk = In X,
b, = ln( &/a~).
When the linear system is underdetermined, the solution is obtained from the
set of linear equations with maximum rank together with a set of nonlinear
constraints that are generated in the recasting procedure.
The final step in recasting a set of nonlinear equations into S-system form
involves reducing multiple sums and differences of terms that involve prod-
ucts of power-law functions to a single difference of terms that involve
products of power-law functions [12].
The reductive step involves associating appropriate terms to yield two equa-
tions in S-system form.
From this result, it is immediately apparent that X, and X, will not exhibit a
steady state even though the original variable X, does. The variable X,
increases continually with time because the right-hand side of its equation is
positive definite. The variable X, decreases continually with time in such a
manner that the product X,X, approaches the steady state of X,.
This awkwardness is due to the manner in which the negative term was
distributed entirely to the equation for X,. Since the manner of distribution is
arbitrary, one can choose the distribution that will produce a true steady state
for each of the new variables X, and X,. For example,
where
x, + x, = 1. (10)
fi = xz f3 = x,
x2 + x3
(12)
XL?
+ x3
ln(l/d = x
2
=zx
ln
3( X2
X2 + X3 I+
x3
X2 + X3
In
i
X3
X2 + X3 I (13)
and in steady state the set of equations (S), (9), and (11) can be w&ten in
Roots of Nonlinear Algebraic Equations 191
linear form as
(14)
The variable elements in the matrix [ A] and the vector b] of (14) can be
determined from an arbitrary distribution over the auxiliary variables X, and
X,. This allows a solution of the linear system, including new estimates for
the auxiliary variables, and the procedure is repeated. This iterative algorithm
converges rapidly to a distribution over the auxiliary variables that yields a
true steady state for the S-system equations produced in the reductive step of
recasting. It also yields a steady-state solution for the original variable X,.
This equation can be recast to S-system form again by letting the original
variable be replaced by the product of new variables, the minimum number
being two. However, four new variables are needed to ensure that the
resulting S-system equations have a true steady state. Let x, = x,xb x,x,!,
differentiate the product, and associate appropriate terms to yield four
equations in S-system form.
192 MICHAEL A. SAVAGEAU
where
x2 + x3 = 1 (20)
x, + x, = 1. (21)
y1x,f2xp = 1 (22)
yzxfix5fs = 1 (23)
where
f2 = x2 j-3 = x3
(24)
x2 + x3 x2 + x3
x, X,
ln( l/rl) = x :, In + In
x2”;,:, (25)
2 3 i x2 + x3 I x2 + x3 i I
f4 = x4 f5 = x5 (26)
X‘I+ x5 x4 + x5
X5 x5
ln(l/r2> = x
4
“:, ln i 5
x4
x* + xs 1 +
x4 + x5
In
i x4 + x5 1 (27)
and in steady state, the set of equations (16)-(181, (22)~(23) can be written in
_I
linear form as
(&h-h) -1 0 1 0 yI - 14 h/c~~>
_
(glllA12) - 1 0 0. 1 yz 14 k42/all)
(g,,2&,,) 0 -1 1 0 Y3 = ln( &/Q) (28)
0 fi f3 0 0 y4 14 l/x)
0 0 0 f4 f5 y5 - ln(l/y2)
Note that (19) is a linear combination of (16)~(18) in steady state, and so has
been eliminated.
Roots of Nonlinear Algebraic Equations 193
The variable elements in the matrix [A] and the vector b] of (28) can be
determined from an arbitrary distribution over the auxiliary variables X, and
X,, and X4 and X5. This allows a solution of the linear system, including new
estimates for the auxiliary variables, and the procedure is repeated. This
iterative algorithm converges rapidly to a distribution over the auxiliary
variables that yields a true steady state for the S-system equations produced
in the reductive step of recasting. It also yields a steady-state solution for the
original variable X,.
3.3. Generalization
The algorithm described above in the context of two simple examples is
readily generalized to treat systems of nonlinear algebraic equations having an
arbitrary number of equations (variables) with an arbitrary number of positive
and negative terms in each equation. The auxiliary variables exist as disjoint
sets numbering at most two for each of the original equations. The number of
auxiliary variables in each of these sets is determined by the number of terms
of a given sign in each of the original equations.
If one assigns an initial value of unity to one of the auxiliary variables in
each set and zero to all the others in each set, then the number of different
initial assignments is
NAS
where
This product also provides an upper bound on the number of positive real
roots. This bound is less than Bezout’s number [l, 81 in the case of
multinomial equations with a high order but few terms per equation; it is
greater than Bezout’s number in the case of multinomial equations with low
order but many terms per equation. Yet another upper bound is less than
either of these in some instances. It is calculated as follows. First, determine
the minimum number of terms of a given sign for each equation. Second,
double these numbers and subtract one from each doubled number that is
greater than two. Finally, take the product of the resulting numbers to give
an upper bound on the number of positive real roots. The search for
additional roots can be terminated once any of these bounds is reached.
The algorithm as currently implemented has found several, and in many
cases all, of the positive real roots for the problems tested when one chooses
194 MICHAEL A. SAVAGEAU
the initial values for the auxiliary variables according to the combinatorial
scheme described above. The method can be generalized in a number of
ways, as will be described below.
4. NUMERICAL EXAMPLES
The following examples are representative of the test problems that have
been examined. In each case a tolerance parameter has been set at l.Oe-8.
The three positive real roots are found as in Table 1. These require five to
eight iterations and the rate of convergence is approximately quadratic. The
TABLE 1
ROOTS WITH NEGATIVE AS WELL AS POSITIVE REAL COMPONENTS FOR THE PROBLEM
IN EXAMPLE 1
Rate of Convergence+
* The tolerance was set at e - 8, but the data are reported with four significant
digits.
’ Rate of convergence between two iterations is calculated as ri = log(ei/ei_r).
*Average is the geometric mean calculated from the iteration that first produces
an error less than 1.
’ Rate of convergence between the last two iterations before the tolerance
condition is met.
Roots of Nonlinear Algebraic Equations 195
method also finds roots with a negative real component. These are obtained
by replacing xi by -xi, which leads to a reversal of sign for odd functions of
xi, and solving for the positive real roots of the resulting equations. This
yields the six roots with a negative real component (Table 1). A similar
number of iterations and a similar rate of convergence are observed for these
solutions. Thus, the method finds nine real solutions, which is the total for
this set of equations [6].
3xl-‘x”.0”
1 2 + x3x6
1 2
+ x3 - 1ox2xP
1 2
- x4 = 0 (32)
The four positive real roots are found as in Table 2. In this example, three to
six iterations are required, and the rate of convergence is again approximately
quadratic. The method also finds all the roots with one or more components
equal to zero (Table 2). I n some of these cases, the equations reduce exactly
to S-systems and the solution is obtained directly without iteration. Otherwise
three to seven iterations are required, and the rate of convergence is nearly
quadratic. The method finds all of the nonnegative real roots.
(38)
TABLE2
ROOTS WITH ZERO AS WELL AS POSITIVE REAL COMPONENTS FOR THE PROBLEM
IN EXAMPLE 2
Rate of
Convergence+
Roots* IterationsAverage* Final'
*The tolerance was setat e - 8, but the data are reportedwith foursignificant
digits.
'Rate of convergence between two iterations iscalculated as T<= logfe,/e,>.
*Averageisthegeometricmean calculated from the iteration that first produces an
errorlessthan 1.
"Rate of convergence between the lasttwo iterations beforethe tolerance condi-
tionismet.
that for physical reasons there can only be a single solution, as in the case of
chemical equilibrium, one could stop the procedure after obtaining the first
solution.
5. DISCUSSION
REFERENCES