4.3 Algebraic and Numerical Methods
In this section, we first consider the task of constructing continued- fraction representations from power series. This is the topic which we considered from a theoretical viewpoint in the previous section. Second, we discuss methods by which the continued fractions which we have constructed, or derived, should be evaluated numerically.
Viskovatov's method  is a handy method for constructing contin- ued-fraction expansions, or equivalently staircase sequences of Pade approximants, from the given power series. It avoids having to reciprocate any series numerically, as was part of the construction of (2.2) and (2.3). It is a labor saving device which is as useful for people as for computers. The method is based on the formal identity
Viskovatov's method is best described by an illustrative example.
(3.2)Thus the generation of staircase sequences of Pade approximants is absolutely straightforward in the nondegenerate case. The other straightforward method of economical construction of continued fractions from power series is the Q.D. algorithm, which we treat fully in the next section.
The other main question we consider in this section is that of how to evaluate a fraction such as (3.2) numerically. There does not seem to be an agreed best method of computing continued fractions, and so we state three principal methods without making any clear recommendation about which is "best". The problem is, given the elements a,, ft, and a value of the variable z, to compute which we assume to be a convergent fraction.
Forward recurrence method. The numerators At-(z) and denominators 5,(z) are calculated from the recurrence relation (2.7). Because of the ease of computing the values of successive convergents, the forward recurrence method is the standard method for computing continued fractions.
It is sometimes the case in practice that the crude approximation
gives a rough and ready estimate of the rate of growth of An and Bn. Equation (3.4) is exact in the special case that the coefficients an ft, in (2.7) are constant. If |'"i|>|'2|, the Y terms are the dominant components and the /2,5 terms become less significant as oo. Such considerations assist in the understanding of the following numerical hazards of the forward recurrence method.
Floating point overflow of A n, Bn. If |r, |> 1, floating-point overflow may occur at some point during the actual computation. Rescaling of An, An_\, Bn, and Bn , at the critical point in the obvious way is recommended. Underflow must likewise be anticipated.
Suppression of required solution by a dominant solution. If the true solutions An, Bn of the exact recurrence relation (2.7) are the subdominant solutions, they may be suppressed numerically. Other solutions, An, Bn, generated by different initial conditions (e.g. A0 ^A0) for which An/An —>0, B„/B„ —>0, exist in this case. Rounding error introduced in the initial stages always becomes a dominant effect in such calculations. The connection with the concepts of Miller's backward recurrence algorithm is obvious.
(c) Accumulation of roundoff error in a dominant solution. If \An\ and \Bn\ are decreasing sequences, rounding error may accumulate using the forward recurrence method. This has been discovered empirically [Jones and Thron, 1974] for the following continued fraction:
using fixed decimal precision in the floating point
operations. (Why would binary operations be different.') the traction is being evaluated on the edge of its domain of convergence, and so it is a natural prototype for investigation of the buildup of roundoff error.
Backward recurrence method. This method of evaluating the wth convergent of (3.3) starts at the "tail" end of the convergent. We define
is the wth convergent of f(z). The major drawback of this
method is that ot deciding in advance which is the appropriate value of n to choose so that the wth convergent is an adequate approximation to f(z).
Summation formula. The wth convergent of (3.3) is given by the formula
where the quantities p,, p2, p3,..., p„ are defined in terms of the elements of the continued fraction by
and recursively by
Proof. We construct a summation formula by using the identity
a formula equivalent to (3.10) is derived as (7.6). A comparison of Equations (3.7), (3.10) leads us to define
Equation (3.8) defines the first two convergents correctly. To establish (3.9). we use the recurrence relation
to prove that
and the recurrence
to prove that
Equation (3.9) is proved by multiplying (3.12) and (3.13).
which is a representation of (5.55)
A consequence of (3.8), (3.9) and the equivalence transformation (1.6) is the new representation
The summation formula expressed by (3.7)-(3.9) has the advantage that it is suitable for iterative computation, which is important if the rate of convergence is not known a priori.
For further details of the numerical methods, and especially for error estimates, we refer to Jones and Thron's companion volume , to Blanch , and to Gautschi .
Exercise 1. Prove that the value of the repeating fraction (assumed conver-
satisfies the equation
Deduce the value of the fraction. Pringsheim  gives a full discussion of the evaluation of periodic continued fractions.
Exercise 2. Solve the recurrence relation for the numerators and denominators of the fraction