JOURNAL OF MATHEMATKAL ANALYSIS AND APPLICATIONS 101, 39-63 (1984)

Lie Series and

Nonlinear Ordinary Differential Equations*

STANLY STEINBERG’

Department of Mathematics and Statistics,

University of New Mexico, Albuquerque, New Mexico 87131
Submitted by C. L. Dolph

Lie series are used to calculate both closed form and approximate solutions for
elementary nonlinear ordinary differential equations.

Contents. 1. Introduction. 2. Classical results. 3. Lie series. 4. One Term

equations. 5. Two term equations, the finite case. 6. Two term equations, infinite
case. 7. General equations. 8. Equations defined by power series.

1. INTRODUCTION

We are interested in solving initial value problems of the form

that is, the initial value problem for a single, first order, autonomous non-
linear ordinary differential equation (NLODE). Thus we assume that y(t)
is a smooth real valued function, y’(t) = dy(t)/dt, y is a given real number,
and f(y) is a smooth function. We will be interested in the particular case
wheref can be written as a finite or infinite sum of powers of y:
f(y) = ,y aiya;+‘.
I
Here aj and cziare real and the form of the exponent was chosen to simplify
some formulas that occur later. To keep the initial value problem
nonsingular at y = 0, in the sensethat f(y) is finite, we require oli > -1.

* This work was partially supported by NSF Grant MCS-8102683.

’ Visiting Professor at the Instituto de Investigaciones en Matematicas Aplicadas y en
Sistemas, Universidad National Autonoma de Mexico, Mexico City. On Sabbatical leave
from the Department of Mathematics and Statistics, University of New Mexico, Albuquerque,
N.M. 87131.
39
0022-241)3/84 $3.00
Copyright 0 1984 by Academic Press, Inc.
All rights of reproduction in any form reserved.
40 STANLY STEINBERG

A Lie series is an exponential of a differential operator,

erf(y)d’dy
= g f (f(y) $) “*
Here t is a real parameter. To make printing our formulas more convenient,
we set
D = d/dy

and then write the Lie series as

effcy)D = i. f (f(y)D)“.

In terms of Lie series, the solution of the above initial value problem is
written
y(t) = efftyjDy.

One advantage of the Lie series notation is that it suggests the use of
noncommuting exponential identities. These identities are frequently called
Campbell-Baker-Hausdorff formulas. To make use of these identities we
need the concept of a commutator. If f (y)D and g(y)D are two differential
operators, then the commutator is given by

[.OYP, &PI =f(v) DdyP - g(y) W-(VP

= U-(Y) g’(y) - g(y) f’(y)P.

As we note in Section 2, the differential equation under consideration can

be analyzed in rather complete detail from a standard point of view. Why
then another discussion of such a simple equation? First, we wish to
acquaint the reader with Lie series techniques and this simple situation
facilitates understanding by avoiding technical detail. Also, in this simple
situation, our results are more complete than in the more general situation
and so provide a standard by which to measure and understand results in a
more general setting. It is also easy, in this situation, to compare the Lie
series results to the standard results for the nonlinear differential equation.
More importantly, in the case where the function f( y) can written as a finite
or infinite sum of powers of y, we can provide some new insight into the
solutions of the differential equation.
Thus if

f(y) = T alP+ ‘,
 LIE SERIES AND NLODEs 41

then we introduce the operators

L i = y “i+ ‘D.

The commutator algebra associated with the differential equation is the

smallest linear space of operators containing the Li and closed under
commutation. In the case that the commutator algebra is finite dimensional,
it is convenient to allow some of the ai in the expansion off(y) to be zero.
This then allows us to say that the commutator algebra is finite dimensional
provided the differential equation can be written using a finite sum and the
associated operators Li satisfy
[Li, Lj] = 2 CiL,

where the ci are real constants. This condition says, if we assumethe Li are
linearly independent, that the Li are a basis for a finite-dimensional Lie
algebra and the CL are the structure constants for this algebra.
It is possible to completely characterize the differential equations for
which the commutator algebra is finite dimensional and in all of these cases
we give an explicit closed form solution of the equation using Lie series.
Incidentally, these equations are all Riccati equations. When the commutator
algebra is not finite dimensional, the noncommuting exponential identities
provide us with a new approximation procedure for the solutions of the
differential equation that we call factored composition product expansions.
We believe that the connection between the commutator algebra and the
form of the solution of the equation is an important insight provided by the
Lie series technique.
Another important aspect of the Lie series technique is that these ideas
generalize to nonautonomous systems of nonlinear ordinary differential
equations [ 15, 16, 341. At present there are but a few applications for this
general theory. However, when these ideas are implemented for Hamiltonian
systems, a beautiful and useful theory results. Among the many areas to
which this theory has been applied are celestial mechanics [6, 12, 17, 181,
laser physics [13], magnetic lenses [8-lo], neutron transport [20, 23, 24, 29,
361, optical lenses [ 111, plasma physics [3], and other areas [ 1, 5, 14, 19,
21, 22, 28, 331. For a more extensive list of references see [34]. In many of
the applications of Lie series extensive algebraic computations are required.
Consequently many workers use computer symbol manipulation programs
[2, 27, 31, 341 to help with these computations. Finally, we note that this
work is related to much of the work in modern Hamiltonian mechanics [30].
This paper is organized as follows. In Section 2 we note the classical
approach to the initial value problem. In Section 3 we state the basic
properties of Lie series. We give no proofs here because the required proofs
can be found in the literature [4, 7, 8-10, 32, 34, 351. We note that except
42 STANLY STEINBERG

for the noncommuting exponential identities all of the results are elementary
[ 341. In Section 4 we note the solution of our initial value problem whenf is
a single power of y and show how the closed form solution can be used to
evaluate certain Lie series.
In Section 5 we solve, in closed form, equations wheref is the sum of two
powers and the commutator algebra is two dimensional. These equations can
be solved by classical means; however, we believe that the Lie series method
is simple, elegant, and demonstrates clearly the connection between the finite
commutator algebra and the closed form solution. It is our impression that
this method of solving these problems has not appeared in the literature
before. In Section 6 we look at two term equations for which the commutator
algebra is infinite dimensional and develop an approximation schemefor the
solution of the initial value problem that uses factored composition product
expansions. Because we deal with nonanalytic solutions of differential
equations, these results appear to have no analog in the literature. Moreover,
we show how to compute all terms in the expansion, not just the first few
terms, another result not recorded in the literature. We also give an estimate
of the degree of approximation to be expected from these expansions and
provide some numerical examples to illustrate that the estimates are good.
We believe that this is the first time such estimates have been given for
factored composition product expansions.
In Section 7 we address the question of finding all finite-dimensional
operator algebras made up of first order differential operators in one
variable. This result is due to Lie [25, 261. We produce a proof here to
illustrate in what circumstances one can obtain linite-dimensional
commutator algebras, and why all such equations are essentially Riccati
equations. We then use Lie series to give the closed form solution for the
typical example.
In the last section, Section 8, we look at the case where f is given as a
power series and write the solution of the initial value problem as an infinite
factored composition product. Again, we show how to compute all of the
terms in the factored composition product. Becausef(y) is analytic, this
situation is directly analogous to the Hamiltonian case. Even in this well-
studied case, we believe that the method for computing all of the terms in the
factored composition product expansion is new. We carry out this expansion
to five terms and then use the expressions to approximate the solution of a
particular initial value problem. Again, we give an estimate of the degree of
approximation that we believe is new and provide some numerical examples
to illustrate that these estimates are good.
 LIESERIESAND NLODEs 43

2. CLASSICAL RESULTS

Our differential equation is a first order autonomous differential equation,

v’(t) =f(Yw)T
wheref(y) is a sufficiently smooth function. If we write this equation in the
form

and separate variables, then

N-(y) = dt
and consequently

This formula has some notable difficulties; the integral may not be
computable in the closed form, and even if it is, the resulting formula may
not be solvable for y.
Next, we note that if we make a change of scale in both y and t, then the
equation
y’ = c a, ye’+ 1

becomes
y’ = C rqai+ la, yai+ ‘.

Thus, if we wish, we can normalize two of the constants in our equation. If

we make the nonlinear change of variable
y=zP, a # 0,
then our equation becomes

Thus we may normalize one of the exponents in our equation. Note that if
ai = 0, then the exponent does not change when we go from y to u.
Another point to note is that if all ai > -1, then y = 0 is a critical point of
the differential equation. If some ai < -1, then y = 0 is not a critical point
44 STANLY STEINBERG

and if some ai < -1, then y = 0 is a singular point of the differential

equation. To have the equation regular at y = 0 in the sensethat af(y) is
bounded near zero, we must require ai > 0.

3. LIE SERIES

Here, for the convenience of the reader, we state the basic properties of
Lie series. As we noted in the Introduction, other accounts can be found in
the literature. All of what we have to say in this section will hold in the case
of n variables. When not all of the oi are integers, the usual analyticity
assumptions are not satisfied.
A Lie series is an exponential of a first order linear analytic partial
differential operator in n-variables. For our purpose we may assumethat we
have only one variable and thus that we have an operator L of the form
L =f(Y)a D = d/dy

wheref(y) is an analytic function near y = 0. A Lie series is then given by

The action of the Lie series on a function g(y), analytic near y = 0, is given
by 00 rllT n
e”gW
= n=o
2 y * g(Y)
= n.o5 (f(v>D)” g(y)

= g(y) +f(yWg(y)) + &f(y) W(Y) Dg(y)) + .a- .

Taylor’s theorem gives us one elementary example of a Lie series that we can
sum,
et”&> = nzo 5 $ g(Y) = & + f).

If g(y) = y, then another example is

etyDg(Y)
=nzo
s (Y$1’
Y
=;o$y
= e’y.
 LIE SERIES AND NLODEs 45

As we will discuss later, this last result implies that

e’yDgW = de39
for any analytic function g.
Properties
We assume that f(y), g(v), and h(y) are analytic functions near y = 0,
that Q and b are real constants, that c(t) is a smooth real valued function,
and that
L =f(yP =f(Y) f.

( 1) Convergence. e”g(y) is a well-defined analytic function of y and t

for y and t small enough.
(2) Time derivative. dec(‘jL/dt = c’(t) LeC(‘jL = cC(‘jLc’(t)L.
(3) Linearity. efL(ag + bh) = aefLg + be*‘h.
(4) Product preservation. e”(gh) = (etLg)(etLh).
(5) Composition. e”g(y) = g(elLy).
We now suppose that P is another first order differential operator and that
F(x) is an analytic function.
(6) Similarity. etLPeptL = e”L,‘lP = C,“=. (t”/n!)[L, .]” P, where
[L, .]OP=P
[L,.]‘P=LP-PL=[L,P]
[L, .I” P = [L, *In-l [L, P], n> 1.
Also,
eKF(P) eptL = F(etLPepfL).

(7) Function multiplier. etLgeptLh = (etLg)h.

(8) Noncommuting exponential identities.

et(L +m = etLetPet2L*ef’L,et4L‘$et5L5 .. . = . . . et5L5e-14L,et3L,e-t2*etPetL,

where
L, = i[L, P]
L,=~[[L,P],Ll--;[[L,P],P]
L=&wJwl,Ll
-d[[[L,Pl,L],P] -Q[[[Lpl,pl,P]?
46 STANLY STEINBERG

and so forth. Also,

where

and so forth.
(9) Dlperential equation property. If y(t) = etf(Y)Dythen y’(t) =f(y(t)),
Y(O) = Y*
Because all of our results are based on this last property, we give its proof.
From the Time Derivative Property we get

-$ y(t) = et’(‘)’ f(y) Dy

= etf(YjD f(y).

The Composition Property then gives

y’(t) =f(etf’Y’Dy)

=f (y(t)>
as was desired.

4. ONE TERM EQUATIONS

We now look at problems of the form

y’(t) = ay”+ l(t), Y(O) = Y*
Separation of variables gives the solution of this problem as

y(t) = (1 _ a;ty-,l,c2

when a # 0. Note that if a > 0, then this solution is singular at aatya = 1.

Even if a < 0, in the case -a is not an integer, some finite derivative of the
solution will be singular. When a = 0 the solution is
y(t) = e”‘y.
 LIE SERIES AND NLODEs 47

PROPOSITION.

lQYDy = e’Qy
e

efCIY*+‘D
y= (1 - IaYaya)l/n 3
a f 0.

ProoJ: This follows directly from the Differential Equations Property of

Lie series and the solutions that we found above.

5. Two TERM EQUATIONS, THE FINITE CASE

We are now interested in problems of the form

y’(t) = ay”“(l) + byb+‘(t), Y(O) = Y,

where a and b are not zero and a and /I are distinct. The nonlinear change of
coordinates given in Section 2 reduces the problem to studying
y’(t) = a + byD+‘(t).

An important object in the use of Lie series is the commutator algebra

associated with all of the terms in the differential equation. Thus, for the last
equation, we compute

P9Y4+‘D] = (p+ 1) y%.

Next compute the higher order commutators,

[D, y4D] = ,L3y4-‘D,

[y’+‘D,y”D] = -y=‘D,

and so forth. If we continue in this fashion, then we will clearly get an

infinite list of distinct operators unless /I = 0.
Let us look at the case /3= 0 first. The equation under consideration is
now a simple linear equation that is easily solved by at least two classical
methods. However, we will use Lie series to solve this problem becausethe
technique used here is the same as that used in the most complicated
problems and it is nice to see it in such a simple setting. Because the
commutator algebra is finite dimensional, repeated applications of the
noncommuting exponential identities tell us that for somef and g,
et’a”+byD’ _- &DetbyD

409/101/1-4
48 STANLY STEINBERG

and

Here f and g are functions of t, a, and b. Note that the position of the
unknown function is determined by the commutation relation
[D, yD] = D.
The use of the noncommuting exponential identities to find a general form
for an expression without finding the exact formula is typical of our use of
these identities. The method we use to find the exact identity also plays an
important role in more general problems.
To find f we differentiate the proposed identity with respect to C:

(UD t byD) et(aD+byD)

=f rDefD@‘YD + &Dby&.tbYDa

Now multiply on the right by the inverse of the proposed identity,

e - t(aD + byD) = e - tbyDe -fD

to obtain
aD t byD =f’D + dDbyDe-fD.
The basic properties of Lie series give us
gDbyDeefD = b(dDye-fD)(dDDdD)
= b(dDy)D
= b(y +f)D.
Thus we need
aD =f’D t bfD
or
f't bf=a.
We also needf = 0 when t = 0, so
f = (a/b)(l - eebt).
We can calculate g in a similar way. However, it is easier to notice that if
we replace a by --Q and b by -b in the formula for the inverse of the identity
that we have already calculated, then we find
g(a, b, t) = -f(-a, -b, t) = (a/b)(eb’ - 1).
 LIE SERIES AND NLODEs 49

This now gives two formulas for the solution of

v’(t) = a + by(t), Y(O) = Y9

namely,
y(t) = dDetbyDy
= gDetby

= etb(y +f)
= yetb + (a/b)(etb - 1)
and
y(t) = etbyDegDy
= etbyD(y + g)
= (etbDy)+ g
= yetb + (u/b)(etb - 1).

This is an elementary illustration of what can be done with Lie series

when the commutator algebra of the terms in the differential equation is
finite dimensional. We can now make a change of variables or repeat the
above calculation in a more general setting to obtain the following results.

PROPOSITION.

f=f(a, b,t) = $ l- e-40t

P
and g = g(a, b, t) = -f(-a, 4, t).
ProoJ This is easy extension of the above discussion.

PROPOSITION. The solution of

y’(t) = uy(t) + byb+l(t), Y(O) = Y
50 STANLY STEINBERG

can be written
e”‘y
y(t) = (1 _ (b/a)($“’ - 1)~~)“’ *

Proof: This follows directly from the basic properties of Lie series as in
the above proposition.

6. Two TERM EQUATIONS, INFINITE CASE

We now look at problems of the form

y’(t) = ay”+ ‘(t) + by4+‘(t), Y(O) = Y

where
p>a>O.

The commutator corresponding to the two terms in the equation is

[ya+‘D,y4+‘D] = @?-a) y”+4+‘D.

It is easy to see that the commutator algebra of these two operators has as a
basis
iatjDt1
Y D, i>O, j>O

where i=O=sj= 1, j=O=xi= 1.

With some effort, it is possible to seefrom the noncommuting exponential
identities that we can write the exponential of (tay”’ ’ + thy”+ ‘)D as a
product of exponentials of multiples of the basis elements. Instead of
proceeding in this way, we use a more direct method to calculate the
required coefftcients. This method is a direct analog of the method used in
the previous section.
Becausep > a > 0, we know that

We also have

a + 2P < ia + $3,
provided j? < 2a and
i+jZ4, i#O, j#O.
 LIE SERIESAND NLODEs 51

However, without knowing more about a and /3, we do not know the
ordering of all of the number of the form ia +j@

THEOREM. Zf p > a > p/2 > 0, then

~,yn+4+1Def2y2~+0+lD~~y~+ZO+lD . ..
et(ayQ+‘+byfi+‘)D =
etaya+lD etbyL’+‘D
. . . e g,yn+2d+l,, g yZa+O+lD g ya+B+lD Ib@+lD toya+lD
= e2 e I e. e

where
f, = -(t2/2)@ - a) ab
f2 = (t3/6)(p - a) /3a2b
f3 = (t”/3)@ - a) aab*

and gi(t, a, b, a, p) = -.(t, -a, -b, a, p). If we only have /I > a > 0, then we
must drop f3 and g,. It is possible to compute, using only elementary
operations, as many terms in this expansion as we wish.
Proof: Differentiate the proposed identity with respect to t and then
multiply on the right by the inverse of the proposed identity to obtain

Cayat1 + byD+‘)D = ay”+‘D

tayQ’t1Dby4+IDe--layQt’D
Se
tay”+‘DefbyO+‘Df; ya+~+lDe-‘by“+‘De-tycP~~D
+e
t . .. .

We will use more terms than written down here but it is clear what they
are. Next, expand each term using the Similarity Property,

eL1L2ewLl=L2 t [L,,L,]+j[L,, [L,,L,]] t ... .

and then retain only those terms with powers of y that interest us. Thus
(aya+’ t by4+‘)D = ay”+‘D
t b(y’+’ + t@ - a) ayat4+’ + tt’(J? - a)
)(~a2yZnf~t' + . ..)D

tf~(y"+"+' + t/lay2ai-4+’ + tabya+*O+‘)D

+fSY 2n+b+1D + . . .
+f;Y rr+2D+1D + . . .
t ... .
52 STANLY STEINBERG

Comparing powers of y gives us

t(p-a)ab+f;=O
~t2~-a)pu2+f~tpa+f~=o
f;tab+f;=O.

Thus
f; = -t@ - a) ub,
f; = ft’Q3 - a) @u’b
f; = t2(j3- a) aub2

and consequently we have the desired formulas by integration.

Let us now use these formulas to approximate the solutions of our
differential equations. First define

Next, given t, y, a, b, a, and j?, define the sequences

Yo=Y y. = a to = ta
Yl =wo9 YOTYO) YI =P t,=tb
Yz=W,9Y,9Yd Y2=a+P t2 =fi (t, a, b, a, P)
Y, = w,, Y2 9 Y2) y3=2a+p t, =f2(t, a, b, a, P)
Y4 = w3 9 73 9 Y3) y4 = a + 2P t, =f3(t, a, b, a, P)
Y, = w4 3Y4 3 Y4)

We now think of yi as a function of t, y,(t), and, as before, let y(t) be the

exact solution of the initial value problem. We define the errors by

ei(t) = I Y(t) - Ydtl-

Our next result will give an estimate for these errors. Before we give the
estimates we note that unless a and /? are integers, we cannot expect that the
solutions of the differential equation will be analytic functions of t or y.
However, the Lie series procedure does give us a series solution of the form
y(t) = C uUkyia+jq+ltk.
i9j.k
 LIE SERIES AND NLODEs 53

We will use the notation z-y ia+jSt ‘tk to indicate the lowest term in a series
of the above form.

THEOREM. For the errors defined above, we have

e,(t) cz ty4+’
e*(t) z ~‘JP+~+’
e,(t) r t3y2”+4+ ’
e,(t) z t3ya+25+1
e,(t) E t4y2a+24+1.

TABLE I"

t Y e, e2 e3 e4 e,

00 .lOOOOO .oOe+OO .oOe+OO .OOe+ 00 .OOe+OO .OOe+OO

.lO .101828 .824z- 03 .42e - 06 .38e - 08 .22e - 08 .41e - 10
.20 .103725 .17e-02 .17e-05 .32e - 07 .19e-07 .72e-09
.30 .I05697 .26e - 02 .41e - 05 .12e-06 .68e - 07 .40e - 08
.40 .107746 .36e-02 .77e - 05 .30e - 06 .17e-06 .14e-07
SO .109878 .46e - 02 .13e-04 .62e - 06 .35e - 06 .37e - 07
.60 .112099 .57e-02 .19e-04 .12e-05 .65e - 06 .83e - 07
.70 .114413 .69e - 02 .28e - 04 .2Oe - 05 .lle-05 .17e-06
.80 .116826 .81e - 02 .38e - 04 .32e - 05 .18e-05 .32e - 06
.90 .119347 .95e - 02 .51e-04 .49e - 05 .27e - 05 .56e - 06
1.00 .121981 .lle-01 .66e - 04 .72e-05 .39e - 05 .94e - 06

a a= 1.0, b= 1.0, a= 1.0,/I= 1.1.

TABLE II"

t Y 5 e2 e3 e4 es

SKI .010000 .ooe + 00 .OOt?+OO .OOt?+OO .OOe+ 00 .OOC?+OO

.lO .010016 .63e - 05 .32e - 09 .25e - 12 .13e- 12 .24e - 15
.20 .010033 .13e-04 .13e-08 .2Oe- 11 .lle- 11 .38e - 14
.30 .010049 .19e-04 .29e - 08 .68e - 11 .36e - 11 .20e - 13
.40 .010066 .26e-04 .5le-08 .16e - 10 .87e- 11 .62e - 13
.50 .010082 .32e-04 .81e - 08 .32e - 10 .17e - 10 .15e- 12
.60 .o 10099 .38e - 04 .12e-07 .56e - 10 .3Oe - 10 .32e- 12
.70 .010116 .45e - 04 .16e-07 .89e - 10 .47e - 10 .6Oe - 12
.80 .010132 .52e - 04 .21e-07 .13e-09 .71e - 10 .lOe- 11
.90 .010149 .58e-04 .27e - 07 .19e-09 .lOe-09 .17e- 11
1.00 .010166 .65e - 04 .33e - 07 .26e - 09 .14e-09 .25e- 11

' a= 1.0, b= 1.0, CI= l.O,D= 1.1.

54 STANLY STEINBERG

Proof: Observe that, in the proof of the previous theorem, we collected

all of the required lower order terms.
Note that becausewe cannot estimate the coefficients of the leading terms
in the errors, we cannot infer that the expansion of the solution of the
differential equation using Lie series converges. However these estimates do
give an accurate picture of how well the composition product expansions
represent the solutions of the differential equation, as the numerical results
presented in Tables I and II indicate.
We also note that when we reverse the order of the composition product
we obtain a similar degree of approximation. Before we leave this subject we
want to look at two limiting cases.First, we note that if a = /I, thenf, = f2 =
f3 = a.. = 0 and our formulas are exact. This was the reason for choosing
,8z a in our computer runs. Next, if a = 0 then the expansion becomes

erayDexp tb-$?ub+$~‘n2b+ ..a j yn+‘Dj.

The second exponent is nothing but the power series expansion of the closed
form formula that we found in the previous section.

7. GENERAL EQUATIONS

We begin by trying to discover which equations of the form

Y’(f)= 2 .fxY’(Q)
i=l

have associated finite-dimensional commutator algebras. Thus let us assume

that the set of operators

.fXyP, l<i<m,

is closed under commutation. This means that

[.fXrP, &(YPI = 2 c:f,(~P

k=l

for some constants c$, 1 < i, j, k < m.

We need to assumethat all of the functions& are analytic near some point
y = a. By translation, we may assumethat a = 0. By relabeling the functions
and perhaps a further translation, we may assumethat
 LIE SERIES AND NLODEs 55

If we now make the change of variables u = g(y) where g’(y) = l/f(v),

then we will obtain a problem where f,(y) s 1. Now subtract a multiple of
the first operator, D, from each of the operators A(y)D so that A(O) = 0,
2<i<m.
We can now relabel things so that&(O) = 0 and f;(O) # 0. If not, then all
&(y), k > 2, have a multiple zero at y = 0. Choosef*(y) to be the function
with a zero of lowest multiplicity. Now

must also be in the algebra and has a zero of lower multiplicity thanf,. Thus
f,(y) must have a zero of multiplicity one. Now multiply fiO by a constant
so that the second operator is in the form yg,(y)D with g2(0) = 1.
By repeating this argument we can show that we may choose a new basis
for our operator algebra of the form

where I= m - 1 and g,(O) = 1, 1 Q i < 1. Now the commutations are given

by

b’g,(W, y’- ‘g,-,(yPl = Y*‘-*WP

where h(y) is some function. However, y*‘-*h(y) must be in the algebra, so
we must have

21- 2 < 1,

that is,

Thus, any algebra of the type we are considering can be at most three dimen-
sional and if it is three dimensional then a basis can be written in the form

where g(0) = 1 and h(0) = 1.

The commutators of these operators are given by

Pvy@l = (g + WP
[D, y*hD] = y(2h + yh’)D
[ ygD, y%D] = y*(hg + ygh’ - &‘)D.
56 STANLY STEINBERG

These commutators must be linear combinations of the given basis. A check

of the coefficients of various powers of y shows that we must have

g+yg= 1 +ayg+by2h,
2h+yh’=2g+cyh,
hg + ygh’ - yhg’ = h,
for some constants a, b, and c.
If we solve the second equation for

2g=2h+yh’-cyh

and then differentiate this and use these formulas to eliminate g from the
equation obtained from the third commutator, then we get
y2hh” -y*(h’)* - 2h(h - 1) = 0, h(0) = 1.

If we divide this equation by y and let y appraoch zero, then we see that

h’(O) = lim h(yie l = 0.

y-0

Now divide the differential equation by y* to obtain

h--l
hh” - (h’)* - 2h yz =O h(0) = 1, h’(0) = 0.
3

This equation is singular for arbitrary initial values; however, for our
particular initial values it is regular and consequently has a unique power
series solution. However,
h(y) E 1

is a solution and so must be this unique solution. We also have

g = 1 - cy/2.
If we now subtract a multiple of the third operator from the second operator,
we end up with the basis D, yD, y2D.
A similar calculation can be carried out in the two dimensional case to
give us the following result due to Lie [25, 261.

THEOREM. Any fmite-dimensional commutator algebra of first order

dlperential operators in one variable has dimension at most three. Up to
 LIE SERIES AND NLODEs 57

change of coordinates any such algebra can have its basis chosenfrom D,
YD, Y’D.
Proof. This follows from the above discussion.
Now we can write the solution of

y’(t) = a + by(t) + cy’(t), Y(O) = Y9

in closed form using Lie series (as well as any equation that can be
transformed into this equation). We do not carry out the solution explicitly
but we will record the necessary exponential formulas to do this. We note
that the above equation is a constant coefftcient Riccati equation.

PROPOSITION.
t(ayI-=tby+cY'+9D __ efy’-QDegyDehy’-“D
e

wheref = f (t), g = g(t), h = h(t) satisfy the following initial value problems:
f’ = a*cf’ - abf + a, f(0) = 0
g’ = b - 2acf, g(O) = 0
h’ = cepag, h(0) = 0.
Proof. This goes just like the proof in the previous section. Note that f
can be computed using integration and then g and h are given by integrals.

8. EQUATIONS DEFINED BY POWER SERIES

We now turn to the study of problems of the form

Y'W =f(y(t)), Y(O)= YT

where f(0) = 0 and f is analytic, that is,

f(Y) = 2 akYk.
k=l

The commutator algebra is now infinite dimensional and has as a basis

ykD, l<k<oo.
THEOREM. If

f(Y)= f akYk9 a,ZO,

k=l
58 STANLY STEINBERG

then
etfC~‘D = eg,yDeg,y2Deg,r’Deg4y4De~sY5D . .,

where

g,=z (1- epatt)

g,=$ (1 -e-zalt)
1

g4 =
aI a4 + a2a3 (1 _ e-Ja”’ - E?$ (1 - e-w)
3af
u~f.2, + 2a,a,a, t aia3 (1 - ee401r)
g, =
4ai
2
-2 v2a4 + a2a3 (1 _ e-3a,t)
3ai

ah
t 2a: (1 - e-za1t).

We can compute as many terms in the expansion as we wish using only

elementary means.
ProoJ We write
,tfW = ew’Deszy2D ...
>

differentiate with respect to t, and then multiply on the right by the inverse of
the proposed identity to obtain

f(yP’ = g; YD
+ eglfDg;y2De-glYD

g,YDew2Dg; y3De-gzY2De-g,YD
+e
+ . ..*

Becausethe commutation by yD does not increase the power of y in the coef-

ficients of operators, we handle terms involving this operator in a special
way. First we write the above equation in the form
epglyDf(y) DeglyD = gi yD + gi y2D
+ @zY’Dgj y3De-g2YzD + ... ,
 LIE SERIES AND NLODEs 59

Next write

= fG-9) yD
e-81Y
= egff(e-gly)D

and

= &(e g2y2Dy)y2D.

Complete the Similarity Property expansions to get

2 ake(‘-k)glykD = g; yD + g; y2D
k=l

+g;(y+g,y2+g:y3+~~~)y2D
+ g;(y* + 2g, y3 + ..a) y*D
+ &(y3 + -.-) y*D
+ **a.

Comparing the coeffkients of the powers of y gives us

g1=a,
gi = a3eezfl
g; = a,e-‘*l
g;+g,g,=a,e -3&T,
g; + 2g: g, + g; g: = a, e p4gl.

We can now compute the desired formulas by integration.

As in the previous section, we can use this expansion to approximate the
solution, y(t) of the differential equation. However, there is one important
difference between this case and the previous case. Because f(y) is an
analytic function, the solution y(t) has a convergent power series expansion
of the form

y(t) = f a,t’y.
i,j=O
60 STANLY STEINBERG

If we define F(t, y, y) as above, g,(t) as in the previous theorem,

YjCt) = F(t9 Y9Y)? YdO = Y

and

ej(t) = I Y(t) - Yj(tl9

then we obtain the following result.

THEOREM. Using the above notation and the hypothesesof the previous
theorem, we obtain
ej(t) E tti+ ‘.
Proof The initial conditions give us eJ0) = 0. The gr(t) are chosen so
that

& (y(t) - yj(t)) = f rt(t) yiDe’f’y’Dy.

i=jtl

Here r[(t) is to be determined, but it is easy to check that, in general, this

function will not be zero at t = 0.
We implemented the previous formulas to approximate the solution of

y'(t) = y(t) + 2y2(t) t 3y3(t) t 4y4(t) t 5YV), Y(O) = Y*

We give te results of three typical computer runs in Tables III-V.
Note that the results in the previous theorem accurately predict the depen-
dence of the errors on t and y in the tables. The dependenceof e, on t should

TABLE III

t Y 6 e2 e3 e4 5

.oo .010000 .oOe+OO .ooe+oo .ooe+oo .OOe+ 00 .OOe+OO

.Ol .010103 .21e-05 .31e-07 .41e - 09 .51e- 11 .22e - 14
.02 .O10206 .42e - 05 .63e - 07 .84e - 09 .lOe - 10 .88e - 14
.03 .0103 11 .64e - 05 .97e - 07 .13e-08 .16e - 10 .20e - 13
.04 .010417 .86e - 05 .13e-06 .17e-08 .21e - 10 .37e - 13
.05 .010524 .lle-04 .17e-06 .22e - 08 .27e - 10 .59e - 13
.06 .010632 .13e-04 .21e - 06 .27e - 08 .33e - 10 .87e - 13
.07 .010741 .16e-04 .25e - 06 .32e-08 .39e - 10 .12e- 12
.08 .01085 1 .18e-04 .29e - 06 .38e - 08 .45e - 10 .16e- 12
.09 .010963 .21e - 04 .33e - 06 .43e - 08 .52e - 10 .21e - 12
.lO .011075 .24e - 04 .37e - 06 .49e - 08 .58e - 10 .26e- 12
 LIE SERIES AND NLODEs 61

TABLE IV

t Y el e2 e3 6
---
SKI .Oloooo .ooe+oo .oOe+OO .OOe+OO .OOe+ 00 .OOe+OO
.I0 .011075 .24e - 04 .37e - 06 .49e - 08 .58e - 10 .29e - 12
.20 .012269 .55e-04 .93e - 06 .12e-07 .14e-09 .14e- 11
.30 .013595 .97e - 04 .17e-05 .22e - 07 .26e - 09 .39e - 11
.40 .015069 .15e-03 .29e - 05 .35e - 07 .43e - 09 .88e - 11
.50 .016708 .22e - 03 .45e - 05 .55e - 07 .72e - 09 .18e - 10
.60 .018533 .31e - 03 .68e - 05 .82e - 07 .12e-08 .34e - 10
.70 .020564 .43e - 03 .99e - 05 .12e-06 .2oe - 08 .64e - 10
.80 .022829 .57e - 03 .14e-04 .18e-06 .35e - 08 .12e-09
.90 SO25356 .76e - 03 .21e - 04 .26e - 06 .6Oe - 08 .21e-09
1.00 .028 180 .lOe-02 .29e - 04 .37e - 06 .lle-07 .39e - 09

TABLE V

t Y el e2 e3 e, e5
.oo .lOOOOO .ooe+oo .OOe+OO .OOe+OO .OOe+OO .oOe+OO
.lO .113344 .28e - 02 .45e - 03 .59e - 04 .63e - 05 .32e - 06
.20 .I29003 .69e - 02 .12e-02 .16e-03 .16e-04 .lEe-05
.30 .147584 .13e-01 .24e - 02 .33e - 03 .33e - 04 .61e - 05
.40 .169956 .21e - 01 .45e - 02 .62e - 03 .64e - 04 .17e-04
.50 .197414 .33e - 01 .8Oe - 02 .12e-02 .13e-03 .44e - 04
.60 .232034 .50e - 01 .14e-01 .22e - 02 .26e - 03 .12e-03
.70 .277449 .76e - 01 .25e - 01 .44e - 02 .61e - 03 .32e - 03
.80 .340959 .12e+OO .46e - 01 .97e - 02 .16e - 02 .loe - 02
.90 .441o40 .2oe+oo .94e - 01 .25e - 01 .54e - 02 .45e - 02
1.00 .655285 .38e + 00 .24e + 00 .99e - 01 .31e-01 .48e-01

be better than predicted because the differential equation has only five
nonzero terms. We believe these results are good and indicate why there is
an interest in using these expansions in magnetic lens and in geometric optics
calculations. We also observe that if we expand each g, in powers of t, then
we get
gkza,t
as the noncommuting exponential formulas predict.
Let us now compare this result to the expansion in Section 6. Thus, in
Section 6, we take a =O, /3= 1, a = a,, b=a, and find that
62 STANLY STEINBERG

Thus the method of Section 6 looks like a truncated power series expansion
of the results in this section. Also note that if we set Q, = 0 in the formulas
of the previous theorem, then (by L’Hospital’s rule) gj = a#. This represents
only the first term obtained by applying the noncommuting exponential iden-
tities.

ACKNOWLEDGMENT

I would like to thank John Filcoff for bringing a problem of the form studied in this paper
to my attention during a seminar I was giving at the Air Force Weapons Laboratory at
K&land Air Force Base.

REFERENCES

1. R. K. BANSALAND R. SUBRAMANIAN,Stability analysis of power systems using Lie series

and pattern-recognition techniques, Proc. Inst. El&r. Engrs. 121 (1974), 623-629.
2. R. A. BROUCKE,A FORTRAN-4 system for the manipulation of symbolic Poisson series
with applications to celestial mechanics, preprint, Institute for Advanced Study in Orbital
Mechanics, University of Texas at Austin.
3. J. R. CARY AND A. N. KAUFMAN, Ponderomotive effects in collisionless plasma: A Lie
transform approach, Phys. Fluids 24(7) (1981), 1238-1250.
4. J. R. CARY, Lie transforms and their use in Hamiltonian perturbation theory, preprint,
Lawrence Berkeley Laboratory, 1978.
5. G. H. CRISTEA,The study of the operating mechanism of Lie series to solve the forward
Kolmogorov equation, Technical Report, Inst. Atomic Physics, Bucharest.
6. A. DEPRIT, J. HENRARD,J. F. PRICE, AND A. ROM, Birkhoffs normalization, Celestial
Mech. 1 (1969), 222-251.
7. R. L. DEWAR,Exact oscillation-centre transformations, .I. Phys. A 11 (1978), 9-26.
8. A. J. DRAGT, A method of transfer maps for linear and nonlinear beam elements, IEEE
Trans. Nuclear Sci. NS-26 (1979), 3601-3603.
9. A. J. DRAGT AND J. M. FINN, Normal form for mirror machine Hamiltonians, J. Math.
Z’hys. 20 (1979), 2649-2660.
10. D. R. DOUGLASAND A. J. DRAGT, Charged particle beam transport using Lie algebraic
methods, IEEE Trans. Nuclear Sci. NS-28 (1981), 2522-2524.
11. A. J. DRAGT, A Lie algebraic theory of geometrical optics and optical aberrations, J. Opt.
Sot. 73(3) (1982), 372-379.
12. B. GARFINKEL,A. JUPP,AND C. WILLIAMS,A recursive von Zeipel algorithm for the ideal
resonance problem, Astronom. J. 76 (1971), 157-166.
13. J. C. GARIUSONAND J. WONG, Perturbation theory for the variable parameter free
electron laser, preprint, Lawrence Livermore Laboratory, 1981.
14. D. M. G~UEGAND M. A. ABD-EL-NABY, Iterative solutions of nonlinear initial value
differential equations in Chebyshev series using Lie series, Numer. Math. 34 (1980)
1-13.
15. W. GROBNER,H. KNAPP, AND G. WANNER,Serie di Lie e loro applicaziom, Cremonese,
Rome, 1973.
16. W. GROBNER AND H. KNOPP, Contributions to the Method of Lie Series,
Bibliographisches Institut, Manuheim, 1967.
 LIE SERIES AND NLODEs

17. G. HORI, Theory of general perturbations, in “Recent Advances in Dynamical

Astronomy” (B. B. Tapley and V. Szebehely, Eds.), pp. 231-249, Reidel, Dordrecht,
1973.
18. R. A. HOWLAND,JR., Anaccelerated elimination technique for the solution of perturbed
Hamiltonian systems, Celestial Mech. 15 (1977), 327-352.
19. T. K. Hu AND J. C. ZINK, Lie series perturbation solution of the invariant embedding
reflection equations, Trans. Amer. Nuclear Sot. 18 (1974), 146-147.
20. T. JUILLEARAT, A three-dimensional multigroup p,-theory with axially variable
parameters, Nukleonik 12 (1969), 117-l 24.
2 I. J. KOIIMANIK AND C. C. Lr, On an algorithm for estimating uniform asymptotic stability
boundary of nonautonomous systems in “Proceedings, 1976 IEEE Conference on
Decison and Control Including the 15th Symposium on Adaptive Processes,” IEEE, New
York, 1976.
22. E. LEIMANIS, “The General Problem of the Motion of Coupled Rigid Bodies about a
Fixed Point,” Springer-Verlag, New York, 1965.
23. M. LEMANSKA,Exact solution of p, time-dependent equations with time-dependent cross-
sections for slab geometry, Z. Angew. Math. Phys. 26 (1975), 701-711.
24. M. LEMANSKA,Y. ILAMED, AND S. YIFTAH, The theory of a one-dimensionalbare reactor,
treated by the Lie series method, J. Nuclear Energy 23 (1969), 361-367.
25. S. LIE, “Sophus 1884 Differential Invariants Paper” (Translation by M. Acherman,
Comments by R. Hermann), Math. Sci. Press, Brookline, Mass., 1976.
26. S. LIE, “Sophus Lie’s 1880 Transformation Group Paper” (Translation by M. Ackerman,
Comments by R. Hermann), Math. Sci. Press, Brookline, Mass., 1976.
27. “MACSYMA Reference Manual,” Project MAC Mathlab Group, MIT, July 1977.
28. B. MCNAMARA, Super-convergent adiabatic invariants with resonant denominators by Lie
series transforms, J. Math. Phys. 19 (1978), 2154-2164.
29. J. MENNIG AND M. LEMANSKA,Exact solution of the p, time-dependent equations with
time-dependent cross-sections for slab and spherical geometry, Z. Angew. Math. Ph.vs. 25
(1974), 433442.
30. J. MOSER, “Stable and Random Motions in Dynamical Systems,” Hermann Weyl
Lectures, Princeton Univ. Press, Princeton, N.J., 1973.
31. R. PAVELLE,M. ROTHSTEIN,AND JOHN FITCH, Computer algebra, ScientiJic American
245, No. 6 (Dec. 1981).
32. V. P. PETRUK,The use of Lie series to study nonlinear Hamiltonian systems. Cosmic Res.
15 (1977), 8W803.
33. J. RAE, Subdynamics in classical mechanics, in “Recent Advances in Celestial
Astromomy” (B. B. Tapley and V. Szegehely, Eds.), pp. 262-269, Reidel, Dordrecht,
1973.
34. S. STEINBERG,Lie series and their applications, in preparation.
35. K. STUMPFF,On the application of Lie-series to the problems of celestial mechanics.
NASA Technical :;ote, NASA TN D-4460, 1968.
36. A. ZURKINDEN,Time dependent neutron transport theory in multigroup p,-approximation.
Z. Angew. Math. Phys. 28 (1977), 393-408.