Anda di halaman 1dari 10

A Sequence of Series for The Lambert W Function

Robert M. Corless1 , David J. Jerey1 , and Donald E. Knuth2


1 Department of Applied Mathematics
University of Western Ontario
London, Canada N6A 5B7
frcorless,djjg@uwo.ca

2 Computer Science Department,


Gates 4B,
Stanford University,
Stanford, CA 94305-9405 USA

Abstract these results to be of direct interest for computer algebra re-


searchers, partly because of their combinatorial applications
We give a uniform treatment of several series expansions for but also because of their intrinsic value. A further point of
the Lambert W function, leading to an innite family of new interest to computer algebraists is that in order to repre-
series. We also discuss standardization, complex branches, a sent these innite series in Maple it is necessary to extend
family of arbitrary-order iterative methods for computation its knowledge of several special families of numbers (specif-
of W , and give a theorem showing how to correctly solve ically Stirling cycle and subset numbers, associated Stirling
another simple and frequently occurring nonlinear equation numbers, and second-order Eulerian numbers).
in terms of W and the unwinding number. We also present some innite products for W (z), some
sequences converging to W (z), the Laplace transform of
1 Introduction W (exp(1 + z)) and the Mellin transform of W (z).
Investigations of the properties of the Lambert W func- 1.1 Denitions
tion are good examples of nontrivial interactions between
computer algebra, mathematics, and applications. To begin A review of the history, theory and applications of the Lam-
with, the standardization of the name W by computer alge- bert W function may be found in 5]. The many-valued
bra (see section 1.2 below) has had several e ects. First, this function W (z) is dened as the root of
standardization has exposed a great variety of applications

second, it has uncovered a signicant history, hitherto unno- W (z)eW (z) = z : (1)
ticed because the lack of a standard name meant that most
researchers were unaware of previous work
and, third, it has One may issue the Maple command
now stimulated current interest in this remarkable function. > plot(  w*exp(w), w, w=-4..1] )
Further, many of the recent investigations have been carried
out themselves using computer algebra, often forcing further to see a graph of W (x) for real x. For x  0 there is only one
development in computer algebra. This process has not yet real branch, but for ;1=e  x < 0 there are two. Complex
reached its nal state with regard to W . branches are discussed briey in section 1.3 of this paper.
Series expansions for W about various points play im- In this paper we also discuss series for the Tree function
portant roles in many studies and applications of W . Most T (z) = ;W (;z), which satises
basically they give starting values for numerical computation
of W
further, they were one of the main factors in deciding T (z)e;T (z) = z : (2)
the branch cuts for W
and in combinatorial applications
series about the origin and about the branch point are im- This form of the function often occurs in combinatorial ap-
portant. Although several series for W have already been plications.
published, it has not been realized that there are systematic
ways of developing them. 1.2 A note on notation
We here give a uniform treatment of several series expan-
sions for W (sections 2.3, 2.4, 4.4, 4.6, 4.8 and 4.11 in par- The notation W (z) and the name `the Lambert W func-
ticular contain results not previously published). We believe tion' have quickly become a standard, since the introduction
of W into Maple sometime in the 1980's and the publica-
tion of 5] in its technical report form (January 1993). As
of Maple V Release 4 the name of the function in Maple is
Intended for ISSAC '97, Maui LambertW. The reasons for naming the function after Lam-
bert are detailed in 5]. Briey, they are, rst, that Euler
clearly acknowledged Lambert in his paper \On a series of
Im(W)

2 Taylor series
4π Branch k= 2 The Taylor series for W (z) about z = 0 has been known
3π since Euler's paper 9]. The series can be derived very simply
2π Branch k= 1 using the Lagrange Inversion Formula (see e.g. 3]). How-
ever, it is of historical interest to note that Lambert's deriva-
π
tion of his series pre-dates the Lagrange Inversion Formula.
We here use more modern tools, as expounded in the very
Principal
Re(W)
elegant paper 17]. But before we give the full derivation,
Branch k= 0
−π

−2π Branch k= −1 note that the rst several terms may be computed very sim-
ply in a computer algebra system by the commands (here
−3π
Branch k= −2
written in Maple for convenience)
−4π
> series(w*exp(w)-z, w)
;z + w + w + 12 w + 16 w + 241 w + O(w ) :
−5π
2 3 4 5 6

Figure 1: The ranges for the branches of the W function. > solve(", w)
Closure information is not indicated on the gure. Branches
are closed on their top boundaries. This means in particu- z ; z2 + 32 z3 ; 38 z4 + 12524 z + O(z )
5 6

lar that only branch 0 and branch ;1 take on real values.


The curves other than the semi-line (;1 ;1] are described To get the complete series, we note that if f (w) =
parametrically by ; cot  + i, ;1 <  < 1. Incidentally, w(w) = w exp w then (w) = exp w and
they form a subset of the curves called `The Quadratrix of X k
Hippias' which can be used both to square the circle and (w)s = esw = sk! wk : (3)
trisect the angle. k 0
This means that the suite of polynomials of binomial
type 17] associated with f (w) is just
Lambert's and some of its signicant properties" 9], where k
Euler developed the Taylor series for ;W (;z) and identi-
ed its radius of convergence, and, second, that naming yet Pk (s) = sk! : (4)
another function after Euler would not be useful. The formal series for f ;1 (z) is in general 17]
The use of the letter W was more or less accidental X Pn (;1 ; n) n
in Maple but turns out to have some signicance in that f ;1 (z) = z
E. M. Wright was the rst person to do signicant work on n0
n+1 z (5)
the complex branches of this function, and used it both in
dynamical and combinatorial applications. so in our case, by using our known Pk (s) we get
In the interests of true standardization, we suggest that X n;1
any CAS implementing the function satisfying (1) use the W (z) = (;nn)! zn : (6)
name LambertW (with arguments z or k,z) and the branch n1
cuts and closures of 5]. For documentation and for use in
papers we recommend the notations W (z) and Wk (z). The ratio test establishes that this series converges if
We recognize that some designers may wish instead (or in jzj < 1=e.
addition) to implement the function satisfying (2), because We might instead try to derive the Taylor series directly,
the minus signs in the series (6) below are a distraction since a formula for the n-th derivative of W for n  1 is
in combinatorial applications. Use of a di erent name and known 5, p. 340]. The formula is
notation is perhaps justied in this case
we suggest T (z), dn W (x) = exp(;nW (x))pn(W (x)) 
Tk (z) (with the branches dened by Tk (z) = ;Wk (;z)), dxn (1 + W (x))2n;1 (7)
TreeT(z), and TreeT(k,z). Note that DEK has suggested
that Maple implement this function in addition to LambertW where the polynomials pn (w) satisfy the recurrence relation
but this has not happened yet. pn+1 (w) = ;(nw + 3n ; 1)pn (w) + (1 + w)p0n (w) : (8)
1.3 Complex branches of W However, nding pn (0) from this recurrence is more dif-
cult than the previous derivation. The initial polyno-
A detailed discussion of the complex branches of W can mial is p1 (w) = 1. The next few polynomials are, by
be found in 5]. The branches are denoted Wk (z) (the Maple, p2 (w2) = ;2 ; w, p33 (w) = 9 + 8 w + 2 w2 , p4 (w) =
exact Maple V Release 4 syntax is LambertW(k,z)) with ;64 ;2 36 w ;379 w ; 46 w , and p5 (w) = 974 w + 625 +
W0 (z) being the principal branch, satisfying W0 (0) = 0 and 622 w + 192 w + 24 w . A simple consequence of the re-
W0 (x) > 0 for x > 0. The only other branch to have real currence is that if W (a) = w is rational, then the Taylor
values is W;1 (x) which takes on values in (;1 ;1] for series of W about x = a has rational coecients, apart from
x 2 ;1=e 0). The ranges of the branches are indicated the factor annwhich occurs if a 6= 0. This uses the fact that
in Figure 1. exp(nw) = a =wn .
We will usually deal with the principal branch in this It is clear that W (a) will be rational if and only if a =
paper, except where otherwise noted, and will write W (z) r exp(r) for some rational r. But W can take on other simple
for W0 (z) when there is no possibility for confusion. values, and this will be the subject of an upcoming paper
by David Je rey and co-workers.

2
2.1 Some series for T (z) = ;W (;z) The number of mappings from f1 2 : : :  ng into itself
We have immediately that having exactly k component cycles is the coecient of yk in
tn (y) where tn (y) is the tree polynomial of order n (see 16])
X nn;1 n and is generated by
T (z) = n! z (9) X
n1 1 = t n (y ) zn : (21)
(1 ; T (z))y n0 n!
but Lagrange inversion (or the suite of polynomials treat-
ment) gives more: Comparison with (18) gives for n  1
X n;1 n+x
T (z)x = x(n + xn) ! z (10) n;1
lim tn (y) = t0n (0) = Q(nn)n! : (22)
n0 y!0 y
(cf. 14, exercise 2.3.4.4{29]). In particular, for integer k we These series and others may be found in 16], where they
have are used to analyze a recurrence related to trees.
X n;k n Finally, note that using equation (13) with
T (z)k = k n (n ; 1)  n !(n ; k + 1)z : (11) g(n) = Q(1 2 32  : : :
n)
n1  
This gives a connection with `Q series' (see 14]): if = F 1 2 11 ; n   ;n (23)
g(n) = Q(a1  a2  a3  : : :
n) where the hypergeometric function F is written using the
notation of 10], allows us to write series expansions for any
= a1 + a2 n ; 1 n;1n;2
n + a3 n n +     (12) rational function of T (z), by rst expanding in partial frac-
tions.
then
X g(n)nn;1 z n 2.2 Taylor series for W (exp x)
a1 T (z) + 21 a2 T (z)2 + 13 a3 T (z)3 +    = n! :
n1 The equation
(13) y + ln y = z (24)
This follows on substitution of equation (11) and rearrang- often occurs in applications. Its solution in terms of W is
ing. Since Q(1 2 3 : : :
n) = n, an immediate consequence (ignoring branches for the moment) y = W (exp z), which in
is some ways is a nicer function than W (z). For one thing, its
1 = X nn z n : (14) derivatives are slightly simpler, and without much diculty
1 ; T (z) n0 n! one can establish by induction that for n  1
Similarly, Q(2  12  3  22  4  32  : : :
n) = 2n2 and hence dn W (ez ) = qn (W (ez ))  (25)
X nn+1 z n dzn (1 + W (ez ))2n;1
T (z)
(1 ; T (z))3 = n1 n! : (15) where the polynomials qn (w) are given by
n;1
X 
Note also that (1+x=n)n;1 = Q(1 x=1! x2 =2! : : :
n), which qn (w) = n;1 (;1)k wk+1 : (26)
yields the identity k=0
k
X n;1 n
exT (z) = x(n + nx!) z : (16) These polynomials have coecients expressed in terms of
second-order Eulerian numbers 10]. See section 4.1 for re-
n0 currence relations for these numbers. The qn (w) may be
We have essentially seenx this before (in equation (10)), since computed from the recurrence relation 5]
exp(xT (z)) = (T (z)=z) . In its new form it makes sense for qn+1 (w) = ;(2n ; 1)wqn (w) + (w + w2 )qn0 (w) : (27)
complex x.
Let Q(n) = Q(1 1 1 : : :
n) be Ramanujan's function Discovering formulas (25{26) in the rst place is another
matter, of course, and it is here that computation of the rst
Q(n) = 1 + n ; 1 n;1n;2
n + n n + : (17) several derivatives can point the way.
Then > alias( w = LambertW( exp(z) ) )
> seq( normal(diff( w, z$n)), n=1..5)
X nn;1 Q(n)z n
ln 1 ; 1T (z) = n! (18)
n1
X nn Q(n)zn  1 +w w  (1 +ww)3  ; w ((1;+
1 + 2 w) 
w)5
T (z) = (19)
(1 ; T (z))2 n1
n! w (1 ; 8 w + 6 w ) 
2

(1 + w)7
T (z) + T (z)2 = X nn+1 Q(n)zn : (20) ; w (;1 + 22(1w +; w58)9w + 24 w ) ]
2 3
(1 ; T (z))4 n1
n!

3
As before, when w = W (exp a) is rational the Taylor Getting Maple to evaluate this integral requires explicit use
series for W (exp z) about z = a has rational coecients. of changevar and assumptions on s. The requisite substi-
For example, this occurs when a = 1, giving tution is w = W (exp(1 + z)) so dz = (1 + w)dw=w and
exp z = w exp(w ; 1).
W (ez ) = 1 + 12 (z ; 1) + 16
1 (z ; 1)2 ; 1 (z ; 1)3 ;
192 The e ect of the Laplace transform, as can be veried
1 (z ; 1)4 + 13 (z ; 1)5 + O((z ; 1)6 ) : (28) with Watson's lemma 1], is to convert an exponential gen-
erating function into an ordinary generating function (in
3072 61440 1=s
one can keep the same
R form of the power series by us-
p ing the Borel transform 01 exp(;t)f (st) dt, but this is only
This series has radius of convergence 4 + 2 , which is
the distance to the nearest singularity at z = ;1 + i . This trivially di erent).
gives the asymptotics of qn (1). Since we can evaluate the Laplace transform of W (exp z)
it is no surprise that we can evaluate the Mellin transform
2.3 Branches in equation (24) of W (z). The result is
Theorem: The unique solution of y + ln y = z is Z1
M (W (z)) = xs;1 W (x) dx
y = W;K(z)(ez ) (29) 0
;s
where K(z) is the unwinding number of z (see 6]), unless
z = t + i for ;1 < t  ;1, in which case there are exactly = (;s) s ;(s)  (34)
two solutions, y = W;1 (exp z) and y = W0 (exp z). for ;1 < <(s) < 0. Again explicit use of changevar and as-
Proof. Taking exponentials of both sides of y + ln y = z sumptions are necessary to get Maple to evaluate this trans-
we see that if y is a solution, then y = Wk (exp z) for some k. form.
To go in the other direction, we use the relation
Wk (z) + ln Wk (z) = ln z + 2 ik (30) 3 Series about the branch point
unless k = ;1 and z 2 ;1=e 0), when W;1 (z) + p
ln W;1 (z) = ln z. For a proof of this relation see 12]. We If we put p = 2(ez + 1) in W exp W = z, and expand in
replace z in the above by exp z, and since powers of 1 + W and revert, we obtain
ln ez = z + 2 iK(z) (31) X
W0 (z) =
` p` = ;1 + p ; 13 p2 + 72 p +  :
11 3
(35)
(indeed this denes the unwinding number K(z), see 6]), we `0
have that
Wk (ez ) + ln Wk (ez ) = z + 2 i (K(z) + k) (32) This series is well-known and has many combinatorial ap-
plications (see for example 16]). It converges for jpj < 21=2 .
unless k = ;1 and exp z 2 ;1=e 0), in which case we The coecients may be computed by the following recur-
replace k by 0 on the right hand side of (32). Thus we rence relations, which were communicated to us by Don
have that Wk (exp z) + ln Wk (exp z) = z if and only if Coppersmith.
k = ;K(z), unless k = ;1, as claimed. It is easy to see
that both k = ;1 and k = 0 work if z = t + i for some

k = kk ; 1 ;1
t 2 (;1 ;1] as claimed, and because the unwinding num-  1
+ 1 2
k;2 + 4 k;2 ; 2 k ; k+1
k;1
1 1

bers for z = t + i(2m + 1) are all di erent from 0 if m 6= 0,


this half-line is the only set of exceptions. k;1
X
Remark. It is very interesting that we may write the k =
j
k+1;j  0 = 2 1 = ;1  (36)
simple formula (29) for a (nearly) single-valued function y(z) j =2
in terms of a (very) multi-valued function Wk , using another
simple function of z to deftly switch branches as required. where
0 = ;1 and
1 = 1. p
This clean formulation, valid except on a single half-line in If =(z)  0 we may take p = ; 2(ez + 1) in this series
the z-plane, conrms that the branch choices made in 5] to get a good approximation to W;1 (z). If instead =(z) < 0,
(which were done so some asymptotic series worked out then the series (with the negative sign on p) gives a good
nicely) were convenient ones. approximation to W1 (z). This `branch splitting' is a result
of the branch choices for W , and is the price we pay for
2.4 Laplace and Mellin transforms convenient asymptotic expansions at 0 and 1.
The series (28) is quite remarkably connected with the
asymptotic expansion for `Airey's convergence factor' (which 3.1 Branch point series for W (exp z)
was used in 20] to improve convergence of the asymptotic The paper 19] makes a delightful connection between Stir-
series for the exponential integral. See 2] for details). This ling's approximation for n! and W . That there is a connec-
is essentially because the Laplace transform of W (exp z) can tion was pointed out to RMC and DJJ some years ago by
be evaluated in terms of the incomplete Gamma function as Bruno Salvy, who said in an e-mail:
follows:
Z1
;  It is possible to use this to prove Stirling's formula.
L W (e z ) =
1+
e;sz W (e1+z ) dz One rst computes the local expansion of W at its
0
singularity by
= 1s + es ss;2 ;(1 ; s s) : (33)

4
> op({solve(series(subs(w=-1+u, z=-exp(-1) and the series for T0 is the same but with alternating signs.
+ t^2* exp(-1),w*exp(w)-z),u),u)}) This gives at last that
p X
n! = 2 nnn e;n 1  3  5  n k(2k + 1) a2k+1 : (44)
p p 3 43 4 769 p 5
2 t ; 32 t2 + 11
36 2 t ; 135 t + 4320 2 t
k 0
1768 t6 + O(t7 )  The following Maple procedure implements the recur-
; 8505 rence relation derived in 19] for the ak .
p p 3 43 4 769 p 5
; 2 t ; 32 t2 ; 11
36 2 t ; 135 t ; 4320 2 t
a := proc(n) option remember
local k
; 1768
8505 t + O(t )
6 7 1/(n+1)/a(1)*(a(n-1) -
add( k*a(k)*a(n+1-k), k=2..n-1))
end:
where t = (1 + ez)1=2 . Now, since the singularity a(0) := 1
is isolated, it is possible to use Darboux's theorem a(1) := 1
which says that the asymptotic expansion of the
n-th Taylor coecient of W at the origin is ob- The choice a(1) = 1 gives the series for v, whilst a(1) = ;1
tained by taking the coecients term by term in gives the series for u. In more mathematical notation, their
the above equation. In other words (considering recurrence relation is
only the rst order term):
a0 = 1 (45)
(;n)n = (p;e)n (37) a1 =
1 (46)
n! 2 n
n;1 !
X
which is Stirling's formula. : : :] I heard it a few an = (n +11)a an;1 ; kak an+1;k : (47)
years ago from Philippe Flajolet. 1
k=2
The connection between series at the branch point for W The rst few terms are v = T;1 (e;1;z2 =2 ) =
and Stirling's formula for n! is made explicit in 19], who use
elementary arguments, and R we paraphrase their results be-
low. Starting from n! = 01 xn exp(;x) dx, changing vari- 1 + z + 13 z2 + 36
1 z3 ; 1 z4
270
ables to y = x=n and splitting the range of integration into 1 1
+ 4320 z + 17010 z ; 5443200
5 6 139 z7
0 1] and 1 1], they get
Z 1 Z1 + 2041201 z8 ; 571 z9
n! = nn+1 e;n (ue1;u )n du + (ve1;v )n dv : (38) 2351462400
0 1 281 163879 z11
; 1515591000 z10 + 2172751257600
They then make ad hoc substitutions
here we use our
knowledge of W and its branches to repeat their work. + O(z12 ) : (48)
Putting Since the nearest other singularity is at z = 1, this series
  converges for all nite z.
u = T0 e;1;z =2
2
and (39)
  3.2 Other branch point series
v = T;1 e;1;z =
2 2
 (40)
The relation between W;1 and W0 near the branch point
we arrive at was studied by Karamata in 13], who studied the coe-
Z1  cients in the power series
n! = nn+1 e;n ze;nz2 =2 1 ; 1
T;1 1 ; T0 dz : (41)
X
0
1 ;
= + 23 2 + 49 3 +    = cn n  (49)
Note that z = 0 corresponds to the branch point of T . We n1
use the T notation to keep minus signs to a minimum. We
being the solution to
now expand the expression in brackets about z = 0 (it is
more convenient to use the series for du and dv directly, as (1 +
)e; = (1 ; )e : (50)
was done in 19], than to rearrange the geometric series for
1=(1 ; T ) in (41)), integrate term by term, and arrive at the In terms of T , the solutions are
asymptotic expansion for n! in the form 8 ; 
X  k+1=2 < T0 (1 ; )e;(1;)
n! = nn+1 e;n (2k + 1)a2k+1 n2 ;(k + 1=2) (42) 1+
=
: T;1 ;(1 ; )e;(1;)
(51)
k 0
where X where if > 0 then T0 simplies to 1 ; and gives
= ;
T;1 (e;1;z2 =2 ) = an zn  (43) but T;1 does not simplify, whilst if < 0 then T;1 simplies
n0 to 1 ; but T0 does not. Karamata has developed a series

5
for the root that does not simplify. Using the di erential We also make use of r-associated Stirling subset numbers
equation  (we use only the case r = 2), which are generated by

d
= 1 ; 1  (1 +
)  (52) r ;1 k
d 1; (ez ;
X z )m = m! X n n o zn : (61)
it is easy to derive the recurrence relation k=0
k! n0
m r n!
" n;1 #
X Finally, we need the second-order Eulerian numbers. As
cn = n +1 1 2 + cj (1 ; jcn;j+1 )  (53) before we follow 10]. The numbers are dened as follows.
j =2
  
which is valid for n  2 if the sum is taken to be 0 if empty. n = 1 if n = 0 (62)
An interesting parametric description of the real n 0 otherwise
branches of W is contained in 18]. If we put  
n = 1 (63)
p = 12 (W0 (z) ; W;1 (z))  (54) 0
   
then 18] gives n = (k + 1) n ; 1
k k
z = ;pe;p coth p cosech p (55)  
n ;
+ (2n ; 1 ; k) k ; 1 :1 (64)
W0 (z) = ;pepcosech p (56)
; p
W;1 (z) = ;pe cosech p (57)
and expands these in series containing Bernoulli numbers. 4.2 The fundamental relation
The connection with Karamata's work is that If we start with W (z) exp(W (z)) = z, then by putting
p = ( +
)=2  (58) W (z) = ln z ; ln ln z + u (which denes u for any z 6= 0,
but which makes particularly good sense if z is large) we get
though it is not clear at this writing if this gives us any new
information. (ln z ; ln ln z + u) lnzz eu = z (65)
and since z 6= 0 we may divide both sides by z, and put
4 Series about Innity = 1= ln z and = ln ln z= ln z to get
We collect and present the previously known series for 1 ; + u = e;u : (66)
W (z) about innity, which go back to de Bruijn 8] and
to Comtet 4]. We then show that these known series are This is what we call the fundamental relation, and we will
members of an innite family of series (some of these were see some of its consequences in what follows. Note that this
exhibited in 5, 11]), and show that they are valid as z ! 0 same equation, mutatis mutandis, arises if we replace Wk (z)
as well as for z ! 1. Finally we exhibit some new se- by lnk z ; ln lnk z + u(k) , where lnk (z) = ln(z) + 2 ik is the
ries, including an expansion for W ( exp(it)) in terms of notation introduced in 12] for the k-th logarithm branch.
W ( ) and it which separates the real and imaginary parts
of W ( exp(it)). 4.3 de Bruijn and Comtet's expansion
4.1 Stirling and Eulerian numbers de Bruijn solved the fundamental equation (66) in series to
show that the asymptotic expansion of W (x) for large x is
To x notation, we summarize
  with generating functions. in fact convergent 8]. Comtet later identied the coe-
Stirling cycle numbers mn are generated (for example) cients explicitly as Stirling numbers 4]. The result is (ex-
by 10] plicitly adding branch information which Comtet did not
use, and specically excluding the case where k = ;1 and
X h n i zn z 2 ;1=e 0) where another expansion holds)
lnm (1 + z) = m! (;1)n+m m n! : (59)
n0
 
Wk (z) = lnk (z) ; ln lnk (z)
The numbers (;1)n+m mn are also called Stirling numbers X X (ln lnk (z ))m
+ c`m  (67)
of the rst kind. See 15] for a discussion of the advantages `0 m1 (lnk z)`+m
of the notation used here.    
The Stirling subset numbers mn , which are also called where c`m = m1! (;1)` ``++1m is expressed in terms of Stir-
Stirling numbers of the second kind, are generated (for ex- ling cycle numbers. This (absolutely convergent for large
ample) by X n n o zn enough z) doubly innite series can then be rearranged to
(ez ; 1)m = m! get
m n! : (60)
n0 Wk (z) = lnk z ; ln lnk z (68)
Recurrence relations (and much more) for these numbers X (;1)n X n h i m
+ (ln z)n (;1)m n ; m n (ln lnk z) :
may be found in 10]. k + 1 m!
n1 m=1

6
These series were further developed and rearranged inn v while the numerator is degree n ; 1, and because the
in 11] using the new variable  = 1=(1 + ) to get p terms grow more slowly). Thus this series again gives
the correct asymptotics. These observations were used in
Wk (z) = lnk z ; ln lnk z (69) section 4.3 to extend Comtet's series to arbitrary branches
X m mX ;1   of W .
+ m ! (;1)p+m;1  p+m p + mp ; 1  The following Maple session shows the rst few terms of
m1 p=0 2 this new series (75).
> basic := 1 - tau + sigma*u - exp(-u)
and a further series which we will discuss below.
basic := 1 ; + u ; e(;u)
4.4 A new series > subs(sigma=1/v,tau=-p/v,")
The fundamental equation (66) may be solved for u, for any
and , not just the values used here, in terms of W . The 1 + vp + uv ; e(;u)
solution is
u = W ( 1 e(1; )= ) ; 1 ; : (70) > series(",u)

If = 1= ln z and = ln ln z= ln z as we started with, then p 1 1 2 1 3 1 4 1 5


this equation gives no new information|apparently. Indeed,
when simplied, equation (70) becomes v + ( v + 1) u ; 2 u + 6 u ; 24 u + 120 u +
 
O(u6 )
W (z) = W 1 e(1; )= (71) > solve(",u):
> series(v + p + ",p):
which really only rewrites z using the new variables.
After some thought, we nd that if, instead of follow- > W(z) = map(normal,")
ing de Bruijn and using the variables and , we use the
variables v = 1= and p = ; = , giving W (z) = v + 1 +v v p + 12 (1 +v v)3 p2 ; 16 v ((1;1++v)25v) p3
; 
W (z) = W ln zeln z e; ln ln z = W (vev ep )  (72) 1 v (6 v2 ; 8 v + 1) p4 ;
+ 24 (1 + v)7
then our formulas become simpler and more comprehensible.
We will also need the unwinding number for W , namely the 1 v (24 v3 ; 58 v2 + 22 v ; 1) p5 + O(p6 )
function U` (v) given by 120 (1 + v)9
W` (vev ) = v + 2 iU` (v) : (73) This is a truncation of (75) and as predicted it contains
second-order Eulerian numbers.
One can show that U` (v) = K(v + ln v) + ` + O(1=v) is
asymptotically an integer. Moreover, if v is in the range of 4.5 Convergence
W` , then U` (v) = 0.
Finally, we must extend the formula for the n-th deriva- The series (75) converges, as a series in p, for all jpj less
tive of W (exp z) to the following: than the distance to the singularity at v exp(v + p^) = ;1=e.
 Because exp(2 ik) = 1, we have to nd the nearest such
dn W (aeaebz )  = bn qn (W (aea)) : singularity. The logarithmic singularity at z = 0 has been
dzn  a 2n;1 (74) moved to p = 1.
z=0 (1 + W (ae ))
This gives us the series expansion 4.6 An innite sequence of series
X 2 iU (v)) pn : Experimentation with the series (75) suggested many rear-
W (z) = (1 +qnv(v++2 i U (v))2n;1 n! (75) rangements (some of which are given in section 4.9). But
n0 our experiments also suggested that equations (72) and (75)
are invariant under the transformations
Remark. This series provides a foundation for an in-
nite number of series for W (z). We rst observe the remark- vn+1 = vn + pn
able fact that if v = ln z and p = ; ln ln z, then this series pn+1 = ; ln(1 + pn =vn ) : (76)
is asymptotic as z ! 1 (and also convergent for large z),
and also asymptotic as z ! 0 for all branches except the We take v0 and p0 to be the v and p already dened. This
principal branch. That is, the same series works for small z invariance was already observed in 11] (in terms of an itera-
and for large! Further, if v = ln(;x) and p = ln(; ln(;x)), tion for and ) but the full signicance was not then com-
then (75) also works for the exceptional case W;1 (x) as pletely understood. It is now clear that this iteration may
x ! 0; . Finally, we may x z and consider the asymptotics be run forward, for n ! 1, and backward, for n ! ;1.
of Wk (z) as jkj gets large. Using relation (30) as a guide, Running the iteration backward gives
we put v = lnk z and p = ; ln lnk z. We can show that the
unwinding numbers Uk (v) are zero for large enough v, and vn;1 = vn epn
that the terms in this series form an asymptotic sequence as pn;1 = vn (1 ; epn ) : (77)
z ! 1 (essentially since the denominator is degree 2n ; 1

7
In the case n ! 1, one can show that vn ! W (z) if and we will have W (z) = W (v exp(v + p)) so we may
jW (z)j > 1, while vn ! W (z) in jW (z)j < 1 if n ! ;1. use (75). This gives us many interesting new series for W ,
In fact this iteration turns out to be completely equiva- but it also gives us a family of iterative formulas for W ,
lent to the well-studied exponential iteration dening which can be of arbitrary order, as follows. We choose k = 0
because we want p ! 0.
zz
z


(78) Choose v0 to be an approximation to W (z). Then for any


vn , dene pn by equation (84), and compute an improved
and this connection is discussed in detail elsewhere 7]. For estimate vn+1 by
our purposes we note that this gives us an innite number N
of series for W (z), since vn+1 = vn + 1 +vnv pn + : : : + (1 +qNv(v)n2N) ;1 pNn! : (85)
n n
z = vn evn epn (79)
If N = 1, then the iteration is quadratically convergent (like
as can easily be established by induction. This gives us a Newton's Method). If N = 2, the iteration is cubically
family of bi-innite sequences of series convergent. In general if vn exp(vn ) = z(1 + ) then p =
X ln(1=(1 + )) = O() and so the error in vn+1 is O(N +1 ),
W` (z) = (1 +qkv(vn++2 i 2 iU` (vn )) pkn provided we are not too near the branch point z = ;1=e.
k 0 n U` (vn ))2k;1 k! (80) We do not pursue here the issues of appropriate initial
guesses v0 and optimal eciency by choice of N and the
for W (z). We may add a further innity of series by us- number of digits used.
ing di erent branches for ln, by putting v0(m) = lnm (z) and 4.9 Rearrangements
p(0m) = ; ln lnm (z). It is this sequence of series that gives
the title of this paper. Some rearrangements include the series from 11] below,
If jW (z)j > 1 then pn =vn ! 0 and vn ! W (z) as n ! 1, which use L = ln(1 ; ) and  = =(1 ; ).
and hence these series converge faster for larger n. If instead
jW (z)j < 1 then pn =vn ! 0 and vn ! W (z) as n ! ;1, Wk (z) = lnk z ; ln lnk z ; L
n h
and hence these series converge faster for more negative n.
Remark. If v0 = ln z and p0 = ; ln ln z, then v;1 = 1
X X
+ (;)n (;1)m n ; m n i Lm (86)
and p;1 = ln(z=e). This makes (80) particularly simple, n1 m=1
+ 1 m!
being a series of polynomials in ln(z=e). This gives
and
W (z) = 1 + 21 ln( ze ) + 161 ln( z )2 ; 1 ln( z )3 ;    (81)
e 192 e Wk (z) = lnk z ; ln lnk z ; L
X 1 m m mX ;1  
which is really just (75) with z = 1  exp(1) exp(ln(z) ; 1). + L  p + m ; 1 (;1)p+m;1 :
A more interesting series comes from v;2 = z=e and m!  p 2 (1 +  )
p+m
p;2 = 1 ; z=e. This new series contains only terms rational m1 p=0
in z, and is quite accurate for small and moderate z. The (87)
rst few terms are
The series converge for large enough real z, though the
z (1 ; ze ) z2 1 (1 ; ze )2 z3 detailed regions of convergence are not yet settled. Curi-
W (z ) = e ; 2 + ously enough (87) is exact at z = e and at z = 1, and
e (1 + ze ) 2 e3 (1 + ze )3 moreover if we truncate it to N terms it agrees with the N
e )3 z3 (;2 z2 + z ) term Taylor series expansion at z = e as well, making one
(1 ; think of `Hermite' interpolation at e and at 1. Convergence
; 16 z e2 e +    (82) is rapid.
e3 (1 + ze )5
4.10 Some Other Asymptotic Series
4.7 A completely di erent series The following are special cases of Theorem 2 in 21].
P
Using vn = nk pn , we may let n ! 1 and obtain the X (ln ln z )
completely di erent series ln W (z) = ln ln z + Bnln (88)
=0
nz
X n0
W (z) = ln z ; ln ln z ; ln(1 + pk =vk )  (83)
k 0 where the polynomials Bn are computable by B0 = 0,
Bn0 +1 (x) = nBn (x) ; Bn0 (x) and Bn (0) = 0.
which converges if jW (z)j > 1. Secondly, for any b and c,
X (ln ln z )
4.8 A continuous family of series and arbitrary- ebW (z) W (z)c = zb lnc;b z Cnln nz (89)
order iterative formulae n0
6 0 we can put
For any v = where the polynomials Cn are computable by C0 = 1,
  Cn0 +1 (x) = (b ; c + n)Cn (x) ; Cn0 (x) and Cn (0) = 0 for
p = ln zv + 2 ik 
ve (84) n > 0.

8
More generally, for any b and c and any power series 6 A Final Pair of Expansions
X
G(x) = ai xi (90) The iterations (76{77) may be used to show that W (z) can
i0 be written as
W (z) = z (98)
with a0 6= 0, the following expansion holds: exp z
  z
1 b c;b X An (ln ln z )
exp

ebW (z) W (z)cG W (z) = z ln z n0 lnn z (91)


exp
z
...
or
where the polynomials An are computable by A0 = a0 , z
A0n+1 (x) = (b ; c + n)An (x) ; A0n (x) and An (0) = an . W (z) = ln (99)
ln zz
4.11 Series for W ( exp(it)) ln
...
If y = W ( exp(it)) for  0 and ; < t  , then dene
v by according as jW (z)j < 1 or jW (z)j > 1. These curi-
W ( eit) = W ( ) + it + v : (92) ous formulae are just the iterated exponential in disguise,
and indeed are naturally discovered from rewriting W (z) =
Then W exp W = exp(it) implies z= exp W (z) and W (z) = ln(z=W (z)) as iterations.
(W ( ) + it + v)eW ()+it+v = eit (93) Acknowledgements
or, using W ( ) exp(W ( )) = , We thank Dave Hare and Bruno Salvy for useful comments
on earlier drafts. We also thank Gaston H. Gonnet for
1 + Wit( ) + W1( ) v = e;v : (94) putting W into Maple in the rst place.
But this is just (70) with = 1=W ( ) and = ;it=W ( ), References
and thus all of our fundamental series solutions to (70) ap-
ply! The nicest one is (75), which splits into separate series 1] Bender, C. M., and Orszag, S. A. Advanced
for the real and imaginary parts of W ( exp(it)). We have Mathematical Methods for Scientists and Engineers.
McGraw-Hill, 1978.
X n
W ( eit) = (Wq(n ()W+ (1) ))2n;1 (itn)! (95) 2] Borwein, J. M., and Corless, R. M. \Emerging
n0 tools for experimental mathematics". in preparation
(1997).
and clearly all the odd terms are purely imaginary and the 3] Caratheodory, C. Theory of Functions of a Complex
even terms are real. Variable. Chelsea, 1954.
5 Innite Products 4] Comtet, L. Advanced Combinatorics. Reidel, 1974.
From the relation W (z) = z exp(;W (z)) it is easy to see 5] Corless, R. M., Gonnet, G. H., Hare, D. E. G.,
that any series for W (z) may be trivially transformed into Jeffrey, D. J., and Knuth, D. E. \On the Lambert
an innite product. For example, from the series (6) we have W function". Advances in Computational Mathematics
5 (1996), 329{359.
Y n;1 n
W (z) = z exp( (;nn)! z) (96) 6] Corless, R. M., and Jeffrey, D. J. \The unwinding
n1 number". Sigsam Bulletin 30, 2 (June 1996), 28{35.
but this of course gives us no essentially new information. 7] Corless, R. M., and Jeffrey, D. J. \The dynamics
However, the series (83) gives us of the Lambert W function". in preparation (1997).
1
Y 8] de Bruijn, N. G. Asymptotic Methods in Analysis.
W (z) = ln(z) (1 + pn =vn ) (97) North-Holland, 1961.
n=0 9] Euler, L. \De serie Lambertina plurimisque eius in-
signibus proprietatibus", vol. 6 of Opera Mathematica
in terms of the iterates of (76), and this is a simpler and 1. 1921 (orig. date 1779), pp. 350{369.
more natural representation if nothing else.
10] Graham, R. L., Knuth, D. E., and Patashnik, O.
Concrete Mathematics. Addison-Wesley, 1994.
11] Jeffrey, D. J., Corless, R. M., Hare, D. E. G.,
and Knuth, D. E. \Sur l'inversion de y  ey au moyen
des nombres de Stirling associ#es". Comptes Rendus de
l'Acad
emie des Sciences, Paris 320 (1995), 1449{1452.

9
12] Jeffrey, D. J., Hare, D. E. G., and Corless,
R. M. \Unwinding the branches of the Lambert W
function". Mathematical Scientist 21 (1996), 1{7.
13] Karamata, J. \Sur quelques probl$emes pos#es par Ra-
manujan". J. Indian Math. Soc. 24 (1960), 343{365.
14] Knuth, D. E. The Art of Computer Programming,
2nd edition, vol. I. Addison-Wesley, 1973.
15] Knuth, D. E. \Two notes on notation". American
Mathematical Monthly 99 (1992), 403{422.
16] Knuth, D. E., and Pittel, B. \A recurrence related
to trees". Proc. Amer. Math. Soc. 105 (1989), 335{349.
17] Labelle, G. \Sur l'inversion et l'it#eration continue des
s#eries formelles". European Journal of Combinatorics
1 (1980), 113{138.
18] Lauwerier, H. A. \The asymptotic expansion
of the statistical distribution of N. V. Smirnov".
Z. Wahrscheinlichkeitstheorie 2 (1963), 61{68.
19] Marsaglia, G., and Marsaglia, J. C. \A new
derivation of Stirling's approximation to n!". Ameri-
can Mathematical Monthly 97 (1990), 826{829.
20] Murnaghan, F. D. \Airey's converging factor". Proc.
Nat. Acad. Sci. USA 69 (1972), 440{441.
21] Salvy, B. \Fast computation of some asymptotic func-
tional inverses". Journal of Symbolic Computation 17
(1994), 227{236.

10