Beruflich Dokumente
Kultur Dokumente
net/publication/266301429
CITATIONS READS
64 478
3 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by David J. Jeffrey on 15 June 2015.
found because of users discovering that Maple had used in the solution of a problem they
W
had posed. In addition to its applicability, the function has rich mathematical properties.
It is de
ned to be the solution ( ) of
W z
()
W z e
W (z )
= z (1)
where is a complex variable. The history, applications and properties of were recently
z W
reviewed in Corless et al.3], and so just two examples of the uses of are given here, in
W
is the evaluation of
zz
h z( )= zz
whenever it makes sense. The function ( ) can be found in closed form by solving the
h z
( ) = ; (log; log )
h z
W
(2)
z
:
z
For more details and references on the problem, see Baker & Rippon 2].
Another application is the solution of linear constant-coecient delay equations 8].
Consider the simple delay equation
_ ( ) = ( ; 1) y t ay t (3)
tution shows that exp( ( ) ) is a solution of (3). The starting condition can be matched
W a t
by noting that (1) does not de
ne uniquely. If complex values are considered, multiple
W
1
X
y = ck exp( Wk a t ( )) (4)
k= ;1
and the k can be determined to match ( ). One sees immediately that the solution will
c f t
grow exponentially if any of the k ( ) have positive real part, and this leads to important
W a
satis
es ; arg , and the principal branch of its (natural) logarithm is de
ned to
< z
implying that ln0 is another representation of the principal branch. (On the matter of
z
notation, notice that although logk means logarithm to the base , there is no standard k
an unspeci
ed branch. The range of each lnk in the complex plane is shown in
gure 1.
z
Plotted horizontally is the real part of the logarithm: < lnk = ln j j. The imaginary part, z z
= lnk = arg +2 , is plotted vertically. The points that form the boundary between two
z z k
branches belong to the region below them, because of the de nition of arg . This closure z
also agrees with the `counter-clockwise continuous' convention (CCC) of Kahan 7].
Figure 2 shows the complex range of each k . The horizontal axis is = < , the
W W
real part of the appropriate branch of , and the vertical axis is = = . The branch
W W
boundaries obey either = 0 or = ; cot and correspond to cuts in the plane along
z
portions of the negative real axis. The points on the boundaries belong to the branch
below them, which closure rule again satis
es the counter-clockwise continuous (CCC)
convention 7]. The negative real axis is divided at = ;1, with ;1 belonging to
branch 0 and ;1 belonging to branch = ;1. It should be noticed that the dashed
< k
asymptotes at odd multiples of coincide with the branch boundaries in
gure 1. This is
2
ℑ(Ln z)
3π
k=1
2π
π Principal
k=0
0
ℜ(Ln z)
Branch
−π
k = −1
−2π
−3π
Figure 1. The ranges of the branches of the multivalued logarithm Ln . The heavy lines
z
are the boundaries between the branches, and the value of for each branch is shown.
k
logarithm.
The current de
nition of the branches is a little long-winded, and it would be a great
convenience if it could be replaced, or at least supplemented, by an algebraic relation that
directly discriminates between the branches of . Section 3 derives such a relation. It is
W
more concise, as a de
nition, although the original de
nition is more graphical. To derive
the result, we have used a method currently being explored in connection with computer
algebra systems.
Computer algebra systems, and humans for that matter, face the problem of how
to handle the fact that in the complex plane ln( ) 6= ln + ln , when logarithms are
zw z w
interpreted as principal branch. Any scheme should be suitable for automatic operation.
One proposed solution is to transfer the problem to the design of the user interface to
the system, and not develop any new mathematics. This possibility has been discussed,
amongst other issues, in Corless & Jerey 4]. Although plausible in the computer-algebra
context, an interface solution does nothing to alleviate the mathematical diculty. Here,
a mathematical idea called the unwinding number is described and applied. Similar ideas
have been discussed informally, particularly in electronic forums such as newsnet, but have
not appeared in print, although an unnamed function similar to the unwinding number
made a cameo appearance in the book by Apostol 1, theorem 1.54].
2. Denition and properties of the unwinding number
The unwinding number K( ) is an integer function de
ned using the principal-value
z
3
η = ℑ(W)
5π
4π Branch k = 2
3π
2π Branch k = 1
π
Principal
ξ = ℜ(W)
Branch k = 0
−π
−2π Branch k = −1
−3π
Branch k = −2
−4π
−5π
Figure 2. The ranges of the branches of . Each branch is given a number, the principal
W
branch being numbered 0. The boundaries of the branches are asymptotic to the dashed
lines, which are horizontal at integer multiples of . The boundary curves obey either
= ; cot , for 2
n < j j (2 + 1) , or = 0 for ;1.
< n
Equivalent de nitions use the oor function or the round function. If = is the imag- z
Here, R is the round function with half-integers rounded down. Particular values of K are
8 1
when ;3 ; ,
>
< 0
< y
when
K( + ) = > ;1 when ; , < y
x iy
: 3 ,
< y
; when (2 ; 1)
n (2 + 1) . n < y n
Only those properties of K that are of immediate relevance will be explored here.
From (6), it is clear that K( ) = K( = ), and also
z i z
K( + 2 ) = K( ) ;
z in z n for an integer.
n (7)
This relation is very useful for removing one unwinding number from the argument of
another. The application of K in this paper is to the manipulation of logarithms, as
summarized by the following formulae.
ln( ) = ln + ln + 2 K(ln + ln )
zw z w i z w : (8)
ln( ) = ln ; ln + 2 K(ln ; ln )
z=w z w i z w : (9)
4
Both of these apply for any complex numbers and . The derivation of (8) is eected z w
K(ln ) = 0 (10)z :
While giving an example of working with K, we can derive a result that will be used
in section 4. If and are complex numbers for which j arg ; arg j , with arg the
a b b a <
a
: (11)
The derivation uses (8), (9) and (7).
ln = ln + ln(
b a b=a ) + 2 K (ln + ln( ))
i a b=a
The rst unwinding number is zero by (10) and the second by the condition on and . a b
( ) + ln ln
( ) = ln z for = ;1 and 2 ;1 0) ,
k z =e
(12)
Wk z Wk z
k z otherwise.
The proof takes logarithms of (1). Omitting the argument of Wk z ( ) for clarity, we have
ln = ln ( k exp k ) = ln
z W W Wk + ln exp k + 2 K(ln k + ln exp
W i W Wk )
= ln k + k + 2 K(
W W i W k ) + 2 K(ln k + k + 2 K( k ))
i W W i W :
this multiple is the same for all and , except possibly when is real and negative,
r W
5
because then arg is discontinuous. Putting the exceptions to one side, we identify
W
lim
!1(arg
r
Wk += Wk ) = rlim
!1 = + 2 k
arg k + = k = + 2 W W k : (15)
The special case = 0 requires = 0 and is trivial the special case = requires either
k
arg ;1 ( ) + = ;1( ) =
W z W z for ; 1 =e z < 0
4. Application
De Bruijn 5] obtained an asymptotic expansion for the real values of 0 ( ), for W x x
real, when ! 1. This was generalized to all branches of ( ) for complex in the case
x W z z
j j ! 1 by Corless et al. 3]. In both cases, the derivations were based on a quantity ,
z v
for which an equation was derived that was valid in the asymptotic region. Here, (12) and
(11) are used to obtain an equation for that is valid for all and all branches.
v z
Generalizing de Bruijn, we de ne k by v
n
ln lnk + k = 2 for = 0 and 2 (0 1),
k + ln 1 ; (17)
z v i k z
v
lnk lnk z z 0 otherwise.
The proof treats separately the three cases arg k = 0 and arg W Wk not equal to these
values. If arg 6= 0 , then substituting (16) into (12) gives
W
lnk ; ln lnk +
z z vk + ln(lnk ; ln lnk + k ) = lnk
z z v z :
For all branches, arg k and arg lnk have the same sign, and so (11) applies, and results
W
in (17).
The special case arg k = 0 implies = 0 and is real and positive, and checking
W k z
this case is straightforward. Finally, if arg = , then can be 0 or ;1, and, for eitherW k
value of , ;1
k =e 0. Since = ;1 is a special case in (12), one might expect to see
z < k
a separate line for it in (17), but a factor 2 is introduced because of the two conditions
i
arg ;1 = but arg ;1 0, and this factor cancels the same factor in (12).
W L <
6
Exponentiating (17) casts it in a more familiar form due to de Bruijn.
;v ; 1 + ln lnk ;
lnk = 0 (18)
z v
e
lnkz z
:
From (17), ; = , and any solution of (18) must satisfy this constraint to be
v
relevant. The last equation has been the starting point for several series expressions, both
asymptotic and convergent, for the Lambert W function 6].
References
1] T.M. Apostol (1974) Mathematical Analysis, 2nd Ed. Addison-Wesley, Reading MA.
2] I.N. Baker and P.J. Rippon (1985) A note on complex iteration. Amer. Math. Monthly,
92, 501{504.
3] R.M. Corless, G.H. Gonnet, D.E.G. Hare, D.J. Jerey and D.E. Knuth (1996) On the
Lambert function. Adv. Comput. Maths, in press.
W
4] R.M. Corless and D.J. Jerey (1992) Well, it isn't quite that simple. SIGSAM Bulletin,
26(3), 2{6.
5] N.G. de Bruijn (1961) Asymptotic Methods in Analysis, North-Holland, Amsterdam.
6] D.J. Jerey, R.M. Corless, D.E.G. Hare & D.E. Knuth (1995) Sur l'inversion de yy e
au moyen de nombres de Stirling associ!es. C. R. Acad. Sc. Paris, Serie I, 320, 1449{
1452.
7] W. Kahan (1986) Branch cuts for complex elementary functions. In The State of the
Art in Numerical Analysis: Proceedings of the Joint IMA/SIAM Conference on the
State of the Art in Numerical Analysis, University of Birmingham, April 14-18, 1986,
Edited by M. J. D. Powell and A. Iserles, Oxford University Press.
8] E.M. Wright (1949) The linear dierence-dierential equation with constant coe-
cients. Proc. Roy. Soc. Edinburgh, A62, 387{393.