Hyperasymptotics (Berry and Howl)

Hyperasymptotics
Author(s): M. V. Berry and C. J. Howls

Source: Proceedings: Mathematical and Physical Sciences, Vol. 430, No. 1880 (Sep. 8, 1990),
pp. 653-668
Published by: Royal Society
Stable URL: http://www.jstor.org/stable/79960
Accessed: 07-11-2016 05:44 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
Royal Society is collaborating with JSTOR to digitize, preserve and extend access to Proceedings:
Mathematical and Physical Sciences
This content downloaded from 139.133.124.87 on Mon, 07 Nov 2016 05:44:18 UTC
All use subject to http://about.jstor.org/terms
Hyperasymptotics
BY M. V. BERRY AND C. J. HOWLS
H. H. Wills Physics Laboratory, Tyndall Avenue, Bristol BS8 1TL,
We develop a technique for systematically reducing the exponentially small

('superasymptotic') remainder of an asymptotic expansion truncated near its least
term, for solutions of ordinary differential equations of Schrodinger type where one
transition point dominates. This is achieved by repeatedly applying Borel summation
to a resurgence formula discovered by Dingle, relating the late to the early terms of
the original expansion. The improvements form a nested sequence of asymptotic
series truncated at their least terms. Each such 'hyperseries' involves the terms of
the original asymptotic series for the particular function being approximated,
together with terminating integrals that are universal in form, and is half the length
of its predecessor. The hyperasymptotic sequence is therefore finite, and leads to an
ultimate approximation whose error is less than the square of the original
superasymptotic remainder. The Stokes phenomenon is automatically and exactly
incorporated into the scheme. Numerical computations confirm the efficacy of the
technique.
1. Introduction
In an asymptotic series, it is common for the terms Yr to decrease at

ultimately increase, so that the series diverges. If A is the (large) asy
parameter (i.e. Y oc A-r) then the smallest error (relative to the first t
from the bare series is exp (-AA), where A is a positive constant, and is ob
truncation at or near the least term (for example, Olver 1974). The order o
term is r oc A. Such 'superasymptotics', in which exponential accuracy
by making the number of terms increase with the large parameter (ra
considering a fixed number of terms as in the asymptotics of Poincare), ha
history. Stokes (1864) used it in the accurate approximation of Bessel f
Nekhoroshev (1976) made it the basis for estimates of long-time predic
nonlinear hamiltonian dynamics. And it is an essential component of detail
(Berry 1989a, b, 1990a; Olver 1990a, b; Boyd 1990; Jones 1990) of t
phenomenon, which is the appearance and disappearance of small expon
certain variables (not A) change.
Here we go beyond superasymptotics, and obtain by repeated resum
sequence of asymptotic series, each approximating the leftover from
truncation of its predecessor. We define hyperasymptotics as the systemat
these approximations to the small exponential error left by truncation of
series. The possibility of such sequential improvement was glimpsed by
(1886), explored in particular cases by Airey (1937) and Miller (1952) (re
Olver 1974), and clearly envisaged by Dingle (1973, hereinafter called D
& Solov'ev (1989) and Boyd (1990). D developed a systematic method, b
Borel resummation, for the first stage of hyperasymptotics, in which the te
Proc. R. Soc. Lond. A (1990) 430, 653-667 Printed in Great Britain
23-2
654 M. V. Berry and C. J. Howls
first hyperseries for a great variety of functions are expressed as certain standard
integrals called 'basic terminants'. Boyd (1990) envisages a hyperasymptotics based
on Stieltjes transforms, and studies its first stage for certain Bessel functions.
In exceptional cases, hyperasymptotics is unnecessary because the exponential
leftover can be expressed exactly as a single basic terminant (an example is Erf(z),
Izl large), or as a convergent series of basic terminants (an example is In F(z), lzl large).
Usually, however, hyperasymptotics is non-trivial in the sense that the resummations
persist in generating asymptotic series which themselves require resummation.
Perhaps the simplest case where this happens is in the approximation of solutions
y(z, A) of the one-dimensional Helmholtz (or Schr6dinger) equation,
d2y(z,A)/d2 = A2Z(z) y(z, A), (1)

where z is a complex variable and Z(z) is analytic. This is the situ
concentrating on the (generic) case of points z where y is domin
specified later, by a single transition point (zero of Z(z)). We are
hyperasymptotic series explicitly (?3) and study in detail the
sequence of remainders that it generates (?4). Our hope was that
converge to the exact solution when each hyperseries is truncate
but this is not the case. It turns out that the successive hyperse
the process terminates naturally after a number (of order ln A) o
this last stage is of order exp {-(1 + 21n2)AA} exp(-2.386A
positive constant already introduced; this error is smaller th
error at the zeroth (superasymptotic) stage. (In principle, arbitr
be obtained by a non-optimal truncation of the hyperseries, but
this procedure would be impractical.)
We achieve this improvement by exploiting a 'resurgence form
by D (p. 300) and rediscovered by Rakovic & Solov'ev (1989),
Yr (r - oo) in the asymptotic series for each of the two 'wave ' so
an expansion giving each Yr for large r as a series over the Yr for
early terms. At first encounter this appears astonishing but in f
because of the following: (i) each of the two wave series is by itse
solution of (1); (ii) both series must diverge, to reflect the fa
constant) there are really two waves in the solution of (1); (iii) the
tail of each series must encode the early terms of the other seri
the two series are the same (apart from signs) namely Yr.
In different regions of the z plane (separated by Stokes lines) t
the solutions y(z, A) contain different combinations of the two
solutions. This is the Stokes phenomenon. Our hyperasym
incorporates these changes of form automatically and exactly
To illustrate the workings of hyperasymptotics we presen
example. This is the Airy function, in which Z(z) in (1) is li
sequential improvement over brute asymptotics (stopping at the
error - A-l) and superasymptotics (error - exp (-AA)), down t
(error - exp (- 2.386AA)) is achieved in all Stokes sectors.
Such extreme numerical precision is not, however, our main re
out this study. After all, it is often the case that where a nume
required, standard differential equation routines (e.g. Runge-Kut
series, are more accurate or faster than asymptotics (ordinary, s
even when A is quite large. Our main motive is rather that t
Proc. R. Soc. Lond. A (1990)
Hyperasymptotics 655
solution of (1) provides a testbed for studying in great detail the structure of a
physical theory as an important parameter (here A) takes a limiting value. One such
limit, of considerable current interest, is the classical limit of quantum mechanics
(Berry 1990b).
2. Dingle's resurgence formula

The two lowest-order 'wave' solutions of (1) are given by the phase-integral (i.e.
WKB) method (Heading 1962; Fr6man & Fr6man 1965) as
y(z) - exp [A f dZ()1]/ Z(z) (2)

in which z* is an arbitrary reference point. A natural variable will be the differ
between the two exponents, namely
F(z) _= 2A dZ(C). (3)

z*
It is convenient to perform the analysis for the solution which is subdominant (i.e.
exponentially small) when ReF > 0, and write this as
y(z) = (exp [-2F]/Zl(z)) Y(F). (4)

A formal asymptotic series for Y, in descending powers o
substitution into (1). We write
00
Y(F) = (-1)rYr(F), where Y0 = I and Yr c A-r. (5)

r=O
Then the Yr satisfy the recurrence relation (D, p. 296)
Y'r+ (F) = -Yr(F) +G(F) Yr(F), (6)

where primes denote derivatives with respect to F and
G(F) = (Z1)"/Z1. (7)
Note that the large parameter A no longer
define the terms Yr in the series representing
An important role is played by the transitio
the corresponding points Fj, the function
depend only on the order of the zero: for an
(7) that
G(F) --m(m + 4)/4(m + 2) (F-Fj)2 as F-> Fj. (8)
To avoid complicating the analysis, we henceforth consider only the gener
m = 1; generalization to arbitrary m is straightforward. When iterating
obtain the Yr, the derivatives magnify the singularities (8), leading as explaine
(p. 299) to the following simple approximate formula for the late terms:
Yr(F)-* (r-1) !/2t(F-Fo)r as r-oo, (9)
where Fo denotes the transition point Fj which is closest to F i
the smallest value of IF-Fjl. For large IF-Fol (ensured by
decrease until r IF-F0o and then increase.
The resurgence formula which will be central to all our subsequent analysis is a
formally exact representation of Yr with (9) as its leading term. By direct substitution,
it can be confirmed that
Y(F) -_= E Y)(F)[-(F--F1)] (r-s- )! (10)

( j )r 8=0
satisfies the recurrence relation (6). The function G(F) enters only th
positions Fj of its poles. It should be remarked that (6) is formally satisfied
more general relations, in which (i) Fj are any points whatever, (ii) on the r
side r is replaced by r + a with c arbitrary, (iii) 1/(27) can be any constant, a
Y4 can be any solutions of (6) (rather than those with the same integration
as the Y, on the left-hand side). The particular choice (10) reproduces th
form (9), which corresponds to the term s = 0, and also satisfies certai
requirements, as will be explained in ?5. The higher terms s > 0 provide
asymptotic expansion for the late Yr in terms of successive early Yr.
In what follows we will use the simplified form of (10) obtained by negle
transition points other than the closest, Fo, which we will henceforth t
origin F = 0 (this is equivalent to choosing z* = z in (3)). This gives
00
hY(F)
Y,(F) -^ S1 = 2nFr s (11)
E(r-s-l1)!(--F)`Y(F).
As they stand, (10) and ( 1) are numerically
s > r-1 are infinite. They can, however, be
and the hope is that these resummed versions
this hope in the case of ( 1), by showing (
equivalent to the Stieltjes transform relation b
basis of the rigorous analysis of Boyd (1990
Thus (11) will be exact (when suitably interp
point. If more transition points are present,
of small exponentials whose order is
exp -rln I(F-F1)/FI}, (12)
where F1 is the distance (3) between the transition
F (by hypothesis, IF-F11 > IFI). We will use (11) on
neglected exponential will never be larger than
exp {-IFI ln l(F--F)/FI}. (13)
Now, the smallest small exponentials in hyperasymptotics
exp{-( + 2 n2)1Fl} (?4) so we can expect our analysis to be v
smaller than these, that is if
IF-Fll > 4elFl (14)
Equation (11) is Dingle's formula (D
to denote F, the distance to the dom
(1989) rediscovered (11), and presented
transition points.
3. Hyperasymptotic resum
We seek accurate approximations to Y
the distance IFI to the nearest transitio
Retain the first No- 1 terms Yr as given by the solution of (6), where No is the least
term, and replace the remaining terms by (11). Thus (interchanging the r and s labels)
y =So+ E Yr(-F)r E (--1)s tF (15)

r=0 s=N0
No-l
where So= _ (-1)r}. (16)
r=0
The truncated sum 80 is the zeroth level of hyperasymptotics, which we have ca

superasymptotics.
The next stage is Borel summation of the s sum in (15), that is replacement of
factorial by its integral representation, followed by interchange of summation
integration and evaluation of the sum. After an elementary change of variable,
obtain
Y= 8+ (-1)rYKrl, (17)
r=0
_l)No Co No-r-1
where Krl N=2-r d exp( 6)1i/F (18)

In ?4 we shall see that the second series in (17) also diverge
say. Therefore as before we retain the first N1-1 term
remaining terms by the resurgence relation (11), to get
y SO+S+ E Yr(-F)r (1) (s-r-1)'KsV (19)

r=O s=N1 2F
N1-1
where 81 = (-1)rYrKrl. (20)

r=O
The truncated sum S1 gives the first level of hyperasymptotics. This was studied i
great detail by D (chapters XXI and XXIV), who called the integrals Krl
'terminants', by Boyd (1990) (for Bessel functions) and by Olver (1990b) (for
confluent hypergeometric functions).
Again we can apply Borel summation to the s sum in (19), and obtain
Y = S + 1? + E (- )ryrKr2, (21)
r=O
where
(-1 )No+N,1 ro No-N -1 e0 Ni-r-1

Kr2 (2d)2eFNoxp-r dglexp F) -2 1 dexp( - (2) 2F(I I)1 -- 2/1)'(22)
Now the third sum in (21) diverges, with a least term at N2, say. Therefore, we again
retain the first N2-1 terms and replace the Yr in the tail by the resurgence relation
(11); the truncated sum constitutes the second level of hyperasymptotics.
Now the pattern is clear: truncation, resurgence and Borel summation can be
repeated, leading to the hyperasymptotic sequence
Y = S0+S+S2+ ..., (23)
Nn-1
where Sn = (- 1)r YK. (24)

r=-
The Yr are the terms in the original asymptotic series. Th

problem as embodied in the function Z(z) in (1), which
through G(F) as defined in (7). The Krn constitute the hy
of Dingle's terminants Kr1. They are universal functions -
detailed form of Z(z) - given by the n-fold multiple integ
Kro= 1,
KrnKrn(FN0,., Nn) I
rn- rn o * * n (25)
_ 1 n i ;Ni_-Ni-1( _ 1)N_I (25)
(2= )nF Nor '= d6i exp ( - i) (-i-1
For these integrals to converge, Ni < Ni_1, so that
fewer terms and the process of hyperasymptotics m
where Nn = 1.
It is clear from equations (23) and (24) that the s
universal renormalization of the Yr, enabling the info
decoded, via resurgence, to yield much higher pr
approximated. This information is however finite, because, since successive
hyperseries get smaller, only the first No terms Y) will participate in the
hyperasymptotics. Therefore we can expect the ultimate error of the scheme, when
it stops, to be finite and not zero.
Of course, the actual numerical extraction of the information contained in the first
No terms Yr requires knowledge of the Kn to sufficient accuracy. In a thoroughgoing
asymptotics, the multiple integrals (25) would themselves be expanded asymp-
totically (and where necessary hyperasymptotically) for large F, and the form of
the expansions would depend on r and the Ni (cf. the formulae for Krl in D, p. 415ff).
We do not carry out this programme, but simply assume that the Krn are known; in
our numerical example (?6) we reduce the required Krn to certain special functions
which can be computed 'exactly' (see Appendix B).
4. Optimal truncations; estimates of ultimate accuracy

In choosing No, N1, ..., the guiding principle is that the successive series So, S, ...,
are truncated near their least terms. To implement this principle for So we use the
estimate (9) for Yr, replacing (r-1)! by Stirling's formula, and thus obtain the
well-known result
N = Int IFI. (26)
For the higher sums we need es
the 6 in the slow varying denomin
rest of the integrands, namely
6* -N_1-Ni-1 (Nn = r) (27)
This uncouples the integrals and gives
n 1
Km n 1 (_ ')Nji
rn (21)nFNo-r i (i +)/ / 1)(2
In S1 the terms are
K (r-l)! ! (r-1)!(IF-r-1)! (29)

lYr rll o Fr KNo_ r - ,[1 (29)
Fr FNO-r IElIFI
whose least term is obvious from symmetry and giv

Boyd 1990)
N -= Int (llFI). (30)
Repetition of this procedure gives
Nn = Int (IlF/2n). (31)
Thus each hyperseries Sn is half the length
hyperasymptotics comes when the Sn has jus
stages, given by Nn = 1 as
nmax = Intlog2lFI. (32)
To study the magnitudes of the terms we use
Stirling's approximation and the denominators
and (31) for large F). The last term of the nth h
reduction, as
( l)No+... Nn 2n(n+7)/4
( ) _n-1 I)NnYN K exp{-
Nn-ln (2)n+l) F exp{- IFl
[1 +2(-2)ln2]. (33) [1 +
The first term of the (n + l)st hyperseries, namely K,n+l, involves exactly
factors, with the additional denominator
n+l Nn -- 1
*n n- l-Nn-1
This shows that not only do the terms decreas

but that each new hyperseries begins with a term
last term in its predecessor. (There is one exce
is not real, the factor is not 1 but 1/[ + exp
The ultimate accuracy of the hyperasymptot
same order of magnitude as the single term
estimate this by substituting nmax from (32) int
Kon - IFI2 [[log,2lF+4-log (3Vi 2)]
This confirms that the error is indeed finite, as

number of participating Yr, and of the orde
Introduction (with AA now identified as IFl).
Now we show that it is possible to devise hyp
accuracy (within the one-transition-point app
optimality requirement that the terms must a
need to recast the scheme (23)-(35) so as to yiel
manipulations (cf. Appendix A) give the remaind
namely n
R(F;No...Nn)--Y- E Sm (36)
m=0
24 Vol. 430. A
as
"'""^^(2t)~ 0' t0(I +Jt0)J t(1 + t)"
nt) exp [-F(t t t1 + +t+ t1 ... t)] Y(Fto t .. t) (37)

Bounding IY(F)I by a constant C, we obtain
IR(F;No, N, ... N)I < C[(No-N1- 1)! (Ni-N2-- 1) .. (- 1)!]/(21)nIFNol. (38)
For given No the smallest bound on the remainder occurs when all the factorials
are unity, and is achieved by making successive truncation limits decrease by 1
(rather than halving as in the optimal scheme). The scheme thus terminates after
Int (No) stages, leaving a remainder
IR(F;No, No-, 1..., 1)I < C/12cFINo. (39)
This can be made arbitrarily small by increasing N0. None of the tr
optimal now; the price of arbitrary accuracy is to represent Y(F)
cancelling terms, just as in convergent series representations, and all th
of asymptotics are lost.
The impracticality of these arbitrarily accurate schemes leads to
appreciation of the 'live now, pay later' philosophy underlying th
hyperasymptotic (and ordinary asymptotic) schemes, in which ultimate
sacrificed as a consequence of the requirement that the terms always de
each hyperseries and from one hyperseries to the next).
5. Exact Stokes relations
In the scheme (23)-(25), the singulant F enters through the Yr and i

denominator of the first integral (over 61) in each K,. F can take any comp
We now show that as F makes a circuit around the origin, i.e. as z en
transition point, the solution to (1) as given by the hyperasymptotic
reproduces the Stokes phenomenon - the birth and death of small exponent
presence of large ones - exactly and to all orders.
Consider first F real and positive. Then (4) represents an exponentially
(subdominant) solution of (1). Now let argF increase to n, so that (4) re
dominant solution. The replacement F = -IFI in the series So gives late
(-1)iy which all have the same sign (cf. the limiting form (9)). This is b
characteristic of a Stokes line (Stokes 1864; D, pp. 7, 414). Near to the t
point in the original z plane the n rotation of F corresponds to a 17 rotatio
In the hyperasymptotic corrections, the rotation produces a pole at 61 = I
first integrand of each Kr. As arg F increases to n, the pole approaches the
from below. Thus, by continuity, the integral splits into two contributions:
principal value and from the negatively traversed infinitesimal semicircle a
pole. Rather than write out these contributions explicitly for each of the s
hyperasymptotics, we carry out the splitting in one step using the resumme
of the scheme, based on (36) and (37) with n = 0. This is
i (0 t N?A exp t)?
Y(F) = o(F) + dt ) Y(t). (40)
The splitting just described can now be easily implemented, and gives
Y(IFI exp (ic)) = S,(-IFl)
1 o t N__t exp (-t)
+2
2nf/dtll Y(l/IFIY(t)
IFI t(1 -tlFI) + 2i Y(IFI) exp (-IF), (41)
that is
Y(IFI exp (ic)) = Yp(- IF) +i Y(IFI) exp (- 1F), (42)

where the subscript p denotes the hyperasymptotic series for Y(-IFl) with all 61
integrals interpreted as their principal values. Note that (42) is an identity for the
function Y(F), in which all reference to the truncation No has disappeared.
Thus on the Stokes line the solution (4) which is subdominant for F positive real
is, exactly,
y(z) = Z-(z) exp [ + 1FI] Y(IFI exp (ic))
- Z-(z) {Yp(- IF) exp [ + IlFl] +i Y(IFI) exp (- lFl)} (43)

showing that the solution has now acquired a subdominant contribution whose
leading term is -n out of phase and half the size of the dominant exponential, as it
must be (D, pp. 8, 414).
The strength of the subdominant contribution grows from 0 to 1 across the Stokes
line, being i on the line itself, as indicated in (43). Berry (1989a, 1990a) showed that
the strength increases smoothly, according to a uniform approximation involving an
error function. In the present hyperasymptotic scheme, this emerges as an
approximation to the first term in the first hyperseries, i.e. to the first terminant
Kol(F,No) (equation (18)) for F near the negative real axis. Here - and only here -
our previous approximation (28) for the generalized terminants (25) breaks down,
because in the 61 integral the pole and the stationary point coincide. For the higher
terminants (r > 0 or n > 1), the pole at 1 =- IFI for F negative real can never
coincide with the stationary point at 61 = * (equation (27)) if the Ni are chosen
optimally (i.e. as IF1/2i), and so the principal-value pole contribution is exponentially
small compared with that from the stationary point. A consequence of this is that the
Stokes phenomenon, regarded as the rapid switching-on of the subdominant
exponential, is completely described by the error function; further hyperasymptotic
corrections vary only slowly across the Stokes line.
Now continue the positive rotation of F until arg F = 3. We are now on an anti-
Stokes line, where both exponentials have equal magnitude and opposite purely
imaginary phases. The pole in the 61 integrals has rotated positively up to the
imaginary axis, and by continuity has dragged the contour with it. Again these
integrals can be split into contributions from the pole and from the original contour
among the positive real axis. From (40) we obtain (cf. (41))
Y(lIF exp (3iC)) = So(lIF exp (--iC))
+2 dtilFl t(1 + it/IFI) 44

that is (cf. (42))
Y(IFl exp (3iT)) = Y(-ilFl)+iY(ilFl)exp (-ilFl) (45)
24-2
in which the functions Y on the right-hand side are evaluated with all Krn contours
along the positive real axis.
Thus on the anti-Stokes line the solution (4) is, exactly,
y(z) = Z-i(z) exp [ + lilFI] Y([FI exp (ji))

= Z(z)e ) 2Re [Y(ixp) exp i{i(-lFI + )}], (46)
showing that the subdominant contribution which appeared on th
equal in magnitude to the original exponential, and y(z) is now pro
oscillatory function, as it must be (D, p. 297).
The function Y(F) is not single valued. We can see this by rot
argF = 27, that is onto the second Stokes line in the z plane, c
171 rotation close to the transition point. Similar reasoning to tha
(cf. (42))
Y(IFI exp (2i7i)) = Y(FI) + iYp(- [F) exp (IFI). (47)
Finally, we rotate to argF = 3c, corresponding to a complete circuit of the
transition point in the z plane. The pole has now dragged the 61 integration contours
around a point on the second sheet, above the real axis. There is a contribution from
this pole at 61 = IFI exp (27i), as well as a half contribution from the pole at 6 = IFl,
leading to
Y(IFI exp (3ii)) = Yp(-IFl) + i[Y(IFI exp (2ii)) + Y(IFl)] exp (-IFl)
= iY(FI) exp (-IFI). (48)
The solution (4) now becomes
Y(z) = ~ jexp [LIFl] Y(IFI e

Z4(Z -> z exp (27ii)
iexp[-IFI/2] Y(IFI). (49)

Z4(z - z exp (2r7i))
This must be exactly the same as at the start of the circuit, becau
valued, and indeed it is, because the factor i is cancelled by the chang
its zero.
6. Numerical illustration: the Airy function

Here we choose Z(z) = z, which is the paradigm for the study of equation (1) near
a single transition point. Then without loss of generality we can set A = 1, because
asymptotics for large A is equivalent to that for large lzl. From (3), the singulant is
F = z. (50)
The solution of (1) wh

Airy function Ai (z)
have
Ai (z) = Y(F) (51)

271 z4
so that Y(F) = 2V/(i(F)6exp (iF)Ai{(F)i}. (52)

total number of terms in hyperasymptotic ser
I i I I 1
10 20 30
-5-
-15-
zeroth stage 1st stage 2nd 3rd4th

Figure 1. Decrease of the terms in the first five hyperseries of Y(16), for the Airy function Ai.
Table 1. Hyperasymptotic approximations to the Airy function Ai (equation (51)) for F = 16
level approximation to Y(F) approx. - exact

lowest 1 8.163x 10-3
SO 0.9918367935113234591100 -5.677 x 10-9
SO+S1 0.9918367991882512550983 -1.134 x 10-14
So+S +S2 0.9918367991882625907500 -8.160 x 10-18
S +... + S3 0.991836 799 188 262 599 868 2 9.584 x 10-19
S + ...+ S 0.9918367991882626006031 1.151 x 10-18
exact 0.9918367991882625989098 0
To compute Y(F) hyperasymptotically we require the terms in its formal

asymptotic expansion. These are
I F(3r + 1)
Yr(F)=
Yr(F) + (53)
= (27F)r F(r + 1)
which satisfy Y0 = and the lim

terminant integrals (25). Krl and
for higher hyperseries we used th
All numerical work was perform
program Mathematica (Wolfram 1
it can be configured to work to
functions for complex arguments.
we checked against the representa
as a convergent series.
First we study Y(F) on the posi
F = 16, that is z = 5.2414827884177932413.... From (35), we can hope to be able
to compute Y(16) with an error 8.4 x 10-19.
Figure 1 shows the decrease of the terms in the first five stages of hyperasymptotics,
that is the terms in S0, S1, S2, S3 and S4 (the last series having only one term). The
optimality of the hyperseries is obvious. Table 1 shows the numerical values of the
successive approximants to Y(F), together with their errors. The improvement with
hyperasymptotics is dramatic. Even the first level reduces the error of super-
Table 2. Hyperasymptotic approximations to the Airy function Bi (equation (55)) for IFI = 16
stage approximation to Yp(- IF) approx.-exact

lowest 1 -9.355 x 10-3
o8 1.009354544224441282011 12 -7.389 x 10-9
0 +81 1.00935455 1613418 76942446 -4.296 x 10-14
80 + 81+S82 1.009354 551 613461695449 -3.012x 10-17
So + ... +. 83 1.009354551 613461721493 -4.078x 10-18
So+... +S4 1.009354551613461 721984 -3.587 x 10-18
exact 1.00935455161346172557054 0
asymptotics (S8) by five orders of m

reduction is nearly ten orders of magn
of the method.
Now we move onto the Stokes line, where arg z = 1: and arg F = n. Because of the
identity (42), it is necessary only to study Yp(- IF), which is the real function
obtained when all 61 integrals in the generalized terminants (25) are interpreted as
principal values. This is equivalent to studying the exponentially increasing Airy
function Bi (z) (Abramowitz & Stegun 1964) for z positive real, because (cf. (51) and
~(52)) Bi (z) =- - exp (1FI) Yp(- FI), (54)

i.e. Yp(- IF) = -/(3jF[)I)exp (-IlFI)Bi {(31FI)2}. (55)
The approximations to Y (- IFl) are shown in table 2, again for IFI = 16. Her
improvement over superasymptotics (So) with the first three stages of
asymptotics is nine orders of magnitude.
Gailitis & Silverstone (1988) have studied the asymptotics of Bi (2.5), whose
value is 6.4816607.... With Pade summation of the first 40 Y, they obta
this real quantity the complex value 6.48166-0.00001 i. With three s
hyperasymptotics -nine terms in all-based on only the first five Yr, w
6.4816598, thereby achieving the ultimate error of the method, which from
0.000001.
Finally, we examine the anti-Stokes line, where arg z = n and the Airy fun
oscillatory. From equation (46),
Ai {-(1IFI))} = Re [Y(iF) exp {i(-I Fj + IT)}]/v7/(IF[)1
Here we choose to study the zeros, rather than the values, of Ai.
The lowest few stages of hyperasymptotic approximation for the first thr
are shown in table 3, with errors displayed as fractions of the asymptotic m
spacing 27. For the crudest approximation (Y ~ 1), the errors are of order 1/
are roughly the same for the three zeros. Superasymptotics (Y So
exponentially small errors which therefore diminish considerably from the fir
third zero. Hyperasymptotics gives further, enormous, improvements: even
first zero, where Fl ?% 4.8 is hardly large, the error diminishes from 10
10-7 (S0+S1+S2); for the third zero (IFlI 17.3) the additional improv
nearly eight orders of magnitude.
7. Discussion
We have given an explicit description of hyperasymptotic correc

corrections beyond the accuracy exp (- IFX) obtained by the usual (supe
procedure of summing the primary asymptotic series down to its l
Table 3. Hyperasymptotic approximations to the values of IF = (31z1/4) for the first three zeros of the
Airy function Ai (-lzl) (equation (56))
Table 3a
stage IFI (lst zero) (approx.-exact)/2n:

Y 1 4.712388980 384689857 694 -8.67 x 10-3
So 4.765396652270803735893 -2.38 x 10=4
SO +8 4.766878592385215365778 -2.33 x 10-6
80 +1 +82 4.76689444457039445853 1.94 x 10-7
exact 4.766893225061654917 199 0
Table 3b
stage [FI (2nd zero) (approx.-exact)/27:

Y 1 10.99557428756427633462 - 3.95 x 10-3
SO 1 1.020393250277312441 71 3.28 x 10-7
S +S1 11.020391 19052794602966 -3.17 x 10-11
+ 81 + 82 11.020391 19072562870 -2.84 x 10-13
exact 11.020391 19072741674417 0
Table 3c
stage IFl (3rd zero) (approx.-exact)/2t

Y - 1 17.27875959474386281154 -2.54 x 10-3
So 17.29471532320673859433 -4.98 x 10-10
So+S1 17.294715326337 19810422 -5.94x 10-16
So+S1 +2 17.29471532633720182004 -2.20 x 10-18
exact 17.294715 326 337 201833 86 0
structure of the scheme is particular

product of a universal factor, namely a
(25)), and a non-universal factor, namely
the particular function being approx
asymptotics of the hyperasymptotic
terminates naturally, leaving an ultim
Much remains to be done. Our argu
formula (11), which applies when the p
a single transition point. While this is
to envisage other situations, in which a
make appreciable contributions. Pres
based on the more general resurgence f
we emphasize that a proper mathematic
Any such generalization of hyperasym
points, would have to be compatible w
Stokes relations of ?5 (cf. Heading 19
We thank Dr W. G. C. Boyd for helpful disc

from the U.K. Science and Engineering Res
Appendix A. Resumming resurgence

Summing both sides of (11), we obtain Y(F) from (5) (after an interchange of
summation labels) as
Y(F) = E (F) (-F)r . (A 1)

2 :r=O s=0 (-F)'
Next, the sum is interpreted by the Borel procedure, wh
I \=exp(-
Y(F) Yr ) (F)
v \ F
-
2 6(1 + 6/F) r=o 6
= dx (x) F) - (A 2)
271j x(l+x) =o 0
Because our treatment is restricted to a single f
separated (in F) from any others, it is consistent
such cases as Z(z) = z exp (z), where the transition
Yr(F) is proportional to F-r (cf. (53)), so that
Yr(F) (- l/x)r = (- )r Yr(Fx). (A 3)
Again using (5) we obtain the resummed resurgence formula
1 (0? exp (-Fx)

Y(F) = 2 dx ( Y(Fx) (A 4)
2n), J x(1 +x)
corresponding precisely to the exact Stielt
order 1, as used by Boyd (1990) in his stu
In the more general situation, where there
not be precisely proportional to F-r, and t
cannot be used. But by construction Yr is
parameter A. We can exploit this by redefi
A, that is as
Y(z) -2fd d6Z(C (A 5)

z*
Then we can resum the conjectured more general relation (10), in

transition points Fj, without assuming (A 3). Thus we obtain a puta
formula for the multiplier Y(F, A) in the solution (4):
Y(YF, A) 1 T d exp{-- h(- FJ) x}Y(Ax

2 J x( +x) X)
relating the solutions of (1) for different values of A. We can
nor exploit this intriguing relation.
Appendix B. Calculation of generalized termin

Our aim here is to express the terminants for the first two stages of
hyperasymptotics, that is the integrals (18) and (22), in terms of functions that can
be evaluated by Mathematica with sufficient precision (up to 24 digits for the
calculations presented in ?6).
Hyperasymptotics
667
The first-stage terminant integrals Kr1 can be expressed in terms of incomplete

gamma functions (cf. D, p. 415, Abramowitz & Stegun 1964):
Kri(F,No) = ((- )No/2r) exp (F)r(No-r) r(r-No + 1,F). (B 1)
Now we use
( _i \N iN-I (_ 1)mm! 1
r(-N, F) = ()rF(0, F)-exp (-F) Fm+l (B 2)
?N! m=Om+0
and obtain
Kri(F, N)- )r+l No-r-2 (--l)mm!

Krl(F,No) = 2
^ exp (F)E1l(F)- E Fm+l (B3)
L m-O-0
where E1 denotes the exponential integr

To simplify the second-stage terminan
the 62 integral in (22):
Kr2(F, No N) (1) )NO+Ni F(N-

(2ic)2 J(1?t)
Next we use (B 2), which gives

Kr2= (-1)No-r-1/(27;)2 (I-J),
where I= dp exp(-P) /Fdt t-r+N 0-

Po ( +Jt) (
N-r-2 (- )m! dt exp (-Ft) t-r+No-m-2
m0 Fm+ J (1 + t)
For I we substitute the expansion
tN (_1\N N-1
( +t)(-
(1+t)- + r=o
(1+t) 1)N N (-t)r (B 6)
and thereby obtain
I dpexp(-P) ln (+ --2 ( (B 7)
l P oF (q + 1)Fq+l '
The integral can be expressed in terms of Euler's constant y and the dig
function - (Prudnikov et al. 1986, vol. 1, p. 530), leading to
I =l[ln?/Ex+ [2/ +1Y 2/l+ (k)-ln F ~
No-r-2 (__l)y
E-- F (ReF> 0). (B8)
o (q? - 1) F+
The sum over Ik converges adequately fast for our purposes.
To simplify J we notice that the integral has the same form as Kr1 and then
substitute (B 1) and (B 2)
N-r-2 mi Ni-r-2 No-r-m-3 (-)Sm! (B9)
= (- --exp (F)E(F) Fm+1+ m+s+2 (B9) m=0 m=O s=0
References
Abramowitz, M. & Stegun, I. A. 1964 Handbook of mathematical functions. Washington

National Bureau of Standards.
Airey, J. R. 1937 Phil. Mag. 24, 521-552.

Berry, M. V. 1989a Proc. R. Soc. Lond. A 422, 7-21.
Berry, M. V. 1989b Publications Mathematiques IHES, Paris 68, 211-221.
Berry, M. V. 1990a Proc. R. Soc. Lond. A 427, 265-280.
Berry, M. V. 1990b Some quantum-to-classical asymptotics. In Chaos and quantum ph
M.-J. Giannoni & A. Voros) (Les Houches Summer School session 52, 1989) North-H
the press.)
Boyd, W. G. C. 1990 Proc. R. Soc. Lond. A 429, 227-246.
Dingle, R. B. 1973 Asymptotic expansions: their derivation and interpretation. New York and
London: Academic Press. (Ref. D of the text.)
Friman, N. & Friman, P. 0. 1965 JWKB approximation; contributions to the theory. North-
Holland: Amsterdam.
Gailitis, M. & Silverstone, H. J. 1988 Theor. chim. Acta 73, 104-114.

Gradshteyn, I. S. & Ryzhik, I. P. 1965 Table of integrals, series, and products. New York and
London: Academic Press.
Heading, J. A. 1962 An introduction to phase-integral methods. London: Methuen.

Heading, J. A. 1977 Q. Jl Math. 30, 281-302.
Jones, D. S. 1990 Proceedings of International Symposium on Asymptotic and Comp
Analysis, Manitoba, Winnipeg 1989 (ed. R. Wong), pp. 241-264. Marcel Dekker: New
Maeder, R. 1990 Programming in Mathematica. Addison-Wesley.
Miller, J. C. P. 1952 Proc. Camb. phil. Soc. 48, 243-254.
Nekhoroshev, N. N. 1977 Russ. math. Surv. 32, 1-65.
Olver, F. W. J. 1974 Asymptotics and special functions. New York and London: Academic Press.
Olver, F. W. J. 1978 Phil. Trans. R. Soc. Lond. 289, 501-548.
Olver, F. W. J. 1990a Proceedings of International Sumposium on Asymptotic and Computational
Analysis, Manitoba, Winnipeg 1989 (ed. R. Wong), pp. 329-355. Marcel Dekker: New York.
Olver, F. W. J. 1990b Preprint from Institute for Physical Science and Technology, University of
Maryland.
Prudnikov, A. P., Brychkov, Yu. A. & Marichev, O.I. 1986 Integrals and series. New York:
Gordon and Breach.
Rakovic, M. J. & Solov'ev, E. A. 1989 Phys. Rev. A 40, 6692-6694.

Sibuya, Y. 1977 Bull. Am. math. Soc. 83, 1075-1077.
Stieltjes, T. J. 1886 Annls l'Ecole Normale 3, 201-258.
Stokes, G. G. 1864 Trans. Cambs. phil. Soc. 10, 106-128. (Reprinted in Mathematical and ph
papers by the late Sir George Gabriel Stokes, vol. iv, pp. 77-109. Cambridge University Press
Voros, A. 1983 Annls Inst. H. Poincare' 39, 211-338.
Wolfram, S. 1988 Mathematica: a system for doing mathematics. Addison-Wesley.
Received 1 May 1990; accepted 6 June 1990

Hyperasymptotics (Berry and Howl)

Uploaded by

Copyright:

Available Formats

Hyperasymptotics (Berry and Howl)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hyperasymptotics (Berry and Howl)

Uploaded by

Copyright:

Available Formats

Hyperasymptotics

Author(s): M. V. Berry and C. J. Howls

H. H. Wills Physics Laboratory, Tyndall Avenue, Bristol BS8 1TL,

We develop a technique for systematically reducing the exponentially small

In an asymptotic series, it is common for the terms Yr to decrease at

d2y(z,A)/d2 = A2Z(z) y(z, A), (1)

2. Dingle's resurgence formula

y(z) - exp [A f dZ()1]/ Z(z) (2)

F(z) _= 2A dZ(C). (3)

y(z) = (exp [-2F]/Zl(z)) Y(F). (4)

Y(F) = (-1)rYr(F), where Y0 = I and Yr c A-r. (5)

Then the Yr satisfy the recurrence relation (D, p. 296)

Y'r+ (F) = -Yr(F) +G(F) Yr(F), (6)

Y(F) -_= E Y)(F)[-(F--F1)] (r-s- )! (10)

y =So+ E Yr(-F)r E (--1)s tF (15)

The truncated sum 80 is the zeroth level of hyperasymptotics, which we have ca

where Krl N=2-r d exp( 6)1i/F (18)

y SO+S+ E Yr(-F)r (1) (s-r-1)'KsV (19)

where 81 = (-1)rYrKrl. (20)

(-1 )No+N,1 ro No-N -1 e0 Ni-r-1

where Sn = (- 1)r YK. (24)

The Yr are the terms in the original asymptotic series. Th

4. Optimal truncations; estimates of ultimate accuracy

In S1 the terms are

K (r-l)! ! (r-1)!(IF-r-1)! (29)

whose least term is obvious from symmetry and giv

This shows that not only do the terms decreas

Kon - IFI2 [[log,2lF+4-log (3Vi 2)]

This confirms that the error is indeed finite, as

Proc. R. Soc. Lond. A (1990)

"'""^^(2t)~ 0' t0(I +Jt0)J t(1 + t)"

nt) exp [-F(t t t1 + +t+ t1 ... t)] Y(Fto t .. t) (37)

5. Exact Stokes relations

In the scheme (23)-(25), the singulant F enters through the Yr and i

Proc. R. Soc. Lond. A (1990)

Y(IFI exp (ic)) = Yp(- IF) +i Y(IFI) exp (- 1F), (42)

y(z) = Z-(z) exp [ + 1FI] Y(IFI exp (ic))

- Z-(z) {Yp(- IF) exp [ + IlFl] +i Y(IFI) exp (- lFl)} (43)

Y(lIF exp (3iC)) = So(lIF exp (--iC))

+2 dtilFl t(1 + it/IFI) 44

y(z) = Z-i(z) exp [ + lilFI] Y([FI exp (ji))

Y(z) = ~ jexp [LIFl] Y(IFI e

iexp[-IFI/2] Y(IFI). (49)

6. Numerical illustration: the Airy function

The solution of (1) wh

Ai (z) = Y(F) (51)

so that Y(F) = 2V/(i(F)6exp (iF)Ai{(F)i}. (52)

zeroth stage 1st stage 2nd 3rd4th

Table 1. Hyperasymptotic approximations to the Airy function Ai (equation (51)) for F = 16

level approximation to Y(F) approx. - exact

To compute Y(F) hyperasymptotically we require the terms in its formal

which satisfy Y0 = and the lim

stage approximation to Yp(- IF) approx.-exact

asymptotics (S8) by five orders of m

~(52)) Bi (z) =- - exp (1FI) Yp(- FI), (54)

We have given an explicit description of hyperasymptotic correc

stage IFI (lst zero) (approx.-exact)/2n:

stage [FI (2nd zero) (approx.-exact)/27:

stage IFl (3rd zero) (approx.-exact)/2t