ASYMPTOTIC METHODS

Prof. N. S. Manton

Introduction
In this course, we will be interested in finding so-called asymptotic series, leading to approximations to the
values of integrals depending on some parameter, or to the solutions of differential equations. For instance,
consider Stirling’s Approximation

 n n  1/2  
2π n n
n! ∼ (2πn)1/2 as n → ∞ or Γ(n) = (n − 1)! ∼
e n e

We will also be interested in the Stieltjes Integral: let ρ(t) be a well-behaved non-negative function that
decays sufficiently fast as t → ∞. Then
Z ∞ Z ∞
ρ(t)
dt ∼ ρ(t) dt for small x > 0
0 1 + xt 0

The modified Bessel function of order zero, I0 , solves the differential equation

d2 y 1 dy
2
+ −y =0
dx x dx

and has the following expansion with infinite radius of convergence:

∞
X x2n
I0 (x) = 2n (n!)2
∼ (2πx)−1/2 ex as x → ∞
n=0
2

Asymptotic expansions are widely used in applied mathematics and theoretical physics, e.g. when deriving
the leading-order contribution to the electrostatic potential corresponding to a charge distribution of limited
spatial extent when viewed from far away. In number theory, the prime number theorem states that the
n
number π(n) of primes at most n satisfies π(n) ∼ log n as n → ∞.

In the above examples, the symbol ∼ means that the ratio of the two sides tends to unity in the appropriate
limit. If we seek better approximations, we can derive asymptotic series, such as the Stirling series
 1/2    
2π n n 1 1
Γ(n) ∼ 1+ + + · · ·
n e 12n 288n2

where the coefficients are known, but not elementary. The coefficients get smaller for a while, but eventually
grow rapidly. The series does not converge for any n. Upon truncating, this series gives a sequence of
approximations:
 
Γ(n) Γ(n) 1 1
 1/2   − 1 = o(1),  1/2   − 1 − 12n = o n
, ... for large n
2π n n 2π n n
n e n e

For n = 4, these approximations give 3! ≈ 5.877 and 3! ≈ 5.999. For n = 4, the approximation using six
terms gives the best answer; with more terms, the error increases, since the series does not converge.

For the Stieltjes integral, we will show that, using the Taylor expansion (1 + xt)−1 = 1 − xt + x2 t2 + · · ·
valid for t < 1/x,
Z ∞ ∞ Z ∞
ρ(t) X
dt ∼ (−1)n cn xn where cn = tn ρ(t) dt
0 1 + xt n=0 0

Mathematical Tripos Part II Page 1 of 20 Asymptotic Methods — Lent 2012

for small x > 0, provided that ρ(t) → 0 as t → ∞ fast enough for the moments in the series to exist. For the
modified Bessel function, it is possible to show that

12 1 12 32 1 12 32 52 1
 
−1/2 x
I0 (x) ∼ (2πx) e 1+ + + + ···
1! 8x 2! (8x)2 3! (8x)3

For instance, the five term approximation for x = 20 has an error of about 0.0009%. As before, the series
does not converge, since the coefficients increase as n!.

The Rule of Thumb is to sum up the series as far as, but not including the smallest term. This term gives
an estimate of the error. The place to stop depends on the value at which one wants to evaluate the series.

Asymptotic series used to be important for precise evaluation of special functions, but, in that area, they have
now been superseded. Often, they give insight because they replace some sophisticated mathematical function
by elementary functions, such as exponential, trigonometric functions or powers. They play a central role in
computation techniques such as shooting methods. The behaviour of these functions can be rich, such as the
behaviour of the Airy function shown below, which exhibits both trigonometric oscillations and exponential
decay.
0.4

0.2

-20 -10 10 20

-0.2

-0.4

Another feature of asymptotic series that we will discuss in this course is Stokes’ phenomenon. Many classical
and special functions, e.g. those satisfying second-order ordinary differential equations, are entire functions
on the complex plane, so, unless they are constant, which is kind of duff, they are unbounded. We therefore
look for asymptotic series of the form

f (z) ∼ c1 es1 (z) + c2 es2 (z) as z → ∞

The functions s1 , s2 are Liouville–Green Functions, and we will discuss them later on. Interestingly, one
finds that these expansions are only valid for a restricted range of the complex argument; the constants
c1 , c2 have different values in different sectors of this kind. The sector boundaries are called Stokes Lines. In
fact, s1 , s2 are often rational functions, i.e. not entire, in which case Stokes lines are inevitable. The analytic
continuation of the asymptotic approximation fails to be an asymptotic approximation outside the sector.

Asymptotic Series
A real-valued function f (x) has an asymptotic series or asymptotic expansion around x0 , written

∞
X
f (x) ∼ an (x − x0 )n as x → x0
n=0

where the series should be understood as a formal power series, if

N
X
f (x) − an (x − x0 )n  (x − x0 )N for each N > 0 and for x − x0 small
n=0

or, more formally, if

N
an (x − x0 )n
P
f (x) −
n=0
→0 as x → x0 , for each N > 0.
(x − x0 )N

Mathematical Tripos Part II Page 2 of 20 Asymptotic Methods — Lent 2012

For a truncated asymptotic series, there is just a finite number of inequalities or limits. Note that the
definition only involves partial sums: the series itself need not converge. Also, we consider N to be fixed in
the limit x → x0 ; this is the other way round for convergent power series. Similarly, a function f (x) has an
asymptotic series about infinity, written
∞
X
f (x) ∼ an x−n as x → ∞
n=0

if it holds that
N
X
f (x) − an x−n  x−N for all N > 0, as x → ∞.
n=0

These definitions extend to complex functions, but it turns out that one needs to be more careful about
regions of validity. One often needs to take out a prefactor to be able to use the definition; for instance,

I0 (x) 1 9
−1/2
∼1+ + + ··· as x → ∞
(2πx) ex 8x 128x2

This definition involves a function and its expansion. The expansion has little meaning by itself, and the
function must be defined in its own right first, by contrast with Taylor series.

Lemma. A function has a most one asymptotic expansion about some given point, or about infinity.

Suppose that a function f (x) has two asymptotic expansions about some point x0 , the argument for expan-
sions about infinity being analogous, viz
∞
X ∞
X
f (x) ∼ an (x − x0 )n and f (x) ∼ bn (x − x0 )n
n=0 n=0

Suppose that, for some m > 0, am 6= bm , but an = bn for 0 6 n < m. Then

m m
an (x − x0 )n bn (x − x0 )n
P P
f (x) − f (x) −
n=0 n=0
→ 0, →0 as x → x0 .
(x − x0 )m (x − x0 )m

Upon subtracting these two limits, am − bm → 0, and thus am = bm , a contradiction. Hence the two
asymptotic expansions are equal.

Note that there are plenty of functions with no asymptotic series, or with only a finite asymptotic series.
For instance, it is easy to see that e−x does not have an asymptotic series around infinity, for all coefficients
in the asymptotic expansion vanish. One says that e−x is ‘small beyond all orders’ in inverse powers of x.
This also shows that two different functions can have the same asymptotic series: if f (x) has an asymptotic
series about infinity, then f (x) + e−x has the same asymptotic series about infinity.

Elementary Properties

Consider two functions f (x) and g(x) with respective asymptotic expansions

∞
X ∞
X
f (x) ∼ an (x − x0 )n and g(x) ∼ bn (x − x0 )n
n=0 n=0

about x0 . Then f (x) + g(x) and f (x)g(x) both have asymptotic expansions,

∞
X ∞
X n
X
f (x) + g(x) ∼ (an + bn )(x − x0 )n and f (x)g(x) ∼ cn (x − x0 )n where cn = ak bn−k
n=0 n=0 k=0

Mathematical Tripos Part II Page 3 of 20 Asymptotic Methods — Lent 2012

Assume furthermore that f (x) is integrable. Then
Z x ∞
X an
f (ξ) dξ ∼ (x − x0 )n+1 as x → x0
x0 n=0
n + 1

Indeed, for any N > 0 and for any ε > 0, there exists η > 0 such that, whenever |x − x0 | < ε,

Z x N Z x
X η
f (ξ) − an (ξ − x0 )n dξ 6 η |ξ − x0 |N dξ = |x − x0 |N +1
x0 n=0 x0 N +1

The claim then follows from the observation that

N N
!
Z x Z x
X an X
f (ξ) dξ − (x − x0 )n+1 = f (ξ) − an (ξ − x0 )n dξ
x0 n=0
n + 1 x 0 n=0
Z x XN
6 f (ξ) − an (ξ − x0 )n dξ
x0 n=0

Asymptotics of Integrals
As an example, we consider the Stieltjes Integral more rigorously. Let ρ(t) be a non-negative function that
does not vanish identically, and such that the moments
Z ∞
cn = tn ρ(t) dt exist for all n > 0.
0

We are interested in deriving an asymptotic series for the integral

Z ∞
ρ(t)
I(x) = dt where x > 0.
0 1 + xt

Note that there is an exact expansion

1 (−1)N +1 xN +1 tN +1
= 1 − xt + x2 t2 + · · · + (−1)N xN tN +
1 + xt 1 + xt

It follows that
N Z ∞ N +1
X t ρ(t)
I(x) = (−1)n cn xn + (−1)N +1 xN +1 dt
n=0 0 1 + xt

But 1 + xt > 1 since x > 0, and thus

Z ∞ N +1
t ρ(t)
x−N (−1)N +1 xN +1 dt 6 xcN +1 → 0 as x → 0.
0 1 + xt

We have thus derived the asymptotic expansion

Z ∞ ∞
ρ(t) X
I(x) = dt ∼ (−1)n cn xn as x → 0.
0 1 + xt n=0

One way to see that this is not a convergent Taylor expansion is to note that the left-hand side is ill-defined
for negative x, due to the pole of the integrand at t = −1/x. Alternatively, one can evaluate the moments
explicitly for simple examples, such as ρ(t) = e−t .

Integration by Parts Formulae

Integration by parts is useful for determining asymptotic expansions of integrals where the end point domi-
nates. Let f (n) (t) denote the n-th derivative of a function f (t). (We shall always, in this course, assume that

Mathematical Tripos Part II Page 4 of 20 Asymptotic Methods — Lent 2012

functions are sufficiently smooth.) Integrating by parts n times gives
Z b n
X h ib Z b
f (t)g (n) (t) dt = (−1)k−1 f (k−1) (t)g (n−k) (t) + (−1)n f (n) (t)g(t) dt
a a a
k=1

Example of an Asymptotic Expansion of an Integral

Let a < b be fixed, and let f (t) be a smooth function such that f (b) 6= 0. We are interested in obtaining an
asymptotic expansion, as x → ∞, for the integral
Z b
I(x) = f (t)ext dt
a

Observe that the endpoint t = b completely dominates in this integral. Integrating by parts gives
n
X h ib (−1)n Z b
I(x) = (−1)k−1 f (k−1) (t)ext x−k + f (n) (t)ext dt
a xn a
k=1
" n n
!#
n Z b
X (−1) X
= exb (−1)k−1 f (k−1) (b)x−k + f (n) (t)ex(t−b) dt − e−x(b−a) (−1) k−1 (k−1)
f (a)x −k
xn a
k=1 k=1

Since b − a > 0, the term in e−x(b−a) is small beyond all orders. What is more,
Z b Z b !
(n) x(t−b) (n) x(t−b) 1 − e−x(b−a)
f (t)e dt 6 max f (t) e dt = max f (n) (t) →0 as x → ∞.
a t∈[a,b] a t∈[a,b] x

We thus have, provided that f (b) 6= 0, the asymptotic expansion

" n
#
Z b X
I(x) = f (t)ext dt ∼ exb (−1)k−1 f (k−1) (b)x−k as x → ∞.
a k=1

Again, this cannot be an exact, or convergent, series, as the result is independent of a. The leading term in
this approximation is
b b
exb
Z Z
f (t)ext dt ∼ exb f (b)x−1 which arises from ext dt =
a −∞ x
More generally, we get any number of terms in this asymptotic series using a Taylor expansion of f (t) around
t = b, and integrating from −∞ to b, making an error that is small beyond all orders.

Laplace’s Integral

The ideas of the above example are generalised by the Laplace Integral. Let f (t) be a smooth function with
f (b) 6= 0. We are interested in expansions of
Z b
I(x) = f (t)exφ(t) dt
a

If φ0 (t) does not change sign on [a, b], the Laplace integral just reduces to the integral in our above example:
assume without loss of generality that φ0 (t) > 0 in [a, b]. Change variables to u = φ(t), so that
Z φ(b)
f (t) xu
I(x) = e du where t = φ−1 (u)
φ(a) φ0 (t)

Having inverted this change of variables (which is pretty straightforward in examples), we obtain an expansion
for the integral, which is rather messy in general. The leading-order approximation is

f (b) 1
I(x) ∼ exφ(b)
φ0 (b) x

Mathematical Tripos Part II Page 5 of 20 Asymptotic Methods — Lent 2012

If φ(t) attains its maximum value in [a, b], at t = b, with φ0 (b) > 0 and f (b) 6= 0 as usual, this leading-order
approximation is still valid: to leading order, the only relevant contribution to the integral comes from a
neighbourhood of t = b.
A variant of the simple Laplace integral we have been considering above is the following: suppose that f (t)
is defined by a convergent power series in some neighbourhood of t = 0, viz
∞
X
f (t) = an tn
n=0

Note that f (n) (0) = n!an . Hence, adapting our previous results, for b > 0, we have the asymptotic series
Z b ∞
X
f (t)e−xt dt ∼ n!an x−(n+1) as n → ∞
0 n=0

Watson’s Lemma

Before generalising the previous results, we need two preliminaries. First, note that, so far, we have on-
ly considered asymptotic expansions with integral powers. Some functions have asymptotic expansions in
fractional powers of their argument, of the form
∞
X
f (x) ∼ cn xα+βn as x → 0+
n=0

where β > 0. Taking out the prefactor xα , and changing variables to ξ = xβ , we recover the standard
definition of an asymptotic series.

Let us now make a few remarks about the Gamma function, which is defined for all complex numbers with
Re z > 0 by the integral
Z ∞
Γ(z) = tz−1 e−t dt
0

Integrating by parts, one obtains

Γ(z + 1) = zΓ(z)

Thus, it is possible to analytically continue the Gamma function to a meromorphic function on the entire
complex plane, with poles at z = 0, −1, −2, . . . .
1

√
Clearly, Γ(1) = 1, so Γ(n+1) = n! by induction. Another special value that is sometimes useful is Γ 2 = π;
it is obtained by evaluating a Gaussian integral.

Theorem (Watson’s Lemma). Let f (t) be a function such that f (t) is bounded on [t, 1], for any 0 < t < 1,
and that f (t) < ect at large t, for some constant c. Let f (t) have the asymptotic series

∞
X
f (t) ∼ cn tα+βn as t → 0+
n=0

where α > −1 and β > 0. Then

Z b ∞
X
f (t)e−xt dt ∼ Γ(α + βn + 1)cn x−(α+βn+1) as x → ∞
0 n=0

Observe that the conditions imposed on f (t) are sufficient for the integral to exist. Also notice that the result
is independent of b, and that the case α = 0, β = 1 corresponds to our previous result. Fix N > 0, and take
ε > 0. Then

ε b b
e−xε
Z Z Z  
−xt −xt −xt
I(x) ∼ f (t)e dt with an error f (t)e dt 6 C e dt = O
0 ε ε x

Mathematical Tripos Part II Page 6 of 20 Asymptotic Methods — Lent 2012

which is small beyond all orders as x → ∞. Taking ε to be small enough,

N
!
Z ε X Z ε 
α+βn −xt α+β(N +1) −xt
I(x) ∼ cn t e dt + O t e dt
0 n=0 0

Watson’s Lemma then follows from the fact that, up to an error that is exponentially small,
Z ε Z εt Z ∞
α+βn −xt −(α+βn+1) α+βn −u −(α+βn+1)
t e dt = x u
| {z e } du ∼ x g(u) du = Γ(α + βn + 1)x−(α+βn+1)
0 0 0
g(u)

5
Γ( 12 ) Γ( 32 )
Z

−1/2
Example. I(x) = t2 + 2t e−xt dt ∼ − .
0 (2x)1/2 2(2x)3/2

Laplace’s Integral with Interior Maximum

We now consider the Laplace integral

Z b
I(x) = f (t)exφ(t) dt as x → ∞
a

where φ(t) has a smooth non-zero quadratic maximum at some c ∈ (a, b), i.e. φ0 (c) = 0 and φ00 (c) < 0, and
where f (c) 6= 0. The integral is dominated by a neighbourhood of c, in which the integral becomes Gaussian
upon writing f (t) ≈ f (c) and φ(t) ≈ φ(c) + 21 φ00 (c)(t − c)2 . The leading-order asymptotics are
Z c+ε h  i Z c+ε
00
1
(c)(t−c)2
I(x) ∼ f (c) exp x φ(c) + 21 φ00 (c)(t − c)2 dt = f (c)exφ(c) e 2 xφ dt
c−ε c−ε

1/2
where the error is exponentially small. Change variables to s = − 12 φ00 (c) (t − c), noting that φ00 (c) < 0.
It follows that

(− 12 xφ00 (c))1/2 ε ∞
f (c)exφ(c) f (c)exφ(c)
Z Z
2 2
I(x) ∼
1/2 e−s ds ∼
1/2 e−s ds
− 12 xφ00 (c) −(− 12 xφ00 (c))1/2 ε − 12 xφ00 (c) −∞

where, again, the error is exponentially small. Thus, finally,

 1/2
2π
I(x) ∼ f (x)exφ(c) x−1/2 as x → ∞
−φ00 (c)

One can similarly deal with the case where φ(t) has a quadratic maximum at one of a, b. Note that, in that
case, the above expansion has an extra factor of one half, since we only pick up half a Gaussian integral.
That case can also be dealt with by changing variables and using Watson’s Lemma, but that approach is
rather fiddly.
π
ex
Z
1
Example: The Modified Bessel Function. I0 (x) = ex cos t dt ∼ √ .
π 0 2πx
We can similarly obtain higher-order terms in the asymptotic expansions using Taylor series. We illustrate
the technique by example. Consider the integral
Z π
2 2
I(x) = e−x sin t
dt as x → ∞
0

Note that φ(t) = − sin2 t has a maximum at t = 0, and sin2 t ≈ t2 − 13 t4 near t = 0. The integral is
exponentially dominated by a region near t = 0, so, for some ε > 0,
Z ε Z ε  
−x sin2 t 2
I(x) ∼ e dt ∼ e−xt 1 + 31 xt4 dt
0 0

Mathematical Tripos Part II Page 7 of 20 Asymptotic Methods — Lent 2012

 √
Changing variables to s = xt takes the upper limit of the integral to infinity in the limit x → ∞, and we
thus obtain
√ √
π −1/2 π −3/2
I(x) ∼ x + x as x → ∞
2 8

Notice that, as before, the leading term arises from half a Gaussian integral. Observe that it is important
not to Taylor expand the leading term in the exponential. If the integral had a prefactor, the latter would
have to be Taylor expanded, too.

Fourier–Type Integrals and the Stationary Phase Method

In this subsection, we study the asymptotics of integrals of the form

Z b Z b
I(ω) = f (t)eiωt dt or, more generally, I(ω) = f (t)eiωφ(t) dt as |ω| → ∞
a a

Here, ω is a real variable, and the function φ(t) is real. Note that the first case corresponds to the Fourier
transform of a localised function. Let us remark that the real and imaginary parts of eiωt are rapidly oscillating
for large |ω|. Heuristically, we therefore expect the negative and positive parts to cancel approximately, so
that I(ω) is small.

In fact, we can make this remark more precise by recalling that a well-behaved function f : [0, 2π] → C has
a Fourier series given by
∞ Z 2π
X 1
f (t) = cn eiωt where cn = f (t)eiωt dt
n=−∞
2π 0

Parseval’s Theorem for Fourier series, as introduced in the Methods course, states that
Z 2π ∞
2 X
f (t) dt = 2π |cn |2
0 n=−∞

Now, the series on the right-hand side only converges if cn → 0 as n → ±∞. This property of the Fourier
coefficients is generalised by the Riemann–Lebesgue Lemma, which will form the basis for our study of the
asymptotics of I(ω):

Theorem (Riemann–Lebesgue Lemma). Let f : [a, b] → C be Riemann integrable on a closed interval

[a, b]. Then it holds that
!
Z b
iωt
lim f (t)e dt =0
|ω|→∞ a

Let us first remark that this result is rather obvious under the stronger assumption of f being continuously
differentiable. In that case, we may integrate by parts to obtain
b b
f (t)eiωt
 Z
1
I(ω) = − f 0 (t)eiωt dt (∗)
iω a iω a

Since continuous functions on a closed bounded interval are bounded, this implies that I(ω) → 0 as |ω| → ∞.
Similarly, the result holds for piecewise constant functions. The result follows in general, because, by definition
of Riemann integrability, any Riemann integrable function can be bounded above and below by piecewise
constant functions, the integrals of which differ by no more than ε, for any ε > 0, i.e. there is a partition of
[a, b] with associated piecewise constant functions m(t) and M (t) such that
Z b Z b
M (t) − m(t) eiωt dt 6



m(t) 6 f (t) 6 M (t) and M (t) − m(t) dt 6 ε.
a a

Mathematical Tripos Part II Page 8 of 20 Asymptotic Methods — Lent 2012

We now consider the asymptotic expansion of integrals of the form in the Riemann-Lebesgue lemma, assuming
f to be a smooth function. By (∗) above, since f 0 is smooth, the remainder integral decays as o(1/ω). Hence
we actually have an asymptotic expansion

f (b)eiωb − f (a)eiωa
I(ω) ∼ as |ω| → ∞
iω

to leading order. Notice that, unlike the case of Laplace integrals, both endpoints contribute to the asymptotic
expansion. Since f is smooth, we can integrate by parts repeatedly to obtain the infinite series
∞
X (−1)k−1 h (k−1) iωb (k−1) iωa
i
I(ω) ∼ f (b)e − f (a)e as |ω| → ∞
(iω)k
k=1

1 ∞
eiωt eiω
Z X (k − 1)!  
Example. I(ω) = dt ∼ −1 as |ω| → ∞.
0 1+t (iω)k 2k
k=0

Integrals on Infinite Intervals and Discontinuities

Suppose that the function f and all its derivatives exist, and vanish as t → ∞. Then, letting the limits in
the integral above tend to ±∞, we find that the integral
Z ∞
I(ω) = f (t)eiωt dt
−∞

vanishes beyond all orders as |ω| → ∞. This is saying that the Fourier transform of f vanishes faster than any
inverse power of ω. This is a familiar result: for example, the Fourier transform of a Gaussian is a Gaussian
(and thus vanishes exponentially fast). Similarly, if f has a finite number of discontinuities, we may split up
the integral, and thus I(ω) = O(1/ω). If f is continuous, with f 0 having finitely many discontinuities, then
I(ω) ∼ O(1/ω 2 ) and so on.

Generalised Fourier Integrals: The Stationary Phase Method

We now consider generalised Fourier integrals of the form
Z b
I(ω) = f (t)eiωφ(t) dt as |ω| → ∞
a

A generalised version of the Riemann–Lebesgue lemma shows that this integral vanishes as |ω| → ±∞,
provided that φ0 (t) is continuous, and that φ(t) is not constant on any interval of positive length. We are
interested in finding the leading asymptotics for smooth f .
The simplest case has φ0 (t) > 0 for all t ∈ [a, b]. The case where
φ0 (t) < 0 for all t ∈ [a, b] is of course completely analogous. Upon
changing variables, our previous analysis applies, and we obtain that,
to leading order,
 
1 f (b) iωφ(b) f (a) iωφ(a)
I(ω) ∼ e − e as |ω| → ∞
iω φ0 (b) φ0 (a)

The case where φ0 (c) = 0 for some c ∈ (a, b) is more interesting. In that
case, the integral is said to have a stationary phase at c. Intuitively, the
condition φ0 (c) = 0 means that cancellation is slower in a neighbourhood
of c, as illustrated on the right. We estimate
Z c+ε  
iωφ(t) 1
I(ω) = f (t)e dt + O
c−ε ω

where ε > 0. Of course, this only makes sense if we show that the leading order term in the first integral
vanishes more slowly that 1/ω. Taylor expand, writing φ(t) = φ(c) + 12 φ00 (c)(t − c)2 , where we assume that

Mathematical Tripos Part II Page 9 of 20 Asymptotic Methods — Lent 2012

φ00 (c) 6= 0. For convenience, suppose that φ00 (c) > 0. Also estimate f (t) ≈ f (c), assuming that f (c) 6= 0. We
then obtain the leading approximation
Z c+ε   12 Z ∞
1 00
(c)(t−c)2 2 2
I(ω) ∼ f (c)eiωφ(c) e 2 iωφ dt ∼ f (c)eiωφ(c) eis ds
c−ε ωφ00 (c) −∞

upon changing variables, and taking the limits of integration to infinity. The so-called Fresnel integral above
actually converges, and it can be evaluated by contour integration. More generally, one can show that
∞
Z r
2 1 π iπ/4
eixs ds = e provided that x > 0
0 2 x

Hence, we obtain the leading-order asymptotic expansion

  12   12
iωφ(c) 2π iπ/4 iωφ(c) 2π 00
I(ω) ∼ f (c)e e or I(ω) ∼ f (c)e e(iπ/4)sign(φ (c))
ωφ00 (c) ω|φ00 (c)|

if we only suppose that φ00 (c) 6= 0. Note that these leading asymptotics are O ω −1/2 , so they dominate the

effects of the endpoints at a and b. There are higher-order corrections from both the endpoints and from the
neighbourhood of c. Obtaining these higher-order terms is quite tricky.
Z ∞ iωt  π  12
e
Example. 2
dt ∼ eiπ/4 .
−∞ 1 + t ω
Of course, we can deal with similar integrals involving cosines and sines by taking real parts in the above
results. The method of stationary phase can be generalised to the case where φ has higher-order stationary
points at some c ∈ (a, b). In that case, one needs to evaluate higher-order Fresnel integrals, by choosing
appropriate contours.

The Method of Steepest Descent

In this section, we take the idea of Laplace-type integrals and the stationary phase method one step further,
by extending the results to integrals in the complex plane of the form
Z
I(x) = f (z)exφ(z) dz as x → ∞
C

where x is a real parameter. Here, C is some curve in the complex plane, and the functions f (z) and φ(z)
are analytic in a domain containing C .

Write z = p + iq, with p = Re z, q = Im z, and let φ(z) = u(p, q) + iv(p, q), where the functions u, v are
real-valued. Since φ is analytic, these functions satisfy the Cauchy–Riemann equations

∂u ∂v ∂u ∂v
= and =−
∂p ∂q ∂q ∂p

which are saying that the gradients of u and v have the same magnitude, but are orthogonal, i.e. the curves
of constant u and v are at right angles. Note that the gradient of u is perpendicular to the curves of constant
v, and vice versa.
The ‘optimal’ contour is one of ‘steepest descent’ of u: in that case, the integral becomes Laplace-like,
dominated by the part where u has its maximum. Such a path of steepest descent is parallel, but in opposite
direction, to ∇u, i.e. parallel to the lines of constant v. Thus the phase is constant on such a path.
By Cauchy’s theorem, we can deform the contour of integration. The idea of the method of steepest descent
is to deform the original contour to some other contour on which the phase is constant as far as possible.
Paths of steepest descent typically start at saddle points of u, or at infinity. Note that, since ∇2 u = 0, the
function u has no local maxima or minima.

Mathematical Tripos Part II Page 10 of 20 Asymptotic Methods — Lent 2012

 u large

u small

u small
A S B

u large

Note that paths of steepest descent are orthogonal to the contours of u. In the figure above, we have chosen
endpoints with the same phase at the saddle. We shall see that it is possible to deal with situations where
the endpoints have different phases. Note that one might think that, in the plot above, the region where u
is large dominates the integral, but there, the integrand oscillates.

To obtain the leading asymptotics, it suffices to note that the integral is Gaussian near the saddle point,
where φ0 (z) = 0. Watson’s lemma allows calculating the complete asymptotic expansion using local data at
the saddle point, but we will not go into that.

If the phases at the endpoints of the integral differ, or if they are different from the phase at the saddle
point, we need two or three paths of steepest descent, joined up at infinity.

Let us now determine the basic contribution from a simple saddle. Let C be a steepest descent path from a
saddle at z0 . Consider the integral
Z
I0 = exφ(z) dz where φ(z) = αeiβ (z − z0 )2
C

where α > 0 and −π < β 6 π. Changing variables to y = −ieiβ/2 (z − z0 ) moves the contour onto the positive
real axis. Then
Z r r
2 i π −iβ/2 π
I0 = ie−iβ/2 e−xαy dy = e =i
C0 2 xα 2φ00 (z0 )x

More generally, by Taylor expanding around a general saddle point, we obtain the leading asymptotics for a
path of steepest descent starting at a saddle point,
Z r
π
I(x) = f (z)exφ(z) dz ∼ if (z0 )eφ(z0 )x as x → ∞
C 2φ00 (z0 )x

Example: Stirling’s Formula for the Gamma Function. Applying these techniques to the Gamma function is
a bit of an overkill. Let us do it nonetheless. We have
Z ∞ Z ∞ Z ∞
Γ(x + 1) = tx e−t dt = ex log t−t dt = xx+1 ex(log t−t) dt
0 0 0

where we have changed variables to bring the integral into the general form of integrals that we have been
considering above. The meromorphic function φ(t) = log t − t has a saddle on the real axis, at t = 1, where
φ(t) viewed as a real function has its maximum.
The method of steepest descent and the Laplace method are the same in this case. Noting that there are
two contributions to the integral, one from each side of the saddle, we obtain the expansion
√ 1
Γ(x + 1) ∼ 2πxx+ 2 e−x as x → ∞

Mathematical Tripos Part II Page 11 of 20 Asymptotic Methods — Lent 2012

In fact, in this case, we can obtain higher-order terms by using Γ(x + 1) = xΓ(x), and Taylor expanding the
expansion. We thus obtain, after ploughing through a certain amount of algebra,

√
 
x+ 12 −x 1 1
Γ(x + 1) ∼ 2πx e 1+ + + ··· as x → ∞
12x 288x2

Example: A Finite-Range Fresnel Integral. Consider the finite-range Fresnel integral

Z 1
2
I(x) = eixt dt as x → ∞
0

Set φ(t) = it2 . Note that Im φ(0) 6= Im φ(1). It is easy to see that the paths of steepest descent or ascent
are hyperbolae with axes y 2 = x2 . We can thus join up any such paths by a small segment at infinity, in the
first quadrant, where the integrand decays. We thus obtain the asymptotic expansion
r
π iπ/4 i ix
I(x) ∼ e − e as x → ∞
4x 2x

The first term comes from the saddle at the origin, while the second term comes from a path of steepest
descent through 1; we obtain the second term using the usual Laplace method. Note that, to be rigorous,
we would need to check that there is no contribution of order O(1/x) from the first path.

Dispersive Waves and Group Velocity

Let us now turn to an application of the method of steepest descent. Consider a wave ei(kx−ωt) with phase
velocity ω/k. The dispersion relation that relates the frequency to the wave number is determined by the
wave equation, e.g. the scaled Schrödinger equation

∂ψ ∂2ψ
i =− 2 with dispersion relation ω = k 2
∂t ∂x

Consider a smooth wavepacket having a narrow maximum at k = k0 , defined by the integral

Z ∞
ψ(x, t) = b(k)ei(kx−ω(k)t) dk
0

so that b(k) is just the Fourier transform of ψ(x, 0). Consider the asymptotics of ψ(x, t) in the large-time limit
t → ∞, with x = V t for some constant V . Use the stationary phase approximation with φ(t) = kV − ω(k).
Note that the phase is stationary when ω 0 (k) = V . Let k 0 be a solution of this equation, and assume, for
convenience, that it is the only solution. For the Schrödinger equation, k 0 = 1
2V . Now, by the stationary
phase approximation,

0 0 A
ψ(x, t) ∼ b(k 0 )ei(k x−ω(k )t) as t → ∞, with x = V t
t1/2

where A is a constant. Hence the wave amplitude decays as t−1/2 , by dispersion (note that there is no
mechanism for energy loss) along the line x = V t, for each V . Note that the wave disturbance is maximal
when k 0 = k0 , i.e. along the line x = ω 0 (k0 )t. The group velocity cg = ω 0 (k) is a function of k.
Hence, by assumption, the wavepacket moves with speed cg (k0 ), and the width of the wavepacket increases
as t → ∞. Note that, in the case of the free Schrödinger equation, the group velocity is related to the
phase velocity c = ω/k by cg = 2c. These approximate results are to compared to the Gaussian wavepacket
considered in the Quantum Mechanics course.

The Airy Function and the Airy Integral

In the section, we first introduce the Airy function, to which we will apply the various techniques introduced
in this course: conveniently, it can be defined in terms of an integral, but also, it satisfies a differential

Mathematical Tripos Part II Page 12 of 20 Asymptotic Methods — Lent 2012

equation (and we shall come back to asymptotics of solutions of differential equations). The Airy function
is important for finding approximate solutions of the Schrödinger equation with non-zero potential. It is
defined by
Z ∞
1 1 3
Ai(x) = ei(xs+ 3 s )
ds
2π −∞

where, for the time being, x is a real variable. Note that the Airy function is real-valued for real x. The
integral converges for reasons similar to those explaining why the Fresnel integral converges. Note that we
can deform the contour, initially the real axis, to any contour such that Im s > 0 and Im s3 > 0 far away
π 2π
from the origin, i.e. such that 0 < θ < 3 or 3 < θ < π for large s, where θ = arg s. We need to impose this
restriction lest the integrand blow up.

Let us thus deform the initial contour to a new contour C 0 contained in the strict upper half-plane and
obeying the above restrictions. Observe that

d2 Ai
Z Z
1 1 3 i d  i(xs+ 1 s3 ) 
− s2 + x ei(xs+ 3 s ) ds =


2
− xAi(x) = e 3 ds = 0
dx 2π C0 2π C ds

Note that we needed to deform the contour, because taking the derivative inside the integral is not well-
defined when we integrate along the real axis. Hence the Airy function is a solution of the simple differential
equation
d2 y
= xy
dx2

This is sometimes called the Airy equation. It has two linearly independent solutions, the Airy function and
the ‘Bairy function’ Bi(x). By the standard method for series solutions, we find that the general solution of
the above has the Taylor series about x = 0,
∞
X an−3
y(x) = an xn where a0 , a1 are arbitrary, a2 = 0, and an = for n > 3,
n=0
n(n − 1)

so the general solution is

x3 x6 x4 x7
   
y(x) = a0 1 + + + · · · + a1 x + + + ···
6 180 12 504
This series has an infinite radius of convergence. The values of the constants a0 and a1 for the Airy function
are determined by Ai(0) and Ai0 (0). One shows that

3−1/6 1

31/6 2


a0 = Γ 3 and a1 = − Γ 3
2π 2π

Asymptotics of the Airy Function

Let us first consider asymptotics of the Airy function for large positive x. Shift the contour to a contour in
the strict upper half-plane as before. Reparametrise the contour by setting s = x1/2 t, and letting y = x3/2 .
Then

y 1/3
Z Z
1 1 3 1 3
Ai(x) = ei(xs+ 3 s )
ds =⇒ Ai y 2/3 = eiy(t+ 3 t )
dt
2π C0 2π C0

Use the method of steepest descent, using the saddle point at i. Steepest descent contours through this saddle
have Im φ = 0, and hence, if t = p + iq, we have p = 0 or q 2 = 1 + 13 p2 along paths of steepest descent.
π
The second path is a hyperbola, the asymptotes of which are at angles of magnitude 6 to the real axis, and
therefore lie in the appropriate sector. We thus obtain

y −1/6 2
Ai y 3/2 ∼ √ e− 3 y as y → ∞
2 π

Mathematical Tripos Part II Page 13 of 20 Asymptotic Methods — Lent 2012

and thus, upon reintroducing the original variable,

x−1/4 − 2 x3/2
Ai(x) ∼ √ e 3 as x → ∞
2 π

Hence the Airy function decays exponentially as x → ∞. This is as expected, because the phase of the
integrand has no points of stationary phase along the real axis.

Let us now seek the asymptotics of Ai(−x) as x → ∞. This is in fact much easier, because the integrand
now has points of stationary phase on the real axis. Making the same change of variables as before,

∞   21 h

y 1/3 y 1/3
Z
1 3 π 2 π 2 π
i
Ai −y 2/3 = eiy(−t+ 3 t )
dt ∼ e−i( 3 y− 4 ) + ei( 3 y− 4 ) as y → ∞
2π −∞ 2π y

It follows that, collecting the complex exponentials into cosines

x−1/4
 
2 3/2 π
Ai(−x) ∼ √ cos x − as x → ∞
π 3 4

As expected, this is a real result. The function now exhibits oscillatory decay; the decay is however no longer
exponential. Note that the frequency of the oscillations increases as x → ∞. This result is in fact not strictly
right, for the cosine on the right-hand side may vanish, in which case we would need to go to higher-order
terms in the series. Notice that this series can be used to find approximations to the zeros of the Airy
function; the effect of the higher-order terms is to slightly change the position of these zeros. Using these
asymptotics, we can sketch the Airy function:

0.4

0.2

-20 -10 10 20

-0.2

-0.4

Note that the Airy equation looks like a (scaled) Schrödinger equation with linear potential V (x) = x. Hence
the Airy function is the stationary state wavefunction for a particle coming in from the left and bouncing
off the potential V (x) = x, being totally reflected in the process.

Asymptotics of Solutions of Differential Equations

In this section, we consider the asymptotics of solutions of ordinary differential equations. Consider a general
second-order differential equation

d2 y dy
+ p(x) + q(x)y = 0
dx2 dx

Use dashes to denote differentiation. Recall the Frobenius series in the neighbourhood of an ordinary point
or a regular singular point x0 , of the form

y(x) = (x − x0 )α a0 + a1 (x − x0 ) + a2 (x − x0 )2 + · · ·

There is at least one solution of this type; possibly, there is a logarithmic solution, too. The expansion near
1
an irregular singular point at x0 is more difficult. This occurs when p(x) is more singular than x−x0 , i.e. the
1
singularity is more complicated than a simple pole, and/or q(x) is more singular than (x−x0 )2 .

Mathematical Tripos Part II Page 14 of 20 Asymptotic Methods — Lent 2012

The Liouville–Green Method

The Liouville–Green Method is a method of obtaining approximate solutions in the neighbourhood of an

irregular singular point. The basic idea is to try a solution of the form y(x) = eS(x) . Substituting into the
differential equation, we obtain the exact result

S 00 + S 02 + p(x)S 0 + q(x) = 0

We assume that S 00  S 02 . The idea is to make this approximation, and to check later on that the solution
one obtains indeed satisfies this inequality. Note that, if S(x) = a(x − x0 )−b , for some b > 0, then we have
S 0 (x)2 = a2 b2 (x − x0 )−2b−2 , while S 00 (x) = ab(b + 1)(x − x0 )−b−2 , so S 00  S 02 near the singularity, as
required for our approximation. Making this approximation, we have

S 02 + p(x)S 0 + q(x) = 0

This is a quadratic equation for S 0 , which we can solve in principle. We can then integrate to get an
approximate solution of the differential equation.

Example. Let us illustrate this result for y 00 = y/x3 . This has an irregular singular point at the origin. We
−1/2 −1/2
obtain approximate solutions y ≈ Ae2x and y ≈ Be−2x . The first solution is exponentially large, while
the second one is exponentially suppressed. It is therefore not very useful to consider a linear combination
of these solutions, for the error in the first solution is very much larger than the second solution. However,
if we require y to be regular at x = 0, the second solution gives a useful approximation.

Improved Approximations
Let us now try to get a better approximation, and in fact get the leading-order asymptotic behaviour. We
will illustrate the method using the above example. Look for corrections of the form
2
S(x) = + C(x) where C(x)  x−1/2
x1/2
in the limit x → 0+ . Note that we have chosen one of the two possible solutions. Substituting into the exact
equation gives
3 −5/2 3 −5/2
x + C 00 − 2x−3/2 C 0 + C 02 = 0 =⇒ x − 2x−3/2 C 0 ≈ 0
2 2
where we make the approximations C 0  x−3/2 , so that C 0 x−3/2  C 02 , and C 00  x−5/2 . In general, one
cannot of course differentiate inequalities in this sense, but we aim to check retrospectively whether our
approximations make sense. Rearranging the above and integrating yields C 0 = log x4/3 up to a constant of
integration, and thus, we have obtained that

−1/2
y(x) ∼ Ax3/4 e2x as x → 0+

where A is a constant. This is in fact an asymptotic approximation, and one can obtain further terms in the
3
asymptotic series by setting C(x) = 4 log x + D(x), where D(x)  log x. It turns out, that, in fact, D(x) is
finite as x → 0 , and all further corrections decay as x → 0+ , and so we get a proper asymptotic series with
+

the above as prefactor.

This Liouville–Green approach outlined above has wide validity, but justifying these results rigorously is
rather difficult.

The Liouville–Green Method for Irregular Singular Points at Infinity

Changing variables to t = 1/x, the general second-order ordinary differential equation considered above
becomes
d2 y
 
2 P (t) dy Q(t) 1

1


+ − 2 + 4 y = 0 where P (t) = p t , Q(t) = q t
dt2 t t dt t

Typically, i.e. unless p(x) and q(x) decay rapidly as x → ∞, this equation has an irregular singular point at
t = 0, i.e. at infinity in x-space.

Mathematical Tripos Part II Page 15 of 20 Asymptotic Methods — Lent 2012

Example: The Airy Equation. The Airy equation y 00 −xy = 0 has an irregular singular point at infinity. Rather
than expanding about the origin after changing variables, we shall develop the Liouville–Green method for
expansions about infinity.
Try a solution of the form y(x) = eS(x) . Then S 00 + S 02 − x = 0. Assuming that S 00  S 02 gives S 0 (x) = ±x1/2
(which enables us to verify our assumption), so S(x) = ± 32 x3/2 . Choose the exponentially decaying solution.
To get an asymptotic approximation, set S(x) = − 23 x3/2 +C(x), with C(x)  x3/2 . We thus obtain, similarly
to before,
2 3/2
y(x) ∼ Ae− 3 x x−1/4

However, there is now way of finding A for the Airy function by this method. Our integral asymptotics
however led to A = 12 π −1/2 .

Observe that, in these examples, we were able to solve the equation for S 0 and hence find S ‘exactly’ once we
had assumed that S 00  S 02 . In other examples, this can be messy or impossible. Instead, it is often easier
to expand S in suitable powers of x.

Equations with Small Leading Parameter: The WKB Method

Let 0 < ε  1 be a small positive parameter. In this subsection, we consider second-order ordinary differential
equations of the form
d2 y
ε2 2 = q(x)y
dx

which do not have singular points. We can still use the ideas of the Liouville–Green method developed
above; in this context, they are referred to as the WKB approximation, after Wenzel, Kramer, Brillouin and
sometimes Jeffreys, or semiclassical method. Note that, often, q(x) has a definite length scale, so we cannot
just rescale x and ignore ε.
As before, we try a solution of the form y(x) = eS(x) , so that ε2 S 00 + S 02 − q(x) = 0. Supposing as before

that S 00  S 02 , we get approximate solutions

Z x
1
S(x) = ± q(ξ)1/2 dξ
ε ?

This shows that there is a major difference between q(x) > 0 and q(x) < 0: in the first case, S(x) is real, so
the solution grows or decays; in the second case, the solution oscillates. Solutions of q(x) = 0 are referred to
as turning points. Seek an improved approximation,
Z x
1
S(x) = ± q(ξ)1/2 dξ + C(x) so that ± 12 εq −1/2 (x)q 0 (x) ± 2εq 1/2 (x)C 0 (x) + ε2 C 02 = 0
ε ?

Balancing the terms that are O(ε),

q 0 (x)
C 0 (x) = − =⇒ C(x) = log q(x)−1/4 assuming that q(x) > 0
4q(x)

Again, this solution breaks down close to turning points. It is remarkable that we can integrate this term.
Hence we obtain a solution
 Z x 
−1/4 1 1/2
y(x) ∼ A|q(x)| exp ± |q(ξ)| dξ
ε a

Note that this expansion is not however valid across or too close to a turning point, where the logarithm
behaves badly.

Connection Across a Turning Point

The question becomes: how can we match asymptotic expansions on either side of a turning point? Assume
that there is a turning point at x = a, with µ = q 0 (a) > 0 for convenience. Choose the solution which is

Mathematical Tripos Part II Page 16 of 20 Asymptotic Methods — Lent 2012

exponentially decaying for large positive x, so that

1 x
  Z 
A 1/2
exp − q(ξ) dξ if x > a


q(x)1/4  Z ε a

y(x) ∼ a 
B 1
cos |q(ξ)|1/2 dξ − γ if x < a


|q(x)|1/4 ε x


where γ is an unknown phase. The above approximations are only valid if |x − α|  ε. The idea is to note
that the slope close to x = a is approximately constant, so we can approximate the solution using the Airy
function: near x = a, q(x) ≈ µ(x − a), so

d2 y d2 y  µ 1/3
ε2 = µ(x − a)y =⇒ = ty where t = (x − a)
dx2 dt2 ε2

Comparing the above asymptotics with those of the Airy function as t → ±∞ implies that we can have the
π
expansions match up provided that B = 2A and γ = 4.

Bounded Solutions with Two Turning Points

Now assume that q(x) has two turning points at a, b, where a > b. We assume that the solution of the
differential equation decays exponentially in x > a and x < b. The resulting problem is essentially an
eigenvalue problem. Approximating the solutions for x < a and x > b by their asymptotic series,
 Z a  Z x
D0
 
D 1 1/2 π 1 1/2 π
y(x) ∼ cos q(ξ) dξ − and y(x) ∼ cos q(ξ) dξ −
|q(x)|1/4 ε x 4 |q(x)|1/4 ε b 4

where D, D0 are constants. The expansions overlap and must thus agree. Note that requiring the arguments
to be equal does not work, for the arguments have oppositive derivatives with respect to x. However, we
may require
 Z a   Z x  Z a
1 1/2 π 1 1/2 π 1
|q(ξ)|1/2 dξ = n + 1


|q(ξ)| dξ − = nπ − |q(ξ)| dξ − =⇒ 2 π
ε x 4 ε b 4 ε b

for some n ∈ {0, 1, 2, . . . }, provided that D0 = (−1)n D. If q(x) satisfies this integral relationship, there is a
bounded solution y(x). Note that the amplitude of the oscillations is largest close to the turning points. This
result is often useful for all natural values of n, but really, to get proper asymptotics, n needs to be suitably
large, because of the 1ε factor on the left-hand side above.

Application: Bound States in Quantum Mechanics

Consider the stationary Schrödinger equation in one-dimensional quantum mechanics, for a particle of mass
m and energy E moving in a potential well V (x),

h̄2 d2 ψ
− + V (x)ψ = Eψ
2m dx2


Taking h̄ as our small parameter, and writing q(x) = 2m V (x) − E , and requiring the bound states to
be normalisable, i.e. exponentially decaying in the region where q(x) > 0, the eigenvalues are, in the WKB
approximation, given by
√ Z a Z b
2m
1/2 1

1/2 h̄π
n + 12



En − V (x) dx = n + 2 π =⇒ En − V (x) dx = √
h̄ b a 2m

This is the Bohr–Sommerfeld Quantisation Condition. Note that a, b are functions of E. The WKB approxi-
mation is generally particularly good for large n, in the so-called semi-classical regime. Note that, classically,
|q(x)|1/2 is the momentum of the particle, so, in the WKB approximation,
I
p dx = (2n + 1)h̄π

Mathematical Tripos Part II Page 17 of 20 Asymptotic Methods — Lent 2012

where the integral is the integral over one classical orbit of the particle, i.e. the area of phase space corre-
sponding to one orbit. The local wavelength satisfies 2π
λ = h̄p , which rearranges to the well-known de Broglie
relation λ = 2πh̄/p. Roughly, there is one quantum state per area 2πh̄ of phase space.
Finally, the probability density averaged over one wavelength is proportional to 1/p, as expected. This is
proportional to the classical time spent in a spatial interval. Note that, as expected, the probability density
is larger close to the turning points, as expected from the classical limit.

Example: The Quantum Harmonic Oscillator. The classical Hamiltonian for a harmonic oscillator is given
p2
by H = 2m + 12 mω 2 x2 . Hence the orbits in phase space are the curves

1 2 mω 2 2
p + x =E
2m 2

1/2
These are ellipses of semi-axes (2mE)1/2 and 2E/mω 2 , and therefore of area 2πE/ω. Plugging this into
the Bohr-Sommerfeld conditions gives
1


En = n + 2 h̄ω

Hence the WKB approximations gives the exact result for the energy levels of the quantum harmonic
oscillator. This is why the WKB method gives a good approximation to all the energy levels, and not only
the high energy levels, for potentials that look roughly like the potential of a harmonic oscillator. By contrast,
the approximations for the energy levels in an infinite square well are not great, because the potential is
infinite at the walls.

Stokes’ Phenomenon
Suppose that f (z) ∼ g(z) as |z| → ∞. Such an asymptotic expansion is generally only valid in a limited
sector or range of arg z, because an exponentially small error, or difference may become exponentially large
when crossing a sector boundary.
For instance, e−z is exponentially small for − π2 6 arg z 6 π
2, but exponentially large for π
2 6 arg z 6 3π
2 .
Such a boundary is called a Stokes line.

Stokes Lines for Liouville–Green Solutions

Consider a Liouville–Green solution f (z) ∼ AeS1 (z) + BeS2 (z) as |z| → ∞, for a second-order linear ordinary
differential equation. Suppose that, in some sector, Re S2 (z)  Re S1 (z), where, for once, signs matter in
the comparison.
In that case, we say that the first term in the asymptotic solution is dominant, while the second term is
recessive and is absent from the asymptotics.
In this setup, we define a Stokes line to be where Re S1 (z) = Re S2 (z). The dominant and recessive terms may
change across a Stokes line. Even keeping A, B, which is good near a Stokes line, the asymptotics expansion
above is still not valid for all values of arg z, because S1 (z) and S2 (z) are not usually entire functions, even
if f is.
The coefficient of the recessive exponential therefore needs to jump in places. The best place for such a jump
is where the real parts of S1 (z) and S2 (z) are maximally different. This occurs on the anti-Stokes lines. They
are often where the imaginary parts of S1 (z) and S2 (z) are equal.

Example: The Airy and Bairy Functions

2 3/2
We have seen previously that the asymptotics of the solutions of y 00 = zy are dominated by e 3 z and
− 32 z 3/2 π 5π 2π 4π
e for large |z|. Stokes lines are therefore arg z = 3 , π, 3 , and anti-Stokes lines are arg z = 0, 3 , 3 .
This requires some care, because the square root functions are not defined on the entire complex plane. The
general solution of the Airy equation can be shown to have the asymptotic expansion
   
2 3/2 5 2 3/2 5
y(z) ∼ Az −1/4 e− 3 z 1 − z −3/2 + · · · + Bz −1/4 e 3 z 1 + z −3/2 + · · ·
48 48

Mathematical Tripos Part II Page 18 of 20 Asymptotic Methods — Lent 2012

The Airy function and the Bairy function (the former of which is exponentially decaying, while the latter is
exponentially increasing as z → ∞) have the expansions
)  
Ai(z) 2 3/2 π
∼ π −1/2 |z|−1/4 cos |z| ∓ as z → −∞
Bi(z) 3 4

The values of the constants A, B for the Airy function and the Bairy function are as follows in the different
sectors. (Again, because the square root function is not defined on the entire complex plane, this requires
some care.) The expansions shown below are valid between the (dashed) anti-Stokes lines (or slightly beyond
them), and overlap in the sectors defined by the Stokes lines. Notice that the Airy function is special: since
it decays exponentially as z → ∞, we must have B = 0 along the positive real axis, so neither A or B can
jump across this anti-Stokes line, and there are only two regimes, rather than three.
i
A= √
2 π A= 1
√
2 π
B= √1
π B=0
i
A= 2
√
π A= 1
√
2 π
Bi(z) Ai(z)
1
B= √
2 π B= i
√
2 π

A = − 2√i π A= 1
√
2 π
B= √1
π B=0

The Airy function also has applications in optics. In ray optics, rays can form a caustic: rays that are
refracted by a lens can be tangent to a curve, the caustic, in such a way that on one side of the ray, there
are two rays through any point, whereas there are no rays on the other side of the caustic. An example of
a caustic in everyday life is a rainbow; note that light is refracted at different angles for different colours,
thence the colouring of the rain bow. The intensity of the light is strongest at the caustic, where it has a
inverse square-root singularity. In wave optics, one ends up with a wave amplitude
Z ∞
1 3
Ψ(ξ) = 2πAi(ξ) = ei(ξs+ 3 s )
ds
−∞

where ξ is some coordinate. The rays correspond to stationary phase points with ξ + s2 = 0. Hence there
are no real solutions for ξ > 0, and two real solutions for ξ < 0. Interestingly, there are additional peaks, or
supernumerary bows, apart from the caustic. By contrast to the caustic predicted by ray theory, the Airy
function smoothes out the caustic.
Sometimes, it is possible for two caustics to meet at a cusp, diving the plane into a region where there is
one ray through each point, and a region where there are three rays through each point. The corresponding
amplitude is represented by a Pearcey function
Z ∞
2
1
+ 14 t4 )
Ψ(ξ, η) = ei(ηt+ 2 ξt dt
−∞

where ξ and η are coordinates. The phase of the integral is stationary when η + ξt + t3 = 0. For real ξ, η, it

1/2
follows that there are three real solutions if |η| 6 4ξ 3 /27 , and one otherwise. This leads to a cusp.

Example: The Modified Bessel Function

Consider the modified Bessel function I0 (z), which is an entire solution of the second-order ordinary diffe-
rential equation zy 00 + y 0 z − y = 0.
Liouville–Green solutions to this differential equations are asymptotic to (2πz)−1/2 ez and (2πz)1/2 e−z as
|z| → ∞. The asymptotic solution for I0 turns out to be
   
1 1
I0 (z) ∼ (2πz)−1/2 ez 1 + + · · · + i(2πz)−1/2 e−z 1 − + ··· as |z| → ∞, for − π2 < arg z < 3π
2
8z 8z

Mathematical Tripos Part II Page 19 of 20 Asymptotic Methods — Lent 2012

This is essentially J0 on the imaginary axis. Notice that we need both terms close to the Stokes line
π
arg z = 2. The asymptotic result is valid except in a small neighbourhood of arg z = − π2 , but the asymptotic
approximations get poorer once we cross the anti-Stokes lines arg z = 0 and arg z = π.

Asymptotics Beyond all Orders

In the previous sections, we have derived results that are valid as |z| → ∞. What about actually using these
expansions to get approximations results for small finite values of z?
Let us illustrate this for the Airy function, in the range 0 6 arg z 6 π. The graphs below suggest that we
need to include the recessive term for accuracy at finite |z|, despite this being completely negligible in the
2π
limit |z| → ∞. Notice that the asymptotic expansion in the sector 0 6 arg z 6 3 used for the first graph
becomes increasingly bad as we approach arg z = π. Also notice that optimal truncation gives are large gain
in accuracy.

Let us make some remarks on the magnitude of the erros in truncated asymptotic series. Divergent series
often have ‘late’ terms of the form n!X −n . Optimal truncation is at the smallest term, where n/X ≈ 1, i.e.
n ≈ X. Hence we estimate the truncation error as X!/X X ≈ X 1/2 e−X using Stirling’s approximation.
Note that, for the Airy function, X = 23 z 3/2 , so the error in the dominant series is exponentially small, and
comparable with the recessive term.
This suggests as above that one can improve the asymptotic approximation by optimal truncation and
inclusion of the recessive term. However, the coefficient of the recessive term jumps. It was only recently
discovered that there is a something better than Stokes jumps: use the optimally truncated dominating and
recessive series with coefficients A and B, and replace the discontinuous |B| by an error function, i.e. an
integral of Gaussian, with width depending on |z|: for large |z|, the Gaussian is narrow, for smaller |z|, it is
wide. This greatly increases the accuracy of the ‘hyperasymptotic’ approximation.

Mathematical Tripos Part II Page 20 of 20 Asymptotic Methods — Lent 2012