Appendix D

Asymptotic series

Asymptotic series play a crucial role in understanding quantum field theory, as Feynman diagram
expansions are typically asymptotic series expansions. As I will occasionally refer to asymptotic
series, I have included in this appendix some basic information on the subject.

See [AW] sections 5.9, 5.10, 7.3 (7.4 in 5th edition), 8.3 (10.3 in 5th edition) for some of the
material below.

D.1 Definition

By now as graduate students you have seen infinite series appear many times. However, in most
of those appearances, you have probably made the assumption that the series converged, or that
the series is only useful when convergent.

Asymptotic series are non-convergent series, that nevertheless can be made useful, and play an
important role in physics. The infinite series one gets in quantum field theory by summing
Feynman diagrams, for example, are asymptotic series.

To be precise, consider a function f (z) with an expansion as

A1 A2
f (z) = A0 + + 2 + ···
z z
where the Ai are numbers. We can think of the series Ai /z i as approximating f (z)/ϕ(z) for
P

large values of z.

We say that the series Ai /z i represents f (z) asymptotically in direction eiφ if, for a given n, the
P

first n terms of the series may be made as close as desired to f (z) by making |z| large enough
with arg z fixed to φ, i.e.  
n
A p
lim z n f (z) −
X
p
= 0
|z|→∞
p=0
z

1
APPENDIX D. ASYMPTOTIC SERIES 2

(Put another way, write z = reiφ , then take the limit as r → ∞ but hold φ fixed.) We shall see
later that as one varies the direction eiφ , one can get different asymptotic series expansions for the
same function – this is known as Stokes’ phenomenon, and we shall study it in section *** CITE
***.

Asymptotic series need not converge; in fact, in typical cases of interest, an asymptotic series will
never converge. (Nonconvergence is sometimes added to the definition of asymptotic series, so
that, in that alternate definition, an asymptotic series can never converge. In our definition here,
convergence is allowed, albeit it is unusual.)

It is important to note that asymptotic series are distinct from convergent series: a convergent
series need not be asymptotic. For example, consider the Taylor series for exp(z). This is a
convergent power series, but the same power series does not define an asymptotic series for
exp(z). After all,  
n
zp
lim z n exp(z) −
X
 → ∞
z→∞
p=0
n!

and so the series is not asymptotic to exp(z), though it does converge to exp(z).

Not all functions have an asymptotic expansion; exp(z) is one such function. If a function does
have an asymptotic expansion, then that asymptotic expansion is unique. However, several
different functions can have the same asymptotic expansion; the map from functions to
asymptotic expansions is many-to-one, when it is well-defined.

Example: Consider the function

Z ∞ exp(−t)
−Ei(−x) = E1 (x) = dt
x t
We can generate a series approximating this function by a series of integrations by parts:

exp(−x)
Z ∞ exp(−t)
E1 (x) = − dt
x x t2
(−)n n!
Z ∞
exp(−x) 1 2! 3! exp(−t)
 
n+1
= 1 − + 2 − 3 + ··· n
+ (−) (n + 1)! dt
x x x x x x tn+2

The series ∞
n!
(−)n
X

n=0
xn
is not convergent, in any standard sense. For example, when x = 1, this is the series

1 − 2! + 3! − 4! + · · ·

More generally, for any fixed x the magnitude of the terms increases as n grows, so this
alternating series necessarily diverges.
APPENDIX D. ASYMPTOTIC SERIES 3

However, although the series diverges, it is asymptotic to E1 (x)x exp(x). To show this, we must
prove that for fixed n,
 
n
exp(−x) X (−)n n!
lim xn E1 (x) −  = 0
x→∞ x p=0
xn

Using our expansion from successive integrations by parts, we have that the limit is given by
 Z ∞ exp(−t)

lim xn (−)n+1 (n + 1)! dt
x→∞ x tn+2
We can evaluate this limit using the fact that
Z ∞ exp(−t) 1
Z ∞ exp(−x)
n+2
dt < n+2 exp(−t)dt =
x t x x xn+2
Since
xn (n + 1)! exp(−x)
lim = 0
x→∞ xn+2
we see that the series is asymptotic.

Example: Consider the ordinary differential equation

1
y′ + y =
x
The solutions of this ODE have an asymptotic expansion, as we shall now verify.

To begin, assume that the solutions have a power series expansion of the form
∞
X an
y(x) =
n=0
xn

for some constants an . Plugging this ansatz into the differential equation above and solving for
the coefficients, we find

a0 = 0
a1 = 1
a2 = a1 = 1
a3 = 2a2 = 2
a4 = 3a3 = 3!
a5 = 4a4 = 4!

and so forth, leading to the expression

∞
X (n − 1)!
y(x) =
n=1
xn
APPENDIX D. ASYMPTOTIC SERIES 4

First, let us check convergence of this series. Apply the ratio test to find
n!/xn+1 n
lim = lim −→ ∞
n→∞ (n − 1)!/xn n→∞ x

In particular, by the ratio test, this series diverges for all x (strictly speaking, all x for which it is
well-defined, i.e. all x 6= 0).

We can derive this asymptotic series in an alternate fashion, which will explain its close
resemblance to the previous example. Recall the method of variation of parameters for solving
inhomogeneous equations: first find the solutions of the associated homogeneous equations, then
make an ansatz that the solution to the inhomogeneous equation is given by multiplying the
solutions to the homogeneous solutions by functions of x. In the present case, the associated
homogeneous equation is given by
y′ + y = 0
which has solution y(x) ∝ exp(−x). Following the method of variation of parameters, we make
the ansatz
y(x) = A(x) exp(−x)
for some function A(x), and plug back into the (inhomogeneous) differential equation to solve for
A(x). In the present case, that yields
1
A′ exp(−x) =
x
which we can solve as Z x
exp(t)
A(x) = dt
−∞ t
(Note that I am implicitly setting a value of the integration constant by setting a lower limit of
integration. Also note that the integral above is ill-defined if x is positive, a matter I will gloss
over for the purposes of this discussion.) Thus, the solution to the inhomogeneous equation is
given by Z x
exp(t)
y(x) = exp(−x) dt
−∞ t
whose resemblance to the previous example should now be obvious.

D.2 The gamma function and the Stirling series

**** See also 0711.4412, which claims to have a very simple proof of Stirling’s formula.

An important example of an asymptotic series is the asymptotic series for the gamma function,
known as the Stirling series. The gamma function is a meromorphic function on the complex
plane that generalizes the factorial function. Denoted Γ(z), it has the properties
Γ(z + 1) = zΓ(z)
√
Γ(1/2) = π
Γ(1) = 1
Γ(n + 1) = n! for n a positive integer
APPENDIX D. ASYMPTOTIC SERIES 5

Because of that last property, because the gamma function generalizes the factorial function,
people sometimes define z! ≡ Γ(z + 1) for any complex number z. The gamma function has simple
poles at z = 0, −1, −2, −3, · · ·. It also obeys numerous curious identities, including
π
Γ(z)Γ(1 − z) =
sin πz
and “Legendre’s duplication formula”
√
Γ(1 + z)Γ(z + 1/2) = 2−2z πΓ(2z + 1)

The gamma function has several equivalent definitions. It can be expressed as an integral, using
an expression due to Euler:
Z ∞
Γ(z) = exp(−t)tz−1 dt, ℜz > 0
0

It can also be expressed as a limit, using another expression also due to Euler:
1 · 2 · 3···n
Γ(z) = lim nz , n 6= 0, −1, −2, −3, · · ·
n→∞ z(z + 1)(z + 2) · · · (z + n)

It can also be expressed as an infinite product, using an expression due to Weierstrass:

∞
1 z
Y  
= z exp(γz) 1 + exp(−z/n)
Γ(z) n=1
n

where γ is the Euler-Mascheroni constant

n
!
X 1
γ ≡ lim − log n
n→∞
m=1
m
≈ 0.577215661901 · · ·

where we use log to denote the natural logarithm. (The regions of validity of each definition are
slightly different; analytic continuation defines the function globally.)

We can also define the digamma and polygamma functions, which are various derivatives of the
gamma function. The digamma function, denoted either ψ or F , is defined by
d
ψ(z + 1) ≡ F (z) ≡ log Γ(z + 1)
dz
It can be shown that
∞
X z
ψ(z + 1) = −γ +
n=1
n(n + z)

The polygamma function ψ (n) is defined to be a higher-order derivative:

dn+1
ψ (n) (z + 1) ≡ F (n) (z) ≡ log Γ(z + 1)
dz n+1
APPENDIX D. ASYMPTOTIC SERIES 6

The Stirling series for the gamma function is derived using the Euler-Maclaurin integration
formula, which we shall digress briefly to explain. (See section 5.9 of [AW] for this background,
and section 8.3 on the resulting asymptotic series, known as the Stirling series.)

First, we derive the Euler-Maclaurin integration formula, which we will use to derive an
asymptotic series for the gamma function. Consider the integral
Z 1 Z 1
f (x)dx = f (x)B0 (x)dx
0 0

where B0 (x) is a Bernoulli function (**** SEE section **** ). The idea is to use the relation
Bn′ (x) = nBn−1 (x) between the Bernoulli functions and successive integrations by parts to
generate an infinite series. At the first step, use B1′ (x) = B0 (x) and integrate by parts to get
Z 1 Z 1
f (x)dx = f (x)B1′ (x)dx
0 0
Z 1
= f (x)B1 (x)]10 − f ′ (x)B1 (x)dx
0
1 1 1 ′
Z
= (f (1) + f (0)) − f (x)B2′ (x)dx
2 2 0
1 1 ′ 1 1 ′′
Z
1
= (f (1) + f (0)) − f (x)B2 (x) 0 + f (x)B2 (x)dx
2 2 2 0
and so forth. Continuing this process, we find
1 q
1 B2p  (2p−1)
Z 
(1) − f (2p−1) (0)
X
f (x)dx = [f (1) + f (0)] − f
0 2 p=1
(2p)!
1
Z 1
+ f (2q) (x)B2q (x)dx
(2q)! 0

where the B2p in the sum above are the Bernoulli numbers, not the functions, and where we have
used the relations
B2n (1) = B2n (0) = B2n
B2n+1 (1) = B2n+1 (0) = 0
(Compare equation (5.168a) in [AW].)

***** Move that to earlier?

By replacing f (x) with f (x + 1) we can shift the integration region from [0, 1] to [1, 2], and by
adding up results for different regions, we finally get the Euler-Maclaurin integration formula:
Z n 1 1
f (x)dx = f (0) + f (1) + f (2) + · · · + f (n − 1) + f (n)
0 2 2
q
B2p  (2p−1) 
(n) − f (2p−1) (0)
X
− f
p=1
(2p)!
n n−1
1
Z
f (2q) (x + s)dx
X
+ B2q (x)
(2q)! 0 s=0
APPENDIX D. ASYMPTOTIC SERIES 7

(Compare equation (5.168b) in [AW].)

***** In the last integral, should the limits be 0 and n, or 0 and 1 ?

Apply the Euler-Maclaurin integration formula above to the right-hand side of the equation
1
Z ∞ 1
= dx
z 0 (z + x)2
to get the series
∞
1 1 X
= f (0) + f (n)
z 2 n=1
∞
B2n  
lim f (2n−1) (x) − f (2n−1) (0)
X
−
n=1
(2n)! x→∞

for f (x) = (z + x)−2 . Thus,

∞ ∞
1 1 1 B2n (2n)!
X X  
= + −
z 2z 2 n=1
(z + n)2 n=1
(2n)! z 2n+1

From equation (10.41) in [AW],

∞
1
= F (1) (z)
X
(z + n)2
n=1
so we can write ∞
(1) d 1 1 X B2n
F (z) = F (z) = − +
dz z 2z 2 n=1
z 2n+1

(Compare equation (10.50) in [AW].) Integrating once we get

∞
1 X B2n
F (z) = C1 + log z + −
2z n=1
2nz 2n

It can be shown (see [AW]) that C1 = 0. Since

d
F (z) = log Γ(z + 1)
dz
we can integrate again to get
∞
1 B2n
  X
log Γ(z + 1) = C2 + z+ log z − z +
2 n=1
(2n)(2n − 1)z 2n−1

for some constant C2 , where we have used the fact that

d
z (log z − 1) = log z.
dz
We can solve for C2 by substituting the expression above into the Legendre duplication formula
1
Γ(z + 1)Γ(z + ) = 2−2z π 1/2 Γ(2z + 1)
2
APPENDIX D. ASYMPTOTIC SERIES 8

(this corrects a minor typo in equation (10.53)) from which one can derive that
C2 = (1/2) log(2π). Thus,
∞
log 2π 1 B2n
  X
log Γ(z + 1) = + z + log z − z +
2 2 n=1
(2n)(2n − 1)z 2n−1

which is Stirling’s series, an asymptotic series for the natural logarithm of the gamma function.

We can also derive a more commonly used expression for Stirling’s series by exponentiating the
series above. We get

√ ∞
!
B2n
2πz z+1/2 exp(−z) exp
X
Γ(z + 1) =
n=1
(2n)(2n − 1)z 2n−1

We can simplify the last factor as follows. Recall the Taylor expansion

x2 x3 x4
log (1 + x) = x − + − + ···
2 3 4
If we find an x such that
∞ n ∞
n+1 x B2n
X X
(−) =
n=1
n n=1
(2n)(2n − 1)z 2n−1

then we can write

∞
!
X B2n
exp = 1+x
n=1
(2n)(2n − 1)z 2n−1
Although finding a closed-form expression is impossible, we can find a series in z for x. From the
first terms, clearly
B2
x = + O(z −2 )
2z
and if we work out the expansion more systematically, we discover

B2 B2
x = + 22 + O(z −3 )
2z 8z
1 1
= + + O(z −3 )
12z 288z 2
Thus, we have that
√ z+1/2

1 1

Γ(z + 1) = 2πz exp(−z) 1 + + + O(z −3 )
12z 288z 2
another form of Stirling’s asymptotic series.
APPENDIX D. ASYMPTOTIC SERIES 9

D.3 Watson’s lemma

Given a power series that converges in some finite domain, when one integrates the power series
term-by-term along a contour that extends outside of the domain, although the result is no longer
convergent, nevertheless it is often still sensible as an asymptotic series.

Formally, this result is known as Watson’s lemma:

Suppose that f (t) is a possibly multivalued function of t, which has the following
expansion near t = 0:
∞
an tn/r−1 , |t| ≤ a, r > 0
X
f (t) =
n=1

Furthermore, suppose that f (t) is of, at most, exponential growth for large t, so that
for suitable positive constants K, b,

|f (t)| < Keb|t| , |t| ≥ a

Then for |z| large, the function

Z ∞
F (z) ≡ e−zt f (t)dt
0

has asymptotic series expansion

∞ Z ∞  ∞
e−zt tn/r−1 dt an Γ(n/r)z −n/r
X X
an = for | arg z| < π/2
n=1 0 n=1

Proof: Let Sk (z) be the finite sum

k
!
Z ∞
−zt n/r−1
X
Sk (z) ≡ e an t dt
0 n=1

Since the sum is finite, we can exchange the sum and integral:
k Z ∞  k
−zt n/r−1
an Γ(n/r)z −n/r
X X
Sk (z) = an e t dt =
n=1 0 n=1

To establish that the result is an asymptotic series expansion of F (t), we need to show that

lim z n/r (F (z) − Sn (z)) = 0

z→∞

for | arg z| < π/2. By virtue of the bound on f (t), there must exist a constant Ck such that
k
an tn/r−1 ≤ Ck eb|t| |t|(k+1)/r−1
X
f (t) −
n=1
APPENDIX D. ASYMPTOTIC SERIES 10

Then,
Z ∞
|z|n/r |F (z) − Sn (z)| ≤ |z|n/r Cn e−(|z| cos φ−b)t t(n+1)/r−1 dt
0
|z|n/r Cn
= Γ((n + 1)/r)
(|z| cos φ − b)(n+1)/r

where φ = arg z. For | arg z| < π/2, cos φ > 0, so for large enough |z| the expression above is
well-defined and goes to zero as |z| → ∞. 2

D.4 Method of steepest descent

Consider a contour integral of the form

Z
G(z) = g(t) exp(zf (t))dt
C

The method of steepest descent is a systematic procedure for generating an asymptotic series that
approximates integrals of this form. Briefly, the method says that for large values of z, the
dominant contribution to the integral G(z) will come from values of t such that f ′ (t) = 0, known
as saddle points, in the sense that such points will make a contribution to the integral that is
proportional to the integrand evaluated at the saddle point.

Let us first see how this works in the special case of a contour located along the real line, so that t
is real, involving real-valued functions f (t), g(t). (This will repeat an argument from section ***
CITE ****.) Let t0 be a point such that f ′ (t0 ) = 0, and for simplicity let us assume f ′′ (t0 ) < 0.
Then, locally, we can approximate the integral by a Gaussian. Expand
1
g(t) = g(t0 ) + (t − t0 )g′ (t0 ) + (t − t0 )2 g′′ (t0 ) + · · ·
2
1
f (t) = f (t0 ) + (t − t0 )f ′ (t0 ) + (t − t0 )2 f ′′ (t0 ) + · · ·
2
1 2 ′′
= f (t0 ) + (t − t0 ) f (t0 ) + · · ·
2
√
Now, define s = z(t − t0 ), so that we can write the integral as
Z ∞  
1

dt g(t0 ) + (t − t0 )g (t0 ) + · · · exp z f (t0 ) + (t − t0 )2 f ′′ (t0 ) + · · ·
′

G(z) =
−∞ 2
Z ∞
ds s ′ 1 2 ′′
   
−1 −1/2
= exp(zf (t0 )) √ g(t0 ) + √ g (t0 ) + O(z ) exp − s |f (t0 )| + O(z )
−∞ z z 2
exp(zf (t0 )) ∞ 1
Z  
= √ ds g(t0 ) exp − s2 |f ′′ (t0 )| + O(z −1 )
z −∞ 2
exp(zf (t0 )) √
= √ g(t0 ) 2|f ′′ (t0 )|π + O(z −1 )
z
APPENDIX D. ASYMPTOTIC SERIES 11

Not only do we see that the dominant contribution comes from t = t0 for large z, in the sense that
the contribution to the integral is proportional to the integrand evaluated at t0 , but we also get
an explicit expression for the contribution to the limit from the t0 point, in the same limit.

The method of steepest descent generalizes to complex integrands and contour integrals. In such
cases, the general idea is that in an integral over a complex exponential of the form exp(zf (t)),
for large z, the part of the integration contour that mostly just changes the phase will not
significantly contribute to the integral, but rather will tend to cancel out. A little more
systematically, if the integral involves integrating over all phases of the complex exponential, then
the different contributions should sum to zero, on the grounds that
Z 2π Z 2π
cos θdθ = sin θdθ = 0
0 0

At this same, level of approximation, the leading contribution to the integral should come from
parts of the contour where the phase does not change significantly. See [AW][section 7.4] or
[MF][section 4.6] for more information on the general complex case.

As described so far here, the method of steepest descent gives the leading-order term in an
asymptotic series expansion of an integral, but can also be generalized to give higher-order terms
in the same expansion.

One application will be to the gamma function. Recall the definite integral description of the
gamma function: Z ∞
Γ(z + 1) = exp(−t)tz dt
0
Change integration variables to t = τ z:
Z ∞
z+1
Γ(z + 1) = z exp(−zτ z)τ z dτ
0

and write
exp(−τ /z)τ z = exp(−zτ + z log τ )
giving an integral of the form Z
exp(zf (t))dt
C
with f (t) = −t + log t.

Another application of the method of steepest descent is to the Feynman path integral description
of quantum mechanics and quantum field theory, where it is used to recover the classical limit.

Let us describe how to derive this leading contribution in the case of the gamma function, then
we shall describe how to more systematically use these general ideas to create an asymptotic
expansion for such contour integrals.

In the case of the gamma function, recall

Z ∞
z+1
Γ(z + 1) = z exp(zf (t))dt
0
APPENDIX D. ASYMPTOTIC SERIES 12

for f (t) = log t − t. Solving ∂f /∂t = 0, we find that the only possible saddle point is at
1
− 1 = 0
t
i.e. t0 = 1. Let us expand f (t) about this saddle point. Write t = 1 + x, then
f (t) = log(1 + x) − (1 + x)
!
x2 x3 x4
= x − + − + ··· − (1 + x)
2 3 4
x2 x3
= −1 − + − ···
2 3
from which we see the leading order approximation to the integral should be proportional to
exp(−z). We can now approximate the gamma function by
Z ∞
z+1
Γ(z + 1) = z exp(zf (1)) exp(zf ′′ (1)x2 /2!)dx
−1
Z ∞
z+1
= z exp(−z) exp(−zx2 /2)dx
−1
Z ∞
∼
= z z+1 exp(−z) exp(−zx2 /2)dx
−∞
r
2π
= z z+1 exp(−z)
√ z
z
= 2πzz exp(−z)
which is the leading term in Stirling’s expansion of the factorial function.

Reference [AW], in section 7.4, discusses how to treat a contour integral in which the contour does
not lie along the real line. An important part of the treatment, which I will omit here but which
is discussed in [AW], involves replacing the contour integral with an integral along an infinite line
that is tangent to a particular direction at the saddle point.

Let us instead outline how to more systematically derive an asymptotic series, not just the
leading term, using these methods. In terms of our original function G(z), if our contour crosses a
single saddle point t0 , and the contour line is along a “path of steepest descent” along which the
imaginary part of zf (t) is constant. Write
f (t) = f (t0 ) − w2
for some variable w, which is real Im f (t) = Im f (t0 ) everywhere along the contour. Then,
dt
Z  
2
G(z) = exp(zf (t0 )) exp(−zw ) dw
C dw
Assume that the contour C is such that the w integral can be taken to run over the real numbers
from −∞ to ∞. Next, we need to write dt/dw as a function of w, rather than t. In general, we
can accomplish such an inversion at the power-series level, so write
∞
dt
an wn
X
=
dw n=0
APPENDIX D. ASYMPTOTIC SERIES 13

for some constants an . (Note that from the definition of w above, one only expects even powers of
w to appear in this power series.) Substituting in one has
Z ∞ ∞
2
an wn dw
X
G(z) = exp(zf (t0 )) exp(−zw )
−∞ n=0
∞
n+1
 
an z −(n+1)/2 Γ
X
= exp(zf (t0 ))
n=0
2
∞   m
exp(zf (t0 )) 1 1
X 
= √ a2m Γ m +
z m=0
2 z

where in the last line we have used the fact that only even powers of w appear in dt/dw.

Let us outline how to derive the Stirling series using this method. Recall there that
Z ∞
Γ(z + 1) = z z+1 exp(zf (t))dt
0

for f (t) = log t − t, with only saddle point at t0 = 1. Expand in a Taylor series about f (t0 ) = −1
to get
(t − 1)2 (t − 1)3 (t − 1)4
f (t) = −1 − + 2 − (3!) + ···
2! 3! 4!
so in the notation above,
(t − 1)2 (t − 1)3
w2 = − + ···
2 3
√
It can √be shown (see [MF][section 4.6] for details, but note their a0 at the bottom of p. 442 is 2
not 1/ 2) that
√
!
dt w2 w4
= 2 1 + + + ···
dw 6 216
and plugging these values into the general expression for the asymptotic series we find

exp(−z) √ 1 1 1 1
 
Γ(z + 1) = z z+1 √ 2 Γ(1/2) + Γ(3/2) + Γ(5/2) 2 + · · ·
z 6 z 216 z
√ z+1/2

1 1

= 2πz exp(−z) 1 + + + ···
12z 288z 2
and so we recover Stirling’s series.

For more information, see [AW][section 7.4] or [MF][section 4.6].

D.5 Uniqueness (or lack thereof)

One very important property of asymptotic series is that they do not uniquely determine a
function. It is easy to check, for example, that if Re(z) > 0, then the same series can be
simultaneously asymptotic to both f (z)/ϕ(z) and f (z)/ϕ(z) + exp(−z).
APPENDIX D. ASYMPTOTIC SERIES 14

This fact is very important in quantum field theory, and is a reflection of nonperturbative effects
in the theory. Summing over Feynman diagrams yields a series in which the coupling constant of
the theory plays the part of 1/z. Now, a typical quantum field theory has ‘nonperturbative
effects,’ which cannot be seen in a (perturbative) Feynman diagram expansion. Nonperturbative
effects, which are not uniquely determined by the perturbative theory, are exponentially small in
the coupling constant, i.e. multiplied by factors of exp(−1/g) = exp(−z). Since the Feynman
diagram expansion is only an asymptotic series, and the nonperturbative effects are exponentially
small, adding nonperturbative effects does not change the asymptotic expansion, i.e. does not
change the Feynman diagram expansion.

Properties of asymptotic series:

Asymptotic series can be added, multiplied, and integrated term-by-term. However, asymptotic
series can only be differentiated term-by-term to obtain an asymptotic expansion for the
derivative only if it is known that the derivative possesses an asymptotic expansion.

D.6 Summation of asymptotic series

How can we sum, in any sense, a divergent series?

One approach is as follows. Given a divergent series

∞
An z n
X
F (z) =
n=0

for some constants An , consider the related series

∞
X zn
B(z) = An
n=0
n!

Depending upon how badly divergent the original series F (z) was, one might hope that the new
series B(z) might actually converge in some region. Assuming that B(z) converges and can be
resummed, how might one recover F (z)? Well, use the formula
Z ∞
exp(−t/z)tn dt = z n+1 n!
0

to show that, formally, Z ∞

zF (z) = exp(−t/z)B(t)dt (D.1)
0

To calculate F (z) using the formal trick above, we need B(t) for real positive values of t less than
or of order z. So long as any singularities in B(t) on the complex t plane are at distances greater
than |z| from the origin, this should be OK. (In quantum field theory, singularities in B(t) are
typically associated with nonperturbative effects – instantons – so again we see that
nonperturbative effects limit the usefulness of resummation methods for the (asymptotic)
Feynman series. See [Weinberg2] for more information.)
APPENDIX D. ASYMPTOTIC SERIES 15

This particular resummation technique is known as Borel summation. If the integral on the
right-hand-side exists at points outside the radius of convergence of the original series, then the
Borel sum is defined to be
1 ∞
Z
exp(−t/z)B(t)dt
z 0
(The original reference on Borel summation is, to our knowledge, Borel, Leçons sur les Séries
Divergentes (1901) pp 97-115.)

As asymptotic series expansions do not uniquely determine functions, the reader should not be
surprised to learn there are additional resummation techniques, which can yield different results.

For example, Euler resummation is to define the sum of the series An to be given by
P

∞
An z n
X
lim
z→1−
n=0

when this limit exists. For example, the Euler sum of the series

1 − 1 + 1 − 1 + 1 − ···

is given by
  1 1
lim− 1 − z + z 2 − z 3 + · · · = lim− =
z→1 z→1 1+z 2

Further resummation methods are discussed in [Hardy, WW].

**** See Zinn-Justin pp 840-, section 37.3, 37.4 for more info on Borel.

D.7 Stokes’ phenomenon

Stokes’ phenomenon is the observation that the operations of analytic continuation and
asymptotic series expansion do not commute with one another.

Let us work through a simple example, following [MF][section 5.3]. Consider the confluent
hypergeometric function defined by
∞
X (a)n z n
M (a, c; z) =
n=0
(c)n n!

where (a)n is the Pochhammer symbol:

(a)n = a(a + 1)(a + 2) · · · (a + n − 1)

(a + n − 1)! Γ(a + n)
= =
(a − 1)! Γ(a)
(a)0 = 1
APPENDIX D. ASYMPTOTIC SERIES 16

First, let us derive the leading term in an asymptotic expansion as z → +∞ along the real axis.
The n term of the series above is given by

(a)n z n (a + n − 1)! (c − 1)! z n

=
(c)n n! (a − 1)! (c + n − 1)! n!

It can be shown [AW][problem 8.3.8] that

(x + a)!
lim xb−a = 1
x→∞ (x + b)!

so we see that for large n, the n term becomes

(c − 1)! n(a−1)−(c−1) n
→ z
(a − 1)! n!
(c − 1)! zn
→
(a − 1)! (n − a + c)!

At large positive z, the terms with large n dominate, so that

∞
X (c − 1)! zn
M (a, c; z) ≈
n=1
(a − 1)! (n − a + c)!

Assume a − c ∈ Z, and ignoring the first few terms (small compared to the rest), we have that
∞
(c − 1)! X z m a−c
M (a, c; z) ≈ z (m = n − a + c)
(a − 1)! m=0 m!
Γ(c) a−c
= z exp(z)
Γ(a)

We have been sloppy, but a more careful analysis reveals this result is correct; the leading term in
an asymptotic series expansion of M (a, c; z) for z → +∞ along the real axis is given by the
expression above.

Next, let us find the leading term in an asymptotic expansion of M (a, c; z) for z → −∞ along the
real axis. The fast way to do this is to use the identity

M (a, c; z) = exp(z)M (c − a, c; −z)

Using this and our previous result, we see that as z → −∞ along the real axis,

Γ(c)
M (a, c; z) → (−z)−a
Γ(c − a)

This is the leading term in an asymptotic series expansion of M (a, c; z) for z → −∞.

But if we compare our results for the two different limits, we see that this is not what we would
have gotten by analytically continuing either separately.
APPENDIX D. ASYMPTOTIC SERIES 17

For example, if we started with the z → +∞ limit, and analytically continued, we would have
found
Γ(c) Γ(c)
(−z)a−c exp(−z) 6= (−z)−a
Γ(a) Γ(c − a)

Thus, analytic continuation does not commute with asymptotic series expansions. This is known
as Stokes’ phenomenon. This is very unlike convergent Taylor series, for example, where analytic
continuation does commute with series expansion.

Let us also understand this more systematically by taking asymptotic series expansions at all
arg(z), not just along the real axis. To do this, we shall use the integral representation
[AW][problem 13.5.10a]

Γ(c)
Z 1
M (a, c; z) = exp(zt)ta−1 (1 − t)c−a−1 dt
Γ(a)Γ(c − a) 0

Deform the integration path to go from first, t = 0 to t = −∞ exp(−iφ), then, from

t = −∞ exp(−iφ) to t = 1, where z = |z| exp(iφ). Then we find that
(Z
Γ(c) −∞ exp(−iφ)
M (a, c; z) = exp(zt)ta−1 (1 − t)c−a−1 dt
Γ(a)Γ(c − a) 0
)
Z 1
+ exp(zt)ta−1 (1 − t)c−a−1 dt
−∞ exp(−iφ)

When 0 < φ < π, make the following changes of variables. In the first integral, define w by

w exp(−iφ) w
t = − = −
|z| z

and in the second integral, define

u exp(−iφ) u
t = 1 − = 1−
|z| z

Plugging these in, we find

(Z a−1  c−a−1 
Γ(c) ∞ −w

w dw

M (a, c; z) = exp(−w) 1+ −
Γ(a)Γ(c − a) 0 z z z
0 a−1  c−a−1  )
u u du
Z 
+ exp(z(1 − u/z)) 1 − −
∞ z z z
( c−a−1
Γ(c) exp(iaπ)
Z ∞ 
w
a−1
= exp(−w)w 1+ dw
Γ(a)Γ(c − a) za 0 z
a−1 )
exp(z)
Z ∞ 
u c−a−1
+ exp(−u) 1 − u du
z c−a 0 z
Γ(c) exp(iaπ) Γ(c)
→ + exp(z)z a−c
Γ(c − a) za Γ(a)
APPENDIX D. ASYMPTOTIC SERIES 18

The first term dominates when φ = π, at which the second term is negligible. When φ = 0, the
opposite is true: the second term dominates, the other term is negligible. For φ = π/2, the two
terms are comparable.

The analysis can be repeated for −π < φ < 0; but as it is very similar, for brevity we shall not
repeat it here.

So, what we have found in general is that for general φ, the leading term is a combination of the
two terms, but in the two limits, one dominates and the other is much less than the corrections,
so that the leading term in the asymptotic series expansion is defined by only term, not both.
Which term dominates, varies as φ changes. Thus, analytic continuation does not commute with
asymptotic series expansion.

D.8 Exercises
1. Show that the function ez has no asymptotic series expansion for z real, positive, and large.
2. Show that the function ez has the asymptotic series expansion
0 + 0 + 0 + 0 + ···
for z real, negative, and of large magnitude. The previous problem plus this one give an
easy example of Stokes’ phenomenon, namely that the same function can have different
asymptotic series expansions as one goes in different directions on the complex plane.
3. (AW 8.3.8) Use the Stirling series for the gamma function to show that
(x + a)!
lim xb−a = 1
x→∞ (x + b)!
4. Use Watson’s lemma to derive an asymptotic series for
Z ∞  1/2
e−zt 1 + t2 dt
0

5. (AW 7.3.1) Using the method of steepest descent, evaluate the second Hankel function given
by
1 0 s 1 dz
Z   
Hν(2) (s) = exp z −
πi −∞ 2 z z ν+1
with contour **** FILL IN ****
6. (WW VIII.6) Show that the series
1 − 2! + 4! − · · ·
is not Borel summable, whereas the series
1 + 0 − 2! + 0 + 4! + · · ·
is Borel summable.

**** Also mention asymptotic series, steepest descent problems buried in the bulk of the text.
Bibliography

[1] B. Schutz, A first course in general relativity, Cambridge University Press, 1990.

[2] D. Gross, E. Witten, “Superstring modifications of Einstein’s equations,” Nucl. Phys. B277
(1986) 1-10.

[3] D. Nemeschansky, A. Sen, “Conformal invariance of supersymmetric sigma models on

Calabi-Yau manifolds,” Phys. Lett. B178 (1985) 365-369.

[PS] M. Peskin, D. Schroeder, An introduction to quantum field theory, Addison-Wesley, 1995.

[4] P. Aspinwall, D. Morrison, “Topological field theory and rational curves,” Comm. Math.
Phys. 151 (1993) 245-262, hep-th/9110048.

[W91] E. Witten, “Mirror manifolds and topological field theory,” hep-th/9112056.

[S05] E. Sharpe, “Notes on certain other (0,2) correlation functions,” hep-th/0605005.

[Atiyah90] M. Atiyah, The geometry and physics of knots, Cambridge University Press, 1990.

[5] K. Costello, “Topological conformal field theories and Calabi-Yau categories,”

math.QA/0412149.

[6] K. Costello, “Topological conformal field theories and gauge theories,” math.QA/0605647.

[Manin88] Y. Manin, Gauge field theory and complex geometry, Springer-Verlag, 1988.

[W93] E. Witten, “Phases of N = 2 theories in two dimensions,” Nucl. Phys. B403 (1993)
159-222, hep-th/9301042.

[AM] N. Ashcroft, N. Mermin, Solid state physics, Harcourt College Publishers, 1976.

[Apostol] T. Apostol, Introduction to analytic number theory, Springer, 1976.

[AW] G. Arfken, H. Weber, Mathematical methods for physicists, Academic Press, 2001.

[Collins] J. Collins, Renormalization, Cambridge University Press, Cambridge, 1984, 1992.

[DFMS] P. Di Francesco, P. Matheiu, D. Sénéchal, Conformal field theory, Springer-Verlag, New

York, 1997.

19
BIBLIOGRAPHY 20

[Fradkin] E. Fradkin, Field theories of condensed matter systems, Westview Press, 1991.

[Goldstein] H. Goldstein, Classical mechanics, second edition, Addison-Wesley, Reading, MA,

1980.

[Huang] K. Huang, Statistical mechanics, second edition, John Wiley & Sons, New York, 1987.

[Ma] S.-K. Ma, Modern theory of critical phenomena, Perseus Publishing, Cambridge,
Massachusetts, 1976, 2000.

[MF] P. Morse and H. Feshbach, Methods of Theoretical Physics volume I, McGraw-Hill, New
York, 1953.

[MTW] C. Misner, K. Thorne, J. Wheeler, Gravitation, W. H. Freeman and Company, New York,
1973.

[PM] J. G. Proakis, D. G. Manolakis, Introduction to digital signal processing, Macmillan

publishing company, New York, 1988.

[SalamSezgin] A. Salam, E. Sezgin, Supergravities in diverse dimensions, volume I,

North-Holland, Amsterdam, 1989.

[Weinberg1] S. Weinberg, The quantum theory of fields, volume I, Cambridge University Press,
1995.

[Weinberg2] S. Weinberg, The Quantum Theory of Fields, volume II, Cambridge University
Press, 1996, section 20.7.

[Wen] X.-G. Wen, Quantum field theory of many-body systems, Oxford University Press, Oxford,
2004.

[Shiv03] B. Shivamoggi, Perturbation methods for differential equations, Birkhäuser, Boston, 2003.

[Hardy] G. H. Hardy, Divergent series, Oxford University Press, 1949.

[Jackson] J. D. Jackson, Classical electrodynamics, third edition, John Wiley & Sons, 1999.

[WW] E. T. Whittaker, G. N. Watson, A course of modern analysis, 4th edition, Cambridge

University Press, 1963.

[ZJ] J. Zinn-Justin, Quantum field theory and critical phenomena, Clarendon Press, Oxford,
1989.

[Zwillinger] D. Zwillinger, Handbook of differential equations, Academic Press, Boston, 1989.