The Saddle Point Method: Brad Klingenberg Complex Variables

The saddle point method
Brad Klingenberg
Complex Variables
April, 2006
Asymptotic Analysis
• Approximate the asymptotic behavior
of integrals
• Example: approximating n! for large n
• Example: integrals such as
Z ∞ Z
enϕ(x)dx and enϕ(z)dz.
−∞ C
Example: Stirling’s approximation
• Approximate n! for large n
• Use the Gamma function,
Z ∞
z−1 −t
Γ(z) = t e dt
0
• Approximate n! for large n
• Use the Gamma function,
Z ∞
z−1 −t
Γ(z) = t e dt
0
Recall for n ∈ N,
Γ(n + 1) = nΓ(n) = . . . = n!Γ(1) = n!

Then
Z ∞
n −t
n! = Γ(n + 1) = t e dt.
0
Introduce the change of variables t = nz.
Z ∞
n! = (nz)n e−nz dz
0
Z ∞
n+1
= ... = n exp(n [ln(z) − z])dz.
0
• ln(z) − z < 0 for z ∈ (0, ∞)
• ln(z) − z has maximum at z = 1
−0.5 ln(z) − z
−1
−1.5
−2
−2.5
−3
0 1 2 3 4
z
• ln(z) − z largest at z = 1
• exp(n [ln(z) − z]) largest at z = 1
• As n becomes large, the difference is
more extreme
• Expect dominant contribution from
z=1
1
n=1
0.8 n=5
n = 75
0.6
0.4
0.2
0
0 1 2 3 4 5
Rescaled integrand for n = 1, 5, 75.

Taylor expand integrand about z = 1, letting
g(z) = ln(z) − z ,
Z ∞
n! = nn+1 exp(n [ln(z) − z])dz
0
Z 1+ǫ
1
≈ n n+1
exp n g(1) + (z − 1)2 g ′′ (1) dz
1−ǫ 2
Z ∞
1
≈ n n+1
exp n g(1) + (z − 1)2 g ′′ (1) dz
0 2
Z ∞ 2 q
s
n+1 −n
= n e exp n ds = nn+1 e−n 2π
−1 2 n
with s = z − 1.
The approximation to the integrand is
very accurate!
−9
x 10 n( ln(z) − z )
2.5
2
n( −1 − (z−1) /2 )
2
1.5
0.5
0
0 0.5 1 1.5 2
z
Integrand approximation for n = 20.

Generalization: Laplace’s Method
Consider integrals of the form
Z
enϕ(x)dx
Approximate as n → ∞.
What have we learned from the example?
Stirling’s approximation provides a strategy:
• Find the point x0 where ϕ(x) attains its maximum
• Expand enϕ(x) in a neighborhood of x0
• Argue
Z x0 +ǫ Z
nϕ(x) nϕ(x)
e dx ≈ e dx
x0 −ǫ R
by the exponential decay of the integrand.

It can be shown that for a < x0 < b,
Z s
b
nϕ(x) nϕ(x0 ) 2π
e dx ≈ e
a n|ϕ′′(x0)|
If x0 is an endpoint, then
Z b r
π
enϕ(x)dx ≈ enϕ(x0)
a 2n|ϕ′′(x0)|
Example 6.2.15
Consider
n
X n
I(n) = k!n−k
k
k=0
Notice
Z ∞ Z ∞
−nx k −k−1 −t k
e x dx = n e t dt
0 0
−1−k −k−1
= Γ(k − 1)n = k!n
where t = nx.
Example 6.2.15
n Z ∞
X n
I(n) = n e−nx xk dx
k 0
k=0
Z ∞ n !
−nx
X n k
= e n x dx
0 k
k=0
Z ∞
−nx n
= e n(1 + x) dx
0
Z ∞
= n exp(n [ln(1 + x) − x])dx
0
| {z }
ϕ(x)
Example 6.2.15
Then ϕ(x) = ln(1 + x) − x,
′ 1
ϕ (x) = −1
1+x
such that x0 = 0 (an endpoint).
ϕ(0) = ln(1) − 0 = 0

′′
−1
|ϕ (0)| = = | − 1| = 1.
(1 + 0)2
Example 6.2.15
So using Laplace’s method,
n
X n
−k
I(n) = k!n
k
k=0
r
nϕ(x0 ) π
≈ ne
2n|ϕ′′(x0)|
r r
π nπ
= n =
2n 2
Extending to the complex plane
Now consider
Z Z
enϕ(z)dz = enReϕ(z)eniImϕ(z)dz
C C
where ϕ(z) is an analytic function on C.

Now consider
Z Z
enϕ(z)dz = enReϕ(z)eniImϕ(z)dz
C C
where ϕ(z) is an analytic function on C.

The strategy for Reϕ is analogous to the
real case. What about Imϕ?
Get rid of it!

Deform the contour C to C ′ so that Imϕ
′
is constant on C . Then
Z Z
nϕ(z) niImϕ(z0 ) nReϕ(z)
e dz = e e dz
C C′
Laplace method applies!
But how do you find C ′ ?

′
Finding C
Require
ϕ(z) = u(x, y) + iv(x, y)

= u(x, y) + iv(x0, y0)
where z0 = x0 + iy0 and ϕ′(z0 ) = 0.

The dominant contribution will come from
z0!
Steepest descent
For v(x, y) = v(x0, y0),
∇v = (∂v/∂x, ∂v/∂y)
= (−∂u/∂y, ∂u/∂x)
The direction tangent to this curve is
(∂u/∂x, ∂u/∂y) = ∇u
- the steepest descent in u!
Example
Consider the asymptotics of
Z
1 1 −z t
= t e dt
Γ(z) 2πi C
Example
The Hankel contour from the notes.

Example
Z
1 1 −z t
= t e dt
Γ(z) 2πi C
Z
1
= exp(t − z ln(t))dt
2πi C
Z
1
= z
exp(z [s − ln(s)])ds
2πiz C | {z }
ϕ(s)
where t = sz.
Example
• Note that
′ 1
ϕ (s) = 1 −
s
such that ϕ′(1) = 0.
• If we deform C to pass through z = 1
and hold Imϕ(z) constant then the
dominant contribution will be around
z = 1.
Example
Plots of Reϕ(z) and Imϕ(z).

Example
• We want to hold Imϕ(z) constant
• What are the directions of steepest

descent from z = 1?
Example
Level curves for Imϕ(z).

Example
Close-up of Reg(z).
Example
• Vary ϕ(z) only in the imaginary
direction
• Let z = 1 + iv, then
ϕ(v) = 1 + iv − ln(1 + iv)

v 2 iv 3
= 1− + + ...
2 3
v2
≈ 1−
2
Example
v2
Letting g(v) = 1 − 2,
ǫZ
1 1
≈ exp(zg(v))dv
Γ(z) 2πz z−1 −ǫ
Z ∞ 2

1 z v
≈ e exp −z dv
2πz z−1 −∞ 2
r
1 z 2π
e z
pz
= z−1
e = z 2π
2πz z
A generalization
It can be shown that to leading order,
Z r
nϕ(z) iθ nϕ(z0 ) π
e dz ≈ e e ′′ (z )|
C k|ϕ 0
where
α π α 3π
θ=− + , − +
2 2 2 2
is the direction of steepest descent, and
′′ ′′ iα
ϕ (z0 ) = |ϕ (z0 )|e .
Conclusions
Laplace’s method is
• General
• Accurate
• Intuitive
• Extended to the complex plane

The Saddle Point Method: Brad Klingenberg Complex Variables

Uploaded by

Copyright:

Available Formats

The Saddle Point Method: Brad Klingenberg Complex Variables

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

The Saddle Point Method: Brad Klingenberg Complex Variables

Uploaded by

Copyright:

Available Formats

The saddle point method

Γ(n + 1) = nΓ(n) = . . . = n!Γ(1) = n!

Introduce the change of variables t = nz.

Rescaled integrand for n = 1, 5, 75.

Integrand approximation for n = 20.

by the exponential decay of the integrand.

where ϕ(z) is an analytic function on C.

where ϕ(z) is an analytic function on C.

Get rid of it!

Laplace method applies!

But how do you find C ′ ?

ϕ(z) = u(x, y) + iv(x, y)

where z0 = x0 + iy0 and ϕ′(z0 ) = 0.

The direction tangent to this curve is

The Hankel contour from the notes.

Plots of Reϕ(z) and Imϕ(z).

• We want to hold Imϕ(z) constant

• What are the directions of steepest

Level curves for Imϕ(z).

ϕ(v) = 1 + iv − ln(1 + iv)

You might also like