Date: Nov. ? , 2004 Nov.

08, 2007 Page: 1

Lecturer: Daniel Panario Jason Z. Gao Scribe:

1 Proof of the Lagrange Inversion Formula

Theorem 1 Lagrange Inversion Formula: Suppose u = u(x) is a power series
in x satisfying x = u/φ(u) where φ(u) is a power series in u with a nonzero
constant term. Then for any power series F (u) of u, we have
1 n−1
[xn ]F (u(x)) = [u ] (F 0 (u)φn (u)) . (1)
n

Proof. Let φ(u) = k≥0 φk uk . We first note that u(x) = φ0 x + O(x2 ) and
P

hence k≥n φk uk does not contribute to either side of (1).

P

Hence we may assume that φ(u) is a polynomial of u, which is analytic every-

where. Similarly we may also assume F (u) is a polynomial in u. Since φ(0) 6= 0,
u/φ(u) is analytic at the origin with a nonzero first derivative at 0, by the im-
plicit function theorem, u(x) is also analytic in a small region enclosing the
origin. Let C be a small circle enclosing the origin in the x-plane and C 0 be the
image of C under the transformation x = u/φ(u). Using the Cauchy formula,
we obtain
Z
n 1 F (u(x))
[x ]F (u(x)) = dx
2πi C xn+1
F (u)φn+1 (u) −1
Z
1
φ (u) − uφ−2 (u) du


= n+1
2πi C 0 u
F (u)φn (u) F (u)φn−1 (u)
Z Z
1 1
= du − du
2πi C 0 un+1 2πi C 0 un
Applying the Cauchy formula again w.r.t. the variable u, we obtain
[xn ]F (u(x)) = [un ]F (u)φn (u) − [un−1 ]F (u)φn−1 (u)
For any power series f (u), we have
1 n−1 0
[un ]f (u) = [u ]f (u),
n
and hence
1 n−1
[xn ]F (u(x)) ] (F 0 (u)φn (u) + nF (u)φn−1 (u) − [un−1 ]F (u)φn−1 (u)


= [u
n
1 n−1
= [u ] ((F 0 (u)φn (u))
n
This completes the proof.

2 Hayman’s Method
Example 1 Stirling’s formula
1 1  e n
∼√ .
n! 2πn n
Date: Nov. ? , 2004 Nov. 08, 2007 Page: 2
Lecturer: Daniel Panario Jason Z. Gao Scribe:

Using Cauchy’s formula with the contour |z| = r, we have

π iθ
ere
Z
1 1
= dθ.
n! 2π −π rn einθ
Since  iθ 
 re 
e  ≤ er , |θ| ≤ π,
we have
1 er
≤ n , for any r > 0.
n! r
To find a least upper bound, we choose r such that
d
(ln(er /rn )) = 0, i.e., r = n.
dr
Hence  e n
1
≤ .
n! n
To obtain a more precise estimate for 1/n!, we need to estimate the above integral
more carefully. To do that we note that | exp(neiθ )| = exp(n cos θ) becomes
exponentially small when θ is away from 0. Hence we only need to estimate the
integral in a small interval |θ| ≤ δ. Expanding eiθ , we obtain
iθ 2
−iθ) /2+O(θ 3 )) 2
en(e = en(1−θ ∼ en(1−θ /2)
,

provided that nδ 3 → 0, or δ = o(n−1/3 ).

Hence
Z δ
1 1 2
∼ (e/n)n e−nθ /2
dθ
n! 2π −δ
Z nδ √
1 2 √
∼ √ (e/n)n √ e−t /2 dt (θ = t/ n)
2π n − nδ
1
∼ √ (e/n)n ,
2πn

provided that nδ 2 → ∞. We may choose δ = n−5/12 .

Hayman (1954) generalized the above estimate to a class of Hayman-admissible

functions.
Suppose f (z) is analytic in |z| < R, and let
f 0 (r)
a(r) = r , b(r) = ra0 (r).
f (r)
It is called Hayman-admissible if there is a positive number R0 and a function
0 < δ(r) < π defined in R0 < r < R such that
Date: Nov. ? , 2004 Nov. 08, 2007 Page: 3
Lecturer: Daniel Panario Jason Z. Gao Scribe:

(a) f (r) > 0 for R0 < r < R;

2
(b) f (reiθ ) ∼ f (r)eia(r)θ−b(r)θ
p
/2
as r → R, uniformly for |θ| ≤ δ(r);
iθ
(c) f (re ) = o(f (r)/ b(r)) as r → R, uniformly for δ(r) ≤ |θ| ≤ π.
(d) b(r) → ∞ as r → R.

Theorem 2 (Hayman) Suppose f (z) is Hayman-admissible, rn satisfies a(rn ) =

n. Then
f (rn )
[z n ]f (z) ∼ p rn−n , as n → ∞.
2πb(rn )

Hayman provided the following large families of admissible functions.

(A) Suppose P (z) is a polynomial with real coefficients, and [z n ]eP (z) > 0 for
all sufficiently large n. Then eP (z) is Hayman admissible in |z| < ∞.
(B) Suppose P (z) is a polynomial with real coefficients, and f (z) is Hayman-
admissible in |z| < R. Then P (f (z)) is also Hayman-admissible in |z| < R.
(C) If f (z) and g(z) are both Hayman-admissible in |z| < R, then f (z)g(z) and
ef (z) are also Hayman-admissible in |z| < R.

Example 2 By (A), ez is Hayman-admissible with R = ∞, and we have a(r) =

r, b(r) = r. Hence rn = n.

Example 3 The EGF for permutations with cycle length at most 2 is f (z) =
2
ez+z /2 . By (A), f (z) is Hayman-admissible with R = ∞, and we have a(r) =
r(1 + r), b(r) = r(1 + 2r). Hence
1/2
√

p 1
rn = n + 1/4 − 1/2 = n 1 + − 1/2
4n
√
 
1 1 
= n 1− √ + + O n−3/2 .
2 n 8n

To obtain the asymptotic expression for [z n ]f (z), we need to estimate rn , exp(rn ),

and rn−n = exp(n ln(1/rn )).

√
f (rn ) ∼ en/2+ n/2−1/4
,
b(rn ) ∼ 2n,
  −n
1 1
rn−n = n −n/2
1− √ − + O(n −3/2
)
2 n 8n
 −1 !!
1 1
= n−n/2 exp n ln 1 − √ − + O(n−3/2
2 n 8n
√
∼ n−n/2 e n/2
.
Date: Nov. ? , 2004 Nov. 08, 2007 Page: 4
Lecturer: Daniel Panario Jason Z. Gao Scribe:

Hence √
en/2+ n−1/4 −n/2
fn /n! ∼ √ n ,
2 nπ
and
1 √
fn ∼ √ nn/2 e−n/2+ n−1/4 .
2

We can also solve r(1 + r) = n using Lagrange inversion formula. Since r → ∞

as n → ∞, it is clear that n ∼ r2 , or 1/r ∼ n−1/2 . We can rewrite r(1 + r) = n
as
1
(1 + 1/r)−1/2 = n−1/2 .
r
Setting u = 1/r and t = n−1/2 , we obtain
u
= t,
(1 + u)1/2

and hence
X 1 X 1  k/2 
u= tk [uk−1 ](1 + u)k/2 = tk ,
k k k−1
k≥1 k≥1
or
1 1 1
= n−1/2 + n−1 + n−3/2 + O n−2 .


r 2 8
This gives
 −1  
rn = n1/2 1 + (1/2)n−1/2 + (1/8)n−1 + O(n−3/2 ) = n1/2 1 − (1/2)n−1/2 + (1/8)n−1 + O(n−3/2 ) ,

1    
ln = ln n−1/2 +ln 1 + (1/2)n−1/2 + (1/8)n−1 + O(n−3/2 ) = −(1/2) ln n+(1/2)n−1/2 +O n−3/2
rn