Академический Документы
Профессиональный Документы
Культура Документы
Abstract
Using the saddle point method and multiseriesexpansions,
we obtain from the generating func
n
tion of the Stirling numbers of the second kind
and Cauchys integral formula, asymptotic
m
results in central and non-central regions. In the central region, we revisit the celebrated Gaussian
theorem
with more precision. In the region m = n n , 1 > > 1/2, we analyze the dependence
n
of
on . This paper fits within the framework of Analytic Combinatorics.
m
Keywords: Stirling numbers of the second kind, Asymptotics, Saddle point method, Multiseries
expansions, Analytic Combinatorics.
Introduction
n
be the Stirling number of the second kind. Their generating function is given by
Let
m
X m! n
n
n!
z n = f (z)m ,
f (z) := ez 1.
In the sequel all asymptotics are meant for n .
Let us summarize the related litterature. The asymptotic Gaussian approximation in the central
region is proved in Harper [7]. See also Bender [1], Sachkov [13] and Hwang [10].
In the non-central region, m = n n , 0 > < 1, most of the previous papers use the solution
of
e
n
= .
(1)
e 1
m
As shown in the next section, this actually corresponds to a Saddle point.
Let us mention
Hsu [8]:
For t = o(n1/2 )
n2t
f1 (t) f2 (t)
n+t
= t 1+
+ 2 + ... ,
n
2 t!
n
n
1
f1 (t) = t(2t + 1).
3
Universite Libre de Bruxelles, Departement dInformatique, CP 212, Boulevard du Triomphe, B-1050 Bruxelles,
Belgium, email: louchard@ulb.ac.be
For t = o( n),
(t)2
n t
(t)2 2
n
1+
=
q
q+
q + ... ,
nt
t
12
288
2
.
q=
nt
For n m , n ,
1
n!(e 1)m
15C32
3C4
n
1
= n
+
.
.
.
,
m
m 162 H
4H 2
2 m!(mH)1/2
e (e 1 )
H=
,
2(e 1)2
C3 , C4 are functions of .
Good [6]:
n+t
t
(t + n)!(e ) 1)t
=
g1 () g2 ()
1+
+ 2 + ... ,
t
t
= 2 ()1/2 ,
1 = , 2 = (1 + 1)( 1 ).
Bender [1]:
n!em
n
,
m
m!n1 (1 + e ) 2n
n
= (1 + e ) ln(1 + e ),
m
= ln(1 + e ),
m 2
2 =
1 e ln(1 + e ) ,
n
It is easy to see that here coincides with the solution of (1). Benders expression is similar to
Moser and Wymans result.
Bleick and Wang [2]:
1 e1
n+1
=
.
e1 1
m
Then
n!(e1 1)m
n
=
m
(2(n + 1))1/2 m!n1 (1 G)1/2
2
2 + 18G 20G2 (e1 + 1) + 3G3 (e21 + 4e1 + 1) + 2G4 (e21 e1 + 1)
2
1
+ O(1/n ) ,
24(n + 1)(1 G)3
1
G = 1
.
e 1
X
n
A nm n
=e m
(1)k fk (t0 )mk ,
m
m
k=0
1/2
t0
n
f0 (t0 ) =
,
(1 + t0 )( t0 ) m
n
t0 =
1,
m
where A is a function of , n, m.
Tsylova [17]:
Let m = tn + o(n2/3 ).
(n)n
n
=
exp (m tn)2 /(2n) (1 + o(1)),
m
2n(n)m
(1 e1/ ) = ,
= e1/ (t e1/ ).
After some algebra, this coincides with Moser and Wymans result.
Chelluri, Richmond and Temme [3]:
They prove, with other techniques, that Moser and Wymans expression is valid if n m =
(n1/3 ) and that Hsus formula is valid for y x = o(n1/3 )
Erdos and Szekeres: see Sachkov [14], p.164:
Let m < n/ ln n,
h n
i
mn
n
=
exp
m en/m (1 + o(1)).
m
m!
m
All these papers simply use as the solution of (1). They dont compute the detailed dependence
of on for our non-central range, neither the precise behaviour of functions of they use. Moreover,
most results are related to the case < 1/2.
We will use multiseries expansions: multiseries are in effect power series (in which the powers may
be non-integral but must tend to infinity) and the variables are elements of a scale: details can be
found in Salvy and Shackell [15]. The scale is a set of variables of increasing order. The series is
computed in terms of the variable of maximum order, the coefficients of which are given in terms of
the next-to-maximum order, etc. Actually we implicitly used multiseries in our analysis of Stirling
numbers of the first kind in [11].
Our work fits within the framework of Analytic Combinatorics.
Let us finally mention that Hsu [9] consider some generalized Stirling numbers.
In Sec.2, we revisit the asymptotic expansion in the central region and in Sec.3, we analyse the
non-central region j = n n , > 1/2. We use Cauchys integral formula and the saddle point
method.
Central region
1.
Bn
Bn
M := E(Jn ) =
=e
1+
+ O(1/G ) .
Zn (m) =
Bn
6G(1 + )3/2
2G
Proof.
(3)
(4)
A5 =
2 + 9/ + 16/ 2 + 6/ 3 + 2/ 4
.
24(1 + 1/)3
This leads to (from now on, we only provide a few terms in our expansions, but of course we use more
terms in our computations), using expansions in G,
G
A3 (1 + )1/2
+
+ O(1/G3 ),
2G
(1 + )1/2
n
G
.
ln(n)
We now use the Saddle point technique (for a good introduction to this method, see Flajolet and
Sedgewick [4], ch.V III). Let be the saddle point and the circle ei . By Cauchys theorem,
Z
f (z)m
n! 1
Zn (m) =
dz
m!Bn 2i z n+1
Z
n!
1
f (ei )m eni d
=
m!Bn n 2
Z
1
n!
i
em ln(f (e ))ni d
=
n
m!Bn 2
Z
n!
f ()m
1
i
2
3
=
exp
m
+
.
.
.
,
(5)
2
3
m!Bn n 2
2
6
i
i () =
ln(f (eu ))|u=0 .
(6)
u
See Good [5] for an ancient but neat description of this technique.
Let us now turn to the Saddle point computation. is the root (of smallest module) of
mf 0 () nf () = 0, i.e.
n
e
= ,
e 1
m
which is, of course identical to (1). After some algebra, this gives
n
n
=
+ W en/m .
m
m
In the central region, we choose
m = M + x = F +
x
xA3 (1 + )1/2
G
+
A
+
+ O(1/G2 ).
1
2G
(1 + )1/2
This leads to
x
+ O(1/G2 ),
1/2
(1 + ) G
n
x
A1 + x2 /(1 + )
=
+
+ O(1/G3 ),
m
G2
(1 + )1/2 G
(A1 + x2 /(1 + ) 1)
x
=
+
+ O(1/G3 ),
G2
(1 + )1/2 G
x
ln() = ln()
+ O(1/G2 ).
(1 + )1/2 G
ln(m) = +
Let us remark that the coefficients of G powers are nice rational expressions in .
(7)
ln(n) = + ln(),
so, by Stirlings formula, with (4), the first part of (5) leads to
n!
f ()m = exp [T2 ] H1 H2 ,
m!Bn n
T2 = m( + ln() ln()) (m + ln(2)/2 + ln(m)/2) F ln() T1 ,
A5 (1 + 1/)1/2
H1 = 1/H0 = (1 + 1/)1/2
+ O(1/G4 ),
2
G
1
1
1
x
3
H2 = 1
1+
+ O(1/G4 ).
+
+
O(1/m
)
=1
+
12m 288m2
12G2 12G3 (1 + )1/2
Note carefully that there is a cancellation of the term m ln(m) in T2 . Using all previous expansions,
we obtain
exp(T2 ) = ex
H3 = 1 +
2 /2+ln()
)H3 ,
x(15
6 2
(8)
x2
2x2
+ O(1/G2 ).
x
e (e + 1 + )
+ O(1/G2 ),
=
2
(e 1)
(1 + )1/2 G
and similar expressions for the next i that we dont detail here. Note that 3 , 5 , . . . are useless for
the precision we attain here. Now we use the classical trick of setting
#
"
X
` (i)` /`! = u2 /2.
m 2 2 /2! +
l=3
1 X
=
ai ui ,
G
1
1
1/2
1/2
+ O(1/G3 ).
2G2
d
Setting d = du
du, we integrate on u = [..]: this extension of the range can be justified as in
Flajolet and Segewick [4], ch.V III. Now, inserting the term coming in (8) as eln() , this gives
1/2
H4 =
2G
3
+ O(1/G ) .
1+
2G2
(9)
R1 = e
x(6 + 2x2 + x2 3)
+ )1/2
2
1+
+ O(1/G ) .
6G(1 + )3/2
2G
x2 /2 (1
Note that the coefficient of the exponential term is asymptotically equivalent to the dominant
1
term of 2
, as expected. More terms in this expression can be obtained if we compute M, 2 , Bn /n!
with more precision. Also, using (2), our result can be put into expansions depending on n, ln n, . . ..
To check the quality of our asymptotic, we have chosen n = 3000. This leads to
= 6.184346264 . . . ,
G = 22.02488900 . . . ,
M = 484.1556441 . . . ,
= 8.156422315 . . . ,
Bn = 0.2574879583 . . . 106965 ,
Bn as = 0.2574880457 . . . 106965 ,
h
1
exp
where Bn as is given by (3). Figure 1 shows Zn (m) and 2
mM 2
. i
2 .
0.04
0.03
0.02
0.01
0 460
470
480
490
500
510
mm
1
2
h
exp
mM 2
. i
2
Figure 3 shows the quotient Zn (m) /R1 . The precision is now between 0.004 and 0.01.
Large deviation, m = n n ,
> 1/2
We set
:= n1 ,
7
1.1
1.05
0.95
460
470
480
Figure 2: Zn (m)
490
1
2
h
exp
500
mM 2
510
. i
2
1.01
1.008
1.006
1.004
1.002
1
460
470
480
490
500
510
1
= n1 n n,
L := ln(n).
The multiseries scale is here {n1 , n , n}.
Our result can be summarized in the following local limit theorem
Theorem 3.1
n
= eT1 R,
m
T1 = n (T11 L + T10 ),
R1 R2
1
3
R0 +
R=
+ 2 + O(1/n ) ,
n
n
2n/2
R01
R0 = R00 + + O(1/n2 ),
n
R11
R1 = R10 + + O(1/n2 ),
n
R21
R2 = R20 + + O(1/n2 ),
n
=1+
2
+
+ O(3 /n3 ).
12( 1)n 288( 1)2 n2
Note again that there are cancellations, in T1 of the terms m ln(m) and ln(2)/2.
Now we turn to the integral part. We obtain, for instance, using (6),
4
13
2 = + 2 + 3 + O(4 ),
3
9
1 X
=
ai ui ,
n 1
1
1 2
1 4
6
a1 = 1 + O( ) .
6
72
(10)
with
1
R1 R2
3
R=
R0 +
+ 2 + O(1/n ) ,
n
n
2n/2
R01
R0 = R00 + + O(1/n2 ),
n
R11
R1 = R10 + + O(1/n2 ),
n
R21
R2 = R20 + + O(1/n2 ),
n
1
R00 = 1 + + O(2 ),
3
1
1
R01 = + O(2 ),
12 36
1
1
R10 = 2 + O(3 ),
12
9
1
1 2
R11 =
+
+ O(3 ),
144
108
1
7 2
R20 =
+
+ O(3 ),
288
864
1
7
2 + O(3 ).
R21 =
3456
10368
Given some desired precision, how many terms must we use in our expansions? It depends on . For
instance, in T1 , n k 1 if k < /(1). Also k in R00 is less than ` /n in R10 /n if k` > 1/(1).
Any number of terms can be computed by almost automatic computer algebra. We use Maple in this
paper.
10
To check the quality of our asymptotic, we have chosen n = 100 and a range [1/2, 9/10],
i.e.
a range m [37, 90]. We use 5 or 6 terms in our final expansions. Figure 4 shows the quotient
n
(eT1 R). The precision is at least 0.0066. Note that the range [M 3, M + 3], where the
m
Gaussian approximation is useful, is here m [21, 36].
1.006
1.004
1.002
0.998
40
50
60
Figure 4:
70
80
90
n
(eT1 R)
m
We finally mention that our non-central range is not sacred: other types of ranges can be analyzed
with similar methods.
References
[1] E.A. Bender. Central and local limit theorems applied to asymptotics enumeration. Journal of
Combinatorial Theory, Series A, 15:91111, 1973.
[2] W. E. Bleick and P. C. C. Wang. Asymptotics of Stirling numbers of the second kind. Proceedings
of the American Mathematical Society, 42(2):575580, 1974.
[3] R. Chelluri, L. B. Richmond, and N. M. Temme. Asymptotic estimates for generalized Stirling
numbers. Technical report, CWI, MAS-R9923, 1999.
[4] P. Flajolet and R. Sedgewick. Analytic combinatorics. Cambridge University press, 2009.
[5] I. J. Good. Saddle-point methods for the multinomial distribution. Annals of Mathematical
Statistics, 28(4):861881, 1957.
[6] I. J. Good. An asymptotic formula for the differences of the power at zero. Annals of Mathematical
Statistics, 32(1):249256, 1961.
[7] L. H. Harper. Stirling behaviour is asymptotically normal. Annals of Mathematical Statistics,
38:410414, 1967.
11
[8] L. C. Hsu. Note on an asymptotic expansion of the nth differences of zero. Annals of Mathematical
Statistics, 19:273277, 1948.
[9] L. C. Hsu. A unified approach to generalized Stirling numbers. Advances in Applied Mathematics,
20:366384, 1998.
[10] H.K. Hwang. On convergence rates in the central limit theorems for combinatorial structures.
European Journal of Combinatorics, 19:329343, 1998.
[11] G. Louchard. Asymptotics of the stirling numbers of the first kind revisited: A saddle point
approach. Discrete Mathematics and Theoretical Computer Science, 18(2):167184, 2010.
[12] L. Moser and M. Wyman. Stirling numbers of the second kind. Duke Mathematical Journal,
25:2948, 1958.
[13] V.N. Sachkov. Probabilistic Methods in Combinatorial Analysis. Cambridge University Press,
1997. Translated and adapted from the Russian original edition, Nauka,1978.
[14] V.N. Sachkov. Probabilistic Methods in Combinatorial Analysis. Cambridge University Press,
1997.
[15] B. Salvy and J. Shackell. Symbolic asymptotics: Multiseries of inverse functions. Journal of
Symbolic Computation, 20(6):543563, 1999.
[16] N.M. Temme. Asymptotic estimates of Stirling numbers. Studies in Applied Mathematics, 89:233
243, 1993.
[17] E. G. Tsylova. Probabilistic methods for obtaining asymptotic formulas for generalized Stirling
numbers. Journal of Mathematical Sciences, 75(2):16071614, 1995.
12