The integrals in the sums over 1 Ξ± < 2 have been already

MATH 7770-01
LIMIT THEOREMS
FALL 2010
Lecture 35. Improvements in limit theorems with power β€œtails”, continued.
The integrals in the sums over 1 < 𝛼𝑗 < 2 have been already evaluated in Lecture 14;
the summands in them are equal to
Ξ“(2 βˆ’ 𝛼𝑗 ) [ +
𝛼𝑗 πœ‹
𝛼𝑗 πœ‹ ] 𝛼𝑗
βˆ’
β‹… (𝑐𝛼𝑗 + π‘βˆ’
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£
𝛼𝑗 ) cos
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑧) β‹… sin
𝛼𝑗 βˆ’ 1
2
2
(35.1)
(see formula (14.12)); and it can be rewritten in exactly the form (34.11) if we understand
the function Ξ“ as being continued analytically to all complex numbers except non-positive
integers (see the text after formula (14.12)). So we can extend the first sum in (34.20) to
all 𝛼𝑗 between 0 and 2.
∫
∞
∫
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯) 𝑑𝐡 + (π‘₯) and
If 1 < 𝛽 < 2, the last two integrals in (34.20) are
0
0
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯) 𝑑𝐡 βˆ’ (π‘₯), and we handle them by integrating by parts; say, for the
βˆ’βˆž
second one:
[ 𝑖𝑑π‘₯
]0
(𝑒 βˆ’ 1 βˆ’ 𝑖𝑑π‘₯) β‹… π΅βˆ’ (π‘₯) βˆ’βˆž βˆ’ 𝑖𝑑
∫
0
(𝑒𝑖𝑑π‘₯ βˆ’ 1) β‹… π΅βˆ’ (π‘₯) 𝑑π‘₯.
(35.2)
βˆ’βˆž
The term outside the integral is equal to 0 because limπ‘₯β†’0βˆ’ ∣(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯) β‹… π΅βˆ’ (π‘₯)∣ ≀
limπ‘₯β†’0βˆ’ (𝑑2 π‘₯2 /2)⋅𝑂(∣π‘₯βˆ£βˆ’π›½ ) = 0, limπ‘₯β†’βˆ’βˆž ∣(𝑒𝑖𝑑π‘₯ βˆ’1βˆ’π‘–π‘‘π‘₯)β‹…π΅βˆ’ (π‘₯)∣ ≀ limπ‘₯β†’βˆ’βˆž (2+βˆ£π‘‘βˆ£βˆ£π‘₯∣)×
× π‘‚(∣π‘₯βˆ£βˆ’π›½ ) = 0; and the absolute value of the integral does not exceed
[∫
∫
βˆ£π‘‘βˆ£βˆ’1
constβ‹…
2β‹…βˆ£π‘₯∣
βˆ’βˆž
βˆ’π›½
0
𝑑π‘₯+
βˆ£π‘‘βˆ£βˆ£π‘₯βˆ£β‹…βˆ£π‘₯∣
βˆ’π›½
]
𝑑π‘₯ = constβ‹…
βˆ£π‘‘βˆ£βˆ’1
(
2
1 ) π›½βˆ’1
+
β‹…βˆ£π‘‘βˆ£
. (35.3)
π›½βˆ’1 2βˆ’π›½
So we have:
𝑓 (𝑑) βˆ’ 1
=βˆ’
βˆ‘
𝑗: 0<𝛼𝑗
[
𝛼𝑗 πœ‹
𝛼𝑗 πœ‹ ] 𝛼𝑗
βˆ’
βˆ’
Ξ“(1 βˆ’ 𝛼𝑗 ) β‹… (𝑐+
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£
𝛼𝑗 + 𝑐𝛼𝑗 ) cos
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑑) β‹… sin
2
2
<2
+ 𝑖𝑑1 𝑑 + 𝑂(βˆ£π‘‘βˆ£π›½ )
(𝑑 β†’ 0)
(35.4)
(the term 𝑖𝑑1 𝑑 is infinitely small compared to all terms with 𝛼𝑗 < 1, and infinitely large
compare to those with 𝛼𝑗 > 1, so perhaps we shouldn’t put all terms with 0 < 𝛼𝑗 < 2
together, but write the terms with 0 < 𝛼𝑗 < 1 first, then 𝑖𝑑1 𝑑, then the terms with
1 < 𝛼𝑗 < 2, and finally 𝑂(βˆ£π‘‘βˆ£π›½ ). But of course (35.4) looks shorter as it is written).
Problem 72 . Find the first pseudo-moment for the distribution of Problem 70
with accuracy to 0.1.
So we can write an asymptotic expansion for 𝑓 (𝑑) for the distribution of Problem 70
(this way I am telling you that we can take 𝛽 > 1 in this problem).
1
If 𝛽 > 2, we subtract from the integrand in the last two integral in (34.20) and
subtract from it βˆ’ 𝑑2 π‘₯2 /2; the second pseudo-moment 𝑑2 is defined as
∫
∞
𝑑2 =
[
π‘₯ 𝑑 𝐹 (π‘₯)βˆ’1+
2
0
βˆ‘
βˆ’π›Όπ‘—
𝑐+
𝛼𝑗 π‘₯
]
∫
0
+
βˆ’βˆž
𝑗: 0<𝛼𝑗 <2
βˆ‘
[
π‘₯2 𝑑 𝐹 (π‘₯)βˆ’
βˆ’π›Όπ‘—
π‘βˆ’
𝛼𝑗 ∣π‘₯∣
]
(35.5)
𝑗: 0<𝛼𝑗 <2
(the second pseudo-moment, in contrast with the second moment, needn’t be nonnegative).
We add to and subtract from 𝑒𝑖𝑑π‘₯ βˆ’1βˆ’π‘–π‘‘π‘₯ in (34.20) the next Maclaurin term βˆ’π‘‘2 π‘₯2/2:
𝑓 (𝑑) βˆ’ 1
=βˆ’
βˆ‘
𝑗: 0<𝛼𝑗
[
𝛼𝑗 πœ‹
𝛼𝑗 πœ‹ ] 𝛼𝑗
βˆ’
βˆ’
Ξ“(1 βˆ’ 𝛼𝑗 ) β‹… (𝑐+
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£
𝛼𝑗 + 𝑐𝛼𝑗 ) cos
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑑) β‹… sin
2
2
<2
+ 𝑖𝑑1 𝑑 βˆ’ 𝑑2 𝑑2/2
∫
∞
+
βˆ‘
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑 𝐹 (π‘₯) βˆ’ 1 +
0
βˆ’π›Όπ‘—
𝑐+
𝛼𝑗 π‘₯
]
𝑗: 0<𝛼𝑗 <2
∫
0
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑 𝐹 (π‘₯) βˆ’
+
βˆ‘
]
βˆ’π›Όπ‘—
π‘βˆ’
.
𝛼𝑗 ∣π‘₯∣
(35.6)
𝑗: 0<𝛼𝑗 <2
Now
we
add
to
and
subtract
from
the
expressions
in
the
brackets
the
sums
βˆ‘
βˆ‘
βˆ’
βˆ’π›Όπ‘—
+ βˆ’π›Όπ‘—
(the only thing we seem to be doing is adding
, 𝑗: 2<𝛼𝑗 <3 𝑐𝛼𝑗 ∣π‘₯∣
𝑗: 2<𝛼𝑗 <3 𝑐𝛼𝑗 π‘₯
and subtracting the same things again and again). We get instead of the last two integrals:
βˆ’βˆž
[
βˆ‘
∫
𝑐+
𝛼𝑗
𝑗: 2<𝛼𝑗 <3
∞
0
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑(βˆ’π‘₯βˆ’π›Όπ‘— )
∫ 0
]
βˆ’
+ 𝑐𝛼𝑗
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) π‘‘βˆ£π‘₯βˆ£βˆ’π›Όπ‘—
βˆ’βˆž
∫
∞
+
βˆ‘
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑 𝐹 (π‘₯) βˆ’ 1 +
0
∫
βˆ’π›Όπ‘—
𝑐+
𝛼𝑗 π‘₯
]
(35.7)
𝑗: 0<𝛼𝑗 <3
0
+
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑 𝐹 (π‘₯) βˆ’
βˆ’βˆž
βˆ‘
]
βˆ’π›Όπ‘—
π‘βˆ’
.
𝛼𝑗 ∣π‘₯∣
𝑗: 0<𝛼𝑗 <3
Let’s evaluate the first tow integrals. For 2 < 𝛼𝑗 < 3, integrating by parts, we get:
∫
∞
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑(βˆ’π‘₯βˆ’π›Όπ‘— )
0
[
]∞
= (𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) β‹… (βˆ’π‘₯βˆ’π›Όπ‘— ) 0 βˆ’ 𝑖𝑑
∫
∞
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯) β‹… (βˆ’π‘₯βˆ’π›Όπ‘— ) 𝑑π‘₯.
0
(35.8)
1
𝑑(βˆ’ π‘₯1βˆ’π›Όπ‘— ),
The term outside the integral is equal to 0. Rewriting (βˆ’π‘₯βˆ’π›Όπ‘— ) 𝑑π‘₯ as
𝛼𝑗 βˆ’ 1
we integrate by parts another time (again with the term outside the integral being equal
2
to 0), and get:
∫ ∞
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2/2) 𝑑(βˆ’π‘₯βˆ’π›Όπ‘— ) = βˆ’
0
2
𝑑
=βˆ’
(1 βˆ’ 𝛼𝑗 )(2 βˆ’ 𝛼𝑗 )
∫
𝑑2
𝛼𝑗 βˆ’ 1
∞
(𝑒
𝑖𝑑π‘₯
∫
∞
(𝑒𝑖𝑑π‘₯ βˆ’ 1) β‹… π‘₯1βˆ’π›Όπ‘— 𝑑π‘₯
0
(35.9)
βˆ’ 1) 𝑑(βˆ’π‘₯
2βˆ’π›Όπ‘—
).
0
Since 0 < 𝛼𝑗 βˆ’ 2 < 1, the integral here is of the kind that we have considered in Lecture
14, and it is equal
(
)
βˆ’Ξ“ 1 βˆ’ (𝛼𝑗 βˆ’ 2) β‹… π‘’βˆ’π‘–(𝛼𝑗 βˆ’2)β‹…sign(𝑑)β‹…πœ‹/2 βˆ£π‘‘βˆ£π›Όπ‘— βˆ’2 .
(35.10)
Multiplying by the factor in front of the integral in (35.9) and taking into account that
Ξ“(3 βˆ’ 𝛼𝑗 )
βˆ’1 = π‘’βˆ’π‘–β‹…2β‹…sign(𝑑)β‹…πœ‹/2 and
= Ξ“(1 βˆ’ 𝛼𝑗 ) (with the analytically-continued
(2 βˆ’ 𝛼𝑗 )(1 βˆ’ 𝛼𝑗 )
Ξ“-function), we get that the 𝑗-th summand in the first sum in (35.7) is equal to
[
𝛼𝑗 πœ‹
𝛼𝑗 πœ‹ ] 𝛼𝑗
βˆ’
βˆ’
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£ ,
βˆ’Ξ“(1 βˆ’ 𝛼𝑗 ) β‹… (𝑐+
𝛼𝑗 + 𝑐𝛼𝑗 ) cos
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑑) β‹… sin
2
2
(35.11)
that is, it is expressed by exactly the same formula as for 0 < 𝛼𝑗 < 1 or 1 < 𝛼𝑗 < 2.
So we get:
𝑓 (𝑑) βˆ’ 1
=βˆ’
βˆ‘
𝑗: 0<𝛼𝑗
[
𝛼𝑗 πœ‹ ] 𝛼𝑗
𝛼𝑗 πœ‹
βˆ’
βˆ’
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£
Ξ“(1 βˆ’ 𝛼𝑗 ) β‹… (𝑐+
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑑) β‹… sin
𝛼𝑗 + 𝑐𝛼𝑗 ) cos
2
2
<3
+ 𝑖𝑑1 𝑑 βˆ’ 𝑑2 𝑑2 /2
∫
∞
+
βˆ‘
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2 /2) 𝑑 𝐹 (π‘₯) βˆ’ 1 +
0
∫
βˆ’π›Όπ‘—
𝑐+
𝛼𝑗 π‘₯
]
𝑗: 0<𝛼𝑗 <3
0
+
βˆ‘
[
(𝑒𝑖𝑑π‘₯ βˆ’ 1 βˆ’ 𝑖𝑑π‘₯ + 𝑑2 π‘₯2 /2) 𝑑 𝐹 (π‘₯) βˆ’
]
βˆ’π›Όπ‘—
π‘βˆ’
∣π‘₯∣
.
𝛼𝑗
(35.12)
Let us formulate what we are getting this way (adding and subtracting the same
things, expressing integrals through the Ξ“-function, and from time to time saying β€œif 𝛽
is so and so” and estimating the remainder term) as a lemma (not a theorem, because it
is not a result about the limiting behavior of the distribution of the normalized sum with
improvements).
Lemma 35.1. Let the distribution function satisfy the conditions (34.3), (34.4) with
non-integer 𝛼𝑗 , 𝛽; let conditions (34.9) be satisfied if 𝛽 < 1. Then
βˆ’βˆž
𝑗: 0<𝛼𝑗 <3
𝑓 (𝑑) βˆ’ 1
βˆ‘
[
𝛼𝑗 πœ‹
𝛼𝑗 πœ‹ ] 𝛼𝑗
βˆ’
βˆ’
=βˆ’
Ξ“(1 βˆ’ 𝛼𝑗 ) β‹… (𝑐+
βˆ’ 𝑖 β‹… (𝑐+
βˆ£π‘‘βˆ£
𝛼𝑗 + 𝑐𝛼𝑗 ) cos
𝛼𝑗 βˆ’ 𝑐𝛼𝑗 ) sign(𝑑) β‹… sin
2
2
(35.13)
𝑗
+
βŒŠπ›½βŒ‹
βˆ‘
𝑖𝑠 𝑑𝑠 𝑑𝑠/𝑠! + 𝑂(βˆ£π‘‘βˆ£π›½ )
𝑠=1
3
(the first sum is taken over all 𝑗).
If the conditions (34.1), (34.2) are satisfied, and condition (34.8) for 𝛽 < 1, then the
asymptotic expansion (35.13) holds with 𝑂(βˆ£π‘‘βˆ£π›½ ) replaced with π‘œ(βˆ£π‘‘βˆ£π›½ ) as 𝑑 β†’ 0.
The proof has in fact been carried out for the remainder term 𝑂(βˆ£π‘‘βˆ£π›½ ); the only thing
that remained was to say β€œetc.” The statement with π‘œ(βˆ£π‘‘βˆ£π›½ ) is proved in the same way,
only ±βˆ£π‘‘βˆ£βˆ’1 as( the limits
of integration are replaced with some functions ± 𝑒(𝑑) = π‘œ(βˆ£π‘‘βˆ£βˆ’1 ),
)
such that 𝐡± ± 𝑒(𝑑) (or the corresponding total variations on the interval from 𝑒(𝑑) to ∞
or from βˆ’ ∞ to βˆ’ 𝑒(𝑑)) are π‘œ(βˆ£π‘‘βˆ£βˆ’π›½ ) as 𝑑 β†’ 0.
We know what we are doing after this: we write the asymptotic expansion for ln 𝑓 (𝑑)
as 𝑑 β†’ 0 using the Maclaurin expansion ln 𝑀 = (𝑀 βˆ’ 1) βˆ’ (𝑀 βˆ’ 1)2 /2 + (𝑀 βˆ’ 1)3 /3 βˆ’ ... ;
( 𝑧 )
after this, the expansion in (negative) powers of 𝑛 of 𝑛 ln 𝑓
, and after that, one for
( 𝑧 )𝑛
{
( 𝑧 )}
𝐡𝑛
π‘“πœπ‘› (𝑧) = 𝑓
= exp 𝑛 ln 𝑓
.
𝐡𝑛
𝐡𝑛
4