Motivation

Denition and basic properties

Abel-Tauber Theorems
A (Gentle) Introduction to Regular Variation
Daniel Tokarev
MASCOS
The University of Melbourne, 20 August, 2008
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
References:
1
Feller, An Introduction to Probability Theory and Its
Applications, Vol II.
2
Bingham, Goldie & Teugels, Regular Variation.
3
D. Tokarev, Growth of Integral Transforms and Extinction in
Critical Galton-Watson Processes, JAP, 2008.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Question: Given a real-valued measurable function F and
an operator (eg integral transform, nth iterate, etc), what is
a relationship between the asymptotic behaviour of the
function and the operator?
Example 1: Laplace Transform: Let U(x) be a monotone
function (Probabilists: think "distribution function" (df) or
"measure") possessing a Laplace Transform (LT)
L
U
(t ) :=
_

0
e
xt
U{dx}
(=
_

0
e
xt
u(x)dx, provided u(x) := U

(x) exists)
Obvious Observation 1: As t in the transform,
what will most likely matter is the behaviour of U(x) as
x 0.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Question: Given a real-valued measurable function F and
an operator (eg integral transform, nth iterate, etc), what is
a relationship between the asymptotic behaviour of the
function and the operator?
Example 1: Laplace Transform: Let U(x) be a monotone
function (Probabilists: think "distribution function" (df) or
"measure") possessing a Laplace Transform (LT)
L
U
(t ) :=
_

0
e
xt
U{dx}
(=
_

0
e
xt
u(x)dx, provided u(x) := U

(x) exists)
Obvious Observation 1: As t in the transform,
what will most likely matter is the behaviour of U(x) as
x 0.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Question: Given a real-valued measurable function F and
an operator (eg integral transform, nth iterate, etc), what is
a relationship between the asymptotic behaviour of the
function and the operator?
Example 1: Laplace Transform: Let U(x) be a monotone
function (Probabilists: think "distribution function" (df) or
"measure") possessing a Laplace Transform (LT)
L
U
(t ) :=
_

0
e
xt
U{dx}
(=
_

0
e
xt
u(x)dx, provided u(x) := U

(x) exists)
Obvious Observation 1: As t in the transform,
what will most likely matter is the behaviour of U(x) as
x 0.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Example 2 (Feller II, pp. 277-8)
Let X
(n)
:= max{X
1
, . . . , X
n
}, where X
i
s are non-negative
iid rvs. Do there exist normalising constants a
n
, such that
lim
n
X
(n)
a
n
= X, in distribution, for some non-degenerate
X?
Let F denote the common df of X
i
s, then X
(n)
has df F
n
and the equation above becomes F(a
n
x)
n
G(x), where
G is the df of X.
Taking log of both sides and using log(1 x) x, when
x is close to 0, we obtain
n(1 F(a
n
x)) logG(x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Example 2 (Feller II, pp. 277-8)
Let X
(n)
:= max{X
1
, . . . , X
n
}, where X
i
s are non-negative
iid rvs. Do there exist normalising constants a
n
, such that
lim
n
X
(n)
a
n
= X, in distribution, for some non-degenerate
X?
Let F denote the common df of X
i
s, then X
(n)
has df F
n
and the equation above becomes F(a
n
x)
n
G(x), where
G is the df of X.
Taking log of both sides and using log(1 x) x, when
x is close to 0, we obtain
n(1 F(a
n
x)) logG(x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Example 2 (Feller II, pp. 277-8)
Let X
(n)
:= max{X
1
, . . . , X
n
}, where X
i
s are non-negative
iid rvs. Do there exist normalising constants a
n
, such that
lim
n
X
(n)
a
n
= X, in distribution, for some non-degenerate
X?
Let F denote the common df of X
i
s, then X
(n)
has df F
n
and the equation above becomes F(a
n
x)
n
G(x), where
G is the df of X.
Taking log of both sides and using log(1 x) x, when
x is close to 0, we obtain
n(1 F(a
n
x)) logG(x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Example 2 (Feller II, pp. 277-8)
Let X
(n)
:= max{X
1
, . . . , X
n
}, where X
i
s are non-negative
iid rvs. Do there exist normalising constants a
n
, such that
lim
n
X
(n)
a
n
= X, in distribution, for some non-degenerate
X?
Let F denote the common df of X
i
s, then X
(n)
has df F
n
and the equation above becomes F(a
n
x)
n
G(x), where
G is the df of X.
Taking log of both sides and using log(1 x) x, when
x is close to 0, we obtain
n(1 F(a
n
x)) logG(x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Continued from before....
n(1 F(a
n
x)) logG(x)
On the LHS, we need to "cancel out" n, ie choose a
n
to
grow appropriately, but this choice has to be independent
of x!
For this it is sufcient (and also necessary! - exercise) if for
all x > 0:
lim
t
1 F(tx)
1 F(t )
= h(x).
Since then 1 F(a
n
x) (1 F(a
n
))h(x) and we are at
liberty to choose the appropriate a
n
(ie a
n
= F

(1 1/n)).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Continued from before....
n(1 F(a
n
x)) logG(x)
On the LHS, we need to "cancel out" n, ie choose a
n
to
grow appropriately, but this choice has to be independent
of x!
For this it is sufcient (and also necessary! - exercise) if for
all x > 0:
lim
t
1 F(tx)
1 F(t )
= h(x).
Since then 1 F(a
n
x) (1 F(a
n
))h(x) and we are at
liberty to choose the appropriate a
n
(ie a
n
= F

(1 1/n)).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Continued from before....
n(1 F(a
n
x)) logG(x)
On the LHS, we need to "cancel out" n, ie choose a
n
to
grow appropriately, but this choice has to be independent
of x!
For this it is sufcient (and also necessary! - exercise) if for
all x > 0:
lim
t
1 F(tx)
1 F(t )
= h(x).
Since then 1 F(a
n
x) (1 F(a
n
))h(x) and we are at
liberty to choose the appropriate a
n
(ie a
n
= F

(1 1/n)).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Relationship between function and transform
Extreme Values
Continued from before....
n(1 F(a
n
x)) logG(x)
On the LHS, we need to "cancel out" n, ie choose a
n
to
grow appropriately, but this choice has to be independent
of x!
For this it is sufcient (and also necessary! - exercise) if for
all x > 0:
lim
t
1 F(tx)
1 F(t )
= h(x).
Since then 1 F(a
n
x) (1 F(a
n
))h(x) and we are at
liberty to choose the appropriate a
n
(ie a
n
= F

(1 1/n)).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Karamata, 1930)
Suppose that for a dense set of x > 0,
lim
t
U(tx)
U(t )
= h(x) <
Then h(x) = x

, for some R and we say that U is regularly

varying of index .
Since
U(x
1
x
2
t )
U(t )
=
U(x
1
x
2
t )
U(x
2
t )
U(x
2
t )
U(t )
, taking the limit x and
applying , we obtain h(x
1
x
2
) = h(x
1
)h(x
2
) and hence
h(x) = x

(see Feller I, p.459 for bounded on nite intervals version,

and BGT p. 2-5 for measurable and Baire versions).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Karamata, 1930)
Suppose that for a dense set of x > 0,
lim
t
U(tx)
U(t )
= h(x) <
Then h(x) = x

, for some R and we say that U is regularly

varying of index .
Since
U(x
1
x
2
t )
U(t )
=
U(x
1
x
2
t )
U(x
2
t )
U(x
2
t )
U(t )
, taking the limit x and
applying , we obtain h(x
1
x
2
) = h(x
1
)h(x
2
) and hence
h(x) = x

(see Feller I, p.459 for bounded on nite intervals version,

and BGT p. 2-5 for measurable and Baire versions).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Karamata, 1930)
Suppose that for a dense set of x > 0,
lim
t
U(tx)
U(t )
= h(x) <
Then h(x) = x

, for some R and we say that U is regularly

varying of index .
Since
U(x
1
x
2
t )
U(t )
=
U(x
1
x
2
t )
U(x
2
t )
U(x
2
t )
U(t )
, taking the limit x and
applying , we obtain h(x
1
x
2
) = h(x
1
)h(x
2
) and hence
h(x) = x

(see Feller I, p.459 for bounded on nite intervals version,

and BGT p. 2-5 for measurable and Baire versions).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
We now know G from before: From
n(1 F(a
n
x)) = G(x), x comes out as x

and we obtain:
logG(x) = x

hence G(x) = e
x

.
Theorem (Gnedenko-Fisher)
Let X
(n)
= max{X
1
, . . . , X
n
}, where X
i
are non-negative iid rvs
with df F. Then there exists a sequence {a
n
} such that X
(n)
/a
n
converges in distribution to a non-degenerate rv X with df G iff
1 F is regularly varying, in which case G(x) = e
x

.
The case = 0 , ie U(xt )/U(t ) 1 is called slowly varying
(usually denoted by L(x)).
We have U(x)/x

is slowly varying and conversely x

L(x)
is regularly varying with index , ie U(x) = x

L(x).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
We now know G from before: From
n(1 F(a
n
x)) = G(x), x comes out as x

and we obtain:
logG(x) = x

hence G(x) = e
x

.
Theorem (Gnedenko-Fisher)
Let X
(n)
= max{X
1
, . . . , X
n
}, where X
i
are non-negative iid rvs
with df F. Then there exists a sequence {a
n
} such that X
(n)
/a
n
converges in distribution to a non-degenerate rv X with df G iff
1 F is regularly varying, in which case G(x) = e
x

.
The case = 0 , ie U(xt )/U(t ) 1 is called slowly varying
(usually denoted by L(x)).
We have U(x)/x

is slowly varying and conversely x

L(x)
is regularly varying with index , ie U(x) = x

L(x).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
We now know G from before: From
n(1 F(a
n
x)) = G(x), x comes out as x

and we obtain:
logG(x) = x

hence G(x) = e
x

.
Theorem (Gnedenko-Fisher)
Let X
(n)
= max{X
1
, . . . , X
n
}, where X
i
are non-negative iid rvs
with df F. Then there exists a sequence {a
n
} such that X
(n)
/a
n
converges in distribution to a non-degenerate rv X with df G iff
1 F is regularly varying, in which case G(x) = e
x

.
The case = 0 , ie U(xt )/U(t ) 1 is called slowly varying
(usually denoted by L(x)).
We have U(x)/x

is slowly varying and conversely x

L(x)
is regularly varying with index , ie U(x) = x

L(x).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
We now know G from before: From
n(1 F(a
n
x)) = G(x), x comes out as x

and we obtain:
logG(x) = x

hence G(x) = e
x

.
Theorem (Gnedenko-Fisher)
Let X
(n)
= max{X
1
, . . . , X
n
}, where X
i
are non-negative iid rvs
with df F. Then there exists a sequence {a
n
} such that X
(n)
/a
n
converges in distribution to a non-degenerate rv X with df G iff
1 F is regularly varying, in which case G(x) = e
x

.
The case = 0 , ie U(xt )/U(t ) 1 is called slowly varying
(usually denoted by L(x)).
We have U(x)/x

is slowly varying and conversely x

L(x)
is regularly varying with index , ie U(x) = x

L(x).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Examples of regularly and slowly varying functions:
Any polynomial is regularly varying; any rational function is
regularly varying; log is slowly varying; iterates and powers
of log are slowly varying
slowly varying functions can be unbounded and oscillate
innitely often, eg L(x) = e
(log x)
1/3
(cos x)
1/3
Also if L
1
, L
2
are slowly varying, then so are L
1
L
2
, L
1
+L
2
and if L
2
as x , also L
1
(L
2
)
Obvious Observation 2: Note that most of the
theory can be applied to an asymptote at 0, or any other
point x
0
by transformation x 1/(x
0
x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Examples of regularly and slowly varying functions:
Any polynomial is regularly varying; any rational function is
regularly varying; log is slowly varying; iterates and powers
of log are slowly varying
slowly varying functions can be unbounded and oscillate
innitely often, eg L(x) = e
(log x)
1/3
(cos x)
1/3
Also if L
1
, L
2
are slowly varying, then so are L
1
L
2
, L
1
+L
2
and if L
2
as x , also L
1
(L
2
)
Obvious Observation 2: Note that most of the
theory can be applied to an asymptote at 0, or any other
point x
0
by transformation x 1/(x
0
x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Examples of regularly and slowly varying functions:
Any polynomial is regularly varying; any rational function is
regularly varying; log is slowly varying; iterates and powers
of log are slowly varying
slowly varying functions can be unbounded and oscillate
innitely often, eg L(x) = e
(log x)
1/3
(cos x)
1/3
Also if L
1
, L
2
are slowly varying, then so are L
1
L
2
, L
1
+L
2
and if L
2
as x , also L
1
(L
2
)
Obvious Observation 2: Note that most of the
theory can be applied to an asymptote at 0, or any other
point x
0
by transformation x 1/(x
0
x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Examples of regularly and slowly varying functions:
Any polynomial is regularly varying; any rational function is
regularly varying; log is slowly varying; iterates and powers
of log are slowly varying
slowly varying functions can be unbounded and oscillate
innitely often, eg L(x) = e
(log x)
1/3
(cos x)
1/3
Also if L
1
, L
2
are slowly varying, then so are L
1
L
2
, L
1
+L
2
and if L
2
as x , also L
1
(L
2
)
Obvious Observation 2: Note that most of the
theory can be applied to an asymptote at 0, or any other
point x
0
by transformation x 1/(x
0
x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Examples of regularly and slowly varying functions:
Any polynomial is regularly varying; any rational function is
regularly varying; log is slowly varying; iterates and powers
of log are slowly varying
slowly varying functions can be unbounded and oscillate
innitely often, eg L(x) = e
(log x)
1/3
(cos x)
1/3
Also if L
1
, L
2
are slowly varying, then so are L
1
L
2
, L
1
+L
2
and if L
2
as x , also L
1
(L
2
)
Obvious Observation 2: Note that most of the
theory can be applied to an asymptote at 0, or any other
point x
0
by transformation x 1/(x
0
x)
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Uniform Convergence, Csiszar & Erds, 1967 (BGT
p.9))
If L is slowly varying then L(tx)/L(t ) 1 as x uniformly
on each compact x-set in (0, ).
Let h(x) := logL(e
x
), then the assumption is
h(x +u) h(x) 0, as x u R
Assertion is uniform convergence on u-sets in R.
Suppose not, ie there exist > 0, x
n
and a bounded
sequence u
n
with |h(x
n
+u
n
) h(x
n
)| > 2, for n = 1, 2, . . ..
Passing to a subsequence if necessary, we may assume
that u
n
u
0
and that |u
n
u
0
| 1 for large enough n.
From , for every y and for large enough n,
|h(x
n
+y) h(x
n
)| , hence |h(x
n
+u
n
) h(x
n
+y)| >
for large enough n.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Uniform Convergence, Csiszar & Erds, 1967 (BGT
p.9))
If L is slowly varying then L(tx)/L(t ) 1 as x uniformly
on each compact x-set in (0, ).
Let h(x) := logL(e
x
), then the assumption is
h(x +u) h(x) 0, as x u R
Assertion is uniform convergence on u-sets in R.
Suppose not, ie there exist > 0, x
n
and a bounded
sequence u
n
with |h(x
n
+u
n
) h(x
n
)| > 2, for n = 1, 2, . . ..
Passing to a subsequence if necessary, we may assume
that u
n
u
0
and that |u
n
u
0
| 1 for large enough n.
From , for every y and for large enough n,
|h(x
n
+y) h(x
n
)| , hence |h(x
n
+u
n
) h(x
n
+y)| >
for large enough n.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Uniform Convergence, Csiszar & Erds, 1967 (BGT
p.9))
If L is slowly varying then L(tx)/L(t ) 1 as x uniformly
on each compact x-set in (0, ).
Let h(x) := logL(e
x
), then the assumption is
h(x +u) h(x) 0, as x u R
Assertion is uniform convergence on u-sets in R.
Suppose not, ie there exist > 0, x
n
and a bounded
sequence u
n
with |h(x
n
+u
n
) h(x
n
)| > 2, for n = 1, 2, . . ..
Passing to a subsequence if necessary, we may assume
that u
n
u
0
and that |u
n
u
0
| 1 for large enough n.
From , for every y and for large enough n,
|h(x
n
+y) h(x
n
)| , hence |h(x
n
+u
n
) h(x
n
+y)| >
for large enough n.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Uniform Convergence, Csiszar & Erds, 1967 (BGT
p.9))
If L is slowly varying then L(tx)/L(t ) 1 as x uniformly
on each compact x-set in (0, ).
Let h(x) := logL(e
x
), then the assumption is
h(x +u) h(x) 0, as x u R
Assertion is uniform convergence on u-sets in R.
Suppose not, ie there exist > 0, x
n
and a bounded
sequence u
n
with |h(x
n
+u
n
) h(x
n
)| > 2, for n = 1, 2, . . ..
Passing to a subsequence if necessary, we may assume
that u
n
u
0
and that |u
n
u
0
| 1 for large enough n.
From , for every y and for large enough n,
|h(x
n
+y) h(x
n
)| , hence |h(x
n
+u
n
) h(x
n
+y)| >
for large enough n.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Uniform Convergence, Csiszar & Erds, 1967 (BGT
p.9))
If L is slowly varying then L(tx)/L(t ) 1 as x uniformly
on each compact x-set in (0, ).
Let h(x) := logL(e
x
), then the assumption is
h(x +u) h(x) 0, as x u R
Assertion is uniform convergence on u-sets in R.
Suppose not, ie there exist > 0, x
n
and a bounded
sequence u
n
with |h(x
n
+u
n
) h(x
n
)| > 2, for n = 1, 2, . . ..
Passing to a subsequence if necessary, we may assume
that u
n
u
0
and that |u
n
u
0
| 1 for large enough n.
From , for every y and for large enough n,
|h(x
n
+y) h(x
n
)| , hence |h(x
n
+u
n
) h(x
n
+y)| >
for large enough n.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
So there are sequences x
n
and u
n
u
0
such that
y R and for large enough n, |h(x
n
+u
n
) h(x
n
+y)| > .
We want to nd z s.t. |h(x
n
+u
n
) h(x
n
+u
n
z)| >
Consider

k=1
I
k
= [1, 1], where
I
k
:= [1, 1] {y : n k, |h(x
n
+y
n
) h(x
n
+y)| > }.
Each I
k
is measurable, hence K s.t. I
K
has +ve measure.
Z
n
:= u
n
I
K
= {u
n
y : y I
K
} and let Z :=

j =1

_
n=1
Z
n
.
Since all Z
n
have measure |I
k
| and we took |u
n
u
0
| 1,
they are subsets of [u
0
2, u
0
+2] and
|Z| = lim
j
|

_
n=j
Z
n
| |I
K
| > 0.
Now take z Z and we have a contradiction as in !
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
So there are sequences x
n
and u
n
u
0
such that
y R and for large enough n, |h(x
n
+u
n
) h(x
n
+y)| > .
We want to nd z s.t. |h(x
n
+u
n
) h(x
n
+u
n
z)| >
Consider

k=1
I
k
= [1, 1], where
I
k
:= [1, 1] {y : n k, |h(x
n
+y
n
) h(x
n
+y)| > }.
Each I
k
is measurable, hence K s.t. I
K
has +ve measure.
Z
n
:= u
n
I
K
= {u
n
y : y I
K
} and let Z :=

j =1

_
n=1
Z
n
.
Since all Z
n
have measure |I
k
| and we took |u
n
u
0
| 1,
they are subsets of [u
0
2, u
0
+2] and
|Z| = lim
j
|

_
n=j
Z
n
| |I
K
| > 0.
Now take z Z and we have a contradiction as in !
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
So there are sequences x
n
and u
n
u
0
such that
y R and for large enough n, |h(x
n
+u
n
) h(x
n
+y)| > .
We want to nd z s.t. |h(x
n
+u
n
) h(x
n
+u
n
z)| >
Consider

k=1
I
k
= [1, 1], where
I
k
:= [1, 1] {y : n k, |h(x
n
+y
n
) h(x
n
+y)| > }.
Each I
k
is measurable, hence K s.t. I
K
has +ve measure.
Z
n
:= u
n
I
K
= {u
n
y : y I
K
} and let Z :=

j =1

_
n=1
Z
n
.
Since all Z
n
have measure |I
k
| and we took |u
n
u
0
| 1,
they are subsets of [u
0
2, u
0
+2] and
|Z| = lim
j
|

_
n=j
Z
n
| |I
K
| > 0.
Now take z Z and we have a contradiction as in !
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
So there are sequences x
n
and u
n
u
0
such that
y R and for large enough n, |h(x
n
+u
n
) h(x
n
+y)| > .
We want to nd z s.t. |h(x
n
+u
n
) h(x
n
+u
n
z)| >
Consider

k=1
I
k
= [1, 1], where
I
k
:= [1, 1] {y : n k, |h(x
n
+y
n
) h(x
n
+y)| > }.
Each I
k
is measurable, hence K s.t. I
K
has +ve measure.
Z
n
:= u
n
I
K
= {u
n
y : y I
K
} and let Z :=

j =1

_
n=1
Z
n
.
Since all Z
n
have measure |I
k
| and we took |u
n
u
0
| 1,
they are subsets of [u
0
2, u
0
+2] and
|Z| = lim
j
|

_
n=j
Z
n
| |I
K
| > 0.
Now take z Z and we have a contradiction as in !
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
So there are sequences x
n
and u
n
u
0
such that
y R and for large enough n, |h(x
n
+u
n
) h(x
n
+y)| > .
We want to nd z s.t. |h(x
n
+u
n
) h(x
n
+u
n
z)| >
Consider

k=1
I
k
= [1, 1], where
I
k
:= [1, 1] {y : n k, |h(x
n
+y
n
) h(x
n
+y)| > }.
Each I
k
is measurable, hence K s.t. I
K
has +ve measure.
Z
n
:= u
n
I
K
= {u
n
y : y I
K
} and let Z :=

j =1

_
n=1
Z
n
.
Since all Z
n
have measure |I
k
| and we took |u
n
u
0
| 1,
they are subsets of [u
0
2, u
0
+2] and
|Z| = lim
j
|

_
n=j
Z
n
| |I
K
| > 0.
Now take z Z and we have a contradiction as in !
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Local Boundedness, Seneta,1973 (BGT, p. 12))
If L is positive and slowly varying, then for large enough X, it is
bounded on nite intervals [X, Y], 0 < X < Y < .
Let h(x) = log(L(e
x
)) as before. From Uniform
Convergence, we know that there exists X such that for
u [0, 1] and x X: |h(u +x) h(x)| < 1
Hence |h(x)| |h(X)| +1 on [X, X +1] and by induction
|h(x)| |h(X)| +n on [X, X +n]. The conclusion for L
follows.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Local Boundedness, Seneta,1973 (BGT, p. 12))
If L is positive and slowly varying, then for large enough X, it is
bounded on nite intervals [X, Y], 0 < X < Y < .
Let h(x) = log(L(e
x
)) as before. From Uniform
Convergence, we know that there exists X such that for
u [0, 1] and x X: |h(u +x) h(x)| < 1
Hence |h(x)| |h(X)| +1 on [X, X +1] and by induction
|h(x)| |h(X)| +n on [X, X +n]. The conclusion for L
follows.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Local Boundedness, Seneta,1973 (BGT, p. 12))
If L is positive and slowly varying, then for large enough X, it is
bounded on nite intervals [X, Y], 0 < X < Y < .
Let h(x) = log(L(e
x
)) as before. From Uniform
Convergence, we know that there exists X such that for
u [0, 1] and x X: |h(u +x) h(x)| < 1
Hence |h(x)| |h(X)| +1 on [X, X +1] and by induction
|h(x)| |h(X)| +n on [X, X +n]. The conclusion for L
follows.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Representation Theorem (Karamata, Korevaar))
L is slowly varying iff
L(x) = c(x) exp(
_
x
a
(u)
u
du), or h(x) = d(x) +
_
x
b
e(u) du,
for some c(x) c; d(x) d; (x), e(x) 0; a, b < x.
Since h is integrable on nite intervals far enough to the
right, we can write
h(x) =
_
x+1
x
h(x)h(u) du+
_
x
X
h(u+1)h(u) du+
_
X+1
X
h(u) du
Observe that the rst term on the right is just
_
1
0
h(x +1) h(u) du which must tend to 0 by Uniform
Convergence and the last term is constant.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Representation Theorem (Karamata, Korevaar))
L is slowly varying iff
L(x) = c(x) exp(
_
x
a
(u)
u
du), or h(x) = d(x) +
_
x
b
e(u) du,
for some c(x) c; d(x) d; (x), e(x) 0; a, b < x.
Since h is integrable on nite intervals far enough to the
right, we can write
h(x) =
_
x+1
x
h(x)h(u) du+
_
x
X
h(u+1)h(u) du+
_
X+1
X
h(u) du
Observe that the rst term on the right is just
_
1
0
h(x +1) h(u) du which must tend to 0 by Uniform
Convergence and the last term is constant.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Denition
Examples
Uniform convergence
Representation Theorem
Theorem (Representation Theorem (Karamata, Korevaar))
L is slowly varying iff
L(x) = c(x) exp(
_
x
a
(u)
u
du), or h(x) = d(x) +
_
x
b
e(u) du,
for some c(x) c; d(x) d; (x), e(x) 0; a, b < x.
Since h is integrable on nite intervals far enough to the
right, we can write
h(x) =
_
x+1
x
h(x)h(u) du+
_
x
X
h(u+1)h(u) du+
_
X+1
X
h(u) du
Observe that the rst term on the right is just
_
1
0
h(x +1) h(u) du which must tend to 0 by Uniform
Convergence and the last term is constant.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
If U() has LT L
U
(t ) then U(x) has LT L
U
(t /x)
U(x
1
)
U(x
1
)
U(x
2
)
U(x
2
)

U(x
i
)
U(x
i
)

, for x
i


L
U
(t /x
1
)
U(x
1
)
L
U
(t /x
2
)
U(x
1
)

L
U
(t /x
1
)
U(x
i
)
t

(1 + )
Need to use Extended Continuity of Laplace Transforms
(provided the transforms remain bounded - Exercise)
This means
L
U
(t /x)
U(x)
t

(1 + ) which for t = 1 gives the

relationship we alluded to at the start and in fact we obtain:
Theorem ((Karamata, Feller II, pp. 442-3))
Let U(x) be a measure with LT L
U
(t ). U is regularly varying at
with index iff L
U
is regularly varying at 0 with index and
each of these is equivalent to
L
U
(1/t )/U(t ) (1 + ), as t .
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
If U() has LT L
U
(t ) then U(x) has LT L
U
(t /x)
U(x
1
)
U(x
1
)
U(x
2
)
U(x
2
)

U(x
i
)
U(x
i
)

, for x
i


L
U
(t /x
1
)
U(x
1
)
L
U
(t /x
2
)
U(x
1
)

L
U
(t /x
1
)
U(x
i
)
t

(1 + )
Need to use Extended Continuity of Laplace Transforms
(provided the transforms remain bounded - Exercise)
This means
L
U
(t /x)
U(x)
t

(1 + ) which for t = 1 gives the

relationship we alluded to at the start and in fact we obtain:
Theorem ((Karamata, Feller II, pp. 442-3))
Let U(x) be a measure with LT L
U
(t ). U is regularly varying at
with index iff L
U
is regularly varying at 0 with index and
each of these is equivalent to
L
U
(1/t )/U(t ) (1 + ), as t .
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
If U() has LT L
U
(t ) then U(x) has LT L
U
(t /x)
U(x
1
)
U(x
1
)
U(x
2
)
U(x
2
)

U(x
i
)
U(x
i
)

, for x
i


L
U
(t /x
1
)
U(x
1
)
L
U
(t /x
2
)
U(x
1
)

L
U
(t /x
1
)
U(x
i
)
t

(1 + )
Need to use Extended Continuity of Laplace Transforms
(provided the transforms remain bounded - Exercise)
This means
L
U
(t /x)
U(x)
t

(1 + ) which for t = 1 gives the

relationship we alluded to at the start and in fact we obtain:
Theorem ((Karamata, Feller II, pp. 442-3))
Let U(x) be a measure with LT L
U
(t ). U is regularly varying at
with index iff L
U
is regularly varying at 0 with index and
each of these is equivalent to
L
U
(1/t )/U(t ) (1 + ), as t .
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
If U() has LT L
U
(t ) then U(x) has LT L
U
(t /x)
U(x
1
)
U(x
1
)
U(x
2
)
U(x
2
)

U(x
i
)
U(x
i
)

, for x
i


L
U
(t /x
1
)
U(x
1
)
L
U
(t /x
2
)
U(x
1
)

L
U
(t /x
1
)
U(x
i
)
t

(1 + )
Need to use Extended Continuity of Laplace Transforms
(provided the transforms remain bounded - Exercise)
This means
L
U
(t /x)
U(x)
t

(1 + ) which for t = 1 gives the

relationship we alluded to at the start and in fact we obtain:
Theorem ((Karamata, Feller II, pp. 442-3))
Let U(x) be a measure with LT L
U
(t ). U is regularly varying at
with index iff L
U
is regularly varying at 0 with index and
each of these is equivalent to
L
U
(1/t )/U(t ) (1 + ), as t .
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
If U() has LT L
U
(t ) then U(x) has LT L
U
(t /x)
U(x
1
)
U(x
1
)
U(x
2
)
U(x
2
)

U(x
i
)
U(x
i
)

, for x
i


L
U
(t /x
1
)
U(x
1
)
L
U
(t /x
2
)
U(x
1
)

L
U
(t /x
1
)
U(x
i
)
t

(1 + )
Need to use Extended Continuity of Laplace Transforms
(provided the transforms remain bounded - Exercise)
This means
L
U
(t /x)
U(x)
t

(1 + ) which for t = 1 gives the

relationship we alluded to at the start and in fact we obtain:
Theorem ((Karamata, Feller II, pp. 442-3))
Let U(x) be a measure with LT L
U
(t ). U is regularly varying at
with index iff L
U
is regularly varying at 0 with index and
each of these is equivalent to
L
U
(1/t )/U(t ) (1 + ), as t .
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Theorem ((Tokarev, JAP 2008))
Let g, h : [0, 1) R
+
, L a slowly varying function at 0, and let
(0, 1) ((1) remains valid for = 0). Dene
G(k) :=
_
1
0
s
k
g(s) ds H(k) :=
_
1
0
(1 s
k
)h(s) ds
Then, as x 1 and k
1
g(x) (1 x)

L(1 x) G(k) (1 )k
1
L
_
1
k
_
2
h(x) (1 x)
1
L(1 x) H(k)
(1)

L
_
1
k
_
Sketch of the proof of (1), (2) is analogous. Let
s = 1
x
k
, then G(k) becomes
k
1
_
k
0
_
1
x
k
_
k
x

L(x/k) ds.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Theorem ((Tokarev, JAP 2008))
Let g, h : [0, 1) R
+
, L a slowly varying function at 0, and let
(0, 1) ((1) remains valid for = 0). Dene
G(k) :=
_
1
0
s
k
g(s) ds H(k) :=
_
1
0
(1 s
k
)h(s) ds
Then, as x 1 and k
1
g(x) (1 x)

L(1 x) G(k) (1 )k
1
L
_
1
k
_
2
h(x) (1 x)
1
L(1 x) H(k)
(1)

L
_
1
k
_
Sketch of the proof of (1), (2) is analogous. Let
s = 1
x
k
, then G(k) becomes
k
1
_
k
0
_
1
x
k
_
k
x

L(x/k) ds.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Theorem ((Tokarev, JAP 2008))
Let g, h : [0, 1) R
+
, L a slowly varying function at 0, and let
(0, 1) ((1) remains valid for = 0). Dene
G(k) :=
_
1
0
s
k
g(s) ds H(k) :=
_
1
0
(1 s
k
)h(s) ds
Then, as x 1 and k
1
g(x) (1 x)

L(1 x) G(k) (1 )k
1
L
_
1
k
_
2
h(x) (1 x)
1
L(1 x) H(k)
(1)

L
_
1
k
_
Sketch of the proof of (1), (2) is analogous. Let
s = 1
x
k
, then G(k) becomes
k
1
_
k
0
_
1
x
k
_
k
x

L(x/k) ds.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Let (0, 1) and N > 0 be a "large" number, divide G(k)
by k
1
L(1/k) and split the integral into two parts:
_
N
0
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx +
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
=
_
N
0
_
1
x
k
_
k
x

_
L(x/k)
L(1/k)
1
_
dx +
_
N
0
_
1
x
k
_
k
x

dx
+
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
We can pick N so that by Uniform Convergence the rst
term on the right is < /3, the second is within /3 of
(1 ) and using Potters Thm (BGT p. 25), the last term
is < /3.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Let (0, 1) and N > 0 be a "large" number, divide G(k)
by k
1
L(1/k) and split the integral into two parts:
_
N
0
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx +
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
=
_
N
0
_
1
x
k
_
k
x

_
L(x/k)
L(1/k)
1
_
dx +
_
N
0
_
1
x
k
_
k
x

dx
+
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
We can pick N so that by Uniform Convergence the rst
term on the right is < /3, the second is within /3 of
(1 ) and using Potters Thm (BGT p. 25), the last term
is < /3.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
Let (0, 1) and N > 0 be a "large" number, divide G(k)
by k
1
L(1/k) and split the integral into two parts:
_
N
0
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx +
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
=
_
N
0
_
1
x
k
_
k
x

_
L(x/k)
L(1/k)
1
_
dx +
_
N
0
_
1
x
k
_
k
x

dx
+
_
k
N
_
1
x
k
_
k
x

L(x/k)
L(1/k)
dx
We can pick N so that by Uniform Convergence the rst
term on the right is < /3, the second is within /3 of
(1 ) and using Potters Thm (BGT p. 25), the last term
is < /3.
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
The above result can be used to easily obtain a Tauberian
theorem for iterates.
Theorem (adapted from Tokarev, JAP 2008)
Let f be a monotone function with a monotone derivative and
f (1) = f

(1) = 1 and let f

n
denote the nth functional iterate of f .
Then f (s) s = (1 s)
1+
L(1 s) iff 1 f
n
= n
1/
L

(n),
where L is slowly varying at 0 and L

is slowly varying at .
The proof connects f and f
n
via the quantity

i =0
1 f
i
(0)
n
(in a special case this is the Expected Time to Extinction of
a Galton-Watson process started with k ancestors and a
probability generating function f ).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
The above result can be used to easily obtain a Tauberian
theorem for iterates.
Theorem (adapted from Tokarev, JAP 2008)
Let f be a monotone function with a monotone derivative and
f (1) = f

(1) = 1 and let f

n
denote the nth functional iterate of f .
Then f (s) s = (1 s)
1+
L(1 s) iff 1 f
n
= n
1/
L

(n),
where L is slowly varying at 0 and L

is slowly varying at .
The proof connects f and f
n
via the quantity

i =0
1 f
i
(0)
n
(in a special case this is the Expected Time to Extinction of
a Galton-Watson process started with k ancestors and a
probability generating function f ).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates
The above result can be used to easily obtain a Tauberian
theorem for iterates.
Theorem (adapted from Tokarev, JAP 2008)
Let f be a monotone function with a monotone derivative and
f (1) = f

(1) = 1 and let f

n
denote the nth functional iterate of f .
Then f (s) s = (1 s)
1+
L(1 s) iff 1 f
n
= n
1/
L

(n),
where L is slowly varying at 0 and L

is slowly varying at .
The proof connects f and f
n
via the quantity

i =0
1 f
i
(0)
n
(in a special case this is the Expected Time to Extinction of
a Galton-Watson process started with k ancestors and a
probability generating function f ).
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates

i =0
1 f
i
(0)
n
can be shown to be asymptotically
equivalent to
_

0
1U(s)
k
ds =
_
1
0
U

(s
1/k
) ds = k
_

0
s
k1
U

(s) ds,
where U(s) = f
s
(0). Regular variation of f
n
is equivalent
to regular variation of U and this can be inverted to obtain
regular variation of U

. The integral is then of the form of

G of the above.

i =0
1 f
i
(0)
n
can also be shown to be asymptotically
equivalent to
_
1
0
1 s
k
f (s) s
ds. This is of the form of H.
Obvious Observation 3:And thats probably enough
for today!
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates

i =0
1 f
i
(0)
n
can be shown to be asymptotically
equivalent to
_

0
1U(s)
k
ds =
_
1
0
U

(s
1/k
) ds = k
_

0
s
k1
U

(s) ds,
where U(s) = f
s
(0). Regular variation of f
n
is equivalent
to regular variation of U and this can be inverted to obtain
regular variation of U

. The integral is then of the form of

G of the above.

i =0
1 f
i
(0)
n
can also be shown to be asymptotically
equivalent to
_
1
0
1 s
k
f (s) s
ds. This is of the form of H.
Obvious Observation 3:And thats probably enough
for today!
Daniel Tokarev A (Gentle) Introduction to Regular Variation
Motivation
Denition and basic properties
Abel-Tauber Theorems
Mellin-type transforms
Tauberian theorem for iterates

i =0
1 f
i
(0)
n
can be shown to be asymptotically
equivalent to
_

0
1U(s)
k
ds =
_
1
0
U

(s
1/k
) ds = k
_

0
s
k1
U

(s) ds,
where U(s) = f
s
(0). Regular variation of f
n
is equivalent
to regular variation of U and this can be inverted to obtain
regular variation of U

. The integral is then of the form of

G of the above.

i =0
1 f
i
(0)
n
can also be shown to be asymptotically
equivalent to
_
1
0
1 s
k
f (s) s
ds. This is of the form of H.
Obvious Observation 3:And thats probably enough
for today!
Daniel Tokarev A (Gentle) Introduction to Regular Variation