Download Zeros of random analytic functions and extreme value theory

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
Transcript
Zeros of random analytic functions and
extreme value theory
Zakhar Kabluchko
University of Ulm
November 9, 2012
A random equation
Statement of the problem
We consider an algebraic equation with random coefficients, for
example
z 2000 − z 1999 + z 1998 + z 1997 − z 1996 − . . . + z 3 + z 2 − z + 1 = 0
What is the distribution of the solutions of this equation in the
complex plane?
2
3
Solutions of the equation
Zeros of a random polynomial of degree n = 2000
Random Polynomials (Kac Ensemble)
Statement of the problem
Let ξ0 , ξ1 , . . . be i.i.d. random variables.
Consider the equation
Pn (z) := ξ0 + ξ1 z + ξ2 z 2 + . . . + ξn z n = 0.
The equation has n complex roots z1 , . . . , zn .
What is the distribution of roots in the complex plane?
Consider the empirical measure
n
1X
δ(zk ).
µn =
n k=1
Problem: Find limn→∞ µn .
Distribution of roots
Theorem (Ibragimov, Zaporozhets, 2011)
The following conditions are equivalent:
1
With probability 1, the sequence µn converges weakly to
the uniform distribution on the unit circle.
2
E log(1 + |ξ0 |) < ∞.
Remark
P∞
k
The series
converges a.s. in the unit circle iff
k=0 ξk z
E log(1 + |ξ0 |) < ∞.
History
Hammersley (1954), Shparo und Shur (1962), Arnold (1966),
Shepp und Vanderbei (1995), ...
5
Logarithmic tails
Problem
What happens if the coefficients ξk are heavy-tailed?
Logarithmic tails
We consider coefficients satisfying
P[|ξ0 | > t] ∼
L(log t)
, t → +∞,
(log t)α
where α > 0 and L is a slowly varying function.
Remark
(
infinite, α < 1,
E log(1 + |ξ0 |) is
finite,
α > 1.
7
Example: Coefficients with logarithmic tails
The coefficients ξk are such that P[|ξk | > t] = 1/(log t)2 and
the degree is n = 2000.
Distribution of roots
Assume for simplicity that
P[|ξ0 | > t] ∼
1
(log t)α
as t → +∞.
(1)
Theorem (Kabluchko, Zaporozhets, 2011)
For coefficients with logarithmic tails, the following weak convergence of random probability measures holds:
n
1
1X
w
hn1− α i
δ(zk
) −→ Πα .
n→∞
n k=1
The limiting random probability measure Πα is a.s. a convex
combination of at most countably many uniform measures concentrated on circles centered at the origin.
9
Light and logarithmically tailed coefficients
Distribution of roots
Example: α = 1
If P[|ξ0 | > t] ∼
1
log t
as t → +∞, then we have
n
1X
w
δ(zk ) −→ Π1 .
n→∞
n k=1
No normalization of roots is needed.
Example: α > 1 (E log |ξ0 | < ∞)
The roots approach the unit circle. The distance between the
1
roots and the unit circle is of order n α −1 .
Example: α < 1 (E log |ξ0 | = ∞)
The roots diverge to ∞ and 0. The absolute values of the roots
1 −1
are of order O(1)n α .
Extremal order statistics of the coefficients
Theorem
Let ξ0 , ξ1 , . . . be i.i.d. random variables with
P[log |ξk | > t] ∼ t −α as t → +∞.
Then, we have the following weak convergence of point processes on [0, 1] × (0, ∞):
n
X
k log |ξk |
w
δ
, 1/α
−→ PPP αv −(α+1) dudv .
n→∞
n n
k=0
11
12
Extremal order statistics of the coefficients
Newton polygon
Idea of the proof
For large n consider the equation
±e nx1 ± . . . ± e nxd = 0,
where xi > 0. The most easy way for this to be true is the
following:
1
Two terms, say e nxk and e nxl , cancel each other.
2
All other terms are much smaller than these two.
Newton polygon
Idea of the proof
Apply this to the equation ξ0 + ξ1 z + . . . + ξn z n = 0.
1
Two terms cancel each other: ξk z k + ξl z l = 0.
2
All other terms are much smaller than these two.
Geometrically: the points (k, log |ξk |) and (l, log |ξl |) are neighboring vertices of the least concave majorant of the set
{(0, log |ξ0 |), . . . , (n, log |ξn |)}.
14
The least concave majorant
Remark
The number of segments is finite a.s. if and only if α < 1.
15
Limiting empirical measure of the roots
Limiting measure Πα
Consider the least concave majorant of the Poisson process with
intensity αv −(α+1) dudv .
1
Radii of circles = exponentials of the slopes of the majorant.
2
Number of roots on a circle = length of the linearity interval.
16
Very heavy tails: α = 0
Let the coefficients be such that
P[|ξ0 | > t] ∼ L(log t) as t → +∞,
where L is a slowly varying function.
The roots concentrate on two circles, one with a small radius,
one with a large radius. The proportion of the roots lying on
the small circle is uniform on [0, 1].
17
Weyl Polynomials
Let ξ0 , ξ1 , . . . be i.i.d. random variables.
Consider the Weyl Polynomials
Pn (z) =
n
X
zk
ξk √ .
k!
k=0
Let z1 , . . . , zn be the zeros of Pn .
Theorem (Kabluchko, Zaporozhets, 2012)
The following conditions are equivalent:
P
1
The sequence of probability measures n1 k=1 δ( √zkn ) converges a.s. to the uniform distribution on the unit disk
{|z| ≤ 1}.
2
E log(1 + |ξ0 |) < ∞.
18
19
Weyl Polynomials
Zeros of a Weyl polynomial: Normally distributed coefficients
20
Weyl Polynomials
Zeros of a Weyl polynomial: Logarithmic tails
Littlewood–Offord Polynomials (1939)
Let ξ0 , ξ1 , . . . be i.i.d. random variables with E log(1 +
|ξ0 |) < ∞.
Consider the Littlewood–Offord polynomials
Pn (z) =
n
X
ξk
k=0
zk
.
(k!)α
Let z1 , . . . , zn be the zeros of Pn .
Theorem (Kabluchko, Zaporozhets, 2012)
With
1, the sequence of random measures
Pn probability
zk
1
δ(
)
converges
to the probability measure with the
k=1
n
nα
density
1
1
|z| α −2 , |z| < 1.
2πα
21
Littlewood–Offord Polynomials
Zeros of the Littlewood–Offord polynomials: Normal
coefficients
Littlewood–Offord Polynomials
Zeros of the Littlewood–Offord polynomials: Logarithmic
coefficients
Szegö Polynomials
Szegö Polynomials: sn (z) =
zk
k=0 k! .
Pn
Theorem (Szegö, 1924)
The zeros of sn (nz) cluster along the curve |ze 1−z | = 1, |z| < 1.
Zeros in the Random Energy Model
Random Energy Model (Derrida, 1981)
A system has N states.
√
The energy of the system in state i is log Nξi .
ξ1 , . . . , ξN are i.i.d. standard Gaussian random variables.
Consider the partition function
ZN (β) =
N
X
eβ
√
log Nξk
, β ∈ C.
k=1
Other motivations
ZN is an empirical Laplace transform.
ZN is a nice random analytic function.
Zeros in the Random Energy Model
Complex zeros of ZN . Source: C. Moukarzel und N. Parga:
Physica A 177 (1991).
26
Zeros in the Random Energy Model
Theorem (Derrida, 1991)
For β = σ + iτ ∈ C it holds that

1 + 1 (σ 2 − τ 2 ), β ∈ B 1 ,
log |ZN (β)| √ 2
lim
=
2|σ|,
β ∈ B 2,
N→∞

log N
1
+ σ2,
β ∈ B 3,
2
where
B1 = C\B2 ∪ B3 ,
B2 = {β ∈ C : 2σ 2 > 1, |σ| + |τ | >
2
2
2
√
2},
B3 = {β ∈ C : 2σ < 1, σ + τ > 1}.
27
Zeros in the Random Energy Model
Theorem (Derrida, 1991)
The random measure
2π
log N
X
δ(β)
β:ZN (β)=0
converges weakly (as N → ∞) to the deterministic measure
Ξ = 2Ξ3 + Ξ12 + Ξ13 . Here,
Ξ3 is the Lebesgue measure on B3 .
Ξ13 is the one-dimensional Lebesgue measure on the
boundary between B1 and B3 .
√
Ξ12 is the measure with density 2|τ | on the boundary
between B1 and B2 .
Rigorous proof, further results: Kabluchko und Klimovsky, 2012.